2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presence of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
101 #define FLAG_ECE 0x40 /* ECE in this ACK */
102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained DSACK info */
108 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
109 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
110 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
111 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
112 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
114 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
115 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
116 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
118 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
120 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
121 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
123 /* Adapt the MSS value used to make delayed ack decision to the
126 static void tcp_measure_rcv_mss(struct sock *sk,
127 const struct sk_buff *skb)
129 struct inet_connection_sock *icsk = inet_csk(sk);
130 const unsigned int lss = icsk->icsk_ack.last_seg_size;
133 icsk->icsk_ack.last_seg_size = 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len = skb_shinfo(skb)->gso_size ?: skb->len;
139 if (len >= icsk->icsk_ack.rcv_mss) {
140 icsk->icsk_ack.rcv_mss = len;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len += skb->data - skb_transport_header(skb);
148 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len -= tcp_sk(sk)->tcp_header_len;
161 icsk->icsk_ack.last_seg_size = len;
163 icsk->icsk_ack.rcv_mss = len;
167 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
173 static void tcp_incr_quickack(struct sock *sk)
175 struct inet_connection_sock *icsk = inet_csk(sk);
176 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
180 if (quickacks > icsk->icsk_ack.quick)
181 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
184 void tcp_enter_quickack_mode(struct sock *sk)
186 struct inet_connection_sock *icsk = inet_csk(sk);
187 tcp_incr_quickack(sk);
188 icsk->icsk_ack.pingpong = 0;
189 icsk->icsk_ack.ato = TCP_ATO_MIN;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline int tcp_in_quickack_mode(const struct sock *sk)
198 const struct inet_connection_sock *icsk = inet_csk(sk);
199 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
202 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
204 if (tp->ecn_flags&TCP_ECN_OK)
205 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
208 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
210 if (tcp_hdr(skb)->cwr)
211 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
216 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
219 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
221 if (tp->ecn_flags&TCP_ECN_OK) {
222 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
223 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
224 /* Funny extension: if ECT is not set on a segment,
225 * it is surely retransmit. It is not in ECN RFC,
226 * but Linux follows this rule. */
227 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
228 tcp_enter_quickack_mode((struct sock *)tp);
232 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
234 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
235 tp->ecn_flags &= ~TCP_ECN_OK;
238 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
240 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
241 tp->ecn_flags &= ~TCP_ECN_OK;
244 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
246 if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
251 /* Buffer size and advertised window tuning.
253 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
256 static void tcp_fixup_sndbuf(struct sock *sk)
258 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
259 sizeof(struct sk_buff);
261 if (sk->sk_sndbuf < 3 * sndmem)
262 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
265 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
267 * All tcp_full_space() is split to two parts: "network" buffer, allocated
268 * forward and advertised in receiver window (tp->rcv_wnd) and
269 * "application buffer", required to isolate scheduling/application
270 * latencies from network.
271 * window_clamp is maximal advertised window. It can be less than
272 * tcp_full_space(), in this case tcp_full_space() - window_clamp
273 * is reserved for "application" buffer. The less window_clamp is
274 * the smoother our behaviour from viewpoint of network, but the lower
275 * throughput and the higher sensitivity of the connection to losses. 8)
277 * rcv_ssthresh is more strict window_clamp used at "slow start"
278 * phase to predict further behaviour of this connection.
279 * It is used for two goals:
280 * - to enforce header prediction at sender, even when application
281 * requires some significant "application buffer". It is check #1.
282 * - to prevent pruning of receive queue because of misprediction
283 * of receiver window. Check #2.
285 * The scheme does not work when sender sends good segments opening
286 * window and then starts to feed us spaghetti. But it should work
287 * in common situations. Otherwise, we have to rely on queue collapsing.
290 /* Slow part of check#2. */
291 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
293 struct tcp_sock *tp = tcp_sk(sk);
295 int truesize = tcp_win_from_space(skb->truesize)/2;
296 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
298 while (tp->rcv_ssthresh <= window) {
299 if (truesize <= skb->len)
300 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
308 static void tcp_grow_window(struct sock *sk,
311 struct tcp_sock *tp = tcp_sk(sk);
314 if (tp->rcv_ssthresh < tp->window_clamp &&
315 (int)tp->rcv_ssthresh < tcp_space(sk) &&
316 !tcp_memory_pressure) {
319 /* Check #2. Increase window, if skb with such overhead
320 * will fit to rcvbuf in future.
322 if (tcp_win_from_space(skb->truesize) <= skb->len)
325 incr = __tcp_grow_window(sk, skb);
328 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
329 inet_csk(sk)->icsk_ack.quick |= 1;
334 /* 3. Tuning rcvbuf, when connection enters established state. */
336 static void tcp_fixup_rcvbuf(struct sock *sk)
338 struct tcp_sock *tp = tcp_sk(sk);
339 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
341 /* Try to select rcvbuf so that 4 mss-sized segments
342 * will fit to window and corresponding skbs will fit to our rcvbuf.
343 * (was 3; 4 is minimum to allow fast retransmit to work.)
345 while (tcp_win_from_space(rcvmem) < tp->advmss)
347 if (sk->sk_rcvbuf < 4 * rcvmem)
348 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
351 /* 4. Try to fixup all. It is made immediately after connection enters
354 static void tcp_init_buffer_space(struct sock *sk)
356 struct tcp_sock *tp = tcp_sk(sk);
359 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
360 tcp_fixup_rcvbuf(sk);
361 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
362 tcp_fixup_sndbuf(sk);
364 tp->rcvq_space.space = tp->rcv_wnd;
366 maxwin = tcp_full_space(sk);
368 if (tp->window_clamp >= maxwin) {
369 tp->window_clamp = maxwin;
371 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
372 tp->window_clamp = max(maxwin -
373 (maxwin >> sysctl_tcp_app_win),
377 /* Force reservation of one segment. */
378 if (sysctl_tcp_app_win &&
379 tp->window_clamp > 2 * tp->advmss &&
380 tp->window_clamp + tp->advmss > maxwin)
381 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
383 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
384 tp->snd_cwnd_stamp = tcp_time_stamp;
387 /* 5. Recalculate window clamp after socket hit its memory bounds. */
388 static void tcp_clamp_window(struct sock *sk)
390 struct tcp_sock *tp = tcp_sk(sk);
391 struct inet_connection_sock *icsk = inet_csk(sk);
393 icsk->icsk_ack.quick = 0;
395 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
396 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
397 !tcp_memory_pressure &&
398 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
399 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
402 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
403 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
407 /* Initialize RCV_MSS value.
408 * RCV_MSS is an our guess about MSS used by the peer.
409 * We haven't any direct information about the MSS.
410 * It's better to underestimate the RCV_MSS rather than overestimate.
411 * Overestimations make us ACKing less frequently than needed.
412 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
414 void tcp_initialize_rcv_mss(struct sock *sk)
416 struct tcp_sock *tp = tcp_sk(sk);
417 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
419 hint = min(hint, tp->rcv_wnd/2);
420 hint = min(hint, TCP_MIN_RCVMSS);
421 hint = max(hint, TCP_MIN_MSS);
423 inet_csk(sk)->icsk_ack.rcv_mss = hint;
426 /* Receiver "autotuning" code.
428 * The algorithm for RTT estimation w/o timestamps is based on
429 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
430 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
432 * More detail on this code can be found at
433 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
434 * though this reference is out of date. A new paper
437 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
439 u32 new_sample = tp->rcv_rtt_est.rtt;
445 if (new_sample != 0) {
446 /* If we sample in larger samples in the non-timestamp
447 * case, we could grossly overestimate the RTT especially
448 * with chatty applications or bulk transfer apps which
449 * are stalled on filesystem I/O.
451 * Also, since we are only going for a minimum in the
452 * non-timestamp case, we do not smooth things out
453 * else with timestamps disabled convergence takes too
457 m -= (new_sample >> 3);
459 } else if (m < new_sample)
462 /* No previous measure. */
466 if (tp->rcv_rtt_est.rtt != new_sample)
467 tp->rcv_rtt_est.rtt = new_sample;
470 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
472 if (tp->rcv_rtt_est.time == 0)
474 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
476 tcp_rcv_rtt_update(tp,
477 jiffies - tp->rcv_rtt_est.time,
481 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
482 tp->rcv_rtt_est.time = tcp_time_stamp;
485 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
487 struct tcp_sock *tp = tcp_sk(sk);
488 if (tp->rx_opt.rcv_tsecr &&
489 (TCP_SKB_CB(skb)->end_seq -
490 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
491 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
495 * This function should be called every time data is copied to user space.
496 * It calculates the appropriate TCP receive buffer space.
498 void tcp_rcv_space_adjust(struct sock *sk)
500 struct tcp_sock *tp = tcp_sk(sk);
504 if (tp->rcvq_space.time == 0)
507 time = tcp_time_stamp - tp->rcvq_space.time;
508 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
509 tp->rcv_rtt_est.rtt == 0)
512 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
514 space = max(tp->rcvq_space.space, space);
516 if (tp->rcvq_space.space != space) {
519 tp->rcvq_space.space = space;
521 if (sysctl_tcp_moderate_rcvbuf &&
522 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
523 int new_clamp = space;
525 /* Receive space grows, normalize in order to
526 * take into account packet headers and sk_buff
527 * structure overhead.
532 rcvmem = (tp->advmss + MAX_TCP_HEADER +
533 16 + sizeof(struct sk_buff));
534 while (tcp_win_from_space(rcvmem) < tp->advmss)
537 space = min(space, sysctl_tcp_rmem[2]);
538 if (space > sk->sk_rcvbuf) {
539 sk->sk_rcvbuf = space;
541 /* Make the window clamp follow along. */
542 tp->window_clamp = new_clamp;
548 tp->rcvq_space.seq = tp->copied_seq;
549 tp->rcvq_space.time = tcp_time_stamp;
552 /* There is something which you must keep in mind when you analyze the
553 * behavior of the tp->ato delayed ack timeout interval. When a
554 * connection starts up, we want to ack as quickly as possible. The
555 * problem is that "good" TCP's do slow start at the beginning of data
556 * transmission. The means that until we send the first few ACK's the
557 * sender will sit on his end and only queue most of his data, because
558 * he can only send snd_cwnd unacked packets at any given time. For
559 * each ACK we send, he increments snd_cwnd and transmits more of his
562 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
564 struct tcp_sock *tp = tcp_sk(sk);
565 struct inet_connection_sock *icsk = inet_csk(sk);
568 inet_csk_schedule_ack(sk);
570 tcp_measure_rcv_mss(sk, skb);
572 tcp_rcv_rtt_measure(tp);
574 now = tcp_time_stamp;
576 if (!icsk->icsk_ack.ato) {
577 /* The _first_ data packet received, initialize
578 * delayed ACK engine.
580 tcp_incr_quickack(sk);
581 icsk->icsk_ack.ato = TCP_ATO_MIN;
583 int m = now - icsk->icsk_ack.lrcvtime;
585 if (m <= TCP_ATO_MIN/2) {
586 /* The fastest case is the first. */
587 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
588 } else if (m < icsk->icsk_ack.ato) {
589 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
590 if (icsk->icsk_ack.ato > icsk->icsk_rto)
591 icsk->icsk_ack.ato = icsk->icsk_rto;
592 } else if (m > icsk->icsk_rto) {
593 /* Too long gap. Apparently sender failed to
594 * restart window, so that we send ACKs quickly.
596 tcp_incr_quickack(sk);
597 sk_stream_mem_reclaim(sk);
600 icsk->icsk_ack.lrcvtime = now;
602 TCP_ECN_check_ce(tp, skb);
605 tcp_grow_window(sk, skb);
608 static u32 tcp_rto_min(struct sock *sk)
610 struct dst_entry *dst = __sk_dst_get(sk);
611 u32 rto_min = TCP_RTO_MIN;
613 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
614 rto_min = dst->metrics[RTAX_RTO_MIN-1];
618 /* Called to compute a smoothed rtt estimate. The data fed to this
619 * routine either comes from timestamps, or from segments that were
620 * known _not_ to have been retransmitted [see Karn/Partridge
621 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
622 * piece by Van Jacobson.
623 * NOTE: the next three routines used to be one big routine.
624 * To save cycles in the RFC 1323 implementation it was better to break
625 * it up into three procedures. -- erics
627 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
629 struct tcp_sock *tp = tcp_sk(sk);
630 long m = mrtt; /* RTT */
632 /* The following amusing code comes from Jacobson's
633 * article in SIGCOMM '88. Note that rtt and mdev
634 * are scaled versions of rtt and mean deviation.
