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 = 2;
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 */
107 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
120 /* Adapt the MSS value used to make delayed ack decision to the
123 static void tcp_measure_rcv_mss(struct sock *sk,
124 const struct sk_buff *skb)
126 struct inet_connection_sock *icsk = inet_csk(sk);
127 const unsigned int lss = icsk->icsk_ack.last_seg_size;
130 icsk->icsk_ack.last_seg_size = 0;
132 /* skb->len may jitter because of SACKs, even if peer
133 * sends good full-sized frames.
135 len = skb_shinfo(skb)->gso_size ?: skb->len;
136 if (len >= icsk->icsk_ack.rcv_mss) {
137 icsk->icsk_ack.rcv_mss = len;
139 /* Otherwise, we make more careful check taking into account,
140 * that SACKs block is variable.
142 * "len" is invariant segment length, including TCP header.
144 len += skb->data - skb_transport_header(skb);
145 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
146 /* If PSH is not set, packet should be
147 * full sized, provided peer TCP is not badly broken.
148 * This observation (if it is correct 8)) allows
149 * to handle super-low mtu links fairly.
151 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
152 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
153 /* Subtract also invariant (if peer is RFC compliant),
154 * tcp header plus fixed timestamp option length.
155 * Resulting "len" is MSS free of SACK jitter.
157 len -= tcp_sk(sk)->tcp_header_len;
158 icsk->icsk_ack.last_seg_size = len;
160 icsk->icsk_ack.rcv_mss = len;
164 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
166 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
170 static void tcp_incr_quickack(struct sock *sk)
172 struct inet_connection_sock *icsk = inet_csk(sk);
173 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
177 if (quickacks > icsk->icsk_ack.quick)
178 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
181 void tcp_enter_quickack_mode(struct sock *sk)
183 struct inet_connection_sock *icsk = inet_csk(sk);
184 tcp_incr_quickack(sk);
185 icsk->icsk_ack.pingpong = 0;
186 icsk->icsk_ack.ato = TCP_ATO_MIN;
189 /* Send ACKs quickly, if "quick" count is not exhausted
190 * and the session is not interactive.
193 static inline int tcp_in_quickack_mode(const struct sock *sk)
195 const struct inet_connection_sock *icsk = inet_csk(sk);
196 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
201 if (tp->ecn_flags&TCP_ECN_OK)
202 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
207 if (tcp_hdr(skb)->cwr)
208 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
218 if (tp->ecn_flags&TCP_ECN_OK) {
219 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
220 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
221 /* Funny extension: if ECT is not set on a segment,
222 * it is surely retransmit. It is not in ECN RFC,
223 * but Linux follows this rule. */
224 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
225 tcp_enter_quickack_mode((struct sock *)tp);
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
231 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
232 tp->ecn_flags &= ~TCP_ECN_OK;
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
237 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
238 tp->ecn_flags &= ~TCP_ECN_OK;
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
243 if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
248 /* Buffer size and advertised window tuning.
250 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
253 static void tcp_fixup_sndbuf(struct sock *sk)
255 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
256 sizeof(struct sk_buff);
258 if (sk->sk_sndbuf < 3 * sndmem)
259 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
264 * All tcp_full_space() is split to two parts: "network" buffer, allocated
265 * forward and advertised in receiver window (tp->rcv_wnd) and
266 * "application buffer", required to isolate scheduling/application
267 * latencies from network.
268 * window_clamp is maximal advertised window. It can be less than
269 * tcp_full_space(), in this case tcp_full_space() - window_clamp
270 * is reserved for "application" buffer. The less window_clamp is
271 * the smoother our behaviour from viewpoint of network, but the lower
272 * throughput and the higher sensitivity of the connection to losses. 8)
274 * rcv_ssthresh is more strict window_clamp used at "slow start"
275 * phase to predict further behaviour of this connection.
276 * It is used for two goals:
277 * - to enforce header prediction at sender, even when application
278 * requires some significant "application buffer". It is check #1.
279 * - to prevent pruning of receive queue because of misprediction
280 * of receiver window. Check #2.
282 * The scheme does not work when sender sends good segments opening
283 * window and then starts to feed us spaghetti. But it should work
284 * in common situations. Otherwise, we have to rely on queue collapsing.
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
290 struct tcp_sock *tp = tcp_sk(sk);
292 int truesize = tcp_win_from_space(skb->truesize)/2;
293 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
295 while (tp->rcv_ssthresh <= window) {
296 if (truesize <= skb->len)
297 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
305 static void tcp_grow_window(struct sock *sk,
308 struct tcp_sock *tp = tcp_sk(sk);
311 if (tp->rcv_ssthresh < tp->window_clamp &&
312 (int)tp->rcv_ssthresh < tcp_space(sk) &&
313 !tcp_memory_pressure) {
316 /* Check #2. Increase window, if skb with such overhead
317 * will fit to rcvbuf in future.
319 if (tcp_win_from_space(skb->truesize) <= skb->len)
322 incr = __tcp_grow_window(sk, skb);
325 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
326 inet_csk(sk)->icsk_ack.quick |= 1;
331 /* 3. Tuning rcvbuf, when connection enters established state. */
333 static void tcp_fixup_rcvbuf(struct sock *sk)
335 struct tcp_sock *tp = tcp_sk(sk);
336 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
338 /* Try to select rcvbuf so that 4 mss-sized segments
339 * will fit to window and corresponding skbs will fit to our rcvbuf.
340 * (was 3; 4 is minimum to allow fast retransmit to work.)
342 while (tcp_win_from_space(rcvmem) < tp->advmss)
344 if (sk->sk_rcvbuf < 4 * rcvmem)
345 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
348 /* 4. Try to fixup all. It is made immediately after connection enters
351 static void tcp_init_buffer_space(struct sock *sk)
353 struct tcp_sock *tp = tcp_sk(sk);
356 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
357 tcp_fixup_rcvbuf(sk);
358 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
359 tcp_fixup_sndbuf(sk);
361 tp->rcvq_space.space = tp->rcv_wnd;
363 maxwin = tcp_full_space(sk);
365 if (tp->window_clamp >= maxwin) {
366 tp->window_clamp = maxwin;
368 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
369 tp->window_clamp = max(maxwin -
370 (maxwin >> sysctl_tcp_app_win),
374 /* Force reservation of one segment. */
375 if (sysctl_tcp_app_win &&
376 tp->window_clamp > 2 * tp->advmss &&
377 tp->window_clamp + tp->advmss > maxwin)
378 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
380 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
381 tp->snd_cwnd_stamp = tcp_time_stamp;
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock *sk)
387 struct tcp_sock *tp = tcp_sk(sk);
388 struct inet_connection_sock *icsk = inet_csk(sk);
390 icsk->icsk_ack.quick = 0;
392 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
393 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
394 !tcp_memory_pressure &&
395 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
396 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
399 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
400 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
404 /* Initialize RCV_MSS value.
405 * RCV_MSS is an our guess about MSS used by the peer.
406 * We haven't any direct information about the MSS.
407 * It's better to underestimate the RCV_MSS rather than overestimate.
408 * Overestimations make us ACKing less frequently than needed.
409 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
411 void tcp_initialize_rcv_mss(struct sock *sk)
413 struct tcp_sock *tp = tcp_sk(sk);
414 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
416 hint = min(hint, tp->rcv_wnd/2);
417 hint = min(hint, TCP_MIN_RCVMSS);
418 hint = max(hint, TCP_MIN_MSS);
420 inet_csk(sk)->icsk_ack.rcv_mss = hint;
423 /* Receiver "autotuning" code.
425 * The algorithm for RTT estimation w/o timestamps is based on
426 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
427 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
429 * More detail on this code can be found at
430 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
431 * though this reference is out of date. A new paper
434 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
436 u32 new_sample = tp->rcv_rtt_est.rtt;
442 if (new_sample != 0) {
443 /* If we sample in larger samples in the non-timestamp
444 * case, we could grossly overestimate the RTT especially
445 * with chatty applications or bulk transfer apps which
446 * are stalled on filesystem I/O.
448 * Also, since we are only going for a minimum in the
449 * non-timestamp case, we do not smooth things out
450 * else with timestamps disabled convergence takes too
454 m -= (new_sample >> 3);
456 } else if (m < new_sample)
459 /* No previous measure. */
463 if (tp->rcv_rtt_est.rtt != new_sample)
464 tp->rcv_rtt_est.rtt = new_sample;
467 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
469 if (tp->rcv_rtt_est.time == 0)
471 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
473 tcp_rcv_rtt_update(tp,
474 jiffies - tp->rcv_rtt_est.time,
478 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
479 tp->rcv_rtt_est.time = tcp_time_stamp;
482 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
484 struct tcp_sock *tp = tcp_sk(sk);
485 if (tp->rx_opt.rcv_tsecr &&
486 (TCP_SKB_CB(skb)->end_seq -
487 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
488 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
492 * This function should be called every time data is copied to user space.
493 * It calculates the appropriate TCP receive buffer space.
495 void tcp_rcv_space_adjust(struct sock *sk)
497 struct tcp_sock *tp = tcp_sk(sk);
501 if (tp->rcvq_space.time == 0)
504 time = tcp_time_stamp - tp->rcvq_space.time;
505 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
506 tp->rcv_rtt_est.rtt == 0)
509 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
511 space = max(tp->rcvq_space.space, space);
513 if (tp->rcvq_space.space != space) {
516 tp->rcvq_space.space = space;
518 if (sysctl_tcp_moderate_rcvbuf &&
519 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
520 int new_clamp = space;
522 /* Receive space grows, normalize in order to
523 * take into account packet headers and sk_buff
524 * structure overhead.
529 rcvmem = (tp->advmss + MAX_TCP_HEADER +
530 16 + sizeof(struct sk_buff));
531 while (tcp_win_from_space(rcvmem) < tp->advmss)
534 space = min(space, sysctl_tcp_rmem[2]);
535 if (space > sk->sk_rcvbuf) {
536 sk->sk_rcvbuf = space;
538 /* Make the window clamp follow along. */
539 tp->window_clamp = new_clamp;
545 tp->rcvq_space.seq = tp->copied_seq;
546 tp->rcvq_space.time = tcp_time_stamp;
549 /* There is something which you must keep in mind when you analyze the
550 * behavior of the tp->ato delayed ack timeout interval. When a
551 * connection starts up, we want to ack as quickly as possible. The
552 * problem is that "good" TCP's do slow start at the beginning of data
553 * transmission. The means that until we send the first few ACK's the
554 * sender will sit on his end and only queue most of his data, because
555 * he can only send snd_cwnd unacked packets at any given time. For
556 * each ACK we send, he increments snd_cwnd and transmits more of his
559 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
561 struct tcp_sock *tp = tcp_sk(sk);
562 struct inet_connection_sock *icsk = inet_csk(sk);
565 inet_csk_schedule_ack(sk);
567 tcp_measure_rcv_mss(sk, skb);
569 tcp_rcv_rtt_measure(tp);
571 now = tcp_time_stamp;
573 if (!icsk->icsk_ack.ato) {
574 /* The _first_ data packet received, initialize
575 * delayed ACK engine.
577 tcp_incr_quickack(sk);
578 icsk->icsk_ack.ato = TCP_ATO_MIN;
580 int m = now - icsk->icsk_ack.lrcvtime;
582 if (m <= TCP_ATO_MIN/2) {
583 /* The fastest case is the first. */
584 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
585 } else if (m < icsk->icsk_ack.ato) {
586 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
587 if (icsk->icsk_ack.ato > icsk->icsk_rto)
588 icsk->icsk_ack.ato = icsk->icsk_rto;
589 } else if (m > icsk->icsk_rto) {
590 /* Too long gap. Apparently sender failed to
591 * restart window, so that we send ACKs quickly.
593 tcp_incr_quickack(sk);
594 sk_stream_mem_reclaim(sk);
597 icsk->icsk_ack.lrcvtime = now;
599 TCP_ECN_check_ce(tp, skb);
602 tcp_grow_window(sk, skb);
605 static u32 tcp_rto_min(struct sock *sk)
607 struct dst_entry *dst = __sk_dst_get(sk);
608 u32 rto_min = TCP_RTO_MIN;
610 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
611 rto_min = dst->metrics[RTAX_RTO_MIN-1];
615 /* Called to compute a smoothed rtt estimate. The data fed to this
616 * routine either comes from timestamps, or from segments that were
617 * known _not_ to have been retransmitted [see Karn/Partridge
618 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
619 * piece by Van Jacobson.
620 * NOTE: the next three routines used to be one big routine.
621 * To save cycles in the RFC 1323 implementation it was better to break
622 * it up into three procedures. -- erics
624 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
626 struct tcp_sock *tp = tcp_sk(sk);
627 long m = mrtt; /* RTT */
629 /* The following amusing code comes from Jacobson's
630 * article in SIGCOMM '88. Note that rtt and mdev
631 * are scaled versions of rtt and mean deviation.
632 * This is designed to be as fast as possible
633 * m stands for "measurement".
635 * On a 1990 paper the rto value is changed to:
636 * RTO = rtt + 4 * mdev
638 * Funny. This algorithm seems to be very broken.
639 * These formulae increase RTO, when it should be decreased, increase
640 * too slowly, when it should be increased quickly, decrease too quickly
641 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
642 * does not matter how to _calculate_ it. Seems, it was trap
643 * that VJ failed to avoid. 8)
648 m -= (tp->srtt >> 3); /* m is now error in rtt est */
649 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
651 m = -m; /* m is now abs(error) */
652 m -= (tp->mdev >> 2); /* similar update on mdev */
653 /* This is similar to one of Eifel findings.
654 * Eifel blocks mdev updates when rtt decreases.
655 * This solution is a bit different: we use finer gain
656 * for mdev in this case (alpha*beta).
657 * Like Eifel it also prevents growth of rto,
658 * but also it limits too fast rto decreases,
659 * happening in pure Eifel.
664 m -= (tp->mdev >> 2); /* similar update on mdev */
666 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
667 if (tp->mdev > tp->mdev_max) {
668 tp->mdev_max = tp->mdev;
669 if (tp->mdev_max > tp->rttvar)
670 tp->rttvar = tp->mdev_max;
672 if (after(tp->snd_una, tp->rtt_seq)) {
673 if (tp->mdev_max < tp->rttvar)
674 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
675 tp->rtt_seq = tp->snd_nxt;
676 tp->mdev_max = tcp_rto_min(sk);
679 /* no previous measure. */
680 tp->srtt = m<<3; /* take the measured time to be rtt */
681 tp->mdev = m<<1; /* make sure rto = 3*rtt */
682 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
683 tp->rtt_seq = tp->snd_nxt;
687 /* Calculate rto without backoff. This is the second half of Van Jacobson's
688 * routine referred to above.
690 static inline void tcp_set_rto(struct sock *sk)
692 const struct tcp_sock *tp = tcp_sk(sk);
693 /* Old crap is replaced with new one. 8)
696 * 1. If rtt variance happened to be less 50msec, it is hallucination.
697 * It cannot be less due to utterly erratic ACK generation made
698 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
699 * to do with delayed acks, because at cwnd>2 true delack timeout
700 * is invisible. Actually, Linux-2.4 also generates erratic
701 * ACKs in some circumstances.
703 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
705 /* 2. Fixups made earlier cannot be right.
706 * If we do not estimate RTO correctly without them,
707 * all the algo is pure shit and should be replaced
708 * with correct one. It is exactly, which we pretend to do.
712 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
713 * guarantees that rto is higher.
715 static inline void tcp_bound_rto(struct sock *sk)
717 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
718 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
721 /* Save metrics learned by this TCP session.
