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 */
106 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
107 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
108 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
109 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
111 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
112 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
113 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
119 /* Adapt the MSS value used to make delayed ack decision to the
122 static void tcp_measure_rcv_mss(struct sock *sk,
123 const struct sk_buff *skb)
125 struct inet_connection_sock *icsk = inet_csk(sk);
126 const unsigned int lss = icsk->icsk_ack.last_seg_size;
129 icsk->icsk_ack.last_seg_size = 0;
131 /* skb->len may jitter because of SACKs, even if peer
132 * sends good full-sized frames.
134 len = skb_shinfo(skb)->gso_size ?: skb->len;
135 if (len >= icsk->icsk_ack.rcv_mss) {
136 icsk->icsk_ack.rcv_mss = len;
138 /* Otherwise, we make more careful check taking into account,
139 * that SACKs block is variable.
141 * "len" is invariant segment length, including TCP header.
143 len += skb->data - skb_transport_header(skb);
144 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
145 /* If PSH is not set, packet should be
146 * full sized, provided peer TCP is not badly broken.
147 * This observation (if it is correct 8)) allows
148 * to handle super-low mtu links fairly.
150 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
151 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
152 /* Subtract also invariant (if peer is RFC compliant),
153 * tcp header plus fixed timestamp option length.
154 * Resulting "len" is MSS free of SACK jitter.
156 len -= tcp_sk(sk)->tcp_header_len;
157 icsk->icsk_ack.last_seg_size = len;
159 icsk->icsk_ack.rcv_mss = len;
163 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
164 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
169 static void tcp_incr_quickack(struct sock *sk)
171 struct inet_connection_sock *icsk = inet_csk(sk);
172 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
176 if (quickacks > icsk->icsk_ack.quick)
177 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
180 void tcp_enter_quickack_mode(struct sock *sk)
182 struct inet_connection_sock *icsk = inet_csk(sk);
183 tcp_incr_quickack(sk);
184 icsk->icsk_ack.pingpong = 0;
185 icsk->icsk_ack.ato = TCP_ATO_MIN;
188 /* Send ACKs quickly, if "quick" count is not exhausted
189 * and the session is not interactive.
192 static inline int tcp_in_quickack_mode(const struct sock *sk)
194 const struct inet_connection_sock *icsk = inet_csk(sk);
195 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
198 /* Buffer size and advertised window tuning.
200 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
203 static void tcp_fixup_sndbuf(struct sock *sk)
205 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
206 sizeof(struct sk_buff);
208 if (sk->sk_sndbuf < 3 * sndmem)
209 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
212 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
214 * All tcp_full_space() is split to two parts: "network" buffer, allocated
215 * forward and advertised in receiver window (tp->rcv_wnd) and
216 * "application buffer", required to isolate scheduling/application
217 * latencies from network.
218 * window_clamp is maximal advertised window. It can be less than
219 * tcp_full_space(), in this case tcp_full_space() - window_clamp
220 * is reserved for "application" buffer. The less window_clamp is
221 * the smoother our behaviour from viewpoint of network, but the lower
222 * throughput and the higher sensitivity of the connection to losses. 8)
224 * rcv_ssthresh is more strict window_clamp used at "slow start"
225 * phase to predict further behaviour of this connection.
226 * It is used for two goals:
227 * - to enforce header prediction at sender, even when application
228 * requires some significant "application buffer". It is check #1.
229 * - to prevent pruning of receive queue because of misprediction
230 * of receiver window. Check #2.
232 * The scheme does not work when sender sends good segments opening
233 * window and then starts to feed us spaghetti. But it should work
234 * in common situations. Otherwise, we have to rely on queue collapsing.
237 /* Slow part of check#2. */
238 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
240 struct tcp_sock *tp = tcp_sk(sk);
242 int truesize = tcp_win_from_space(skb->truesize)/2;
243 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
245 while (tp->rcv_ssthresh <= window) {
246 if (truesize <= skb->len)
247 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
255 static void tcp_grow_window(struct sock *sk,
258 struct tcp_sock *tp = tcp_sk(sk);
261 if (tp->rcv_ssthresh < tp->window_clamp &&
262 (int)tp->rcv_ssthresh < tcp_space(sk) &&
263 !tcp_memory_pressure) {
266 /* Check #2. Increase window, if skb with such overhead
267 * will fit to rcvbuf in future.
269 if (tcp_win_from_space(skb->truesize) <= skb->len)
272 incr = __tcp_grow_window(sk, skb);
275 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
276 inet_csk(sk)->icsk_ack.quick |= 1;
281 /* 3. Tuning rcvbuf, when connection enters established state. */
283 static void tcp_fixup_rcvbuf(struct sock *sk)
285 struct tcp_sock *tp = tcp_sk(sk);
286 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
288 /* Try to select rcvbuf so that 4 mss-sized segments
289 * will fit to window and corresponding skbs will fit to our rcvbuf.
290 * (was 3; 4 is minimum to allow fast retransmit to work.)
292 while (tcp_win_from_space(rcvmem) < tp->advmss)
294 if (sk->sk_rcvbuf < 4 * rcvmem)
295 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
298 /* 4. Try to fixup all. It is made immediately after connection enters
301 static void tcp_init_buffer_space(struct sock *sk)
303 struct tcp_sock *tp = tcp_sk(sk);
306 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
307 tcp_fixup_rcvbuf(sk);
308 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
309 tcp_fixup_sndbuf(sk);
311 tp->rcvq_space.space = tp->rcv_wnd;
313 maxwin = tcp_full_space(sk);
315 if (tp->window_clamp >= maxwin) {
316 tp->window_clamp = maxwin;
318 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
319 tp->window_clamp = max(maxwin -
320 (maxwin >> sysctl_tcp_app_win),
324 /* Force reservation of one segment. */
325 if (sysctl_tcp_app_win &&
326 tp->window_clamp > 2 * tp->advmss &&
327 tp->window_clamp + tp->advmss > maxwin)
328 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
330 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
331 tp->snd_cwnd_stamp = tcp_time_stamp;
334 /* 5. Recalculate window clamp after socket hit its memory bounds. */
335 static void tcp_clamp_window(struct sock *sk)
337 struct tcp_sock *tp = tcp_sk(sk);
338 struct inet_connection_sock *icsk = inet_csk(sk);
340 icsk->icsk_ack.quick = 0;
342 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
343 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
344 !tcp_memory_pressure &&
345 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
346 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
349 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
350 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
354 /* Initialize RCV_MSS value.
355 * RCV_MSS is an our guess about MSS used by the peer.
356 * We haven't any direct information about the MSS.
357 * It's better to underestimate the RCV_MSS rather than overestimate.
358 * Overestimations make us ACKing less frequently than needed.
359 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
361 void tcp_initialize_rcv_mss(struct sock *sk)
363 struct tcp_sock *tp = tcp_sk(sk);
364 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
366 hint = min(hint, tp->rcv_wnd/2);
367 hint = min(hint, TCP_MIN_RCVMSS);
368 hint = max(hint, TCP_MIN_MSS);
370 inet_csk(sk)->icsk_ack.rcv_mss = hint;
373 /* Receiver "autotuning" code.
375 * The algorithm for RTT estimation w/o timestamps is based on
376 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
377 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
379 * More detail on this code can be found at
380 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
381 * though this reference is out of date. A new paper
384 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
386 u32 new_sample = tp->rcv_rtt_est.rtt;
392 if (new_sample != 0) {
393 /* If we sample in larger samples in the non-timestamp
394 * case, we could grossly overestimate the RTT especially
395 * with chatty applications or bulk transfer apps which
396 * are stalled on filesystem I/O.
398 * Also, since we are only going for a minimum in the
399 * non-timestamp case, we do not smooth things out
400 * else with timestamps disabled convergence takes too
404 m -= (new_sample >> 3);
406 } else if (m < new_sample)
409 /* No previous measure. */
413 if (tp->rcv_rtt_est.rtt != new_sample)
414 tp->rcv_rtt_est.rtt = new_sample;
417 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
419 if (tp->rcv_rtt_est.time == 0)
421 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
423 tcp_rcv_rtt_update(tp,
424 jiffies - tp->rcv_rtt_est.time,
428 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
429 tp->rcv_rtt_est.time = tcp_time_stamp;
432 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
434 struct tcp_sock *tp = tcp_sk(sk);
435 if (tp->rx_opt.rcv_tsecr &&
436 (TCP_SKB_CB(skb)->end_seq -
437 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
438 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
442 * This function should be called every time data is copied to user space.
443 * It calculates the appropriate TCP receive buffer space.
445 void tcp_rcv_space_adjust(struct sock *sk)
447 struct tcp_sock *tp = tcp_sk(sk);
451 if (tp->rcvq_space.time == 0)
454 time = tcp_time_stamp - tp->rcvq_space.time;
455 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
456 tp->rcv_rtt_est.rtt == 0)
459 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
461 space = max(tp->rcvq_space.space, space);
463 if (tp->rcvq_space.space != space) {
466 tp->rcvq_space.space = space;
468 if (sysctl_tcp_moderate_rcvbuf &&
469 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
470 int new_clamp = space;
472 /* Receive space grows, normalize in order to
473 * take into account packet headers and sk_buff
474 * structure overhead.
479 rcvmem = (tp->advmss + MAX_TCP_HEADER +
480 16 + sizeof(struct sk_buff));
481 while (tcp_win_from_space(rcvmem) < tp->advmss)
484 space = min(space, sysctl_tcp_rmem[2]);
485 if (space > sk->sk_rcvbuf) {
486 sk->sk_rcvbuf = space;
488 /* Make the window clamp follow along. */
489 tp->window_clamp = new_clamp;
495 tp->rcvq_space.seq = tp->copied_seq;
496 tp->rcvq_space.time = tcp_time_stamp;
499 /* There is something which you must keep in mind when you analyze the
500 * behavior of the tp->ato delayed ack timeout interval. When a
501 * connection starts up, we want to ack as quickly as possible. The
502 * problem is that "good" TCP's do slow start at the beginning of data
503 * transmission. The means that until we send the first few ACK's the
504 * sender will sit on his end and only queue most of his data, because
505 * he can only send snd_cwnd unacked packets at any given time. For
506 * each ACK we send, he increments snd_cwnd and transmits more of his
509 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
511 struct tcp_sock *tp = tcp_sk(sk);
512 struct inet_connection_sock *icsk = inet_csk(sk);
515 inet_csk_schedule_ack(sk);
517 tcp_measure_rcv_mss(sk, skb);
519 tcp_rcv_rtt_measure(tp);
521 now = tcp_time_stamp;
523 if (!icsk->icsk_ack.ato) {
524 /* The _first_ data packet received, initialize
525 * delayed ACK engine.
527 tcp_incr_quickack(sk);
528 icsk->icsk_ack.ato = TCP_ATO_MIN;
530 int m = now - icsk->icsk_ack.lrcvtime;
532 if (m <= TCP_ATO_MIN/2) {
533 /* The fastest case is the first. */
534 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
535 } else if (m < icsk->icsk_ack.ato) {
536 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
537 if (icsk->icsk_ack.ato > icsk->icsk_rto)
538 icsk->icsk_ack.ato = icsk->icsk_rto;
539 } else if (m > icsk->icsk_rto) {
540 /* Too long gap. Apparently sender failed to
541 * restart window, so that we send ACKs quickly.
543 tcp_incr_quickack(sk);
544 sk_stream_mem_reclaim(sk);
547 icsk->icsk_ack.lrcvtime = now;
549 TCP_ECN_check_ce(tp, skb);
552 tcp_grow_window(sk, skb);
555 /* Called to compute a smoothed rtt estimate. The data fed to this
556 * routine either comes from timestamps, or from segments that were
557 * known _not_ to have been retransmitted [see Karn/Partridge
558 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
559 * piece by Van Jacobson.
560 * NOTE: the next three routines used to be one big routine.
561 * To save cycles in the RFC 1323 implementation it was better to break
562 * it up into three procedures. -- erics
564 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
566 struct tcp_sock *tp = tcp_sk(sk);
567 long m = mrtt; /* RTT */
569 /* The following amusing code comes from Jacobson's
570 * article in SIGCOMM '88. Note that rtt and mdev
571 * are scaled versions of rtt and mean deviation.
572 * This is designed to be as fast as possible
573 * m stands for "measurement".
575 * On a 1990 paper the rto value is changed to:
576 * RTO = rtt + 4 * mdev
578 * Funny. This algorithm seems to be very broken.
579 * These formulae increase RTO, when it should be decreased, increase
580 * too slowly, when it should be increased quickly, decrease too quickly
581 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
582 * does not matter how to _calculate_ it. Seems, it was trap
583 * that VJ failed to avoid. 8)
588 m -= (tp->srtt >> 3); /* m is now error in rtt est */
589 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
591 m = -m; /* m is now abs(error) */
592 m -= (tp->mdev >> 2); /* similar update on mdev */
593 /* This is similar to one of Eifel findings.
594 * Eifel blocks mdev updates when rtt decreases.
595 * This solution is a bit different: we use finer gain
596 * for mdev in this case (alpha*beta).
597 * Like Eifel it also prevents growth of rto,
598 * but also it limits too fast rto decreases,
599 * happening in pure Eifel.
604 m -= (tp->mdev >> 2); /* similar update on mdev */
606 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
607 if (tp->mdev > tp->mdev_max) {
608 tp->mdev_max = tp->mdev;
609 if (tp->mdev_max > tp->rttvar)
610 tp->rttvar = tp->mdev_max;
612 if (after(tp->snd_una, tp->rtt_seq)) {
613 if (tp->mdev_max < tp->rttvar)
614 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
615 tp->rtt_seq = tp->snd_nxt;
616 tp->mdev_max = TCP_RTO_MIN;
619 /* no previous measure. */
620 tp->srtt = m<<3; /* take the measured time to be rtt */
621 tp->mdev = m<<1; /* make sure rto = 3*rtt */
622 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
623 tp->rtt_seq = tp->snd_nxt;
627 /* Calculate rto without backoff. This is the second half of Van Jacobson's
628 * routine referred to above.
630 static inline void tcp_set_rto(struct sock *sk)
632 const struct tcp_sock *tp = tcp_sk(sk);
633 /* Old crap is replaced with new one. 8)
636 * 1. If rtt variance happened to be less 50msec, it is hallucination.
637 * It cannot be less due to utterly erratic ACK generation made
638 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
639 * to do with delayed acks, because at cwnd>2 true delack timeout
640 * is invisible. Actually, Linux-2.4 also generates erratic
641 * ACKs in some circumstances.
643 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
645 /* 2. Fixups made earlier cannot be right.
646 * If we do not estimate RTO correctly without them,
647 * all the algo is pure shit and should be replaced
648 * with correct one. It is exactly, which we pretend to do.
652 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
653 * guarantees that rto is higher.
655 static inline void tcp_bound_rto(struct sock *sk)
657 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
658 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
661 /* Save metrics learned by this TCP session.
