net: convert BUG_TRAP to generic WARN_ON
[linux-3.10.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:     Ross Biro
9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *              Florian La Roche, <flla@stud.uni-sb.de>
13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *              Jorge Cwik, <jorge@laser.satlink.net>
19  */
20
21 /*
22  * Changes:
23  *              Pedro Roque     :       Fast Retransmit/Recovery.
24  *                                      Two receive queues.
25  *                                      Retransmit queue handled by TCP.
26  *                                      Better retransmit timer handling.
27  *                                      New congestion avoidance.
28  *                                      Header prediction.
29  *                                      Variable renaming.
30  *
31  *              Eric            :       Fast Retransmit.
32  *              Randy Scott     :       MSS option defines.
33  *              Eric Schenk     :       Fixes to slow start algorithm.
34  *              Eric Schenk     :       Yet another double ACK bug.
35  *              Eric Schenk     :       Delayed ACK bug fixes.
36  *              Eric Schenk     :       Floyd style fast retrans war avoidance.
37  *              David S. Miller :       Don't allow zero congestion window.
38  *              Eric Schenk     :       Fix retransmitter so that it sends
39  *                                      next packet on ack of previous packet.
40  *              Andi Kleen      :       Moved open_request checking here
41  *                                      and process RSTs for open_requests.
42  *              Andi Kleen      :       Better prune_queue, and other fixes.
43  *              Andrey Savochkin:       Fix RTT measurements in the presence of
44  *                                      timestamps.
45  *              Andrey Savochkin:       Check sequence numbers correctly when
46  *                                      removing SACKs due to in sequence incoming
47  *                                      data segments.
48  *              Andi Kleen:             Make sure we never ack data there is not
49  *                                      enough room for. Also make this condition
50  *                                      a fatal error if it might still happen.
51  *              Andi Kleen:             Add tcp_measure_rcv_mss to make
52  *                                      connections with MSS<min(MTU,ann. MSS)
53  *                                      work without delayed acks.
54  *              Andi Kleen:             Process packets with PSH set in the
55  *                                      fast path.
56  *              J Hadi Salim:           ECN support
57  *              Andrei Gurtov,
58  *              Pasi Sarolahti,
59  *              Panu Kuhlberg:          Experimental audit of TCP (re)transmission
60  *                                      engine. Lots of bugs are found.
61  *              Pasi Sarolahti:         F-RTO for dealing with spurious RTOs
62  */
63
64 #include <linux/mm.h>
65 #include <linux/module.h>
66 #include <linux/sysctl.h>
67 #include <net/dst.h>
68 #include <net/tcp.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71 #include <asm/unaligned.h>
72 #include <net/netdma.h>
73
74 int sysctl_tcp_timestamps __read_mostly = 1;
75 int sysctl_tcp_window_scaling __read_mostly = 1;
76 int sysctl_tcp_sack __read_mostly = 1;
77 int sysctl_tcp_fack __read_mostly = 1;
78 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
79 int sysctl_tcp_ecn __read_mostly;
80 int sysctl_tcp_dsack __read_mostly = 1;
81 int sysctl_tcp_app_win __read_mostly = 31;
82 int sysctl_tcp_adv_win_scale __read_mostly = 2;
83
84 int sysctl_tcp_stdurg __read_mostly;
85 int sysctl_tcp_rfc1337 __read_mostly;
86 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
87 int sysctl_tcp_frto __read_mostly = 2;
88 int sysctl_tcp_frto_response __read_mostly;
89 int sysctl_tcp_nometrics_save __read_mostly;
90
91 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
92 int sysctl_tcp_abc __read_mostly;
93
94 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
95 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
96 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
98 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
99 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
100 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
101 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
102 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
104 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
105 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
106 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
107 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
108
109 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
114
115 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
117
118 /* Adapt the MSS value used to make delayed ack decision to the
119  * real world.
120  */
121 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
122 {
123         struct inet_connection_sock *icsk = inet_csk(sk);
124         const unsigned int lss = icsk->icsk_ack.last_seg_size;
125         unsigned int len;
126
127         icsk->icsk_ack.last_seg_size = 0;
128
129         /* skb->len may jitter because of SACKs, even if peer
130          * sends good full-sized frames.
131          */
132         len = skb_shinfo(skb)->gso_size ? : skb->len;
133         if (len >= icsk->icsk_ack.rcv_mss) {
134                 icsk->icsk_ack.rcv_mss = len;
135         } else {
136                 /* Otherwise, we make more careful check taking into account,
137                  * that SACKs block is variable.
138                  *
139                  * "len" is invariant segment length, including TCP header.
140                  */
141                 len += skb->data - skb_transport_header(skb);
142                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
143                     /* If PSH is not set, packet should be
144                      * full sized, provided peer TCP is not badly broken.
145                      * This observation (if it is correct 8)) allows
146                      * to handle super-low mtu links fairly.
147                      */
148                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
149                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
150                         /* Subtract also invariant (if peer is RFC compliant),
151                          * tcp header plus fixed timestamp option length.
152                          * Resulting "len" is MSS free of SACK jitter.
153                          */
154                         len -= tcp_sk(sk)->tcp_header_len;
155                         icsk->icsk_ack.last_seg_size = len;
156                         if (len == lss) {
157                                 icsk->icsk_ack.rcv_mss = len;
158                                 return;
159                         }
160                 }
161                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
162                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
163                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
164         }
165 }
166
167 static void tcp_incr_quickack(struct sock *sk)
168 {
169         struct inet_connection_sock *icsk = inet_csk(sk);
170         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
171
172         if (quickacks == 0)
173                 quickacks = 2;
174         if (quickacks > icsk->icsk_ack.quick)
175                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
176 }
177
178 void tcp_enter_quickack_mode(struct sock *sk)
179 {
180         struct inet_connection_sock *icsk = inet_csk(sk);
181         tcp_incr_quickack(sk);
182         icsk->icsk_ack.pingpong = 0;
183         icsk->icsk_ack.ato = TCP_ATO_MIN;
184 }
185
186 /* Send ACKs quickly, if "quick" count is not exhausted
187  * and the session is not interactive.
188  */
189
190 static inline int tcp_in_quickack_mode(const struct sock *sk)
191 {
192         const struct inet_connection_sock *icsk = inet_csk(sk);
193         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
194 }
195
196 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
197 {
198         if (tp->ecn_flags & TCP_ECN_OK)
199                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
200 }
201
202 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
203 {
204         if (tcp_hdr(skb)->cwr)
205                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
206 }
207
208 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
209 {
210         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
211 }
212
213 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
214 {
215         if (tp->ecn_flags & TCP_ECN_OK) {
216                 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
217                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
218                 /* Funny extension: if ECT is not set on a segment,
219                  * it is surely retransmit. It is not in ECN RFC,
220                  * but Linux follows this rule. */
221                 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
222                         tcp_enter_quickack_mode((struct sock *)tp);
223         }
224 }
225
226 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
227 {
228         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
229                 tp->ecn_flags &= ~TCP_ECN_OK;
230 }
231
232 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
233 {
234         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
235                 tp->ecn_flags &= ~TCP_ECN_OK;
236 }
237
238 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
239 {
240         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
241                 return 1;
242         return 0;
243 }
244
245 /* Buffer size and advertised window tuning.
246  *
247  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
248  */
249
250 static void tcp_fixup_sndbuf(struct sock *sk)
251 {
252         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
253                      sizeof(struct sk_buff);
254
255         if (sk->sk_sndbuf < 3 * sndmem)
256                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
257 }
258
259 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
260  *
261  * All tcp_full_space() is split to two parts: "network" buffer, allocated
262  * forward and advertised in receiver window (tp->rcv_wnd) and
263  * "application buffer", required to isolate scheduling/application
264  * latencies from network.
265  * window_clamp is maximal advertised window. It can be less than
266  * tcp_full_space(), in this case tcp_full_space() - window_clamp
267  * is reserved for "application" buffer. The less window_clamp is
268  * the smoother our behaviour from viewpoint of network, but the lower
269  * throughput and the higher sensitivity of the connection to losses. 8)
270  *
271  * rcv_ssthresh is more strict window_clamp used at "slow start"
272  * phase to predict further behaviour of this connection.
273  * It is used for two goals:
274  * - to enforce header prediction at sender, even when application
275  *   requires some significant "application buffer". It is check #1.
276  * - to prevent pruning of receive queue because of misprediction
277  *   of receiver window. Check #2.
278  *
279  * The scheme does not work when sender sends good segments opening
280  * window and then starts to feed us spaghetti. But it should work
281  * in common situations. Otherwise, we have to rely on queue collapsing.
282  */
283
284 /* Slow part of check#2. */
285 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
286 {
287         struct tcp_sock *tp = tcp_sk(sk);
288         /* Optimize this! */
289         int truesize = tcp_win_from_space(skb->truesize) >> 1;
290         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
291
292         while (tp->rcv_ssthresh <= window) {
293                 if (truesize <= skb->len)
294                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
295
296                 truesize >>= 1;
297                 window >>= 1;
298         }
299         return 0;
300 }
301
302 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
303 {
304         struct tcp_sock *tp = tcp_sk(sk);
305
306         /* Check #1 */
307         if (tp->rcv_ssthresh < tp->window_clamp &&
308             (int)tp->rcv_ssthresh < tcp_space(sk) &&
309             !tcp_memory_pressure) {
310                 int incr;
311
312                 /* Check #2. Increase window, if skb with such overhead
313                  * will fit to rcvbuf in future.
314                  */
315                 if (tcp_win_from_space(skb->truesize) <= skb->len)
316                         incr = 2 * tp->advmss;
317                 else
318                         incr = __tcp_grow_window(sk, skb);
319
320                 if (incr) {
321                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
322                                                tp->window_clamp);
323                         inet_csk(sk)->icsk_ack.quick |= 1;
324                 }
325         }
326 }
327
328 /* 3. Tuning rcvbuf, when connection enters established state. */
329
330 static void tcp_fixup_rcvbuf(struct sock *sk)
331 {
332         struct tcp_sock *tp = tcp_sk(sk);
333         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
334
335         /* Try to select rcvbuf so that 4 mss-sized segments
336          * will fit to window and corresponding skbs will fit to our rcvbuf.
337          * (was 3; 4 is minimum to allow fast retransmit to work.)
338          */
339         while (tcp_win_from_space(rcvmem) < tp->advmss)
340                 rcvmem += 128;
341         if (sk->sk_rcvbuf < 4 * rcvmem)
342                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
343 }
344
345 /* 4. Try to fixup all. It is made immediately after connection enters
346  *    established state.
347  */
348 static void tcp_init_buffer_space(struct sock *sk)
349 {
350         struct tcp_sock *tp = tcp_sk(sk);
351         int maxwin;
352
353         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
354                 tcp_fixup_rcvbuf(sk);
355         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
356                 tcp_fixup_sndbuf(sk);
357
358         tp->rcvq_space.space = tp->rcv_wnd;
359
360         maxwin = tcp_full_space(sk);
361
362         if (tp->window_clamp >= maxwin) {
363                 tp->window_clamp = maxwin;
364
365                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
366                         tp->window_clamp = max(maxwin -
367                                                (maxwin >> sysctl_tcp_app_win),
368                                                4 * tp->advmss);
369         }
370
371         /* Force reservation of one segment. */
372         if (sysctl_tcp_app_win &&
373             tp->window_clamp > 2 * tp->advmss &&
374             tp->window_clamp + tp->advmss > maxwin)
375                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
376
377         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
378         tp->snd_cwnd_stamp = tcp_time_stamp;
379 }
380
381 /* 5. Recalculate window clamp after socket hit its memory bounds. */
382 static void tcp_clamp_window(struct sock *sk)
383 {
384         struct tcp_sock *tp = tcp_sk(sk);
385         struct inet_connection_sock *icsk = inet_csk(sk);
386
387         icsk->icsk_ack.quick = 0;
388
389         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
390             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
391             !tcp_memory_pressure &&
392             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
393                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
394                                     sysctl_tcp_rmem[2]);
395         }
396         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
397                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
398 }
399
400 /* Initialize RCV_MSS value.
401  * RCV_MSS is an our guess about MSS used by the peer.
402  * We haven't any direct information about the MSS.
403  * It's better to underestimate the RCV_MSS rather than overestimate.
404  * Overestimations make us ACKing less frequently than needed.
405  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
406  */
407 void tcp_initialize_rcv_mss(struct sock *sk)
408 {
409         struct tcp_sock *tp = tcp_sk(sk);
410         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
411
412         hint = min(hint, tp->rcv_wnd / 2);
413         hint = min(hint, TCP_MIN_RCVMSS);
414         hint = max(hint, TCP_MIN_MSS);
415
416         inet_csk(sk)->icsk_ack.rcv_mss = hint;
417 }
418
419 /* Receiver "autotuning" code.
420  *
421  * The algorithm for RTT estimation w/o timestamps is based on
422  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
423  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
424  *
425  * More detail on this code can be found at
426  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
427  * though this reference is out of date.  A new paper
428  * is pending.
429  */
430 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
431 {
432         u32 new_sample = tp->rcv_rtt_est.rtt;
433         long m = sample;
434
435         if (m == 0)
436                 m = 1;
437
438         if (new_sample != 0) {
439                 /* If we sample in larger samples in the non-timestamp
440                  * case, we could grossly overestimate the RTT especially
441                  * with chatty applications or bulk transfer apps which
442                  * are stalled on filesystem I/O.
443                  *
444                  * Also, since we are only going for a minimum in the
445                  * non-timestamp case, we do not smooth things out
446                  * else with timestamps disabled convergence takes too
447                  * long.
448                  */
449                 if (!win_dep) {
450                         m -= (new_sample >> 3);
451                         new_sample += m;
452                 } else if (m < new_sample)
453                         new_sample = m << 3;
454         } else {
455                 /* No previous measure. */
456                 new_sample = m << 3;
457         }
458
459         if (tp->rcv_rtt_est.rtt != new_sample)
460                 tp->rcv_rtt_est.rtt = new_sample;
461 }
462
463 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
464 {
465         if (tp->rcv_rtt_est.time == 0)
466                 goto new_measure;
467         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
468                 return;
469         tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
470
471 new_measure:
472         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
473         tp->rcv_rtt_est.time = tcp_time_stamp;
474 }
475
476 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
477                                           const struct sk_buff *skb)
478 {
479         struct tcp_sock *tp = tcp_sk(sk);
480         if (tp->rx_opt.rcv_tsecr &&
481             (TCP_SKB_CB(skb)->end_seq -
482              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
483                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
484 }
485
486 /*
487  * This function should be called every time data is copied to user space.
488  * It calculates the appropriate TCP receive buffer space.
489  */
490 void tcp_rcv_space_adjust(struct sock *sk)
491 {
492         struct tcp_sock *tp = tcp_sk(sk);
493         int time;
494         int space;
495
496         if (tp->rcvq_space.time == 0)
497                 goto new_measure;
498
499         time = tcp_time_stamp - tp->rcvq_space.time;
500         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
501                 return;
502
503         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
504
505         space = max(tp->rcvq_space.space, space);
506
507         if (tp->rcvq_space.space != space) {
508                 int rcvmem;
509
510                 tp->rcvq_space.space = space;
511
512                 if (sysctl_tcp_moderate_rcvbuf &&
513                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
514                         int new_clamp = space;
515
516                         /* Receive space grows, normalize in order to
517                          * take into account packet headers and sk_buff
518                          * structure overhead.
519                          */
520                         space /= tp->advmss;
521                         if (!space)
522                                 space = 1;
523                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
524                                   16 + sizeof(struct sk_buff));
525                         while (tcp_win_from_space(rcvmem) < tp->advmss)
526                                 rcvmem += 128;
527                         space *= rcvmem;
528                         space = min(space, sysctl_tcp_rmem[2]);
529                         if (space > sk->sk_rcvbuf) {
530                                 sk->sk_rcvbuf = space;
531
532                                 /* Make the window clamp follow along.  */
533                                 tp->window_clamp = new_clamp;
534                         }
535                 }
536         }
537
538 new_measure:
539         tp->rcvq_space.seq = tp->copied_seq;
540         tp->rcvq_space.time = tcp_time_stamp;
541 }
542
543 /* There is something which you must keep in mind when you analyze the
544  * behavior of the tp->ato delayed ack timeout interval.  When a
545  * connection starts up, we want to ack as quickly as possible.  The
546  * problem is that "good" TCP's do slow start at the beginning of data
547  * transmission.  The means that until we send the first few ACK's the
548  * sender will sit on his end and only queue most of his data, because
549  * he can only send snd_cwnd unacked packets at any given time.  For
550  * each ACK we send, he increments snd_cwnd and transmits more of his
551  * queue.  -DaveM
552  */
553 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
554 {
555         struct tcp_sock *tp = tcp_sk(sk);
556         struct inet_connection_sock *icsk = inet_csk(sk);
557         u32 now;
558
559         inet_csk_schedule_ack(sk);
560
561         tcp_measure_rcv_mss(sk, skb);
562
563         tcp_rcv_rtt_measure(tp);
564
565         now = tcp_time_stamp;
566
567         if (!icsk->icsk_ack.ato) {
568                 /* The _first_ data packet received, initialize
569                  * delayed ACK engine.
570                  */
571                 tcp_incr_quickack(sk);
572                 icsk->icsk_ack.ato = TCP_ATO_MIN;
573         } else {
574                 int m = now - icsk->icsk_ack.lrcvtime;
575
576                 if (m <= TCP_ATO_MIN / 2) {
577                         /* The fastest case is the first. */
578                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
579                 } else if (m < icsk->icsk_ack.ato) {
580                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
581                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
582                                 icsk->icsk_ack.ato = icsk->icsk_rto;
583                 } else if (m > icsk->icsk_rto) {
584                         /* Too long gap. Apparently sender failed to
585                          * restart window, so that we send ACKs quickly.
586                          */
587                         tcp_incr_quickack(sk);
588                         sk_mem_reclaim(sk);
589                 }
590         }
591         icsk->icsk_ack.lrcvtime = now;
592
593         TCP_ECN_check_ce(tp, skb);
594
595         if (skb->len >= 128)
596                 tcp_grow_window(sk, skb);
597 }
598
599 static u32 tcp_rto_min(struct sock *sk)
600 {
601         struct dst_entry *dst = __sk_dst_get(sk);
602         u32 rto_min = TCP_RTO_MIN;
603
604         if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
605                 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
606         return rto_min;
607 }
608
609 /* Called to compute a smoothed rtt estimate. The data fed to this
610  * routine either comes from timestamps, or from segments that were
611  * known _not_ to have been retransmitted [see Karn/Partridge
612  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
613  * piece by Van Jacobson.
614  * NOTE: the next three routines used to be one big routine.
615  * To save cycles in the RFC 1323 implementation it was better to break
616  * it up into three procedures. -- erics
617  */
618 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
619 {
620         struct tcp_sock *tp = tcp_sk(sk);
621         long m = mrtt; /* RTT */
622
623         /*      The following amusing code comes from Jacobson's
624          *      article in SIGCOMM '88.  Note that rtt and mdev
625          *      are scaled versions of rtt and mean deviation.
626          *      This is designed to be as fast as possible
627          *      m stands for "measurement".
628          *
629          *      On a 1990 paper the rto value is changed to:
630          *      RTO = rtt + 4 * mdev
631          *
632          * Funny. This algorithm seems to be very broken.
633          * These formulae increase RTO, when it should be decreased, increase
634          * too slowly, when it should be increased quickly, decrease too quickly
635          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
636          * does not matter how to _calculate_ it. Seems, it was trap
637          * that VJ failed to avoid. 8)
638          */
639         if (m == 0)
640                 m = 1;
641         if (tp->srtt != 0) {
642                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
643                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
644                 if (m < 0) {
645                         m = -m;         /* m is now abs(error) */
646                         m -= (tp->mdev >> 2);   /* similar update on mdev */
647                         /* This is similar to one of Eifel findings.
648                          * Eifel blocks mdev updates when rtt decreases.
649                          * This solution is a bit different: we use finer gain
650                          * for mdev in this case (alpha*beta).
651                          * Like Eifel it also prevents growth of rto,
652                          * but also it limits too fast rto decreases,
653                          * happening in pure Eifel.
654                          */
655                         if (m > 0)
656                                 m >>= 3;
657                 } else {
658                         m -= (tp->mdev >> 2);   /* similar update on mdev */
659                 }
660                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
661                 if (tp->mdev > tp->mdev_max) {
662                         tp->mdev_max = tp->mdev;
663                         if (tp->mdev_max > tp->rttvar)
664                                 tp->rttvar = tp->mdev_max;
665                 }
666                 if (after(tp->snd_una, tp->rtt_seq)) {
667                         if (tp->mdev_max < tp->rttvar)
668                                 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
669                         tp->rtt_seq = tp->snd_nxt;
670                         tp->mdev_max = tcp_rto_min(sk);
671                 }
672         } else {
673                 /* no previous measure. */
674                 tp->srtt = m << 3;      /* take the measured time to be rtt */
675                 tp->mdev = m << 1;      /* make sure rto = 3*rtt */
676                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
677                 tp->rtt_seq = tp->snd_nxt;
678         }
679 }
680
681 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
682  * routine referred to above.
683  */
684 static inline void tcp_set_rto(struct sock *sk)
685 {
686         const struct tcp_sock *tp = tcp_sk(sk);
687         /* Old crap is replaced with new one. 8)
688          *
689          * More seriously:
690          * 1. If rtt variance happened to be less 50msec, it is hallucination.
691          *    It cannot be less due to utterly erratic ACK generation made
692          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
693          *    to do with delayed acks, because at cwnd>2 true delack timeout
694          *    is invisible. Actually, Linux-2.4 also generates erratic
695          *    ACKs in some circumstances.
