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 static int tcp_mark_lost_retrans(struct sock *sk, u32 lost_retrans)
1115 struct tcp_sock *tp = tcp_sk(sk);
1116 struct sk_buff *skb;
1119 tcp_for_write_queue(skb, sk) {
1120 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1122 if (skb == tcp_send_head(sk))
1124 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1126 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1129 if ((TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) &&
1130 after(lost_retrans, ack_seq) &&
1132 !before(lost_retrans,
1133 ack_seq + tp->reordering * tp->mss_cache))) {
1134 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1135 tp->retrans_out -= tcp_skb_pcount(skb);
1137 /* clear lost hint */
1138 tp->retransmit_skb_hint = NULL;
1140 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1141 tp->lost_out += tcp_skb_pcount(skb);
1142 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1143 flag |= FLAG_DATA_SACKED;
1144 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1151 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1152 struct tcp_sack_block_wire *sp, int num_sacks,
1155 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1156 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1159 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1162 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1163 } else if (num_sacks > 1) {
1164 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1165 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1167 if (!after(end_seq_0, end_seq_1) &&
1168 !before(start_seq_0, start_seq_1)) {
1171 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1175 /* D-SACK for already forgotten data... Do dumb counting. */
1177 !after(end_seq_0, prior_snd_una) &&
1178 after(end_seq_0, tp->undo_marker))
1185 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1187 const struct inet_connection_sock *icsk = inet_csk(sk);
1188 struct tcp_sock *tp = tcp_sk(sk);
1189 unsigned char *ptr = (skb_transport_header(ack_skb) +
1190 TCP_SKB_CB(ack_skb)->sacked);
1191 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1192 struct sk_buff *cached_skb;
1193 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1194 int reord = tp->packets_out;
1196 u32 lost_retrans = 0;
1198 int found_dup_sack = 0;
1199 int cached_fack_count;
1201 int first_sack_index;
1203 if (!tp->sacked_out) {
1204 if (WARN_ON(tp->fackets_out))
1205 tp->fackets_out = 0;
1206 tp->highest_sack = tp->snd_una;
1208 prior_fackets = tp->fackets_out;
1210 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1211 num_sacks, prior_snd_una);
1213 flag |= FLAG_DSACKING_ACK;
1215 /* Eliminate too old ACKs, but take into
1216 * account more or less fresh ones, they can
1217 * contain valid SACK info.
1219 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1223 * if the only SACK change is the increase of the end_seq of
1224 * the first block then only apply that SACK block
1225 * and use retrans queue hinting otherwise slowpath */
1227 for (i = 0; i < num_sacks; i++) {
1228 __be32 start_seq = sp[i].start_seq;
1229 __be32 end_seq = sp[i].end_seq;
1232 if (tp->recv_sack_cache[i].start_seq != start_seq)
1235 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1236 (tp->recv_sack_cache[i].end_seq != end_seq))
1239 tp->recv_sack_cache[i].start_seq = start_seq;
1240 tp->recv_sack_cache[i].end_seq = end_seq;
1242 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1243 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1244 tp->recv_sack_cache[i].start_seq = 0;
1245 tp->recv_sack_cache[i].end_seq = 0;
1248 first_sack_index = 0;
1253 tp->fastpath_skb_hint = NULL;
1255 /* order SACK blocks to allow in order walk of the retrans queue */
1256 for (i = num_sacks-1; i > 0; i--) {
1257 for (j = 0; j < i; j++){
1258 if (after(ntohl(sp[j].start_seq),
1259 ntohl(sp[j+1].start_seq))){
1260 struct tcp_sack_block_wire tmp;
1266 /* Track where the first SACK block goes to */
1267 if (j == first_sack_index)
1268 first_sack_index = j+1;
1275 /* clear flag as used for different purpose in following code */
1278 /* Use SACK fastpath hint if valid */
1279 cached_skb = tp->fastpath_skb_hint;
1280 cached_fack_count = tp->fastpath_cnt_hint;
1282 cached_skb = tcp_write_queue_head(sk);
1283 cached_fack_count = 0;
1286 for (i=0; i<num_sacks; i++, sp++) {
1287 struct sk_buff *skb;
1288 __u32 start_seq = ntohl(sp->start_seq);
1289 __u32 end_seq = ntohl(sp->end_seq);
1291 int dup_sack = (found_dup_sack && (i == first_sack_index));
1293 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) {
1295 if (!tp->undo_marker)
1296 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1298 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1300 /* Don't count olds caused by ACK reordering */
1301 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1302 !after(end_seq, tp->snd_una))
1304 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1310 fack_count = cached_fack_count;
1312 /* Event "B" in the comment above. */
1313 if (after(end_seq, tp->high_seq))
1314 flag |= FLAG_DATA_LOST;
1316 tcp_for_write_queue_from(skb, sk) {
1317 int in_sack, pcount;
1320 if (skb == tcp_send_head(sk))
1324 cached_fack_count = fack_count;
1325 if (i == first_sack_index) {
1326 tp->fastpath_skb_hint = skb;
1327 tp->fastpath_cnt_hint = fack_count;
1330 /* The retransmission queue is always in order, so
1331 * we can short-circuit the walk early.
1333 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1336 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1337 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1339 pcount = tcp_skb_pcount(skb);
1341 if (pcount > 1 && !in_sack &&
1342 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1343 unsigned int pkt_len;
1345 in_sack = !after(start_seq,
1346 TCP_SKB_CB(skb)->seq);
1349 pkt_len = (start_seq -
1350 TCP_SKB_CB(skb)->seq);
1352 pkt_len = (end_seq -
1353 TCP_SKB_CB(skb)->seq);
1354 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1356 pcount = tcp_skb_pcount(skb);
1359 fack_count += pcount;
1361 sacked = TCP_SKB_CB(skb)->sacked;
1363 /* Account D-SACK for retransmitted packet. */
1364 if ((dup_sack && in_sack) &&
1365 (sacked & TCPCB_RETRANS) &&
1366 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1369 /* The frame is ACKed. */
1370 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1371 if (sacked&TCPCB_RETRANS) {
1372 if ((dup_sack && in_sack) &&
1373 (sacked&TCPCB_SACKED_ACKED))
1374 reord = min(fack_count, reord);
1376 /* If it was in a hole, we detected reordering. */
1377 if (fack_count < prior_fackets &&
1378 !(sacked&TCPCB_SACKED_ACKED))
1379 reord = min(fack_count, reord);
1382 /* Nothing to do; acked frame is about to be dropped. */
1386 if ((sacked&TCPCB_SACKED_RETRANS) &&
1387 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1388 (!lost_retrans || after(end_seq, lost_retrans)))
1389 lost_retrans = end_seq;
1394 if (!(sacked&TCPCB_SACKED_ACKED)) {
1395 if (sacked & TCPCB_SACKED_RETRANS) {
1396 /* If the segment is not tagged as lost,
1397 * we do not clear RETRANS, believing
1398 * that retransmission is still in flight.
1400 if (sacked & TCPCB_LOST) {
1401 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1402 tp->lost_out -= tcp_skb_pcount(skb);
1403 tp->retrans_out -= tcp_skb_pcount(skb);
1405 /* clear lost hint */
1406 tp->retransmit_skb_hint = NULL;
1409 /* New sack for not retransmitted frame,
1410 * which was in hole. It is reordering.
1412 if (!(sacked & TCPCB_RETRANS) &&
1413 fack_count < prior_fackets)
1414 reord = min(fack_count, reord);
1416 if (sacked & TCPCB_LOST) {
1417 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1418 tp->lost_out -= tcp_skb_pcount(skb);
1420 /* clear lost hint */
1421 tp->retransmit_skb_hint = NULL;
1423 /* SACK enhanced F-RTO detection.
1424 * Set flag if and only if non-rexmitted
1425 * segments below frto_highmark are
1426 * SACKed (RFC4138; Appendix B).
1427 * Clearing correct due to in-order walk
1429 if (after(end_seq, tp->frto_highmark)) {
1430 flag &= ~FLAG_ONLY_ORIG_SACKED;
1432 if (!(sacked & TCPCB_RETRANS))
1433 flag |= FLAG_ONLY_ORIG_SACKED;
1437 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1438 flag |= FLAG_DATA_SACKED;
1439 tp->sacked_out += tcp_skb_pcount(skb);
1441 if (fack_count > tp->fackets_out)
1442 tp->fackets_out = fack_count;
1444 if (after(TCP_SKB_CB(skb)->seq,
1446 tp->highest_sack = TCP_SKB_CB(skb)->seq;
1448 if (dup_sack && (sacked&TCPCB_RETRANS))
1449 reord = min(fack_count, reord);
1452 /* D-SACK. We can detect redundant retransmission
1453 * in S|R and plain R frames and clear it.
1454 * undo_retrans is decreased above, L|R frames
1455 * are accounted above as well.
1458 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1459 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1460 tp->retrans_out -= tcp_skb_pcount(skb);
1461 tp->retransmit_skb_hint = NULL;
1466 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery)
1467 flag |= tcp_mark_lost_retrans(sk, lost_retrans);
1469 tcp_verify_left_out(tp);
1471 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1472 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1473 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1475 #if FASTRETRANS_DEBUG > 0
1476 BUG_TRAP((int)tp->sacked_out >= 0);
1477 BUG_TRAP((int)tp->lost_out >= 0);
1478 BUG_TRAP((int)tp->retrans_out >= 0);
1479 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1484 /* If we receive more dupacks than we expected counting segments
1485 * in assumption of absent reordering, interpret this as reordering.
1486 * The only another reason could be bug in receiver TCP.
1488 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1490 struct tcp_sock *tp = tcp_sk(sk);
1493 holes = max(tp->lost_out, 1U);
1494 holes = min(holes, tp->packets_out);
1496 if ((tp->sacked_out + holes) > tp->packets_out) {
1497 tp->sacked_out = tp->packets_out - holes;
1498 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1502 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1504 static void tcp_add_reno_sack(struct sock *sk)
1506 struct tcp_sock *tp = tcp_sk(sk);
1508 tcp_check_reno_reordering(sk, 0);
1509 tcp_verify_left_out(tp);
1512 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1514 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1516 struct tcp_sock *tp = tcp_sk(sk);
1519 /* One ACK acked hole. The rest eat duplicate ACKs. */
1520 if (acked-1 >= tp->sacked_out)
1523 tp->sacked_out -= acked-1;
1525 tcp_check_reno_reordering(sk, acked);
1526 tcp_verify_left_out(tp);
1529 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1534 /* F-RTO can only be used if TCP has never retransmitted anything other than
1535 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1537 int tcp_use_frto(struct sock *sk)
1539 const struct tcp_sock *tp = tcp_sk(sk);
1540 struct sk_buff *skb;
1542 if (!sysctl_tcp_frto)
1548 /* Avoid expensive walking of rexmit queue if possible */
1549 if (tp->retrans_out > 1)
1552 skb = tcp_write_queue_head(sk);
1553 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1554 tcp_for_write_queue_from(skb, sk) {
1555 if (skb == tcp_send_head(sk))
1557 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1559 /* Short-circuit when first non-SACKed skb has been checked */
1560 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1566 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1567 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1568 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1569 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1570 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1571 * bits are handled if the Loss state is really to be entered (in
1572 * tcp_enter_frto_loss).
1574 * Do like tcp_enter_loss() would; when RTO expires the second time it
1576 * "Reduce ssthresh if it has not yet been made inside this window."
1578 void tcp_enter_frto(struct sock *sk)
1580 const struct inet_connection_sock *icsk = inet_csk(sk);
1581 struct tcp_sock *tp = tcp_sk(sk);
1582 struct sk_buff *skb;
1584 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1585 tp->snd_una == tp->high_seq ||
1586 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1587 !icsk->icsk_retransmits)) {
1588 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1589 /* Our state is too optimistic in ssthresh() call because cwnd
1590 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1591 * recovery has not yet completed. Pattern would be this: RTO,
1592 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1594 * RFC4138 should be more specific on what to do, even though
1595 * RTO is quite unlikely to occur after the first Cumulative ACK
1596 * due to back-off and complexity of triggering events ...