635 * This is designed to be as fast as possible
636 * m stands for "measurement".
638 * On a 1990 paper the rto value is changed to:
639 * RTO = rtt + 4 * mdev
641 * Funny. This algorithm seems to be very broken.
642 * These formulae increase RTO, when it should be decreased, increase
643 * too slowly, when it should be increased quickly, decrease too quickly
644 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
645 * does not matter how to _calculate_ it. Seems, it was trap
646 * that VJ failed to avoid. 8)
651 m -= (tp->srtt >> 3); /* m is now error in rtt est */
652 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
654 m = -m; /* m is now abs(error) */
655 m -= (tp->mdev >> 2); /* similar update on mdev */
656 /* This is similar to one of Eifel findings.
657 * Eifel blocks mdev updates when rtt decreases.
658 * This solution is a bit different: we use finer gain
659 * for mdev in this case (alpha*beta).
660 * Like Eifel it also prevents growth of rto,
661 * but also it limits too fast rto decreases,
662 * happening in pure Eifel.
667 m -= (tp->mdev >> 2); /* similar update on mdev */
669 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
670 if (tp->mdev > tp->mdev_max) {
671 tp->mdev_max = tp->mdev;
672 if (tp->mdev_max > tp->rttvar)
673 tp->rttvar = tp->mdev_max;
675 if (after(tp->snd_una, tp->rtt_seq)) {
676 if (tp->mdev_max < tp->rttvar)
677 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
678 tp->rtt_seq = tp->snd_nxt;
679 tp->mdev_max = tcp_rto_min(sk);
682 /* no previous measure. */
683 tp->srtt = m<<3; /* take the measured time to be rtt */
684 tp->mdev = m<<1; /* make sure rto = 3*rtt */
685 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
686 tp->rtt_seq = tp->snd_nxt;
690 /* Calculate rto without backoff. This is the second half of Van Jacobson's
691 * routine referred to above.
693 static inline void tcp_set_rto(struct sock *sk)
695 const struct tcp_sock *tp = tcp_sk(sk);
696 /* Old crap is replaced with new one. 8)
699 * 1. If rtt variance happened to be less 50msec, it is hallucination.
700 * It cannot be less due to utterly erratic ACK generation made
701 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
702 * to do with delayed acks, because at cwnd>2 true delack timeout
703 * is invisible. Actually, Linux-2.4 also generates erratic
704 * ACKs in some circumstances.
706 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
708 /* 2. Fixups made earlier cannot be right.
709 * If we do not estimate RTO correctly without them,
710 * all the algo is pure shit and should be replaced
711 * with correct one. It is exactly, which we pretend to do.
715 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
716 * guarantees that rto is higher.
718 static inline void tcp_bound_rto(struct sock *sk)
720 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
721 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
724 /* Save metrics learned by this TCP session.
725 This function is called only, when TCP finishes successfully
726 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
728 void tcp_update_metrics(struct sock *sk)
730 struct tcp_sock *tp = tcp_sk(sk);
731 struct dst_entry *dst = __sk_dst_get(sk);
733 if (sysctl_tcp_nometrics_save)
738 if (dst && (dst->flags&DST_HOST)) {
739 const struct inet_connection_sock *icsk = inet_csk(sk);
742 if (icsk->icsk_backoff || !tp->srtt) {
743 /* This session failed to estimate rtt. Why?
744 * Probably, no packets returned in time.
747 if (!(dst_metric_locked(dst, RTAX_RTT)))
748 dst->metrics[RTAX_RTT-1] = 0;
752 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
754 /* If newly calculated rtt larger than stored one,
755 * store new one. Otherwise, use EWMA. Remember,
756 * rtt overestimation is always better than underestimation.
758 if (!(dst_metric_locked(dst, RTAX_RTT))) {
760 dst->metrics[RTAX_RTT-1] = tp->srtt;
762 dst->metrics[RTAX_RTT-1] -= (m>>3);
765 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
769 /* Scale deviation to rttvar fixed point */
774 if (m >= dst_metric(dst, RTAX_RTTVAR))
775 dst->metrics[RTAX_RTTVAR-1] = m;
777 dst->metrics[RTAX_RTTVAR-1] -=
778 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
781 if (tp->snd_ssthresh >= 0xFFFF) {
782 /* Slow start still did not finish. */
783 if (dst_metric(dst, RTAX_SSTHRESH) &&
784 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
785 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
786 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
787 if (!dst_metric_locked(dst, RTAX_CWND) &&
788 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
789 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
790 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
791 icsk->icsk_ca_state == TCP_CA_Open) {
792 /* Cong. avoidance phase, cwnd is reliable. */
793 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
794 dst->metrics[RTAX_SSTHRESH-1] =
795 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
796 if (!dst_metric_locked(dst, RTAX_CWND))
797 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
799 /* Else slow start did not finish, cwnd is non-sense,
800 ssthresh may be also invalid.
802 if (!dst_metric_locked(dst, RTAX_CWND))
803 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
804 if (dst->metrics[RTAX_SSTHRESH-1] &&
805 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
806 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
807 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
810 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
811 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
812 tp->reordering != sysctl_tcp_reordering)
813 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
818 /* Numbers are taken from RFC3390.
820 * John Heffner states:
822 * The RFC specifies a window of no more than 4380 bytes
823 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
824 * is a bit misleading because they use a clamp at 4380 bytes
825 * rather than use a multiplier in the relevant range.
827 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
829 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
832 if (tp->mss_cache > 1460)
835 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
837 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
840 /* Set slow start threshold and cwnd not falling to slow start */
841 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
843 struct tcp_sock *tp = tcp_sk(sk);
844 const struct inet_connection_sock *icsk = inet_csk(sk);
846 tp->prior_ssthresh = 0;
848 if (icsk->icsk_ca_state < TCP_CA_CWR) {
851 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
852 tp->snd_cwnd = min(tp->snd_cwnd,
853 tcp_packets_in_flight(tp) + 1U);
854 tp->snd_cwnd_cnt = 0;
855 tp->high_seq = tp->snd_nxt;
856 tp->snd_cwnd_stamp = tcp_time_stamp;
857 TCP_ECN_queue_cwr(tp);
859 tcp_set_ca_state(sk, TCP_CA_CWR);
863 /* Initialize metrics on socket. */
865 static void tcp_init_metrics(struct sock *sk)
867 struct tcp_sock *tp = tcp_sk(sk);
868 struct dst_entry *dst = __sk_dst_get(sk);
875 if (dst_metric_locked(dst, RTAX_CWND))
876 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
877 if (dst_metric(dst, RTAX_SSTHRESH)) {
878 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
879 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
880 tp->snd_ssthresh = tp->snd_cwnd_clamp;
882 if (dst_metric(dst, RTAX_REORDERING) &&
883 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
884 tp->rx_opt.sack_ok &= ~2;
885 tp->reordering = dst_metric(dst, RTAX_REORDERING);
888 if (dst_metric(dst, RTAX_RTT) == 0)
891 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
894 /* Initial rtt is determined from SYN,SYN-ACK.
895 * The segment is small and rtt may appear much
896 * less than real one. Use per-dst memory
897 * to make it more realistic.
899 * A bit of theory. RTT is time passed after "normal" sized packet
900 * is sent until it is ACKed. In normal circumstances sending small
901 * packets force peer to delay ACKs and calculation is correct too.
902 * The algorithm is adaptive and, provided we follow specs, it
903 * NEVER underestimate RTT. BUT! If peer tries to make some clever
904 * tricks sort of "quick acks" for time long enough to decrease RTT
905 * to low value, and then abruptly stops to do it and starts to delay
906 * ACKs, wait for troubles.
908 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
909 tp->srtt = dst_metric(dst, RTAX_RTT);
910 tp->rtt_seq = tp->snd_nxt;
912 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
913 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
914 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
918 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
920 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
921 tp->snd_cwnd_stamp = tcp_time_stamp;
925 /* Play conservative. If timestamps are not
926 * supported, TCP will fail to recalculate correct
927 * rtt, if initial rto is too small. FORGET ALL AND RESET!
929 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
931 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
932 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
936 static void tcp_update_reordering(struct sock *sk, const int metric,
939 struct tcp_sock *tp = tcp_sk(sk);
940 if (metric > tp->reordering) {
941 tp->reordering = min(TCP_MAX_REORDERING, metric);
943 /* This exciting event is worth to be remembered. 8) */
945 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
947 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
949 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
951 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
952 #if FASTRETRANS_DEBUG > 1
953 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
954 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
958 tp->undo_marker ? tp->undo_retrans : 0);
960 /* Disable FACK yet. */
961 tp->rx_opt.sack_ok &= ~2;
965 /* This procedure tags the retransmission queue when SACKs arrive.
967 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
968 * Packets in queue with these bits set are counted in variables
969 * sacked_out, retrans_out and lost_out, correspondingly.
971 * Valid combinations are:
972 * Tag InFlight Description
973 * 0 1 - orig segment is in flight.
974 * S 0 - nothing flies, orig reached receiver.
975 * L 0 - nothing flies, orig lost by net.
976 * R 2 - both orig and retransmit are in flight.
977 * L|R 1 - orig is lost, retransmit is in flight.
978 * S|R 1 - orig reached receiver, retrans is still in flight.
979 * (L|S|R is logically valid, it could occur when L|R is sacked,
980 * but it is equivalent to plain S and code short-curcuits it to S.
981 * L|S is logically invalid, it would mean -1 packet in flight 8))
983 * These 6 states form finite state machine, controlled by the following events:
984 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
985 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
986 * 3. Loss detection event of one of three flavors:
987 * A. Scoreboard estimator decided the packet is lost.
988 * A'. Reno "three dupacks" marks head of queue lost.
989 * A''. Its FACK modfication, head until snd.fack is lost.
990 * B. SACK arrives sacking data transmitted after never retransmitted
992 * C. SACK arrives sacking SND.NXT at the moment, when the
993 * segment was retransmitted.
994 * 4. D-SACK added new rule: D-SACK changes any tag to S.
996 * It is pleasant to note, that state diagram turns out to be commutative,
997 * so that we are allowed not to be bothered by order of our actions,
998 * when multiple events arrive simultaneously. (see the function below).
1000 * Reordering detection.
1001 * --------------------
1002 * Reordering metric is maximal distance, which a packet can be displaced
1003 * in packet stream. With SACKs we can estimate it:
1005 * 1. SACK fills old hole and the corresponding segment was not
1006 * ever retransmitted -> reordering. Alas, we cannot use it
1007 * when segment was retransmitted.
1008 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1009 * for retransmitted and already SACKed segment -> reordering..
1010 * Both of these heuristics are not used in Loss state, when we cannot
1011 * account for retransmits accurately.
1013 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1014 struct tcp_sack_block_wire *sp, int num_sacks,
1017 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1018 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1021 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1023 tp->rx_opt.sack_ok |= 4;
1024 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1025 } else if (num_sacks > 1) {
1026 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1027 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1029 if (!after(end_seq_0, end_seq_1) &&
1030 !before(start_seq_0, start_seq_1)) {
1032 tp->rx_opt.sack_ok |= 4;
1033 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1037 /* D-SACK for already forgotten data... Do dumb counting. */
1039 !after(end_seq_0, prior_snd_una) &&
1040 after(end_seq_0, tp->undo_marker))
1047 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
1049 const struct inet_connection_sock *icsk = inet_csk(sk);
1050 struct tcp_sock *tp = tcp_sk(sk);
1051 unsigned char *ptr = (skb_transport_header(ack_skb) +
1052 TCP_SKB_CB(ack_skb)->sacked);
1053 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1054 struct sk_buff *cached_skb;
1055 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1056 int reord = tp->packets_out;
1058 u32 lost_retrans = 0;
1060 int found_dup_sack = 0;
1061 int cached_fack_count;
1063 int first_sack_index;
1065 if (!tp->sacked_out) {
1066 tp->fackets_out = 0;
1067 tp->highest_sack = tp->snd_una;
1069 prior_fackets = tp->fackets_out;
1071 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1072 num_sacks, prior_snd_una);
1074 flag |= FLAG_DSACKING_ACK;
1076 /* Eliminate too old ACKs, but take into
1077 * account more or less fresh ones, they can
1078 * contain valid SACK info.