722 This function is called only, when TCP finishes successfully
723 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
725 void tcp_update_metrics(struct sock *sk)
727 struct tcp_sock *tp = tcp_sk(sk);
728 struct dst_entry *dst = __sk_dst_get(sk);
730 if (sysctl_tcp_nometrics_save)
735 if (dst && (dst->flags&DST_HOST)) {
736 const struct inet_connection_sock *icsk = inet_csk(sk);
739 if (icsk->icsk_backoff || !tp->srtt) {
740 /* This session failed to estimate rtt. Why?
741 * Probably, no packets returned in time.
744 if (!(dst_metric_locked(dst, RTAX_RTT)))
745 dst->metrics[RTAX_RTT-1] = 0;
749 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
751 /* If newly calculated rtt larger than stored one,
752 * store new one. Otherwise, use EWMA. Remember,
753 * rtt overestimation is always better than underestimation.
755 if (!(dst_metric_locked(dst, RTAX_RTT))) {
757 dst->metrics[RTAX_RTT-1] = tp->srtt;
759 dst->metrics[RTAX_RTT-1] -= (m>>3);
762 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
766 /* Scale deviation to rttvar fixed point */
771 if (m >= dst_metric(dst, RTAX_RTTVAR))
772 dst->metrics[RTAX_RTTVAR-1] = m;
774 dst->metrics[RTAX_RTTVAR-1] -=
775 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
778 if (tp->snd_ssthresh >= 0xFFFF) {
779 /* Slow start still did not finish. */
780 if (dst_metric(dst, RTAX_SSTHRESH) &&
781 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
782 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
783 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
784 if (!dst_metric_locked(dst, RTAX_CWND) &&
785 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
786 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
787 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
788 icsk->icsk_ca_state == TCP_CA_Open) {
789 /* Cong. avoidance phase, cwnd is reliable. */
790 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
791 dst->metrics[RTAX_SSTHRESH-1] =
792 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
793 if (!dst_metric_locked(dst, RTAX_CWND))
794 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
796 /* Else slow start did not finish, cwnd is non-sense,
797 ssthresh may be also invalid.
799 if (!dst_metric_locked(dst, RTAX_CWND))
800 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
801 if (dst->metrics[RTAX_SSTHRESH-1] &&
802 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
803 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
804 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
807 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
808 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
809 tp->reordering != sysctl_tcp_reordering)
810 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
815 /* Numbers are taken from RFC3390.
817 * John Heffner states:
819 * The RFC specifies a window of no more than 4380 bytes
820 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
821 * is a bit misleading because they use a clamp at 4380 bytes
822 * rather than use a multiplier in the relevant range.
824 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
826 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
829 if (tp->mss_cache > 1460)
832 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
834 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
837 /* Set slow start threshold and cwnd not falling to slow start */
838 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
840 struct tcp_sock *tp = tcp_sk(sk);
841 const struct inet_connection_sock *icsk = inet_csk(sk);
843 tp->prior_ssthresh = 0;
845 if (icsk->icsk_ca_state < TCP_CA_CWR) {
848 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
849 tp->snd_cwnd = min(tp->snd_cwnd,
850 tcp_packets_in_flight(tp) + 1U);
851 tp->snd_cwnd_cnt = 0;
852 tp->high_seq = tp->snd_nxt;
853 tp->snd_cwnd_stamp = tcp_time_stamp;
854 TCP_ECN_queue_cwr(tp);
856 tcp_set_ca_state(sk, TCP_CA_CWR);
861 * Packet counting of FACK is based on in-order assumptions, therefore TCP
862 * disables it when reordering is detected
864 static void tcp_disable_fack(struct tcp_sock *tp)
866 tp->rx_opt.sack_ok &= ~2;
869 /* Take a notice that peer is sending DSACKs */
870 static void tcp_dsack_seen(struct tcp_sock *tp)
872 tp->rx_opt.sack_ok |= 4;
875 /* Initialize metrics on socket. */
877 static void tcp_init_metrics(struct sock *sk)
879 struct tcp_sock *tp = tcp_sk(sk);
880 struct dst_entry *dst = __sk_dst_get(sk);
887 if (dst_metric_locked(dst, RTAX_CWND))
888 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
889 if (dst_metric(dst, RTAX_SSTHRESH)) {
890 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
891 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
892 tp->snd_ssthresh = tp->snd_cwnd_clamp;
894 if (dst_metric(dst, RTAX_REORDERING) &&
895 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
896 tcp_disable_fack(tp);
897 tp->reordering = dst_metric(dst, RTAX_REORDERING);
900 if (dst_metric(dst, RTAX_RTT) == 0)
903 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
906 /* Initial rtt is determined from SYN,SYN-ACK.
907 * The segment is small and rtt may appear much
908 * less than real one. Use per-dst memory
909 * to make it more realistic.
911 * A bit of theory. RTT is time passed after "normal" sized packet
912 * is sent until it is ACKed. In normal circumstances sending small
913 * packets force peer to delay ACKs and calculation is correct too.
914 * The algorithm is adaptive and, provided we follow specs, it
915 * NEVER underestimate RTT. BUT! If peer tries to make some clever
916 * tricks sort of "quick acks" for time long enough to decrease RTT
917 * to low value, and then abruptly stops to do it and starts to delay
918 * ACKs, wait for troubles.
920 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
921 tp->srtt = dst_metric(dst, RTAX_RTT);
922 tp->rtt_seq = tp->snd_nxt;
924 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
925 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
926 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
930 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
932 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
933 tp->snd_cwnd_stamp = tcp_time_stamp;
937 /* Play conservative. If timestamps are not
938 * supported, TCP will fail to recalculate correct
939 * rtt, if initial rto is too small. FORGET ALL AND RESET!
941 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
943 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
944 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
948 static void tcp_update_reordering(struct sock *sk, const int metric,
951 struct tcp_sock *tp = tcp_sk(sk);
952 if (metric > tp->reordering) {
953 tp->reordering = min(TCP_MAX_REORDERING, metric);
955 /* This exciting event is worth to be remembered. 8) */
957 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
958 else if (tcp_is_reno(tp))
959 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
960 else if (tcp_is_fack(tp))
961 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
963 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
964 #if FASTRETRANS_DEBUG > 1
965 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
966 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
970 tp->undo_marker ? tp->undo_retrans : 0);
972 tcp_disable_fack(tp);
976 /* This procedure tags the retransmission queue when SACKs arrive.
978 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
979 * Packets in queue with these bits set are counted in variables
980 * sacked_out, retrans_out and lost_out, correspondingly.
982 * Valid combinations are:
983 * Tag InFlight Description
984 * 0 1 - orig segment is in flight.
985 * S 0 - nothing flies, orig reached receiver.
986 * L 0 - nothing flies, orig lost by net.
987 * R 2 - both orig and retransmit are in flight.
988 * L|R 1 - orig is lost, retransmit is in flight.
989 * S|R 1 - orig reached receiver, retrans is still in flight.
990 * (L|S|R is logically valid, it could occur when L|R is sacked,
991 * but it is equivalent to plain S and code short-curcuits it to S.
992 * L|S is logically invalid, it would mean -1 packet in flight 8))
994 * These 6 states form finite state machine, controlled by the following events:
995 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
996 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
997 * 3. Loss detection event of one of three flavors:
998 * A. Scoreboard estimator decided the packet is lost.
999 * A'. Reno "three dupacks" marks head of queue lost.
1000 * A''. Its FACK modfication, head until snd.fack is lost.
1001 * B. SACK arrives sacking data transmitted after never retransmitted
1002 * hole was sent out.
1003 * C. SACK arrives sacking SND.NXT at the moment, when the
1004 * segment was retransmitted.
1005 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1007 * It is pleasant to note, that state diagram turns out to be commutative,
1008 * so that we are allowed not to be bothered by order of our actions,
1009 * when multiple events arrive simultaneously. (see the function below).
1011 * Reordering detection.
1012 * --------------------
1013 * Reordering metric is maximal distance, which a packet can be displaced
1014 * in packet stream. With SACKs we can estimate it:
1016 * 1. SACK fills old hole and the corresponding segment was not
1017 * ever retransmitted -> reordering. Alas, we cannot use it
1018 * when segment was retransmitted.
1019 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1020 * for retransmitted and already SACKed segment -> reordering..
1021 * Both of these heuristics are not used in Loss state, when we cannot
1022 * account for retransmits accurately.
1024 * SACK block validation.
1025 * ----------------------
1027 * SACK block range validation checks that the received SACK block fits to
1028 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1029 * Note that SND.UNA is not included to the range though being valid because
1030 * it means that the receiver is rather inconsistent with itself reporting
1031 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1032 * perfectly valid, however, in light of RFC2018 which explicitly states
1033 * that "SACK block MUST reflect the newest segment. Even if the newest
1034 * segment is going to be discarded ...", not that it looks very clever
1035 * in case of head skb. Due to potentional receiver driven attacks, we
1036 * choose to avoid immediate execution of a walk in write queue due to
1037 * reneging and defer head skb's loss recovery to standard loss recovery
1038 * procedure that will eventually trigger (nothing forbids us doing this).
1040 * Implements also blockage to start_seq wrap-around. Problem lies in the
1041 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1042 * there's no guarantee that it will be before snd_nxt (n). The problem
1043 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1046 * <- outs wnd -> <- wrapzone ->
1047 * u e n u_w e_w s n_w
1049 * |<------------+------+----- TCP seqno space --------------+---------->|
1050 * ...-- <2^31 ->| |<--------...
1051 * ...---- >2^31 ------>| |<--------...
1053 * Current code wouldn't be vulnerable but it's better still to discard such
1054 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1055 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1056 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1057 * equal to the ideal case (infinite seqno space without wrap caused issues).
1059 * With D-SACK the lower bound is extended to cover sequence space below
1060 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1061 * again, DSACK block must not to go across snd_una (for the same reason as
1062 * for the normal SACK blocks, explained above). But there all simplicity
1063 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1064 * fully below undo_marker they do not affect behavior in anyway and can
1065 * therefore be safely ignored. In rare cases (which are more or less
1066 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1067 * fragmentation and packet reordering past skb's retransmission. To consider
1068 * them correctly, the acceptable range must be extended even more though
1069 * the exact amount is rather hard to quantify. However, tp->max_window can
1070 * be used as an exaggerated estimate.
1072 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1073 u32 start_seq, u32 end_seq)
1075 /* Too far in future, or reversed (interpretation is ambiguous) */
1076 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1079 /* Nasty start_seq wrap-around check (see comments above) */
1080 if (!before(start_seq, tp->snd_nxt))
1083 /* In outstanding window? ...This is valid exit for DSACKs too.
1084 * start_seq == snd_una is non-sensical (see comments above)
1086 if (after(start_seq, tp->snd_una))
1089 if (!is_dsack || !tp->undo_marker)
1092 /* ...Then it's D-SACK, and must reside below snd_una completely */
1093 if (!after(end_seq, tp->snd_una))
1096 if (!before(start_seq, tp->undo_marker))
1100 if (!after(end_seq, tp->undo_marker))
1103 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1104 * start_seq < undo_marker and end_seq >= undo_marker.
1106 return !before(start_seq, end_seq - tp->max_window);
1109 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1110 * Event "C". Later note: FACK people cheated me again 8), we have to account
1111 * for reordering! Ugly, but should help.
1113 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1114 * less than what is now known to be received by the other end (derived from
1115 * SACK blocks by the caller).
1117 static int tcp_mark_lost_retrans(struct sock *sk, u32 received_upto)
1119 struct tcp_sock *tp = tcp_sk(sk);
1120 struct sk_buff *skb;
1124 tcp_for_write_queue(skb, sk) {
1125 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1127 if (skb == tcp_send_head(sk))
1129 if (cnt == tp->retrans_out)
1131 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1134 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1137 if (after(received_upto, ack_seq) &&
1139 !before(received_upto,
1140 ack_seq + tp->reordering * tp->mss_cache))) {
1141 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1142 tp->retrans_out -= tcp_skb_pcount(skb);
1144 /* clear lost hint */
1145 tp->retransmit_skb_hint = NULL;
1147 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1148 tp->lost_out += tcp_skb_pcount(skb);
1149 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1150 flag |= FLAG_DATA_SACKED;
1151 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1154 cnt += tcp_skb_pcount(skb);
1160 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1161 struct tcp_sack_block_wire *sp, int num_sacks,
1164 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1165 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1168 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1171 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1172 } else if (num_sacks > 1) {
1173 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1174 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1176 if (!after(end_seq_0, end_seq_1) &&
1177 !before(start_seq_0, start_seq_1)) {
1180 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1184 /* D-SACK for already forgotten data... Do dumb counting. */
1186 !after(end_seq_0, prior_snd_una) &&
1187 after(end_seq_0, tp->undo_marker))
1193 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1194 * the incoming SACK may not exactly match but we can find smaller MSS
1195 * aligned portion of it that matches. Therefore we might need to fragment
1196 * which may fail and creates some hassle (caller must handle error case
1199 int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1200 u32 start_seq, u32 end_seq)
1203 unsigned int pkt_len;
1205 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1206 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1208 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1209 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1211 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1214 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1216 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1217 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1226 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1228 const struct inet_connection_sock *icsk = inet_csk(sk);
1229 struct tcp_sock *tp = tcp_sk(sk);
1230 unsigned char *ptr = (skb_transport_header(ack_skb) +
1231 TCP_SKB_CB(ack_skb)->sacked);
1232 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1233 struct sk_buff *cached_skb;
1234 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1235 int reord = tp->packets_out;
1237 u32 highest_sack_end_seq = 0;
1239 int found_dup_sack = 0;
1240 int cached_fack_count;
1242 int first_sack_index;
1244 if (!tp->sacked_out) {
1245 if (WARN_ON(tp->fackets_out))
1246 tp->fackets_out = 0;
1247 tp->highest_sack = tp->snd_una;
1249 prior_fackets = tp->fackets_out;
1251 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1252 num_sacks, prior_snd_una);
1254 flag |= FLAG_DSACKING_ACK;
1256 /* Eliminate too old ACKs, but take into
1257 * account more or less fresh ones, they can
1258 * contain valid SACK info.
1260 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1264 * if the only SACK change is the increase of the end_seq of
1265 * the first block then only apply that SACK block
1266 * and use retrans queue hinting otherwise slowpath */
1268 for (i = 0; i < num_sacks; i++) {
1269 __be32 start_seq = sp[i].start_seq;
1270 __be32 end_seq = sp[i].end_seq;
1273 if (tp->recv_sack_cache[i].start_seq != start_seq)
1276 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1277 (tp->recv_sack_cache[i].end_seq != end_seq))
1280 tp->recv_sack_cache[i].start_seq = start_seq;
1281 tp->recv_sack_cache[i].end_seq = end_seq;
1283 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1284 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1285 tp->recv_sack_cache[i].start_seq = 0;
1286 tp->recv_sack_cache[i].end_seq = 0;
1289 first_sack_index = 0;
1294 tp->fastpath_skb_hint = NULL;
1296 /* order SACK blocks to allow in order walk of the retrans queue */
1297 for (i = num_sacks-1; i > 0; i--) {
1298 for (j = 0; j < i; j++){
1299 if (after(ntohl(sp[j].start_seq),
1300 ntohl(sp[j+1].start_seq))){
1301 struct tcp_sack_block_wire tmp;
1307 /* Track where the first SACK block goes to */
1308 if (j == first_sack_index)
1309 first_sack_index = j+1;
1316 /* clear flag as used for different purpose in following code */
1319 /* Use SACK fastpath hint if valid */
1320 cached_skb = tp->fastpath_skb_hint;
1321 cached_fack_count = tp->fastpath_cnt_hint;
1323 cached_skb = tcp_write_queue_head(sk);
1324 cached_fack_count = 0;
1327 for (i=0; i<num_sacks; i++, sp++) {
1328 struct sk_buff *skb;
1329 __u32 start_seq = ntohl(sp->start_seq);
1330 __u32 end_seq = ntohl(sp->end_seq);
1332 int dup_sack = (found_dup_sack && (i == first_sack_index));
1334 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) {
1336 if (!tp->undo_marker)
1337 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1339 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1341 /* Don't count olds caused by ACK reordering */
1342 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1343 !after(end_seq, tp->snd_una))
1345 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1351 fack_count = cached_fack_count;
1353 /* Event "B" in the comment above. */
1354 if (after(end_seq, tp->high_seq))
1355 flag |= FLAG_DATA_LOST;
1357 tcp_for_write_queue_from(skb, sk) {
1361 if (skb == tcp_send_head(sk))
1365 cached_fack_count = fack_count;
1366 if (i == first_sack_index) {
1367 tp->fastpath_skb_hint = skb;
1368 tp->fastpath_cnt_hint = fack_count;
1371 /* The retransmission queue is always in order, so
1372 * we can short-circuit the walk early.
