Merge branch 'topic/hda' into for-linus
[linux-2.6] / net / ipv4 / tcp_input.c
1 /*
2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
3  *              operating system.  INET is implemented using the  BSD Socket
4  *              interface as the means of communication with the user level.
5  *
6  *              Implementation of the Transmission Control Protocol(TCP).
7  *
8  * 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 <linux/kernel.h>
68 #include <net/dst.h>
69 #include <net/tcp.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
74
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
91
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
94
95 #define FLAG_DATA               0x01 /* Incoming frame contained data.          */
96 #define FLAG_WIN_UPDATE         0x02 /* Incoming ACK was a window update.       */
97 #define FLAG_DATA_ACKED         0x04 /* This ACK acknowledged new data.         */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted.  */
99 #define FLAG_SYN_ACKED          0x10 /* This ACK acknowledged SYN.              */
100 #define FLAG_DATA_SACKED        0x20 /* New SACK.                               */
101 #define FLAG_ECE                0x40 /* ECE in this ACK                         */
102 #define FLAG_DATA_LOST          0x80 /* SACK detected data lossage.             */
103 #define FLAG_SLOWPATH           0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED   0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED   0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK       0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED      0x1000 /* Non-head rexmitted data was ACKed */
108 #define FLAG_SACK_RENEGING      0x2000 /* snd_una advanced to a sacked seq */
109
110 #define FLAG_ACKED              (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
111 #define FLAG_NOT_DUP            (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
112 #define FLAG_CA_ALERT           (FLAG_DATA_SACKED|FLAG_ECE)
113 #define FLAG_FORWARD_PROGRESS   (FLAG_ACKED|FLAG_DATA_SACKED)
114 #define FLAG_ANY_PROGRESS       (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115
116 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
117 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
118
119 /* Adapt the MSS value used to make delayed ack decision to the
120  * real world.
121  */
122 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
123 {
124         struct inet_connection_sock *icsk = inet_csk(sk);
125         const unsigned int lss = icsk->icsk_ack.last_seg_size;
126         unsigned int len;
127
128         icsk->icsk_ack.last_seg_size = 0;
129
130         /* skb->len may jitter because of SACKs, even if peer
131          * sends good full-sized frames.
132          */
133         len = skb_shinfo(skb)->gso_size ? : skb->len;
134         if (len >= icsk->icsk_ack.rcv_mss) {
135                 icsk->icsk_ack.rcv_mss = len;
136         } else {
137                 /* Otherwise, we make more careful check taking into account,
138                  * that SACKs block is variable.
139                  *
140                  * "len" is invariant segment length, including TCP header.
141                  */
142                 len += skb->data - skb_transport_header(skb);
143                 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
144                     /* If PSH is not set, packet should be
145                      * full sized, provided peer TCP is not badly broken.
146                      * This observation (if it is correct 8)) allows
147                      * to handle super-low mtu links fairly.
148                      */
149                     (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
150                      !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
151                         /* Subtract also invariant (if peer is RFC compliant),
152                          * tcp header plus fixed timestamp option length.
153                          * Resulting "len" is MSS free of SACK jitter.
154                          */
155                         len -= tcp_sk(sk)->tcp_header_len;
156                         icsk->icsk_ack.last_seg_size = len;
157                         if (len == lss) {
158                                 icsk->icsk_ack.rcv_mss = len;
159                                 return;
160                         }
161                 }
162                 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
163                         icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
164                 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
165         }
166 }
167
168 static void tcp_incr_quickack(struct sock *sk)
169 {
170         struct inet_connection_sock *icsk = inet_csk(sk);
171         unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
172
173         if (quickacks == 0)
174                 quickacks = 2;
175         if (quickacks > icsk->icsk_ack.quick)
176                 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
177 }
178
179 void tcp_enter_quickack_mode(struct sock *sk)
180 {
181         struct inet_connection_sock *icsk = inet_csk(sk);
182         tcp_incr_quickack(sk);
183         icsk->icsk_ack.pingpong = 0;
184         icsk->icsk_ack.ato = TCP_ATO_MIN;
185 }
186
187 /* Send ACKs quickly, if "quick" count is not exhausted
188  * and the session is not interactive.
189  */
190
191 static inline int tcp_in_quickack_mode(const struct sock *sk)
192 {
193         const struct inet_connection_sock *icsk = inet_csk(sk);
194         return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
195 }
196
197 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
198 {
199         if (tp->ecn_flags & TCP_ECN_OK)
200                 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
201 }
202
203 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
204 {
205         if (tcp_hdr(skb)->cwr)
206                 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
207 }
208
209 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
210 {
211         tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
212 }
213
214 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
215 {
216         if (tp->ecn_flags & TCP_ECN_OK) {
217                 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
218                         tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
219                 /* Funny extension: if ECT is not set on a segment,
220                  * it is surely retransmit. It is not in ECN RFC,
221                  * but Linux follows this rule. */
222                 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
223                         tcp_enter_quickack_mode((struct sock *)tp);
224         }
225 }
226
227 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
228 {
229         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
230                 tp->ecn_flags &= ~TCP_ECN_OK;
231 }
232
233 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
234 {
235         if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
236                 tp->ecn_flags &= ~TCP_ECN_OK;
237 }
238
239 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
240 {
241         if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
242                 return 1;
243         return 0;
244 }
245
246 /* Buffer size and advertised window tuning.
247  *
248  * 1. Tuning sk->sk_sndbuf, when connection enters established state.
249  */
250
251 static void tcp_fixup_sndbuf(struct sock *sk)
252 {
253         int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
254                      sizeof(struct sk_buff);
255
256         if (sk->sk_sndbuf < 3 * sndmem)
257                 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
258 }
259
260 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261  *
262  * All tcp_full_space() is split to two parts: "network" buffer, allocated
263  * forward and advertised in receiver window (tp->rcv_wnd) and
264  * "application buffer", required to isolate scheduling/application
265  * latencies from network.
266  * window_clamp is maximal advertised window. It can be less than
267  * tcp_full_space(), in this case tcp_full_space() - window_clamp
268  * is reserved for "application" buffer. The less window_clamp is
269  * the smoother our behaviour from viewpoint of network, but the lower
270  * throughput and the higher sensitivity of the connection to losses. 8)
271  *
272  * rcv_ssthresh is more strict window_clamp used at "slow start"
273  * phase to predict further behaviour of this connection.
274  * It is used for two goals:
275  * - to enforce header prediction at sender, even when application
276  *   requires some significant "application buffer". It is check #1.
277  * - to prevent pruning of receive queue because of misprediction
278  *   of receiver window. Check #2.
279  *
280  * The scheme does not work when sender sends good segments opening
281  * window and then starts to feed us spaghetti. But it should work
282  * in common situations. Otherwise, we have to rely on queue collapsing.
283  */
284
285 /* Slow part of check#2. */
286 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
287 {
288         struct tcp_sock *tp = tcp_sk(sk);
289         /* Optimize this! */
290         int truesize = tcp_win_from_space(skb->truesize) >> 1;
291         int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
292
293         while (tp->rcv_ssthresh <= window) {
294                 if (truesize <= skb->len)
295                         return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
296
297                 truesize >>= 1;
298                 window >>= 1;
299         }
300         return 0;
301 }
302
303 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
304 {
305         struct tcp_sock *tp = tcp_sk(sk);
306
307         /* Check #1 */
308         if (tp->rcv_ssthresh < tp->window_clamp &&
309             (int)tp->rcv_ssthresh < tcp_space(sk) &&
310             !tcp_memory_pressure) {
311                 int incr;
312
313                 /* Check #2. Increase window, if skb with such overhead
314                  * will fit to rcvbuf in future.
315                  */
316                 if (tcp_win_from_space(skb->truesize) <= skb->len)
317                         incr = 2 * tp->advmss;
318                 else
319                         incr = __tcp_grow_window(sk, skb);
320
321                 if (incr) {
322                         tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
323                                                tp->window_clamp);
324                         inet_csk(sk)->icsk_ack.quick |= 1;
325                 }
326         }
327 }
328
329 /* 3. Tuning rcvbuf, when connection enters established state. */
330
331 static void tcp_fixup_rcvbuf(struct sock *sk)
332 {
333         struct tcp_sock *tp = tcp_sk(sk);
334         int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
335
336         /* Try to select rcvbuf so that 4 mss-sized segments
337          * will fit to window and corresponding skbs will fit to our rcvbuf.
338          * (was 3; 4 is minimum to allow fast retransmit to work.)
339          */
340         while (tcp_win_from_space(rcvmem) < tp->advmss)
341                 rcvmem += 128;
342         if (sk->sk_rcvbuf < 4 * rcvmem)
343                 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
344 }
345
346 /* 4. Try to fixup all. It is made immediately after connection enters
347  *    established state.
348  */
349 static void tcp_init_buffer_space(struct sock *sk)
350 {
351         struct tcp_sock *tp = tcp_sk(sk);
352         int maxwin;
353
354         if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
355                 tcp_fixup_rcvbuf(sk);
356         if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
357                 tcp_fixup_sndbuf(sk);
358
359         tp->rcvq_space.space = tp->rcv_wnd;
360
361         maxwin = tcp_full_space(sk);
362
363         if (tp->window_clamp >= maxwin) {
364                 tp->window_clamp = maxwin;
365
366                 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
367                         tp->window_clamp = max(maxwin -
368                                                (maxwin >> sysctl_tcp_app_win),
369                                                4 * tp->advmss);
370         }
371
372         /* Force reservation of one segment. */
373         if (sysctl_tcp_app_win &&
374             tp->window_clamp > 2 * tp->advmss &&
375             tp->window_clamp + tp->advmss > maxwin)
376                 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
377
378         tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
379         tp->snd_cwnd_stamp = tcp_time_stamp;
380 }
381
382 /* 5. Recalculate window clamp after socket hit its memory bounds. */
383 static void tcp_clamp_window(struct sock *sk)
384 {
385         struct tcp_sock *tp = tcp_sk(sk);
386         struct inet_connection_sock *icsk = inet_csk(sk);
387
388         icsk->icsk_ack.quick = 0;
389
390         if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
391             !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
392             !tcp_memory_pressure &&
393             atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
394                 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
395                                     sysctl_tcp_rmem[2]);
396         }
397         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
398                 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
399 }
400
401 /* Initialize RCV_MSS value.
402  * RCV_MSS is an our guess about MSS used by the peer.
403  * We haven't any direct information about the MSS.
404  * It's better to underestimate the RCV_MSS rather than overestimate.
405  * Overestimations make us ACKing less frequently than needed.
406  * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407  */
408 void tcp_initialize_rcv_mss(struct sock *sk)
409 {
410         struct tcp_sock *tp = tcp_sk(sk);
411         unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
412
413         hint = min(hint, tp->rcv_wnd / 2);
414         hint = min(hint, TCP_MIN_RCVMSS);
415         hint = max(hint, TCP_MIN_MSS);
416
417         inet_csk(sk)->icsk_ack.rcv_mss = hint;
418 }
419
420 /* Receiver "autotuning" code.
421  *
422  * The algorithm for RTT estimation w/o timestamps is based on
423  * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
424  * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425  *
426  * More detail on this code can be found at
427  * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
428  * though this reference is out of date.  A new paper
429  * is pending.
430  */
431 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
432 {
433         u32 new_sample = tp->rcv_rtt_est.rtt;
434         long m = sample;
435
436         if (m == 0)
437                 m = 1;
438
439         if (new_sample != 0) {
440                 /* If we sample in larger samples in the non-timestamp
441                  * case, we could grossly overestimate the RTT especially
442                  * with chatty applications or bulk transfer apps which
443                  * are stalled on filesystem I/O.
444                  *
445                  * Also, since we are only going for a minimum in the
446                  * non-timestamp case, we do not smooth things out
447                  * else with timestamps disabled convergence takes too
448                  * long.
449                  */
450                 if (!win_dep) {
451                         m -= (new_sample >> 3);
452                         new_sample += m;
453                 } else if (m < new_sample)
454                         new_sample = m << 3;
455         } else {
456                 /* No previous measure. */
457                 new_sample = m << 3;
458         }
459
460         if (tp->rcv_rtt_est.rtt != new_sample)
461                 tp->rcv_rtt_est.rtt = new_sample;
462 }
463
464 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
465 {
466         if (tp->rcv_rtt_est.time == 0)
467                 goto new_measure;
468         if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
469                 return;
470         tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
471
472 new_measure:
473         tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
474         tp->rcv_rtt_est.time = tcp_time_stamp;
475 }
476
477 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
478                                           const struct sk_buff *skb)
479 {
480         struct tcp_sock *tp = tcp_sk(sk);
481         if (tp->rx_opt.rcv_tsecr &&
482             (TCP_SKB_CB(skb)->end_seq -
483              TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
484                 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
485 }
486
487 /*
488  * This function should be called every time data is copied to user space.
489  * It calculates the appropriate TCP receive buffer space.
490  */
491 void tcp_rcv_space_adjust(struct sock *sk)
492 {
493         struct tcp_sock *tp = tcp_sk(sk);
494         int time;
495         int space;
496
497         if (tp->rcvq_space.time == 0)
498                 goto new_measure;
499
500         time = tcp_time_stamp - tp->rcvq_space.time;
501         if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
502                 return;
503
504         space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
505
506         space = max(tp->rcvq_space.space, space);
507
508         if (tp->rcvq_space.space != space) {
509                 int rcvmem;
510
511                 tp->rcvq_space.space = space;
512
513                 if (sysctl_tcp_moderate_rcvbuf &&
514                     !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
515                         int new_clamp = space;
516
517                         /* Receive space grows, normalize in order to
518                          * take into account packet headers and sk_buff
519                          * structure overhead.
520                          */
521                         space /= tp->advmss;
522                         if (!space)
523                                 space = 1;
524                         rcvmem = (tp->advmss + MAX_TCP_HEADER +
525                                   16 + sizeof(struct sk_buff));
526                         while (tcp_win_from_space(rcvmem) < tp->advmss)
527                                 rcvmem += 128;
528                         space *= rcvmem;
529                         space = min(space, sysctl_tcp_rmem[2]);
530                         if (space > sk->sk_rcvbuf) {
531                                 sk->sk_rcvbuf = space;
532
533                                 /* Make the window clamp follow along.  */
534                                 tp->window_clamp = new_clamp;
535                         }
536                 }
537         }
538
539 new_measure:
540         tp->rcvq_space.seq = tp->copied_seq;
541         tp->rcvq_space.time = tcp_time_stamp;
542 }
543
544 /* There is something which you must keep in mind when you analyze the
545  * behavior of the tp->ato delayed ack timeout interval.  When a
546  * connection starts up, we want to ack as quickly as possible.  The
547  * problem is that "good" TCP's do slow start at the beginning of data
548  * transmission.  The means that until we send the first few ACK's the
549  * sender will sit on his end and only queue most of his data, because
550  * he can only send snd_cwnd unacked packets at any given time.  For
551  * each ACK we send, he increments snd_cwnd and transmits more of his
552  * queue.  -DaveM
553  */
554 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
555 {
556         struct tcp_sock *tp = tcp_sk(sk);
557         struct inet_connection_sock *icsk = inet_csk(sk);
558         u32 now;
559
560         inet_csk_schedule_ack(sk);
561
562         tcp_measure_rcv_mss(sk, skb);
563
564         tcp_rcv_rtt_measure(tp);
565
566         now = tcp_time_stamp;
567
568         if (!icsk->icsk_ack.ato) {
569                 /* The _first_ data packet received, initialize
570                  * delayed ACK engine.
571                  */
572                 tcp_incr_quickack(sk);
573                 icsk->icsk_ack.ato = TCP_ATO_MIN;
574         } else {
575                 int m = now - icsk->icsk_ack.lrcvtime;
576
577                 if (m <= TCP_ATO_MIN / 2) {
578                         /* The fastest case is the first. */
579                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
580                 } else if (m < icsk->icsk_ack.ato) {
581                         icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
582                         if (icsk->icsk_ack.ato > icsk->icsk_rto)
583                                 icsk->icsk_ack.ato = icsk->icsk_rto;
584                 } else if (m > icsk->icsk_rto) {
585                         /* Too long gap. Apparently sender failed to
586                          * restart window, so that we send ACKs quickly.
587                          */
588                         tcp_incr_quickack(sk);
589                         sk_mem_reclaim(sk);
590                 }
591         }
592         icsk->icsk_ack.lrcvtime = now;
593
594         TCP_ECN_check_ce(tp, skb);
595
596         if (skb->len >= 128)
597                 tcp_grow_window(sk, skb);
598 }
599
600 /* Called to compute a smoothed rtt estimate. The data fed to this
601  * routine either comes from timestamps, or from segments that were
602  * known _not_ to have been retransmitted [see Karn/Partridge
603  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
604  * piece by Van Jacobson.
605  * NOTE: the next three routines used to be one big routine.
606  * To save cycles in the RFC 1323 implementation it was better to break
607  * it up into three procedures. -- erics
608  */
609 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
610 {
611         struct tcp_sock *tp = tcp_sk(sk);
612         long m = mrtt; /* RTT */
613
614         /*      The following amusing code comes from Jacobson's
615          *      article in SIGCOMM '88.  Note that rtt and mdev
616          *      are scaled versions of rtt and mean deviation.
617          *      This is designed to be as fast as possible
618          *      m stands for "measurement".
619          *
620          *      On a 1990 paper the rto value is changed to:
621          *      RTO = rtt + 4 * mdev
622          *
623          * Funny. This algorithm seems to be very broken.
624          * These formulae increase RTO, when it should be decreased, increase
625          * too slowly, when it should be increased quickly, decrease too quickly
626          * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
627          * does not matter how to _calculate_ it. Seems, it was trap
628          * that VJ failed to avoid. 8)
629          */
630         if (m == 0)
631                 m = 1;
632         if (tp->srtt != 0) {
633                 m -= (tp->srtt >> 3);   /* m is now error in rtt est */
634                 tp->srtt += m;          /* rtt = 7/8 rtt + 1/8 new */
635                 if (m < 0) {
636                         m = -m;         /* m is now abs(error) */
637                         m -= (tp->mdev >> 2);   /* similar update on mdev */
638                         /* This is similar to one of Eifel findings.
639                          * Eifel blocks mdev updates when rtt decreases.
640                          * This solution is a bit different: we use finer gain
641                          * for mdev in this case (alpha*beta).
642                          * Like Eifel it also prevents growth of rto,
643                          * but also it limits too fast rto decreases,
644                          * happening in pure Eifel.
645                          */
646                         if (m > 0)
647                                 m >>= 3;
648                 } else {
649                         m -= (tp->mdev >> 2);   /* similar update on mdev */
650                 }
651                 tp->mdev += m;          /* mdev = 3/4 mdev + 1/4 new */
652                 if (tp->mdev > tp->mdev_max) {
653                         tp->mdev_max = tp->mdev;
654                         if (tp->mdev_max > tp->rttvar)
655                                 tp->rttvar = tp->mdev_max;
656                 }
657                 if (after(tp->snd_una, tp->rtt_seq)) {
658                         if (tp->mdev_max < tp->rttvar)
659                                 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
660                         tp->rtt_seq = tp->snd_nxt;
661                         tp->mdev_max = tcp_rto_min(sk);
662                 }
663         } else {
664                 /* no previous measure. */
665                 tp->srtt = m << 3;      /* take the measured time to be rtt */
666                 tp->mdev = m << 1;      /* make sure rto = 3*rtt */
667                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
668                 tp->rtt_seq = tp->snd_nxt;
669         }
670 }
671
672 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
673  * routine referred to above.
674  */
675 static inline void tcp_set_rto(struct sock *sk)
676 {
677         const struct tcp_sock *tp = tcp_sk(sk);
678         /* Old crap is replaced with new one. 8)
679          *
680          * More seriously:
681          * 1. If rtt variance happened to be less 50msec, it is hallucination.
682          *    It cannot be less due to utterly erratic ACK generation made
683          *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
684          *    to do with delayed acks, because at cwnd>2 true delack timeout
685          *    is invisible. Actually, Linux-2.4 also generates erratic
686          *    ACKs in some circumstances.
687          */
688         inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
689
690         /* 2. Fixups made earlier cannot be right.
691          *    If we do not estimate RTO correctly without them,
692          *    all the algo is pure shit and should be replaced
693          *    with correct one. It is exactly, which we pretend to do.
694          */
695
696         /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
697          * guarantees that rto is higher.
698          */
699         if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
700                 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
701 }
702
703 /* Save metrics learned by this TCP session.
704    This function is called only, when TCP finishes successfully
705    i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
706  */
707 void tcp_update_metrics(struct sock *sk)
708 {
709         struct tcp_sock *tp = tcp_sk(sk);
710         struct dst_entry *dst = __sk_dst_get(sk);
711
712         if (sysctl_tcp_nometrics_save)
713                 return;
714
715         dst_confirm(dst);
716
717         if (dst && (dst->flags & DST_HOST)) {
718                 const struct inet_connection_sock *icsk = inet_csk(sk);
719                 int m;
720                 unsigned long rtt;
721
722                 if (icsk->icsk_backoff || !tp->srtt) {
723                         /* This session failed to estimate rtt. Why?
724                          * Probably, no packets returned in time.
725                          * Reset our results.
726                          */
727                         if (!(dst_metric_locked(dst, RTAX_RTT)))
728                                 dst->metrics[RTAX_RTT - 1] = 0;
729                         return;
730                 }
731
732                 rtt = dst_metric_rtt(dst, RTAX_RTT);
733                 m = rtt - tp->srtt;
734
735                 /* If newly calculated rtt larger than stored one,
736                  * store new one. Otherwise, use EWMA. Remember,
737                  * rtt overestimation is always better than underestimation.
738                  */
739                 if (!(dst_metric_locked(dst, RTAX_RTT))) {
740                         if (m <= 0)
741                                 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
742                         else
743                                 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
744                 }
745
746                 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
747                         unsigned long var;
748                         if (m < 0)
749                                 m = -m;
750
751                         /* Scale deviation to rttvar fixed point */
752                         m >>= 1;
753                         if (m < tp->mdev)
754                                 m = tp->mdev;
755
756                         var = dst_metric_rtt(dst, RTAX_RTTVAR);
757                         if (m >= var)
758                                 var = m;
759                         else
760                                 var -= (var - m) >> 2;
761
762                         set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
763                 }
764
765                 if (tp->snd_ssthresh >= 0xFFFF) {
766                         /* Slow start still did not finish. */
767                         if (dst_metric(dst, RTAX_SSTHRESH) &&
768                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
769                             (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
770                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
771                         if (!dst_metric_locked(dst, RTAX_CWND) &&
772                             tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
773                                 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
774                 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
775                            icsk->icsk_ca_state == TCP_CA_Open) {
776                         /* Cong. avoidance phase, cwnd is reliable. */
777                         if (!dst_metric_locked(dst, RTAX_SSTHRESH))
778                                 dst->metrics[RTAX_SSTHRESH-1] =
779                                         max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
780                         if (!dst_metric_locked(dst, RTAX_CWND))
781                                 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
782                 } else {
783                         /* Else slow start did not finish, cwnd is non-sense,
784                            ssthresh may be also invalid.
785                          */
786                         if (!dst_metric_locked(dst, RTAX_CWND))
787                                 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
788                         if (dst_metric(dst, RTAX_SSTHRESH) &&
789                             !dst_metric_locked(dst, RTAX_SSTHRESH) &&
790                             tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
791                                 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
792                 }
793
794                 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
795                         if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
796                             tp->reordering != sysctl_tcp_reordering)
797                                 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
798                 }
799         }
800 }
801
802 /* Numbers are taken from RFC3390.