1598 if (tp->frto_counter) {
1600 stored_cwnd = tp->snd_cwnd;
1602 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1603 tp->snd_cwnd = stored_cwnd;
1605 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1607 /* ... in theory, cong.control module could do "any tricks" in
1608 * ssthresh(), which means that ca_state, lost bits and lost_out
1609 * counter would have to be faked before the call occurs. We
1610 * consider that too expensive, unlikely and hacky, so modules
1611 * using these in ssthresh() must deal these incompatibility
1612 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1614 tcp_ca_event(sk, CA_EVENT_FRTO);
1617 tp->undo_marker = tp->snd_una;
1618 tp->undo_retrans = 0;
1620 skb = tcp_write_queue_head(sk);
1621 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1622 tp->undo_marker = 0;
1623 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1624 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1625 tp->retrans_out -= tcp_skb_pcount(skb);
1627 tcp_verify_left_out(tp);
1629 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1630 * The last condition is necessary at least in tp->frto_counter case.
1632 if (IsSackFrto() && (tp->frto_counter ||
1633 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1634 after(tp->high_seq, tp->snd_una)) {
1635 tp->frto_highmark = tp->high_seq;
1637 tp->frto_highmark = tp->snd_nxt;
1639 tcp_set_ca_state(sk, TCP_CA_Disorder);
1640 tp->high_seq = tp->snd_nxt;
1641 tp->frto_counter = 1;
1644 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1645 * which indicates that we should follow the traditional RTO recovery,
1646 * i.e. mark everything lost and do go-back-N retransmission.
1648 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1650 struct tcp_sock *tp = tcp_sk(sk);
1651 struct sk_buff *skb;
1654 tp->retrans_out = 0;
1655 if (tcp_is_reno(tp))
1656 tcp_reset_reno_sack(tp);
1658 tcp_for_write_queue(skb, sk) {
1659 if (skb == tcp_send_head(sk))
1662 * Count the retransmission made on RTO correctly (only when
1663 * waiting for the first ACK and did not get it)...
1665 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1666 /* For some reason this R-bit might get cleared? */
1667 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1668 tp->retrans_out += tcp_skb_pcount(skb);
1669 /* ...enter this if branch just for the first segment */
1670 flag |= FLAG_DATA_ACKED;
1672 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1673 tp->undo_marker = 0;
1674 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1677 /* Don't lost mark skbs that were fwd transmitted after RTO */
1678 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1679 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1680 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1681 tp->lost_out += tcp_skb_pcount(skb);
1684 tcp_verify_left_out(tp);
1686 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1687 tp->snd_cwnd_cnt = 0;
1688 tp->snd_cwnd_stamp = tcp_time_stamp;
1689 tp->frto_counter = 0;
1690 tp->bytes_acked = 0;
1692 tp->reordering = min_t(unsigned int, tp->reordering,
1693 sysctl_tcp_reordering);
1694 tcp_set_ca_state(sk, TCP_CA_Loss);
1695 tp->high_seq = tp->frto_highmark;
1696 TCP_ECN_queue_cwr(tp);
1698 tcp_clear_retrans_hints_partial(tp);
1701 void tcp_clear_retrans(struct tcp_sock *tp)
1703 tp->retrans_out = 0;
1705 tp->fackets_out = 0;
1709 tp->undo_marker = 0;
1710 tp->undo_retrans = 0;
1713 /* Enter Loss state. If "how" is not zero, forget all SACK information
1714 * and reset tags completely, otherwise preserve SACKs. If receiver
1715 * dropped its ofo queue, we will know this due to reneging detection.
1717 void tcp_enter_loss(struct sock *sk, int how)
1719 const struct inet_connection_sock *icsk = inet_csk(sk);
1720 struct tcp_sock *tp = tcp_sk(sk);
1721 struct sk_buff *skb;
1724 /* Reduce ssthresh if it has not yet been made inside this window. */
1725 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1726 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1727 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1728 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1729 tcp_ca_event(sk, CA_EVENT_LOSS);
1732 tp->snd_cwnd_cnt = 0;
1733 tp->snd_cwnd_stamp = tcp_time_stamp;
1735 tp->bytes_acked = 0;
1736 tcp_clear_retrans(tp);
1739 /* Push undo marker, if it was plain RTO and nothing
1740 * was retransmitted. */
1741 tp->undo_marker = tp->snd_una;
1742 tcp_clear_retrans_hints_partial(tp);
1744 tcp_clear_all_retrans_hints(tp);
1747 tcp_for_write_queue(skb, sk) {
1748 if (skb == tcp_send_head(sk))
1750 cnt += tcp_skb_pcount(skb);
1751 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1752 tp->undo_marker = 0;
1753 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1754 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1755 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1756 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1757 tp->lost_out += tcp_skb_pcount(skb);
1759 tp->sacked_out += tcp_skb_pcount(skb);
1760 tp->fackets_out = cnt;
1763 tcp_verify_left_out(tp);
1765 tp->reordering = min_t(unsigned int, tp->reordering,
1766 sysctl_tcp_reordering);
1767 tcp_set_ca_state(sk, TCP_CA_Loss);
1768 tp->high_seq = tp->snd_nxt;
1769 TCP_ECN_queue_cwr(tp);
1770 /* Abort FRTO algorithm if one is in progress */
1771 tp->frto_counter = 0;
1774 static int tcp_check_sack_reneging(struct sock *sk)
1776 struct sk_buff *skb;
1778 /* If ACK arrived pointing to a remembered SACK,
1779 * it means that our remembered SACKs do not reflect
1780 * real state of receiver i.e.
1781 * receiver _host_ is heavily congested (or buggy).
1782 * Do processing similar to RTO timeout.
1784 if ((skb = tcp_write_queue_head(sk)) != NULL &&
1785 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1786 struct inet_connection_sock *icsk = inet_csk(sk);
1787 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1789 tcp_enter_loss(sk, 1);
1790 icsk->icsk_retransmits++;
1791 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1792 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1793 icsk->icsk_rto, TCP_RTO_MAX);
1799 static inline int tcp_fackets_out(struct tcp_sock *tp)
1801 return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1804 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1806 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1809 static inline int tcp_head_timedout(struct sock *sk)
1811 struct tcp_sock *tp = tcp_sk(sk);
1813 return tp->packets_out &&
1814 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1817 /* Linux NewReno/SACK/FACK/ECN state machine.
1818 * --------------------------------------
1820 * "Open" Normal state, no dubious events, fast path.
1821 * "Disorder" In all the respects it is "Open",
1822 * but requires a bit more attention. It is entered when
1823 * we see some SACKs or dupacks. It is split of "Open"
1824 * mainly to move some processing from fast path to slow one.
1825 * "CWR" CWND was reduced due to some Congestion Notification event.
1826 * It can be ECN, ICMP source quench, local device congestion.
1827 * "Recovery" CWND was reduced, we are fast-retransmitting.
1828 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1830 * tcp_fastretrans_alert() is entered:
1831 * - each incoming ACK, if state is not "Open"
1832 * - when arrived ACK is unusual, namely:
1837 * Counting packets in flight is pretty simple.
1839 * in_flight = packets_out - left_out + retrans_out
1841 * packets_out is SND.NXT-SND.UNA counted in packets.
1843 * retrans_out is number of retransmitted segments.
1845 * left_out is number of segments left network, but not ACKed yet.
1847 * left_out = sacked_out + lost_out
1849 * sacked_out: Packets, which arrived to receiver out of order
1850 * and hence not ACKed. With SACKs this number is simply
1851 * amount of SACKed data. Even without SACKs
1852 * it is easy to give pretty reliable estimate of this number,
1853 * counting duplicate ACKs.
1855 * lost_out: Packets lost by network. TCP has no explicit
1856 * "loss notification" feedback from network (for now).
1857 * It means that this number can be only _guessed_.
1858 * Actually, it is the heuristics to predict lossage that
1859 * distinguishes different algorithms.
1861 * F.e. after RTO, when all the queue is considered as lost,
1862 * lost_out = packets_out and in_flight = retrans_out.
1864 * Essentially, we have now two algorithms counting
1867 * FACK: It is the simplest heuristics. As soon as we decided
1868 * that something is lost, we decide that _all_ not SACKed
1869 * packets until the most forward SACK are lost. I.e.
1870 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1871 * It is absolutely correct estimate, if network does not reorder
1872 * packets. And it loses any connection to reality when reordering
1873 * takes place. We use FACK by default until reordering
1874 * is suspected on the path to this destination.
1876 * NewReno: when Recovery is entered, we assume that one segment
1877 * is lost (classic Reno). While we are in Recovery and
1878 * a partial ACK arrives, we assume that one more packet
1879 * is lost (NewReno). This heuristics are the same in NewReno
1882 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1883 * deflation etc. CWND is real congestion window, never inflated, changes
1884 * only according to classic VJ rules.
1886 * Really tricky (and requiring careful tuning) part of algorithm
1887 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1888 * The first determines the moment _when_ we should reduce CWND and,
1889 * hence, slow down forward transmission. In fact, it determines the moment
1890 * when we decide that hole is caused by loss, rather than by a reorder.
1892 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1893 * holes, caused by lost packets.
1895 * And the most logically complicated part of algorithm is undo
1896 * heuristics. We detect false retransmits due to both too early
1897 * fast retransmit (reordering) and underestimated RTO, analyzing
1898 * timestamps and D-SACKs. When we detect that some segments were
1899 * retransmitted by mistake and CWND reduction was wrong, we undo
1900 * window reduction and abort recovery phase. This logic is hidden
1901 * inside several functions named tcp_try_undo_<something>.
1904 /* This function decides, when we should leave Disordered state
1905 * and enter Recovery phase, reducing congestion window.
1907 * Main question: may we further continue forward transmission
1908 * with the same cwnd?
1910 static int tcp_time_to_recover(struct sock *sk)
1912 struct tcp_sock *tp = tcp_sk(sk);
1915 /* Do not perform any recovery during FRTO algorithm */
1916 if (tp->frto_counter)
1919 /* Trick#1: The loss is proven. */
1923 /* Not-A-Trick#2 : Classic rule... */
1924 if (tcp_fackets_out(tp) > tp->reordering)
1927 /* Trick#3 : when we use RFC2988 timer restart, fast
1928 * retransmit can be triggered by timeout of queue head.
1930 if (tcp_head_timedout(sk))
1933 /* Trick#4: It is still not OK... But will it be useful to delay
1936 packets_out = tp->packets_out;
1937 if (packets_out <= tp->reordering &&
1938 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1939 !tcp_may_send_now(sk)) {
1940 /* We have nothing to send. This connection is limited
1941 * either by receiver window or by application.
1949 /* RFC: This is from the original, I doubt that this is necessary at all:
1950 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1951 * retransmitted past LOST markings in the first place? I'm not fully sure
1952 * about undo and end of connection cases, which can cause R without L?
1954 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
1955 struct sk_buff *skb)
1957 if ((tp->retransmit_skb_hint != NULL) &&
1958 before(TCP_SKB_CB(skb)->seq,
1959 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1960 tp->retransmit_skb_hint = NULL;
1963 /* Mark head of queue up as lost. */
1964 static void tcp_mark_head_lost(struct sock *sk,
1965 int packets, u32 high_seq)
1967 struct tcp_sock *tp = tcp_sk(sk);
1968 struct sk_buff *skb;
1971 BUG_TRAP(packets <= tp->packets_out);
1972 if (tp->lost_skb_hint) {
1973 skb = tp->lost_skb_hint;
1974 cnt = tp->lost_cnt_hint;
1976 skb = tcp_write_queue_head(sk);
1980 tcp_for_write_queue_from(skb, sk) {
1981 if (skb == tcp_send_head(sk))
1983 /* TODO: do this better */
1984 /* this is not the most efficient way to do this... */
1985 tp->lost_skb_hint = skb;
1986 tp->lost_cnt_hint = cnt;
1987 cnt += tcp_skb_pcount(skb);
1988 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1990 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1991 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1992 tp->lost_out += tcp_skb_pcount(skb);
1993 tcp_verify_retransmit_hint(tp, skb);
1996 tcp_verify_left_out(tp);
1999 /* Account newly detected lost packet(s) */
2001 static void tcp_update_scoreboard(struct sock *sk)
2003 struct tcp_sock *tp = tcp_sk(sk);
2005 if (tcp_is_fack(tp)) {
2006 int lost = tp->fackets_out - tp->reordering;
2009 tcp_mark_head_lost(sk, lost, tp->high_seq);
2011 tcp_mark_head_lost(sk, 1, tp->high_seq);
2014 /* New heuristics: it is possible only after we switched
2015 * to restart timer each time when something is ACKed.
