net: replace uses of NIP6_FMT with %p6
[linux-2.6.git] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * 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 must be called before lost_out is incremented */
983 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
984 {
985         if ((tp->retransmit_skb_hint == NULL) ||
986             before(TCP_SKB_CB(skb)->seq,
987                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
988                 tp->retransmit_skb_hint = skb;
989
990         if (!tp->lost_out ||
991             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
992                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
993 }
994
995 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
996 {
997         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
998                 tcp_verify_retransmit_hint(tp, skb);
999
1000                 tp->lost_out += tcp_skb_pcount(skb);
1001                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1002         }
1003 }
1004
1005 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
1006 {
1007         tcp_verify_retransmit_hint(tp, skb);
1008
1009         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1010                 tp->lost_out += tcp_skb_pcount(skb);
1011                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1012         }
1013 }
1014
1015 /* This procedure tags the retransmission queue when SACKs arrive.
1016  *
1017  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1018  * Packets in queue with these bits set are counted in variables
1019  * sacked_out, retrans_out and lost_out, correspondingly.
1020  *
1021  * Valid combinations are:
1022  * Tag  InFlight        Description
1023  * 0    1               - orig segment is in flight.
1024  * S    0               - nothing flies, orig reached receiver.
1025  * L    0               - nothing flies, orig lost by net.
1026  * R    2               - both orig and retransmit are in flight.
1027  * L|R  1               - orig is lost, retransmit is in flight.
1028  * S|R  1               - orig reached receiver, retrans is still in flight.
1029  * (L|S|R is logically valid, it could occur when L|R is sacked,
1030  *  but it is equivalent to plain S and code short-curcuits it to S.
1031  *  L|S is logically invalid, it would mean -1 packet in flight 8))
1032  *
1033  * These 6 states form finite state machine, controlled by the following events:
1034  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1035  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1036  * 3. Loss detection event of one of three flavors:
1037  *      A. Scoreboard estimator decided the packet is lost.
1038  *         A'. Reno "three dupacks" marks head of queue lost.
1039  *         A''. Its FACK modfication, head until snd.fack is lost.
1040  *      B. SACK arrives sacking data transmitted after never retransmitted
1041  *         hole was sent out.
1042  *      C. SACK arrives sacking SND.NXT at the moment, when the
1043  *         segment was retransmitted.
1044  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1045  *
1046  * It is pleasant to note, that state diagram turns out to be commutative,
1047  * so that we are allowed not to be bothered by order of our actions,
1048  * when multiple events arrive simultaneously. (see the function below).
1049  *
1050  * Reordering detection.
1051  * --------------------
1052  * Reordering metric is maximal distance, which a packet can be displaced
1053  * in packet stream. With SACKs we can estimate it:
1054  *
1055  * 1. SACK fills old hole and the corresponding segment was not
1056  *    ever retransmitted -> reordering. Alas, we cannot use it
1057  *    when segment was retransmitted.
1058  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1059  *    for retransmitted and already SACKed segment -> reordering..
1060  * Both of these heuristics are not used in Loss state, when we cannot
1061  * account for retransmits accurately.
1062  *
1063  * SACK block validation.
1064  * ----------------------
1065  *
1066  * SACK block range validation checks that the received SACK block fits to
1067  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1068  * Note that SND.UNA is not included to the range though being valid because
1069  * it means that the receiver is rather inconsistent with itself reporting
1070  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1071  * perfectly valid, however, in light of RFC2018 which explicitly states
1072  * that "SACK block MUST reflect the newest segment.  Even if the newest
1073  * segment is going to be discarded ...", not that it looks very clever
1074  * in case of head skb. Due to potentional receiver driven attacks, we
1075  * choose to avoid immediate execution of a walk in write queue due to
1076  * reneging and defer head skb's loss recovery to standard loss recovery
1077  * procedure that will eventually trigger (nothing forbids us doing this).
1078  *
1079  * Implements also blockage to start_seq wrap-around. Problem lies in the
1080  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1081  * there's no guarantee that it will be before snd_nxt (n). The problem
1082  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1083  * wrap (s_w):
1084  *
1085  *         <- outs wnd ->                          <- wrapzone ->
1086  *         u     e      n                         u_w   e_w  s n_w
1087  *         |     |      |                          |     |   |  |
1088  * |<------------+------+----- TCP seqno space --------------+---------->|
1089  * ...-- <2^31 ->|                                           |<--------...
1090  * ...---- >2^31 ------>|                                    |<--------...
1091  *
1092  * Current code wouldn't be vulnerable but it's better still to discard such
1093  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1094  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1095  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1096  * equal to the ideal case (infinite seqno space without wrap caused issues).
1097  *
1098  * With D-SACK the lower bound is extended to cover sequence space below
1099  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1100  * again, D-SACK block must not to go across snd_una (for the same reason as
1101  * for the normal SACK blocks, explained above). But there all simplicity
1102  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1103  * fully below undo_marker they do not affect behavior in anyway and can
1104  * therefore be safely ignored. In rare cases (which are more or less
1105  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1106  * fragmentation and packet reordering past skb's retransmission. To consider
1107  * them correctly, the acceptable range must be extended even more though
1108  * the exact amount is rather hard to quantify. However, tp->max_window can
1109  * be used as an exaggerated estimate.
1110  */
1111 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1112                                   u32 start_seq, u32 end_seq)
1113 {
1114         /* Too far in future, or reversed (interpretation is ambiguous) */
1115         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1116                 return 0;
1117
1118         /* Nasty start_seq wrap-around check (see comments above) */
1119         if (!before(start_seq, tp->snd_nxt))
1120                 return 0;
1121
1122         /* In outstanding window? ...This is valid exit for D-SACKs too.
1123          * start_seq == snd_una is non-sensical (see comments above)
1124          */
1125         if (after(start_seq, tp->snd_una))
1126                 return 1;
1127
1128         if (!is_dsack || !tp->undo_marker)
1129                 return 0;
1130
1131         /* ...Then it's D-SACK, and must reside below snd_una completely */
1132         if (!after(end_seq, tp->snd_una))
1133                 return 0;
1134
1135         if (!before(start_seq, tp->undo_marker))
1136                 return 1;
1137
1138         /* Too old */
1139         if (!after(end_seq, tp->undo_marker))
1140                 return 0;
1141
1142         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1143          *   start_seq < undo_marker and end_seq >= undo_marker.
1144          */
1145         return !before(start_seq, end_seq - tp->max_window);
1146 }
1147
1148 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1149  * Event "C". Later note: FACK people cheated me again 8), we have to account
1150  * for reordering! Ugly, but should help.
1151  *
1152  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1153  * less than what is now known to be received by the other end (derived from
1154  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1155  * retransmitted skbs to avoid some costly processing per ACKs.
1156  */
1157 static void tcp_mark_lost_retrans(struct sock *sk)
1158 {
1159         const struct inet_connection_sock *icsk = inet_csk(sk);
1160         struct tcp_sock *tp = tcp_sk(sk);
1161         struct sk_buff *skb;
1162         int cnt = 0;
1163         u32 new_low_seq = tp->snd_nxt;
1164         u32 received_upto = tcp_highest_sack_seq(tp);
1165
1166         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1167             !after(received_upto, tp->lost_retrans_low) ||
1168             icsk->icsk_ca_state != TCP_CA_Recovery)
1169                 return;
1170
1171         tcp_for_write_queue(skb, sk) {
1172                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1173
1174                 if (skb == tcp_send_head(sk))
1175                         break;
1176                 if (cnt == tp->retrans_out)
1177                         break;
1178                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1179                         continue;
1180
1181                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1182                         continue;
1183
1184                 if (after(received_upto, ack_seq) &&
1185                     (tcp_is_fack(tp) ||
1186                      !before(received_upto,
1187                              ack_seq + tp->reordering * tp->mss_cache))) {
1188                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1189                         tp->retrans_out -= tcp_skb_pcount(skb);
1190
1191                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1192                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1193                 } else {
1194                         if (before(ack_seq, new_low_seq))
1195                                 new_low_seq = ack_seq;
1196                         cnt += tcp_skb_pcount(skb);
1197                 }
1198         }
1199
1200         if (tp->retrans_out)
1201                 tp->lost_retrans_low = new_low_seq;
1202 }
1203
1204 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1205                            struct tcp_sack_block_wire *sp, int num_sacks,
1206                            u32 prior_snd_una)
1207 {
1208         struct tcp_sock *tp = tcp_sk(sk);
1209         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1210         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1211         int dup_sack = 0;
1212
1213         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1214                 dup_sack = 1;
1215                 tcp_dsack_seen(tp);
1216                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1217         } else if (num_sacks > 1) {
1218                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1219                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1220
1221                 if (!after(end_seq_0, end_seq_1) &&
1222                     !before(start_seq_0, start_seq_1)) {
1223                         dup_sack = 1;
1224                         tcp_dsack_seen(tp);
1225                         NET_INC_STATS_BH(sock_net(sk),
1226                                         LINUX_MIB_TCPDSACKOFORECV);
1227                 }
1228         }
1229
1230         /* D-SACK for already forgotten data... Do dumb counting. */
1231         if (dup_sack &&
1232             !after(end_seq_0, prior_snd_una) &&
1233             after(end_seq_0, tp->undo_marker))
1234                 tp->undo_retrans--;
1235
1236         return dup_sack;
1237 }
1238
1239 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1240  * the incoming SACK may not exactly match but we can find smaller MSS
1241  * aligned portion of it that matches. Therefore we might need to fragment
1242  * which may fail and creates some hassle (caller must handle error case
1243  * returns).
1244  */
1245 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1246                                  u32 start_seq, u32 end_seq)
1247 {
1248         int in_sack, err;
1249         unsigned int pkt_len;
1250
1251         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1252                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1253
1254         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1255             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1256
1257                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1258
1259                 if (!in_sack)
1260                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1261                 else
1262                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1263                 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1264                 if (err < 0)
1265                         return err;
1266         }
1267
1268         return in_sack;
1269 }
1270
1271 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1272                            int *reord, int dup_sack, int fack_count)
1273 {
1274         struct tcp_sock *tp = tcp_sk(sk);
1275         u8 sacked = TCP_SKB_CB(skb)->sacked;
1276         int flag = 0;
1277
1278         /* Account D-SACK for retransmitted packet. */
1279         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1280                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1281                         tp->undo_retrans--;
1282                 if (sacked & TCPCB_SACKED_ACKED)
1283                         *reord = min(fack_count, *reord);
1284         }
1285
1286         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1287         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1288                 return flag;
1289
1290         if (!(sacked & TCPCB_SACKED_ACKED)) {
1291                 if (sacked & TCPCB_SACKED_RETRANS) {
1292                         /* If the segment is not tagged as lost,
1293                          * we do not clear RETRANS, believing
1294                          * that retransmission is still in flight.
1295                          */
1296                         if (sacked & TCPCB_LOST) {
1297                                 TCP_SKB_CB(skb)->sacked &=
1298                                         ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1299                                 tp->lost_out -= tcp_skb_pcount(skb);
1300                                 tp->retrans_out -= tcp_skb_pcount(skb);
1301                         }
1302                 } else {
1303                         if (!(sacked & TCPCB_RETRANS)) {
1304                                 /* New sack for not retransmitted frame,
1305                                  * which was in hole. It is reordering.
1306                                  */
1307                                 if (before(TCP_SKB_CB(skb)->seq,
1308                                            tcp_highest_sack_seq(tp)))
1309                                         *reord = min(fack_count, *reord);
1310
1311                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1312                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1313                                         flag |= FLAG_ONLY_ORIG_SACKED;
1314                         }
1315
1316                         if (sacked & TCPCB_LOST) {
1317                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1318                                 tp->lost_out -= tcp_skb_pcount(skb);
1319                         }
1320                 }
1321
1322                 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1323                 flag |= FLAG_DATA_SACKED;
1324                 tp->sacked_out += tcp_skb_pcount(skb);
1325
1326                 fack_count += tcp_skb_pcount(skb);
1327
1328                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1329                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1330                     before(TCP_SKB_CB(skb)->seq,
1331                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1332                         tp->lost_cnt_hint += tcp_skb_pcount(skb);
1333
1334                 if (fack_count > tp->fackets_out)
1335                         tp->fackets_out = fack_count;
1336
1337                 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1338                         tcp_advance_highest_sack(sk, skb);
1339         }
1340
1341         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1342          * frames and clear it. undo_retrans is decreased above, L|R frames
1343          * are accounted above as well.
1344          */
1345         if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) {
1346                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1347                 tp->retrans_out -= tcp_skb_pcount(skb);
1348         }
1349
1350         return flag;
1351 }
1352
1353 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1354                                         struct tcp_sack_block *next_dup,
1355                                         u32 start_seq, u32 end_seq,
1356                                         int dup_sack_in, int *fack_count,
1357                                         int *reord, int *flag)
1358 {
1359         tcp_for_write_queue_from(skb, sk) {
1360                 int in_sack = 0;
1361                 int dup_sack = dup_sack_in;
1362
1363                 if (skb == tcp_send_head(sk))
1364                         break;
1365
1366                 /* queue is in-order => we can short-circuit the walk early */
1367                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1368                         break;
1369
1370                 if ((next_dup != NULL) &&
1371                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1372                         in_sack = tcp_match_skb_to_sack(sk, skb,
1373                                                         next_dup->start_seq,
1374                                                         next_dup->end_seq);
1375                         if (in_sack > 0)
1376                                 dup_sack = 1;
1377                 }
1378
1379                 if (in_sack <= 0)
1380                         in_sack = tcp_match_skb_to_sack(sk, skb, start_seq,
1381                                                         end_seq);
1382                 if (unlikely(in_sack < 0))
1383                         break;
1384
1385                 if (in_sack)
1386                         *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack,
1387                                                  *fack_count);
1388
1389                 *fack_count += tcp_skb_pcount(skb);
1390         }
1391         return skb;
1392 }
1393
1394 /* Avoid all extra work that is being done by sacktag while walking in
1395  * a normal way
1396  */
1397 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1398                                         u32 skip_to_seq, int *fack_count)
1399 {
1400         tcp_for_write_queue_from(skb, sk) {
1401                 if (skb == tcp_send_head(sk))
1402                         break;
1403
1404                 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1405                         break;
1406
1407                 *fack_count += tcp_skb_pcount(skb);
1408         }
1409         return skb;
1410 }
1411
1412 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1413                                                 struct sock *sk,
1414                                                 struct tcp_sack_block *next_dup,
1415                                                 u32 skip_to_seq,
1416                                                 int *fack_count, int *reord,
1417                                                 int *flag)
1418 {
1419         if (next_dup == NULL)
1420                 return skb;
1421
1422         if (before(next_dup->start_seq, skip_to_seq)) {
1423                 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count);
1424                 skb = tcp_sacktag_walk(skb, sk, NULL,
1425                                      next_dup->start_seq, next_dup->end_seq,
1426                                      1, fack_count, reord, flag);
1427         }
1428
1429         return skb;
1430 }
1431
1432 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1433 {
1434         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1435 }
1436
1437 static int
1438 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1439                         u32 prior_snd_una)
1440 {
1441         const struct inet_connection_sock *icsk = inet_csk(sk);
1442         struct tcp_sock *tp = tcp_sk(sk);
1443         unsigned char *ptr = (skb_transport_header(ack_skb) +
1444                               TCP_SKB_CB(ack_skb)->sacked);
1445         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1446         struct tcp_sack_block sp[TCP_NUM_SACKS];
1447         struct tcp_sack_block *cache;
1448         struct sk_buff *skb;
1449         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1450         int used_sacks;
1451         int reord = tp->packets_out;
1452         int flag = 0;
1453         int found_dup_sack = 0;
1454         int fack_count;
1455         int i, j;
1456         int first_sack_index;
1457
1458         if (!tp->sacked_out) {
1459                 if (WARN_ON(tp->fackets_out))
1460                         tp->fackets_out = 0;
1461                 tcp_highest_sack_reset(sk);
1462         }
1463
1464         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1465                                          num_sacks, prior_snd_una);
1466         if (found_dup_sack)
1467                 flag |= FLAG_DSACKING_ACK;
1468
1469         /* Eliminate too old ACKs, but take into
1470          * account more or less fresh ones, they can
1471          * contain valid SACK info.