803  *
804  * John Heffner states:
805  *
806  *      The RFC specifies a window of no more than 4380 bytes
807  *      unless 2*MSS > 4380.  Reading the pseudocode in the RFC
808  *      is a bit misleading because they use a clamp at 4380 bytes
809  *      rather than use a multiplier in the relevant range.
810  */
811 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
812 {
813         __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
814
815         if (!cwnd) {
816                 if (tp->mss_cache > 1460)
817                         cwnd = 2;
818                 else
819                         cwnd = (tp->mss_cache > 1095) ? 3 : 4;
820         }
821         return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
822 }
823
824 /* Set slow start threshold and cwnd not falling to slow start */
825 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
826 {
827         struct tcp_sock *tp = tcp_sk(sk);
828         const struct inet_connection_sock *icsk = inet_csk(sk);
829
830         tp->prior_ssthresh = 0;
831         tp->bytes_acked = 0;
832         if (icsk->icsk_ca_state < TCP_CA_CWR) {
833                 tp->undo_marker = 0;
834                 if (set_ssthresh)
835                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
836                 tp->snd_cwnd = min(tp->snd_cwnd,
837                                    tcp_packets_in_flight(tp) + 1U);
838                 tp->snd_cwnd_cnt = 0;
839                 tp->high_seq = tp->snd_nxt;
840                 tp->snd_cwnd_stamp = tcp_time_stamp;
841                 TCP_ECN_queue_cwr(tp);
842
843                 tcp_set_ca_state(sk, TCP_CA_CWR);
844         }
845 }
846
847 /*
848  * Packet counting of FACK is based on in-order assumptions, therefore TCP
849  * disables it when reordering is detected
850  */
851 static void tcp_disable_fack(struct tcp_sock *tp)
852 {
853         /* RFC3517 uses different metric in lost marker => reset on change */
854         if (tcp_is_fack(tp))
855                 tp->lost_skb_hint = NULL;
856         tp->rx_opt.sack_ok &= ~2;
857 }
858
859 /* Take a notice that peer is sending D-SACKs */
860 static void tcp_dsack_seen(struct tcp_sock *tp)
861 {
862         tp->rx_opt.sack_ok |= 4;
863 }
864
865 /* Initialize metrics on socket. */
866
867 static void tcp_init_metrics(struct sock *sk)
868 {
869         struct tcp_sock *tp = tcp_sk(sk);
870         struct dst_entry *dst = __sk_dst_get(sk);
871
872         if (dst == NULL)
873                 goto reset;
874
875         dst_confirm(dst);
876
877         if (dst_metric_locked(dst, RTAX_CWND))
878                 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
879         if (dst_metric(dst, RTAX_SSTHRESH)) {
880                 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
881                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
882                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
883         }
884         if (dst_metric(dst, RTAX_REORDERING) &&
885             tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
886                 tcp_disable_fack(tp);
887                 tp->reordering = dst_metric(dst, RTAX_REORDERING);
888         }
889
890         if (dst_metric(dst, RTAX_RTT) == 0)
891                 goto reset;
892
893         if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
894                 goto reset;
895
896         /* Initial rtt is determined from SYN,SYN-ACK.
897          * The segment is small and rtt may appear much
898          * less than real one. Use per-dst memory
899          * to make it more realistic.
900          *
901          * A bit of theory. RTT is time passed after "normal" sized packet
902          * is sent until it is ACKed. In normal circumstances sending small
903          * packets force peer to delay ACKs and calculation is correct too.
904          * The algorithm is adaptive and, provided we follow specs, it
905          * NEVER underestimate RTT. BUT! If peer tries to make some clever
906          * tricks sort of "quick acks" for time long enough to decrease RTT
907          * to low value, and then abruptly stops to do it and starts to delay
908          * ACKs, wait for troubles.
909          */
910         if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
911                 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
912                 tp->rtt_seq = tp->snd_nxt;
913         }
914         if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
915                 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
916                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
917         }
918         tcp_set_rto(sk);
919         if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
920                 goto reset;
921
922 cwnd:
923         tp->snd_cwnd = tcp_init_cwnd(tp, dst);
924         tp->snd_cwnd_stamp = tcp_time_stamp;
925         return;
926
927 reset:
928         /* Play conservative. If timestamps are not
929          * supported, TCP will fail to recalculate correct
930          * rtt, if initial rto is too small. FORGET ALL AND RESET!
931          */
932         if (!tp->rx_opt.saw_tstamp && tp->srtt) {
933                 tp->srtt = 0;
934                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
935                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
936         }
937         goto cwnd;
938 }
939
940 static void tcp_update_reordering(struct sock *sk, const int metric,
941                                   const int ts)
942 {
943         struct tcp_sock *tp = tcp_sk(sk);
944         if (metric > tp->reordering) {
945                 int mib_idx;
946
947                 tp->reordering = min(TCP_MAX_REORDERING, metric);
948
949                 /* This exciting event is worth to be remembered. 8) */
950                 if (ts)
951                         mib_idx = LINUX_MIB_TCPTSREORDER;
952                 else if (tcp_is_reno(tp))
953                         mib_idx = LINUX_MIB_TCPRENOREORDER;
954                 else if (tcp_is_fack(tp))
955                         mib_idx = LINUX_MIB_TCPFACKREORDER;
956                 else
957                         mib_idx = LINUX_MIB_TCPSACKREORDER;
958
959                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
960 #if FASTRETRANS_DEBUG > 1
961                 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
962                        tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
963                        tp->reordering,
964                        tp->fackets_out,
965                        tp->sacked_out,
966                        tp->undo_marker ? tp->undo_retrans : 0);
967 #endif
968                 tcp_disable_fack(tp);
969         }
970 }
971
972 /* This must be called before lost_out is incremented */
973 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
974 {
975         if ((tp->retransmit_skb_hint == NULL) ||
976             before(TCP_SKB_CB(skb)->seq,
977                    TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
978                 tp->retransmit_skb_hint = skb;
979
980         if (!tp->lost_out ||
981             after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
982                 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
983 }
984
985 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
986 {
987         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
988                 tcp_verify_retransmit_hint(tp, skb);
989
990                 tp->lost_out += tcp_skb_pcount(skb);
991                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
992         }
993 }
994
995 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
996                                             struct sk_buff *skb)
997 {
998         tcp_verify_retransmit_hint(tp, skb);
999
1000         if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1001                 tp->lost_out += tcp_skb_pcount(skb);
1002                 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1003         }
1004 }
1005
1006 /* This procedure tags the retransmission queue when SACKs arrive.
1007  *
1008  * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1009  * Packets in queue with these bits set are counted in variables
1010  * sacked_out, retrans_out and lost_out, correspondingly.
1011  *
1012  * Valid combinations are:
1013  * Tag  InFlight        Description
1014  * 0    1               - orig segment is in flight.
1015  * S    0               - nothing flies, orig reached receiver.
1016  * L    0               - nothing flies, orig lost by net.
1017  * R    2               - both orig and retransmit are in flight.
1018  * L|R  1               - orig is lost, retransmit is in flight.
1019  * S|R  1               - orig reached receiver, retrans is still in flight.
1020  * (L|S|R is logically valid, it could occur when L|R is sacked,
1021  *  but it is equivalent to plain S and code short-curcuits it to S.
1022  *  L|S is logically invalid, it would mean -1 packet in flight 8))
1023  *
1024  * These 6 states form finite state machine, controlled by the following events:
1025  * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1026  * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1027  * 3. Loss detection event of one of three flavors:
1028  *      A. Scoreboard estimator decided the packet is lost.
1029  *         A'. Reno "three dupacks" marks head of queue lost.
1030  *         A''. Its FACK modfication, head until snd.fack is lost.
1031  *      B. SACK arrives sacking data transmitted after never retransmitted
1032  *         hole was sent out.
1033  *      C. SACK arrives sacking SND.NXT at the moment, when the
1034  *         segment was retransmitted.
1035  * 4. D-SACK added new rule: D-SACK changes any tag to S.
1036  *
1037  * It is pleasant to note, that state diagram turns out to be commutative,
1038  * so that we are allowed not to be bothered by order of our actions,
1039  * when multiple events arrive simultaneously. (see the function below).
1040  *
1041  * Reordering detection.
1042  * --------------------
1043  * Reordering metric is maximal distance, which a packet can be displaced
1044  * in packet stream. With SACKs we can estimate it:
1045  *
1046  * 1. SACK fills old hole and the corresponding segment was not
1047  *    ever retransmitted -> reordering. Alas, we cannot use it
1048  *    when segment was retransmitted.
1049  * 2. The last flaw is solved with D-SACK. D-SACK arrives
1050  *    for retransmitted and already SACKed segment -> reordering..
1051  * Both of these heuristics are not used in Loss state, when we cannot
1052  * account for retransmits accurately.
1053  *
1054  * SACK block validation.
1055  * ----------------------
1056  *
1057  * SACK block range validation checks that the received SACK block fits to
1058  * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1059  * Note that SND.UNA is not included to the range though being valid because
1060  * it means that the receiver is rather inconsistent with itself reporting
1061  * SACK reneging when it should advance SND.UNA. Such SACK block this is
1062  * perfectly valid, however, in light of RFC2018 which explicitly states
1063  * that "SACK block MUST reflect the newest segment.  Even if the newest
1064  * segment is going to be discarded ...", not that it looks very clever
1065  * in case of head skb. Due to potentional receiver driven attacks, we
1066  * choose to avoid immediate execution of a walk in write queue due to
1067  * reneging and defer head skb's loss recovery to standard loss recovery
1068  * procedure that will eventually trigger (nothing forbids us doing this).
1069  *
1070  * Implements also blockage to start_seq wrap-around. Problem lies in the
1071  * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1072  * there's no guarantee that it will be before snd_nxt (n). The problem
1073  * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1074  * wrap (s_w):
1075  *
1076  *         <- outs wnd ->                          <- wrapzone ->
1077  *         u     e      n                         u_w   e_w  s n_w
1078  *         |     |      |                          |     |   |  |
1079  * |<------------+------+----- TCP seqno space --------------+---------->|
1080  * ...-- <2^31 ->|                                           |<--------...
1081  * ...---- >2^31 ------>|                                    |<--------...
1082  *
1083  * Current code wouldn't be vulnerable but it's better still to discard such
1084  * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1085  * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1086  * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1087  * equal to the ideal case (infinite seqno space without wrap caused issues).
1088  *
1089  * With D-SACK the lower bound is extended to cover sequence space below
1090  * SND.UNA down to undo_marker, which is the last point of interest. Yet
1091  * again, D-SACK block must not to go across snd_una (for the same reason as
1092  * for the normal SACK blocks, explained above). But there all simplicity
1093  * ends, TCP might receive valid D-SACKs below that. As long as they reside
1094  * fully below undo_marker they do not affect behavior in anyway and can
1095  * therefore be safely ignored. In rare cases (which are more or less
1096  * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1097  * fragmentation and packet reordering past skb's retransmission. To consider
1098  * them correctly, the acceptable range must be extended even more though
1099  * the exact amount is rather hard to quantify. However, tp->max_window can
1100  * be used as an exaggerated estimate.
1101  */
1102 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1103                                   u32 start_seq, u32 end_seq)
1104 {
1105         /* Too far in future, or reversed (interpretation is ambiguous) */
1106         if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1107                 return 0;
1108
1109         /* Nasty start_seq wrap-around check (see comments above) */
1110         if (!before(start_seq, tp->snd_nxt))
1111                 return 0;
1112
1113         /* In outstanding window? ...This is valid exit for D-SACKs too.
1114          * start_seq == snd_una is non-sensical (see comments above)
1115          */
1116         if (after(start_seq, tp->snd_una))
1117                 return 1;
1118
1119         if (!is_dsack || !tp->undo_marker)
1120                 return 0;
1121
1122         /* ...Then it's D-SACK, and must reside below snd_una completely */
1123         if (!after(end_seq, tp->snd_una))
1124                 return 0;
1125
1126         if (!before(start_seq, tp->undo_marker))
1127                 return 1;
1128
1129         /* Too old */
1130         if (!after(end_seq, tp->undo_marker))
1131                 return 0;
1132
1133         /* Undo_marker boundary crossing (overestimates a lot). Known already:
1134          *   start_seq < undo_marker and end_seq >= undo_marker.
1135          */
1136         return !before(start_seq, end_seq - tp->max_window);
1137 }
1138
1139 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1140  * Event "C". Later note: FACK people cheated me again 8), we have to account
1141  * for reordering! Ugly, but should help.
1142  *
1143  * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1144  * less than what is now known to be received by the other end (derived from
1145  * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1146  * retransmitted skbs to avoid some costly processing per ACKs.
1147  */
1148 static void tcp_mark_lost_retrans(struct sock *sk)
1149 {
1150         const struct inet_connection_sock *icsk = inet_csk(sk);
1151         struct tcp_sock *tp = tcp_sk(sk);
1152         struct sk_buff *skb;
1153         int cnt = 0;
1154         u32 new_low_seq = tp->snd_nxt;
1155         u32 received_upto = tcp_highest_sack_seq(tp);
1156
1157         if (!tcp_is_fack(tp) || !tp->retrans_out ||
1158             !after(received_upto, tp->lost_retrans_low) ||
1159             icsk->icsk_ca_state != TCP_CA_Recovery)
1160                 return;
1161
1162         tcp_for_write_queue(skb, sk) {
1163                 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1164
1165                 if (skb == tcp_send_head(sk))
1166                         break;
1167                 if (cnt == tp->retrans_out)
1168                         break;
1169                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1170                         continue;
1171
1172                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1173                         continue;
1174
1175                 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1176                  * constraint here (see above) but figuring out that at
1177                  * least tp->reordering SACK blocks reside between ack_seq
1178                  * and received_upto is not easy task to do cheaply with
1179                  * the available datastructures.
1180                  *
1181                  * Whether FACK should check here for tp->reordering segs
1182                  * in-between one could argue for either way (it would be
1183                  * rather simple to implement as we could count fack_count
1184                  * during the walk and do tp->fackets_out - fack_count).
1185                  */
1186                 if (after(received_upto, ack_seq)) {
1187                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1188                         tp->retrans_out -= tcp_skb_pcount(skb);
1189
1190                         tcp_skb_mark_lost_uncond_verify(tp, skb);
1191                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1192                 } else {
1193                         if (before(ack_seq, new_low_seq))
1194                                 new_low_seq = ack_seq;
1195                         cnt += tcp_skb_pcount(skb);
1196                 }
1197         }
1198
1199         if (tp->retrans_out)
1200                 tp->lost_retrans_low = new_low_seq;
1201 }
1202
1203 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1204                            struct tcp_sack_block_wire *sp, int num_sacks,
1205                            u32 prior_snd_una)
1206 {
1207         struct tcp_sock *tp = tcp_sk(sk);
1208         u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1209         u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1210         int dup_sack = 0;
1211
1212         if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1213                 dup_sack = 1;
1214                 tcp_dsack_seen(tp);
1215                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1216         } else if (num_sacks > 1) {
1217                 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1218                 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1219
1220                 if (!after(end_seq_0, end_seq_1) &&
1221                     !before(start_seq_0, start_seq_1)) {
1222                         dup_sack = 1;
1223                         tcp_dsack_seen(tp);
1224                         NET_INC_STATS_BH(sock_net(sk),
1225                                         LINUX_MIB_TCPDSACKOFORECV);
1226                 }
1227         }
1228
1229         /* D-SACK for already forgotten data... Do dumb counting. */
1230         if (dup_sack &&
1231             !after(end_seq_0, prior_snd_una) &&
1232             after(end_seq_0, tp->undo_marker))
1233                 tp->undo_retrans--;
1234
1235         return dup_sack;
1236 }
1237
1238 struct tcp_sacktag_state {
1239         int reord;
1240         int fack_count;
1241         int flag;
1242 };
1243
1244 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1245  * the incoming SACK may not exactly match but we can find smaller MSS
1246  * aligned portion of it that matches. Therefore we might need to fragment
1247  * which may fail and creates some hassle (caller must handle error case
1248  * returns).
1249  *
1250  * FIXME: this could be merged to shift decision code
1251  */
1252 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1253                                  u32 start_seq, u32 end_seq)
1254 {
1255         int in_sack, err;
1256         unsigned int pkt_len;
1257         unsigned int mss;
1258
1259         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1260                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1261
1262         if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1263             after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1264                 mss = tcp_skb_mss(skb);
1265                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1266
1267                 if (!in_sack) {
1268                         pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1269                         if (pkt_len < mss)
1270                                 pkt_len = mss;
1271                 } else {
1272                         pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1273                         if (pkt_len < mss)
1274                                 return -EINVAL;
1275                 }
1276
1277                 /* Round if necessary so that SACKs cover only full MSSes
1278                  * and/or the remaining small portion (if present)
1279                  */
1280                 if (pkt_len > mss) {
1281                         unsigned int new_len = (pkt_len / mss) * mss;
1282                         if (!in_sack && new_len < pkt_len) {
1283                                 new_len += mss;
1284                                 if (new_len > skb->len)
1285                                         return 0;
1286                         }
1287                         pkt_len = new_len;
1288                 }
1289                 err = tcp_fragment(sk, skb, pkt_len, mss);
1290                 if (err < 0)
1291                         return err;
1292         }
1293
1294         return in_sack;
1295 }
1296
1297 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1298                           struct tcp_sacktag_state *state,
1299                           int dup_sack, int pcount)
1300 {
1301         struct tcp_sock *tp = tcp_sk(sk);
1302         u8 sacked = TCP_SKB_CB(skb)->sacked;
1303         int fack_count = state->fack_count;
1304
1305         /* Account D-SACK for retransmitted packet. */
1306         if (dup_sack && (sacked & TCPCB_RETRANS)) {
1307                 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1308                         tp->undo_retrans--;
1309                 if (sacked & TCPCB_SACKED_ACKED)
1310                         state->reord = min(fack_count, state->reord);
1311         }
1312
1313         /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1314         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1315                 return sacked;
1316
1317         if (!(sacked & TCPCB_SACKED_ACKED)) {
1318                 if (sacked & TCPCB_SACKED_RETRANS) {
1319                         /* If the segment is not tagged as lost,
1320                          * we do not clear RETRANS, believing
1321                          * that retransmission is still in flight.
1322                          */
1323                         if (sacked & TCPCB_LOST) {
1324                                 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1325                                 tp->lost_out -= pcount;
1326                                 tp->retrans_out -= pcount;
1327                         }
1328                 } else {
1329                         if (!(sacked & TCPCB_RETRANS)) {
1330                                 /* New sack for not retransmitted frame,
1331                                  * which was in hole. It is reordering.
1332                                  */
1333                                 if (before(TCP_SKB_CB(skb)->seq,
1334                                            tcp_highest_sack_seq(tp)))
1335                                         state->reord = min(fack_count,
1336                                                            state->reord);
1337
1338                                 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1339                                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1340                                         state->flag |= FLAG_ONLY_ORIG_SACKED;
1341                         }
1342
1343                         if (sacked & TCPCB_LOST) {
1344                                 sacked &= ~TCPCB_LOST;
1345                                 tp->lost_out -= pcount;
1346                         }
1347                 }
1348
1349                 sacked |= TCPCB_SACKED_ACKED;
1350                 state->flag |= FLAG_DATA_SACKED;
1351                 tp->sacked_out += pcount;
1352
1353                 fack_count += pcount;
1354
1355                 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1356                 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1357                     before(TCP_SKB_CB(skb)->seq,
1358                            TCP_SKB_CB(tp->lost_skb_hint)->seq))
1359                         tp->lost_cnt_hint += pcount;
1360
1361                 if (fack_count > tp->fackets_out)
1362                         tp->fackets_out = fack_count;
1363         }
1364
1365         /* D-SACK. We can detect redundant retransmission in S|R and plain R
1366          * frames and clear it. undo_retrans is decreased above, L|R frames
1367          * are accounted above as well.
1368          */
1369         if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1370                 sacked &= ~TCPCB_SACKED_RETRANS;
1371                 tp->retrans_out -= pcount;
1372         }
1373
1374         return sacked;
1375 }
1376
1377 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1378                            struct tcp_sacktag_state *state,
1379                            unsigned int pcount, int shifted, int mss,
1380                            int dup_sack)
1381 {
1382         struct tcp_sock *tp = tcp_sk(sk);
1383         struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1384
1385         BUG_ON(!pcount);
1386
1387         /* Tweak before seqno plays */
1388         if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1389             !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1390                 tp->lost_cnt_hint += pcount;
1391
1392         TCP_SKB_CB(prev)->end_seq += shifted;
1393         TCP_SKB_CB(skb)->seq += shifted;
1394
1395         skb_shinfo(prev)->gso_segs += pcount;
1396         BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1397         skb_shinfo(skb)->gso_segs -= pcount;
1398
1399         /* When we're adding to gso_segs == 1, gso_size will be zero,
1400          * in theory this shouldn't be necessary but as long as DSACK
1401          * code can come after this skb later on it's better to keep
1402          * setting gso_size to something.
1403          */
1404         if (!skb_shinfo(prev)->gso_size) {
1405                 skb_shinfo(prev)->gso_size = mss;
1406                 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1407         }
1408
1409         /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1410         if (skb_shinfo(skb)->gso_segs <= 1) {
1411                 skb_shinfo(skb)->gso_size = 0;
1412                 skb_shinfo(skb)->gso_type = 0;
1413         }
1414
1415         /* We discard results */
1416         tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1417
1418         /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1419         TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1420
1421         if (skb->len > 0) {
1422                 BUG_ON(!tcp_skb_pcount(skb));
1423                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1424                 return 0;
1425         }
1426
1427         /* Whole SKB was eaten :-) */
1428
1429         if (skb == tp->retransmit_skb_hint)
1430                 tp->retransmit_skb_hint = prev;
1431         if (skb == tp->scoreboard_skb_hint)
1432                 tp->scoreboard_skb_hint = prev;
1433         if (skb == tp->lost_skb_hint) {
1434                 tp->lost_skb_hint = prev;
1435                 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1436         }
1437
1438         TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1439         if (skb == tcp_highest_sack(sk))
1440                 tcp_advance_highest_sack(sk, skb);
1441
1442         tcp_unlink_write_queue(skb, sk);
1443         sk_wmem_free_skb(sk, skb);
1444
1445         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1446
1447         return 1;
1448 }
1449
1450 /* I wish gso_size would have a bit more sane initialization than
1451  * something-or-zero which complicates things
1452  */
1453 static int tcp_skb_seglen(struct sk_buff *skb)
1454 {
1455         return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1456 }
1457
1458 /* Shifting pages past head area doesn't work */
1459 static int skb_can_shift(struct sk_buff *skb)
1460 {
1461         return !skb_headlen(skb) && skb_is_nonlinear(skb);
1462 }
1463
1464 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1465  * skb.