662 This function is called only, when TCP finishes successfully
663 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
665 void tcp_update_metrics(struct sock *sk)
667 struct tcp_sock *tp = tcp_sk(sk);
668 struct dst_entry *dst = __sk_dst_get(sk);
670 if (sysctl_tcp_nometrics_save)
675 if (dst && (dst->flags&DST_HOST)) {
676 const struct inet_connection_sock *icsk = inet_csk(sk);
679 if (icsk->icsk_backoff || !tp->srtt) {
680 /* This session failed to estimate rtt. Why?
681 * Probably, no packets returned in time.
684 if (!(dst_metric_locked(dst, RTAX_RTT)))
685 dst->metrics[RTAX_RTT-1] = 0;
689 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
691 /* If newly calculated rtt larger than stored one,
692 * store new one. Otherwise, use EWMA. Remember,
693 * rtt overestimation is always better than underestimation.
695 if (!(dst_metric_locked(dst, RTAX_RTT))) {
697 dst->metrics[RTAX_RTT-1] = tp->srtt;
699 dst->metrics[RTAX_RTT-1] -= (m>>3);
702 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
706 /* Scale deviation to rttvar fixed point */
711 if (m >= dst_metric(dst, RTAX_RTTVAR))
712 dst->metrics[RTAX_RTTVAR-1] = m;
714 dst->metrics[RTAX_RTTVAR-1] -=
715 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
718 if (tp->snd_ssthresh >= 0xFFFF) {
719 /* Slow start still did not finish. */
720 if (dst_metric(dst, RTAX_SSTHRESH) &&
721 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
722 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
723 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
724 if (!dst_metric_locked(dst, RTAX_CWND) &&
725 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
726 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
727 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
728 icsk->icsk_ca_state == TCP_CA_Open) {
729 /* Cong. avoidance phase, cwnd is reliable. */
730 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
731 dst->metrics[RTAX_SSTHRESH-1] =
732 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
733 if (!dst_metric_locked(dst, RTAX_CWND))
734 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
736 /* Else slow start did not finish, cwnd is non-sense,
737 ssthresh may be also invalid.
739 if (!dst_metric_locked(dst, RTAX_CWND))
740 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
741 if (dst->metrics[RTAX_SSTHRESH-1] &&
742 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
743 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
744 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
747 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
748 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
749 tp->reordering != sysctl_tcp_reordering)
750 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
755 /* Numbers are taken from RFC2414. */
756 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
758 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
761 if (tp->mss_cache > 1460)
764 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
766 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
769 /* Set slow start threshold and cwnd not falling to slow start */
770 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
772 struct tcp_sock *tp = tcp_sk(sk);
773 const struct inet_connection_sock *icsk = inet_csk(sk);
775 tp->prior_ssthresh = 0;
777 if (icsk->icsk_ca_state < TCP_CA_CWR) {
780 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
781 tp->snd_cwnd = min(tp->snd_cwnd,
782 tcp_packets_in_flight(tp) + 1U);
783 tp->snd_cwnd_cnt = 0;
784 tp->high_seq = tp->snd_nxt;
785 tp->snd_cwnd_stamp = tcp_time_stamp;
786 TCP_ECN_queue_cwr(tp);
788 tcp_set_ca_state(sk, TCP_CA_CWR);
792 /* Initialize metrics on socket. */
794 static void tcp_init_metrics(struct sock *sk)
796 struct tcp_sock *tp = tcp_sk(sk);
797 struct dst_entry *dst = __sk_dst_get(sk);
804 if (dst_metric_locked(dst, RTAX_CWND))
805 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
806 if (dst_metric(dst, RTAX_SSTHRESH)) {
807 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
808 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
809 tp->snd_ssthresh = tp->snd_cwnd_clamp;
811 if (dst_metric(dst, RTAX_REORDERING) &&
812 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
813 tp->rx_opt.sack_ok &= ~2;
814 tp->reordering = dst_metric(dst, RTAX_REORDERING);
817 if (dst_metric(dst, RTAX_RTT) == 0)
820 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
823 /* Initial rtt is determined from SYN,SYN-ACK.
824 * The segment is small and rtt may appear much
825 * less than real one. Use per-dst memory
826 * to make it more realistic.
828 * A bit of theory. RTT is time passed after "normal" sized packet
829 * is sent until it is ACKed. In normal circumstances sending small
830 * packets force peer to delay ACKs and calculation is correct too.
831 * The algorithm is adaptive and, provided we follow specs, it
832 * NEVER underestimate RTT. BUT! If peer tries to make some clever
833 * tricks sort of "quick acks" for time long enough to decrease RTT
834 * to low value, and then abruptly stops to do it and starts to delay
835 * ACKs, wait for troubles.
837 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
838 tp->srtt = dst_metric(dst, RTAX_RTT);
839 tp->rtt_seq = tp->snd_nxt;
841 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
842 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
843 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
847 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
849 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
850 tp->snd_cwnd_stamp = tcp_time_stamp;
854 /* Play conservative. If timestamps are not
855 * supported, TCP will fail to recalculate correct
856 * rtt, if initial rto is too small. FORGET ALL AND RESET!
858 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
860 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
861 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
865 static void tcp_update_reordering(struct sock *sk, const int metric,
868 struct tcp_sock *tp = tcp_sk(sk);
869 if (metric > tp->reordering) {
870 tp->reordering = min(TCP_MAX_REORDERING, metric);
872 /* This exciting event is worth to be remembered. 8) */
874 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
876 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
878 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
880 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
881 #if FASTRETRANS_DEBUG > 1
882 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
883 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
887 tp->undo_marker ? tp->undo_retrans : 0);
889 /* Disable FACK yet. */
890 tp->rx_opt.sack_ok &= ~2;
894 /* This procedure tags the retransmission queue when SACKs arrive.
896 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
897 * Packets in queue with these bits set are counted in variables
898 * sacked_out, retrans_out and lost_out, correspondingly.
900 * Valid combinations are:
901 * Tag InFlight Description
902 * 0 1 - orig segment is in flight.
903 * S 0 - nothing flies, orig reached receiver.
904 * L 0 - nothing flies, orig lost by net.
905 * R 2 - both orig and retransmit are in flight.
906 * L|R 1 - orig is lost, retransmit is in flight.
907 * S|R 1 - orig reached receiver, retrans is still in flight.
908 * (L|S|R is logically valid, it could occur when L|R is sacked,
909 * but it is equivalent to plain S and code short-curcuits it to S.
910 * L|S is logically invalid, it would mean -1 packet in flight 8))
912 * These 6 states form finite state machine, controlled by the following events:
913 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
914 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
915 * 3. Loss detection event of one of three flavors:
916 * A. Scoreboard estimator decided the packet is lost.
917 * A'. Reno "three dupacks" marks head of queue lost.
918 * A''. Its FACK modfication, head until snd.fack is lost.
919 * B. SACK arrives sacking data transmitted after never retransmitted
921 * C. SACK arrives sacking SND.NXT at the moment, when the
922 * segment was retransmitted.
923 * 4. D-SACK added new rule: D-SACK changes any tag to S.
925 * It is pleasant to note, that state diagram turns out to be commutative,
926 * so that we are allowed not to be bothered by order of our actions,
927 * when multiple events arrive simultaneously. (see the function below).
929 * Reordering detection.
930 * --------------------
931 * Reordering metric is maximal distance, which a packet can be displaced
932 * in packet stream. With SACKs we can estimate it:
934 * 1. SACK fills old hole and the corresponding segment was not
935 * ever retransmitted -> reordering. Alas, we cannot use it
936 * when segment was retransmitted.
937 * 2. The last flaw is solved with D-SACK. D-SACK arrives
938 * for retransmitted and already SACKed segment -> reordering..
939 * Both of these heuristics are not used in Loss state, when we cannot
940 * account for retransmits accurately.
943 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
945 const struct inet_connection_sock *icsk = inet_csk(sk);
946 struct tcp_sock *tp = tcp_sk(sk);
947 unsigned char *ptr = (skb_transport_header(ack_skb) +
948 TCP_SKB_CB(ack_skb)->sacked);
949 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
950 struct sk_buff *cached_skb;
951 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
952 int reord = tp->packets_out;
954 u32 lost_retrans = 0;
956 int found_dup_sack = 0;
957 int cached_fack_count;
959 int first_sack_index;
963 prior_fackets = tp->fackets_out;
965 /* Check for D-SACK. */
966 if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
968 tp->rx_opt.sack_ok |= 4;
969 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
970 } else if (num_sacks > 1 &&
971 !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
972 !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
974 tp->rx_opt.sack_ok |= 4;
975 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
978 /* D-SACK for already forgotten data...
979 * Do dumb counting. */
980 if (found_dup_sack &&
981 !after(ntohl(sp[0].end_seq), prior_snd_una) &&
982 after(ntohl(sp[0].end_seq), tp->undo_marker))
985 /* Eliminate too old ACKs, but take into
986 * account more or less fresh ones, they can
987 * contain valid SACK info.
989 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
993 * if the only SACK change is the increase of the end_seq of
994 * the first block then only apply that SACK block
995 * and use retrans queue hinting otherwise slowpath */
997 for (i = 0; i < num_sacks; i++) {
998 __be32 start_seq = sp[i].start_seq;
999 __be32 end_seq = sp[i].end_seq;
1002 if (tp->recv_sack_cache[i].start_seq != start_seq)
1005 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1006 (tp->recv_sack_cache[i].end_seq != end_seq))
1009 tp->recv_sack_cache[i].start_seq = start_seq;
1010 tp->recv_sack_cache[i].end_seq = end_seq;
1012 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1013 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1014 tp->recv_sack_cache[i].start_seq = 0;
1015 tp->recv_sack_cache[i].end_seq = 0;
1018 first_sack_index = 0;
1023 tp->fastpath_skb_hint = NULL;
1025 /* order SACK blocks to allow in order walk of the retrans queue */
1026 for (i = num_sacks-1; i > 0; i--) {
1027 for (j = 0; j < i; j++){
1028 if (after(ntohl(sp[j].start_seq),
1029 ntohl(sp[j+1].start_seq))){
1030 struct tcp_sack_block_wire tmp;
1036 /* Track where the first SACK block goes to */
1037 if (j == first_sack_index)
1038 first_sack_index = j+1;
1045 /* clear flag as used for different purpose in following code */
1048 /* Use SACK fastpath hint if valid */
1049 cached_skb = tp->fastpath_skb_hint;
1050 cached_fack_count = tp->fastpath_cnt_hint;
1052 cached_skb = tcp_write_queue_head(sk);
1053 cached_fack_count = 0;
1056 for (i=0; i<num_sacks; i++, sp++) {
1057 struct sk_buff *skb;
1058 __u32 start_seq = ntohl(sp->start_seq);
1059 __u32 end_seq = ntohl(sp->end_seq);
1061 int dup_sack = (found_dup_sack && (i == first_sack_index));
1064 fack_count = cached_fack_count;
1066 /* Event "B" in the comment above. */
1067 if (after(end_seq, tp->high_seq))
1068 flag |= FLAG_DATA_LOST;
1070 tcp_for_write_queue_from(skb, sk) {
1071 int in_sack, pcount;
1074 if (skb == tcp_send_head(sk))
1078 cached_fack_count = fack_count;
1079 if (i == first_sack_index) {
1080 tp->fastpath_skb_hint = skb;
1081 tp->fastpath_cnt_hint = fack_count;
1084 /* The retransmission queue is always in order, so
1085 * we can short-circuit the walk early.
1087 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1090 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1091 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1093 pcount = tcp_skb_pcount(skb);
1095 if (pcount > 1 && !in_sack &&
1096 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1097 unsigned int pkt_len;
1099 in_sack = !after(start_seq,
1100 TCP_SKB_CB(skb)->seq);
1103 pkt_len = (start_seq -
1104 TCP_SKB_CB(skb)->seq);
1106 pkt_len = (end_seq -
1107 TCP_SKB_CB(skb)->seq);
1108 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1110 pcount = tcp_skb_pcount(skb);
1113 fack_count += pcount;
1115 sacked = TCP_SKB_CB(skb)->sacked;
1117 /* Account D-SACK for retransmitted packet. */
1118 if ((dup_sack && in_sack) &&
1119 (sacked & TCPCB_RETRANS) &&
1120 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1123 /* The frame is ACKed. */
1124 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1125 if (sacked&TCPCB_RETRANS) {
1126 if ((dup_sack && in_sack) &&
1127 (sacked&TCPCB_SACKED_ACKED))
1128 reord = min(fack_count, reord);
1130 /* If it was in a hole, we detected reordering. */
1131 if (fack_count < prior_fackets &&
1132 !(sacked&TCPCB_SACKED_ACKED))
1133 reord = min(fack_count, reord);
1136 /* Nothing to do; acked frame is about to be dropped. */
1140 if ((sacked&TCPCB_SACKED_RETRANS) &&
1141 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1142 (!lost_retrans || after(end_seq, lost_retrans)))
1143 lost_retrans = end_seq;
1148 if (!(sacked&TCPCB_SACKED_ACKED)) {
1149 if (sacked & TCPCB_SACKED_RETRANS) {
1150 /* If the segment is not tagged as lost,
1151 * we do not clear RETRANS, believing
1152 * that retransmission is still in flight.
1154 if (sacked & TCPCB_LOST) {
1155 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1156 tp->lost_out -= tcp_skb_pcount(skb);
1157 tp->retrans_out -= tcp_skb_pcount(skb);
1159 /* clear lost hint */
1160 tp->retransmit_skb_hint = NULL;
1163 /* New sack for not retransmitted frame,
1164 * which was in hole. It is reordering.
1166 if (!(sacked & TCPCB_RETRANS) &&
1167 fack_count < prior_fackets)
1168 reord = min(fack_count, reord);
1170 if (sacked & TCPCB_LOST) {
1171 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1172 tp->lost_out -= tcp_skb_pcount(skb);
1174 /* clear lost hint */
1175 tp->retransmit_skb_hint = NULL;
1177 /* SACK enhanced F-RTO detection.
1178 * Set flag if and only if non-rexmitted
1179 * segments below frto_highmark are
1180 * SACKed (RFC4138; Appendix B).
1181 * Clearing correct due to in-order walk
1183 if (after(end_seq, tp->frto_highmark)) {
1184 flag &= ~FLAG_ONLY_ORIG_SACKED;
1186 if (!(sacked & TCPCB_RETRANS))
1187 flag |= FLAG_ONLY_ORIG_SACKED;
1191 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1192 flag |= FLAG_DATA_SACKED;
1193 tp->sacked_out += tcp_skb_pcount(skb);
1195 if (fack_count > tp->fackets_out)
1196 tp->fackets_out = fack_count;
1198 if (dup_sack && (sacked&TCPCB_RETRANS))
1199 reord = min(fack_count, reord);
1202 /* D-SACK. We can detect redundant retransmission
1203 * in S|R and plain R frames and clear it.
1204 * undo_retrans is decreased above, L|R frames
1205 * are accounted above as well.
1208 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1209 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1210 tp->retrans_out -= tcp_skb_pcount(skb);
1211 tp->retransmit_skb_hint = NULL;
1216 /* Check for lost retransmit. This superb idea is
1217 * borrowed from "ratehalving". Event "C".