1472          */
1473         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1474                 return 0;
1475
1476         if (!tp->packets_out)
1477                 goto out;
1478
1479         used_sacks = 0;
1480         first_sack_index = 0;
1481         for (i = 0; i < num_sacks; i++) {
1482                 int dup_sack = !i && found_dup_sack;
1483
1484                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1485                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1486
1487                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1488                                             sp[used_sacks].start_seq,
1489                                             sp[used_sacks].end_seq)) {
1490                         int mib_idx;
1491
1492                         if (dup_sack) {
1493                                 if (!tp->undo_marker)
1494                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1495                                 else
1496                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1497                         } else {
1498                                 /* Don't count olds caused by ACK reordering */
1499                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1500                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1501                                         continue;
1502                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1503                         }
1504
1505                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1506                         if (i == 0)
1507                                 first_sack_index = -1;
1508                         continue;
1509                 }
1510
1511                 /* Ignore very old stuff early */
1512                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1513                         continue;
1514
1515                 used_sacks++;
1516         }
1517
1518         /* order SACK blocks to allow in order walk of the retrans queue */
1519         for (i = used_sacks - 1; i > 0; i--) {
1520                 for (j = 0; j < i; j++) {
1521                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1522                                 struct tcp_sack_block tmp;
1523
1524                                 tmp = sp[j];
1525                                 sp[j] = sp[j + 1];
1526                                 sp[j + 1] = tmp;
1527
1528                                 /* Track where the first SACK block goes to */
1529                                 if (j == first_sack_index)
1530                                         first_sack_index = j + 1;
1531                         }
1532                 }
1533         }
1534
1535         skb = tcp_write_queue_head(sk);
1536         fack_count = 0;
1537         i = 0;
1538
1539         if (!tp->sacked_out) {
1540                 /* It's already past, so skip checking against it */
1541                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1542         } else {
1543                 cache = tp->recv_sack_cache;
1544                 /* Skip empty blocks in at head of the cache */
1545                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1546                        !cache->end_seq)
1547                         cache++;
1548         }
1549
1550         while (i < used_sacks) {
1551                 u32 start_seq = sp[i].start_seq;
1552                 u32 end_seq = sp[i].end_seq;
1553                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1554                 struct tcp_sack_block *next_dup = NULL;
1555
1556                 if (found_dup_sack && ((i + 1) == first_sack_index))
1557                         next_dup = &sp[i + 1];
1558
1559                 /* Event "B" in the comment above. */
1560                 if (after(end_seq, tp->high_seq))
1561                         flag |= FLAG_DATA_LOST;
1562
1563                 /* Skip too early cached blocks */
1564                 while (tcp_sack_cache_ok(tp, cache) &&
1565                        !before(start_seq, cache->end_seq))
1566                         cache++;
1567
1568                 /* Can skip some work by looking recv_sack_cache? */
1569                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1570                     after(end_seq, cache->start_seq)) {
1571
1572                         /* Head todo? */
1573                         if (before(start_seq, cache->start_seq)) {
1574                                 skb = tcp_sacktag_skip(skb, sk, start_seq,
1575                                                        &fack_count);
1576                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1577                                                        start_seq,
1578                                                        cache->start_seq,
1579                                                        dup_sack, &fack_count,
1580                                                        &reord, &flag);
1581                         }
1582
1583                         /* Rest of the block already fully processed? */
1584                         if (!after(end_seq, cache->end_seq))
1585                                 goto advance_sp;
1586
1587                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1588                                                        cache->end_seq,
1589                                                        &fack_count, &reord,
1590                                                        &flag);
1591
1592                         /* ...tail remains todo... */
1593                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1594                                 /* ...but better entrypoint exists! */
1595                                 skb = tcp_highest_sack(sk);
1596                                 if (skb == NULL)
1597                                         break;
1598                                 fack_count = tp->fackets_out;
1599                                 cache++;
1600                                 goto walk;
1601                         }
1602
1603                         skb = tcp_sacktag_skip(skb, sk, cache->end_seq,
1604                                                &fack_count);
1605                         /* Check overlap against next cached too (past this one already) */
1606                         cache++;
1607                         continue;
1608                 }
1609
1610                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1611                         skb = tcp_highest_sack(sk);
1612                         if (skb == NULL)
1613                                 break;
1614                         fack_count = tp->fackets_out;
1615                 }
1616                 skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count);
1617
1618 walk:
1619                 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1620                                        dup_sack, &fack_count, &reord, &flag);
1621
1622 advance_sp:
1623                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1624                  * due to in-order walk
1625                  */
1626                 if (after(end_seq, tp->frto_highmark))
1627                         flag &= ~FLAG_ONLY_ORIG_SACKED;
1628
1629                 i++;
1630         }
1631
1632         /* Clear the head of the cache sack blocks so we can skip it next time */
1633         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1634                 tp->recv_sack_cache[i].start_seq = 0;
1635                 tp->recv_sack_cache[i].end_seq = 0;
1636         }
1637         for (j = 0; j < used_sacks; j++)
1638                 tp->recv_sack_cache[i++] = sp[j];
1639
1640         tcp_mark_lost_retrans(sk);
1641
1642         tcp_verify_left_out(tp);
1643
1644         if ((reord < tp->fackets_out) &&
1645             ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1646             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1647                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1648
1649 out:
1650
1651 #if FASTRETRANS_DEBUG > 0
1652         WARN_ON((int)tp->sacked_out < 0);
1653         WARN_ON((int)tp->lost_out < 0);
1654         WARN_ON((int)tp->retrans_out < 0);
1655         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1656 #endif
1657         return flag;
1658 }
1659
1660 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1661  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1662  */
1663 int tcp_limit_reno_sacked(struct tcp_sock *tp)
1664 {
1665         u32 holes;
1666
1667         holes = max(tp->lost_out, 1U);
1668         holes = min(holes, tp->packets_out);
1669
1670         if ((tp->sacked_out + holes) > tp->packets_out) {
1671                 tp->sacked_out = tp->packets_out - holes;
1672                 return 1;
1673         }
1674         return 0;
1675 }
1676
1677 /* If we receive more dupacks than we expected counting segments
1678  * in assumption of absent reordering, interpret this as reordering.
1679  * The only another reason could be bug in receiver TCP.
1680  */
1681 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1682 {
1683         struct tcp_sock *tp = tcp_sk(sk);
1684         if (tcp_limit_reno_sacked(tp))
1685                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1686 }
1687
1688 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1689
1690 static void tcp_add_reno_sack(struct sock *sk)
1691 {
1692         struct tcp_sock *tp = tcp_sk(sk);
1693         tp->sacked_out++;
1694         tcp_check_reno_reordering(sk, 0);
1695         tcp_verify_left_out(tp);
1696 }
1697
1698 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1699
1700 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1701 {
1702         struct tcp_sock *tp = tcp_sk(sk);
1703
1704         if (acked > 0) {
1705                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1706                 if (acked - 1 >= tp->sacked_out)
1707                         tp->sacked_out = 0;
1708                 else
1709                         tp->sacked_out -= acked - 1;
1710         }
1711         tcp_check_reno_reordering(sk, acked);
1712         tcp_verify_left_out(tp);
1713 }
1714
1715 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1716 {
1717         tp->sacked_out = 0;
1718 }
1719
1720 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1721 {
1722         return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1723 }
1724
1725 /* F-RTO can only be used if TCP has never retransmitted anything other than
1726  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1727  */
1728 int tcp_use_frto(struct sock *sk)
1729 {
1730         const struct tcp_sock *tp = tcp_sk(sk);
1731         const struct inet_connection_sock *icsk = inet_csk(sk);
1732         struct sk_buff *skb;
1733
1734         if (!sysctl_tcp_frto)
1735                 return 0;
1736
1737         /* MTU probe and F-RTO won't really play nicely along currently */
1738         if (icsk->icsk_mtup.probe_size)
1739                 return 0;
1740
1741         if (tcp_is_sackfrto(tp))
1742                 return 1;
1743
1744         /* Avoid expensive walking of rexmit queue if possible */
1745         if (tp->retrans_out > 1)
1746                 return 0;
1747
1748         skb = tcp_write_queue_head(sk);
1749         if (tcp_skb_is_last(sk, skb))
1750                 return 1;
1751         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
1752         tcp_for_write_queue_from(skb, sk) {
1753                 if (skb == tcp_send_head(sk))
1754                         break;
1755                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1756                         return 0;
1757                 /* Short-circuit when first non-SACKed skb has been checked */
1758                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1759                         break;
1760         }
1761         return 1;
1762 }
1763
1764 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1765  * recovery a bit and use heuristics in tcp_process_frto() to detect if
1766  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1767  * keep retrans_out counting accurate (with SACK F-RTO, other than head
1768  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1769  * bits are handled if the Loss state is really to be entered (in
1770  * tcp_enter_frto_loss).
1771  *
1772  * Do like tcp_enter_loss() would; when RTO expires the second time it
1773  * does:
1774  *  "Reduce ssthresh if it has not yet been made inside this window."
1775  */
1776 void tcp_enter_frto(struct sock *sk)
1777 {
1778         const struct inet_connection_sock *icsk = inet_csk(sk);
1779         struct tcp_sock *tp = tcp_sk(sk);
1780         struct sk_buff *skb;
1781
1782         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1783             tp->snd_una == tp->high_seq ||
1784             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1785              !icsk->icsk_retransmits)) {
1786                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1787                 /* Our state is too optimistic in ssthresh() call because cwnd
1788                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1789                  * recovery has not yet completed. Pattern would be this: RTO,
1790                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
1791                  * up here twice).
1792                  * RFC4138 should be more specific on what to do, even though
1793                  * RTO is quite unlikely to occur after the first Cumulative ACK
1794                  * due to back-off and complexity of triggering events ...
1795                  */
1796                 if (tp->frto_counter) {
1797                         u32 stored_cwnd;
1798                         stored_cwnd = tp->snd_cwnd;
1799                         tp->snd_cwnd = 2;
1800                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1801                         tp->snd_cwnd = stored_cwnd;
1802                 } else {
1803                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1804                 }
1805                 /* ... in theory, cong.control module could do "any tricks" in
1806                  * ssthresh(), which means that ca_state, lost bits and lost_out
1807                  * counter would have to be faked before the call occurs. We
1808                  * consider that too expensive, unlikely and hacky, so modules
1809                  * using these in ssthresh() must deal these incompatibility
1810                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1811                  */
1812                 tcp_ca_event(sk, CA_EVENT_FRTO);
1813         }
1814
1815         tp->undo_marker = tp->snd_una;
1816         tp->undo_retrans = 0;
1817
1818         skb = tcp_write_queue_head(sk);
1819         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1820                 tp->undo_marker = 0;
1821         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1822                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1823                 tp->retrans_out -= tcp_skb_pcount(skb);
1824         }
1825         tcp_verify_left_out(tp);
1826
1827         /* Too bad if TCP was application limited */
1828         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1829
1830         /* Earlier loss recovery underway (see RFC4138; Appendix B).
1831          * The last condition is necessary at least in tp->frto_counter case.
1832          */
1833         if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
1834             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1835             after(tp->high_seq, tp->snd_una)) {
1836                 tp->frto_highmark = tp->high_seq;
1837         } else {
1838                 tp->frto_highmark = tp->snd_nxt;
1839         }
1840         tcp_set_ca_state(sk, TCP_CA_Disorder);
1841         tp->high_seq = tp->snd_nxt;
1842         tp->frto_counter = 1;
1843 }
1844
1845 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1846  * which indicates that we should follow the traditional RTO recovery,
1847  * i.e. mark everything lost and do go-back-N retransmission.
1848  */
1849 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1850 {
1851         struct tcp_sock *tp = tcp_sk(sk);
1852         struct sk_buff *skb;
1853
1854         tp->lost_out = 0;
1855         tp->retrans_out = 0;
1856         if (tcp_is_reno(tp))
1857                 tcp_reset_reno_sack(tp);
1858
1859         tcp_for_write_queue(skb, sk) {
1860                 if (skb == tcp_send_head(sk))
1861                         break;
1862
1863                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1864                 /*
1865                  * Count the retransmission made on RTO correctly (only when
1866                  * waiting for the first ACK and did not get it)...
1867                  */
1868                 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1869                         /* For some reason this R-bit might get cleared? */
1870                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1871                                 tp->retrans_out += tcp_skb_pcount(skb);
1872                         /* ...enter this if branch just for the first segment */
1873                         flag |= FLAG_DATA_ACKED;
1874                 } else {
1875                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1876                                 tp->undo_marker = 0;
1877                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1878                 }
1879
1880                 /* Marking forward transmissions that were made after RTO lost
1881                  * can cause unnecessary retransmissions in some scenarios,
1882                  * SACK blocks will mitigate that in some but not in all cases.
1883                  * We used to not mark them but it was causing break-ups with
1884                  * receivers that do only in-order receival.
1885                  *
1886                  * TODO: we could detect presence of such receiver and select
1887                  * different behavior per flow.
1888                  */
1889                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1890                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1891                         tp->lost_out += tcp_skb_pcount(skb);
1892                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1893                 }
1894         }
1895         tcp_verify_left_out(tp);
1896
1897         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1898         tp->snd_cwnd_cnt = 0;
1899         tp->snd_cwnd_stamp = tcp_time_stamp;
1900         tp->frto_counter = 0;
1901         tp->bytes_acked = 0;
1902
1903         tp->reordering = min_t(unsigned int, tp->reordering,
1904                                sysctl_tcp_reordering);
1905         tcp_set_ca_state(sk, TCP_CA_Loss);
1906         tp->high_seq = tp->snd_nxt;
1907         TCP_ECN_queue_cwr(tp);
1908
1909         tcp_clear_all_retrans_hints(tp);
1910 }
1911
1912 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1913 {
1914         tp->retrans_out = 0;
1915         tp->lost_out = 0;
1916
1917         tp->undo_marker = 0;
1918         tp->undo_retrans = 0;
1919 }
1920
1921 void tcp_clear_retrans(struct tcp_sock *tp)
1922 {
1923         tcp_clear_retrans_partial(tp);
1924
1925         tp->fackets_out = 0;
1926         tp->sacked_out = 0;
1927 }
1928
1929 /* Enter Loss state. If "how" is not zero, forget all SACK information
1930  * and reset tags completely, otherwise preserve SACKs. If receiver
1931  * dropped its ofo queue, we will know this due to reneging detection.
1932  */
1933 void tcp_enter_loss(struct sock *sk, int how)
1934 {
1935         const struct inet_connection_sock *icsk = inet_csk(sk);
1936         struct tcp_sock *tp = tcp_sk(sk);
1937         struct sk_buff *skb;
1938
1939         /* Reduce ssthresh if it has not yet been made inside this window. */
1940         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1941             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1942                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1943                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1944                 tcp_ca_event(sk, CA_EVENT_LOSS);
1945         }
1946         tp->snd_cwnd       = 1;
1947         tp->snd_cwnd_cnt   = 0;
1948         tp->snd_cwnd_stamp = tcp_time_stamp;
1949
1950         tp->bytes_acked = 0;
1951         tcp_clear_retrans_partial(tp);
1952
1953         if (tcp_is_reno(tp))
1954                 tcp_reset_reno_sack(tp);
1955
1956         if (!how) {
1957                 /* Push undo marker, if it was plain RTO and nothing
1958                  * was retransmitted. */
1959                 tp->undo_marker = tp->snd_una;
1960         } else {
1961                 tp->sacked_out = 0;
1962                 tp->fackets_out = 0;
1963         }
1964         tcp_clear_all_retrans_hints(tp);
1965
1966         tcp_for_write_queue(skb, sk) {
1967                 if (skb == tcp_send_head(sk))
1968                         break;
1969
1970                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1971                         tp->undo_marker = 0;
1972                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1973                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1974                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1975                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1976                         tp->lost_out += tcp_skb_pcount(skb);
1977                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1978                 }
1979         }
1980         tcp_verify_left_out(tp);
1981
1982         tp->reordering = min_t(unsigned int, tp->reordering,
1983                                sysctl_tcp_reordering);
1984         tcp_set_ca_state(sk, TCP_CA_Loss);
1985         tp->high_seq = tp->snd_nxt;
1986         TCP_ECN_queue_cwr(tp);
1987         /* Abort F-RTO algorithm if one is in progress */
1988         tp->frto_counter = 0;
1989 }
1990
1991 /* If ACK arrived pointing to a remembered SACK, it means that our
1992  * remembered SACKs do not reflect real state of receiver i.e.
1993  * receiver _host_ is heavily congested (or buggy).
1994  *
1995  * Do processing similar to RTO timeout.
1996  */
1997 static int tcp_check_sack_reneging(struct sock *sk, int flag)
1998 {
1999         if (flag & FLAG_SACK_RENEGING) {
2000                 struct inet_connection_sock *icsk = inet_csk(sk);
2001                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2002
2003                 tcp_enter_loss(sk, 1);
2004                 icsk->icsk_retransmits++;
2005                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2006                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2007                                           icsk->icsk_rto, TCP_RTO_MAX);
2008                 return 1;
2009         }
2010         return 0;
2011 }
2012
2013 static inline int tcp_fackets_out(struct tcp_sock *tp)
2014 {
2015         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2016 }
2017
2018 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2019  * counter when SACK is enabled (without SACK, sacked_out is used for
2020  * that purpose).
2021  *
2022  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2023  * segments up to the highest received SACK block so far and holes in
2024  * between them.
2025  *
2026  * With reordering, holes may still be in flight, so RFC3517 recovery
2027  * uses pure sacked_out (total number of SACKed segments) even though
2028  * it violates the RFC that uses duplicate ACKs, often these are equal
2029  * but when e.g. out-of-window ACKs or packet duplication occurs,
2030  * they differ. Since neither occurs due to loss, TCP should really
2031  * ignore them.
2032  */
2033 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2034 {
2035         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2036 }
2037
2038 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2039 {
2040         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2041 }
2042
2043 static inline int tcp_head_timedout(struct sock *sk)
2044 {
2045         struct tcp_sock *tp = tcp_sk(sk);
2046
2047         return tp->packets_out &&
2048                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2049 }
2050
2051 /* Linux NewReno/SACK/FACK/ECN state machine.
2052  * --------------------------------------
2053  *
2054  * "Open"       Normal state, no dubious events, fast path.
2055  * "Disorder"   In all the respects it is "Open",
2056  *              but requires a bit more attention. It is entered when
2057  *              we see some SACKs or dupacks. It is split of "Open"
2058  *              mainly to move some processing from fast path to slow one.
2059  * "CWR"        CWND was reduced due to some Congestion Notification event.
2060  *              It can be ECN, ICMP source quench, local device congestion.
2061  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2062  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2063  *
2064  * tcp_fastretrans_alert() is entered:
2065  * - each incoming ACK, if state is not "Open"
2066  * - when arrived ACK is unusual, namely:
2067  *      * SACK
2068  *      * Duplicate ACK.
2069  *      * ECN ECE.
2070  *
2071  * Counting packets in flight is pretty simple.
2072  *
2073  *      in_flight = packets_out - left_out + retrans_out
2074  *
2075  *      packets_out is SND.NXT-SND.UNA counted in packets.
2076  *
2077  *      retrans_out is number of retransmitted segments.
2078  *
2079  *      left_out is number of segments left network, but not ACKed yet.