1080 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1084 * if the only SACK change is the increase of the end_seq of
1085 * the first block then only apply that SACK block
1086 * and use retrans queue hinting otherwise slowpath */
1088 for (i = 0; i < num_sacks; i++) {
1089 __be32 start_seq = sp[i].start_seq;
1090 __be32 end_seq = sp[i].end_seq;
1093 if (tp->recv_sack_cache[i].start_seq != start_seq)
1096 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1097 (tp->recv_sack_cache[i].end_seq != end_seq))
1100 tp->recv_sack_cache[i].start_seq = start_seq;
1101 tp->recv_sack_cache[i].end_seq = end_seq;
1103 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1104 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1105 tp->recv_sack_cache[i].start_seq = 0;
1106 tp->recv_sack_cache[i].end_seq = 0;
1109 first_sack_index = 0;
1114 tp->fastpath_skb_hint = NULL;
1116 /* order SACK blocks to allow in order walk of the retrans queue */
1117 for (i = num_sacks-1; i > 0; i--) {
1118 for (j = 0; j < i; j++){
1119 if (after(ntohl(sp[j].start_seq),
1120 ntohl(sp[j+1].start_seq))){
1121 struct tcp_sack_block_wire tmp;
1127 /* Track where the first SACK block goes to */
1128 if (j == first_sack_index)
1129 first_sack_index = j+1;
1136 /* clear flag as used for different purpose in following code */
1139 /* Use SACK fastpath hint if valid */
1140 cached_skb = tp->fastpath_skb_hint;
1141 cached_fack_count = tp->fastpath_cnt_hint;
1143 cached_skb = tcp_write_queue_head(sk);
1144 cached_fack_count = 0;
1147 for (i=0; i<num_sacks; i++, sp++) {
1148 struct sk_buff *skb;
1149 __u32 start_seq = ntohl(sp->start_seq);
1150 __u32 end_seq = ntohl(sp->end_seq);
1152 int dup_sack = (found_dup_sack && (i == first_sack_index));
1155 fack_count = cached_fack_count;
1157 /* Event "B" in the comment above. */
1158 if (after(end_seq, tp->high_seq))
1159 flag |= FLAG_DATA_LOST;
1161 tcp_for_write_queue_from(skb, sk) {
1162 int in_sack, pcount;
1165 if (skb == tcp_send_head(sk))
1169 cached_fack_count = fack_count;
1170 if (i == first_sack_index) {
1171 tp->fastpath_skb_hint = skb;
1172 tp->fastpath_cnt_hint = fack_count;
1175 /* The retransmission queue is always in order, so
1176 * we can short-circuit the walk early.
1178 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1181 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1182 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1184 pcount = tcp_skb_pcount(skb);
1186 if (pcount > 1 && !in_sack &&
1187 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1188 unsigned int pkt_len;
1190 in_sack = !after(start_seq,
1191 TCP_SKB_CB(skb)->seq);
1194 pkt_len = (start_seq -
1195 TCP_SKB_CB(skb)->seq);
1197 pkt_len = (end_seq -
1198 TCP_SKB_CB(skb)->seq);
1199 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1201 pcount = tcp_skb_pcount(skb);
1204 fack_count += pcount;
1206 sacked = TCP_SKB_CB(skb)->sacked;
1208 /* Account D-SACK for retransmitted packet. */
1209 if ((dup_sack && in_sack) &&
1210 (sacked & TCPCB_RETRANS) &&
1211 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1214 /* The frame is ACKed. */
1215 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1216 if (sacked&TCPCB_RETRANS) {
1217 if ((dup_sack && in_sack) &&
1218 (sacked&TCPCB_SACKED_ACKED))
1219 reord = min(fack_count, reord);
1221 /* If it was in a hole, we detected reordering. */
1222 if (fack_count < prior_fackets &&
1223 !(sacked&TCPCB_SACKED_ACKED))
1224 reord = min(fack_count, reord);
1227 /* Nothing to do; acked frame is about to be dropped. */
1231 if ((sacked&TCPCB_SACKED_RETRANS) &&
1232 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1233 (!lost_retrans || after(end_seq, lost_retrans)))
1234 lost_retrans = end_seq;
1239 if (!(sacked&TCPCB_SACKED_ACKED)) {
1240 if (sacked & TCPCB_SACKED_RETRANS) {
1241 /* If the segment is not tagged as lost,
1242 * we do not clear RETRANS, believing
1243 * that retransmission is still in flight.
1245 if (sacked & TCPCB_LOST) {
1246 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1247 tp->lost_out -= tcp_skb_pcount(skb);
1248 tp->retrans_out -= tcp_skb_pcount(skb);
1250 /* clear lost hint */
1251 tp->retransmit_skb_hint = NULL;
1254 /* New sack for not retransmitted frame,
1255 * which was in hole. It is reordering.
1257 if (!(sacked & TCPCB_RETRANS) &&
1258 fack_count < prior_fackets)
1259 reord = min(fack_count, reord);
1261 if (sacked & TCPCB_LOST) {
1262 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1263 tp->lost_out -= tcp_skb_pcount(skb);
1265 /* clear lost hint */
1266 tp->retransmit_skb_hint = NULL;
1268 /* SACK enhanced F-RTO detection.
1269 * Set flag if and only if non-rexmitted
1270 * segments below frto_highmark are
1271 * SACKed (RFC4138; Appendix B).
1272 * Clearing correct due to in-order walk
1274 if (after(end_seq, tp->frto_highmark)) {
1275 flag &= ~FLAG_ONLY_ORIG_SACKED;
1277 if (!(sacked & TCPCB_RETRANS))
1278 flag |= FLAG_ONLY_ORIG_SACKED;
1282 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1283 flag |= FLAG_DATA_SACKED;
1284 tp->sacked_out += tcp_skb_pcount(skb);
1286 if (fack_count > tp->fackets_out)
1287 tp->fackets_out = fack_count;
1289 if (after(TCP_SKB_CB(skb)->seq,
1291 tp->highest_sack = TCP_SKB_CB(skb)->seq;
1293 if (dup_sack && (sacked&TCPCB_RETRANS))
1294 reord = min(fack_count, reord);
1297 /* D-SACK. We can detect redundant retransmission
1298 * in S|R and plain R frames and clear it.
1299 * undo_retrans is decreased above, L|R frames
1300 * are accounted above as well.
1303 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1304 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1305 tp->retrans_out -= tcp_skb_pcount(skb);
1306 tp->retransmit_skb_hint = NULL;
1311 /* Check for lost retransmit. This superb idea is
1312 * borrowed from "ratehalving". Event "C".
1313 * Later note: FACK people cheated me again 8),
1314 * we have to account for reordering! Ugly,
1317 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1318 struct sk_buff *skb;
1320 tcp_for_write_queue(skb, sk) {
1321 if (skb == tcp_send_head(sk))
1323 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1325 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1327 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1328 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1330 !before(lost_retrans,
1331 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1333 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1334 tp->retrans_out -= tcp_skb_pcount(skb);
1336 /* clear lost hint */
1337 tp->retransmit_skb_hint = NULL;
1339 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1340 tp->lost_out += tcp_skb_pcount(skb);
1341 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1342 flag |= FLAG_DATA_SACKED;
1343 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1349 tp->left_out = tp->sacked_out + tp->lost_out;
1351 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1352 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1353 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1355 #if FASTRETRANS_DEBUG > 0
1356 BUG_TRAP((int)tp->sacked_out >= 0);
1357 BUG_TRAP((int)tp->lost_out >= 0);
1358 BUG_TRAP((int)tp->retrans_out >= 0);
1359 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1364 /* F-RTO can only be used if TCP has never retransmitted anything other than
1365 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1367 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1369 struct tcp_sock *tp = tcp_sk(sk);
1372 holes = max(tp->lost_out, 1U);
1373 holes = min(holes, tp->packets_out);
1375 if ((tp->sacked_out + holes) > tp->packets_out) {
1376 tp->sacked_out = tp->packets_out - holes;
1377 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1381 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1383 static void tcp_add_reno_sack(struct sock *sk)
1385 struct tcp_sock *tp = tcp_sk(sk);
1387 tcp_check_reno_reordering(sk, 0);
1388 tcp_sync_left_out(tp);
1391 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1393 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1395 struct tcp_sock *tp = tcp_sk(sk);
1398 /* One ACK acked hole. The rest eat duplicate ACKs. */
1399 if (acked-1 >= tp->sacked_out)
1402 tp->sacked_out -= acked-1;
1404 tcp_check_reno_reordering(sk, acked);
1405 tcp_sync_left_out(tp);
1408 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1411 tp->left_out = tp->lost_out;
1414 int tcp_use_frto(struct sock *sk)
1416 const struct tcp_sock *tp = tcp_sk(sk);
1417 struct sk_buff *skb;
1419 if (!sysctl_tcp_frto)
1425 /* Avoid expensive walking of rexmit queue if possible */
1426 if (tp->retrans_out > 1)
1429 skb = tcp_write_queue_head(sk);
1430 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1431 tcp_for_write_queue_from(skb, sk) {
1432 if (skb == tcp_send_head(sk))
1434 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1436 /* Short-circuit when first non-SACKed skb has been checked */
1437 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1443 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1444 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1445 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1446 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1447 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1448 * bits are handled if the Loss state is really to be entered (in
1449 * tcp_enter_frto_loss).
1451 * Do like tcp_enter_loss() would; when RTO expires the second time it
1453 * "Reduce ssthresh if it has not yet been made inside this window."
1455 void tcp_enter_frto(struct sock *sk)
1457 const struct inet_connection_sock *icsk = inet_csk(sk);
1458 struct tcp_sock *tp = tcp_sk(sk);
1459 struct sk_buff *skb;
1461 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1462 tp->snd_una == tp->high_seq ||
1463 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1464 !icsk->icsk_retransmits)) {
1465 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1466 /* Our state is too optimistic in ssthresh() call because cwnd
1467 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1468 * recovery has not yet completed. Pattern would be this: RTO,
1469 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1471 * RFC4138 should be more specific on what to do, even though
1472 * RTO is quite unlikely to occur after the first Cumulative ACK
1473 * due to back-off and complexity of triggering events ...
1475 if (tp->frto_counter) {
1477 stored_cwnd = tp->snd_cwnd;
1479 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1480 tp->snd_cwnd = stored_cwnd;
1482 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1484 /* ... in theory, cong.control module could do "any tricks" in
1485 * ssthresh(), which means that ca_state, lost bits and lost_out
1486 * counter would have to be faked before the call occurs. We
1487 * consider that too expensive, unlikely and hacky, so modules
1488 * using these in ssthresh() must deal these incompatibility
1489 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1491 tcp_ca_event(sk, CA_EVENT_FRTO);
1494 tp->undo_marker = tp->snd_una;
1495 tp->undo_retrans = 0;
1497 skb = tcp_write_queue_head(sk);
1498 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1499 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1500 tp->retrans_out -= tcp_skb_pcount(skb);
1502 tcp_sync_left_out(tp);
1504 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1505 * The last condition is necessary at least in tp->frto_counter case.
1507 if (IsSackFrto() && (tp->frto_counter ||
1508 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1509 after(tp->high_seq, tp->snd_una)) {
1510 tp->frto_highmark = tp->high_seq;
1512 tp->frto_highmark = tp->snd_nxt;
1514 tcp_set_ca_state(sk, TCP_CA_Disorder);
1515 tp->high_seq = tp->snd_nxt;
1516 tp->frto_counter = 1;
1519 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1520 * which indicates that we should follow the traditional RTO recovery,
1521 * i.e. mark everything lost and do go-back-N retransmission.
1523 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1525 struct tcp_sock *tp = tcp_sk(sk);
1526 struct sk_buff *skb;
1529 tp->retrans_out = 0;
1531 tcp_reset_reno_sack(tp);
1533 tcp_for_write_queue(skb, sk) {
1534 if (skb == tcp_send_head(sk))
1537 * Count the retransmission made on RTO correctly (only when
1538 * waiting for the first ACK and did not get it)...