1374 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1377 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq);
1381 fack_count += tcp_skb_pcount(skb);
1383 sacked = TCP_SKB_CB(skb)->sacked;
1385 /* Account D-SACK for retransmitted packet. */
1386 if ((dup_sack && in_sack) &&
1387 (sacked & TCPCB_RETRANS) &&
1388 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1391 /* The frame is ACKed. */
1392 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1393 if (sacked&TCPCB_RETRANS) {
1394 if ((dup_sack && in_sack) &&
1395 (sacked&TCPCB_SACKED_ACKED))
1396 reord = min(fack_count, reord);
1398 /* If it was in a hole, we detected reordering. */
1399 if (fack_count < prior_fackets &&
1400 !(sacked&TCPCB_SACKED_ACKED))
1401 reord = min(fack_count, reord);
1404 /* Nothing to do; acked frame is about to be dropped. */
1411 if (!(sacked&TCPCB_SACKED_ACKED)) {
1412 if (sacked & TCPCB_SACKED_RETRANS) {
1413 /* If the segment is not tagged as lost,
1414 * we do not clear RETRANS, believing
1415 * that retransmission is still in flight.
1417 if (sacked & TCPCB_LOST) {
1418 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1419 tp->lost_out -= tcp_skb_pcount(skb);
1420 tp->retrans_out -= tcp_skb_pcount(skb);
1422 /* clear lost hint */
1423 tp->retransmit_skb_hint = NULL;
1426 /* New sack for not retransmitted frame,
1427 * which was in hole. It is reordering.
1429 if (!(sacked & TCPCB_RETRANS) &&
1430 fack_count < prior_fackets)
1431 reord = min(fack_count, reord);
1433 if (sacked & TCPCB_LOST) {
1434 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1435 tp->lost_out -= tcp_skb_pcount(skb);
1437 /* clear lost hint */
1438 tp->retransmit_skb_hint = NULL;
1440 /* SACK enhanced F-RTO detection.
1441 * Set flag if and only if non-rexmitted
1442 * segments below frto_highmark are
1443 * SACKed (RFC4138; Appendix B).
1444 * Clearing correct due to in-order walk
1446 if (after(end_seq, tp->frto_highmark)) {
1447 flag &= ~FLAG_ONLY_ORIG_SACKED;
1449 if (!(sacked & TCPCB_RETRANS))
1450 flag |= FLAG_ONLY_ORIG_SACKED;
1454 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1455 flag |= FLAG_DATA_SACKED;
1456 tp->sacked_out += tcp_skb_pcount(skb);
1458 if (fack_count > tp->fackets_out)
1459 tp->fackets_out = fack_count;
1461 if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack)) {
1462 tp->highest_sack = TCP_SKB_CB(skb)->seq;
1463 highest_sack_end_seq = TCP_SKB_CB(skb)->end_seq;
1466 if (dup_sack && (sacked&TCPCB_RETRANS))
1467 reord = min(fack_count, reord);
1470 /* D-SACK. We can detect redundant retransmission
1471 * in S|R and plain R frames and clear it.
1472 * undo_retrans is decreased above, L|R frames
1473 * are accounted above as well.
1476 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1477 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1478 tp->retrans_out -= tcp_skb_pcount(skb);
1479 tp->retransmit_skb_hint = NULL;
1484 if (tp->retrans_out && highest_sack_end_seq &&
1485 after(highest_sack_end_seq, tp->high_seq) &&
1486 icsk->icsk_ca_state == TCP_CA_Recovery)
1487 flag |= tcp_mark_lost_retrans(sk, highest_sack_end_seq);
1489 tcp_verify_left_out(tp);
1491 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1492 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1493 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1495 #if FASTRETRANS_DEBUG > 0
1496 BUG_TRAP((int)tp->sacked_out >= 0);
1497 BUG_TRAP((int)tp->lost_out >= 0);
1498 BUG_TRAP((int)tp->retrans_out >= 0);
1499 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1504 /* If we receive more dupacks than we expected counting segments
1505 * in assumption of absent reordering, interpret this as reordering.
1506 * The only another reason could be bug in receiver TCP.
1508 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1510 struct tcp_sock *tp = tcp_sk(sk);
1513 holes = max(tp->lost_out, 1U);
1514 holes = min(holes, tp->packets_out);
1516 if ((tp->sacked_out + holes) > tp->packets_out) {
1517 tp->sacked_out = tp->packets_out - holes;
1518 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1522 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1524 static void tcp_add_reno_sack(struct sock *sk)
1526 struct tcp_sock *tp = tcp_sk(sk);
1528 tcp_check_reno_reordering(sk, 0);
1529 tcp_verify_left_out(tp);
1532 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1534 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1536 struct tcp_sock *tp = tcp_sk(sk);
1539 /* One ACK acked hole. The rest eat duplicate ACKs. */
1540 if (acked-1 >= tp->sacked_out)
1543 tp->sacked_out -= acked-1;
1545 tcp_check_reno_reordering(sk, acked);
1546 tcp_verify_left_out(tp);
1549 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1554 /* F-RTO can only be used if TCP has never retransmitted anything other than
1555 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1557 int tcp_use_frto(struct sock *sk)
1559 const struct tcp_sock *tp = tcp_sk(sk);
1560 struct sk_buff *skb;
1562 if (!sysctl_tcp_frto)
1568 /* Avoid expensive walking of rexmit queue if possible */
1569 if (tp->retrans_out > 1)
1572 skb = tcp_write_queue_head(sk);
1573 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1574 tcp_for_write_queue_from(skb, sk) {
1575 if (skb == tcp_send_head(sk))
1577 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1579 /* Short-circuit when first non-SACKed skb has been checked */
1580 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1586 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1587 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1588 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1589 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1590 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1591 * bits are handled if the Loss state is really to be entered (in
1592 * tcp_enter_frto_loss).
1594 * Do like tcp_enter_loss() would; when RTO expires the second time it
1596 * "Reduce ssthresh if it has not yet been made inside this window."
1598 void tcp_enter_frto(struct sock *sk)
1600 const struct inet_connection_sock *icsk = inet_csk(sk);
1601 struct tcp_sock *tp = tcp_sk(sk);
1602 struct sk_buff *skb;
1604 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1605 tp->snd_una == tp->high_seq ||
1606 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1607 !icsk->icsk_retransmits)) {
1608 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1609 /* Our state is too optimistic in ssthresh() call because cwnd
1610 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1611 * recovery has not yet completed. Pattern would be this: RTO,
1612 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1614 * RFC4138 should be more specific on what to do, even though
1615 * RTO is quite unlikely to occur after the first Cumulative ACK
1616 * due to back-off and complexity of triggering events ...
1618 if (tp->frto_counter) {
1620 stored_cwnd = tp->snd_cwnd;
1622 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1623 tp->snd_cwnd = stored_cwnd;
1625 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1627 /* ... in theory, cong.control module could do "any tricks" in
1628 * ssthresh(), which means that ca_state, lost bits and lost_out
1629 * counter would have to be faked before the call occurs. We
1630 * consider that too expensive, unlikely and hacky, so modules
1631 * using these in ssthresh() must deal these incompatibility
1632 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1634 tcp_ca_event(sk, CA_EVENT_FRTO);
1637 tp->undo_marker = tp->snd_una;
1638 tp->undo_retrans = 0;
1640 skb = tcp_write_queue_head(sk);
1641 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1642 tp->undo_marker = 0;
1643 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1644 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1645 tp->retrans_out -= tcp_skb_pcount(skb);
1647 tcp_verify_left_out(tp);
1649 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1650 * The last condition is necessary at least in tp->frto_counter case.
1652 if (IsSackFrto() && (tp->frto_counter ||
1653 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1654 after(tp->high_seq, tp->snd_una)) {
1655 tp->frto_highmark = tp->high_seq;
1657 tp->frto_highmark = tp->snd_nxt;
1659 tcp_set_ca_state(sk, TCP_CA_Disorder);
1660 tp->high_seq = tp->snd_nxt;
1661 tp->frto_counter = 1;
1664 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1665 * which indicates that we should follow the traditional RTO recovery,
1666 * i.e. mark everything lost and do go-back-N retransmission.
1668 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1670 struct tcp_sock *tp = tcp_sk(sk);
1671 struct sk_buff *skb;
1674 tp->retrans_out = 0;
1675 if (tcp_is_reno(tp))
1676 tcp_reset_reno_sack(tp);
1678 tcp_for_write_queue(skb, sk) {
1679 if (skb == tcp_send_head(sk))
1682 * Count the retransmission made on RTO correctly (only when
1683 * waiting for the first ACK and did not get it)...
1685 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1686 /* For some reason this R-bit might get cleared? */
1687 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1688 tp->retrans_out += tcp_skb_pcount(skb);
1689 /* ...enter this if branch just for the first segment */
1690 flag |= FLAG_DATA_ACKED;
1692 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1693 tp->undo_marker = 0;
1694 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1697 /* Don't lost mark skbs that were fwd transmitted after RTO */
1698 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1699 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1700 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1701 tp->lost_out += tcp_skb_pcount(skb);
1704 tcp_verify_left_out(tp);
1706 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1707 tp->snd_cwnd_cnt = 0;
1708 tp->snd_cwnd_stamp = tcp_time_stamp;
1709 tp->frto_counter = 0;
1710 tp->bytes_acked = 0;
1712 tp->reordering = min_t(unsigned int, tp->reordering,
1713 sysctl_tcp_reordering);
1714 tcp_set_ca_state(sk, TCP_CA_Loss);
1715 tp->high_seq = tp->frto_highmark;
1716 TCP_ECN_queue_cwr(tp);
1718 tcp_clear_retrans_hints_partial(tp);
1721 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1723 tp->retrans_out = 0;
1726 tp->undo_marker = 0;
1727 tp->undo_retrans = 0;
1730 void tcp_clear_retrans(struct tcp_sock *tp)
1732 tcp_clear_retrans_partial(tp);
1734 tp->fackets_out = 0;
1738 /* Enter Loss state. If "how" is not zero, forget all SACK information
1739 * and reset tags completely, otherwise preserve SACKs. If receiver
1740 * dropped its ofo queue, we will know this due to reneging detection.
1742 void tcp_enter_loss(struct sock *sk, int how)
1744 const struct inet_connection_sock *icsk = inet_csk(sk);
1745 struct tcp_sock *tp = tcp_sk(sk);
1746 struct sk_buff *skb;
1748 /* Reduce ssthresh if it has not yet been made inside this window. */
1749 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1750 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1751 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1752 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1753 tcp_ca_event(sk, CA_EVENT_LOSS);
1756 tp->snd_cwnd_cnt = 0;
1757 tp->snd_cwnd_stamp = tcp_time_stamp;
1759 tp->bytes_acked = 0;
1760 tcp_clear_retrans_partial(tp);
1762 if (tcp_is_reno(tp))
1763 tcp_reset_reno_sack(tp);
1766 /* Push undo marker, if it was plain RTO and nothing
1767 * was retransmitted. */
1768 tp->undo_marker = tp->snd_una;
1769 tcp_clear_retrans_hints_partial(tp);
1772 tp->fackets_out = 0;
1773 tcp_clear_all_retrans_hints(tp);
1776 tcp_for_write_queue(skb, sk) {
1777 if (skb == tcp_send_head(sk))
1780 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1781 tp->undo_marker = 0;
1782 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1783 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1784 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1785 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1786 tp->lost_out += tcp_skb_pcount(skb);
1789 tcp_verify_left_out(tp);
1791 tp->reordering = min_t(unsigned int, tp->reordering,
1792 sysctl_tcp_reordering);
1793 tcp_set_ca_state(sk, TCP_CA_Loss);
1794 tp->high_seq = tp->snd_nxt;
1795 TCP_ECN_queue_cwr(tp);
1796 /* Abort FRTO algorithm if one is in progress */
1797 tp->frto_counter = 0;
1800 static int tcp_check_sack_reneging(struct sock *sk)
1802 struct sk_buff *skb;
1804 /* If ACK arrived pointing to a remembered SACK,
1805 * it means that our remembered SACKs do not reflect
1806 * real state of receiver i.e.
1807 * receiver _host_ is heavily congested (or buggy).
1808 * Do processing similar to RTO timeout.
1810 if ((skb = tcp_write_queue_head(sk)) != NULL &&
1811 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1812 struct inet_connection_sock *icsk = inet_csk(sk);
1813 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1815 tcp_enter_loss(sk, 1);
1816 icsk->icsk_retransmits++;
1817 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1818 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1819 icsk->icsk_rto, TCP_RTO_MAX);
1825 static inline int tcp_fackets_out(struct tcp_sock *tp)
1827 return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1830 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1832 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1835 static inline int tcp_head_timedout(struct sock *sk)
1837 struct tcp_sock *tp = tcp_sk(sk);
1839 return tp->packets_out &&
1840 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1843 /* Linux NewReno/SACK/FACK/ECN state machine.
1844 * --------------------------------------
1846 * "Open" Normal state, no dubious events, fast path.
1847 * "Disorder" In all the respects it is "Open",
1848 * but requires a bit more attention. It is entered when
1849 * we see some SACKs or dupacks. It is split of "Open"
1850 * mainly to move some processing from fast path to slow one.
1851 * "CWR" CWND was reduced due to some Congestion Notification event.
1852 * It can be ECN, ICMP source quench, local device congestion.
1853 * "Recovery" CWND was reduced, we are fast-retransmitting.
1854 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1856 * tcp_fastretrans_alert() is entered:
1857 * - each incoming ACK, if state is not "Open"
1858 * - when arrived ACK is unusual, namely:
1863 * Counting packets in flight is pretty simple.
1865 * in_flight = packets_out - left_out + retrans_out
1867 * packets_out is SND.NXT-SND.UNA counted in packets.
1869 * retrans_out is number of retransmitted segments.
1871 * left_out is number of segments left network, but not ACKed yet.
1873 * left_out = sacked_out + lost_out
1875 * sacked_out: Packets, which arrived to receiver out of order
1876 * and hence not ACKed. With SACKs this number is simply
1877 * amount of SACKed data. Even without SACKs
1878 * it is easy to give pretty reliable estimate of this number,
1879 * counting duplicate ACKs.
1881 * lost_out: Packets lost by network. TCP has no explicit
1882 * "loss notification" feedback from network (for now).
1883 * It means that this number can be only _guessed_.
1884 * Actually, it is the heuristics to predict lossage that
1885 * distinguishes different algorithms.
1887 * F.e. after RTO, when all the queue is considered as lost,
1888 * lost_out = packets_out and in_flight = retrans_out.
1890 * Essentially, we have now two algorithms counting
1893 * FACK: It is the simplest heuristics. As soon as we decided
1894 * that something is lost, we decide that _all_ not SACKed
1895 * packets until the most forward SACK are lost. I.e.