696          */
697         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
698
699         /* 2. Fixups made earlier cannot be right.
700          *    If we do not estimate RTO correctly without them,
701          *    all the algo is pure shit and should be replaced
702          *    with correct one. It is exactly, which we pretend to do.
703          */
704 }
705
706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707  * guarantees that rto is higher.
708  */
709 static inline void tcp_bound_rto(struct sock *sk)
710 {
711         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
712                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
713 }
714
715 /* Save metrics learned by this TCP session.
716    This function is called only, when TCP finishes successfully
717    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
718  */
719 void tcp_update_metrics(struct sock *sk)
720 {
721         struct tcp_sock *tp = tcp_sk(sk);
722         struct dst_entry *dst = __sk_dst_get(sk);
723
724         if (sysctl_tcp_nometrics_save)
725                 return;
726
727         dst_confirm(dst);
728
729         if (dst && (dst->flags & DST_HOST)) {
730                 const struct inet_connection_sock *icsk = inet_csk(sk);
731                 int m;
732                 unsigned long rtt;
733
734                 if (icsk->icsk_backoff || !tp->srtt) {
735                         /* This session failed to estimate rtt. Why?
736                          * Probably, no packets returned in time.
737                          * Reset our results.
738                          */
739                         if (!(dst_metric_locked(dst, RTAX_RTT)))
740                                 dst->metrics[RTAX_RTT - 1] = 0;
741                         return;
742                 }
743
744                 rtt = dst_metric_rtt(dst, RTAX_RTT);
745                 m = rtt - tp->srtt;
746
747                 /* If newly calculated rtt larger than stored one,
748                  * store new one. Otherwise, use EWMA. Remember,
749                  * rtt overestimation is always better than underestimation.
750                  */
751                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
752                         if (m <= 0)
753                                 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
754                         else
755                                 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
756                 }
757
758                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
759                         unsigned long var;
760                         if (m < 0)
761                                 m = -m;
762
763                         /* Scale deviation to rttvar fixed point */
764                         m >>= 1;
765                         if (m < tp->mdev)
766                                 m = tp->mdev;
767
768                         var = dst_metric_rtt(dst, RTAX_RTTVAR);
769                         if (m >= var)
770                                 var = m;
771                         else
772                                 var -= (var - m) >> 2;
773
774                         set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
775                 }
776
777                 if (tp->snd_ssthresh >= 0xFFFF) {
778                         /* Slow start still did not finish. */
779                         if (dst_metric(dst, RTAX_SSTHRESH) &&
780                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
781                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
782                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
783                         if (!dst_metric_locked(dst, RTAX_CWND) &&
784                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
785                                 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
786                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
787                            icsk->icsk_ca_state == TCP_CA_Open) {
788                         /* Cong. avoidance phase, cwnd is reliable. */
789                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
790                                 dst->metrics[RTAX_SSTHRESH-1] =
791                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
792                         if (!dst_metric_locked(dst, RTAX_CWND))
793                                 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
794                 } else {
795                         /* Else slow start did not finish, cwnd is non-sense,
796                            ssthresh may be also invalid.
797                          */
798                         if (!dst_metric_locked(dst, RTAX_CWND))
799                                 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
800                         if (dst_metric(dst, RTAX_SSTHRESH) &&
801                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
802                             tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
803                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
804                 }
805
806                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
807                         if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
808                             tp->reordering != sysctl_tcp_reordering)
809                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
810                 }
811         }
812 }
813
814 /* Numbers are taken from RFC3390.
815  *
816  * John Heffner states:
817  *
818  *      The RFC specifies a window of no more than 4380 bytes
819  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
820  *      is a bit misleading because they use a clamp at 4380 bytes
821  *      rather than use a multiplier in the relevant range.
822  */
823 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
824 {
825         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
826
827         if (!cwnd) {
828                 if (tp->mss_cache > 1460)
829                         cwnd = 2;
830                 else
831                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
832         }
833         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
834 }
835
836 /* Set slow start threshold and cwnd not falling to slow start */
837 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
838 {
839         struct tcp_sock *tp = tcp_sk(sk);
840         const struct inet_connection_sock *icsk = inet_csk(sk);
841
842         tp->prior_ssthresh = 0;
843         tp->bytes_acked = 0;
844         if (icsk->icsk_ca_state < TCP_CA_CWR) {
845                 tp->undo_marker = 0;
846                 if (set_ssthresh)
847                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
848                 tp->snd_cwnd = min(tp->snd_cwnd,
849                                    tcp_packets_in_flight(tp) + 1U);
850                 tp->snd_cwnd_cnt = 0;
851                 tp->high_seq = tp->snd_nxt;
852                 tp->snd_cwnd_stamp = tcp_time_stamp;
853                 TCP_ECN_queue_cwr(tp);
854
855                 tcp_set_ca_state(sk, TCP_CA_CWR);
856         }
857 }
858
859 /*
860  * Packet counting of FACK is based on in-order assumptions, therefore TCP
861  * disables it when reordering is detected
862  */
863 static void tcp_disable_fack(struct tcp_sock *tp)
864 {
865         /* RFC3517 uses different metric in lost marker => reset on change */
866         if (tcp_is_fack(tp))
867                 tp->lost_skb_hint = NULL;
868         tp->rx_opt.sack_ok &= ~2;
869 }
870
871 /* Take a notice that peer is sending D-SACKs */
872 static void tcp_dsack_seen(struct tcp_sock *tp)
873 {
874         tp->rx_opt.sack_ok |= 4;
875 }
876
877 /* Initialize metrics on socket. */
878
879 static void tcp_init_metrics(struct sock *sk)
880 {
881         struct tcp_sock *tp = tcp_sk(sk);
882         struct dst_entry *dst = __sk_dst_get(sk);
883
884         if (dst == NULL)
885                 goto reset;
886
887         dst_confirm(dst);
888
889         if (dst_metric_locked(dst, RTAX_CWND))
890                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
891         if (dst_metric(dst, RTAX_SSTHRESH)) {
892                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
893                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
894                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
895         }
896         if (dst_metric(dst, RTAX_REORDERING) &&
897             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
898                 tcp_disable_fack(tp);
899                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
900         }
901
902         if (dst_metric(dst, RTAX_RTT) == 0)
903                 goto reset;
904
905         if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
906                 goto reset;
907
908         /* Initial rtt is determined from SYN,SYN-ACK.
909          * The segment is small and rtt may appear much
910          * less than real one. Use per-dst memory
911          * to make it more realistic.
912          *
913          * A bit of theory. RTT is time passed after "normal" sized packet
914          * is sent until it is ACKed. In normal circumstances sending small
915          * packets force peer to delay ACKs and calculation is correct too.
916          * The algorithm is adaptive and, provided we follow specs, it
917          * NEVER underestimate RTT. BUT! If peer tries to make some clever
918          * tricks sort of "quick acks" for time long enough to decrease RTT
919          * to low value, and then abruptly stops to do it and starts to delay
920          * ACKs, wait for troubles.
921          */
922         if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
923                 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
924                 tp->rtt_seq = tp->snd_nxt;
925         }
926         if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
927                 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
928                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
929         }
930         tcp_set_rto(sk);
931         tcp_bound_rto(sk);
932         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
933                 goto reset;
934         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
935         tp->snd_cwnd_stamp = tcp_time_stamp;
936         return;
937
938 reset:
939         /* Play conservative. If timestamps are not
940          * supported, TCP will fail to recalculate correct
941          * rtt, if initial rto is too small. FORGET ALL AND RESET!
942          */
943         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
944                 tp->srtt = 0;
945                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
946                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
947         }
948 }
949
950 static void tcp_update_reordering(struct sock *sk, const int metric,
951                                   const int ts)
952 {
953         struct tcp_sock *tp = tcp_sk(sk);
954         if (metric > tp->reordering) {
955                 int mib_idx;
956
957                 tp->reordering = min(TCP_MAX_REORDERING, metric);
958
959                 /* This exciting event is worth to be remembered. 8) */
960                 if (ts)
961                         mib_idx = LINUX_MIB_TCPTSREORDER;
962                 else if (tcp_is_reno(tp))
963                         mib_idx = LINUX_MIB_TCPRENOREORDER;
964                 else if (tcp_is_fack(tp))
965                         mib_idx = LINUX_MIB_TCPFACKREORDER;
966                 else
967                         mib_idx = LINUX_MIB_TCPSACKREORDER;
968
969                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
970 #if FASTRETRANS_DEBUG > 1
971                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
972                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
973                        tp->reordering,
974                        tp->fackets_out,
975                        tp->sacked_out,
976                        tp->undo_marker ? tp->undo_retrans : 0);
977 #endif
978                 tcp_disable_fack(tp);
979         }
980 }
981
982 /* This procedure tags the retransmission queue when SACKs arrive.
983  *
984  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
985  * Packets in queue with these bits set are counted in variables
986  * sacked_out, retrans_out and lost_out, correspondingly.
987  *
988  * Valid combinations are:
989  * Tag  InFlight        Description
990  * 0    1               - orig segment is in flight.
991  * S    0               - nothing flies, orig reached receiver.
992  * L    0               - nothing flies, orig lost by net.
993  * R    2               - both orig and retransmit are in flight.
994  * L|R  1               - orig is lost, retransmit is in flight.
995  * S|R  1               - orig reached receiver, retrans is still in flight.
996  * (L|S|R is logically valid, it could occur when L|R is sacked,
997  *  but it is equivalent to plain S and code short-curcuits it to S.
998  *  L|S is logically invalid, it would mean -1 packet in flight 8))
999  *
1000  * These 6 states form finite state machine, controlled by the following events:
1001  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1002  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1003  * 3. Loss detection event of one of three flavors:
1004  *      A. Scoreboard estimator decided the packet is lost.
1005  *         A'. Reno "three dupacks" marks head of queue lost.
1006  *         A''. Its FACK modfication, head until snd.fack is lost.
1007  *      B. SACK arrives sacking data transmitted after never retransmitted
1008  *         hole was sent out.
1009  *      C. SACK arrives sacking SND.NXT at the moment, when the
1010  *         segment was retransmitted.
1011  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1012  *
1013  * It is pleasant to note, that state diagram turns out to be commutative,
1014  * so that we are allowed not to be bothered by order of our actions,
1015  * when multiple events arrive simultaneously. (see the function below).
1016  *
1017  * Reordering detection.
1018  * --------------------
1019  * Reordering metric is maximal distance, which a packet can be displaced
1020  * in packet stream. With SACKs we can estimate it:
1021  *
1022  * 1. SACK fills old hole and the corresponding segment was not
1023  *    ever retransmitted -> reordering. Alas, we cannot use it
1024  *    when segment was retransmitted.
1025  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1026  *    for retransmitted and already SACKed segment -> reordering..
1027  * Both of these heuristics are not used in Loss state, when we cannot
1028  * account for retransmits accurately.
1029  *
1030  * SACK block validation.
1031  * ----------------------
1032  *
1033  * SACK block range validation checks that the received SACK block fits to
1034  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1035  * Note that SND.UNA is not included to the range though being valid because
1036  * it means that the receiver is rather inconsistent with itself reporting
1037  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1038  * perfectly valid, however, in light of RFC2018 which explicitly states
1039  * that "SACK block MUST reflect the newest segment.  Even if the newest
1040  * segment is going to be discarded ...", not that it looks very clever
1041  * in case of head skb. Due to potentional receiver driven attacks, we
1042  * choose to avoid immediate execution of a walk in write queue due to
1043  * reneging and defer head skb's loss recovery to standard loss recovery
1044  * procedure that will eventually trigger (nothing forbids us doing this).
1045  *
1046  * Implements also blockage to start_seq wrap-around. Problem lies in the
1047  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1048  * there's no guarantee that it will be before snd_nxt (n). The problem
1049  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1050  * wrap (s_w):
1051  *
1052  *         <- outs wnd ->                          <- wrapzone ->
1053  *         u     e      n                         u_w   e_w  s n_w
1054  *         |     |      |                          |     |   |  |
1055  * |<------------+------+----- TCP seqno space --------------+---------->|
1056  * ...-- <2^31 ->|                                           |<--------...
1057  * ...---- >2^31 ------>|                                    |<--------...
1058  *
1059  * Current code wouldn't be vulnerable but it's better still to discard such
1060  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1061  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1062  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1063  * equal to the ideal case (infinite seqno space without wrap caused issues).
1064  *
1065  * With D-SACK the lower bound is extended to cover sequence space below
1066  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1067  * again, D-SACK block must not to go across snd_una (for the same reason as
1068  * for the normal SACK blocks, explained above). But there all simplicity
1069  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1070  * fully below undo_marker they do not affect behavior in anyway and can
1071  * therefore be safely ignored. In rare cases (which are more or less
1072  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1073  * fragmentation and packet reordering past skb's retransmission. To consider
1074  * them correctly, the acceptable range must be extended even more though
1075  * the exact amount is rather hard to quantify. However, tp->max_window can
1076  * be used as an exaggerated estimate.
1077  */
1078 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1079                                   u32 start_seq, u32 end_seq)
1080 {
1081         /* Too far in future, or reversed (interpretation is ambiguous) */
1082         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1083                 return 0;
1084
1085         /* Nasty start_seq wrap-around check (see comments above) */
1086         if (!before(start_seq, tp->snd_nxt))
1087                 return 0;
1088
1089         /* In outstanding window? ...This is valid exit for D-SACKs too.
1090          * start_seq == snd_una is non-sensical (see comments above)
1091          */
1092         if (after(start_seq, tp->snd_una))
1093                 return 1;
1094
1095         if (!is_dsack || !tp->undo_marker)
1096                 return 0;
1097
1098         /* ...Then it's D-SACK, and must reside below snd_una completely */
1099         if (!after(end_seq, tp->snd_una))
1100                 return 0;
1101
1102         if (!before(start_seq, tp->undo_marker))
1103                 return 1;
1104
1105         /* Too old */
1106         if (!after(end_seq, tp->undo_marker))
1107                 return 0;
1108
1109         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1110          *   start_seq < undo_marker and end_seq >= undo_marker.
1111          */
1112         return !before(start_seq, end_seq - tp->max_window);
1113 }
1114
1115 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1116  * Event "C". Later note: FACK people cheated me again 8), we have to account
1117  * for reordering! Ugly, but should help.
1118  *
1119  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1120  * less than what is now known to be received by the other end (derived from
1121  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1122  * retransmitted skbs to avoid some costly processing per ACKs.
1123  */
1124 static void tcp_mark_lost_retrans(struct sock *sk)
1125 {
1126         const struct inet_connection_sock *icsk = inet_csk(sk);
1127         struct tcp_sock *tp = tcp_sk(sk);
1128         struct sk_buff *skb;
1129         int cnt = 0;
1130         u32 new_low_seq = tp->snd_nxt;
1131         u32 received_upto = tcp_highest_sack_seq(tp);
1132
1133         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1134             !after(received_upto, tp->lost_retrans_low) ||
1135             icsk->icsk_ca_state != TCP_CA_Recovery)
1136                 return;
1137
1138         tcp_for_write_queue(skb, sk) {
1139                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1140
1141                 if (skb == tcp_send_head(sk))
1142                         break;
1143                 if (cnt == tp->retrans_out)
1144                         break;
1145                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1146                         continue;
1147
1148                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1149                         continue;
1150
1151                 if (after(received_upto, ack_seq) &&
1152                     (tcp_is_fack(tp) ||
1153                      !before(received_upto,
1154                              ack_seq + tp->reordering * tp->mss_cache))) {
1155                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1156                         tp->retrans_out -= tcp_skb_pcount(skb);
1157
1158                         /* clear lost hint */
1159                         tp->retransmit_skb_hint = NULL;
1160
1161                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1162                                 tp->lost_out += tcp_skb_pcount(skb);
1163                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1164                         }
1165                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1166                 } else {
1167                         if (before(ack_seq, new_low_seq))
1168                                 new_low_seq = ack_seq;
1169                         cnt += tcp_skb_pcount(skb);
1170                 }
1171         }
1172
1173         if (tp->retrans_out)
1174                 tp->lost_retrans_low = new_low_seq;
1175 }
1176
1177 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1178                            struct tcp_sack_block_wire *sp, int num_sacks,
1179                            u32 prior_snd_una)
1180 {
1181         struct tcp_sock *tp = tcp_sk(sk);
1182         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1183         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1184         int dup_sack = 0;
1185
1186         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1187                 dup_sack = 1;
1188                 tcp_dsack_seen(tp);
1189                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1190         } else if (num_sacks > 1) {
1191                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1192                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1193
1194                 if (!after(end_seq_0, end_seq_1) &&
1195                     !before(start_seq_0, start_seq_1)) {
1196                         dup_sack = 1;
1197                         tcp_dsack_seen(tp);
1198                         NET_INC_STATS_BH(sock_net(sk),
1199                                         LINUX_MIB_TCPDSACKOFORECV);
1200                 }
1201         }
1202
1203         /* D-SACK for already forgotten data... Do dumb counting. */
1204         if (dup_sack &&
1205             !after(end_seq_0, prior_snd_una) &&
1206             after(end_seq_0, tp->undo_marker))
1207                 tp->undo_retrans--;
1208
1209         return dup_sack;
1210 }
1211
1212 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1213  * the incoming SACK may not exactly match but we can find smaller MSS
1214  * aligned portion of it that matches. Therefore we might need to fragment
1215  * which may fail and creates some hassle (caller must handle error case
1216  * returns).
1217  */
1218 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1219                                  u32 start_seq, u32 end_seq)
1220 {
1221         int in_sack, err;
1222         unsigned int pkt_len;
1223
1224         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1225                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1226
1227         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1228             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1229
1230                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1231
1232                 if (!in_sack)
1233                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1234                 else
1235                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1236                 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1237                 if (err < 0)
1238                         return err;
1239         }
1240
1241         return in_sack;
1242 }
1243
1244 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1245                            int *reord, int dup_sack, int fack_count)
1246 {
1247         struct tcp_sock *tp = tcp_sk(sk);
1248         u8 sacked = TCP_SKB_CB(skb)->sacked;
1249         int flag = 0;
1250
1251         /* Account D-SACK for retransmitted packet. */
1252         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1253                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1254                         tp->undo_retrans--;
1255                 if (sacked & TCPCB_SACKED_ACKED)
1256                         *reord = min(fack_count, *reord);
1257         }
1258
1259         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1260         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1261                 return flag;
1262
1263         if (!(sacked & TCPCB_SACKED_ACKED)) {
1264                 if (sacked & TCPCB_SACKED_RETRANS) {
1265                         /* If the segment is not tagged as lost,
1266                          * we do not clear RETRANS, believing
1267                          * that retransmission is still in flight.
1268                          */
1269                         if (sacked & TCPCB_LOST) {
1270                                 TCP_SKB_CB(skb)->sacked &=
1271                                         ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1272                                 tp->lost_out -= tcp_skb_pcount(skb);
1273                                 tp->retrans_out -= tcp_skb_pcount(skb);
1274
1275                                 /* clear lost hint */
1276                                 tp->retransmit_skb_hint = NULL;
1277                         }
1278                 } else {
1279                         if (!(sacked & TCPCB_RETRANS)) {
1280                                 /* New sack for not retransmitted frame,
1281                                  * which was in hole. It is reordering.
1282                                  */
1283                                 if (before(TCP_SKB_CB(skb)->seq,
1284                                            tcp_highest_sack_seq(tp)))
1285                                         *reord = min(fack_count, *reord);
1286
1287                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1288                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1289                                         flag |= FLAG_ONLY_ORIG_SACKED;
1290                         }
1291
1292                         if (sacked & TCPCB_LOST) {
1293                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1294                                 tp->lost_out -= tcp_skb_pcount(skb);
1295
1296                                 /* clear lost hint */
1297                                 tp->retransmit_skb_hint = NULL;
1298                         }
1299                 }
1300
1301                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1302                 flag |= FLAG_DATA_SACKED;
1303                 tp->sacked_out += tcp_skb_pcount(skb);
1304
1305                 fack_count += tcp_skb_pcount(skb);
1306
1307                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1308                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1309                     before(TCP_SKB_CB(skb)->seq,
1310                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1311                         tp->lost_cnt_hint += tcp_skb_pcount(skb);
1312
1313                 if (fack_count > tp->fackets_out)
1314                         tp->fackets_out = fack_count;
1315
1316                 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1317                         tcp_advance_highest_sack(sk, skb);
1318         }
1319
1320         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1321          * frames and clear it. undo_retrans is decreased above, L|R frames
1322          * are accounted above as well.