2016 * Hence, we can detect timed out packets during fast
2017 * retransmit without falling to slow start.
2019 if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) {
2020 struct sk_buff *skb;
2022 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2023 : tcp_write_queue_head(sk);
2025 tcp_for_write_queue_from(skb, sk) {
2026 if (skb == tcp_send_head(sk))
2028 if (!tcp_skb_timedout(sk, skb))
2031 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
2032 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2033 tp->lost_out += tcp_skb_pcount(skb);
2034 tcp_verify_retransmit_hint(tp, skb);
2038 tp->scoreboard_skb_hint = skb;
2040 tcp_verify_left_out(tp);
2044 /* CWND moderation, preventing bursts due to too big ACKs
2045 * in dubious situations.
2047 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2049 tp->snd_cwnd = min(tp->snd_cwnd,
2050 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
2051 tp->snd_cwnd_stamp = tcp_time_stamp;
2054 /* Lower bound on congestion window is slow start threshold
2055 * unless congestion avoidance choice decides to overide it.
2057 static inline u32 tcp_cwnd_min(const struct sock *sk)
2059 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2061 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2064 /* Decrease cwnd each second ack. */
2065 static void tcp_cwnd_down(struct sock *sk, int flag)
2067 struct tcp_sock *tp = tcp_sk(sk);
2068 int decr = tp->snd_cwnd_cnt + 1;
2070 if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
2071 (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
2072 tp->snd_cwnd_cnt = decr&1;
2075 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2076 tp->snd_cwnd -= decr;
2078 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
2079 tp->snd_cwnd_stamp = tcp_time_stamp;
2083 /* Nothing was retransmitted or returned timestamp is less
2084 * than timestamp of the first retransmission.
2086 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2088 return !tp->retrans_stamp ||
2089 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2090 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2093 /* Undo procedures. */
2095 #if FASTRETRANS_DEBUG > 1
2096 static void DBGUNDO(struct sock *sk, const char *msg)
2098 struct tcp_sock *tp = tcp_sk(sk);
2099 struct inet_sock *inet = inet_sk(sk);
2101 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2103 NIPQUAD(inet->daddr), ntohs(inet->dport),
2104 tp->snd_cwnd, tcp_left_out(tp),
2105 tp->snd_ssthresh, tp->prior_ssthresh,
2109 #define DBGUNDO(x...) do { } while (0)
2112 static void tcp_undo_cwr(struct sock *sk, const int undo)
2114 struct tcp_sock *tp = tcp_sk(sk);
2116 if (tp->prior_ssthresh) {
2117 const struct inet_connection_sock *icsk = inet_csk(sk);
2119 if (icsk->icsk_ca_ops->undo_cwnd)
2120 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2122 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
2124 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2125 tp->snd_ssthresh = tp->prior_ssthresh;
2126 TCP_ECN_withdraw_cwr(tp);
2129 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2131 tcp_moderate_cwnd(tp);
2132 tp->snd_cwnd_stamp = tcp_time_stamp;
2134 /* There is something screwy going on with the retrans hints after
2136 tcp_clear_all_retrans_hints(tp);
2139 static inline int tcp_may_undo(struct tcp_sock *tp)
2141 return tp->undo_marker &&
2142 (!tp->undo_retrans || tcp_packet_delayed(tp));
2145 /* People celebrate: "We love our President!" */
2146 static int tcp_try_undo_recovery(struct sock *sk)
2148 struct tcp_sock *tp = tcp_sk(sk);
2150 if (tcp_may_undo(tp)) {
2151 /* Happy end! We did not retransmit anything
2152 * or our original transmission succeeded.
2154 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2155 tcp_undo_cwr(sk, 1);
2156 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2157 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2159 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2160 tp->undo_marker = 0;
2162 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2163 /* Hold old state until something *above* high_seq
2164 * is ACKed. For Reno it is MUST to prevent false
2165 * fast retransmits (RFC2582). SACK TCP is safe. */
2166 tcp_moderate_cwnd(tp);
2169 tcp_set_ca_state(sk, TCP_CA_Open);
2173 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2174 static void tcp_try_undo_dsack(struct sock *sk)
2176 struct tcp_sock *tp = tcp_sk(sk);
2178 if (tp->undo_marker && !tp->undo_retrans) {
2179 DBGUNDO(sk, "D-SACK");
2180 tcp_undo_cwr(sk, 1);
2181 tp->undo_marker = 0;
2182 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2186 /* Undo during fast recovery after partial ACK. */
2188 static int tcp_try_undo_partial(struct sock *sk, int acked)
2190 struct tcp_sock *tp = tcp_sk(sk);
2191 /* Partial ACK arrived. Force Hoe's retransmit. */
2192 int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering;
2194 if (tcp_may_undo(tp)) {
2195 /* Plain luck! Hole if filled with delayed
2196 * packet, rather than with a retransmit.
2198 if (tp->retrans_out == 0)
2199 tp->retrans_stamp = 0;
2201 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2204 tcp_undo_cwr(sk, 0);
2205 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2207 /* So... Do not make Hoe's retransmit yet.
2208 * If the first packet was delayed, the rest
2209 * ones are most probably delayed as well.
2216 /* Undo during loss recovery after partial ACK. */
2217 static int tcp_try_undo_loss(struct sock *sk)
2219 struct tcp_sock *tp = tcp_sk(sk);
2221 if (tcp_may_undo(tp)) {
2222 struct sk_buff *skb;
2223 tcp_for_write_queue(skb, sk) {
2224 if (skb == tcp_send_head(sk))
2226 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2229 tcp_clear_all_retrans_hints(tp);
2231 DBGUNDO(sk, "partial loss");
2233 tcp_undo_cwr(sk, 1);
2234 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2235 inet_csk(sk)->icsk_retransmits = 0;
2236 tp->undo_marker = 0;
2237 if (tcp_is_sack(tp))
2238 tcp_set_ca_state(sk, TCP_CA_Open);
2244 static inline void tcp_complete_cwr(struct sock *sk)
2246 struct tcp_sock *tp = tcp_sk(sk);
2247 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2248 tp->snd_cwnd_stamp = tcp_time_stamp;
2249 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2252 static void tcp_try_to_open(struct sock *sk, int flag)
2254 struct tcp_sock *tp = tcp_sk(sk);
2256 tcp_verify_left_out(tp);
2258 if (tp->retrans_out == 0)
2259 tp->retrans_stamp = 0;
2262 tcp_enter_cwr(sk, 1);
2264 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2265 int state = TCP_CA_Open;
2267 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2268 state = TCP_CA_Disorder;
2270 if (inet_csk(sk)->icsk_ca_state != state) {
2271 tcp_set_ca_state(sk, state);
2272 tp->high_seq = tp->snd_nxt;
2274 tcp_moderate_cwnd(tp);
2276 tcp_cwnd_down(sk, flag);
2280 static void tcp_mtup_probe_failed(struct sock *sk)
2282 struct inet_connection_sock *icsk = inet_csk(sk);
2284 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2285 icsk->icsk_mtup.probe_size = 0;
2288 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2290 struct tcp_sock *tp = tcp_sk(sk);
2291 struct inet_connection_sock *icsk = inet_csk(sk);
2293 /* FIXME: breaks with very large cwnd */
2294 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2295 tp->snd_cwnd = tp->snd_cwnd *
2296 tcp_mss_to_mtu(sk, tp->mss_cache) /
2297 icsk->icsk_mtup.probe_size;
2298 tp->snd_cwnd_cnt = 0;
2299 tp->snd_cwnd_stamp = tcp_time_stamp;
2300 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2302 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2303 icsk->icsk_mtup.probe_size = 0;
2304 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2308 /* Process an event, which can update packets-in-flight not trivially.
2309 * Main goal of this function is to calculate new estimate for left_out,
2310 * taking into account both packets sitting in receiver's buffer and
2311 * packets lost by network.
2313 * Besides that it does CWND reduction, when packet loss is detected
2314 * and changes state of machine.
2316 * It does _not_ decide what to send, it is made in function
2317 * tcp_xmit_retransmit_queue().
2320 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2322 struct inet_connection_sock *icsk = inet_csk(sk);
2323 struct tcp_sock *tp = tcp_sk(sk);
2324 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2325 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2326 (tp->fackets_out > tp->reordering));
2328 /* Some technical things:
2329 * 1. Reno does not count dupacks (sacked_out) automatically. */
2330 if (!tp->packets_out)
2333 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2334 tp->fackets_out = 0;
2336 /* Now state machine starts.
2337 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2339 tp->prior_ssthresh = 0;
2341 /* B. In all the states check for reneging SACKs. */
2342 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2345 /* C. Process data loss notification, provided it is valid. */
2346 if ((flag&FLAG_DATA_LOST) &&
2347 before(tp->snd_una, tp->high_seq) &&
2348 icsk->icsk_ca_state != TCP_CA_Open &&
2349 tp->fackets_out > tp->reordering) {
2350 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2351 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2354 /* D. Check consistency of the current state. */
2355 tcp_verify_left_out(tp);
2357 /* E. Check state exit conditions. State can be terminated
2358 * when high_seq is ACKed. */
2359 if (icsk->icsk_ca_state == TCP_CA_Open) {
2360 BUG_TRAP(tp->retrans_out == 0);
2361 tp->retrans_stamp = 0;
2362 } else if (!before(tp->snd_una, tp->high_seq)) {
2363 switch (icsk->icsk_ca_state) {
2365 icsk->icsk_retransmits = 0;
2366 if (tcp_try_undo_recovery(sk))
2371 /* CWR is to be held something *above* high_seq
2372 * is ACKed for CWR bit to reach receiver. */
2373 if (tp->snd_una != tp->high_seq) {
2374 tcp_complete_cwr(sk);
2375 tcp_set_ca_state(sk, TCP_CA_Open);
2379 case TCP_CA_Disorder:
2380 tcp_try_undo_dsack(sk);
2381 if (!tp->undo_marker ||
2382 /* For SACK case do not Open to allow to undo
2383 * catching for all duplicate ACKs. */
2384 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2385 tp->undo_marker = 0;
2386 tcp_set_ca_state(sk, TCP_CA_Open);
2390 case TCP_CA_Recovery:
2391 if (tcp_is_reno(tp))
2392 tcp_reset_reno_sack(tp);
2393 if (tcp_try_undo_recovery(sk))
2395 tcp_complete_cwr(sk);
2400 /* F. Process state. */
2401 switch (icsk->icsk_ca_state) {
2402 case TCP_CA_Recovery:
2403 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2404 if (tcp_is_reno(tp) && is_dupack)
2405 tcp_add_reno_sack(sk);
2407 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2410 if (flag&FLAG_DATA_ACKED)
2411 icsk->icsk_retransmits = 0;
2412 if (!tcp_try_undo_loss(sk)) {
2413 tcp_moderate_cwnd(tp);
2414 tcp_xmit_retransmit_queue(sk);
2417 if (icsk->icsk_ca_state != TCP_CA_Open)
2419 /* Loss is undone; fall through to processing in Open state. */
2421 if (tcp_is_reno(tp)) {
2422 if (flag & FLAG_SND_UNA_ADVANCED)
2423 tcp_reset_reno_sack(tp);
2425 tcp_add_reno_sack(sk);
2428 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2429 tcp_try_undo_dsack(sk);
2431 if (!tcp_time_to_recover(sk)) {
2432 tcp_try_to_open(sk, flag);
2436 /* MTU probe failure: don't reduce cwnd */
2437 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2438 icsk->icsk_mtup.probe_size &&
2439 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2440 tcp_mtup_probe_failed(sk);
2441 /* Restores the reduction we did in tcp_mtup_probe() */
2443 tcp_simple_retransmit(sk);
2447 /* Otherwise enter Recovery state */
2449 if (tcp_is_reno(tp))
2450 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2452 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2454 tp->high_seq = tp->snd_nxt;
2455 tp->prior_ssthresh = 0;
2456 tp->undo_marker = tp->snd_una;
2457 tp->undo_retrans = tp->retrans_out;
2459 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2460 if (!(flag&FLAG_ECE))
2461 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2462 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2463 TCP_ECN_queue_cwr(tp);
2466 tp->bytes_acked = 0;
2467 tp->snd_cwnd_cnt = 0;
2468 tcp_set_ca_state(sk, TCP_CA_Recovery);
2471 if (do_lost || tcp_head_timedout(sk))
2472 tcp_update_scoreboard(sk);
2473 tcp_cwnd_down(sk, flag);
2474 tcp_xmit_retransmit_queue(sk);
2477 /* Read draft-ietf-tcplw-high-performance before mucking
2478 * with this code. (Supersedes RFC1323)
2480 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2482 /* RTTM Rule: A TSecr value received in a segment is used to
2483 * update the averaged RTT measurement only if the segment
2484 * acknowledges some new data, i.e., only if it advances the
2485 * left edge of the send window.
