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.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presence of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
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;
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;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
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) */
107 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
108 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
109 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
110 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
111 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
113 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
114 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
115 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
117 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
119 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
121 /* Adapt the MSS value used to make delayed ack decision to the
124 static void tcp_measure_rcv_mss(struct sock *sk,
125 const struct sk_buff *skb)
127 struct inet_connection_sock *icsk = inet_csk(sk);
128 const unsigned int lss = icsk->icsk_ack.last_seg_size;
131 icsk->icsk_ack.last_seg_size = 0;
133 /* skb->len may jitter because of SACKs, even if peer
134 * sends good full-sized frames.
136 len = skb_shinfo(skb)->gso_size ?: skb->len;
137 if (len >= icsk->icsk_ack.rcv_mss) {
138 icsk->icsk_ack.rcv_mss = len;
140 /* Otherwise, we make more careful check taking into account,
141 * that SACKs block is variable.
143 * "len" is invariant segment length, including TCP header.
145 len += skb->data - skb_transport_header(skb);
146 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
147 /* If PSH is not set, packet should be
148 * full sized, provided peer TCP is not badly broken.
149 * This observation (if it is correct 8)) allows
150 * to handle super-low mtu links fairly.
152 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
153 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
154 /* Subtract also invariant (if peer is RFC compliant),
155 * tcp header plus fixed timestamp option length.
156 * Resulting "len" is MSS free of SACK jitter.
158 len -= tcp_sk(sk)->tcp_header_len;
159 icsk->icsk_ack.last_seg_size = len;
161 icsk->icsk_ack.rcv_mss = len;
165 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
166 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
167 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
171 static void tcp_incr_quickack(struct sock *sk)
173 struct inet_connection_sock *icsk = inet_csk(sk);
174 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
178 if (quickacks > icsk->icsk_ack.quick)
179 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
182 void tcp_enter_quickack_mode(struct sock *sk)
184 struct inet_connection_sock *icsk = inet_csk(sk);
185 tcp_incr_quickack(sk);
186 icsk->icsk_ack.pingpong = 0;
187 icsk->icsk_ack.ato = TCP_ATO_MIN;
190 /* Send ACKs quickly, if "quick" count is not exhausted
191 * and the session is not interactive.
194 static inline int tcp_in_quickack_mode(const struct sock *sk)
196 const struct inet_connection_sock *icsk = inet_csk(sk);
197 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
200 /* Buffer size and advertised window tuning.
202 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
205 static void tcp_fixup_sndbuf(struct sock *sk)
207 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
208 sizeof(struct sk_buff);
210 if (sk->sk_sndbuf < 3 * sndmem)
211 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
214 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
216 * All tcp_full_space() is split to two parts: "network" buffer, allocated
217 * forward and advertised in receiver window (tp->rcv_wnd) and
218 * "application buffer", required to isolate scheduling/application
219 * latencies from network.
220 * window_clamp is maximal advertised window. It can be less than
221 * tcp_full_space(), in this case tcp_full_space() - window_clamp
222 * is reserved for "application" buffer. The less window_clamp is
223 * the smoother our behaviour from viewpoint of network, but the lower
224 * throughput and the higher sensitivity of the connection to losses. 8)
226 * rcv_ssthresh is more strict window_clamp used at "slow start"
227 * phase to predict further behaviour of this connection.
228 * It is used for two goals:
229 * - to enforce header prediction at sender, even when application
230 * requires some significant "application buffer". It is check #1.
231 * - to prevent pruning of receive queue because of misprediction
232 * of receiver window. Check #2.
234 * The scheme does not work when sender sends good segments opening
235 * window and then starts to feed us spaghetti. But it should work
236 * in common situations. Otherwise, we have to rely on queue collapsing.
239 /* Slow part of check#2. */
240 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
242 struct tcp_sock *tp = tcp_sk(sk);
244 int truesize = tcp_win_from_space(skb->truesize)/2;
245 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
247 while (tp->rcv_ssthresh <= window) {
248 if (truesize <= skb->len)
249 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
257 static void tcp_grow_window(struct sock *sk,
260 struct tcp_sock *tp = tcp_sk(sk);
263 if (tp->rcv_ssthresh < tp->window_clamp &&
264 (int)tp->rcv_ssthresh < tcp_space(sk) &&
265 !tcp_memory_pressure) {
268 /* Check #2. Increase window, if skb with such overhead
269 * will fit to rcvbuf in future.
271 if (tcp_win_from_space(skb->truesize) <= skb->len)
274 incr = __tcp_grow_window(sk, skb);
277 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
278 inet_csk(sk)->icsk_ack.quick |= 1;
283 /* 3. Tuning rcvbuf, when connection enters established state. */
285 static void tcp_fixup_rcvbuf(struct sock *sk)
287 struct tcp_sock *tp = tcp_sk(sk);
288 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
290 /* Try to select rcvbuf so that 4 mss-sized segments
291 * will fit to window and corresponding skbs will fit to our rcvbuf.
292 * (was 3; 4 is minimum to allow fast retransmit to work.)
294 while (tcp_win_from_space(rcvmem) < tp->advmss)
296 if (sk->sk_rcvbuf < 4 * rcvmem)
297 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
300 /* 4. Try to fixup all. It is made immediately after connection enters
303 static void tcp_init_buffer_space(struct sock *sk)
305 struct tcp_sock *tp = tcp_sk(sk);
308 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
309 tcp_fixup_rcvbuf(sk);
310 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
311 tcp_fixup_sndbuf(sk);
313 tp->rcvq_space.space = tp->rcv_wnd;
315 maxwin = tcp_full_space(sk);
317 if (tp->window_clamp >= maxwin) {
318 tp->window_clamp = maxwin;
320 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
321 tp->window_clamp = max(maxwin -
322 (maxwin >> sysctl_tcp_app_win),
326 /* Force reservation of one segment. */
327 if (sysctl_tcp_app_win &&
328 tp->window_clamp > 2 * tp->advmss &&
329 tp->window_clamp + tp->advmss > maxwin)
330 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
332 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
333 tp->snd_cwnd_stamp = tcp_time_stamp;
336 /* 5. Recalculate window clamp after socket hit its memory bounds. */
337 static void tcp_clamp_window(struct sock *sk)
339 struct tcp_sock *tp = tcp_sk(sk);
340 struct inet_connection_sock *icsk = inet_csk(sk);
342 icsk->icsk_ack.quick = 0;
344 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
345 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
346 !tcp_memory_pressure &&
347 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
348 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
351 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
352 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
356 /* Initialize RCV_MSS value.
357 * RCV_MSS is an our guess about MSS used by the peer.
358 * We haven't any direct information about the MSS.
359 * It's better to underestimate the RCV_MSS rather than overestimate.
360 * Overestimations make us ACKing less frequently than needed.
361 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
363 void tcp_initialize_rcv_mss(struct sock *sk)
365 struct tcp_sock *tp = tcp_sk(sk);
366 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
368 hint = min(hint, tp->rcv_wnd/2);
369 hint = min(hint, TCP_MIN_RCVMSS);
370 hint = max(hint, TCP_MIN_MSS);
372 inet_csk(sk)->icsk_ack.rcv_mss = hint;
375 /* Receiver "autotuning" code.
377 * The algorithm for RTT estimation w/o timestamps is based on
378 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
379 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
381 * More detail on this code can be found at
382 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
383 * though this reference is out of date. A new paper
386 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
388 u32 new_sample = tp->rcv_rtt_est.rtt;
394 if (new_sample != 0) {
395 /* If we sample in larger samples in the non-timestamp
396 * case, we could grossly overestimate the RTT especially
397 * with chatty applications or bulk transfer apps which
398 * are stalled on filesystem I/O.
400 * Also, since we are only going for a minimum in the
401 * non-timestamp case, we do not smooth things out
402 * else with timestamps disabled convergence takes too
406 m -= (new_sample >> 3);
408 } else if (m < new_sample)
411 /* No previous measure. */
415 if (tp->rcv_rtt_est.rtt != new_sample)
416 tp->rcv_rtt_est.rtt = new_sample;
419 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
421 if (tp->rcv_rtt_est.time == 0)
423 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
425 tcp_rcv_rtt_update(tp,
426 jiffies - tp->rcv_rtt_est.time,
430 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
431 tp->rcv_rtt_est.time = tcp_time_stamp;
434 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
436 struct tcp_sock *tp = tcp_sk(sk);
437 if (tp->rx_opt.rcv_tsecr &&
438 (TCP_SKB_CB(skb)->end_seq -
439 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
440 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
444 * This function should be called every time data is copied to user space.
445 * It calculates the appropriate TCP receive buffer space.
447 void tcp_rcv_space_adjust(struct sock *sk)
449 struct tcp_sock *tp = tcp_sk(sk);
453 if (tp->rcvq_space.time == 0)
456 time = tcp_time_stamp - tp->rcvq_space.time;
457 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
458 tp->rcv_rtt_est.rtt == 0)
461 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
463 space = max(tp->rcvq_space.space, space);
465 if (tp->rcvq_space.space != space) {
468 tp->rcvq_space.space = space;
470 if (sysctl_tcp_moderate_rcvbuf &&
471 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
472 int new_clamp = space;
474 /* Receive space grows, normalize in order to
475 * take into account packet headers and sk_buff
476 * structure overhead.
481 rcvmem = (tp->advmss + MAX_TCP_HEADER +
482 16 + sizeof(struct sk_buff));
483 while (tcp_win_from_space(rcvmem) < tp->advmss)
486 space = min(space, sysctl_tcp_rmem[2]);
487 if (space > sk->sk_rcvbuf) {
488 sk->sk_rcvbuf = space;
490 /* Make the window clamp follow along. */
491 tp->window_clamp = new_clamp;
497 tp->rcvq_space.seq = tp->copied_seq;
498 tp->rcvq_space.time = tcp_time_stamp;
501 /* There is something which you must keep in mind when you analyze the
502 * behavior of the tp->ato delayed ack timeout interval. When a
503 * connection starts up, we want to ack as quickly as possible. The
504 * problem is that "good" TCP's do slow start at the beginning of data
505 * transmission. The means that until we send the first few ACK's the
506 * sender will sit on his end and only queue most of his data, because
507 * he can only send snd_cwnd unacked packets at any given time. For
508 * each ACK we send, he increments snd_cwnd and transmits more of his
511 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
513 struct tcp_sock *tp = tcp_sk(sk);
514 struct inet_connection_sock *icsk = inet_csk(sk);
517 inet_csk_schedule_ack(sk);
519 tcp_measure_rcv_mss(sk, skb);
521 tcp_rcv_rtt_measure(tp);
523 now = tcp_time_stamp;
525 if (!icsk->icsk_ack.ato) {
526 /* The _first_ data packet received, initialize
527 * delayed ACK engine.
529 tcp_incr_quickack(sk);
530 icsk->icsk_ack.ato = TCP_ATO_MIN;
532 int m = now - icsk->icsk_ack.lrcvtime;
534 if (m <= TCP_ATO_MIN/2) {
535 /* The fastest case is the first. */
536 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
537 } else if (m < icsk->icsk_ack.ato) {
538 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
539 if (icsk->icsk_ack.ato > icsk->icsk_rto)
540 icsk->icsk_ack.ato = icsk->icsk_rto;
541 } else if (m > icsk->icsk_rto) {
542 /* Too long gap. Apparently sender failed to
543 * restart window, so that we send ACKs quickly.
545 tcp_incr_quickack(sk);
546 sk_stream_mem_reclaim(sk);
549 icsk->icsk_ack.lrcvtime = now;
551 TCP_ECN_check_ce(tp, skb);
554 tcp_grow_window(sk, skb);
557 /* Called to compute a smoothed rtt estimate. The data fed to this
558 * routine either comes from timestamps, or from segments that were
559 * known _not_ to have been retransmitted [see Karn/Partridge
560 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
561 * piece by Van Jacobson.
562 * NOTE: the next three routines used to be one big routine.
563 * To save cycles in the RFC 1323 implementation it was better to break
564 * it up into three procedures. -- erics
566 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
568 struct tcp_sock *tp = tcp_sk(sk);
569 long m = mrtt; /* RTT */
571 /* The following amusing code comes from Jacobson's
572 * article in SIGCOMM '88. Note that rtt and mdev
573 * are scaled versions of rtt and mean deviation.
574 * This is designed to be as fast as possible
575 * m stands for "measurement".
577 * On a 1990 paper the rto value is changed to:
578 * RTO = rtt + 4 * mdev
580 * Funny. This algorithm seems to be very broken.
581 * These formulae increase RTO, when it should be decreased, increase
582 * too slowly, when it should be increased quickly, decrease too quickly
583 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
584 * does not matter how to _calculate_ it. Seems, it was trap
585 * that VJ failed to avoid. 8)
590 m -= (tp->srtt >> 3); /* m is now error in rtt est */
591 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
593 m = -m; /* m is now abs(error) */
594 m -= (tp->mdev >> 2); /* similar update on mdev */
595 /* This is similar to one of Eifel findings.
596 * Eifel blocks mdev updates when rtt decreases.
597 * This solution is a bit different: we use finer gain
598 * for mdev in this case (alpha*beta).
599 * Like Eifel it also prevents growth of rto,
600 * but also it limits too fast rto decreases,
601 * happening in pure Eifel.
606 m -= (tp->mdev >> 2); /* similar update on mdev */
608 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
609 if (tp->mdev > tp->mdev_max) {
610 tp->mdev_max = tp->mdev;
611 if (tp->mdev_max > tp->rttvar)
612 tp->rttvar = tp->mdev_max;
614 if (after(tp->snd_una, tp->rtt_seq)) {
615 if (tp->mdev_max < tp->rttvar)
616 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
617 tp->rtt_seq = tp->snd_nxt;
618 tp->mdev_max = TCP_RTO_MIN;
621 /* no previous measure. */
622 tp->srtt = m<<3; /* take the measured time to be rtt */
623 tp->mdev = m<<1; /* make sure rto = 3*rtt */
624 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
625 tp->rtt_seq = tp->snd_nxt;
629 /* Calculate rto without backoff. This is the second half of Van Jacobson's
630 * routine referred to above.
632 static inline void tcp_set_rto(struct sock *sk)
634 const struct tcp_sock *tp = tcp_sk(sk);
635 /* Old crap is replaced with new one. 8)
638 * 1. If rtt variance happened to be less 50msec, it is hallucination.
639 * It cannot be less due to utterly erratic ACK generation made
640 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
641 * to do with delayed acks, because at cwnd>2 true delack timeout
642 * is invisible. Actually, Linux-2.4 also generates erratic
643 * ACKs in some circumstances.
645 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
647 /* 2. Fixups made earlier cannot be right.
648 * If we do not estimate RTO correctly without them,
649 * all the algo is pure shit and should be replaced
650 * with correct one. It is exactly, which we pretend to do.
654 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
655 * guarantees that rto is higher.
657 static inline void tcp_bound_rto(struct sock *sk)
659 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
660 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
663 /* Save metrics learned by this TCP session.