1466  */
1467 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1468                                           struct tcp_sacktag_state *state,
1469                                           u32 start_seq, u32 end_seq,
1470                                           int dup_sack)
1471 {
1472         struct tcp_sock *tp = tcp_sk(sk);
1473         struct sk_buff *prev;
1474         int mss;
1475         int pcount = 0;
1476         int len;
1477         int in_sack;
1478
1479         if (!sk_can_gso(sk))
1480                 goto fallback;
1481
1482         /* Normally R but no L won't result in plain S */
1483         if (!dup_sack &&
1484             (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1485                 goto fallback;
1486         if (!skb_can_shift(skb))
1487                 goto fallback;
1488         /* This frame is about to be dropped (was ACKed). */
1489         if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1490                 goto fallback;
1491
1492         /* Can only happen with delayed DSACK + discard craziness */
1493         if (unlikely(skb == tcp_write_queue_head(sk)))
1494                 goto fallback;
1495         prev = tcp_write_queue_prev(sk, skb);
1496
1497         if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1498                 goto fallback;
1499
1500         in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1501                   !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1502
1503         if (in_sack) {
1504                 len = skb->len;
1505                 pcount = tcp_skb_pcount(skb);
1506                 mss = tcp_skb_seglen(skb);
1507
1508                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1509                  * drop this restriction as unnecessary
1510                  */
1511                 if (mss != tcp_skb_seglen(prev))
1512                         goto fallback;
1513         } else {
1514                 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1515                         goto noop;
1516                 /* CHECKME: This is non-MSS split case only?, this will
1517                  * cause skipped skbs due to advancing loop btw, original
1518                  * has that feature too
1519                  */
1520                 if (tcp_skb_pcount(skb) <= 1)
1521                         goto noop;
1522
1523                 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1524                 if (!in_sack) {
1525                         /* TODO: head merge to next could be attempted here
1526                          * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1527                          * though it might not be worth of the additional hassle
1528                          *
1529                          * ...we can probably just fallback to what was done
1530                          * previously. We could try merging non-SACKed ones
1531                          * as well but it probably isn't going to buy off
1532                          * because later SACKs might again split them, and
1533                          * it would make skb timestamp tracking considerably
1534                          * harder problem.
1535                          */
1536                         goto fallback;
1537                 }
1538
1539                 len = end_seq - TCP_SKB_CB(skb)->seq;
1540                 BUG_ON(len < 0);
1541                 BUG_ON(len > skb->len);
1542
1543                 /* MSS boundaries should be honoured or else pcount will
1544                  * severely break even though it makes things bit trickier.
1545                  * Optimize common case to avoid most of the divides
1546                  */
1547                 mss = tcp_skb_mss(skb);
1548
1549                 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1550                  * drop this restriction as unnecessary
1551                  */
1552                 if (mss != tcp_skb_seglen(prev))
1553                         goto fallback;
1554
1555                 if (len == mss) {
1556                         pcount = 1;
1557                 } else if (len < mss) {
1558                         goto noop;
1559                 } else {
1560                         pcount = len / mss;
1561                         len = pcount * mss;
1562                 }
1563         }
1564
1565         if (!skb_shift(prev, skb, len))
1566                 goto fallback;
1567         if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1568                 goto out;
1569
1570         /* Hole filled allows collapsing with the next as well, this is very
1571          * useful when hole on every nth skb pattern happens
1572          */
1573         if (prev == tcp_write_queue_tail(sk))
1574                 goto out;
1575         skb = tcp_write_queue_next(sk, prev);
1576
1577         if (!skb_can_shift(skb) ||
1578             (skb == tcp_send_head(sk)) ||
1579             ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1580             (mss != tcp_skb_seglen(skb)))
1581                 goto out;
1582
1583         len = skb->len;
1584         if (skb_shift(prev, skb, len)) {
1585                 pcount += tcp_skb_pcount(skb);
1586                 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1587         }
1588
1589 out:
1590         state->fack_count += pcount;
1591         return prev;
1592
1593 noop:
1594         return skb;
1595
1596 fallback:
1597         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1598         return NULL;
1599 }
1600
1601 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1602                                         struct tcp_sack_block *next_dup,
1603                                         struct tcp_sacktag_state *state,
1604                                         u32 start_seq, u32 end_seq,
1605                                         int dup_sack_in)
1606 {
1607         struct tcp_sock *tp = tcp_sk(sk);
1608         struct sk_buff *tmp;
1609
1610         tcp_for_write_queue_from(skb, sk) {
1611                 int in_sack = 0;
1612                 int dup_sack = dup_sack_in;
1613
1614                 if (skb == tcp_send_head(sk))
1615                         break;
1616
1617                 /* queue is in-order => we can short-circuit the walk early */
1618                 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1619                         break;
1620
1621                 if ((next_dup != NULL) &&
1622                     before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1623                         in_sack = tcp_match_skb_to_sack(sk, skb,
1624                                                         next_dup->start_seq,
1625                                                         next_dup->end_seq);
1626                         if (in_sack > 0)
1627                                 dup_sack = 1;
1628                 }
1629
1630                 /* skb reference here is a bit tricky to get right, since
1631                  * shifting can eat and free both this skb and the next,
1632                  * so not even _safe variant of the loop is enough.
1633                  */
1634                 if (in_sack <= 0) {
1635                         tmp = tcp_shift_skb_data(sk, skb, state,
1636                                                  start_seq, end_seq, dup_sack);
1637                         if (tmp != NULL) {
1638                                 if (tmp != skb) {
1639                                         skb = tmp;
1640                                         continue;
1641                                 }
1642
1643                                 in_sack = 0;
1644                         } else {
1645                                 in_sack = tcp_match_skb_to_sack(sk, skb,
1646                                                                 start_seq,
1647                                                                 end_seq);
1648                         }
1649                 }
1650
1651                 if (unlikely(in_sack < 0))
1652                         break;
1653
1654                 if (in_sack) {
1655                         TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1656                                                                   state,
1657                                                                   dup_sack,
1658                                                                   tcp_skb_pcount(skb));
1659
1660                         if (!before(TCP_SKB_CB(skb)->seq,
1661                                     tcp_highest_sack_seq(tp)))
1662                                 tcp_advance_highest_sack(sk, skb);
1663                 }
1664
1665                 state->fack_count += tcp_skb_pcount(skb);
1666         }
1667         return skb;
1668 }
1669
1670 /* Avoid all extra work that is being done by sacktag while walking in
1671  * a normal way
1672  */
1673 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1674                                         struct tcp_sacktag_state *state,
1675                                         u32 skip_to_seq)
1676 {
1677         tcp_for_write_queue_from(skb, sk) {
1678                 if (skb == tcp_send_head(sk))
1679                         break;
1680
1681                 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1682                         break;
1683
1684                 state->fack_count += tcp_skb_pcount(skb);
1685         }
1686         return skb;
1687 }
1688
1689 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1690                                                 struct sock *sk,
1691                                                 struct tcp_sack_block *next_dup,
1692                                                 struct tcp_sacktag_state *state,
1693                                                 u32 skip_to_seq)
1694 {
1695         if (next_dup == NULL)
1696                 return skb;
1697
1698         if (before(next_dup->start_seq, skip_to_seq)) {
1699                 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1700                 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1701                                        next_dup->start_seq, next_dup->end_seq,
1702                                        1);
1703         }
1704
1705         return skb;
1706 }
1707
1708 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1709 {
1710         return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1711 }
1712
1713 static int
1714 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1715                         u32 prior_snd_una)
1716 {
1717         const struct inet_connection_sock *icsk = inet_csk(sk);
1718         struct tcp_sock *tp = tcp_sk(sk);
1719         unsigned char *ptr = (skb_transport_header(ack_skb) +
1720                               TCP_SKB_CB(ack_skb)->sacked);
1721         struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1722         struct tcp_sack_block sp[TCP_NUM_SACKS];
1723         struct tcp_sack_block *cache;
1724         struct tcp_sacktag_state state;
1725         struct sk_buff *skb;
1726         int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1727         int used_sacks;
1728         int found_dup_sack = 0;
1729         int i, j;
1730         int first_sack_index;
1731
1732         state.flag = 0;
1733         state.reord = tp->packets_out;
1734
1735         if (!tp->sacked_out) {
1736                 if (WARN_ON(tp->fackets_out))
1737                         tp->fackets_out = 0;
1738                 tcp_highest_sack_reset(sk);
1739         }
1740
1741         found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1742                                          num_sacks, prior_snd_una);
1743         if (found_dup_sack)
1744                 state.flag |= FLAG_DSACKING_ACK;
1745
1746         /* Eliminate too old ACKs, but take into
1747          * account more or less fresh ones, they can
1748          * contain valid SACK info.
1749          */
1750         if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1751                 return 0;
1752
1753         if (!tp->packets_out)
1754                 goto out;
1755
1756         used_sacks = 0;
1757         first_sack_index = 0;
1758         for (i = 0; i < num_sacks; i++) {
1759                 int dup_sack = !i && found_dup_sack;
1760
1761                 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1762                 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1763
1764                 if (!tcp_is_sackblock_valid(tp, dup_sack,
1765                                             sp[used_sacks].start_seq,
1766                                             sp[used_sacks].end_seq)) {
1767                         int mib_idx;
1768
1769                         if (dup_sack) {
1770                                 if (!tp->undo_marker)
1771                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1772                                 else
1773                                         mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1774                         } else {
1775                                 /* Don't count olds caused by ACK reordering */
1776                                 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1777                                     !after(sp[used_sacks].end_seq, tp->snd_una))
1778                                         continue;
1779                                 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1780                         }
1781
1782                         NET_INC_STATS_BH(sock_net(sk), mib_idx);
1783                         if (i == 0)
1784                                 first_sack_index = -1;
1785                         continue;
1786                 }
1787
1788                 /* Ignore very old stuff early */
1789                 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1790                         continue;
1791
1792                 used_sacks++;
1793         }
1794
1795         /* order SACK blocks to allow in order walk of the retrans queue */
1796         for (i = used_sacks - 1; i > 0; i--) {
1797                 for (j = 0; j < i; j++) {
1798                         if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1799                                 swap(sp[j], sp[j + 1]);
1800
1801                                 /* Track where the first SACK block goes to */
1802                                 if (j == first_sack_index)
1803                                         first_sack_index = j + 1;
1804                         }
1805                 }
1806         }
1807
1808         skb = tcp_write_queue_head(sk);
1809         state.fack_count = 0;
1810         i = 0;
1811
1812         if (!tp->sacked_out) {
1813                 /* It's already past, so skip checking against it */
1814                 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1815         } else {
1816                 cache = tp->recv_sack_cache;
1817                 /* Skip empty blocks in at head of the cache */
1818                 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1819                        !cache->end_seq)
1820                         cache++;
1821         }
1822
1823         while (i < used_sacks) {
1824                 u32 start_seq = sp[i].start_seq;
1825                 u32 end_seq = sp[i].end_seq;
1826                 int dup_sack = (found_dup_sack && (i == first_sack_index));
1827                 struct tcp_sack_block *next_dup = NULL;
1828
1829                 if (found_dup_sack && ((i + 1) == first_sack_index))
1830                         next_dup = &sp[i + 1];
1831
1832                 /* Event "B" in the comment above. */
1833                 if (after(end_seq, tp->high_seq))
1834                         state.flag |= FLAG_DATA_LOST;
1835
1836                 /* Skip too early cached blocks */
1837                 while (tcp_sack_cache_ok(tp, cache) &&
1838                        !before(start_seq, cache->end_seq))
1839                         cache++;
1840
1841                 /* Can skip some work by looking recv_sack_cache? */
1842                 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1843                     after(end_seq, cache->start_seq)) {
1844
1845                         /* Head todo? */
1846                         if (before(start_seq, cache->start_seq)) {
1847                                 skb = tcp_sacktag_skip(skb, sk, &state,
1848                                                        start_seq);
1849                                 skb = tcp_sacktag_walk(skb, sk, next_dup,
1850                                                        &state,
1851                                                        start_seq,
1852                                                        cache->start_seq,
1853                                                        dup_sack);
1854                         }
1855
1856                         /* Rest of the block already fully processed? */
1857                         if (!after(end_seq, cache->end_seq))
1858                                 goto advance_sp;
1859
1860                         skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1861                                                        &state,
1862                                                        cache->end_seq);
1863
1864                         /* ...tail remains todo... */
1865                         if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1866                                 /* ...but better entrypoint exists! */
1867                                 skb = tcp_highest_sack(sk);
1868                                 if (skb == NULL)
1869                                         break;
1870                                 state.fack_count = tp->fackets_out;
1871                                 cache++;
1872                                 goto walk;
1873                         }
1874
1875                         skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1876                         /* Check overlap against next cached too (past this one already) */
1877                         cache++;
1878                         continue;
1879                 }
1880
1881                 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1882                         skb = tcp_highest_sack(sk);
1883                         if (skb == NULL)
1884                                 break;
1885                         state.fack_count = tp->fackets_out;
1886                 }
1887                 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1888
1889 walk:
1890                 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1891                                        start_seq, end_seq, dup_sack);
1892
1893 advance_sp:
1894                 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1895                  * due to in-order walk
1896                  */
1897                 if (after(end_seq, tp->frto_highmark))
1898                         state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1899
1900                 i++;
1901         }
1902
1903         /* Clear the head of the cache sack blocks so we can skip it next time */
1904         for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1905                 tp->recv_sack_cache[i].start_seq = 0;
1906                 tp->recv_sack_cache[i].end_seq = 0;
1907         }
1908         for (j = 0; j < used_sacks; j++)
1909                 tp->recv_sack_cache[i++] = sp[j];
1910
1911         tcp_mark_lost_retrans(sk);
1912
1913         tcp_verify_left_out(tp);
1914
1915         if ((state.reord < tp->fackets_out) &&
1916             ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1917             (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1918                 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1919
1920 out:
1921
1922 #if FASTRETRANS_DEBUG > 0
1923         WARN_ON((int)tp->sacked_out < 0);
1924         WARN_ON((int)tp->lost_out < 0);
1925         WARN_ON((int)tp->retrans_out < 0);
1926         WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1927 #endif
1928         return state.flag;
1929 }
1930
1931 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1932  * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1933  */
1934 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1935 {
1936         u32 holes;
1937
1938         holes = max(tp->lost_out, 1U);
1939         holes = min(holes, tp->packets_out);
1940
1941         if ((tp->sacked_out + holes) > tp->packets_out) {
1942                 tp->sacked_out = tp->packets_out - holes;
1943                 return 1;
1944         }
1945         return 0;
1946 }
1947
1948 /* If we receive more dupacks than we expected counting segments
1949  * in assumption of absent reordering, interpret this as reordering.
1950  * The only another reason could be bug in receiver TCP.
1951  */
1952 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1953 {
1954         struct tcp_sock *tp = tcp_sk(sk);
1955         if (tcp_limit_reno_sacked(tp))
1956                 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1957 }
1958
1959 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1960
1961 static void tcp_add_reno_sack(struct sock *sk)
1962 {
1963         struct tcp_sock *tp = tcp_sk(sk);
1964         tp->sacked_out++;
1965         tcp_check_reno_reordering(sk, 0);
1966         tcp_verify_left_out(tp);
1967 }
1968
1969 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1970
1971 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1972 {
1973         struct tcp_sock *tp = tcp_sk(sk);
1974
1975         if (acked > 0) {
1976                 /* One ACK acked hole. The rest eat duplicate ACKs. */
1977                 if (acked - 1 >= tp->sacked_out)
1978                         tp->sacked_out = 0;
1979                 else
1980                         tp->sacked_out -= acked - 1;
1981         }
1982         tcp_check_reno_reordering(sk, acked);
1983         tcp_verify_left_out(tp);
1984 }
1985
1986 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1987 {
1988         tp->sacked_out = 0;
1989 }
1990
1991 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1992 {
1993         return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1994 }
1995
1996 /* F-RTO can only be used if TCP has never retransmitted anything other than
1997  * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1998  */
1999 int tcp_use_frto(struct sock *sk)
2000 {
2001         const struct tcp_sock *tp = tcp_sk(sk);
2002         const struct inet_connection_sock *icsk = inet_csk(sk);
2003         struct sk_buff *skb;
2004
2005         if (!sysctl_tcp_frto)
2006                 return 0;
2007
2008         /* MTU probe and F-RTO won't really play nicely along currently */
2009         if (icsk->icsk_mtup.probe_size)
2010                 return 0;
2011
2012         if (tcp_is_sackfrto(tp))
2013                 return 1;
2014
2015         /* Avoid expensive walking of rexmit queue if possible */
2016         if (tp->retrans_out > 1)
2017                 return 0;
2018
2019         skb = tcp_write_queue_head(sk);
2020         if (tcp_skb_is_last(sk, skb))
2021                 return 1;
2022         skb = tcp_write_queue_next(sk, skb);    /* Skips head */
2023         tcp_for_write_queue_from(skb, sk) {
2024                 if (skb == tcp_send_head(sk))
2025                         break;
2026                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2027                         return 0;
2028                 /* Short-circuit when first non-SACKed skb has been checked */
2029                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2030                         break;
2031         }
2032         return 1;
2033 }
2034
2035 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2036  * recovery a bit and use heuristics in tcp_process_frto() to detect if
2037  * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2038  * keep retrans_out counting accurate (with SACK F-RTO, other than head
2039  * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2040  * bits are handled if the Loss state is really to be entered (in
2041  * tcp_enter_frto_loss).
2042  *
2043  * Do like tcp_enter_loss() would; when RTO expires the second time it
2044  * does:
2045  *  "Reduce ssthresh if it has not yet been made inside this window."
2046  */
2047 void tcp_enter_frto(struct sock *sk)
2048 {
2049         const struct inet_connection_sock *icsk = inet_csk(sk);
2050         struct tcp_sock *tp = tcp_sk(sk);
2051         struct sk_buff *skb;
2052
2053         if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2054             tp->snd_una == tp->high_seq ||
2055             ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2056              !icsk->icsk_retransmits)) {
2057                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2058                 /* Our state is too optimistic in ssthresh() call because cwnd
2059                  * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2060                  * recovery has not yet completed. Pattern would be this: RTO,
2061                  * Cumulative ACK, RTO (2xRTO for the same segment does not end
2062                  * up here twice).
2063                  * RFC4138 should be more specific on what to do, even though
2064                  * RTO is quite unlikely to occur after the first Cumulative ACK
2065                  * due to back-off and complexity of triggering events ...
2066                  */
2067                 if (tp->frto_counter) {
2068                         u32 stored_cwnd;
2069                         stored_cwnd = tp->snd_cwnd;
2070                         tp->snd_cwnd = 2;
2071                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2072                         tp->snd_cwnd = stored_cwnd;
2073                 } else {
2074                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2075                 }
2076                 /* ... in theory, cong.control module could do "any tricks" in
2077                  * ssthresh(), which means that ca_state, lost bits and lost_out
2078                  * counter would have to be faked before the call occurs. We
2079                  * consider that too expensive, unlikely and hacky, so modules
2080                  * using these in ssthresh() must deal these incompatibility
2081                  * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2082                  */
2083                 tcp_ca_event(sk, CA_EVENT_FRTO);
2084         }
2085
2086         tp->undo_marker = tp->snd_una;
2087         tp->undo_retrans = 0;
2088
2089         skb = tcp_write_queue_head(sk);
2090         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2091                 tp->undo_marker = 0;
2092         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2093                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2094                 tp->retrans_out -= tcp_skb_pcount(skb);
2095         }
2096         tcp_verify_left_out(tp);
2097
2098         /* Too bad if TCP was application limited */
2099         tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2100
2101         /* Earlier loss recovery underway (see RFC4138; Appendix B).
2102          * The last condition is necessary at least in tp->frto_counter case.
2103          */
2104         if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2105             ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2106             after(tp->high_seq, tp->snd_una)) {
2107                 tp->frto_highmark = tp->high_seq;
2108         } else {
2109                 tp->frto_highmark = tp->snd_nxt;
2110         }
2111         tcp_set_ca_state(sk, TCP_CA_Disorder);
2112         tp->high_seq = tp->snd_nxt;
2113         tp->frto_counter = 1;
2114 }
2115
2116 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2117  * which indicates that we should follow the traditional RTO recovery,
2118  * i.e. mark everything lost and do go-back-N retransmission.
2119  */
2120 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2121 {
2122         struct tcp_sock *tp = tcp_sk(sk);
2123         struct sk_buff *skb;
2124
2125         tp->lost_out = 0;
2126         tp->retrans_out = 0;
2127         if (tcp_is_reno(tp))
2128                 tcp_reset_reno_sack(tp);
2129
2130         tcp_for_write_queue(skb, sk) {
2131                 if (skb == tcp_send_head(sk))
2132                         break;
2133
2134                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2135                 /*
2136                  * Count the retransmission made on RTO correctly (only when
2137                  * waiting for the first ACK and did not get it)...
2138                  */
2139                 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2140                         /* For some reason this R-bit might get cleared? */
2141                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2142                                 tp->retrans_out += tcp_skb_pcount(skb);
2143                         /* ...enter this if branch just for the first segment */
2144                         flag |= FLAG_DATA_ACKED;
2145                 } else {
2146                         if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2147                                 tp->undo_marker = 0;
2148                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2149                 }
2150
2151                 /* Marking forward transmissions that were made after RTO lost
2152                  * can cause unnecessary retransmissions in some scenarios,
2153                  * SACK blocks will mitigate that in some but not in all cases.
2154                  * We used to not mark them but it was causing break-ups with
2155                  * receivers that do only in-order receival.
2156                  *
2157                  * TODO: we could detect presence of such receiver and select
2158                  * different behavior per flow.
2159                  */
2160                 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2161                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2162                         tp->lost_out += tcp_skb_pcount(skb);
2163                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2164                 }
2165         }
2166         tcp_verify_left_out(tp);
2167
2168         tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2169         tp->snd_cwnd_cnt = 0;
2170         tp->snd_cwnd_stamp = tcp_time_stamp;
2171         tp->frto_counter = 0;
2172         tp->bytes_acked = 0;
2173
2174         tp->reordering = min_t(unsigned int, tp->reordering,
2175                                sysctl_tcp_reordering);
2176         tcp_set_ca_state(sk, TCP_CA_Loss);
2177         tp->high_seq = tp->snd_nxt;
2178         TCP_ECN_queue_cwr(tp);
2179
2180         tcp_clear_all_retrans_hints(tp);
2181 }
2182
2183 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2184 {
2185         tp->retrans_out = 0;
2186         tp->lost_out = 0;
2187
2188         tp->undo_marker = 0;
2189         tp->undo_retrans = 0;
2190 }
2191
2192 void tcp_clear_retrans(struct tcp_sock *tp)
2193 {
2194         tcp_clear_retrans_partial(tp);
2195
2196         tp->fackets_out = 0;
2197         tp->sacked_out = 0;
2198 }
2199
2200 /* Enter Loss state. If "how" is not zero, forget all SACK information
2201  * and reset tags completely, otherwise preserve SACKs. If receiver
2202  * dropped its ofo queue, we will know this due to reneging detection.
2203  */
2204 void tcp_enter_loss(struct sock *sk, int how)
2205 {
2206         const struct inet_connection_sock *icsk = inet_csk(sk);
2207         struct tcp_sock *tp = tcp_sk(sk);
2208         struct sk_buff *skb;
2209
2210         /* Reduce ssthresh if it has not yet been made inside this window. */
2211         if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2212             (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2213                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2214                 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2215                 tcp_ca_event(sk, CA_EVENT_LOSS);
2216         }
2217         tp->snd_cwnd       = 1;
2218         tp->snd_cwnd_cnt   = 0;
2219         tp->snd_cwnd_stamp = tcp_time_stamp;
2220
2221         tp->bytes_acked = 0;
2222         tcp_clear_retrans_partial(tp);
2223
2224         if (tcp_is_reno(tp))
2225                 tcp_reset_reno_sack(tp);
2226
2227         if (!how) {
2228                 /* Push undo marker, if it was plain RTO and nothing
2229                  * was retransmitted. */
2230                 tp->undo_marker = tp->snd_una;
2231         } else {
2232                 tp->sacked_out = 0;
2233                 tp->fackets_out = 0;
2234         }
2235         tcp_clear_all_retrans_hints(tp);
2236
2237         tcp_for_write_queue(skb, sk) {
2238                 if (skb == tcp_send_head(sk))
2239                         break;
2240
2241                 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2242                         tp->undo_marker = 0;
2243                 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2244                 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2245                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2246                         TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2247                         tp->lost_out += tcp_skb_pcount(skb);
2248                         tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2249                 }
2250         }
2251         tcp_verify_left_out(tp);
2252
2253         tp->reordering = min_t(unsigned int, tp->reordering,
2254                                sysctl_tcp_reordering);
2255         tcp_set_ca_state(sk, TCP_CA_Loss);
2256         tp->high_seq = tp->snd_nxt;
2257         TCP_ECN_queue_cwr(tp);
2258         /* Abort F-RTO algorithm if one is in progress */
2259         tp->frto_counter = 0;
2260 }
2261
2262 /* If ACK arrived pointing to a remembered SACK, it means that our
2263  * remembered SACKs do not reflect real state of receiver i.e.