2487 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2488 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2490 * Changed: reset backoff as soon as we see the first valid sample.
2491 * If we do not, we get strongly overestimated rto. With timestamps
2492 * samples are accepted even from very old segments: f.e., when rtt=1
2493 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2494 * answer arrives rto becomes 120 seconds! If at least one of segments
2495 * in window is lost... Voila. --ANK (010210)
2497 struct tcp_sock *tp = tcp_sk(sk);
2498 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2499 tcp_rtt_estimator(sk, seq_rtt);
2501 inet_csk(sk)->icsk_backoff = 0;
2505 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2507 /* We don't have a timestamp. Can only use
2508 * packets that are not retransmitted to determine
2509 * rtt estimates. Also, we must not reset the
2510 * backoff for rto until we get a non-retransmitted
2511 * packet. This allows us to deal with a situation
2512 * where the network delay has increased suddenly.
2513 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2516 if (flag & FLAG_RETRANS_DATA_ACKED)
2519 tcp_rtt_estimator(sk, seq_rtt);
2521 inet_csk(sk)->icsk_backoff = 0;
2525 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2528 const struct tcp_sock *tp = tcp_sk(sk);
2529 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2530 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2531 tcp_ack_saw_tstamp(sk, flag);
2532 else if (seq_rtt >= 0)
2533 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2536 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2537 u32 in_flight, int good)
2539 const struct inet_connection_sock *icsk = inet_csk(sk);
2540 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2541 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2544 /* Restart timer after forward progress on connection.
2545 * RFC2988 recommends to restart timer to now+rto.
2547 static void tcp_rearm_rto(struct sock *sk)
2549 struct tcp_sock *tp = tcp_sk(sk);
2551 if (!tp->packets_out) {
2552 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2554 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2558 /* If we get here, the whole TSO packet has not been acked. */
2559 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2561 struct tcp_sock *tp = tcp_sk(sk);
2564 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2566 packets_acked = tcp_skb_pcount(skb);
2567 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2569 packets_acked -= tcp_skb_pcount(skb);
2571 if (packets_acked) {
2572 BUG_ON(tcp_skb_pcount(skb) == 0);
2573 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2576 return packets_acked;
2579 /* Remove acknowledged frames from the retransmission queue. If our packet
2580 * is before the ack sequence we can discard it as it's confirmed to have
2581 * arrived at the other end.
2583 static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p)
2585 struct tcp_sock *tp = tcp_sk(sk);
2586 const struct inet_connection_sock *icsk = inet_csk(sk);
2587 struct sk_buff *skb;
2588 u32 now = tcp_time_stamp;
2589 int fully_acked = 1;
2591 int prior_packets = tp->packets_out;
2593 ktime_t last_ackt = net_invalid_timestamp();
2595 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2596 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2599 u8 sacked = scb->sacked;
2601 if (after(scb->end_seq, tp->snd_una)) {
2602 if (tcp_skb_pcount(skb) == 1 ||
2603 !after(tp->snd_una, scb->seq))
2606 packets_acked = tcp_tso_acked(sk, skb);
2611 end_seq = tp->snd_una;
2613 packets_acked = tcp_skb_pcount(skb);
2614 end_seq = scb->end_seq;
2617 /* MTU probing checks */
2618 if (fully_acked && icsk->icsk_mtup.probe_size &&
2619 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2620 tcp_mtup_probe_success(sk, skb);
2624 if (sacked & TCPCB_RETRANS) {
2625 if (sacked & TCPCB_SACKED_RETRANS)
2626 tp->retrans_out -= packets_acked;
2627 flag |= FLAG_RETRANS_DATA_ACKED;
2629 if ((flag & FLAG_DATA_ACKED) ||
2630 (packets_acked > 1))
2631 flag |= FLAG_NONHEAD_RETRANS_ACKED;
2632 } else if (seq_rtt < 0) {
2633 seq_rtt = now - scb->when;
2635 last_ackt = skb->tstamp;
2638 if (sacked & TCPCB_SACKED_ACKED)
2639 tp->sacked_out -= packets_acked;
2640 if (sacked & TCPCB_LOST)
2641 tp->lost_out -= packets_acked;
2643 if ((sacked & TCPCB_URG) && tp->urg_mode &&
2644 !before(end_seq, tp->snd_up))
2646 } else if (seq_rtt < 0) {
2647 seq_rtt = now - scb->when;
2649 last_ackt = skb->tstamp;
2651 tp->packets_out -= packets_acked;
2653 /* Initial outgoing SYN's get put onto the write_queue
2654 * just like anything else we transmit. It is not
2655 * true data, and if we misinform our callers that
2656 * this ACK acks real data, we will erroneously exit
2657 * connection startup slow start one packet too
2658 * quickly. This is severely frowned upon behavior.
2660 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2661 flag |= FLAG_DATA_ACKED;
2663 flag |= FLAG_SYN_ACKED;
2664 tp->retrans_stamp = 0;
2670 tcp_unlink_write_queue(skb, sk);
2671 sk_stream_free_skb(sk, skb);
2672 tcp_clear_all_retrans_hints(tp);
2675 if (flag & FLAG_ACKED) {
2676 u32 pkts_acked = prior_packets - tp->packets_out;
2677 const struct tcp_congestion_ops *ca_ops
2678 = inet_csk(sk)->icsk_ca_ops;
2680 tcp_ack_update_rtt(sk, flag, seq_rtt);
2683 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2684 /* hint's skb might be NULL but we don't need to care */
2685 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
2686 tp->fastpath_cnt_hint);
2687 if (tcp_is_reno(tp))
2688 tcp_remove_reno_sacks(sk, pkts_acked);
2690 if (ca_ops->pkts_acked) {
2693 /* Is the ACK triggering packet unambiguous? */
2694 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2695 /* High resolution needed and available? */
2696 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2697 !ktime_equal(last_ackt,
2698 net_invalid_timestamp()))
2699 rtt_us = ktime_us_delta(ktime_get_real(),
2701 else if (seq_rtt > 0)
2702 rtt_us = jiffies_to_usecs(seq_rtt);
2705 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2709 #if FASTRETRANS_DEBUG > 0
2710 BUG_TRAP((int)tp->sacked_out >= 0);
2711 BUG_TRAP((int)tp->lost_out >= 0);
2712 BUG_TRAP((int)tp->retrans_out >= 0);
2713 if (!tp->packets_out && tcp_is_sack(tp)) {
2714 icsk = inet_csk(sk);
2716 printk(KERN_DEBUG "Leak l=%u %d\n",
2717 tp->lost_out, icsk->icsk_ca_state);
2720 if (tp->sacked_out) {
2721 printk(KERN_DEBUG "Leak s=%u %d\n",
2722 tp->sacked_out, icsk->icsk_ca_state);
2725 if (tp->retrans_out) {
2726 printk(KERN_DEBUG "Leak r=%u %d\n",
2727 tp->retrans_out, icsk->icsk_ca_state);
2728 tp->retrans_out = 0;
2732 *seq_rtt_p = seq_rtt;
2736 static void tcp_ack_probe(struct sock *sk)
2738 const struct tcp_sock *tp = tcp_sk(sk);
2739 struct inet_connection_sock *icsk = inet_csk(sk);
2741 /* Was it a usable window open? */
2743 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2744 tp->snd_una + tp->snd_wnd)) {
2745 icsk->icsk_backoff = 0;
2746 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2747 /* Socket must be waked up by subsequent tcp_data_snd_check().
2748 * This function is not for random using!
2751 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2752 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2757 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2759 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2760 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2763 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2765 const struct tcp_sock *tp = tcp_sk(sk);
2766 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2767 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2770 /* Check that window update is acceptable.
2771 * The function assumes that snd_una<=ack<=snd_next.
2773 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2774 const u32 ack_seq, const u32 nwin)
2776 return (after(ack, tp->snd_una) ||
2777 after(ack_seq, tp->snd_wl1) ||
2778 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2781 /* Update our send window.
2783 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2784 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2786 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2789 struct tcp_sock *tp = tcp_sk(sk);
2791 u32 nwin = ntohs(tcp_hdr(skb)->window);
2793 if (likely(!tcp_hdr(skb)->syn))
2794 nwin <<= tp->rx_opt.snd_wscale;
2796 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2797 flag |= FLAG_WIN_UPDATE;
2798 tcp_update_wl(tp, ack, ack_seq);
2800 if (tp->snd_wnd != nwin) {
2803 /* Note, it is the only place, where
2804 * fast path is recovered for sending TCP.
2807 tcp_fast_path_check(sk);
2809 if (nwin > tp->max_window) {
2810 tp->max_window = nwin;
2811 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2821 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2822 * continue in congestion avoidance.
2824 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2826 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2827 tp->snd_cwnd_cnt = 0;
2828 tp->bytes_acked = 0;
2829 TCP_ECN_queue_cwr(tp);
2830 tcp_moderate_cwnd(tp);
2833 /* A conservative spurious RTO response algorithm: reduce cwnd using
2834 * rate halving and continue in congestion avoidance.
2836 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2838 tcp_enter_cwr(sk, 0);
2841 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2844 tcp_ratehalving_spur_to_response(sk);
2846 tcp_undo_cwr(sk, 1);
2849 /* F-RTO spurious RTO detection algorithm (RFC4138)
2851 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2852 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2853 * window (but not to or beyond highest sequence sent before RTO):
2854 * On First ACK, send two new segments out.
2855 * On Second ACK, RTO was likely spurious. Do spurious response (response
2856 * algorithm is not part of the F-RTO detection algorithm
2857 * given in RFC4138 but can be selected separately).
2858 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2859 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2860 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2861 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2863 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2864 * original window even after we transmit two new data segments.
2867 * on first step, wait until first cumulative ACK arrives, then move to
2868 * the second step. In second step, the next ACK decides.
2870 * F-RTO is implemented (mainly) in four functions:
2871 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2872 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2873 * called when tcp_use_frto() showed green light
2874 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2875 * - tcp_enter_frto_loss() is called if there is not enough evidence
2876 * to prove that the RTO is indeed spurious. It transfers the control
2877 * from F-RTO to the conventional RTO recovery
2879 static int tcp_process_frto(struct sock *sk, int flag)
2881 struct tcp_sock *tp = tcp_sk(sk);
2883 tcp_verify_left_out(tp);
2885 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2886 if (flag&FLAG_DATA_ACKED)
2887 inet_csk(sk)->icsk_retransmits = 0;
2889 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
2890 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
2891 tp->undo_marker = 0;
2893 if (!before(tp->snd_una, tp->frto_highmark)) {
2894 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2898 if (!IsSackFrto() || tcp_is_reno(tp)) {
2899 /* RFC4138 shortcoming in step 2; should also have case c):
2900 * ACK isn't duplicate nor advances window, e.g., opposite dir
2903 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2906 if (!(flag&FLAG_DATA_ACKED)) {
2907 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2912 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2913 /* Prevent sending of new data. */
2914 tp->snd_cwnd = min(tp->snd_cwnd,
2915 tcp_packets_in_flight(tp));
2919 if ((tp->frto_counter >= 2) &&
2920 (!(flag&FLAG_FORWARD_PROGRESS) ||
2921 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2922 /* RFC4138 shortcoming (see comment above) */
2923 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2926 tcp_enter_frto_loss(sk, 3, flag);
2931 if (tp->frto_counter == 1) {
2932 /* Sending of the next skb must be allowed or no FRTO */
2933 if (!tcp_send_head(sk) ||
2934 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2935 tp->snd_una + tp->snd_wnd)) {
2936 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2941 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2942 tp->frto_counter = 2;
2945 switch (sysctl_tcp_frto_response) {
2947 tcp_undo_spur_to_response(sk, flag);
2950 tcp_conservative_spur_to_response(tp);
2953 tcp_ratehalving_spur_to_response(sk);
2956 tp->frto_counter = 0;
2957 tp->undo_marker = 0;
2958 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
2963 /* This routine deals with incoming acks, but not outgoing ones. */
2964 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2966 struct inet_connection_sock *icsk = inet_csk(sk);
2967 struct tcp_sock *tp = tcp_sk(sk);
2968 u32 prior_snd_una = tp->snd_una;
2969 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2970 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2971 u32 prior_in_flight;
2976 /* If the ack is newer than sent or older than previous acks
2977 * then we can probably ignore it.