664 This function is called only, when TCP finishes successfully
665 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
667 void tcp_update_metrics(struct sock *sk)
669 struct tcp_sock *tp = tcp_sk(sk);
670 struct dst_entry *dst = __sk_dst_get(sk);
672 if (sysctl_tcp_nometrics_save)
677 if (dst && (dst->flags&DST_HOST)) {
678 const struct inet_connection_sock *icsk = inet_csk(sk);
681 if (icsk->icsk_backoff || !tp->srtt) {
682 /* This session failed to estimate rtt. Why?
683 * Probably, no packets returned in time.
686 if (!(dst_metric_locked(dst, RTAX_RTT)))
687 dst->metrics[RTAX_RTT-1] = 0;
691 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
693 /* If newly calculated rtt larger than stored one,
694 * store new one. Otherwise, use EWMA. Remember,
695 * rtt overestimation is always better than underestimation.
697 if (!(dst_metric_locked(dst, RTAX_RTT))) {
699 dst->metrics[RTAX_RTT-1] = tp->srtt;
701 dst->metrics[RTAX_RTT-1] -= (m>>3);
704 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
708 /* Scale deviation to rttvar fixed point */
713 if (m >= dst_metric(dst, RTAX_RTTVAR))
714 dst->metrics[RTAX_RTTVAR-1] = m;
716 dst->metrics[RTAX_RTTVAR-1] -=
717 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
720 if (tp->snd_ssthresh >= 0xFFFF) {
721 /* Slow start still did not finish. */
722 if (dst_metric(dst, RTAX_SSTHRESH) &&
723 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
724 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
725 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
726 if (!dst_metric_locked(dst, RTAX_CWND) &&
727 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
728 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
729 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
730 icsk->icsk_ca_state == TCP_CA_Open) {
731 /* Cong. avoidance phase, cwnd is reliable. */
732 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
733 dst->metrics[RTAX_SSTHRESH-1] =
734 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
735 if (!dst_metric_locked(dst, RTAX_CWND))
736 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
738 /* Else slow start did not finish, cwnd is non-sense,
739 ssthresh may be also invalid.
741 if (!dst_metric_locked(dst, RTAX_CWND))
742 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
743 if (dst->metrics[RTAX_SSTHRESH-1] &&
744 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
745 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
746 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
749 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
750 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
751 tp->reordering != sysctl_tcp_reordering)
752 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
757 /* Numbers are taken from RFC2414. */
758 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
760 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
763 if (tp->mss_cache > 1460)
766 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
768 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
771 /* Set slow start threshold and cwnd not falling to slow start */
772 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
774 struct tcp_sock *tp = tcp_sk(sk);
775 const struct inet_connection_sock *icsk = inet_csk(sk);
777 tp->prior_ssthresh = 0;
779 if (icsk->icsk_ca_state < TCP_CA_CWR) {
782 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
783 tp->snd_cwnd = min(tp->snd_cwnd,
784 tcp_packets_in_flight(tp) + 1U);
785 tp->snd_cwnd_cnt = 0;
786 tp->high_seq = tp->snd_nxt;
787 tp->snd_cwnd_stamp = tcp_time_stamp;
788 TCP_ECN_queue_cwr(tp);
790 tcp_set_ca_state(sk, TCP_CA_CWR);
794 /* Initialize metrics on socket. */
796 static void tcp_init_metrics(struct sock *sk)
798 struct tcp_sock *tp = tcp_sk(sk);
799 struct dst_entry *dst = __sk_dst_get(sk);
806 if (dst_metric_locked(dst, RTAX_CWND))
807 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
808 if (dst_metric(dst, RTAX_SSTHRESH)) {
809 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
810 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
811 tp->snd_ssthresh = tp->snd_cwnd_clamp;
813 if (dst_metric(dst, RTAX_REORDERING) &&
814 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
815 tp->rx_opt.sack_ok &= ~2;
816 tp->reordering = dst_metric(dst, RTAX_REORDERING);
819 if (dst_metric(dst, RTAX_RTT) == 0)
822 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
825 /* Initial rtt is determined from SYN,SYN-ACK.
826 * The segment is small and rtt may appear much
827 * less than real one. Use per-dst memory
828 * to make it more realistic.
830 * A bit of theory. RTT is time passed after "normal" sized packet
831 * is sent until it is ACKed. In normal circumstances sending small
832 * packets force peer to delay ACKs and calculation is correct too.
833 * The algorithm is adaptive and, provided we follow specs, it
834 * NEVER underestimate RTT. BUT! If peer tries to make some clever
835 * tricks sort of "quick acks" for time long enough to decrease RTT
836 * to low value, and then abruptly stops to do it and starts to delay
837 * ACKs, wait for troubles.
839 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
840 tp->srtt = dst_metric(dst, RTAX_RTT);
841 tp->rtt_seq = tp->snd_nxt;
843 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
844 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
845 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
849 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
851 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
852 tp->snd_cwnd_stamp = tcp_time_stamp;
856 /* Play conservative. If timestamps are not
857 * supported, TCP will fail to recalculate correct
858 * rtt, if initial rto is too small. FORGET ALL AND RESET!
860 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
862 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
863 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
867 static void tcp_update_reordering(struct sock *sk, const int metric,
870 struct tcp_sock *tp = tcp_sk(sk);
871 if (metric > tp->reordering) {
872 tp->reordering = min(TCP_MAX_REORDERING, metric);
874 /* This exciting event is worth to be remembered. 8) */
876 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
878 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
880 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
882 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
883 #if FASTRETRANS_DEBUG > 1
884 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
885 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
889 tp->undo_marker ? tp->undo_retrans : 0);
891 /* Disable FACK yet. */
892 tp->rx_opt.sack_ok &= ~2;
896 /* This procedure tags the retransmission queue when SACKs arrive.
898 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
899 * Packets in queue with these bits set are counted in variables
900 * sacked_out, retrans_out and lost_out, correspondingly.
902 * Valid combinations are:
903 * Tag InFlight Description
904 * 0 1 - orig segment is in flight.
905 * S 0 - nothing flies, orig reached receiver.
906 * L 0 - nothing flies, orig lost by net.
907 * R 2 - both orig and retransmit are in flight.
908 * L|R 1 - orig is lost, retransmit is in flight.
909 * S|R 1 - orig reached receiver, retrans is still in flight.
910 * (L|S|R is logically valid, it could occur when L|R is sacked,
911 * but it is equivalent to plain S and code short-curcuits it to S.
912 * L|S is logically invalid, it would mean -1 packet in flight 8))
914 * These 6 states form finite state machine, controlled by the following events:
915 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
916 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
917 * 3. Loss detection event of one of three flavors:
918 * A. Scoreboard estimator decided the packet is lost.
919 * A'. Reno "three dupacks" marks head of queue lost.
920 * A''. Its FACK modfication, head until snd.fack is lost.
921 * B. SACK arrives sacking data transmitted after never retransmitted
923 * C. SACK arrives sacking SND.NXT at the moment, when the
924 * segment was retransmitted.
925 * 4. D-SACK added new rule: D-SACK changes any tag to S.
927 * It is pleasant to note, that state diagram turns out to be commutative,
928 * so that we are allowed not to be bothered by order of our actions,
929 * when multiple events arrive simultaneously. (see the function below).
931 * Reordering detection.
932 * --------------------
933 * Reordering metric is maximal distance, which a packet can be displaced
934 * in packet stream. With SACKs we can estimate it:
936 * 1. SACK fills old hole and the corresponding segment was not
937 * ever retransmitted -> reordering. Alas, we cannot use it
938 * when segment was retransmitted.
939 * 2. The last flaw is solved with D-SACK. D-SACK arrives
940 * for retransmitted and already SACKed segment -> reordering..
941 * Both of these heuristics are not used in Loss state, when we cannot
942 * account for retransmits accurately.
945 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
947 const struct inet_connection_sock *icsk = inet_csk(sk);
948 struct tcp_sock *tp = tcp_sk(sk);
949 unsigned char *ptr = (skb_transport_header(ack_skb) +
950 TCP_SKB_CB(ack_skb)->sacked);
951 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
952 struct sk_buff *cached_skb;
953 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
954 int reord = tp->packets_out;
956 u32 lost_retrans = 0;
958 int found_dup_sack = 0;
959 int cached_fack_count;
961 int first_sack_index;
965 prior_fackets = tp->fackets_out;
967 /* Check for D-SACK. */
968 if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
970 tp->rx_opt.sack_ok |= 4;
971 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
972 } else if (num_sacks > 1 &&
973 !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
974 !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
976 tp->rx_opt.sack_ok |= 4;
977 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
980 /* D-SACK for already forgotten data...
981 * Do dumb counting. */
982 if (found_dup_sack &&
983 !after(ntohl(sp[0].end_seq), prior_snd_una) &&
984 after(ntohl(sp[0].end_seq), tp->undo_marker))
987 /* Eliminate too old ACKs, but take into
988 * account more or less fresh ones, they can
989 * contain valid SACK info.
991 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
995 * if the only SACK change is the increase of the end_seq of
996 * the first block then only apply that SACK block
997 * and use retrans queue hinting otherwise slowpath */
999 for (i = 0; i < num_sacks; i++) {
1000 __be32 start_seq = sp[i].start_seq;
1001 __be32 end_seq = sp[i].end_seq;
1004 if (tp->recv_sack_cache[i].start_seq != start_seq)
1007 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1008 (tp->recv_sack_cache[i].end_seq != end_seq))
1011 tp->recv_sack_cache[i].start_seq = start_seq;
1012 tp->recv_sack_cache[i].end_seq = end_seq;
1014 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1015 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1016 tp->recv_sack_cache[i].start_seq = 0;
1017 tp->recv_sack_cache[i].end_seq = 0;
1020 first_sack_index = 0;
1025 tp->fastpath_skb_hint = NULL;
1027 /* order SACK blocks to allow in order walk of the retrans queue */
1028 for (i = num_sacks-1; i > 0; i--) {
1029 for (j = 0; j < i; j++){
1030 if (after(ntohl(sp[j].start_seq),
1031 ntohl(sp[j+1].start_seq))){
1032 struct tcp_sack_block_wire tmp;
1038 /* Track where the first SACK block goes to */
1039 if (j == first_sack_index)
1040 first_sack_index = j+1;
1047 /* clear flag as used for different purpose in following code */
1050 /* Use SACK fastpath hint if valid */
1051 cached_skb = tp->fastpath_skb_hint;
1052 cached_fack_count = tp->fastpath_cnt_hint;
1054 cached_skb = tcp_write_queue_head(sk);
1055 cached_fack_count = 0;
1058 for (i=0; i<num_sacks; i++, sp++) {
1059 struct sk_buff *skb;
1060 __u32 start_seq = ntohl(sp->start_seq);
1061 __u32 end_seq = ntohl(sp->end_seq);
1063 int dup_sack = (found_dup_sack && (i == first_sack_index));
1066 fack_count = cached_fack_count;
1068 /* Event "B" in the comment above. */
1069 if (after(end_seq, tp->high_seq))
1070 flag |= FLAG_DATA_LOST;
1072 tcp_for_write_queue_from(skb, sk) {
1073 int in_sack, pcount;
1076 if (skb == tcp_send_head(sk))
1080 cached_fack_count = fack_count;
1081 if (i == first_sack_index) {
1082 tp->fastpath_skb_hint = skb;
1083 tp->fastpath_cnt_hint = fack_count;
1086 /* The retransmission queue is always in order, so
1087 * we can short-circuit the walk early.
1089 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1092 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1093 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1095 pcount = tcp_skb_pcount(skb);
1097 if (pcount > 1 && !in_sack &&
1098 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1099 unsigned int pkt_len;
1101 in_sack = !after(start_seq,
1102 TCP_SKB_CB(skb)->seq);
1105 pkt_len = (start_seq -
1106 TCP_SKB_CB(skb)->seq);
1108 pkt_len = (end_seq -
1109 TCP_SKB_CB(skb)->seq);
1110 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1112 pcount = tcp_skb_pcount(skb);
1115 fack_count += pcount;
1117 sacked = TCP_SKB_CB(skb)->sacked;
1119 /* Account D-SACK for retransmitted packet. */
1120 if ((dup_sack && in_sack) &&
1121 (sacked & TCPCB_RETRANS) &&
1122 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1125 /* The frame is ACKed. */
1126 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1127 if (sacked&TCPCB_RETRANS) {
1128 if ((dup_sack && in_sack) &&
1129 (sacked&TCPCB_SACKED_ACKED))
1130 reord = min(fack_count, reord);
1132 /* If it was in a hole, we detected reordering. */
1133 if (fack_count < prior_fackets &&
1134 !(sacked&TCPCB_SACKED_ACKED))
1135 reord = min(fack_count, reord);
1138 /* Nothing to do; acked frame is about to be dropped. */
1142 if ((sacked&TCPCB_SACKED_RETRANS) &&
1143 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1144 (!lost_retrans || after(end_seq, lost_retrans)))
1145 lost_retrans = end_seq;
1150 if (!(sacked&TCPCB_SACKED_ACKED)) {
1151 if (sacked & TCPCB_SACKED_RETRANS) {
1152 /* If the segment is not tagged as lost,
1153 * we do not clear RETRANS, believing
1154 * that retransmission is still in flight.
1156 if (sacked & TCPCB_LOST) {
1157 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1158 tp->lost_out -= tcp_skb_pcount(skb);
1159 tp->retrans_out -= tcp_skb_pcount(skb);
1161 /* clear lost hint */
1162 tp->retransmit_skb_hint = NULL;
1165 /* New sack for not retransmitted frame,
1166 * which was in hole. It is reordering.
1168 if (!(sacked & TCPCB_RETRANS) &&
1169 fack_count < prior_fackets)
1170 reord = min(fack_count, reord);
1172 if (sacked & TCPCB_LOST) {
1173 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1174 tp->lost_out -= tcp_skb_pcount(skb);
1176 /* clear lost hint */
1177 tp->retransmit_skb_hint = NULL;
1179 /* SACK enhanced F-RTO detection.
1180 * Set flag if and only if non-rexmitted
1181 * segments below frto_highmark are
1182 * SACKed (RFC4138; Appendix B).
1183 * Clearing correct due to in-order walk
1185 if (after(end_seq, tp->frto_highmark)) {
1186 flag &= ~FLAG_ONLY_ORIG_SACKED;
1188 if (!(sacked & TCPCB_RETRANS))
1189 flag |= FLAG_ONLY_ORIG_SACKED;
1193 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1194 flag |= FLAG_DATA_SACKED;
1195 tp->sacked_out += tcp_skb_pcount(skb);
1197 if (fack_count > tp->fackets_out)
1198 tp->fackets_out = fack_count;
1200 if (dup_sack && (sacked&TCPCB_RETRANS))
1201 reord = min(fack_count, reord);
1204 /* D-SACK. We can detect redundant retransmission
1205 * in S|R and plain R frames and clear it.
1206 * undo_retrans is decreased above, L|R frames
1207 * are accounted above as well.
1210 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1211 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1212 tp->retrans_out -= tcp_skb_pcount(skb);
1213 tp->retransmit_skb_hint = NULL;
1218 /* Check for lost retransmit. This superb idea is
1219 * borrowed from "ratehalving". Event "C".