2264  * receiver _host_ is heavily congested (or buggy).
2265  *
2266  * Do processing similar to RTO timeout.
2267  */
2268 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2269 {
2270         if (flag & FLAG_SACK_RENEGING) {
2271                 struct inet_connection_sock *icsk = inet_csk(sk);
2272                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2273
2274                 tcp_enter_loss(sk, 1);
2275                 icsk->icsk_retransmits++;
2276                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2277                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2278                                           icsk->icsk_rto, TCP_RTO_MAX);
2279                 return 1;
2280         }
2281         return 0;
2282 }
2283
2284 static inline int tcp_fackets_out(struct tcp_sock *tp)
2285 {
2286         return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2287 }
2288
2289 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2290  * counter when SACK is enabled (without SACK, sacked_out is used for
2291  * that purpose).
2292  *
2293  * Instead, with FACK TCP uses fackets_out that includes both SACKed
2294  * segments up to the highest received SACK block so far and holes in
2295  * between them.
2296  *
2297  * With reordering, holes may still be in flight, so RFC3517 recovery
2298  * uses pure sacked_out (total number of SACKed segments) even though
2299  * it violates the RFC that uses duplicate ACKs, often these are equal
2300  * but when e.g. out-of-window ACKs or packet duplication occurs,
2301  * they differ. Since neither occurs due to loss, TCP should really
2302  * ignore them.
2303  */
2304 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2305 {
2306         return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2307 }
2308
2309 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2310 {
2311         return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2312 }
2313
2314 static inline int tcp_head_timedout(struct sock *sk)
2315 {
2316         struct tcp_sock *tp = tcp_sk(sk);
2317
2318         return tp->packets_out &&
2319                tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2320 }
2321
2322 /* Linux NewReno/SACK/FACK/ECN state machine.
2323  * --------------------------------------
2324  *
2325  * "Open"       Normal state, no dubious events, fast path.
2326  * "Disorder"   In all the respects it is "Open",
2327  *              but requires a bit more attention. It is entered when
2328  *              we see some SACKs or dupacks. It is split of "Open"
2329  *              mainly to move some processing from fast path to slow one.
2330  * "CWR"        CWND was reduced due to some Congestion Notification event.
2331  *              It can be ECN, ICMP source quench, local device congestion.
2332  * "Recovery"   CWND was reduced, we are fast-retransmitting.
2333  * "Loss"       CWND was reduced due to RTO timeout or SACK reneging.
2334  *
2335  * tcp_fastretrans_alert() is entered:
2336  * - each incoming ACK, if state is not "Open"
2337  * - when arrived ACK is unusual, namely:
2338  *      * SACK
2339  *      * Duplicate ACK.
2340  *      * ECN ECE.
2341  *
2342  * Counting packets in flight is pretty simple.
2343  *
2344  *      in_flight = packets_out - left_out + retrans_out
2345  *
2346  *      packets_out is SND.NXT-SND.UNA counted in packets.
2347  *
2348  *      retrans_out is number of retransmitted segments.
2349  *
2350  *      left_out is number of segments left network, but not ACKed yet.
2351  *
2352  *              left_out = sacked_out + lost_out
2353  *
2354  *     sacked_out: Packets, which arrived to receiver out of order
2355  *                 and hence not ACKed. With SACKs this number is simply
2356  *                 amount of SACKed data. Even without SACKs
2357  *                 it is easy to give pretty reliable estimate of this number,
2358  *                 counting duplicate ACKs.
2359  *
2360  *       lost_out: Packets lost by network. TCP has no explicit
2361  *                 "loss notification" feedback from network (for now).
2362  *                 It means that this number can be only _guessed_.
2363  *                 Actually, it is the heuristics to predict lossage that
2364  *                 distinguishes different algorithms.
2365  *
2366  *      F.e. after RTO, when all the queue is considered as lost,
2367  *      lost_out = packets_out and in_flight = retrans_out.
2368  *
2369  *              Essentially, we have now two algorithms counting
2370  *              lost packets.
2371  *
2372  *              FACK: It is the simplest heuristics. As soon as we decided
2373  *              that something is lost, we decide that _all_ not SACKed
2374  *              packets until the most forward SACK are lost. I.e.
2375  *              lost_out = fackets_out - sacked_out and left_out = fackets_out.
2376  *              It is absolutely correct estimate, if network does not reorder
2377  *              packets. And it loses any connection to reality when reordering
2378  *              takes place. We use FACK by default until reordering
2379  *              is suspected on the path to this destination.
2380  *
2381  *              NewReno: when Recovery is entered, we assume that one segment
2382  *              is lost (classic Reno). While we are in Recovery and
2383  *              a partial ACK arrives, we assume that one more packet
2384  *              is lost (NewReno). This heuristics are the same in NewReno
2385  *              and SACK.
2386  *
2387  *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2388  *  deflation etc. CWND is real congestion window, never inflated, changes
2389  *  only according to classic VJ rules.
2390  *
2391  * Really tricky (and requiring careful tuning) part of algorithm
2392  * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2393  * The first determines the moment _when_ we should reduce CWND and,
2394  * hence, slow down forward transmission. In fact, it determines the moment
2395  * when we decide that hole is caused by loss, rather than by a reorder.
2396  *
2397  * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2398  * holes, caused by lost packets.
2399  *
2400  * And the most logically complicated part of algorithm is undo
2401  * heuristics. We detect false retransmits due to both too early
2402  * fast retransmit (reordering) and underestimated RTO, analyzing
2403  * timestamps and D-SACKs. When we detect that some segments were
2404  * retransmitted by mistake and CWND reduction was wrong, we undo
2405  * window reduction and abort recovery phase. This logic is hidden
2406  * inside several functions named tcp_try_undo_<something>.
2407  */
2408
2409 /* This function decides, when we should leave Disordered state
2410  * and enter Recovery phase, reducing congestion window.
2411  *
2412  * Main question: may we further continue forward transmission
2413  * with the same cwnd?
2414  */
2415 static int tcp_time_to_recover(struct sock *sk)
2416 {
2417         struct tcp_sock *tp = tcp_sk(sk);
2418         __u32 packets_out;
2419
2420         /* Do not perform any recovery during F-RTO algorithm */
2421         if (tp->frto_counter)
2422                 return 0;
2423
2424         /* Trick#1: The loss is proven. */
2425         if (tp->lost_out)
2426                 return 1;
2427
2428         /* Not-A-Trick#2 : Classic rule... */
2429         if (tcp_dupack_heurestics(tp) > tp->reordering)
2430                 return 1;
2431
2432         /* Trick#3 : when we use RFC2988 timer restart, fast
2433          * retransmit can be triggered by timeout of queue head.
2434          */
2435         if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2436                 return 1;
2437
2438         /* Trick#4: It is still not OK... But will it be useful to delay
2439          * recovery more?
2440          */
2441         packets_out = tp->packets_out;
2442         if (packets_out <= tp->reordering &&
2443             tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2444             !tcp_may_send_now(sk)) {
2445                 /* We have nothing to send. This connection is limited
2446                  * either by receiver window or by application.
2447                  */
2448                 return 1;
2449         }
2450
2451         return 0;
2452 }
2453
2454 /* New heuristics: it is possible only after we switched to restart timer
2455  * each time when something is ACKed. Hence, we can detect timed out packets
2456  * during fast retransmit without falling to slow start.
2457  *
2458  * Usefulness of this as is very questionable, since we should know which of
2459  * the segments is the next to timeout which is relatively expensive to find
2460  * in general case unless we add some data structure just for that. The
2461  * current approach certainly won't find the right one too often and when it
2462  * finally does find _something_ it usually marks large part of the window
2463  * right away (because a retransmission with a larger timestamp blocks the
2464  * loop from advancing). -ij
2465  */
2466 static void tcp_timeout_skbs(struct sock *sk)
2467 {
2468         struct tcp_sock *tp = tcp_sk(sk);
2469         struct sk_buff *skb;
2470
2471         if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2472                 return;
2473
2474         skb = tp->scoreboard_skb_hint;
2475         if (tp->scoreboard_skb_hint == NULL)
2476                 skb = tcp_write_queue_head(sk);
2477
2478         tcp_for_write_queue_from(skb, sk) {
2479                 if (skb == tcp_send_head(sk))
2480                         break;
2481                 if (!tcp_skb_timedout(sk, skb))
2482                         break;
2483
2484                 tcp_skb_mark_lost(tp, skb);
2485         }
2486
2487         tp->scoreboard_skb_hint = skb;
2488
2489         tcp_verify_left_out(tp);
2490 }
2491
2492 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2493  * is against sacked "cnt", otherwise it's against facked "cnt"
2494  */
2495 static void tcp_mark_head_lost(struct sock *sk, int packets)
2496 {
2497         struct tcp_sock *tp = tcp_sk(sk);
2498         struct sk_buff *skb;
2499         int cnt, oldcnt;
2500         int err;
2501         unsigned int mss;
2502
2503         WARN_ON(packets > tp->packets_out);
2504         if (tp->lost_skb_hint) {
2505                 skb = tp->lost_skb_hint;
2506                 cnt = tp->lost_cnt_hint;
2507         } else {
2508                 skb = tcp_write_queue_head(sk);
2509                 cnt = 0;
2510         }
2511
2512         tcp_for_write_queue_from(skb, sk) {
2513                 if (skb == tcp_send_head(sk))
2514                         break;
2515                 /* TODO: do this better */
2516                 /* this is not the most efficient way to do this... */
2517                 tp->lost_skb_hint = skb;
2518                 tp->lost_cnt_hint = cnt;
2519
2520                 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2521                         break;
2522
2523                 oldcnt = cnt;
2524                 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2525                     (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2526                         cnt += tcp_skb_pcount(skb);
2527
2528                 if (cnt > packets) {
2529                         if (tcp_is_sack(tp) || (oldcnt >= packets))
2530                                 break;
2531
2532                         mss = skb_shinfo(skb)->gso_size;
2533                         err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2534                         if (err < 0)
2535                                 break;
2536                         cnt = packets;
2537                 }
2538
2539                 tcp_skb_mark_lost(tp, skb);
2540         }
2541         tcp_verify_left_out(tp);
2542 }
2543
2544 /* Account newly detected lost packet(s) */
2545
2546 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2547 {
2548         struct tcp_sock *tp = tcp_sk(sk);
2549
2550         if (tcp_is_reno(tp)) {
2551                 tcp_mark_head_lost(sk, 1);
2552         } else if (tcp_is_fack(tp)) {
2553                 int lost = tp->fackets_out - tp->reordering;
2554                 if (lost <= 0)
2555                         lost = 1;
2556                 tcp_mark_head_lost(sk, lost);
2557         } else {
2558                 int sacked_upto = tp->sacked_out - tp->reordering;
2559                 if (sacked_upto < fast_rexmit)
2560                         sacked_upto = fast_rexmit;
2561                 tcp_mark_head_lost(sk, sacked_upto);
2562         }
2563
2564         tcp_timeout_skbs(sk);
2565 }
2566
2567 /* CWND moderation, preventing bursts due to too big ACKs
2568  * in dubious situations.
2569  */
2570 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2571 {
2572         tp->snd_cwnd = min(tp->snd_cwnd,
2573                            tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2574         tp->snd_cwnd_stamp = tcp_time_stamp;
2575 }
2576
2577 /* Lower bound on congestion window is slow start threshold
2578  * unless congestion avoidance choice decides to overide it.
2579  */
2580 static inline u32 tcp_cwnd_min(const struct sock *sk)
2581 {
2582         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2583
2584         return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2585 }
2586
2587 /* Decrease cwnd each second ack. */
2588 static void tcp_cwnd_down(struct sock *sk, int flag)
2589 {
2590         struct tcp_sock *tp = tcp_sk(sk);
2591         int decr = tp->snd_cwnd_cnt + 1;
2592
2593         if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2594             (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2595                 tp->snd_cwnd_cnt = decr & 1;
2596                 decr >>= 1;
2597
2598                 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2599                         tp->snd_cwnd -= decr;
2600
2601                 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2602                 tp->snd_cwnd_stamp = tcp_time_stamp;
2603         }
2604 }
2605
2606 /* Nothing was retransmitted or returned timestamp is less
2607  * than timestamp of the first retransmission.
2608  */
2609 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2610 {
2611         return !tp->retrans_stamp ||
2612                 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2613                  before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2614 }
2615
2616 /* Undo procedures. */
2617
2618 #if FASTRETRANS_DEBUG > 1
2619 static void DBGUNDO(struct sock *sk, const char *msg)
2620 {
2621         struct tcp_sock *tp = tcp_sk(sk);
2622         struct inet_sock *inet = inet_sk(sk);
2623
2624         if (sk->sk_family == AF_INET) {
2625                 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2626                        msg,
2627                        &inet->daddr, ntohs(inet->dport),
2628                        tp->snd_cwnd, tcp_left_out(tp),
2629                        tp->snd_ssthresh, tp->prior_ssthresh,
2630                        tp->packets_out);
2631         }
2632 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2633         else if (sk->sk_family == AF_INET6) {
2634                 struct ipv6_pinfo *np = inet6_sk(sk);
2635                 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2636                        msg,
2637                        &np->daddr, ntohs(inet->dport),
2638                        tp->snd_cwnd, tcp_left_out(tp),
2639                        tp->snd_ssthresh, tp->prior_ssthresh,
2640                        tp->packets_out);
2641         }
2642 #endif
2643 }
2644 #else
2645 #define DBGUNDO(x...) do { } while (0)
2646 #endif
2647
2648 static void tcp_undo_cwr(struct sock *sk, const int undo)
2649 {
2650         struct tcp_sock *tp = tcp_sk(sk);
2651
2652         if (tp->prior_ssthresh) {
2653                 const struct inet_connection_sock *icsk = inet_csk(sk);
2654
2655                 if (icsk->icsk_ca_ops->undo_cwnd)
2656                         tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2657                 else
2658                         tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2659
2660                 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2661                         tp->snd_ssthresh = tp->prior_ssthresh;
2662                         TCP_ECN_withdraw_cwr(tp);
2663                 }
2664         } else {
2665                 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2666         }
2667         tcp_moderate_cwnd(tp);
2668         tp->snd_cwnd_stamp = tcp_time_stamp;
2669 }
2670
2671 static inline int tcp_may_undo(struct tcp_sock *tp)
2672 {
2673         return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2674 }
2675
2676 /* People celebrate: "We love our President!" */
2677 static int tcp_try_undo_recovery(struct sock *sk)
2678 {
2679         struct tcp_sock *tp = tcp_sk(sk);
2680
2681         if (tcp_may_undo(tp)) {
2682                 int mib_idx;
2683
2684                 /* Happy end! We did not retransmit anything
2685                  * or our original transmission succeeded.
2686                  */
2687                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2688                 tcp_undo_cwr(sk, 1);
2689                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2690                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
2691                 else
2692                         mib_idx = LINUX_MIB_TCPFULLUNDO;
2693
2694                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2695                 tp->undo_marker = 0;
2696         }
2697         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2698                 /* Hold old state until something *above* high_seq
2699                  * is ACKed. For Reno it is MUST to prevent false
2700                  * fast retransmits (RFC2582). SACK TCP is safe. */
2701                 tcp_moderate_cwnd(tp);
2702                 return 1;
2703         }
2704         tcp_set_ca_state(sk, TCP_CA_Open);
2705         return 0;
2706 }
2707
2708 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2709 static void tcp_try_undo_dsack(struct sock *sk)
2710 {
2711         struct tcp_sock *tp = tcp_sk(sk);
2712
2713         if (tp->undo_marker && !tp->undo_retrans) {
2714                 DBGUNDO(sk, "D-SACK");
2715                 tcp_undo_cwr(sk, 1);
2716                 tp->undo_marker = 0;
2717                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2718         }
2719 }
2720
2721 /* Undo during fast recovery after partial ACK. */
2722
2723 static int tcp_try_undo_partial(struct sock *sk, int acked)
2724 {
2725         struct tcp_sock *tp = tcp_sk(sk);
2726         /* Partial ACK arrived. Force Hoe's retransmit. */
2727         int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2728
2729         if (tcp_may_undo(tp)) {
2730                 /* Plain luck! Hole if filled with delayed
2731                  * packet, rather than with a retransmit.
2732                  */
2733                 if (tp->retrans_out == 0)
2734                         tp->retrans_stamp = 0;
2735
2736                 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2737
2738                 DBGUNDO(sk, "Hoe");
2739                 tcp_undo_cwr(sk, 0);
2740                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2741
2742                 /* So... Do not make Hoe's retransmit yet.
2743                  * If the first packet was delayed, the rest
2744                  * ones are most probably delayed as well.
2745                  */
2746                 failed = 0;
2747         }
2748         return failed;
2749 }
2750
2751 /* Undo during loss recovery after partial ACK. */
2752 static int tcp_try_undo_loss(struct sock *sk)
2753 {
2754         struct tcp_sock *tp = tcp_sk(sk);
2755
2756         if (tcp_may_undo(tp)) {
2757                 struct sk_buff *skb;
2758                 tcp_for_write_queue(skb, sk) {
2759                         if (skb == tcp_send_head(sk))
2760                                 break;
2761                         TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2762                 }
2763
2764                 tcp_clear_all_retrans_hints(tp);
2765
2766                 DBGUNDO(sk, "partial loss");
2767                 tp->lost_out = 0;
2768                 tcp_undo_cwr(sk, 1);
2769                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2770                 inet_csk(sk)->icsk_retransmits = 0;
2771                 tp->undo_marker = 0;
2772                 if (tcp_is_sack(tp))
2773                         tcp_set_ca_state(sk, TCP_CA_Open);
2774                 return 1;
2775         }
2776         return 0;
2777 }
2778
2779 static inline void tcp_complete_cwr(struct sock *sk)
2780 {
2781         struct tcp_sock *tp = tcp_sk(sk);
2782         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2783         tp->snd_cwnd_stamp = tcp_time_stamp;
2784         tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2785 }
2786
2787 static void tcp_try_keep_open(struct sock *sk)
2788 {
2789         struct tcp_sock *tp = tcp_sk(sk);
2790         int state = TCP_CA_Open;
2791
2792         if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2793                 state = TCP_CA_Disorder;
2794
2795         if (inet_csk(sk)->icsk_ca_state != state) {
2796                 tcp_set_ca_state(sk, state);
2797                 tp->high_seq = tp->snd_nxt;
2798         }
2799 }
2800
2801 static void tcp_try_to_open(struct sock *sk, int flag)
2802 {
2803         struct tcp_sock *tp = tcp_sk(sk);
2804
2805         tcp_verify_left_out(tp);
2806
2807         if (!tp->frto_counter && tp->retrans_out == 0)
2808                 tp->retrans_stamp = 0;
2809
2810         if (flag & FLAG_ECE)
2811                 tcp_enter_cwr(sk, 1);
2812
2813         if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2814                 tcp_try_keep_open(sk);
2815                 tcp_moderate_cwnd(tp);
2816         } else {
2817                 tcp_cwnd_down(sk, flag);
2818         }
2819 }
2820
2821 static void tcp_mtup_probe_failed(struct sock *sk)
2822 {
2823         struct inet_connection_sock *icsk = inet_csk(sk);
2824
2825         icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2826         icsk->icsk_mtup.probe_size = 0;
2827 }
2828
2829 static void tcp_mtup_probe_success(struct sock *sk)
2830 {
2831         struct tcp_sock *tp = tcp_sk(sk);
2832         struct inet_connection_sock *icsk = inet_csk(sk);
2833
2834         /* FIXME: breaks with very large cwnd */
2835         tp->prior_ssthresh = tcp_current_ssthresh(sk);
2836         tp->snd_cwnd = tp->snd_cwnd *
2837                        tcp_mss_to_mtu(sk, tp->mss_cache) /
2838                        icsk->icsk_mtup.probe_size;
2839         tp->snd_cwnd_cnt = 0;
2840         tp->snd_cwnd_stamp = tcp_time_stamp;
2841         tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2842
2843         icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2844         icsk->icsk_mtup.probe_size = 0;
2845         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2846 }
2847
2848 /* Do a simple retransmit without using the backoff mechanisms in
2849  * tcp_timer. This is used for path mtu discovery.
2850  * The socket is already locked here.
2851  */
2852 void tcp_simple_retransmit(struct sock *sk)
2853 {
2854         const struct inet_connection_sock *icsk = inet_csk(sk);
2855         struct tcp_sock *tp = tcp_sk(sk);
2856         struct sk_buff *skb;
2857         unsigned int mss = tcp_current_mss(sk);
2858         u32 prior_lost = tp->lost_out;
2859
2860         tcp_for_write_queue(skb, sk) {
2861                 if (skb == tcp_send_head(sk))
2862                         break;
2863                 if (tcp_skb_seglen(skb) > mss &&
2864                     !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2865                         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2866                                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2867                                 tp->retrans_out -= tcp_skb_pcount(skb);
2868                         }
2869                         tcp_skb_mark_lost_uncond_verify(tp, skb);
2870                 }
2871         }
2872
2873         tcp_clear_retrans_hints_partial(tp);
2874
2875         if (prior_lost == tp->lost_out)
2876                 return;
2877
2878         if (tcp_is_reno(tp))
2879                 tcp_limit_reno_sacked(tp);
2880
2881         tcp_verify_left_out(tp);
2882
2883         /* Don't muck with the congestion window here.
2884          * Reason is that we do not increase amount of _data_
2885          * in network, but units changed and effective
2886          * cwnd/ssthresh really reduced now.
2887          */
2888         if (icsk->icsk_ca_state != TCP_CA_Loss) {
2889                 tp->high_seq = tp->snd_nxt;
2890                 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2891                 tp->prior_ssthresh = 0;
2892                 tp->undo_marker = 0;
2893                 tcp_set_ca_state(sk, TCP_CA_Loss);
2894         }
2895         tcp_xmit_retransmit_queue(sk);
2896 }
2897
2898 /* Process an event, which can update packets-in-flight not trivially.
2899  * Main goal of this function is to calculate new estimate for left_out,
2900  * taking into account both packets sitting in receiver's buffer and
2901  * packets lost by network.
2902  *
2903  * Besides that it does CWND reduction, when packet loss is detected
2904  * and changes state of machine.
2905  *
2906  * It does _not_ decide what to send, it is made in function
2907  * tcp_xmit_retransmit_queue().
2908  */
2909 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2910 {
2911         struct inet_connection_sock *icsk = inet_csk(sk);
2912         struct tcp_sock *tp = tcp_sk(sk);
2913         int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2914         int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2915                                     (tcp_fackets_out(tp) > tp->reordering));
2916         int fast_rexmit = 0, mib_idx;
2917
2918         if (WARN_ON(!tp->packets_out && tp->sacked_out))
2919                 tp->sacked_out = 0;
2920         if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2921                 tp->fackets_out = 0;
2922
2923         /* Now state machine starts.