1896 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1897 * It is absolutely correct estimate, if network does not reorder
1898 * packets. And it loses any connection to reality when reordering
1899 * takes place. We use FACK by default until reordering
1900 * is suspected on the path to this destination.
1902 * NewReno: when Recovery is entered, we assume that one segment
1903 * is lost (classic Reno). While we are in Recovery and
1904 * a partial ACK arrives, we assume that one more packet
1905 * is lost (NewReno). This heuristics are the same in NewReno
1908 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1909 * deflation etc. CWND is real congestion window, never inflated, changes
1910 * only according to classic VJ rules.
1912 * Really tricky (and requiring careful tuning) part of algorithm
1913 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1914 * The first determines the moment _when_ we should reduce CWND and,
1915 * hence, slow down forward transmission. In fact, it determines the moment
1916 * when we decide that hole is caused by loss, rather than by a reorder.
1918 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1919 * holes, caused by lost packets.
1921 * And the most logically complicated part of algorithm is undo
1922 * heuristics. We detect false retransmits due to both too early
1923 * fast retransmit (reordering) and underestimated RTO, analyzing
1924 * timestamps and D-SACKs. When we detect that some segments were
1925 * retransmitted by mistake and CWND reduction was wrong, we undo
1926 * window reduction and abort recovery phase. This logic is hidden
1927 * inside several functions named tcp_try_undo_<something>.
1930 /* This function decides, when we should leave Disordered state
1931 * and enter Recovery phase, reducing congestion window.
1933 * Main question: may we further continue forward transmission
1934 * with the same cwnd?
1936 static int tcp_time_to_recover(struct sock *sk)
1938 struct tcp_sock *tp = tcp_sk(sk);
1941 /* Do not perform any recovery during FRTO algorithm */
1942 if (tp->frto_counter)
1945 /* Trick#1: The loss is proven. */
1949 /* Not-A-Trick#2 : Classic rule... */
1950 if (tcp_fackets_out(tp) > tp->reordering)
1953 /* Trick#3 : when we use RFC2988 timer restart, fast
1954 * retransmit can be triggered by timeout of queue head.
1956 if (tcp_head_timedout(sk))
1959 /* Trick#4: It is still not OK... But will it be useful to delay
1962 packets_out = tp->packets_out;
1963 if (packets_out <= tp->reordering &&
1964 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1965 !tcp_may_send_now(sk)) {
1966 /* We have nothing to send. This connection is limited
1967 * either by receiver window or by application.
1975 /* RFC: This is from the original, I doubt that this is necessary at all:
1976 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1977 * retransmitted past LOST markings in the first place? I'm not fully sure
1978 * about undo and end of connection cases, which can cause R without L?
1980 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
1981 struct sk_buff *skb)
1983 if ((tp->retransmit_skb_hint != NULL) &&
1984 before(TCP_SKB_CB(skb)->seq,
1985 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1986 tp->retransmit_skb_hint = NULL;
1989 /* Mark head of queue up as lost. */
1990 static void tcp_mark_head_lost(struct sock *sk,
1991 int packets, u32 high_seq)
1993 struct tcp_sock *tp = tcp_sk(sk);
1994 struct sk_buff *skb;
1997 BUG_TRAP(packets <= tp->packets_out);
1998 if (tp->lost_skb_hint) {
1999 skb = tp->lost_skb_hint;
2000 cnt = tp->lost_cnt_hint;
2002 skb = tcp_write_queue_head(sk);
2006 tcp_for_write_queue_from(skb, sk) {
2007 if (skb == tcp_send_head(sk))
2009 /* TODO: do this better */
2010 /* this is not the most efficient way to do this... */
2011 tp->lost_skb_hint = skb;
2012 tp->lost_cnt_hint = cnt;
2013 cnt += tcp_skb_pcount(skb);
2014 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
2016 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2017 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2018 tp->lost_out += tcp_skb_pcount(skb);
2019 tcp_verify_retransmit_hint(tp, skb);
2022 tcp_verify_left_out(tp);
2025 /* Account newly detected lost packet(s) */
2027 static void tcp_update_scoreboard(struct sock *sk)
2029 struct tcp_sock *tp = tcp_sk(sk);
2031 if (tcp_is_fack(tp)) {
2032 int lost = tp->fackets_out - tp->reordering;
2035 tcp_mark_head_lost(sk, lost, tp->high_seq);
2037 tcp_mark_head_lost(sk, 1, tp->high_seq);
2040 /* New heuristics: it is possible only after we switched
2041 * to restart timer each time when something is ACKed.
2042 * Hence, we can detect timed out packets during fast
2043 * retransmit without falling to slow start.
2045 if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) {
2046 struct sk_buff *skb;
2048 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2049 : tcp_write_queue_head(sk);
2051 tcp_for_write_queue_from(skb, sk) {
2052 if (skb == tcp_send_head(sk))
2054 if (!tcp_skb_timedout(sk, skb))
2057 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
2058 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2059 tp->lost_out += tcp_skb_pcount(skb);
2060 tcp_verify_retransmit_hint(tp, skb);
2064 tp->scoreboard_skb_hint = skb;
2066 tcp_verify_left_out(tp);
2070 /* CWND moderation, preventing bursts due to too big ACKs
2071 * in dubious situations.
2073 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2075 tp->snd_cwnd = min(tp->snd_cwnd,
2076 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
2077 tp->snd_cwnd_stamp = tcp_time_stamp;
2080 /* Lower bound on congestion window is slow start threshold
2081 * unless congestion avoidance choice decides to overide it.
2083 static inline u32 tcp_cwnd_min(const struct sock *sk)
2085 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2087 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2090 /* Decrease cwnd each second ack. */
2091 static void tcp_cwnd_down(struct sock *sk, int flag)
2093 struct tcp_sock *tp = tcp_sk(sk);
2094 int decr = tp->snd_cwnd_cnt + 1;
2096 if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
2097 (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
2098 tp->snd_cwnd_cnt = decr&1;
2101 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2102 tp->snd_cwnd -= decr;
2104 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
2105 tp->snd_cwnd_stamp = tcp_time_stamp;
2109 /* Nothing was retransmitted or returned timestamp is less
2110 * than timestamp of the first retransmission.
2112 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2114 return !tp->retrans_stamp ||
2115 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2116 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2119 /* Undo procedures. */
2121 #if FASTRETRANS_DEBUG > 1
2122 static void DBGUNDO(struct sock *sk, const char *msg)
2124 struct tcp_sock *tp = tcp_sk(sk);
2125 struct inet_sock *inet = inet_sk(sk);
2127 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2129 NIPQUAD(inet->daddr), ntohs(inet->dport),
2130 tp->snd_cwnd, tcp_left_out(tp),
2131 tp->snd_ssthresh, tp->prior_ssthresh,
2135 #define DBGUNDO(x...) do { } while (0)
2138 static void tcp_undo_cwr(struct sock *sk, const int undo)
2140 struct tcp_sock *tp = tcp_sk(sk);
2142 if (tp->prior_ssthresh) {
2143 const struct inet_connection_sock *icsk = inet_csk(sk);
2145 if (icsk->icsk_ca_ops->undo_cwnd)
2146 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2148 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
2150 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2151 tp->snd_ssthresh = tp->prior_ssthresh;
2152 TCP_ECN_withdraw_cwr(tp);
2155 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2157 tcp_moderate_cwnd(tp);
2158 tp->snd_cwnd_stamp = tcp_time_stamp;
2160 /* There is something screwy going on with the retrans hints after
2162 tcp_clear_all_retrans_hints(tp);
2165 static inline int tcp_may_undo(struct tcp_sock *tp)
2167 return tp->undo_marker &&
2168 (!tp->undo_retrans || tcp_packet_delayed(tp));
2171 /* People celebrate: "We love our President!" */
2172 static int tcp_try_undo_recovery(struct sock *sk)
2174 struct tcp_sock *tp = tcp_sk(sk);
2176 if (tcp_may_undo(tp)) {
2177 /* Happy end! We did not retransmit anything
2178 * or our original transmission succeeded.
2180 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2181 tcp_undo_cwr(sk, 1);
2182 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2183 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2185 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2186 tp->undo_marker = 0;
2188 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2189 /* Hold old state until something *above* high_seq
2190 * is ACKed. For Reno it is MUST to prevent false
2191 * fast retransmits (RFC2582). SACK TCP is safe. */
2192 tcp_moderate_cwnd(tp);
2195 tcp_set_ca_state(sk, TCP_CA_Open);
2199 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2200 static void tcp_try_undo_dsack(struct sock *sk)
2202 struct tcp_sock *tp = tcp_sk(sk);
2204 if (tp->undo_marker && !tp->undo_retrans) {
2205 DBGUNDO(sk, "D-SACK");
2206 tcp_undo_cwr(sk, 1);
2207 tp->undo_marker = 0;
2208 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2212 /* Undo during fast recovery after partial ACK. */
2214 static int tcp_try_undo_partial(struct sock *sk, int acked)
2216 struct tcp_sock *tp = tcp_sk(sk);
2217 /* Partial ACK arrived. Force Hoe's retransmit. */
2218 int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering;
2220 if (tcp_may_undo(tp)) {
2221 /* Plain luck! Hole if filled with delayed
2222 * packet, rather than with a retransmit.
2224 if (tp->retrans_out == 0)
2225 tp->retrans_stamp = 0;
2227 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2230 tcp_undo_cwr(sk, 0);
2231 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2233 /* So... Do not make Hoe's retransmit yet.
2234 * If the first packet was delayed, the rest
2235 * ones are most probably delayed as well.
2242 /* Undo during loss recovery after partial ACK. */
2243 static int tcp_try_undo_loss(struct sock *sk)
2245 struct tcp_sock *tp = tcp_sk(sk);
2247 if (tcp_may_undo(tp)) {
2248 struct sk_buff *skb;
2249 tcp_for_write_queue(skb, sk) {
2250 if (skb == tcp_send_head(sk))
2252 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2255 tcp_clear_all_retrans_hints(tp);
2257 DBGUNDO(sk, "partial loss");
2259 tcp_undo_cwr(sk, 1);
2260 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2261 inet_csk(sk)->icsk_retransmits = 0;
2262 tp->undo_marker = 0;
2263 if (tcp_is_sack(tp))
2264 tcp_set_ca_state(sk, TCP_CA_Open);
2270 static inline void tcp_complete_cwr(struct sock *sk)
2272 struct tcp_sock *tp = tcp_sk(sk);
2273 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2274 tp->snd_cwnd_stamp = tcp_time_stamp;
2275 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2278 static void tcp_try_to_open(struct sock *sk, int flag)
2280 struct tcp_sock *tp = tcp_sk(sk);
2282 tcp_verify_left_out(tp);
2284 if (tp->retrans_out == 0)
2285 tp->retrans_stamp = 0;
2288 tcp_enter_cwr(sk, 1);
2290 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2291 int state = TCP_CA_Open;
2293 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2294 state = TCP_CA_Disorder;
2296 if (inet_csk(sk)->icsk_ca_state != state) {
2297 tcp_set_ca_state(sk, state);
2298 tp->high_seq = tp->snd_nxt;
2300 tcp_moderate_cwnd(tp);
2302 tcp_cwnd_down(sk, flag);
2306 static void tcp_mtup_probe_failed(struct sock *sk)
2308 struct inet_connection_sock *icsk = inet_csk(sk);
2310 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2311 icsk->icsk_mtup.probe_size = 0;
2314 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2316 struct tcp_sock *tp = tcp_sk(sk);
2317 struct inet_connection_sock *icsk = inet_csk(sk);
2319 /* FIXME: breaks with very large cwnd */
2320 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2321 tp->snd_cwnd = tp->snd_cwnd *
2322 tcp_mss_to_mtu(sk, tp->mss_cache) /
2323 icsk->icsk_mtup.probe_size;
2324 tp->snd_cwnd_cnt = 0;
2325 tp->snd_cwnd_stamp = tcp_time_stamp;
2326 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2328 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2329 icsk->icsk_mtup.probe_size = 0;
2330 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2334 /* Process an event, which can update packets-in-flight not trivially.
2335 * Main goal of this function is to calculate new estimate for left_out,
2336 * taking into account both packets sitting in receiver's buffer and
2337 * packets lost by network.
2339 * Besides that it does CWND reduction, when packet loss is detected
2340 * and changes state of machine.
2342 * It does _not_ decide what to send, it is made in function
2343 * tcp_xmit_retransmit_queue().
2346 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2348 struct inet_connection_sock *icsk = inet_csk(sk);
2349 struct tcp_sock *tp = tcp_sk(sk);
2350 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2351 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2352 (tp->fackets_out > tp->reordering));
2354 /* Some technical things:
2355 * 1. Reno does not count dupacks (sacked_out) automatically. */
2356 if (!tp->packets_out)
2359 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2360 tp->fackets_out = 0;
2362 /* Now state machine starts.
2363 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2365 tp->prior_ssthresh = 0;
2367 /* B. In all the states check for reneging SACKs. */
2368 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2371 /* C. Process data loss notification, provided it is valid. */
2372 if ((flag&FLAG_DATA_LOST) &&
2373 before(tp->snd_una, tp->high_seq) &&
2374 icsk->icsk_ca_state != TCP_CA_Open &&
2375 tp->fackets_out > tp->reordering) {
2376 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2377 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2380 /* D. Check consistency of the current state. */
2381 tcp_verify_left_out(tp);
2383 /* E. Check state exit conditions. State can be terminated
2384 * when high_seq is ACKed. */
2385 if (icsk->icsk_ca_state == TCP_CA_Open) {
2386 BUG_TRAP(tp->retrans_out == 0);
2387 tp->retrans_stamp = 0;
2388 } else if (!before(tp->snd_una, tp->high_seq)) {
2389 switch (icsk->icsk_ca_state) {
2391 icsk->icsk_retransmits = 0;
2392 if (tcp_try_undo_recovery(sk))
2397 /* CWR is to be held something *above* high_seq
2398 * is ACKed for CWR bit to reach receiver. */
2399 if (tp->snd_una != tp->high_seq) {
2400 tcp_complete_cwr(sk);
2401 tcp_set_ca_state(sk, TCP_CA_Open);
2405 case TCP_CA_Disorder:
2406 tcp_try_undo_dsack(sk);
2407 if (!tp->undo_marker ||
2408 /* For SACK case do not Open to allow to undo
2409 * catching for all duplicate ACKs. */
2410 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2411 tp->undo_marker = 0;
2412 tcp_set_ca_state(sk, TCP_CA_Open);
2416 case TCP_CA_Recovery:
2417 if (tcp_is_reno(tp))
2418 tcp_reset_reno_sack(tp);
2419 if (tcp_try_undo_recovery(sk))
2421 tcp_complete_cwr(sk);
2426 /* F. Process state. */
2427 switch (icsk->icsk_ca_state) {
2428 case TCP_CA_Recovery:
2429 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2430 if (tcp_is_reno(tp) && is_dupack)
2431 tcp_add_reno_sack(sk);
2433 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2436 if (flag&FLAG_DATA_ACKED)
2437 icsk->icsk_retransmits = 0;
2438 if (!tcp_try_undo_loss(sk)) {
2439 tcp_moderate_cwnd(tp);
2440 tcp_xmit_retransmit_queue(sk);
2443 if (icsk->icsk_ca_state != TCP_CA_Open)
2445 /* Loss is undone; fall through to processing in Open state. */
2447 if (tcp_is_reno(tp)) {
2448 if (flag & FLAG_SND_UNA_ADVANCED)
2449 tcp_reset_reno_sack(tp);
2451 tcp_add_reno_sack(sk);
2454 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2455 tcp_try_undo_dsack(sk);
2457 if (!tcp_time_to_recover(sk)) {
2458 tcp_try_to_open(sk, flag);
2462 /* MTU probe failure: don't reduce cwnd */
2463 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2464 icsk->icsk_mtup.probe_size &&
2465 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2466 tcp_mtup_probe_failed(sk);
2467 /* Restores the reduction we did in tcp_mtup_probe() */
2469 tcp_simple_retransmit(sk);
2473 /* Otherwise enter Recovery state */
2475 if (tcp_is_reno(tp))
2476 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2478 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2480 tp->high_seq = tp->snd_nxt;
2481 tp->prior_ssthresh = 0;
2482 tp->undo_marker = tp->snd_una;
2483 tp->undo_retrans = tp->retrans_out;
2485 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2486 if (!(flag&FLAG_ECE))
2487 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2488 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2489 TCP_ECN_queue_cwr(tp);
2492 tp->bytes_acked = 0;
2493 tp->snd_cwnd_cnt = 0;
2494 tcp_set_ca_state(sk, TCP_CA_Recovery);
2497 if (do_lost || tcp_head_timedout(sk))
2498 tcp_update_scoreboard(sk);
2499 tcp_cwnd_down(sk, flag);
2500 tcp_xmit_retransmit_queue(sk);
2503 /* Read draft-ietf-tcplw-high-performance before mucking
2504 * with this code. (Supersedes RFC1323)
2506 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2508 /* RTTM Rule: A TSecr value received in a segment is used to
2509 * update the averaged RTT measurement only if the segment
2510 * acknowledges some new data, i.e., only if it advances the
2511 * left edge of the send window.
