2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presence of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
101 #define FLAG_ECE 0x40 /* ECE in this ACK */
102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained DSACK info */
108 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
109 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
110 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
111 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
112 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
114 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
116 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
117 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
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 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
200 if (tp->ecn_flags&TCP_ECN_OK)
201 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
204 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
206 if (tcp_hdr(skb)->cwr)
207 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
210 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
217 if (tp->ecn_flags&TCP_ECN_OK) {
218 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
219 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
220 /* Funny extension: if ECT is not set on a segment,
221 * it is surely retransmit. It is not in ECN RFC,
222 * but Linux follows this rule. */
223 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
224 tcp_enter_quickack_mode((struct sock *)tp);
228 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
230 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
231 tp->ecn_flags &= ~TCP_ECN_OK;
234 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
236 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
237 tp->ecn_flags &= ~TCP_ECN_OK;
240 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
242 if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
247 /* Buffer size and advertised window tuning.
249 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
252 static void tcp_fixup_sndbuf(struct sock *sk)
254 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
255 sizeof(struct sk_buff);
257 if (sk->sk_sndbuf < 3 * sndmem)
258 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
261 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
263 * All tcp_full_space() is split to two parts: "network" buffer, allocated
264 * forward and advertised in receiver window (tp->rcv_wnd) and
265 * "application buffer", required to isolate scheduling/application
266 * latencies from network.
267 * window_clamp is maximal advertised window. It can be less than
268 * tcp_full_space(), in this case tcp_full_space() - window_clamp
269 * is reserved for "application" buffer. The less window_clamp is
270 * the smoother our behaviour from viewpoint of network, but the lower
271 * throughput and the higher sensitivity of the connection to losses. 8)
273 * rcv_ssthresh is more strict window_clamp used at "slow start"
274 * phase to predict further behaviour of this connection.
275 * It is used for two goals:
276 * - to enforce header prediction at sender, even when application
277 * requires some significant "application buffer". It is check #1.
278 * - to prevent pruning of receive queue because of misprediction
279 * of receiver window. Check #2.
281 * The scheme does not work when sender sends good segments opening
282 * window and then starts to feed us spaghetti. But it should work
283 * in common situations. Otherwise, we have to rely on queue collapsing.
286 /* Slow part of check#2. */
287 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
289 struct tcp_sock *tp = tcp_sk(sk);
291 int truesize = tcp_win_from_space(skb->truesize)/2;
292 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
294 while (tp->rcv_ssthresh <= window) {
295 if (truesize <= skb->len)
296 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
304 static void tcp_grow_window(struct sock *sk,
307 struct tcp_sock *tp = tcp_sk(sk);
310 if (tp->rcv_ssthresh < tp->window_clamp &&
311 (int)tp->rcv_ssthresh < tcp_space(sk) &&
312 !tcp_memory_pressure) {
315 /* Check #2. Increase window, if skb with such overhead
316 * will fit to rcvbuf in future.
318 if (tcp_win_from_space(skb->truesize) <= skb->len)
321 incr = __tcp_grow_window(sk, skb);
324 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
325 inet_csk(sk)->icsk_ack.quick |= 1;
330 /* 3. Tuning rcvbuf, when connection enters established state. */
332 static void tcp_fixup_rcvbuf(struct sock *sk)
334 struct tcp_sock *tp = tcp_sk(sk);
335 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
337 /* Try to select rcvbuf so that 4 mss-sized segments
338 * will fit to window and corresponding skbs will fit to our rcvbuf.
339 * (was 3; 4 is minimum to allow fast retransmit to work.)
341 while (tcp_win_from_space(rcvmem) < tp->advmss)
343 if (sk->sk_rcvbuf < 4 * rcvmem)
344 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
347 /* 4. Try to fixup all. It is made immediately after connection enters
350 static void tcp_init_buffer_space(struct sock *sk)
352 struct tcp_sock *tp = tcp_sk(sk);
355 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
356 tcp_fixup_rcvbuf(sk);
357 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
358 tcp_fixup_sndbuf(sk);
360 tp->rcvq_space.space = tp->rcv_wnd;
362 maxwin = tcp_full_space(sk);
364 if (tp->window_clamp >= maxwin) {
365 tp->window_clamp = maxwin;
367 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
368 tp->window_clamp = max(maxwin -
369 (maxwin >> sysctl_tcp_app_win),
373 /* Force reservation of one segment. */
374 if (sysctl_tcp_app_win &&
375 tp->window_clamp > 2 * tp->advmss &&
376 tp->window_clamp + tp->advmss > maxwin)
377 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
379 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
380 tp->snd_cwnd_stamp = tcp_time_stamp;
383 /* 5. Recalculate window clamp after socket hit its memory bounds. */
384 static void tcp_clamp_window(struct sock *sk)
386 struct tcp_sock *tp = tcp_sk(sk);
387 struct inet_connection_sock *icsk = inet_csk(sk);
389 icsk->icsk_ack.quick = 0;
391 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
392 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
393 !tcp_memory_pressure &&
394 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
395 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
398 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
399 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
403 /* Initialize RCV_MSS value.
404 * RCV_MSS is an our guess about MSS used by the peer.
405 * We haven't any direct information about the MSS.
406 * It's better to underestimate the RCV_MSS rather than overestimate.
407 * Overestimations make us ACKing less frequently than needed.
408 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
410 void tcp_initialize_rcv_mss(struct sock *sk)
412 struct tcp_sock *tp = tcp_sk(sk);
413 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
415 hint = min(hint, tp->rcv_wnd/2);
416 hint = min(hint, TCP_MIN_RCVMSS);
417 hint = max(hint, TCP_MIN_MSS);
419 inet_csk(sk)->icsk_ack.rcv_mss = hint;
422 /* Receiver "autotuning" code.
424 * The algorithm for RTT estimation w/o timestamps is based on
425 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
426 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
428 * More detail on this code can be found at
429 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
430 * though this reference is out of date. A new paper
433 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
435 u32 new_sample = tp->rcv_rtt_est.rtt;
441 if (new_sample != 0) {
442 /* If we sample in larger samples in the non-timestamp
443 * case, we could grossly overestimate the RTT especially
444 * with chatty applications or bulk transfer apps which
445 * are stalled on filesystem I/O.
447 * Also, since we are only going for a minimum in the
448 * non-timestamp case, we do not smooth things out
449 * else with timestamps disabled convergence takes too
453 m -= (new_sample >> 3);
455 } else if (m < new_sample)
458 /* No previous measure. */
462 if (tp->rcv_rtt_est.rtt != new_sample)
463 tp->rcv_rtt_est.rtt = new_sample;
466 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
468 if (tp->rcv_rtt_est.time == 0)
470 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
472 tcp_rcv_rtt_update(tp,
473 jiffies - tp->rcv_rtt_est.time,
477 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
478 tp->rcv_rtt_est.time = tcp_time_stamp;
481 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
483 struct tcp_sock *tp = tcp_sk(sk);
484 if (tp->rx_opt.rcv_tsecr &&
485 (TCP_SKB_CB(skb)->end_seq -
486 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
487 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
491 * This function should be called every time data is copied to user space.
492 * It calculates the appropriate TCP receive buffer space.
494 void tcp_rcv_space_adjust(struct sock *sk)
496 struct tcp_sock *tp = tcp_sk(sk);
500 if (tp->rcvq_space.time == 0)
503 time = tcp_time_stamp - tp->rcvq_space.time;
504 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
505 tp->rcv_rtt_est.rtt == 0)
508 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
510 space = max(tp->rcvq_space.space, space);
512 if (tp->rcvq_space.space != space) {
515 tp->rcvq_space.space = space;
517 if (sysctl_tcp_moderate_rcvbuf &&
518 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
519 int new_clamp = space;
521 /* Receive space grows, normalize in order to
522 * take into account packet headers and sk_buff
523 * structure overhead.
528 rcvmem = (tp->advmss + MAX_TCP_HEADER +
529 16 + sizeof(struct sk_buff));
530 while (tcp_win_from_space(rcvmem) < tp->advmss)
533 space = min(space, sysctl_tcp_rmem[2]);
534 if (space > sk->sk_rcvbuf) {
535 sk->sk_rcvbuf = space;
537 /* Make the window clamp follow along. */
538 tp->window_clamp = new_clamp;
544 tp->rcvq_space.seq = tp->copied_seq;
545 tp->rcvq_space.time = tcp_time_stamp;
548 /* There is something which you must keep in mind when you analyze the
549 * behavior of the tp->ato delayed ack timeout interval. When a
550 * connection starts up, we want to ack as quickly as possible. The
551 * problem is that "good" TCP's do slow start at the beginning of data
552 * transmission. The means that until we send the first few ACK's the
553 * sender will sit on his end and only queue most of his data, because
554 * he can only send snd_cwnd unacked packets at any given time. For
555 * each ACK we send, he increments snd_cwnd and transmits more of his
558 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
560 struct tcp_sock *tp = tcp_sk(sk);
561 struct inet_connection_sock *icsk = inet_csk(sk);
564 inet_csk_schedule_ack(sk);
566 tcp_measure_rcv_mss(sk, skb);
568 tcp_rcv_rtt_measure(tp);
570 now = tcp_time_stamp;
572 if (!icsk->icsk_ack.ato) {
573 /* The _first_ data packet received, initialize
574 * delayed ACK engine.
576 tcp_incr_quickack(sk);
577 icsk->icsk_ack.ato = TCP_ATO_MIN;
579 int m = now - icsk->icsk_ack.lrcvtime;
581 if (m <= TCP_ATO_MIN/2) {
582 /* The fastest case is the first. */
583 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
584 } else if (m < icsk->icsk_ack.ato) {
585 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
586 if (icsk->icsk_ack.ato > icsk->icsk_rto)
587 icsk->icsk_ack.ato = icsk->icsk_rto;
588 } else if (m > icsk->icsk_rto) {
589 /* Too long gap. Apparently sender failed to
590 * restart window, so that we send ACKs quickly.
592 tcp_incr_quickack(sk);
593 sk_stream_mem_reclaim(sk);
596 icsk->icsk_ack.lrcvtime = now;
598 TCP_ECN_check_ce(tp, skb);
601 tcp_grow_window(sk, skb);
604 static u32 tcp_rto_min(struct sock *sk)
606 struct dst_entry *dst = __sk_dst_get(sk);
607 u32 rto_min = TCP_RTO_MIN;
609 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
610 rto_min = dst->metrics[RTAX_RTO_MIN-1];
614 /* Called to compute a smoothed rtt estimate. The data fed to this
615 * routine either comes from timestamps, or from segments that were
616 * known _not_ to have been retransmitted [see Karn/Partridge
617 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
618 * piece by Van Jacobson.
619 * NOTE: the next three routines used to be one big routine.
620 * To save cycles in the RFC 1323 implementation it was better to break
621 * it up into three procedures. -- erics
623 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
625 struct tcp_sock *tp = tcp_sk(sk);
626 long m = mrtt; /* RTT */
628 /* The following amusing code comes from Jacobson's
629 * article in SIGCOMM '88. Note that rtt and mdev
630 * are scaled versions of rtt and mean deviation.
631 * This is designed to be as fast as possible
632 * m stands for "measurement".
634 * On a 1990 paper the rto value is changed to:
635 * RTO = rtt + 4 * mdev
637 * Funny. This algorithm seems to be very broken.
638 * These formulae increase RTO, when it should be decreased, increase
639 * too slowly, when it should be increased quickly, decrease too quickly
640 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
641 * does not matter how to _calculate_ it. Seems, it was trap
642 * that VJ failed to avoid. 8)
647 m -= (tp->srtt >> 3); /* m is now error in rtt est */
648 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
650 m = -m; /* m is now abs(error) */
651 m -= (tp->mdev >> 2); /* similar update on mdev */
652 /* This is similar to one of Eifel findings.
653 * Eifel blocks mdev updates when rtt decreases.
654 * This solution is a bit different: we use finer gain
655 * for mdev in this case (alpha*beta).
656 * Like Eifel it also prevents growth of rto,
657 * but also it limits too fast rto decreases,
658 * happening in pure Eifel.
663 m -= (tp->mdev >> 2); /* similar update on mdev */
665 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
666 if (tp->mdev > tp->mdev_max) {
667 tp->mdev_max = tp->mdev;
668 if (tp->mdev_max > tp->rttvar)
669 tp->rttvar = tp->mdev_max;
671 if (after(tp->snd_una, tp->rtt_seq)) {
672 if (tp->mdev_max < tp->rttvar)
673 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
674 tp->rtt_seq = tp->snd_nxt;
675 tp->mdev_max = tcp_rto_min(sk);
678 /* no previous measure. */
679 tp->srtt = m<<3; /* take the measured time to be rtt */
680 tp->mdev = m<<1; /* make sure rto = 3*rtt */
681 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
682 tp->rtt_seq = tp->snd_nxt;
686 /* Calculate rto without backoff. This is the second half of Van Jacobson's
687 * routine referred to above.
689 static inline void tcp_set_rto(struct sock *sk)
691 const struct tcp_sock *tp = tcp_sk(sk);
692 /* Old crap is replaced with new one. 8)
695 * 1. If rtt variance happened to be less 50msec, it is hallucination.
696 * It cannot be less due to utterly erratic ACK generation made
697 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
698 * to do with delayed acks, because at cwnd>2 true delack timeout
699 * is invisible. Actually, Linux-2.4 also generates erratic
700 * ACKs in some circumstances.
702 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
704 /* 2. Fixups made earlier cannot be right.
705 * If we do not estimate RTO correctly without them,
706 * all the algo is pure shit and should be replaced
707 * with correct one. It is exactly, which we pretend to do.
711 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
712 * guarantees that rto is higher.
714 static inline void tcp_bound_rto(struct sock *sk)
716 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
717 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
720 /* Save metrics learned by this TCP session.
721 This function is called only, when TCP finishes successfully
722 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
724 void tcp_update_metrics(struct sock *sk)
726 struct tcp_sock *tp = tcp_sk(sk);
727 struct dst_entry *dst = __sk_dst_get(sk);
729 if (sysctl_tcp_nometrics_save)
734 if (dst && (dst->flags&DST_HOST)) {
735 const struct inet_connection_sock *icsk = inet_csk(sk);
738 if (icsk->icsk_backoff || !tp->srtt) {
739 /* This session failed to estimate rtt. Why?
740 * Probably, no packets returned in time.
743 if (!(dst_metric_locked(dst, RTAX_RTT)))
744 dst->metrics[RTAX_RTT-1] = 0;
748 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
750 /* If newly calculated rtt larger than stored one,
751 * store new one. Otherwise, use EWMA. Remember,
752 * rtt overestimation is always better than underestimation.
754 if (!(dst_metric_locked(dst, RTAX_RTT))) {
756 dst->metrics[RTAX_RTT-1] = tp->srtt;
758 dst->metrics[RTAX_RTT-1] -= (m>>3);
761 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
765 /* Scale deviation to rttvar fixed point */
770 if (m >= dst_metric(dst, RTAX_RTTVAR))
771 dst->metrics[RTAX_RTTVAR-1] = m;
773 dst->metrics[RTAX_RTTVAR-1] -=
774 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
777 if (tp->snd_ssthresh >= 0xFFFF) {
778 /* Slow start still did not finish. */
779 if (dst_metric(dst, RTAX_SSTHRESH) &&
780 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
781 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
782 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
783 if (!dst_metric_locked(dst, RTAX_CWND) &&
784 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
785 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
786 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
787 icsk->icsk_ca_state == TCP_CA_Open) {
788 /* Cong. avoidance phase, cwnd is reliable. */
789 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
790 dst->metrics[RTAX_SSTHRESH-1] =
791 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
792 if (!dst_metric_locked(dst, RTAX_CWND))
793 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
795 /* Else slow start did not finish, cwnd is non-sense,
796 ssthresh may be also invalid.