2080  *
2081  *              left_out = sacked_out + lost_out
2082  *
2083  *     sacked_out: Packets, which arrived to receiver out of order
2084  *                 and hence not ACKed. With SACKs this number is simply
2085  *                 amount of SACKed data. Even without SACKs
2086  *                 it is easy to give pretty reliable estimate of this number,
2087  *                 counting duplicate ACKs.
2088  *
2089  *       lost_out: Packets lost by network. TCP has no explicit
2090  *                 "loss notification" feedback from network (for now).
2091  *                 It means that this number can be only _guessed_.
2092  *                 Actually, it is the heuristics to predict lossage that
2093  *                 distinguishes different algorithms.
2094  *
2095  *      F.e. after RTO, when all the queue is considered as lost,
2096  *      lost_out = packets_out and in_flight = retrans_out.
2097  *
2098  *              Essentially, we have now two algorithms counting
2099  *              lost packets.
2100  *
2101  *              FACK: It is the simplest heuristics. As soon as we decided
2102  *              that something is lost, we decide that _all_ not SACKed
2103  *              packets until the most forward SACK are lost. I.e.
2104  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2105  *              It is absolutely correct estimate, if network does not reorder
2106  *              packets. And it loses any connection to reality when reordering
2107  *              takes place. We use FACK by default until reordering
2108  *              is suspected on the path to this destination.
2109  *
2110  *              NewReno: when Recovery is entered, we assume that one segment
2111  *              is lost (classic Reno). While we are in Recovery and
2112  *              a partial ACK arrives, we assume that one more packet
2113  *              is lost (NewReno). This heuristics are the same in NewReno
2114  *              and SACK.
2115  *
2116  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2117  *  deflation etc. CWND is real congestion window, never inflated, changes
2118  *  only according to classic VJ rules.
2119  *
2120  * Really tricky (and requiring careful tuning) part of algorithm
2121  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2122  * The first determines the moment _when_ we should reduce CWND and,
2123  * hence, slow down forward transmission. In fact, it determines the moment
2124  * when we decide that hole is caused by loss, rather than by a reorder.
2125  *
2126  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2127  * holes, caused by lost packets.
2128  *
2129  * And the most logically complicated part of algorithm is undo
2130  * heuristics. We detect false retransmits due to both too early
2131  * fast retransmit (reordering) and underestimated RTO, analyzing
2132  * timestamps and D-SACKs. When we detect that some segments were
2133  * retransmitted by mistake and CWND reduction was wrong, we undo
2134  * window reduction and abort recovery phase. This logic is hidden
2135  * inside several functions named tcp_try_undo_<something>.
2136  */
2137
2138 /* This function decides, when we should leave Disordered state
2139  * and enter Recovery phase, reducing congestion window.
2140  *
2141  * Main question: may we further continue forward transmission
2142  * with the same cwnd?
2143  */
2144 static int tcp_time_to_recover(struct sock *sk)
2145 {
2146         struct tcp_sock *tp = tcp_sk(sk);
2147         __u32 packets_out;
2148
2149         /* Do not perform any recovery during F-RTO algorithm */
2150         if (tp->frto_counter)
2151                 return 0;
2152
2153         /* Trick#1: The loss is proven. */
2154         if (tp->lost_out)
2155                 return 1;
2156
2157         /* Not-A-Trick#2 : Classic rule... */
2158         if (tcp_dupack_heurestics(tp) > tp->reordering)
2159                 return 1;
2160
2161         /* Trick#3 : when we use RFC2988 timer restart, fast
2162          * retransmit can be triggered by timeout of queue head.
2163          */
2164         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2165                 return 1;
2166
2167         /* Trick#4: It is still not OK... But will it be useful to delay
2168          * recovery more?
2169          */
2170         packets_out = tp->packets_out;
2171         if (packets_out <= tp->reordering &&
2172             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2173             !tcp_may_send_now(sk)) {
2174                 /* We have nothing to send. This connection is limited
2175                  * either by receiver window or by application.
2176                  */
2177                 return 1;
2178         }
2179
2180         return 0;
2181 }
2182
2183 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2184  * is against sacked "cnt", otherwise it's against facked "cnt"
2185  */
2186 static void tcp_mark_head_lost(struct sock *sk, int packets)
2187 {
2188         struct tcp_sock *tp = tcp_sk(sk);
2189         struct sk_buff *skb;
2190         int cnt, oldcnt;
2191         int err;
2192         unsigned int mss;
2193
2194         WARN_ON(packets > tp->packets_out);
2195         if (tp->lost_skb_hint) {
2196                 skb = tp->lost_skb_hint;
2197                 cnt = tp->lost_cnt_hint;
2198         } else {
2199                 skb = tcp_write_queue_head(sk);
2200                 cnt = 0;
2201         }
2202
2203         tcp_for_write_queue_from(skb, sk) {
2204                 if (skb == tcp_send_head(sk))
2205                         break;
2206                 /* TODO: do this better */
2207                 /* this is not the most efficient way to do this... */
2208                 tp->lost_skb_hint = skb;
2209                 tp->lost_cnt_hint = cnt;
2210
2211                 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2212                         break;
2213
2214                 oldcnt = cnt;
2215                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2216                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2217                         cnt += tcp_skb_pcount(skb);
2218
2219                 if (cnt > packets) {
2220                         if (tcp_is_sack(tp) || (oldcnt >= packets))
2221                                 break;
2222
2223                         mss = skb_shinfo(skb)->gso_size;
2224                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2225                         if (err < 0)
2226                                 break;
2227                         cnt = packets;
2228                 }
2229
2230                 tcp_skb_mark_lost(tp, skb);
2231         }
2232         tcp_verify_left_out(tp);
2233 }
2234
2235 /* Account newly detected lost packet(s) */
2236
2237 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2238 {
2239         struct tcp_sock *tp = tcp_sk(sk);
2240
2241         if (tcp_is_reno(tp)) {
2242                 tcp_mark_head_lost(sk, 1);
2243         } else if (tcp_is_fack(tp)) {
2244                 int lost = tp->fackets_out - tp->reordering;
2245                 if (lost <= 0)
2246                         lost = 1;
2247                 tcp_mark_head_lost(sk, lost);
2248         } else {
2249                 int sacked_upto = tp->sacked_out - tp->reordering;
2250                 if (sacked_upto < fast_rexmit)
2251                         sacked_upto = fast_rexmit;
2252                 tcp_mark_head_lost(sk, sacked_upto);
2253         }
2254
2255         /* New heuristics: it is possible only after we switched
2256          * to restart timer each time when something is ACKed.
2257          * Hence, we can detect timed out packets during fast
2258          * retransmit without falling to slow start.
2259          */
2260         if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2261                 struct sk_buff *skb;
2262
2263                 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2264                         : tcp_write_queue_head(sk);
2265
2266                 tcp_for_write_queue_from(skb, sk) {
2267                         if (skb == tcp_send_head(sk))
2268                                 break;
2269                         if (!tcp_skb_timedout(sk, skb))
2270                                 break;
2271
2272                         tcp_skb_mark_lost(tp, skb);
2273                 }
2274
2275                 tp->scoreboard_skb_hint = skb;
2276
2277                 tcp_verify_left_out(tp);
2278         }
2279 }
2280
2281 /* CWND moderation, preventing bursts due to too big ACKs
2282  * in dubious situations.
2283  */
2284 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2285 {
2286         tp->snd_cwnd = min(tp->snd_cwnd,
2287                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2288         tp->snd_cwnd_stamp = tcp_time_stamp;
2289 }
2290
2291 /* Lower bound on congestion window is slow start threshold
2292  * unless congestion avoidance choice decides to overide it.
2293  */
2294 static inline u32 tcp_cwnd_min(const struct sock *sk)
2295 {
2296         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2297
2298         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2299 }
2300
2301 /* Decrease cwnd each second ack. */
2302 static void tcp_cwnd_down(struct sock *sk, int flag)
2303 {
2304         struct tcp_sock *tp = tcp_sk(sk);
2305         int decr = tp->snd_cwnd_cnt + 1;
2306
2307         if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2308             (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2309                 tp->snd_cwnd_cnt = decr & 1;
2310                 decr >>= 1;
2311
2312                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2313                         tp->snd_cwnd -= decr;
2314
2315                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2316                 tp->snd_cwnd_stamp = tcp_time_stamp;
2317         }
2318 }
2319
2320 /* Nothing was retransmitted or returned timestamp is less
2321  * than timestamp of the first retransmission.
2322  */
2323 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2324 {
2325         return !tp->retrans_stamp ||
2326                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2327                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2328 }
2329
2330 /* Undo procedures. */
2331
2332 #if FASTRETRANS_DEBUG > 1
2333 static void DBGUNDO(struct sock *sk, const char *msg)
2334 {
2335         struct tcp_sock *tp = tcp_sk(sk);
2336         struct inet_sock *inet = inet_sk(sk);
2337
2338         if (sk->sk_family == AF_INET) {
2339                 printk(KERN_DEBUG "Undo %s " NIPQUAD_FMT "/%u c%u l%u ss%u/%u p%u\n",
2340                        msg,
2341                        NIPQUAD(inet->daddr), ntohs(inet->dport),
2342                        tp->snd_cwnd, tcp_left_out(tp),
2343                        tp->snd_ssthresh, tp->prior_ssthresh,
2344                        tp->packets_out);
2345         }
2346 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2347         else if (sk->sk_family == AF_INET6) {
2348                 struct ipv6_pinfo *np = inet6_sk(sk);
2349                 printk(KERN_DEBUG "Undo %s %p6/%u c%u l%u ss%u/%u p%u\n",
2350                        msg,
2351                        &np->daddr, ntohs(inet->dport),
2352                        tp->snd_cwnd, tcp_left_out(tp),
2353                        tp->snd_ssthresh, tp->prior_ssthresh,
2354                        tp->packets_out);
2355         }
2356 #endif
2357 }
2358 #else
2359 #define DBGUNDO(x...) do { } while (0)
2360 #endif
2361
2362 static void tcp_undo_cwr(struct sock *sk, const int undo)
2363 {
2364         struct tcp_sock *tp = tcp_sk(sk);
2365
2366         if (tp->prior_ssthresh) {
2367                 const struct inet_connection_sock *icsk = inet_csk(sk);
2368
2369                 if (icsk->icsk_ca_ops->undo_cwnd)
2370                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2371                 else
2372                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2373
2374                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2375                         tp->snd_ssthresh = tp->prior_ssthresh;
2376                         TCP_ECN_withdraw_cwr(tp);
2377                 }
2378         } else {
2379                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2380         }
2381         tcp_moderate_cwnd(tp);
2382         tp->snd_cwnd_stamp = tcp_time_stamp;
2383 }
2384
2385 static inline int tcp_may_undo(struct tcp_sock *tp)
2386 {
2387         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2388 }
2389
2390 /* People celebrate: "We love our President!" */
2391 static int tcp_try_undo_recovery(struct sock *sk)
2392 {
2393         struct tcp_sock *tp = tcp_sk(sk);
2394
2395         if (tcp_may_undo(tp)) {
2396                 int mib_idx;
2397
2398                 /* Happy end! We did not retransmit anything
2399                  * or our original transmission succeeded.
2400                  */
2401                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2402                 tcp_undo_cwr(sk, 1);
2403                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2404                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2405                 else
2406                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2407
2408                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2409                 tp->undo_marker = 0;
2410         }
2411         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2412                 /* Hold old state until something *above* high_seq
2413                  * is ACKed. For Reno it is MUST to prevent false
2414                  * fast retransmits (RFC2582). SACK TCP is safe. */
2415                 tcp_moderate_cwnd(tp);
2416                 return 1;
2417         }
2418         tcp_set_ca_state(sk, TCP_CA_Open);
2419         return 0;
2420 }
2421
2422 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2423 static void tcp_try_undo_dsack(struct sock *sk)
2424 {
2425         struct tcp_sock *tp = tcp_sk(sk);
2426
2427         if (tp->undo_marker && !tp->undo_retrans) {
2428                 DBGUNDO(sk, "D-SACK");
2429                 tcp_undo_cwr(sk, 1);
2430                 tp->undo_marker = 0;
2431                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2432         }
2433 }
2434
2435 /* Undo during fast recovery after partial ACK. */
2436
2437 static int tcp_try_undo_partial(struct sock *sk, int acked)
2438 {
2439         struct tcp_sock *tp = tcp_sk(sk);
2440         /* Partial ACK arrived. Force Hoe's retransmit. */
2441         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2442
2443         if (tcp_may_undo(tp)) {
2444                 /* Plain luck! Hole if filled with delayed
2445                  * packet, rather than with a retransmit.
2446                  */
2447                 if (tp->retrans_out == 0)
2448                         tp->retrans_stamp = 0;
2449
2450                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2451
2452                 DBGUNDO(sk, "Hoe");
2453                 tcp_undo_cwr(sk, 0);
2454                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2455
2456                 /* So... Do not make Hoe's retransmit yet.
2457                  * If the first packet was delayed, the rest
2458                  * ones are most probably delayed as well.
2459                  */
2460                 failed = 0;
2461         }
2462         return failed;
2463 }
2464
2465 /* Undo during loss recovery after partial ACK. */
2466 static int tcp_try_undo_loss(struct sock *sk)
2467 {
2468         struct tcp_sock *tp = tcp_sk(sk);
2469
2470         if (tcp_may_undo(tp)) {
2471                 struct sk_buff *skb;
2472                 tcp_for_write_queue(skb, sk) {
2473                         if (skb == tcp_send_head(sk))
2474                                 break;
2475                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2476                 }
2477
2478                 tcp_clear_all_retrans_hints(tp);
2479
2480                 DBGUNDO(sk, "partial loss");
2481                 tp->lost_out = 0;
2482                 tcp_undo_cwr(sk, 1);
2483                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2484                 inet_csk(sk)->icsk_retransmits = 0;
2485                 tp->undo_marker = 0;
2486                 if (tcp_is_sack(tp))
2487                         tcp_set_ca_state(sk, TCP_CA_Open);
2488                 return 1;
2489         }
2490         return 0;
2491 }
2492
2493 static inline void tcp_complete_cwr(struct sock *sk)
2494 {
2495         struct tcp_sock *tp = tcp_sk(sk);
2496         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2497         tp->snd_cwnd_stamp = tcp_time_stamp;
2498         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2499 }
2500
2501 static void tcp_try_keep_open(struct sock *sk)
2502 {
2503         struct tcp_sock *tp = tcp_sk(sk);
2504         int state = TCP_CA_Open;
2505
2506         if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2507                 state = TCP_CA_Disorder;
2508
2509         if (inet_csk(sk)->icsk_ca_state != state) {
2510                 tcp_set_ca_state(sk, state);
2511                 tp->high_seq = tp->snd_nxt;
2512         }
2513 }
2514
2515 static void tcp_try_to_open(struct sock *sk, int flag)
2516 {
2517         struct tcp_sock *tp = tcp_sk(sk);
2518
2519         tcp_verify_left_out(tp);
2520
2521         if (!tp->frto_counter && tp->retrans_out == 0)
2522                 tp->retrans_stamp = 0;
2523
2524         if (flag & FLAG_ECE)
2525                 tcp_enter_cwr(sk, 1);
2526
2527         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2528                 tcp_try_keep_open(sk);
2529                 tcp_moderate_cwnd(tp);
2530         } else {
2531                 tcp_cwnd_down(sk, flag);
2532         }
2533 }
2534
2535 static void tcp_mtup_probe_failed(struct sock *sk)
2536 {
2537         struct inet_connection_sock *icsk = inet_csk(sk);
2538
2539         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2540         icsk->icsk_mtup.probe_size = 0;
2541 }
2542
2543 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2544 {
2545         struct tcp_sock *tp = tcp_sk(sk);
2546         struct inet_connection_sock *icsk = inet_csk(sk);
2547
2548         /* FIXME: breaks with very large cwnd */
2549         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2550         tp->snd_cwnd = tp->snd_cwnd *
2551                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2552                        icsk->icsk_mtup.probe_size;
2553         tp->snd_cwnd_cnt = 0;
2554         tp->snd_cwnd_stamp = tcp_time_stamp;
2555         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2556
2557         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2558         icsk->icsk_mtup.probe_size = 0;
2559         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2560 }
2561
2562 /* Process an event, which can update packets-in-flight not trivially.
2563  * Main goal of this function is to calculate new estimate for left_out,
2564  * taking into account both packets sitting in receiver's buffer and
2565  * packets lost by network.
2566  *
2567  * Besides that it does CWND reduction, when packet loss is detected
2568  * and changes state of machine.
2569  *
2570  * It does _not_ decide what to send, it is made in function
2571  * tcp_xmit_retransmit_queue().
2572  */
2573 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2574 {
2575         struct inet_connection_sock *icsk = inet_csk(sk);
2576         struct tcp_sock *tp = tcp_sk(sk);
2577         int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2578         int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2579                                     (tcp_fackets_out(tp) > tp->reordering));
2580         int fast_rexmit = 0, mib_idx;
2581
2582         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2583                 tp->sacked_out = 0;
2584         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2585                 tp->fackets_out = 0;
2586
2587         /* Now state machine starts.