1220 * Later note: FACK people cheated me again 8),
1221 * we have to account for reordering! Ugly,
1224 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1225 struct sk_buff *skb;
1227 tcp_for_write_queue(skb, sk) {
1228 if (skb == tcp_send_head(sk))
1230 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1232 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1234 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1235 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1237 !before(lost_retrans,
1238 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1240 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1241 tp->retrans_out -= tcp_skb_pcount(skb);
1243 /* clear lost hint */
1244 tp->retransmit_skb_hint = NULL;
1246 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1247 tp->lost_out += tcp_skb_pcount(skb);
1248 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1249 flag |= FLAG_DATA_SACKED;
1250 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1256 tp->left_out = tp->sacked_out + tp->lost_out;
1258 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1259 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1260 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1262 #if FASTRETRANS_DEBUG > 0
1263 BUG_TRAP((int)tp->sacked_out >= 0);
1264 BUG_TRAP((int)tp->lost_out >= 0);
1265 BUG_TRAP((int)tp->retrans_out >= 0);
1266 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1271 /* F-RTO can only be used if TCP has never retransmitted anything other than
1272 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1274 int tcp_use_frto(struct sock *sk)
1276 const struct tcp_sock *tp = tcp_sk(sk);
1277 struct sk_buff *skb;
1279 if (!sysctl_tcp_frto)
1285 /* Avoid expensive walking of rexmit queue if possible */
1286 if (tp->retrans_out > 1)
1289 skb = tcp_write_queue_head(sk);
1290 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1291 tcp_for_write_queue_from(skb, sk) {
1292 if (skb == tcp_send_head(sk))
1294 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1296 /* Short-circuit when first non-SACKed skb has been checked */
1297 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1303 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1304 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1305 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1306 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1307 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1308 * bits are handled if the Loss state is really to be entered (in
1309 * tcp_enter_frto_loss).
1311 * Do like tcp_enter_loss() would; when RTO expires the second time it
1313 * "Reduce ssthresh if it has not yet been made inside this window."
1315 void tcp_enter_frto(struct sock *sk)
1317 const struct inet_connection_sock *icsk = inet_csk(sk);
1318 struct tcp_sock *tp = tcp_sk(sk);
1319 struct sk_buff *skb;
1321 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1322 tp->snd_una == tp->high_seq ||
1323 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1324 !icsk->icsk_retransmits)) {
1325 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1326 /* Our state is too optimistic in ssthresh() call because cwnd
1327 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1328 * recovery has not yet completed. Pattern would be this: RTO,
1329 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1331 * RFC4138 should be more specific on what to do, even though
1332 * RTO is quite unlikely to occur after the first Cumulative ACK
1333 * due to back-off and complexity of triggering events ...
1335 if (tp->frto_counter) {
1337 stored_cwnd = tp->snd_cwnd;
1339 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1340 tp->snd_cwnd = stored_cwnd;
1342 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1344 /* ... in theory, cong.control module could do "any tricks" in
1345 * ssthresh(), which means that ca_state, lost bits and lost_out
1346 * counter would have to be faked before the call occurs. We
1347 * consider that too expensive, unlikely and hacky, so modules
1348 * using these in ssthresh() must deal these incompatibility
1349 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1351 tcp_ca_event(sk, CA_EVENT_FRTO);
1354 tp->undo_marker = tp->snd_una;
1355 tp->undo_retrans = 0;
1357 skb = tcp_write_queue_head(sk);
1358 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1359 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1360 tp->retrans_out -= tcp_skb_pcount(skb);
1362 tcp_sync_left_out(tp);
1364 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1365 * The last condition is necessary at least in tp->frto_counter case.
1367 if (IsSackFrto() && (tp->frto_counter ||
1368 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1369 after(tp->high_seq, tp->snd_una)) {
1370 tp->frto_highmark = tp->high_seq;
1372 tp->frto_highmark = tp->snd_nxt;
1374 tcp_set_ca_state(sk, TCP_CA_Disorder);
1375 tp->high_seq = tp->snd_nxt;
1376 tp->frto_counter = 1;
1379 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1380 * which indicates that we should follow the traditional RTO recovery,
1381 * i.e. mark everything lost and do go-back-N retransmission.
1383 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1385 struct tcp_sock *tp = tcp_sk(sk);
1386 struct sk_buff *skb;
1391 tp->fackets_out = 0;
1392 tp->retrans_out = 0;
1394 tcp_for_write_queue(skb, sk) {
1395 if (skb == tcp_send_head(sk))
1397 cnt += tcp_skb_pcount(skb);
1399 * Count the retransmission made on RTO correctly (only when
1400 * waiting for the first ACK and did not get it)...
1402 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1403 /* For some reason this R-bit might get cleared? */
1404 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1405 tp->retrans_out += tcp_skb_pcount(skb);
1406 /* ...enter this if branch just for the first segment */
1407 flag |= FLAG_DATA_ACKED;
1409 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1411 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1413 /* Do not mark those segments lost that were
1414 * forward transmitted after RTO
1416 if (!after(TCP_SKB_CB(skb)->end_seq,
1417 tp->frto_highmark)) {
1418 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1419 tp->lost_out += tcp_skb_pcount(skb);
1422 tp->sacked_out += tcp_skb_pcount(skb);
1423 tp->fackets_out = cnt;
1426 tcp_sync_left_out(tp);
1428 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1429 tp->snd_cwnd_cnt = 0;
1430 tp->snd_cwnd_stamp = tcp_time_stamp;
1431 tp->undo_marker = 0;
1432 tp->frto_counter = 0;
1434 tp->reordering = min_t(unsigned int, tp->reordering,
1435 sysctl_tcp_reordering);
1436 tcp_set_ca_state(sk, TCP_CA_Loss);
1437 tp->high_seq = tp->frto_highmark;
1438 TCP_ECN_queue_cwr(tp);
1440 clear_all_retrans_hints(tp);
1443 void tcp_clear_retrans(struct tcp_sock *tp)
1446 tp->retrans_out = 0;
1448 tp->fackets_out = 0;
1452 tp->undo_marker = 0;
1453 tp->undo_retrans = 0;
1456 /* Enter Loss state. If "how" is not zero, forget all SACK information
1457 * and reset tags completely, otherwise preserve SACKs. If receiver
1458 * dropped its ofo queue, we will know this due to reneging detection.
1460 void tcp_enter_loss(struct sock *sk, int how)
1462 const struct inet_connection_sock *icsk = inet_csk(sk);
1463 struct tcp_sock *tp = tcp_sk(sk);
1464 struct sk_buff *skb;
1467 /* Reduce ssthresh if it has not yet been made inside this window. */
1468 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1469 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1470 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1471 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1472 tcp_ca_event(sk, CA_EVENT_LOSS);
1475 tp->snd_cwnd_cnt = 0;
1476 tp->snd_cwnd_stamp = tcp_time_stamp;
1478 tp->bytes_acked = 0;
1479 tcp_clear_retrans(tp);
1481 /* Push undo marker, if it was plain RTO and nothing
1482 * was retransmitted. */
1484 tp->undo_marker = tp->snd_una;
1486 tcp_for_write_queue(skb, sk) {
1487 if (skb == tcp_send_head(sk))
1489 cnt += tcp_skb_pcount(skb);
1490 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1491 tp->undo_marker = 0;
1492 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1493 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1494 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1495 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1496 tp->lost_out += tcp_skb_pcount(skb);
1498 tp->sacked_out += tcp_skb_pcount(skb);
1499 tp->fackets_out = cnt;
1502 tcp_sync_left_out(tp);
1504 tp->reordering = min_t(unsigned int, tp->reordering,
1505 sysctl_tcp_reordering);
1506 tcp_set_ca_state(sk, TCP_CA_Loss);
1507 tp->high_seq = tp->snd_nxt;
1508 TCP_ECN_queue_cwr(tp);
1509 /* Abort FRTO algorithm if one is in progress */
1510 tp->frto_counter = 0;
1512 clear_all_retrans_hints(tp);
1515 static int tcp_check_sack_reneging(struct sock *sk)
1517 struct sk_buff *skb;
1519 /* If ACK arrived pointing to a remembered SACK,
1520 * it means that our remembered SACKs do not reflect
1521 * real state of receiver i.e.
1522 * receiver _host_ is heavily congested (or buggy).
1523 * Do processing similar to RTO timeout.
1525 if ((skb = tcp_write_queue_head(sk)) != NULL &&
1526 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1527 struct inet_connection_sock *icsk = inet_csk(sk);
1528 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1530 tcp_enter_loss(sk, 1);
1531 icsk->icsk_retransmits++;
1532 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1533 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1534 icsk->icsk_rto, TCP_RTO_MAX);
1540 static inline int tcp_fackets_out(struct tcp_sock *tp)
1542 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1545 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1547 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1550 static inline int tcp_head_timedout(struct sock *sk)
1552 struct tcp_sock *tp = tcp_sk(sk);
1554 return tp->packets_out &&
1555 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1558 /* Linux NewReno/SACK/FACK/ECN state machine.
1559 * --------------------------------------
1561 * "Open" Normal state, no dubious events, fast path.
1562 * "Disorder" In all the respects it is "Open",
1563 * but requires a bit more attention. It is entered when
1564 * we see some SACKs or dupacks. It is split of "Open"
1565 * mainly to move some processing from fast path to slow one.
1566 * "CWR" CWND was reduced due to some Congestion Notification event.
1567 * It can be ECN, ICMP source quench, local device congestion.
1568 * "Recovery" CWND was reduced, we are fast-retransmitting.
1569 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1571 * tcp_fastretrans_alert() is entered:
1572 * - each incoming ACK, if state is not "Open"
1573 * - when arrived ACK is unusual, namely:
1578 * Counting packets in flight is pretty simple.
1580 * in_flight = packets_out - left_out + retrans_out
1582 * packets_out is SND.NXT-SND.UNA counted in packets.
1584 * retrans_out is number of retransmitted segments.
1586 * left_out is number of segments left network, but not ACKed yet.
1588 * left_out = sacked_out + lost_out
1590 * sacked_out: Packets, which arrived to receiver out of order
1591 * and hence not ACKed. With SACKs this number is simply
1592 * amount of SACKed data. Even without SACKs
1593 * it is easy to give pretty reliable estimate of this number,
1594 * counting duplicate ACKs.
1596 * lost_out: Packets lost by network. TCP has no explicit
1597 * "loss notification" feedback from network (for now).
1598 * It means that this number can be only _guessed_.
1599 * Actually, it is the heuristics to predict lossage that
1600 * distinguishes different algorithms.
1602 * F.e. after RTO, when all the queue is considered as lost,
1603 * lost_out = packets_out and in_flight = retrans_out.
1605 * Essentially, we have now two algorithms counting
1608 * FACK: It is the simplest heuristics. As soon as we decided
1609 * that something is lost, we decide that _all_ not SACKed
1610 * packets until the most forward SACK are lost. I.e.
1611 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1612 * It is absolutely correct estimate, if network does not reorder
1613 * packets. And it loses any connection to reality when reordering
1614 * takes place. We use FACK by default until reordering
1615 * is suspected on the path to this destination.
1617 * NewReno: when Recovery is entered, we assume that one segment
1618 * is lost (classic Reno). While we are in Recovery and
1619 * a partial ACK arrives, we assume that one more packet
1620 * is lost (NewReno). This heuristics are the same in NewReno
1623 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1624 * deflation etc. CWND is real congestion window, never inflated, changes
1625 * only according to classic VJ rules.
1627 * Really tricky (and requiring careful tuning) part of algorithm
1628 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1629 * The first determines the moment _when_ we should reduce CWND and,
1630 * hence, slow down forward transmission. In fact, it determines the moment
1631 * when we decide that hole is caused by loss, rather than by a reorder.
1633 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1634 * holes, caused by lost packets.
1636 * And the most logically complicated part of algorithm is undo
1637 * heuristics. We detect false retransmits due to both too early
1638 * fast retransmit (reordering) and underestimated RTO, analyzing
1639 * timestamps and D-SACKs. When we detect that some segments were
1640 * retransmitted by mistake and CWND reduction was wrong, we undo
1641 * window reduction and abort recovery phase. This logic is hidden
1642 * inside several functions named tcp_try_undo_<something>.
1645 /* This function decides, when we should leave Disordered state
1646 * and enter Recovery phase, reducing congestion window.
1648 * Main question: may we further continue forward transmission
1649 * with the same cwnd?
1651 static int tcp_time_to_recover(struct sock *sk)
1653 struct tcp_sock *tp = tcp_sk(sk);
1656 /* Do not perform any recovery during FRTO algorithm */
1657 if (tp->frto_counter)
1660 /* Trick#1: The loss is proven. */
1664 /* Not-A-Trick#2 : Classic rule... */
1665 if (tcp_fackets_out(tp) > tp->reordering)
1668 /* Trick#3 : when we use RFC2988 timer restart, fast
1669 * retransmit can be triggered by timeout of queue head.
1671 if (tcp_head_timedout(sk))
1674 /* Trick#4: It is still not OK... But will it be useful to delay
1677 packets_out = tp->packets_out;
1678 if (packets_out <= tp->reordering &&
1679 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1680 !tcp_may_send_now(sk)) {
1681 /* We have nothing to send. This connection is limited
1682 * either by receiver window or by application.
1690 /* If we receive more dupacks than we expected counting segments
1691 * in assumption of absent reordering, interpret this as reordering.
1692 * The only another reason could be bug in receiver TCP.
1694 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1696 struct tcp_sock *tp = tcp_sk(sk);
1699 holes = max(tp->lost_out, 1U);
1700 holes = min(holes, tp->packets_out);
1702 if ((tp->sacked_out + holes) > tp->packets_out) {
1703 tp->sacked_out = tp->packets_out - holes;
1704 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1708 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1710 static void tcp_add_reno_sack(struct sock *sk)
1712 struct tcp_sock *tp = tcp_sk(sk);
1714 tcp_check_reno_reordering(sk, 0);
1715 tcp_sync_left_out(tp);
1718 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1720 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1722 struct tcp_sock *tp = tcp_sk(sk);
1725 /* One ACK acked hole. The rest eat duplicate ACKs. */
1726 if (acked-1 >= tp->sacked_out)
1729 tp->sacked_out -= acked-1;
1731 tcp_check_reno_reordering(sk, acked);
1732 tcp_sync_left_out(tp);
1735 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1738 tp->left_out = tp->lost_out;
1741 /* Mark head of queue up as lost. */
1742 static void tcp_mark_head_lost(struct sock *sk,
1743 int packets, u32 high_seq)
1745 struct tcp_sock *tp = tcp_sk(sk);
1746 struct sk_buff *skb;
1749 BUG_TRAP(packets <= tp->packets_out);
1750 if (tp->lost_skb_hint) {
1751 skb = tp->lost_skb_hint;
1752 cnt = tp->lost_cnt_hint;
1754 skb = tcp_write_queue_head(sk);
1758 tcp_for_write_queue_from(skb, sk) {
1759 if (skb == tcp_send_head(sk))
1761 /* TODO: do this better */
1762 /* this is not the most efficient way to do this... */
1763 tp->lost_skb_hint = skb;
1764 tp->lost_cnt_hint = cnt;
1765 cnt += tcp_skb_pcount(skb);
1766 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1768 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1769 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1770 tp->lost_out += tcp_skb_pcount(skb);
1772 /* clear xmit_retransmit_queue hints
1773 * if this is beyond hint */
1774 if (tp->retransmit_skb_hint != NULL &&
1775 before(TCP_SKB_CB(skb)->seq,
1776 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1777 tp->retransmit_skb_hint = NULL;
1781 tcp_sync_left_out(tp);
1784 /* Account newly detected lost packet(s) */
1786 static void tcp_update_scoreboard(struct sock *sk)
1788 struct tcp_sock *tp = tcp_sk(sk);
1791 int lost = tp->fackets_out - tp->reordering;
1794 tcp_mark_head_lost(sk, lost, tp->high_seq);
1796 tcp_mark_head_lost(sk, 1, tp->high_seq);
1799 /* New heuristics: it is possible only after we switched
1800 * to restart timer each time when something is ACKed.