2924          * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2925         if (flag & FLAG_ECE)
2926                 tp->prior_ssthresh = 0;
2927
2928         /* B. In all the states check for reneging SACKs. */
2929         if (tcp_check_sack_reneging(sk, flag))
2930                 return;
2931
2932         /* C. Process data loss notification, provided it is valid. */
2933         if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2934             before(tp->snd_una, tp->high_seq) &&
2935             icsk->icsk_ca_state != TCP_CA_Open &&
2936             tp->fackets_out > tp->reordering) {
2937                 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2938                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2939         }
2940
2941         /* D. Check consistency of the current state. */
2942         tcp_verify_left_out(tp);
2943
2944         /* E. Check state exit conditions. State can be terminated
2945          *    when high_seq is ACKed. */
2946         if (icsk->icsk_ca_state == TCP_CA_Open) {
2947                 WARN_ON(tp->retrans_out != 0);
2948                 tp->retrans_stamp = 0;
2949         } else if (!before(tp->snd_una, tp->high_seq)) {
2950                 switch (icsk->icsk_ca_state) {
2951                 case TCP_CA_Loss:
2952                         icsk->icsk_retransmits = 0;
2953                         if (tcp_try_undo_recovery(sk))
2954                                 return;
2955                         break;
2956
2957                 case TCP_CA_CWR:
2958                         /* CWR is to be held something *above* high_seq
2959                          * is ACKed for CWR bit to reach receiver. */
2960                         if (tp->snd_una != tp->high_seq) {
2961                                 tcp_complete_cwr(sk);
2962                                 tcp_set_ca_state(sk, TCP_CA_Open);
2963                         }
2964                         break;
2965
2966                 case TCP_CA_Disorder:
2967                         tcp_try_undo_dsack(sk);
2968                         if (!tp->undo_marker ||
2969                             /* For SACK case do not Open to allow to undo
2970                              * catching for all duplicate ACKs. */
2971                             tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2972                                 tp->undo_marker = 0;
2973                                 tcp_set_ca_state(sk, TCP_CA_Open);
2974                         }
2975                         break;
2976
2977                 case TCP_CA_Recovery:
2978                         if (tcp_is_reno(tp))
2979                                 tcp_reset_reno_sack(tp);
2980                         if (tcp_try_undo_recovery(sk))
2981                                 return;
2982                         tcp_complete_cwr(sk);
2983                         break;
2984                 }
2985         }
2986
2987         /* F. Process state. */
2988         switch (icsk->icsk_ca_state) {
2989         case TCP_CA_Recovery:
2990                 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2991                         if (tcp_is_reno(tp) && is_dupack)
2992                                 tcp_add_reno_sack(sk);
2993                 } else
2994                         do_lost = tcp_try_undo_partial(sk, pkts_acked);
2995                 break;
2996         case TCP_CA_Loss:
2997                 if (flag & FLAG_DATA_ACKED)
2998                         icsk->icsk_retransmits = 0;
2999                 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3000                         tcp_reset_reno_sack(tp);
3001                 if (!tcp_try_undo_loss(sk)) {
3002                         tcp_moderate_cwnd(tp);
3003                         tcp_xmit_retransmit_queue(sk);
3004                         return;
3005                 }
3006                 if (icsk->icsk_ca_state != TCP_CA_Open)
3007                         return;
3008                 /* Loss is undone; fall through to processing in Open state. */
3009         default:
3010                 if (tcp_is_reno(tp)) {
3011                         if (flag & FLAG_SND_UNA_ADVANCED)
3012                                 tcp_reset_reno_sack(tp);
3013                         if (is_dupack)
3014                                 tcp_add_reno_sack(sk);
3015                 }
3016
3017                 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3018                         tcp_try_undo_dsack(sk);
3019
3020                 if (!tcp_time_to_recover(sk)) {
3021                         tcp_try_to_open(sk, flag);
3022                         return;
3023                 }
3024
3025                 /* MTU probe failure: don't reduce cwnd */
3026                 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3027                     icsk->icsk_mtup.probe_size &&
3028                     tp->snd_una == tp->mtu_probe.probe_seq_start) {
3029                         tcp_mtup_probe_failed(sk);
3030                         /* Restores the reduction we did in tcp_mtup_probe() */
3031                         tp->snd_cwnd++;
3032                         tcp_simple_retransmit(sk);
3033                         return;
3034                 }
3035
3036                 /* Otherwise enter Recovery state */
3037
3038                 if (tcp_is_reno(tp))
3039                         mib_idx = LINUX_MIB_TCPRENORECOVERY;
3040                 else
3041                         mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3042
3043                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3044
3045                 tp->high_seq = tp->snd_nxt;
3046                 tp->prior_ssthresh = 0;
3047                 tp->undo_marker = tp->snd_una;
3048                 tp->undo_retrans = tp->retrans_out;
3049
3050                 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3051                         if (!(flag & FLAG_ECE))
3052                                 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3053                         tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3054                         TCP_ECN_queue_cwr(tp);
3055                 }
3056
3057                 tp->bytes_acked = 0;
3058                 tp->snd_cwnd_cnt = 0;
3059                 tcp_set_ca_state(sk, TCP_CA_Recovery);
3060                 fast_rexmit = 1;
3061         }
3062
3063         if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3064                 tcp_update_scoreboard(sk, fast_rexmit);
3065         tcp_cwnd_down(sk, flag);
3066         tcp_xmit_retransmit_queue(sk);
3067 }
3068
3069 static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3070 {
3071         tcp_rtt_estimator(sk, seq_rtt);
3072         tcp_set_rto(sk);
3073         inet_csk(sk)->icsk_backoff = 0;
3074 }
3075
3076 /* Read draft-ietf-tcplw-high-performance before mucking
3077  * with this code. (Supersedes RFC1323)
3078  */
3079 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3080 {
3081         /* RTTM Rule: A TSecr value received in a segment is used to
3082          * update the averaged RTT measurement only if the segment
3083          * acknowledges some new data, i.e., only if it advances the
3084          * left edge of the send window.
3085          *
3086          * See draft-ietf-tcplw-high-performance-00, section 3.3.
3087          * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3088          *
3089          * Changed: reset backoff as soon as we see the first valid sample.
3090          * If we do not, we get strongly overestimated rto. With timestamps
3091          * samples are accepted even from very old segments: f.e., when rtt=1
3092          * increases to 8, we retransmit 5 times and after 8 seconds delayed
3093          * answer arrives rto becomes 120 seconds! If at least one of segments
3094          * in window is lost... Voila.                          --ANK (010210)
3095          */
3096         struct tcp_sock *tp = tcp_sk(sk);
3097
3098         tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3099 }
3100
3101 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3102 {
3103         /* We don't have a timestamp. Can only use
3104          * packets that are not retransmitted to determine
3105          * rtt estimates. Also, we must not reset the
3106          * backoff for rto until we get a non-retransmitted
3107          * packet. This allows us to deal with a situation
3108          * where the network delay has increased suddenly.
3109          * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3110          */
3111
3112         if (flag & FLAG_RETRANS_DATA_ACKED)
3113                 return;
3114
3115         tcp_valid_rtt_meas(sk, seq_rtt);
3116 }
3117
3118 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3119                                       const s32 seq_rtt)
3120 {
3121         const struct tcp_sock *tp = tcp_sk(sk);
3122         /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3123         if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3124                 tcp_ack_saw_tstamp(sk, flag);
3125         else if (seq_rtt >= 0)
3126                 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3127 }
3128
3129 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3130 {
3131         const struct inet_connection_sock *icsk = inet_csk(sk);
3132         icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3133         tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3134 }
3135
3136 /* Restart timer after forward progress on connection.
3137  * RFC2988 recommends to restart timer to now+rto.
3138  */
3139 static void tcp_rearm_rto(struct sock *sk)
3140 {
3141         struct tcp_sock *tp = tcp_sk(sk);
3142
3143         if (!tp->packets_out) {
3144                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3145         } else {
3146                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3147                                           inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3148         }
3149 }
3150
3151 /* If we get here, the whole TSO packet has not been acked. */
3152 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3153 {
3154         struct tcp_sock *tp = tcp_sk(sk);
3155         u32 packets_acked;
3156
3157         BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3158
3159         packets_acked = tcp_skb_pcount(skb);
3160         if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3161                 return 0;
3162         packets_acked -= tcp_skb_pcount(skb);
3163
3164         if (packets_acked) {
3165                 BUG_ON(tcp_skb_pcount(skb) == 0);
3166                 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3167         }
3168
3169         return packets_acked;
3170 }
3171
3172 /* Remove acknowledged frames from the retransmission queue. If our packet
3173  * is before the ack sequence we can discard it as it's confirmed to have
3174  * arrived at the other end.
3175  */
3176 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3177                                u32 prior_snd_una)
3178 {
3179         struct tcp_sock *tp = tcp_sk(sk);
3180         const struct inet_connection_sock *icsk = inet_csk(sk);
3181         struct sk_buff *skb;
3182         u32 now = tcp_time_stamp;
3183         int fully_acked = 1;
3184         int flag = 0;
3185         u32 pkts_acked = 0;
3186         u32 reord = tp->packets_out;
3187         u32 prior_sacked = tp->sacked_out;
3188         s32 seq_rtt = -1;
3189         s32 ca_seq_rtt = -1;
3190         ktime_t last_ackt = net_invalid_timestamp();
3191
3192         while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3193                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3194                 u32 acked_pcount;
3195                 u8 sacked = scb->sacked;
3196
3197                 /* Determine how many packets and what bytes were acked, tso and else */
3198                 if (after(scb->end_seq, tp->snd_una)) {
3199                         if (tcp_skb_pcount(skb) == 1 ||
3200                             !after(tp->snd_una, scb->seq))
3201                                 break;
3202
3203                         acked_pcount = tcp_tso_acked(sk, skb);
3204                         if (!acked_pcount)
3205                                 break;
3206
3207                         fully_acked = 0;
3208                 } else {
3209                         acked_pcount = tcp_skb_pcount(skb);
3210                 }
3211
3212                 if (sacked & TCPCB_RETRANS) {
3213                         if (sacked & TCPCB_SACKED_RETRANS)
3214                                 tp->retrans_out -= acked_pcount;
3215                         flag |= FLAG_RETRANS_DATA_ACKED;
3216                         ca_seq_rtt = -1;
3217                         seq_rtt = -1;
3218                         if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3219                                 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3220                 } else {
3221                         ca_seq_rtt = now - scb->when;
3222                         last_ackt = skb->tstamp;
3223                         if (seq_rtt < 0) {
3224                                 seq_rtt = ca_seq_rtt;
3225                         }
3226                         if (!(sacked & TCPCB_SACKED_ACKED))
3227                                 reord = min(pkts_acked, reord);
3228                 }
3229
3230                 if (sacked & TCPCB_SACKED_ACKED)
3231                         tp->sacked_out -= acked_pcount;
3232                 if (sacked & TCPCB_LOST)
3233                         tp->lost_out -= acked_pcount;
3234
3235                 tp->packets_out -= acked_pcount;
3236                 pkts_acked += acked_pcount;
3237
3238                 /* Initial outgoing SYN's get put onto the write_queue
3239                  * just like anything else we transmit.  It is not
3240                  * true data, and if we misinform our callers that
3241                  * this ACK acks real data, we will erroneously exit
3242                  * connection startup slow start one packet too
3243                  * quickly.  This is severely frowned upon behavior.
3244                  */
3245                 if (!(scb->flags & TCPCB_FLAG_SYN)) {
3246                         flag |= FLAG_DATA_ACKED;
3247                 } else {
3248                         flag |= FLAG_SYN_ACKED;
3249                         tp->retrans_stamp = 0;
3250                 }
3251
3252                 if (!fully_acked)
3253                         break;
3254
3255                 tcp_unlink_write_queue(skb, sk);
3256                 sk_wmem_free_skb(sk, skb);
3257                 tp->scoreboard_skb_hint = NULL;
3258                 if (skb == tp->retransmit_skb_hint)
3259                         tp->retransmit_skb_hint = NULL;
3260                 if (skb == tp->lost_skb_hint)
3261                         tp->lost_skb_hint = NULL;
3262         }
3263
3264         if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3265                 tp->snd_up = tp->snd_una;
3266
3267         if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3268                 flag |= FLAG_SACK_RENEGING;
3269
3270         if (flag & FLAG_ACKED) {
3271                 const struct tcp_congestion_ops *ca_ops
3272                         = inet_csk(sk)->icsk_ca_ops;
3273
3274                 if (unlikely(icsk->icsk_mtup.probe_size &&
3275                              !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3276                         tcp_mtup_probe_success(sk);
3277                 }
3278
3279                 tcp_ack_update_rtt(sk, flag, seq_rtt);
3280                 tcp_rearm_rto(sk);
3281
3282                 if (tcp_is_reno(tp)) {
3283                         tcp_remove_reno_sacks(sk, pkts_acked);
3284                 } else {
3285                         int delta;
3286
3287                         /* Non-retransmitted hole got filled? That's reordering */
3288                         if (reord < prior_fackets)
3289                                 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3290
3291                         delta = tcp_is_fack(tp) ? pkts_acked :
3292                                                   prior_sacked - tp->sacked_out;
3293                         tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3294                 }
3295
3296                 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3297
3298                 if (ca_ops->pkts_acked) {
3299                         s32 rtt_us = -1;
3300
3301                         /* Is the ACK triggering packet unambiguous? */
3302                         if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3303                                 /* High resolution needed and available? */
3304                                 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3305                                     !ktime_equal(last_ackt,
3306                                                  net_invalid_timestamp()))
3307                                         rtt_us = ktime_us_delta(ktime_get_real(),
3308                                                                 last_ackt);
3309                                 else if (ca_seq_rtt > 0)
3310                                         rtt_us = jiffies_to_usecs(ca_seq_rtt);
3311                         }
3312
3313                         ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3314                 }
3315         }
3316
3317 #if FASTRETRANS_DEBUG > 0
3318         WARN_ON((int)tp->sacked_out < 0);
3319         WARN_ON((int)tp->lost_out < 0);
3320         WARN_ON((int)tp->retrans_out < 0);
3321         if (!tp->packets_out && tcp_is_sack(tp)) {
3322                 icsk = inet_csk(sk);
3323                 if (tp->lost_out) {
3324                         printk(KERN_DEBUG "Leak l=%u %d\n",
3325                                tp->lost_out, icsk->icsk_ca_state);
3326                         tp->lost_out = 0;
3327                 }
3328                 if (tp->sacked_out) {
3329                         printk(KERN_DEBUG "Leak s=%u %d\n",
3330                                tp->sacked_out, icsk->icsk_ca_state);
3331                         tp->sacked_out = 0;
3332                 }
3333                 if (tp->retrans_out) {
3334                         printk(KERN_DEBUG "Leak r=%u %d\n",
3335                                tp->retrans_out, icsk->icsk_ca_state);
3336                         tp->retrans_out = 0;
3337                 }
3338         }
3339 #endif
3340         return flag;
3341 }
3342
3343 static void tcp_ack_probe(struct sock *sk)
3344 {
3345         const struct tcp_sock *tp = tcp_sk(sk);
3346         struct inet_connection_sock *icsk = inet_csk(sk);
3347
3348         /* Was it a usable window open? */
3349
3350         if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3351                 icsk->icsk_backoff = 0;
3352                 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3353                 /* Socket must be waked up by subsequent tcp_data_snd_check().
3354                  * This function is not for random using!
3355                  */
3356         } else {
3357                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3358                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3359                                           TCP_RTO_MAX);
3360         }
3361 }
3362
3363 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3364 {
3365         return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3366                 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3367 }
3368
3369 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3370 {
3371         const struct tcp_sock *tp = tcp_sk(sk);
3372         return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3373                 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3374 }
3375
3376 /* Check that window update is acceptable.
3377  * The function assumes that snd_una<=ack<=snd_next.
3378  */
3379 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3380                                         const u32 ack, const u32 ack_seq,
3381                                         const u32 nwin)
3382 {
3383         return (after(ack, tp->snd_una) ||
3384                 after(ack_seq, tp->snd_wl1) ||
3385                 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3386 }
3387
3388 /* Update our send window.
3389  *
3390  * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3391  * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3392  */
3393 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3394                                  u32 ack_seq)
3395 {
3396         struct tcp_sock *tp = tcp_sk(sk);
3397         int flag = 0;
3398         u32 nwin = ntohs(tcp_hdr(skb)->window);
3399
3400         if (likely(!tcp_hdr(skb)->syn))
3401                 nwin <<= tp->rx_opt.snd_wscale;
3402
3403         if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3404                 flag |= FLAG_WIN_UPDATE;
3405                 tcp_update_wl(tp, ack_seq);
3406
3407                 if (tp->snd_wnd != nwin) {
3408                         tp->snd_wnd = nwin;
3409
3410                         /* Note, it is the only place, where
3411                          * fast path is recovered for sending TCP.
3412                          */
3413                         tp->pred_flags = 0;
3414                         tcp_fast_path_check(sk);
3415
3416                         if (nwin > tp->max_window) {
3417                                 tp->max_window = nwin;
3418                                 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3419                         }
3420                 }
3421         }
3422
3423         tp->snd_una = ack;
3424
3425         return flag;
3426 }
3427
3428 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3429  * continue in congestion avoidance.
3430  */
3431 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3432 {
3433         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3434         tp->snd_cwnd_cnt = 0;
3435         tp->bytes_acked = 0;
3436         TCP_ECN_queue_cwr(tp);
3437         tcp_moderate_cwnd(tp);
3438 }
3439
3440 /* A conservative spurious RTO response algorithm: reduce cwnd using
3441  * rate halving and continue in congestion avoidance.
3442  */
3443 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3444 {
3445         tcp_enter_cwr(sk, 0);
3446 }
3447
3448 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3449 {
3450         if (flag & FLAG_ECE)
3451                 tcp_ratehalving_spur_to_response(sk);
3452         else
3453                 tcp_undo_cwr(sk, 1);
3454 }
3455
3456 /* F-RTO spurious RTO detection algorithm (RFC4138)
3457  *
3458  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3459  * comments). State (ACK number) is kept in frto_counter. When ACK advances
3460  * window (but not to or beyond highest sequence sent before RTO):
3461  *   On First ACK,  send two new segments out.
3462  *   On Second ACK, RTO was likely spurious. Do spurious response (response
3463  *                  algorithm is not part of the F-RTO detection algorithm
3464  *                  given in RFC4138 but can be selected separately).
3465  * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3466  * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3467  * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3468  * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3469  *
3470  * Rationale: if the RTO was spurious, new ACKs should arrive from the
3471  * original window even after we transmit two new data segments.
3472  *
3473  * SACK version:
3474  *   on first step, wait until first cumulative ACK arrives, then move to
3475  *   the second step. In second step, the next ACK decides.
3476  *
3477  * F-RTO is implemented (mainly) in four functions:
3478  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
3479  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3480  *     called when tcp_use_frto() showed green light
3481  *   - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3482  *   - tcp_enter_frto_loss() is called if there is not enough evidence
3483  *     to prove that the RTO is indeed spurious. It transfers the control
3484  *     from F-RTO to the conventional RTO recovery
3485  */
3486 static int tcp_process_frto(struct sock *sk, int flag)
3487 {
3488         struct tcp_sock *tp = tcp_sk(sk);
3489
3490         tcp_verify_left_out(tp);
3491
3492         /* Duplicate the behavior from Loss state (fastretrans_alert) */
3493         if (flag & FLAG_DATA_ACKED)
3494                 inet_csk(sk)->icsk_retransmits = 0;
3495
3496         if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3497             ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3498                 tp->undo_marker = 0;
3499
3500         if (!before(tp->snd_una, tp->frto_highmark)) {
3501                 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3502                 return 1;
3503         }
3504
3505         if (!tcp_is_sackfrto(tp)) {
3506                 /* RFC4138 shortcoming in step 2; should also have case c):
3507                  * ACK isn't duplicate nor advances window, e.g., opposite dir
3508                  * data, winupdate
3509                  */
3510                 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3511                         return 1;
3512
3513                 if (!(flag & FLAG_DATA_ACKED)) {
3514                         tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3515                                             flag);
3516                         return 1;
3517                 }
3518         } else {
3519                 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3520                         /* Prevent sending of new data. */
3521                         tp->snd_cwnd = min(tp->snd_cwnd,
3522                                            tcp_packets_in_flight(tp));
3523                         return 1;
3524                 }
3525
3526                 if ((tp->frto_counter >= 2) &&
3527                     (!(flag & FLAG_FORWARD_PROGRESS) ||
3528                      ((flag & FLAG_DATA_SACKED) &&
3529                       !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3530                         /* RFC4138 shortcoming (see comment above) */
3531                         if (!(flag & FLAG_FORWARD_PROGRESS) &&
3532                             (flag & FLAG_NOT_DUP))
3533                                 return 1;
3534
3535                         tcp_enter_frto_loss(sk, 3, flag);
3536                         return 1;
3537                 }
3538         }
3539
3540         if (tp->frto_counter == 1) {
3541                 /* tcp_may_send_now needs to see updated state */
3542                 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3543                 tp->frto_counter = 2;
3544
3545                 if (!tcp_may_send_now(sk))
3546                         tcp_enter_frto_loss(sk, 2, flag);
3547
3548                 return 1;
3549         } else {
3550                 switch (sysctl_tcp_frto_response) {
3551                 case 2:
3552                         tcp_undo_spur_to_response(sk, flag);
3553                         break;
3554                 case 1:
3555                         tcp_conservative_spur_to_response(tp);
3556                         break;
3557                 default:
3558                         tcp_ratehalving_spur_to_response(sk);
3559                         break;
3560                 }
3561                 tp->frto_counter = 0;
3562                 tp->undo_marker = 0;
3563                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3564         }
3565         return 0;
3566 }
3567
3568 /* This routine deals with incoming acks, but not outgoing ones. */
3569 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3570 {
3571         struct inet_connection_sock *icsk = inet_csk(sk);
3572         struct tcp_sock *tp = tcp_sk(sk);
3573         u32 prior_snd_una = tp->snd_una;
3574         u32 ack_seq = TCP_SKB_CB(skb)->seq;
3575         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3576         u32 prior_in_flight;
3577         u32 prior_fackets;
3578         int prior_packets;
3579         int frto_cwnd = 0;
3580
3581         /* If the ack is older than previous acks
3582          * then we can probably ignore it.
3583          */
3584         if (before(ack, prior_snd_una))
3585                 goto old_ack;
3586
3587         /* If the ack includes data we haven't sent yet, discard
3588          * this segment (RFC793 Section 3.9).
3589          */
3590         if (after(ack, tp->snd_nxt))
3591                 goto invalid_ack;
3592
3593         if (after(ack, prior_snd_una))
3594                 flag |= FLAG_SND_UNA_ADVANCED;
3595
3596         if (sysctl_tcp_abc) {
3597                 if (icsk->icsk_ca_state < TCP_CA_CWR)
3598                         tp->bytes_acked += ack - prior_snd_una;
3599                 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3600                         /* we assume just one segment left network */
3601                         tp->bytes_acked += min(ack - prior_snd_una,
3602                                                tp->mss_cache);
3603         }
3604
3605         prior_fackets = tp->fackets_out;
3606         prior_in_flight = tcp_packets_in_flight(tp);
3607
3608         if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3609                 /* Window is constant, pure forward advance.