1323          */
1324         if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) {
1325                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1326                 tp->retrans_out -= tcp_skb_pcount(skb);
1327                 tp->retransmit_skb_hint = NULL;
1328         }
1329
1330         return flag;
1331 }
1332
1333 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1334                                         struct tcp_sack_block *next_dup,
1335                                         u32 start_seq, u32 end_seq,
1336                                         int dup_sack_in, int *fack_count,
1337                                         int *reord, int *flag)
1338 {
1339         tcp_for_write_queue_from(skb, sk) {
1340                 int in_sack = 0;
1341                 int dup_sack = dup_sack_in;
1342
1343                 if (skb == tcp_send_head(sk))
1344                         break;
1345
1346                 /* queue is in-order => we can short-circuit the walk early */
1347                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1348                         break;
1349
1350                 if ((next_dup != NULL) &&
1351                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1352                         in_sack = tcp_match_skb_to_sack(sk, skb,
1353                                                         next_dup->start_seq,
1354                                                         next_dup->end_seq);
1355                         if (in_sack > 0)
1356                                 dup_sack = 1;
1357                 }
1358
1359                 if (in_sack <= 0)
1360                         in_sack = tcp_match_skb_to_sack(sk, skb, start_seq,
1361                                                         end_seq);
1362                 if (unlikely(in_sack < 0))
1363                         break;
1364
1365                 if (in_sack)
1366                         *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack,
1367                                                  *fack_count);
1368
1369                 *fack_count += tcp_skb_pcount(skb);
1370         }
1371         return skb;
1372 }
1373
1374 /* Avoid all extra work that is being done by sacktag while walking in
1375  * a normal way
1376  */
1377 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1378                                         u32 skip_to_seq, int *fack_count)
1379 {
1380         tcp_for_write_queue_from(skb, sk) {
1381                 if (skb == tcp_send_head(sk))
1382                         break;
1383
1384                 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1385                         break;
1386
1387                 *fack_count += tcp_skb_pcount(skb);
1388         }
1389         return skb;
1390 }
1391
1392 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1393                                                 struct sock *sk,
1394                                                 struct tcp_sack_block *next_dup,
1395                                                 u32 skip_to_seq,
1396                                                 int *fack_count, int *reord,
1397                                                 int *flag)
1398 {
1399         if (next_dup == NULL)
1400                 return skb;
1401
1402         if (before(next_dup->start_seq, skip_to_seq)) {
1403                 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count);
1404                 skb = tcp_sacktag_walk(skb, sk, NULL,
1405                                      next_dup->start_seq, next_dup->end_seq,
1406                                      1, fack_count, reord, flag);
1407         }
1408
1409         return skb;
1410 }
1411
1412 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1413 {
1414         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1415 }
1416
1417 static int
1418 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1419                         u32 prior_snd_una)
1420 {
1421         const struct inet_connection_sock *icsk = inet_csk(sk);
1422         struct tcp_sock *tp = tcp_sk(sk);
1423         unsigned char *ptr = (skb_transport_header(ack_skb) +
1424                               TCP_SKB_CB(ack_skb)->sacked);
1425         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1426         struct tcp_sack_block sp[TCP_NUM_SACKS];
1427         struct tcp_sack_block *cache;
1428         struct sk_buff *skb;
1429         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1430         int used_sacks;
1431         int reord = tp->packets_out;
1432         int flag = 0;
1433         int found_dup_sack = 0;
1434         int fack_count;
1435         int i, j;
1436         int first_sack_index;
1437
1438         if (!tp->sacked_out) {
1439                 if (WARN_ON(tp->fackets_out))
1440                         tp->fackets_out = 0;
1441                 tcp_highest_sack_reset(sk);
1442         }
1443
1444         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1445                                          num_sacks, prior_snd_una);
1446         if (found_dup_sack)
1447                 flag |= FLAG_DSACKING_ACK;
1448
1449         /* Eliminate too old ACKs, but take into
1450          * account more or less fresh ones, they can
1451          * contain valid SACK info.
1452          */
1453         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1454                 return 0;
1455
1456         if (!tp->packets_out)
1457                 goto out;
1458
1459         used_sacks = 0;
1460         first_sack_index = 0;
1461         for (i = 0; i < num_sacks; i++) {
1462                 int dup_sack = !i && found_dup_sack;
1463
1464                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1465                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1466
1467                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1468                                             sp[used_sacks].start_seq,
1469                                             sp[used_sacks].end_seq)) {
1470                         int mib_idx;
1471
1472                         if (dup_sack) {
1473                                 if (!tp->undo_marker)
1474                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1475                                 else
1476                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1477                         } else {
1478                                 /* Don't count olds caused by ACK reordering */
1479                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1480                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1481                                         continue;
1482                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1483                         }
1484
1485                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1486                         if (i == 0)
1487                                 first_sack_index = -1;
1488                         continue;
1489                 }
1490
1491                 /* Ignore very old stuff early */
1492                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1493                         continue;
1494
1495                 used_sacks++;
1496         }
1497
1498         /* order SACK blocks to allow in order walk of the retrans queue */
1499         for (i = used_sacks - 1; i > 0; i--) {
1500                 for (j = 0; j < i; j++) {
1501                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1502                                 struct tcp_sack_block tmp;
1503
1504                                 tmp = sp[j];
1505                                 sp[j] = sp[j + 1];
1506                                 sp[j + 1] = tmp;
1507
1508                                 /* Track where the first SACK block goes to */
1509                                 if (j == first_sack_index)
1510                                         first_sack_index = j + 1;
1511                         }
1512                 }
1513         }
1514
1515         skb = tcp_write_queue_head(sk);
1516         fack_count = 0;
1517         i = 0;
1518
1519         if (!tp->sacked_out) {
1520                 /* It's already past, so skip checking against it */
1521                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1522         } else {
1523                 cache = tp->recv_sack_cache;
1524                 /* Skip empty blocks in at head of the cache */
1525                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1526                        !cache->end_seq)
1527                         cache++;
1528         }
1529
1530         while (i < used_sacks) {
1531                 u32 start_seq = sp[i].start_seq;
1532                 u32 end_seq = sp[i].end_seq;
1533                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1534                 struct tcp_sack_block *next_dup = NULL;
1535
1536                 if (found_dup_sack && ((i + 1) == first_sack_index))
1537                         next_dup = &sp[i + 1];
1538
1539                 /* Event "B" in the comment above. */
1540                 if (after(end_seq, tp->high_seq))
1541                         flag |= FLAG_DATA_LOST;
1542
1543                 /* Skip too early cached blocks */
1544                 while (tcp_sack_cache_ok(tp, cache) &&
1545                        !before(start_seq, cache->end_seq))
1546                         cache++;
1547
1548                 /* Can skip some work by looking recv_sack_cache? */
1549                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1550                     after(end_seq, cache->start_seq)) {
1551
1552                         /* Head todo? */
1553                         if (before(start_seq, cache->start_seq)) {
1554                                 skb = tcp_sacktag_skip(skb, sk, start_seq,
1555                                                        &fack_count);
1556                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1557                                                        start_seq,
1558                                                        cache->start_seq,
1559                                                        dup_sack, &fack_count,
1560                                                        &reord, &flag);
1561                         }
1562
1563                         /* Rest of the block already fully processed? */
1564                         if (!after(end_seq, cache->end_seq))
1565                                 goto advance_sp;
1566
1567                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1568                                                        cache->end_seq,
1569                                                        &fack_count, &reord,
1570                                                        &flag);
1571
1572                         /* ...tail remains todo... */
1573                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1574                                 /* ...but better entrypoint exists! */
1575                                 skb = tcp_highest_sack(sk);
1576                                 if (skb == NULL)
1577                                         break;
1578                                 fack_count = tp->fackets_out;
1579                                 cache++;
1580                                 goto walk;
1581                         }
1582
1583                         skb = tcp_sacktag_skip(skb, sk, cache->end_seq,
1584                                                &fack_count);
1585                         /* Check overlap against next cached too (past this one already) */
1586                         cache++;
1587                         continue;
1588                 }
1589
1590                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1591                         skb = tcp_highest_sack(sk);
1592                         if (skb == NULL)
1593                                 break;
1594                         fack_count = tp->fackets_out;
1595                 }
1596                 skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count);
1597
1598 walk:
1599                 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1600                                        dup_sack, &fack_count, &reord, &flag);
1601
1602 advance_sp:
1603                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1604                  * due to in-order walk
1605                  */
1606                 if (after(end_seq, tp->frto_highmark))
1607                         flag &= ~FLAG_ONLY_ORIG_SACKED;
1608
1609                 i++;
1610         }
1611
1612         /* Clear the head of the cache sack blocks so we can skip it next time */
1613         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1614                 tp->recv_sack_cache[i].start_seq = 0;
1615                 tp->recv_sack_cache[i].end_seq = 0;
1616         }
1617         for (j = 0; j < used_sacks; j++)
1618                 tp->recv_sack_cache[i++] = sp[j];
1619
1620         tcp_mark_lost_retrans(sk);
1621
1622         tcp_verify_left_out(tp);
1623
1624         if ((reord < tp->fackets_out) &&
1625             ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1626             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1627                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1628
1629 out:
1630
1631 #if FASTRETRANS_DEBUG > 0
1632         WARN_ON((int)tp->sacked_out < 0);
1633         WARN_ON((int)tp->lost_out < 0);
1634         WARN_ON((int)tp->retrans_out < 0);
1635         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1636 #endif
1637         return flag;
1638 }
1639
1640 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1641  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1642  */
1643 int tcp_limit_reno_sacked(struct tcp_sock *tp)
1644 {
1645         u32 holes;
1646
1647         holes = max(tp->lost_out, 1U);
1648         holes = min(holes, tp->packets_out);
1649
1650         if ((tp->sacked_out + holes) > tp->packets_out) {
1651                 tp->sacked_out = tp->packets_out - holes;
1652                 return 1;
1653         }
1654         return 0;
1655 }
1656
1657 /* If we receive more dupacks than we expected counting segments
1658  * in assumption of absent reordering, interpret this as reordering.
1659  * The only another reason could be bug in receiver TCP.
1660  */
1661 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1662 {
1663         struct tcp_sock *tp = tcp_sk(sk);
1664         if (tcp_limit_reno_sacked(tp))
1665                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1666 }
1667
1668 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1669
1670 static void tcp_add_reno_sack(struct sock *sk)
1671 {
1672         struct tcp_sock *tp = tcp_sk(sk);
1673         tp->sacked_out++;
1674         tcp_check_reno_reordering(sk, 0);
1675         tcp_verify_left_out(tp);
1676 }
1677
1678 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1679
1680 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1681 {
1682         struct tcp_sock *tp = tcp_sk(sk);
1683
1684         if (acked > 0) {
1685                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1686                 if (acked - 1 >= tp->sacked_out)
1687                         tp->sacked_out = 0;
1688                 else
1689                         tp->sacked_out -= acked - 1;
1690         }
1691         tcp_check_reno_reordering(sk, acked);
1692         tcp_verify_left_out(tp);
1693 }
1694
1695 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1696 {
1697         tp->sacked_out = 0;
1698 }
1699
1700 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1701 {
1702         return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1703 }
1704
1705 /* F-RTO can only be used if TCP has never retransmitted anything other than
1706  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1707  */
1708 int tcp_use_frto(struct sock *sk)
1709 {
1710         const struct tcp_sock *tp = tcp_sk(sk);
1711         const struct inet_connection_sock *icsk = inet_csk(sk);
1712         struct sk_buff *skb;
1713
1714         if (!sysctl_tcp_frto)
1715                 return 0;
1716
1717         /* MTU probe and F-RTO won't really play nicely along currently */
1718         if (icsk->icsk_mtup.probe_size)
1719                 return 0;
1720
1721         if (tcp_is_sackfrto(tp))
1722                 return 1;
1723
1724         /* Avoid expensive walking of rexmit queue if possible */
1725         if (tp->retrans_out > 1)
1726                 return 0;
1727
1728         skb = tcp_write_queue_head(sk);
1729         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1730         tcp_for_write_queue_from(skb, sk) {
1731                 if (skb == tcp_send_head(sk))
1732                         break;
1733                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1734                         return 0;
1735                 /* Short-circuit when first non-SACKed skb has been checked */
1736                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1737                         break;
1738         }
1739         return 1;
1740 }
1741
1742 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1743  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1744  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1745  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1746  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1747  * bits are handled if the Loss state is really to be entered (in
1748  * tcp_enter_frto_loss).
1749  *
1750  * Do like tcp_enter_loss() would; when RTO expires the second time it
1751  * does:
1752  *  "Reduce ssthresh if it has not yet been made inside this window."
1753  */
1754 void tcp_enter_frto(struct sock *sk)
1755 {
1756         const struct inet_connection_sock *icsk = inet_csk(sk);
1757         struct tcp_sock *tp = tcp_sk(sk);
1758         struct sk_buff *skb;
1759
1760         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1761             tp->snd_una == tp->high_seq ||
1762             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1763              !icsk->icsk_retransmits)) {
1764                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1765                 /* Our state is too optimistic in ssthresh() call because cwnd
1766                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1767                  * recovery has not yet completed. Pattern would be this: RTO,
1768                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1769                  * up here twice).
1770                  * RFC4138 should be more specific on what to do, even though
1771                  * RTO is quite unlikely to occur after the first Cumulative ACK
1772                  * due to back-off and complexity of triggering events ...
1773                  */
1774                 if (tp->frto_counter) {
1775                         u32 stored_cwnd;
1776                         stored_cwnd = tp->snd_cwnd;
1777                         tp->snd_cwnd = 2;
1778                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1779                         tp->snd_cwnd = stored_cwnd;
1780                 } else {
1781                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1782                 }
1783                 /* ... in theory, cong.control module could do "any tricks" in
1784                  * ssthresh(), which means that ca_state, lost bits and lost_out
1785                  * counter would have to be faked before the call occurs. We
1786                  * consider that too expensive, unlikely and hacky, so modules
1787                  * using these in ssthresh() must deal these incompatibility
1788                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1789                  */
1790                 tcp_ca_event(sk, CA_EVENT_FRTO);
1791         }
1792
1793         tp->undo_marker = tp->snd_una;
1794         tp->undo_retrans = 0;
1795
1796         skb = tcp_write_queue_head(sk);
1797         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1798                 tp->undo_marker = 0;
1799         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1800                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1801                 tp->retrans_out -= tcp_skb_pcount(skb);
1802         }
1803         tcp_verify_left_out(tp);
1804
1805         /* Too bad if TCP was application limited */
1806         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1807
1808         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1809          * The last condition is necessary at least in tp->frto_counter case.
1810          */
1811         if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
1812             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1813             after(tp->high_seq, tp->snd_una)) {
1814                 tp->frto_highmark = tp->high_seq;
1815         } else {
1816                 tp->frto_highmark = tp->snd_nxt;
1817         }
1818         tcp_set_ca_state(sk, TCP_CA_Disorder);
1819         tp->high_seq = tp->snd_nxt;
1820         tp->frto_counter = 1;
1821 }
1822
1823 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1824  * which indicates that we should follow the traditional RTO recovery,
1825  * i.e. mark everything lost and do go-back-N retransmission.
1826  */
1827 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1828 {
1829         struct tcp_sock *tp = tcp_sk(sk);
1830         struct sk_buff *skb;
1831
1832         tp->lost_out = 0;
1833         tp->retrans_out = 0;
1834         if (tcp_is_reno(tp))
1835                 tcp_reset_reno_sack(tp);
1836
1837         tcp_for_write_queue(skb, sk) {
1838                 if (skb == tcp_send_head(sk))
1839                         break;
1840
1841                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1842                 /*
1843                  * Count the retransmission made on RTO correctly (only when
1844                  * waiting for the first ACK and did not get it)...
1845                  */
1846                 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1847                         /* For some reason this R-bit might get cleared? */
1848                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1849                                 tp->retrans_out += tcp_skb_pcount(skb);
1850                         /* ...enter this if branch just for the first segment */
1851                         flag |= FLAG_DATA_ACKED;
1852                 } else {
1853                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1854                                 tp->undo_marker = 0;
1855                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1856                 }
1857
1858                 /* Marking forward transmissions that were made after RTO lost
1859                  * can cause unnecessary retransmissions in some scenarios,
1860                  * SACK blocks will mitigate that in some but not in all cases.
1861                  * We used to not mark them but it was causing break-ups with
1862                  * receivers that do only in-order receival.
1863                  *
1864                  * TODO: we could detect presence of such receiver and select
1865                  * different behavior per flow.
1866                  */
1867                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1868                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1869                         tp->lost_out += tcp_skb_pcount(skb);
1870                 }
1871         }
1872         tcp_verify_left_out(tp);
1873
1874         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1875         tp->snd_cwnd_cnt = 0;
1876         tp->snd_cwnd_stamp = tcp_time_stamp;
1877         tp->frto_counter = 0;
1878         tp->bytes_acked = 0;
1879
1880         tp->reordering = min_t(unsigned int, tp->reordering,
1881                                sysctl_tcp_reordering);
1882         tcp_set_ca_state(sk, TCP_CA_Loss);
1883         tp->high_seq = tp->snd_nxt;
1884         TCP_ECN_queue_cwr(tp);
1885
1886         tcp_clear_retrans_hints_partial(tp);
1887 }
1888
1889 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1890 {
1891         tp->retrans_out = 0;
1892         tp->lost_out = 0;
1893
1894         tp->undo_marker = 0;
1895         tp->undo_retrans = 0;
1896 }
1897
1898 void tcp_clear_retrans(struct tcp_sock *tp)
1899 {
1900         tcp_clear_retrans_partial(tp);
1901
1902         tp->fackets_out = 0;
1903         tp->sacked_out = 0;
1904 }
1905
1906 /* Enter Loss state. If "how" is not zero, forget all SACK information
1907  * and reset tags completely, otherwise preserve SACKs. If receiver
1908  * dropped its ofo queue, we will know this due to reneging detection.
1909  */
1910 void tcp_enter_loss(struct sock *sk, int how)
1911 {
1912         const struct inet_connection_sock *icsk = inet_csk(sk);
1913         struct tcp_sock *tp = tcp_sk(sk);
1914         struct sk_buff *skb;
1915
1916         /* Reduce ssthresh if it has not yet been made inside this window. */
1917         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1918             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1919                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1920                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1921                 tcp_ca_event(sk, CA_EVENT_LOSS);
1922         }
1923         tp->snd_cwnd       = 1;
1924         tp->snd_cwnd_cnt   = 0;
1925         tp->snd_cwnd_stamp = tcp_time_stamp;
1926
1927         tp->bytes_acked = 0;
1928         tcp_clear_retrans_partial(tp);
1929
1930         if (tcp_is_reno(tp))
1931                 tcp_reset_reno_sack(tp);
1932
1933         if (!how) {
1934                 /* Push undo marker, if it was plain RTO and nothing
1935                  * was retransmitted. */
1936                 tp->undo_marker = tp->snd_una;
1937                 tcp_clear_retrans_hints_partial(tp);
1938         } else {
1939                 tp->sacked_out = 0;
1940                 tp->fackets_out = 0;
1941                 tcp_clear_all_retrans_hints(tp);
1942         }
1943
1944         tcp_for_write_queue(skb, sk) {
1945                 if (skb == tcp_send_head(sk))
1946                         break;
1947
1948                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1949                         tp->undo_marker = 0;
1950                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1951                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1952                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1953                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1954                         tp->lost_out += tcp_skb_pcount(skb);
1955                 }
1956         }
1957         tcp_verify_left_out(tp);
1958
1959         tp->reordering = min_t(unsigned int, tp->reordering,
1960                                sysctl_tcp_reordering);
1961         tcp_set_ca_state(sk, TCP_CA_Loss);
1962         tp->high_seq = tp->snd_nxt;
1963         TCP_ECN_queue_cwr(tp);
1964         /* Abort F-RTO algorithm if one is in progress */
1965         tp->frto_counter = 0;
1966 }
1967
1968 /* If ACK arrived pointing to a remembered SACK, it means that our
1969  * remembered SACKs do not reflect real state of receiver i.e.
1970  * receiver _host_ is heavily congested (or buggy).
1971  *
1972  * Do processing similar to RTO timeout.
1973  */
1974 static int tcp_check_sack_reneging(struct sock *sk, int flag)
1975 {
1976         if (flag & FLAG_SACK_RENEGING) {
1977                 struct inet_connection_sock *icsk = inet_csk(sk);
1978                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1979
1980                 tcp_enter_loss(sk, 1);
1981                 icsk->icsk_retransmits++;
1982                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1983                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1984                                           icsk->icsk_rto, TCP_RTO_MAX);
1985                 return 1;
1986         }
1987         return 0;
1988 }
1989
1990 static inline int tcp_fackets_out(struct tcp_sock *tp)
1991 {
1992         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1993 }
1994
1995 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1996  * counter when SACK is enabled (without SACK, sacked_out is used for
1997  * that purpose).
1998  *
1999  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2000  * segments up to the highest received SACK block so far and holes in
2001  * between them.
2002  *
2003  * With reordering, holes may still be in flight, so RFC3517 recovery
2004  * uses pure sacked_out (total number of SACKed segments) even though
2005  * it violates the RFC that uses duplicate ACKs, often these are equal
2006  * but when e.g. out-of-window ACKs or packet duplication occurs,
2007  * they differ. Since neither occurs due to loss, TCP should really
2008  * ignore them.
2009  */
2010 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2011 {
2012         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2013 }
2014
2015 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2016 {
2017         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2018 }
2019
2020 static inline int tcp_head_timedout(struct sock *sk)
2021 {
2022         struct tcp_sock *tp = tcp_sk(sk);
2023
2024         return tp->packets_out &&
2025                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2026 }
2027
2028 /* Linux NewReno/SACK/FACK/ECN state machine.
2029  * --------------------------------------
2030  *
2031  * "Open"       Normal state, no dubious events, fast path.
2032  * "Disorder"   In all the respects it is "Open",
2033  *              but requires a bit more attention. It is entered when
2034  *              we see some SACKs or dupacks. It is split of "Open"
2035  *              mainly to move some processing from fast path to slow one.
2036  * "CWR"        CWND was reduced due to some Congestion Notification event.
2037  *              It can be ECN, ICMP source quench, local device congestion.
2038  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2039  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2040  *
2041  * tcp_fastretrans_alert() is entered:
2042  * - each incoming ACK, if state is not "Open"
2043  * - when arrived ACK is unusual, namely:
2044  *      * SACK
2045  *      * Duplicate ACK.
2046  *      * ECN ECE.
2047  *
2048  * Counting packets in flight is pretty simple.
2049  *
2050  *      in_flight = packets_out - left_out + retrans_out
2051  *
2052  *      packets_out is SND.NXT-SND.UNA counted in packets.
2053  *
2054  *      retrans_out is number of retransmitted segments.