2588          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2589         if (flag & FLAG_ECE)
2590                 tp->prior_ssthresh = 0;
2591
2592         /* B. In all the states check for reneging SACKs. */
2593         if (tcp_check_sack_reneging(sk, flag))
2594                 return;
2595
2596         /* C. Process data loss notification, provided it is valid. */
2597         if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2598             before(tp->snd_una, tp->high_seq) &&
2599             icsk->icsk_ca_state != TCP_CA_Open &&
2600             tp->fackets_out > tp->reordering) {
2601                 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2602                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2603         }
2604
2605         /* D. Check consistency of the current state. */
2606         tcp_verify_left_out(tp);
2607
2608         /* E. Check state exit conditions. State can be terminated
2609          *    when high_seq is ACKed. */
2610         if (icsk->icsk_ca_state == TCP_CA_Open) {
2611                 WARN_ON(tp->retrans_out != 0);
2612                 tp->retrans_stamp = 0;
2613         } else if (!before(tp->snd_una, tp->high_seq)) {
2614                 switch (icsk->icsk_ca_state) {
2615                 case TCP_CA_Loss:
2616                         icsk->icsk_retransmits = 0;
2617                         if (tcp_try_undo_recovery(sk))
2618                                 return;
2619                         break;
2620
2621                 case TCP_CA_CWR:
2622                         /* CWR is to be held something *above* high_seq
2623                          * is ACKed for CWR bit to reach receiver. */
2624                         if (tp->snd_una != tp->high_seq) {
2625                                 tcp_complete_cwr(sk);
2626                                 tcp_set_ca_state(sk, TCP_CA_Open);
2627                         }
2628                         break;
2629
2630                 case TCP_CA_Disorder:
2631                         tcp_try_undo_dsack(sk);
2632                         if (!tp->undo_marker ||
2633                             /* For SACK case do not Open to allow to undo
2634                              * catching for all duplicate ACKs. */
2635                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2636                                 tp->undo_marker = 0;
2637                                 tcp_set_ca_state(sk, TCP_CA_Open);
2638                         }
2639                         break;
2640
2641                 case TCP_CA_Recovery:
2642                         if (tcp_is_reno(tp))
2643                                 tcp_reset_reno_sack(tp);
2644                         if (tcp_try_undo_recovery(sk))
2645                                 return;
2646                         tcp_complete_cwr(sk);
2647                         break;
2648                 }
2649         }
2650
2651         /* F. Process state. */
2652         switch (icsk->icsk_ca_state) {
2653         case TCP_CA_Recovery:
2654                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2655                         if (tcp_is_reno(tp) && is_dupack)
2656                                 tcp_add_reno_sack(sk);
2657                 } else
2658                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2659                 break;
2660         case TCP_CA_Loss:
2661                 if (flag & FLAG_DATA_ACKED)
2662                         icsk->icsk_retransmits = 0;
2663                 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2664                         tcp_reset_reno_sack(tp);
2665                 if (!tcp_try_undo_loss(sk)) {
2666                         tcp_moderate_cwnd(tp);
2667                         tcp_xmit_retransmit_queue(sk);
2668                         return;
2669                 }
2670                 if (icsk->icsk_ca_state != TCP_CA_Open)
2671                         return;
2672                 /* Loss is undone; fall through to processing in Open state. */
2673         default:
2674                 if (tcp_is_reno(tp)) {
2675                         if (flag & FLAG_SND_UNA_ADVANCED)
2676                                 tcp_reset_reno_sack(tp);
2677                         if (is_dupack)
2678                                 tcp_add_reno_sack(sk);
2679                 }
2680
2681                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2682                         tcp_try_undo_dsack(sk);
2683
2684                 if (!tcp_time_to_recover(sk)) {
2685                         tcp_try_to_open(sk, flag);
2686                         return;
2687                 }
2688
2689                 /* MTU probe failure: don't reduce cwnd */
2690                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2691                     icsk->icsk_mtup.probe_size &&
2692                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
2693                         tcp_mtup_probe_failed(sk);
2694                         /* Restores the reduction we did in tcp_mtup_probe() */
2695                         tp->snd_cwnd++;
2696                         tcp_simple_retransmit(sk);
2697                         return;
2698                 }
2699
2700                 /* Otherwise enter Recovery state */
2701
2702                 if (tcp_is_reno(tp))
2703                         mib_idx = LINUX_MIB_TCPRENORECOVERY;
2704                 else
2705                         mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2706
2707                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2708
2709                 tp->high_seq = tp->snd_nxt;
2710                 tp->prior_ssthresh = 0;
2711                 tp->undo_marker = tp->snd_una;
2712                 tp->undo_retrans = tp->retrans_out;
2713
2714                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2715                         if (!(flag & FLAG_ECE))
2716                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2717                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2718                         TCP_ECN_queue_cwr(tp);
2719                 }
2720
2721                 tp->bytes_acked = 0;
2722                 tp->snd_cwnd_cnt = 0;
2723                 tcp_set_ca_state(sk, TCP_CA_Recovery);
2724                 fast_rexmit = 1;
2725         }
2726
2727         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2728                 tcp_update_scoreboard(sk, fast_rexmit);
2729         tcp_cwnd_down(sk, flag);
2730         tcp_xmit_retransmit_queue(sk);
2731 }
2732
2733 /* Read draft-ietf-tcplw-high-performance before mucking
2734  * with this code. (Supersedes RFC1323)
2735  */
2736 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2737 {
2738         /* RTTM Rule: A TSecr value received in a segment is used to
2739          * update the averaged RTT measurement only if the segment
2740          * acknowledges some new data, i.e., only if it advances the
2741          * left edge of the send window.
2742          *
2743          * See draft-ietf-tcplw-high-performance-00, section 3.3.
2744          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2745          *
2746          * Changed: reset backoff as soon as we see the first valid sample.
2747          * If we do not, we get strongly overestimated rto. With timestamps
2748          * samples are accepted even from very old segments: f.e., when rtt=1
2749          * increases to 8, we retransmit 5 times and after 8 seconds delayed
2750          * answer arrives rto becomes 120 seconds! If at least one of segments
2751          * in window is lost... Voila.                          --ANK (010210)
2752          */
2753         struct tcp_sock *tp = tcp_sk(sk);
2754         const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2755         tcp_rtt_estimator(sk, seq_rtt);
2756         tcp_set_rto(sk);
2757         inet_csk(sk)->icsk_backoff = 0;
2758         tcp_bound_rto(sk);
2759 }
2760
2761 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2762 {
2763         /* We don't have a timestamp. Can only use
2764          * packets that are not retransmitted to determine
2765          * rtt estimates. Also, we must not reset the
2766          * backoff for rto until we get a non-retransmitted
2767          * packet. This allows us to deal with a situation
2768          * where the network delay has increased suddenly.
2769          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2770          */
2771
2772         if (flag & FLAG_RETRANS_DATA_ACKED)
2773                 return;
2774
2775         tcp_rtt_estimator(sk, seq_rtt);
2776         tcp_set_rto(sk);
2777         inet_csk(sk)->icsk_backoff = 0;
2778         tcp_bound_rto(sk);
2779 }
2780
2781 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2782                                       const s32 seq_rtt)
2783 {
2784         const struct tcp_sock *tp = tcp_sk(sk);
2785         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2786         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2787                 tcp_ack_saw_tstamp(sk, flag);
2788         else if (seq_rtt >= 0)
2789                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2790 }
2791
2792 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2793 {
2794         const struct inet_connection_sock *icsk = inet_csk(sk);
2795         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2796         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2797 }
2798
2799 /* Restart timer after forward progress on connection.
2800  * RFC2988 recommends to restart timer to now+rto.
2801  */
2802 static void tcp_rearm_rto(struct sock *sk)
2803 {
2804         struct tcp_sock *tp = tcp_sk(sk);
2805
2806         if (!tp->packets_out) {
2807                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2808         } else {
2809                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2810                                           inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2811         }
2812 }
2813
2814 /* If we get here, the whole TSO packet has not been acked. */
2815 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2816 {
2817         struct tcp_sock *tp = tcp_sk(sk);
2818         u32 packets_acked;
2819
2820         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2821
2822         packets_acked = tcp_skb_pcount(skb);
2823         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2824                 return 0;
2825         packets_acked -= tcp_skb_pcount(skb);
2826
2827         if (packets_acked) {
2828                 BUG_ON(tcp_skb_pcount(skb) == 0);
2829                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2830         }
2831
2832         return packets_acked;
2833 }
2834
2835 /* Remove acknowledged frames from the retransmission queue. If our packet
2836  * is before the ack sequence we can discard it as it's confirmed to have
2837  * arrived at the other end.
2838  */
2839 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
2840                                u32 prior_snd_una)
2841 {
2842         struct tcp_sock *tp = tcp_sk(sk);
2843         const struct inet_connection_sock *icsk = inet_csk(sk);
2844         struct sk_buff *skb;
2845         u32 now = tcp_time_stamp;
2846         int fully_acked = 1;
2847         int flag = 0;
2848         u32 pkts_acked = 0;
2849         u32 reord = tp->packets_out;
2850         u32 prior_sacked = tp->sacked_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                 tp->packets_out -= acked_pcount;
2908                 pkts_acked += acked_pcount;
2909
2910                 /* Initial outgoing SYN's get put onto the write_queue
2911                  * just like anything else we transmit.  It is not
2912                  * true data, and if we misinform our callers that
2913                  * this ACK acks real data, we will erroneously exit
2914                  * connection startup slow start one packet too
2915                  * quickly.  This is severely frowned upon behavior.
2916                  */
2917                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2918                         flag |= FLAG_DATA_ACKED;
2919                 } else {
2920                         flag |= FLAG_SYN_ACKED;
2921                         tp->retrans_stamp = 0;
2922                 }
2923
2924                 if (!fully_acked)
2925                         break;
2926
2927                 tcp_unlink_write_queue(skb, sk);
2928                 sk_wmem_free_skb(sk, skb);
2929                 tp->scoreboard_skb_hint = NULL;
2930                 if (skb == tp->retransmit_skb_hint)
2931                         tp->retransmit_skb_hint = NULL;
2932                 if (skb == tp->lost_skb_hint)
2933                         tp->lost_skb_hint = NULL;
2934         }
2935
2936         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
2937                 tp->snd_up = tp->snd_una;
2938
2939         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2940                 flag |= FLAG_SACK_RENEGING;
2941
2942         if (flag & FLAG_ACKED) {
2943                 const struct tcp_congestion_ops *ca_ops
2944                         = inet_csk(sk)->icsk_ca_ops;
2945
2946                 tcp_ack_update_rtt(sk, flag, seq_rtt);
2947                 tcp_rearm_rto(sk);
2948
2949                 if (tcp_is_reno(tp)) {
2950                         tcp_remove_reno_sacks(sk, pkts_acked);
2951                 } else {
2952                         /* Non-retransmitted hole got filled? That's reordering */
2953                         if (reord < prior_fackets)
2954                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2955
2956                         /* No need to care for underflows here because
2957                          * the lost_skb_hint gets NULLed if we're past it
2958                          * (or something non-trivial happened)
2959                          */
2960                         if (tcp_is_fack(tp))
2961                                 tp->lost_cnt_hint -= pkts_acked;
2962                         else
2963                                 tp->lost_cnt_hint -= prior_sacked - tp->sacked_out;
2964                 }
2965
2966                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2967
2968                 if (ca_ops->pkts_acked) {
2969                         s32 rtt_us = -1;
2970
2971                         /* Is the ACK triggering packet unambiguous? */
2972                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2973                                 /* High resolution needed and available? */
2974                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2975                                     !ktime_equal(last_ackt,
2976                                                  net_invalid_timestamp()))
2977                                         rtt_us = ktime_us_delta(ktime_get_real(),
2978                                                                 last_ackt);
2979                                 else if (ca_seq_rtt > 0)
2980                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
2981                         }
2982
2983                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2984                 }
2985         }
2986
2987 #if FASTRETRANS_DEBUG > 0
2988         WARN_ON((int)tp->sacked_out < 0);
2989         WARN_ON((int)tp->lost_out < 0);
2990         WARN_ON((int)tp->retrans_out < 0);
2991         if (!tp->packets_out && tcp_is_sack(tp)) {
2992                 icsk = inet_csk(sk);
2993                 if (tp->lost_out) {
2994                         printk(KERN_DEBUG "Leak l=%u %d\n",
2995                                tp->lost_out, icsk->icsk_ca_state);
2996                         tp->lost_out = 0;
2997                 }
2998                 if (tp->sacked_out) {
2999                         printk(KERN_DEBUG "Leak s=%u %d\n",
3000                                tp->sacked_out, icsk->icsk_ca_state);
3001                         tp->sacked_out = 0;
3002                 }
3003                 if (tp->retrans_out) {
3004                         printk(KERN_DEBUG "Leak r=%u %d\n",
3005                                tp->retrans_out, icsk->icsk_ca_state);
3006                         tp->retrans_out = 0;
3007                 }
3008         }
3009 #endif
3010         return flag;
3011 }
3012
3013 static void tcp_ack_probe(struct sock *sk)
3014 {
3015         const struct tcp_sock *tp = tcp_sk(sk);
3016         struct inet_connection_sock *icsk = inet_csk(sk);
3017
3018         /* Was it a usable window open? */
3019
3020         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3021                 icsk->icsk_backoff = 0;
3022                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3023                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3024                  * This function is not for random using!
3025                  */
3026         } else {
3027                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3028                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3029                                           TCP_RTO_MAX);
3030         }
3031 }
3032
3033 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3034 {
3035         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3036                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3037 }
3038
3039 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3040 {
3041         const struct tcp_sock *tp = tcp_sk(sk);
3042         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3043                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3044 }
3045
3046 /* Check that window update is acceptable.
3047  * The function assumes that snd_una<=ack<=snd_next.
3048  */
3049 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3050                                         const u32 ack, const u32 ack_seq,
3051                                         const u32 nwin)
3052 {
3053         return (after(ack, tp->snd_una) ||
3054                 after(ack_seq, tp->snd_wl1) ||
3055                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3056 }
3057
3058 /* Update our send window.
3059  *
3060  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3061  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3062  */
3063 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3064                                  u32 ack_seq)
3065 {
3066         struct tcp_sock *tp = tcp_sk(sk);
3067         int flag = 0;
3068         u32 nwin = ntohs(tcp_hdr(skb)->window);
3069
3070         if (likely(!tcp_hdr(skb)->syn))
3071                 nwin <<= tp->rx_opt.snd_wscale;
3072
3073         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3074                 flag |= FLAG_WIN_UPDATE;
3075                 tcp_update_wl(tp, ack, ack_seq);
3076
3077                 if (tp->snd_wnd != nwin) {
3078                         tp->snd_wnd = nwin;
3079
3080                         /* Note, it is the only place, where
3081                          * fast path is recovered for sending TCP.
3082                          */
3083                         tp->pred_flags = 0;
3084                         tcp_fast_path_check(sk);
3085
3086                         if (nwin > tp->max_window) {
3087                                 tp->max_window = nwin;
3088                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3089                         }
3090                 }
3091         }
3092
3093         tp->snd_una = ack;
3094
3095         return flag;
3096 }
3097
3098 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3099  * continue in congestion avoidance.
3100  */
3101 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3102 {
3103         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3104         tp->snd_cwnd_cnt = 0;
3105         tp->bytes_acked = 0;
3106         TCP_ECN_queue_cwr(tp);
3107         tcp_moderate_cwnd(tp);
3108 }
3109
3110 /* A conservative spurious RTO response algorithm: reduce cwnd using
3111  * rate halving and continue in congestion avoidance.
3112  */
3113 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3114 {
3115         tcp_enter_cwr(sk, 0);
3116 }
3117
3118 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3119 {
3120         if (flag & FLAG_ECE)
3121                 tcp_ratehalving_spur_to_response(sk);
3122         else
3123                 tcp_undo_cwr(sk, 1);
3124 }
3125
3126 /* F-RTO spurious RTO detection algorithm (RFC4138)
3127  *
3128  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3129  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3130  * window (but not to or beyond highest sequence sent before RTO):
3131  *   On First ACK,  send two new segments out.
3132  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3133  *                  algorithm is not part of the F-RTO detection algorithm
3134  *                  given in RFC4138 but can be selected separately).
3135  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3136  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3137  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3138  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3139  *
3140  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3141  * original window even after we transmit two new data segments.
3142  *
3143  * SACK version:
3144  *   on first step, wait until first cumulative ACK arrives, then move to
3145  *   the second step. In second step, the next ACK decides.
3146  *
3147  * F-RTO is implemented (mainly) in four functions:
3148  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3149  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3150  *     called when tcp_use_frto() showed green light
3151  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3152  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3153  *     to prove that the RTO is indeed spurious. It transfers the control
3154  *     from F-RTO to the conventional RTO recovery
3155  */
3156 static int tcp_process_frto(struct sock *sk, int flag)
3157 {
3158         struct tcp_sock *tp = tcp_sk(sk);
3159
3160         tcp_verify_left_out(tp);
3161
3162         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3163         if (flag & FLAG_DATA_ACKED)
3164                 inet_csk(sk)->icsk_retransmits = 0;
3165
3166         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3167             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3168                 tp->undo_marker = 0;
3169
3170         if (!before(tp->snd_una, tp->frto_highmark)) {
3171                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3172                 return 1;
3173         }
3174
3175         if (!tcp_is_sackfrto(tp)) {
3176                 /* RFC4138 shortcoming in step 2; should also have case c):
3177                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3178                  * data, winupdate
3179                  */
3180                 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3181                         return 1;
3182
3183                 if (!(flag & FLAG_DATA_ACKED)) {
3184                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3185                                             flag);
3186                         return 1;
3187                 }
3188         } else {
3189                 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3190                         /* Prevent sending of new data. */
3191                         tp->snd_cwnd = min(tp->snd_cwnd,
3192                                            tcp_packets_in_flight(tp));
3193                         return 1;
3194                 }
3195
3196                 if ((tp->frto_counter >= 2) &&
3197                     (!(flag & FLAG_FORWARD_PROGRESS) ||
3198                      ((flag & FLAG_DATA_SACKED) &&
3199                       !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3200                         /* RFC4138 shortcoming (see comment above) */
3201                         if (!(flag & FLAG_FORWARD_PROGRESS) &&
3202                             (flag & FLAG_NOT_DUP))
3203                                 return 1;
3204
3205                         tcp_enter_frto_loss(sk, 3, flag);
3206                         return 1;
3207                 }
3208         }
3209
3210         if (tp->frto_counter == 1) {
3211                 /* tcp_may_send_now needs to see updated state */
3212                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3213                 tp->frto_counter = 2;
3214
3215                 if (!tcp_may_send_now(sk))
3216                         tcp_enter_frto_loss(sk, 2, flag);
3217
3218                 return 1;
3219         } else {
3220                 switch (sysctl_tcp_frto_response) {
3221                 case 2:
3222                         tcp_undo_spur_to_response(sk, flag);
3223                         break;
3224                 case 1:
3225                         tcp_conservative_spur_to_response(tp);
3226                         break;
3227                 default:
3228                         tcp_ratehalving_spur_to_response(sk);
3229                         break;
3230                 }
3231                 tp->frto_counter = 0;
3232                 tp->undo_marker = 0;
3233                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3234         }
3235         return 0;
3236 }
3237
3238 /* This routine deals with incoming acks, but not outgoing ones. */
3239 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3240 {
3241         struct inet_connection_sock *icsk = inet_csk(sk);
3242         struct tcp_sock *tp = tcp_sk(sk);
3243         u32 prior_snd_una = tp->snd_una;
3244         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3245         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3246         u32 prior_in_flight;
3247         u32 prior_fackets;
3248         int prior_packets;
3249         int frto_cwnd = 0;
3250
3251         /* If the ack is newer than sent or older than previous acks
3252          * then we can probably ignore it.