3610                  * No more checks are required.
3611                  * Note, we use the fact that SND.UNA>=SND.WL2.
3612                  */
3613                 tcp_update_wl(tp, ack_seq);
3614                 tp->snd_una = ack;
3615                 flag |= FLAG_WIN_UPDATE;
3616
3617                 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3618
3619                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3620         } else {
3621                 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3622                         flag |= FLAG_DATA;
3623                 else
3624                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3625
3626                 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3627
3628                 if (TCP_SKB_CB(skb)->sacked)
3629                         flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3630
3631                 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3632                         flag |= FLAG_ECE;
3633
3634                 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3635         }
3636
3637         /* We passed data and got it acked, remove any soft error
3638          * log. Something worked...
3639          */
3640         sk->sk_err_soft = 0;
3641         icsk->icsk_probes_out = 0;
3642         tp->rcv_tstamp = tcp_time_stamp;
3643         prior_packets = tp->packets_out;
3644         if (!prior_packets)
3645                 goto no_queue;
3646
3647         /* See if we can take anything off of the retransmit queue. */
3648         flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3649
3650         if (tp->frto_counter)
3651                 frto_cwnd = tcp_process_frto(sk, flag);
3652         /* Guarantee sacktag reordering detection against wrap-arounds */
3653         if (before(tp->frto_highmark, tp->snd_una))
3654                 tp->frto_highmark = 0;
3655
3656         if (tcp_ack_is_dubious(sk, flag)) {
3657                 /* Advance CWND, if state allows this. */
3658                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3659                     tcp_may_raise_cwnd(sk, flag))
3660                         tcp_cong_avoid(sk, ack, prior_in_flight);
3661                 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3662                                       flag);
3663         } else {
3664                 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3665                         tcp_cong_avoid(sk, ack, prior_in_flight);
3666         }
3667
3668         if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3669                 dst_confirm(sk->sk_dst_cache);
3670
3671         return 1;
3672
3673 no_queue:
3674         /* If this ack opens up a zero window, clear backoff.  It was
3675          * being used to time the probes, and is probably far higher than
3676          * it needs to be for normal retransmission.
3677          */
3678         if (tcp_send_head(sk))
3679                 tcp_ack_probe(sk);
3680         return 1;
3681
3682 invalid_ack:
3683         SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3684         return -1;
3685
3686 old_ack:
3687         if (TCP_SKB_CB(skb)->sacked) {
3688                 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3689                 if (icsk->icsk_ca_state == TCP_CA_Open)
3690                         tcp_try_keep_open(sk);
3691         }
3692
3693         SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3694         return 0;
3695 }
3696
3697 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3698  * But, this can also be called on packets in the established flow when
3699  * the fast version below fails.
3700  */
3701 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3702                        int estab)
3703 {
3704         unsigned char *ptr;
3705         struct tcphdr *th = tcp_hdr(skb);
3706         int length = (th->doff * 4) - sizeof(struct tcphdr);
3707
3708         ptr = (unsigned char *)(th + 1);
3709         opt_rx->saw_tstamp = 0;
3710
3711         while (length > 0) {
3712                 int opcode = *ptr++;
3713                 int opsize;
3714
3715                 switch (opcode) {
3716                 case TCPOPT_EOL:
3717                         return;
3718                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
3719                         length--;
3720                         continue;
3721                 default:
3722                         opsize = *ptr++;
3723                         if (opsize < 2) /* "silly options" */
3724                                 return;
3725                         if (opsize > length)
3726                                 return; /* don't parse partial options */
3727                         switch (opcode) {
3728                         case TCPOPT_MSS:
3729                                 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3730                                         u16 in_mss = get_unaligned_be16(ptr);
3731                                         if (in_mss) {
3732                                                 if (opt_rx->user_mss &&
3733                                                     opt_rx->user_mss < in_mss)
3734                                                         in_mss = opt_rx->user_mss;
3735                                                 opt_rx->mss_clamp = in_mss;
3736                                         }
3737                                 }
3738                                 break;
3739                         case TCPOPT_WINDOW:
3740                                 if (opsize == TCPOLEN_WINDOW && th->syn &&
3741                                     !estab && sysctl_tcp_window_scaling) {
3742                                         __u8 snd_wscale = *(__u8 *)ptr;
3743                                         opt_rx->wscale_ok = 1;
3744                                         if (snd_wscale > 14) {
3745                                                 if (net_ratelimit())
3746                                                         printk(KERN_INFO "tcp_parse_options: Illegal window "
3747                                                                "scaling value %d >14 received.\n",
3748                                                                snd_wscale);
3749                                                 snd_wscale = 14;
3750                                         }
3751                                         opt_rx->snd_wscale = snd_wscale;
3752                                 }
3753                                 break;
3754                         case TCPOPT_TIMESTAMP:
3755                                 if ((opsize == TCPOLEN_TIMESTAMP) &&
3756                                     ((estab && opt_rx->tstamp_ok) ||
3757                                      (!estab && sysctl_tcp_timestamps))) {
3758                                         opt_rx->saw_tstamp = 1;
3759                                         opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3760                                         opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3761                                 }
3762                                 break;
3763                         case TCPOPT_SACK_PERM:
3764                                 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3765                                     !estab && sysctl_tcp_sack) {
3766                                         opt_rx->sack_ok = 1;
3767                                         tcp_sack_reset(opt_rx);
3768                                 }
3769                                 break;
3770
3771                         case TCPOPT_SACK:
3772                                 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3773                                    !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3774                                    opt_rx->sack_ok) {
3775                                         TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3776                                 }
3777                                 break;
3778 #ifdef CONFIG_TCP_MD5SIG
3779                         case TCPOPT_MD5SIG:
3780                                 /*
3781                                  * The MD5 Hash has already been
3782                                  * checked (see tcp_v{4,6}_do_rcv()).
3783                                  */
3784                                 break;
3785 #endif
3786                         }
3787
3788                         ptr += opsize-2;
3789                         length -= opsize;
3790                 }
3791         }
3792 }
3793
3794 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3795 {
3796         __be32 *ptr = (__be32 *)(th + 1);
3797
3798         if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3799                           | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3800                 tp->rx_opt.saw_tstamp = 1;
3801                 ++ptr;
3802                 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3803                 ++ptr;
3804                 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3805                 return 1;
3806         }
3807         return 0;
3808 }
3809
3810 /* Fast parse options. This hopes to only see timestamps.
3811  * If it is wrong it falls back on tcp_parse_options().
3812  */
3813 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3814                                   struct tcp_sock *tp)
3815 {
3816         if (th->doff == sizeof(struct tcphdr) >> 2) {
3817                 tp->rx_opt.saw_tstamp = 0;
3818                 return 0;
3819         } else if (tp->rx_opt.tstamp_ok &&
3820                    th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3821                 if (tcp_parse_aligned_timestamp(tp, th))
3822                         return 1;
3823         }
3824         tcp_parse_options(skb, &tp->rx_opt, 1);
3825         return 1;
3826 }
3827
3828 #ifdef CONFIG_TCP_MD5SIG
3829 /*
3830  * Parse MD5 Signature option
3831  */
3832 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3833 {
3834         int length = (th->doff << 2) - sizeof (*th);
3835         u8 *ptr = (u8*)(th + 1);
3836
3837         /* If the TCP option is too short, we can short cut */
3838         if (length < TCPOLEN_MD5SIG)
3839                 return NULL;
3840
3841         while (length > 0) {
3842                 int opcode = *ptr++;
3843                 int opsize;
3844
3845                 switch(opcode) {
3846                 case TCPOPT_EOL:
3847                         return NULL;
3848                 case TCPOPT_NOP:
3849                         length--;
3850                         continue;
3851                 default:
3852                         opsize = *ptr++;
3853                         if (opsize < 2 || opsize > length)
3854                                 return NULL;
3855                         if (opcode == TCPOPT_MD5SIG)
3856                                 return ptr;
3857                 }
3858                 ptr += opsize - 2;
3859                 length -= opsize;
3860         }
3861         return NULL;
3862 }
3863 #endif
3864
3865 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3866 {
3867         tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3868         tp->rx_opt.ts_recent_stamp = get_seconds();
3869 }
3870
3871 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3872 {
3873         if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3874                 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3875                  * extra check below makes sure this can only happen
3876                  * for pure ACK frames.  -DaveM
3877                  *
3878                  * Not only, also it occurs for expired timestamps.
3879                  */
3880
3881                 if (tcp_paws_check(&tp->rx_opt, 0))
3882                         tcp_store_ts_recent(tp);
3883         }
3884 }
3885
3886 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3887  *
3888  * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3889  * it can pass through stack. So, the following predicate verifies that
3890  * this segment is not used for anything but congestion avoidance or
3891  * fast retransmit. Moreover, we even are able to eliminate most of such
3892  * second order effects, if we apply some small "replay" window (~RTO)
3893  * to timestamp space.
3894  *
3895  * All these measures still do not guarantee that we reject wrapped ACKs
3896  * on networks with high bandwidth, when sequence space is recycled fastly,
3897  * but it guarantees that such events will be very rare and do not affect
3898  * connection seriously. This doesn't look nice, but alas, PAWS is really
3899  * buggy extension.
3900  *
3901  * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3902  * states that events when retransmit arrives after original data are rare.
3903  * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3904  * the biggest problem on large power networks even with minor reordering.
3905  * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3906  * up to bandwidth of 18Gigabit/sec. 8) ]
3907  */
3908
3909 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3910 {
3911         struct tcp_sock *tp = tcp_sk(sk);
3912         struct tcphdr *th = tcp_hdr(skb);
3913         u32 seq = TCP_SKB_CB(skb)->seq;
3914         u32 ack = TCP_SKB_CB(skb)->ack_seq;
3915
3916         return (/* 1. Pure ACK with correct sequence number. */
3917                 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3918
3919                 /* 2. ... and duplicate ACK. */
3920                 ack == tp->snd_una &&
3921
3922                 /* 3. ... and does not update window. */
3923                 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3924
3925                 /* 4. ... and sits in replay window. */
3926                 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3927 }
3928
3929 static inline int tcp_paws_discard(const struct sock *sk,
3930                                    const struct sk_buff *skb)
3931 {
3932         const struct tcp_sock *tp = tcp_sk(sk);
3933
3934         return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3935                !tcp_disordered_ack(sk, skb);
3936 }
3937
3938 /* Check segment sequence number for validity.
3939  *
3940  * Segment controls are considered valid, if the segment
3941  * fits to the window after truncation to the window. Acceptability
3942  * of data (and SYN, FIN, of course) is checked separately.
3943  * See tcp_data_queue(), for example.
3944  *
3945  * Also, controls (RST is main one) are accepted using RCV.WUP instead
3946  * of RCV.NXT. Peer still did not advance his SND.UNA when we
3947  * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3948  * (borrowed from freebsd)
3949  */
3950
3951 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3952 {
3953         return  !before(end_seq, tp->rcv_wup) &&
3954                 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3955 }
3956
3957 /* When we get a reset we do this. */
3958 static void tcp_reset(struct sock *sk)
3959 {
3960         /* We want the right error as BSD sees it (and indeed as we do). */
3961         switch (sk->sk_state) {
3962         case TCP_SYN_SENT:
3963                 sk->sk_err = ECONNREFUSED;
3964                 break;
3965         case TCP_CLOSE_WAIT:
3966                 sk->sk_err = EPIPE;
3967                 break;
3968         case TCP_CLOSE:
3969                 return;
3970         default:
3971                 sk->sk_err = ECONNRESET;
3972         }
3973
3974         if (!sock_flag(sk, SOCK_DEAD))
3975                 sk->sk_error_report(sk);
3976
3977         tcp_done(sk);
3978 }
3979
3980 /*
3981  *      Process the FIN bit. This now behaves as it is supposed to work
3982  *      and the FIN takes effect when it is validly part of sequence
3983  *      space. Not before when we get holes.
3984  *
3985  *      If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3986  *      (and thence onto LAST-ACK and finally, CLOSE, we never enter
3987  *      TIME-WAIT)
3988  *
3989  *      If we are in FINWAIT-1, a received FIN indicates simultaneous
3990  *      close and we go into CLOSING (and later onto TIME-WAIT)
3991  *
3992  *      If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3993  */
3994 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3995 {
3996         struct tcp_sock *tp = tcp_sk(sk);
3997
3998         inet_csk_schedule_ack(sk);
3999
4000         sk->sk_shutdown |= RCV_SHUTDOWN;
4001         sock_set_flag(sk, SOCK_DONE);
4002
4003         switch (sk->sk_state) {
4004         case TCP_SYN_RECV:
4005         case TCP_ESTABLISHED:
4006                 /* Move to CLOSE_WAIT */
4007                 tcp_set_state(sk, TCP_CLOSE_WAIT);
4008                 inet_csk(sk)->icsk_ack.pingpong = 1;
4009                 break;
4010
4011         case TCP_CLOSE_WAIT:
4012         case TCP_CLOSING:
4013                 /* Received a retransmission of the FIN, do
4014                  * nothing.
4015                  */
4016                 break;
4017         case TCP_LAST_ACK:
4018                 /* RFC793: Remain in the LAST-ACK state. */
4019                 break;
4020
4021         case TCP_FIN_WAIT1:
4022                 /* This case occurs when a simultaneous close
4023                  * happens, we must ack the received FIN and
4024                  * enter the CLOSING state.
4025                  */
4026                 tcp_send_ack(sk);
4027                 tcp_set_state(sk, TCP_CLOSING);
4028                 break;
4029         case TCP_FIN_WAIT2:
4030                 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4031                 tcp_send_ack(sk);
4032                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4033                 break;
4034         default:
4035                 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4036                  * cases we should never reach this piece of code.
4037                  */
4038                 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4039                        __func__, sk->sk_state);
4040                 break;
4041         }
4042
4043         /* It _is_ possible, that we have something out-of-order _after_ FIN.
4044          * Probably, we should reset in this case. For now drop them.
4045          */
4046         __skb_queue_purge(&tp->out_of_order_queue);
4047         if (tcp_is_sack(tp))
4048                 tcp_sack_reset(&tp->rx_opt);
4049         sk_mem_reclaim(sk);
4050
4051         if (!sock_flag(sk, SOCK_DEAD)) {
4052                 sk->sk_state_change(sk);
4053
4054                 /* Do not send POLL_HUP for half duplex close. */
4055                 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4056                     sk->sk_state == TCP_CLOSE)
4057                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4058                 else
4059                         sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4060         }
4061 }
4062
4063 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4064                                   u32 end_seq)
4065 {
4066         if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4067                 if (before(seq, sp->start_seq))
4068                         sp->start_seq = seq;
4069                 if (after(end_seq, sp->end_seq))
4070                         sp->end_seq = end_seq;
4071                 return 1;
4072         }
4073         return 0;
4074 }
4075
4076 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4077 {
4078         struct tcp_sock *tp = tcp_sk(sk);
4079
4080         if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4081                 int mib_idx;
4082
4083                 if (before(seq, tp->rcv_nxt))
4084                         mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4085                 else
4086                         mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4087
4088                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4089
4090                 tp->rx_opt.dsack = 1;
4091                 tp->duplicate_sack[0].start_seq = seq;
4092                 tp->duplicate_sack[0].end_seq = end_seq;
4093         }
4094 }
4095
4096 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4097 {
4098         struct tcp_sock *tp = tcp_sk(sk);
4099
4100         if (!tp->rx_opt.dsack)
4101                 tcp_dsack_set(sk, seq, end_seq);
4102         else
4103                 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4104 }
4105
4106 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4107 {
4108         struct tcp_sock *tp = tcp_sk(sk);
4109
4110         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4111             before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4112                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4113                 tcp_enter_quickack_mode(sk);
4114
4115                 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4116                         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4117
4118                         if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4119                                 end_seq = tp->rcv_nxt;
4120                         tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4121                 }
4122         }
4123
4124         tcp_send_ack(sk);
4125 }
4126
4127 /* These routines update the SACK block as out-of-order packets arrive or
4128  * in-order packets close up the sequence space.
4129  */
4130 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4131 {
4132         int this_sack;
4133         struct tcp_sack_block *sp = &tp->selective_acks[0];
4134         struct tcp_sack_block *swalk = sp + 1;
4135
4136         /* See if the recent change to the first SACK eats into
4137          * or hits the sequence space of other SACK blocks, if so coalesce.
4138          */
4139         for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4140                 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4141                         int i;
4142
4143                         /* Zap SWALK, by moving every further SACK up by one slot.
4144                          * Decrease num_sacks.
4145                          */
4146                         tp->rx_opt.num_sacks--;
4147                         for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4148                                 sp[i] = sp[i + 1];
4149                         continue;
4150                 }
4151                 this_sack++, swalk++;
4152         }
4153 }
4154
4155 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4156 {
4157         struct tcp_sock *tp = tcp_sk(sk);
4158         struct tcp_sack_block *sp = &tp->selective_acks[0];
4159         int cur_sacks = tp->rx_opt.num_sacks;
4160         int this_sack;
4161
4162         if (!cur_sacks)
4163                 goto new_sack;
4164
4165         for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4166                 if (tcp_sack_extend(sp, seq, end_seq)) {
4167                         /* Rotate this_sack to the first one. */
4168                         for (; this_sack > 0; this_sack--, sp--)
4169                                 swap(*sp, *(sp - 1));
4170                         if (cur_sacks > 1)
4171                                 tcp_sack_maybe_coalesce(tp);
4172                         return;
4173                 }
4174         }
4175
4176         /* Could not find an adjacent existing SACK, build a new one,
4177          * put it at the front, and shift everyone else down.  We
4178          * always know there is at least one SACK present already here.
4179          *
4180          * If the sack array is full, forget about the last one.
4181          */
4182         if (this_sack >= TCP_NUM_SACKS) {
4183                 this_sack--;
4184                 tp->rx_opt.num_sacks--;
4185                 sp--;
4186         }
4187         for (; this_sack > 0; this_sack--, sp--)
4188                 *sp = *(sp - 1);
4189
4190 new_sack:
4191         /* Build the new head SACK, and we're done. */
4192         sp->start_seq = seq;
4193         sp->end_seq = end_seq;
4194         tp->rx_opt.num_sacks++;
4195 }
4196
4197 /* RCV.NXT advances, some SACKs should be eaten. */
4198
4199 static void tcp_sack_remove(struct tcp_sock *tp)
4200 {
4201         struct tcp_sack_block *sp = &tp->selective_acks[0];
4202         int num_sacks = tp->rx_opt.num_sacks;
4203         int this_sack;
4204
4205         /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4206         if (skb_queue_empty(&tp->out_of_order_queue)) {
4207                 tp->rx_opt.num_sacks = 0;
4208                 return;
4209         }
4210
4211         for (this_sack = 0; this_sack < num_sacks;) {
4212                 /* Check if the start of the sack is covered by RCV.NXT. */
4213                 if (!before(tp->rcv_nxt, sp->start_seq)) {
4214                         int i;
4215
4216                         /* RCV.NXT must cover all the block! */
4217                         WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4218
4219                         /* Zap this SACK, by moving forward any other SACKS. */
4220                         for (i=this_sack+1; i < num_sacks; i++)
4221                                 tp->selective_acks[i-1] = tp->selective_acks[i];
4222                         num_sacks--;
4223                         continue;
4224                 }
4225                 this_sack++;
4226                 sp++;
4227         }
4228         tp->rx_opt.num_sacks = num_sacks;
4229 }
4230
4231 /* This one checks to see if we can put data from the
4232  * out_of_order queue into the receive_queue.
4233  */
4234 static void tcp_ofo_queue(struct sock *sk)
4235 {
4236         struct tcp_sock *tp = tcp_sk(sk);
4237         __u32 dsack_high = tp->rcv_nxt;
4238         struct sk_buff *skb;
4239
4240         while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4241                 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4242                         break;
4243
4244                 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4245                         __u32 dsack = dsack_high;
4246                         if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4247                                 dsack_high = TCP_SKB_CB(skb)->end_seq;
4248                         tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4249                 }
4250
4251                 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4252                         SOCK_DEBUG(sk, "ofo packet was already received \n");
4253                         __skb_unlink(skb, &tp->out_of_order_queue);
4254                         __kfree_skb(skb);
4255                         continue;
4256                 }
4257                 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4258                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4259                            TCP_SKB_CB(skb)->end_seq);
4260
4261                 __skb_unlink(skb, &tp->out_of_order_queue);
4262                 __skb_queue_tail(&sk->sk_receive_queue, skb);
4263                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4264                 if (tcp_hdr(skb)->fin)
4265                         tcp_fin(skb, sk, tcp_hdr(skb));
4266         }
4267 }
4268
4269 static int tcp_prune_ofo_queue(struct sock *sk);
4270 static int tcp_prune_queue(struct sock *sk);
4271
4272 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4273 {
4274         if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4275             !sk_rmem_schedule(sk, size)) {
4276
4277                 if (tcp_prune_queue(sk) < 0)
4278                         return -1;
4279
4280                 if (!sk_rmem_schedule(sk, size)) {
4281                         if (!tcp_prune_ofo_queue(sk))
4282                                 return -1;
4283
4284                         if (!sk_rmem_schedule(sk, size))
4285                                 return -1;
4286                 }
4287         }
4288         return 0;
4289 }
4290
4291 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4292 {
4293         struct tcphdr *th = tcp_hdr(skb);
4294         struct tcp_sock *tp = tcp_sk(sk);
4295         int eaten = -1;
4296
4297         if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4298                 goto drop;
4299
4300         __skb_pull(skb, th->doff * 4);
4301
4302         TCP_ECN_accept_cwr(tp, skb);
4303
4304         tp->rx_opt.dsack = 0;
4305
4306         /*  Queue data for delivery to the user.
4307          *  Packets in sequence go to the receive queue.
4308          *  Out of sequence packets to the out_of_order_queue.
4309          */
4310         if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4311                 if (tcp_receive_window(tp) == 0)
4312                         goto out_of_window;
4313
4314                 /* Ok. In sequence. In window. */
4315                 if (tp->ucopy.task == current &&
4316                     tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4317                     sock_owned_by_user(sk) && !tp->urg_data) {
4318                         int chunk = min_t(unsigned int, skb->len,
4319                                           tp->ucopy.len);
4320
4321                         __set_current_state(TASK_RUNNING);
4322
4323                         local_bh_enable();
4324                         if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4325                                 tp->ucopy.len -= chunk;
4326                                 tp->copied_seq += chunk;
4327                                 eaten = (chunk == skb->len && !th->fin);
4328                                 tcp_rcv_space_adjust(sk);
4329                         }
4330                         local_bh_disable();
4331                 }
4332
4333                 if (eaten <= 0) {
4334 queue_and_out:
4335                         if (eaten < 0 &&
4336                             tcp_try_rmem_schedule(sk, skb->truesize))
4337                                 goto drop;
4338
4339                         skb_set_owner_r(skb, sk);
4340                         __skb_queue_tail(&sk->sk_receive_queue, skb);
4341                 }
4342                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4343                 if (skb->len)
4344                         tcp_event_data_recv(sk, skb);
4345                 if (th->fin)
4346                         tcp_fin(skb, sk, th);
4347
4348                 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4349                         tcp_ofo_queue(sk);
4350
4351                         /* RFC2581. 4.2. SHOULD send immediate ACK, when
4352                          * gap in queue is filled.