2055  *
2056  *      left_out is number of segments left network, but not ACKed yet.
2057  *
2058  *              left_out = sacked_out + lost_out
2059  *
2060  *     sacked_out: Packets, which arrived to receiver out of order
2061  *                 and hence not ACKed. With SACKs this number is simply
2062  *                 amount of SACKed data. Even without SACKs
2063  *                 it is easy to give pretty reliable estimate of this number,
2064  *                 counting duplicate ACKs.
2065  *
2066  *       lost_out: Packets lost by network. TCP has no explicit
2067  *                 "loss notification" feedback from network (for now).
2068  *                 It means that this number can be only _guessed_.
2069  *                 Actually, it is the heuristics to predict lossage that
2070  *                 distinguishes different algorithms.
2071  *
2072  *      F.e. after RTO, when all the queue is considered as lost,
2073  *      lost_out = packets_out and in_flight = retrans_out.
2074  *
2075  *              Essentially, we have now two algorithms counting
2076  *              lost packets.
2077  *
2078  *              FACK: It is the simplest heuristics. As soon as we decided
2079  *              that something is lost, we decide that _all_ not SACKed
2080  *              packets until the most forward SACK are lost. I.e.
2081  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2082  *              It is absolutely correct estimate, if network does not reorder
2083  *              packets. And it loses any connection to reality when reordering
2084  *              takes place. We use FACK by default until reordering
2085  *              is suspected on the path to this destination.
2086  *
2087  *              NewReno: when Recovery is entered, we assume that one segment
2088  *              is lost (classic Reno). While we are in Recovery and
2089  *              a partial ACK arrives, we assume that one more packet
2090  *              is lost (NewReno). This heuristics are the same in NewReno
2091  *              and SACK.
2092  *
2093  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2094  *  deflation etc. CWND is real congestion window, never inflated, changes
2095  *  only according to classic VJ rules.
2096  *
2097  * Really tricky (and requiring careful tuning) part of algorithm
2098  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2099  * The first determines the moment _when_ we should reduce CWND and,
2100  * hence, slow down forward transmission. In fact, it determines the moment
2101  * when we decide that hole is caused by loss, rather than by a reorder.
2102  *
2103  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2104  * holes, caused by lost packets.
2105  *
2106  * And the most logically complicated part of algorithm is undo
2107  * heuristics. We detect false retransmits due to both too early
2108  * fast retransmit (reordering) and underestimated RTO, analyzing
2109  * timestamps and D-SACKs. When we detect that some segments were
2110  * retransmitted by mistake and CWND reduction was wrong, we undo
2111  * window reduction and abort recovery phase. This logic is hidden
2112  * inside several functions named tcp_try_undo_<something>.
2113  */
2114
2115 /* This function decides, when we should leave Disordered state
2116  * and enter Recovery phase, reducing congestion window.
2117  *
2118  * Main question: may we further continue forward transmission
2119  * with the same cwnd?
2120  */
2121 static int tcp_time_to_recover(struct sock *sk)
2122 {
2123         struct tcp_sock *tp = tcp_sk(sk);
2124         __u32 packets_out;
2125
2126         /* Do not perform any recovery during F-RTO algorithm */
2127         if (tp->frto_counter)
2128                 return 0;
2129
2130         /* Trick#1: The loss is proven. */
2131         if (tp->lost_out)
2132                 return 1;
2133
2134         /* Not-A-Trick#2 : Classic rule... */
2135         if (tcp_dupack_heurestics(tp) > tp->reordering)
2136                 return 1;
2137
2138         /* Trick#3 : when we use RFC2988 timer restart, fast
2139          * retransmit can be triggered by timeout of queue head.
2140          */
2141         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2142                 return 1;
2143
2144         /* Trick#4: It is still not OK... But will it be useful to delay
2145          * recovery more?
2146          */
2147         packets_out = tp->packets_out;
2148         if (packets_out <= tp->reordering &&
2149             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2150             !tcp_may_send_now(sk)) {
2151                 /* We have nothing to send. This connection is limited
2152                  * either by receiver window or by application.
2153                  */
2154                 return 1;
2155         }
2156
2157         return 0;
2158 }
2159
2160 /* RFC: This is from the original, I doubt that this is necessary at all:
2161  * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2162  * retransmitted past LOST markings in the first place? I'm not fully sure
2163  * about undo and end of connection cases, which can cause R without L?
2164  */
2165 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
2166 {
2167         if ((tp->retransmit_skb_hint != NULL) &&
2168             before(TCP_SKB_CB(skb)->seq,
2169                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
2170                 tp->retransmit_skb_hint = NULL;
2171 }
2172
2173 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2174  * is against sacked "cnt", otherwise it's against facked "cnt"
2175  */
2176 static void tcp_mark_head_lost(struct sock *sk, int packets)
2177 {
2178         struct tcp_sock *tp = tcp_sk(sk);
2179         struct sk_buff *skb;
2180         int cnt, oldcnt;
2181         int err;
2182         unsigned int mss;
2183
2184         WARN_ON(packets > tp->packets_out);
2185         if (tp->lost_skb_hint) {
2186                 skb = tp->lost_skb_hint;
2187                 cnt = tp->lost_cnt_hint;
2188         } else {
2189                 skb = tcp_write_queue_head(sk);
2190                 cnt = 0;
2191         }
2192
2193         tcp_for_write_queue_from(skb, sk) {
2194                 if (skb == tcp_send_head(sk))
2195                         break;
2196                 /* TODO: do this better */
2197                 /* this is not the most efficient way to do this... */
2198                 tp->lost_skb_hint = skb;
2199                 tp->lost_cnt_hint = cnt;
2200
2201                 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2202                         break;
2203
2204                 oldcnt = cnt;
2205                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2206                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2207                         cnt += tcp_skb_pcount(skb);
2208
2209                 if (cnt > packets) {
2210                         if (tcp_is_sack(tp) || (oldcnt >= packets))
2211                                 break;
2212
2213                         mss = skb_shinfo(skb)->gso_size;
2214                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2215                         if (err < 0)
2216                                 break;
2217                         cnt = packets;
2218                 }
2219
2220                 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2221                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2222                         tp->lost_out += tcp_skb_pcount(skb);
2223                         tcp_verify_retransmit_hint(tp, skb);
2224                 }
2225         }
2226         tcp_verify_left_out(tp);
2227 }
2228
2229 /* Account newly detected lost packet(s) */
2230
2231 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2232 {
2233         struct tcp_sock *tp = tcp_sk(sk);
2234
2235         if (tcp_is_reno(tp)) {
2236                 tcp_mark_head_lost(sk, 1);
2237         } else if (tcp_is_fack(tp)) {
2238                 int lost = tp->fackets_out - tp->reordering;
2239                 if (lost <= 0)
2240                         lost = 1;
2241                 tcp_mark_head_lost(sk, lost);
2242         } else {
2243                 int sacked_upto = tp->sacked_out - tp->reordering;
2244                 if (sacked_upto < fast_rexmit)
2245                         sacked_upto = fast_rexmit;
2246                 tcp_mark_head_lost(sk, sacked_upto);
2247         }
2248
2249         /* New heuristics: it is possible only after we switched
2250          * to restart timer each time when something is ACKed.
2251          * Hence, we can detect timed out packets during fast
2252          * retransmit without falling to slow start.
2253          */
2254         if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2255                 struct sk_buff *skb;
2256
2257                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2258                         : tcp_write_queue_head(sk);
2259
2260                 tcp_for_write_queue_from(skb, sk) {
2261                         if (skb == tcp_send_head(sk))
2262                                 break;
2263                         if (!tcp_skb_timedout(sk, skb))
2264                                 break;
2265
2266                         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2267                                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2268                                 tp->lost_out += tcp_skb_pcount(skb);
2269                                 tcp_verify_retransmit_hint(tp, skb);
2270                         }
2271                 }
2272
2273                 tp->scoreboard_skb_hint = skb;
2274
2275                 tcp_verify_left_out(tp);
2276         }
2277 }
2278
2279 /* CWND moderation, preventing bursts due to too big ACKs
2280  * in dubious situations.
2281  */
2282 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2283 {
2284         tp->snd_cwnd = min(tp->snd_cwnd,
2285                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2286         tp->snd_cwnd_stamp = tcp_time_stamp;
2287 }
2288
2289 /* Lower bound on congestion window is slow start threshold
2290  * unless congestion avoidance choice decides to overide it.
2291  */
2292 static inline u32 tcp_cwnd_min(const struct sock *sk)
2293 {
2294         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2295
2296         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2297 }
2298
2299 /* Decrease cwnd each second ack. */
2300 static void tcp_cwnd_down(struct sock *sk, int flag)
2301 {
2302         struct tcp_sock *tp = tcp_sk(sk);
2303         int decr = tp->snd_cwnd_cnt + 1;
2304
2305         if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2306             (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2307                 tp->snd_cwnd_cnt = decr & 1;
2308                 decr >>= 1;
2309
2310                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2311                         tp->snd_cwnd -= decr;
2312
2313                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2314                 tp->snd_cwnd_stamp = tcp_time_stamp;
2315         }
2316 }
2317
2318 /* Nothing was retransmitted or returned timestamp is less
2319  * than timestamp of the first retransmission.
2320  */
2321 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2322 {
2323         return !tp->retrans_stamp ||
2324                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2325                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2326 }
2327
2328 /* Undo procedures. */
2329
2330 #if FASTRETRANS_DEBUG > 1
2331 static void DBGUNDO(struct sock *sk, const char *msg)
2332 {
2333         struct tcp_sock *tp = tcp_sk(sk);
2334         struct inet_sock *inet = inet_sk(sk);
2335
2336         if (sk->sk_family == AF_INET) {
2337                 printk(KERN_DEBUG "Undo %s " NIPQUAD_FMT "/%u c%u l%u ss%u/%u p%u\n",
2338                        msg,
2339                        NIPQUAD(inet->daddr), ntohs(inet->dport),
2340                        tp->snd_cwnd, tcp_left_out(tp),
2341                        tp->snd_ssthresh, tp->prior_ssthresh,
2342                        tp->packets_out);
2343         }
2344 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2345         else if (sk->sk_family == AF_INET6) {
2346                 struct ipv6_pinfo *np = inet6_sk(sk);
2347                 printk(KERN_DEBUG "Undo %s " NIP6_FMT "/%u c%u l%u ss%u/%u p%u\n",
2348                        msg,
2349                        NIP6(np->daddr), ntohs(inet->dport),
2350                        tp->snd_cwnd, tcp_left_out(tp),
2351                        tp->snd_ssthresh, tp->prior_ssthresh,
2352                        tp->packets_out);
2353         }
2354 #endif
2355 }
2356 #else
2357 #define DBGUNDO(x...) do { } while (0)
2358 #endif
2359
2360 static void tcp_undo_cwr(struct sock *sk, const int undo)
2361 {
2362         struct tcp_sock *tp = tcp_sk(sk);
2363
2364         if (tp->prior_ssthresh) {
2365                 const struct inet_connection_sock *icsk = inet_csk(sk);
2366
2367                 if (icsk->icsk_ca_ops->undo_cwnd)
2368                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2369                 else
2370                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2371
2372                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2373                         tp->snd_ssthresh = tp->prior_ssthresh;
2374                         TCP_ECN_withdraw_cwr(tp);
2375                 }
2376         } else {
2377                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2378         }
2379         tcp_moderate_cwnd(tp);
2380         tp->snd_cwnd_stamp = tcp_time_stamp;
2381
2382         /* There is something screwy going on with the retrans hints after
2383            an undo */
2384         tcp_clear_all_retrans_hints(tp);
2385 }
2386
2387 static inline int tcp_may_undo(struct tcp_sock *tp)
2388 {
2389         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2390 }
2391
2392 /* People celebrate: "We love our President!" */
2393 static int tcp_try_undo_recovery(struct sock *sk)
2394 {
2395         struct tcp_sock *tp = tcp_sk(sk);
2396
2397         if (tcp_may_undo(tp)) {
2398                 int mib_idx;
2399
2400                 /* Happy end! We did not retransmit anything
2401                  * or our original transmission succeeded.
2402                  */
2403                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2404                 tcp_undo_cwr(sk, 1);
2405                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2406                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2407                 else
2408                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2409
2410                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2411                 tp->undo_marker = 0;
2412         }
2413         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2414                 /* Hold old state until something *above* high_seq
2415                  * is ACKed. For Reno it is MUST to prevent false
2416                  * fast retransmits (RFC2582). SACK TCP is safe. */
2417                 tcp_moderate_cwnd(tp);
2418                 return 1;
2419         }
2420         tcp_set_ca_state(sk, TCP_CA_Open);
2421         return 0;
2422 }
2423
2424 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2425 static void tcp_try_undo_dsack(struct sock *sk)
2426 {
2427         struct tcp_sock *tp = tcp_sk(sk);
2428
2429         if (tp->undo_marker && !tp->undo_retrans) {
2430                 DBGUNDO(sk, "D-SACK");
2431                 tcp_undo_cwr(sk, 1);
2432                 tp->undo_marker = 0;
2433                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2434         }
2435 }
2436
2437 /* Undo during fast recovery after partial ACK. */
2438
2439 static int tcp_try_undo_partial(struct sock *sk, int acked)
2440 {
2441         struct tcp_sock *tp = tcp_sk(sk);
2442         /* Partial ACK arrived. Force Hoe's retransmit. */
2443         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2444
2445         if (tcp_may_undo(tp)) {
2446                 /* Plain luck! Hole if filled with delayed
2447                  * packet, rather than with a retransmit.
2448                  */
2449                 if (tp->retrans_out == 0)
2450                         tp->retrans_stamp = 0;
2451
2452                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2453
2454                 DBGUNDO(sk, "Hoe");
2455                 tcp_undo_cwr(sk, 0);
2456                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2457
2458                 /* So... Do not make Hoe's retransmit yet.
2459                  * If the first packet was delayed, the rest
2460                  * ones are most probably delayed as well.
2461                  */
2462                 failed = 0;
2463         }
2464         return failed;
2465 }
2466
2467 /* Undo during loss recovery after partial ACK. */
2468 static int tcp_try_undo_loss(struct sock *sk)
2469 {
2470         struct tcp_sock *tp = tcp_sk(sk);
2471
2472         if (tcp_may_undo(tp)) {
2473                 struct sk_buff *skb;
2474                 tcp_for_write_queue(skb, sk) {
2475                         if (skb == tcp_send_head(sk))
2476                                 break;
2477                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2478                 }
2479
2480                 tcp_clear_all_retrans_hints(tp);
2481
2482                 DBGUNDO(sk, "partial loss");
2483                 tp->lost_out = 0;
2484                 tcp_undo_cwr(sk, 1);
2485                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2486                 inet_csk(sk)->icsk_retransmits = 0;
2487                 tp->undo_marker = 0;
2488                 if (tcp_is_sack(tp))
2489                         tcp_set_ca_state(sk, TCP_CA_Open);
2490                 return 1;
2491         }
2492         return 0;
2493 }
2494
2495 static inline void tcp_complete_cwr(struct sock *sk)
2496 {
2497         struct tcp_sock *tp = tcp_sk(sk);
2498         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2499         tp->snd_cwnd_stamp = tcp_time_stamp;
2500         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2501 }
2502
2503 static void tcp_try_keep_open(struct sock *sk)
2504 {
2505         struct tcp_sock *tp = tcp_sk(sk);
2506         int state = TCP_CA_Open;
2507
2508         if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2509                 state = TCP_CA_Disorder;
2510
2511         if (inet_csk(sk)->icsk_ca_state != state) {
2512                 tcp_set_ca_state(sk, state);
2513                 tp->high_seq = tp->snd_nxt;
2514         }
2515 }
2516
2517 static void tcp_try_to_open(struct sock *sk, int flag)
2518 {
2519         struct tcp_sock *tp = tcp_sk(sk);
2520
2521         tcp_verify_left_out(tp);
2522
2523         if (!tp->frto_counter && tp->retrans_out == 0)
2524                 tp->retrans_stamp = 0;
2525
2526         if (flag & FLAG_ECE)
2527                 tcp_enter_cwr(sk, 1);
2528
2529         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2530                 tcp_try_keep_open(sk);
2531                 tcp_moderate_cwnd(tp);
2532         } else {
2533                 tcp_cwnd_down(sk, flag);
2534         }
2535 }
2536
2537 static void tcp_mtup_probe_failed(struct sock *sk)
2538 {
2539         struct inet_connection_sock *icsk = inet_csk(sk);
2540
2541         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2542         icsk->icsk_mtup.probe_size = 0;
2543 }
2544
2545 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2546 {
2547         struct tcp_sock *tp = tcp_sk(sk);
2548         struct inet_connection_sock *icsk = inet_csk(sk);
2549
2550         /* FIXME: breaks with very large cwnd */
2551         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2552         tp->snd_cwnd = tp->snd_cwnd *
2553                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2554                        icsk->icsk_mtup.probe_size;
2555         tp->snd_cwnd_cnt = 0;
2556         tp->snd_cwnd_stamp = tcp_time_stamp;
2557         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2558
2559         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2560         icsk->icsk_mtup.probe_size = 0;
2561         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2562 }
2563
2564 /* Process an event, which can update packets-in-flight not trivially.
2565  * Main goal of this function is to calculate new estimate for left_out,
2566  * taking into account both packets sitting in receiver's buffer and
2567  * packets lost by network.
2568  *
2569  * Besides that it does CWND reduction, when packet loss is detected
2570  * and changes state of machine.
2571  *
2572  * It does _not_ decide what to send, it is made in function
2573  * tcp_xmit_retransmit_queue().
2574  */
2575 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2576 {
2577         struct inet_connection_sock *icsk = inet_csk(sk);
2578         struct tcp_sock *tp = tcp_sk(sk);
2579         int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2580         int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2581                                     (tcp_fackets_out(tp) > tp->reordering));
2582         int fast_rexmit = 0, mib_idx;
2583
2584         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2585                 tp->sacked_out = 0;
2586         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2587                 tp->fackets_out = 0;
2588
2589         /* Now state machine starts.
2590          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2591         if (flag & FLAG_ECE)
2592                 tp->prior_ssthresh = 0;
2593
2594         /* B. In all the states check for reneging SACKs. */
2595         if (tcp_check_sack_reneging(sk, flag))
2596                 return;
2597
2598         /* C. Process data loss notification, provided it is valid. */
2599         if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2600             before(tp->snd_una, tp->high_seq) &&
2601             icsk->icsk_ca_state != TCP_CA_Open &&
2602             tp->fackets_out > tp->reordering) {
2603                 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2604                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2605         }
2606
2607         /* D. Check consistency of the current state. */
2608         tcp_verify_left_out(tp);
2609
2610         /* E. Check state exit conditions. State can be terminated
2611          *    when high_seq is ACKed. */
2612         if (icsk->icsk_ca_state == TCP_CA_Open) {
2613                 WARN_ON(tp->retrans_out != 0);
2614                 tp->retrans_stamp = 0;
2615         } else if (!before(tp->snd_una, tp->high_seq)) {
2616                 switch (icsk->icsk_ca_state) {
2617                 case TCP_CA_Loss:
2618                         icsk->icsk_retransmits = 0;
2619                         if (tcp_try_undo_recovery(sk))
2620                                 return;
2621                         break;
2622
2623                 case TCP_CA_CWR:
2624                         /* CWR is to be held something *above* high_seq
2625                          * is ACKed for CWR bit to reach receiver. */
2626                         if (tp->snd_una != tp->high_seq) {
2627                                 tcp_complete_cwr(sk);
2628                                 tcp_set_ca_state(sk, TCP_CA_Open);
2629                         }
2630                         break;
2631
2632                 case TCP_CA_Disorder:
2633                         tcp_try_undo_dsack(sk);
2634                         if (!tp->undo_marker ||
2635                             /* For SACK case do not Open to allow to undo
2636                              * catching for all duplicate ACKs. */
2637                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2638                                 tp->undo_marker = 0;
2639                                 tcp_set_ca_state(sk, TCP_CA_Open);
2640                         }
2641                         break;
2642
2643                 case TCP_CA_Recovery:
2644                         if (tcp_is_reno(tp))
2645                                 tcp_reset_reno_sack(tp);
2646                         if (tcp_try_undo_recovery(sk))
2647                                 return;
2648                         tcp_complete_cwr(sk);
2649                         break;
2650                 }
2651         }
2652
2653         /* F. Process state. */
2654         switch (icsk->icsk_ca_state) {
2655         case TCP_CA_Recovery:
2656                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2657                         if (tcp_is_reno(tp) && is_dupack)
2658                                 tcp_add_reno_sack(sk);
2659                 } else
2660                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2661                 break;
2662         case TCP_CA_Loss:
2663                 if (flag & FLAG_DATA_ACKED)
2664                         icsk->icsk_retransmits = 0;
2665                 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2666                         tcp_reset_reno_sack(tp);
2667                 if (!tcp_try_undo_loss(sk)) {
2668                         tcp_moderate_cwnd(tp);
2669                         tcp_xmit_retransmit_queue(sk);
2670                         return;
2671                 }
2672                 if (icsk->icsk_ca_state != TCP_CA_Open)
2673                         return;
2674                 /* Loss is undone; fall through to processing in Open state. */
2675         default:
2676                 if (tcp_is_reno(tp)) {
2677                         if (flag & FLAG_SND_UNA_ADVANCED)
2678                                 tcp_reset_reno_sack(tp);
2679                         if (is_dupack)
2680                                 tcp_add_reno_sack(sk);
2681                 }
2682
2683                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2684                         tcp_try_undo_dsack(sk);
2685
2686                 if (!tcp_time_to_recover(sk)) {
2687                         tcp_try_to_open(sk, flag);
2688                         return;
2689                 }
2690
2691                 /* MTU probe failure: don't reduce cwnd */
2692                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2693                     icsk->icsk_mtup.probe_size &&
2694                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2695                         tcp_mtup_probe_failed(sk);
2696                         /* Restores the reduction we did in tcp_mtup_probe() */
2697                         tp->snd_cwnd++;
2698                         tcp_simple_retransmit(sk);
2699                         return;
2700                 }
2701
2702                 /* Otherwise enter Recovery state */
2703
2704                 if (tcp_is_reno(tp))
2705                         mib_idx = LINUX_MIB_TCPRENORECOVERY;
2706                 else
2707                         mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2708
2709                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2710
2711                 tp->high_seq = tp->snd_nxt;
2712                 tp->prior_ssthresh = 0;
2713                 tp->undo_marker = tp->snd_una;
2714                 tp->undo_retrans = tp->retrans_out;
2715
2716                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2717                         if (!(flag & FLAG_ECE))
2718                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2719                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2720                         TCP_ECN_queue_cwr(tp);
2721                 }
2722
2723                 tp->bytes_acked = 0;
2724                 tp->snd_cwnd_cnt = 0;
2725                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2726                 fast_rexmit = 1;
2727         }
2728
2729         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2730                 tcp_update_scoreboard(sk, fast_rexmit);
2731         tcp_cwnd_down(sk, flag);
2732         tcp_xmit_retransmit_queue(sk);
2733 }
2734
2735 /* Read draft-ietf-tcplw-high-performance before mucking
2736  * with this code. (Supersedes RFC1323)
2737  */
2738 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2739 {
2740         /* RTTM Rule: A TSecr value received in a segment is used to
2741          * update the averaged RTT measurement only if the segment
2742          * acknowledges some new data, i.e., only if it advances the
2743          * left edge of the send window.