3253          */
3254         if (after(ack, tp->snd_nxt))
3255                 goto uninteresting_ack;
3256
3257         if (before(ack, prior_snd_una))
3258                 goto old_ack;
3259
3260         if (after(ack, prior_snd_una))
3261                 flag |= FLAG_SND_UNA_ADVANCED;
3262
3263         if (sysctl_tcp_abc) {
3264                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3265                         tp->bytes_acked += ack - prior_snd_una;
3266                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3267                         /* we assume just one segment left network */
3268                         tp->bytes_acked += min(ack - prior_snd_una,
3269                                                tp->mss_cache);
3270         }
3271
3272         prior_fackets = tp->fackets_out;
3273         prior_in_flight = tcp_packets_in_flight(tp);
3274
3275         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3276                 /* Window is constant, pure forward advance.
3277                  * No more checks are required.
3278                  * Note, we use the fact that SND.UNA>=SND.WL2.
3279                  */
3280                 tcp_update_wl(tp, ack, ack_seq);
3281                 tp->snd_una = ack;
3282                 flag |= FLAG_WIN_UPDATE;
3283
3284                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3285
3286                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3287         } else {
3288                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3289                         flag |= FLAG_DATA;
3290                 else
3291                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3292
3293                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3294
3295                 if (TCP_SKB_CB(skb)->sacked)
3296                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3297
3298                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3299                         flag |= FLAG_ECE;
3300
3301                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3302         }
3303
3304         /* We passed data and got it acked, remove any soft error
3305          * log. Something worked...
3306          */
3307         sk->sk_err_soft = 0;
3308         icsk->icsk_probes_out = 0;
3309         tp->rcv_tstamp = tcp_time_stamp;
3310         prior_packets = tp->packets_out;
3311         if (!prior_packets)
3312                 goto no_queue;
3313
3314         /* See if we can take anything off of the retransmit queue. */
3315         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3316
3317         if (tp->frto_counter)
3318                 frto_cwnd = tcp_process_frto(sk, flag);
3319         /* Guarantee sacktag reordering detection against wrap-arounds */
3320         if (before(tp->frto_highmark, tp->snd_una))
3321                 tp->frto_highmark = 0;
3322
3323         if (tcp_ack_is_dubious(sk, flag)) {
3324                 /* Advance CWND, if state allows this. */
3325                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3326                     tcp_may_raise_cwnd(sk, flag))
3327                         tcp_cong_avoid(sk, ack, prior_in_flight);
3328                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3329                                       flag);
3330         } else {
3331                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3332                         tcp_cong_avoid(sk, ack, prior_in_flight);
3333         }
3334
3335         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3336                 dst_confirm(sk->sk_dst_cache);
3337
3338         return 1;
3339
3340 no_queue:
3341         /* If this ack opens up a zero window, clear backoff.  It was
3342          * being used to time the probes, and is probably far higher than
3343          * it needs to be for normal retransmission.
3344          */
3345         if (tcp_send_head(sk))
3346                 tcp_ack_probe(sk);
3347         return 1;
3348
3349 old_ack:
3350         if (TCP_SKB_CB(skb)->sacked) {
3351                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3352                 if (icsk->icsk_ca_state == TCP_CA_Open)
3353                         tcp_try_keep_open(sk);
3354         }
3355
3356 uninteresting_ack:
3357         SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3358         return 0;
3359 }
3360
3361 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3362  * But, this can also be called on packets in the established flow when
3363  * the fast version below fails.
3364  */
3365 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3366                        int estab)
3367 {
3368         unsigned char *ptr;
3369         struct tcphdr *th = tcp_hdr(skb);
3370         int length = (th->doff * 4) - sizeof(struct tcphdr);
3371
3372         ptr = (unsigned char *)(th + 1);
3373         opt_rx->saw_tstamp = 0;
3374
3375         while (length > 0) {
3376                 int opcode = *ptr++;
3377                 int opsize;
3378
3379                 switch (opcode) {
3380                 case TCPOPT_EOL:
3381                         return;
3382                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3383                         length--;
3384                         continue;
3385                 default:
3386                         opsize = *ptr++;
3387                         if (opsize < 2) /* "silly options" */
3388                                 return;
3389                         if (opsize > length)
3390                                 return; /* don't parse partial options */
3391                         switch (opcode) {
3392                         case TCPOPT_MSS:
3393                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3394                                         u16 in_mss = get_unaligned_be16(ptr);
3395                                         if (in_mss) {
3396                                                 if (opt_rx->user_mss &&
3397                                                     opt_rx->user_mss < in_mss)
3398                                                         in_mss = opt_rx->user_mss;
3399                                                 opt_rx->mss_clamp = in_mss;
3400                                         }
3401                                 }
3402                                 break;
3403                         case TCPOPT_WINDOW:
3404                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3405                                     !estab && sysctl_tcp_window_scaling) {
3406                                         __u8 snd_wscale = *(__u8 *)ptr;
3407                                         opt_rx->wscale_ok = 1;
3408                                         if (snd_wscale > 14) {
3409                                                 if (net_ratelimit())
3410                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3411                                                                "scaling value %d >14 received.\n",
3412                                                                snd_wscale);
3413                                                 snd_wscale = 14;
3414                                         }
3415                                         opt_rx->snd_wscale = snd_wscale;
3416                                 }
3417                                 break;
3418                         case TCPOPT_TIMESTAMP:
3419                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3420                                     ((estab && opt_rx->tstamp_ok) ||
3421                                      (!estab && sysctl_tcp_timestamps))) {
3422                                         opt_rx->saw_tstamp = 1;
3423                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3424                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3425                                 }
3426                                 break;
3427                         case TCPOPT_SACK_PERM:
3428                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3429                                     !estab && sysctl_tcp_sack) {
3430                                         opt_rx->sack_ok = 1;
3431                                         tcp_sack_reset(opt_rx);
3432                                 }
3433                                 break;
3434
3435                         case TCPOPT_SACK:
3436                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3437                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3438                                    opt_rx->sack_ok) {
3439                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3440                                 }
3441                                 break;
3442 #ifdef CONFIG_TCP_MD5SIG
3443                         case TCPOPT_MD5SIG:
3444                                 /*
3445                                  * The MD5 Hash has already been
3446                                  * checked (see tcp_v{4,6}_do_rcv()).
3447                                  */
3448                                 break;
3449 #endif
3450                         }
3451
3452                         ptr += opsize-2;
3453                         length -= opsize;
3454                 }
3455         }
3456 }
3457
3458 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3459 {
3460         __be32 *ptr = (__be32 *)(th + 1);
3461
3462         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3463                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3464                 tp->rx_opt.saw_tstamp = 1;
3465                 ++ptr;
3466                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3467                 ++ptr;
3468                 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3469                 return 1;
3470         }
3471         return 0;
3472 }
3473
3474 /* Fast parse options. This hopes to only see timestamps.
3475  * If it is wrong it falls back on tcp_parse_options().
3476  */
3477 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3478                                   struct tcp_sock *tp)
3479 {
3480         if (th->doff == sizeof(struct tcphdr) >> 2) {
3481                 tp->rx_opt.saw_tstamp = 0;
3482                 return 0;
3483         } else if (tp->rx_opt.tstamp_ok &&
3484                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3485                 if (tcp_parse_aligned_timestamp(tp, th))
3486                         return 1;
3487         }
3488         tcp_parse_options(skb, &tp->rx_opt, 1);
3489         return 1;
3490 }
3491
3492 #ifdef CONFIG_TCP_MD5SIG
3493 /*
3494  * Parse MD5 Signature option
3495  */
3496 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3497 {
3498         int length = (th->doff << 2) - sizeof (*th);
3499         u8 *ptr = (u8*)(th + 1);
3500
3501         /* If the TCP option is too short, we can short cut */
3502         if (length < TCPOLEN_MD5SIG)
3503                 return NULL;
3504
3505         while (length > 0) {
3506                 int opcode = *ptr++;
3507                 int opsize;
3508
3509                 switch(opcode) {
3510                 case TCPOPT_EOL:
3511                         return NULL;
3512                 case TCPOPT_NOP:
3513                         length--;
3514                         continue;
3515                 default:
3516                         opsize = *ptr++;
3517                         if (opsize < 2 || opsize > length)
3518                                 return NULL;
3519                         if (opcode == TCPOPT_MD5SIG)
3520                                 return ptr;
3521                 }
3522                 ptr += opsize - 2;
3523                 length -= opsize;
3524         }
3525         return NULL;
3526 }
3527 #endif
3528
3529 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3530 {
3531         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3532         tp->rx_opt.ts_recent_stamp = get_seconds();
3533 }
3534
3535 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3536 {
3537         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3538                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3539                  * extra check below makes sure this can only happen
3540                  * for pure ACK frames.  -DaveM
3541                  *
3542                  * Not only, also it occurs for expired timestamps.
3543                  */
3544
3545                 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3546                    get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3547                         tcp_store_ts_recent(tp);
3548         }
3549 }
3550
3551 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3552  *
3553  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3554  * it can pass through stack. So, the following predicate verifies that
3555  * this segment is not used for anything but congestion avoidance or
3556  * fast retransmit. Moreover, we even are able to eliminate most of such
3557  * second order effects, if we apply some small "replay" window (~RTO)
3558  * to timestamp space.
3559  *
3560  * All these measures still do not guarantee that we reject wrapped ACKs
3561  * on networks with high bandwidth, when sequence space is recycled fastly,
3562  * but it guarantees that such events will be very rare and do not affect
3563  * connection seriously. This doesn't look nice, but alas, PAWS is really
3564  * buggy extension.
3565  *
3566  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3567  * states that events when retransmit arrives after original data are rare.
3568  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3569  * the biggest problem on large power networks even with minor reordering.
3570  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3571  * up to bandwidth of 18Gigabit/sec. 8) ]
3572  */
3573
3574 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3575 {
3576         struct tcp_sock *tp = tcp_sk(sk);
3577         struct tcphdr *th = tcp_hdr(skb);
3578         u32 seq = TCP_SKB_CB(skb)->seq;
3579         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3580
3581         return (/* 1. Pure ACK with correct sequence number. */
3582                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3583
3584                 /* 2. ... and duplicate ACK. */
3585                 ack == tp->snd_una &&
3586
3587                 /* 3. ... and does not update window. */
3588                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3589
3590                 /* 4. ... and sits in replay window. */
3591                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3592 }
3593
3594 static inline int tcp_paws_discard(const struct sock *sk,
3595                                    const struct sk_buff *skb)
3596 {
3597         const struct tcp_sock *tp = tcp_sk(sk);
3598         return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3599                 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3600                 !tcp_disordered_ack(sk, skb));
3601 }
3602
3603 /* Check segment sequence number for validity.
3604  *
3605  * Segment controls are considered valid, if the segment
3606  * fits to the window after truncation to the window. Acceptability
3607  * of data (and SYN, FIN, of course) is checked separately.
3608  * See tcp_data_queue(), for example.
3609  *
3610  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3611  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3612  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3613  * (borrowed from freebsd)
3614  */
3615
3616 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3617 {
3618         return  !before(end_seq, tp->rcv_wup) &&
3619                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3620 }
3621
3622 /* When we get a reset we do this. */
3623 static void tcp_reset(struct sock *sk)
3624 {
3625         /* We want the right error as BSD sees it (and indeed as we do). */
3626         switch (sk->sk_state) {
3627         case TCP_SYN_SENT:
3628                 sk->sk_err = ECONNREFUSED;
3629                 break;
3630         case TCP_CLOSE_WAIT:
3631                 sk->sk_err = EPIPE;
3632                 break;
3633         case TCP_CLOSE:
3634                 return;
3635         default:
3636                 sk->sk_err = ECONNRESET;
3637         }
3638
3639         if (!sock_flag(sk, SOCK_DEAD))
3640                 sk->sk_error_report(sk);
3641
3642         tcp_done(sk);
3643 }
3644
3645 /*
3646  *      Process the FIN bit. This now behaves as it is supposed to work
3647  *      and the FIN takes effect when it is validly part of sequence
3648  *      space. Not before when we get holes.
3649  *
3650  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3651  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3652  *      TIME-WAIT)
3653  *
3654  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3655  *      close and we go into CLOSING (and later onto TIME-WAIT)
3656  *
3657  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3658  */
3659 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3660 {
3661         struct tcp_sock *tp = tcp_sk(sk);
3662
3663         inet_csk_schedule_ack(sk);
3664
3665         sk->sk_shutdown |= RCV_SHUTDOWN;
3666         sock_set_flag(sk, SOCK_DONE);
3667
3668         switch (sk->sk_state) {
3669         case TCP_SYN_RECV:
3670         case TCP_ESTABLISHED:
3671                 /* Move to CLOSE_WAIT */
3672                 tcp_set_state(sk, TCP_CLOSE_WAIT);
3673                 inet_csk(sk)->icsk_ack.pingpong = 1;
3674                 break;
3675
3676         case TCP_CLOSE_WAIT:
3677         case TCP_CLOSING:
3678                 /* Received a retransmission of the FIN, do
3679                  * nothing.
3680                  */
3681                 break;
3682         case TCP_LAST_ACK:
3683                 /* RFC793: Remain in the LAST-ACK state. */
3684                 break;
3685
3686         case TCP_FIN_WAIT1:
3687                 /* This case occurs when a simultaneous close
3688                  * happens, we must ack the received FIN and
3689                  * enter the CLOSING state.
3690                  */
3691                 tcp_send_ack(sk);
3692                 tcp_set_state(sk, TCP_CLOSING);
3693                 break;
3694         case TCP_FIN_WAIT2:
3695                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3696                 tcp_send_ack(sk);
3697                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3698                 break;
3699         default:
3700                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3701                  * cases we should never reach this piece of code.
3702                  */
3703                 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3704                        __func__, sk->sk_state);
3705                 break;
3706         }
3707
3708         /* It _is_ possible, that we have something out-of-order _after_ FIN.
3709          * Probably, we should reset in this case. For now drop them.
3710          */
3711         __skb_queue_purge(&tp->out_of_order_queue);
3712         if (tcp_is_sack(tp))
3713                 tcp_sack_reset(&tp->rx_opt);
3714         sk_mem_reclaim(sk);
3715
3716         if (!sock_flag(sk, SOCK_DEAD)) {
3717                 sk->sk_state_change(sk);
3718
3719                 /* Do not send POLL_HUP for half duplex close. */
3720                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3721                     sk->sk_state == TCP_CLOSE)
3722                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3723                 else
3724                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3725         }
3726 }
3727
3728 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3729                                   u32 end_seq)
3730 {
3731         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3732                 if (before(seq, sp->start_seq))
3733                         sp->start_seq = seq;
3734                 if (after(end_seq, sp->end_seq))
3735                         sp->end_seq = end_seq;
3736                 return 1;
3737         }
3738         return 0;
3739 }
3740
3741 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3742 {
3743         struct tcp_sock *tp = tcp_sk(sk);
3744
3745         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3746                 int mib_idx;
3747
3748                 if (before(seq, tp->rcv_nxt))
3749                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3750                 else
3751                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3752
3753                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3754
3755                 tp->rx_opt.dsack = 1;
3756                 tp->duplicate_sack[0].start_seq = seq;
3757                 tp->duplicate_sack[0].end_seq = end_seq;
3758                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 1;
3759         }
3760 }
3761
3762 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3763 {
3764         struct tcp_sock *tp = tcp_sk(sk);
3765
3766         if (!tp->rx_opt.dsack)
3767                 tcp_dsack_set(sk, seq, end_seq);
3768         else
3769                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3770 }
3771
3772 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3773 {
3774         struct tcp_sock *tp = tcp_sk(sk);
3775
3776         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3777             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3778                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3779                 tcp_enter_quickack_mode(sk);
3780
3781                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3782                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3783
3784                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3785                                 end_seq = tp->rcv_nxt;
3786                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3787                 }
3788         }
3789
3790         tcp_send_ack(sk);
3791 }
3792
3793 /* These routines update the SACK block as out-of-order packets arrive or
3794  * in-order packets close up the sequence space.
3795  */
3796 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3797 {
3798         int this_sack;
3799         struct tcp_sack_block *sp = &tp->selective_acks[0];
3800         struct tcp_sack_block *swalk = sp + 1;
3801
3802         /* See if the recent change to the first SACK eats into
3803          * or hits the sequence space of other SACK blocks, if so coalesce.
3804          */
3805         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3806                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3807                         int i;
3808
3809                         /* Zap SWALK, by moving every further SACK up by one slot.