1540 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1541 /* For some reason this R-bit might get cleared? */
1542 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1543 tp->retrans_out += tcp_skb_pcount(skb);
1544 /* ...enter this if branch just for the first segment */
1545 flag |= FLAG_DATA_ACKED;
1547 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1550 /* Don't lost mark skbs that were fwd transmitted after RTO */
1551 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1552 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1553 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1554 tp->lost_out += tcp_skb_pcount(skb);
1557 tcp_sync_left_out(tp);
1559 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1560 tp->snd_cwnd_cnt = 0;
1561 tp->snd_cwnd_stamp = tcp_time_stamp;
1562 tp->undo_marker = 0;
1563 tp->frto_counter = 0;
1565 tp->reordering = min_t(unsigned int, tp->reordering,
1566 sysctl_tcp_reordering);
1567 tcp_set_ca_state(sk, TCP_CA_Loss);
1568 tp->high_seq = tp->frto_highmark;
1569 TCP_ECN_queue_cwr(tp);
1571 clear_all_retrans_hints(tp);
1574 void tcp_clear_retrans(struct tcp_sock *tp)
1577 tp->retrans_out = 0;
1579 tp->fackets_out = 0;
1583 tp->undo_marker = 0;
1584 tp->undo_retrans = 0;
1587 /* Enter Loss state. If "how" is not zero, forget all SACK information
1588 * and reset tags completely, otherwise preserve SACKs. If receiver
1589 * dropped its ofo queue, we will know this due to reneging detection.
1591 void tcp_enter_loss(struct sock *sk, int how)
1593 const struct inet_connection_sock *icsk = inet_csk(sk);
1594 struct tcp_sock *tp = tcp_sk(sk);
1595 struct sk_buff *skb;
1598 /* Reduce ssthresh if it has not yet been made inside this window. */
1599 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1600 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1601 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1602 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1603 tcp_ca_event(sk, CA_EVENT_LOSS);
1606 tp->snd_cwnd_cnt = 0;
1607 tp->snd_cwnd_stamp = tcp_time_stamp;
1609 tp->bytes_acked = 0;
1610 tcp_clear_retrans(tp);
1612 /* Push undo marker, if it was plain RTO and nothing
1613 * was retransmitted. */
1615 tp->undo_marker = tp->snd_una;
1617 tcp_for_write_queue(skb, sk) {
1618 if (skb == tcp_send_head(sk))
1620 cnt += tcp_skb_pcount(skb);
1621 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1622 tp->undo_marker = 0;
1623 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1624 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1625 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1626 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1627 tp->lost_out += tcp_skb_pcount(skb);
1629 tp->sacked_out += tcp_skb_pcount(skb);
1630 tp->fackets_out = cnt;
1633 tcp_sync_left_out(tp);
1635 tp->reordering = min_t(unsigned int, tp->reordering,
1636 sysctl_tcp_reordering);
1637 tcp_set_ca_state(sk, TCP_CA_Loss);
1638 tp->high_seq = tp->snd_nxt;
1639 TCP_ECN_queue_cwr(tp);
1640 /* Abort FRTO algorithm if one is in progress */
1641 tp->frto_counter = 0;
1643 clear_all_retrans_hints(tp);
1646 static int tcp_check_sack_reneging(struct sock *sk)
1648 struct sk_buff *skb;
1650 /* If ACK arrived pointing to a remembered SACK,
1651 * it means that our remembered SACKs do not reflect
1652 * real state of receiver i.e.
1653 * receiver _host_ is heavily congested (or buggy).
1654 * Do processing similar to RTO timeout.
1656 if ((skb = tcp_write_queue_head(sk)) != NULL &&
1657 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1658 struct inet_connection_sock *icsk = inet_csk(sk);
1659 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1661 tcp_enter_loss(sk, 1);
1662 icsk->icsk_retransmits++;
1663 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1664 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1665 icsk->icsk_rto, TCP_RTO_MAX);
1671 static inline int tcp_fackets_out(struct tcp_sock *tp)
1673 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1676 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1678 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1681 static inline int tcp_head_timedout(struct sock *sk)
1683 struct tcp_sock *tp = tcp_sk(sk);
1685 return tp->packets_out &&
1686 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1689 /* Linux NewReno/SACK/FACK/ECN state machine.
1690 * --------------------------------------
1692 * "Open" Normal state, no dubious events, fast path.
1693 * "Disorder" In all the respects it is "Open",
1694 * but requires a bit more attention. It is entered when
1695 * we see some SACKs or dupacks. It is split of "Open"
1696 * mainly to move some processing from fast path to slow one.
1697 * "CWR" CWND was reduced due to some Congestion Notification event.
1698 * It can be ECN, ICMP source quench, local device congestion.
1699 * "Recovery" CWND was reduced, we are fast-retransmitting.
1700 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1702 * tcp_fastretrans_alert() is entered:
1703 * - each incoming ACK, if state is not "Open"
1704 * - when arrived ACK is unusual, namely:
1709 * Counting packets in flight is pretty simple.
1711 * in_flight = packets_out - left_out + retrans_out
1713 * packets_out is SND.NXT-SND.UNA counted in packets.
1715 * retrans_out is number of retransmitted segments.
1717 * left_out is number of segments left network, but not ACKed yet.
1719 * left_out = sacked_out + lost_out
1721 * sacked_out: Packets, which arrived to receiver out of order
1722 * and hence not ACKed. With SACKs this number is simply
1723 * amount of SACKed data. Even without SACKs
1724 * it is easy to give pretty reliable estimate of this number,
1725 * counting duplicate ACKs.
1727 * lost_out: Packets lost by network. TCP has no explicit
1728 * "loss notification" feedback from network (for now).
1729 * It means that this number can be only _guessed_.
1730 * Actually, it is the heuristics to predict lossage that
1731 * distinguishes different algorithms.
1733 * F.e. after RTO, when all the queue is considered as lost,
1734 * lost_out = packets_out and in_flight = retrans_out.
1736 * Essentially, we have now two algorithms counting
1739 * FACK: It is the simplest heuristics. As soon as we decided
1740 * that something is lost, we decide that _all_ not SACKed
1741 * packets until the most forward SACK are lost. I.e.
1742 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1743 * It is absolutely correct estimate, if network does not reorder
1744 * packets. And it loses any connection to reality when reordering
1745 * takes place. We use FACK by default until reordering
1746 * is suspected on the path to this destination.
1748 * NewReno: when Recovery is entered, we assume that one segment
1749 * is lost (classic Reno). While we are in Recovery and
1750 * a partial ACK arrives, we assume that one more packet
1751 * is lost (NewReno). This heuristics are the same in NewReno
1754 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1755 * deflation etc. CWND is real congestion window, never inflated, changes
1756 * only according to classic VJ rules.
1758 * Really tricky (and requiring careful tuning) part of algorithm
1759 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1760 * The first determines the moment _when_ we should reduce CWND and,
1761 * hence, slow down forward transmission. In fact, it determines the moment
1762 * when we decide that hole is caused by loss, rather than by a reorder.
1764 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1765 * holes, caused by lost packets.
1767 * And the most logically complicated part of algorithm is undo
1768 * heuristics. We detect false retransmits due to both too early
1769 * fast retransmit (reordering) and underestimated RTO, analyzing
1770 * timestamps and D-SACKs. When we detect that some segments were
1771 * retransmitted by mistake and CWND reduction was wrong, we undo
1772 * window reduction and abort recovery phase. This logic is hidden
1773 * inside several functions named tcp_try_undo_<something>.
1776 /* This function decides, when we should leave Disordered state
1777 * and enter Recovery phase, reducing congestion window.
1779 * Main question: may we further continue forward transmission
1780 * with the same cwnd?
1782 static int tcp_time_to_recover(struct sock *sk)
1784 struct tcp_sock *tp = tcp_sk(sk);
1787 /* Do not perform any recovery during FRTO algorithm */
1788 if (tp->frto_counter)
1791 /* Trick#1: The loss is proven. */
1795 /* Not-A-Trick#2 : Classic rule... */
1796 if (tcp_fackets_out(tp) > tp->reordering)
1799 /* Trick#3 : when we use RFC2988 timer restart, fast
1800 * retransmit can be triggered by timeout of queue head.
1802 if (tcp_head_timedout(sk))
1805 /* Trick#4: It is still not OK... But will it be useful to delay
1808 packets_out = tp->packets_out;
1809 if (packets_out <= tp->reordering &&
1810 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1811 !tcp_may_send_now(sk)) {
1812 /* We have nothing to send. This connection is limited
1813 * either by receiver window or by application.
1821 /* RFC: This is from the original, I doubt that this is necessary at all:
1822 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1823 * retransmitted past LOST markings in the first place? I'm not fully sure
1824 * about undo and end of connection cases, which can cause R without L?
1826 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
1827 struct sk_buff *skb)
1829 if ((tp->retransmit_skb_hint != NULL) &&
1830 before(TCP_SKB_CB(skb)->seq,
1831 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1832 tp->retransmit_skb_hint = NULL;
1835 /* Mark head of queue up as lost. */
1836 static void tcp_mark_head_lost(struct sock *sk,
1837 int packets, u32 high_seq)
1839 struct tcp_sock *tp = tcp_sk(sk);
1840 struct sk_buff *skb;
1843 BUG_TRAP(packets <= tp->packets_out);
1844 if (tp->lost_skb_hint) {
1845 skb = tp->lost_skb_hint;
1846 cnt = tp->lost_cnt_hint;
1848 skb = tcp_write_queue_head(sk);
1852 tcp_for_write_queue_from(skb, sk) {
1853 if (skb == tcp_send_head(sk))
1855 /* TODO: do this better */
1856 /* this is not the most efficient way to do this... */
1857 tp->lost_skb_hint = skb;
1858 tp->lost_cnt_hint = cnt;
1859 cnt += tcp_skb_pcount(skb);
1860 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1862 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1863 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1864 tp->lost_out += tcp_skb_pcount(skb);
1865 tcp_verify_retransmit_hint(tp, skb);
1868 tcp_sync_left_out(tp);
1871 /* Account newly detected lost packet(s) */
1873 static void tcp_update_scoreboard(struct sock *sk)
1875 struct tcp_sock *tp = tcp_sk(sk);
1878 int lost = tp->fackets_out - tp->reordering;
1881 tcp_mark_head_lost(sk, lost, tp->high_seq);
1883 tcp_mark_head_lost(sk, 1, tp->high_seq);
1886 /* New heuristics: it is possible only after we switched
1887 * to restart timer each time when something is ACKed.
1888 * Hence, we can detect timed out packets during fast
1889 * retransmit without falling to slow start.
1891 if (!IsReno(tp) && tcp_head_timedout(sk)) {
1892 struct sk_buff *skb;
1894 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1895 : tcp_write_queue_head(sk);
1897 tcp_for_write_queue_from(skb, sk) {
1898 if (skb == tcp_send_head(sk))
1900 if (!tcp_skb_timedout(sk, skb))
1903 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1904 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1905 tp->lost_out += tcp_skb_pcount(skb);
1906 tcp_verify_retransmit_hint(tp, skb);
1910 tp->scoreboard_skb_hint = skb;
1912 tcp_sync_left_out(tp);
1916 /* CWND moderation, preventing bursts due to too big ACKs
1917 * in dubious situations.
1919 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1921 tp->snd_cwnd = min(tp->snd_cwnd,
1922 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1923 tp->snd_cwnd_stamp = tcp_time_stamp;
1926 /* Lower bound on congestion window is slow start threshold
1927 * unless congestion avoidance choice decides to overide it.
1929 static inline u32 tcp_cwnd_min(const struct sock *sk)
1931 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1933 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1936 /* Decrease cwnd each second ack. */
1937 static void tcp_cwnd_down(struct sock *sk, int flag)
1939 struct tcp_sock *tp = tcp_sk(sk);
1940 int decr = tp->snd_cwnd_cnt + 1;
1942 if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
1943 (IsReno(tp) && !(flag&FLAG_NOT_DUP))) {
1944 tp->snd_cwnd_cnt = decr&1;
1947 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1948 tp->snd_cwnd -= decr;
1950 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1951 tp->snd_cwnd_stamp = tcp_time_stamp;
1955 /* Nothing was retransmitted or returned timestamp is less
1956 * than timestamp of the first retransmission.
1958 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1960 return !tp->retrans_stamp ||
1961 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1962 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1965 /* Undo procedures. */
1967 #if FASTRETRANS_DEBUG > 1
1968 static void DBGUNDO(struct sock *sk, const char *msg)
1970 struct tcp_sock *tp = tcp_sk(sk);
1971 struct inet_sock *inet = inet_sk(sk);
1973 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1975 NIPQUAD(inet->daddr), ntohs(inet->dport),
1976 tp->snd_cwnd, tp->left_out,
1977 tp->snd_ssthresh, tp->prior_ssthresh,
1981 #define DBGUNDO(x...) do { } while (0)
1984 static void tcp_undo_cwr(struct sock *sk, const int undo)
1986 struct tcp_sock *tp = tcp_sk(sk);
1988 if (tp->prior_ssthresh) {
1989 const struct inet_connection_sock *icsk = inet_csk(sk);
1991 if (icsk->icsk_ca_ops->undo_cwnd)
1992 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1994 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1996 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1997 tp->snd_ssthresh = tp->prior_ssthresh;
1998 TCP_ECN_withdraw_cwr(tp);
2001 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2003 tcp_moderate_cwnd(tp);
2004 tp->snd_cwnd_stamp = tcp_time_stamp;
2006 /* There is something screwy going on with the retrans hints after
2008 clear_all_retrans_hints(tp);
2011 static inline int tcp_may_undo(struct tcp_sock *tp)
2013 return tp->undo_marker &&
2014 (!tp->undo_retrans || tcp_packet_delayed(tp));
2017 /* People celebrate: "We love our President!" */
2018 static int tcp_try_undo_recovery(struct sock *sk)
2020 struct tcp_sock *tp = tcp_sk(sk);
2022 if (tcp_may_undo(tp)) {
2023 /* Happy end! We did not retransmit anything
2024 * or our original transmission succeeded.