4353                          */
4354                         if (skb_queue_empty(&tp->out_of_order_queue))
4355                                 inet_csk(sk)->icsk_ack.pingpong = 0;
4356                 }
4357
4358                 if (tp->rx_opt.num_sacks)
4359                         tcp_sack_remove(tp);
4360
4361                 tcp_fast_path_check(sk);
4362
4363                 if (eaten > 0)
4364                         __kfree_skb(skb);
4365                 else if (!sock_flag(sk, SOCK_DEAD))
4366                         sk->sk_data_ready(sk, 0);
4367                 return;
4368         }
4369
4370         if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4371                 /* A retransmit, 2nd most common case.  Force an immediate ack. */
4372                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4373                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4374
4375 out_of_window:
4376                 tcp_enter_quickack_mode(sk);
4377                 inet_csk_schedule_ack(sk);
4378 drop:
4379                 __kfree_skb(skb);
4380                 return;
4381         }
4382
4383         /* Out of window. F.e. zero window probe. */
4384         if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4385                 goto out_of_window;
4386
4387         tcp_enter_quickack_mode(sk);
4388
4389         if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4390                 /* Partial packet, seq < rcv_next < end_seq */
4391                 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4392                            tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4393                            TCP_SKB_CB(skb)->end_seq);
4394
4395                 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4396
4397                 /* If window is closed, drop tail of packet. But after
4398                  * remembering D-SACK for its head made in previous line.
4399                  */
4400                 if (!tcp_receive_window(tp))
4401                         goto out_of_window;
4402                 goto queue_and_out;
4403         }
4404
4405         TCP_ECN_check_ce(tp, skb);
4406
4407         if (tcp_try_rmem_schedule(sk, skb->truesize))
4408                 goto drop;
4409
4410         /* Disable header prediction. */
4411         tp->pred_flags = 0;
4412         inet_csk_schedule_ack(sk);
4413
4414         SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4415                    tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4416
4417         skb_set_owner_r(skb, sk);
4418
4419         if (!skb_peek(&tp->out_of_order_queue)) {
4420                 /* Initial out of order segment, build 1 SACK. */
4421                 if (tcp_is_sack(tp)) {
4422                         tp->rx_opt.num_sacks = 1;
4423                         tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4424                         tp->selective_acks[0].end_seq =
4425                                                 TCP_SKB_CB(skb)->end_seq;
4426                 }
4427                 __skb_queue_head(&tp->out_of_order_queue, skb);
4428         } else {
4429                 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4430                 u32 seq = TCP_SKB_CB(skb)->seq;
4431                 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4432
4433                 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4434                         __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4435
4436                         if (!tp->rx_opt.num_sacks ||
4437                             tp->selective_acks[0].end_seq != seq)
4438                                 goto add_sack;
4439
4440                         /* Common case: data arrive in order after hole. */
4441                         tp->selective_acks[0].end_seq = end_seq;
4442                         return;
4443                 }
4444
4445                 /* Find place to insert this segment. */
4446                 do {
4447                         if (!after(TCP_SKB_CB(skb1)->seq, seq))
4448                                 break;
4449                 } while ((skb1 = skb1->prev) !=
4450                          (struct sk_buff *)&tp->out_of_order_queue);
4451
4452                 /* Do skb overlap to previous one? */
4453                 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4454                     before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4455                         if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4456                                 /* All the bits are present. Drop. */
4457                                 __kfree_skb(skb);
4458                                 tcp_dsack_set(sk, seq, end_seq);
4459                                 goto add_sack;
4460                         }
4461                         if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4462                                 /* Partial overlap. */
4463                                 tcp_dsack_set(sk, seq,
4464                                               TCP_SKB_CB(skb1)->end_seq);
4465                         } else {
4466                                 skb1 = skb1->prev;
4467                         }
4468                 }
4469                 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4470
4471                 /* And clean segments covered by new one as whole. */
4472                 while ((skb1 = skb->next) !=
4473                        (struct sk_buff *)&tp->out_of_order_queue &&
4474                        after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4475                         if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4476                                 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4477                                                  end_seq);
4478                                 break;
4479                         }
4480                         __skb_unlink(skb1, &tp->out_of_order_queue);
4481                         tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4482                                          TCP_SKB_CB(skb1)->end_seq);
4483                         __kfree_skb(skb1);
4484                 }
4485
4486 add_sack:
4487                 if (tcp_is_sack(tp))
4488                         tcp_sack_new_ofo_skb(sk, seq, end_seq);
4489         }
4490 }
4491
4492 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4493                                         struct sk_buff_head *list)
4494 {
4495         struct sk_buff *next = skb->next;
4496
4497         __skb_unlink(skb, list);
4498         __kfree_skb(skb);
4499         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4500
4501         return next;
4502 }
4503
4504 /* Collapse contiguous sequence of skbs head..tail with
4505  * sequence numbers start..end.
4506  * Segments with FIN/SYN are not collapsed (only because this
4507  * simplifies code)
4508  */
4509 static void
4510 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4511              struct sk_buff *head, struct sk_buff *tail,
4512              u32 start, u32 end)
4513 {
4514         struct sk_buff *skb;
4515
4516         /* First, check that queue is collapsible and find
4517          * the point where collapsing can be useful. */
4518         for (skb = head; skb != tail;) {
4519                 /* No new bits? It is possible on ofo queue. */
4520                 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4521                         skb = tcp_collapse_one(sk, skb, list);
4522                         continue;
4523                 }
4524
4525                 /* The first skb to collapse is:
4526                  * - not SYN/FIN and
4527                  * - bloated or contains data before "start" or
4528                  *   overlaps to the next one.
4529                  */
4530                 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4531                     (tcp_win_from_space(skb->truesize) > skb->len ||
4532                      before(TCP_SKB_CB(skb)->seq, start) ||
4533                      (skb->next != tail &&
4534                       TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4535                         break;
4536
4537                 /* Decided to skip this, advance start seq. */
4538                 start = TCP_SKB_CB(skb)->end_seq;
4539                 skb = skb->next;
4540         }
4541         if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4542                 return;
4543
4544         while (before(start, end)) {
4545                 struct sk_buff *nskb;
4546                 unsigned int header = skb_headroom(skb);
4547                 int copy = SKB_MAX_ORDER(header, 0);
4548
4549                 /* Too big header? This can happen with IPv6. */
4550                 if (copy < 0)
4551                         return;
4552                 if (end - start < copy)
4553                         copy = end - start;
4554                 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4555                 if (!nskb)
4556                         return;
4557
4558                 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4559                 skb_set_network_header(nskb, (skb_network_header(skb) -
4560                                               skb->head));
4561                 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4562                                                 skb->head));
4563                 skb_reserve(nskb, header);
4564                 memcpy(nskb->head, skb->head, header);
4565                 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4566                 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4567                 __skb_queue_before(list, skb, nskb);
4568                 skb_set_owner_r(nskb, sk);
4569
4570                 /* Copy data, releasing collapsed skbs. */
4571                 while (copy > 0) {
4572                         int offset = start - TCP_SKB_CB(skb)->seq;
4573                         int size = TCP_SKB_CB(skb)->end_seq - start;
4574
4575                         BUG_ON(offset < 0);
4576                         if (size > 0) {
4577                                 size = min(copy, size);
4578                                 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4579                                         BUG();
4580                                 TCP_SKB_CB(nskb)->end_seq += size;
4581                                 copy -= size;
4582                                 start += size;
4583                         }
4584                         if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4585                                 skb = tcp_collapse_one(sk, skb, list);
4586                                 if (skb == tail ||
4587                                     tcp_hdr(skb)->syn ||
4588                                     tcp_hdr(skb)->fin)
4589                                         return;
4590                         }
4591                 }
4592         }
4593 }
4594
4595 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4596  * and tcp_collapse() them until all the queue is collapsed.
4597  */
4598 static void tcp_collapse_ofo_queue(struct sock *sk)
4599 {
4600         struct tcp_sock *tp = tcp_sk(sk);
4601         struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4602         struct sk_buff *head;
4603         u32 start, end;
4604
4605         if (skb == NULL)
4606                 return;
4607
4608         start = TCP_SKB_CB(skb)->seq;
4609         end = TCP_SKB_CB(skb)->end_seq;
4610         head = skb;
4611
4612         for (;;) {
4613                 skb = skb->next;
4614
4615                 /* Segment is terminated when we see gap or when
4616                  * we are at the end of all the queue. */
4617                 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4618                     after(TCP_SKB_CB(skb)->seq, end) ||
4619                     before(TCP_SKB_CB(skb)->end_seq, start)) {
4620                         tcp_collapse(sk, &tp->out_of_order_queue,
4621                                      head, skb, start, end);
4622                         head = skb;
4623                         if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4624                                 break;
4625                         /* Start new segment */
4626                         start = TCP_SKB_CB(skb)->seq;
4627                         end = TCP_SKB_CB(skb)->end_seq;
4628                 } else {
4629                         if (before(TCP_SKB_CB(skb)->seq, start))
4630                                 start = TCP_SKB_CB(skb)->seq;
4631                         if (after(TCP_SKB_CB(skb)->end_seq, end))
4632                                 end = TCP_SKB_CB(skb)->end_seq;
4633                 }
4634         }
4635 }
4636
4637 /*
4638  * Purge the out-of-order queue.
4639  * Return true if queue was pruned.
4640  */
4641 static int tcp_prune_ofo_queue(struct sock *sk)
4642 {
4643         struct tcp_sock *tp = tcp_sk(sk);
4644         int res = 0;
4645
4646         if (!skb_queue_empty(&tp->out_of_order_queue)) {
4647                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4648                 __skb_queue_purge(&tp->out_of_order_queue);
4649
4650                 /* Reset SACK state.  A conforming SACK implementation will
4651                  * do the same at a timeout based retransmit.  When a connection
4652                  * is in a sad state like this, we care only about integrity
4653                  * of the connection not performance.
4654                  */
4655                 if (tp->rx_opt.sack_ok)
4656                         tcp_sack_reset(&tp->rx_opt);
4657                 sk_mem_reclaim(sk);
4658                 res = 1;
4659         }
4660         return res;
4661 }
4662
4663 /* Reduce allocated memory if we can, trying to get
4664  * the socket within its memory limits again.
4665  *
4666  * Return less than zero if we should start dropping frames
4667  * until the socket owning process reads some of the data
4668  * to stabilize the situation.
4669  */
4670 static int tcp_prune_queue(struct sock *sk)
4671 {
4672         struct tcp_sock *tp = tcp_sk(sk);
4673
4674         SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4675
4676         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4677
4678         if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4679                 tcp_clamp_window(sk);
4680         else if (tcp_memory_pressure)
4681                 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4682
4683         tcp_collapse_ofo_queue(sk);
4684         tcp_collapse(sk, &sk->sk_receive_queue,
4685                      sk->sk_receive_queue.next,
4686                      (struct sk_buff *)&sk->sk_receive_queue,
4687                      tp->copied_seq, tp->rcv_nxt);
4688         sk_mem_reclaim(sk);
4689
4690         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4691                 return 0;
4692
4693         /* Collapsing did not help, destructive actions follow.
4694          * This must not ever occur. */
4695
4696         tcp_prune_ofo_queue(sk);
4697
4698         if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4699                 return 0;
4700
4701         /* If we are really being abused, tell the caller to silently
4702          * drop receive data on the floor.  It will get retransmitted
4703          * and hopefully then we'll have sufficient space.
4704          */
4705         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4706
4707         /* Massive buffer overcommit. */
4708         tp->pred_flags = 0;
4709         return -1;
4710 }
4711
4712 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4713  * As additional protections, we do not touch cwnd in retransmission phases,
4714  * and if application hit its sndbuf limit recently.
4715  */
4716 void tcp_cwnd_application_limited(struct sock *sk)
4717 {
4718         struct tcp_sock *tp = tcp_sk(sk);
4719
4720         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4721             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4722                 /* Limited by application or receiver window. */
4723                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4724                 u32 win_used = max(tp->snd_cwnd_used, init_win);
4725                 if (win_used < tp->snd_cwnd) {
4726                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
4727                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4728                 }
4729                 tp->snd_cwnd_used = 0;
4730         }
4731         tp->snd_cwnd_stamp = tcp_time_stamp;
4732 }
4733
4734 static int tcp_should_expand_sndbuf(struct sock *sk)
4735 {
4736         struct tcp_sock *tp = tcp_sk(sk);
4737
4738         /* If the user specified a specific send buffer setting, do
4739          * not modify it.
4740          */
4741         if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4742                 return 0;
4743
4744         /* If we are under global TCP memory pressure, do not expand.  */
4745         if (tcp_memory_pressure)
4746                 return 0;
4747
4748         /* If we are under soft global TCP memory pressure, do not expand.  */
4749         if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4750                 return 0;
4751
4752         /* If we filled the congestion window, do not expand.  */
4753         if (tp->packets_out >= tp->snd_cwnd)
4754                 return 0;
4755
4756         return 1;
4757 }
4758
4759 /* When incoming ACK allowed to free some skb from write_queue,
4760  * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4761  * on the exit from tcp input handler.
4762  *
4763  * PROBLEM: sndbuf expansion does not work well with largesend.
4764  */
4765 static void tcp_new_space(struct sock *sk)
4766 {
4767         struct tcp_sock *tp = tcp_sk(sk);
4768
4769         if (tcp_should_expand_sndbuf(sk)) {
4770                 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4771                         MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4772                 int demanded = max_t(unsigned int, tp->snd_cwnd,
4773                                      tp->reordering + 1);
4774                 sndmem *= 2 * demanded;
4775                 if (sndmem > sk->sk_sndbuf)
4776                         sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4777                 tp->snd_cwnd_stamp = tcp_time_stamp;
4778         }
4779
4780         sk->sk_write_space(sk);
4781 }
4782
4783 static void tcp_check_space(struct sock *sk)
4784 {
4785         if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4786                 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4787                 if (sk->sk_socket &&
4788                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4789                         tcp_new_space(sk);
4790         }
4791 }
4792
4793 static inline void tcp_data_snd_check(struct sock *sk)
4794 {
4795         tcp_push_pending_frames(sk);
4796         tcp_check_space(sk);
4797 }
4798
4799 /*
4800  * Check if sending an ack is needed.
4801  */
4802 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4803 {
4804         struct tcp_sock *tp = tcp_sk(sk);
4805
4806             /* More than one full frame received... */
4807         if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4808              /* ... and right edge of window advances far enough.
4809               * (tcp_recvmsg() will send ACK otherwise). Or...
4810               */
4811              && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4812             /* We ACK each frame or... */
4813             tcp_in_quickack_mode(sk) ||
4814             /* We have out of order data. */
4815             (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4816                 /* Then ack it now */
4817                 tcp_send_ack(sk);
4818         } else {
4819                 /* Else, send delayed ack. */
4820                 tcp_send_delayed_ack(sk);
4821         }
4822 }
4823
4824 static inline void tcp_ack_snd_check(struct sock *sk)
4825 {
4826         if (!inet_csk_ack_scheduled(sk)) {
4827                 /* We sent a data segment already. */
4828                 return;
4829         }
4830         __tcp_ack_snd_check(sk, 1);
4831 }
4832
4833 /*
4834  *      This routine is only called when we have urgent data
4835  *      signaled. Its the 'slow' part of tcp_urg. It could be
4836  *      moved inline now as tcp_urg is only called from one
4837  *      place. We handle URGent data wrong. We have to - as
4838  *      BSD still doesn't use the correction from RFC961.
4839  *      For 1003.1g we should support a new option TCP_STDURG to permit
4840  *      either form (or just set the sysctl tcp_stdurg).
4841  */
4842
4843 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4844 {
4845         struct tcp_sock *tp = tcp_sk(sk);
4846         u32 ptr = ntohs(th->urg_ptr);
4847
4848         if (ptr && !sysctl_tcp_stdurg)
4849                 ptr--;
4850         ptr += ntohl(th->seq);
4851
4852         /* Ignore urgent data that we've already seen and read. */
4853         if (after(tp->copied_seq, ptr))
4854                 return;
4855
4856         /* Do not replay urg ptr.
4857          *
4858          * NOTE: interesting situation not covered by specs.
4859          * Misbehaving sender may send urg ptr, pointing to segment,
4860          * which we already have in ofo queue. We are not able to fetch
4861          * such data and will stay in TCP_URG_NOTYET until will be eaten
4862          * by recvmsg(). Seems, we are not obliged to handle such wicked
4863          * situations. But it is worth to think about possibility of some
4864          * DoSes using some hypothetical application level deadlock.
4865          */
4866         if (before(ptr, tp->rcv_nxt))
4867                 return;
4868
4869         /* Do we already have a newer (or duplicate) urgent pointer? */
4870         if (tp->urg_data && !after(ptr, tp->urg_seq))
4871                 return;
4872
4873         /* Tell the world about our new urgent pointer. */
4874         sk_send_sigurg(sk);
4875
4876         /* We may be adding urgent data when the last byte read was
4877          * urgent. To do this requires some care. We cannot just ignore
4878          * tp->copied_seq since we would read the last urgent byte again
4879          * as data, nor can we alter copied_seq until this data arrives
4880          * or we break the semantics of SIOCATMARK (and thus sockatmark())
4881          *
4882          * NOTE. Double Dutch. Rendering to plain English: author of comment
4883          * above did something sort of  send("A", MSG_OOB); send("B", MSG_OOB);
4884          * and expect that both A and B disappear from stream. This is _wrong_.
4885          * Though this happens in BSD with high probability, this is occasional.
4886          * Any application relying on this is buggy. Note also, that fix "works"
4887          * only in this artificial test. Insert some normal data between A and B and we will
4888          * decline of BSD again. Verdict: it is better to remove to trap
4889          * buggy users.
4890          */
4891         if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4892             !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4893                 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4894                 tp->copied_seq++;
4895                 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4896                         __skb_unlink(skb, &sk->sk_receive_queue);
4897                         __kfree_skb(skb);
4898                 }
4899         }
4900
4901         tp->urg_data = TCP_URG_NOTYET;
4902         tp->urg_seq = ptr;
4903
4904         /* Disable header prediction. */
4905         tp->pred_flags = 0;
4906 }
4907
4908 /* This is the 'fast' part of urgent handling. */
4909 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4910 {
4911         struct tcp_sock *tp = tcp_sk(sk);
4912
4913         /* Check if we get a new urgent pointer - normally not. */
4914         if (th->urg)
4915                 tcp_check_urg(sk, th);
4916
4917         /* Do we wait for any urgent data? - normally not... */
4918         if (tp->urg_data == TCP_URG_NOTYET) {
4919                 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4920                           th->syn;
4921
4922                 /* Is the urgent pointer pointing into this packet? */
4923                 if (ptr < skb->len) {
4924                         u8 tmp;
4925                         if (skb_copy_bits(skb, ptr, &tmp, 1))
4926                                 BUG();
4927                         tp->urg_data = TCP_URG_VALID | tmp;
4928                         if (!sock_flag(sk, SOCK_DEAD))
4929                                 sk->sk_data_ready(sk, 0);
4930                 }
4931         }
4932 }
4933
4934 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4935 {
4936         struct tcp_sock *tp = tcp_sk(sk);
4937         int chunk = skb->len - hlen;
4938         int err;
4939
4940         local_bh_enable();
4941         if (skb_csum_unnecessary(skb))
4942                 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4943         else
4944                 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4945                                                        tp->ucopy.iov);
4946
4947         if (!err) {
4948                 tp->ucopy.len -= chunk;
4949                 tp->copied_seq += chunk;
4950                 tcp_rcv_space_adjust(sk);
4951         }
4952
4953         local_bh_disable();
4954         return err;
4955 }
4956
4957 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4958                                             struct sk_buff *skb)
4959 {
4960         __sum16 result;
4961
4962         if (sock_owned_by_user(sk)) {
4963                 local_bh_enable();
4964                 result = __tcp_checksum_complete(skb);
4965                 local_bh_disable();
4966         } else {
4967                 result = __tcp_checksum_complete(skb);
4968         }
4969         return result;
4970 }
4971
4972 static inline int tcp_checksum_complete_user(struct sock *sk,
4973                                              struct sk_buff *skb)
4974 {
4975         return !skb_csum_unnecessary(skb) &&
4976                __tcp_checksum_complete_user(sk, skb);
4977 }
4978
4979 #ifdef CONFIG_NET_DMA
4980 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4981                                   int hlen)
4982 {
4983         struct tcp_sock *tp = tcp_sk(sk);
4984         int chunk = skb->len - hlen;
4985         int dma_cookie;
4986         int copied_early = 0;
4987
4988         if (tp->ucopy.wakeup)
4989                 return 0;
4990
4991         if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4992                 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
4993
4994         if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4995
4996                 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4997                                                          skb, hlen,
4998                                                          tp->ucopy.iov, chunk,
4999                                                          tp->ucopy.pinned_list);
5000
5001                 if (dma_cookie < 0)
5002                         goto out;
5003
5004                 tp->ucopy.dma_cookie = dma_cookie;
5005                 copied_early = 1;
5006
5007                 tp->ucopy.len -= chunk;
5008                 tp->copied_seq += chunk;
5009                 tcp_rcv_space_adjust(sk);
5010
5011                 if ((tp->ucopy.len == 0) ||
5012                     (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5013                     (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5014                         tp->ucopy.wakeup = 1;
5015                         sk->sk_data_ready(sk, 0);
5016                 }
5017         } else if (chunk > 0) {
5018                 tp->ucopy.wakeup = 1;
5019                 sk->sk_data_ready(sk, 0);
5020         }
5021 out:
5022         return copied_early;
5023 }
5024 #endif /* CONFIG_NET_DMA */
5025
5026 /* Does PAWS and seqno based validation of an incoming segment, flags will
5027  * play significant role here.
5028  */
5029 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5030                               struct tcphdr *th, int syn_inerr)
5031 {
5032         struct tcp_sock *tp = tcp_sk(sk);
5033
5034         /* RFC1323: H1. Apply PAWS check first. */
5035         if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5036             tcp_paws_discard(sk, skb)) {
5037                 if (!th->rst) {
5038                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5039                         tcp_send_dupack(sk, skb);
5040                         goto discard;
5041                 }
5042                 /* Reset is accepted even if it did not pass PAWS. */
5043         }
5044
5045         /* Step 1: check sequence number */
5046         if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5047                 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5048                  * (RST) segments are validated by checking their SEQ-fields."
5049                  * And page 69: "If an incoming segment is not acceptable,
5050                  * an acknowledgment should be sent in reply (unless the RST
5051                  * bit is set, if so drop the segment and return)".
5052                  */
5053                 if (!th->rst)
5054                         tcp_send_dupack(sk, skb);
5055                 goto discard;
5056         }
5057
5058         /* Step 2: check RST bit */
5059         if (th->rst) {
5060                 tcp_reset(sk);
5061                 goto discard;
5062         }
5063
5064         /* ts_recent update must be made after we are sure that the packet
5065          * is in window.
5066          */
5067         tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5068
5069         /* step 3: check security and precedence [ignored] */
5070
5071         /* step 4: Check for a SYN in window. */
5072         if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5073                 if (syn_inerr)
5074                         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5075                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5076                 tcp_reset(sk);
5077                 return -1;
5078         }
5079
5080         return 1;
5081
5082 discard:
5083         __kfree_skb(skb);
5084         return 0;
5085 }
5086
5087 /*
5088  *      TCP receive function for the ESTABLISHED state.
5089  *
5090  *      It is split into a fast path and a slow path. The fast path is
5091  *      disabled when:
5092  *      - A zero window was announced from us - zero window probing
5093  *        is only handled properly in the slow path.
5094  *      - Out of order segments arrived.
5095  *      - Urgent data is expected.