3810                          * Decrease num_sacks.
3811                          */
3812                         tp->rx_opt.num_sacks--;
3813                         tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
3814                                                tp->rx_opt.dsack;
3815                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3816                                 sp[i] = sp[i + 1];
3817                         continue;
3818                 }
3819                 this_sack++, swalk++;
3820         }
3821 }
3822
3823 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
3824                                  struct tcp_sack_block *sack2)
3825 {
3826         __u32 tmp;
3827
3828         tmp = sack1->start_seq;
3829         sack1->start_seq = sack2->start_seq;
3830         sack2->start_seq = tmp;
3831
3832         tmp = sack1->end_seq;
3833         sack1->end_seq = sack2->end_seq;
3834         sack2->end_seq = tmp;
3835 }
3836
3837 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3838 {
3839         struct tcp_sock *tp = tcp_sk(sk);
3840         struct tcp_sack_block *sp = &tp->selective_acks[0];
3841         int cur_sacks = tp->rx_opt.num_sacks;
3842         int this_sack;
3843
3844         if (!cur_sacks)
3845                 goto new_sack;
3846
3847         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3848                 if (tcp_sack_extend(sp, seq, end_seq)) {
3849                         /* Rotate this_sack to the first one. */
3850                         for (; this_sack > 0; this_sack--, sp--)
3851                                 tcp_sack_swap(sp, sp - 1);
3852                         if (cur_sacks > 1)
3853                                 tcp_sack_maybe_coalesce(tp);
3854                         return;
3855                 }
3856         }
3857
3858         /* Could not find an adjacent existing SACK, build a new one,
3859          * put it at the front, and shift everyone else down.  We
3860          * always know there is at least one SACK present already here.
3861          *
3862          * If the sack array is full, forget about the last one.
3863          */
3864         if (this_sack >= TCP_NUM_SACKS) {
3865                 this_sack--;
3866                 tp->rx_opt.num_sacks--;
3867                 sp--;
3868         }
3869         for (; this_sack > 0; this_sack--, sp--)
3870                 *sp = *(sp - 1);
3871
3872 new_sack:
3873         /* Build the new head SACK, and we're done. */
3874         sp->start_seq = seq;
3875         sp->end_seq = end_seq;
3876         tp->rx_opt.num_sacks++;
3877         tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
3878 }
3879
3880 /* RCV.NXT advances, some SACKs should be eaten. */
3881
3882 static void tcp_sack_remove(struct tcp_sock *tp)
3883 {
3884         struct tcp_sack_block *sp = &tp->selective_acks[0];
3885         int num_sacks = tp->rx_opt.num_sacks;
3886         int this_sack;
3887
3888         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3889         if (skb_queue_empty(&tp->out_of_order_queue)) {
3890                 tp->rx_opt.num_sacks = 0;
3891                 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3892                 return;
3893         }
3894
3895         for (this_sack = 0; this_sack < num_sacks;) {
3896                 /* Check if the start of the sack is covered by RCV.NXT. */
3897                 if (!before(tp->rcv_nxt, sp->start_seq)) {
3898                         int i;
3899
3900                         /* RCV.NXT must cover all the block! */
3901                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
3902
3903                         /* Zap this SACK, by moving forward any other SACKS. */
3904                         for (i=this_sack+1; i < num_sacks; i++)
3905                                 tp->selective_acks[i-1] = tp->selective_acks[i];
3906                         num_sacks--;
3907                         continue;
3908                 }
3909                 this_sack++;
3910                 sp++;
3911         }
3912         if (num_sacks != tp->rx_opt.num_sacks) {
3913                 tp->rx_opt.num_sacks = num_sacks;
3914                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
3915                                        tp->rx_opt.dsack;
3916         }
3917 }
3918
3919 /* This one checks to see if we can put data from the
3920  * out_of_order queue into the receive_queue.
3921  */
3922 static void tcp_ofo_queue(struct sock *sk)
3923 {
3924         struct tcp_sock *tp = tcp_sk(sk);
3925         __u32 dsack_high = tp->rcv_nxt;
3926         struct sk_buff *skb;
3927
3928         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3929                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3930                         break;
3931
3932                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3933                         __u32 dsack = dsack_high;
3934                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3935                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
3936                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
3937                 }
3938
3939                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3940                         SOCK_DEBUG(sk, "ofo packet was already received \n");
3941                         __skb_unlink(skb, &tp->out_of_order_queue);
3942                         __kfree_skb(skb);
3943                         continue;
3944                 }
3945                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3946                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3947                            TCP_SKB_CB(skb)->end_seq);
3948
3949                 __skb_unlink(skb, &tp->out_of_order_queue);
3950                 __skb_queue_tail(&sk->sk_receive_queue, skb);
3951                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3952                 if (tcp_hdr(skb)->fin)
3953                         tcp_fin(skb, sk, tcp_hdr(skb));
3954         }
3955 }
3956
3957 static int tcp_prune_ofo_queue(struct sock *sk);
3958 static int tcp_prune_queue(struct sock *sk);
3959
3960 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
3961 {
3962         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3963             !sk_rmem_schedule(sk, size)) {
3964
3965                 if (tcp_prune_queue(sk) < 0)
3966                         return -1;
3967
3968                 if (!sk_rmem_schedule(sk, size)) {
3969                         if (!tcp_prune_ofo_queue(sk))
3970                                 return -1;
3971
3972                         if (!sk_rmem_schedule(sk, size))
3973                                 return -1;
3974                 }
3975         }
3976         return 0;
3977 }
3978
3979 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3980 {
3981         struct tcphdr *th = tcp_hdr(skb);
3982         struct tcp_sock *tp = tcp_sk(sk);
3983         int eaten = -1;
3984
3985         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3986                 goto drop;
3987
3988         __skb_pull(skb, th->doff * 4);
3989
3990         TCP_ECN_accept_cwr(tp, skb);
3991
3992         if (tp->rx_opt.dsack) {
3993                 tp->rx_opt.dsack = 0;
3994                 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
3995         }
3996
3997         /*  Queue data for delivery to the user.
3998          *  Packets in sequence go to the receive queue.
3999          *  Out of sequence packets to the out_of_order_queue.
4000          */
4001         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4002                 if (tcp_receive_window(tp) == 0)
4003                         goto out_of_window;
4004
4005                 /* Ok. In sequence. In window. */
4006                 if (tp->ucopy.task == current &&
4007                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4008                     sock_owned_by_user(sk) && !tp->urg_data) {
4009                         int chunk = min_t(unsigned int, skb->len,
4010                                           tp->ucopy.len);
4011
4012                         __set_current_state(TASK_RUNNING);
4013
4014                         local_bh_enable();
4015                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4016                                 tp->ucopy.len -= chunk;
4017                                 tp->copied_seq += chunk;
4018                                 eaten = (chunk == skb->len && !th->fin);
4019                                 tcp_rcv_space_adjust(sk);
4020                         }
4021                         local_bh_disable();
4022                 }
4023
4024                 if (eaten <= 0) {
4025 queue_and_out:
4026                         if (eaten < 0 &&
4027                             tcp_try_rmem_schedule(sk, skb->truesize))
4028                                 goto drop;
4029
4030                         skb_set_owner_r(skb, sk);
4031                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4032                 }
4033                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4034                 if (skb->len)
4035                         tcp_event_data_recv(sk, skb);
4036                 if (th->fin)
4037                         tcp_fin(skb, sk, th);
4038
4039                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4040                         tcp_ofo_queue(sk);
4041
4042                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4043                          * gap in queue is filled.
4044                          */
4045                         if (skb_queue_empty(&tp->out_of_order_queue))
4046                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4047                 }
4048
4049                 if (tp->rx_opt.num_sacks)
4050                         tcp_sack_remove(tp);
4051
4052                 tcp_fast_path_check(sk);
4053
4054                 if (eaten > 0)
4055                         __kfree_skb(skb);
4056                 else if (!sock_flag(sk, SOCK_DEAD))
4057                         sk->sk_data_ready(sk, 0);
4058                 return;
4059         }
4060
4061         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4062                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4063                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4064                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4065
4066 out_of_window:
4067                 tcp_enter_quickack_mode(sk);
4068                 inet_csk_schedule_ack(sk);
4069 drop:
4070                 __kfree_skb(skb);
4071                 return;
4072         }
4073
4074         /* Out of window. F.e. zero window probe. */
4075         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4076                 goto out_of_window;
4077
4078         tcp_enter_quickack_mode(sk);
4079
4080         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4081                 /* Partial packet, seq < rcv_next < end_seq */
4082                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4083                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4084                            TCP_SKB_CB(skb)->end_seq);
4085
4086                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4087
4088                 /* If window is closed, drop tail of packet. But after
4089                  * remembering D-SACK for its head made in previous line.
4090                  */
4091                 if (!tcp_receive_window(tp))
4092                         goto out_of_window;
4093                 goto queue_and_out;
4094         }
4095
4096         TCP_ECN_check_ce(tp, skb);
4097
4098         if (tcp_try_rmem_schedule(sk, skb->truesize))
4099                 goto drop;
4100
4101         /* Disable header prediction. */
4102         tp->pred_flags = 0;
4103         inet_csk_schedule_ack(sk);
4104
4105         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4106                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4107
4108         skb_set_owner_r(skb, sk);
4109
4110         if (!skb_peek(&tp->out_of_order_queue)) {
4111                 /* Initial out of order segment, build 1 SACK. */
4112                 if (tcp_is_sack(tp)) {
4113                         tp->rx_opt.num_sacks = 1;
4114                         tp->rx_opt.dsack     = 0;
4115                         tp->rx_opt.eff_sacks = 1;
4116                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4117                         tp->selective_acks[0].end_seq =
4118                                                 TCP_SKB_CB(skb)->end_seq;
4119                 }
4120                 __skb_queue_head(&tp->out_of_order_queue, skb);
4121         } else {
4122                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4123                 u32 seq = TCP_SKB_CB(skb)->seq;
4124                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4125
4126                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4127                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4128
4129                         if (!tp->rx_opt.num_sacks ||
4130                             tp->selective_acks[0].end_seq != seq)
4131                                 goto add_sack;
4132
4133                         /* Common case: data arrive in order after hole. */
4134                         tp->selective_acks[0].end_seq = end_seq;
4135                         return;
4136                 }
4137
4138                 /* Find place to insert this segment. */
4139                 do {
4140                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
4141                                 break;
4142                 } while ((skb1 = skb1->prev) !=
4143                          (struct sk_buff *)&tp->out_of_order_queue);
4144
4145                 /* Do skb overlap to previous one? */
4146                 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4147                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4148                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4149                                 /* All the bits are present. Drop. */
4150                                 __kfree_skb(skb);
4151                                 tcp_dsack_set(sk, seq, end_seq);
4152                                 goto add_sack;
4153                         }
4154                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4155                                 /* Partial overlap. */
4156                                 tcp_dsack_set(sk, seq,
4157                                               TCP_SKB_CB(skb1)->end_seq);
4158                         } else {
4159                                 skb1 = skb1->prev;
4160                         }
4161                 }
4162                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4163
4164                 /* And clean segments covered by new one as whole. */
4165                 while ((skb1 = skb->next) !=
4166                        (struct sk_buff *)&tp->out_of_order_queue &&
4167                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4168                         if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4169                                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4170                                                  end_seq);
4171                                 break;
4172                         }
4173                         __skb_unlink(skb1, &tp->out_of_order_queue);
4174                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4175                                          TCP_SKB_CB(skb1)->end_seq);
4176                         __kfree_skb(skb1);
4177                 }
4178
4179 add_sack:
4180                 if (tcp_is_sack(tp))
4181                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
4182         }
4183 }
4184
4185 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4186                                         struct sk_buff_head *list)
4187 {
4188         struct sk_buff *next = skb->next;
4189
4190         __skb_unlink(skb, list);
4191         __kfree_skb(skb);
4192         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4193
4194         return next;
4195 }
4196
4197 /* Collapse contiguous sequence of skbs head..tail with
4198  * sequence numbers start..end.
4199  * Segments with FIN/SYN are not collapsed (only because this
4200  * simplifies code)
4201  */
4202 static void
4203 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4204              struct sk_buff *head, struct sk_buff *tail,
4205              u32 start, u32 end)
4206 {
4207         struct sk_buff *skb;
4208
4209         /* First, check that queue is collapsible and find
4210          * the point where collapsing can be useful. */
4211         for (skb = head; skb != tail;) {
4212                 /* No new bits? It is possible on ofo queue. */
4213                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4214                         skb = tcp_collapse_one(sk, skb, list);
4215                         continue;
4216                 }
4217
4218                 /* The first skb to collapse is:
4219                  * - not SYN/FIN and
4220                  * - bloated or contains data before "start" or
4221                  *   overlaps to the next one.
4222                  */
4223                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4224                     (tcp_win_from_space(skb->truesize) > skb->len ||
4225                      before(TCP_SKB_CB(skb)->seq, start) ||
4226                      (skb->next != tail &&
4227                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4228                         break;
4229
4230                 /* Decided to skip this, advance start seq. */
4231                 start = TCP_SKB_CB(skb)->end_seq;
4232                 skb = skb->next;
4233         }
4234         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4235                 return;
4236
4237         while (before(start, end)) {
4238                 struct sk_buff *nskb;
4239                 unsigned int header = skb_headroom(skb);
4240                 int copy = SKB_MAX_ORDER(header, 0);
4241
4242                 /* Too big header? This can happen with IPv6. */
4243                 if (copy < 0)
4244                         return;
4245                 if (end - start < copy)
4246                         copy = end - start;
4247                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4248                 if (!nskb)
4249                         return;
4250
4251                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4252                 skb_set_network_header(nskb, (skb_network_header(skb) -
4253                                               skb->head));
4254                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4255                                                 skb->head));
4256                 skb_reserve(nskb, header);
4257                 memcpy(nskb->head, skb->head, header);
4258                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4259                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4260                 __skb_queue_before(list, skb, nskb);
4261                 skb_set_owner_r(nskb, sk);
4262
4263                 /* Copy data, releasing collapsed skbs. */
4264                 while (copy > 0) {
4265                         int offset = start - TCP_SKB_CB(skb)->seq;
4266                         int size = TCP_SKB_CB(skb)->end_seq - start;
4267
4268                         BUG_ON(offset < 0);
4269                         if (size > 0) {
4270                                 size = min(copy, size);
4271                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4272                                         BUG();
4273                                 TCP_SKB_CB(nskb)->end_seq += size;
4274                                 copy -= size;
4275                                 start += size;
4276                         }
4277                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4278                                 skb = tcp_collapse_one(sk, skb, list);
4279                                 if (skb == tail ||
4280                                     tcp_hdr(skb)->syn ||
4281                                     tcp_hdr(skb)->fin)
4282                                         return;
4283                         }
4284                 }
4285         }
4286 }
4287
4288 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4289  * and tcp_collapse() them until all the queue is collapsed.
4290  */
4291 static void tcp_collapse_ofo_queue(struct sock *sk)
4292 {
4293         struct tcp_sock *tp = tcp_sk(sk);
4294         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4295         struct sk_buff *head;
4296         u32 start, end;
4297
4298         if (skb == NULL)
4299                 return;
4300
4301         start = TCP_SKB_CB(skb)->seq;
4302         end = TCP_SKB_CB(skb)->end_seq;
4303         head = skb;
4304
4305         for (;;) {
4306                 skb = skb->next;
4307
4308                 /* Segment is terminated when we see gap or when
4309                  * we are at the end of all the queue. */
4310                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4311                     after(TCP_SKB_CB(skb)->seq, end) ||
4312                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4313                         tcp_collapse(sk, &tp->out_of_order_queue,
4314                                      head, skb, start, end);
4315                         head = skb;
4316                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4317                                 break;
4318                         /* Start new segment */
4319                         start = TCP_SKB_CB(skb)->seq;
4320                         end = TCP_SKB_CB(skb)->end_seq;
4321                 } else {
4322                         if (before(TCP_SKB_CB(skb)->seq, start))
4323                                 start = TCP_SKB_CB(skb)->seq;
4324                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4325                                 end = TCP_SKB_CB(skb)->end_seq;
4326                 }
4327         }
4328 }
4329
4330 /*
4331  * Purge the out-of-order queue.
4332  * Return true if queue was pruned.
4333  */
4334 static int tcp_prune_ofo_queue(struct sock *sk)
4335 {
4336         struct tcp_sock *tp = tcp_sk(sk);
4337         int res = 0;
4338
4339         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4340                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4341                 __skb_queue_purge(&tp->out_of_order_queue);
4342
4343                 /* Reset SACK state.  A conforming SACK implementation will
4344                  * do the same at a timeout based retransmit.  When a connection
4345                  * is in a sad state like this, we care only about integrity
4346                  * of the connection not performance.
4347                  */
4348                 if (tp->rx_opt.sack_ok)
4349                         tcp_sack_reset(&tp->rx_opt);
4350                 sk_mem_reclaim(sk);
4351                 res = 1;
4352         }
4353         return res;
4354 }
4355
4356 /* Reduce allocated memory if we can, trying to get
4357  * the socket within its memory limits again.
4358  *
4359  * Return less than zero if we should start dropping frames
4360  * until the socket owning process reads some of the data
4361  * to stabilize the situation.
4362  */
4363 static int tcp_prune_queue(struct sock *sk)
4364 {
4365         struct tcp_sock *tp = tcp_sk(sk);
4366
4367         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4368
4369         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4370
4371         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4372                 tcp_clamp_window(sk);
4373         else if (tcp_memory_pressure)
4374                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4375
4376         tcp_collapse_ofo_queue(sk);
4377         tcp_collapse(sk, &sk->sk_receive_queue,
4378                      sk->sk_receive_queue.next,
4379                      (struct sk_buff *)&sk->sk_receive_queue,
4380                      tp->copied_seq, tp->rcv_nxt);
4381         sk_mem_reclaim(sk);
4382
4383         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4384                 return 0;
4385
4386         /* Collapsing did not help, destructive actions follow.