2979 if (after(ack, tp->snd_nxt))
2980 goto uninteresting_ack;
2982 if (before(ack, prior_snd_una))
2985 if (after(ack, prior_snd_una))
2986 flag |= FLAG_SND_UNA_ADVANCED;
2988 if (sysctl_tcp_abc) {
2989 if (icsk->icsk_ca_state < TCP_CA_CWR)
2990 tp->bytes_acked += ack - prior_snd_una;
2991 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2992 /* we assume just one segment left network */
2993 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2996 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2997 /* Window is constant, pure forward advance.
2998 * No more checks are required.
2999 * Note, we use the fact that SND.UNA>=SND.WL2.
3001 tcp_update_wl(tp, ack, ack_seq);
3003 flag |= FLAG_WIN_UPDATE;
3005 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3007 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3009 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3012 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3014 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3016 if (TCP_SKB_CB(skb)->sacked)
3017 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3019 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3022 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3025 /* We passed data and got it acked, remove any soft error
3026 * log. Something worked...
3028 sk->sk_err_soft = 0;
3029 tp->rcv_tstamp = tcp_time_stamp;
3030 prior_packets = tp->packets_out;
3034 prior_in_flight = tcp_packets_in_flight(tp);
3036 /* See if we can take anything off of the retransmit queue. */
3037 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
3039 /* Guarantee sacktag reordering detection against wrap-arounds */
3040 if (before(tp->frto_highmark, tp->snd_una))
3041 tp->frto_highmark = 0;
3042 if (tp->frto_counter)
3043 frto_cwnd = tcp_process_frto(sk, flag);
3045 if (tcp_ack_is_dubious(sk, flag)) {
3046 /* Advance CWND, if state allows this. */
3047 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3048 tcp_may_raise_cwnd(sk, flag))
3049 tcp_cong_avoid(sk, ack, prior_in_flight, 0);
3050 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
3052 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3053 tcp_cong_avoid(sk, ack, prior_in_flight, 1);
3056 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
3057 dst_confirm(sk->sk_dst_cache);
3062 icsk->icsk_probes_out = 0;
3064 /* If this ack opens up a zero window, clear backoff. It was
3065 * being used to time the probes, and is probably far higher than
3066 * it needs to be for normal retransmission.
3068 if (tcp_send_head(sk))
3073 if (TCP_SKB_CB(skb)->sacked)
3074 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3077 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3082 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3083 * But, this can also be called on packets in the established flow when
3084 * the fast version below fails.
3086 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
3089 struct tcphdr *th = tcp_hdr(skb);
3090 int length=(th->doff*4)-sizeof(struct tcphdr);
3092 ptr = (unsigned char *)(th + 1);
3093 opt_rx->saw_tstamp = 0;
3095 while (length > 0) {
3102 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3107 if (opsize < 2) /* "silly options" */
3109 if (opsize > length)
3110 return; /* don't parse partial options */
3113 if (opsize==TCPOLEN_MSS && th->syn && !estab) {
3114 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3116 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3117 in_mss = opt_rx->user_mss;
3118 opt_rx->mss_clamp = in_mss;
3123 if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3124 if (sysctl_tcp_window_scaling) {
3125 __u8 snd_wscale = *(__u8 *) ptr;
3126 opt_rx->wscale_ok = 1;
3127 if (snd_wscale > 14) {
3128 if (net_ratelimit())
3129 printk(KERN_INFO "tcp_parse_options: Illegal window "
3130 "scaling value %d >14 received.\n",
3134 opt_rx->snd_wscale = snd_wscale;
3137 case TCPOPT_TIMESTAMP:
3138 if (opsize==TCPOLEN_TIMESTAMP) {
3139 if ((estab && opt_rx->tstamp_ok) ||
3140 (!estab && sysctl_tcp_timestamps)) {
3141 opt_rx->saw_tstamp = 1;
3142 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3143 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3147 case TCPOPT_SACK_PERM:
3148 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3149 if (sysctl_tcp_sack) {
3150 opt_rx->sack_ok = 1;
3151 tcp_sack_reset(opt_rx);
3157 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3158 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3160 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3163 #ifdef CONFIG_TCP_MD5SIG
3166 * The MD5 Hash has already been
3167 * checked (see tcp_v{4,6}_do_rcv()).
3179 /* Fast parse options. This hopes to only see timestamps.
3180 * If it is wrong it falls back on tcp_parse_options().
3182 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3183 struct tcp_sock *tp)
3185 if (th->doff == sizeof(struct tcphdr)>>2) {
3186 tp->rx_opt.saw_tstamp = 0;
3188 } else if (tp->rx_opt.tstamp_ok &&
3189 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3190 __be32 *ptr = (__be32 *)(th + 1);
3191 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3192 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3193 tp->rx_opt.saw_tstamp = 1;
3195 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3197 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3201 tcp_parse_options(skb, &tp->rx_opt, 1);
3205 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3207 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3208 tp->rx_opt.ts_recent_stamp = get_seconds();
3211 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3213 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3214 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3215 * extra check below makes sure this can only happen
3216 * for pure ACK frames. -DaveM
3218 * Not only, also it occurs for expired timestamps.
3221 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3222 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3223 tcp_store_ts_recent(tp);
3227 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3229 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3230 * it can pass through stack. So, the following predicate verifies that
3231 * this segment is not used for anything but congestion avoidance or
3232 * fast retransmit. Moreover, we even are able to eliminate most of such
3233 * second order effects, if we apply some small "replay" window (~RTO)
3234 * to timestamp space.
3236 * All these measures still do not guarantee that we reject wrapped ACKs
3237 * on networks with high bandwidth, when sequence space is recycled fastly,
3238 * but it guarantees that such events will be very rare and do not affect
3239 * connection seriously. This doesn't look nice, but alas, PAWS is really
3242 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3243 * states that events when retransmit arrives after original data are rare.
3244 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3245 * the biggest problem on large power networks even with minor reordering.
3246 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3247 * up to bandwidth of 18Gigabit/sec. 8) ]
3250 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3252 struct tcp_sock *tp = tcp_sk(sk);
3253 struct tcphdr *th = tcp_hdr(skb);
3254 u32 seq = TCP_SKB_CB(skb)->seq;
3255 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3257 return (/* 1. Pure ACK with correct sequence number. */
3258 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3260 /* 2. ... and duplicate ACK. */
3261 ack == tp->snd_una &&
3263 /* 3. ... and does not update window. */
3264 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3266 /* 4. ... and sits in replay window. */
3267 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3270 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3272 const struct tcp_sock *tp = tcp_sk(sk);
3273 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3274 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3275 !tcp_disordered_ack(sk, skb));
3278 /* Check segment sequence number for validity.
3280 * Segment controls are considered valid, if the segment
3281 * fits to the window after truncation to the window. Acceptability
3282 * of data (and SYN, FIN, of course) is checked separately.
3283 * See tcp_data_queue(), for example.
3285 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3286 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3287 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3288 * (borrowed from freebsd)
3291 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3293 return !before(end_seq, tp->rcv_wup) &&
3294 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3297 /* When we get a reset we do this. */
3298 static void tcp_reset(struct sock *sk)
3300 /* We want the right error as BSD sees it (and indeed as we do). */
3301 switch (sk->sk_state) {
3303 sk->sk_err = ECONNREFUSED;
3305 case TCP_CLOSE_WAIT:
3311 sk->sk_err = ECONNRESET;
3314 if (!sock_flag(sk, SOCK_DEAD))
3315 sk->sk_error_report(sk);
3321 * Process the FIN bit. This now behaves as it is supposed to work
3322 * and the FIN takes effect when it is validly part of sequence
3323 * space. Not before when we get holes.
3325 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3326 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3329 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3330 * close and we go into CLOSING (and later onto TIME-WAIT)
3332 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3334 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3336 struct tcp_sock *tp = tcp_sk(sk);
3338 inet_csk_schedule_ack(sk);
3340 sk->sk_shutdown |= RCV_SHUTDOWN;
3341 sock_set_flag(sk, SOCK_DONE);
3343 switch (sk->sk_state) {
3345 case TCP_ESTABLISHED:
3346 /* Move to CLOSE_WAIT */
3347 tcp_set_state(sk, TCP_CLOSE_WAIT);
3348 inet_csk(sk)->icsk_ack.pingpong = 1;
3351 case TCP_CLOSE_WAIT:
3353 /* Received a retransmission of the FIN, do
3358 /* RFC793: Remain in the LAST-ACK state. */
3362 /* This case occurs when a simultaneous close
3363 * happens, we must ack the received FIN and
3364 * enter the CLOSING state.
3367 tcp_set_state(sk, TCP_CLOSING);
3370 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3372 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3375 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3376 * cases we should never reach this piece of code.
3378 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3379 __FUNCTION__, sk->sk_state);
3383 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3384 * Probably, we should reset in this case. For now drop them.
3386 __skb_queue_purge(&tp->out_of_order_queue);
3387 if (tcp_is_sack(tp))
3388 tcp_sack_reset(&tp->rx_opt);
3389 sk_stream_mem_reclaim(sk);
3391 if (!sock_flag(sk, SOCK_DEAD)) {
3392 sk->sk_state_change(sk);
3394 /* Do not send POLL_HUP for half duplex close. */
3395 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3396 sk->sk_state == TCP_CLOSE)
3397 sk_wake_async(sk, 1, POLL_HUP);
3399 sk_wake_async(sk, 1, POLL_IN);
3403 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3405 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3406 if (before(seq, sp->start_seq))
3407 sp->start_seq = seq;
3408 if (after(end_seq, sp->end_seq))
3409 sp->end_seq = end_seq;
3415 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3417 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3418 if (before(seq, tp->rcv_nxt))
3419 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3421 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3423 tp->rx_opt.dsack = 1;
3424 tp->duplicate_sack[0].start_seq = seq;
3425 tp->duplicate_sack[0].end_seq = end_seq;
3426 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3430 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3432 if (!tp->rx_opt.dsack)
3433 tcp_dsack_set(tp, seq, end_seq);
3435 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3438 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3440 struct tcp_sock *tp = tcp_sk(sk);
3442 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3443 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3444 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3445 tcp_enter_quickack_mode(sk);
3447 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3448 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3450 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3451 end_seq = tp->rcv_nxt;
3452 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3459 /* These routines update the SACK block as out-of-order packets arrive or
3460 * in-order packets close up the sequence space.
3462 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3465 struct tcp_sack_block *sp = &tp->selective_acks[0];
3466 struct tcp_sack_block *swalk = sp+1;
3468 /* See if the recent change to the first SACK eats into
3469 * or hits the sequence space of other SACK blocks, if so coalesce.
3471 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3472 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3475 /* Zap SWALK, by moving every further SACK up by one slot.
3476 * Decrease num_sacks.
3478 tp->rx_opt.num_sacks--;
3479 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3480 for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3484 this_sack++, swalk++;
3488 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3492 tmp = sack1->start_seq;
3493 sack1->start_seq = sack2->start_seq;
3494 sack2->start_seq = tmp;
3496 tmp = sack1->end_seq;
3497 sack1->end_seq = sack2->end_seq;
3498 sack2->end_seq = tmp;
3501 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3503 struct tcp_sock *tp = tcp_sk(sk);
3504 struct tcp_sack_block *sp = &tp->selective_acks[0];
3505 int cur_sacks = tp->rx_opt.num_sacks;
3511 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3512 if (tcp_sack_extend(sp, seq, end_seq)) {
3513 /* Rotate this_sack to the first one. */
3514 for (; this_sack>0; this_sack--, sp--)
3515 tcp_sack_swap(sp, sp-1);
3517 tcp_sack_maybe_coalesce(tp);
3522 /* Could not find an adjacent existing SACK, build a new one,
3523 * put it at the front, and shift everyone else down. We
3524 * always know there is at least one SACK present already here.
3526 * If the sack array is full, forget about the last one.