1218 * Later note: FACK people cheated me again 8),
1219 * we have to account for reordering! Ugly,
1222 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1223 struct sk_buff *skb;
1225 tcp_for_write_queue(skb, sk) {
1226 if (skb == tcp_send_head(sk))
1228 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1230 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1232 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1233 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1235 !before(lost_retrans,
1236 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1238 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1239 tp->retrans_out -= tcp_skb_pcount(skb);
1241 /* clear lost hint */
1242 tp->retransmit_skb_hint = NULL;
1244 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1245 tp->lost_out += tcp_skb_pcount(skb);
1246 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1247 flag |= FLAG_DATA_SACKED;
1248 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1254 tp->left_out = tp->sacked_out + tp->lost_out;
1256 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1257 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1258 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1260 #if FASTRETRANS_DEBUG > 0
1261 BUG_TRAP((int)tp->sacked_out >= 0);
1262 BUG_TRAP((int)tp->lost_out >= 0);
1263 BUG_TRAP((int)tp->retrans_out >= 0);
1264 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1269 /* F-RTO can only be used if TCP has never retransmitted anything other than
1270 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1272 int tcp_use_frto(struct sock *sk)
1274 const struct tcp_sock *tp = tcp_sk(sk);
1275 struct sk_buff *skb;
1277 if (!sysctl_tcp_frto)
1283 /* Avoid expensive walking of rexmit queue if possible */
1284 if (tp->retrans_out > 1)
1287 skb = tcp_write_queue_head(sk);
1288 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1289 tcp_for_write_queue_from(skb, sk) {
1290 if (skb == tcp_send_head(sk))
1292 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1294 /* Short-circuit when first non-SACKed skb has been checked */
1295 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1301 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1302 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1303 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1304 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1305 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1306 * bits are handled if the Loss state is really to be entered (in
1307 * tcp_enter_frto_loss).
1309 * Do like tcp_enter_loss() would; when RTO expires the second time it
1311 * "Reduce ssthresh if it has not yet been made inside this window."
1313 void tcp_enter_frto(struct sock *sk)
1315 const struct inet_connection_sock *icsk = inet_csk(sk);
1316 struct tcp_sock *tp = tcp_sk(sk);
1317 struct sk_buff *skb;
1319 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1320 tp->snd_una == tp->high_seq ||
1321 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1322 !icsk->icsk_retransmits)) {
1323 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1324 /* Our state is too optimistic in ssthresh() call because cwnd
1325 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1326 * recovery has not yet completed. Pattern would be this: RTO,
1327 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1329 * RFC4138 should be more specific on what to do, even though
1330 * RTO is quite unlikely to occur after the first Cumulative ACK
1331 * due to back-off and complexity of triggering events ...
1333 if (tp->frto_counter) {
1335 stored_cwnd = tp->snd_cwnd;
1337 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1338 tp->snd_cwnd = stored_cwnd;
1340 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1342 /* ... in theory, cong.control module could do "any tricks" in
1343 * ssthresh(), which means that ca_state, lost bits and lost_out
1344 * counter would have to be faked before the call occurs. We
1345 * consider that too expensive, unlikely and hacky, so modules
1346 * using these in ssthresh() must deal these incompatibility
1347 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1349 tcp_ca_event(sk, CA_EVENT_FRTO);
1352 tp->undo_marker = tp->snd_una;
1353 tp->undo_retrans = 0;
1355 skb = tcp_write_queue_head(sk);
1356 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1357 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1358 tp->retrans_out -= tcp_skb_pcount(skb);
1360 tcp_sync_left_out(tp);
1362 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1363 * The last condition is necessary at least in tp->frto_counter case.
1365 if (IsSackFrto() && (tp->frto_counter ||
1366 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1367 after(tp->high_seq, tp->snd_una)) {
1368 tp->frto_highmark = tp->high_seq;
1370 tp->frto_highmark = tp->snd_nxt;
1372 tcp_set_ca_state(sk, TCP_CA_Disorder);
1373 tp->high_seq = tp->snd_nxt;
1374 tp->frto_counter = 1;
1377 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1378 * which indicates that we should follow the traditional RTO recovery,
1379 * i.e. mark everything lost and do go-back-N retransmission.
1381 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1383 struct tcp_sock *tp = tcp_sk(sk);
1384 struct sk_buff *skb;
1389 tp->fackets_out = 0;
1390 tp->retrans_out = 0;
1392 tcp_for_write_queue(skb, sk) {
1393 if (skb == tcp_send_head(sk))
1395 cnt += tcp_skb_pcount(skb);
1397 * Count the retransmission made on RTO correctly (only when
1398 * waiting for the first ACK and did not get it)...
1400 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1401 /* For some reason this R-bit might get cleared? */
1402 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1403 tp->retrans_out += tcp_skb_pcount(skb);
1404 /* ...enter this if branch just for the first segment */
1405 flag |= FLAG_DATA_ACKED;
1407 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1409 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1411 /* Do not mark those segments lost that were
1412 * forward transmitted after RTO
1414 if (!after(TCP_SKB_CB(skb)->end_seq,
1415 tp->frto_highmark)) {
1416 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1417 tp->lost_out += tcp_skb_pcount(skb);
1420 tp->sacked_out += tcp_skb_pcount(skb);
1421 tp->fackets_out = cnt;
1424 tcp_sync_left_out(tp);
1426 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1427 tp->snd_cwnd_cnt = 0;
1428 tp->snd_cwnd_stamp = tcp_time_stamp;
1429 tp->undo_marker = 0;
1430 tp->frto_counter = 0;
1432 tp->reordering = min_t(unsigned int, tp->reordering,
1433 sysctl_tcp_reordering);
1434 tcp_set_ca_state(sk, TCP_CA_Loss);
1435 tp->high_seq = tp->frto_highmark;
1436 TCP_ECN_queue_cwr(tp);
1438 clear_all_retrans_hints(tp);
1441 void tcp_clear_retrans(struct tcp_sock *tp)
1444 tp->retrans_out = 0;
1446 tp->fackets_out = 0;
1450 tp->undo_marker = 0;
1451 tp->undo_retrans = 0;
1454 /* Enter Loss state. If "how" is not zero, forget all SACK information
1455 * and reset tags completely, otherwise preserve SACKs. If receiver
1456 * dropped its ofo queue, we will know this due to reneging detection.
1458 void tcp_enter_loss(struct sock *sk, int how)
1460 const struct inet_connection_sock *icsk = inet_csk(sk);
1461 struct tcp_sock *tp = tcp_sk(sk);
1462 struct sk_buff *skb;
1465 /* Reduce ssthresh if it has not yet been made inside this window. */
1466 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1467 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1468 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1469 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1470 tcp_ca_event(sk, CA_EVENT_LOSS);
1473 tp->snd_cwnd_cnt = 0;
1474 tp->snd_cwnd_stamp = tcp_time_stamp;
1476 tp->bytes_acked = 0;
1477 tcp_clear_retrans(tp);
1479 /* Push undo marker, if it was plain RTO and nothing
1480 * was retransmitted. */
1482 tp->undo_marker = tp->snd_una;
1484 tcp_for_write_queue(skb, sk) {
1485 if (skb == tcp_send_head(sk))
1487 cnt += tcp_skb_pcount(skb);
1488 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1489 tp->undo_marker = 0;
1490 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1491 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1492 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1493 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1494 tp->lost_out += tcp_skb_pcount(skb);
1496 tp->sacked_out += tcp_skb_pcount(skb);
1497 tp->fackets_out = cnt;
1500 tcp_sync_left_out(tp);
1502 tp->reordering = min_t(unsigned int, tp->reordering,
1503 sysctl_tcp_reordering);
1504 tcp_set_ca_state(sk, TCP_CA_Loss);
1505 tp->high_seq = tp->snd_nxt;
1506 TCP_ECN_queue_cwr(tp);
1507 /* Abort FRTO algorithm if one is in progress */
1508 tp->frto_counter = 0;
1510 clear_all_retrans_hints(tp);
1513 static int tcp_check_sack_reneging(struct sock *sk)
1515 struct sk_buff *skb;
1517 /* If ACK arrived pointing to a remembered SACK,
1518 * it means that our remembered SACKs do not reflect
1519 * real state of receiver i.e.
1520 * receiver _host_ is heavily congested (or buggy).
1521 * Do processing similar to RTO timeout.
1523 if ((skb = tcp_write_queue_head(sk)) != NULL &&
1524 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1525 struct inet_connection_sock *icsk = inet_csk(sk);
1526 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1528 tcp_enter_loss(sk, 1);
1529 icsk->icsk_retransmits++;
1530 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1531 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1532 icsk->icsk_rto, TCP_RTO_MAX);
1538 static inline int tcp_fackets_out(struct tcp_sock *tp)
1540 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1543 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1545 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1548 static inline int tcp_head_timedout(struct sock *sk)
1550 struct tcp_sock *tp = tcp_sk(sk);
1552 return tp->packets_out &&
1553 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1556 /* Linux NewReno/SACK/FACK/ECN state machine.
1557 * --------------------------------------
1559 * "Open" Normal state, no dubious events, fast path.
1560 * "Disorder" In all the respects it is "Open",
1561 * but requires a bit more attention. It is entered when
1562 * we see some SACKs or dupacks. It is split of "Open"
1563 * mainly to move some processing from fast path to slow one.
1564 * "CWR" CWND was reduced due to some Congestion Notification event.
1565 * It can be ECN, ICMP source quench, local device congestion.
1566 * "Recovery" CWND was reduced, we are fast-retransmitting.
1567 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1569 * tcp_fastretrans_alert() is entered:
1570 * - each incoming ACK, if state is not "Open"
1571 * - when arrived ACK is unusual, namely:
1576 * Counting packets in flight is pretty simple.
1578 * in_flight = packets_out - left_out + retrans_out
1580 * packets_out is SND.NXT-SND.UNA counted in packets.
1582 * retrans_out is number of retransmitted segments.
1584 * left_out is number of segments left network, but not ACKed yet.
1586 * left_out = sacked_out + lost_out
1588 * sacked_out: Packets, which arrived to receiver out of order
1589 * and hence not ACKed. With SACKs this number is simply
1590 * amount of SACKed data. Even without SACKs
1591 * it is easy to give pretty reliable estimate of this number,
1592 * counting duplicate ACKs.
1594 * lost_out: Packets lost by network. TCP has no explicit
1595 * "loss notification" feedback from network (for now).
1596 * It means that this number can be only _guessed_.
1597 * Actually, it is the heuristics to predict lossage that
1598 * distinguishes different algorithms.
1600 * F.e. after RTO, when all the queue is considered as lost,
1601 * lost_out = packets_out and in_flight = retrans_out.
1603 * Essentially, we have now two algorithms counting
1606 * FACK: It is the simplest heuristics. As soon as we decided
1607 * that something is lost, we decide that _all_ not SACKed
1608 * packets until the most forward SACK are lost. I.e.
1609 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1610 * It is absolutely correct estimate, if network does not reorder
1611 * packets. And it loses any connection to reality when reordering
1612 * takes place. We use FACK by default until reordering
1613 * is suspected on the path to this destination.
1615 * NewReno: when Recovery is entered, we assume that one segment
1616 * is lost (classic Reno). While we are in Recovery and
1617 * a partial ACK arrives, we assume that one more packet
1618 * is lost (NewReno). This heuristics are the same in NewReno
1621 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1622 * deflation etc. CWND is real congestion window, never inflated, changes
1623 * only according to classic VJ rules.
1625 * Really tricky (and requiring careful tuning) part of algorithm
1626 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1627 * The first determines the moment _when_ we should reduce CWND and,
1628 * hence, slow down forward transmission. In fact, it determines the moment
1629 * when we decide that hole is caused by loss, rather than by a reorder.
1631 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1632 * holes, caused by lost packets.
1634 * And the most logically complicated part of algorithm is undo
1635 * heuristics. We detect false retransmits due to both too early
1636 * fast retransmit (reordering) and underestimated RTO, analyzing
1637 * timestamps and D-SACKs. When we detect that some segments were
1638 * retransmitted by mistake and CWND reduction was wrong, we undo
1639 * window reduction and abort recovery phase. This logic is hidden
1640 * inside several functions named tcp_try_undo_<something>.
1643 /* This function decides, when we should leave Disordered state
1644 * and enter Recovery phase, reducing congestion window.
1646 * Main question: may we further continue forward transmission
1647 * with the same cwnd?
1649 static int tcp_time_to_recover(struct sock *sk)
1651 struct tcp_sock *tp = tcp_sk(sk);
1654 /* Do not perform any recovery during FRTO algorithm */
1655 if (tp->frto_counter)
1658 /* Trick#1: The loss is proven. */
1662 /* Not-A-Trick#2 : Classic rule... */
1663 if (tcp_fackets_out(tp) > tp->reordering)
1666 /* Trick#3 : when we use RFC2988 timer restart, fast
1667 * retransmit can be triggered by timeout of queue head.
1669 if (tcp_head_timedout(sk))
1672 /* Trick#4: It is still not OK... But will it be useful to delay
1675 packets_out = tp->packets_out;
1676 if (packets_out <= tp->reordering &&
1677 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1678 !tcp_may_send_now(sk)) {
1679 /* We have nothing to send. This connection is limited
1680 * either by receiver window or by application.
1688 /* If we receive more dupacks than we expected counting segments
1689 * in assumption of absent reordering, interpret this as reordering.
1690 * The only another reason could be bug in receiver TCP.
1692 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1694 struct tcp_sock *tp = tcp_sk(sk);
1697 holes = max(tp->lost_out, 1U);
1698 holes = min(holes, tp->packets_out);
1700 if ((tp->sacked_out + holes) > tp->packets_out) {
1701 tp->sacked_out = tp->packets_out - holes;
1702 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1706 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1708 static void tcp_add_reno_sack(struct sock *sk)
1710 struct tcp_sock *tp = tcp_sk(sk);
1712 tcp_check_reno_reordering(sk, 0);
1713 tcp_sync_left_out(tp);
1716 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1718 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1720 struct tcp_sock *tp = tcp_sk(sk);
1723 /* One ACK acked hole. The rest eat duplicate ACKs. */
1724 if (acked-1 >= tp->sacked_out)
1727 tp->sacked_out -= acked-1;
1729 tcp_check_reno_reordering(sk, acked);
1730 tcp_sync_left_out(tp);
1733 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1736 tp->left_out = tp->lost_out;
1739 /* Mark head of queue up as lost. */
1740 static void tcp_mark_head_lost(struct sock *sk,
1741 int packets, u32 high_seq)
1743 struct tcp_sock *tp = tcp_sk(sk);
1744 struct sk_buff *skb;
1747 BUG_TRAP(packets <= tp->packets_out);
1748 if (tp->lost_skb_hint) {
1749 skb = tp->lost_skb_hint;
1750 cnt = tp->lost_cnt_hint;
1752 skb = tcp_write_queue_head(sk);
1756 tcp_for_write_queue_from(skb, sk) {
1757 if (skb == tcp_send_head(sk))
1759 /* TODO: do this better */
1760 /* this is not the most efficient way to do this... */
1761 tp->lost_skb_hint = skb;
1762 tp->lost_cnt_hint = cnt;
1763 cnt += tcp_skb_pcount(skb);
1764 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1766 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1767 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1768 tp->lost_out += tcp_skb_pcount(skb);
1770 /* clear xmit_retransmit_queue hints
1771 * if this is beyond hint */
1772 if (tp->retransmit_skb_hint != NULL &&
1773 before(TCP_SKB_CB(skb)->seq,
1774 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1775 tp->retransmit_skb_hint = NULL;
1779 tcp_sync_left_out(tp);
1782 /* Account newly detected lost packet(s) */
1784 static void tcp_update_scoreboard(struct sock *sk)
1786 struct tcp_sock *tp = tcp_sk(sk);
1789 int lost = tp->fackets_out - tp->reordering;
1792 tcp_mark_head_lost(sk, lost, tp->high_seq);
1794 tcp_mark_head_lost(sk, 1, tp->high_seq);
1797 /* New heuristics: it is possible only after we switched
1798 * to restart timer each time when something is ACKed.