4387          * This must not ever occur. */
4388
4389         tcp_prune_ofo_queue(sk);
4390
4391         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4392                 return 0;
4393
4394         /* If we are really being abused, tell the caller to silently
4395          * drop receive data on the floor.  It will get retransmitted
4396          * and hopefully then we'll have sufficient space.
4397          */
4398         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4399
4400         /* Massive buffer overcommit. */
4401         tp->pred_flags = 0;
4402         return -1;
4403 }
4404
4405 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4406  * As additional protections, we do not touch cwnd in retransmission phases,
4407  * and if application hit its sndbuf limit recently.
4408  */
4409 void tcp_cwnd_application_limited(struct sock *sk)
4410 {
4411         struct tcp_sock *tp = tcp_sk(sk);
4412
4413         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4414             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4415                 /* Limited by application or receiver window. */
4416                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4417                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4418                 if (win_used < tp->snd_cwnd) {
4419                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4420                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4421                 }
4422                 tp->snd_cwnd_used = 0;
4423         }
4424         tp->snd_cwnd_stamp = tcp_time_stamp;
4425 }
4426
4427 static int tcp_should_expand_sndbuf(struct sock *sk)
4428 {
4429         struct tcp_sock *tp = tcp_sk(sk);
4430
4431         /* If the user specified a specific send buffer setting, do
4432          * not modify it.
4433          */
4434         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4435                 return 0;
4436
4437         /* If we are under global TCP memory pressure, do not expand.  */
4438         if (tcp_memory_pressure)
4439                 return 0;
4440
4441         /* If we are under soft global TCP memory pressure, do not expand.  */
4442         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4443                 return 0;
4444
4445         /* If we filled the congestion window, do not expand.  */
4446         if (tp->packets_out >= tp->snd_cwnd)
4447                 return 0;
4448
4449         return 1;
4450 }
4451
4452 /* When incoming ACK allowed to free some skb from write_queue,
4453  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4454  * on the exit from tcp input handler.
4455  *
4456  * PROBLEM: sndbuf expansion does not work well with largesend.
4457  */
4458 static void tcp_new_space(struct sock *sk)
4459 {
4460         struct tcp_sock *tp = tcp_sk(sk);
4461
4462         if (tcp_should_expand_sndbuf(sk)) {
4463                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4464                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4465                 int demanded = max_t(unsigned int, tp->snd_cwnd,
4466                                      tp->reordering + 1);
4467                 sndmem *= 2 * demanded;
4468                 if (sndmem > sk->sk_sndbuf)
4469                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4470                 tp->snd_cwnd_stamp = tcp_time_stamp;
4471         }
4472
4473         sk->sk_write_space(sk);
4474 }
4475
4476 static void tcp_check_space(struct sock *sk)
4477 {
4478         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4479                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4480                 if (sk->sk_socket &&
4481                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4482                         tcp_new_space(sk);
4483         }
4484 }
4485
4486 static inline void tcp_data_snd_check(struct sock *sk)
4487 {
4488         tcp_push_pending_frames(sk);
4489         tcp_check_space(sk);
4490 }
4491
4492 /*
4493  * Check if sending an ack is needed.
4494  */
4495 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4496 {
4497         struct tcp_sock *tp = tcp_sk(sk);
4498
4499             /* More than one full frame received... */
4500         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4501              /* ... and right edge of window advances far enough.
4502               * (tcp_recvmsg() will send ACK otherwise). Or...
4503               */
4504              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4505             /* We ACK each frame or... */
4506             tcp_in_quickack_mode(sk) ||
4507             /* We have out of order data. */
4508             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4509                 /* Then ack it now */
4510                 tcp_send_ack(sk);
4511         } else {
4512                 /* Else, send delayed ack. */
4513                 tcp_send_delayed_ack(sk);
4514         }
4515 }
4516
4517 static inline void tcp_ack_snd_check(struct sock *sk)
4518 {
4519         if (!inet_csk_ack_scheduled(sk)) {
4520                 /* We sent a data segment already. */
4521                 return;
4522         }
4523         __tcp_ack_snd_check(sk, 1);
4524 }
4525
4526 /*
4527  *      This routine is only called when we have urgent data
4528  *      signaled. Its the 'slow' part of tcp_urg. It could be
4529  *      moved inline now as tcp_urg is only called from one
4530  *      place. We handle URGent data wrong. We have to - as
4531  *      BSD still doesn't use the correction from RFC961.
4532  *      For 1003.1g we should support a new option TCP_STDURG to permit
4533  *      either form (or just set the sysctl tcp_stdurg).
4534  */
4535
4536 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4537 {
4538         struct tcp_sock *tp = tcp_sk(sk);
4539         u32 ptr = ntohs(th->urg_ptr);
4540
4541         if (ptr && !sysctl_tcp_stdurg)
4542                 ptr--;
4543         ptr += ntohl(th->seq);
4544
4545         /* Ignore urgent data that we've already seen and read. */
4546         if (after(tp->copied_seq, ptr))
4547                 return;
4548
4549         /* Do not replay urg ptr.
4550          *
4551          * NOTE: interesting situation not covered by specs.
4552          * Misbehaving sender may send urg ptr, pointing to segment,
4553          * which we already have in ofo queue. We are not able to fetch
4554          * such data and will stay in TCP_URG_NOTYET until will be eaten
4555          * by recvmsg(). Seems, we are not obliged to handle such wicked
4556          * situations. But it is worth to think about possibility of some
4557          * DoSes using some hypothetical application level deadlock.
4558          */
4559         if (before(ptr, tp->rcv_nxt))
4560                 return;
4561
4562         /* Do we already have a newer (or duplicate) urgent pointer? */
4563         if (tp->urg_data && !after(ptr, tp->urg_seq))
4564                 return;
4565
4566         /* Tell the world about our new urgent pointer. */
4567         sk_send_sigurg(sk);
4568
4569         /* We may be adding urgent data when the last byte read was
4570          * urgent. To do this requires some care. We cannot just ignore
4571          * tp->copied_seq since we would read the last urgent byte again
4572          * as data, nor can we alter copied_seq until this data arrives
4573          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4574          *
4575          * NOTE. Double Dutch. Rendering to plain English: author of comment
4576          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4577          * and expect that both A and B disappear from stream. This is _wrong_.
4578          * Though this happens in BSD with high probability, this is occasional.
4579          * Any application relying on this is buggy. Note also, that fix "works"
4580          * only in this artificial test. Insert some normal data between A and B and we will
4581          * decline of BSD again. Verdict: it is better to remove to trap
4582          * buggy users.
4583          */
4584         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4585             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4586                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4587                 tp->copied_seq++;
4588                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4589                         __skb_unlink(skb, &sk->sk_receive_queue);
4590                         __kfree_skb(skb);
4591                 }
4592         }
4593
4594         tp->urg_data = TCP_URG_NOTYET;
4595         tp->urg_seq = ptr;
4596
4597         /* Disable header prediction. */
4598         tp->pred_flags = 0;
4599 }
4600
4601 /* This is the 'fast' part of urgent handling. */
4602 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4603 {
4604         struct tcp_sock *tp = tcp_sk(sk);
4605
4606         /* Check if we get a new urgent pointer - normally not. */
4607         if (th->urg)
4608                 tcp_check_urg(sk, th);
4609
4610         /* Do we wait for any urgent data? - normally not... */
4611         if (tp->urg_data == TCP_URG_NOTYET) {
4612                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4613                           th->syn;
4614
4615                 /* Is the urgent pointer pointing into this packet? */
4616                 if (ptr < skb->len) {
4617                         u8 tmp;
4618                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4619                                 BUG();
4620                         tp->urg_data = TCP_URG_VALID | tmp;
4621                         if (!sock_flag(sk, SOCK_DEAD))
4622                                 sk->sk_data_ready(sk, 0);
4623                 }
4624         }
4625 }
4626
4627 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4628 {
4629         struct tcp_sock *tp = tcp_sk(sk);
4630         int chunk = skb->len - hlen;
4631         int err;
4632
4633         local_bh_enable();
4634         if (skb_csum_unnecessary(skb))
4635                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4636         else
4637                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4638                                                        tp->ucopy.iov);
4639
4640         if (!err) {
4641                 tp->ucopy.len -= chunk;
4642                 tp->copied_seq += chunk;
4643                 tcp_rcv_space_adjust(sk);
4644         }
4645
4646         local_bh_disable();
4647         return err;
4648 }
4649
4650 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4651                                             struct sk_buff *skb)
4652 {
4653         __sum16 result;
4654
4655         if (sock_owned_by_user(sk)) {
4656                 local_bh_enable();
4657                 result = __tcp_checksum_complete(skb);
4658                 local_bh_disable();
4659         } else {
4660                 result = __tcp_checksum_complete(skb);
4661         }
4662         return result;
4663 }
4664
4665 static inline int tcp_checksum_complete_user(struct sock *sk,
4666                                              struct sk_buff *skb)
4667 {
4668         return !skb_csum_unnecessary(skb) &&
4669                __tcp_checksum_complete_user(sk, skb);
4670 }
4671
4672 #ifdef CONFIG_NET_DMA
4673 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4674                                   int hlen)
4675 {
4676         struct tcp_sock *tp = tcp_sk(sk);
4677         int chunk = skb->len - hlen;
4678         int dma_cookie;
4679         int copied_early = 0;
4680
4681         if (tp->ucopy.wakeup)
4682                 return 0;
4683
4684         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4685                 tp->ucopy.dma_chan = get_softnet_dma();
4686
4687         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4688
4689                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4690                                                          skb, hlen,
4691                                                          tp->ucopy.iov, chunk,
4692                                                          tp->ucopy.pinned_list);
4693
4694                 if (dma_cookie < 0)
4695                         goto out;
4696
4697                 tp->ucopy.dma_cookie = dma_cookie;
4698                 copied_early = 1;
4699
4700                 tp->ucopy.len -= chunk;
4701                 tp->copied_seq += chunk;
4702                 tcp_rcv_space_adjust(sk);
4703
4704                 if ((tp->ucopy.len == 0) ||
4705                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4706                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4707                         tp->ucopy.wakeup = 1;
4708                         sk->sk_data_ready(sk, 0);
4709                 }
4710         } else if (chunk > 0) {
4711                 tp->ucopy.wakeup = 1;
4712                 sk->sk_data_ready(sk, 0);
4713         }
4714 out:
4715         return copied_early;
4716 }
4717 #endif /* CONFIG_NET_DMA */
4718
4719 /* Does PAWS and seqno based validation of an incoming segment, flags will
4720  * play significant role here.
4721  */
4722 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4723                               struct tcphdr *th, int syn_inerr)
4724 {
4725         struct tcp_sock *tp = tcp_sk(sk);
4726
4727         /* RFC1323: H1. Apply PAWS check first. */
4728         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4729             tcp_paws_discard(sk, skb)) {
4730                 if (!th->rst) {
4731                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4732                         tcp_send_dupack(sk, skb);
4733                         goto discard;
4734                 }
4735                 /* Reset is accepted even if it did not pass PAWS. */
4736         }
4737
4738         /* Step 1: check sequence number */
4739         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4740                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4741                  * (RST) segments are validated by checking their SEQ-fields."
4742                  * And page 69: "If an incoming segment is not acceptable,
4743                  * an acknowledgment should be sent in reply (unless the RST
4744                  * bit is set, if so drop the segment and return)".
4745                  */
4746                 if (!th->rst)
4747                         tcp_send_dupack(sk, skb);
4748                 goto discard;
4749         }
4750
4751         /* Step 2: check RST bit */
4752         if (th->rst) {
4753                 tcp_reset(sk);
4754                 goto discard;
4755         }
4756
4757         /* ts_recent update must be made after we are sure that the packet
4758          * is in window.
4759          */
4760         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4761
4762         /* step 3: check security and precedence [ignored] */
4763
4764         /* step 4: Check for a SYN in window. */
4765         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4766                 if (syn_inerr)
4767                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4768                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
4769                 tcp_reset(sk);
4770                 return -1;
4771         }
4772
4773         return 1;
4774
4775 discard:
4776         __kfree_skb(skb);
4777         return 0;
4778 }
4779
4780 /*
4781  *      TCP receive function for the ESTABLISHED state.
4782  *
4783  *      It is split into a fast path and a slow path. The fast path is
4784  *      disabled when:
4785  *      - A zero window was announced from us - zero window probing
4786  *        is only handled properly in the slow path.
4787  *      - Out of order segments arrived.
4788  *      - Urgent data is expected.
4789  *      - There is no buffer space left
4790  *      - Unexpected TCP flags/window values/header lengths are received
4791  *        (detected by checking the TCP header against pred_flags)
4792  *      - Data is sent in both directions. Fast path only supports pure senders
4793  *        or pure receivers (this means either the sequence number or the ack
4794  *        value must stay constant)
4795  *      - Unexpected TCP option.
4796  *
4797  *      When these conditions are not satisfied it drops into a standard
4798  *      receive procedure patterned after RFC793 to handle all cases.
4799  *      The first three cases are guaranteed by proper pred_flags setting,
4800  *      the rest is checked inline. Fast processing is turned on in
4801  *      tcp_data_queue when everything is OK.
4802  */
4803 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4804                         struct tcphdr *th, unsigned len)
4805 {
4806         struct tcp_sock *tp = tcp_sk(sk);
4807         int res;
4808
4809         /*
4810          *      Header prediction.
4811          *      The code loosely follows the one in the famous
4812          *      "30 instruction TCP receive" Van Jacobson mail.
4813          *
4814          *      Van's trick is to deposit buffers into socket queue
4815          *      on a device interrupt, to call tcp_recv function
4816          *      on the receive process context and checksum and copy
4817          *      the buffer to user space. smart...
4818          *
4819          *      Our current scheme is not silly either but we take the
4820          *      extra cost of the net_bh soft interrupt processing...
4821          *      We do checksum and copy also but from device to kernel.
4822          */
4823
4824         tp->rx_opt.saw_tstamp = 0;
4825
4826         /*      pred_flags is 0xS?10 << 16 + snd_wnd
4827          *      if header_prediction is to be made
4828          *      'S' will always be tp->tcp_header_len >> 2
4829          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4830          *  turn it off (when there are holes in the receive
4831          *       space for instance)
4832          *      PSH flag is ignored.
4833          */
4834
4835         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4836             TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4837                 int tcp_header_len = tp->tcp_header_len;
4838
4839                 /* Timestamp header prediction: tcp_header_len
4840                  * is automatically equal to th->doff*4 due to pred_flags
4841                  * match.
4842                  */
4843
4844                 /* Check timestamp */
4845                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4846                         /* No? Slow path! */
4847                         if (!tcp_parse_aligned_timestamp(tp, th))
4848                                 goto slow_path;
4849
4850                         /* If PAWS failed, check it more carefully in slow path */
4851                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4852                                 goto slow_path;
4853
4854                         /* DO NOT update ts_recent here, if checksum fails
4855                          * and timestamp was corrupted part, it will result
4856                          * in a hung connection since we will drop all
4857                          * future packets due to the PAWS test.
4858                          */
4859                 }
4860
4861                 if (len <= tcp_header_len) {
4862                         /* Bulk data transfer: sender */
4863                         if (len == tcp_header_len) {
4864                                 /* Predicted packet is in window by definition.
4865                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4866                                  * Hence, check seq<=rcv_wup reduces to:
4867                                  */
4868                                 if (tcp_header_len ==
4869                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4870                                     tp->rcv_nxt == tp->rcv_wup)
4871                                         tcp_store_ts_recent(tp);
4872
4873                                 /* We know that such packets are checksummed
4874                                  * on entry.
4875                                  */
4876                                 tcp_ack(sk, skb, 0);
4877                                 __kfree_skb(skb);
4878                                 tcp_data_snd_check(sk);
4879                                 return 0;
4880                         } else { /* Header too small */
4881                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4882                                 goto discard;
4883                         }
4884                 } else {
4885                         int eaten = 0;
4886                         int copied_early = 0;
4887
4888                         if (tp->copied_seq == tp->rcv_nxt &&
4889                             len - tcp_header_len <= tp->ucopy.len) {
4890 #ifdef CONFIG_NET_DMA
4891                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4892                                         copied_early = 1;
4893                                         eaten = 1;
4894                                 }
4895 #endif
4896                                 if (tp->ucopy.task == current &&
4897                                     sock_owned_by_user(sk) && !copied_early) {
4898                                         __set_current_state(TASK_RUNNING);
4899
4900                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4901                                                 eaten = 1;
4902                                 }
4903                                 if (eaten) {
4904                                         /* Predicted packet is in window by definition.
4905                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4906                                          * Hence, check seq<=rcv_wup reduces to:
4907                                          */
4908                                         if (tcp_header_len ==
4909                                             (sizeof(struct tcphdr) +
4910                                              TCPOLEN_TSTAMP_ALIGNED) &&
4911                                             tp->rcv_nxt == tp->rcv_wup)
4912                                                 tcp_store_ts_recent(tp);
4913
4914                                         tcp_rcv_rtt_measure_ts(sk, skb);
4915
4916                                         __skb_pull(skb, tcp_header_len);
4917                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4918                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
4919                                 }
4920                                 if (copied_early)
4921                                         tcp_cleanup_rbuf(sk, skb->len);
4922                         }
4923                         if (!eaten) {
4924                                 if (tcp_checksum_complete_user(sk, skb))
4925                                         goto csum_error;
4926
4927                                 /* Predicted packet is in window by definition.