798 if (!dst_metric_locked(dst, RTAX_CWND))
799 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
800 if (dst->metrics[RTAX_SSTHRESH-1] &&
801 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
802 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
803 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
806 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
807 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
808 tp->reordering != sysctl_tcp_reordering)
809 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
814 /* Numbers are taken from RFC3390.
816 * John Heffner states:
818 * The RFC specifies a window of no more than 4380 bytes
819 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
820 * is a bit misleading because they use a clamp at 4380 bytes
821 * rather than use a multiplier in the relevant range.
823 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
825 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
828 if (tp->mss_cache > 1460)
831 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
833 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
836 /* Set slow start threshold and cwnd not falling to slow start */
837 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
839 struct tcp_sock *tp = tcp_sk(sk);
840 const struct inet_connection_sock *icsk = inet_csk(sk);
842 tp->prior_ssthresh = 0;
844 if (icsk->icsk_ca_state < TCP_CA_CWR) {
847 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
848 tp->snd_cwnd = min(tp->snd_cwnd,
849 tcp_packets_in_flight(tp) + 1U);
850 tp->snd_cwnd_cnt = 0;
851 tp->high_seq = tp->snd_nxt;
852 tp->snd_cwnd_stamp = tcp_time_stamp;
853 TCP_ECN_queue_cwr(tp);
855 tcp_set_ca_state(sk, TCP_CA_CWR);
860 * Packet counting of FACK is based on in-order assumptions, therefore TCP
861 * disables it when reordering is detected
863 static void tcp_disable_fack(struct tcp_sock *tp)
865 tp->rx_opt.sack_ok &= ~2;
868 /* Take a notice that peer is sending DSACKs */
869 static void tcp_dsack_seen(struct tcp_sock *tp)
871 tp->rx_opt.sack_ok |= 4;
874 /* Initialize metrics on socket. */
876 static void tcp_init_metrics(struct sock *sk)
878 struct tcp_sock *tp = tcp_sk(sk);
879 struct dst_entry *dst = __sk_dst_get(sk);
886 if (dst_metric_locked(dst, RTAX_CWND))
887 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
888 if (dst_metric(dst, RTAX_SSTHRESH)) {
889 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
890 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
891 tp->snd_ssthresh = tp->snd_cwnd_clamp;
893 if (dst_metric(dst, RTAX_REORDERING) &&
894 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
895 tcp_disable_fack(tp);
896 tp->reordering = dst_metric(dst, RTAX_REORDERING);
899 if (dst_metric(dst, RTAX_RTT) == 0)
902 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
905 /* Initial rtt is determined from SYN,SYN-ACK.
906 * The segment is small and rtt may appear much
907 * less than real one. Use per-dst memory
908 * to make it more realistic.
910 * A bit of theory. RTT is time passed after "normal" sized packet
911 * is sent until it is ACKed. In normal circumstances sending small
912 * packets force peer to delay ACKs and calculation is correct too.
913 * The algorithm is adaptive and, provided we follow specs, it
914 * NEVER underestimate RTT. BUT! If peer tries to make some clever
915 * tricks sort of "quick acks" for time long enough to decrease RTT
916 * to low value, and then abruptly stops to do it and starts to delay
917 * ACKs, wait for troubles.
919 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
920 tp->srtt = dst_metric(dst, RTAX_RTT);
921 tp->rtt_seq = tp->snd_nxt;
923 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
924 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
925 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
929 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
931 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
932 tp->snd_cwnd_stamp = tcp_time_stamp;
936 /* Play conservative. If timestamps are not
937 * supported, TCP will fail to recalculate correct
938 * rtt, if initial rto is too small. FORGET ALL AND RESET!
940 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
942 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
943 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
947 static void tcp_update_reordering(struct sock *sk, const int metric,
950 struct tcp_sock *tp = tcp_sk(sk);
951 if (metric > tp->reordering) {
952 tp->reordering = min(TCP_MAX_REORDERING, metric);
954 /* This exciting event is worth to be remembered. 8) */
956 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
957 else if (tcp_is_reno(tp))
958 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
959 else if (tcp_is_fack(tp))
960 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
962 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
963 #if FASTRETRANS_DEBUG > 1
964 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
965 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
969 tp->undo_marker ? tp->undo_retrans : 0);
971 tcp_disable_fack(tp);
975 /* This procedure tags the retransmission queue when SACKs arrive.
977 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
978 * Packets in queue with these bits set are counted in variables
979 * sacked_out, retrans_out and lost_out, correspondingly.
981 * Valid combinations are:
982 * Tag InFlight Description
983 * 0 1 - orig segment is in flight.
984 * S 0 - nothing flies, orig reached receiver.
985 * L 0 - nothing flies, orig lost by net.
986 * R 2 - both orig and retransmit are in flight.
987 * L|R 1 - orig is lost, retransmit is in flight.
988 * S|R 1 - orig reached receiver, retrans is still in flight.
989 * (L|S|R is logically valid, it could occur when L|R is sacked,
990 * but it is equivalent to plain S and code short-curcuits it to S.
991 * L|S is logically invalid, it would mean -1 packet in flight 8))
993 * These 6 states form finite state machine, controlled by the following events:
994 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
995 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
996 * 3. Loss detection event of one of three flavors:
997 * A. Scoreboard estimator decided the packet is lost.
998 * A'. Reno "three dupacks" marks head of queue lost.
999 * A''. Its FACK modfication, head until snd.fack is lost.
1000 * B. SACK arrives sacking data transmitted after never retransmitted
1001 * hole was sent out.
1002 * C. SACK arrives sacking SND.NXT at the moment, when the
1003 * segment was retransmitted.
1004 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1006 * It is pleasant to note, that state diagram turns out to be commutative,
1007 * so that we are allowed not to be bothered by order of our actions,
1008 * when multiple events arrive simultaneously. (see the function below).
1010 * Reordering detection.
1011 * --------------------
1012 * Reordering metric is maximal distance, which a packet can be displaced
1013 * in packet stream. With SACKs we can estimate it:
1015 * 1. SACK fills old hole and the corresponding segment was not
1016 * ever retransmitted -> reordering. Alas, we cannot use it
1017 * when segment was retransmitted.
1018 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1019 * for retransmitted and already SACKed segment -> reordering..
1020 * Both of these heuristics are not used in Loss state, when we cannot
1021 * account for retransmits accurately.
1023 * SACK block validation.
1024 * ----------------------
1026 * SACK block range validation checks that the received SACK block fits to
1027 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1028 * Note that SND.UNA is not included to the range though being valid because
1029 * it means that the receiver is rather inconsistent with itself (reports
1030 * SACK reneging when it should advance SND.UNA).
1032 * Implements also blockage to start_seq wrap-around. Problem lies in the
1033 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1034 * there's no guarantee that it will be before snd_nxt (n). The problem
1035 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1038 * <- outs wnd -> <- wrapzone ->
1039 * u e n u_w e_w s n_w
1041 * |<------------+------+----- TCP seqno space --------------+---------->|
1042 * ...-- <2^31 ->| |<--------...
1043 * ...---- >2^31 ------>| |<--------...
1045 * Current code wouldn't be vulnerable but it's better still to discard such
1046 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1047 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1048 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1049 * equal to the ideal case (infinite seqno space without wrap caused issues).
1051 * With D-SACK the lower bound is extended to cover sequence space below
1052 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1053 * again, DSACK block must not to go across snd_una (for the same reason as
1054 * for the normal SACK blocks, explained above). But there all simplicity
1055 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1056 * fully below undo_marker they do not affect behavior in anyway and can
1057 * therefore be safely ignored. In rare cases (which are more or less
1058 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1059 * fragmentation and packet reordering past skb's retransmission. To consider
1060 * them correctly, the acceptable range must be extended even more though
1061 * the exact amount is rather hard to quantify. However, tp->max_window can
1062 * be used as an exaggerated estimate.
1064 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1065 u32 start_seq, u32 end_seq)
1067 /* Too far in future, or reversed (interpretation is ambiguous) */
1068 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1071 /* Nasty start_seq wrap-around check (see comments above) */
1072 if (!before(start_seq, tp->snd_nxt))
1075 /* In outstanding window? ...This is valid exit for DSACKs too.
1076 * start_seq == snd_una is non-sensical (see comments above)
1078 if (after(start_seq, tp->snd_una))
1081 if (!is_dsack || !tp->undo_marker)
1084 /* ...Then it's D-SACK, and must reside below snd_una completely */
1085 if (!after(end_seq, tp->snd_una))
1088 if (!before(start_seq, tp->undo_marker))
1092 if (!after(end_seq, tp->undo_marker))
1095 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1096 * start_seq < undo_marker and end_seq >= undo_marker.
1098 return !before(start_seq, end_seq - tp->max_window);
1102 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1103 struct tcp_sack_block_wire *sp, int num_sacks,
1106 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1107 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1110 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1113 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1114 } else if (num_sacks > 1) {
1115 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1116 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1118 if (!after(end_seq_0, end_seq_1) &&
1119 !before(start_seq_0, start_seq_1)) {
1122 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1126 /* D-SACK for already forgotten data... Do dumb counting. */
1128 !after(end_seq_0, prior_snd_una) &&
1129 after(end_seq_0, tp->undo_marker))
1136 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1138 const struct inet_connection_sock *icsk = inet_csk(sk);
1139 struct tcp_sock *tp = tcp_sk(sk);
1140 unsigned char *ptr = (skb_transport_header(ack_skb) +
1141 TCP_SKB_CB(ack_skb)->sacked);
1142 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1143 struct sk_buff *cached_skb;
1144 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1145 int reord = tp->packets_out;
1147 u32 lost_retrans = 0;
1149 int found_dup_sack = 0;
1150 int cached_fack_count;
1152 int first_sack_index;
1154 if (!tp->sacked_out) {
1155 tp->fackets_out = 0;
1156 tp->highest_sack = tp->snd_una;
1158 prior_fackets = tp->fackets_out;
1160 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1161 num_sacks, prior_snd_una);
1163 flag |= FLAG_DSACKING_ACK;
1165 /* Eliminate too old ACKs, but take into
1166 * account more or less fresh ones, they can
1167 * contain valid SACK info.
1169 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1173 * if the only SACK change is the increase of the end_seq of
1174 * the first block then only apply that SACK block
1175 * and use retrans queue hinting otherwise slowpath */
1177 for (i = 0; i < num_sacks; i++) {
1178 __be32 start_seq = sp[i].start_seq;
1179 __be32 end_seq = sp[i].end_seq;
1182 if (tp->recv_sack_cache[i].start_seq != start_seq)
1185 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1186 (tp->recv_sack_cache[i].end_seq != end_seq))
1189 tp->recv_sack_cache[i].start_seq = start_seq;
1190 tp->recv_sack_cache[i].end_seq = end_seq;
1192 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1193 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1194 tp->recv_sack_cache[i].start_seq = 0;
1195 tp->recv_sack_cache[i].end_seq = 0;
1198 first_sack_index = 0;
1203 tp->fastpath_skb_hint = NULL;
1205 /* order SACK blocks to allow in order walk of the retrans queue */
1206 for (i = num_sacks-1; i > 0; i--) {
1207 for (j = 0; j < i; j++){
1208 if (after(ntohl(sp[j].start_seq),
1209 ntohl(sp[j+1].start_seq))){
1210 struct tcp_sack_block_wire tmp;
1216 /* Track where the first SACK block goes to */
1217 if (j == first_sack_index)
1218 first_sack_index = j+1;
1225 /* clear flag as used for different purpose in following code */
1228 /* Use SACK fastpath hint if valid */
1229 cached_skb = tp->fastpath_skb_hint;
1230 cached_fack_count = tp->fastpath_cnt_hint;
1232 cached_skb = tcp_write_queue_head(sk);
1233 cached_fack_count = 0;
1236 for (i=0; i<num_sacks; i++, sp++) {
1237 struct sk_buff *skb;
1238 __u32 start_seq = ntohl(sp->start_seq);
1239 __u32 end_seq = ntohl(sp->end_seq);
1241 int dup_sack = (found_dup_sack && (i == first_sack_index));
1243 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) {
1245 if (!tp->undo_marker)
1246 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1248 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1250 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1255 fack_count = cached_fack_count;
1257 /* Event "B" in the comment above. */
1258 if (after(end_seq, tp->high_seq))
1259 flag |= FLAG_DATA_LOST;
1261 tcp_for_write_queue_from(skb, sk) {
1262 int in_sack, pcount;
1265 if (skb == tcp_send_head(sk))
1269 cached_fack_count = fack_count;
1270 if (i == first_sack_index) {
1271 tp->fastpath_skb_hint = skb;
1272 tp->fastpath_cnt_hint = fack_count;
1275 /* The retransmission queue is always in order, so
1276 * we can short-circuit the walk early.
1278 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1281 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1282 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1284 pcount = tcp_skb_pcount(skb);
1286 if (pcount > 1 && !in_sack &&
1287 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1288 unsigned int pkt_len;
1290 in_sack = !after(start_seq,
1291 TCP_SKB_CB(skb)->seq);
1294 pkt_len = (start_seq -
1295 TCP_SKB_CB(skb)->seq);
1297 pkt_len = (end_seq -
1298 TCP_SKB_CB(skb)->seq);
1299 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1301 pcount = tcp_skb_pcount(skb);
1304 fack_count += pcount;
1306 sacked = TCP_SKB_CB(skb)->sacked;
1308 /* Account D-SACK for retransmitted packet. */
1309 if ((dup_sack && in_sack) &&
1310 (sacked & TCPCB_RETRANS) &&
1311 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1314 /* The frame is ACKed. */
1315 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1316 if (sacked&TCPCB_RETRANS) {
1317 if ((dup_sack && in_sack) &&
1318 (sacked&TCPCB_SACKED_ACKED))
1319 reord = min(fack_count, reord);
1321 /* If it was in a hole, we detected reordering. */
1322 if (fack_count < prior_fackets &&
1323 !(sacked&TCPCB_SACKED_ACKED))
1324 reord = min(fack_count, reord);
1327 /* Nothing to do; acked frame is about to be dropped. */
1331 if ((sacked&TCPCB_SACKED_RETRANS) &&
1332 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1333 (!lost_retrans || after(end_seq, lost_retrans)))
1334 lost_retrans = end_seq;
1339 if (!(sacked&TCPCB_SACKED_ACKED)) {
1340 if (sacked & TCPCB_SACKED_RETRANS) {
1341 /* If the segment is not tagged as lost,
1342 * we do not clear RETRANS, believing
1343 * that retransmission is still in flight.