1799 * Hence, we can detect timed out packets during fast
1800 * retransmit without falling to slow start.
1802 if (!IsReno(tp) && tcp_head_timedout(sk)) {
1803 struct sk_buff *skb;
1805 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1806 : tcp_write_queue_head(sk);
1808 tcp_for_write_queue_from(skb, sk) {
1809 if (skb == tcp_send_head(sk))
1811 if (!tcp_skb_timedout(sk, skb))
1814 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1815 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1816 tp->lost_out += tcp_skb_pcount(skb);
1818 /* clear xmit_retrans hint */
1819 if (tp->retransmit_skb_hint &&
1820 before(TCP_SKB_CB(skb)->seq,
1821 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1823 tp->retransmit_skb_hint = NULL;
1827 tp->scoreboard_skb_hint = skb;
1829 tcp_sync_left_out(tp);
1833 /* CWND moderation, preventing bursts due to too big ACKs
1834 * in dubious situations.
1836 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1838 tp->snd_cwnd = min(tp->snd_cwnd,
1839 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1840 tp->snd_cwnd_stamp = tcp_time_stamp;
1843 /* Lower bound on congestion window is slow start threshold
1844 * unless congestion avoidance choice decides to overide it.
1846 static inline u32 tcp_cwnd_min(const struct sock *sk)
1848 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1850 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1853 /* Decrease cwnd each second ack. */
1854 static void tcp_cwnd_down(struct sock *sk)
1856 struct tcp_sock *tp = tcp_sk(sk);
1857 int decr = tp->snd_cwnd_cnt + 1;
1859 tp->snd_cwnd_cnt = decr&1;
1862 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1863 tp->snd_cwnd -= decr;
1865 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1866 tp->snd_cwnd_stamp = tcp_time_stamp;
1869 /* Nothing was retransmitted or returned timestamp is less
1870 * than timestamp of the first retransmission.
1872 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1874 return !tp->retrans_stamp ||
1875 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1876 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1879 /* Undo procedures. */
1881 #if FASTRETRANS_DEBUG > 1
1882 static void DBGUNDO(struct sock *sk, const char *msg)
1884 struct tcp_sock *tp = tcp_sk(sk);
1885 struct inet_sock *inet = inet_sk(sk);
1887 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1889 NIPQUAD(inet->daddr), ntohs(inet->dport),
1890 tp->snd_cwnd, tp->left_out,
1891 tp->snd_ssthresh, tp->prior_ssthresh,
1895 #define DBGUNDO(x...) do { } while (0)
1898 static void tcp_undo_cwr(struct sock *sk, const int undo)
1900 struct tcp_sock *tp = tcp_sk(sk);
1902 if (tp->prior_ssthresh) {
1903 const struct inet_connection_sock *icsk = inet_csk(sk);
1905 if (icsk->icsk_ca_ops->undo_cwnd)
1906 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1908 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1910 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1911 tp->snd_ssthresh = tp->prior_ssthresh;
1912 TCP_ECN_withdraw_cwr(tp);
1915 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1917 tcp_moderate_cwnd(tp);
1918 tp->snd_cwnd_stamp = tcp_time_stamp;
1920 /* There is something screwy going on with the retrans hints after
1922 clear_all_retrans_hints(tp);
1925 static inline int tcp_may_undo(struct tcp_sock *tp)
1927 return tp->undo_marker &&
1928 (!tp->undo_retrans || tcp_packet_delayed(tp));
1931 /* People celebrate: "We love our President!" */
1932 static int tcp_try_undo_recovery(struct sock *sk)
1934 struct tcp_sock *tp = tcp_sk(sk);
1936 if (tcp_may_undo(tp)) {
1937 /* Happy end! We did not retransmit anything
1938 * or our original transmission succeeded.
1940 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1941 tcp_undo_cwr(sk, 1);
1942 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1943 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1945 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1946 tp->undo_marker = 0;
1948 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1949 /* Hold old state until something *above* high_seq
1950 * is ACKed. For Reno it is MUST to prevent false
1951 * fast retransmits (RFC2582). SACK TCP is safe. */
1952 tcp_moderate_cwnd(tp);
1955 tcp_set_ca_state(sk, TCP_CA_Open);
1959 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1960 static void tcp_try_undo_dsack(struct sock *sk)
1962 struct tcp_sock *tp = tcp_sk(sk);
1964 if (tp->undo_marker && !tp->undo_retrans) {
1965 DBGUNDO(sk, "D-SACK");
1966 tcp_undo_cwr(sk, 1);
1967 tp->undo_marker = 0;
1968 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1972 /* Undo during fast recovery after partial ACK. */
1974 static int tcp_try_undo_partial(struct sock *sk, int acked)
1976 struct tcp_sock *tp = tcp_sk(sk);
1977 /* Partial ACK arrived. Force Hoe's retransmit. */
1978 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1980 if (tcp_may_undo(tp)) {
1981 /* Plain luck! Hole if filled with delayed
1982 * packet, rather than with a retransmit.
1984 if (tp->retrans_out == 0)
1985 tp->retrans_stamp = 0;
1987 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1990 tcp_undo_cwr(sk, 0);
1991 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1993 /* So... Do not make Hoe's retransmit yet.
1994 * If the first packet was delayed, the rest
1995 * ones are most probably delayed as well.
2002 /* Undo during loss recovery after partial ACK. */
2003 static int tcp_try_undo_loss(struct sock *sk)
2005 struct tcp_sock *tp = tcp_sk(sk);
2007 if (tcp_may_undo(tp)) {
2008 struct sk_buff *skb;
2009 tcp_for_write_queue(skb, sk) {
2010 if (skb == tcp_send_head(sk))
2012 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2015 clear_all_retrans_hints(tp);
2017 DBGUNDO(sk, "partial loss");
2019 tp->left_out = tp->sacked_out;
2020 tcp_undo_cwr(sk, 1);
2021 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2022 inet_csk(sk)->icsk_retransmits = 0;
2023 tp->undo_marker = 0;
2025 tcp_set_ca_state(sk, TCP_CA_Open);
2031 static inline void tcp_complete_cwr(struct sock *sk)
2033 struct tcp_sock *tp = tcp_sk(sk);
2034 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2035 tp->snd_cwnd_stamp = tcp_time_stamp;
2036 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2039 static void tcp_try_to_open(struct sock *sk, int flag)
2041 struct tcp_sock *tp = tcp_sk(sk);
2043 tcp_sync_left_out(tp);
2045 if (tp->retrans_out == 0)
2046 tp->retrans_stamp = 0;
2049 tcp_enter_cwr(sk, 1);
2051 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2052 int state = TCP_CA_Open;
2054 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2055 state = TCP_CA_Disorder;
2057 if (inet_csk(sk)->icsk_ca_state != state) {
2058 tcp_set_ca_state(sk, state);
2059 tp->high_seq = tp->snd_nxt;
2061 tcp_moderate_cwnd(tp);
2067 static void tcp_mtup_probe_failed(struct sock *sk)
2069 struct inet_connection_sock *icsk = inet_csk(sk);
2071 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2072 icsk->icsk_mtup.probe_size = 0;
2075 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2077 struct tcp_sock *tp = tcp_sk(sk);
2078 struct inet_connection_sock *icsk = inet_csk(sk);
2080 /* FIXME: breaks with very large cwnd */
2081 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2082 tp->snd_cwnd = tp->snd_cwnd *
2083 tcp_mss_to_mtu(sk, tp->mss_cache) /
2084 icsk->icsk_mtup.probe_size;
2085 tp->snd_cwnd_cnt = 0;
2086 tp->snd_cwnd_stamp = tcp_time_stamp;
2087 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2089 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2090 icsk->icsk_mtup.probe_size = 0;
2091 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2095 /* Process an event, which can update packets-in-flight not trivially.
2096 * Main goal of this function is to calculate new estimate for left_out,
2097 * taking into account both packets sitting in receiver's buffer and
2098 * packets lost by network.
2100 * Besides that it does CWND reduction, when packet loss is detected
2101 * and changes state of machine.
2103 * It does _not_ decide what to send, it is made in function
2104 * tcp_xmit_retransmit_queue().
2107 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
2108 int prior_packets, int flag)
2110 struct inet_connection_sock *icsk = inet_csk(sk);
2111 struct tcp_sock *tp = tcp_sk(sk);
2112 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
2114 /* Some technical things:
2115 * 1. Reno does not count dupacks (sacked_out) automatically. */
2116 if (!tp->packets_out)
2118 /* 2. SACK counts snd_fack in packets inaccurately. */
2119 if (tp->sacked_out == 0)
2120 tp->fackets_out = 0;
2122 /* Now state machine starts.
2123 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2125 tp->prior_ssthresh = 0;
2127 /* B. In all the states check for reneging SACKs. */
2128 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2131 /* C. Process data loss notification, provided it is valid. */
2132 if ((flag&FLAG_DATA_LOST) &&
2133 before(tp->snd_una, tp->high_seq) &&
2134 icsk->icsk_ca_state != TCP_CA_Open &&
2135 tp->fackets_out > tp->reordering) {
2136 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2137 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2140 /* D. Synchronize left_out to current state. */
2141 tcp_sync_left_out(tp);
2143 /* E. Check state exit conditions. State can be terminated
2144 * when high_seq is ACKed. */
2145 if (icsk->icsk_ca_state == TCP_CA_Open) {
2146 BUG_TRAP(tp->retrans_out == 0);
2147 tp->retrans_stamp = 0;
2148 } else if (!before(tp->snd_una, tp->high_seq)) {
2149 switch (icsk->icsk_ca_state) {
2151 icsk->icsk_retransmits = 0;
2152 if (tcp_try_undo_recovery(sk))
2157 /* CWR is to be held something *above* high_seq
2158 * is ACKed for CWR bit to reach receiver. */
2159 if (tp->snd_una != tp->high_seq) {
2160 tcp_complete_cwr(sk);
2161 tcp_set_ca_state(sk, TCP_CA_Open);
2165 case TCP_CA_Disorder:
2166 tcp_try_undo_dsack(sk);
2167 if (!tp->undo_marker ||
2168 /* For SACK case do not Open to allow to undo
2169 * catching for all duplicate ACKs. */
2170 IsReno(tp) || tp->snd_una != tp->high_seq) {
2171 tp->undo_marker = 0;
2172 tcp_set_ca_state(sk, TCP_CA_Open);
2176 case TCP_CA_Recovery:
2178 tcp_reset_reno_sack(tp);
2179 if (tcp_try_undo_recovery(sk))
2181 tcp_complete_cwr(sk);
2186 /* F. Process state. */
2187 switch (icsk->icsk_ca_state) {
2188 case TCP_CA_Recovery:
2189 if (prior_snd_una == tp->snd_una) {
2190 if (IsReno(tp) && is_dupack)
2191 tcp_add_reno_sack(sk);
2193 int acked = prior_packets - tp->packets_out;
2195 tcp_remove_reno_sacks(sk, acked);
2196 is_dupack = tcp_try_undo_partial(sk, acked);
2200 if (flag&FLAG_DATA_ACKED)
2201 icsk->icsk_retransmits = 0;
2202 if (!tcp_try_undo_loss(sk)) {
2203 tcp_moderate_cwnd(tp);
2204 tcp_xmit_retransmit_queue(sk);
2207 if (icsk->icsk_ca_state != TCP_CA_Open)
2209 /* Loss is undone; fall through to processing in Open state. */
2212 if (tp->snd_una != prior_snd_una)
2213 tcp_reset_reno_sack(tp);
2215 tcp_add_reno_sack(sk);
2218 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2219 tcp_try_undo_dsack(sk);
2221 if (!tcp_time_to_recover(sk)) {
2222 tcp_try_to_open(sk, flag);
2226 /* MTU probe failure: don't reduce cwnd */
2227 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2228 icsk->icsk_mtup.probe_size &&
2229 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2230 tcp_mtup_probe_failed(sk);
2231 /* Restores the reduction we did in tcp_mtup_probe() */
2233 tcp_simple_retransmit(sk);
2237 /* Otherwise enter Recovery state */
2240 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2242 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2244 tp->high_seq = tp->snd_nxt;
2245 tp->prior_ssthresh = 0;
2246 tp->undo_marker = tp->snd_una;
2247 tp->undo_retrans = tp->retrans_out;
2249 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2250 if (!(flag&FLAG_ECE))
2251 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2252 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2253 TCP_ECN_queue_cwr(tp);
2256 tp->bytes_acked = 0;
2257 tp->snd_cwnd_cnt = 0;
2258 tcp_set_ca_state(sk, TCP_CA_Recovery);
2261 if (is_dupack || tcp_head_timedout(sk))
2262 tcp_update_scoreboard(sk);
2264 tcp_xmit_retransmit_queue(sk);
2267 /* Read draft-ietf-tcplw-high-performance before mucking
2268 * with this code. (Supersedes RFC1323)
2270 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2272 /* RTTM Rule: A TSecr value received in a segment is used to
2273 * update the averaged RTT measurement only if the segment
2274 * acknowledges some new data, i.e., only if it advances the
2275 * left edge of the send window.
2277 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2278 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2280 * Changed: reset backoff as soon as we see the first valid sample.
2281 * If we do not, we get strongly overestimated rto. With timestamps
2282 * samples are accepted even from very old segments: f.e., when rtt=1
2283 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2284 * answer arrives rto becomes 120 seconds! If at least one of segments
2285 * in window is lost... Voila. --ANK (010210)
2287 struct tcp_sock *tp = tcp_sk(sk);
2288 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2289 tcp_rtt_estimator(sk, seq_rtt);
2291 inet_csk(sk)->icsk_backoff = 0;
2295 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2297 /* We don't have a timestamp. Can only use
2298 * packets that are not retransmitted to determine
2299 * rtt estimates. Also, we must not reset the
2300 * backoff for rto until we get a non-retransmitted
2301 * packet. This allows us to deal with a situation
2302 * where the network delay has increased suddenly.
2303 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2306 if (flag & FLAG_RETRANS_DATA_ACKED)
2309 tcp_rtt_estimator(sk, seq_rtt);
2311 inet_csk(sk)->icsk_backoff = 0;
2315 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2318 const struct tcp_sock *tp = tcp_sk(sk);
2319 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2320 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2321 tcp_ack_saw_tstamp(sk, flag);
2322 else if (seq_rtt >= 0)
2323 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2326 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2327 u32 in_flight, int good)
2329 const struct inet_connection_sock *icsk = inet_csk(sk);
2330 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2331 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2334 /* Restart timer after forward progress on connection.