3528 if (this_sack >= 4) {
3530 tp->rx_opt.num_sacks--;
3533 for (; this_sack > 0; this_sack--, sp--)
3537 /* Build the new head SACK, and we're done. */
3538 sp->start_seq = seq;
3539 sp->end_seq = end_seq;
3540 tp->rx_opt.num_sacks++;
3541 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3544 /* RCV.NXT advances, some SACKs should be eaten. */
3546 static void tcp_sack_remove(struct tcp_sock *tp)
3548 struct tcp_sack_block *sp = &tp->selective_acks[0];
3549 int num_sacks = tp->rx_opt.num_sacks;
3552 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3553 if (skb_queue_empty(&tp->out_of_order_queue)) {
3554 tp->rx_opt.num_sacks = 0;
3555 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3559 for (this_sack = 0; this_sack < num_sacks; ) {
3560 /* Check if the start of the sack is covered by RCV.NXT. */
3561 if (!before(tp->rcv_nxt, sp->start_seq)) {
3564 /* RCV.NXT must cover all the block! */
3565 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3567 /* Zap this SACK, by moving forward any other SACKS. */
3568 for (i=this_sack+1; i < num_sacks; i++)
3569 tp->selective_acks[i-1] = tp->selective_acks[i];
3576 if (num_sacks != tp->rx_opt.num_sacks) {
3577 tp->rx_opt.num_sacks = num_sacks;
3578 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3582 /* This one checks to see if we can put data from the
3583 * out_of_order queue into the receive_queue.
3585 static void tcp_ofo_queue(struct sock *sk)
3587 struct tcp_sock *tp = tcp_sk(sk);
3588 __u32 dsack_high = tp->rcv_nxt;
3589 struct sk_buff *skb;
3591 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3592 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3595 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3596 __u32 dsack = dsack_high;
3597 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3598 dsack_high = TCP_SKB_CB(skb)->end_seq;
3599 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3602 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3603 SOCK_DEBUG(sk, "ofo packet was already received \n");
3604 __skb_unlink(skb, &tp->out_of_order_queue);
3608 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3609 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3610 TCP_SKB_CB(skb)->end_seq);
3612 __skb_unlink(skb, &tp->out_of_order_queue);
3613 __skb_queue_tail(&sk->sk_receive_queue, skb);
3614 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3615 if (tcp_hdr(skb)->fin)
3616 tcp_fin(skb, sk, tcp_hdr(skb));
3620 static int tcp_prune_queue(struct sock *sk);
3622 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3624 struct tcphdr *th = tcp_hdr(skb);
3625 struct tcp_sock *tp = tcp_sk(sk);
3628 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3631 __skb_pull(skb, th->doff*4);
3633 TCP_ECN_accept_cwr(tp, skb);
3635 if (tp->rx_opt.dsack) {
3636 tp->rx_opt.dsack = 0;
3637 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3638 4 - tp->rx_opt.tstamp_ok);
3641 /* Queue data for delivery to the user.
3642 * Packets in sequence go to the receive queue.
3643 * Out of sequence packets to the out_of_order_queue.
3645 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3646 if (tcp_receive_window(tp) == 0)
3649 /* Ok. In sequence. In window. */
3650 if (tp->ucopy.task == current &&
3651 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3652 sock_owned_by_user(sk) && !tp->urg_data) {
3653 int chunk = min_t(unsigned int, skb->len,
3656 __set_current_state(TASK_RUNNING);
3659 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3660 tp->ucopy.len -= chunk;
3661 tp->copied_seq += chunk;
3662 eaten = (chunk == skb->len && !th->fin);
3663 tcp_rcv_space_adjust(sk);
3671 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3672 !sk_stream_rmem_schedule(sk, skb))) {
3673 if (tcp_prune_queue(sk) < 0 ||
3674 !sk_stream_rmem_schedule(sk, skb))
3677 sk_stream_set_owner_r(skb, sk);
3678 __skb_queue_tail(&sk->sk_receive_queue, skb);
3680 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3682 tcp_event_data_recv(sk, skb);
3684 tcp_fin(skb, sk, th);
3686 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3689 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3690 * gap in queue is filled.
3692 if (skb_queue_empty(&tp->out_of_order_queue))
3693 inet_csk(sk)->icsk_ack.pingpong = 0;
3696 if (tp->rx_opt.num_sacks)
3697 tcp_sack_remove(tp);
3699 tcp_fast_path_check(sk);
3703 else if (!sock_flag(sk, SOCK_DEAD))
3704 sk->sk_data_ready(sk, 0);
3708 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3709 /* A retransmit, 2nd most common case. Force an immediate ack. */
3710 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3711 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3714 tcp_enter_quickack_mode(sk);
3715 inet_csk_schedule_ack(sk);
3721 /* Out of window. F.e. zero window probe. */
3722 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3725 tcp_enter_quickack_mode(sk);
3727 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3728 /* Partial packet, seq < rcv_next < end_seq */
3729 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3730 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3731 TCP_SKB_CB(skb)->end_seq);
3733 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3735 /* If window is closed, drop tail of packet. But after
3736 * remembering D-SACK for its head made in previous line.
3738 if (!tcp_receive_window(tp))
3743 TCP_ECN_check_ce(tp, skb);
3745 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3746 !sk_stream_rmem_schedule(sk, skb)) {
3747 if (tcp_prune_queue(sk) < 0 ||
3748 !sk_stream_rmem_schedule(sk, skb))
3752 /* Disable header prediction. */
3754 inet_csk_schedule_ack(sk);
3756 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3757 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3759 sk_stream_set_owner_r(skb, sk);
3761 if (!skb_peek(&tp->out_of_order_queue)) {
3762 /* Initial out of order segment, build 1 SACK. */
3763 if (tcp_is_sack(tp)) {
3764 tp->rx_opt.num_sacks = 1;
3765 tp->rx_opt.dsack = 0;
3766 tp->rx_opt.eff_sacks = 1;
3767 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3768 tp->selective_acks[0].end_seq =
3769 TCP_SKB_CB(skb)->end_seq;
3771 __skb_queue_head(&tp->out_of_order_queue,skb);
3773 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3774 u32 seq = TCP_SKB_CB(skb)->seq;
3775 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3777 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3778 __skb_append(skb1, skb, &tp->out_of_order_queue);
3780 if (!tp->rx_opt.num_sacks ||
3781 tp->selective_acks[0].end_seq != seq)
3784 /* Common case: data arrive in order after hole. */
3785 tp->selective_acks[0].end_seq = end_seq;
3789 /* Find place to insert this segment. */
3791 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3793 } while ((skb1 = skb1->prev) !=
3794 (struct sk_buff*)&tp->out_of_order_queue);
3796 /* Do skb overlap to previous one? */
3797 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3798 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3799 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3800 /* All the bits are present. Drop. */
3802 tcp_dsack_set(tp, seq, end_seq);
3805 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3806 /* Partial overlap. */
3807 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3812 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3814 /* And clean segments covered by new one as whole. */
3815 while ((skb1 = skb->next) !=
3816 (struct sk_buff*)&tp->out_of_order_queue &&
3817 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3818 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3819 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3822 __skb_unlink(skb1, &tp->out_of_order_queue);
3823 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3828 if (tcp_is_sack(tp))
3829 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3833 /* Collapse contiguous sequence of skbs head..tail with
3834 * sequence numbers start..end.
3835 * Segments with FIN/SYN are not collapsed (only because this
3839 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3840 struct sk_buff *head, struct sk_buff *tail,
3843 struct sk_buff *skb;
3845 /* First, check that queue is collapsible and find
3846 * the point where collapsing can be useful. */
3847 for (skb = head; skb != tail; ) {
3848 /* No new bits? It is possible on ofo queue. */
3849 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3850 struct sk_buff *next = skb->next;
3851 __skb_unlink(skb, list);
3853 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3858 /* The first skb to collapse is:
3860 * - bloated or contains data before "start" or
3861 * overlaps to the next one.
3863 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3864 (tcp_win_from_space(skb->truesize) > skb->len ||
3865 before(TCP_SKB_CB(skb)->seq, start) ||
3866 (skb->next != tail &&
3867 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3870 /* Decided to skip this, advance start seq. */
3871 start = TCP_SKB_CB(skb)->end_seq;
3874 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3877 while (before(start, end)) {
3878 struct sk_buff *nskb;
3879 int header = skb_headroom(skb);
3880 int copy = SKB_MAX_ORDER(header, 0);
3882 /* Too big header? This can happen with IPv6. */
3885 if (end-start < copy)
3887 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3891 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3892 skb_set_network_header(nskb, (skb_network_header(skb) -
3894 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3896 skb_reserve(nskb, header);
3897 memcpy(nskb->head, skb->head, header);
3898 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3899 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3900 __skb_insert(nskb, skb->prev, skb, list);
3901 sk_stream_set_owner_r(nskb, sk);
3903 /* Copy data, releasing collapsed skbs. */
3905 int offset = start - TCP_SKB_CB(skb)->seq;
3906 int size = TCP_SKB_CB(skb)->end_seq - start;
3910 size = min(copy, size);
3911 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3913 TCP_SKB_CB(nskb)->end_seq += size;
3917 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3918 struct sk_buff *next = skb->next;
3919 __skb_unlink(skb, list);
3921 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3924 tcp_hdr(skb)->syn ||
3932 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3933 * and tcp_collapse() them until all the queue is collapsed.
3935 static void tcp_collapse_ofo_queue(struct sock *sk)
3937 struct tcp_sock *tp = tcp_sk(sk);
3938 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3939 struct sk_buff *head;
3945 start = TCP_SKB_CB(skb)->seq;
3946 end = TCP_SKB_CB(skb)->end_seq;
3952 /* Segment is terminated when we see gap or when
3953 * we are at the end of all the queue. */
3954 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3955 after(TCP_SKB_CB(skb)->seq, end) ||
3956 before(TCP_SKB_CB(skb)->end_seq, start)) {
3957 tcp_collapse(sk, &tp->out_of_order_queue,
3958 head, skb, start, end);
3960 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3962 /* Start new segment */
3963 start = TCP_SKB_CB(skb)->seq;
3964 end = TCP_SKB_CB(skb)->end_seq;
3966 if (before(TCP_SKB_CB(skb)->seq, start))
3967 start = TCP_SKB_CB(skb)->seq;
3968 if (after(TCP_SKB_CB(skb)->end_seq, end))
3969 end = TCP_SKB_CB(skb)->end_seq;
3974 /* Reduce allocated memory if we can, trying to get
3975 * the socket within its memory limits again.
3977 * Return less than zero if we should start dropping frames
3978 * until the socket owning process reads some of the data
3979 * to stabilize the situation.
3981 static int tcp_prune_queue(struct sock *sk)
3983 struct tcp_sock *tp = tcp_sk(sk);
3985 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3987 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3989 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3990 tcp_clamp_window(sk);
3991 else if (tcp_memory_pressure)
3992 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3994 tcp_collapse_ofo_queue(sk);
3995 tcp_collapse(sk, &sk->sk_receive_queue,
3996 sk->sk_receive_queue.next,
3997 (struct sk_buff*)&sk->sk_receive_queue,
3998 tp->copied_seq, tp->rcv_nxt);
3999 sk_stream_mem_reclaim(sk);
4001 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4004 /* Collapsing did not help, destructive actions follow.
4005 * This must not ever occur. */
4007 /* First, purge the out_of_order queue. */
4008 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4009 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
4010 __skb_queue_purge(&tp->out_of_order_queue);
4012 /* Reset SACK state. A conforming SACK implementation will
4013 * do the same at a timeout based retransmit. When a connection
4014 * is in a sad state like this, we care only about integrity
4015 * of the connection not performance.
4017 if (tcp_is_sack(tp))
4018 tcp_sack_reset(&tp->rx_opt);
4019 sk_stream_mem_reclaim(sk);
4022 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4025 /* If we are really being abused, tell the caller to silently
4026 * drop receive data on the floor. It will get retransmitted
4027 * and hopefully then we'll have sufficient space.
4029 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4031 /* Massive buffer overcommit. */
4037 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4038 * As additional protections, we do not touch cwnd in retransmission phases,
4039 * and if application hit its sndbuf limit recently.