2744          *
2745          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2746          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2747          *
2748          * Changed: reset backoff as soon as we see the first valid sample.
2749          * If we do not, we get strongly overestimated rto. With timestamps
2750          * samples are accepted even from very old segments: f.e., when rtt=1
2751          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2752          * answer arrives rto becomes 120 seconds! If at least one of segments
2753          * in window is lost... Voila.                          --ANK (010210)
2754          */
2755         struct tcp_sock *tp = tcp_sk(sk);
2756         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2757         tcp_rtt_estimator(sk, seq_rtt);
2758         tcp_set_rto(sk);
2759         inet_csk(sk)->icsk_backoff = 0;
2760         tcp_bound_rto(sk);
2761 }
2762
2763 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2764 {
2765         /* We don't have a timestamp. Can only use
2766          * packets that are not retransmitted to determine
2767          * rtt estimates. Also, we must not reset the
2768          * backoff for rto until we get a non-retransmitted
2769          * packet. This allows us to deal with a situation
2770          * where the network delay has increased suddenly.
2771          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2772          */
2773
2774         if (flag & FLAG_RETRANS_DATA_ACKED)
2775                 return;
2776
2777         tcp_rtt_estimator(sk, seq_rtt);
2778         tcp_set_rto(sk);
2779         inet_csk(sk)->icsk_backoff = 0;
2780         tcp_bound_rto(sk);
2781 }
2782
2783 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2784                                       const s32 seq_rtt)
2785 {
2786         const struct tcp_sock *tp = tcp_sk(sk);
2787         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2788         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2789                 tcp_ack_saw_tstamp(sk, flag);
2790         else if (seq_rtt >= 0)
2791                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2792 }
2793
2794 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2795 {
2796         const struct inet_connection_sock *icsk = inet_csk(sk);
2797         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2798         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2799 }
2800
2801 /* Restart timer after forward progress on connection.
2802  * RFC2988 recommends to restart timer to now+rto.
2803  */
2804 static void tcp_rearm_rto(struct sock *sk)
2805 {
2806         struct tcp_sock *tp = tcp_sk(sk);
2807
2808         if (!tp->packets_out) {
2809                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2810         } else {
2811                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2812                                           inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2813         }
2814 }
2815
2816 /* If we get here, the whole TSO packet has not been acked. */
2817 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2818 {
2819         struct tcp_sock *tp = tcp_sk(sk);
2820         u32 packets_acked;
2821
2822         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2823
2824         packets_acked = tcp_skb_pcount(skb);
2825         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2826                 return 0;
2827         packets_acked -= tcp_skb_pcount(skb);
2828
2829         if (packets_acked) {
2830                 BUG_ON(tcp_skb_pcount(skb) == 0);
2831                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2832         }
2833
2834         return packets_acked;
2835 }
2836
2837 /* Remove acknowledged frames from the retransmission queue. If our packet
2838  * is before the ack sequence we can discard it as it's confirmed to have
2839  * arrived at the other end.
2840  */
2841 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
2842 {
2843         struct tcp_sock *tp = tcp_sk(sk);
2844         const struct inet_connection_sock *icsk = inet_csk(sk);
2845         struct sk_buff *skb;
2846         u32 now = tcp_time_stamp;
2847         int fully_acked = 1;
2848         int flag = 0;
2849         u32 pkts_acked = 0;
2850         u32 reord = tp->packets_out;
2851         s32 seq_rtt = -1;
2852         s32 ca_seq_rtt = -1;
2853         ktime_t last_ackt = net_invalid_timestamp();
2854
2855         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2856                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2857                 u32 end_seq;
2858                 u32 acked_pcount;
2859                 u8 sacked = scb->sacked;
2860
2861                 /* Determine how many packets and what bytes were acked, tso and else */
2862                 if (after(scb->end_seq, tp->snd_una)) {
2863                         if (tcp_skb_pcount(skb) == 1 ||
2864                             !after(tp->snd_una, scb->seq))
2865                                 break;
2866
2867                         acked_pcount = tcp_tso_acked(sk, skb);
2868                         if (!acked_pcount)
2869                                 break;
2870
2871                         fully_acked = 0;
2872                         end_seq = tp->snd_una;
2873                 } else {
2874                         acked_pcount = tcp_skb_pcount(skb);
2875                         end_seq = scb->end_seq;
2876                 }
2877
2878                 /* MTU probing checks */
2879                 if (fully_acked && icsk->icsk_mtup.probe_size &&
2880                     !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2881                         tcp_mtup_probe_success(sk, skb);
2882                 }
2883
2884                 if (sacked & TCPCB_RETRANS) {
2885                         if (sacked & TCPCB_SACKED_RETRANS)
2886                                 tp->retrans_out -= acked_pcount;
2887                         flag |= FLAG_RETRANS_DATA_ACKED;
2888                         ca_seq_rtt = -1;
2889                         seq_rtt = -1;
2890                         if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
2891                                 flag |= FLAG_NONHEAD_RETRANS_ACKED;
2892                 } else {
2893                         ca_seq_rtt = now - scb->when;
2894                         last_ackt = skb->tstamp;
2895                         if (seq_rtt < 0) {
2896                                 seq_rtt = ca_seq_rtt;
2897                         }
2898                         if (!(sacked & TCPCB_SACKED_ACKED))
2899                                 reord = min(pkts_acked, reord);
2900                 }
2901
2902                 if (sacked & TCPCB_SACKED_ACKED)
2903                         tp->sacked_out -= acked_pcount;
2904                 if (sacked & TCPCB_LOST)
2905                         tp->lost_out -= acked_pcount;
2906
2907                 if (unlikely(tp->urg_mode && !before(end_seq, tp->snd_up)))
2908                         tp->urg_mode = 0;
2909
2910                 tp->packets_out -= acked_pcount;
2911                 pkts_acked += acked_pcount;
2912
2913                 /* Initial outgoing SYN's get put onto the write_queue
2914                  * just like anything else we transmit.  It is not
2915                  * true data, and if we misinform our callers that
2916                  * this ACK acks real data, we will erroneously exit
2917                  * connection startup slow start one packet too
2918                  * quickly.  This is severely frowned upon behavior.
2919                  */
2920                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2921                         flag |= FLAG_DATA_ACKED;
2922                 } else {
2923                         flag |= FLAG_SYN_ACKED;
2924                         tp->retrans_stamp = 0;
2925                 }
2926
2927                 if (!fully_acked)
2928                         break;
2929
2930                 tcp_unlink_write_queue(skb, sk);
2931                 sk_wmem_free_skb(sk, skb);
2932                 tcp_clear_all_retrans_hints(tp);
2933         }
2934
2935         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2936                 flag |= FLAG_SACK_RENEGING;
2937
2938         if (flag & FLAG_ACKED) {
2939                 const struct tcp_congestion_ops *ca_ops
2940                         = inet_csk(sk)->icsk_ca_ops;
2941
2942                 tcp_ack_update_rtt(sk, flag, seq_rtt);
2943                 tcp_rearm_rto(sk);
2944
2945                 if (tcp_is_reno(tp)) {
2946                         tcp_remove_reno_sacks(sk, pkts_acked);
2947                 } else {
2948                         /* Non-retransmitted hole got filled? That's reordering */
2949                         if (reord < prior_fackets)
2950                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2951                 }
2952
2953                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2954
2955                 if (ca_ops->pkts_acked) {
2956                         s32 rtt_us = -1;
2957
2958                         /* Is the ACK triggering packet unambiguous? */
2959                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2960                                 /* High resolution needed and available? */
2961                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2962                                     !ktime_equal(last_ackt,
2963                                                  net_invalid_timestamp()))
2964                                         rtt_us = ktime_us_delta(ktime_get_real(),
2965                                                                 last_ackt);
2966                                 else if (ca_seq_rtt > 0)
2967                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
2968                         }
2969
2970                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2971                 }
2972         }
2973
2974 #if FASTRETRANS_DEBUG > 0
2975         WARN_ON((int)tp->sacked_out < 0);
2976         WARN_ON((int)tp->lost_out < 0);
2977         WARN_ON((int)tp->retrans_out < 0);
2978         if (!tp->packets_out && tcp_is_sack(tp)) {
2979                 icsk = inet_csk(sk);
2980                 if (tp->lost_out) {
2981                         printk(KERN_DEBUG "Leak l=%u %d\n",
2982                                tp->lost_out, icsk->icsk_ca_state);
2983                         tp->lost_out = 0;
2984                 }
2985                 if (tp->sacked_out) {
2986                         printk(KERN_DEBUG "Leak s=%u %d\n",
2987                                tp->sacked_out, icsk->icsk_ca_state);
2988                         tp->sacked_out = 0;
2989                 }
2990                 if (tp->retrans_out) {
2991                         printk(KERN_DEBUG "Leak r=%u %d\n",
2992                                tp->retrans_out, icsk->icsk_ca_state);
2993                         tp->retrans_out = 0;
2994                 }
2995         }
2996 #endif
2997         return flag;
2998 }
2999
3000 static void tcp_ack_probe(struct sock *sk)
3001 {
3002         const struct tcp_sock *tp = tcp_sk(sk);
3003         struct inet_connection_sock *icsk = inet_csk(sk);
3004
3005         /* Was it a usable window open? */
3006
3007         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3008                 icsk->icsk_backoff = 0;
3009                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3010                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3011                  * This function is not for random using!
3012                  */
3013         } else {
3014                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3015                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3016                                           TCP_RTO_MAX);
3017         }
3018 }
3019
3020 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3021 {
3022         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3023                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3024 }
3025
3026 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3027 {
3028         const struct tcp_sock *tp = tcp_sk(sk);
3029         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3030                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3031 }
3032
3033 /* Check that window update is acceptable.
3034  * The function assumes that snd_una<=ack<=snd_next.
3035  */
3036 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3037                                         const u32 ack, const u32 ack_seq,
3038                                         const u32 nwin)
3039 {
3040         return (after(ack, tp->snd_una) ||
3041                 after(ack_seq, tp->snd_wl1) ||
3042                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3043 }
3044
3045 /* Update our send window.
3046  *
3047  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3048  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3049  */
3050 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3051                                  u32 ack_seq)
3052 {
3053         struct tcp_sock *tp = tcp_sk(sk);
3054         int flag = 0;
3055         u32 nwin = ntohs(tcp_hdr(skb)->window);
3056
3057         if (likely(!tcp_hdr(skb)->syn))
3058                 nwin <<= tp->rx_opt.snd_wscale;
3059
3060         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3061                 flag |= FLAG_WIN_UPDATE;
3062                 tcp_update_wl(tp, ack, ack_seq);
3063
3064                 if (tp->snd_wnd != nwin) {
3065                         tp->snd_wnd = nwin;
3066
3067                         /* Note, it is the only place, where
3068                          * fast path is recovered for sending TCP.
3069                          */
3070                         tp->pred_flags = 0;
3071                         tcp_fast_path_check(sk);
3072
3073                         if (nwin > tp->max_window) {
3074                                 tp->max_window = nwin;
3075                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3076                         }
3077                 }
3078         }
3079
3080         tp->snd_una = ack;
3081
3082         return flag;
3083 }
3084
3085 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3086  * continue in congestion avoidance.
3087  */
3088 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3089 {
3090         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3091         tp->snd_cwnd_cnt = 0;
3092         tp->bytes_acked = 0;
3093         TCP_ECN_queue_cwr(tp);
3094         tcp_moderate_cwnd(tp);
3095 }
3096
3097 /* A conservative spurious RTO response algorithm: reduce cwnd using
3098  * rate halving and continue in congestion avoidance.
3099  */
3100 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3101 {
3102         tcp_enter_cwr(sk, 0);
3103 }
3104
3105 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3106 {
3107         if (flag & FLAG_ECE)
3108                 tcp_ratehalving_spur_to_response(sk);
3109         else
3110                 tcp_undo_cwr(sk, 1);
3111 }
3112
3113 /* F-RTO spurious RTO detection algorithm (RFC4138)
3114  *
3115  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3116  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3117  * window (but not to or beyond highest sequence sent before RTO):
3118  *   On First ACK,  send two new segments out.
3119  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3120  *                  algorithm is not part of the F-RTO detection algorithm
3121  *                  given in RFC4138 but can be selected separately).
3122  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3123  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3124  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3125  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3126  *
3127  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3128  * original window even after we transmit two new data segments.
3129  *
3130  * SACK version:
3131  *   on first step, wait until first cumulative ACK arrives, then move to
3132  *   the second step. In second step, the next ACK decides.
3133  *
3134  * F-RTO is implemented (mainly) in four functions:
3135  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3136  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3137  *     called when tcp_use_frto() showed green light
3138  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3139  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3140  *     to prove that the RTO is indeed spurious. It transfers the control
3141  *     from F-RTO to the conventional RTO recovery
3142  */
3143 static int tcp_process_frto(struct sock *sk, int flag)
3144 {
3145         struct tcp_sock *tp = tcp_sk(sk);
3146
3147         tcp_verify_left_out(tp);
3148
3149         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3150         if (flag & FLAG_DATA_ACKED)
3151                 inet_csk(sk)->icsk_retransmits = 0;
3152
3153         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3154             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3155                 tp->undo_marker = 0;
3156
3157         if (!before(tp->snd_una, tp->frto_highmark)) {
3158                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3159                 return 1;
3160         }
3161
3162         if (!tcp_is_sackfrto(tp)) {
3163                 /* RFC4138 shortcoming in step 2; should also have case c):
3164                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3165                  * data, winupdate
3166                  */
3167                 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3168                         return 1;
3169
3170                 if (!(flag & FLAG_DATA_ACKED)) {
3171                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3172                                             flag);
3173                         return 1;
3174                 }
3175         } else {
3176                 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3177                         /* Prevent sending of new data. */
3178                         tp->snd_cwnd = min(tp->snd_cwnd,
3179                                            tcp_packets_in_flight(tp));
3180                         return 1;
3181                 }
3182
3183                 if ((tp->frto_counter >= 2) &&
3184                     (!(flag & FLAG_FORWARD_PROGRESS) ||
3185                      ((flag & FLAG_DATA_SACKED) &&
3186                       !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3187                         /* RFC4138 shortcoming (see comment above) */
3188                         if (!(flag & FLAG_FORWARD_PROGRESS) &&
3189                             (flag & FLAG_NOT_DUP))
3190                                 return 1;
3191
3192                         tcp_enter_frto_loss(sk, 3, flag);
3193                         return 1;
3194                 }
3195         }
3196
3197         if (tp->frto_counter == 1) {
3198                 /* tcp_may_send_now needs to see updated state */
3199                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3200                 tp->frto_counter = 2;
3201
3202                 if (!tcp_may_send_now(sk))
3203                         tcp_enter_frto_loss(sk, 2, flag);
3204
3205                 return 1;
3206         } else {
3207                 switch (sysctl_tcp_frto_response) {
3208                 case 2:
3209                         tcp_undo_spur_to_response(sk, flag);
3210                         break;
3211                 case 1:
3212                         tcp_conservative_spur_to_response(tp);
3213                         break;
3214                 default:
3215                         tcp_ratehalving_spur_to_response(sk);
3216                         break;
3217                 }
3218                 tp->frto_counter = 0;
3219                 tp->undo_marker = 0;
3220                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3221         }
3222         return 0;
3223 }
3224
3225 /* This routine deals with incoming acks, but not outgoing ones. */
3226 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3227 {
3228         struct inet_connection_sock *icsk = inet_csk(sk);
3229         struct tcp_sock *tp = tcp_sk(sk);
3230         u32 prior_snd_una = tp->snd_una;
3231         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3232         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3233         u32 prior_in_flight;
3234         u32 prior_fackets;
3235         int prior_packets;
3236         int frto_cwnd = 0;
3237
3238         /* If the ack is newer than sent or older than previous acks
3239          * then we can probably ignore it.
3240          */
3241         if (after(ack, tp->snd_nxt))
3242                 goto uninteresting_ack;
3243
3244         if (before(ack, prior_snd_una))
3245                 goto old_ack;
3246
3247         if (after(ack, prior_snd_una))
3248                 flag |= FLAG_SND_UNA_ADVANCED;
3249
3250         if (sysctl_tcp_abc) {
3251                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3252                         tp->bytes_acked += ack - prior_snd_una;
3253                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3254                         /* we assume just one segment left network */
3255                         tp->bytes_acked += min(ack - prior_snd_una,
3256                                                tp->mss_cache);
3257         }
3258
3259         prior_fackets = tp->fackets_out;
3260         prior_in_flight = tcp_packets_in_flight(tp);
3261
3262         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3263                 /* Window is constant, pure forward advance.
3264                  * No more checks are required.
3265                  * Note, we use the fact that SND.UNA>=SND.WL2.
3266                  */
3267                 tcp_update_wl(tp, ack, ack_seq);
3268                 tp->snd_una = ack;
3269                 flag |= FLAG_WIN_UPDATE;
3270
3271                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3272
3273                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3274         } else {
3275                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3276                         flag |= FLAG_DATA;
3277                 else
3278                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3279
3280                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3281
3282                 if (TCP_SKB_CB(skb)->sacked)
3283                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3284
3285                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3286                         flag |= FLAG_ECE;
3287
3288                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3289         }
3290
3291         /* We passed data and got it acked, remove any soft error
3292          * log. Something worked...
3293          */
3294         sk->sk_err_soft = 0;
3295         icsk->icsk_probes_out = 0;
3296         tp->rcv_tstamp = tcp_time_stamp;
3297         prior_packets = tp->packets_out;
3298         if (!prior_packets)
3299                 goto no_queue;
3300
3301         /* See if we can take anything off of the retransmit queue. */
3302         flag |= tcp_clean_rtx_queue(sk, prior_fackets);
3303
3304         if (tp->frto_counter)
3305                 frto_cwnd = tcp_process_frto(sk, flag);
3306         /* Guarantee sacktag reordering detection against wrap-arounds */
3307         if (before(tp->frto_highmark, tp->snd_una))
3308                 tp->frto_highmark = 0;
3309
3310         if (tcp_ack_is_dubious(sk, flag)) {
3311                 /* Advance CWND, if state allows this. */
3312                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3313                     tcp_may_raise_cwnd(sk, flag))
3314                         tcp_cong_avoid(sk, ack, prior_in_flight);
3315                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3316                                       flag);
3317         } else {
3318                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3319                         tcp_cong_avoid(sk, ack, prior_in_flight);
3320         }
3321
3322         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3323                 dst_confirm(sk->sk_dst_cache);
3324
3325         return 1;
3326
3327 no_queue:
3328         /* If this ack opens up a zero window, clear backoff.  It was
3329          * being used to time the probes, and is probably far higher than
3330          * it needs to be for normal retransmission.
3331          */
3332         if (tcp_send_head(sk))
3333                 tcp_ack_probe(sk);
3334         return 1;
3335
3336 old_ack:
3337         if (TCP_SKB_CB(skb)->sacked) {
3338                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3339                 if (icsk->icsk_ca_state == TCP_CA_Open)
3340                         tcp_try_keep_open(sk);
3341         }
3342
3343 uninteresting_ack:
3344         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3345         return 0;
3346 }
3347
3348 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3349  * But, this can also be called on packets in the established flow when
3350  * the fast version below fails.