1801 * Hence, we can detect timed out packets during fast
1802 * retransmit without falling to slow start.
1804 if (!IsReno(tp) && tcp_head_timedout(sk)) {
1805 struct sk_buff *skb;
1807 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1808 : tcp_write_queue_head(sk);
1810 tcp_for_write_queue_from(skb, sk) {
1811 if (skb == tcp_send_head(sk))
1813 if (!tcp_skb_timedout(sk, skb))
1816 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1817 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1818 tp->lost_out += tcp_skb_pcount(skb);
1820 /* clear xmit_retrans hint */
1821 if (tp->retransmit_skb_hint &&
1822 before(TCP_SKB_CB(skb)->seq,
1823 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1825 tp->retransmit_skb_hint = NULL;
1829 tp->scoreboard_skb_hint = skb;
1831 tcp_sync_left_out(tp);
1835 /* CWND moderation, preventing bursts due to too big ACKs
1836 * in dubious situations.
1838 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1840 tp->snd_cwnd = min(tp->snd_cwnd,
1841 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1842 tp->snd_cwnd_stamp = tcp_time_stamp;
1845 /* Lower bound on congestion window is slow start threshold
1846 * unless congestion avoidance choice decides to overide it.
1848 static inline u32 tcp_cwnd_min(const struct sock *sk)
1850 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1852 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1855 /* Decrease cwnd each second ack. */
1856 static void tcp_cwnd_down(struct sock *sk, int flag)
1858 struct tcp_sock *tp = tcp_sk(sk);
1859 int decr = tp->snd_cwnd_cnt + 1;
1861 if ((flag&FLAG_ANY_PROGRESS) ||
1862 (IsReno(tp) && !(flag&FLAG_NOT_DUP))) {
1863 tp->snd_cwnd_cnt = decr&1;
1866 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1867 tp->snd_cwnd -= decr;
1869 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1870 tp->snd_cwnd_stamp = tcp_time_stamp;
1874 /* Nothing was retransmitted or returned timestamp is less
1875 * than timestamp of the first retransmission.
1877 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1879 return !tp->retrans_stamp ||
1880 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1881 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1884 /* Undo procedures. */
1886 #if FASTRETRANS_DEBUG > 1
1887 static void DBGUNDO(struct sock *sk, const char *msg)
1889 struct tcp_sock *tp = tcp_sk(sk);
1890 struct inet_sock *inet = inet_sk(sk);
1892 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1894 NIPQUAD(inet->daddr), ntohs(inet->dport),
1895 tp->snd_cwnd, tp->left_out,
1896 tp->snd_ssthresh, tp->prior_ssthresh,
1900 #define DBGUNDO(x...) do { } while (0)
1903 static void tcp_undo_cwr(struct sock *sk, const int undo)
1905 struct tcp_sock *tp = tcp_sk(sk);
1907 if (tp->prior_ssthresh) {
1908 const struct inet_connection_sock *icsk = inet_csk(sk);
1910 if (icsk->icsk_ca_ops->undo_cwnd)
1911 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1913 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1915 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1916 tp->snd_ssthresh = tp->prior_ssthresh;
1917 TCP_ECN_withdraw_cwr(tp);
1920 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1922 tcp_moderate_cwnd(tp);
1923 tp->snd_cwnd_stamp = tcp_time_stamp;
1925 /* There is something screwy going on with the retrans hints after
1927 clear_all_retrans_hints(tp);
1930 static inline int tcp_may_undo(struct tcp_sock *tp)
1932 return tp->undo_marker &&
1933 (!tp->undo_retrans || tcp_packet_delayed(tp));
1936 /* People celebrate: "We love our President!" */
1937 static int tcp_try_undo_recovery(struct sock *sk)
1939 struct tcp_sock *tp = tcp_sk(sk);
1941 if (tcp_may_undo(tp)) {
1942 /* Happy end! We did not retransmit anything
1943 * or our original transmission succeeded.
1945 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1946 tcp_undo_cwr(sk, 1);
1947 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1948 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1950 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1951 tp->undo_marker = 0;
1953 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1954 /* Hold old state until something *above* high_seq
1955 * is ACKed. For Reno it is MUST to prevent false
1956 * fast retransmits (RFC2582). SACK TCP is safe. */
1957 tcp_moderate_cwnd(tp);
1960 tcp_set_ca_state(sk, TCP_CA_Open);
1964 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1965 static void tcp_try_undo_dsack(struct sock *sk)
1967 struct tcp_sock *tp = tcp_sk(sk);
1969 if (tp->undo_marker && !tp->undo_retrans) {
1970 DBGUNDO(sk, "D-SACK");
1971 tcp_undo_cwr(sk, 1);
1972 tp->undo_marker = 0;
1973 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1977 /* Undo during fast recovery after partial ACK. */
1979 static int tcp_try_undo_partial(struct sock *sk, int acked)
1981 struct tcp_sock *tp = tcp_sk(sk);
1982 /* Partial ACK arrived. Force Hoe's retransmit. */
1983 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1985 if (tcp_may_undo(tp)) {
1986 /* Plain luck! Hole if filled with delayed
1987 * packet, rather than with a retransmit.
1989 if (tp->retrans_out == 0)
1990 tp->retrans_stamp = 0;
1992 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1995 tcp_undo_cwr(sk, 0);
1996 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1998 /* So... Do not make Hoe's retransmit yet.
1999 * If the first packet was delayed, the rest
2000 * ones are most probably delayed as well.
2007 /* Undo during loss recovery after partial ACK. */
2008 static int tcp_try_undo_loss(struct sock *sk)
2010 struct tcp_sock *tp = tcp_sk(sk);
2012 if (tcp_may_undo(tp)) {
2013 struct sk_buff *skb;
2014 tcp_for_write_queue(skb, sk) {
2015 if (skb == tcp_send_head(sk))
2017 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2020 clear_all_retrans_hints(tp);
2022 DBGUNDO(sk, "partial loss");
2024 tp->left_out = tp->sacked_out;
2025 tcp_undo_cwr(sk, 1);
2026 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2027 inet_csk(sk)->icsk_retransmits = 0;
2028 tp->undo_marker = 0;
2030 tcp_set_ca_state(sk, TCP_CA_Open);
2036 static inline void tcp_complete_cwr(struct sock *sk)
2038 struct tcp_sock *tp = tcp_sk(sk);
2039 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2040 tp->snd_cwnd_stamp = tcp_time_stamp;
2041 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2044 static void tcp_try_to_open(struct sock *sk, int flag)
2046 struct tcp_sock *tp = tcp_sk(sk);
2048 tcp_sync_left_out(tp);
2050 if (tp->retrans_out == 0)
2051 tp->retrans_stamp = 0;
2054 tcp_enter_cwr(sk, 1);
2056 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2057 int state = TCP_CA_Open;
2059 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2060 state = TCP_CA_Disorder;
2062 if (inet_csk(sk)->icsk_ca_state != state) {
2063 tcp_set_ca_state(sk, state);
2064 tp->high_seq = tp->snd_nxt;
2066 tcp_moderate_cwnd(tp);
2068 tcp_cwnd_down(sk, flag);
2072 static void tcp_mtup_probe_failed(struct sock *sk)
2074 struct inet_connection_sock *icsk = inet_csk(sk);
2076 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2077 icsk->icsk_mtup.probe_size = 0;
2080 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2082 struct tcp_sock *tp = tcp_sk(sk);
2083 struct inet_connection_sock *icsk = inet_csk(sk);
2085 /* FIXME: breaks with very large cwnd */
2086 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2087 tp->snd_cwnd = tp->snd_cwnd *
2088 tcp_mss_to_mtu(sk, tp->mss_cache) /
2089 icsk->icsk_mtup.probe_size;
2090 tp->snd_cwnd_cnt = 0;
2091 tp->snd_cwnd_stamp = tcp_time_stamp;
2092 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2094 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2095 icsk->icsk_mtup.probe_size = 0;
2096 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2100 /* Process an event, which can update packets-in-flight not trivially.
2101 * Main goal of this function is to calculate new estimate for left_out,
2102 * taking into account both packets sitting in receiver's buffer and
2103 * packets lost by network.
2105 * Besides that it does CWND reduction, when packet loss is detected
2106 * and changes state of machine.
2108 * It does _not_ decide what to send, it is made in function
2109 * tcp_xmit_retransmit_queue().
2112 tcp_fastretrans_alert(struct sock *sk, int prior_packets, int flag)
2114 struct inet_connection_sock *icsk = inet_csk(sk);
2115 struct tcp_sock *tp = tcp_sk(sk);
2116 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2117 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2118 (tp->fackets_out > tp->reordering));
2120 /* Some technical things:
2121 * 1. Reno does not count dupacks (sacked_out) automatically. */
2122 if (!tp->packets_out)
2124 /* 2. SACK counts snd_fack in packets inaccurately. */
2125 if (tp->sacked_out == 0)
2126 tp->fackets_out = 0;
2128 /* Now state machine starts.
2129 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2131 tp->prior_ssthresh = 0;
2133 /* B. In all the states check for reneging SACKs. */
2134 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2137 /* C. Process data loss notification, provided it is valid. */
2138 if ((flag&FLAG_DATA_LOST) &&
2139 before(tp->snd_una, tp->high_seq) &&
2140 icsk->icsk_ca_state != TCP_CA_Open &&
2141 tp->fackets_out > tp->reordering) {
2142 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2143 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2146 /* D. Synchronize left_out to current state. */
2147 tcp_sync_left_out(tp);
2149 /* E. Check state exit conditions. State can be terminated
2150 * when high_seq is ACKed. */
2151 if (icsk->icsk_ca_state == TCP_CA_Open) {
2152 BUG_TRAP(tp->retrans_out == 0);
2153 tp->retrans_stamp = 0;
2154 } else if (!before(tp->snd_una, tp->high_seq)) {
2155 switch (icsk->icsk_ca_state) {
2157 icsk->icsk_retransmits = 0;
2158 if (tcp_try_undo_recovery(sk))
2163 /* CWR is to be held something *above* high_seq
2164 * is ACKed for CWR bit to reach receiver. */
2165 if (tp->snd_una != tp->high_seq) {
2166 tcp_complete_cwr(sk);
2167 tcp_set_ca_state(sk, TCP_CA_Open);
2171 case TCP_CA_Disorder:
2172 tcp_try_undo_dsack(sk);
2173 if (!tp->undo_marker ||
2174 /* For SACK case do not Open to allow to undo
2175 * catching for all duplicate ACKs. */
2176 IsReno(tp) || tp->snd_una != tp->high_seq) {
2177 tp->undo_marker = 0;
2178 tcp_set_ca_state(sk, TCP_CA_Open);
2182 case TCP_CA_Recovery:
2184 tcp_reset_reno_sack(tp);
2185 if (tcp_try_undo_recovery(sk))
2187 tcp_complete_cwr(sk);
2192 /* F. Process state. */
2193 switch (icsk->icsk_ca_state) {
2194 case TCP_CA_Recovery:
2195 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2196 if (IsReno(tp) && is_dupack)
2197 tcp_add_reno_sack(sk);
2199 int acked = prior_packets - tp->packets_out;
2201 tcp_remove_reno_sacks(sk, acked);
2202 do_lost = tcp_try_undo_partial(sk, acked);
2206 if (flag&FLAG_DATA_ACKED)
2207 icsk->icsk_retransmits = 0;
2208 if (!tcp_try_undo_loss(sk)) {
2209 tcp_moderate_cwnd(tp);
2210 tcp_xmit_retransmit_queue(sk);
2213 if (icsk->icsk_ca_state != TCP_CA_Open)
2215 /* Loss is undone; fall through to processing in Open state. */
2218 if (flag & FLAG_SND_UNA_ADVANCED)
2219 tcp_reset_reno_sack(tp);
2221 tcp_add_reno_sack(sk);
2224 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2225 tcp_try_undo_dsack(sk);
2227 if (!tcp_time_to_recover(sk)) {
2228 tcp_try_to_open(sk, flag);
2232 /* MTU probe failure: don't reduce cwnd */
2233 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2234 icsk->icsk_mtup.probe_size &&
2235 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2236 tcp_mtup_probe_failed(sk);
2237 /* Restores the reduction we did in tcp_mtup_probe() */
2239 tcp_simple_retransmit(sk);
2243 /* Otherwise enter Recovery state */
2246 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2248 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2250 tp->high_seq = tp->snd_nxt;
2251 tp->prior_ssthresh = 0;
2252 tp->undo_marker = tp->snd_una;
2253 tp->undo_retrans = tp->retrans_out;
2255 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2256 if (!(flag&FLAG_ECE))
2257 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2258 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2259 TCP_ECN_queue_cwr(tp);
2262 tp->bytes_acked = 0;
2263 tp->snd_cwnd_cnt = 0;
2264 tcp_set_ca_state(sk, TCP_CA_Recovery);
2267 if (do_lost || tcp_head_timedout(sk))
2268 tcp_update_scoreboard(sk);
2269 tcp_cwnd_down(sk, flag);
2270 tcp_xmit_retransmit_queue(sk);
2273 /* Read draft-ietf-tcplw-high-performance before mucking
2274 * with this code. (Supersedes RFC1323)
2276 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2278 /* RTTM Rule: A TSecr value received in a segment is used to
2279 * update the averaged RTT measurement only if the segment
2280 * acknowledges some new data, i.e., only if it advances the
2281 * left edge of the send window.
2283 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2284 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2286 * Changed: reset backoff as soon as we see the first valid sample.
2287 * If we do not, we get strongly overestimated rto. With timestamps
2288 * samples are accepted even from very old segments: f.e., when rtt=1
2289 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2290 * answer arrives rto becomes 120 seconds! If at least one of segments
2291 * in window is lost... Voila. --ANK (010210)
2293 struct tcp_sock *tp = tcp_sk(sk);
2294 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2295 tcp_rtt_estimator(sk, seq_rtt);
2297 inet_csk(sk)->icsk_backoff = 0;
2301 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2303 /* We don't have a timestamp. Can only use
2304 * packets that are not retransmitted to determine
2305 * rtt estimates. Also, we must not reset the
2306 * backoff for rto until we get a non-retransmitted
2307 * packet. This allows us to deal with a situation
2308 * where the network delay has increased suddenly.