4041 void tcp_cwnd_application_limited(struct sock *sk)
4043 struct tcp_sock *tp = tcp_sk(sk);
4045 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4046 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4047 /* Limited by application or receiver window. */
4048 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4049 u32 win_used = max(tp->snd_cwnd_used, init_win);
4050 if (win_used < tp->snd_cwnd) {
4051 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4052 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4054 tp->snd_cwnd_used = 0;
4056 tp->snd_cwnd_stamp = tcp_time_stamp;
4059 static int tcp_should_expand_sndbuf(struct sock *sk)
4061 struct tcp_sock *tp = tcp_sk(sk);
4063 /* If the user specified a specific send buffer setting, do
4066 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4069 /* If we are under global TCP memory pressure, do not expand. */
4070 if (tcp_memory_pressure)
4073 /* If we are under soft global TCP memory pressure, do not expand. */
4074 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4077 /* If we filled the congestion window, do not expand. */
4078 if (tp->packets_out >= tp->snd_cwnd)
4084 /* When incoming ACK allowed to free some skb from write_queue,
4085 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4086 * on the exit from tcp input handler.
4088 * PROBLEM: sndbuf expansion does not work well with largesend.
4090 static void tcp_new_space(struct sock *sk)
4092 struct tcp_sock *tp = tcp_sk(sk);
4094 if (tcp_should_expand_sndbuf(sk)) {
4095 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4096 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4097 demanded = max_t(unsigned int, tp->snd_cwnd,
4098 tp->reordering + 1);
4099 sndmem *= 2*demanded;
4100 if (sndmem > sk->sk_sndbuf)
4101 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4102 tp->snd_cwnd_stamp = tcp_time_stamp;
4105 sk->sk_write_space(sk);
4108 static void tcp_check_space(struct sock *sk)
4110 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4111 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4112 if (sk->sk_socket &&
4113 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4118 static inline void tcp_data_snd_check(struct sock *sk)
4120 tcp_push_pending_frames(sk);
4121 tcp_check_space(sk);
4125 * Check if sending an ack is needed.
4127 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4129 struct tcp_sock *tp = tcp_sk(sk);
4131 /* More than one full frame received... */
4132 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4133 /* ... and right edge of window advances far enough.
4134 * (tcp_recvmsg() will send ACK otherwise). Or...
4136 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4137 /* We ACK each frame or... */
4138 tcp_in_quickack_mode(sk) ||
4139 /* We have out of order data. */
4141 skb_peek(&tp->out_of_order_queue))) {
4142 /* Then ack it now */
4145 /* Else, send delayed ack. */
4146 tcp_send_delayed_ack(sk);
4150 static inline void tcp_ack_snd_check(struct sock *sk)
4152 if (!inet_csk_ack_scheduled(sk)) {
4153 /* We sent a data segment already. */
4156 __tcp_ack_snd_check(sk, 1);
4160 * This routine is only called when we have urgent data
4161 * signaled. Its the 'slow' part of tcp_urg. It could be
4162 * moved inline now as tcp_urg is only called from one
4163 * place. We handle URGent data wrong. We have to - as
4164 * BSD still doesn't use the correction from RFC961.
4165 * For 1003.1g we should support a new option TCP_STDURG to permit
4166 * either form (or just set the sysctl tcp_stdurg).
4169 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4171 struct tcp_sock *tp = tcp_sk(sk);
4172 u32 ptr = ntohs(th->urg_ptr);
4174 if (ptr && !sysctl_tcp_stdurg)
4176 ptr += ntohl(th->seq);
4178 /* Ignore urgent data that we've already seen and read. */
4179 if (after(tp->copied_seq, ptr))
4182 /* Do not replay urg ptr.
4184 * NOTE: interesting situation not covered by specs.
4185 * Misbehaving sender may send urg ptr, pointing to segment,
4186 * which we already have in ofo queue. We are not able to fetch
4187 * such data and will stay in TCP_URG_NOTYET until will be eaten
4188 * by recvmsg(). Seems, we are not obliged to handle such wicked
4189 * situations. But it is worth to think about possibility of some
4190 * DoSes using some hypothetical application level deadlock.
4192 if (before(ptr, tp->rcv_nxt))
4195 /* Do we already have a newer (or duplicate) urgent pointer? */
4196 if (tp->urg_data && !after(ptr, tp->urg_seq))
4199 /* Tell the world about our new urgent pointer. */
4202 /* We may be adding urgent data when the last byte read was
4203 * urgent. To do this requires some care. We cannot just ignore
4204 * tp->copied_seq since we would read the last urgent byte again
4205 * as data, nor can we alter copied_seq until this data arrives
4206 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4208 * NOTE. Double Dutch. Rendering to plain English: author of comment
4209 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4210 * and expect that both A and B disappear from stream. This is _wrong_.
4211 * Though this happens in BSD with high probability, this is occasional.
4212 * Any application relying on this is buggy. Note also, that fix "works"
4213 * only in this artificial test. Insert some normal data between A and B and we will
4214 * decline of BSD again. Verdict: it is better to remove to trap
4217 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4218 !sock_flag(sk, SOCK_URGINLINE) &&
4219 tp->copied_seq != tp->rcv_nxt) {
4220 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4222 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4223 __skb_unlink(skb, &sk->sk_receive_queue);
4228 tp->urg_data = TCP_URG_NOTYET;
4231 /* Disable header prediction. */
4235 /* This is the 'fast' part of urgent handling. */
4236 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4238 struct tcp_sock *tp = tcp_sk(sk);
4240 /* Check if we get a new urgent pointer - normally not. */
4242 tcp_check_urg(sk,th);
4244 /* Do we wait for any urgent data? - normally not... */
4245 if (tp->urg_data == TCP_URG_NOTYET) {
4246 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4249 /* Is the urgent pointer pointing into this packet? */
4250 if (ptr < skb->len) {
4252 if (skb_copy_bits(skb, ptr, &tmp, 1))
4254 tp->urg_data = TCP_URG_VALID | tmp;
4255 if (!sock_flag(sk, SOCK_DEAD))
4256 sk->sk_data_ready(sk, 0);
4261 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4263 struct tcp_sock *tp = tcp_sk(sk);
4264 int chunk = skb->len - hlen;
4268 if (skb_csum_unnecessary(skb))
4269 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4271 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4275 tp->ucopy.len -= chunk;
4276 tp->copied_seq += chunk;
4277 tcp_rcv_space_adjust(sk);
4284 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4288 if (sock_owned_by_user(sk)) {
4290 result = __tcp_checksum_complete(skb);
4293 result = __tcp_checksum_complete(skb);
4298 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4300 return !skb_csum_unnecessary(skb) &&
4301 __tcp_checksum_complete_user(sk, skb);
4304 #ifdef CONFIG_NET_DMA
4305 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4307 struct tcp_sock *tp = tcp_sk(sk);
4308 int chunk = skb->len - hlen;
4310 int copied_early = 0;
4312 if (tp->ucopy.wakeup)
4315 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4316 tp->ucopy.dma_chan = get_softnet_dma();
4318 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4320 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4321 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4326 tp->ucopy.dma_cookie = dma_cookie;
4329 tp->ucopy.len -= chunk;
4330 tp->copied_seq += chunk;
4331 tcp_rcv_space_adjust(sk);
4333 if ((tp->ucopy.len == 0) ||
4334 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4335 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4336 tp->ucopy.wakeup = 1;
4337 sk->sk_data_ready(sk, 0);
4339 } else if (chunk > 0) {
4340 tp->ucopy.wakeup = 1;
4341 sk->sk_data_ready(sk, 0);
4344 return copied_early;
4346 #endif /* CONFIG_NET_DMA */
4349 * TCP receive function for the ESTABLISHED state.
4351 * It is split into a fast path and a slow path. The fast path is
4353 * - A zero window was announced from us - zero window probing
4354 * is only handled properly in the slow path.
4355 * - Out of order segments arrived.
4356 * - Urgent data is expected.
4357 * - There is no buffer space left
4358 * - Unexpected TCP flags/window values/header lengths are received
4359 * (detected by checking the TCP header against pred_flags)
4360 * - Data is sent in both directions. Fast path only supports pure senders
4361 * or pure receivers (this means either the sequence number or the ack
4362 * value must stay constant)
4363 * - Unexpected TCP option.
4365 * When these conditions are not satisfied it drops into a standard
4366 * receive procedure patterned after RFC793 to handle all cases.
4367 * The first three cases are guaranteed by proper pred_flags setting,
4368 * the rest is checked inline. Fast processing is turned on in
4369 * tcp_data_queue when everything is OK.
4371 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4372 struct tcphdr *th, unsigned len)
4374 struct tcp_sock *tp = tcp_sk(sk);
4377 * Header prediction.
4378 * The code loosely follows the one in the famous
4379 * "30 instruction TCP receive" Van Jacobson mail.
4381 * Van's trick is to deposit buffers into socket queue
4382 * on a device interrupt, to call tcp_recv function
4383 * on the receive process context and checksum and copy
4384 * the buffer to user space. smart...
4386 * Our current scheme is not silly either but we take the
4387 * extra cost of the net_bh soft interrupt processing...
4388 * We do checksum and copy also but from device to kernel.
4391 tp->rx_opt.saw_tstamp = 0;
4393 /* pred_flags is 0xS?10 << 16 + snd_wnd
4394 * if header_prediction is to be made
4395 * 'S' will always be tp->tcp_header_len >> 2
4396 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4397 * turn it off (when there are holes in the receive
4398 * space for instance)
4399 * PSH flag is ignored.
4402 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4403 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4404 int tcp_header_len = tp->tcp_header_len;
4406 /* Timestamp header prediction: tcp_header_len
4407 * is automatically equal to th->doff*4 due to pred_flags
4411 /* Check timestamp */
4412 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4413 __be32 *ptr = (__be32 *)(th + 1);
4415 /* No? Slow path! */
4416 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4417 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4420 tp->rx_opt.saw_tstamp = 1;
4422 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4424 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4426 /* If PAWS failed, check it more carefully in slow path */
4427 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4430 /* DO NOT update ts_recent here, if checksum fails
4431 * and timestamp was corrupted part, it will result
4432 * in a hung connection since we will drop all
4433 * future packets due to the PAWS test.
4437 if (len <= tcp_header_len) {
4438 /* Bulk data transfer: sender */
4439 if (len == tcp_header_len) {
4440 /* Predicted packet is in window by definition.
4441 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4442 * Hence, check seq<=rcv_wup reduces to:
4444 if (tcp_header_len ==
4445 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4446 tp->rcv_nxt == tp->rcv_wup)
4447 tcp_store_ts_recent(tp);
4449 /* We know that such packets are checksummed
4452 tcp_ack(sk, skb, 0);
4454 tcp_data_snd_check(sk);
4456 } else { /* Header too small */
4457 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4462 int copied_early = 0;
4464 if (tp->copied_seq == tp->rcv_nxt &&
4465 len - tcp_header_len <= tp->ucopy.len) {
4466 #ifdef CONFIG_NET_DMA
4467 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4472 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4473 __set_current_state(TASK_RUNNING);
4475 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4479 /* Predicted packet is in window by definition.
4480 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4481 * Hence, check seq<=rcv_wup reduces to:
4483 if (tcp_header_len ==
4484 (sizeof(struct tcphdr) +
4485 TCPOLEN_TSTAMP_ALIGNED) &&
4486 tp->rcv_nxt == tp->rcv_wup)
4487 tcp_store_ts_recent(tp);
4489 tcp_rcv_rtt_measure_ts(sk, skb);
4491 __skb_pull(skb, tcp_header_len);
4492 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4493 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4496 tcp_cleanup_rbuf(sk, skb->len);
4499 if (tcp_checksum_complete_user(sk, skb))
4502 /* Predicted packet is in window by definition.
4503 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4504 * Hence, check seq<=rcv_wup reduces to:
4506 if (tcp_header_len ==
4507 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4508 tp->rcv_nxt == tp->rcv_wup)
4509 tcp_store_ts_recent(tp);
4511 tcp_rcv_rtt_measure_ts(sk, skb);
4513 if ((int)skb->truesize > sk->sk_forward_alloc)
4516 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4518 /* Bulk data transfer: receiver */
4519 __skb_pull(skb,tcp_header_len);
4520 __skb_queue_tail(&sk->sk_receive_queue, skb);
4521 sk_stream_set_owner_r(skb, sk);
4522 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4525 tcp_event_data_recv(sk, skb);
4527 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4528 /* Well, only one small jumplet in fast path... */
4529 tcp_ack(sk, skb, FLAG_DATA);
4530 tcp_data_snd_check(sk);
4531 if (!inet_csk_ack_scheduled(sk))
4535 __tcp_ack_snd_check(sk, 0);
4537 #ifdef CONFIG_NET_DMA
4539 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4545 sk->sk_data_ready(sk, 0);
4551 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4555 * RFC1323: H1. Apply PAWS check first.