4928                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4929                                  * Hence, check seq<=rcv_wup reduces to:
4930                                  */
4931                                 if (tcp_header_len ==
4932                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4933                                     tp->rcv_nxt == tp->rcv_wup)
4934                                         tcp_store_ts_recent(tp);
4935
4936                                 tcp_rcv_rtt_measure_ts(sk, skb);
4937
4938                                 if ((int)skb->truesize > sk->sk_forward_alloc)
4939                                         goto step5;
4940
4941                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
4942
4943                                 /* Bulk data transfer: receiver */
4944                                 __skb_pull(skb, tcp_header_len);
4945                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4946                                 skb_set_owner_r(skb, sk);
4947                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4948                         }
4949
4950                         tcp_event_data_recv(sk, skb);
4951
4952                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4953                                 /* Well, only one small jumplet in fast path... */
4954                                 tcp_ack(sk, skb, FLAG_DATA);
4955                                 tcp_data_snd_check(sk);
4956                                 if (!inet_csk_ack_scheduled(sk))
4957                                         goto no_ack;
4958                         }
4959
4960                         if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
4961                                 __tcp_ack_snd_check(sk, 0);
4962 no_ack:
4963 #ifdef CONFIG_NET_DMA
4964                         if (copied_early)
4965                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4966                         else
4967 #endif
4968                         if (eaten)
4969                                 __kfree_skb(skb);
4970                         else
4971                                 sk->sk_data_ready(sk, 0);
4972                         return 0;
4973                 }
4974         }
4975
4976 slow_path:
4977         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
4978                 goto csum_error;
4979
4980         /*
4981          *      Standard slow path.
4982          */
4983
4984         res = tcp_validate_incoming(sk, skb, th, 1);
4985         if (res <= 0)
4986                 return -res;
4987
4988 step5:
4989         if (th->ack)
4990                 tcp_ack(sk, skb, FLAG_SLOWPATH);
4991
4992         tcp_rcv_rtt_measure_ts(sk, skb);
4993
4994         /* Process urgent data. */
4995         tcp_urg(sk, skb, th);
4996
4997         /* step 7: process the segment text */
4998         tcp_data_queue(sk, skb);
4999
5000         tcp_data_snd_check(sk);
5001         tcp_ack_snd_check(sk);
5002         return 0;
5003
5004 csum_error:
5005         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5006
5007 discard:
5008         __kfree_skb(skb);
5009         return 0;
5010 }
5011
5012 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5013                                          struct tcphdr *th, unsigned len)
5014 {
5015         struct tcp_sock *tp = tcp_sk(sk);
5016         struct inet_connection_sock *icsk = inet_csk(sk);
5017         int saved_clamp = tp->rx_opt.mss_clamp;
5018
5019         tcp_parse_options(skb, &tp->rx_opt, 0);
5020
5021         if (th->ack) {
5022                 /* rfc793:
5023                  * "If the state is SYN-SENT then
5024                  *    first check the ACK bit
5025                  *      If the ACK bit is set
5026                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5027                  *        a reset (unless the RST bit is set, if so drop
5028                  *        the segment and return)"
5029                  *
5030                  *  We do not send data with SYN, so that RFC-correct
5031                  *  test reduces to:
5032                  */
5033                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5034                         goto reset_and_undo;
5035
5036                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5037                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5038                              tcp_time_stamp)) {
5039                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5040                         goto reset_and_undo;
5041                 }
5042
5043                 /* Now ACK is acceptable.
5044                  *
5045                  * "If the RST bit is set
5046                  *    If the ACK was acceptable then signal the user "error:
5047                  *    connection reset", drop the segment, enter CLOSED state,
5048                  *    delete TCB, and return."
5049                  */
5050
5051                 if (th->rst) {
5052                         tcp_reset(sk);
5053                         goto discard;
5054                 }
5055
5056                 /* rfc793:
5057                  *   "fifth, if neither of the SYN or RST bits is set then
5058                  *    drop the segment and return."
5059                  *
5060                  *    See note below!
5061                  *                                        --ANK(990513)
5062                  */
5063                 if (!th->syn)
5064                         goto discard_and_undo;
5065
5066                 /* rfc793:
5067                  *   "If the SYN bit is on ...
5068                  *    are acceptable then ...
5069                  *    (our SYN has been ACKed), change the connection
5070                  *    state to ESTABLISHED..."
5071                  */
5072
5073                 TCP_ECN_rcv_synack(tp, th);
5074
5075                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5076                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5077
5078                 /* Ok.. it's good. Set up sequence numbers and
5079                  * move to established.
5080                  */
5081                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5082                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5083
5084                 /* RFC1323: The window in SYN & SYN/ACK segments is
5085                  * never scaled.
5086                  */
5087                 tp->snd_wnd = ntohs(th->window);
5088                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
5089
5090                 if (!tp->rx_opt.wscale_ok) {
5091                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5092                         tp->window_clamp = min(tp->window_clamp, 65535U);
5093                 }
5094
5095                 if (tp->rx_opt.saw_tstamp) {
5096                         tp->rx_opt.tstamp_ok       = 1;
5097                         tp->tcp_header_len =
5098                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5099                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5100                         tcp_store_ts_recent(tp);
5101                 } else {
5102                         tp->tcp_header_len = sizeof(struct tcphdr);
5103                 }
5104
5105                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5106                         tcp_enable_fack(tp);
5107
5108                 tcp_mtup_init(sk);
5109                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5110                 tcp_initialize_rcv_mss(sk);
5111
5112                 /* Remember, tcp_poll() does not lock socket!
5113                  * Change state from SYN-SENT only after copied_seq
5114                  * is initialized. */
5115                 tp->copied_seq = tp->rcv_nxt;
5116                 smp_mb();
5117                 tcp_set_state(sk, TCP_ESTABLISHED);
5118
5119                 security_inet_conn_established(sk, skb);
5120
5121                 /* Make sure socket is routed, for correct metrics.  */
5122                 icsk->icsk_af_ops->rebuild_header(sk);
5123
5124                 tcp_init_metrics(sk);
5125
5126                 tcp_init_congestion_control(sk);
5127
5128                 /* Prevent spurious tcp_cwnd_restart() on first data
5129                  * packet.
5130                  */
5131                 tp->lsndtime = tcp_time_stamp;
5132
5133                 tcp_init_buffer_space(sk);
5134
5135                 if (sock_flag(sk, SOCK_KEEPOPEN))
5136                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5137
5138                 if (!tp->rx_opt.snd_wscale)
5139                         __tcp_fast_path_on(tp, tp->snd_wnd);
5140                 else
5141                         tp->pred_flags = 0;
5142
5143                 if (!sock_flag(sk, SOCK_DEAD)) {
5144                         sk->sk_state_change(sk);
5145                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5146                 }
5147
5148                 if (sk->sk_write_pending ||
5149                     icsk->icsk_accept_queue.rskq_defer_accept ||
5150                     icsk->icsk_ack.pingpong) {
5151                         /* Save one ACK. Data will be ready after
5152                          * several ticks, if write_pending is set.
5153                          *
5154                          * It may be deleted, but with this feature tcpdumps
5155                          * look so _wonderfully_ clever, that I was not able
5156                          * to stand against the temptation 8)     --ANK
5157                          */
5158                         inet_csk_schedule_ack(sk);
5159                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5160                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
5161                         tcp_incr_quickack(sk);
5162                         tcp_enter_quickack_mode(sk);
5163                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5164                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5165
5166 discard:
5167                         __kfree_skb(skb);
5168                         return 0;
5169                 } else {
5170                         tcp_send_ack(sk);
5171                 }
5172                 return -1;
5173         }
5174
5175         /* No ACK in the segment */
5176
5177         if (th->rst) {
5178                 /* rfc793:
5179                  * "If the RST bit is set
5180                  *
5181                  *      Otherwise (no ACK) drop the segment and return."
5182                  */
5183
5184                 goto discard_and_undo;
5185         }
5186
5187         /* PAWS check. */
5188         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5189             tcp_paws_check(&tp->rx_opt, 0))
5190                 goto discard_and_undo;
5191
5192         if (th->syn) {
5193                 /* We see SYN without ACK. It is attempt of
5194                  * simultaneous connect with crossed SYNs.
5195                  * Particularly, it can be connect to self.
5196                  */
5197                 tcp_set_state(sk, TCP_SYN_RECV);
5198
5199                 if (tp->rx_opt.saw_tstamp) {
5200                         tp->rx_opt.tstamp_ok = 1;
5201                         tcp_store_ts_recent(tp);
5202                         tp->tcp_header_len =
5203                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5204                 } else {
5205                         tp->tcp_header_len = sizeof(struct tcphdr);
5206                 }
5207
5208                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5209                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5210
5211                 /* RFC1323: The window in SYN & SYN/ACK segments is
5212                  * never scaled.
5213                  */
5214                 tp->snd_wnd    = ntohs(th->window);
5215                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5216                 tp->max_window = tp->snd_wnd;
5217
5218                 TCP_ECN_rcv_syn(tp, th);
5219
5220                 tcp_mtup_init(sk);
5221                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5222                 tcp_initialize_rcv_mss(sk);
5223
5224                 tcp_send_synack(sk);
5225 #if 0
5226                 /* Note, we could accept data and URG from this segment.
5227                  * There are no obstacles to make this.
5228                  *
5229                  * However, if we ignore data in ACKless segments sometimes,
5230                  * we have no reasons to accept it sometimes.
5231                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5232                  * is not flawless. So, discard packet for sanity.
5233                  * Uncomment this return to process the data.
5234                  */
5235                 return -1;
5236 #else
5237                 goto discard;
5238 #endif
5239         }
5240         /* "fifth, if neither of the SYN or RST bits is set then
5241          * drop the segment and return."
5242          */
5243
5244 discard_and_undo:
5245         tcp_clear_options(&tp->rx_opt);
5246         tp->rx_opt.mss_clamp = saved_clamp;
5247         goto discard;
5248
5249 reset_and_undo:
5250         tcp_clear_options(&tp->rx_opt);
5251         tp->rx_opt.mss_clamp = saved_clamp;
5252         return 1;
5253 }
5254
5255 /*
5256  *      This function implements the receiving procedure of RFC 793 for
5257  *      all states except ESTABLISHED and TIME_WAIT.
5258  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5259  *      address independent.
5260  */
5261
5262 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5263                           struct tcphdr *th, unsigned len)
5264 {
5265         struct tcp_sock *tp = tcp_sk(sk);
5266         struct inet_connection_sock *icsk = inet_csk(sk);
5267         int queued = 0;
5268         int res;
5269
5270         tp->rx_opt.saw_tstamp = 0;
5271
5272         switch (sk->sk_state) {
5273         case TCP_CLOSE:
5274                 goto discard;
5275
5276         case TCP_LISTEN:
5277                 if (th->ack)
5278                         return 1;
5279
5280                 if (th->rst)
5281                         goto discard;
5282
5283                 if (th->syn) {
5284                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5285                                 return 1;
5286
5287                         /* Now we have several options: In theory there is
5288                          * nothing else in the frame. KA9Q has an option to
5289                          * send data with the syn, BSD accepts data with the
5290                          * syn up to the [to be] advertised window and
5291                          * Solaris 2.1 gives you a protocol error. For now
5292                          * we just ignore it, that fits the spec precisely
5293                          * and avoids incompatibilities. It would be nice in
5294                          * future to drop through and process the data.
5295                          *
5296                          * Now that TTCP is starting to be used we ought to
5297                          * queue this data.
5298                          * But, this leaves one open to an easy denial of
5299                          * service attack, and SYN cookies can't defend
5300                          * against this problem. So, we drop the data
5301                          * in the interest of security over speed unless
5302                          * it's still in use.
5303                          */
5304                         kfree_skb(skb);
5305                         return 0;
5306                 }
5307                 goto discard;
5308
5309         case TCP_SYN_SENT:
5310                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5311                 if (queued >= 0)
5312                         return queued;
5313
5314                 /* Do step6 onward by hand. */
5315                 tcp_urg(sk, skb, th);
5316                 __kfree_skb(skb);
5317                 tcp_data_snd_check(sk);
5318                 return 0;
5319         }
5320
5321         res = tcp_validate_incoming(sk, skb, th, 0);
5322         if (res <= 0)
5323                 return -res;
5324
5325         /* step 5: check the ACK field */
5326         if (th->ack) {
5327                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5328
5329                 switch (sk->sk_state) {
5330                 case TCP_SYN_RECV:
5331                         if (acceptable) {
5332                                 tp->copied_seq = tp->rcv_nxt;
5333                                 smp_mb();
5334                                 tcp_set_state(sk, TCP_ESTABLISHED);
5335                                 sk->sk_state_change(sk);
5336
5337                                 /* Note, that this wakeup is only for marginal
5338                                  * crossed SYN case. Passively open sockets
5339                                  * are not waked up, because sk->sk_sleep ==
5340                                  * NULL and sk->sk_socket == NULL.
5341                                  */
5342                                 if (sk->sk_socket)
5343                                         sk_wake_async(sk,
5344                                                       SOCK_WAKE_IO, POLL_OUT);
5345
5346                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5347                                 tp->snd_wnd = ntohs(th->window) <<
5348                                               tp->rx_opt.snd_wscale;
5349                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5350                                             TCP_SKB_CB(skb)->seq);
5351
5352                                 /* tcp_ack considers this ACK as duplicate
5353                                  * and does not calculate rtt.
5354                                  * Fix it at least with timestamps.
5355                                  */
5356                                 if (tp->rx_opt.saw_tstamp &&
5357                                     tp->rx_opt.rcv_tsecr && !tp->srtt)
5358                                         tcp_ack_saw_tstamp(sk, 0);
5359
5360                                 if (tp->rx_opt.tstamp_ok)
5361                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5362
5363                                 /* Make sure socket is routed, for
5364                                  * correct metrics.
5365                                  */
5366                                 icsk->icsk_af_ops->rebuild_header(sk);
5367
5368                                 tcp_init_metrics(sk);
5369
5370                                 tcp_init_congestion_control(sk);
5371
5372                                 /* Prevent spurious tcp_cwnd_restart() on
5373                                  * first data packet.
5374                                  */
5375                                 tp->lsndtime = tcp_time_stamp;
5376
5377                                 tcp_mtup_init(sk);
5378                                 tcp_initialize_rcv_mss(sk);
5379                                 tcp_init_buffer_space(sk);
5380                                 tcp_fast_path_on(tp);
5381                         } else {
5382                                 return 1;
5383                         }
5384                         break;
5385
5386                 case TCP_FIN_WAIT1:
5387                         if (tp->snd_una == tp->write_seq) {
5388                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5389                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5390                                 dst_confirm(sk->sk_dst_cache);
5391
5392                                 if (!sock_flag(sk, SOCK_DEAD))
5393                                         /* Wake up lingering close() */
5394                                         sk->sk_state_change(sk);
5395                                 else {
5396                                         int tmo;
5397
5398                                         if (tp->linger2 < 0 ||
5399                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5400                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5401                                                 tcp_done(sk);
5402                                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5403                                                 return 1;
5404                                         }
5405
5406                                         tmo = tcp_fin_time(sk);
5407                                         if (tmo > TCP_TIMEWAIT_LEN) {
5408                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5409                                         } else if (th->fin || sock_owned_by_user(sk)) {
5410                                                 /* Bad case. We could lose such FIN otherwise.
5411                                                  * It is not a big problem, but it looks confusing
5412                                                  * and not so rare event. We still can lose it now,
5413                                                  * if it spins in bh_lock_sock(), but it is really
5414                                                  * marginal case.
5415                                                  */
5416                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5417                                         } else {
5418                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5419                                                 goto discard;
5420                                         }
5421                                 }
5422                         }
5423                         break;
5424
5425                 case TCP_CLOSING:
5426                         if (tp->snd_una == tp->write_seq) {
5427                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5428                                 goto discard;
5429                         }
5430                         break;
5431
5432                 case TCP_LAST_ACK:
5433                         if (tp->snd_una == tp->write_seq) {
5434                                 tcp_update_metrics(sk);
5435                                 tcp_done(sk);
5436                                 goto discard;
5437                         }
5438                         break;
5439                 }
5440         } else
5441                 goto discard;
5442
5443         /* step 6: check the URG bit */
5444         tcp_urg(sk, skb, th);
5445
5446         /* step 7: process the segment text */
5447         switch (sk->sk_state) {
5448         case TCP_CLOSE_WAIT:
5449         case TCP_CLOSING:
5450         case TCP_LAST_ACK:
5451                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5452                         break;
5453         case TCP_FIN_WAIT1:
5454         case TCP_FIN_WAIT2:
5455                 /* RFC 793 says to queue data in these states,
5456                  * RFC 1122 says we MUST send a reset.
5457                  * BSD 4.4 also does reset.
5458                  */
5459                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5460                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5461                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5462                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5463                                 tcp_reset(sk);
5464                                 return 1;
5465                         }
5466                 }
5467                 /* Fall through */
5468         case TCP_ESTABLISHED:
5469                 tcp_data_queue(sk, skb);
5470                 queued = 1;
5471                 break;
5472         }
5473
5474         /* tcp_data could move socket to TIME-WAIT */
5475         if (sk->sk_state != TCP_CLOSE) {
5476                 tcp_data_snd_check(sk);
5477                 tcp_ack_snd_check(sk);
5478         }
5479
5480         if (!queued) {
5481 discard:
5482                 __kfree_skb(skb);
5483         }
5484         return 0;
5485 }
5486
5487 EXPORT_SYMBOL(sysctl_tcp_ecn);
5488 EXPORT_SYMBOL(sysctl_tcp_reordering);
5489 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5490 EXPORT_SYMBOL(tcp_parse_options);
5491 #ifdef CONFIG_TCP_MD5SIG
5492 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5493 #endif
5494 EXPORT_SYMBOL(tcp_rcv_established);
5495 EXPORT_SYMBOL(tcp_rcv_state_process);
5496 EXPORT_SYMBOL(tcp_initialize_rcv_mss);