2513 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2514 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2516 * Changed: reset backoff as soon as we see the first valid sample.
2517 * If we do not, we get strongly overestimated rto. With timestamps
2518 * samples are accepted even from very old segments: f.e., when rtt=1
2519 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2520 * answer arrives rto becomes 120 seconds! If at least one of segments
2521 * in window is lost... Voila. --ANK (010210)
2523 struct tcp_sock *tp = tcp_sk(sk);
2524 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2525 tcp_rtt_estimator(sk, seq_rtt);
2527 inet_csk(sk)->icsk_backoff = 0;
2531 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2533 /* We don't have a timestamp. Can only use
2534 * packets that are not retransmitted to determine
2535 * rtt estimates. Also, we must not reset the
2536 * backoff for rto until we get a non-retransmitted
2537 * packet. This allows us to deal with a situation
2538 * where the network delay has increased suddenly.
2539 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2542 if (flag & FLAG_RETRANS_DATA_ACKED)
2545 tcp_rtt_estimator(sk, seq_rtt);
2547 inet_csk(sk)->icsk_backoff = 0;
2551 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2554 const struct tcp_sock *tp = tcp_sk(sk);
2555 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2556 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2557 tcp_ack_saw_tstamp(sk, flag);
2558 else if (seq_rtt >= 0)
2559 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2562 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2563 u32 in_flight, int good)
2565 const struct inet_connection_sock *icsk = inet_csk(sk);
2566 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2567 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2570 /* Restart timer after forward progress on connection.
2571 * RFC2988 recommends to restart timer to now+rto.
2573 static void tcp_rearm_rto(struct sock *sk)
2575 struct tcp_sock *tp = tcp_sk(sk);
2577 if (!tp->packets_out) {
2578 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2580 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2584 /* If we get here, the whole TSO packet has not been acked. */
2585 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2587 struct tcp_sock *tp = tcp_sk(sk);
2590 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2592 packets_acked = tcp_skb_pcount(skb);
2593 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2595 packets_acked -= tcp_skb_pcount(skb);
2597 if (packets_acked) {
2598 BUG_ON(tcp_skb_pcount(skb) == 0);
2599 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2602 return packets_acked;
2605 /* Remove acknowledged frames from the retransmission queue. If our packet
2606 * is before the ack sequence we can discard it as it's confirmed to have
2607 * arrived at the other end.
2609 static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p)
2611 struct tcp_sock *tp = tcp_sk(sk);
2612 const struct inet_connection_sock *icsk = inet_csk(sk);
2613 struct sk_buff *skb;
2614 u32 now = tcp_time_stamp;
2615 int fully_acked = 1;
2617 int prior_packets = tp->packets_out;
2619 ktime_t last_ackt = net_invalid_timestamp();
2621 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2622 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2625 u8 sacked = scb->sacked;
2627 if (after(scb->end_seq, tp->snd_una)) {
2628 if (tcp_skb_pcount(skb) == 1 ||
2629 !after(tp->snd_una, scb->seq))
2632 packets_acked = tcp_tso_acked(sk, skb);
2637 end_seq = tp->snd_una;
2639 packets_acked = tcp_skb_pcount(skb);
2640 end_seq = scb->end_seq;
2643 /* MTU probing checks */
2644 if (fully_acked && icsk->icsk_mtup.probe_size &&
2645 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2646 tcp_mtup_probe_success(sk, skb);
2650 if (sacked & TCPCB_RETRANS) {
2651 if (sacked & TCPCB_SACKED_RETRANS)
2652 tp->retrans_out -= packets_acked;
2653 flag |= FLAG_RETRANS_DATA_ACKED;
2655 if ((flag & FLAG_DATA_ACKED) ||
2656 (packets_acked > 1))
2657 flag |= FLAG_NONHEAD_RETRANS_ACKED;
2658 } else if (seq_rtt < 0) {
2659 seq_rtt = now - scb->when;
2661 last_ackt = skb->tstamp;
2664 if (sacked & TCPCB_SACKED_ACKED)
2665 tp->sacked_out -= packets_acked;
2666 if (sacked & TCPCB_LOST)
2667 tp->lost_out -= packets_acked;
2669 if ((sacked & TCPCB_URG) && tp->urg_mode &&
2670 !before(end_seq, tp->snd_up))
2672 } else if (seq_rtt < 0) {
2673 seq_rtt = now - scb->when;
2675 last_ackt = skb->tstamp;
2677 tp->packets_out -= packets_acked;
2679 /* Initial outgoing SYN's get put onto the write_queue
2680 * just like anything else we transmit. It is not
2681 * true data, and if we misinform our callers that
2682 * this ACK acks real data, we will erroneously exit
2683 * connection startup slow start one packet too
2684 * quickly. This is severely frowned upon behavior.
2686 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2687 flag |= FLAG_DATA_ACKED;
2689 flag |= FLAG_SYN_ACKED;
2690 tp->retrans_stamp = 0;
2696 tcp_unlink_write_queue(skb, sk);
2697 sk_stream_free_skb(sk, skb);
2698 tcp_clear_all_retrans_hints(tp);
2701 if (flag & FLAG_ACKED) {
2702 u32 pkts_acked = prior_packets - tp->packets_out;
2703 const struct tcp_congestion_ops *ca_ops
2704 = inet_csk(sk)->icsk_ca_ops;
2706 tcp_ack_update_rtt(sk, flag, seq_rtt);
2709 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2710 /* hint's skb might be NULL but we don't need to care */
2711 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
2712 tp->fastpath_cnt_hint);
2713 if (tcp_is_reno(tp))
2714 tcp_remove_reno_sacks(sk, pkts_acked);
2716 if (ca_ops->pkts_acked) {
2719 /* Is the ACK triggering packet unambiguous? */
2720 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2721 /* High resolution needed and available? */
2722 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2723 !ktime_equal(last_ackt,
2724 net_invalid_timestamp()))
2725 rtt_us = ktime_us_delta(ktime_get_real(),
2727 else if (seq_rtt > 0)
2728 rtt_us = jiffies_to_usecs(seq_rtt);
2731 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2735 #if FASTRETRANS_DEBUG > 0
2736 BUG_TRAP((int)tp->sacked_out >= 0);
2737 BUG_TRAP((int)tp->lost_out >= 0);
2738 BUG_TRAP((int)tp->retrans_out >= 0);
2739 if (!tp->packets_out && tcp_is_sack(tp)) {
2740 icsk = inet_csk(sk);
2742 printk(KERN_DEBUG "Leak l=%u %d\n",
2743 tp->lost_out, icsk->icsk_ca_state);
2746 if (tp->sacked_out) {
2747 printk(KERN_DEBUG "Leak s=%u %d\n",
2748 tp->sacked_out, icsk->icsk_ca_state);
2751 if (tp->retrans_out) {
2752 printk(KERN_DEBUG "Leak r=%u %d\n",
2753 tp->retrans_out, icsk->icsk_ca_state);
2754 tp->retrans_out = 0;
2758 *seq_rtt_p = seq_rtt;
2762 static void tcp_ack_probe(struct sock *sk)
2764 const struct tcp_sock *tp = tcp_sk(sk);
2765 struct inet_connection_sock *icsk = inet_csk(sk);
2767 /* Was it a usable window open? */
2769 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2770 tp->snd_una + tp->snd_wnd)) {
2771 icsk->icsk_backoff = 0;
2772 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2773 /* Socket must be waked up by subsequent tcp_data_snd_check().
2774 * This function is not for random using!
2777 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2778 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2783 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2785 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2786 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2789 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2791 const struct tcp_sock *tp = tcp_sk(sk);
2792 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2793 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2796 /* Check that window update is acceptable.
2797 * The function assumes that snd_una<=ack<=snd_next.
2799 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2800 const u32 ack_seq, const u32 nwin)
2802 return (after(ack, tp->snd_una) ||
2803 after(ack_seq, tp->snd_wl1) ||
2804 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2807 /* Update our send window.
2809 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2810 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2812 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2815 struct tcp_sock *tp = tcp_sk(sk);
2817 u32 nwin = ntohs(tcp_hdr(skb)->window);
2819 if (likely(!tcp_hdr(skb)->syn))
2820 nwin <<= tp->rx_opt.snd_wscale;
2822 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2823 flag |= FLAG_WIN_UPDATE;
2824 tcp_update_wl(tp, ack, ack_seq);
2826 if (tp->snd_wnd != nwin) {
2829 /* Note, it is the only place, where
2830 * fast path is recovered for sending TCP.
2833 tcp_fast_path_check(sk);
2835 if (nwin > tp->max_window) {
2836 tp->max_window = nwin;
2837 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2847 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2848 * continue in congestion avoidance.
2850 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2852 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2853 tp->snd_cwnd_cnt = 0;
2854 tp->bytes_acked = 0;
2855 TCP_ECN_queue_cwr(tp);
2856 tcp_moderate_cwnd(tp);
2859 /* A conservative spurious RTO response algorithm: reduce cwnd using
2860 * rate halving and continue in congestion avoidance.
2862 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2864 tcp_enter_cwr(sk, 0);
2867 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2870 tcp_ratehalving_spur_to_response(sk);
2872 tcp_undo_cwr(sk, 1);
2875 /* F-RTO spurious RTO detection algorithm (RFC4138)
2877 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2878 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2879 * window (but not to or beyond highest sequence sent before RTO):
2880 * On First ACK, send two new segments out.
2881 * On Second ACK, RTO was likely spurious. Do spurious response (response
2882 * algorithm is not part of the F-RTO detection algorithm
2883 * given in RFC4138 but can be selected separately).
2884 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2885 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2886 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2887 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2889 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2890 * original window even after we transmit two new data segments.
2893 * on first step, wait until first cumulative ACK arrives, then move to
2894 * the second step. In second step, the next ACK decides.
2896 * F-RTO is implemented (mainly) in four functions:
2897 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2898 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2899 * called when tcp_use_frto() showed green light
2900 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2901 * - tcp_enter_frto_loss() is called if there is not enough evidence
2902 * to prove that the RTO is indeed spurious. It transfers the control
2903 * from F-RTO to the conventional RTO recovery
2905 static int tcp_process_frto(struct sock *sk, int flag)
2907 struct tcp_sock *tp = tcp_sk(sk);
2909 tcp_verify_left_out(tp);
2911 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2912 if (flag&FLAG_DATA_ACKED)
2913 inet_csk(sk)->icsk_retransmits = 0;
2915 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
2916 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
2917 tp->undo_marker = 0;
2919 if (!before(tp->snd_una, tp->frto_highmark)) {
2920 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2924 if (!IsSackFrto() || tcp_is_reno(tp)) {
2925 /* RFC4138 shortcoming in step 2; should also have case c):
2926 * ACK isn't duplicate nor advances window, e.g., opposite dir
2929 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2932 if (!(flag&FLAG_DATA_ACKED)) {
2933 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2938 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2939 /* Prevent sending of new data. */
2940 tp->snd_cwnd = min(tp->snd_cwnd,
2941 tcp_packets_in_flight(tp));
2945 if ((tp->frto_counter >= 2) &&
2946 (!(flag&FLAG_FORWARD_PROGRESS) ||
2947 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2948 /* RFC4138 shortcoming (see comment above) */
2949 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2952 tcp_enter_frto_loss(sk, 3, flag);
2957 if (tp->frto_counter == 1) {
2958 /* Sending of the next skb must be allowed or no FRTO */
2959 if (!tcp_send_head(sk) ||
2960 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2961 tp->snd_una + tp->snd_wnd)) {
2962 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2967 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2968 tp->frto_counter = 2;
2971 switch (sysctl_tcp_frto_response) {
2973 tcp_undo_spur_to_response(sk, flag);
2976 tcp_conservative_spur_to_response(tp);
2979 tcp_ratehalving_spur_to_response(sk);
2982 tp->frto_counter = 0;
2983 tp->undo_marker = 0;
2984 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
2989 /* This routine deals with incoming acks, but not outgoing ones. */
2990 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2992 struct inet_connection_sock *icsk = inet_csk(sk);
2993 struct tcp_sock *tp = tcp_sk(sk);
2994 u32 prior_snd_una = tp->snd_una;
2995 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2996 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2997 u32 prior_in_flight;
3002 /* If the ack is newer than sent or older than previous acks
3003 * then we can probably ignore it.
3005 if (after(ack, tp->snd_nxt))
3006 goto uninteresting_ack;
3008 if (before(ack, prior_snd_una))
3011 if (after(ack, prior_snd_una))
3012 flag |= FLAG_SND_UNA_ADVANCED;
3014 if (sysctl_tcp_abc) {
3015 if (icsk->icsk_ca_state < TCP_CA_CWR)
3016 tp->bytes_acked += ack - prior_snd_una;
3017 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3018 /* we assume just one segment left network */
3019 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
3022 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3023 /* Window is constant, pure forward advance.
3024 * No more checks are required.
3025 * Note, we use the fact that SND.UNA>=SND.WL2.
3027 tcp_update_wl(tp, ack, ack_seq);
3029 flag |= FLAG_WIN_UPDATE;
3031 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3033 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3035 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3038 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3040 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3042 if (TCP_SKB_CB(skb)->sacked)
3043 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3045 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3048 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3051 /* We passed data and got it acked, remove any soft error
3052 * log. Something worked...
3054 sk->sk_err_soft = 0;
3055 tp->rcv_tstamp = tcp_time_stamp;
3056 prior_packets = tp->packets_out;
3060 prior_in_flight = tcp_packets_in_flight(tp);
3062 /* See if we can take anything off of the retransmit queue. */
3063 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
3065 /* Guarantee sacktag reordering detection against wrap-arounds */
3066 if (before(tp->frto_highmark, tp->snd_una))
3067 tp->frto_highmark = 0;
3068 if (tp->frto_counter)
3069 frto_cwnd = tcp_process_frto(sk, flag);
3071 if (tcp_ack_is_dubious(sk, flag)) {
3072 /* Advance CWND, if state allows this. */
3073 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3074 tcp_may_raise_cwnd(sk, flag))
3075 tcp_cong_avoid(sk, ack, prior_in_flight, 0);
3076 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
3078 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3079 tcp_cong_avoid(sk, ack, prior_in_flight, 1);
3082 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
3083 dst_confirm(sk->sk_dst_cache);
3088 icsk->icsk_probes_out = 0;
3090 /* If this ack opens up a zero window, clear backoff. It was
3091 * being used to time the probes, and is probably far higher than
3092 * it needs to be for normal retransmission.