2335 * RFC2988 recommends to restart timer to now+rto.
2338 static void tcp_ack_packets_out(struct sock *sk)
2340 struct tcp_sock *tp = tcp_sk(sk);
2342 if (!tp->packets_out) {
2343 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2345 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2349 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2350 __u32 now, __s32 *seq_rtt)
2352 struct tcp_sock *tp = tcp_sk(sk);
2353 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2354 __u32 seq = tp->snd_una;
2355 __u32 packets_acked;
2358 /* If we get here, the whole TSO packet has not been
2361 BUG_ON(!after(scb->end_seq, seq));
2363 packets_acked = tcp_skb_pcount(skb);
2364 if (tcp_trim_head(sk, skb, seq - scb->seq))
2366 packets_acked -= tcp_skb_pcount(skb);
2368 if (packets_acked) {
2369 __u8 sacked = scb->sacked;
2371 acked |= FLAG_DATA_ACKED;
2373 if (sacked & TCPCB_RETRANS) {
2374 if (sacked & TCPCB_SACKED_RETRANS)
2375 tp->retrans_out -= packets_acked;
2376 acked |= FLAG_RETRANS_DATA_ACKED;
2378 } else if (*seq_rtt < 0)
2379 *seq_rtt = now - scb->when;
2380 if (sacked & TCPCB_SACKED_ACKED)
2381 tp->sacked_out -= packets_acked;
2382 if (sacked & TCPCB_LOST)
2383 tp->lost_out -= packets_acked;
2384 if (sacked & TCPCB_URG) {
2386 !before(seq, tp->snd_up))
2389 } else if (*seq_rtt < 0)
2390 *seq_rtt = now - scb->when;
2392 if (tp->fackets_out) {
2393 __u32 dval = min(tp->fackets_out, packets_acked);
2394 tp->fackets_out -= dval;
2396 tp->packets_out -= packets_acked;
2398 BUG_ON(tcp_skb_pcount(skb) == 0);
2399 BUG_ON(!before(scb->seq, scb->end_seq));
2405 /* Remove acknowledged frames from the retransmission queue. */
2406 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2408 struct tcp_sock *tp = tcp_sk(sk);
2409 const struct inet_connection_sock *icsk = inet_csk(sk);
2410 struct sk_buff *skb;
2411 __u32 now = tcp_time_stamp;
2413 int prior_packets = tp->packets_out;
2415 ktime_t last_ackt = net_invalid_timestamp();
2417 while ((skb = tcp_write_queue_head(sk)) &&
2418 skb != tcp_send_head(sk)) {
2419 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2420 __u8 sacked = scb->sacked;
2422 /* If our packet is before the ack sequence we can
2423 * discard it as it's confirmed to have arrived at
2426 if (after(scb->end_seq, tp->snd_una)) {
2427 if (tcp_skb_pcount(skb) > 1 &&
2428 after(tp->snd_una, scb->seq))
2429 acked |= tcp_tso_acked(sk, skb,
2434 /* Initial outgoing SYN's get put onto the write_queue
2435 * just like anything else we transmit. It is not
2436 * true data, and if we misinform our callers that
2437 * this ACK acks real data, we will erroneously exit
2438 * connection startup slow start one packet too
2439 * quickly. This is severely frowned upon behavior.
2441 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2442 acked |= FLAG_DATA_ACKED;
2444 acked |= FLAG_SYN_ACKED;
2445 tp->retrans_stamp = 0;
2448 /* MTU probing checks */
2449 if (icsk->icsk_mtup.probe_size) {
2450 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2451 tcp_mtup_probe_success(sk, skb);
2456 if (sacked & TCPCB_RETRANS) {
2457 if (sacked & TCPCB_SACKED_RETRANS)
2458 tp->retrans_out -= tcp_skb_pcount(skb);
2459 acked |= FLAG_RETRANS_DATA_ACKED;
2461 } else if (seq_rtt < 0) {
2462 seq_rtt = now - scb->when;
2463 last_ackt = skb->tstamp;
2465 if (sacked & TCPCB_SACKED_ACKED)
2466 tp->sacked_out -= tcp_skb_pcount(skb);
2467 if (sacked & TCPCB_LOST)
2468 tp->lost_out -= tcp_skb_pcount(skb);
2469 if (sacked & TCPCB_URG) {
2471 !before(scb->end_seq, tp->snd_up))
2474 } else if (seq_rtt < 0) {
2475 seq_rtt = now - scb->when;
2476 last_ackt = skb->tstamp;
2478 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2479 tcp_packets_out_dec(tp, skb);
2480 tcp_unlink_write_queue(skb, sk);
2481 sk_stream_free_skb(sk, skb);
2482 clear_all_retrans_hints(tp);
2485 if (acked&FLAG_ACKED) {
2486 u32 pkts_acked = prior_packets - tp->packets_out;
2487 const struct tcp_congestion_ops *ca_ops
2488 = inet_csk(sk)->icsk_ca_ops;
2490 tcp_ack_update_rtt(sk, acked, seq_rtt);
2491 tcp_ack_packets_out(sk);
2493 if (ca_ops->pkts_acked) {
2496 /* Is the ACK triggering packet unambiguous? */
2497 if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
2498 /* High resolution needed and available? */
2499 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2500 !ktime_equal(last_ackt,
2501 net_invalid_timestamp()))
2502 rtt_us = ktime_us_delta(ktime_get_real(),
2504 else if (seq_rtt > 0)
2505 rtt_us = jiffies_to_usecs(seq_rtt);
2508 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2512 #if FASTRETRANS_DEBUG > 0
2513 BUG_TRAP((int)tp->sacked_out >= 0);
2514 BUG_TRAP((int)tp->lost_out >= 0);
2515 BUG_TRAP((int)tp->retrans_out >= 0);
2516 if (!tp->packets_out && tp->rx_opt.sack_ok) {
2517 const struct inet_connection_sock *icsk = inet_csk(sk);
2519 printk(KERN_DEBUG "Leak l=%u %d\n",
2520 tp->lost_out, icsk->icsk_ca_state);
2523 if (tp->sacked_out) {
2524 printk(KERN_DEBUG "Leak s=%u %d\n",
2525 tp->sacked_out, icsk->icsk_ca_state);
2528 if (tp->retrans_out) {
2529 printk(KERN_DEBUG "Leak r=%u %d\n",
2530 tp->retrans_out, icsk->icsk_ca_state);
2531 tp->retrans_out = 0;
2535 *seq_rtt_p = seq_rtt;
2539 static void tcp_ack_probe(struct sock *sk)
2541 const struct tcp_sock *tp = tcp_sk(sk);
2542 struct inet_connection_sock *icsk = inet_csk(sk);
2544 /* Was it a usable window open? */
2546 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2547 tp->snd_una + tp->snd_wnd)) {
2548 icsk->icsk_backoff = 0;
2549 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2550 /* Socket must be waked up by subsequent tcp_data_snd_check().
2551 * This function is not for random using!
2554 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2555 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2560 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2562 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2563 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2566 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2568 const struct tcp_sock *tp = tcp_sk(sk);
2569 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2570 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2573 /* Check that window update is acceptable.
2574 * The function assumes that snd_una<=ack<=snd_next.
2576 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2577 const u32 ack_seq, const u32 nwin)
2579 return (after(ack, tp->snd_una) ||
2580 after(ack_seq, tp->snd_wl1) ||
2581 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2584 /* Update our send window.
2586 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2587 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2589 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2592 struct tcp_sock *tp = tcp_sk(sk);
2594 u32 nwin = ntohs(tcp_hdr(skb)->window);
2596 if (likely(!tcp_hdr(skb)->syn))
2597 nwin <<= tp->rx_opt.snd_wscale;
2599 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2600 flag |= FLAG_WIN_UPDATE;
2601 tcp_update_wl(tp, ack, ack_seq);
2603 if (tp->snd_wnd != nwin) {
2606 /* Note, it is the only place, where
2607 * fast path is recovered for sending TCP.
2610 tcp_fast_path_check(sk);
2612 if (nwin > tp->max_window) {
2613 tp->max_window = nwin;
2614 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2624 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2625 * continue in congestion avoidance.
2627 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2629 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2630 tp->snd_cwnd_cnt = 0;
2631 TCP_ECN_queue_cwr(tp);
2632 tcp_moderate_cwnd(tp);
2635 /* A conservative spurious RTO response algorithm: reduce cwnd using
2636 * rate halving and continue in congestion avoidance.
2638 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2640 tcp_enter_cwr(sk, 0);
2643 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2646 tcp_ratehalving_spur_to_response(sk);
2648 tcp_undo_cwr(sk, 1);
2651 /* F-RTO spurious RTO detection algorithm (RFC4138)
2653 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2654 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2655 * window (but not to or beyond highest sequence sent before RTO):
2656 * On First ACK, send two new segments out.
2657 * On Second ACK, RTO was likely spurious. Do spurious response (response
2658 * algorithm is not part of the F-RTO detection algorithm
2659 * given in RFC4138 but can be selected separately).
2660 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2661 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2662 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2663 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2665 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2666 * original window even after we transmit two new data segments.
2669 * on first step, wait until first cumulative ACK arrives, then move to
2670 * the second step. In second step, the next ACK decides.
2672 * F-RTO is implemented (mainly) in four functions:
2673 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2674 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2675 * called when tcp_use_frto() showed green light
2676 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2677 * - tcp_enter_frto_loss() is called if there is not enough evidence
2678 * to prove that the RTO is indeed spurious. It transfers the control
2679 * from F-RTO to the conventional RTO recovery
2681 static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
2683 struct tcp_sock *tp = tcp_sk(sk);
2685 tcp_sync_left_out(tp);
2687 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2688 if (flag&FLAG_DATA_ACKED)
2689 inet_csk(sk)->icsk_retransmits = 0;
2691 if (!before(tp->snd_una, tp->frto_highmark)) {
2692 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2696 if (!IsSackFrto() || IsReno(tp)) {
2697 /* RFC4138 shortcoming in step 2; should also have case c):
2698 * ACK isn't duplicate nor advances window, e.g., opposite dir
2701 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
2702 !(flag&FLAG_FORWARD_PROGRESS))
2705 if (!(flag&FLAG_DATA_ACKED)) {
2706 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2711 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2712 /* Prevent sending of new data. */
2713 tp->snd_cwnd = min(tp->snd_cwnd,
2714 tcp_packets_in_flight(tp));
2718 if ((tp->frto_counter >= 2) &&
2719 (!(flag&FLAG_FORWARD_PROGRESS) ||
2720 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2721 /* RFC4138 shortcoming (see comment above) */
2722 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2725 tcp_enter_frto_loss(sk, 3, flag);
2730 if (tp->frto_counter == 1) {
2731 /* Sending of the next skb must be allowed or no FRTO */
2732 if (!tcp_send_head(sk) ||
2733 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2734 tp->snd_una + tp->snd_wnd)) {
2735 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2740 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2741 tp->frto_counter = 2;
2744 switch (sysctl_tcp_frto_response) {
2746 tcp_undo_spur_to_response(sk, flag);
2749 tcp_conservative_spur_to_response(tp);
2752 tcp_ratehalving_spur_to_response(sk);
2755 tp->frto_counter = 0;
2760 /* This routine deals with incoming acks, but not outgoing ones. */
2761 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2763 struct inet_connection_sock *icsk = inet_csk(sk);
2764 struct tcp_sock *tp = tcp_sk(sk);
2765 u32 prior_snd_una = tp->snd_una;
2766 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2767 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2768 u32 prior_in_flight;
2773 /* If the ack is newer than sent or older than previous acks
2774 * then we can probably ignore it.
2776 if (after(ack, tp->snd_nxt))
2777 goto uninteresting_ack;
2779 if (before(ack, prior_snd_una))
2782 if (sysctl_tcp_abc) {
2783 if (icsk->icsk_ca_state < TCP_CA_CWR)
2784 tp->bytes_acked += ack - prior_snd_una;
2785 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2786 /* we assume just one segment left network */
2787 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2790 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2791 /* Window is constant, pure forward advance.
2792 * No more checks are required.
2793 * Note, we use the fact that SND.UNA>=SND.WL2.
2795 tcp_update_wl(tp, ack, ack_seq);
2797 flag |= FLAG_WIN_UPDATE;
2799 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2801 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2803 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2806 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2808 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2810 if (TCP_SKB_CB(skb)->sacked)
2811 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2813 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2816 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2819 /* We passed data and got it acked, remove any soft error
2820 * log. Something worked...
2822 sk->sk_err_soft = 0;
2823 tp->rcv_tstamp = tcp_time_stamp;
2824 prior_packets = tp->packets_out;
2828 prior_in_flight = tcp_packets_in_flight(tp);
2830 /* See if we can take anything off of the retransmit queue. */
2831 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2833 if (tp->frto_counter)
2834 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag);
2836 if (tcp_ack_is_dubious(sk, flag)) {
2837 /* Advance CWND, if state allows this. */
2838 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2839 tcp_may_raise_cwnd(sk, flag))
2840 tcp_cong_avoid(sk, ack, prior_in_flight, 0);
2841 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2843 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2844 tcp_cong_avoid(sk, ack, prior_in_flight, 1);
2847 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2848 dst_confirm(sk->sk_dst_cache);
2853 icsk->icsk_probes_out = 0;
2855 /* If this ack opens up a zero window, clear backoff. It was
2856 * being used to time the probes, and is probably far higher than
2857 * it needs to be for normal retransmission.
2859 if (tcp_send_head(sk))
2864 if (TCP_SKB_CB(skb)->sacked)
2865 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2868 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2873 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2874 * But, this can also be called on packets in the established flow when
2875 * the fast version below fails.
2877 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2880 struct tcphdr *th = tcp_hdr(skb);
2881 int length=(th->doff*4)-sizeof(struct tcphdr);
2883 ptr = (unsigned char *)(th + 1);
2884 opt_rx->saw_tstamp = 0;
2886 while (length > 0) {
2893 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2898 if (opsize < 2) /* "silly options" */
2900 if (opsize > length)
2901 return; /* don't parse partial options */
2904 if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2905 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2907 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2908 in_mss = opt_rx->user_mss;
2909 opt_rx->mss_clamp = in_mss;
2914 if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2915 if (sysctl_tcp_window_scaling) {
2916 __u8 snd_wscale = *(__u8 *) ptr;
2917 opt_rx->wscale_ok = 1;
2918 if (snd_wscale > 14) {
2919 if (net_ratelimit())
2920 printk(KERN_INFO "tcp_parse_options: Illegal window "
2921 "scaling value %d >14 received.\n",
2925 opt_rx->snd_wscale = snd_wscale;
2928 case TCPOPT_TIMESTAMP:
2929 if (opsize==TCPOLEN_TIMESTAMP) {
2930 if ((estab && opt_rx->tstamp_ok) ||
2931 (!estab && sysctl_tcp_timestamps)) {
2932 opt_rx->saw_tstamp = 1;
2933 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2934 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2938 case TCPOPT_SACK_PERM:
2939 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2940 if (sysctl_tcp_sack) {
2941 opt_rx->sack_ok = 1;
2942 tcp_sack_reset(opt_rx);
2948 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2949 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2951 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2954 #ifdef CONFIG_TCP_MD5SIG
2957 * The MD5 Hash has already been
2958 * checked (see tcp_v{4,6}_do_rcv()).