1345 if (sacked & TCPCB_LOST) {
1346 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1347 tp->lost_out -= tcp_skb_pcount(skb);
1348 tp->retrans_out -= tcp_skb_pcount(skb);
1350 /* clear lost hint */
1351 tp->retransmit_skb_hint = NULL;
1354 /* New sack for not retransmitted frame,
1355 * which was in hole. It is reordering.
1357 if (!(sacked & TCPCB_RETRANS) &&
1358 fack_count < prior_fackets)
1359 reord = min(fack_count, reord);
1361 if (sacked & TCPCB_LOST) {
1362 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1363 tp->lost_out -= tcp_skb_pcount(skb);
1365 /* clear lost hint */
1366 tp->retransmit_skb_hint = NULL;
1368 /* SACK enhanced F-RTO detection.
1369 * Set flag if and only if non-rexmitted
1370 * segments below frto_highmark are
1371 * SACKed (RFC4138; Appendix B).
1372 * Clearing correct due to in-order walk
1374 if (after(end_seq, tp->frto_highmark)) {
1375 flag &= ~FLAG_ONLY_ORIG_SACKED;
1377 if (!(sacked & TCPCB_RETRANS))
1378 flag |= FLAG_ONLY_ORIG_SACKED;
1382 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1383 flag |= FLAG_DATA_SACKED;
1384 tp->sacked_out += tcp_skb_pcount(skb);
1386 if (fack_count > tp->fackets_out)
1387 tp->fackets_out = fack_count;
1389 if (after(TCP_SKB_CB(skb)->seq,
1391 tp->highest_sack = TCP_SKB_CB(skb)->seq;
1393 if (dup_sack && (sacked&TCPCB_RETRANS))
1394 reord = min(fack_count, reord);
1397 /* D-SACK. We can detect redundant retransmission
1398 * in S|R and plain R frames and clear it.
1399 * undo_retrans is decreased above, L|R frames
1400 * are accounted above as well.
1403 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1404 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1405 tp->retrans_out -= tcp_skb_pcount(skb);
1406 tp->retransmit_skb_hint = NULL;
1411 /* Check for lost retransmit. This superb idea is
1412 * borrowed from "ratehalving". Event "C".
1413 * Later note: FACK people cheated me again 8),
1414 * we have to account for reordering! Ugly,
1417 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1418 struct sk_buff *skb;
1420 tcp_for_write_queue(skb, sk) {
1421 if (skb == tcp_send_head(sk))
1423 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1425 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1427 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1428 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1430 !before(lost_retrans,
1431 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1433 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1434 tp->retrans_out -= tcp_skb_pcount(skb);
1436 /* clear lost hint */
1437 tp->retransmit_skb_hint = NULL;
1439 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1440 tp->lost_out += tcp_skb_pcount(skb);
1441 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1442 flag |= FLAG_DATA_SACKED;
1443 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1449 tcp_verify_left_out(tp);
1451 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1452 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1453 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1455 #if FASTRETRANS_DEBUG > 0
1456 BUG_TRAP((int)tp->sacked_out >= 0);
1457 BUG_TRAP((int)tp->lost_out >= 0);
1458 BUG_TRAP((int)tp->retrans_out >= 0);
1459 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1464 /* F-RTO can only be used if TCP has never retransmitted anything other than
1465 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1467 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1469 struct tcp_sock *tp = tcp_sk(sk);
1472 holes = max(tp->lost_out, 1U);
1473 holes = min(holes, tp->packets_out);
1475 if ((tp->sacked_out + holes) > tp->packets_out) {
1476 tp->sacked_out = tp->packets_out - holes;
1477 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1481 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1483 static void tcp_add_reno_sack(struct sock *sk)
1485 struct tcp_sock *tp = tcp_sk(sk);
1487 tcp_check_reno_reordering(sk, 0);
1488 tcp_verify_left_out(tp);
1491 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1493 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1495 struct tcp_sock *tp = tcp_sk(sk);
1498 /* One ACK acked hole. The rest eat duplicate ACKs. */
1499 if (acked-1 >= tp->sacked_out)
1502 tp->sacked_out -= acked-1;
1504 tcp_check_reno_reordering(sk, acked);
1505 tcp_verify_left_out(tp);
1508 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1513 int tcp_use_frto(struct sock *sk)
1515 const struct tcp_sock *tp = tcp_sk(sk);
1516 struct sk_buff *skb;
1518 if (!sysctl_tcp_frto)
1524 /* Avoid expensive walking of rexmit queue if possible */
1525 if (tp->retrans_out > 1)
1528 skb = tcp_write_queue_head(sk);
1529 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1530 tcp_for_write_queue_from(skb, sk) {
1531 if (skb == tcp_send_head(sk))
1533 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1535 /* Short-circuit when first non-SACKed skb has been checked */
1536 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1542 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1543 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1544 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1545 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1546 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1547 * bits are handled if the Loss state is really to be entered (in
1548 * tcp_enter_frto_loss).
1550 * Do like tcp_enter_loss() would; when RTO expires the second time it
1552 * "Reduce ssthresh if it has not yet been made inside this window."
1554 void tcp_enter_frto(struct sock *sk)
1556 const struct inet_connection_sock *icsk = inet_csk(sk);
1557 struct tcp_sock *tp = tcp_sk(sk);
1558 struct sk_buff *skb;
1560 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1561 tp->snd_una == tp->high_seq ||
1562 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1563 !icsk->icsk_retransmits)) {
1564 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1565 /* Our state is too optimistic in ssthresh() call because cwnd
1566 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1567 * recovery has not yet completed. Pattern would be this: RTO,
1568 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1570 * RFC4138 should be more specific on what to do, even though
1571 * RTO is quite unlikely to occur after the first Cumulative ACK
1572 * due to back-off and complexity of triggering events ...
1574 if (tp->frto_counter) {
1576 stored_cwnd = tp->snd_cwnd;
1578 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1579 tp->snd_cwnd = stored_cwnd;
1581 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1583 /* ... in theory, cong.control module could do "any tricks" in
1584 * ssthresh(), which means that ca_state, lost bits and lost_out
1585 * counter would have to be faked before the call occurs. We
1586 * consider that too expensive, unlikely and hacky, so modules
1587 * using these in ssthresh() must deal these incompatibility
1588 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1590 tcp_ca_event(sk, CA_EVENT_FRTO);
1593 tp->undo_marker = tp->snd_una;
1594 tp->undo_retrans = 0;
1596 skb = tcp_write_queue_head(sk);
1597 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1598 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1599 tp->retrans_out -= tcp_skb_pcount(skb);
1601 tcp_verify_left_out(tp);
1603 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1604 * The last condition is necessary at least in tp->frto_counter case.
1606 if (IsSackFrto() && (tp->frto_counter ||
1607 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1608 after(tp->high_seq, tp->snd_una)) {
1609 tp->frto_highmark = tp->high_seq;
1611 tp->frto_highmark = tp->snd_nxt;
1613 tcp_set_ca_state(sk, TCP_CA_Disorder);
1614 tp->high_seq = tp->snd_nxt;
1615 tp->frto_counter = 1;
1618 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1619 * which indicates that we should follow the traditional RTO recovery,
1620 * i.e. mark everything lost and do go-back-N retransmission.
1622 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1624 struct tcp_sock *tp = tcp_sk(sk);
1625 struct sk_buff *skb;
1628 tp->retrans_out = 0;
1629 if (tcp_is_reno(tp))
1630 tcp_reset_reno_sack(tp);
1632 tcp_for_write_queue(skb, sk) {
1633 if (skb == tcp_send_head(sk))
1636 * Count the retransmission made on RTO correctly (only when
1637 * waiting for the first ACK and did not get it)...
1639 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1640 /* For some reason this R-bit might get cleared? */
1641 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1642 tp->retrans_out += tcp_skb_pcount(skb);
1643 /* ...enter this if branch just for the first segment */
1644 flag |= FLAG_DATA_ACKED;
1646 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1649 /* Don't lost mark skbs that were fwd transmitted after RTO */
1650 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1651 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1652 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1653 tp->lost_out += tcp_skb_pcount(skb);
1656 tcp_verify_left_out(tp);
1658 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1659 tp->snd_cwnd_cnt = 0;
1660 tp->snd_cwnd_stamp = tcp_time_stamp;
1661 tp->undo_marker = 0;
1662 tp->frto_counter = 0;
1664 tp->reordering = min_t(unsigned int, tp->reordering,
1665 sysctl_tcp_reordering);
1666 tcp_set_ca_state(sk, TCP_CA_Loss);
1667 tp->high_seq = tp->frto_highmark;
1668 TCP_ECN_queue_cwr(tp);
1670 tcp_clear_all_retrans_hints(tp);
1673 void tcp_clear_retrans(struct tcp_sock *tp)
1675 tp->retrans_out = 0;
1677 tp->fackets_out = 0;
1681 tp->undo_marker = 0;
1682 tp->undo_retrans = 0;
1685 /* Enter Loss state. If "how" is not zero, forget all SACK information
1686 * and reset tags completely, otherwise preserve SACKs. If receiver
1687 * dropped its ofo queue, we will know this due to reneging detection.
1689 void tcp_enter_loss(struct sock *sk, int how)
1691 const struct inet_connection_sock *icsk = inet_csk(sk);
1692 struct tcp_sock *tp = tcp_sk(sk);
1693 struct sk_buff *skb;
1696 /* Reduce ssthresh if it has not yet been made inside this window. */
1697 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1698 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1699 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1700 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1701 tcp_ca_event(sk, CA_EVENT_LOSS);
1704 tp->snd_cwnd_cnt = 0;
1705 tp->snd_cwnd_stamp = tcp_time_stamp;
1707 tp->bytes_acked = 0;
1708 tcp_clear_retrans(tp);
1710 /* Push undo marker, if it was plain RTO and nothing
1711 * was retransmitted. */
1713 tp->undo_marker = tp->snd_una;
1715 tcp_for_write_queue(skb, sk) {
1716 if (skb == tcp_send_head(sk))
1718 cnt += tcp_skb_pcount(skb);
1719 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1720 tp->undo_marker = 0;
1721 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1722 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1723 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1724 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1725 tp->lost_out += tcp_skb_pcount(skb);
1727 tp->sacked_out += tcp_skb_pcount(skb);
1728 tp->fackets_out = cnt;
1731 tcp_verify_left_out(tp);
1733 tp->reordering = min_t(unsigned int, tp->reordering,
1734 sysctl_tcp_reordering);
1735 tcp_set_ca_state(sk, TCP_CA_Loss);
1736 tp->high_seq = tp->snd_nxt;
1737 TCP_ECN_queue_cwr(tp);
1738 /* Abort FRTO algorithm if one is in progress */
1739 tp->frto_counter = 0;
1741 tcp_clear_all_retrans_hints(tp);
1744 static int tcp_check_sack_reneging(struct sock *sk)
1746 struct sk_buff *skb;
1748 /* If ACK arrived pointing to a remembered SACK,
1749 * it means that our remembered SACKs do not reflect
1750 * real state of receiver i.e.
1751 * receiver _host_ is heavily congested (or buggy).
1752 * Do processing similar to RTO timeout.
1754 if ((skb = tcp_write_queue_head(sk)) != NULL &&
1755 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1756 struct inet_connection_sock *icsk = inet_csk(sk);
1757 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1759 tcp_enter_loss(sk, 1);
1760 icsk->icsk_retransmits++;
1761 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1762 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1763 icsk->icsk_rto, TCP_RTO_MAX);
1769 static inline int tcp_fackets_out(struct tcp_sock *tp)
1771 return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1774 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1776 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1779 static inline int tcp_head_timedout(struct sock *sk)
1781 struct tcp_sock *tp = tcp_sk(sk);
1783 return tp->packets_out &&
1784 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1787 /* Linux NewReno/SACK/FACK/ECN state machine.
1788 * --------------------------------------
1790 * "Open" Normal state, no dubious events, fast path.
1791 * "Disorder" In all the respects it is "Open",
1792 * but requires a bit more attention. It is entered when
1793 * we see some SACKs or dupacks. It is split of "Open"
1794 * mainly to move some processing from fast path to slow one.
1795 * "CWR" CWND was reduced due to some Congestion Notification event.
1796 * It can be ECN, ICMP source quench, local device congestion.
1797 * "Recovery" CWND was reduced, we are fast-retransmitting.
1798 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1800 * tcp_fastretrans_alert() is entered:
1801 * - each incoming ACK, if state is not "Open"
1802 * - when arrived ACK is unusual, namely:
1807 * Counting packets in flight is pretty simple.
1809 * in_flight = packets_out - left_out + retrans_out
1811 * packets_out is SND.NXT-SND.UNA counted in packets.
1813 * retrans_out is number of retransmitted segments.
1815 * left_out is number of segments left network, but not ACKed yet.
1817 * left_out = sacked_out + lost_out
1819 * sacked_out: Packets, which arrived to receiver out of order
1820 * and hence not ACKed. With SACKs this number is simply
1821 * amount of SACKed data. Even without SACKs
1822 * it is easy to give pretty reliable estimate of this number,
1823 * counting duplicate ACKs.
1825 * lost_out: Packets lost by network. TCP has no explicit
1826 * "loss notification" feedback from network (for now).
1827 * It means that this number can be only _guessed_.
1828 * Actually, it is the heuristics to predict lossage that
1829 * distinguishes different algorithms.
1831 * F.e. after RTO, when all the queue is considered as lost,
1832 * lost_out = packets_out and in_flight = retrans_out.
1834 * Essentially, we have now two algorithms counting
1837 * FACK: It is the simplest heuristics. As soon as we decided
1838 * that something is lost, we decide that _all_ not SACKed
1839 * packets until the most forward SACK are lost. I.e.
1840 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1841 * It is absolutely correct estimate, if network does not reorder
1842 * packets. And it loses any connection to reality when reordering
1843 * takes place. We use FACK by default until reordering
1844 * is suspected on the path to this destination.
1846 * NewReno: when Recovery is entered, we assume that one segment
1847 * is lost (classic Reno). While we are in Recovery and
1848 * a partial ACK arrives, we assume that one more packet
1849 * is lost (NewReno). This heuristics are the same in NewReno
1852 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1853 * deflation etc. CWND is real congestion window, never inflated, changes
1854 * only according to classic VJ rules.
1856 * Really tricky (and requiring careful tuning) part of algorithm
1857 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1858 * The first determines the moment _when_ we should reduce CWND and,
1859 * hence, slow down forward transmission. In fact, it determines the moment
1860 * when we decide that hole is caused by loss, rather than by a reorder.
1862 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1863 * holes, caused by lost packets.
1865 * And the most logically complicated part of algorithm is undo
1866 * heuristics. We detect false retransmits due to both too early
1867 * fast retransmit (reordering) and underestimated RTO, analyzing
1868 * timestamps and D-SACKs. When we detect that some segments were
1869 * retransmitted by mistake and CWND reduction was wrong, we undo
1870 * window reduction and abort recovery phase. This logic is hidden
1871 * inside several functions named tcp_try_undo_<something>.
1874 /* This function decides, when we should leave Disordered state
1875 * and enter Recovery phase, reducing congestion window.
1877 * Main question: may we further continue forward transmission
1878 * with the same cwnd?