3351  */
3352 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3353                        int estab)
3354 {
3355         unsigned char *ptr;
3356         struct tcphdr *th = tcp_hdr(skb);
3357         int length = (th->doff * 4) - sizeof(struct tcphdr);
3358
3359         ptr = (unsigned char *)(th + 1);
3360         opt_rx->saw_tstamp = 0;
3361
3362         while (length > 0) {
3363                 int opcode = *ptr++;
3364                 int opsize;
3365
3366                 switch (opcode) {
3367                 case TCPOPT_EOL:
3368                         return;
3369                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3370                         length--;
3371                         continue;
3372                 default:
3373                         opsize = *ptr++;
3374                         if (opsize < 2) /* "silly options" */
3375                                 return;
3376                         if (opsize > length)
3377                                 return; /* don't parse partial options */
3378                         switch (opcode) {
3379                         case TCPOPT_MSS:
3380                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3381                                         u16 in_mss = get_unaligned_be16(ptr);
3382                                         if (in_mss) {
3383                                                 if (opt_rx->user_mss &&
3384                                                     opt_rx->user_mss < in_mss)
3385                                                         in_mss = opt_rx->user_mss;
3386                                                 opt_rx->mss_clamp = in_mss;
3387                                         }
3388                                 }
3389                                 break;
3390                         case TCPOPT_WINDOW:
3391                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3392                                     !estab && sysctl_tcp_window_scaling) {
3393                                         __u8 snd_wscale = *(__u8 *)ptr;
3394                                         opt_rx->wscale_ok = 1;
3395                                         if (snd_wscale > 14) {
3396                                                 if (net_ratelimit())
3397                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3398                                                                "scaling value %d >14 received.\n",
3399                                                                snd_wscale);
3400                                                 snd_wscale = 14;
3401                                         }
3402                                         opt_rx->snd_wscale = snd_wscale;
3403                                 }
3404                                 break;
3405                         case TCPOPT_TIMESTAMP:
3406                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3407                                     ((estab && opt_rx->tstamp_ok) ||
3408                                      (!estab && sysctl_tcp_timestamps))) {
3409                                         opt_rx->saw_tstamp = 1;
3410                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3411                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3412                                 }
3413                                 break;
3414                         case TCPOPT_SACK_PERM:
3415                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3416                                     !estab && sysctl_tcp_sack) {
3417                                         opt_rx->sack_ok = 1;
3418                                         tcp_sack_reset(opt_rx);
3419                                 }
3420                                 break;
3421
3422                         case TCPOPT_SACK:
3423                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3424                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3425                                    opt_rx->sack_ok) {
3426                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3427                                 }
3428                                 break;
3429 #ifdef CONFIG_TCP_MD5SIG
3430                         case TCPOPT_MD5SIG:
3431                                 /*
3432                                  * The MD5 Hash has already been
3433                                  * checked (see tcp_v{4,6}_do_rcv()).
3434                                  */
3435                                 break;
3436 #endif
3437                         }
3438
3439                         ptr += opsize-2;
3440                         length -= opsize;
3441                 }
3442         }
3443 }
3444
3445 /* Fast parse options. This hopes to only see timestamps.
3446  * If it is wrong it falls back on tcp_parse_options().
3447  */
3448 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3449                                   struct tcp_sock *tp)
3450 {
3451         if (th->doff == sizeof(struct tcphdr) >> 2) {
3452                 tp->rx_opt.saw_tstamp = 0;
3453                 return 0;
3454         } else if (tp->rx_opt.tstamp_ok &&
3455                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3456                 __be32 *ptr = (__be32 *)(th + 1);
3457                 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3458                                   | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3459                         tp->rx_opt.saw_tstamp = 1;
3460                         ++ptr;
3461                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
3462                         ++ptr;
3463                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3464                         return 1;
3465                 }
3466         }
3467         tcp_parse_options(skb, &tp->rx_opt, 1);
3468         return 1;
3469 }
3470
3471 #ifdef CONFIG_TCP_MD5SIG
3472 /*
3473  * Parse MD5 Signature option
3474  */
3475 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3476 {
3477         int length = (th->doff << 2) - sizeof (*th);
3478         u8 *ptr = (u8*)(th + 1);
3479
3480         /* If the TCP option is too short, we can short cut */
3481         if (length < TCPOLEN_MD5SIG)
3482                 return NULL;
3483
3484         while (length > 0) {
3485                 int opcode = *ptr++;
3486                 int opsize;
3487
3488                 switch(opcode) {
3489                 case TCPOPT_EOL:
3490                         return NULL;
3491                 case TCPOPT_NOP:
3492                         length--;
3493                         continue;
3494                 default:
3495                         opsize = *ptr++;
3496                         if (opsize < 2 || opsize > length)
3497                                 return NULL;
3498                         if (opcode == TCPOPT_MD5SIG)
3499                                 return ptr;
3500                 }
3501                 ptr += opsize - 2;
3502                 length -= opsize;
3503         }
3504         return NULL;
3505 }
3506 #endif
3507
3508 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3509 {
3510         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3511         tp->rx_opt.ts_recent_stamp = get_seconds();
3512 }
3513
3514 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3515 {
3516         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3517                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3518                  * extra check below makes sure this can only happen
3519                  * for pure ACK frames.  -DaveM
3520                  *
3521                  * Not only, also it occurs for expired timestamps.
3522                  */
3523
3524                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3525                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3526                         tcp_store_ts_recent(tp);
3527         }
3528 }
3529
3530 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3531  *
3532  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3533  * it can pass through stack. So, the following predicate verifies that
3534  * this segment is not used for anything but congestion avoidance or
3535  * fast retransmit. Moreover, we even are able to eliminate most of such
3536  * second order effects, if we apply some small "replay" window (~RTO)
3537  * to timestamp space.
3538  *
3539  * All these measures still do not guarantee that we reject wrapped ACKs
3540  * on networks with high bandwidth, when sequence space is recycled fastly,
3541  * but it guarantees that such events will be very rare and do not affect
3542  * connection seriously. This doesn't look nice, but alas, PAWS is really
3543  * buggy extension.
3544  *
3545  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3546  * states that events when retransmit arrives after original data are rare.
3547  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3548  * the biggest problem on large power networks even with minor reordering.
3549  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3550  * up to bandwidth of 18Gigabit/sec. 8) ]
3551  */
3552
3553 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3554 {
3555         struct tcp_sock *tp = tcp_sk(sk);
3556         struct tcphdr *th = tcp_hdr(skb);
3557         u32 seq = TCP_SKB_CB(skb)->seq;
3558         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3559
3560         return (/* 1. Pure ACK with correct sequence number. */
3561                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3562
3563                 /* 2. ... and duplicate ACK. */
3564                 ack == tp->snd_una &&
3565
3566                 /* 3. ... and does not update window. */
3567                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3568
3569                 /* 4. ... and sits in replay window. */
3570                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3571 }
3572
3573 static inline int tcp_paws_discard(const struct sock *sk,
3574                                    const struct sk_buff *skb)
3575 {
3576         const struct tcp_sock *tp = tcp_sk(sk);
3577         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3578                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3579                 !tcp_disordered_ack(sk, skb));
3580 }
3581
3582 /* Check segment sequence number for validity.
3583  *
3584  * Segment controls are considered valid, if the segment
3585  * fits to the window after truncation to the window. Acceptability
3586  * of data (and SYN, FIN, of course) is checked separately.
3587  * See tcp_data_queue(), for example.
3588  *
3589  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3590  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3591  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3592  * (borrowed from freebsd)
3593  */
3594
3595 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3596 {
3597         return  !before(end_seq, tp->rcv_wup) &&
3598                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3599 }
3600
3601 /* When we get a reset we do this. */
3602 static void tcp_reset(struct sock *sk)
3603 {
3604         /* We want the right error as BSD sees it (and indeed as we do). */
3605         switch (sk->sk_state) {
3606         case TCP_SYN_SENT:
3607                 sk->sk_err = ECONNREFUSED;
3608                 break;
3609         case TCP_CLOSE_WAIT:
3610                 sk->sk_err = EPIPE;
3611                 break;
3612         case TCP_CLOSE:
3613                 return;
3614         default:
3615                 sk->sk_err = ECONNRESET;
3616         }
3617
3618         if (!sock_flag(sk, SOCK_DEAD))
3619                 sk->sk_error_report(sk);
3620
3621         tcp_done(sk);
3622 }
3623
3624 /*
3625  *      Process the FIN bit. This now behaves as it is supposed to work
3626  *      and the FIN takes effect when it is validly part of sequence
3627  *      space. Not before when we get holes.
3628  *
3629  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3630  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3631  *      TIME-WAIT)
3632  *
3633  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3634  *      close and we go into CLOSING (and later onto TIME-WAIT)
3635  *
3636  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3637  */
3638 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3639 {
3640         struct tcp_sock *tp = tcp_sk(sk);
3641
3642         inet_csk_schedule_ack(sk);
3643
3644         sk->sk_shutdown |= RCV_SHUTDOWN;
3645         sock_set_flag(sk, SOCK_DONE);
3646
3647         switch (sk->sk_state) {
3648         case TCP_SYN_RECV:
3649         case TCP_ESTABLISHED:
3650                 /* Move to CLOSE_WAIT */
3651                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3652                 inet_csk(sk)->icsk_ack.pingpong = 1;
3653                 break;
3654
3655         case TCP_CLOSE_WAIT:
3656         case TCP_CLOSING:
3657                 /* Received a retransmission of the FIN, do
3658                  * nothing.
3659                  */
3660                 break;
3661         case TCP_LAST_ACK:
3662                 /* RFC793: Remain in the LAST-ACK state. */
3663                 break;
3664
3665         case TCP_FIN_WAIT1:
3666                 /* This case occurs when a simultaneous close
3667                  * happens, we must ack the received FIN and
3668                  * enter the CLOSING state.
3669                  */
3670                 tcp_send_ack(sk);
3671                 tcp_set_state(sk, TCP_CLOSING);
3672                 break;
3673         case TCP_FIN_WAIT2:
3674                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3675                 tcp_send_ack(sk);
3676                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3677                 break;
3678         default:
3679                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3680                  * cases we should never reach this piece of code.
3681                  */
3682                 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3683                        __func__, sk->sk_state);
3684                 break;
3685         }
3686
3687         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3688          * Probably, we should reset in this case. For now drop them.
3689          */
3690         __skb_queue_purge(&tp->out_of_order_queue);
3691         if (tcp_is_sack(tp))
3692                 tcp_sack_reset(&tp->rx_opt);
3693         sk_mem_reclaim(sk);
3694
3695         if (!sock_flag(sk, SOCK_DEAD)) {
3696                 sk->sk_state_change(sk);
3697
3698                 /* Do not send POLL_HUP for half duplex close. */
3699                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3700                     sk->sk_state == TCP_CLOSE)
3701                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3702                 else
3703                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3704         }
3705 }
3706
3707 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3708                                   u32 end_seq)
3709 {
3710         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3711                 if (before(seq, sp->start_seq))
3712                         sp->start_seq = seq;
3713                 if (after(end_seq, sp->end_seq))
3714                         sp->end_seq = end_seq;
3715                 return 1;
3716         }
3717         return 0;
3718 }
3719
3720 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3721 {
3722         struct tcp_sock *tp = tcp_sk(sk);
3723
3724         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3725                 int mib_idx;
3726
3727                 if (before(seq, tp->rcv_nxt))
3728                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3729                 else
3730                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3731
3732                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3733
3734                 tp->rx_opt.dsack = 1;
3735                 tp->duplicate_sack[0].start_seq = seq;
3736                 tp->duplicate_sack[0].end_seq = end_seq;
3737                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 1;
3738         }
3739 }
3740
3741 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3742 {
3743         struct tcp_sock *tp = tcp_sk(sk);
3744
3745         if (!tp->rx_opt.dsack)
3746                 tcp_dsack_set(sk, seq, end_seq);
3747         else
3748                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3749 }
3750
3751 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3752 {
3753         struct tcp_sock *tp = tcp_sk(sk);
3754
3755         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3756             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3757                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3758                 tcp_enter_quickack_mode(sk);
3759
3760                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3761                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3762
3763                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3764                                 end_seq = tp->rcv_nxt;
3765                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3766                 }
3767         }
3768
3769         tcp_send_ack(sk);
3770 }
3771
3772 /* These routines update the SACK block as out-of-order packets arrive or
3773  * in-order packets close up the sequence space.
3774  */
3775 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3776 {
3777         int this_sack;
3778         struct tcp_sack_block *sp = &tp->selective_acks[0];
3779         struct tcp_sack_block *swalk = sp + 1;
3780
3781         /* See if the recent change to the first SACK eats into
3782          * or hits the sequence space of other SACK blocks, if so coalesce.
3783          */
3784         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3785                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3786                         int i;
3787
3788                         /* Zap SWALK, by moving every further SACK up by one slot.
3789                          * Decrease num_sacks.
3790                          */
3791                         tp->rx_opt.num_sacks--;
3792                         tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
3793                                                tp->rx_opt.dsack;
3794                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3795                                 sp[i] = sp[i + 1];
3796                         continue;
3797                 }
3798                 this_sack++, swalk++;
3799         }
3800 }
3801
3802 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
3803                                  struct tcp_sack_block *sack2)
3804 {
3805         __u32 tmp;
3806
3807         tmp = sack1->start_seq;
3808         sack1->start_seq = sack2->start_seq;
3809         sack2->start_seq = tmp;
3810
3811         tmp = sack1->end_seq;
3812         sack1->end_seq = sack2->end_seq;
3813         sack2->end_seq = tmp;
3814 }
3815
3816 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3817 {
3818         struct tcp_sock *tp = tcp_sk(sk);
3819         struct tcp_sack_block *sp = &tp->selective_acks[0];
3820         int cur_sacks = tp->rx_opt.num_sacks;
3821         int this_sack;
3822
3823         if (!cur_sacks)
3824                 goto new_sack;
3825
3826         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3827                 if (tcp_sack_extend(sp, seq, end_seq)) {
3828                         /* Rotate this_sack to the first one. */
3829                         for (; this_sack > 0; this_sack--, sp--)
3830                                 tcp_sack_swap(sp, sp - 1);
3831                         if (cur_sacks > 1)
3832                                 tcp_sack_maybe_coalesce(tp);
3833                         return;
3834                 }
3835         }
3836
3837         /* Could not find an adjacent existing SACK, build a new one,
3838          * put it at the front, and shift everyone else down.  We
3839          * always know there is at least one SACK present already here.
3840          *
3841          * If the sack array is full, forget about the last one.
3842          */
3843         if (this_sack >= TCP_NUM_SACKS) {
3844                 this_sack--;
3845                 tp->rx_opt.num_sacks--;
3846                 sp--;
3847         }
3848         for (; this_sack > 0; this_sack--, sp--)
3849                 *sp = *(sp - 1);
3850
3851 new_sack:
3852         /* Build the new head SACK, and we're done. */
3853         sp->start_seq = seq;
3854         sp->end_seq = end_seq;
3855         tp->rx_opt.num_sacks++;
3856         tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
3857 }
3858
3859 /* RCV.NXT advances, some SACKs should be eaten. */
3860
3861 static void tcp_sack_remove(struct tcp_sock *tp)
3862 {
3863         struct tcp_sack_block *sp = &tp->selective_acks[0];
3864         int num_sacks = tp->rx_opt.num_sacks;
3865         int this_sack;
3866
3867         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3868         if (skb_queue_empty(&tp->out_of_order_queue)) {
3869                 tp->rx_opt.num_sacks = 0;
3870                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3871                 return;
3872         }
3873
3874         for (this_sack = 0; this_sack < num_sacks;) {
3875                 /* Check if the start of the sack is covered by RCV.NXT. */
3876                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3877                         int i;
3878
3879                         /* RCV.NXT must cover all the block! */
3880                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
3881
3882                         /* Zap this SACK, by moving forward any other SACKS. */
3883                         for (i=this_sack+1; i < num_sacks; i++)
3884                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3885                         num_sacks--;
3886                         continue;
3887                 }
3888                 this_sack++;
3889                 sp++;
3890         }
3891         if (num_sacks != tp->rx_opt.num_sacks) {
3892                 tp->rx_opt.num_sacks = num_sacks;
3893                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
3894                                        tp->rx_opt.dsack;
3895         }
3896 }
3897
3898 /* This one checks to see if we can put data from the
3899  * out_of_order queue into the receive_queue.
3900  */
3901 static void tcp_ofo_queue(struct sock *sk)
3902 {
3903         struct tcp_sock *tp = tcp_sk(sk);
3904         __u32 dsack_high = tp->rcv_nxt;
3905         struct sk_buff *skb;
3906
3907         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3908                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3909                         break;
3910
3911                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3912                         __u32 dsack = dsack_high;
3913                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3914                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3915                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
3916                 }
3917
3918                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3919                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3920                         __skb_unlink(skb, &tp->out_of_order_queue);
3921                         __kfree_skb(skb);
3922                         continue;
3923                 }
3924                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3925                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3926                            TCP_SKB_CB(skb)->end_seq);
3927
3928                 __skb_unlink(skb, &tp->out_of_order_queue);
3929                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3930                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3931                 if (tcp_hdr(skb)->fin)
3932                         tcp_fin(skb, sk, tcp_hdr(skb));
3933         }
3934 }
3935
3936 static int tcp_prune_ofo_queue(struct sock *sk);
3937 static int tcp_prune_queue(struct sock *sk);
3938
3939 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
3940 {
3941         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3942             !sk_rmem_schedule(sk, size)) {
3943
3944                 if (tcp_prune_queue(sk) < 0)
3945                         return -1;
3946
3947                 if (!sk_rmem_schedule(sk, size)) {
3948                         if (!tcp_prune_ofo_queue(sk))
3949                                 return -1;
3950
3951                         if (!sk_rmem_schedule(sk, size))
3952                                 return -1;
3953                 }
3954         }
3955         return 0;
3956 }
3957
3958 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3959 {
3960         struct tcphdr *th = tcp_hdr(skb);
3961         struct tcp_sock *tp = tcp_sk(sk);
3962         int eaten = -1;
3963
3964         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3965                 goto drop;
3966
3967         __skb_pull(skb, th->doff * 4);
3968
3969         TCP_ECN_accept_cwr(tp, skb);
3970
3971         if (tp->rx_opt.dsack) {
3972                 tp->rx_opt.dsack = 0;
3973                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
3974         }
3975
3976         /*  Queue data for delivery to the user.
3977          *  Packets in sequence go to the receive queue.
3978          *  Out of sequence packets to the out_of_order_queue.
3979          */
3980         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3981                 if (tcp_receive_window(tp) == 0)
3982                         goto out_of_window;
3983
3984                 /* Ok. In sequence. In window. */
3985                 if (tp->ucopy.task == current &&
3986                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3987                     sock_owned_by_user(sk) && !tp->urg_data) {
3988                         int chunk = min_t(unsigned int, skb->len,
3989                                           tp->ucopy.len);
3990
3991                         __set_current_state(TASK_RUNNING);
3992
3993                         local_bh_enable();
3994                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3995                                 tp->ucopy.len -= chunk;
3996                                 tp->copied_seq += chunk;
3997                                 eaten = (chunk == skb->len && !th->fin);
3998                                 tcp_rcv_space_adjust(sk);
3999                         }
4000                         local_bh_disable();
4001                 }
4002
4003                 if (eaten <= 0) {
4004 queue_and_out:
4005                         if (eaten < 0 &&
4006                             tcp_try_rmem_schedule(sk, skb->truesize))
4007                                 goto drop;
4008
4009                         skb_set_owner_r(skb, sk);
4010                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4011                 }
4012                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4013                 if (skb->len)
4014                         tcp_event_data_recv(sk, skb);
4015                 if (th->fin)
4016                         tcp_fin(skb, sk, th);
4017
4018                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4019                         tcp_ofo_queue(sk);
4020
4021                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4022                          * gap in queue is filled.
4023                          */
4024                         if (skb_queue_empty(&tp->out_of_order_queue))
4025                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4026                 }
4027
4028                 if (tp->rx_opt.num_sacks)
4029                         tcp_sack_remove(tp);
4030
4031                 tcp_fast_path_check(sk);
4032
4033                 if (eaten > 0)
4034                         __kfree_skb(skb);
4035                 else if (!sock_flag(sk, SOCK_DEAD))
4036                         sk->sk_data_ready(sk, 0);
4037                 return;
4038         }
4039
4040         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4041                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4042                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4043                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4044
4045 out_of_window:
4046                 tcp_enter_quickack_mode(sk);
4047                 inet_csk_schedule_ack(sk);
4048 drop:
4049                 __kfree_skb(skb);
4050                 return;
4051         }
4052
4053         /* Out of window. F.e. zero window probe. */
4054         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4055                 goto out_of_window;
4056
4057         tcp_enter_quickack_mode(sk);
4058
4059         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4060                 /* Partial packet, seq < rcv_next < end_seq */
4061                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4062                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4063                            TCP_SKB_CB(skb)->end_seq);
4064
4065                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4066
4067                 /* If window is closed, drop tail of packet. But after
4068                  * remembering D-SACK for its head made in previous line.