2970 /* Fast parse options. This hopes to only see timestamps.
2971 * If it is wrong it falls back on tcp_parse_options().
2973 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2974 struct tcp_sock *tp)
2976 if (th->doff == sizeof(struct tcphdr)>>2) {
2977 tp->rx_opt.saw_tstamp = 0;
2979 } else if (tp->rx_opt.tstamp_ok &&
2980 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2981 __be32 *ptr = (__be32 *)(th + 1);
2982 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2983 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2984 tp->rx_opt.saw_tstamp = 1;
2986 tp->rx_opt.rcv_tsval = ntohl(*ptr);
2988 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2992 tcp_parse_options(skb, &tp->rx_opt, 1);
2996 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2998 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2999 tp->rx_opt.ts_recent_stamp = get_seconds();
3002 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3004 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3005 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3006 * extra check below makes sure this can only happen
3007 * for pure ACK frames. -DaveM
3009 * Not only, also it occurs for expired timestamps.
3012 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3013 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3014 tcp_store_ts_recent(tp);
3018 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3020 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3021 * it can pass through stack. So, the following predicate verifies that
3022 * this segment is not used for anything but congestion avoidance or
3023 * fast retransmit. Moreover, we even are able to eliminate most of such
3024 * second order effects, if we apply some small "replay" window (~RTO)
3025 * to timestamp space.
3027 * All these measures still do not guarantee that we reject wrapped ACKs
3028 * on networks with high bandwidth, when sequence space is recycled fastly,
3029 * but it guarantees that such events will be very rare and do not affect
3030 * connection seriously. This doesn't look nice, but alas, PAWS is really
3033 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3034 * states that events when retransmit arrives after original data are rare.
3035 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3036 * the biggest problem on large power networks even with minor reordering.
3037 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3038 * up to bandwidth of 18Gigabit/sec. 8) ]
3041 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3043 struct tcp_sock *tp = tcp_sk(sk);
3044 struct tcphdr *th = tcp_hdr(skb);
3045 u32 seq = TCP_SKB_CB(skb)->seq;
3046 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3048 return (/* 1. Pure ACK with correct sequence number. */
3049 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3051 /* 2. ... and duplicate ACK. */
3052 ack == tp->snd_una &&
3054 /* 3. ... and does not update window. */
3055 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3057 /* 4. ... and sits in replay window. */
3058 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3061 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3063 const struct tcp_sock *tp = tcp_sk(sk);
3064 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3065 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3066 !tcp_disordered_ack(sk, skb));
3069 /* Check segment sequence number for validity.
3071 * Segment controls are considered valid, if the segment
3072 * fits to the window after truncation to the window. Acceptability
3073 * of data (and SYN, FIN, of course) is checked separately.
3074 * See tcp_data_queue(), for example.
3076 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3077 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3078 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3079 * (borrowed from freebsd)
3082 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3084 return !before(end_seq, tp->rcv_wup) &&
3085 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3088 /* When we get a reset we do this. */
3089 static void tcp_reset(struct sock *sk)
3091 /* We want the right error as BSD sees it (and indeed as we do). */
3092 switch (sk->sk_state) {
3094 sk->sk_err = ECONNREFUSED;
3096 case TCP_CLOSE_WAIT:
3102 sk->sk_err = ECONNRESET;
3105 if (!sock_flag(sk, SOCK_DEAD))
3106 sk->sk_error_report(sk);
3112 * Process the FIN bit. This now behaves as it is supposed to work
3113 * and the FIN takes effect when it is validly part of sequence
3114 * space. Not before when we get holes.
3116 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3117 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3120 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3121 * close and we go into CLOSING (and later onto TIME-WAIT)
3123 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3125 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3127 struct tcp_sock *tp = tcp_sk(sk);
3129 inet_csk_schedule_ack(sk);
3131 sk->sk_shutdown |= RCV_SHUTDOWN;
3132 sock_set_flag(sk, SOCK_DONE);
3134 switch (sk->sk_state) {
3136 case TCP_ESTABLISHED:
3137 /* Move to CLOSE_WAIT */
3138 tcp_set_state(sk, TCP_CLOSE_WAIT);
3139 inet_csk(sk)->icsk_ack.pingpong = 1;
3142 case TCP_CLOSE_WAIT:
3144 /* Received a retransmission of the FIN, do
3149 /* RFC793: Remain in the LAST-ACK state. */
3153 /* This case occurs when a simultaneous close
3154 * happens, we must ack the received FIN and
3155 * enter the CLOSING state.
3158 tcp_set_state(sk, TCP_CLOSING);
3161 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3163 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3166 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3167 * cases we should never reach this piece of code.
3169 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3170 __FUNCTION__, sk->sk_state);
3174 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3175 * Probably, we should reset in this case. For now drop them.
3177 __skb_queue_purge(&tp->out_of_order_queue);
3178 if (tp->rx_opt.sack_ok)
3179 tcp_sack_reset(&tp->rx_opt);
3180 sk_stream_mem_reclaim(sk);
3182 if (!sock_flag(sk, SOCK_DEAD)) {
3183 sk->sk_state_change(sk);
3185 /* Do not send POLL_HUP for half duplex close. */
3186 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3187 sk->sk_state == TCP_CLOSE)
3188 sk_wake_async(sk, 1, POLL_HUP);
3190 sk_wake_async(sk, 1, POLL_IN);
3194 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3196 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3197 if (before(seq, sp->start_seq))
3198 sp->start_seq = seq;
3199 if (after(end_seq, sp->end_seq))
3200 sp->end_seq = end_seq;
3206 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3208 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3209 if (before(seq, tp->rcv_nxt))
3210 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3212 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3214 tp->rx_opt.dsack = 1;
3215 tp->duplicate_sack[0].start_seq = seq;
3216 tp->duplicate_sack[0].end_seq = end_seq;
3217 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3221 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3223 if (!tp->rx_opt.dsack)
3224 tcp_dsack_set(tp, seq, end_seq);
3226 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3229 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3231 struct tcp_sock *tp = tcp_sk(sk);
3233 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3234 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3235 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3236 tcp_enter_quickack_mode(sk);
3238 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3239 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3241 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3242 end_seq = tp->rcv_nxt;
3243 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3250 /* These routines update the SACK block as out-of-order packets arrive or
3251 * in-order packets close up the sequence space.
3253 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3256 struct tcp_sack_block *sp = &tp->selective_acks[0];
3257 struct tcp_sack_block *swalk = sp+1;
3259 /* See if the recent change to the first SACK eats into
3260 * or hits the sequence space of other SACK blocks, if so coalesce.
3262 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3263 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3266 /* Zap SWALK, by moving every further SACK up by one slot.
3267 * Decrease num_sacks.
3269 tp->rx_opt.num_sacks--;
3270 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3271 for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3275 this_sack++, swalk++;
3279 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3283 tmp = sack1->start_seq;
3284 sack1->start_seq = sack2->start_seq;
3285 sack2->start_seq = tmp;
3287 tmp = sack1->end_seq;
3288 sack1->end_seq = sack2->end_seq;
3289 sack2->end_seq = tmp;
3292 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3294 struct tcp_sock *tp = tcp_sk(sk);
3295 struct tcp_sack_block *sp = &tp->selective_acks[0];
3296 int cur_sacks = tp->rx_opt.num_sacks;
3302 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3303 if (tcp_sack_extend(sp, seq, end_seq)) {
3304 /* Rotate this_sack to the first one. */
3305 for (; this_sack>0; this_sack--, sp--)
3306 tcp_sack_swap(sp, sp-1);
3308 tcp_sack_maybe_coalesce(tp);
3313 /* Could not find an adjacent existing SACK, build a new one,
3314 * put it at the front, and shift everyone else down. We
3315 * always know there is at least one SACK present already here.
3317 * If the sack array is full, forget about the last one.
3319 if (this_sack >= 4) {
3321 tp->rx_opt.num_sacks--;
3324 for (; this_sack > 0; this_sack--, sp--)
3328 /* Build the new head SACK, and we're done. */
3329 sp->start_seq = seq;
3330 sp->end_seq = end_seq;
3331 tp->rx_opt.num_sacks++;
3332 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3335 /* RCV.NXT advances, some SACKs should be eaten. */
3337 static void tcp_sack_remove(struct tcp_sock *tp)
3339 struct tcp_sack_block *sp = &tp->selective_acks[0];
3340 int num_sacks = tp->rx_opt.num_sacks;
3343 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3344 if (skb_queue_empty(&tp->out_of_order_queue)) {
3345 tp->rx_opt.num_sacks = 0;
3346 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3350 for (this_sack = 0; this_sack < num_sacks; ) {
3351 /* Check if the start of the sack is covered by RCV.NXT. */
3352 if (!before(tp->rcv_nxt, sp->start_seq)) {
3355 /* RCV.NXT must cover all the block! */
3356 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3358 /* Zap this SACK, by moving forward any other SACKS. */
3359 for (i=this_sack+1; i < num_sacks; i++)
3360 tp->selective_acks[i-1] = tp->selective_acks[i];
3367 if (num_sacks != tp->rx_opt.num_sacks) {
3368 tp->rx_opt.num_sacks = num_sacks;
3369 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3373 /* This one checks to see if we can put data from the
3374 * out_of_order queue into the receive_queue.
3376 static void tcp_ofo_queue(struct sock *sk)
3378 struct tcp_sock *tp = tcp_sk(sk);
3379 __u32 dsack_high = tp->rcv_nxt;
3380 struct sk_buff *skb;
3382 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3383 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3386 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3387 __u32 dsack = dsack_high;
3388 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3389 dsack_high = TCP_SKB_CB(skb)->end_seq;
3390 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3393 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3394 SOCK_DEBUG(sk, "ofo packet was already received \n");
3395 __skb_unlink(skb, &tp->out_of_order_queue);
3399 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3400 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3401 TCP_SKB_CB(skb)->end_seq);
3403 __skb_unlink(skb, &tp->out_of_order_queue);
3404 __skb_queue_tail(&sk->sk_receive_queue, skb);
3405 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3406 if (tcp_hdr(skb)->fin)
3407 tcp_fin(skb, sk, tcp_hdr(skb));
3411 static int tcp_prune_queue(struct sock *sk);
3413 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3415 struct tcphdr *th = tcp_hdr(skb);
3416 struct tcp_sock *tp = tcp_sk(sk);
3419 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3422 __skb_pull(skb, th->doff*4);
3424 TCP_ECN_accept_cwr(tp, skb);
3426 if (tp->rx_opt.dsack) {
3427 tp->rx_opt.dsack = 0;
3428 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3429 4 - tp->rx_opt.tstamp_ok);
3432 /* Queue data for delivery to the user.
3433 * Packets in sequence go to the receive queue.
3434 * Out of sequence packets to the out_of_order_queue.
3436 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3437 if (tcp_receive_window(tp) == 0)
3440 /* Ok. In sequence. In window. */
3441 if (tp->ucopy.task == current &&
3442 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3443 sock_owned_by_user(sk) && !tp->urg_data) {
3444 int chunk = min_t(unsigned int, skb->len,
3447 __set_current_state(TASK_RUNNING);
3450 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3451 tp->ucopy.len -= chunk;
3452 tp->copied_seq += chunk;
3453 eaten = (chunk == skb->len && !th->fin);
3454 tcp_rcv_space_adjust(sk);
3462 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3463 !sk_stream_rmem_schedule(sk, skb))) {
3464 if (tcp_prune_queue(sk) < 0 ||
3465 !sk_stream_rmem_schedule(sk, skb))
3468 sk_stream_set_owner_r(skb, sk);
3469 __skb_queue_tail(&sk->sk_receive_queue, skb);
3471 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3473 tcp_event_data_recv(sk, skb);
3475 tcp_fin(skb, sk, th);
3477 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3480 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3481 * gap in queue is filled.
3483 if (skb_queue_empty(&tp->out_of_order_queue))
3484 inet_csk(sk)->icsk_ack.pingpong = 0;
3487 if (tp->rx_opt.num_sacks)
3488 tcp_sack_remove(tp);
3490 tcp_fast_path_check(sk);
3494 else if (!sock_flag(sk, SOCK_DEAD))
3495 sk->sk_data_ready(sk, 0);
3499 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3500 /* A retransmit, 2nd most common case. Force an immediate ack. */
3501 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3502 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3505 tcp_enter_quickack_mode(sk);
3506 inet_csk_schedule_ack(sk);
3512 /* Out of window. F.e. zero window probe. */
3513 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3516 tcp_enter_quickack_mode(sk);
3518 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3519 /* Partial packet, seq < rcv_next < end_seq */
3520 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3521 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3522 TCP_SKB_CB(skb)->end_seq);
3524 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3526 /* If window is closed, drop tail of packet. But after
3527 * remembering D-SACK for its head made in previous line.
3529 if (!tcp_receive_window(tp))
3534 TCP_ECN_check_ce(tp, skb);
3536 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3537 !sk_stream_rmem_schedule(sk, skb)) {
3538 if (tcp_prune_queue(sk) < 0 ||
3539 !sk_stream_rmem_schedule(sk, skb))
3543 /* Disable header prediction. */
3545 inet_csk_schedule_ack(sk);
3547 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3548 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3550 sk_stream_set_owner_r(skb, sk);
3552 if (!skb_peek(&tp->out_of_order_queue)) {
3553 /* Initial out of order segment, build 1 SACK. */
3554 if (tp->rx_opt.sack_ok) {
3555 tp->rx_opt.num_sacks = 1;
3556 tp->rx_opt.dsack = 0;
3557 tp->rx_opt.eff_sacks = 1;
3558 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3559 tp->selective_acks[0].end_seq =
3560 TCP_SKB_CB(skb)->end_seq;
3562 __skb_queue_head(&tp->out_of_order_queue,skb);
3564 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3565 u32 seq = TCP_SKB_CB(skb)->seq;
3566 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3568 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3569 __skb_append(skb1, skb, &tp->out_of_order_queue);
3571 if (!tp->rx_opt.num_sacks ||
3572 tp->selective_acks[0].end_seq != seq)
3575 /* Common case: data arrive in order after hole. */
3576 tp->selective_acks[0].end_seq = end_seq;
3580 /* Find place to insert this segment. */
3582 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3584 } while ((skb1 = skb1->prev) !=
3585 (struct sk_buff*)&tp->out_of_order_queue);
3587 /* Do skb overlap to previous one? */
3588 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3589 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3590 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3591 /* All the bits are present. Drop. */
3593 tcp_dsack_set(tp, seq, end_seq);
3596 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3597 /* Partial overlap. */
3598 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3603 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3605 /* And clean segments covered by new one as whole. */
3606 while ((skb1 = skb->next) !=
3607 (struct sk_buff*)&tp->out_of_order_queue &&
3608 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3609 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3610 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3613 __skb_unlink(skb1, &tp->out_of_order_queue);
3614 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3619 if (tp->rx_opt.sack_ok)
3620 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3624 /* Collapse contiguous sequence of skbs head..tail with
3625 * sequence numbers start..end.