1880 static int tcp_time_to_recover(struct sock *sk)
1882 struct tcp_sock *tp = tcp_sk(sk);
1885 /* Do not perform any recovery during FRTO algorithm */
1886 if (tp->frto_counter)
1889 /* Trick#1: The loss is proven. */
1893 /* Not-A-Trick#2 : Classic rule... */
1894 if (tcp_fackets_out(tp) > tp->reordering)
1897 /* Trick#3 : when we use RFC2988 timer restart, fast
1898 * retransmit can be triggered by timeout of queue head.
1900 if (tcp_head_timedout(sk))
1903 /* Trick#4: It is still not OK... But will it be useful to delay
1906 packets_out = tp->packets_out;
1907 if (packets_out <= tp->reordering &&
1908 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1909 !tcp_may_send_now(sk)) {
1910 /* We have nothing to send. This connection is limited
1911 * either by receiver window or by application.
1919 /* RFC: This is from the original, I doubt that this is necessary at all:
1920 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1921 * retransmitted past LOST markings in the first place? I'm not fully sure
1922 * about undo and end of connection cases, which can cause R without L?
1924 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
1925 struct sk_buff *skb)
1927 if ((tp->retransmit_skb_hint != NULL) &&
1928 before(TCP_SKB_CB(skb)->seq,
1929 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1930 tp->retransmit_skb_hint = NULL;
1933 /* Mark head of queue up as lost. */
1934 static void tcp_mark_head_lost(struct sock *sk,
1935 int packets, u32 high_seq)
1937 struct tcp_sock *tp = tcp_sk(sk);
1938 struct sk_buff *skb;
1941 BUG_TRAP(packets <= tp->packets_out);
1942 if (tp->lost_skb_hint) {
1943 skb = tp->lost_skb_hint;
1944 cnt = tp->lost_cnt_hint;
1946 skb = tcp_write_queue_head(sk);
1950 tcp_for_write_queue_from(skb, sk) {
1951 if (skb == tcp_send_head(sk))
1953 /* TODO: do this better */
1954 /* this is not the most efficient way to do this... */
1955 tp->lost_skb_hint = skb;
1956 tp->lost_cnt_hint = cnt;
1957 cnt += tcp_skb_pcount(skb);
1958 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1960 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1961 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1962 tp->lost_out += tcp_skb_pcount(skb);
1963 tcp_verify_retransmit_hint(tp, skb);
1966 tcp_verify_left_out(tp);
1969 /* Account newly detected lost packet(s) */
1971 static void tcp_update_scoreboard(struct sock *sk)
1973 struct tcp_sock *tp = tcp_sk(sk);
1975 if (tcp_is_fack(tp)) {
1976 int lost = tp->fackets_out - tp->reordering;
1979 tcp_mark_head_lost(sk, lost, tp->high_seq);
1981 tcp_mark_head_lost(sk, 1, tp->high_seq);
1984 /* New heuristics: it is possible only after we switched
1985 * to restart timer each time when something is ACKed.
1986 * Hence, we can detect timed out packets during fast
1987 * retransmit without falling to slow start.
1989 if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) {
1990 struct sk_buff *skb;
1992 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1993 : tcp_write_queue_head(sk);
1995 tcp_for_write_queue_from(skb, sk) {
1996 if (skb == tcp_send_head(sk))
1998 if (!tcp_skb_timedout(sk, skb))
2001 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
2002 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2003 tp->lost_out += tcp_skb_pcount(skb);
2004 tcp_verify_retransmit_hint(tp, skb);
2008 tp->scoreboard_skb_hint = skb;
2010 tcp_verify_left_out(tp);
2014 /* CWND moderation, preventing bursts due to too big ACKs
2015 * in dubious situations.
2017 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2019 tp->snd_cwnd = min(tp->snd_cwnd,
2020 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
2021 tp->snd_cwnd_stamp = tcp_time_stamp;
2024 /* Lower bound on congestion window is slow start threshold
2025 * unless congestion avoidance choice decides to overide it.
2027 static inline u32 tcp_cwnd_min(const struct sock *sk)
2029 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2031 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2034 /* Decrease cwnd each second ack. */
2035 static void tcp_cwnd_down(struct sock *sk, int flag)
2037 struct tcp_sock *tp = tcp_sk(sk);
2038 int decr = tp->snd_cwnd_cnt + 1;
2040 if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
2041 (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
2042 tp->snd_cwnd_cnt = decr&1;
2045 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2046 tp->snd_cwnd -= decr;
2048 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
2049 tp->snd_cwnd_stamp = tcp_time_stamp;
2053 /* Nothing was retransmitted or returned timestamp is less
2054 * than timestamp of the first retransmission.
2056 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2058 return !tp->retrans_stamp ||
2059 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2060 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2063 /* Undo procedures. */
2065 #if FASTRETRANS_DEBUG > 1
2066 static void DBGUNDO(struct sock *sk, const char *msg)
2068 struct tcp_sock *tp = tcp_sk(sk);
2069 struct inet_sock *inet = inet_sk(sk);
2071 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2073 NIPQUAD(inet->daddr), ntohs(inet->dport),
2074 tp->snd_cwnd, tcp_left_out(tp),
2075 tp->snd_ssthresh, tp->prior_ssthresh,
2079 #define DBGUNDO(x...) do { } while (0)
2082 static void tcp_undo_cwr(struct sock *sk, const int undo)
2084 struct tcp_sock *tp = tcp_sk(sk);
2086 if (tp->prior_ssthresh) {
2087 const struct inet_connection_sock *icsk = inet_csk(sk);
2089 if (icsk->icsk_ca_ops->undo_cwnd)
2090 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2092 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
2094 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2095 tp->snd_ssthresh = tp->prior_ssthresh;
2096 TCP_ECN_withdraw_cwr(tp);
2099 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2101 tcp_moderate_cwnd(tp);
2102 tp->snd_cwnd_stamp = tcp_time_stamp;
2104 /* There is something screwy going on with the retrans hints after
2106 tcp_clear_all_retrans_hints(tp);
2109 static inline int tcp_may_undo(struct tcp_sock *tp)
2111 return tp->undo_marker &&
2112 (!tp->undo_retrans || tcp_packet_delayed(tp));
2115 /* People celebrate: "We love our President!" */
2116 static int tcp_try_undo_recovery(struct sock *sk)
2118 struct tcp_sock *tp = tcp_sk(sk);
2120 if (tcp_may_undo(tp)) {
2121 /* Happy end! We did not retransmit anything
2122 * or our original transmission succeeded.
2124 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2125 tcp_undo_cwr(sk, 1);
2126 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2127 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2129 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2130 tp->undo_marker = 0;
2132 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2133 /* Hold old state until something *above* high_seq
2134 * is ACKed. For Reno it is MUST to prevent false
2135 * fast retransmits (RFC2582). SACK TCP is safe. */
2136 tcp_moderate_cwnd(tp);
2139 tcp_set_ca_state(sk, TCP_CA_Open);
2143 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2144 static void tcp_try_undo_dsack(struct sock *sk)
2146 struct tcp_sock *tp = tcp_sk(sk);
2148 if (tp->undo_marker && !tp->undo_retrans) {
2149 DBGUNDO(sk, "D-SACK");
2150 tcp_undo_cwr(sk, 1);
2151 tp->undo_marker = 0;
2152 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2156 /* Undo during fast recovery after partial ACK. */
2158 static int tcp_try_undo_partial(struct sock *sk, int acked)
2160 struct tcp_sock *tp = tcp_sk(sk);
2161 /* Partial ACK arrived. Force Hoe's retransmit. */
2162 int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering;
2164 if (tcp_may_undo(tp)) {
2165 /* Plain luck! Hole if filled with delayed
2166 * packet, rather than with a retransmit.
2168 if (tp->retrans_out == 0)
2169 tp->retrans_stamp = 0;
2171 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2174 tcp_undo_cwr(sk, 0);
2175 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2177 /* So... Do not make Hoe's retransmit yet.
2178 * If the first packet was delayed, the rest
2179 * ones are most probably delayed as well.
2186 /* Undo during loss recovery after partial ACK. */
2187 static int tcp_try_undo_loss(struct sock *sk)
2189 struct tcp_sock *tp = tcp_sk(sk);
2191 if (tcp_may_undo(tp)) {
2192 struct sk_buff *skb;
2193 tcp_for_write_queue(skb, sk) {
2194 if (skb == tcp_send_head(sk))
2196 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2199 tcp_clear_all_retrans_hints(tp);
2201 DBGUNDO(sk, "partial loss");
2203 tcp_undo_cwr(sk, 1);
2204 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2205 inet_csk(sk)->icsk_retransmits = 0;
2206 tp->undo_marker = 0;
2207 if (tcp_is_sack(tp))
2208 tcp_set_ca_state(sk, TCP_CA_Open);
2214 static inline void tcp_complete_cwr(struct sock *sk)
2216 struct tcp_sock *tp = tcp_sk(sk);
2217 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2218 tp->snd_cwnd_stamp = tcp_time_stamp;
2219 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2222 static void tcp_try_to_open(struct sock *sk, int flag)
2224 struct tcp_sock *tp = tcp_sk(sk);
2226 tcp_verify_left_out(tp);
2228 if (tp->retrans_out == 0)
2229 tp->retrans_stamp = 0;
2232 tcp_enter_cwr(sk, 1);
2234 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2235 int state = TCP_CA_Open;
2237 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2238 state = TCP_CA_Disorder;
2240 if (inet_csk(sk)->icsk_ca_state != state) {
2241 tcp_set_ca_state(sk, state);
2242 tp->high_seq = tp->snd_nxt;
2244 tcp_moderate_cwnd(tp);
2246 tcp_cwnd_down(sk, flag);
2250 static void tcp_mtup_probe_failed(struct sock *sk)
2252 struct inet_connection_sock *icsk = inet_csk(sk);
2254 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2255 icsk->icsk_mtup.probe_size = 0;
2258 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2260 struct tcp_sock *tp = tcp_sk(sk);
2261 struct inet_connection_sock *icsk = inet_csk(sk);
2263 /* FIXME: breaks with very large cwnd */
2264 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2265 tp->snd_cwnd = tp->snd_cwnd *
2266 tcp_mss_to_mtu(sk, tp->mss_cache) /
2267 icsk->icsk_mtup.probe_size;
2268 tp->snd_cwnd_cnt = 0;
2269 tp->snd_cwnd_stamp = tcp_time_stamp;
2270 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2272 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2273 icsk->icsk_mtup.probe_size = 0;
2274 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2278 /* Process an event, which can update packets-in-flight not trivially.
2279 * Main goal of this function is to calculate new estimate for left_out,
2280 * taking into account both packets sitting in receiver's buffer and
2281 * packets lost by network.
2283 * Besides that it does CWND reduction, when packet loss is detected
2284 * and changes state of machine.
2286 * It does _not_ decide what to send, it is made in function
2287 * tcp_xmit_retransmit_queue().
2290 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2292 struct inet_connection_sock *icsk = inet_csk(sk);
2293 struct tcp_sock *tp = tcp_sk(sk);
2294 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2295 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2296 (tp->fackets_out > tp->reordering));
2298 /* Some technical things:
2299 * 1. Reno does not count dupacks (sacked_out) automatically. */
2300 if (!tp->packets_out)
2303 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2304 tp->fackets_out = 0;
2306 /* Now state machine starts.
2307 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2309 tp->prior_ssthresh = 0;
2311 /* B. In all the states check for reneging SACKs. */
2312 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2315 /* C. Process data loss notification, provided it is valid. */
2316 if ((flag&FLAG_DATA_LOST) &&
2317 before(tp->snd_una, tp->high_seq) &&
2318 icsk->icsk_ca_state != TCP_CA_Open &&
2319 tp->fackets_out > tp->reordering) {
2320 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2321 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2324 /* D. Check consistency of the current state. */
2325 tcp_verify_left_out(tp);
2327 /* E. Check state exit conditions. State can be terminated
2328 * when high_seq is ACKed. */
2329 if (icsk->icsk_ca_state == TCP_CA_Open) {
2330 BUG_TRAP(tp->retrans_out == 0);
2331 tp->retrans_stamp = 0;
2332 } else if (!before(tp->snd_una, tp->high_seq)) {
2333 switch (icsk->icsk_ca_state) {
2335 icsk->icsk_retransmits = 0;
2336 if (tcp_try_undo_recovery(sk))
2341 /* CWR is to be held something *above* high_seq
2342 * is ACKed for CWR bit to reach receiver. */
2343 if (tp->snd_una != tp->high_seq) {
2344 tcp_complete_cwr(sk);
2345 tcp_set_ca_state(sk, TCP_CA_Open);
2349 case TCP_CA_Disorder:
2350 tcp_try_undo_dsack(sk);
2351 if (!tp->undo_marker ||
2352 /* For SACK case do not Open to allow to undo
2353 * catching for all duplicate ACKs. */
2354 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2355 tp->undo_marker = 0;
2356 tcp_set_ca_state(sk, TCP_CA_Open);
2360 case TCP_CA_Recovery:
2361 if (tcp_is_reno(tp))
2362 tcp_reset_reno_sack(tp);
2363 if (tcp_try_undo_recovery(sk))
2365 tcp_complete_cwr(sk);
2370 /* F. Process state. */
2371 switch (icsk->icsk_ca_state) {
2372 case TCP_CA_Recovery:
2373 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2374 if (tcp_is_reno(tp) && is_dupack)
2375 tcp_add_reno_sack(sk);
2377 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2380 if (flag&FLAG_DATA_ACKED)
2381 icsk->icsk_retransmits = 0;
2382 if (!tcp_try_undo_loss(sk)) {
2383 tcp_moderate_cwnd(tp);
2384 tcp_xmit_retransmit_queue(sk);
2387 if (icsk->icsk_ca_state != TCP_CA_Open)
2389 /* Loss is undone; fall through to processing in Open state. */
2391 if (tcp_is_reno(tp)) {
2392 if (flag & FLAG_SND_UNA_ADVANCED)
2393 tcp_reset_reno_sack(tp);
2395 tcp_add_reno_sack(sk);
2398 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2399 tcp_try_undo_dsack(sk);
2401 if (!tcp_time_to_recover(sk)) {
2402 tcp_try_to_open(sk, flag);
2406 /* MTU probe failure: don't reduce cwnd */
2407 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2408 icsk->icsk_mtup.probe_size &&
2409 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2410 tcp_mtup_probe_failed(sk);
2411 /* Restores the reduction we did in tcp_mtup_probe() */
2413 tcp_simple_retransmit(sk);
2417 /* Otherwise enter Recovery state */
2419 if (tcp_is_reno(tp))
2420 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2422 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2424 tp->high_seq = tp->snd_nxt;
2425 tp->prior_ssthresh = 0;
2426 tp->undo_marker = tp->snd_una;
2427 tp->undo_retrans = tp->retrans_out;
2429 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2430 if (!(flag&FLAG_ECE))
2431 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2432 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2433 TCP_ECN_queue_cwr(tp);
2436 tp->bytes_acked = 0;
2437 tp->snd_cwnd_cnt = 0;
2438 tcp_set_ca_state(sk, TCP_CA_Recovery);
2441 if (do_lost || tcp_head_timedout(sk))
2442 tcp_update_scoreboard(sk);
2443 tcp_cwnd_down(sk, flag);
2444 tcp_xmit_retransmit_queue(sk);
2447 /* Read draft-ietf-tcplw-high-performance before mucking
2448 * with this code. (Supersedes RFC1323)
2450 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2452 /* RTTM Rule: A TSecr value received in a segment is used to
2453 * update the averaged RTT measurement only if the segment
2454 * acknowledges some new data, i.e., only if it advances the
2455 * left edge of the send window.