2026 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2027 tcp_undo_cwr(sk, 1);
2028 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2029 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2031 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2032 tp->undo_marker = 0;
2034 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
2035 /* Hold old state until something *above* high_seq
2036 * is ACKed. For Reno it is MUST to prevent false
2037 * fast retransmits (RFC2582). SACK TCP is safe. */
2038 tcp_moderate_cwnd(tp);
2041 tcp_set_ca_state(sk, TCP_CA_Open);
2045 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2046 static void tcp_try_undo_dsack(struct sock *sk)
2048 struct tcp_sock *tp = tcp_sk(sk);
2050 if (tp->undo_marker && !tp->undo_retrans) {
2051 DBGUNDO(sk, "D-SACK");
2052 tcp_undo_cwr(sk, 1);
2053 tp->undo_marker = 0;
2054 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2058 /* Undo during fast recovery after partial ACK. */
2060 static int tcp_try_undo_partial(struct sock *sk, int acked)
2062 struct tcp_sock *tp = tcp_sk(sk);
2063 /* Partial ACK arrived. Force Hoe's retransmit. */
2064 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
2066 if (tcp_may_undo(tp)) {
2067 /* Plain luck! Hole if filled with delayed
2068 * packet, rather than with a retransmit.
2070 if (tp->retrans_out == 0)
2071 tp->retrans_stamp = 0;
2073 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2076 tcp_undo_cwr(sk, 0);
2077 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2079 /* So... Do not make Hoe's retransmit yet.
2080 * If the first packet was delayed, the rest
2081 * ones are most probably delayed as well.
2088 /* Undo during loss recovery after partial ACK. */
2089 static int tcp_try_undo_loss(struct sock *sk)
2091 struct tcp_sock *tp = tcp_sk(sk);
2093 if (tcp_may_undo(tp)) {
2094 struct sk_buff *skb;
2095 tcp_for_write_queue(skb, sk) {
2096 if (skb == tcp_send_head(sk))
2098 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2101 clear_all_retrans_hints(tp);
2103 DBGUNDO(sk, "partial loss");
2105 tp->left_out = tp->sacked_out;
2106 tcp_undo_cwr(sk, 1);
2107 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2108 inet_csk(sk)->icsk_retransmits = 0;
2109 tp->undo_marker = 0;
2111 tcp_set_ca_state(sk, TCP_CA_Open);
2117 static inline void tcp_complete_cwr(struct sock *sk)
2119 struct tcp_sock *tp = tcp_sk(sk);
2120 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2121 tp->snd_cwnd_stamp = tcp_time_stamp;
2122 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2125 static void tcp_try_to_open(struct sock *sk, int flag)
2127 struct tcp_sock *tp = tcp_sk(sk);
2129 tcp_sync_left_out(tp);
2131 if (tp->retrans_out == 0)
2132 tp->retrans_stamp = 0;
2135 tcp_enter_cwr(sk, 1);
2137 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2138 int state = TCP_CA_Open;
2140 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2141 state = TCP_CA_Disorder;
2143 if (inet_csk(sk)->icsk_ca_state != state) {
2144 tcp_set_ca_state(sk, state);
2145 tp->high_seq = tp->snd_nxt;
2147 tcp_moderate_cwnd(tp);
2149 tcp_cwnd_down(sk, flag);
2153 static void tcp_mtup_probe_failed(struct sock *sk)
2155 struct inet_connection_sock *icsk = inet_csk(sk);
2157 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2158 icsk->icsk_mtup.probe_size = 0;
2161 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2163 struct tcp_sock *tp = tcp_sk(sk);
2164 struct inet_connection_sock *icsk = inet_csk(sk);
2166 /* FIXME: breaks with very large cwnd */
2167 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2168 tp->snd_cwnd = tp->snd_cwnd *
2169 tcp_mss_to_mtu(sk, tp->mss_cache) /
2170 icsk->icsk_mtup.probe_size;
2171 tp->snd_cwnd_cnt = 0;
2172 tp->snd_cwnd_stamp = tcp_time_stamp;
2173 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2175 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2176 icsk->icsk_mtup.probe_size = 0;
2177 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2181 /* Process an event, which can update packets-in-flight not trivially.
2182 * Main goal of this function is to calculate new estimate for left_out,
2183 * taking into account both packets sitting in receiver's buffer and
2184 * packets lost by network.
2186 * Besides that it does CWND reduction, when packet loss is detected
2187 * and changes state of machine.
2189 * It does _not_ decide what to send, it is made in function
2190 * tcp_xmit_retransmit_queue().
2193 tcp_fastretrans_alert(struct sock *sk, int prior_packets, int flag)
2195 struct inet_connection_sock *icsk = inet_csk(sk);
2196 struct tcp_sock *tp = tcp_sk(sk);
2197 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2198 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2199 (tp->fackets_out > tp->reordering));
2201 /* Some technical things:
2202 * 1. Reno does not count dupacks (sacked_out) automatically. */
2203 if (!tp->packets_out)
2205 /* 2. SACK counts snd_fack in packets inaccurately. */
2206 if (tp->sacked_out == 0)
2207 tp->fackets_out = 0;
2209 /* Now state machine starts.
2210 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2212 tp->prior_ssthresh = 0;
2214 /* B. In all the states check for reneging SACKs. */
2215 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2218 /* C. Process data loss notification, provided it is valid. */
2219 if ((flag&FLAG_DATA_LOST) &&
2220 before(tp->snd_una, tp->high_seq) &&
2221 icsk->icsk_ca_state != TCP_CA_Open &&
2222 tp->fackets_out > tp->reordering) {
2223 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2224 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2227 /* D. Synchronize left_out to current state. */
2228 tcp_sync_left_out(tp);
2230 /* E. Check state exit conditions. State can be terminated
2231 * when high_seq is ACKed. */
2232 if (icsk->icsk_ca_state == TCP_CA_Open) {
2233 BUG_TRAP(tp->retrans_out == 0);
2234 tp->retrans_stamp = 0;
2235 } else if (!before(tp->snd_una, tp->high_seq)) {
2236 switch (icsk->icsk_ca_state) {
2238 icsk->icsk_retransmits = 0;
2239 if (tcp_try_undo_recovery(sk))
2244 /* CWR is to be held something *above* high_seq
2245 * is ACKed for CWR bit to reach receiver. */
2246 if (tp->snd_una != tp->high_seq) {
2247 tcp_complete_cwr(sk);
2248 tcp_set_ca_state(sk, TCP_CA_Open);
2252 case TCP_CA_Disorder:
2253 tcp_try_undo_dsack(sk);
2254 if (!tp->undo_marker ||
2255 /* For SACK case do not Open to allow to undo
2256 * catching for all duplicate ACKs. */
2257 IsReno(tp) || tp->snd_una != tp->high_seq) {
2258 tp->undo_marker = 0;
2259 tcp_set_ca_state(sk, TCP_CA_Open);
2263 case TCP_CA_Recovery:
2265 tcp_reset_reno_sack(tp);
2266 if (tcp_try_undo_recovery(sk))
2268 tcp_complete_cwr(sk);
2273 /* F. Process state. */
2274 switch (icsk->icsk_ca_state) {
2275 case TCP_CA_Recovery:
2276 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2277 if (IsReno(tp) && is_dupack)
2278 tcp_add_reno_sack(sk);
2280 int acked = prior_packets - tp->packets_out;
2282 tcp_remove_reno_sacks(sk, acked);
2283 do_lost = tcp_try_undo_partial(sk, acked);
2287 if (flag&FLAG_DATA_ACKED)
2288 icsk->icsk_retransmits = 0;
2289 if (!tcp_try_undo_loss(sk)) {
2290 tcp_moderate_cwnd(tp);
2291 tcp_xmit_retransmit_queue(sk);
2294 if (icsk->icsk_ca_state != TCP_CA_Open)
2296 /* Loss is undone; fall through to processing in Open state. */
2299 if (flag & FLAG_SND_UNA_ADVANCED)
2300 tcp_reset_reno_sack(tp);
2302 tcp_add_reno_sack(sk);
2305 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2306 tcp_try_undo_dsack(sk);
2308 if (!tcp_time_to_recover(sk)) {
2309 tcp_try_to_open(sk, flag);
2313 /* MTU probe failure: don't reduce cwnd */
2314 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2315 icsk->icsk_mtup.probe_size &&
2316 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2317 tcp_mtup_probe_failed(sk);
2318 /* Restores the reduction we did in tcp_mtup_probe() */
2320 tcp_simple_retransmit(sk);
2324 /* Otherwise enter Recovery state */
2327 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2329 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2331 tp->high_seq = tp->snd_nxt;
2332 tp->prior_ssthresh = 0;
2333 tp->undo_marker = tp->snd_una;
2334 tp->undo_retrans = tp->retrans_out;
2336 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2337 if (!(flag&FLAG_ECE))
2338 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2339 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2340 TCP_ECN_queue_cwr(tp);
2343 tp->bytes_acked = 0;
2344 tp->snd_cwnd_cnt = 0;
2345 tcp_set_ca_state(sk, TCP_CA_Recovery);
2348 if (do_lost || tcp_head_timedout(sk))
2349 tcp_update_scoreboard(sk);
2350 tcp_cwnd_down(sk, flag);
2351 tcp_xmit_retransmit_queue(sk);
2354 /* Read draft-ietf-tcplw-high-performance before mucking
2355 * with this code. (Supersedes RFC1323)
2357 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2359 /* RTTM Rule: A TSecr value received in a segment is used to
2360 * update the averaged RTT measurement only if the segment
2361 * acknowledges some new data, i.e., only if it advances the
2362 * left edge of the send window.
2364 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2365 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2367 * Changed: reset backoff as soon as we see the first valid sample.
2368 * If we do not, we get strongly overestimated rto. With timestamps
2369 * samples are accepted even from very old segments: f.e., when rtt=1
2370 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2371 * answer arrives rto becomes 120 seconds! If at least one of segments
2372 * in window is lost... Voila. --ANK (010210)
2374 struct tcp_sock *tp = tcp_sk(sk);
2375 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2376 tcp_rtt_estimator(sk, seq_rtt);
2378 inet_csk(sk)->icsk_backoff = 0;
2382 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2384 /* We don't have a timestamp. Can only use
2385 * packets that are not retransmitted to determine
2386 * rtt estimates. Also, we must not reset the
2387 * backoff for rto until we get a non-retransmitted
2388 * packet. This allows us to deal with a situation
2389 * where the network delay has increased suddenly.
2390 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2393 if (flag & FLAG_RETRANS_DATA_ACKED)
2396 tcp_rtt_estimator(sk, seq_rtt);
2398 inet_csk(sk)->icsk_backoff = 0;
2402 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2405 const struct tcp_sock *tp = tcp_sk(sk);
2406 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2407 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2408 tcp_ack_saw_tstamp(sk, flag);
2409 else if (seq_rtt >= 0)
2410 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2413 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2414 u32 in_flight, int good)
2416 const struct inet_connection_sock *icsk = inet_csk(sk);
2417 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2418 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2421 /* Restart timer after forward progress on connection.
2422 * RFC2988 recommends to restart timer to now+rto.