3626 * Segments with FIN/SYN are not collapsed (only because this
3630 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3631 struct sk_buff *head, struct sk_buff *tail,
3634 struct sk_buff *skb;
3636 /* First, check that queue is collapsible and find
3637 * the point where collapsing can be useful. */
3638 for (skb = head; skb != tail; ) {
3639 /* No new bits? It is possible on ofo queue. */
3640 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3641 struct sk_buff *next = skb->next;
3642 __skb_unlink(skb, list);
3644 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3649 /* The first skb to collapse is:
3651 * - bloated or contains data before "start" or
3652 * overlaps to the next one.
3654 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3655 (tcp_win_from_space(skb->truesize) > skb->len ||
3656 before(TCP_SKB_CB(skb)->seq, start) ||
3657 (skb->next != tail &&
3658 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3661 /* Decided to skip this, advance start seq. */
3662 start = TCP_SKB_CB(skb)->end_seq;
3665 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3668 while (before(start, end)) {
3669 struct sk_buff *nskb;
3670 int header = skb_headroom(skb);
3671 int copy = SKB_MAX_ORDER(header, 0);
3673 /* Too big header? This can happen with IPv6. */
3676 if (end-start < copy)
3678 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3682 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3683 skb_set_network_header(nskb, (skb_network_header(skb) -
3685 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3687 skb_reserve(nskb, header);
3688 memcpy(nskb->head, skb->head, header);
3689 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3690 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3691 __skb_insert(nskb, skb->prev, skb, list);
3692 sk_stream_set_owner_r(nskb, sk);
3694 /* Copy data, releasing collapsed skbs. */
3696 int offset = start - TCP_SKB_CB(skb)->seq;
3697 int size = TCP_SKB_CB(skb)->end_seq - start;
3701 size = min(copy, size);
3702 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3704 TCP_SKB_CB(nskb)->end_seq += size;
3708 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3709 struct sk_buff *next = skb->next;
3710 __skb_unlink(skb, list);
3712 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3715 tcp_hdr(skb)->syn ||
3723 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3724 * and tcp_collapse() them until all the queue is collapsed.
3726 static void tcp_collapse_ofo_queue(struct sock *sk)
3728 struct tcp_sock *tp = tcp_sk(sk);
3729 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3730 struct sk_buff *head;
3736 start = TCP_SKB_CB(skb)->seq;
3737 end = TCP_SKB_CB(skb)->end_seq;
3743 /* Segment is terminated when we see gap or when
3744 * we are at the end of all the queue. */
3745 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3746 after(TCP_SKB_CB(skb)->seq, end) ||
3747 before(TCP_SKB_CB(skb)->end_seq, start)) {
3748 tcp_collapse(sk, &tp->out_of_order_queue,
3749 head, skb, start, end);
3751 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3753 /* Start new segment */
3754 start = TCP_SKB_CB(skb)->seq;
3755 end = TCP_SKB_CB(skb)->end_seq;
3757 if (before(TCP_SKB_CB(skb)->seq, start))
3758 start = TCP_SKB_CB(skb)->seq;
3759 if (after(TCP_SKB_CB(skb)->end_seq, end))
3760 end = TCP_SKB_CB(skb)->end_seq;
3765 /* Reduce allocated memory if we can, trying to get
3766 * the socket within its memory limits again.
3768 * Return less than zero if we should start dropping frames
3769 * until the socket owning process reads some of the data
3770 * to stabilize the situation.
3772 static int tcp_prune_queue(struct sock *sk)
3774 struct tcp_sock *tp = tcp_sk(sk);
3776 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3778 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3780 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3781 tcp_clamp_window(sk);
3782 else if (tcp_memory_pressure)
3783 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3785 tcp_collapse_ofo_queue(sk);
3786 tcp_collapse(sk, &sk->sk_receive_queue,
3787 sk->sk_receive_queue.next,
3788 (struct sk_buff*)&sk->sk_receive_queue,
3789 tp->copied_seq, tp->rcv_nxt);
3790 sk_stream_mem_reclaim(sk);
3792 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3795 /* Collapsing did not help, destructive actions follow.
3796 * This must not ever occur. */
3798 /* First, purge the out_of_order queue. */
3799 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3800 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3801 __skb_queue_purge(&tp->out_of_order_queue);
3803 /* Reset SACK state. A conforming SACK implementation will
3804 * do the same at a timeout based retransmit. When a connection
3805 * is in a sad state like this, we care only about integrity
3806 * of the connection not performance.
3808 if (tp->rx_opt.sack_ok)
3809 tcp_sack_reset(&tp->rx_opt);
3810 sk_stream_mem_reclaim(sk);
3813 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3816 /* If we are really being abused, tell the caller to silently
3817 * drop receive data on the floor. It will get retransmitted
3818 * and hopefully then we'll have sufficient space.
3820 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3822 /* Massive buffer overcommit. */
3828 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3829 * As additional protections, we do not touch cwnd in retransmission phases,
3830 * and if application hit its sndbuf limit recently.
3832 void tcp_cwnd_application_limited(struct sock *sk)
3834 struct tcp_sock *tp = tcp_sk(sk);
3836 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3837 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3838 /* Limited by application or receiver window. */
3839 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3840 u32 win_used = max(tp->snd_cwnd_used, init_win);
3841 if (win_used < tp->snd_cwnd) {
3842 tp->snd_ssthresh = tcp_current_ssthresh(sk);
3843 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3845 tp->snd_cwnd_used = 0;
3847 tp->snd_cwnd_stamp = tcp_time_stamp;
3850 static int tcp_should_expand_sndbuf(struct sock *sk)
3852 struct tcp_sock *tp = tcp_sk(sk);
3854 /* If the user specified a specific send buffer setting, do
3857 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3860 /* If we are under global TCP memory pressure, do not expand. */
3861 if (tcp_memory_pressure)
3864 /* If we are under soft global TCP memory pressure, do not expand. */
3865 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3868 /* If we filled the congestion window, do not expand. */
3869 if (tp->packets_out >= tp->snd_cwnd)
3875 /* When incoming ACK allowed to free some skb from write_queue,
3876 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3877 * on the exit from tcp input handler.
3879 * PROBLEM: sndbuf expansion does not work well with largesend.
3881 static void tcp_new_space(struct sock *sk)
3883 struct tcp_sock *tp = tcp_sk(sk);
3885 if (tcp_should_expand_sndbuf(sk)) {
3886 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3887 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3888 demanded = max_t(unsigned int, tp->snd_cwnd,
3889 tp->reordering + 1);
3890 sndmem *= 2*demanded;
3891 if (sndmem > sk->sk_sndbuf)
3892 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3893 tp->snd_cwnd_stamp = tcp_time_stamp;
3896 sk->sk_write_space(sk);
3899 static void tcp_check_space(struct sock *sk)
3901 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3902 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3903 if (sk->sk_socket &&
3904 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3909 static inline void tcp_data_snd_check(struct sock *sk)
3911 tcp_push_pending_frames(sk);
3912 tcp_check_space(sk);
3916 * Check if sending an ack is needed.
3918 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3920 struct tcp_sock *tp = tcp_sk(sk);
3922 /* More than one full frame received... */
3923 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3924 /* ... and right edge of window advances far enough.
3925 * (tcp_recvmsg() will send ACK otherwise). Or...
3927 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3928 /* We ACK each frame or... */
3929 tcp_in_quickack_mode(sk) ||
3930 /* We have out of order data. */
3932 skb_peek(&tp->out_of_order_queue))) {
3933 /* Then ack it now */
3936 /* Else, send delayed ack. */
3937 tcp_send_delayed_ack(sk);
3941 static inline void tcp_ack_snd_check(struct sock *sk)
3943 if (!inet_csk_ack_scheduled(sk)) {
3944 /* We sent a data segment already. */
3947 __tcp_ack_snd_check(sk, 1);
3951 * This routine is only called when we have urgent data
3952 * signaled. Its the 'slow' part of tcp_urg. It could be
3953 * moved inline now as tcp_urg is only called from one
3954 * place. We handle URGent data wrong. We have to - as
3955 * BSD still doesn't use the correction from RFC961.
3956 * For 1003.1g we should support a new option TCP_STDURG to permit
3957 * either form (or just set the sysctl tcp_stdurg).
3960 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3962 struct tcp_sock *tp = tcp_sk(sk);
3963 u32 ptr = ntohs(th->urg_ptr);
3965 if (ptr && !sysctl_tcp_stdurg)
3967 ptr += ntohl(th->seq);
3969 /* Ignore urgent data that we've already seen and read. */
3970 if (after(tp->copied_seq, ptr))
3973 /* Do not replay urg ptr.
3975 * NOTE: interesting situation not covered by specs.
3976 * Misbehaving sender may send urg ptr, pointing to segment,
3977 * which we already have in ofo queue. We are not able to fetch
3978 * such data and will stay in TCP_URG_NOTYET until will be eaten
3979 * by recvmsg(). Seems, we are not obliged to handle such wicked
3980 * situations. But it is worth to think about possibility of some
3981 * DoSes using some hypothetical application level deadlock.
3983 if (before(ptr, tp->rcv_nxt))
3986 /* Do we already have a newer (or duplicate) urgent pointer? */
3987 if (tp->urg_data && !after(ptr, tp->urg_seq))
3990 /* Tell the world about our new urgent pointer. */
3993 /* We may be adding urgent data when the last byte read was
3994 * urgent. To do this requires some care. We cannot just ignore
3995 * tp->copied_seq since we would read the last urgent byte again
3996 * as data, nor can we alter copied_seq until this data arrives
3997 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3999 * NOTE. Double Dutch. Rendering to plain English: author of comment
4000 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4001 * and expect that both A and B disappear from stream. This is _wrong_.
4002 * Though this happens in BSD with high probability, this is occasional.
4003 * Any application relying on this is buggy. Note also, that fix "works"
4004 * only in this artificial test. Insert some normal data between A and B and we will
4005 * decline of BSD again. Verdict: it is better to remove to trap
4008 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4009 !sock_flag(sk, SOCK_URGINLINE) &&
4010 tp->copied_seq != tp->rcv_nxt) {
4011 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4013 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4014 __skb_unlink(skb, &sk->sk_receive_queue);
4019 tp->urg_data = TCP_URG_NOTYET;
4022 /* Disable header prediction. */
4026 /* This is the 'fast' part of urgent handling. */
4027 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4029 struct tcp_sock *tp = tcp_sk(sk);
4031 /* Check if we get a new urgent pointer - normally not. */
4033 tcp_check_urg(sk,th);
4035 /* Do we wait for any urgent data? - normally not... */
4036 if (tp->urg_data == TCP_URG_NOTYET) {
4037 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4040 /* Is the urgent pointer pointing into this packet? */
4041 if (ptr < skb->len) {
4043 if (skb_copy_bits(skb, ptr, &tmp, 1))
4045 tp->urg_data = TCP_URG_VALID | tmp;
4046 if (!sock_flag(sk, SOCK_DEAD))
4047 sk->sk_data_ready(sk, 0);
4052 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4054 struct tcp_sock *tp = tcp_sk(sk);
4055 int chunk = skb->len - hlen;
4059 if (skb_csum_unnecessary(skb))
4060 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4062 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4066 tp->ucopy.len -= chunk;
4067 tp->copied_seq += chunk;
4068 tcp_rcv_space_adjust(sk);
4075 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4079 if (sock_owned_by_user(sk)) {
4081 result = __tcp_checksum_complete(skb);
4084 result = __tcp_checksum_complete(skb);
4089 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4091 return !skb_csum_unnecessary(skb) &&
4092 __tcp_checksum_complete_user(sk, skb);
4095 #ifdef CONFIG_NET_DMA
4096 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4098 struct tcp_sock *tp = tcp_sk(sk);
4099 int chunk = skb->len - hlen;
4101 int copied_early = 0;
4103 if (tp->ucopy.wakeup)
4106 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4107 tp->ucopy.dma_chan = get_softnet_dma();
4109 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4111 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4112 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4117 tp->ucopy.dma_cookie = dma_cookie;
4120 tp->ucopy.len -= chunk;
4121 tp->copied_seq += chunk;
4122 tcp_rcv_space_adjust(sk);
4124 if ((tp->ucopy.len == 0) ||
4125 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4126 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4127 tp->ucopy.wakeup = 1;
4128 sk->sk_data_ready(sk, 0);
4130 } else if (chunk > 0) {
4131 tp->ucopy.wakeup = 1;
4132 sk->sk_data_ready(sk, 0);
4135 return copied_early;
4137 #endif /* CONFIG_NET_DMA */
4140 * TCP receive function for the ESTABLISHED state.
4142 * It is split into a fast path and a slow path. The fast path is
4144 * - A zero window was announced from us - zero window probing
4145 * is only handled properly in the slow path.
4146 * - Out of order segments arrived.
4147 * - Urgent data is expected.
4148 * - There is no buffer space left
4149 * - Unexpected TCP flags/window values/header lengths are received
4150 * (detected by checking the TCP header against pred_flags)
4151 * - Data is sent in both directions. Fast path only supports pure senders
4152 * or pure receivers (this means either the sequence number or the ack
4153 * value must stay constant)
4154 * - Unexpected TCP option.
4156 * When these conditions are not satisfied it drops into a standard
4157 * receive procedure patterned after RFC793 to handle all cases.
4158 * The first three cases are guaranteed by proper pred_flags setting,
4159 * the rest is checked inline. Fast processing is turned on in
4160 * tcp_data_queue when everything is OK.
4162 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4163 struct tcphdr *th, unsigned len)
4165 struct tcp_sock *tp = tcp_sk(sk);
4168 * Header prediction.
4169 * The code loosely follows the one in the famous
4170 * "30 instruction TCP receive" Van Jacobson mail.
4172 * Van's trick is to deposit buffers into socket queue
4173 * on a device interrupt, to call tcp_recv function
4174 * on the receive process context and checksum and copy
4175 * the buffer to user space. smart...
4177 * Our current scheme is not silly either but we take the
4178 * extra cost of the net_bh soft interrupt processing...
4179 * We do checksum and copy also but from device to kernel.
4182 tp->rx_opt.saw_tstamp = 0;
4184 /* pred_flags is 0xS?10 << 16 + snd_wnd
4185 * if header_prediction is to be made
4186 * 'S' will always be tp->tcp_header_len >> 2
4187 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4188 * turn it off (when there are holes in the receive
4189 * space for instance)
4190 * PSH flag is ignored.
4193 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4194 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4195 int tcp_header_len = tp->tcp_header_len;
4197 /* Timestamp header prediction: tcp_header_len
4198 * is automatically equal to th->doff*4 due to pred_flags
4202 /* Check timestamp */
4203 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4204 __be32 *ptr = (__be32 *)(th + 1);
4206 /* No? Slow path! */
4207 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4208 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4211 tp->rx_opt.saw_tstamp = 1;
4213 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4215 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4217 /* If PAWS failed, check it more carefully in slow path */
4218 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4221 /* DO NOT update ts_recent here, if checksum fails
4222 * and timestamp was corrupted part, it will result
4223 * in a hung connection since we will drop all
4224 * future packets due to the PAWS test.