2457 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2458 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2460 * Changed: reset backoff as soon as we see the first valid sample.
2461 * If we do not, we get strongly overestimated rto. With timestamps
2462 * samples are accepted even from very old segments: f.e., when rtt=1
2463 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2464 * answer arrives rto becomes 120 seconds! If at least one of segments
2465 * in window is lost... Voila. --ANK (010210)
2467 struct tcp_sock *tp = tcp_sk(sk);
2468 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2469 tcp_rtt_estimator(sk, seq_rtt);
2471 inet_csk(sk)->icsk_backoff = 0;
2475 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2477 /* We don't have a timestamp. Can only use
2478 * packets that are not retransmitted to determine
2479 * rtt estimates. Also, we must not reset the
2480 * backoff for rto until we get a non-retransmitted
2481 * packet. This allows us to deal with a situation
2482 * where the network delay has increased suddenly.
2483 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2486 if (flag & FLAG_RETRANS_DATA_ACKED)
2489 tcp_rtt_estimator(sk, seq_rtt);
2491 inet_csk(sk)->icsk_backoff = 0;
2495 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2498 const struct tcp_sock *tp = tcp_sk(sk);
2499 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2500 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2501 tcp_ack_saw_tstamp(sk, flag);
2502 else if (seq_rtt >= 0)
2503 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2506 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2507 u32 in_flight, int good)
2509 const struct inet_connection_sock *icsk = inet_csk(sk);
2510 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2511 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2514 /* Restart timer after forward progress on connection.
2515 * RFC2988 recommends to restart timer to now+rto.
2517 static void tcp_rearm_rto(struct sock *sk)
2519 struct tcp_sock *tp = tcp_sk(sk);
2521 if (!tp->packets_out) {
2522 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2524 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2528 /* If we get here, the whole TSO packet has not been acked. */
2529 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2531 struct tcp_sock *tp = tcp_sk(sk);
2534 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2536 packets_acked = tcp_skb_pcount(skb);
2537 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2539 packets_acked -= tcp_skb_pcount(skb);
2541 if (packets_acked) {
2542 BUG_ON(tcp_skb_pcount(skb) == 0);
2543 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2546 return packets_acked;
2549 /* Remove acknowledged frames from the retransmission queue. If our packet
2550 * is before the ack sequence we can discard it as it's confirmed to have
2551 * arrived at the other end.
2553 static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p)
2555 struct tcp_sock *tp = tcp_sk(sk);
2556 const struct inet_connection_sock *icsk = inet_csk(sk);
2557 struct sk_buff *skb;
2558 u32 now = tcp_time_stamp;
2559 int fully_acked = 1;
2561 int prior_packets = tp->packets_out;
2563 ktime_t last_ackt = net_invalid_timestamp();
2565 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2566 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2569 u8 sacked = scb->sacked;
2571 if (after(scb->end_seq, tp->snd_una)) {
2572 if (tcp_skb_pcount(skb) == 1 ||
2573 !after(tp->snd_una, scb->seq))
2576 packets_acked = tcp_tso_acked(sk, skb);
2581 end_seq = tp->snd_una;
2583 packets_acked = tcp_skb_pcount(skb);
2584 end_seq = scb->end_seq;
2587 /* Initial outgoing SYN's get put onto the write_queue
2588 * just like anything else we transmit. It is not
2589 * true data, and if we misinform our callers that
2590 * this ACK acks real data, we will erroneously exit
2591 * connection startup slow start one packet too
2592 * quickly. This is severely frowned upon behavior.
2594 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2595 flag |= FLAG_DATA_ACKED;
2597 flag |= FLAG_SYN_ACKED;
2598 tp->retrans_stamp = 0;
2601 /* MTU probing checks */
2602 if (fully_acked && icsk->icsk_mtup.probe_size &&
2603 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2604 tcp_mtup_probe_success(sk, skb);
2608 if (sacked & TCPCB_RETRANS) {
2609 if (sacked & TCPCB_SACKED_RETRANS)
2610 tp->retrans_out -= packets_acked;
2611 flag |= FLAG_RETRANS_DATA_ACKED;
2613 } else if (seq_rtt < 0) {
2614 seq_rtt = now - scb->when;
2616 last_ackt = skb->tstamp;
2619 if (sacked & TCPCB_SACKED_ACKED)
2620 tp->sacked_out -= packets_acked;
2621 if (sacked & TCPCB_LOST)
2622 tp->lost_out -= packets_acked;
2624 if ((sacked & TCPCB_URG) && tp->urg_mode &&
2625 !before(end_seq, tp->snd_up))
2627 } else if (seq_rtt < 0) {
2628 seq_rtt = now - scb->when;
2630 last_ackt = skb->tstamp;
2632 tp->packets_out -= packets_acked;
2637 tcp_unlink_write_queue(skb, sk);
2638 sk_stream_free_skb(sk, skb);
2639 tcp_clear_all_retrans_hints(tp);
2642 if (flag & FLAG_ACKED) {
2643 u32 pkts_acked = prior_packets - tp->packets_out;
2644 const struct tcp_congestion_ops *ca_ops
2645 = inet_csk(sk)->icsk_ca_ops;
2647 tcp_ack_update_rtt(sk, flag, seq_rtt);
2650 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2651 /* hint's skb might be NULL but we don't need to care */
2652 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
2653 tp->fastpath_cnt_hint);
2654 if (tcp_is_reno(tp))
2655 tcp_remove_reno_sacks(sk, pkts_acked);
2657 if (ca_ops->pkts_acked) {
2660 /* Is the ACK triggering packet unambiguous? */
2661 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2662 /* High resolution needed and available? */
2663 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2664 !ktime_equal(last_ackt,
2665 net_invalid_timestamp()))
2666 rtt_us = ktime_us_delta(ktime_get_real(),
2668 else if (seq_rtt > 0)
2669 rtt_us = jiffies_to_usecs(seq_rtt);
2672 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2676 #if FASTRETRANS_DEBUG > 0
2677 BUG_TRAP((int)tp->sacked_out >= 0);
2678 BUG_TRAP((int)tp->lost_out >= 0);
2679 BUG_TRAP((int)tp->retrans_out >= 0);
2680 if (!tp->packets_out && tcp_is_sack(tp)) {
2681 const struct inet_connection_sock *icsk = inet_csk(sk);
2683 printk(KERN_DEBUG "Leak l=%u %d\n",
2684 tp->lost_out, icsk->icsk_ca_state);
2687 if (tp->sacked_out) {
2688 printk(KERN_DEBUG "Leak s=%u %d\n",
2689 tp->sacked_out, icsk->icsk_ca_state);
2692 if (tp->retrans_out) {
2693 printk(KERN_DEBUG "Leak r=%u %d\n",
2694 tp->retrans_out, icsk->icsk_ca_state);
2695 tp->retrans_out = 0;
2699 *seq_rtt_p = seq_rtt;
2703 static void tcp_ack_probe(struct sock *sk)
2705 const struct tcp_sock *tp = tcp_sk(sk);
2706 struct inet_connection_sock *icsk = inet_csk(sk);
2708 /* Was it a usable window open? */
2710 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2711 tp->snd_una + tp->snd_wnd)) {
2712 icsk->icsk_backoff = 0;
2713 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2714 /* Socket must be waked up by subsequent tcp_data_snd_check().
2715 * This function is not for random using!
2718 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2719 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2724 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2726 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2727 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2730 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2732 const struct tcp_sock *tp = tcp_sk(sk);
2733 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2734 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2737 /* Check that window update is acceptable.
2738 * The function assumes that snd_una<=ack<=snd_next.
2740 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2741 const u32 ack_seq, const u32 nwin)
2743 return (after(ack, tp->snd_una) ||
2744 after(ack_seq, tp->snd_wl1) ||
2745 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2748 /* Update our send window.
2750 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2751 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2753 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2756 struct tcp_sock *tp = tcp_sk(sk);
2758 u32 nwin = ntohs(tcp_hdr(skb)->window);
2760 if (likely(!tcp_hdr(skb)->syn))
2761 nwin <<= tp->rx_opt.snd_wscale;
2763 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2764 flag |= FLAG_WIN_UPDATE;
2765 tcp_update_wl(tp, ack, ack_seq);
2767 if (tp->snd_wnd != nwin) {
2770 /* Note, it is the only place, where
2771 * fast path is recovered for sending TCP.
2774 tcp_fast_path_check(sk);
2776 if (nwin > tp->max_window) {
2777 tp->max_window = nwin;
2778 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2788 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2789 * continue in congestion avoidance.
2791 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2793 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2794 tp->snd_cwnd_cnt = 0;
2795 TCP_ECN_queue_cwr(tp);
2796 tcp_moderate_cwnd(tp);
2799 /* A conservative spurious RTO response algorithm: reduce cwnd using
2800 * rate halving and continue in congestion avoidance.
2802 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2804 tcp_enter_cwr(sk, 0);
2807 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2810 tcp_ratehalving_spur_to_response(sk);
2812 tcp_undo_cwr(sk, 1);
2815 /* F-RTO spurious RTO detection algorithm (RFC4138)
2817 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2818 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2819 * window (but not to or beyond highest sequence sent before RTO):
2820 * On First ACK, send two new segments out.
2821 * On Second ACK, RTO was likely spurious. Do spurious response (response
2822 * algorithm is not part of the F-RTO detection algorithm
2823 * given in RFC4138 but can be selected separately).
2824 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2825 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2826 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2827 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2829 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2830 * original window even after we transmit two new data segments.
2833 * on first step, wait until first cumulative ACK arrives, then move to
2834 * the second step. In second step, the next ACK decides.
2836 * F-RTO is implemented (mainly) in four functions:
2837 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2838 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2839 * called when tcp_use_frto() showed green light
2840 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2841 * - tcp_enter_frto_loss() is called if there is not enough evidence
2842 * to prove that the RTO is indeed spurious. It transfers the control
2843 * from F-RTO to the conventional RTO recovery
2845 static int tcp_process_frto(struct sock *sk, int flag)
2847 struct tcp_sock *tp = tcp_sk(sk);
2849 tcp_verify_left_out(tp);
2851 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2852 if (flag&FLAG_DATA_ACKED)
2853 inet_csk(sk)->icsk_retransmits = 0;
2855 if (!before(tp->snd_una, tp->frto_highmark)) {
2856 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2860 if (!IsSackFrto() || tcp_is_reno(tp)) {
2861 /* RFC4138 shortcoming in step 2; should also have case c):
2862 * ACK isn't duplicate nor advances window, e.g., opposite dir
2865 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2868 if (!(flag&FLAG_DATA_ACKED)) {
2869 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2874 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2875 /* Prevent sending of new data. */
2876 tp->snd_cwnd = min(tp->snd_cwnd,
2877 tcp_packets_in_flight(tp));
2881 if ((tp->frto_counter >= 2) &&
2882 (!(flag&FLAG_FORWARD_PROGRESS) ||
2883 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2884 /* RFC4138 shortcoming (see comment above) */
2885 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2888 tcp_enter_frto_loss(sk, 3, flag);
2893 if (tp->frto_counter == 1) {
2894 /* Sending of the next skb must be allowed or no FRTO */
2895 if (!tcp_send_head(sk) ||
2896 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2897 tp->snd_una + tp->snd_wnd)) {
2898 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2903 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2904 tp->frto_counter = 2;
2907 switch (sysctl_tcp_frto_response) {
2909 tcp_undo_spur_to_response(sk, flag);
2912 tcp_conservative_spur_to_response(tp);
2915 tcp_ratehalving_spur_to_response(sk);
2918 tp->frto_counter = 0;
2923 /* This routine deals with incoming acks, but not outgoing ones. */
2924 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2926 struct inet_connection_sock *icsk = inet_csk(sk);
2927 struct tcp_sock *tp = tcp_sk(sk);
2928 u32 prior_snd_una = tp->snd_una;
2929 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2930 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2931 u32 prior_in_flight;
2936 /* If the ack is newer than sent or older than previous acks
2937 * then we can probably ignore it.
2939 if (after(ack, tp->snd_nxt))
2940 goto uninteresting_ack;
2942 if (before(ack, prior_snd_una))
2945 if (after(ack, prior_snd_una))
2946 flag |= FLAG_SND_UNA_ADVANCED;
2948 if (sysctl_tcp_abc) {
2949 if (icsk->icsk_ca_state < TCP_CA_CWR)
2950 tp->bytes_acked += ack - prior_snd_una;
2951 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2952 /* we assume just one segment left network */
2953 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2956 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2957 /* Window is constant, pure forward advance.
2958 * No more checks are required.
2959 * Note, we use the fact that SND.UNA>=SND.WL2.
2961 tcp_update_wl(tp, ack, ack_seq);
2963 flag |= FLAG_WIN_UPDATE;
2965 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2967 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2969 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2972 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2974 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2976 if (TCP_SKB_CB(skb)->sacked)
2977 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2979 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2982 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2985 /* We passed data and got it acked, remove any soft error
2986 * log. Something worked...
2988 sk->sk_err_soft = 0;
2989 tp->rcv_tstamp = tcp_time_stamp;
2990 prior_packets = tp->packets_out;
2994 prior_in_flight = tcp_packets_in_flight(tp);
2996 /* See if we can take anything off of the retransmit queue. */
2997 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2999 if (tp->frto_counter)
3000 frto_cwnd = tcp_process_frto(sk, flag);
3002 if (tcp_ack_is_dubious(sk, flag)) {
3003 /* Advance CWND, if state allows this. */
3004 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3005 tcp_may_raise_cwnd(sk, flag))
3006 tcp_cong_avoid(sk, ack, prior_in_flight, 0);
3007 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
3009 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3010 tcp_cong_avoid(sk, ack, prior_in_flight, 1);
3013 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
3014 dst_confirm(sk->sk_dst_cache);
3019 icsk->icsk_probes_out = 0;
3021 /* If this ack opens up a zero window, clear backoff. It was
3022 * being used to time the probes, and is probably far higher than
3023 * it needs to be for normal retransmission.
3025 if (tcp_send_head(sk))
3030 if (TCP_SKB_CB(skb)->sacked)
3031 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3034 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3039 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3040 * But, this can also be called on packets in the established flow when
3041 * the fast version below fails.