2309 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2312 if (flag & FLAG_RETRANS_DATA_ACKED)
2315 tcp_rtt_estimator(sk, seq_rtt);
2317 inet_csk(sk)->icsk_backoff = 0;
2321 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2324 const struct tcp_sock *tp = tcp_sk(sk);
2325 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2326 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2327 tcp_ack_saw_tstamp(sk, flag);
2328 else if (seq_rtt >= 0)
2329 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2332 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2333 u32 in_flight, int good)
2335 const struct inet_connection_sock *icsk = inet_csk(sk);
2336 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2337 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2340 /* Restart timer after forward progress on connection.
2341 * RFC2988 recommends to restart timer to now+rto.
2344 static void tcp_ack_packets_out(struct sock *sk)
2346 struct tcp_sock *tp = tcp_sk(sk);
2348 if (!tp->packets_out) {
2349 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2351 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2355 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2356 __u32 now, __s32 *seq_rtt)
2358 struct tcp_sock *tp = tcp_sk(sk);
2359 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2360 __u32 seq = tp->snd_una;
2361 __u32 packets_acked;
2364 /* If we get here, the whole TSO packet has not been
2367 BUG_ON(!after(scb->end_seq, seq));
2369 packets_acked = tcp_skb_pcount(skb);
2370 if (tcp_trim_head(sk, skb, seq - scb->seq))
2372 packets_acked -= tcp_skb_pcount(skb);
2374 if (packets_acked) {
2375 __u8 sacked = scb->sacked;
2377 acked |= FLAG_DATA_ACKED;
2379 if (sacked & TCPCB_RETRANS) {
2380 if (sacked & TCPCB_SACKED_RETRANS)
2381 tp->retrans_out -= packets_acked;
2382 acked |= FLAG_RETRANS_DATA_ACKED;
2384 } else if (*seq_rtt < 0)
2385 *seq_rtt = now - scb->when;
2386 if (sacked & TCPCB_SACKED_ACKED)
2387 tp->sacked_out -= packets_acked;
2388 if (sacked & TCPCB_LOST)
2389 tp->lost_out -= packets_acked;
2390 if (sacked & TCPCB_URG) {
2392 !before(seq, tp->snd_up))
2395 } else if (*seq_rtt < 0)
2396 *seq_rtt = now - scb->when;
2398 if (tp->fackets_out) {
2399 __u32 dval = min(tp->fackets_out, packets_acked);
2400 tp->fackets_out -= dval;
2402 tp->packets_out -= packets_acked;
2404 BUG_ON(tcp_skb_pcount(skb) == 0);
2405 BUG_ON(!before(scb->seq, scb->end_seq));
2411 /* Remove acknowledged frames from the retransmission queue. */
2412 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2414 struct tcp_sock *tp = tcp_sk(sk);
2415 const struct inet_connection_sock *icsk = inet_csk(sk);
2416 struct sk_buff *skb;
2417 __u32 now = tcp_time_stamp;
2419 int prior_packets = tp->packets_out;
2421 ktime_t last_ackt = net_invalid_timestamp();
2423 while ((skb = tcp_write_queue_head(sk)) &&
2424 skb != tcp_send_head(sk)) {
2425 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2426 __u8 sacked = scb->sacked;
2428 /* If our packet is before the ack sequence we can
2429 * discard it as it's confirmed to have arrived at
2432 if (after(scb->end_seq, tp->snd_una)) {
2433 if (tcp_skb_pcount(skb) > 1 &&
2434 after(tp->snd_una, scb->seq))
2435 acked |= tcp_tso_acked(sk, skb,
2440 /* Initial outgoing SYN's get put onto the write_queue
2441 * just like anything else we transmit. It is not
2442 * true data, and if we misinform our callers that
2443 * this ACK acks real data, we will erroneously exit
2444 * connection startup slow start one packet too
2445 * quickly. This is severely frowned upon behavior.
2447 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2448 acked |= FLAG_DATA_ACKED;
2450 acked |= FLAG_SYN_ACKED;
2451 tp->retrans_stamp = 0;
2454 /* MTU probing checks */
2455 if (icsk->icsk_mtup.probe_size) {
2456 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2457 tcp_mtup_probe_success(sk, skb);
2462 if (sacked & TCPCB_RETRANS) {
2463 if (sacked & TCPCB_SACKED_RETRANS)
2464 tp->retrans_out -= tcp_skb_pcount(skb);
2465 acked |= FLAG_RETRANS_DATA_ACKED;
2467 } else if (seq_rtt < 0) {
2468 seq_rtt = now - scb->when;
2469 last_ackt = skb->tstamp;
2471 if (sacked & TCPCB_SACKED_ACKED)
2472 tp->sacked_out -= tcp_skb_pcount(skb);
2473 if (sacked & TCPCB_LOST)
2474 tp->lost_out -= tcp_skb_pcount(skb);
2475 if (sacked & TCPCB_URG) {
2477 !before(scb->end_seq, tp->snd_up))
2480 } else if (seq_rtt < 0) {
2481 seq_rtt = now - scb->when;
2482 last_ackt = skb->tstamp;
2484 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2485 tcp_packets_out_dec(tp, skb);
2486 tcp_unlink_write_queue(skb, sk);
2487 sk_stream_free_skb(sk, skb);
2488 clear_all_retrans_hints(tp);
2491 if (acked&FLAG_ACKED) {
2492 u32 pkts_acked = prior_packets - tp->packets_out;
2493 const struct tcp_congestion_ops *ca_ops
2494 = inet_csk(sk)->icsk_ca_ops;
2496 tcp_ack_update_rtt(sk, acked, seq_rtt);
2497 tcp_ack_packets_out(sk);
2499 if (ca_ops->pkts_acked) {
2502 /* Is the ACK triggering packet unambiguous? */
2503 if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
2504 /* High resolution needed and available? */
2505 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2506 !ktime_equal(last_ackt,
2507 net_invalid_timestamp()))
2508 rtt_us = ktime_us_delta(ktime_get_real(),
2510 else if (seq_rtt > 0)
2511 rtt_us = jiffies_to_usecs(seq_rtt);
2514 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2518 #if FASTRETRANS_DEBUG > 0
2519 BUG_TRAP((int)tp->sacked_out >= 0);
2520 BUG_TRAP((int)tp->lost_out >= 0);
2521 BUG_TRAP((int)tp->retrans_out >= 0);
2522 if (!tp->packets_out && tp->rx_opt.sack_ok) {
2523 const struct inet_connection_sock *icsk = inet_csk(sk);
2525 printk(KERN_DEBUG "Leak l=%u %d\n",
2526 tp->lost_out, icsk->icsk_ca_state);
2529 if (tp->sacked_out) {
2530 printk(KERN_DEBUG "Leak s=%u %d\n",
2531 tp->sacked_out, icsk->icsk_ca_state);
2534 if (tp->retrans_out) {
2535 printk(KERN_DEBUG "Leak r=%u %d\n",
2536 tp->retrans_out, icsk->icsk_ca_state);
2537 tp->retrans_out = 0;
2541 *seq_rtt_p = seq_rtt;
2545 static void tcp_ack_probe(struct sock *sk)
2547 const struct tcp_sock *tp = tcp_sk(sk);
2548 struct inet_connection_sock *icsk = inet_csk(sk);
2550 /* Was it a usable window open? */
2552 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2553 tp->snd_una + tp->snd_wnd)) {
2554 icsk->icsk_backoff = 0;
2555 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2556 /* Socket must be waked up by subsequent tcp_data_snd_check().
2557 * This function is not for random using!
2560 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2561 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2566 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2568 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2569 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2572 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2574 const struct tcp_sock *tp = tcp_sk(sk);
2575 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2576 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2579 /* Check that window update is acceptable.
2580 * The function assumes that snd_una<=ack<=snd_next.
2582 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2583 const u32 ack_seq, const u32 nwin)
2585 return (after(ack, tp->snd_una) ||
2586 after(ack_seq, tp->snd_wl1) ||
2587 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2590 /* Update our send window.
2592 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2593 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2595 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2598 struct tcp_sock *tp = tcp_sk(sk);
2600 u32 nwin = ntohs(tcp_hdr(skb)->window);
2602 if (likely(!tcp_hdr(skb)->syn))
2603 nwin <<= tp->rx_opt.snd_wscale;
2605 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2606 flag |= FLAG_WIN_UPDATE;
2607 tcp_update_wl(tp, ack, ack_seq);
2609 if (tp->snd_wnd != nwin) {
2612 /* Note, it is the only place, where
2613 * fast path is recovered for sending TCP.
2616 tcp_fast_path_check(sk);
2618 if (nwin > tp->max_window) {
2619 tp->max_window = nwin;
2620 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2630 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2631 * continue in congestion avoidance.
2633 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2635 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2636 tp->snd_cwnd_cnt = 0;
2637 TCP_ECN_queue_cwr(tp);
2638 tcp_moderate_cwnd(tp);
2641 /* A conservative spurious RTO response algorithm: reduce cwnd using
2642 * rate halving and continue in congestion avoidance.
2644 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2646 tcp_enter_cwr(sk, 0);
2649 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2652 tcp_ratehalving_spur_to_response(sk);
2654 tcp_undo_cwr(sk, 1);
2657 /* F-RTO spurious RTO detection algorithm (RFC4138)
2659 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2660 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2661 * window (but not to or beyond highest sequence sent before RTO):
2662 * On First ACK, send two new segments out.
2663 * On Second ACK, RTO was likely spurious. Do spurious response (response
2664 * algorithm is not part of the F-RTO detection algorithm
2665 * given in RFC4138 but can be selected separately).
2666 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2667 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2668 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2669 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2671 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2672 * original window even after we transmit two new data segments.
2675 * on first step, wait until first cumulative ACK arrives, then move to
2676 * the second step. In second step, the next ACK decides.
2678 * F-RTO is implemented (mainly) in four functions:
2679 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2680 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2681 * called when tcp_use_frto() showed green light
2682 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2683 * - tcp_enter_frto_loss() is called if there is not enough evidence
2684 * to prove that the RTO is indeed spurious. It transfers the control
2685 * from F-RTO to the conventional RTO recovery
2687 static int tcp_process_frto(struct sock *sk, int flag)
2689 struct tcp_sock *tp = tcp_sk(sk);
2691 tcp_sync_left_out(tp);
2693 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2694 if (flag&FLAG_DATA_ACKED)
2695 inet_csk(sk)->icsk_retransmits = 0;
2697 if (!before(tp->snd_una, tp->frto_highmark)) {
2698 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2702 if (!IsSackFrto() || IsReno(tp)) {
2703 /* RFC4138 shortcoming in step 2; should also have case c):
2704 * ACK isn't duplicate nor advances window, e.g., opposite dir
2707 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2710 if (!(flag&FLAG_DATA_ACKED)) {
2711 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2716 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2717 /* Prevent sending of new data. */
2718 tp->snd_cwnd = min(tp->snd_cwnd,
2719 tcp_packets_in_flight(tp));
2723 if ((tp->frto_counter >= 2) &&
2724 (!(flag&FLAG_FORWARD_PROGRESS) ||
2725 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2726 /* RFC4138 shortcoming (see comment above) */
2727 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2730 tcp_enter_frto_loss(sk, 3, flag);
2735 if (tp->frto_counter == 1) {
2736 /* Sending of the next skb must be allowed or no FRTO */
2737 if (!tcp_send_head(sk) ||
2738 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2739 tp->snd_una + tp->snd_wnd)) {
2740 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2745 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2746 tp->frto_counter = 2;
2749 switch (sysctl_tcp_frto_response) {
2751 tcp_undo_spur_to_response(sk, flag);
2754 tcp_conservative_spur_to_response(tp);
2757 tcp_ratehalving_spur_to_response(sk);
2760 tp->frto_counter = 0;
2765 /* This routine deals with incoming acks, but not outgoing ones. */
2766 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2768 struct inet_connection_sock *icsk = inet_csk(sk);
2769 struct tcp_sock *tp = tcp_sk(sk);
2770 u32 prior_snd_una = tp->snd_una;
2771 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2772 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2773 u32 prior_in_flight;
2778 /* If the ack is newer than sent or older than previous acks
2779 * then we can probably ignore it.
2781 if (after(ack, tp->snd_nxt))
2782 goto uninteresting_ack;
2784 if (before(ack, prior_snd_una))
2787 if (after(ack, prior_snd_una))
2788 flag |= FLAG_SND_UNA_ADVANCED;
2790 if (sysctl_tcp_abc) {
2791 if (icsk->icsk_ca_state < TCP_CA_CWR)
2792 tp->bytes_acked += ack - prior_snd_una;
2793 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2794 /* we assume just one segment left network */
2795 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2798 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2799 /* Window is constant, pure forward advance.
2800 * No more checks are required.
2801 * Note, we use the fact that SND.UNA>=SND.WL2.
2803 tcp_update_wl(tp, ack, ack_seq);
2805 flag |= FLAG_WIN_UPDATE;
2807 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2809 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2811 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2814 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2816 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2818 if (TCP_SKB_CB(skb)->sacked)
2819 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2821 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2824 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2827 /* We passed data and got it acked, remove any soft error
2828 * log. Something worked...
2830 sk->sk_err_soft = 0;
2831 tp->rcv_tstamp = tcp_time_stamp;
2832 prior_packets = tp->packets_out;
2836 prior_in_flight = tcp_packets_in_flight(tp);
2838 /* See if we can take anything off of the retransmit queue. */
2839 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2841 if (tp->frto_counter)
2842 frto_cwnd = tcp_process_frto(sk, flag);
2844 if (tcp_ack_is_dubious(sk, flag)) {
2845 /* Advance CWND, if state allows this. */
2846 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2847 tcp_may_raise_cwnd(sk, flag))
2848 tcp_cong_avoid(sk, ack, prior_in_flight, 0);
2849 tcp_fastretrans_alert(sk, prior_packets, flag);
2851 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2852 tcp_cong_avoid(sk, ack, prior_in_flight, 1);
2855 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2856 dst_confirm(sk->sk_dst_cache);
2861 icsk->icsk_probes_out = 0;
2863 /* If this ack opens up a zero window, clear backoff. It was
2864 * being used to time the probes, and is probably far higher than
2865 * it needs to be for normal retransmission.
2867 if (tcp_send_head(sk))
2872 if (TCP_SKB_CB(skb)->sacked)
2873 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2876 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2881 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2882 * But, this can also be called on packets in the established flow when
2883 * the fast version below fails.
2885 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2888 struct tcphdr *th = tcp_hdr(skb);
2889 int length=(th->doff*4)-sizeof(struct tcphdr);
2891 ptr = (unsigned char *)(th + 1);
2892 opt_rx->saw_tstamp = 0;
2894 while (length > 0) {
2901 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2906 if (opsize < 2) /* "silly options" */
2908 if (opsize > length)
2909 return; /* don't parse partial options */
2912 if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2913 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2915 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2916 in_mss = opt_rx->user_mss;
2917 opt_rx->mss_clamp = in_mss;
2922 if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2923 if (sysctl_tcp_window_scaling) {
2924 __u8 snd_wscale = *(__u8 *) ptr;
2925 opt_rx->wscale_ok = 1;
2926 if (snd_wscale > 14) {
2927 if (net_ratelimit())
2928 printk(KERN_INFO "tcp_parse_options: Illegal window "
2929 "scaling value %d >14 received.\n",
2933 opt_rx->snd_wscale = snd_wscale;
2936 case TCPOPT_TIMESTAMP:
2937 if (opsize==TCPOLEN_TIMESTAMP) {
2938 if ((estab && opt_rx->tstamp_ok) ||
2939 (!estab && sysctl_tcp_timestamps)) {
2940 opt_rx->saw_tstamp = 1;
2941 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2942 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2946 case TCPOPT_SACK_PERM:
2947 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2948 if (sysctl_tcp_sack) {
2949 opt_rx->sack_ok = 1;
2950 tcp_sack_reset(opt_rx);
2956 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2957 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2959 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2962 #ifdef CONFIG_TCP_MD5SIG
2965 * The MD5 Hash has already been
2966 * checked (see tcp_v{4,6}_do_rcv()).