2425 static void tcp_ack_packets_out(struct sock *sk)
2427 struct tcp_sock *tp = tcp_sk(sk);
2429 if (!tp->packets_out) {
2430 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2432 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2436 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2437 __u32 now, __s32 *seq_rtt)
2439 struct tcp_sock *tp = tcp_sk(sk);
2440 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2441 __u32 seq = tp->snd_una;
2442 __u32 packets_acked;
2445 /* If we get here, the whole TSO packet has not been
2448 BUG_ON(!after(scb->end_seq, seq));
2450 packets_acked = tcp_skb_pcount(skb);
2451 if (tcp_trim_head(sk, skb, seq - scb->seq))
2453 packets_acked -= tcp_skb_pcount(skb);
2455 if (packets_acked) {
2456 __u8 sacked = scb->sacked;
2458 acked |= FLAG_DATA_ACKED;
2460 if (sacked & TCPCB_RETRANS) {
2461 if (sacked & TCPCB_SACKED_RETRANS)
2462 tp->retrans_out -= packets_acked;
2463 acked |= FLAG_RETRANS_DATA_ACKED;
2465 } else if (*seq_rtt < 0)
2466 *seq_rtt = now - scb->when;
2467 if (sacked & TCPCB_SACKED_ACKED)
2468 tp->sacked_out -= packets_acked;
2469 if (sacked & TCPCB_LOST)
2470 tp->lost_out -= packets_acked;
2471 if (sacked & TCPCB_URG) {
2473 !before(seq, tp->snd_up))
2476 } else if (*seq_rtt < 0)
2477 *seq_rtt = now - scb->when;
2479 if (tp->fackets_out) {
2480 __u32 dval = min(tp->fackets_out, packets_acked);
2481 tp->fackets_out -= dval;
2483 /* hint's skb might be NULL but we don't need to care */
2484 tp->fastpath_cnt_hint -= min_t(u32, packets_acked,
2485 tp->fastpath_cnt_hint);
2486 tp->packets_out -= packets_acked;
2488 BUG_ON(tcp_skb_pcount(skb) == 0);
2489 BUG_ON(!before(scb->seq, scb->end_seq));
2495 /* Remove acknowledged frames from the retransmission queue. */
2496 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2498 struct tcp_sock *tp = tcp_sk(sk);
2499 const struct inet_connection_sock *icsk = inet_csk(sk);
2500 struct sk_buff *skb;
2501 __u32 now = tcp_time_stamp;
2503 int prior_packets = tp->packets_out;
2505 ktime_t last_ackt = net_invalid_timestamp();
2507 while ((skb = tcp_write_queue_head(sk)) &&
2508 skb != tcp_send_head(sk)) {
2509 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2510 __u8 sacked = scb->sacked;
2512 /* If our packet is before the ack sequence we can
2513 * discard it as it's confirmed to have arrived at
2516 if (after(scb->end_seq, tp->snd_una)) {
2517 if (tcp_skb_pcount(skb) > 1 &&
2518 after(tp->snd_una, scb->seq))
2519 acked |= tcp_tso_acked(sk, skb,
2524 /* Initial outgoing SYN's get put onto the write_queue
2525 * just like anything else we transmit. It is not
2526 * true data, and if we misinform our callers that
2527 * this ACK acks real data, we will erroneously exit
2528 * connection startup slow start one packet too
2529 * quickly. This is severely frowned upon behavior.
2531 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2532 acked |= FLAG_DATA_ACKED;
2534 acked |= FLAG_SYN_ACKED;
2535 tp->retrans_stamp = 0;
2538 /* MTU probing checks */
2539 if (icsk->icsk_mtup.probe_size) {
2540 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2541 tcp_mtup_probe_success(sk, skb);
2546 if (sacked & TCPCB_RETRANS) {
2547 if (sacked & TCPCB_SACKED_RETRANS)
2548 tp->retrans_out -= tcp_skb_pcount(skb);
2549 acked |= FLAG_RETRANS_DATA_ACKED;
2551 } else if (seq_rtt < 0) {
2552 seq_rtt = now - scb->when;
2553 last_ackt = skb->tstamp;
2555 if (sacked & TCPCB_SACKED_ACKED)
2556 tp->sacked_out -= tcp_skb_pcount(skb);
2557 if (sacked & TCPCB_LOST)
2558 tp->lost_out -= tcp_skb_pcount(skb);
2559 if (sacked & TCPCB_URG) {
2561 !before(scb->end_seq, tp->snd_up))
2564 } else if (seq_rtt < 0) {
2565 seq_rtt = now - scb->when;
2566 last_ackt = skb->tstamp;
2568 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2569 tcp_packets_out_dec(tp, skb);
2570 tcp_unlink_write_queue(skb, sk);
2571 sk_stream_free_skb(sk, skb);
2572 clear_all_retrans_hints(tp);
2575 if (acked&FLAG_ACKED) {
2576 u32 pkts_acked = prior_packets - tp->packets_out;
2577 const struct tcp_congestion_ops *ca_ops
2578 = inet_csk(sk)->icsk_ca_ops;
2580 tcp_ack_update_rtt(sk, acked, seq_rtt);
2581 tcp_ack_packets_out(sk);
2583 if (ca_ops->pkts_acked) {
2586 /* Is the ACK triggering packet unambiguous? */
2587 if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
2588 /* High resolution needed and available? */
2589 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2590 !ktime_equal(last_ackt,
2591 net_invalid_timestamp()))
2592 rtt_us = ktime_us_delta(ktime_get_real(),
2594 else if (seq_rtt > 0)
2595 rtt_us = jiffies_to_usecs(seq_rtt);
2598 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2602 #if FASTRETRANS_DEBUG > 0
2603 BUG_TRAP((int)tp->sacked_out >= 0);
2604 BUG_TRAP((int)tp->lost_out >= 0);
2605 BUG_TRAP((int)tp->retrans_out >= 0);
2606 if (!tp->packets_out && tp->rx_opt.sack_ok) {
2607 const struct inet_connection_sock *icsk = inet_csk(sk);
2609 printk(KERN_DEBUG "Leak l=%u %d\n",
2610 tp->lost_out, icsk->icsk_ca_state);
2613 if (tp->sacked_out) {
2614 printk(KERN_DEBUG "Leak s=%u %d\n",
2615 tp->sacked_out, icsk->icsk_ca_state);
2618 if (tp->retrans_out) {
2619 printk(KERN_DEBUG "Leak r=%u %d\n",
2620 tp->retrans_out, icsk->icsk_ca_state);
2621 tp->retrans_out = 0;
2625 *seq_rtt_p = seq_rtt;
2629 static void tcp_ack_probe(struct sock *sk)
2631 const struct tcp_sock *tp = tcp_sk(sk);
2632 struct inet_connection_sock *icsk = inet_csk(sk);
2634 /* Was it a usable window open? */
2636 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2637 tp->snd_una + tp->snd_wnd)) {
2638 icsk->icsk_backoff = 0;
2639 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2640 /* Socket must be waked up by subsequent tcp_data_snd_check().
2641 * This function is not for random using!
2644 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2645 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2650 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2652 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2653 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2656 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2658 const struct tcp_sock *tp = tcp_sk(sk);
2659 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2660 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2663 /* Check that window update is acceptable.
2664 * The function assumes that snd_una<=ack<=snd_next.
2666 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2667 const u32 ack_seq, const u32 nwin)
2669 return (after(ack, tp->snd_una) ||
2670 after(ack_seq, tp->snd_wl1) ||
2671 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2674 /* Update our send window.
2676 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2677 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2679 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2682 struct tcp_sock *tp = tcp_sk(sk);
2684 u32 nwin = ntohs(tcp_hdr(skb)->window);
2686 if (likely(!tcp_hdr(skb)->syn))
2687 nwin <<= tp->rx_opt.snd_wscale;
2689 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2690 flag |= FLAG_WIN_UPDATE;
2691 tcp_update_wl(tp, ack, ack_seq);
2693 if (tp->snd_wnd != nwin) {
2696 /* Note, it is the only place, where
2697 * fast path is recovered for sending TCP.
2700 tcp_fast_path_check(sk);
2702 if (nwin > tp->max_window) {
2703 tp->max_window = nwin;
2704 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2714 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2715 * continue in congestion avoidance.
2717 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2719 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2720 tp->snd_cwnd_cnt = 0;
2721 TCP_ECN_queue_cwr(tp);
2722 tcp_moderate_cwnd(tp);
2725 /* A conservative spurious RTO response algorithm: reduce cwnd using
2726 * rate halving and continue in congestion avoidance.
2728 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2730 tcp_enter_cwr(sk, 0);
2733 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2736 tcp_ratehalving_spur_to_response(sk);
2738 tcp_undo_cwr(sk, 1);
2741 /* F-RTO spurious RTO detection algorithm (RFC4138)
2743 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2744 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2745 * window (but not to or beyond highest sequence sent before RTO):
2746 * On First ACK, send two new segments out.
2747 * On Second ACK, RTO was likely spurious. Do spurious response (response
2748 * algorithm is not part of the F-RTO detection algorithm
2749 * given in RFC4138 but can be selected separately).
2750 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2751 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2752 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2753 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2755 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2756 * original window even after we transmit two new data segments.
2759 * on first step, wait until first cumulative ACK arrives, then move to
2760 * the second step. In second step, the next ACK decides.
2762 * F-RTO is implemented (mainly) in four functions:
2763 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2764 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2765 * called when tcp_use_frto() showed green light
2766 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2767 * - tcp_enter_frto_loss() is called if there is not enough evidence
2768 * to prove that the RTO is indeed spurious. It transfers the control
2769 * from F-RTO to the conventional RTO recovery
2771 static int tcp_process_frto(struct sock *sk, int flag)
2773 struct tcp_sock *tp = tcp_sk(sk);
2775 tcp_sync_left_out(tp);
2777 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2778 if (flag&FLAG_DATA_ACKED)
2779 inet_csk(sk)->icsk_retransmits = 0;
2781 if (!before(tp->snd_una, tp->frto_highmark)) {
2782 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2786 if (!IsSackFrto() || IsReno(tp)) {
2787 /* RFC4138 shortcoming in step 2; should also have case c):
2788 * ACK isn't duplicate nor advances window, e.g., opposite dir
2791 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2794 if (!(flag&FLAG_DATA_ACKED)) {
2795 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2800 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2801 /* Prevent sending of new data. */
2802 tp->snd_cwnd = min(tp->snd_cwnd,
2803 tcp_packets_in_flight(tp));
2807 if ((tp->frto_counter >= 2) &&
2808 (!(flag&FLAG_FORWARD_PROGRESS) ||
2809 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2810 /* RFC4138 shortcoming (see comment above) */
2811 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2814 tcp_enter_frto_loss(sk, 3, flag);
2819 if (tp->frto_counter == 1) {
2820 /* Sending of the next skb must be allowed or no FRTO */
2821 if (!tcp_send_head(sk) ||
2822 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2823 tp->snd_una + tp->snd_wnd)) {
2824 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2829 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2830 tp->frto_counter = 2;
2833 switch (sysctl_tcp_frto_response) {
2835 tcp_undo_spur_to_response(sk, flag);
2838 tcp_conservative_spur_to_response(tp);
2841 tcp_ratehalving_spur_to_response(sk);
2844 tp->frto_counter = 0;
2849 /* This routine deals with incoming acks, but not outgoing ones. */
2850 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2852 struct inet_connection_sock *icsk = inet_csk(sk);
2853 struct tcp_sock *tp = tcp_sk(sk);
2854 u32 prior_snd_una = tp->snd_una;
2855 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2856 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2857 u32 prior_in_flight;
2862 /* If the ack is newer than sent or older than previous acks
2863 * then we can probably ignore it.
2865 if (after(ack, tp->snd_nxt))
2866 goto uninteresting_ack;
2868 if (before(ack, prior_snd_una))
2871 if (after(ack, prior_snd_una))
2872 flag |= FLAG_SND_UNA_ADVANCED;
2874 if (sysctl_tcp_abc) {
2875 if (icsk->icsk_ca_state < TCP_CA_CWR)
2876 tp->bytes_acked += ack - prior_snd_una;
2877 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2878 /* we assume just one segment left network */
2879 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2882 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2883 /* Window is constant, pure forward advance.
2884 * No more checks are required.
2885 * Note, we use the fact that SND.UNA>=SND.WL2.
2887 tcp_update_wl(tp, ack, ack_seq);
2889 flag |= FLAG_WIN_UPDATE;
2891 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2893 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2895 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2898 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2900 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2902 if (TCP_SKB_CB(skb)->sacked)
2903 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2905 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2908 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2911 /* We passed data and got it acked, remove any soft error
2912 * log. Something worked...