3043 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
3046 struct tcphdr *th = tcp_hdr(skb);
3047 int length=(th->doff*4)-sizeof(struct tcphdr);
3049 ptr = (unsigned char *)(th + 1);
3050 opt_rx->saw_tstamp = 0;
3052 while (length > 0) {
3059 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3064 if (opsize < 2) /* "silly options" */
3066 if (opsize > length)
3067 return; /* don't parse partial options */
3070 if (opsize==TCPOLEN_MSS && th->syn && !estab) {
3071 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3073 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3074 in_mss = opt_rx->user_mss;
3075 opt_rx->mss_clamp = in_mss;
3080 if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3081 if (sysctl_tcp_window_scaling) {
3082 __u8 snd_wscale = *(__u8 *) ptr;
3083 opt_rx->wscale_ok = 1;
3084 if (snd_wscale > 14) {
3085 if (net_ratelimit())
3086 printk(KERN_INFO "tcp_parse_options: Illegal window "
3087 "scaling value %d >14 received.\n",
3091 opt_rx->snd_wscale = snd_wscale;
3094 case TCPOPT_TIMESTAMP:
3095 if (opsize==TCPOLEN_TIMESTAMP) {
3096 if ((estab && opt_rx->tstamp_ok) ||
3097 (!estab && sysctl_tcp_timestamps)) {
3098 opt_rx->saw_tstamp = 1;
3099 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3100 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3104 case TCPOPT_SACK_PERM:
3105 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3106 if (sysctl_tcp_sack) {
3107 opt_rx->sack_ok = 1;
3108 tcp_sack_reset(opt_rx);
3114 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3115 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3117 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3120 #ifdef CONFIG_TCP_MD5SIG
3123 * The MD5 Hash has already been
3124 * checked (see tcp_v{4,6}_do_rcv()).
3136 /* Fast parse options. This hopes to only see timestamps.
3137 * If it is wrong it falls back on tcp_parse_options().
3139 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3140 struct tcp_sock *tp)
3142 if (th->doff == sizeof(struct tcphdr)>>2) {
3143 tp->rx_opt.saw_tstamp = 0;
3145 } else if (tp->rx_opt.tstamp_ok &&
3146 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3147 __be32 *ptr = (__be32 *)(th + 1);
3148 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3149 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3150 tp->rx_opt.saw_tstamp = 1;
3152 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3154 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3158 tcp_parse_options(skb, &tp->rx_opt, 1);
3162 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3164 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3165 tp->rx_opt.ts_recent_stamp = get_seconds();
3168 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3170 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3171 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3172 * extra check below makes sure this can only happen
3173 * for pure ACK frames. -DaveM
3175 * Not only, also it occurs for expired timestamps.
3178 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3179 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3180 tcp_store_ts_recent(tp);
3184 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3186 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3187 * it can pass through stack. So, the following predicate verifies that
3188 * this segment is not used for anything but congestion avoidance or
3189 * fast retransmit. Moreover, we even are able to eliminate most of such
3190 * second order effects, if we apply some small "replay" window (~RTO)
3191 * to timestamp space.
3193 * All these measures still do not guarantee that we reject wrapped ACKs
3194 * on networks with high bandwidth, when sequence space is recycled fastly,
3195 * but it guarantees that such events will be very rare and do not affect
3196 * connection seriously. This doesn't look nice, but alas, PAWS is really
3199 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3200 * states that events when retransmit arrives after original data are rare.
3201 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3202 * the biggest problem on large power networks even with minor reordering.
3203 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3204 * up to bandwidth of 18Gigabit/sec. 8) ]
3207 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3209 struct tcp_sock *tp = tcp_sk(sk);
3210 struct tcphdr *th = tcp_hdr(skb);
3211 u32 seq = TCP_SKB_CB(skb)->seq;
3212 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3214 return (/* 1. Pure ACK with correct sequence number. */
3215 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3217 /* 2. ... and duplicate ACK. */
3218 ack == tp->snd_una &&
3220 /* 3. ... and does not update window. */
3221 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3223 /* 4. ... and sits in replay window. */
3224 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3227 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3229 const struct tcp_sock *tp = tcp_sk(sk);
3230 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3231 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3232 !tcp_disordered_ack(sk, skb));
3235 /* Check segment sequence number for validity.
3237 * Segment controls are considered valid, if the segment
3238 * fits to the window after truncation to the window. Acceptability
3239 * of data (and SYN, FIN, of course) is checked separately.
3240 * See tcp_data_queue(), for example.
3242 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3243 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3244 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3245 * (borrowed from freebsd)
3248 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3250 return !before(end_seq, tp->rcv_wup) &&
3251 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3254 /* When we get a reset we do this. */
3255 static void tcp_reset(struct sock *sk)
3257 /* We want the right error as BSD sees it (and indeed as we do). */
3258 switch (sk->sk_state) {
3260 sk->sk_err = ECONNREFUSED;
3262 case TCP_CLOSE_WAIT:
3268 sk->sk_err = ECONNRESET;
3271 if (!sock_flag(sk, SOCK_DEAD))
3272 sk->sk_error_report(sk);
3278 * Process the FIN bit. This now behaves as it is supposed to work
3279 * and the FIN takes effect when it is validly part of sequence
3280 * space. Not before when we get holes.
3282 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3283 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3286 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3287 * close and we go into CLOSING (and later onto TIME-WAIT)
3289 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3291 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3293 struct tcp_sock *tp = tcp_sk(sk);
3295 inet_csk_schedule_ack(sk);
3297 sk->sk_shutdown |= RCV_SHUTDOWN;
3298 sock_set_flag(sk, SOCK_DONE);
3300 switch (sk->sk_state) {
3302 case TCP_ESTABLISHED:
3303 /* Move to CLOSE_WAIT */
3304 tcp_set_state(sk, TCP_CLOSE_WAIT);
3305 inet_csk(sk)->icsk_ack.pingpong = 1;
3308 case TCP_CLOSE_WAIT:
3310 /* Received a retransmission of the FIN, do
3315 /* RFC793: Remain in the LAST-ACK state. */
3319 /* This case occurs when a simultaneous close
3320 * happens, we must ack the received FIN and
3321 * enter the CLOSING state.
3324 tcp_set_state(sk, TCP_CLOSING);
3327 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3329 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3332 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3333 * cases we should never reach this piece of code.
3335 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3336 __FUNCTION__, sk->sk_state);
3340 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3341 * Probably, we should reset in this case. For now drop them.
3343 __skb_queue_purge(&tp->out_of_order_queue);
3344 if (tcp_is_sack(tp))
3345 tcp_sack_reset(&tp->rx_opt);
3346 sk_stream_mem_reclaim(sk);
3348 if (!sock_flag(sk, SOCK_DEAD)) {
3349 sk->sk_state_change(sk);
3351 /* Do not send POLL_HUP for half duplex close. */
3352 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3353 sk->sk_state == TCP_CLOSE)
3354 sk_wake_async(sk, 1, POLL_HUP);
3356 sk_wake_async(sk, 1, POLL_IN);
3360 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3362 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3363 if (before(seq, sp->start_seq))
3364 sp->start_seq = seq;
3365 if (after(end_seq, sp->end_seq))
3366 sp->end_seq = end_seq;
3372 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3374 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3375 if (before(seq, tp->rcv_nxt))
3376 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3378 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3380 tp->rx_opt.dsack = 1;
3381 tp->duplicate_sack[0].start_seq = seq;
3382 tp->duplicate_sack[0].end_seq = end_seq;
3383 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3387 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3389 if (!tp->rx_opt.dsack)
3390 tcp_dsack_set(tp, seq, end_seq);
3392 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3395 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3397 struct tcp_sock *tp = tcp_sk(sk);
3399 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3400 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3401 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3402 tcp_enter_quickack_mode(sk);
3404 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3405 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3407 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3408 end_seq = tp->rcv_nxt;
3409 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3416 /* These routines update the SACK block as out-of-order packets arrive or
3417 * in-order packets close up the sequence space.
3419 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3422 struct tcp_sack_block *sp = &tp->selective_acks[0];
3423 struct tcp_sack_block *swalk = sp+1;
3425 /* See if the recent change to the first SACK eats into
3426 * or hits the sequence space of other SACK blocks, if so coalesce.
3428 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3429 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3432 /* Zap SWALK, by moving every further SACK up by one slot.
3433 * Decrease num_sacks.
3435 tp->rx_opt.num_sacks--;
3436 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3437 for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3441 this_sack++, swalk++;
3445 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3449 tmp = sack1->start_seq;
3450 sack1->start_seq = sack2->start_seq;
3451 sack2->start_seq = tmp;
3453 tmp = sack1->end_seq;
3454 sack1->end_seq = sack2->end_seq;
3455 sack2->end_seq = tmp;
3458 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3460 struct tcp_sock *tp = tcp_sk(sk);
3461 struct tcp_sack_block *sp = &tp->selective_acks[0];
3462 int cur_sacks = tp->rx_opt.num_sacks;
3468 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3469 if (tcp_sack_extend(sp, seq, end_seq)) {
3470 /* Rotate this_sack to the first one. */
3471 for (; this_sack>0; this_sack--, sp--)
3472 tcp_sack_swap(sp, sp-1);
3474 tcp_sack_maybe_coalesce(tp);
3479 /* Could not find an adjacent existing SACK, build a new one,
3480 * put it at the front, and shift everyone else down. We
3481 * always know there is at least one SACK present already here.
3483 * If the sack array is full, forget about the last one.
3485 if (this_sack >= 4) {
3487 tp->rx_opt.num_sacks--;
3490 for (; this_sack > 0; this_sack--, sp--)
3494 /* Build the new head SACK, and we're done. */
3495 sp->start_seq = seq;
3496 sp->end_seq = end_seq;
3497 tp->rx_opt.num_sacks++;
3498 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3501 /* RCV.NXT advances, some SACKs should be eaten. */
3503 static void tcp_sack_remove(struct tcp_sock *tp)
3505 struct tcp_sack_block *sp = &tp->selective_acks[0];
3506 int num_sacks = tp->rx_opt.num_sacks;
3509 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3510 if (skb_queue_empty(&tp->out_of_order_queue)) {
3511 tp->rx_opt.num_sacks = 0;
3512 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3516 for (this_sack = 0; this_sack < num_sacks; ) {
3517 /* Check if the start of the sack is covered by RCV.NXT. */
3518 if (!before(tp->rcv_nxt, sp->start_seq)) {
3521 /* RCV.NXT must cover all the block! */
3522 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3524 /* Zap this SACK, by moving forward any other SACKS. */
3525 for (i=this_sack+1; i < num_sacks; i++)
3526 tp->selective_acks[i-1] = tp->selective_acks[i];
3533 if (num_sacks != tp->rx_opt.num_sacks) {
3534 tp->rx_opt.num_sacks = num_sacks;
3535 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3539 /* This one checks to see if we can put data from the
3540 * out_of_order queue into the receive_queue.
3542 static void tcp_ofo_queue(struct sock *sk)
3544 struct tcp_sock *tp = tcp_sk(sk);
3545 __u32 dsack_high = tp->rcv_nxt;
3546 struct sk_buff *skb;
3548 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3549 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3552 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3553 __u32 dsack = dsack_high;
3554 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3555 dsack_high = TCP_SKB_CB(skb)->end_seq;
3556 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3559 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3560 SOCK_DEBUG(sk, "ofo packet was already received \n");
3561 __skb_unlink(skb, &tp->out_of_order_queue);
3565 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3566 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3567 TCP_SKB_CB(skb)->end_seq);
3569 __skb_unlink(skb, &tp->out_of_order_queue);
3570 __skb_queue_tail(&sk->sk_receive_queue, skb);
3571 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3572 if (tcp_hdr(skb)->fin)
3573 tcp_fin(skb, sk, tcp_hdr(skb));
3577 static int tcp_prune_queue(struct sock *sk);
3579 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3581 struct tcphdr *th = tcp_hdr(skb);
3582 struct tcp_sock *tp = tcp_sk(sk);
3585 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3588 __skb_pull(skb, th->doff*4);
3590 TCP_ECN_accept_cwr(tp, skb);
3592 if (tp->rx_opt.dsack) {
3593 tp->rx_opt.dsack = 0;
3594 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3595 4 - tp->rx_opt.tstamp_ok);
3598 /* Queue data for delivery to the user.
3599 * Packets in sequence go to the receive queue.
3600 * Out of sequence packets to the out_of_order_queue.
3602 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3603 if (tcp_receive_window(tp) == 0)
3606 /* Ok. In sequence. In window. */
3607 if (tp->ucopy.task == current &&
3608 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3609 sock_owned_by_user(sk) && !tp->urg_data) {
3610 int chunk = min_t(unsigned int, skb->len,
3613 __set_current_state(TASK_RUNNING);
3616 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3617 tp->ucopy.len -= chunk;
3618 tp->copied_seq += chunk;
3619 eaten = (chunk == skb->len && !th->fin);
3620 tcp_rcv_space_adjust(sk);
3628 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3629 !sk_stream_rmem_schedule(sk, skb))) {
3630 if (tcp_prune_queue(sk) < 0 ||
3631 !sk_stream_rmem_schedule(sk, skb))
3634 sk_stream_set_owner_r(skb, sk);
3635 __skb_queue_tail(&sk->sk_receive_queue, skb);
3637 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3639 tcp_event_data_recv(sk, skb);
3641 tcp_fin(skb, sk, th);
3643 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3646 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3647 * gap in queue is filled.
3649 if (skb_queue_empty(&tp->out_of_order_queue))
3650 inet_csk(sk)->icsk_ack.pingpong = 0;
3653 if (tp->rx_opt.num_sacks)
3654 tcp_sack_remove(tp);
3656 tcp_fast_path_check(sk);
3660 else if (!sock_flag(sk, SOCK_DEAD))
3661 sk->sk_data_ready(sk, 0);
3665 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3666 /* A retransmit, 2nd most common case. Force an immediate ack. */
3667 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3668 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3671 tcp_enter_quickack_mode(sk);
3672 inet_csk_schedule_ack(sk);
3678 /* Out of window. F.e. zero window probe. */
3679 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3682 tcp_enter_quickack_mode(sk);
3684 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3685 /* Partial packet, seq < rcv_next < end_seq */
3686 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3687 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3688 TCP_SKB_CB(skb)->end_seq);
3690 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3692 /* If window is closed, drop tail of packet. But after
3693 * remembering D-SACK for its head made in previous line.
3695 if (!tcp_receive_window(tp))
3700 TCP_ECN_check_ce(tp, skb);
3702 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3703 !sk_stream_rmem_schedule(sk, skb)) {
3704 if (tcp_prune_queue(sk) < 0 ||
3705 !sk_stream_rmem_schedule(sk, skb))
3709 /* Disable header prediction. */
3711 inet_csk_schedule_ack(sk);
3713 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3714 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3716 sk_stream_set_owner_r(skb, sk);
3718 if (!skb_peek(&tp->out_of_order_queue)) {
3719 /* Initial out of order segment, build 1 SACK. */
3720 if (tcp_is_sack(tp)) {
3721 tp->rx_opt.num_sacks = 1;
3722 tp->rx_opt.dsack = 0;
3723 tp->rx_opt.eff_sacks = 1;
3724 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3725 tp->selective_acks[0].end_seq =
3726 TCP_SKB_CB(skb)->end_seq;
3728 __skb_queue_head(&tp->out_of_order_queue,skb);
3730 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3731 u32 seq = TCP_SKB_CB(skb)->seq;
3732 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3734 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3735 __skb_append(skb1, skb, &tp->out_of_order_queue);
3737 if (!tp->rx_opt.num_sacks ||
3738 tp->selective_acks[0].end_seq != seq)
3741 /* Common case: data arrive in order after hole. */
3742 tp->selective_acks[0].end_seq = end_seq;
3746 /* Find place to insert this segment. */
3748 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3750 } while ((skb1 = skb1->prev) !=
3751 (struct sk_buff*)&tp->out_of_order_queue);
3753 /* Do skb overlap to previous one? */
3754 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3755 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3756 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3757 /* All the bits are present. Drop. */
3759 tcp_dsack_set(tp, seq, end_seq);
3762 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3763 /* Partial overlap. */
3764 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3769 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3771 /* And clean segments covered by new one as whole. */
3772 while ((skb1 = skb->next) !=
3773 (struct sk_buff*)&tp->out_of_order_queue &&
3774 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3775 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3776 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3779 __skb_unlink(skb1, &tp->out_of_order_queue);
3780 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3785 if (tcp_is_sack(tp))
3786 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3790 /* Collapse contiguous sequence of skbs head..tail with
3791 * sequence numbers start..end.