4228 if (len <= tcp_header_len) {
4229 /* Bulk data transfer: sender */
4230 if (len == tcp_header_len) {
4231 /* Predicted packet is in window by definition.
4232 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4233 * Hence, check seq<=rcv_wup reduces to:
4235 if (tcp_header_len ==
4236 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4237 tp->rcv_nxt == tp->rcv_wup)
4238 tcp_store_ts_recent(tp);
4240 /* We know that such packets are checksummed
4243 tcp_ack(sk, skb, 0);
4245 tcp_data_snd_check(sk);
4247 } else { /* Header too small */
4248 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4253 int copied_early = 0;
4255 if (tp->copied_seq == tp->rcv_nxt &&
4256 len - tcp_header_len <= tp->ucopy.len) {
4257 #ifdef CONFIG_NET_DMA
4258 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4263 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4264 __set_current_state(TASK_RUNNING);
4266 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4270 /* Predicted packet is in window by definition.
4271 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4272 * Hence, check seq<=rcv_wup reduces to:
4274 if (tcp_header_len ==
4275 (sizeof(struct tcphdr) +
4276 TCPOLEN_TSTAMP_ALIGNED) &&
4277 tp->rcv_nxt == tp->rcv_wup)
4278 tcp_store_ts_recent(tp);
4280 tcp_rcv_rtt_measure_ts(sk, skb);
4282 __skb_pull(skb, tcp_header_len);
4283 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4284 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4287 tcp_cleanup_rbuf(sk, skb->len);
4290 if (tcp_checksum_complete_user(sk, skb))
4293 /* Predicted packet is in window by definition.
4294 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4295 * Hence, check seq<=rcv_wup reduces to:
4297 if (tcp_header_len ==
4298 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4299 tp->rcv_nxt == tp->rcv_wup)
4300 tcp_store_ts_recent(tp);
4302 tcp_rcv_rtt_measure_ts(sk, skb);
4304 if ((int)skb->truesize > sk->sk_forward_alloc)
4307 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4309 /* Bulk data transfer: receiver */
4310 __skb_pull(skb,tcp_header_len);
4311 __skb_queue_tail(&sk->sk_receive_queue, skb);
4312 sk_stream_set_owner_r(skb, sk);
4313 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4316 tcp_event_data_recv(sk, skb);
4318 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4319 /* Well, only one small jumplet in fast path... */
4320 tcp_ack(sk, skb, FLAG_DATA);
4321 tcp_data_snd_check(sk);
4322 if (!inet_csk_ack_scheduled(sk))
4326 __tcp_ack_snd_check(sk, 0);
4328 #ifdef CONFIG_NET_DMA
4330 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4336 sk->sk_data_ready(sk, 0);
4342 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4346 * RFC1323: H1. Apply PAWS check first.
4348 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4349 tcp_paws_discard(sk, skb)) {
4351 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4352 tcp_send_dupack(sk, skb);
4355 /* Resets are accepted even if PAWS failed.
4357 ts_recent update must be made after we are sure
4358 that the packet is in window.
4363 * Standard slow path.
4366 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4367 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4368 * (RST) segments are validated by checking their SEQ-fields."
4369 * And page 69: "If an incoming segment is not acceptable,
4370 * an acknowledgment should be sent in reply (unless the RST bit
4371 * is set, if so drop the segment and return)".
4374 tcp_send_dupack(sk, skb);
4383 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4385 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4386 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4387 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4394 tcp_ack(sk, skb, FLAG_SLOWPATH);
4396 tcp_rcv_rtt_measure_ts(sk, skb);
4398 /* Process urgent data. */
4399 tcp_urg(sk, skb, th);
4401 /* step 7: process the segment text */
4402 tcp_data_queue(sk, skb);
4404 tcp_data_snd_check(sk);
4405 tcp_ack_snd_check(sk);
4409 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4416 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4417 struct tcphdr *th, unsigned len)
4419 struct tcp_sock *tp = tcp_sk(sk);
4420 struct inet_connection_sock *icsk = inet_csk(sk);
4421 int saved_clamp = tp->rx_opt.mss_clamp;
4423 tcp_parse_options(skb, &tp->rx_opt, 0);
4427 * "If the state is SYN-SENT then
4428 * first check the ACK bit
4429 * If the ACK bit is set
4430 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4431 * a reset (unless the RST bit is set, if so drop
4432 * the segment and return)"
4434 * We do not send data with SYN, so that RFC-correct
4437 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4438 goto reset_and_undo;
4440 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4441 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4443 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4444 goto reset_and_undo;
4447 /* Now ACK is acceptable.
4449 * "If the RST bit is set
4450 * If the ACK was acceptable then signal the user "error:
4451 * connection reset", drop the segment, enter CLOSED state,
4452 * delete TCB, and return."
4461 * "fifth, if neither of the SYN or RST bits is set then
4462 * drop the segment and return."
4468 goto discard_and_undo;
4471 * "If the SYN bit is on ...
4472 * are acceptable then ...
4473 * (our SYN has been ACKed), change the connection
4474 * state to ESTABLISHED..."
4477 TCP_ECN_rcv_synack(tp, th);
4479 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4480 tcp_ack(sk, skb, FLAG_SLOWPATH);
4482 /* Ok.. it's good. Set up sequence numbers and
4483 * move to established.
4485 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4486 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4488 /* RFC1323: The window in SYN & SYN/ACK segments is
4491 tp->snd_wnd = ntohs(th->window);
4492 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4494 if (!tp->rx_opt.wscale_ok) {
4495 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4496 tp->window_clamp = min(tp->window_clamp, 65535U);
4499 if (tp->rx_opt.saw_tstamp) {
4500 tp->rx_opt.tstamp_ok = 1;
4501 tp->tcp_header_len =
4502 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4503 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4504 tcp_store_ts_recent(tp);
4506 tp->tcp_header_len = sizeof(struct tcphdr);
4509 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4510 tp->rx_opt.sack_ok |= 2;
4513 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4514 tcp_initialize_rcv_mss(sk);
4516 /* Remember, tcp_poll() does not lock socket!
4517 * Change state from SYN-SENT only after copied_seq
4518 * is initialized. */
4519 tp->copied_seq = tp->rcv_nxt;
4521 tcp_set_state(sk, TCP_ESTABLISHED);
4523 security_inet_conn_established(sk, skb);
4525 /* Make sure socket is routed, for correct metrics. */
4526 icsk->icsk_af_ops->rebuild_header(sk);
4528 tcp_init_metrics(sk);
4530 tcp_init_congestion_control(sk);
4532 /* Prevent spurious tcp_cwnd_restart() on first data
4535 tp->lsndtime = tcp_time_stamp;
4537 tcp_init_buffer_space(sk);
4539 if (sock_flag(sk, SOCK_KEEPOPEN))
4540 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4542 if (!tp->rx_opt.snd_wscale)
4543 __tcp_fast_path_on(tp, tp->snd_wnd);
4547 if (!sock_flag(sk, SOCK_DEAD)) {
4548 sk->sk_state_change(sk);
4549 sk_wake_async(sk, 0, POLL_OUT);
4552 if (sk->sk_write_pending ||
4553 icsk->icsk_accept_queue.rskq_defer_accept ||
4554 icsk->icsk_ack.pingpong) {
4555 /* Save one ACK. Data will be ready after
4556 * several ticks, if write_pending is set.
4558 * It may be deleted, but with this feature tcpdumps
4559 * look so _wonderfully_ clever, that I was not able
4560 * to stand against the temptation 8) --ANK
4562 inet_csk_schedule_ack(sk);
4563 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4564 icsk->icsk_ack.ato = TCP_ATO_MIN;
4565 tcp_incr_quickack(sk);
4566 tcp_enter_quickack_mode(sk);
4567 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4568 TCP_DELACK_MAX, TCP_RTO_MAX);
4579 /* No ACK in the segment */
4583 * "If the RST bit is set
4585 * Otherwise (no ACK) drop the segment and return."
4588 goto discard_and_undo;
4592 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4593 goto discard_and_undo;
4596 /* We see SYN without ACK. It is attempt of
4597 * simultaneous connect with crossed SYNs.
4598 * Particularly, it can be connect to self.
4600 tcp_set_state(sk, TCP_SYN_RECV);
4602 if (tp->rx_opt.saw_tstamp) {
4603 tp->rx_opt.tstamp_ok = 1;
4604 tcp_store_ts_recent(tp);
4605 tp->tcp_header_len =
4606 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4608 tp->tcp_header_len = sizeof(struct tcphdr);
4611 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4612 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4614 /* RFC1323: The window in SYN & SYN/ACK segments is
4617 tp->snd_wnd = ntohs(th->window);
4618 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4619 tp->max_window = tp->snd_wnd;
4621 TCP_ECN_rcv_syn(tp, th);
4624 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4625 tcp_initialize_rcv_mss(sk);
4628 tcp_send_synack(sk);
4630 /* Note, we could accept data and URG from this segment.
4631 * There are no obstacles to make this.
4633 * However, if we ignore data in ACKless segments sometimes,
4634 * we have no reasons to accept it sometimes.
4635 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4636 * is not flawless. So, discard packet for sanity.
4637 * Uncomment this return to process the data.
4644 /* "fifth, if neither of the SYN or RST bits is set then
4645 * drop the segment and return."
4649 tcp_clear_options(&tp->rx_opt);
4650 tp->rx_opt.mss_clamp = saved_clamp;
4654 tcp_clear_options(&tp->rx_opt);
4655 tp->rx_opt.mss_clamp = saved_clamp;
4661 * This function implements the receiving procedure of RFC 793 for
4662 * all states except ESTABLISHED and TIME_WAIT.
4663 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4664 * address independent.
4667 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4668 struct tcphdr *th, unsigned len)
4670 struct tcp_sock *tp = tcp_sk(sk);
4671 struct inet_connection_sock *icsk = inet_csk(sk);
4674 tp->rx_opt.saw_tstamp = 0;
4676 switch (sk->sk_state) {
4688 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4691 /* Now we have several options: In theory there is
4692 * nothing else in the frame. KA9Q has an option to
4693 * send data with the syn, BSD accepts data with the
4694 * syn up to the [to be] advertised window and
4695 * Solaris 2.1 gives you a protocol error. For now
4696 * we just ignore it, that fits the spec precisely
4697 * and avoids incompatibilities. It would be nice in
4698 * future to drop through and process the data.
4700 * Now that TTCP is starting to be used we ought to
4702 * But, this leaves one open to an easy denial of
4703 * service attack, and SYN cookies can't defend
4704 * against this problem. So, we drop the data
4705 * in the interest of security over speed unless
4706 * it's still in use.
4714 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4718 /* Do step6 onward by hand. */
4719 tcp_urg(sk, skb, th);
4721 tcp_data_snd_check(sk);
4725 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4726 tcp_paws_discard(sk, skb)) {
4728 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4729 tcp_send_dupack(sk, skb);
4732 /* Reset is accepted even if it did not pass PAWS. */
4735 /* step 1: check sequence number */
4736 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4738 tcp_send_dupack(sk, skb);
4742 /* step 2: check RST bit */
4748 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4750 /* step 3: check security and precedence [ignored] */
4754 * Check for a SYN in window.
4756 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4757 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4762 /* step 5: check the ACK field */
4764 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4766 switch (sk->sk_state) {
4769 tp->copied_seq = tp->rcv_nxt;
4771 tcp_set_state(sk, TCP_ESTABLISHED);
4772 sk->sk_state_change(sk);
4774 /* Note, that this wakeup is only for marginal
4775 * crossed SYN case. Passively open sockets
4776 * are not waked up, because sk->sk_sleep ==
4777 * NULL and sk->sk_socket == NULL.
4779 if (sk->sk_socket) {
4780 sk_wake_async(sk,0,POLL_OUT);
4783 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4784 tp->snd_wnd = ntohs(th->window) <<
4785 tp->rx_opt.snd_wscale;
4786 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4787 TCP_SKB_CB(skb)->seq);
4789 /* tcp_ack considers this ACK as duplicate
4790 * and does not calculate rtt.
4791 * Fix it at least with timestamps.
4793 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4795 tcp_ack_saw_tstamp(sk, 0);
4797 if (tp->rx_opt.tstamp_ok)
4798 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4800 /* Make sure socket is routed, for
4803 icsk->icsk_af_ops->rebuild_header(sk);
4805 tcp_init_metrics(sk);
4807 tcp_init_congestion_control(sk);
4809 /* Prevent spurious tcp_cwnd_restart() on
4810 * first data packet.
4812 tp->lsndtime = tcp_time_stamp;
4815 tcp_initialize_rcv_mss(sk);
4816 tcp_init_buffer_space(sk);
4817 tcp_fast_path_on(tp);
4824 if (tp->snd_una == tp->write_seq) {
4825 tcp_set_state(sk, TCP_FIN_WAIT2);
4826 sk->sk_shutdown |= SEND_SHUTDOWN;
4827 dst_confirm(sk->sk_dst_cache);
4829 if (!sock_flag(sk, SOCK_DEAD))
4830 /* Wake up lingering close() */
4831 sk->sk_state_change(sk);
4835 if (tp->linger2 < 0 ||
4836 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4837 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4839 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4843 tmo = tcp_fin_time(sk);
4844 if (tmo > TCP_TIMEWAIT_LEN) {
4845 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4846 } else if (th->fin || sock_owned_by_user(sk)) {
4847 /* Bad case. We could lose such FIN otherwise.
4848 * It is not a big problem, but it looks confusing
4849 * and not so rare event. We still can lose it now,
4850 * if it spins in bh_lock_sock(), but it is really
4853 inet_csk_reset_keepalive_timer(sk, tmo);
4855 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4863 if (tp->snd_una == tp->write_seq) {
4864 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4870 if (tp->snd_una == tp->write_seq) {
4871 tcp_update_metrics(sk);
4880 /* step 6: check the URG bit */
4881 tcp_urg(sk, skb, th);
4883 /* step 7: process the segment text */
4884 switch (sk->sk_state) {
4885 case TCP_CLOSE_WAIT:
4888 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4892 /* RFC 793 says to queue data in these states,
4893 * RFC 1122 says we MUST send a reset.
4894 * BSD 4.4 also does reset.
4896 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4897 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4898 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4899 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4905 case TCP_ESTABLISHED:
4906 tcp_data_queue(sk, skb);
4911 /* tcp_data could move socket to TIME-WAIT */
4912 if (sk->sk_state != TCP_CLOSE) {
4913 tcp_data_snd_check(sk);
4914 tcp_ack_snd_check(sk);
4924 EXPORT_SYMBOL(sysctl_tcp_ecn);
4925 EXPORT_SYMBOL(sysctl_tcp_reordering);
4926 EXPORT_SYMBOL(tcp_parse_options);
4927 EXPORT_SYMBOL(tcp_rcv_established);
4928 EXPORT_SYMBOL(tcp_rcv_state_process);
4929 EXPORT_SYMBOL(tcp_initialize_rcv_mss);