4069                  */
4070                 if (!tcp_receive_window(tp))
4071                         goto out_of_window;
4072                 goto queue_and_out;
4073         }
4074
4075         TCP_ECN_check_ce(tp, skb);
4076
4077         if (tcp_try_rmem_schedule(sk, skb->truesize))
4078                 goto drop;
4079
4080         /* Disable header prediction. */
4081         tp->pred_flags = 0;
4082         inet_csk_schedule_ack(sk);
4083
4084         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4085                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4086
4087         skb_set_owner_r(skb, sk);
4088
4089         if (!skb_peek(&tp->out_of_order_queue)) {
4090                 /* Initial out of order segment, build 1 SACK. */
4091                 if (tcp_is_sack(tp)) {
4092                         tp->rx_opt.num_sacks = 1;
4093                         tp->rx_opt.dsack     = 0;
4094                         tp->rx_opt.eff_sacks = 1;
4095                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4096                         tp->selective_acks[0].end_seq =
4097                                                 TCP_SKB_CB(skb)->end_seq;
4098                 }
4099                 __skb_queue_head(&tp->out_of_order_queue, skb);
4100         } else {
4101                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4102                 u32 seq = TCP_SKB_CB(skb)->seq;
4103                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4104
4105                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4106                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4107
4108                         if (!tp->rx_opt.num_sacks ||
4109                             tp->selective_acks[0].end_seq != seq)
4110                                 goto add_sack;
4111
4112                         /* Common case: data arrive in order after hole. */
4113                         tp->selective_acks[0].end_seq = end_seq;
4114                         return;
4115                 }
4116
4117                 /* Find place to insert this segment. */
4118                 do {
4119                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
4120                                 break;
4121                 } while ((skb1 = skb1->prev) !=
4122                          (struct sk_buff *)&tp->out_of_order_queue);
4123
4124                 /* Do skb overlap to previous one? */
4125                 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4126                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4127                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4128                                 /* All the bits are present. Drop. */
4129                                 __kfree_skb(skb);
4130                                 tcp_dsack_set(sk, seq, end_seq);
4131                                 goto add_sack;
4132                         }
4133                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4134                                 /* Partial overlap. */
4135                                 tcp_dsack_set(sk, seq,
4136                                               TCP_SKB_CB(skb1)->end_seq);
4137                         } else {
4138                                 skb1 = skb1->prev;
4139                         }
4140                 }
4141                 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
4142
4143                 /* And clean segments covered by new one as whole. */
4144                 while ((skb1 = skb->next) !=
4145                        (struct sk_buff *)&tp->out_of_order_queue &&
4146                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4147                         if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4148                                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4149                                                  end_seq);
4150                                 break;
4151                         }
4152                         __skb_unlink(skb1, &tp->out_of_order_queue);
4153                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4154                                          TCP_SKB_CB(skb1)->end_seq);
4155                         __kfree_skb(skb1);
4156                 }
4157
4158 add_sack:
4159                 if (tcp_is_sack(tp))
4160                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
4161         }
4162 }
4163
4164 /* Collapse contiguous sequence of skbs head..tail with
4165  * sequence numbers start..end.
4166  * Segments with FIN/SYN are not collapsed (only because this
4167  * simplifies code)
4168  */
4169 static void
4170 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4171              struct sk_buff *head, struct sk_buff *tail,
4172              u32 start, u32 end)
4173 {
4174         struct sk_buff *skb;
4175
4176         /* First, check that queue is collapsible and find
4177          * the point where collapsing can be useful. */
4178         for (skb = head; skb != tail;) {
4179                 /* No new bits? It is possible on ofo queue. */
4180                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4181                         struct sk_buff *next = skb->next;
4182                         __skb_unlink(skb, list);
4183                         __kfree_skb(skb);
4184                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4185                         skb = next;
4186                         continue;
4187                 }
4188
4189                 /* The first skb to collapse is:
4190                  * - not SYN/FIN and
4191                  * - bloated or contains data before "start" or
4192                  *   overlaps to the next one.
4193                  */
4194                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4195                     (tcp_win_from_space(skb->truesize) > skb->len ||
4196                      before(TCP_SKB_CB(skb)->seq, start) ||
4197                      (skb->next != tail &&
4198                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4199                         break;
4200
4201                 /* Decided to skip this, advance start seq. */
4202                 start = TCP_SKB_CB(skb)->end_seq;
4203                 skb = skb->next;
4204         }
4205         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4206                 return;
4207
4208         while (before(start, end)) {
4209                 struct sk_buff *nskb;
4210                 unsigned int header = skb_headroom(skb);
4211                 int copy = SKB_MAX_ORDER(header, 0);
4212
4213                 /* Too big header? This can happen with IPv6. */
4214                 if (copy < 0)
4215                         return;
4216                 if (end - start < copy)
4217                         copy = end - start;
4218                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4219                 if (!nskb)
4220                         return;
4221
4222                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4223                 skb_set_network_header(nskb, (skb_network_header(skb) -
4224                                               skb->head));
4225                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4226                                                 skb->head));
4227                 skb_reserve(nskb, header);
4228                 memcpy(nskb->head, skb->head, header);
4229                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4230                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4231                 __skb_insert(nskb, skb->prev, skb, list);
4232                 skb_set_owner_r(nskb, sk);
4233
4234                 /* Copy data, releasing collapsed skbs. */
4235                 while (copy > 0) {
4236                         int offset = start - TCP_SKB_CB(skb)->seq;
4237                         int size = TCP_SKB_CB(skb)->end_seq - start;
4238
4239                         BUG_ON(offset < 0);
4240                         if (size > 0) {
4241                                 size = min(copy, size);
4242                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4243                                         BUG();
4244                                 TCP_SKB_CB(nskb)->end_seq += size;
4245                                 copy -= size;
4246                                 start += size;
4247                         }
4248                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4249                                 struct sk_buff *next = skb->next;
4250                                 __skb_unlink(skb, list);
4251                                 __kfree_skb(skb);
4252                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4253                                 skb = next;
4254                                 if (skb == tail ||
4255                                     tcp_hdr(skb)->syn ||
4256                                     tcp_hdr(skb)->fin)
4257                                         return;
4258                         }
4259                 }
4260         }
4261 }
4262
4263 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4264  * and tcp_collapse() them until all the queue is collapsed.
4265  */
4266 static void tcp_collapse_ofo_queue(struct sock *sk)
4267 {
4268         struct tcp_sock *tp = tcp_sk(sk);
4269         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4270         struct sk_buff *head;
4271         u32 start, end;
4272
4273         if (skb == NULL)
4274                 return;
4275
4276         start = TCP_SKB_CB(skb)->seq;
4277         end = TCP_SKB_CB(skb)->end_seq;
4278         head = skb;
4279
4280         for (;;) {
4281                 skb = skb->next;
4282
4283                 /* Segment is terminated when we see gap or when
4284                  * we are at the end of all the queue. */
4285                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4286                     after(TCP_SKB_CB(skb)->seq, end) ||
4287                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4288                         tcp_collapse(sk, &tp->out_of_order_queue,
4289                                      head, skb, start, end);
4290                         head = skb;
4291                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4292                                 break;
4293                         /* Start new segment */
4294                         start = TCP_SKB_CB(skb)->seq;
4295                         end = TCP_SKB_CB(skb)->end_seq;
4296                 } else {
4297                         if (before(TCP_SKB_CB(skb)->seq, start))
4298                                 start = TCP_SKB_CB(skb)->seq;
4299                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4300                                 end = TCP_SKB_CB(skb)->end_seq;
4301                 }
4302         }
4303 }
4304
4305 /*
4306  * Purge the out-of-order queue.
4307  * Return true if queue was pruned.
4308  */
4309 static int tcp_prune_ofo_queue(struct sock *sk)
4310 {
4311         struct tcp_sock *tp = tcp_sk(sk);
4312         int res = 0;
4313
4314         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4315                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4316                 __skb_queue_purge(&tp->out_of_order_queue);
4317
4318                 /* Reset SACK state.  A conforming SACK implementation will
4319                  * do the same at a timeout based retransmit.  When a connection
4320                  * is in a sad state like this, we care only about integrity
4321                  * of the connection not performance.
4322                  */
4323                 if (tp->rx_opt.sack_ok)
4324                         tcp_sack_reset(&tp->rx_opt);
4325                 sk_mem_reclaim(sk);
4326                 res = 1;
4327         }
4328         return res;
4329 }
4330
4331 /* Reduce allocated memory if we can, trying to get
4332  * the socket within its memory limits again.
4333  *
4334  * Return less than zero if we should start dropping frames
4335  * until the socket owning process reads some of the data
4336  * to stabilize the situation.
4337  */
4338 static int tcp_prune_queue(struct sock *sk)
4339 {
4340         struct tcp_sock *tp = tcp_sk(sk);
4341
4342         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4343
4344         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4345
4346         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4347                 tcp_clamp_window(sk);
4348         else if (tcp_memory_pressure)
4349                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4350
4351         tcp_collapse_ofo_queue(sk);
4352         tcp_collapse(sk, &sk->sk_receive_queue,
4353                      sk->sk_receive_queue.next,
4354                      (struct sk_buff *)&sk->sk_receive_queue,
4355                      tp->copied_seq, tp->rcv_nxt);
4356         sk_mem_reclaim(sk);
4357
4358         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4359                 return 0;
4360
4361         /* Collapsing did not help, destructive actions follow.
4362          * This must not ever occur. */
4363
4364         tcp_prune_ofo_queue(sk);
4365
4366         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4367                 return 0;
4368
4369         /* If we are really being abused, tell the caller to silently
4370          * drop receive data on the floor.  It will get retransmitted
4371          * and hopefully then we'll have sufficient space.
4372          */
4373         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4374
4375         /* Massive buffer overcommit. */
4376         tp->pred_flags = 0;
4377         return -1;
4378 }
4379
4380 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4381  * As additional protections, we do not touch cwnd in retransmission phases,
4382  * and if application hit its sndbuf limit recently.
4383  */
4384 void tcp_cwnd_application_limited(struct sock *sk)
4385 {
4386         struct tcp_sock *tp = tcp_sk(sk);
4387
4388         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4389             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4390                 /* Limited by application or receiver window. */
4391                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4392                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4393                 if (win_used < tp->snd_cwnd) {
4394                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4395                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4396                 }
4397                 tp->snd_cwnd_used = 0;
4398         }
4399         tp->snd_cwnd_stamp = tcp_time_stamp;
4400 }
4401
4402 static int tcp_should_expand_sndbuf(struct sock *sk)
4403 {
4404         struct tcp_sock *tp = tcp_sk(sk);
4405
4406         /* If the user specified a specific send buffer setting, do
4407          * not modify it.
4408          */
4409         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4410                 return 0;
4411
4412         /* If we are under global TCP memory pressure, do not expand.  */
4413         if (tcp_memory_pressure)
4414                 return 0;
4415
4416         /* If we are under soft global TCP memory pressure, do not expand.  */
4417         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4418                 return 0;
4419
4420         /* If we filled the congestion window, do not expand.  */
4421         if (tp->packets_out >= tp->snd_cwnd)
4422                 return 0;
4423
4424         return 1;
4425 }
4426
4427 /* When incoming ACK allowed to free some skb from write_queue,
4428  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4429  * on the exit from tcp input handler.
4430  *
4431  * PROBLEM: sndbuf expansion does not work well with largesend.
4432  */
4433 static void tcp_new_space(struct sock *sk)
4434 {
4435         struct tcp_sock *tp = tcp_sk(sk);
4436
4437         if (tcp_should_expand_sndbuf(sk)) {
4438                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4439                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4440                     demanded = max_t(unsigned int, tp->snd_cwnd,
4441                                      tp->reordering + 1);
4442                 sndmem *= 2 * demanded;
4443                 if (sndmem > sk->sk_sndbuf)
4444                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4445                 tp->snd_cwnd_stamp = tcp_time_stamp;
4446         }
4447
4448         sk->sk_write_space(sk);
4449 }
4450
4451 static void tcp_check_space(struct sock *sk)
4452 {
4453         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4454                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4455                 if (sk->sk_socket &&
4456                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4457                         tcp_new_space(sk);
4458         }
4459 }
4460
4461 static inline void tcp_data_snd_check(struct sock *sk)
4462 {
4463         tcp_push_pending_frames(sk);
4464         tcp_check_space(sk);
4465 }
4466
4467 /*
4468  * Check if sending an ack is needed.
4469  */
4470 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4471 {
4472         struct tcp_sock *tp = tcp_sk(sk);
4473
4474             /* More than one full frame received... */
4475         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4476              /* ... and right edge of window advances far enough.
4477               * (tcp_recvmsg() will send ACK otherwise). Or...
4478               */
4479              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4480             /* We ACK each frame or... */
4481             tcp_in_quickack_mode(sk) ||
4482             /* We have out of order data. */
4483             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4484                 /* Then ack it now */
4485                 tcp_send_ack(sk);
4486         } else {
4487                 /* Else, send delayed ack. */
4488                 tcp_send_delayed_ack(sk);
4489         }
4490 }
4491
4492 static inline void tcp_ack_snd_check(struct sock *sk)
4493 {
4494         if (!inet_csk_ack_scheduled(sk)) {
4495                 /* We sent a data segment already. */
4496                 return;
4497         }
4498         __tcp_ack_snd_check(sk, 1);
4499 }
4500
4501 /*
4502  *      This routine is only called when we have urgent data
4503  *      signaled. Its the 'slow' part of tcp_urg. It could be
4504  *      moved inline now as tcp_urg is only called from one
4505  *      place. We handle URGent data wrong. We have to - as
4506  *      BSD still doesn't use the correction from RFC961.
4507  *      For 1003.1g we should support a new option TCP_STDURG to permit
4508  *      either form (or just set the sysctl tcp_stdurg).
4509  */
4510
4511 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4512 {
4513         struct tcp_sock *tp = tcp_sk(sk);
4514         u32 ptr = ntohs(th->urg_ptr);
4515
4516         if (ptr && !sysctl_tcp_stdurg)
4517                 ptr--;
4518         ptr += ntohl(th->seq);
4519
4520         /* Ignore urgent data that we've already seen and read. */
4521         if (after(tp->copied_seq, ptr))
4522                 return;
4523
4524         /* Do not replay urg ptr.
4525          *
4526          * NOTE: interesting situation not covered by specs.
4527          * Misbehaving sender may send urg ptr, pointing to segment,
4528          * which we already have in ofo queue. We are not able to fetch
4529          * such data and will stay in TCP_URG_NOTYET until will be eaten
4530          * by recvmsg(). Seems, we are not obliged to handle such wicked
4531          * situations. But it is worth to think about possibility of some
4532          * DoSes using some hypothetical application level deadlock.
4533          */
4534         if (before(ptr, tp->rcv_nxt))
4535                 return;
4536
4537         /* Do we already have a newer (or duplicate) urgent pointer? */
4538         if (tp->urg_data && !after(ptr, tp->urg_seq))
4539                 return;
4540
4541         /* Tell the world about our new urgent pointer. */
4542         sk_send_sigurg(sk);
4543
4544         /* We may be adding urgent data when the last byte read was
4545          * urgent. To do this requires some care. We cannot just ignore
4546          * tp->copied_seq since we would read the last urgent byte again
4547          * as data, nor can we alter copied_seq until this data arrives
4548          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4549          *
4550          * NOTE. Double Dutch. Rendering to plain English: author of comment
4551          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4552          * and expect that both A and B disappear from stream. This is _wrong_.
4553          * Though this happens in BSD with high probability, this is occasional.
4554          * Any application relying on this is buggy. Note also, that fix "works"
4555          * only in this artificial test. Insert some normal data between A and B and we will
4556          * decline of BSD again. Verdict: it is better to remove to trap
4557          * buggy users.
4558          */
4559         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4560             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4561                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4562                 tp->copied_seq++;
4563                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4564                         __skb_unlink(skb, &sk->sk_receive_queue);
4565                         __kfree_skb(skb);
4566                 }
4567         }
4568
4569         tp->urg_data = TCP_URG_NOTYET;
4570         tp->urg_seq = ptr;
4571
4572         /* Disable header prediction. */
4573         tp->pred_flags = 0;
4574 }
4575
4576 /* This is the 'fast' part of urgent handling. */
4577 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4578 {
4579         struct tcp_sock *tp = tcp_sk(sk);
4580
4581         /* Check if we get a new urgent pointer - normally not. */
4582         if (th->urg)
4583                 tcp_check_urg(sk, th);
4584
4585         /* Do we wait for any urgent data? - normally not... */
4586         if (tp->urg_data == TCP_URG_NOTYET) {
4587                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4588                           th->syn;
4589
4590                 /* Is the urgent pointer pointing into this packet? */
4591                 if (ptr < skb->len) {
4592                         u8 tmp;
4593                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4594                                 BUG();
4595                         tp->urg_data = TCP_URG_VALID | tmp;
4596                         if (!sock_flag(sk, SOCK_DEAD))
4597                                 sk->sk_data_ready(sk, 0);
4598                 }
4599         }
4600 }
4601
4602 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4603 {
4604         struct tcp_sock *tp = tcp_sk(sk);
4605         int chunk = skb->len - hlen;
4606         int err;
4607
4608         local_bh_enable();
4609         if (skb_csum_unnecessary(skb))
4610                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4611         else
4612                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4613                                                        tp->ucopy.iov);
4614
4615         if (!err) {
4616                 tp->ucopy.len -= chunk;
4617                 tp->copied_seq += chunk;
4618                 tcp_rcv_space_adjust(sk);
4619         }
4620
4621         local_bh_disable();
4622         return err;
4623 }
4624
4625 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4626                                             struct sk_buff *skb)
4627 {
4628         __sum16 result;
4629
4630         if (sock_owned_by_user(sk)) {
4631                 local_bh_enable();
4632                 result = __tcp_checksum_complete(skb);
4633                 local_bh_disable();
4634         } else {
4635                 result = __tcp_checksum_complete(skb);
4636         }
4637         return result;
4638 }
4639
4640 static inline int tcp_checksum_complete_user(struct sock *sk,
4641                                              struct sk_buff *skb)
4642 {
4643         return !skb_csum_unnecessary(skb) &&
4644                __tcp_checksum_complete_user(sk, skb);
4645 }
4646
4647 #ifdef CONFIG_NET_DMA
4648 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4649                                   int hlen)
4650 {
4651         struct tcp_sock *tp = tcp_sk(sk);
4652         int chunk = skb->len - hlen;
4653         int dma_cookie;
4654         int copied_early = 0;
4655
4656         if (tp->ucopy.wakeup)
4657                 return 0;
4658
4659         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4660                 tp->ucopy.dma_chan = get_softnet_dma();
4661
4662         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4663
4664                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4665                                                          skb, hlen,
4666                                                          tp->ucopy.iov, chunk,
4667                                                          tp->ucopy.pinned_list);
4668
4669                 if (dma_cookie < 0)
4670                         goto out;
4671
4672                 tp->ucopy.dma_cookie = dma_cookie;
4673                 copied_early = 1;
4674
4675                 tp->ucopy.len -= chunk;
4676                 tp->copied_seq += chunk;
4677                 tcp_rcv_space_adjust(sk);
4678
4679                 if ((tp->ucopy.len == 0) ||
4680                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4681                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4682                         tp->ucopy.wakeup = 1;
4683                         sk->sk_data_ready(sk, 0);
4684                 }
4685         } else if (chunk > 0) {
4686                 tp->ucopy.wakeup = 1;
4687                 sk->sk_data_ready(sk, 0);
4688         }
4689 out:
4690         return copied_early;
4691 }
4692 #endif /* CONFIG_NET_DMA */
4693
4694 /*
4695  *      TCP receive function for the ESTABLISHED state.
4696  *
4697  *      It is split into a fast path and a slow path. The fast path is
4698  *      disabled when:
4699  *      - A zero window was announced from us - zero window probing
4700  *        is only handled properly in the slow path.
4701  *      - Out of order segments arrived.
4702  *      - Urgent data is expected.
4703  *      - There is no buffer space left
4704  *      - Unexpected TCP flags/window values/header lengths are received
4705  *        (detected by checking the TCP header against pred_flags)
4706  *      - Data is sent in both directions. Fast path only supports pure senders
4707  *        or pure receivers (this means either the sequence number or the ack
4708  *        value must stay constant)
4709  *      - Unexpected TCP option.
4710  *
4711  *      When these conditions are not satisfied it drops into a standard
4712  *      receive procedure patterned after RFC793 to handle all cases.
4713  *      The first three cases are guaranteed by proper pred_flags setting,
4714  *      the rest is checked inline. Fast processing is turned on in
4715  *      tcp_data_queue when everything is OK.
4716  */
4717 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4718                         struct tcphdr *th, unsigned len)
4719 {
4720         struct tcp_sock *tp = tcp_sk(sk);
4721
4722         /*
4723          *      Header prediction.
4724          *      The code loosely follows the one in the famous
4725          *      "30 instruction TCP receive" Van Jacobson mail.
4726          *
4727          *      Van's trick is to deposit buffers into socket queue
4728          *      on a device interrupt, to call tcp_recv function
4729          *      on the receive process context and checksum and copy
4730          *      the buffer to user space. smart...
4731          *
4732          *      Our current scheme is not silly either but we take the
4733          *      extra cost of the net_bh soft interrupt processing...
4734          *      We do checksum and copy also but from device to kernel.
4735          */
4736
4737         tp->rx_opt.saw_tstamp = 0;
4738
4739         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4740          *      if header_prediction is to be made
4741          *      'S' will always be tp->tcp_header_len >> 2
4742          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4743          *  turn it off (when there are holes in the receive
4744          *       space for instance)
4745          *      PSH flag is ignored.
4746          */
4747
4748         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4749             TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4750                 int tcp_header_len = tp->tcp_header_len;
4751
4752                 /* Timestamp header prediction: tcp_header_len
4753                  * is automatically equal to th->doff*4 due to pred_flags
4754                  * match.
4755                  */
4756
4757                 /* Check timestamp */
4758                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4759                         __be32 *ptr = (__be32 *)(th + 1);
4760
4761                         /* No? Slow path! */
4762                         if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4763                                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4764                                 goto slow_path;
4765
4766                         tp->rx_opt.saw_tstamp = 1;
4767                         ++ptr;
4768                         tp->rx_opt.rcv_tsval = ntohl(*ptr);
4769                         ++ptr;
4770                         tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4771
4772                         /* If PAWS failed, check it more carefully in slow path */
4773                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4774                                 goto slow_path;
4775
4776                         /* DO NOT update ts_recent here, if checksum fails
4777                          * and timestamp was corrupted part, it will result
4778                          * in a hung connection since we will drop all
4779                          * future packets due to the PAWS test.
4780                          */
4781                 }
4782
4783                 if (len <= tcp_header_len) {
4784                         /* Bulk data transfer: sender */
4785                         if (len == tcp_header_len) {
4786                                 /* Predicted packet is in window by definition.