3792 * Segments with FIN/SYN are not collapsed (only because this
3796 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3797 struct sk_buff *head, struct sk_buff *tail,
3800 struct sk_buff *skb;
3802 /* First, check that queue is collapsible and find
3803 * the point where collapsing can be useful. */
3804 for (skb = head; skb != tail; ) {
3805 /* No new bits? It is possible on ofo queue. */
3806 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3807 struct sk_buff *next = skb->next;
3808 __skb_unlink(skb, list);
3810 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3815 /* The first skb to collapse is:
3817 * - bloated or contains data before "start" or
3818 * overlaps to the next one.
3820 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3821 (tcp_win_from_space(skb->truesize) > skb->len ||
3822 before(TCP_SKB_CB(skb)->seq, start) ||
3823 (skb->next != tail &&
3824 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3827 /* Decided to skip this, advance start seq. */
3828 start = TCP_SKB_CB(skb)->end_seq;
3831 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3834 while (before(start, end)) {
3835 struct sk_buff *nskb;
3836 int header = skb_headroom(skb);
3837 int copy = SKB_MAX_ORDER(header, 0);
3839 /* Too big header? This can happen with IPv6. */
3842 if (end-start < copy)
3844 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3848 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3849 skb_set_network_header(nskb, (skb_network_header(skb) -
3851 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3853 skb_reserve(nskb, header);
3854 memcpy(nskb->head, skb->head, header);
3855 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3856 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3857 __skb_insert(nskb, skb->prev, skb, list);
3858 sk_stream_set_owner_r(nskb, sk);
3860 /* Copy data, releasing collapsed skbs. */
3862 int offset = start - TCP_SKB_CB(skb)->seq;
3863 int size = TCP_SKB_CB(skb)->end_seq - start;
3867 size = min(copy, size);
3868 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3870 TCP_SKB_CB(nskb)->end_seq += size;
3874 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3875 struct sk_buff *next = skb->next;
3876 __skb_unlink(skb, list);
3878 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3881 tcp_hdr(skb)->syn ||
3889 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3890 * and tcp_collapse() them until all the queue is collapsed.
3892 static void tcp_collapse_ofo_queue(struct sock *sk)
3894 struct tcp_sock *tp = tcp_sk(sk);
3895 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3896 struct sk_buff *head;
3902 start = TCP_SKB_CB(skb)->seq;
3903 end = TCP_SKB_CB(skb)->end_seq;
3909 /* Segment is terminated when we see gap or when
3910 * we are at the end of all the queue. */
3911 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3912 after(TCP_SKB_CB(skb)->seq, end) ||
3913 before(TCP_SKB_CB(skb)->end_seq, start)) {
3914 tcp_collapse(sk, &tp->out_of_order_queue,
3915 head, skb, start, end);
3917 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3919 /* Start new segment */
3920 start = TCP_SKB_CB(skb)->seq;
3921 end = TCP_SKB_CB(skb)->end_seq;
3923 if (before(TCP_SKB_CB(skb)->seq, start))
3924 start = TCP_SKB_CB(skb)->seq;
3925 if (after(TCP_SKB_CB(skb)->end_seq, end))
3926 end = TCP_SKB_CB(skb)->end_seq;
3931 /* Reduce allocated memory if we can, trying to get
3932 * the socket within its memory limits again.
3934 * Return less than zero if we should start dropping frames
3935 * until the socket owning process reads some of the data
3936 * to stabilize the situation.
3938 static int tcp_prune_queue(struct sock *sk)
3940 struct tcp_sock *tp = tcp_sk(sk);
3942 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3944 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3946 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3947 tcp_clamp_window(sk);
3948 else if (tcp_memory_pressure)
3949 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3951 tcp_collapse_ofo_queue(sk);
3952 tcp_collapse(sk, &sk->sk_receive_queue,
3953 sk->sk_receive_queue.next,
3954 (struct sk_buff*)&sk->sk_receive_queue,
3955 tp->copied_seq, tp->rcv_nxt);
3956 sk_stream_mem_reclaim(sk);
3958 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3961 /* Collapsing did not help, destructive actions follow.
3962 * This must not ever occur. */
3964 /* First, purge the out_of_order queue. */
3965 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3966 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3967 __skb_queue_purge(&tp->out_of_order_queue);
3969 /* Reset SACK state. A conforming SACK implementation will
3970 * do the same at a timeout based retransmit. When a connection
3971 * is in a sad state like this, we care only about integrity
3972 * of the connection not performance.
3974 if (tcp_is_sack(tp))
3975 tcp_sack_reset(&tp->rx_opt);
3976 sk_stream_mem_reclaim(sk);
3979 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3982 /* If we are really being abused, tell the caller to silently
3983 * drop receive data on the floor. It will get retransmitted
3984 * and hopefully then we'll have sufficient space.
3986 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3988 /* Massive buffer overcommit. */
3994 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3995 * As additional protections, we do not touch cwnd in retransmission phases,
3996 * and if application hit its sndbuf limit recently.
3998 void tcp_cwnd_application_limited(struct sock *sk)
4000 struct tcp_sock *tp = tcp_sk(sk);
4002 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4003 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4004 /* Limited by application or receiver window. */
4005 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4006 u32 win_used = max(tp->snd_cwnd_used, init_win);
4007 if (win_used < tp->snd_cwnd) {
4008 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4009 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4011 tp->snd_cwnd_used = 0;
4013 tp->snd_cwnd_stamp = tcp_time_stamp;
4016 static int tcp_should_expand_sndbuf(struct sock *sk)
4018 struct tcp_sock *tp = tcp_sk(sk);
4020 /* If the user specified a specific send buffer setting, do
4023 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4026 /* If we are under global TCP memory pressure, do not expand. */
4027 if (tcp_memory_pressure)
4030 /* If we are under soft global TCP memory pressure, do not expand. */
4031 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4034 /* If we filled the congestion window, do not expand. */
4035 if (tp->packets_out >= tp->snd_cwnd)
4041 /* When incoming ACK allowed to free some skb from write_queue,
4042 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4043 * on the exit from tcp input handler.
4045 * PROBLEM: sndbuf expansion does not work well with largesend.
4047 static void tcp_new_space(struct sock *sk)
4049 struct tcp_sock *tp = tcp_sk(sk);
4051 if (tcp_should_expand_sndbuf(sk)) {
4052 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4053 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4054 demanded = max_t(unsigned int, tp->snd_cwnd,
4055 tp->reordering + 1);
4056 sndmem *= 2*demanded;
4057 if (sndmem > sk->sk_sndbuf)
4058 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4059 tp->snd_cwnd_stamp = tcp_time_stamp;
4062 sk->sk_write_space(sk);
4065 static void tcp_check_space(struct sock *sk)
4067 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4068 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4069 if (sk->sk_socket &&
4070 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4075 static inline void tcp_data_snd_check(struct sock *sk)
4077 tcp_push_pending_frames(sk);
4078 tcp_check_space(sk);
4082 * Check if sending an ack is needed.
4084 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4086 struct tcp_sock *tp = tcp_sk(sk);
4088 /* More than one full frame received... */
4089 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4090 /* ... and right edge of window advances far enough.
4091 * (tcp_recvmsg() will send ACK otherwise). Or...
4093 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4094 /* We ACK each frame or... */
4095 tcp_in_quickack_mode(sk) ||
4096 /* We have out of order data. */
4098 skb_peek(&tp->out_of_order_queue))) {
4099 /* Then ack it now */
4102 /* Else, send delayed ack. */
4103 tcp_send_delayed_ack(sk);
4107 static inline void tcp_ack_snd_check(struct sock *sk)
4109 if (!inet_csk_ack_scheduled(sk)) {
4110 /* We sent a data segment already. */
4113 __tcp_ack_snd_check(sk, 1);
4117 * This routine is only called when we have urgent data
4118 * signaled. Its the 'slow' part of tcp_urg. It could be
4119 * moved inline now as tcp_urg is only called from one
4120 * place. We handle URGent data wrong. We have to - as
4121 * BSD still doesn't use the correction from RFC961.
4122 * For 1003.1g we should support a new option TCP_STDURG to permit
4123 * either form (or just set the sysctl tcp_stdurg).
4126 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4128 struct tcp_sock *tp = tcp_sk(sk);
4129 u32 ptr = ntohs(th->urg_ptr);
4131 if (ptr && !sysctl_tcp_stdurg)
4133 ptr += ntohl(th->seq);
4135 /* Ignore urgent data that we've already seen and read. */
4136 if (after(tp->copied_seq, ptr))
4139 /* Do not replay urg ptr.
4141 * NOTE: interesting situation not covered by specs.
4142 * Misbehaving sender may send urg ptr, pointing to segment,
4143 * which we already have in ofo queue. We are not able to fetch
4144 * such data and will stay in TCP_URG_NOTYET until will be eaten
4145 * by recvmsg(). Seems, we are not obliged to handle such wicked
4146 * situations. But it is worth to think about possibility of some
4147 * DoSes using some hypothetical application level deadlock.
4149 if (before(ptr, tp->rcv_nxt))
4152 /* Do we already have a newer (or duplicate) urgent pointer? */
4153 if (tp->urg_data && !after(ptr, tp->urg_seq))
4156 /* Tell the world about our new urgent pointer. */
4159 /* We may be adding urgent data when the last byte read was
4160 * urgent. To do this requires some care. We cannot just ignore
4161 * tp->copied_seq since we would read the last urgent byte again
4162 * as data, nor can we alter copied_seq until this data arrives
4163 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4165 * NOTE. Double Dutch. Rendering to plain English: author of comment
4166 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4167 * and expect that both A and B disappear from stream. This is _wrong_.
4168 * Though this happens in BSD with high probability, this is occasional.
4169 * Any application relying on this is buggy. Note also, that fix "works"
4170 * only in this artificial test. Insert some normal data between A and B and we will
4171 * decline of BSD again. Verdict: it is better to remove to trap
4174 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4175 !sock_flag(sk, SOCK_URGINLINE) &&
4176 tp->copied_seq != tp->rcv_nxt) {
4177 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4179 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4180 __skb_unlink(skb, &sk->sk_receive_queue);
4185 tp->urg_data = TCP_URG_NOTYET;
4188 /* Disable header prediction. */
4192 /* This is the 'fast' part of urgent handling. */
4193 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4195 struct tcp_sock *tp = tcp_sk(sk);
4197 /* Check if we get a new urgent pointer - normally not. */
4199 tcp_check_urg(sk,th);
4201 /* Do we wait for any urgent data? - normally not... */
4202 if (tp->urg_data == TCP_URG_NOTYET) {
4203 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4206 /* Is the urgent pointer pointing into this packet? */
4207 if (ptr < skb->len) {
4209 if (skb_copy_bits(skb, ptr, &tmp, 1))
4211 tp->urg_data = TCP_URG_VALID | tmp;
4212 if (!sock_flag(sk, SOCK_DEAD))
4213 sk->sk_data_ready(sk, 0);
4218 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4220 struct tcp_sock *tp = tcp_sk(sk);
4221 int chunk = skb->len - hlen;
4225 if (skb_csum_unnecessary(skb))
4226 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4228 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4232 tp->ucopy.len -= chunk;
4233 tp->copied_seq += chunk;
4234 tcp_rcv_space_adjust(sk);
4241 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4245 if (sock_owned_by_user(sk)) {
4247 result = __tcp_checksum_complete(skb);
4250 result = __tcp_checksum_complete(skb);
4255 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4257 return !skb_csum_unnecessary(skb) &&
4258 __tcp_checksum_complete_user(sk, skb);
4261 #ifdef CONFIG_NET_DMA
4262 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4264 struct tcp_sock *tp = tcp_sk(sk);
4265 int chunk = skb->len - hlen;
4267 int copied_early = 0;
4269 if (tp->ucopy.wakeup)
4272 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4273 tp->ucopy.dma_chan = get_softnet_dma();
4275 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4277 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4278 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4283 tp->ucopy.dma_cookie = dma_cookie;
4286 tp->ucopy.len -= chunk;
4287 tp->copied_seq += chunk;
4288 tcp_rcv_space_adjust(sk);
4290 if ((tp->ucopy.len == 0) ||
4291 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4292 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4293 tp->ucopy.wakeup = 1;
4294 sk->sk_data_ready(sk, 0);
4296 } else if (chunk > 0) {
4297 tp->ucopy.wakeup = 1;
4298 sk->sk_data_ready(sk, 0);
4301 return copied_early;
4303 #endif /* CONFIG_NET_DMA */
4306 * TCP receive function for the ESTABLISHED state.
4308 * It is split into a fast path and a slow path. The fast path is
4310 * - A zero window was announced from us - zero window probing
4311 * is only handled properly in the slow path.
4312 * - Out of order segments arrived.
4313 * - Urgent data is expected.
4314 * - There is no buffer space left
4315 * - Unexpected TCP flags/window values/header lengths are received
4316 * (detected by checking the TCP header against pred_flags)
4317 * - Data is sent in both directions. Fast path only supports pure senders
4318 * or pure receivers (this means either the sequence number or the ack
4319 * value must stay constant)
4320 * - Unexpected TCP option.
4322 * When these conditions are not satisfied it drops into a standard
4323 * receive procedure patterned after RFC793 to handle all cases.
4324 * The first three cases are guaranteed by proper pred_flags setting,
4325 * the rest is checked inline. Fast processing is turned on in
4326 * tcp_data_queue when everything is OK.
4328 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4329 struct tcphdr *th, unsigned len)
4331 struct tcp_sock *tp = tcp_sk(sk);
4334 * Header prediction.
4335 * The code loosely follows the one in the famous
4336 * "30 instruction TCP receive" Van Jacobson mail.
4338 * Van's trick is to deposit buffers into socket queue
4339 * on a device interrupt, to call tcp_recv function
4340 * on the receive process context and checksum and copy
4341 * the buffer to user space. smart...
4343 * Our current scheme is not silly either but we take the
4344 * extra cost of the net_bh soft interrupt processing...
4345 * We do checksum and copy also but from device to kernel.
4348 tp->rx_opt.saw_tstamp = 0;
4350 /* pred_flags is 0xS?10 << 16 + snd_wnd
4351 * if header_prediction is to be made
4352 * 'S' will always be tp->tcp_header_len >> 2
4353 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4354 * turn it off (when there are holes in the receive
4355 * space for instance)
4356 * PSH flag is ignored.
4359 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4360 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4361 int tcp_header_len = tp->tcp_header_len;
4363 /* Timestamp header prediction: tcp_header_len
4364 * is automatically equal to th->doff*4 due to pred_flags
4368 /* Check timestamp */
4369 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4370 __be32 *ptr = (__be32 *)(th + 1);
4372 /* No? Slow path! */
4373 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4374 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4377 tp->rx_opt.saw_tstamp = 1;
4379 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4381 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4383 /* If PAWS failed, check it more carefully in slow path */
4384 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4387 /* DO NOT update ts_recent here, if checksum fails
4388 * and timestamp was corrupted part, it will result
4389 * in a hung connection since we will drop all
4390 * future packets due to the PAWS test.