2914 sk->sk_err_soft = 0;
2915 tp->rcv_tstamp = tcp_time_stamp;
2916 prior_packets = tp->packets_out;
2920 prior_in_flight = tcp_packets_in_flight(tp);
2922 /* See if we can take anything off of the retransmit queue. */
2923 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2925 if (tp->frto_counter)
2926 frto_cwnd = tcp_process_frto(sk, flag);
2928 if (tcp_ack_is_dubious(sk, flag)) {
2929 /* Advance CWND, if state allows this. */
2930 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2931 tcp_may_raise_cwnd(sk, flag))
2932 tcp_cong_avoid(sk, ack, prior_in_flight, 0);
2933 tcp_fastretrans_alert(sk, prior_packets, flag);
2935 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2936 tcp_cong_avoid(sk, ack, prior_in_flight, 1);
2939 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2940 dst_confirm(sk->sk_dst_cache);
2945 icsk->icsk_probes_out = 0;
2947 /* If this ack opens up a zero window, clear backoff. It was
2948 * being used to time the probes, and is probably far higher than
2949 * it needs to be for normal retransmission.
2951 if (tcp_send_head(sk))
2956 if (TCP_SKB_CB(skb)->sacked)
2957 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2960 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2965 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2966 * But, this can also be called on packets in the established flow when
2967 * the fast version below fails.
2969 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2972 struct tcphdr *th = tcp_hdr(skb);
2973 int length=(th->doff*4)-sizeof(struct tcphdr);
2975 ptr = (unsigned char *)(th + 1);
2976 opt_rx->saw_tstamp = 0;
2978 while (length > 0) {
2985 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2990 if (opsize < 2) /* "silly options" */
2992 if (opsize > length)
2993 return; /* don't parse partial options */
2996 if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2997 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2999 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3000 in_mss = opt_rx->user_mss;
3001 opt_rx->mss_clamp = in_mss;
3006 if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
3007 if (sysctl_tcp_window_scaling) {
3008 __u8 snd_wscale = *(__u8 *) ptr;
3009 opt_rx->wscale_ok = 1;
3010 if (snd_wscale > 14) {
3011 if (net_ratelimit())
3012 printk(KERN_INFO "tcp_parse_options: Illegal window "
3013 "scaling value %d >14 received.\n",
3017 opt_rx->snd_wscale = snd_wscale;
3020 case TCPOPT_TIMESTAMP:
3021 if (opsize==TCPOLEN_TIMESTAMP) {
3022 if ((estab && opt_rx->tstamp_ok) ||
3023 (!estab && sysctl_tcp_timestamps)) {
3024 opt_rx->saw_tstamp = 1;
3025 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3026 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3030 case TCPOPT_SACK_PERM:
3031 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3032 if (sysctl_tcp_sack) {
3033 opt_rx->sack_ok = 1;
3034 tcp_sack_reset(opt_rx);
3040 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3041 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3043 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3046 #ifdef CONFIG_TCP_MD5SIG
3049 * The MD5 Hash has already been
3050 * checked (see tcp_v{4,6}_do_rcv()).
3062 /* Fast parse options. This hopes to only see timestamps.
3063 * If it is wrong it falls back on tcp_parse_options().
3065 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3066 struct tcp_sock *tp)
3068 if (th->doff == sizeof(struct tcphdr)>>2) {
3069 tp->rx_opt.saw_tstamp = 0;
3071 } else if (tp->rx_opt.tstamp_ok &&
3072 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3073 __be32 *ptr = (__be32 *)(th + 1);
3074 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3075 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3076 tp->rx_opt.saw_tstamp = 1;
3078 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3080 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3084 tcp_parse_options(skb, &tp->rx_opt, 1);
3088 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3090 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3091 tp->rx_opt.ts_recent_stamp = get_seconds();
3094 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3096 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3097 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3098 * extra check below makes sure this can only happen
3099 * for pure ACK frames. -DaveM
3101 * Not only, also it occurs for expired timestamps.
3104 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3105 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3106 tcp_store_ts_recent(tp);
3110 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3112 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3113 * it can pass through stack. So, the following predicate verifies that
3114 * this segment is not used for anything but congestion avoidance or
3115 * fast retransmit. Moreover, we even are able to eliminate most of such
3116 * second order effects, if we apply some small "replay" window (~RTO)
3117 * to timestamp space.
3119 * All these measures still do not guarantee that we reject wrapped ACKs
3120 * on networks with high bandwidth, when sequence space is recycled fastly,
3121 * but it guarantees that such events will be very rare and do not affect
3122 * connection seriously. This doesn't look nice, but alas, PAWS is really
3125 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3126 * states that events when retransmit arrives after original data are rare.
3127 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3128 * the biggest problem on large power networks even with minor reordering.
3129 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3130 * up to bandwidth of 18Gigabit/sec. 8) ]
3133 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3135 struct tcp_sock *tp = tcp_sk(sk);
3136 struct tcphdr *th = tcp_hdr(skb);
3137 u32 seq = TCP_SKB_CB(skb)->seq;
3138 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3140 return (/* 1. Pure ACK with correct sequence number. */
3141 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3143 /* 2. ... and duplicate ACK. */
3144 ack == tp->snd_una &&
3146 /* 3. ... and does not update window. */
3147 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3149 /* 4. ... and sits in replay window. */
3150 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3153 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3155 const struct tcp_sock *tp = tcp_sk(sk);
3156 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3157 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3158 !tcp_disordered_ack(sk, skb));
3161 /* Check segment sequence number for validity.
3163 * Segment controls are considered valid, if the segment
3164 * fits to the window after truncation to the window. Acceptability
3165 * of data (and SYN, FIN, of course) is checked separately.
3166 * See tcp_data_queue(), for example.
3168 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3169 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3170 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3171 * (borrowed from freebsd)
3174 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3176 return !before(end_seq, tp->rcv_wup) &&
3177 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3180 /* When we get a reset we do this. */
3181 static void tcp_reset(struct sock *sk)
3183 /* We want the right error as BSD sees it (and indeed as we do). */
3184 switch (sk->sk_state) {
3186 sk->sk_err = ECONNREFUSED;
3188 case TCP_CLOSE_WAIT:
3194 sk->sk_err = ECONNRESET;
3197 if (!sock_flag(sk, SOCK_DEAD))
3198 sk->sk_error_report(sk);
3204 * Process the FIN bit. This now behaves as it is supposed to work
3205 * and the FIN takes effect when it is validly part of sequence
3206 * space. Not before when we get holes.
3208 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3209 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3212 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3213 * close and we go into CLOSING (and later onto TIME-WAIT)
3215 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3217 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3219 struct tcp_sock *tp = tcp_sk(sk);
3221 inet_csk_schedule_ack(sk);
3223 sk->sk_shutdown |= RCV_SHUTDOWN;
3224 sock_set_flag(sk, SOCK_DONE);
3226 switch (sk->sk_state) {
3228 case TCP_ESTABLISHED:
3229 /* Move to CLOSE_WAIT */
3230 tcp_set_state(sk, TCP_CLOSE_WAIT);
3231 inet_csk(sk)->icsk_ack.pingpong = 1;
3234 case TCP_CLOSE_WAIT:
3236 /* Received a retransmission of the FIN, do
3241 /* RFC793: Remain in the LAST-ACK state. */
3245 /* This case occurs when a simultaneous close
3246 * happens, we must ack the received FIN and
3247 * enter the CLOSING state.
3250 tcp_set_state(sk, TCP_CLOSING);
3253 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3255 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3258 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3259 * cases we should never reach this piece of code.
3261 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3262 __FUNCTION__, sk->sk_state);
3266 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3267 * Probably, we should reset in this case. For now drop them.
3269 __skb_queue_purge(&tp->out_of_order_queue);
3270 if (tp->rx_opt.sack_ok)
3271 tcp_sack_reset(&tp->rx_opt);
3272 sk_stream_mem_reclaim(sk);
3274 if (!sock_flag(sk, SOCK_DEAD)) {
3275 sk->sk_state_change(sk);
3277 /* Do not send POLL_HUP for half duplex close. */
3278 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3279 sk->sk_state == TCP_CLOSE)
3280 sk_wake_async(sk, 1, POLL_HUP);
3282 sk_wake_async(sk, 1, POLL_IN);
3286 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3288 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3289 if (before(seq, sp->start_seq))
3290 sp->start_seq = seq;
3291 if (after(end_seq, sp->end_seq))
3292 sp->end_seq = end_seq;
3298 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3300 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3301 if (before(seq, tp->rcv_nxt))
3302 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3304 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3306 tp->rx_opt.dsack = 1;
3307 tp->duplicate_sack[0].start_seq = seq;
3308 tp->duplicate_sack[0].end_seq = end_seq;
3309 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3313 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3315 if (!tp->rx_opt.dsack)
3316 tcp_dsack_set(tp, seq, end_seq);
3318 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3321 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3323 struct tcp_sock *tp = tcp_sk(sk);
3325 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3326 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3327 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3328 tcp_enter_quickack_mode(sk);
3330 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3331 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3333 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3334 end_seq = tp->rcv_nxt;
3335 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3342 /* These routines update the SACK block as out-of-order packets arrive or
3343 * in-order packets close up the sequence space.
3345 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3348 struct tcp_sack_block *sp = &tp->selective_acks[0];
3349 struct tcp_sack_block *swalk = sp+1;
3351 /* See if the recent change to the first SACK eats into
3352 * or hits the sequence space of other SACK blocks, if so coalesce.
3354 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3355 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3358 /* Zap SWALK, by moving every further SACK up by one slot.
3359 * Decrease num_sacks.
3361 tp->rx_opt.num_sacks--;
3362 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3363 for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3367 this_sack++, swalk++;
3371 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3375 tmp = sack1->start_seq;
3376 sack1->start_seq = sack2->start_seq;
3377 sack2->start_seq = tmp;
3379 tmp = sack1->end_seq;
3380 sack1->end_seq = sack2->end_seq;
3381 sack2->end_seq = tmp;
3384 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3386 struct tcp_sock *tp = tcp_sk(sk);
3387 struct tcp_sack_block *sp = &tp->selective_acks[0];
3388 int cur_sacks = tp->rx_opt.num_sacks;
3394 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3395 if (tcp_sack_extend(sp, seq, end_seq)) {
3396 /* Rotate this_sack to the first one. */
3397 for (; this_sack>0; this_sack--, sp--)
3398 tcp_sack_swap(sp, sp-1);
3400 tcp_sack_maybe_coalesce(tp);
3405 /* Could not find an adjacent existing SACK, build a new one,
3406 * put it at the front, and shift everyone else down. We
3407 * always know there is at least one SACK present already here.
3409 * If the sack array is full, forget about the last one.
3411 if (this_sack >= 4) {
3413 tp->rx_opt.num_sacks--;
3416 for (; this_sack > 0; this_sack--, sp--)
3420 /* Build the new head SACK, and we're done. */
3421 sp->start_seq = seq;
3422 sp->end_seq = end_seq;
3423 tp->rx_opt.num_sacks++;
3424 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3427 /* RCV.NXT advances, some SACKs should be eaten. */
3429 static void tcp_sack_remove(struct tcp_sock *tp)
3431 struct tcp_sack_block *sp = &tp->selective_acks[0];
3432 int num_sacks = tp->rx_opt.num_sacks;
3435 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3436 if (skb_queue_empty(&tp->out_of_order_queue)) {
3437 tp->rx_opt.num_sacks = 0;
3438 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3442 for (this_sack = 0; this_sack < num_sacks; ) {
3443 /* Check if the start of the sack is covered by RCV.NXT. */
3444 if (!before(tp->rcv_nxt, sp->start_seq)) {
3447 /* RCV.NXT must cover all the block! */
3448 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3450 /* Zap this SACK, by moving forward any other SACKS. */
3451 for (i=this_sack+1; i < num_sacks; i++)
3452 tp->selective_acks[i-1] = tp->selective_acks[i];
3459 if (num_sacks != tp->rx_opt.num_sacks) {
3460 tp->rx_opt.num_sacks = num_sacks;
3461 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3465 /* This one checks to see if we can put data from the
3466 * out_of_order queue into the receive_queue.