2978 /* Fast parse options. This hopes to only see timestamps.
2979 * If it is wrong it falls back on tcp_parse_options().
2981 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2982 struct tcp_sock *tp)
2984 if (th->doff == sizeof(struct tcphdr)>>2) {
2985 tp->rx_opt.saw_tstamp = 0;
2987 } else if (tp->rx_opt.tstamp_ok &&
2988 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2989 __be32 *ptr = (__be32 *)(th + 1);
2990 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2991 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2992 tp->rx_opt.saw_tstamp = 1;
2994 tp->rx_opt.rcv_tsval = ntohl(*ptr);
2996 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3000 tcp_parse_options(skb, &tp->rx_opt, 1);
3004 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3006 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3007 tp->rx_opt.ts_recent_stamp = get_seconds();
3010 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3012 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3013 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3014 * extra check below makes sure this can only happen
3015 * for pure ACK frames. -DaveM
3017 * Not only, also it occurs for expired timestamps.
3020 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3021 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3022 tcp_store_ts_recent(tp);
3026 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3028 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3029 * it can pass through stack. So, the following predicate verifies that
3030 * this segment is not used for anything but congestion avoidance or
3031 * fast retransmit. Moreover, we even are able to eliminate most of such
3032 * second order effects, if we apply some small "replay" window (~RTO)
3033 * to timestamp space.
3035 * All these measures still do not guarantee that we reject wrapped ACKs
3036 * on networks with high bandwidth, when sequence space is recycled fastly,
3037 * but it guarantees that such events will be very rare and do not affect
3038 * connection seriously. This doesn't look nice, but alas, PAWS is really
3041 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3042 * states that events when retransmit arrives after original data are rare.
3043 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3044 * the biggest problem on large power networks even with minor reordering.
3045 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3046 * up to bandwidth of 18Gigabit/sec. 8) ]
3049 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3051 struct tcp_sock *tp = tcp_sk(sk);
3052 struct tcphdr *th = tcp_hdr(skb);
3053 u32 seq = TCP_SKB_CB(skb)->seq;
3054 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3056 return (/* 1. Pure ACK with correct sequence number. */
3057 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3059 /* 2. ... and duplicate ACK. */
3060 ack == tp->snd_una &&
3062 /* 3. ... and does not update window. */
3063 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3065 /* 4. ... and sits in replay window. */
3066 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3069 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3071 const struct tcp_sock *tp = tcp_sk(sk);
3072 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3073 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3074 !tcp_disordered_ack(sk, skb));
3077 /* Check segment sequence number for validity.
3079 * Segment controls are considered valid, if the segment
3080 * fits to the window after truncation to the window. Acceptability
3081 * of data (and SYN, FIN, of course) is checked separately.
3082 * See tcp_data_queue(), for example.
3084 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3085 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3086 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3087 * (borrowed from freebsd)
3090 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3092 return !before(end_seq, tp->rcv_wup) &&
3093 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3096 /* When we get a reset we do this. */
3097 static void tcp_reset(struct sock *sk)
3099 /* We want the right error as BSD sees it (and indeed as we do). */
3100 switch (sk->sk_state) {
3102 sk->sk_err = ECONNREFUSED;
3104 case TCP_CLOSE_WAIT:
3110 sk->sk_err = ECONNRESET;
3113 if (!sock_flag(sk, SOCK_DEAD))
3114 sk->sk_error_report(sk);
3120 * Process the FIN bit. This now behaves as it is supposed to work
3121 * and the FIN takes effect when it is validly part of sequence
3122 * space. Not before when we get holes.
3124 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3125 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3128 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3129 * close and we go into CLOSING (and later onto TIME-WAIT)
3131 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3133 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3135 struct tcp_sock *tp = tcp_sk(sk);
3137 inet_csk_schedule_ack(sk);
3139 sk->sk_shutdown |= RCV_SHUTDOWN;
3140 sock_set_flag(sk, SOCK_DONE);
3142 switch (sk->sk_state) {
3144 case TCP_ESTABLISHED:
3145 /* Move to CLOSE_WAIT */
3146 tcp_set_state(sk, TCP_CLOSE_WAIT);
3147 inet_csk(sk)->icsk_ack.pingpong = 1;
3150 case TCP_CLOSE_WAIT:
3152 /* Received a retransmission of the FIN, do
3157 /* RFC793: Remain in the LAST-ACK state. */
3161 /* This case occurs when a simultaneous close
3162 * happens, we must ack the received FIN and
3163 * enter the CLOSING state.
3166 tcp_set_state(sk, TCP_CLOSING);
3169 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3171 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3174 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3175 * cases we should never reach this piece of code.
3177 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3178 __FUNCTION__, sk->sk_state);
3182 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3183 * Probably, we should reset in this case. For now drop them.
3185 __skb_queue_purge(&tp->out_of_order_queue);
3186 if (tp->rx_opt.sack_ok)
3187 tcp_sack_reset(&tp->rx_opt);
3188 sk_stream_mem_reclaim(sk);
3190 if (!sock_flag(sk, SOCK_DEAD)) {
3191 sk->sk_state_change(sk);
3193 /* Do not send POLL_HUP for half duplex close. */
3194 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3195 sk->sk_state == TCP_CLOSE)
3196 sk_wake_async(sk, 1, POLL_HUP);
3198 sk_wake_async(sk, 1, POLL_IN);
3202 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3204 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3205 if (before(seq, sp->start_seq))
3206 sp->start_seq = seq;
3207 if (after(end_seq, sp->end_seq))
3208 sp->end_seq = end_seq;
3214 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3216 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3217 if (before(seq, tp->rcv_nxt))
3218 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3220 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3222 tp->rx_opt.dsack = 1;
3223 tp->duplicate_sack[0].start_seq = seq;
3224 tp->duplicate_sack[0].end_seq = end_seq;
3225 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3229 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3231 if (!tp->rx_opt.dsack)
3232 tcp_dsack_set(tp, seq, end_seq);
3234 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3237 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3239 struct tcp_sock *tp = tcp_sk(sk);
3241 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3242 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3243 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3244 tcp_enter_quickack_mode(sk);
3246 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3247 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3249 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3250 end_seq = tp->rcv_nxt;
3251 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3258 /* These routines update the SACK block as out-of-order packets arrive or
3259 * in-order packets close up the sequence space.
3261 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3264 struct tcp_sack_block *sp = &tp->selective_acks[0];
3265 struct tcp_sack_block *swalk = sp+1;
3267 /* See if the recent change to the first SACK eats into
3268 * or hits the sequence space of other SACK blocks, if so coalesce.
3270 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3271 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3274 /* Zap SWALK, by moving every further SACK up by one slot.
3275 * Decrease num_sacks.
3277 tp->rx_opt.num_sacks--;
3278 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3279 for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3283 this_sack++, swalk++;
3287 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3291 tmp = sack1->start_seq;
3292 sack1->start_seq = sack2->start_seq;
3293 sack2->start_seq = tmp;
3295 tmp = sack1->end_seq;
3296 sack1->end_seq = sack2->end_seq;
3297 sack2->end_seq = tmp;
3300 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3302 struct tcp_sock *tp = tcp_sk(sk);
3303 struct tcp_sack_block *sp = &tp->selective_acks[0];
3304 int cur_sacks = tp->rx_opt.num_sacks;
3310 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3311 if (tcp_sack_extend(sp, seq, end_seq)) {
3312 /* Rotate this_sack to the first one. */
3313 for (; this_sack>0; this_sack--, sp--)
3314 tcp_sack_swap(sp, sp-1);
3316 tcp_sack_maybe_coalesce(tp);
3321 /* Could not find an adjacent existing SACK, build a new one,
3322 * put it at the front, and shift everyone else down. We
3323 * always know there is at least one SACK present already here.
3325 * If the sack array is full, forget about the last one.
3327 if (this_sack >= 4) {
3329 tp->rx_opt.num_sacks--;
3332 for (; this_sack > 0; this_sack--, sp--)
3336 /* Build the new head SACK, and we're done. */
3337 sp->start_seq = seq;
3338 sp->end_seq = end_seq;
3339 tp->rx_opt.num_sacks++;
3340 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3343 /* RCV.NXT advances, some SACKs should be eaten. */
3345 static void tcp_sack_remove(struct tcp_sock *tp)
3347 struct tcp_sack_block *sp = &tp->selective_acks[0];
3348 int num_sacks = tp->rx_opt.num_sacks;
3351 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3352 if (skb_queue_empty(&tp->out_of_order_queue)) {
3353 tp->rx_opt.num_sacks = 0;
3354 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3358 for (this_sack = 0; this_sack < num_sacks; ) {
3359 /* Check if the start of the sack is covered by RCV.NXT. */
3360 if (!before(tp->rcv_nxt, sp->start_seq)) {
3363 /* RCV.NXT must cover all the block! */
3364 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3366 /* Zap this SACK, by moving forward any other SACKS. */
3367 for (i=this_sack+1; i < num_sacks; i++)
3368 tp->selective_acks[i-1] = tp->selective_acks[i];
3375 if (num_sacks != tp->rx_opt.num_sacks) {
3376 tp->rx_opt.num_sacks = num_sacks;
3377 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3381 /* This one checks to see if we can put data from the
3382 * out_of_order queue into the receive_queue.
3384 static void tcp_ofo_queue(struct sock *sk)
3386 struct tcp_sock *tp = tcp_sk(sk);
3387 __u32 dsack_high = tp->rcv_nxt;
3388 struct sk_buff *skb;
3390 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3391 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3394 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3395 __u32 dsack = dsack_high;
3396 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3397 dsack_high = TCP_SKB_CB(skb)->end_seq;
3398 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3401 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3402 SOCK_DEBUG(sk, "ofo packet was already received \n");
3403 __skb_unlink(skb, &tp->out_of_order_queue);
3407 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3408 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3409 TCP_SKB_CB(skb)->end_seq);
3411 __skb_unlink(skb, &tp->out_of_order_queue);
3412 __skb_queue_tail(&sk->sk_receive_queue, skb);
3413 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3414 if (tcp_hdr(skb)->fin)
3415 tcp_fin(skb, sk, tcp_hdr(skb));
3419 static int tcp_prune_queue(struct sock *sk);
3421 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3423 struct tcphdr *th = tcp_hdr(skb);
3424 struct tcp_sock *tp = tcp_sk(sk);
3427 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3430 __skb_pull(skb, th->doff*4);
3432 TCP_ECN_accept_cwr(tp, skb);
3434 if (tp->rx_opt.dsack) {
3435 tp->rx_opt.dsack = 0;
3436 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3437 4 - tp->rx_opt.tstamp_ok);
3440 /* Queue data for delivery to the user.
3441 * Packets in sequence go to the receive queue.
3442 * Out of sequence packets to the out_of_order_queue.
3444 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3445 if (tcp_receive_window(tp) == 0)
3448 /* Ok. In sequence. In window. */
3449 if (tp->ucopy.task == current &&
3450 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3451 sock_owned_by_user(sk) && !tp->urg_data) {
3452 int chunk = min_t(unsigned int, skb->len,
3455 __set_current_state(TASK_RUNNING);
3458 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3459 tp->ucopy.len -= chunk;
3460 tp->copied_seq += chunk;
3461 eaten = (chunk == skb->len && !th->fin);
3462 tcp_rcv_space_adjust(sk);
3470 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3471 !sk_stream_rmem_schedule(sk, skb))) {
3472 if (tcp_prune_queue(sk) < 0 ||
3473 !sk_stream_rmem_schedule(sk, skb))
3476 sk_stream_set_owner_r(skb, sk);
3477 __skb_queue_tail(&sk->sk_receive_queue, skb);
3479 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3481 tcp_event_data_recv(sk, skb);
3483 tcp_fin(skb, sk, th);
3485 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3488 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3489 * gap in queue is filled.
3491 if (skb_queue_empty(&tp->out_of_order_queue))
3492 inet_csk(sk)->icsk_ack.pingpong = 0;
3495 if (tp->rx_opt.num_sacks)
3496 tcp_sack_remove(tp);
3498 tcp_fast_path_check(sk);
3502 else if (!sock_flag(sk, SOCK_DEAD))
3503 sk->sk_data_ready(sk, 0);
3507 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3508 /* A retransmit, 2nd most common case. Force an immediate ack. */
3509 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3510 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3513 tcp_enter_quickack_mode(sk);
3514 inet_csk_schedule_ack(sk);
3520 /* Out of window. F.e. zero window probe. */
3521 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3524 tcp_enter_quickack_mode(sk);
3526 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3527 /* Partial packet, seq < rcv_next < end_seq */
3528 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3529 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3530 TCP_SKB_CB(skb)->end_seq);
3532 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3534 /* If window is closed, drop tail of packet. But after
3535 * remembering D-SACK for its head made in previous line.
3537 if (!tcp_receive_window(tp))
3542 TCP_ECN_check_ce(tp, skb);
3544 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3545 !sk_stream_rmem_schedule(sk, skb)) {
3546 if (tcp_prune_queue(sk) < 0 ||
3547 !sk_stream_rmem_schedule(sk, skb))
3551 /* Disable header prediction. */
3553 inet_csk_schedule_ack(sk);
3555 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3556 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3558 sk_stream_set_owner_r(skb, sk);
3560 if (!skb_peek(&tp->out_of_order_queue)) {
3561 /* Initial out of order segment, build 1 SACK. */
3562 if (tp->rx_opt.sack_ok) {
3563 tp->rx_opt.num_sacks = 1;
3564 tp->rx_opt.dsack = 0;
3565 tp->rx_opt.eff_sacks = 1;
3566 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3567 tp->selective_acks[0].end_seq =
3568 TCP_SKB_CB(skb)->end_seq;
3570 __skb_queue_head(&tp->out_of_order_queue,skb);
3572 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3573 u32 seq = TCP_SKB_CB(skb)->seq;
3574 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3576 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3577 __skb_append(skb1, skb, &tp->out_of_order_queue);
3579 if (!tp->rx_opt.num_sacks ||
3580 tp->selective_acks[0].end_seq != seq)
3583 /* Common case: data arrive in order after hole. */
3584 tp->selective_acks[0].end_seq = end_seq;
3588 /* Find place to insert this segment. */
3590 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3592 } while ((skb1 = skb1->prev) !=
3593 (struct sk_buff*)&tp->out_of_order_queue);
3595 /* Do skb overlap to previous one? */
3596 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3597 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3598 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3599 /* All the bits are present. Drop. */
3601 tcp_dsack_set(tp, seq, end_seq);
3604 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3605 /* Partial overlap. */
3606 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3611 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3613 /* And clean segments covered by new one as whole. */
3614 while ((skb1 = skb->next) !=
3615 (struct sk_buff*)&tp->out_of_order_queue &&
3616 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3617 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3618 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3621 __skb_unlink(skb1, &tp->out_of_order_queue);
3622 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3627 if (tp->rx_opt.sack_ok)
3628 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3632 /* Collapse contiguous sequence of skbs head..tail with
3633 * sequence numbers start..end.