4787                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4788                                  * Hence, check seq<=rcv_wup reduces to:
4789                                  */
4790                                 if (tcp_header_len ==
4791                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4792                                     tp->rcv_nxt == tp->rcv_wup)
4793                                         tcp_store_ts_recent(tp);
4794
4795                                 /* We know that such packets are checksummed
4796                                  * on entry.
4797                                  */
4798                                 tcp_ack(sk, skb, 0);
4799                                 __kfree_skb(skb);
4800                                 tcp_data_snd_check(sk);
4801                                 return 0;
4802                         } else { /* Header too small */
4803                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4804                                 goto discard;
4805                         }
4806                 } else {
4807                         int eaten = 0;
4808                         int copied_early = 0;
4809
4810                         if (tp->copied_seq == tp->rcv_nxt &&
4811                             len - tcp_header_len <= tp->ucopy.len) {
4812 #ifdef CONFIG_NET_DMA
4813                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4814                                         copied_early = 1;
4815                                         eaten = 1;
4816                                 }
4817 #endif
4818                                 if (tp->ucopy.task == current &&
4819                                     sock_owned_by_user(sk) && !copied_early) {
4820                                         __set_current_state(TASK_RUNNING);
4821
4822                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4823                                                 eaten = 1;
4824                                 }
4825                                 if (eaten) {
4826                                         /* Predicted packet is in window by definition.
4827                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4828                                          * Hence, check seq<=rcv_wup reduces to:
4829                                          */
4830                                         if (tcp_header_len ==
4831                                             (sizeof(struct tcphdr) +
4832                                              TCPOLEN_TSTAMP_ALIGNED) &&
4833                                             tp->rcv_nxt == tp->rcv_wup)
4834                                                 tcp_store_ts_recent(tp);
4835
4836                                         tcp_rcv_rtt_measure_ts(sk, skb);
4837
4838                                         __skb_pull(skb, tcp_header_len);
4839                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4840                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
4841                                 }
4842                                 if (copied_early)
4843                                         tcp_cleanup_rbuf(sk, skb->len);
4844                         }
4845                         if (!eaten) {
4846                                 if (tcp_checksum_complete_user(sk, skb))
4847                                         goto csum_error;
4848
4849                                 /* Predicted packet is in window by definition.
4850                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4851                                  * Hence, check seq<=rcv_wup reduces to:
4852                                  */
4853                                 if (tcp_header_len ==
4854                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4855                                     tp->rcv_nxt == tp->rcv_wup)
4856                                         tcp_store_ts_recent(tp);
4857
4858                                 tcp_rcv_rtt_measure_ts(sk, skb);
4859
4860                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4861                                         goto step5;
4862
4863                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
4864
4865                                 /* Bulk data transfer: receiver */
4866                                 __skb_pull(skb, tcp_header_len);
4867                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4868                                 skb_set_owner_r(skb, sk);
4869                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4870                         }
4871
4872                         tcp_event_data_recv(sk, skb);
4873
4874                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4875                                 /* Well, only one small jumplet in fast path... */
4876                                 tcp_ack(sk, skb, FLAG_DATA);
4877                                 tcp_data_snd_check(sk);
4878                                 if (!inet_csk_ack_scheduled(sk))
4879                                         goto no_ack;
4880                         }
4881
4882                         __tcp_ack_snd_check(sk, 0);
4883 no_ack:
4884 #ifdef CONFIG_NET_DMA
4885                         if (copied_early)
4886                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4887                         else
4888 #endif
4889                         if (eaten)
4890                                 __kfree_skb(skb);
4891                         else
4892                                 sk->sk_data_ready(sk, 0);
4893                         return 0;
4894                 }
4895         }
4896
4897 slow_path:
4898         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
4899                 goto csum_error;
4900
4901         /*
4902          * RFC1323: H1. Apply PAWS check first.
4903          */
4904         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4905             tcp_paws_discard(sk, skb)) {
4906                 if (!th->rst) {
4907                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4908                         tcp_send_dupack(sk, skb);
4909                         goto discard;
4910                 }
4911                 /* Resets are accepted even if PAWS failed.
4912
4913                    ts_recent update must be made after we are sure
4914                    that the packet is in window.
4915                  */
4916         }
4917
4918         /*
4919          *      Standard slow path.
4920          */
4921
4922         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4923                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4924                  * (RST) segments are validated by checking their SEQ-fields."
4925                  * And page 69: "If an incoming segment is not acceptable,
4926                  * an acknowledgment should be sent in reply (unless the RST bit
4927                  * is set, if so drop the segment and return)".
4928                  */
4929                 if (!th->rst)
4930                         tcp_send_dupack(sk, skb);
4931                 goto discard;
4932         }
4933
4934         if (th->rst) {
4935                 tcp_reset(sk);
4936                 goto discard;
4937         }
4938
4939         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4940
4941         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4942                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4943                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
4944                 tcp_reset(sk);
4945                 return 1;
4946         }
4947
4948 step5:
4949         if (th->ack)
4950                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4951
4952         tcp_rcv_rtt_measure_ts(sk, skb);
4953
4954         /* Process urgent data. */
4955         tcp_urg(sk, skb, th);
4956
4957         /* step 7: process the segment text */
4958         tcp_data_queue(sk, skb);
4959
4960         tcp_data_snd_check(sk);
4961         tcp_ack_snd_check(sk);
4962         return 0;
4963
4964 csum_error:
4965         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4966
4967 discard:
4968         __kfree_skb(skb);
4969         return 0;
4970 }
4971
4972 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4973                                          struct tcphdr *th, unsigned len)
4974 {
4975         struct tcp_sock *tp = tcp_sk(sk);
4976         struct inet_connection_sock *icsk = inet_csk(sk);
4977         int saved_clamp = tp->rx_opt.mss_clamp;
4978
4979         tcp_parse_options(skb, &tp->rx_opt, 0);
4980
4981         if (th->ack) {
4982                 /* rfc793:
4983                  * "If the state is SYN-SENT then
4984                  *    first check the ACK bit
4985                  *      If the ACK bit is set
4986                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4987                  *        a reset (unless the RST bit is set, if so drop
4988                  *        the segment and return)"
4989                  *
4990                  *  We do not send data with SYN, so that RFC-correct
4991                  *  test reduces to:
4992                  */
4993                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4994                         goto reset_and_undo;
4995
4996                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4997                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4998                              tcp_time_stamp)) {
4999                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5000                         goto reset_and_undo;
5001                 }
5002
5003                 /* Now ACK is acceptable.
5004                  *
5005                  * "If the RST bit is set
5006                  *    If the ACK was acceptable then signal the user "error:
5007                  *    connection reset", drop the segment, enter CLOSED state,
5008                  *    delete TCB, and return."
5009                  */
5010
5011                 if (th->rst) {
5012                         tcp_reset(sk);
5013                         goto discard;
5014                 }
5015
5016                 /* rfc793:
5017                  *   "fifth, if neither of the SYN or RST bits is set then
5018                  *    drop the segment and return."
5019                  *
5020                  *    See note below!
5021                  *                                        --ANK(990513)
5022                  */
5023                 if (!th->syn)
5024                         goto discard_and_undo;
5025
5026                 /* rfc793:
5027                  *   "If the SYN bit is on ...
5028                  *    are acceptable then ...
5029                  *    (our SYN has been ACKed), change the connection
5030                  *    state to ESTABLISHED..."
5031                  */
5032
5033                 TCP_ECN_rcv_synack(tp, th);
5034
5035                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5036                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5037
5038                 /* Ok.. it's good. Set up sequence numbers and
5039                  * move to established.
5040                  */
5041                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5042                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5043
5044                 /* RFC1323: The window in SYN & SYN/ACK segments is
5045                  * never scaled.
5046                  */
5047                 tp->snd_wnd = ntohs(th->window);
5048                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
5049
5050                 if (!tp->rx_opt.wscale_ok) {
5051                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5052                         tp->window_clamp = min(tp->window_clamp, 65535U);
5053                 }
5054
5055                 if (tp->rx_opt.saw_tstamp) {
5056                         tp->rx_opt.tstamp_ok       = 1;
5057                         tp->tcp_header_len =
5058                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5059                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5060                         tcp_store_ts_recent(tp);
5061                 } else {
5062                         tp->tcp_header_len = sizeof(struct tcphdr);
5063                 }
5064
5065                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5066                         tcp_enable_fack(tp);
5067
5068                 tcp_mtup_init(sk);
5069                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5070                 tcp_initialize_rcv_mss(sk);
5071
5072                 /* Remember, tcp_poll() does not lock socket!
5073                  * Change state from SYN-SENT only after copied_seq
5074                  * is initialized. */
5075                 tp->copied_seq = tp->rcv_nxt;
5076                 smp_mb();
5077                 tcp_set_state(sk, TCP_ESTABLISHED);
5078
5079                 security_inet_conn_established(sk, skb);
5080
5081                 /* Make sure socket is routed, for correct metrics.  */
5082                 icsk->icsk_af_ops->rebuild_header(sk);
5083
5084                 tcp_init_metrics(sk);
5085
5086                 tcp_init_congestion_control(sk);
5087
5088                 /* Prevent spurious tcp_cwnd_restart() on first data
5089                  * packet.
5090                  */
5091                 tp->lsndtime = tcp_time_stamp;
5092
5093                 tcp_init_buffer_space(sk);
5094
5095                 if (sock_flag(sk, SOCK_KEEPOPEN))
5096                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5097
5098                 if (!tp->rx_opt.snd_wscale)
5099                         __tcp_fast_path_on(tp, tp->snd_wnd);
5100                 else
5101                         tp->pred_flags = 0;
5102
5103                 if (!sock_flag(sk, SOCK_DEAD)) {
5104                         sk->sk_state_change(sk);
5105                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5106                 }
5107
5108                 if (sk->sk_write_pending ||
5109                     icsk->icsk_accept_queue.rskq_defer_accept ||
5110                     icsk->icsk_ack.pingpong) {
5111                         /* Save one ACK. Data will be ready after
5112                          * several ticks, if write_pending is set.
5113                          *
5114                          * It may be deleted, but with this feature tcpdumps
5115                          * look so _wonderfully_ clever, that I was not able
5116                          * to stand against the temptation 8)     --ANK
5117                          */
5118                         inet_csk_schedule_ack(sk);
5119                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5120                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
5121                         tcp_incr_quickack(sk);
5122                         tcp_enter_quickack_mode(sk);
5123                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5124                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5125
5126 discard:
5127                         __kfree_skb(skb);
5128                         return 0;
5129                 } else {
5130                         tcp_send_ack(sk);
5131                 }
5132                 return -1;
5133         }
5134
5135         /* No ACK in the segment */
5136
5137         if (th->rst) {
5138                 /* rfc793:
5139                  * "If the RST bit is set
5140                  *
5141                  *      Otherwise (no ACK) drop the segment and return."
5142                  */
5143
5144                 goto discard_and_undo;
5145         }
5146
5147         /* PAWS check. */
5148         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5149             tcp_paws_check(&tp->rx_opt, 0))
5150                 goto discard_and_undo;
5151
5152         if (th->syn) {
5153                 /* We see SYN without ACK. It is attempt of
5154                  * simultaneous connect with crossed SYNs.
5155                  * Particularly, it can be connect to self.
5156                  */
5157                 tcp_set_state(sk, TCP_SYN_RECV);
5158
5159                 if (tp->rx_opt.saw_tstamp) {
5160                         tp->rx_opt.tstamp_ok = 1;
5161                         tcp_store_ts_recent(tp);
5162                         tp->tcp_header_len =
5163                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5164                 } else {
5165                         tp->tcp_header_len = sizeof(struct tcphdr);
5166                 }
5167
5168                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5169                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5170
5171                 /* RFC1323: The window in SYN & SYN/ACK segments is
5172                  * never scaled.
5173                  */
5174                 tp->snd_wnd    = ntohs(th->window);
5175                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5176                 tp->max_window = tp->snd_wnd;
5177
5178                 TCP_ECN_rcv_syn(tp, th);
5179
5180                 tcp_mtup_init(sk);
5181                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5182                 tcp_initialize_rcv_mss(sk);
5183
5184                 tcp_send_synack(sk);
5185 #if 0
5186                 /* Note, we could accept data and URG from this segment.
5187                  * There are no obstacles to make this.
5188                  *
5189                  * However, if we ignore data in ACKless segments sometimes,
5190                  * we have no reasons to accept it sometimes.
5191                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5192                  * is not flawless. So, discard packet for sanity.
5193                  * Uncomment this return to process the data.
5194                  */
5195                 return -1;
5196 #else
5197                 goto discard;
5198 #endif
5199         }
5200         /* "fifth, if neither of the SYN or RST bits is set then
5201          * drop the segment and return."
5202          */
5203
5204 discard_and_undo:
5205         tcp_clear_options(&tp->rx_opt);
5206         tp->rx_opt.mss_clamp = saved_clamp;
5207         goto discard;
5208
5209 reset_and_undo:
5210         tcp_clear_options(&tp->rx_opt);
5211         tp->rx_opt.mss_clamp = saved_clamp;
5212         return 1;
5213 }
5214
5215 /*
5216  *      This function implements the receiving procedure of RFC 793 for
5217  *      all states except ESTABLISHED and TIME_WAIT.
5218  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5219  *      address independent.
5220  */
5221
5222 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5223                           struct tcphdr *th, unsigned len)
5224 {
5225         struct tcp_sock *tp = tcp_sk(sk);
5226         struct inet_connection_sock *icsk = inet_csk(sk);
5227         int queued = 0;
5228
5229         tp->rx_opt.saw_tstamp = 0;
5230
5231         switch (sk->sk_state) {
5232         case TCP_CLOSE:
5233                 goto discard;
5234
5235         case TCP_LISTEN:
5236                 if (th->ack)
5237                         return 1;
5238
5239                 if (th->rst)
5240                         goto discard;
5241
5242                 if (th->syn) {
5243                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5244                                 return 1;
5245
5246                         /* Now we have several options: In theory there is
5247                          * nothing else in the frame. KA9Q has an option to
5248                          * send data with the syn, BSD accepts data with the
5249                          * syn up to the [to be] advertised window and
5250                          * Solaris 2.1 gives you a protocol error. For now
5251                          * we just ignore it, that fits the spec precisely
5252                          * and avoids incompatibilities. It would be nice in
5253                          * future to drop through and process the data.
5254                          *
5255                          * Now that TTCP is starting to be used we ought to
5256                          * queue this data.
5257                          * But, this leaves one open to an easy denial of
5258                          * service attack, and SYN cookies can't defend
5259                          * against this problem. So, we drop the data
5260                          * in the interest of security over speed unless
5261                          * it's still in use.
5262                          */
5263                         kfree_skb(skb);
5264                         return 0;
5265                 }
5266                 goto discard;
5267
5268         case TCP_SYN_SENT:
5269                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5270                 if (queued >= 0)
5271                         return queued;
5272
5273                 /* Do step6 onward by hand. */
5274                 tcp_urg(sk, skb, th);
5275                 __kfree_skb(skb);
5276                 tcp_data_snd_check(sk);
5277                 return 0;
5278         }
5279
5280         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5281             tcp_paws_discard(sk, skb)) {
5282                 if (!th->rst) {
5283                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5284                         tcp_send_dupack(sk, skb);
5285                         goto discard;
5286                 }
5287                 /* Reset is accepted even if it did not pass PAWS. */
5288         }
5289
5290         /* step 1: check sequence number */
5291         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5292                 if (!th->rst)
5293                         tcp_send_dupack(sk, skb);
5294                 goto discard;
5295         }
5296
5297         /* step 2: check RST bit */
5298         if (th->rst) {
5299                 tcp_reset(sk);
5300                 goto discard;
5301         }
5302
5303         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5304
5305         /* step 3: check security and precedence [ignored] */
5306
5307         /*      step 4:
5308          *
5309          *      Check for a SYN in window.
5310          */
5311         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5312                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5313                 tcp_reset(sk);
5314                 return 1;
5315         }
5316
5317         /* step 5: check the ACK field */
5318         if (th->ack) {
5319                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5320
5321                 switch (sk->sk_state) {
5322                 case TCP_SYN_RECV:
5323                         if (acceptable) {
5324                                 tp->copied_seq = tp->rcv_nxt;
5325                                 smp_mb();
5326                                 tcp_set_state(sk, TCP_ESTABLISHED);
5327                                 sk->sk_state_change(sk);
5328
5329                                 /* Note, that this wakeup is only for marginal
5330                                  * crossed SYN case. Passively open sockets
5331                                  * are not waked up, because sk->sk_sleep ==
5332                                  * NULL and sk->sk_socket == NULL.
5333                                  */
5334                                 if (sk->sk_socket)
5335                                         sk_wake_async(sk,
5336                                                       SOCK_WAKE_IO, POLL_OUT);
5337
5338                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5339                                 tp->snd_wnd = ntohs(th->window) <<
5340                                               tp->rx_opt.snd_wscale;
5341                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5342                                             TCP_SKB_CB(skb)->seq);
5343
5344                                 /* tcp_ack considers this ACK as duplicate
5345                                  * and does not calculate rtt.
5346                                  * Fix it at least with timestamps.
5347                                  */
5348                                 if (tp->rx_opt.saw_tstamp &&
5349                                     tp->rx_opt.rcv_tsecr && !tp->srtt)
5350                                         tcp_ack_saw_tstamp(sk, 0);
5351
5352                                 if (tp->rx_opt.tstamp_ok)
5353                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5354
5355                                 /* Make sure socket is routed, for
5356                                  * correct metrics.
5357                                  */
5358                                 icsk->icsk_af_ops->rebuild_header(sk);
5359
5360                                 tcp_init_metrics(sk);
5361
5362                                 tcp_init_congestion_control(sk);
5363
5364                                 /* Prevent spurious tcp_cwnd_restart() on
5365                                  * first data packet.
5366                                  */
5367                                 tp->lsndtime = tcp_time_stamp;
5368
5369                                 tcp_mtup_init(sk);
5370                                 tcp_initialize_rcv_mss(sk);
5371                                 tcp_init_buffer_space(sk);
5372                                 tcp_fast_path_on(tp);
5373                         } else {
5374                                 return 1;
5375                         }
5376                         break;
5377
5378                 case TCP_FIN_WAIT1:
5379                         if (tp->snd_una == tp->write_seq) {
5380                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5381                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5382                                 dst_confirm(sk->sk_dst_cache);
5383
5384                                 if (!sock_flag(sk, SOCK_DEAD))
5385                                         /* Wake up lingering close() */
5386                                         sk->sk_state_change(sk);
5387                                 else {
5388                                         int tmo;
5389
5390                                         if (tp->linger2 < 0 ||
5391                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5392                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5393                                                 tcp_done(sk);
5394                                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5395                                                 return 1;
5396                                         }
5397
5398                                         tmo = tcp_fin_time(sk);
5399                                         if (tmo > TCP_TIMEWAIT_LEN) {
5400                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5401                                         } else if (th->fin || sock_owned_by_user(sk)) {
5402                                                 /* Bad case. We could lose such FIN otherwise.
5403                                                  * It is not a big problem, but it looks confusing
5404                                                  * and not so rare event. We still can lose it now,
5405                                                  * if it spins in bh_lock_sock(), but it is really
5406                                                  * marginal case.
5407                                                  */
5408                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5409                                         } else {
5410                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5411                                                 goto discard;
5412                                         }
5413                                 }
5414                         }
5415                         break;
5416
5417                 case TCP_CLOSING:
5418                         if (tp->snd_una == tp->write_seq) {
5419                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5420                                 goto discard;
5421                         }
5422                         break;
5423
5424                 case TCP_LAST_ACK:
5425                         if (tp->snd_una == tp->write_seq) {
5426                                 tcp_update_metrics(sk);
5427                                 tcp_done(sk);
5428                                 goto discard;
5429                         }
5430                         break;
5431                 }
5432         } else
5433                 goto discard;
5434
5435         /* step 6: check the URG bit */
5436         tcp_urg(sk, skb, th);
5437
5438         /* step 7: process the segment text */
5439         switch (sk->sk_state) {
5440         case TCP_CLOSE_WAIT:
5441         case TCP_CLOSING:
5442         case TCP_LAST_ACK:
5443                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5444                         break;
5445         case TCP_FIN_WAIT1:
5446         case TCP_FIN_WAIT2:
5447                 /* RFC 793 says to queue data in these states,
5448                  * RFC 1122 says we MUST send a reset.
5449                  * BSD 4.4 also does reset.
5450                  */
5451                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5452                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5453                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5454                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5455                                 tcp_reset(sk);
5456                                 return 1;
5457                         }
5458                 }
5459                 /* Fall through */
5460         case TCP_ESTABLISHED:
5461                 tcp_data_queue(sk, skb);
5462                 queued = 1;
5463                 break;
5464         }
5465
5466         /* tcp_data could move socket to TIME-WAIT */
5467         if (sk->sk_state != TCP_CLOSE) {
5468                 tcp_data_snd_check(sk);
5469                 tcp_ack_snd_check(sk);
5470         }
5471
5472         if (!queued) {
5473 discard:
5474                 __kfree_skb(skb);
5475         }
5476         return 0;
5477 }
5478
5479 EXPORT_SYMBOL(sysctl_tcp_ecn);
5480 EXPORT_SYMBOL(sysctl_tcp_reordering);
5481 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5482 EXPORT_SYMBOL(tcp_parse_options);
5483 #ifdef CONFIG_TCP_MD5SIG
5484 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5485 #endif
5486 EXPORT_SYMBOL(tcp_rcv_established);
5487 EXPORT_SYMBOL(tcp_rcv_state_process);
5488 EXPORT_SYMBOL(tcp_initialize_rcv_mss);