5096  *      - There is no buffer space left
5097  *      - Unexpected TCP flags/window values/header lengths are received
5098  *        (detected by checking the TCP header against pred_flags)
5099  *      - Data is sent in both directions. Fast path only supports pure senders
5100  *        or pure receivers (this means either the sequence number or the ack
5101  *        value must stay constant)
5102  *      - Unexpected TCP option.
5103  *
5104  *      When these conditions are not satisfied it drops into a standard
5105  *      receive procedure patterned after RFC793 to handle all cases.
5106  *      The first three cases are guaranteed by proper pred_flags setting,
5107  *      the rest is checked inline. Fast processing is turned on in
5108  *      tcp_data_queue when everything is OK.
5109  */
5110 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5111                         struct tcphdr *th, unsigned len)
5112 {
5113         struct tcp_sock *tp = tcp_sk(sk);
5114         int res;
5115
5116         /*
5117          *      Header prediction.
5118          *      The code loosely follows the one in the famous
5119          *      "30 instruction TCP receive" Van Jacobson mail.
5120          *
5121          *      Van's trick is to deposit buffers into socket queue
5122          *      on a device interrupt, to call tcp_recv function
5123          *      on the receive process context and checksum and copy
5124          *      the buffer to user space. smart...
5125          *
5126          *      Our current scheme is not silly either but we take the
5127          *      extra cost of the net_bh soft interrupt processing...
5128          *      We do checksum and copy also but from device to kernel.
5129          */
5130
5131         tp->rx_opt.saw_tstamp = 0;
5132
5133         /*      pred_flags is 0xS?10 << 16 + snd_wnd
5134          *      if header_prediction is to be made
5135          *      'S' will always be tp->tcp_header_len >> 2
5136          *      '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5137          *  turn it off (when there are holes in the receive
5138          *       space for instance)
5139          *      PSH flag is ignored.
5140          */
5141
5142         if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5143             TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5144             !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5145                 int tcp_header_len = tp->tcp_header_len;
5146
5147                 /* Timestamp header prediction: tcp_header_len
5148                  * is automatically equal to th->doff*4 due to pred_flags
5149                  * match.
5150                  */
5151
5152                 /* Check timestamp */
5153                 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5154                         /* No? Slow path! */
5155                         if (!tcp_parse_aligned_timestamp(tp, th))
5156                                 goto slow_path;
5157
5158                         /* If PAWS failed, check it more carefully in slow path */
5159                         if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5160                                 goto slow_path;
5161
5162                         /* DO NOT update ts_recent here, if checksum fails
5163                          * and timestamp was corrupted part, it will result
5164                          * in a hung connection since we will drop all
5165                          * future packets due to the PAWS test.
5166                          */
5167                 }
5168
5169                 if (len <= tcp_header_len) {
5170                         /* Bulk data transfer: sender */
5171                         if (len == tcp_header_len) {
5172                                 /* Predicted packet is in window by definition.
5173                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5174                                  * Hence, check seq<=rcv_wup reduces to:
5175                                  */
5176                                 if (tcp_header_len ==
5177                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5178                                     tp->rcv_nxt == tp->rcv_wup)
5179                                         tcp_store_ts_recent(tp);
5180
5181                                 /* We know that such packets are checksummed
5182                                  * on entry.
5183                                  */
5184                                 tcp_ack(sk, skb, 0);
5185                                 __kfree_skb(skb);
5186                                 tcp_data_snd_check(sk);
5187                                 return 0;
5188                         } else { /* Header too small */
5189                                 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5190                                 goto discard;
5191                         }
5192                 } else {
5193                         int eaten = 0;
5194                         int copied_early = 0;
5195
5196                         if (tp->copied_seq == tp->rcv_nxt &&
5197                             len - tcp_header_len <= tp->ucopy.len) {
5198 #ifdef CONFIG_NET_DMA
5199                                 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5200                                         copied_early = 1;
5201                                         eaten = 1;
5202                                 }
5203 #endif
5204                                 if (tp->ucopy.task == current &&
5205                                     sock_owned_by_user(sk) && !copied_early) {
5206                                         __set_current_state(TASK_RUNNING);
5207
5208                                         if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5209                                                 eaten = 1;
5210                                 }
5211                                 if (eaten) {
5212                                         /* Predicted packet is in window by definition.
5213                                          * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5214                                          * Hence, check seq<=rcv_wup reduces to:
5215                                          */
5216                                         if (tcp_header_len ==
5217                                             (sizeof(struct tcphdr) +
5218                                              TCPOLEN_TSTAMP_ALIGNED) &&
5219                                             tp->rcv_nxt == tp->rcv_wup)
5220                                                 tcp_store_ts_recent(tp);
5221
5222                                         tcp_rcv_rtt_measure_ts(sk, skb);
5223
5224                                         __skb_pull(skb, tcp_header_len);
5225                                         tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5226                                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5227                                 }
5228                                 if (copied_early)
5229                                         tcp_cleanup_rbuf(sk, skb->len);
5230                         }
5231                         if (!eaten) {
5232                                 if (tcp_checksum_complete_user(sk, skb))
5233                                         goto csum_error;
5234
5235                                 /* Predicted packet is in window by definition.
5236                                  * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5237                                  * Hence, check seq<=rcv_wup reduces to:
5238                                  */
5239                                 if (tcp_header_len ==
5240                                     (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5241                                     tp->rcv_nxt == tp->rcv_wup)
5242                                         tcp_store_ts_recent(tp);
5243
5244                                 tcp_rcv_rtt_measure_ts(sk, skb);
5245
5246                                 if ((int)skb->truesize > sk->sk_forward_alloc)
5247                                         goto step5;
5248
5249                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5250
5251                                 /* Bulk data transfer: receiver */
5252                                 __skb_pull(skb, tcp_header_len);
5253                                 __skb_queue_tail(&sk->sk_receive_queue, skb);
5254                                 skb_set_owner_r(skb, sk);
5255                                 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5256                         }
5257
5258                         tcp_event_data_recv(sk, skb);
5259
5260                         if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5261                                 /* Well, only one small jumplet in fast path... */
5262                                 tcp_ack(sk, skb, FLAG_DATA);
5263                                 tcp_data_snd_check(sk);
5264                                 if (!inet_csk_ack_scheduled(sk))
5265                                         goto no_ack;
5266                         }
5267
5268                         if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5269                                 __tcp_ack_snd_check(sk, 0);
5270 no_ack:
5271 #ifdef CONFIG_NET_DMA
5272                         if (copied_early)
5273                                 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5274                         else
5275 #endif
5276                         if (eaten)
5277                                 __kfree_skb(skb);
5278                         else
5279                                 sk->sk_data_ready(sk, 0);
5280                         return 0;
5281                 }
5282         }
5283
5284 slow_path:
5285         if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5286                 goto csum_error;
5287
5288         /*
5289          *      Standard slow path.
5290          */
5291
5292         res = tcp_validate_incoming(sk, skb, th, 1);
5293         if (res <= 0)
5294                 return -res;
5295
5296 step5:
5297         if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5298                 goto discard;
5299
5300         tcp_rcv_rtt_measure_ts(sk, skb);
5301
5302         /* Process urgent data. */
5303         tcp_urg(sk, skb, th);
5304
5305         /* step 7: process the segment text */
5306         tcp_data_queue(sk, skb);
5307
5308         tcp_data_snd_check(sk);
5309         tcp_ack_snd_check(sk);
5310         return 0;
5311
5312 csum_error:
5313         TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5314
5315 discard:
5316         __kfree_skb(skb);
5317         return 0;
5318 }
5319
5320 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5321                                          struct tcphdr *th, unsigned len)
5322 {
5323         struct tcp_sock *tp = tcp_sk(sk);
5324         struct inet_connection_sock *icsk = inet_csk(sk);
5325         int saved_clamp = tp->rx_opt.mss_clamp;
5326
5327         tcp_parse_options(skb, &tp->rx_opt, 0);
5328
5329         if (th->ack) {
5330                 /* rfc793:
5331                  * "If the state is SYN-SENT then
5332                  *    first check the ACK bit
5333                  *      If the ACK bit is set
5334                  *        If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5335                  *        a reset (unless the RST bit is set, if so drop
5336                  *        the segment and return)"
5337                  *
5338                  *  We do not send data with SYN, so that RFC-correct
5339                  *  test reduces to:
5340                  */
5341                 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5342                         goto reset_and_undo;
5343
5344                 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5345                     !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5346                              tcp_time_stamp)) {
5347                         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5348                         goto reset_and_undo;
5349                 }
5350
5351                 /* Now ACK is acceptable.
5352                  *
5353                  * "If the RST bit is set
5354                  *    If the ACK was acceptable then signal the user "error:
5355                  *    connection reset", drop the segment, enter CLOSED state,
5356                  *    delete TCB, and return."
5357                  */
5358
5359                 if (th->rst) {
5360                         tcp_reset(sk);
5361                         goto discard;
5362                 }
5363
5364                 /* rfc793:
5365                  *   "fifth, if neither of the SYN or RST bits is set then
5366                  *    drop the segment and return."
5367                  *
5368                  *    See note below!
5369                  *                                        --ANK(990513)
5370                  */
5371                 if (!th->syn)
5372                         goto discard_and_undo;
5373
5374                 /* rfc793:
5375                  *   "If the SYN bit is on ...
5376                  *    are acceptable then ...
5377                  *    (our SYN has been ACKed), change the connection
5378                  *    state to ESTABLISHED..."
5379                  */
5380
5381                 TCP_ECN_rcv_synack(tp, th);
5382
5383                 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5384                 tcp_ack(sk, skb, FLAG_SLOWPATH);
5385
5386                 /* Ok.. it's good. Set up sequence numbers and
5387                  * move to established.
5388                  */
5389                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5390                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5391
5392                 /* RFC1323: The window in SYN & SYN/ACK segments is
5393                  * never scaled.
5394                  */
5395                 tp->snd_wnd = ntohs(th->window);
5396                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5397
5398                 if (!tp->rx_opt.wscale_ok) {
5399                         tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5400                         tp->window_clamp = min(tp->window_clamp, 65535U);
5401                 }
5402
5403                 if (tp->rx_opt.saw_tstamp) {
5404                         tp->rx_opt.tstamp_ok       = 1;
5405                         tp->tcp_header_len =
5406                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5407                         tp->advmss          -= TCPOLEN_TSTAMP_ALIGNED;
5408                         tcp_store_ts_recent(tp);
5409                 } else {
5410                         tp->tcp_header_len = sizeof(struct tcphdr);
5411                 }
5412
5413                 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5414                         tcp_enable_fack(tp);
5415
5416                 tcp_mtup_init(sk);
5417                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5418                 tcp_initialize_rcv_mss(sk);
5419
5420                 /* Remember, tcp_poll() does not lock socket!
5421                  * Change state from SYN-SENT only after copied_seq
5422                  * is initialized. */
5423                 tp->copied_seq = tp->rcv_nxt;
5424                 smp_mb();
5425                 tcp_set_state(sk, TCP_ESTABLISHED);
5426
5427                 security_inet_conn_established(sk, skb);
5428
5429                 /* Make sure socket is routed, for correct metrics.  */
5430                 icsk->icsk_af_ops->rebuild_header(sk);
5431
5432                 tcp_init_metrics(sk);
5433
5434                 tcp_init_congestion_control(sk);
5435
5436                 /* Prevent spurious tcp_cwnd_restart() on first data
5437                  * packet.
5438                  */
5439                 tp->lsndtime = tcp_time_stamp;
5440
5441                 tcp_init_buffer_space(sk);
5442
5443                 if (sock_flag(sk, SOCK_KEEPOPEN))
5444                         inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5445
5446                 if (!tp->rx_opt.snd_wscale)
5447                         __tcp_fast_path_on(tp, tp->snd_wnd);
5448                 else
5449                         tp->pred_flags = 0;
5450
5451                 if (!sock_flag(sk, SOCK_DEAD)) {
5452                         sk->sk_state_change(sk);
5453                         sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5454                 }
5455
5456                 if (sk->sk_write_pending ||
5457                     icsk->icsk_accept_queue.rskq_defer_accept ||
5458                     icsk->icsk_ack.pingpong) {
5459                         /* Save one ACK. Data will be ready after
5460                          * several ticks, if write_pending is set.
5461                          *
5462                          * It may be deleted, but with this feature tcpdumps
5463                          * look so _wonderfully_ clever, that I was not able
5464                          * to stand against the temptation 8)     --ANK
5465                          */
5466                         inet_csk_schedule_ack(sk);
5467                         icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5468                         icsk->icsk_ack.ato       = TCP_ATO_MIN;
5469                         tcp_incr_quickack(sk);
5470                         tcp_enter_quickack_mode(sk);
5471                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5472                                                   TCP_DELACK_MAX, TCP_RTO_MAX);
5473
5474 discard:
5475                         __kfree_skb(skb);
5476                         return 0;
5477                 } else {
5478                         tcp_send_ack(sk);
5479                 }
5480                 return -1;
5481         }
5482
5483         /* No ACK in the segment */
5484
5485         if (th->rst) {
5486                 /* rfc793:
5487                  * "If the RST bit is set
5488                  *
5489                  *      Otherwise (no ACK) drop the segment and return."
5490                  */
5491
5492                 goto discard_and_undo;
5493         }
5494
5495         /* PAWS check. */
5496         if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5497             tcp_paws_reject(&tp->rx_opt, 0))
5498                 goto discard_and_undo;
5499
5500         if (th->syn) {
5501                 /* We see SYN without ACK. It is attempt of
5502                  * simultaneous connect with crossed SYNs.
5503                  * Particularly, it can be connect to self.
5504                  */
5505                 tcp_set_state(sk, TCP_SYN_RECV);
5506
5507                 if (tp->rx_opt.saw_tstamp) {
5508                         tp->rx_opt.tstamp_ok = 1;
5509                         tcp_store_ts_recent(tp);
5510                         tp->tcp_header_len =
5511                                 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5512                 } else {
5513                         tp->tcp_header_len = sizeof(struct tcphdr);
5514                 }
5515
5516                 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5517                 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5518
5519                 /* RFC1323: The window in SYN & SYN/ACK segments is
5520                  * never scaled.
5521                  */
5522                 tp->snd_wnd    = ntohs(th->window);
5523                 tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5524                 tp->max_window = tp->snd_wnd;
5525
5526                 TCP_ECN_rcv_syn(tp, th);
5527
5528                 tcp_mtup_init(sk);
5529                 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5530                 tcp_initialize_rcv_mss(sk);
5531
5532                 tcp_send_synack(sk);
5533 #if 0
5534                 /* Note, we could accept data and URG from this segment.
5535                  * There are no obstacles to make this.
5536                  *
5537                  * However, if we ignore data in ACKless segments sometimes,
5538                  * we have no reasons to accept it sometimes.
5539                  * Also, seems the code doing it in step6 of tcp_rcv_state_process
5540                  * is not flawless. So, discard packet for sanity.
5541                  * Uncomment this return to process the data.
5542                  */
5543                 return -1;
5544 #else
5545                 goto discard;
5546 #endif
5547         }
5548         /* "fifth, if neither of the SYN or RST bits is set then
5549          * drop the segment and return."
5550          */
5551
5552 discard_and_undo:
5553         tcp_clear_options(&tp->rx_opt);
5554         tp->rx_opt.mss_clamp = saved_clamp;
5555         goto discard;
5556
5557 reset_and_undo:
5558         tcp_clear_options(&tp->rx_opt);
5559         tp->rx_opt.mss_clamp = saved_clamp;
5560         return 1;
5561 }
5562
5563 /*
5564  *      This function implements the receiving procedure of RFC 793 for
5565  *      all states except ESTABLISHED and TIME_WAIT.
5566  *      It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5567  *      address independent.
5568  */
5569
5570 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5571                           struct tcphdr *th, unsigned len)
5572 {
5573         struct tcp_sock *tp = tcp_sk(sk);
5574         struct inet_connection_sock *icsk = inet_csk(sk);
5575         int queued = 0;
5576         int res;
5577
5578         tp->rx_opt.saw_tstamp = 0;
5579
5580         switch (sk->sk_state) {
5581         case TCP_CLOSE:
5582                 goto discard;
5583
5584         case TCP_LISTEN:
5585                 if (th->ack)
5586                         return 1;
5587
5588                 if (th->rst)
5589                         goto discard;
5590
5591                 if (th->syn) {
5592                         if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5593                                 return 1;
5594
5595                         /* Now we have several options: In theory there is
5596                          * nothing else in the frame. KA9Q has an option to
5597                          * send data with the syn, BSD accepts data with the
5598                          * syn up to the [to be] advertised window and
5599                          * Solaris 2.1 gives you a protocol error. For now
5600                          * we just ignore it, that fits the spec precisely
5601                          * and avoids incompatibilities. It would be nice in
5602                          * future to drop through and process the data.
5603                          *
5604                          * Now that TTCP is starting to be used we ought to
5605                          * queue this data.
5606                          * But, this leaves one open to an easy denial of
5607                          * service attack, and SYN cookies can't defend
5608                          * against this problem. So, we drop the data
5609                          * in the interest of security over speed unless
5610                          * it's still in use.
5611                          */
5612                         kfree_skb(skb);
5613                         return 0;
5614                 }
5615                 goto discard;
5616
5617         case TCP_SYN_SENT:
5618                 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5619                 if (queued >= 0)
5620                         return queued;
5621
5622                 /* Do step6 onward by hand. */
5623                 tcp_urg(sk, skb, th);
5624                 __kfree_skb(skb);
5625                 tcp_data_snd_check(sk);
5626                 return 0;
5627         }
5628
5629         res = tcp_validate_incoming(sk, skb, th, 0);
5630         if (res <= 0)
5631                 return -res;
5632
5633         /* step 5: check the ACK field */
5634         if (th->ack) {
5635                 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5636
5637                 switch (sk->sk_state) {
5638                 case TCP_SYN_RECV:
5639                         if (acceptable) {
5640                                 tp->copied_seq = tp->rcv_nxt;
5641                                 smp_mb();
5642                                 tcp_set_state(sk, TCP_ESTABLISHED);
5643                                 sk->sk_state_change(sk);
5644
5645                                 /* Note, that this wakeup is only for marginal
5646                                  * crossed SYN case. Passively open sockets
5647                                  * are not waked up, because sk->sk_sleep ==
5648                                  * NULL and sk->sk_socket == NULL.
5649                                  */
5650                                 if (sk->sk_socket)
5651                                         sk_wake_async(sk,
5652                                                       SOCK_WAKE_IO, POLL_OUT);
5653
5654                                 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5655                                 tp->snd_wnd = ntohs(th->window) <<
5656                                               tp->rx_opt.snd_wscale;
5657                                 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5658
5659                                 /* tcp_ack considers this ACK as duplicate
5660                                  * and does not calculate rtt.
5661                                  * Fix it at least with timestamps.
5662                                  */
5663                                 if (tp->rx_opt.saw_tstamp &&
5664                                     tp->rx_opt.rcv_tsecr && !tp->srtt)
5665                                         tcp_ack_saw_tstamp(sk, 0);
5666
5667                                 if (tp->rx_opt.tstamp_ok)
5668                                         tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5669
5670                                 /* Make sure socket is routed, for
5671                                  * correct metrics.
5672                                  */
5673                                 icsk->icsk_af_ops->rebuild_header(sk);
5674
5675                                 tcp_init_metrics(sk);
5676
5677                                 tcp_init_congestion_control(sk);
5678
5679                                 /* Prevent spurious tcp_cwnd_restart() on
5680                                  * first data packet.
5681                                  */
5682                                 tp->lsndtime = tcp_time_stamp;
5683
5684                                 tcp_mtup_init(sk);
5685                                 tcp_initialize_rcv_mss(sk);
5686                                 tcp_init_buffer_space(sk);
5687                                 tcp_fast_path_on(tp);
5688                         } else {
5689                                 return 1;
5690                         }
5691                         break;
5692
5693                 case TCP_FIN_WAIT1:
5694                         if (tp->snd_una == tp->write_seq) {
5695                                 tcp_set_state(sk, TCP_FIN_WAIT2);
5696                                 sk->sk_shutdown |= SEND_SHUTDOWN;
5697                                 dst_confirm(sk->sk_dst_cache);
5698
5699                                 if (!sock_flag(sk, SOCK_DEAD))
5700                                         /* Wake up lingering close() */
5701                                         sk->sk_state_change(sk);
5702                                 else {
5703                                         int tmo;
5704
5705                                         if (tp->linger2 < 0 ||
5706                                             (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5707                                              after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5708                                                 tcp_done(sk);
5709                                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5710                                                 return 1;
5711                                         }
5712
5713                                         tmo = tcp_fin_time(sk);
5714                                         if (tmo > TCP_TIMEWAIT_LEN) {
5715                                                 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5716                                         } else if (th->fin || sock_owned_by_user(sk)) {
5717                                                 /* Bad case. We could lose such FIN otherwise.
5718                                                  * It is not a big problem, but it looks confusing
5719                                                  * and not so rare event. We still can lose it now,
5720                                                  * if it spins in bh_lock_sock(), but it is really
5721                                                  * marginal case.
5722                                                  */
5723                                                 inet_csk_reset_keepalive_timer(sk, tmo);
5724                                         } else {
5725                                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5726                                                 goto discard;
5727                                         }
5728                                 }
5729                         }
5730                         break;
5731
5732                 case TCP_CLOSING:
5733                         if (tp->snd_una == tp->write_seq) {
5734                                 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5735                                 goto discard;
5736                         }
5737                         break;
5738
5739                 case TCP_LAST_ACK:
5740                         if (tp->snd_una == tp->write_seq) {
5741                                 tcp_update_metrics(sk);
5742                                 tcp_done(sk);
5743                                 goto discard;
5744                         }
5745                         break;
5746                 }
5747         } else
5748                 goto discard;
5749
5750         /* step 6: check the URG bit */
5751         tcp_urg(sk, skb, th);
5752
5753         /* step 7: process the segment text */
5754         switch (sk->sk_state) {
5755         case TCP_CLOSE_WAIT:
5756         case TCP_CLOSING:
5757         case TCP_LAST_ACK:
5758                 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5759                         break;
5760         case TCP_FIN_WAIT1:
5761         case TCP_FIN_WAIT2:
5762                 /* RFC 793 says to queue data in these states,
5763                  * RFC 1122 says we MUST send a reset.
5764                  * BSD 4.4 also does reset.
5765                  */
5766                 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5767                         if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5768                             after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5769                                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5770                                 tcp_reset(sk);
5771                                 return 1;
5772                         }
5773                 }
5774                 /* Fall through */
5775         case TCP_ESTABLISHED:
5776                 tcp_data_queue(sk, skb);
5777                 queued = 1;
5778                 break;
5779         }
5780
5781         /* tcp_data could move socket to TIME-WAIT */
5782         if (sk->sk_state != TCP_CLOSE) {
5783                 tcp_data_snd_check(sk);
5784                 tcp_ack_snd_check(sk);
5785         }
5786
5787         if (!queued) {
5788 discard:
5789                 __kfree_skb(skb);
5790         }
5791         return 0;
5792 }
5793
5794 EXPORT_SYMBOL(sysctl_tcp_ecn);
5795 EXPORT_SYMBOL(sysctl_tcp_reordering);
5796 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5797 EXPORT_SYMBOL(tcp_parse_options);
5798 #ifdef CONFIG_TCP_MD5SIG
5799 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5800 #endif
5801 EXPORT_SYMBOL(tcp_rcv_established);
5802 EXPORT_SYMBOL(tcp_rcv_state_process);
5803 EXPORT_SYMBOL(tcp_initialize_rcv_mss);