3634 * Segments with FIN/SYN are not collapsed (only because this
3638 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3639 struct sk_buff *head, struct sk_buff *tail,
3642 struct sk_buff *skb;
3644 /* First, check that queue is collapsible and find
3645 * the point where collapsing can be useful. */
3646 for (skb = head; skb != tail; ) {
3647 /* No new bits? It is possible on ofo queue. */
3648 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3649 struct sk_buff *next = skb->next;
3650 __skb_unlink(skb, list);
3652 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3657 /* The first skb to collapse is:
3659 * - bloated or contains data before "start" or
3660 * overlaps to the next one.
3662 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3663 (tcp_win_from_space(skb->truesize) > skb->len ||
3664 before(TCP_SKB_CB(skb)->seq, start) ||
3665 (skb->next != tail &&
3666 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3669 /* Decided to skip this, advance start seq. */
3670 start = TCP_SKB_CB(skb)->end_seq;
3673 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3676 while (before(start, end)) {
3677 struct sk_buff *nskb;
3678 int header = skb_headroom(skb);
3679 int copy = SKB_MAX_ORDER(header, 0);
3681 /* Too big header? This can happen with IPv6. */
3684 if (end-start < copy)
3686 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3690 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3691 skb_set_network_header(nskb, (skb_network_header(skb) -
3693 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3695 skb_reserve(nskb, header);
3696 memcpy(nskb->head, skb->head, header);
3697 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3698 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3699 __skb_insert(nskb, skb->prev, skb, list);
3700 sk_stream_set_owner_r(nskb, sk);
3702 /* Copy data, releasing collapsed skbs. */
3704 int offset = start - TCP_SKB_CB(skb)->seq;
3705 int size = TCP_SKB_CB(skb)->end_seq - start;
3709 size = min(copy, size);
3710 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3712 TCP_SKB_CB(nskb)->end_seq += size;
3716 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3717 struct sk_buff *next = skb->next;
3718 __skb_unlink(skb, list);
3720 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3723 tcp_hdr(skb)->syn ||
3731 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3732 * and tcp_collapse() them until all the queue is collapsed.
3734 static void tcp_collapse_ofo_queue(struct sock *sk)
3736 struct tcp_sock *tp = tcp_sk(sk);
3737 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3738 struct sk_buff *head;
3744 start = TCP_SKB_CB(skb)->seq;
3745 end = TCP_SKB_CB(skb)->end_seq;
3751 /* Segment is terminated when we see gap or when
3752 * we are at the end of all the queue. */
3753 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3754 after(TCP_SKB_CB(skb)->seq, end) ||
3755 before(TCP_SKB_CB(skb)->end_seq, start)) {
3756 tcp_collapse(sk, &tp->out_of_order_queue,
3757 head, skb, start, end);
3759 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3761 /* Start new segment */
3762 start = TCP_SKB_CB(skb)->seq;
3763 end = TCP_SKB_CB(skb)->end_seq;
3765 if (before(TCP_SKB_CB(skb)->seq, start))
3766 start = TCP_SKB_CB(skb)->seq;
3767 if (after(TCP_SKB_CB(skb)->end_seq, end))
3768 end = TCP_SKB_CB(skb)->end_seq;
3773 /* Reduce allocated memory if we can, trying to get
3774 * the socket within its memory limits again.
3776 * Return less than zero if we should start dropping frames
3777 * until the socket owning process reads some of the data
3778 * to stabilize the situation.
3780 static int tcp_prune_queue(struct sock *sk)
3782 struct tcp_sock *tp = tcp_sk(sk);
3784 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3786 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3788 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3789 tcp_clamp_window(sk);
3790 else if (tcp_memory_pressure)
3791 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3793 tcp_collapse_ofo_queue(sk);
3794 tcp_collapse(sk, &sk->sk_receive_queue,
3795 sk->sk_receive_queue.next,
3796 (struct sk_buff*)&sk->sk_receive_queue,
3797 tp->copied_seq, tp->rcv_nxt);
3798 sk_stream_mem_reclaim(sk);
3800 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3803 /* Collapsing did not help, destructive actions follow.
3804 * This must not ever occur. */
3806 /* First, purge the out_of_order queue. */
3807 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3808 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3809 __skb_queue_purge(&tp->out_of_order_queue);
3811 /* Reset SACK state. A conforming SACK implementation will
3812 * do the same at a timeout based retransmit. When a connection
3813 * is in a sad state like this, we care only about integrity
3814 * of the connection not performance.
3816 if (tp->rx_opt.sack_ok)
3817 tcp_sack_reset(&tp->rx_opt);
3818 sk_stream_mem_reclaim(sk);
3821 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3824 /* If we are really being abused, tell the caller to silently
3825 * drop receive data on the floor. It will get retransmitted
3826 * and hopefully then we'll have sufficient space.
3828 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3830 /* Massive buffer overcommit. */
3836 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3837 * As additional protections, we do not touch cwnd in retransmission phases,
3838 * and if application hit its sndbuf limit recently.
3840 void tcp_cwnd_application_limited(struct sock *sk)
3842 struct tcp_sock *tp = tcp_sk(sk);
3844 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3845 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3846 /* Limited by application or receiver window. */
3847 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3848 u32 win_used = max(tp->snd_cwnd_used, init_win);
3849 if (win_used < tp->snd_cwnd) {
3850 tp->snd_ssthresh = tcp_current_ssthresh(sk);
3851 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3853 tp->snd_cwnd_used = 0;
3855 tp->snd_cwnd_stamp = tcp_time_stamp;
3858 static int tcp_should_expand_sndbuf(struct sock *sk)
3860 struct tcp_sock *tp = tcp_sk(sk);
3862 /* If the user specified a specific send buffer setting, do
3865 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3868 /* If we are under global TCP memory pressure, do not expand. */
3869 if (tcp_memory_pressure)
3872 /* If we are under soft global TCP memory pressure, do not expand. */
3873 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3876 /* If we filled the congestion window, do not expand. */
3877 if (tp->packets_out >= tp->snd_cwnd)
3883 /* When incoming ACK allowed to free some skb from write_queue,
3884 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3885 * on the exit from tcp input handler.
3887 * PROBLEM: sndbuf expansion does not work well with largesend.
3889 static void tcp_new_space(struct sock *sk)
3891 struct tcp_sock *tp = tcp_sk(sk);
3893 if (tcp_should_expand_sndbuf(sk)) {
3894 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3895 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3896 demanded = max_t(unsigned int, tp->snd_cwnd,
3897 tp->reordering + 1);
3898 sndmem *= 2*demanded;
3899 if (sndmem > sk->sk_sndbuf)
3900 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3901 tp->snd_cwnd_stamp = tcp_time_stamp;
3904 sk->sk_write_space(sk);
3907 static void tcp_check_space(struct sock *sk)
3909 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3910 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3911 if (sk->sk_socket &&
3912 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3917 static inline void tcp_data_snd_check(struct sock *sk)
3919 tcp_push_pending_frames(sk);
3920 tcp_check_space(sk);
3924 * Check if sending an ack is needed.
3926 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3928 struct tcp_sock *tp = tcp_sk(sk);
3930 /* More than one full frame received... */
3931 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3932 /* ... and right edge of window advances far enough.
3933 * (tcp_recvmsg() will send ACK otherwise). Or...
3935 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3936 /* We ACK each frame or... */
3937 tcp_in_quickack_mode(sk) ||
3938 /* We have out of order data. */
3940 skb_peek(&tp->out_of_order_queue))) {
3941 /* Then ack it now */
3944 /* Else, send delayed ack. */
3945 tcp_send_delayed_ack(sk);
3949 static inline void tcp_ack_snd_check(struct sock *sk)
3951 if (!inet_csk_ack_scheduled(sk)) {
3952 /* We sent a data segment already. */
3955 __tcp_ack_snd_check(sk, 1);
3959 * This routine is only called when we have urgent data
3960 * signaled. Its the 'slow' part of tcp_urg. It could be
3961 * moved inline now as tcp_urg is only called from one
3962 * place. We handle URGent data wrong. We have to - as
3963 * BSD still doesn't use the correction from RFC961.
3964 * For 1003.1g we should support a new option TCP_STDURG to permit
3965 * either form (or just set the sysctl tcp_stdurg).
3968 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3970 struct tcp_sock *tp = tcp_sk(sk);
3971 u32 ptr = ntohs(th->urg_ptr);
3973 if (ptr && !sysctl_tcp_stdurg)
3975 ptr += ntohl(th->seq);
3977 /* Ignore urgent data that we've already seen and read. */
3978 if (after(tp->copied_seq, ptr))
3981 /* Do not replay urg ptr.
3983 * NOTE: interesting situation not covered by specs.
3984 * Misbehaving sender may send urg ptr, pointing to segment,
3985 * which we already have in ofo queue. We are not able to fetch
3986 * such data and will stay in TCP_URG_NOTYET until will be eaten
3987 * by recvmsg(). Seems, we are not obliged to handle such wicked
3988 * situations. But it is worth to think about possibility of some
3989 * DoSes using some hypothetical application level deadlock.
3991 if (before(ptr, tp->rcv_nxt))
3994 /* Do we already have a newer (or duplicate) urgent pointer? */
3995 if (tp->urg_data && !after(ptr, tp->urg_seq))
3998 /* Tell the world about our new urgent pointer. */
4001 /* We may be adding urgent data when the last byte read was
4002 * urgent. To do this requires some care. We cannot just ignore
4003 * tp->copied_seq since we would read the last urgent byte again
4004 * as data, nor can we alter copied_seq until this data arrives
4005 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4007 * NOTE. Double Dutch. Rendering to plain English: author of comment
4008 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4009 * and expect that both A and B disappear from stream. This is _wrong_.
4010 * Though this happens in BSD with high probability, this is occasional.
4011 * Any application relying on this is buggy. Note also, that fix "works"
4012 * only in this artificial test. Insert some normal data between A and B and we will
4013 * decline of BSD again. Verdict: it is better to remove to trap
4016 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4017 !sock_flag(sk, SOCK_URGINLINE) &&
4018 tp->copied_seq != tp->rcv_nxt) {
4019 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4021 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4022 __skb_unlink(skb, &sk->sk_receive_queue);
4027 tp->urg_data = TCP_URG_NOTYET;
4030 /* Disable header prediction. */
4034 /* This is the 'fast' part of urgent handling. */
4035 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4037 struct tcp_sock *tp = tcp_sk(sk);
4039 /* Check if we get a new urgent pointer - normally not. */
4041 tcp_check_urg(sk,th);
4043 /* Do we wait for any urgent data? - normally not... */
4044 if (tp->urg_data == TCP_URG_NOTYET) {
4045 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4048 /* Is the urgent pointer pointing into this packet? */
4049 if (ptr < skb->len) {
4051 if (skb_copy_bits(skb, ptr, &tmp, 1))
4053 tp->urg_data = TCP_URG_VALID | tmp;
4054 if (!sock_flag(sk, SOCK_DEAD))
4055 sk->sk_data_ready(sk, 0);
4060 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4062 struct tcp_sock *tp = tcp_sk(sk);
4063 int chunk = skb->len - hlen;
4067 if (skb_csum_unnecessary(skb))
4068 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4070 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4074 tp->ucopy.len -= chunk;
4075 tp->copied_seq += chunk;
4076 tcp_rcv_space_adjust(sk);
4083 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4087 if (sock_owned_by_user(sk)) {
4089 result = __tcp_checksum_complete(skb);
4092 result = __tcp_checksum_complete(skb);
4097 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4099 return !skb_csum_unnecessary(skb) &&
4100 __tcp_checksum_complete_user(sk, skb);
4103 #ifdef CONFIG_NET_DMA
4104 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4106 struct tcp_sock *tp = tcp_sk(sk);
4107 int chunk = skb->len - hlen;
4109 int copied_early = 0;
4111 if (tp->ucopy.wakeup)
4114 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4115 tp->ucopy.dma_chan = get_softnet_dma();
4117 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4119 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4120 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4125 tp->ucopy.dma_cookie = dma_cookie;
4128 tp->ucopy.len -= chunk;
4129 tp->copied_seq += chunk;
4130 tcp_rcv_space_adjust(sk);
4132 if ((tp->ucopy.len == 0) ||
4133 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4134 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4135 tp->ucopy.wakeup = 1;
4136 sk->sk_data_ready(sk, 0);
4138 } else if (chunk > 0) {
4139 tp->ucopy.wakeup = 1;
4140 sk->sk_data_ready(sk, 0);
4143 return copied_early;
4145 #endif /* CONFIG_NET_DMA */
4148 * TCP receive function for the ESTABLISHED state.
4150 * It is split into a fast path and a slow path. The fast path is
4152 * - A zero window was announced from us - zero window probing
4153 * is only handled properly in the slow path.
4154 * - Out of order segments arrived.
4155 * - Urgent data is expected.
4156 * - There is no buffer space left
4157 * - Unexpected TCP flags/window values/header lengths are received
4158 * (detected by checking the TCP header against pred_flags)
4159 * - Data is sent in both directions. Fast path only supports pure senders
4160 * or pure receivers (this means either the sequence number or the ack
4161 * value must stay constant)
4162 * - Unexpected TCP option.
4164 * When these conditions are not satisfied it drops into a standard
4165 * receive procedure patterned after RFC793 to handle all cases.
4166 * The first three cases are guaranteed by proper pred_flags setting,
4167 * the rest is checked inline. Fast processing is turned on in
4168 * tcp_data_queue when everything is OK.
4170 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4171 struct tcphdr *th, unsigned len)
4173 struct tcp_sock *tp = tcp_sk(sk);
4176 * Header prediction.
4177 * The code loosely follows the one in the famous
4178 * "30 instruction TCP receive" Van Jacobson mail.
4180 * Van's trick is to deposit buffers into socket queue
4181 * on a device interrupt, to call tcp_recv function
4182 * on the receive process context and checksum and copy
4183 * the buffer to user space. smart...
4185 * Our current scheme is not silly either but we take the
4186 * extra cost of the net_bh soft interrupt processing...
4187 * We do checksum and copy also but from device to kernel.
4190 tp->rx_opt.saw_tstamp = 0;
4192 /* pred_flags is 0xS?10 << 16 + snd_wnd
4193 * if header_prediction is to be made
4194 * 'S' will always be tp->tcp_header_len >> 2
4195 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4196 * turn it off (when there are holes in the receive
4197 * space for instance)
4198 * PSH flag is ignored.
4201 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4202 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4203 int tcp_header_len = tp->tcp_header_len;
4205 /* Timestamp header prediction: tcp_header_len
4206 * is automatically equal to th->doff*4 due to pred_flags
4210 /* Check timestamp */
4211 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4212 __be32 *ptr = (__be32 *)(th + 1);
4214 /* No? Slow path! */
4215 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4216 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4219 tp->rx_opt.saw_tstamp = 1;
4221 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4223 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4225 /* If PAWS failed, check it more carefully in slow path */
4226 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4229 /* DO NOT update ts_recent here, if checksum fails
4230 * and timestamp was corrupted part, it will result
4231 * in a hung connection since we will drop all
4232 * future packets due to the PAWS test.