4394 if (len <= tcp_header_len) {
4395 /* Bulk data transfer: sender */
4396 if (len == tcp_header_len) {
4397 /* Predicted packet is in window by definition.
4398 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4399 * Hence, check seq<=rcv_wup reduces to:
4401 if (tcp_header_len ==
4402 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4403 tp->rcv_nxt == tp->rcv_wup)
4404 tcp_store_ts_recent(tp);
4406 /* We know that such packets are checksummed
4409 tcp_ack(sk, skb, 0);
4411 tcp_data_snd_check(sk);
4413 } else { /* Header too small */
4414 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4419 int copied_early = 0;
4421 if (tp->copied_seq == tp->rcv_nxt &&
4422 len - tcp_header_len <= tp->ucopy.len) {
4423 #ifdef CONFIG_NET_DMA
4424 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4429 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4430 __set_current_state(TASK_RUNNING);
4432 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4436 /* Predicted packet is in window by definition.
4437 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4438 * Hence, check seq<=rcv_wup reduces to:
4440 if (tcp_header_len ==
4441 (sizeof(struct tcphdr) +
4442 TCPOLEN_TSTAMP_ALIGNED) &&
4443 tp->rcv_nxt == tp->rcv_wup)
4444 tcp_store_ts_recent(tp);
4446 tcp_rcv_rtt_measure_ts(sk, skb);
4448 __skb_pull(skb, tcp_header_len);
4449 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4450 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4453 tcp_cleanup_rbuf(sk, skb->len);
4456 if (tcp_checksum_complete_user(sk, skb))
4459 /* Predicted packet is in window by definition.
4460 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4461 * Hence, check seq<=rcv_wup reduces to:
4463 if (tcp_header_len ==
4464 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4465 tp->rcv_nxt == tp->rcv_wup)
4466 tcp_store_ts_recent(tp);
4468 tcp_rcv_rtt_measure_ts(sk, skb);
4470 if ((int)skb->truesize > sk->sk_forward_alloc)
4473 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4475 /* Bulk data transfer: receiver */
4476 __skb_pull(skb,tcp_header_len);
4477 __skb_queue_tail(&sk->sk_receive_queue, skb);
4478 sk_stream_set_owner_r(skb, sk);
4479 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4482 tcp_event_data_recv(sk, skb);
4484 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4485 /* Well, only one small jumplet in fast path... */
4486 tcp_ack(sk, skb, FLAG_DATA);
4487 tcp_data_snd_check(sk);
4488 if (!inet_csk_ack_scheduled(sk))
4492 __tcp_ack_snd_check(sk, 0);
4494 #ifdef CONFIG_NET_DMA
4496 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4502 sk->sk_data_ready(sk, 0);
4508 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4512 * RFC1323: H1. Apply PAWS check first.
4514 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4515 tcp_paws_discard(sk, skb)) {
4517 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4518 tcp_send_dupack(sk, skb);
4521 /* Resets are accepted even if PAWS failed.
4523 ts_recent update must be made after we are sure
4524 that the packet is in window.
4529 * Standard slow path.
4532 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4533 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4534 * (RST) segments are validated by checking their SEQ-fields."
4535 * And page 69: "If an incoming segment is not acceptable,
4536 * an acknowledgment should be sent in reply (unless the RST bit
4537 * is set, if so drop the segment and return)".
4540 tcp_send_dupack(sk, skb);
4549 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4551 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4552 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4553 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4560 tcp_ack(sk, skb, FLAG_SLOWPATH);
4562 tcp_rcv_rtt_measure_ts(sk, skb);
4564 /* Process urgent data. */
4565 tcp_urg(sk, skb, th);
4567 /* step 7: process the segment text */
4568 tcp_data_queue(sk, skb);
4570 tcp_data_snd_check(sk);
4571 tcp_ack_snd_check(sk);
4575 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4582 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4583 struct tcphdr *th, unsigned len)
4585 struct tcp_sock *tp = tcp_sk(sk);
4586 struct inet_connection_sock *icsk = inet_csk(sk);
4587 int saved_clamp = tp->rx_opt.mss_clamp;
4589 tcp_parse_options(skb, &tp->rx_opt, 0);
4593 * "If the state is SYN-SENT then
4594 * first check the ACK bit
4595 * If the ACK bit is set
4596 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4597 * a reset (unless the RST bit is set, if so drop
4598 * the segment and return)"
4600 * We do not send data with SYN, so that RFC-correct
4603 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4604 goto reset_and_undo;
4606 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4607 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4609 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4610 goto reset_and_undo;
4613 /* Now ACK is acceptable.
4615 * "If the RST bit is set
4616 * If the ACK was acceptable then signal the user "error:
4617 * connection reset", drop the segment, enter CLOSED state,
4618 * delete TCB, and return."
4627 * "fifth, if neither of the SYN or RST bits is set then
4628 * drop the segment and return."
4634 goto discard_and_undo;
4637 * "If the SYN bit is on ...
4638 * are acceptable then ...
4639 * (our SYN has been ACKed), change the connection
4640 * state to ESTABLISHED..."
4643 TCP_ECN_rcv_synack(tp, th);
4645 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4646 tcp_ack(sk, skb, FLAG_SLOWPATH);
4648 /* Ok.. it's good. Set up sequence numbers and
4649 * move to established.
4651 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4652 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4654 /* RFC1323: The window in SYN & SYN/ACK segments is
4657 tp->snd_wnd = ntohs(th->window);
4658 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4660 if (!tp->rx_opt.wscale_ok) {
4661 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4662 tp->window_clamp = min(tp->window_clamp, 65535U);
4665 if (tp->rx_opt.saw_tstamp) {
4666 tp->rx_opt.tstamp_ok = 1;
4667 tp->tcp_header_len =
4668 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4669 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4670 tcp_store_ts_recent(tp);
4672 tp->tcp_header_len = sizeof(struct tcphdr);
4675 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4676 tcp_enable_fack(tp);
4679 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4680 tcp_initialize_rcv_mss(sk);
4682 /* Remember, tcp_poll() does not lock socket!
4683 * Change state from SYN-SENT only after copied_seq
4684 * is initialized. */
4685 tp->copied_seq = tp->rcv_nxt;
4687 tcp_set_state(sk, TCP_ESTABLISHED);
4689 security_inet_conn_established(sk, skb);
4691 /* Make sure socket is routed, for correct metrics. */
4692 icsk->icsk_af_ops->rebuild_header(sk);
4694 tcp_init_metrics(sk);
4696 tcp_init_congestion_control(sk);
4698 /* Prevent spurious tcp_cwnd_restart() on first data
4701 tp->lsndtime = tcp_time_stamp;
4703 tcp_init_buffer_space(sk);
4705 if (sock_flag(sk, SOCK_KEEPOPEN))
4706 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4708 if (!tp->rx_opt.snd_wscale)
4709 __tcp_fast_path_on(tp, tp->snd_wnd);
4713 if (!sock_flag(sk, SOCK_DEAD)) {
4714 sk->sk_state_change(sk);
4715 sk_wake_async(sk, 0, POLL_OUT);
4718 if (sk->sk_write_pending ||
4719 icsk->icsk_accept_queue.rskq_defer_accept ||
4720 icsk->icsk_ack.pingpong) {
4721 /* Save one ACK. Data will be ready after
4722 * several ticks, if write_pending is set.
4724 * It may be deleted, but with this feature tcpdumps
4725 * look so _wonderfully_ clever, that I was not able
4726 * to stand against the temptation 8) --ANK
4728 inet_csk_schedule_ack(sk);
4729 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4730 icsk->icsk_ack.ato = TCP_ATO_MIN;
4731 tcp_incr_quickack(sk);
4732 tcp_enter_quickack_mode(sk);
4733 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4734 TCP_DELACK_MAX, TCP_RTO_MAX);
4745 /* No ACK in the segment */
4749 * "If the RST bit is set
4751 * Otherwise (no ACK) drop the segment and return."
4754 goto discard_and_undo;
4758 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4759 goto discard_and_undo;
4762 /* We see SYN without ACK. It is attempt of
4763 * simultaneous connect with crossed SYNs.
4764 * Particularly, it can be connect to self.
4766 tcp_set_state(sk, TCP_SYN_RECV);
4768 if (tp->rx_opt.saw_tstamp) {
4769 tp->rx_opt.tstamp_ok = 1;
4770 tcp_store_ts_recent(tp);
4771 tp->tcp_header_len =
4772 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4774 tp->tcp_header_len = sizeof(struct tcphdr);
4777 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4778 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4780 /* RFC1323: The window in SYN & SYN/ACK segments is
4783 tp->snd_wnd = ntohs(th->window);
4784 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4785 tp->max_window = tp->snd_wnd;
4787 TCP_ECN_rcv_syn(tp, th);
4790 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4791 tcp_initialize_rcv_mss(sk);
4794 tcp_send_synack(sk);
4796 /* Note, we could accept data and URG from this segment.
4797 * There are no obstacles to make this.
4799 * However, if we ignore data in ACKless segments sometimes,
4800 * we have no reasons to accept it sometimes.
4801 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4802 * is not flawless. So, discard packet for sanity.
4803 * Uncomment this return to process the data.
4810 /* "fifth, if neither of the SYN or RST bits is set then
4811 * drop the segment and return."
4815 tcp_clear_options(&tp->rx_opt);
4816 tp->rx_opt.mss_clamp = saved_clamp;
4820 tcp_clear_options(&tp->rx_opt);
4821 tp->rx_opt.mss_clamp = saved_clamp;
4827 * This function implements the receiving procedure of RFC 793 for
4828 * all states except ESTABLISHED and TIME_WAIT.
4829 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4830 * address independent.
4833 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4834 struct tcphdr *th, unsigned len)
4836 struct tcp_sock *tp = tcp_sk(sk);
4837 struct inet_connection_sock *icsk = inet_csk(sk);
4840 tp->rx_opt.saw_tstamp = 0;
4842 switch (sk->sk_state) {
4854 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4857 /* Now we have several options: In theory there is
4858 * nothing else in the frame. KA9Q has an option to
4859 * send data with the syn, BSD accepts data with the
4860 * syn up to the [to be] advertised window and
4861 * Solaris 2.1 gives you a protocol error. For now
4862 * we just ignore it, that fits the spec precisely
4863 * and avoids incompatibilities. It would be nice in
4864 * future to drop through and process the data.
4866 * Now that TTCP is starting to be used we ought to
4868 * But, this leaves one open to an easy denial of
4869 * service attack, and SYN cookies can't defend
4870 * against this problem. So, we drop the data
4871 * in the interest of security over speed unless
4872 * it's still in use.
4880 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4884 /* Do step6 onward by hand. */
4885 tcp_urg(sk, skb, th);
4887 tcp_data_snd_check(sk);
4891 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4892 tcp_paws_discard(sk, skb)) {
4894 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4895 tcp_send_dupack(sk, skb);
4898 /* Reset is accepted even if it did not pass PAWS. */
4901 /* step 1: check sequence number */
4902 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4904 tcp_send_dupack(sk, skb);
4908 /* step 2: check RST bit */
4914 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4916 /* step 3: check security and precedence [ignored] */
4920 * Check for a SYN in window.
4922 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4923 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4928 /* step 5: check the ACK field */
4930 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4932 switch (sk->sk_state) {
4935 tp->copied_seq = tp->rcv_nxt;
4937 tcp_set_state(sk, TCP_ESTABLISHED);
4938 sk->sk_state_change(sk);
4940 /* Note, that this wakeup is only for marginal
4941 * crossed SYN case. Passively open sockets
4942 * are not waked up, because sk->sk_sleep ==
4943 * NULL and sk->sk_socket == NULL.
4945 if (sk->sk_socket) {
4946 sk_wake_async(sk,0,POLL_OUT);
4949 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4950 tp->snd_wnd = ntohs(th->window) <<
4951 tp->rx_opt.snd_wscale;
4952 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4953 TCP_SKB_CB(skb)->seq);
4955 /* tcp_ack considers this ACK as duplicate
4956 * and does not calculate rtt.
4957 * Fix it at least with timestamps.
4959 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4961 tcp_ack_saw_tstamp(sk, 0);
4963 if (tp->rx_opt.tstamp_ok)
4964 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4966 /* Make sure socket is routed, for
4969 icsk->icsk_af_ops->rebuild_header(sk);
4971 tcp_init_metrics(sk);
4973 tcp_init_congestion_control(sk);
4975 /* Prevent spurious tcp_cwnd_restart() on
4976 * first data packet.
4978 tp->lsndtime = tcp_time_stamp;
4981 tcp_initialize_rcv_mss(sk);
4982 tcp_init_buffer_space(sk);
4983 tcp_fast_path_on(tp);
4990 if (tp->snd_una == tp->write_seq) {
4991 tcp_set_state(sk, TCP_FIN_WAIT2);
4992 sk->sk_shutdown |= SEND_SHUTDOWN;
4993 dst_confirm(sk->sk_dst_cache);
4995 if (!sock_flag(sk, SOCK_DEAD))
4996 /* Wake up lingering close() */
4997 sk->sk_state_change(sk);
5001 if (tp->linger2 < 0 ||
5002 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5003 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5005 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5009 tmo = tcp_fin_time(sk);
5010 if (tmo > TCP_TIMEWAIT_LEN) {
5011 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5012 } else if (th->fin || sock_owned_by_user(sk)) {
5013 /* Bad case. We could lose such FIN otherwise.
5014 * It is not a big problem, but it looks confusing
5015 * and not so rare event. We still can lose it now,
5016 * if it spins in bh_lock_sock(), but it is really
5019 inet_csk_reset_keepalive_timer(sk, tmo);
5021 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5029 if (tp->snd_una == tp->write_seq) {
5030 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5036 if (tp->snd_una == tp->write_seq) {
5037 tcp_update_metrics(sk);
5046 /* step 6: check the URG bit */
5047 tcp_urg(sk, skb, th);
5049 /* step 7: process the segment text */
5050 switch (sk->sk_state) {
5051 case TCP_CLOSE_WAIT:
5054 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5058 /* RFC 793 says to queue data in these states,
5059 * RFC 1122 says we MUST send a reset.
5060 * BSD 4.4 also does reset.
5062 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5063 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5064 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5065 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5071 case TCP_ESTABLISHED:
5072 tcp_data_queue(sk, skb);
5077 /* tcp_data could move socket to TIME-WAIT */
5078 if (sk->sk_state != TCP_CLOSE) {
5079 tcp_data_snd_check(sk);
5080 tcp_ack_snd_check(sk);
5090 EXPORT_SYMBOL(sysctl_tcp_ecn);
5091 EXPORT_SYMBOL(sysctl_tcp_reordering);
5092 EXPORT_SYMBOL(tcp_parse_options);
5093 EXPORT_SYMBOL(tcp_rcv_established);
5094 EXPORT_SYMBOL(tcp_rcv_state_process);
5095 EXPORT_SYMBOL(tcp_initialize_rcv_mss);