3094 if (tcp_send_head(sk))
3099 if (TCP_SKB_CB(skb)->sacked)
3100 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3103 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3108 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3109 * But, this can also be called on packets in the established flow when
3110 * the fast version below fails.
3112 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
3115 struct tcphdr *th = tcp_hdr(skb);
3116 int length=(th->doff*4)-sizeof(struct tcphdr);
3118 ptr = (unsigned char *)(th + 1);
3119 opt_rx->saw_tstamp = 0;
3121 while (length > 0) {
3128 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3133 if (opsize < 2) /* "silly options" */
3135 if (opsize > length)
3136 return; /* don't parse partial options */
3139 if (opsize==TCPOLEN_MSS && th->syn && !estab) {
3140 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3142 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3143 in_mss = opt_rx->user_mss;
3144 opt_rx->mss_clamp = in_mss;
3149 if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3150 if (sysctl_tcp_window_scaling) {
3151 __u8 snd_wscale = *(__u8 *) ptr;
3152 opt_rx->wscale_ok = 1;
3153 if (snd_wscale > 14) {
3154 if (net_ratelimit())
3155 printk(KERN_INFO "tcp_parse_options: Illegal window "
3156 "scaling value %d >14 received.\n",
3160 opt_rx->snd_wscale = snd_wscale;
3163 case TCPOPT_TIMESTAMP:
3164 if (opsize==TCPOLEN_TIMESTAMP) {
3165 if ((estab && opt_rx->tstamp_ok) ||
3166 (!estab && sysctl_tcp_timestamps)) {
3167 opt_rx->saw_tstamp = 1;
3168 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3169 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3173 case TCPOPT_SACK_PERM:
3174 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3175 if (sysctl_tcp_sack) {
3176 opt_rx->sack_ok = 1;
3177 tcp_sack_reset(opt_rx);
3183 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3184 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3186 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3189 #ifdef CONFIG_TCP_MD5SIG
3192 * The MD5 Hash has already been
3193 * checked (see tcp_v{4,6}_do_rcv()).
3205 /* Fast parse options. This hopes to only see timestamps.
3206 * If it is wrong it falls back on tcp_parse_options().
3208 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3209 struct tcp_sock *tp)
3211 if (th->doff == sizeof(struct tcphdr)>>2) {
3212 tp->rx_opt.saw_tstamp = 0;
3214 } else if (tp->rx_opt.tstamp_ok &&
3215 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3216 __be32 *ptr = (__be32 *)(th + 1);
3217 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3218 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3219 tp->rx_opt.saw_tstamp = 1;
3221 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3223 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3227 tcp_parse_options(skb, &tp->rx_opt, 1);
3231 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3233 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3234 tp->rx_opt.ts_recent_stamp = get_seconds();
3237 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3239 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3240 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3241 * extra check below makes sure this can only happen
3242 * for pure ACK frames. -DaveM
3244 * Not only, also it occurs for expired timestamps.
3247 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3248 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3249 tcp_store_ts_recent(tp);
3253 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3255 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3256 * it can pass through stack. So, the following predicate verifies that
3257 * this segment is not used for anything but congestion avoidance or
3258 * fast retransmit. Moreover, we even are able to eliminate most of such
3259 * second order effects, if we apply some small "replay" window (~RTO)
3260 * to timestamp space.
3262 * All these measures still do not guarantee that we reject wrapped ACKs
3263 * on networks with high bandwidth, when sequence space is recycled fastly,
3264 * but it guarantees that such events will be very rare and do not affect
3265 * connection seriously. This doesn't look nice, but alas, PAWS is really
3268 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3269 * states that events when retransmit arrives after original data are rare.
3270 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3271 * the biggest problem on large power networks even with minor reordering.
3272 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3273 * up to bandwidth of 18Gigabit/sec. 8) ]
3276 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3278 struct tcp_sock *tp = tcp_sk(sk);
3279 struct tcphdr *th = tcp_hdr(skb);
3280 u32 seq = TCP_SKB_CB(skb)->seq;
3281 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3283 return (/* 1. Pure ACK with correct sequence number. */
3284 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3286 /* 2. ... and duplicate ACK. */
3287 ack == tp->snd_una &&
3289 /* 3. ... and does not update window. */
3290 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3292 /* 4. ... and sits in replay window. */
3293 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3296 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3298 const struct tcp_sock *tp = tcp_sk(sk);
3299 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3300 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3301 !tcp_disordered_ack(sk, skb));
3304 /* Check segment sequence number for validity.
3306 * Segment controls are considered valid, if the segment
3307 * fits to the window after truncation to the window. Acceptability
3308 * of data (and SYN, FIN, of course) is checked separately.
3309 * See tcp_data_queue(), for example.
3311 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3312 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3313 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3314 * (borrowed from freebsd)
3317 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3319 return !before(end_seq, tp->rcv_wup) &&
3320 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3323 /* When we get a reset we do this. */
3324 static void tcp_reset(struct sock *sk)
3326 /* We want the right error as BSD sees it (and indeed as we do). */
3327 switch (sk->sk_state) {
3329 sk->sk_err = ECONNREFUSED;
3331 case TCP_CLOSE_WAIT:
3337 sk->sk_err = ECONNRESET;
3340 if (!sock_flag(sk, SOCK_DEAD))
3341 sk->sk_error_report(sk);
3347 * Process the FIN bit. This now behaves as it is supposed to work
3348 * and the FIN takes effect when it is validly part of sequence
3349 * space. Not before when we get holes.
3351 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3352 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3355 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3356 * close and we go into CLOSING (and later onto TIME-WAIT)
3358 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3360 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3362 struct tcp_sock *tp = tcp_sk(sk);
3364 inet_csk_schedule_ack(sk);
3366 sk->sk_shutdown |= RCV_SHUTDOWN;
3367 sock_set_flag(sk, SOCK_DONE);
3369 switch (sk->sk_state) {
3371 case TCP_ESTABLISHED:
3372 /* Move to CLOSE_WAIT */
3373 tcp_set_state(sk, TCP_CLOSE_WAIT);
3374 inet_csk(sk)->icsk_ack.pingpong = 1;
3377 case TCP_CLOSE_WAIT:
3379 /* Received a retransmission of the FIN, do
3384 /* RFC793: Remain in the LAST-ACK state. */
3388 /* This case occurs when a simultaneous close
3389 * happens, we must ack the received FIN and
3390 * enter the CLOSING state.
3393 tcp_set_state(sk, TCP_CLOSING);
3396 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3398 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3401 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3402 * cases we should never reach this piece of code.
3404 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3405 __FUNCTION__, sk->sk_state);
3409 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3410 * Probably, we should reset in this case. For now drop them.
3412 __skb_queue_purge(&tp->out_of_order_queue);
3413 if (tcp_is_sack(tp))
3414 tcp_sack_reset(&tp->rx_opt);
3415 sk_stream_mem_reclaim(sk);
3417 if (!sock_flag(sk, SOCK_DEAD)) {
3418 sk->sk_state_change(sk);
3420 /* Do not send POLL_HUP for half duplex close. */
3421 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3422 sk->sk_state == TCP_CLOSE)
3423 sk_wake_async(sk, 1, POLL_HUP);
3425 sk_wake_async(sk, 1, POLL_IN);
3429 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3431 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3432 if (before(seq, sp->start_seq))
3433 sp->start_seq = seq;
3434 if (after(end_seq, sp->end_seq))
3435 sp->end_seq = end_seq;
3441 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3443 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3444 if (before(seq, tp->rcv_nxt))
3445 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3447 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3449 tp->rx_opt.dsack = 1;
3450 tp->duplicate_sack[0].start_seq = seq;
3451 tp->duplicate_sack[0].end_seq = end_seq;
3452 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3456 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3458 if (!tp->rx_opt.dsack)
3459 tcp_dsack_set(tp, seq, end_seq);
3461 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3464 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3466 struct tcp_sock *tp = tcp_sk(sk);
3468 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3469 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3470 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3471 tcp_enter_quickack_mode(sk);
3473 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3474 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3476 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3477 end_seq = tp->rcv_nxt;
3478 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3485 /* These routines update the SACK block as out-of-order packets arrive or
3486 * in-order packets close up the sequence space.
3488 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3491 struct tcp_sack_block *sp = &tp->selective_acks[0];
3492 struct tcp_sack_block *swalk = sp+1;
3494 /* See if the recent change to the first SACK eats into
3495 * or hits the sequence space of other SACK blocks, if so coalesce.
3497 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3498 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3501 /* Zap SWALK, by moving every further SACK up by one slot.
3502 * Decrease num_sacks.
3504 tp->rx_opt.num_sacks--;
3505 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3506 for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3510 this_sack++, swalk++;
3514 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3518 tmp = sack1->start_seq;
3519 sack1->start_seq = sack2->start_seq;
3520 sack2->start_seq = tmp;
3522 tmp = sack1->end_seq;
3523 sack1->end_seq = sack2->end_seq;
3524 sack2->end_seq = tmp;
3527 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3529 struct tcp_sock *tp = tcp_sk(sk);
3530 struct tcp_sack_block *sp = &tp->selective_acks[0];
3531 int cur_sacks = tp->rx_opt.num_sacks;
3537 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3538 if (tcp_sack_extend(sp, seq, end_seq)) {
3539 /* Rotate this_sack to the first one. */
3540 for (; this_sack>0; this_sack--, sp--)
3541 tcp_sack_swap(sp, sp-1);
3543 tcp_sack_maybe_coalesce(tp);
3548 /* Could not find an adjacent existing SACK, build a new one,
3549 * put it at the front, and shift everyone else down. We
3550 * always know there is at least one SACK present already here.
3552 * If the sack array is full, forget about the last one.
3554 if (this_sack >= 4) {
3556 tp->rx_opt.num_sacks--;
3559 for (; this_sack > 0; this_sack--, sp--)
3563 /* Build the new head SACK, and we're done. */
3564 sp->start_seq = seq;
3565 sp->end_seq = end_seq;
3566 tp->rx_opt.num_sacks++;
3567 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3570 /* RCV.NXT advances, some SACKs should be eaten. */
3572 static void tcp_sack_remove(struct tcp_sock *tp)
3574 struct tcp_sack_block *sp = &tp->selective_acks[0];
3575 int num_sacks = tp->rx_opt.num_sacks;
3578 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3579 if (skb_queue_empty(&tp->out_of_order_queue)) {
3580 tp->rx_opt.num_sacks = 0;
3581 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3585 for (this_sack = 0; this_sack < num_sacks; ) {
3586 /* Check if the start of the sack is covered by RCV.NXT. */
3587 if (!before(tp->rcv_nxt, sp->start_seq)) {
3590 /* RCV.NXT must cover all the block! */
3591 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3593 /* Zap this SACK, by moving forward any other SACKS. */
3594 for (i=this_sack+1; i < num_sacks; i++)
3595 tp->selective_acks[i-1] = tp->selective_acks[i];
3602 if (num_sacks != tp->rx_opt.num_sacks) {
3603 tp->rx_opt.num_sacks = num_sacks;
3604 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3608 /* This one checks to see if we can put data from the
3609 * out_of_order queue into the receive_queue.
3611 static void tcp_ofo_queue(struct sock *sk)
3613 struct tcp_sock *tp = tcp_sk(sk);
3614 __u32 dsack_high = tp->rcv_nxt;
3615 struct sk_buff *skb;
3617 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3618 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3621 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3622 __u32 dsack = dsack_high;
3623 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3624 dsack_high = TCP_SKB_CB(skb)->end_seq;
3625 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3628 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3629 SOCK_DEBUG(sk, "ofo packet was already received \n");
3630 __skb_unlink(skb, &tp->out_of_order_queue);
3634 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3635 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3636 TCP_SKB_CB(skb)->end_seq);
3638 __skb_unlink(skb, &tp->out_of_order_queue);
3639 __skb_queue_tail(&sk->sk_receive_queue, skb);
3640 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3641 if (tcp_hdr(skb)->fin)
3642 tcp_fin(skb, sk, tcp_hdr(skb));
3646 static int tcp_prune_queue(struct sock *sk);
3648 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3650 struct tcphdr *th = tcp_hdr(skb);
3651 struct tcp_sock *tp = tcp_sk(sk);
3654 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3657 __skb_pull(skb, th->doff*4);
3659 TCP_ECN_accept_cwr(tp, skb);
3661 if (tp->rx_opt.dsack) {
3662 tp->rx_opt.dsack = 0;
3663 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3664 4 - tp->rx_opt.tstamp_ok);
3667 /* Queue data for delivery to the user.
3668 * Packets in sequence go to the receive queue.
3669 * Out of sequence packets to the out_of_order_queue.
3671 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3672 if (tcp_receive_window(tp) == 0)
3675 /* Ok. In sequence. In window. */
3676 if (tp->ucopy.task == current &&
3677 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3678 sock_owned_by_user(sk) && !tp->urg_data) {
3679 int chunk = min_t(unsigned int, skb->len,
3682 __set_current_state(TASK_RUNNING);
3685 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3686 tp->ucopy.len -= chunk;
3687 tp->copied_seq += chunk;
3688 eaten = (chunk == skb->len && !th->fin);
3689 tcp_rcv_space_adjust(sk);
3697 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3698 !sk_stream_rmem_schedule(sk, skb))) {
3699 if (tcp_prune_queue(sk) < 0 ||
3700 !sk_stream_rmem_schedule(sk, skb))
3703 sk_stream_set_owner_r(skb, sk);
3704 __skb_queue_tail(&sk->sk_receive_queue, skb);
3706 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3708 tcp_event_data_recv(sk, skb);
3710 tcp_fin(skb, sk, th);
3712 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3715 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3716 * gap in queue is filled.
3718 if (skb_queue_empty(&tp->out_of_order_queue))
3719 inet_csk(sk)->icsk_ack.pingpong = 0;
3722 if (tp->rx_opt.num_sacks)
3723 tcp_sack_remove(tp);
3725 tcp_fast_path_check(sk);
3729 else if (!sock_flag(sk, SOCK_DEAD))
3730 sk->sk_data_ready(sk, 0);
3734 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3735 /* A retransmit, 2nd most common case. Force an immediate ack. */
3736 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3737 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3740 tcp_enter_quickack_mode(sk);
3741 inet_csk_schedule_ack(sk);
3747 /* Out of window. F.e. zero window probe. */
3748 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3751 tcp_enter_quickack_mode(sk);
3753 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3754 /* Partial packet, seq < rcv_next < end_seq */
3755 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3756 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3757 TCP_SKB_CB(skb)->end_seq);
3759 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3761 /* If window is closed, drop tail of packet. But after
3762 * remembering D-SACK for its head made in previous line.