4236 if (len <= tcp_header_len) {
4237 /* Bulk data transfer: sender */
4238 if (len == tcp_header_len) {
4239 /* Predicted packet is in window by definition.
4240 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4241 * Hence, check seq<=rcv_wup reduces to:
4243 if (tcp_header_len ==
4244 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4245 tp->rcv_nxt == tp->rcv_wup)
4246 tcp_store_ts_recent(tp);
4248 /* We know that such packets are checksummed
4251 tcp_ack(sk, skb, 0);
4253 tcp_data_snd_check(sk);
4255 } else { /* Header too small */
4256 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4261 int copied_early = 0;
4263 if (tp->copied_seq == tp->rcv_nxt &&
4264 len - tcp_header_len <= tp->ucopy.len) {
4265 #ifdef CONFIG_NET_DMA
4266 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4271 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4272 __set_current_state(TASK_RUNNING);
4274 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4278 /* Predicted packet is in window by definition.
4279 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4280 * Hence, check seq<=rcv_wup reduces to:
4282 if (tcp_header_len ==
4283 (sizeof(struct tcphdr) +
4284 TCPOLEN_TSTAMP_ALIGNED) &&
4285 tp->rcv_nxt == tp->rcv_wup)
4286 tcp_store_ts_recent(tp);
4288 tcp_rcv_rtt_measure_ts(sk, skb);
4290 __skb_pull(skb, tcp_header_len);
4291 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4292 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4295 tcp_cleanup_rbuf(sk, skb->len);
4298 if (tcp_checksum_complete_user(sk, skb))
4301 /* Predicted packet is in window by definition.
4302 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4303 * Hence, check seq<=rcv_wup reduces to:
4305 if (tcp_header_len ==
4306 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4307 tp->rcv_nxt == tp->rcv_wup)
4308 tcp_store_ts_recent(tp);
4310 tcp_rcv_rtt_measure_ts(sk, skb);
4312 if ((int)skb->truesize > sk->sk_forward_alloc)
4315 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4317 /* Bulk data transfer: receiver */
4318 __skb_pull(skb,tcp_header_len);
4319 __skb_queue_tail(&sk->sk_receive_queue, skb);
4320 sk_stream_set_owner_r(skb, sk);
4321 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4324 tcp_event_data_recv(sk, skb);
4326 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4327 /* Well, only one small jumplet in fast path... */
4328 tcp_ack(sk, skb, FLAG_DATA);
4329 tcp_data_snd_check(sk);
4330 if (!inet_csk_ack_scheduled(sk))
4334 __tcp_ack_snd_check(sk, 0);
4336 #ifdef CONFIG_NET_DMA
4338 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4344 sk->sk_data_ready(sk, 0);
4350 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4354 * RFC1323: H1. Apply PAWS check first.
4356 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4357 tcp_paws_discard(sk, skb)) {
4359 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4360 tcp_send_dupack(sk, skb);
4363 /* Resets are accepted even if PAWS failed.
4365 ts_recent update must be made after we are sure
4366 that the packet is in window.
4371 * Standard slow path.
4374 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4375 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4376 * (RST) segments are validated by checking their SEQ-fields."
4377 * And page 69: "If an incoming segment is not acceptable,
4378 * an acknowledgment should be sent in reply (unless the RST bit
4379 * is set, if so drop the segment and return)".
4382 tcp_send_dupack(sk, skb);
4391 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4393 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4394 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4395 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4402 tcp_ack(sk, skb, FLAG_SLOWPATH);
4404 tcp_rcv_rtt_measure_ts(sk, skb);
4406 /* Process urgent data. */
4407 tcp_urg(sk, skb, th);
4409 /* step 7: process the segment text */
4410 tcp_data_queue(sk, skb);
4412 tcp_data_snd_check(sk);
4413 tcp_ack_snd_check(sk);
4417 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4424 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4425 struct tcphdr *th, unsigned len)
4427 struct tcp_sock *tp = tcp_sk(sk);
4428 struct inet_connection_sock *icsk = inet_csk(sk);
4429 int saved_clamp = tp->rx_opt.mss_clamp;
4431 tcp_parse_options(skb, &tp->rx_opt, 0);
4435 * "If the state is SYN-SENT then
4436 * first check the ACK bit
4437 * If the ACK bit is set
4438 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4439 * a reset (unless the RST bit is set, if so drop
4440 * the segment and return)"
4442 * We do not send data with SYN, so that RFC-correct
4445 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4446 goto reset_and_undo;
4448 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4449 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4451 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4452 goto reset_and_undo;
4455 /* Now ACK is acceptable.
4457 * "If the RST bit is set
4458 * If the ACK was acceptable then signal the user "error:
4459 * connection reset", drop the segment, enter CLOSED state,
4460 * delete TCB, and return."
4469 * "fifth, if neither of the SYN or RST bits is set then
4470 * drop the segment and return."
4476 goto discard_and_undo;
4479 * "If the SYN bit is on ...
4480 * are acceptable then ...
4481 * (our SYN has been ACKed), change the connection
4482 * state to ESTABLISHED..."
4485 TCP_ECN_rcv_synack(tp, th);
4487 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4488 tcp_ack(sk, skb, FLAG_SLOWPATH);
4490 /* Ok.. it's good. Set up sequence numbers and
4491 * move to established.
4493 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4494 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4496 /* RFC1323: The window in SYN & SYN/ACK segments is
4499 tp->snd_wnd = ntohs(th->window);
4500 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4502 if (!tp->rx_opt.wscale_ok) {
4503 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4504 tp->window_clamp = min(tp->window_clamp, 65535U);
4507 if (tp->rx_opt.saw_tstamp) {
4508 tp->rx_opt.tstamp_ok = 1;
4509 tp->tcp_header_len =
4510 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4511 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4512 tcp_store_ts_recent(tp);
4514 tp->tcp_header_len = sizeof(struct tcphdr);
4517 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4518 tp->rx_opt.sack_ok |= 2;
4521 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4522 tcp_initialize_rcv_mss(sk);
4524 /* Remember, tcp_poll() does not lock socket!
4525 * Change state from SYN-SENT only after copied_seq
4526 * is initialized. */
4527 tp->copied_seq = tp->rcv_nxt;
4529 tcp_set_state(sk, TCP_ESTABLISHED);
4531 security_inet_conn_established(sk, skb);
4533 /* Make sure socket is routed, for correct metrics. */
4534 icsk->icsk_af_ops->rebuild_header(sk);
4536 tcp_init_metrics(sk);
4538 tcp_init_congestion_control(sk);
4540 /* Prevent spurious tcp_cwnd_restart() on first data
4543 tp->lsndtime = tcp_time_stamp;
4545 tcp_init_buffer_space(sk);
4547 if (sock_flag(sk, SOCK_KEEPOPEN))
4548 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4550 if (!tp->rx_opt.snd_wscale)
4551 __tcp_fast_path_on(tp, tp->snd_wnd);
4555 if (!sock_flag(sk, SOCK_DEAD)) {
4556 sk->sk_state_change(sk);
4557 sk_wake_async(sk, 0, POLL_OUT);
4560 if (sk->sk_write_pending ||
4561 icsk->icsk_accept_queue.rskq_defer_accept ||
4562 icsk->icsk_ack.pingpong) {
4563 /* Save one ACK. Data will be ready after
4564 * several ticks, if write_pending is set.
4566 * It may be deleted, but with this feature tcpdumps
4567 * look so _wonderfully_ clever, that I was not able
4568 * to stand against the temptation 8) --ANK
4570 inet_csk_schedule_ack(sk);
4571 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4572 icsk->icsk_ack.ato = TCP_ATO_MIN;
4573 tcp_incr_quickack(sk);
4574 tcp_enter_quickack_mode(sk);
4575 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4576 TCP_DELACK_MAX, TCP_RTO_MAX);
4587 /* No ACK in the segment */
4591 * "If the RST bit is set
4593 * Otherwise (no ACK) drop the segment and return."
4596 goto discard_and_undo;
4600 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4601 goto discard_and_undo;
4604 /* We see SYN without ACK. It is attempt of
4605 * simultaneous connect with crossed SYNs.
4606 * Particularly, it can be connect to self.
4608 tcp_set_state(sk, TCP_SYN_RECV);
4610 if (tp->rx_opt.saw_tstamp) {
4611 tp->rx_opt.tstamp_ok = 1;
4612 tcp_store_ts_recent(tp);
4613 tp->tcp_header_len =
4614 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4616 tp->tcp_header_len = sizeof(struct tcphdr);
4619 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4620 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4622 /* RFC1323: The window in SYN & SYN/ACK segments is
4625 tp->snd_wnd = ntohs(th->window);
4626 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4627 tp->max_window = tp->snd_wnd;
4629 TCP_ECN_rcv_syn(tp, th);
4632 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4633 tcp_initialize_rcv_mss(sk);
4636 tcp_send_synack(sk);
4638 /* Note, we could accept data and URG from this segment.
4639 * There are no obstacles to make this.
4641 * However, if we ignore data in ACKless segments sometimes,
4642 * we have no reasons to accept it sometimes.
4643 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4644 * is not flawless. So, discard packet for sanity.
4645 * Uncomment this return to process the data.
4652 /* "fifth, if neither of the SYN or RST bits is set then
4653 * drop the segment and return."
4657 tcp_clear_options(&tp->rx_opt);
4658 tp->rx_opt.mss_clamp = saved_clamp;
4662 tcp_clear_options(&tp->rx_opt);
4663 tp->rx_opt.mss_clamp = saved_clamp;
4669 * This function implements the receiving procedure of RFC 793 for
4670 * all states except ESTABLISHED and TIME_WAIT.
4671 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4672 * address independent.
4675 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4676 struct tcphdr *th, unsigned len)
4678 struct tcp_sock *tp = tcp_sk(sk);
4679 struct inet_connection_sock *icsk = inet_csk(sk);
4682 tp->rx_opt.saw_tstamp = 0;
4684 switch (sk->sk_state) {
4696 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4699 /* Now we have several options: In theory there is
4700 * nothing else in the frame. KA9Q has an option to
4701 * send data with the syn, BSD accepts data with the
4702 * syn up to the [to be] advertised window and
4703 * Solaris 2.1 gives you a protocol error. For now
4704 * we just ignore it, that fits the spec precisely
4705 * and avoids incompatibilities. It would be nice in
4706 * future to drop through and process the data.
4708 * Now that TTCP is starting to be used we ought to
4710 * But, this leaves one open to an easy denial of
4711 * service attack, and SYN cookies can't defend
4712 * against this problem. So, we drop the data
4713 * in the interest of security over speed unless
4714 * it's still in use.
4722 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4726 /* Do step6 onward by hand. */
4727 tcp_urg(sk, skb, th);
4729 tcp_data_snd_check(sk);
4733 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4734 tcp_paws_discard(sk, skb)) {
4736 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4737 tcp_send_dupack(sk, skb);
4740 /* Reset is accepted even if it did not pass PAWS. */
4743 /* step 1: check sequence number */
4744 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4746 tcp_send_dupack(sk, skb);
4750 /* step 2: check RST bit */
4756 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4758 /* step 3: check security and precedence [ignored] */
4762 * Check for a SYN in window.
4764 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4765 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4770 /* step 5: check the ACK field */
4772 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4774 switch (sk->sk_state) {
4777 tp->copied_seq = tp->rcv_nxt;
4779 tcp_set_state(sk, TCP_ESTABLISHED);
4780 sk->sk_state_change(sk);
4782 /* Note, that this wakeup is only for marginal
4783 * crossed SYN case. Passively open sockets
4784 * are not waked up, because sk->sk_sleep ==
4785 * NULL and sk->sk_socket == NULL.
4787 if (sk->sk_socket) {
4788 sk_wake_async(sk,0,POLL_OUT);
4791 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4792 tp->snd_wnd = ntohs(th->window) <<
4793 tp->rx_opt.snd_wscale;
4794 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4795 TCP_SKB_CB(skb)->seq);
4797 /* tcp_ack considers this ACK as duplicate
4798 * and does not calculate rtt.
4799 * Fix it at least with timestamps.
4801 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4803 tcp_ack_saw_tstamp(sk, 0);
4805 if (tp->rx_opt.tstamp_ok)
4806 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4808 /* Make sure socket is routed, for
4811 icsk->icsk_af_ops->rebuild_header(sk);
4813 tcp_init_metrics(sk);
4815 tcp_init_congestion_control(sk);
4817 /* Prevent spurious tcp_cwnd_restart() on
4818 * first data packet.
4820 tp->lsndtime = tcp_time_stamp;
4823 tcp_initialize_rcv_mss(sk);
4824 tcp_init_buffer_space(sk);
4825 tcp_fast_path_on(tp);
4832 if (tp->snd_una == tp->write_seq) {
4833 tcp_set_state(sk, TCP_FIN_WAIT2);
4834 sk->sk_shutdown |= SEND_SHUTDOWN;
4835 dst_confirm(sk->sk_dst_cache);
4837 if (!sock_flag(sk, SOCK_DEAD))
4838 /* Wake up lingering close() */
4839 sk->sk_state_change(sk);
4843 if (tp->linger2 < 0 ||
4844 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4845 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4847 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4851 tmo = tcp_fin_time(sk);
4852 if (tmo > TCP_TIMEWAIT_LEN) {
4853 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4854 } else if (th->fin || sock_owned_by_user(sk)) {
4855 /* Bad case. We could lose such FIN otherwise.
4856 * It is not a big problem, but it looks confusing
4857 * and not so rare event. We still can lose it now,
4858 * if it spins in bh_lock_sock(), but it is really
4861 inet_csk_reset_keepalive_timer(sk, tmo);
4863 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4871 if (tp->snd_una == tp->write_seq) {
4872 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4878 if (tp->snd_una == tp->write_seq) {
4879 tcp_update_metrics(sk);
4888 /* step 6: check the URG bit */
4889 tcp_urg(sk, skb, th);
4891 /* step 7: process the segment text */
4892 switch (sk->sk_state) {
4893 case TCP_CLOSE_WAIT:
4896 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4900 /* RFC 793 says to queue data in these states,
4901 * RFC 1122 says we MUST send a reset.
4902 * BSD 4.4 also does reset.
4904 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4905 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4906 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4907 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4913 case TCP_ESTABLISHED:
4914 tcp_data_queue(sk, skb);
4919 /* tcp_data could move socket to TIME-WAIT */
4920 if (sk->sk_state != TCP_CLOSE) {
4921 tcp_data_snd_check(sk);
4922 tcp_ack_snd_check(sk);
4932 EXPORT_SYMBOL(sysctl_tcp_ecn);
4933 EXPORT_SYMBOL(sysctl_tcp_reordering);
4934 EXPORT_SYMBOL(tcp_parse_options);
4935 EXPORT_SYMBOL(tcp_rcv_established);
4936 EXPORT_SYMBOL(tcp_rcv_state_process);
4937 EXPORT_SYMBOL(tcp_initialize_rcv_mss);