4557 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4558 tcp_paws_discard(sk, skb)) {
4560 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4561 tcp_send_dupack(sk, skb);
4564 /* Resets are accepted even if PAWS failed.
4566 ts_recent update must be made after we are sure
4567 that the packet is in window.
4572 * Standard slow path.
4575 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4576 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4577 * (RST) segments are validated by checking their SEQ-fields."
4578 * And page 69: "If an incoming segment is not acceptable,
4579 * an acknowledgment should be sent in reply (unless the RST bit
4580 * is set, if so drop the segment and return)".
4583 tcp_send_dupack(sk, skb);
4592 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4594 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4595 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4596 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4603 tcp_ack(sk, skb, FLAG_SLOWPATH);
4605 tcp_rcv_rtt_measure_ts(sk, skb);
4607 /* Process urgent data. */
4608 tcp_urg(sk, skb, th);
4610 /* step 7: process the segment text */
4611 tcp_data_queue(sk, skb);
4613 tcp_data_snd_check(sk);
4614 tcp_ack_snd_check(sk);
4618 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4625 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4626 struct tcphdr *th, unsigned len)
4628 struct tcp_sock *tp = tcp_sk(sk);
4629 struct inet_connection_sock *icsk = inet_csk(sk);
4630 int saved_clamp = tp->rx_opt.mss_clamp;
4632 tcp_parse_options(skb, &tp->rx_opt, 0);
4636 * "If the state is SYN-SENT then
4637 * first check the ACK bit
4638 * If the ACK bit is set
4639 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4640 * a reset (unless the RST bit is set, if so drop
4641 * the segment and return)"
4643 * We do not send data with SYN, so that RFC-correct
4646 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4647 goto reset_and_undo;
4649 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4650 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4652 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4653 goto reset_and_undo;
4656 /* Now ACK is acceptable.
4658 * "If the RST bit is set
4659 * If the ACK was acceptable then signal the user "error:
4660 * connection reset", drop the segment, enter CLOSED state,
4661 * delete TCB, and return."
4670 * "fifth, if neither of the SYN or RST bits is set then
4671 * drop the segment and return."
4677 goto discard_and_undo;
4680 * "If the SYN bit is on ...
4681 * are acceptable then ...
4682 * (our SYN has been ACKed), change the connection
4683 * state to ESTABLISHED..."
4686 TCP_ECN_rcv_synack(tp, th);
4688 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4689 tcp_ack(sk, skb, FLAG_SLOWPATH);
4691 /* Ok.. it's good. Set up sequence numbers and
4692 * move to established.
4694 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4695 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4697 /* RFC1323: The window in SYN & SYN/ACK segments is
4700 tp->snd_wnd = ntohs(th->window);
4701 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4703 if (!tp->rx_opt.wscale_ok) {
4704 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4705 tp->window_clamp = min(tp->window_clamp, 65535U);
4708 if (tp->rx_opt.saw_tstamp) {
4709 tp->rx_opt.tstamp_ok = 1;
4710 tp->tcp_header_len =
4711 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4712 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4713 tcp_store_ts_recent(tp);
4715 tp->tcp_header_len = sizeof(struct tcphdr);
4718 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4719 tcp_enable_fack(tp);
4722 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4723 tcp_initialize_rcv_mss(sk);
4725 /* Remember, tcp_poll() does not lock socket!
4726 * Change state from SYN-SENT only after copied_seq
4727 * is initialized. */
4728 tp->copied_seq = tp->rcv_nxt;
4730 tcp_set_state(sk, TCP_ESTABLISHED);
4732 security_inet_conn_established(sk, skb);
4734 /* Make sure socket is routed, for correct metrics. */
4735 icsk->icsk_af_ops->rebuild_header(sk);
4737 tcp_init_metrics(sk);
4739 tcp_init_congestion_control(sk);
4741 /* Prevent spurious tcp_cwnd_restart() on first data
4744 tp->lsndtime = tcp_time_stamp;
4746 tcp_init_buffer_space(sk);
4748 if (sock_flag(sk, SOCK_KEEPOPEN))
4749 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4751 if (!tp->rx_opt.snd_wscale)
4752 __tcp_fast_path_on(tp, tp->snd_wnd);
4756 if (!sock_flag(sk, SOCK_DEAD)) {
4757 sk->sk_state_change(sk);
4758 sk_wake_async(sk, 0, POLL_OUT);
4761 if (sk->sk_write_pending ||
4762 icsk->icsk_accept_queue.rskq_defer_accept ||
4763 icsk->icsk_ack.pingpong) {
4764 /* Save one ACK. Data will be ready after
4765 * several ticks, if write_pending is set.
4767 * It may be deleted, but with this feature tcpdumps
4768 * look so _wonderfully_ clever, that I was not able
4769 * to stand against the temptation 8) --ANK
4771 inet_csk_schedule_ack(sk);
4772 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4773 icsk->icsk_ack.ato = TCP_ATO_MIN;
4774 tcp_incr_quickack(sk);
4775 tcp_enter_quickack_mode(sk);
4776 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4777 TCP_DELACK_MAX, TCP_RTO_MAX);
4788 /* No ACK in the segment */
4792 * "If the RST bit is set
4794 * Otherwise (no ACK) drop the segment and return."
4797 goto discard_and_undo;
4801 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4802 goto discard_and_undo;
4805 /* We see SYN without ACK. It is attempt of
4806 * simultaneous connect with crossed SYNs.
4807 * Particularly, it can be connect to self.
4809 tcp_set_state(sk, TCP_SYN_RECV);
4811 if (tp->rx_opt.saw_tstamp) {
4812 tp->rx_opt.tstamp_ok = 1;
4813 tcp_store_ts_recent(tp);
4814 tp->tcp_header_len =
4815 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4817 tp->tcp_header_len = sizeof(struct tcphdr);
4820 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4821 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4823 /* RFC1323: The window in SYN & SYN/ACK segments is
4826 tp->snd_wnd = ntohs(th->window);
4827 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4828 tp->max_window = tp->snd_wnd;
4830 TCP_ECN_rcv_syn(tp, th);
4833 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4834 tcp_initialize_rcv_mss(sk);
4837 tcp_send_synack(sk);
4839 /* Note, we could accept data and URG from this segment.
4840 * There are no obstacles to make this.
4842 * However, if we ignore data in ACKless segments sometimes,
4843 * we have no reasons to accept it sometimes.
4844 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4845 * is not flawless. So, discard packet for sanity.
4846 * Uncomment this return to process the data.
4853 /* "fifth, if neither of the SYN or RST bits is set then
4854 * drop the segment and return."
4858 tcp_clear_options(&tp->rx_opt);
4859 tp->rx_opt.mss_clamp = saved_clamp;
4863 tcp_clear_options(&tp->rx_opt);
4864 tp->rx_opt.mss_clamp = saved_clamp;
4870 * This function implements the receiving procedure of RFC 793 for
4871 * all states except ESTABLISHED and TIME_WAIT.
4872 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4873 * address independent.
4876 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4877 struct tcphdr *th, unsigned len)
4879 struct tcp_sock *tp = tcp_sk(sk);
4880 struct inet_connection_sock *icsk = inet_csk(sk);
4883 tp->rx_opt.saw_tstamp = 0;
4885 switch (sk->sk_state) {
4897 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4900 /* Now we have several options: In theory there is
4901 * nothing else in the frame. KA9Q has an option to
4902 * send data with the syn, BSD accepts data with the
4903 * syn up to the [to be] advertised window and
4904 * Solaris 2.1 gives you a protocol error. For now
4905 * we just ignore it, that fits the spec precisely
4906 * and avoids incompatibilities. It would be nice in
4907 * future to drop through and process the data.
4909 * Now that TTCP is starting to be used we ought to
4911 * But, this leaves one open to an easy denial of
4912 * service attack, and SYN cookies can't defend
4913 * against this problem. So, we drop the data
4914 * in the interest of security over speed unless
4915 * it's still in use.
4923 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4927 /* Do step6 onward by hand. */
4928 tcp_urg(sk, skb, th);
4930 tcp_data_snd_check(sk);
4934 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4935 tcp_paws_discard(sk, skb)) {
4937 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4938 tcp_send_dupack(sk, skb);
4941 /* Reset is accepted even if it did not pass PAWS. */
4944 /* step 1: check sequence number */
4945 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4947 tcp_send_dupack(sk, skb);
4951 /* step 2: check RST bit */
4957 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4959 /* step 3: check security and precedence [ignored] */
4963 * Check for a SYN in window.
4965 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4966 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4971 /* step 5: check the ACK field */
4973 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4975 switch (sk->sk_state) {
4978 tp->copied_seq = tp->rcv_nxt;
4980 tcp_set_state(sk, TCP_ESTABLISHED);
4981 sk->sk_state_change(sk);
4983 /* Note, that this wakeup is only for marginal
4984 * crossed SYN case. Passively open sockets
4985 * are not waked up, because sk->sk_sleep ==
4986 * NULL and sk->sk_socket == NULL.
4988 if (sk->sk_socket) {
4989 sk_wake_async(sk,0,POLL_OUT);
4992 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4993 tp->snd_wnd = ntohs(th->window) <<
4994 tp->rx_opt.snd_wscale;
4995 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4996 TCP_SKB_CB(skb)->seq);
4998 /* tcp_ack considers this ACK as duplicate
4999 * and does not calculate rtt.
5000 * Fix it at least with timestamps.
5002 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5004 tcp_ack_saw_tstamp(sk, 0);
5006 if (tp->rx_opt.tstamp_ok)
5007 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5009 /* Make sure socket is routed, for
5012 icsk->icsk_af_ops->rebuild_header(sk);
5014 tcp_init_metrics(sk);
5016 tcp_init_congestion_control(sk);
5018 /* Prevent spurious tcp_cwnd_restart() on
5019 * first data packet.
5021 tp->lsndtime = tcp_time_stamp;
5024 tcp_initialize_rcv_mss(sk);
5025 tcp_init_buffer_space(sk);
5026 tcp_fast_path_on(tp);
5033 if (tp->snd_una == tp->write_seq) {
5034 tcp_set_state(sk, TCP_FIN_WAIT2);
5035 sk->sk_shutdown |= SEND_SHUTDOWN;
5036 dst_confirm(sk->sk_dst_cache);
5038 if (!sock_flag(sk, SOCK_DEAD))
5039 /* Wake up lingering close() */
5040 sk->sk_state_change(sk);
5044 if (tp->linger2 < 0 ||
5045 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5046 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5048 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5052 tmo = tcp_fin_time(sk);
5053 if (tmo > TCP_TIMEWAIT_LEN) {
5054 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5055 } else if (th->fin || sock_owned_by_user(sk)) {
5056 /* Bad case. We could lose such FIN otherwise.
5057 * It is not a big problem, but it looks confusing
5058 * and not so rare event. We still can lose it now,
5059 * if it spins in bh_lock_sock(), but it is really
5062 inet_csk_reset_keepalive_timer(sk, tmo);
5064 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5072 if (tp->snd_una == tp->write_seq) {
5073 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5079 if (tp->snd_una == tp->write_seq) {
5080 tcp_update_metrics(sk);
5089 /* step 6: check the URG bit */
5090 tcp_urg(sk, skb, th);
5092 /* step 7: process the segment text */
5093 switch (sk->sk_state) {
5094 case TCP_CLOSE_WAIT:
5097 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5101 /* RFC 793 says to queue data in these states,
5102 * RFC 1122 says we MUST send a reset.
5103 * BSD 4.4 also does reset.
5105 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5106 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5107 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5108 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5114 case TCP_ESTABLISHED:
5115 tcp_data_queue(sk, skb);
5120 /* tcp_data could move socket to TIME-WAIT */
5121 if (sk->sk_state != TCP_CLOSE) {
5122 tcp_data_snd_check(sk);
5123 tcp_ack_snd_check(sk);
5133 EXPORT_SYMBOL(sysctl_tcp_ecn);
5134 EXPORT_SYMBOL(sysctl_tcp_reordering);
5135 EXPORT_SYMBOL(tcp_parse_options);
5136 EXPORT_SYMBOL(tcp_rcv_established);
5137 EXPORT_SYMBOL(tcp_rcv_state_process);
5138 EXPORT_SYMBOL(tcp_initialize_rcv_mss);