3764 if (!tcp_receive_window(tp))
3769 TCP_ECN_check_ce(tp, skb);
3771 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3772 !sk_stream_rmem_schedule(sk, skb)) {
3773 if (tcp_prune_queue(sk) < 0 ||
3774 !sk_stream_rmem_schedule(sk, skb))
3778 /* Disable header prediction. */
3780 inet_csk_schedule_ack(sk);
3782 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3783 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3785 sk_stream_set_owner_r(skb, sk);
3787 if (!skb_peek(&tp->out_of_order_queue)) {
3788 /* Initial out of order segment, build 1 SACK. */
3789 if (tcp_is_sack(tp)) {
3790 tp->rx_opt.num_sacks = 1;
3791 tp->rx_opt.dsack = 0;
3792 tp->rx_opt.eff_sacks = 1;
3793 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3794 tp->selective_acks[0].end_seq =
3795 TCP_SKB_CB(skb)->end_seq;
3797 __skb_queue_head(&tp->out_of_order_queue,skb);
3799 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3800 u32 seq = TCP_SKB_CB(skb)->seq;
3801 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3803 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3804 __skb_append(skb1, skb, &tp->out_of_order_queue);
3806 if (!tp->rx_opt.num_sacks ||
3807 tp->selective_acks[0].end_seq != seq)
3810 /* Common case: data arrive in order after hole. */
3811 tp->selective_acks[0].end_seq = end_seq;
3815 /* Find place to insert this segment. */
3817 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3819 } while ((skb1 = skb1->prev) !=
3820 (struct sk_buff*)&tp->out_of_order_queue);
3822 /* Do skb overlap to previous one? */
3823 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3824 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3825 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3826 /* All the bits are present. Drop. */
3828 tcp_dsack_set(tp, seq, end_seq);
3831 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3832 /* Partial overlap. */
3833 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3838 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3840 /* And clean segments covered by new one as whole. */
3841 while ((skb1 = skb->next) !=
3842 (struct sk_buff*)&tp->out_of_order_queue &&
3843 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3844 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3845 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3848 __skb_unlink(skb1, &tp->out_of_order_queue);
3849 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3854 if (tcp_is_sack(tp))
3855 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3859 /* Collapse contiguous sequence of skbs head..tail with
3860 * sequence numbers start..end.
3861 * Segments with FIN/SYN are not collapsed (only because this
3865 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3866 struct sk_buff *head, struct sk_buff *tail,
3869 struct sk_buff *skb;
3871 /* First, check that queue is collapsible and find
3872 * the point where collapsing can be useful. */
3873 for (skb = head; skb != tail; ) {
3874 /* No new bits? It is possible on ofo queue. */
3875 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3876 struct sk_buff *next = skb->next;
3877 __skb_unlink(skb, list);
3879 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3884 /* The first skb to collapse is:
3886 * - bloated or contains data before "start" or
3887 * overlaps to the next one.
3889 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3890 (tcp_win_from_space(skb->truesize) > skb->len ||
3891 before(TCP_SKB_CB(skb)->seq, start) ||
3892 (skb->next != tail &&
3893 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3896 /* Decided to skip this, advance start seq. */
3897 start = TCP_SKB_CB(skb)->end_seq;
3900 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3903 while (before(start, end)) {
3904 struct sk_buff *nskb;
3905 int header = skb_headroom(skb);
3906 int copy = SKB_MAX_ORDER(header, 0);
3908 /* Too big header? This can happen with IPv6. */
3911 if (end-start < copy)
3913 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3917 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3918 skb_set_network_header(nskb, (skb_network_header(skb) -
3920 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3922 skb_reserve(nskb, header);
3923 memcpy(nskb->head, skb->head, header);
3924 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3925 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3926 __skb_insert(nskb, skb->prev, skb, list);
3927 sk_stream_set_owner_r(nskb, sk);
3929 /* Copy data, releasing collapsed skbs. */
3931 int offset = start - TCP_SKB_CB(skb)->seq;
3932 int size = TCP_SKB_CB(skb)->end_seq - start;
3936 size = min(copy, size);
3937 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3939 TCP_SKB_CB(nskb)->end_seq += size;
3943 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3944 struct sk_buff *next = skb->next;
3945 __skb_unlink(skb, list);
3947 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3950 tcp_hdr(skb)->syn ||
3958 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3959 * and tcp_collapse() them until all the queue is collapsed.
3961 static void tcp_collapse_ofo_queue(struct sock *sk)
3963 struct tcp_sock *tp = tcp_sk(sk);
3964 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3965 struct sk_buff *head;
3971 start = TCP_SKB_CB(skb)->seq;
3972 end = TCP_SKB_CB(skb)->end_seq;
3978 /* Segment is terminated when we see gap or when
3979 * we are at the end of all the queue. */
3980 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3981 after(TCP_SKB_CB(skb)->seq, end) ||
3982 before(TCP_SKB_CB(skb)->end_seq, start)) {
3983 tcp_collapse(sk, &tp->out_of_order_queue,
3984 head, skb, start, end);
3986 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3988 /* Start new segment */
3989 start = TCP_SKB_CB(skb)->seq;
3990 end = TCP_SKB_CB(skb)->end_seq;
3992 if (before(TCP_SKB_CB(skb)->seq, start))
3993 start = TCP_SKB_CB(skb)->seq;
3994 if (after(TCP_SKB_CB(skb)->end_seq, end))
3995 end = TCP_SKB_CB(skb)->end_seq;
4000 /* Reduce allocated memory if we can, trying to get
4001 * the socket within its memory limits again.
4003 * Return less than zero if we should start dropping frames
4004 * until the socket owning process reads some of the data
4005 * to stabilize the situation.
4007 static int tcp_prune_queue(struct sock *sk)
4009 struct tcp_sock *tp = tcp_sk(sk);
4011 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4013 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
4015 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4016 tcp_clamp_window(sk);
4017 else if (tcp_memory_pressure)
4018 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4020 tcp_collapse_ofo_queue(sk);
4021 tcp_collapse(sk, &sk->sk_receive_queue,
4022 sk->sk_receive_queue.next,
4023 (struct sk_buff*)&sk->sk_receive_queue,
4024 tp->copied_seq, tp->rcv_nxt);
4025 sk_stream_mem_reclaim(sk);
4027 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4030 /* Collapsing did not help, destructive actions follow.
4031 * This must not ever occur. */
4033 /* First, purge the out_of_order queue. */
4034 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4035 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
4036 __skb_queue_purge(&tp->out_of_order_queue);
4038 /* Reset SACK state. A conforming SACK implementation will
4039 * do the same at a timeout based retransmit. When a connection
4040 * is in a sad state like this, we care only about integrity
4041 * of the connection not performance.
4043 if (tcp_is_sack(tp))
4044 tcp_sack_reset(&tp->rx_opt);
4045 sk_stream_mem_reclaim(sk);
4048 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4051 /* If we are really being abused, tell the caller to silently
4052 * drop receive data on the floor. It will get retransmitted
4053 * and hopefully then we'll have sufficient space.
4055 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4057 /* Massive buffer overcommit. */
4063 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4064 * As additional protections, we do not touch cwnd in retransmission phases,
4065 * and if application hit its sndbuf limit recently.
4067 void tcp_cwnd_application_limited(struct sock *sk)
4069 struct tcp_sock *tp = tcp_sk(sk);
4071 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4072 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4073 /* Limited by application or receiver window. */
4074 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4075 u32 win_used = max(tp->snd_cwnd_used, init_win);
4076 if (win_used < tp->snd_cwnd) {
4077 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4078 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4080 tp->snd_cwnd_used = 0;
4082 tp->snd_cwnd_stamp = tcp_time_stamp;
4085 static int tcp_should_expand_sndbuf(struct sock *sk)
4087 struct tcp_sock *tp = tcp_sk(sk);
4089 /* If the user specified a specific send buffer setting, do
4092 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4095 /* If we are under global TCP memory pressure, do not expand. */
4096 if (tcp_memory_pressure)
4099 /* If we are under soft global TCP memory pressure, do not expand. */
4100 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4103 /* If we filled the congestion window, do not expand. */
4104 if (tp->packets_out >= tp->snd_cwnd)
4110 /* When incoming ACK allowed to free some skb from write_queue,
4111 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4112 * on the exit from tcp input handler.
4114 * PROBLEM: sndbuf expansion does not work well with largesend.
4116 static void tcp_new_space(struct sock *sk)
4118 struct tcp_sock *tp = tcp_sk(sk);
4120 if (tcp_should_expand_sndbuf(sk)) {
4121 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4122 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4123 demanded = max_t(unsigned int, tp->snd_cwnd,
4124 tp->reordering + 1);
4125 sndmem *= 2*demanded;
4126 if (sndmem > sk->sk_sndbuf)
4127 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4128 tp->snd_cwnd_stamp = tcp_time_stamp;
4131 sk->sk_write_space(sk);
4134 static void tcp_check_space(struct sock *sk)
4136 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4137 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4138 if (sk->sk_socket &&
4139 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4144 static inline void tcp_data_snd_check(struct sock *sk)
4146 tcp_push_pending_frames(sk);
4147 tcp_check_space(sk);
4151 * Check if sending an ack is needed.
4153 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4155 struct tcp_sock *tp = tcp_sk(sk);
4157 /* More than one full frame received... */
4158 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4159 /* ... and right edge of window advances far enough.
4160 * (tcp_recvmsg() will send ACK otherwise). Or...
4162 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4163 /* We ACK each frame or... */
4164 tcp_in_quickack_mode(sk) ||
4165 /* We have out of order data. */
4167 skb_peek(&tp->out_of_order_queue))) {
4168 /* Then ack it now */
4171 /* Else, send delayed ack. */
4172 tcp_send_delayed_ack(sk);
4176 static inline void tcp_ack_snd_check(struct sock *sk)
4178 if (!inet_csk_ack_scheduled(sk)) {
4179 /* We sent a data segment already. */
4182 __tcp_ack_snd_check(sk, 1);
4186 * This routine is only called when we have urgent data
4187 * signaled. Its the 'slow' part of tcp_urg. It could be
4188 * moved inline now as tcp_urg is only called from one
4189 * place. We handle URGent data wrong. We have to - as
4190 * BSD still doesn't use the correction from RFC961.
4191 * For 1003.1g we should support a new option TCP_STDURG to permit
4192 * either form (or just set the sysctl tcp_stdurg).
4195 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4197 struct tcp_sock *tp = tcp_sk(sk);
4198 u32 ptr = ntohs(th->urg_ptr);
4200 if (ptr && !sysctl_tcp_stdurg)
4202 ptr += ntohl(th->seq);
4204 /* Ignore urgent data that we've already seen and read. */
4205 if (after(tp->copied_seq, ptr))
4208 /* Do not replay urg ptr.
4210 * NOTE: interesting situation not covered by specs.
4211 * Misbehaving sender may send urg ptr, pointing to segment,
4212 * which we already have in ofo queue. We are not able to fetch
4213 * such data and will stay in TCP_URG_NOTYET until will be eaten
4214 * by recvmsg(). Seems, we are not obliged to handle such wicked
4215 * situations. But it is worth to think about possibility of some
4216 * DoSes using some hypothetical application level deadlock.
4218 if (before(ptr, tp->rcv_nxt))
4221 /* Do we already have a newer (or duplicate) urgent pointer? */
4222 if (tp->urg_data && !after(ptr, tp->urg_seq))
4225 /* Tell the world about our new urgent pointer. */
4228 /* We may be adding urgent data when the last byte read was
4229 * urgent. To do this requires some care. We cannot just ignore
4230 * tp->copied_seq since we would read the last urgent byte again
4231 * as data, nor can we alter copied_seq until this data arrives
4232 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4234 * NOTE. Double Dutch. Rendering to plain English: author of comment
4235 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4236 * and expect that both A and B disappear from stream. This is _wrong_.
4237 * Though this happens in BSD with high probability, this is occasional.
4238 * Any application relying on this is buggy. Note also, that fix "works"
4239 * only in this artificial test. Insert some normal data between A and B and we will
4240 * decline of BSD again. Verdict: it is better to remove to trap
4243 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4244 !sock_flag(sk, SOCK_URGINLINE) &&
4245 tp->copied_seq != tp->rcv_nxt) {
4246 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4248 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4249 __skb_unlink(skb, &sk->sk_receive_queue);
4254 tp->urg_data = TCP_URG_NOTYET;
4257 /* Disable header prediction. */
4261 /* This is the 'fast' part of urgent handling. */
4262 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4264 struct tcp_sock *tp = tcp_sk(sk);
4266 /* Check if we get a new urgent pointer - normally not. */
4268 tcp_check_urg(sk,th);
4270 /* Do we wait for any urgent data? - normally not... */
4271 if (tp->urg_data == TCP_URG_NOTYET) {
4272 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4275 /* Is the urgent pointer pointing into this packet? */
4276 if (ptr < skb->len) {
4278 if (skb_copy_bits(skb, ptr, &tmp, 1))
4280 tp->urg_data = TCP_URG_VALID | tmp;
4281 if (!sock_flag(sk, SOCK_DEAD))
4282 sk->sk_data_ready(sk, 0);
4287 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4289 struct tcp_sock *tp = tcp_sk(sk);
4290 int chunk = skb->len - hlen;
4294 if (skb_csum_unnecessary(skb))
4295 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4297 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4301 tp->ucopy.len -= chunk;
4302 tp->copied_seq += chunk;
4303 tcp_rcv_space_adjust(sk);
4310 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4314 if (sock_owned_by_user(sk)) {
4316 result = __tcp_checksum_complete(skb);
4319 result = __tcp_checksum_complete(skb);
4324 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4326 return !skb_csum_unnecessary(skb) &&
4327 __tcp_checksum_complete_user(sk, skb);
4330 #ifdef CONFIG_NET_DMA
4331 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4333 struct tcp_sock *tp = tcp_sk(sk);
4334 int chunk = skb->len - hlen;
4336 int copied_early = 0;
4338 if (tp->ucopy.wakeup)
4341 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4342 tp->ucopy.dma_chan = get_softnet_dma();
4344 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4346 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4347 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4352 tp->ucopy.dma_cookie = dma_cookie;
4355 tp->ucopy.len -= chunk;
4356 tp->copied_seq += chunk;
4357 tcp_rcv_space_adjust(sk);
4359 if ((tp->ucopy.len == 0) ||
4360 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4361 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4362 tp->ucopy.wakeup = 1;
4363 sk->sk_data_ready(sk, 0);
4365 } else if (chunk > 0) {
4366 tp->ucopy.wakeup = 1;
4367 sk->sk_data_ready(sk, 0);
4370 return copied_early;
4372 #endif /* CONFIG_NET_DMA */
4375 * TCP receive function for the ESTABLISHED state.
4377 * It is split into a fast path and a slow path. The fast path is
4379 * - A zero window was announced from us - zero window probing
4380 * is only handled properly in the slow path.
4381 * - Out of order segments arrived.
4382 * - Urgent data is expected.
4383 * - There is no buffer space left
4384 * - Unexpected TCP flags/window values/header lengths are received
4385 * (detected by checking the TCP header against pred_flags)
4386 * - Data is sent in both directions. Fast path only supports pure senders
4387 * or pure receivers (this means either the sequence number or the ack
4388 * value must stay constant)
4389 * - Unexpected TCP option.
4391 * When these conditions are not satisfied it drops into a standard
4392 * receive procedure patterned after RFC793 to handle all cases.
4393 * The first three cases are guaranteed by proper pred_flags setting,
4394 * the rest is checked inline. Fast processing is turned on in
4395 * tcp_data_queue when everything is OK.
4397 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4398 struct tcphdr *th, unsigned len)
4400 struct tcp_sock *tp = tcp_sk(sk);
4403 * Header prediction.
4404 * The code loosely follows the one in the famous
4405 * "30 instruction TCP receive" Van Jacobson mail.
4407 * Van's trick is to deposit buffers into socket queue
4408 * on a device interrupt, to call tcp_recv function
4409 * on the receive process context and checksum and copy
4410 * the buffer to user space. smart...
4412 * Our current scheme is not silly either but we take the
4413 * extra cost of the net_bh soft interrupt processing...
4414 * We do checksum and copy also but from device to kernel.
4417 tp->rx_opt.saw_tstamp = 0;
4419 /* pred_flags is 0xS?10 << 16 + snd_wnd
4420 * if header_prediction is to be made
4421 * 'S' will always be tp->tcp_header_len >> 2
4422 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4423 * turn it off (when there are holes in the receive
4424 * space for instance)
4425 * PSH flag is ignored.
4428 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4429 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4430 int tcp_header_len = tp->tcp_header_len;
4432 /* Timestamp header prediction: tcp_header_len
4433 * is automatically equal to th->doff*4 due to pred_flags
4437 /* Check timestamp */
4438 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4439 __be32 *ptr = (__be32 *)(th + 1);
4441 /* No? Slow path! */
4442 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4443 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4446 tp->rx_opt.saw_tstamp = 1;
4448 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4450 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4452 /* If PAWS failed, check it more carefully in slow path */
4453 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4456 /* DO NOT update ts_recent here, if checksum fails
4457 * and timestamp was corrupted part, it will result
4458 * in a hung connection since we will drop all
4459 * future packets due to the PAWS test.
4463 if (len <= tcp_header_len) {
4464 /* Bulk data transfer: sender */
4465 if (len == tcp_header_len) {
4466 /* Predicted packet is in window by definition.
4467 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4468 * Hence, check seq<=rcv_wup reduces to:
4470 if (tcp_header_len ==
4471 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4472 tp->rcv_nxt == tp->rcv_wup)
4473 tcp_store_ts_recent(tp);
4475 /* We know that such packets are checksummed
4478 tcp_ack(sk, skb, 0);
4480 tcp_data_snd_check(sk);
4482 } else { /* Header too small */
4483 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4488 int copied_early = 0;
4490 if (tp->copied_seq == tp->rcv_nxt &&
4491 len - tcp_header_len <= tp->ucopy.len) {
4492 #ifdef CONFIG_NET_DMA
4493 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4498 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4499 __set_current_state(TASK_RUNNING);
4501 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4505 /* Predicted packet is in window by definition.
4506 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4507 * Hence, check seq<=rcv_wup reduces to:
4509 if (tcp_header_len ==
4510 (sizeof(struct tcphdr) +
4511 TCPOLEN_TSTAMP_ALIGNED) &&
4512 tp->rcv_nxt == tp->rcv_wup)
4513 tcp_store_ts_recent(tp);
4515 tcp_rcv_rtt_measure_ts(sk, skb);
4517 __skb_pull(skb, tcp_header_len);
4518 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4519 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4522 tcp_cleanup_rbuf(sk, skb->len);
4525 if (tcp_checksum_complete_user(sk, skb))
4528 /* Predicted packet is in window by definition.
4529 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4530 * Hence, check seq<=rcv_wup reduces to:
4532 if (tcp_header_len ==
4533 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4534 tp->rcv_nxt == tp->rcv_wup)
4535 tcp_store_ts_recent(tp);
4537 tcp_rcv_rtt_measure_ts(sk, skb);
4539 if ((int)skb->truesize > sk->sk_forward_alloc)
4542 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4544 /* Bulk data transfer: receiver */
4545 __skb_pull(skb,tcp_header_len);
4546 __skb_queue_tail(&sk->sk_receive_queue, skb);
4547 sk_stream_set_owner_r(skb, sk);
4548 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4551 tcp_event_data_recv(sk, skb);
4553 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4554 /* Well, only one small jumplet in fast path... */
4555 tcp_ack(sk, skb, FLAG_DATA);
4556 tcp_data_snd_check(sk);
4557 if (!inet_csk_ack_scheduled(sk))
4561 __tcp_ack_snd_check(sk, 0);
4563 #ifdef CONFIG_NET_DMA
4565 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4571 sk->sk_data_ready(sk, 0);
4577 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4581 * RFC1323: H1. Apply PAWS check first.
4583 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4584 tcp_paws_discard(sk, skb)) {
4586 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4587 tcp_send_dupack(sk, skb);
4590 /* Resets are accepted even if PAWS failed.
4592 ts_recent update must be made after we are sure
4593 that the packet is in window.
4598 * Standard slow path.
4601 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4602 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4603 * (RST) segments are validated by checking their SEQ-fields."
4604 * And page 69: "If an incoming segment is not acceptable,
4605 * an acknowledgment should be sent in reply (unless the RST bit
4606 * is set, if so drop the segment and return)".
4609 tcp_send_dupack(sk, skb);
4618 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4620 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4621 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4622 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4629 tcp_ack(sk, skb, FLAG_SLOWPATH);
4631 tcp_rcv_rtt_measure_ts(sk, skb);
4633 /* Process urgent data. */
4634 tcp_urg(sk, skb, th);
4636 /* step 7: process the segment text */
4637 tcp_data_queue(sk, skb);
4639 tcp_data_snd_check(sk);
4640 tcp_ack_snd_check(sk);
4644 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4651 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4652 struct tcphdr *th, unsigned len)
4654 struct tcp_sock *tp = tcp_sk(sk);
4655 struct inet_connection_sock *icsk = inet_csk(sk);
4656 int saved_clamp = tp->rx_opt.mss_clamp;
4658 tcp_parse_options(skb, &tp->rx_opt, 0);
4662 * "If the state is SYN-SENT then
4663 * first check the ACK bit
4664 * If the ACK bit is set
4665 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4666 * a reset (unless the RST bit is set, if so drop
4667 * the segment and return)"
4669 * We do not send data with SYN, so that RFC-correct
4672 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4673 goto reset_and_undo;
4675 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4676 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4678 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4679 goto reset_and_undo;
4682 /* Now ACK is acceptable.
4684 * "If the RST bit is set
4685 * If the ACK was acceptable then signal the user "error:
4686 * connection reset", drop the segment, enter CLOSED state,
4687 * delete TCB, and return."
4696 * "fifth, if neither of the SYN or RST bits is set then
4697 * drop the segment and return."
4703 goto discard_and_undo;
4706 * "If the SYN bit is on ...
4707 * are acceptable then ...
4708 * (our SYN has been ACKed), change the connection
4709 * state to ESTABLISHED..."
4712 TCP_ECN_rcv_synack(tp, th);
4714 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4715 tcp_ack(sk, skb, FLAG_SLOWPATH);
4717 /* Ok.. it's good. Set up sequence numbers and
4718 * move to established.
4720 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4721 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4723 /* RFC1323: The window in SYN & SYN/ACK segments is
4726 tp->snd_wnd = ntohs(th->window);
4727 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4729 if (!tp->rx_opt.wscale_ok) {
4730 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4731 tp->window_clamp = min(tp->window_clamp, 65535U);
4734 if (tp->rx_opt.saw_tstamp) {
4735 tp->rx_opt.tstamp_ok = 1;
4736 tp->tcp_header_len =
4737 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4738 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4739 tcp_store_ts_recent(tp);
4741 tp->tcp_header_len = sizeof(struct tcphdr);
4744 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4745 tcp_enable_fack(tp);
4748 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4749 tcp_initialize_rcv_mss(sk);
4751 /* Remember, tcp_poll() does not lock socket!
4752 * Change state from SYN-SENT only after copied_seq
4753 * is initialized. */
4754 tp->copied_seq = tp->rcv_nxt;
4756 tcp_set_state(sk, TCP_ESTABLISHED);
4758 security_inet_conn_established(sk, skb);
4760 /* Make sure socket is routed, for correct metrics. */
4761 icsk->icsk_af_ops->rebuild_header(sk);
4763 tcp_init_metrics(sk);
4765 tcp_init_congestion_control(sk);
4767 /* Prevent spurious tcp_cwnd_restart() on first data
4770 tp->lsndtime = tcp_time_stamp;
4772 tcp_init_buffer_space(sk);
4774 if (sock_flag(sk, SOCK_KEEPOPEN))
4775 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4777 if (!tp->rx_opt.snd_wscale)
4778 __tcp_fast_path_on(tp, tp->snd_wnd);
4782 if (!sock_flag(sk, SOCK_DEAD)) {
4783 sk->sk_state_change(sk);
4784 sk_wake_async(sk, 0, POLL_OUT);
4787 if (sk->sk_write_pending ||
4788 icsk->icsk_accept_queue.rskq_defer_accept ||
4789 icsk->icsk_ack.pingpong) {
4790 /* Save one ACK. Data will be ready after
4791 * several ticks, if write_pending is set.
4793 * It may be deleted, but with this feature tcpdumps
4794 * look so _wonderfully_ clever, that I was not able
4795 * to stand against the temptation 8) --ANK
4797 inet_csk_schedule_ack(sk);
4798 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4799 icsk->icsk_ack.ato = TCP_ATO_MIN;
4800 tcp_incr_quickack(sk);
4801 tcp_enter_quickack_mode(sk);
4802 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4803 TCP_DELACK_MAX, TCP_RTO_MAX);
4814 /* No ACK in the segment */
4818 * "If the RST bit is set
4820 * Otherwise (no ACK) drop the segment and return."
4823 goto discard_and_undo;
4827 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4828 goto discard_and_undo;
4831 /* We see SYN without ACK. It is attempt of
4832 * simultaneous connect with crossed SYNs.
4833 * Particularly, it can be connect to self.
4835 tcp_set_state(sk, TCP_SYN_RECV);
4837 if (tp->rx_opt.saw_tstamp) {
4838 tp->rx_opt.tstamp_ok = 1;
4839 tcp_store_ts_recent(tp);
4840 tp->tcp_header_len =
4841 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4843 tp->tcp_header_len = sizeof(struct tcphdr);
4846 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4847 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4849 /* RFC1323: The window in SYN & SYN/ACK segments is
4852 tp->snd_wnd = ntohs(th->window);
4853 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4854 tp->max_window = tp->snd_wnd;
4856 TCP_ECN_rcv_syn(tp, th);
4859 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4860 tcp_initialize_rcv_mss(sk);
4863 tcp_send_synack(sk);
4865 /* Note, we could accept data and URG from this segment.
4866 * There are no obstacles to make this.
4868 * However, if we ignore data in ACKless segments sometimes,
4869 * we have no reasons to accept it sometimes.
4870 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4871 * is not flawless. So, discard packet for sanity.
4872 * Uncomment this return to process the data.
4879 /* "fifth, if neither of the SYN or RST bits is set then
4880 * drop the segment and return."
4884 tcp_clear_options(&tp->rx_opt);
4885 tp->rx_opt.mss_clamp = saved_clamp;
4889 tcp_clear_options(&tp->rx_opt);
4890 tp->rx_opt.mss_clamp = saved_clamp;
4896 * This function implements the receiving procedure of RFC 793 for
4897 * all states except ESTABLISHED and TIME_WAIT.
4898 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4899 * address independent.
4902 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4903 struct tcphdr *th, unsigned len)
4905 struct tcp_sock *tp = tcp_sk(sk);
4906 struct inet_connection_sock *icsk = inet_csk(sk);
4909 tp->rx_opt.saw_tstamp = 0;
4911 switch (sk->sk_state) {
4923 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4926 /* Now we have several options: In theory there is
4927 * nothing else in the frame. KA9Q has an option to
4928 * send data with the syn, BSD accepts data with the
4929 * syn up to the [to be] advertised window and
4930 * Solaris 2.1 gives you a protocol error. For now
4931 * we just ignore it, that fits the spec precisely
4932 * and avoids incompatibilities. It would be nice in
4933 * future to drop through and process the data.
4935 * Now that TTCP is starting to be used we ought to
4937 * But, this leaves one open to an easy denial of
4938 * service attack, and SYN cookies can't defend
4939 * against this problem. So, we drop the data
4940 * in the interest of security over speed unless
4941 * it's still in use.
4949 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4953 /* Do step6 onward by hand. */
4954 tcp_urg(sk, skb, th);
4956 tcp_data_snd_check(sk);
4960 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4961 tcp_paws_discard(sk, skb)) {
4963 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4964 tcp_send_dupack(sk, skb);
4967 /* Reset is accepted even if it did not pass PAWS. */
4970 /* step 1: check sequence number */
4971 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4973 tcp_send_dupack(sk, skb);
4977 /* step 2: check RST bit */
4983 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4985 /* step 3: check security and precedence [ignored] */
4989 * Check for a SYN in window.
4991 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4992 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4997 /* step 5: check the ACK field */
4999 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5001 switch (sk->sk_state) {
5004 tp->copied_seq = tp->rcv_nxt;
5006 tcp_set_state(sk, TCP_ESTABLISHED);
5007 sk->sk_state_change(sk);
5009 /* Note, that this wakeup is only for marginal
5010 * crossed SYN case. Passively open sockets
5011 * are not waked up, because sk->sk_sleep ==
5012 * NULL and sk->sk_socket == NULL.
5014 if (sk->sk_socket) {
5015 sk_wake_async(sk,0,POLL_OUT);
5018 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5019 tp->snd_wnd = ntohs(th->window) <<
5020 tp->rx_opt.snd_wscale;
5021 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5022 TCP_SKB_CB(skb)->seq);
5024 /* tcp_ack considers this ACK as duplicate
5025 * and does not calculate rtt.
5026 * Fix it at least with timestamps.
5028 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5030 tcp_ack_saw_tstamp(sk, 0);
5032 if (tp->rx_opt.tstamp_ok)
5033 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5035 /* Make sure socket is routed, for
5038 icsk->icsk_af_ops->rebuild_header(sk);
5040 tcp_init_metrics(sk);
5042 tcp_init_congestion_control(sk);
5044 /* Prevent spurious tcp_cwnd_restart() on
5045 * first data packet.
5047 tp->lsndtime = tcp_time_stamp;
5050 tcp_initialize_rcv_mss(sk);
5051 tcp_init_buffer_space(sk);
5052 tcp_fast_path_on(tp);
5059 if (tp->snd_una == tp->write_seq) {
5060 tcp_set_state(sk, TCP_FIN_WAIT2);
5061 sk->sk_shutdown |= SEND_SHUTDOWN;
5062 dst_confirm(sk->sk_dst_cache);
5064 if (!sock_flag(sk, SOCK_DEAD))
5065 /* Wake up lingering close() */
5066 sk->sk_state_change(sk);
5070 if (tp->linger2 < 0 ||
5071 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5072 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5074 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5078 tmo = tcp_fin_time(sk);
5079 if (tmo > TCP_TIMEWAIT_LEN) {
5080 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5081 } else if (th->fin || sock_owned_by_user(sk)) {
5082 /* Bad case. We could lose such FIN otherwise.
5083 * It is not a big problem, but it looks confusing
5084 * and not so rare event. We still can lose it now,
5085 * if it spins in bh_lock_sock(), but it is really
5088 inet_csk_reset_keepalive_timer(sk, tmo);
5090 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5098 if (tp->snd_una == tp->write_seq) {
5099 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5105 if (tp->snd_una == tp->write_seq) {
5106 tcp_update_metrics(sk);
5115 /* step 6: check the URG bit */
5116 tcp_urg(sk, skb, th);
5118 /* step 7: process the segment text */
5119 switch (sk->sk_state) {
5120 case TCP_CLOSE_WAIT:
5123 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5127 /* RFC 793 says to queue data in these states,
5128 * RFC 1122 says we MUST send a reset.
5129 * BSD 4.4 also does reset.
5131 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5132 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5133 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5134 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5140 case TCP_ESTABLISHED:
5141 tcp_data_queue(sk, skb);
5146 /* tcp_data could move socket to TIME-WAIT */
5147 if (sk->sk_state != TCP_CLOSE) {
5148 tcp_data_snd_check(sk);
5149 tcp_ack_snd_check(sk);
5159 EXPORT_SYMBOL(sysctl_tcp_ecn);
5160 EXPORT_SYMBOL(sysctl_tcp_reordering);
5161 EXPORT_SYMBOL(tcp_parse_options);
5162 EXPORT_SYMBOL(tcp_rcv_established);
5163 EXPORT_SYMBOL(tcp_rcv_state_process);
5164 EXPORT_SYMBOL(tcp_initialize_rcv_mss);