3468 static void tcp_ofo_queue(struct sock *sk)
3470 struct tcp_sock *tp = tcp_sk(sk);
3471 __u32 dsack_high = tp->rcv_nxt;
3472 struct sk_buff *skb;
3474 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3475 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3478 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3479 __u32 dsack = dsack_high;
3480 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3481 dsack_high = TCP_SKB_CB(skb)->end_seq;
3482 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3485 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3486 SOCK_DEBUG(sk, "ofo packet was already received \n");
3487 __skb_unlink(skb, &tp->out_of_order_queue);
3491 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3492 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3493 TCP_SKB_CB(skb)->end_seq);
3495 __skb_unlink(skb, &tp->out_of_order_queue);
3496 __skb_queue_tail(&sk->sk_receive_queue, skb);
3497 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3498 if (tcp_hdr(skb)->fin)
3499 tcp_fin(skb, sk, tcp_hdr(skb));
3503 static int tcp_prune_queue(struct sock *sk);
3505 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3507 struct tcphdr *th = tcp_hdr(skb);
3508 struct tcp_sock *tp = tcp_sk(sk);
3511 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3514 __skb_pull(skb, th->doff*4);
3516 TCP_ECN_accept_cwr(tp, skb);
3518 if (tp->rx_opt.dsack) {
3519 tp->rx_opt.dsack = 0;
3520 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3521 4 - tp->rx_opt.tstamp_ok);
3524 /* Queue data for delivery to the user.
3525 * Packets in sequence go to the receive queue.
3526 * Out of sequence packets to the out_of_order_queue.
3528 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3529 if (tcp_receive_window(tp) == 0)
3532 /* Ok. In sequence. In window. */
3533 if (tp->ucopy.task == current &&
3534 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3535 sock_owned_by_user(sk) && !tp->urg_data) {
3536 int chunk = min_t(unsigned int, skb->len,
3539 __set_current_state(TASK_RUNNING);
3542 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3543 tp->ucopy.len -= chunk;
3544 tp->copied_seq += chunk;
3545 eaten = (chunk == skb->len && !th->fin);
3546 tcp_rcv_space_adjust(sk);
3554 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3555 !sk_stream_rmem_schedule(sk, skb))) {
3556 if (tcp_prune_queue(sk) < 0 ||
3557 !sk_stream_rmem_schedule(sk, skb))
3560 sk_stream_set_owner_r(skb, sk);
3561 __skb_queue_tail(&sk->sk_receive_queue, skb);
3563 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3565 tcp_event_data_recv(sk, skb);
3567 tcp_fin(skb, sk, th);
3569 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3572 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3573 * gap in queue is filled.
3575 if (skb_queue_empty(&tp->out_of_order_queue))
3576 inet_csk(sk)->icsk_ack.pingpong = 0;
3579 if (tp->rx_opt.num_sacks)
3580 tcp_sack_remove(tp);
3582 tcp_fast_path_check(sk);
3586 else if (!sock_flag(sk, SOCK_DEAD))
3587 sk->sk_data_ready(sk, 0);
3591 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3592 /* A retransmit, 2nd most common case. Force an immediate ack. */
3593 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3594 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3597 tcp_enter_quickack_mode(sk);
3598 inet_csk_schedule_ack(sk);
3604 /* Out of window. F.e. zero window probe. */
3605 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3608 tcp_enter_quickack_mode(sk);
3610 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3611 /* Partial packet, seq < rcv_next < end_seq */
3612 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3613 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3614 TCP_SKB_CB(skb)->end_seq);
3616 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3618 /* If window is closed, drop tail of packet. But after
3619 * remembering D-SACK for its head made in previous line.
3621 if (!tcp_receive_window(tp))
3626 TCP_ECN_check_ce(tp, skb);
3628 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3629 !sk_stream_rmem_schedule(sk, skb)) {
3630 if (tcp_prune_queue(sk) < 0 ||
3631 !sk_stream_rmem_schedule(sk, skb))
3635 /* Disable header prediction. */
3637 inet_csk_schedule_ack(sk);
3639 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3640 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3642 sk_stream_set_owner_r(skb, sk);
3644 if (!skb_peek(&tp->out_of_order_queue)) {
3645 /* Initial out of order segment, build 1 SACK. */
3646 if (tp->rx_opt.sack_ok) {
3647 tp->rx_opt.num_sacks = 1;
3648 tp->rx_opt.dsack = 0;
3649 tp->rx_opt.eff_sacks = 1;
3650 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3651 tp->selective_acks[0].end_seq =
3652 TCP_SKB_CB(skb)->end_seq;
3654 __skb_queue_head(&tp->out_of_order_queue,skb);
3656 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3657 u32 seq = TCP_SKB_CB(skb)->seq;
3658 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3660 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3661 __skb_append(skb1, skb, &tp->out_of_order_queue);
3663 if (!tp->rx_opt.num_sacks ||
3664 tp->selective_acks[0].end_seq != seq)
3667 /* Common case: data arrive in order after hole. */
3668 tp->selective_acks[0].end_seq = end_seq;
3672 /* Find place to insert this segment. */
3674 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3676 } while ((skb1 = skb1->prev) !=
3677 (struct sk_buff*)&tp->out_of_order_queue);
3679 /* Do skb overlap to previous one? */
3680 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3681 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3682 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3683 /* All the bits are present. Drop. */
3685 tcp_dsack_set(tp, seq, end_seq);
3688 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3689 /* Partial overlap. */
3690 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3695 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3697 /* And clean segments covered by new one as whole. */
3698 while ((skb1 = skb->next) !=
3699 (struct sk_buff*)&tp->out_of_order_queue &&
3700 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3701 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3702 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3705 __skb_unlink(skb1, &tp->out_of_order_queue);
3706 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3711 if (tp->rx_opt.sack_ok)
3712 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3716 /* Collapse contiguous sequence of skbs head..tail with
3717 * sequence numbers start..end.
3718 * Segments with FIN/SYN are not collapsed (only because this
3722 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3723 struct sk_buff *head, struct sk_buff *tail,
3726 struct sk_buff *skb;
3728 /* First, check that queue is collapsible and find
3729 * the point where collapsing can be useful. */
3730 for (skb = head; skb != tail; ) {
3731 /* No new bits? It is possible on ofo queue. */
3732 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3733 struct sk_buff *next = skb->next;
3734 __skb_unlink(skb, list);
3736 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3741 /* The first skb to collapse is:
3743 * - bloated or contains data before "start" or
3744 * overlaps to the next one.
3746 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3747 (tcp_win_from_space(skb->truesize) > skb->len ||
3748 before(TCP_SKB_CB(skb)->seq, start) ||
3749 (skb->next != tail &&
3750 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3753 /* Decided to skip this, advance start seq. */
3754 start = TCP_SKB_CB(skb)->end_seq;
3757 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3760 while (before(start, end)) {
3761 struct sk_buff *nskb;
3762 int header = skb_headroom(skb);
3763 int copy = SKB_MAX_ORDER(header, 0);
3765 /* Too big header? This can happen with IPv6. */
3768 if (end-start < copy)
3770 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3774 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3775 skb_set_network_header(nskb, (skb_network_header(skb) -
3777 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3779 skb_reserve(nskb, header);
3780 memcpy(nskb->head, skb->head, header);
3781 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3782 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3783 __skb_insert(nskb, skb->prev, skb, list);
3784 sk_stream_set_owner_r(nskb, sk);
3786 /* Copy data, releasing collapsed skbs. */
3788 int offset = start - TCP_SKB_CB(skb)->seq;
3789 int size = TCP_SKB_CB(skb)->end_seq - start;
3793 size = min(copy, size);
3794 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3796 TCP_SKB_CB(nskb)->end_seq += size;
3800 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3801 struct sk_buff *next = skb->next;
3802 __skb_unlink(skb, list);
3804 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3807 tcp_hdr(skb)->syn ||
3815 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3816 * and tcp_collapse() them until all the queue is collapsed.
3818 static void tcp_collapse_ofo_queue(struct sock *sk)
3820 struct tcp_sock *tp = tcp_sk(sk);
3821 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3822 struct sk_buff *head;
3828 start = TCP_SKB_CB(skb)->seq;
3829 end = TCP_SKB_CB(skb)->end_seq;
3835 /* Segment is terminated when we see gap or when
3836 * we are at the end of all the queue. */
3837 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3838 after(TCP_SKB_CB(skb)->seq, end) ||
3839 before(TCP_SKB_CB(skb)->end_seq, start)) {
3840 tcp_collapse(sk, &tp->out_of_order_queue,
3841 head, skb, start, end);
3843 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3845 /* Start new segment */
3846 start = TCP_SKB_CB(skb)->seq;
3847 end = TCP_SKB_CB(skb)->end_seq;
3849 if (before(TCP_SKB_CB(skb)->seq, start))
3850 start = TCP_SKB_CB(skb)->seq;
3851 if (after(TCP_SKB_CB(skb)->end_seq, end))
3852 end = TCP_SKB_CB(skb)->end_seq;
3857 /* Reduce allocated memory if we can, trying to get
3858 * the socket within its memory limits again.
3860 * Return less than zero if we should start dropping frames
3861 * until the socket owning process reads some of the data
3862 * to stabilize the situation.
3864 static int tcp_prune_queue(struct sock *sk)
3866 struct tcp_sock *tp = tcp_sk(sk);
3868 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3870 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3872 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3873 tcp_clamp_window(sk);
3874 else if (tcp_memory_pressure)
3875 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3877 tcp_collapse_ofo_queue(sk);
3878 tcp_collapse(sk, &sk->sk_receive_queue,
3879 sk->sk_receive_queue.next,
3880 (struct sk_buff*)&sk->sk_receive_queue,
3881 tp->copied_seq, tp->rcv_nxt);
3882 sk_stream_mem_reclaim(sk);
3884 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3887 /* Collapsing did not help, destructive actions follow.
3888 * This must not ever occur. */
3890 /* First, purge the out_of_order queue. */
3891 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3892 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3893 __skb_queue_purge(&tp->out_of_order_queue);
3895 /* Reset SACK state. A conforming SACK implementation will
3896 * do the same at a timeout based retransmit. When a connection
3897 * is in a sad state like this, we care only about integrity
3898 * of the connection not performance.
3900 if (tp->rx_opt.sack_ok)
3901 tcp_sack_reset(&tp->rx_opt);
3902 sk_stream_mem_reclaim(sk);
3905 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3908 /* If we are really being abused, tell the caller to silently
3909 * drop receive data on the floor. It will get retransmitted
3910 * and hopefully then we'll have sufficient space.
3912 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3914 /* Massive buffer overcommit. */
3920 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3921 * As additional protections, we do not touch cwnd in retransmission phases,
3922 * and if application hit its sndbuf limit recently.
3924 void tcp_cwnd_application_limited(struct sock *sk)
3926 struct tcp_sock *tp = tcp_sk(sk);
3928 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3929 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3930 /* Limited by application or receiver window. */
3931 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3932 u32 win_used = max(tp->snd_cwnd_used, init_win);
3933 if (win_used < tp->snd_cwnd) {
3934 tp->snd_ssthresh = tcp_current_ssthresh(sk);
3935 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3937 tp->snd_cwnd_used = 0;
3939 tp->snd_cwnd_stamp = tcp_time_stamp;
3942 static int tcp_should_expand_sndbuf(struct sock *sk)
3944 struct tcp_sock *tp = tcp_sk(sk);
3946 /* If the user specified a specific send buffer setting, do
3949 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3952 /* If we are under global TCP memory pressure, do not expand. */
3953 if (tcp_memory_pressure)
3956 /* If we are under soft global TCP memory pressure, do not expand. */
3957 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3960 /* If we filled the congestion window, do not expand. */
3961 if (tp->packets_out >= tp->snd_cwnd)
3967 /* When incoming ACK allowed to free some skb from write_queue,
3968 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3969 * on the exit from tcp input handler.
3971 * PROBLEM: sndbuf expansion does not work well with largesend.
3973 static void tcp_new_space(struct sock *sk)
3975 struct tcp_sock *tp = tcp_sk(sk);
3977 if (tcp_should_expand_sndbuf(sk)) {
3978 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3979 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3980 demanded = max_t(unsigned int, tp->snd_cwnd,
3981 tp->reordering + 1);
3982 sndmem *= 2*demanded;
3983 if (sndmem > sk->sk_sndbuf)
3984 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3985 tp->snd_cwnd_stamp = tcp_time_stamp;
3988 sk->sk_write_space(sk);
3991 static void tcp_check_space(struct sock *sk)
3993 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3994 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3995 if (sk->sk_socket &&
3996 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4001 static inline void tcp_data_snd_check(struct sock *sk)
4003 tcp_push_pending_frames(sk);
4004 tcp_check_space(sk);
4008 * Check if sending an ack is needed.
4010 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4012 struct tcp_sock *tp = tcp_sk(sk);
4014 /* More than one full frame received... */
4015 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4016 /* ... and right edge of window advances far enough.
4017 * (tcp_recvmsg() will send ACK otherwise). Or...
4019 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4020 /* We ACK each frame or... */
4021 tcp_in_quickack_mode(sk) ||
4022 /* We have out of order data. */
4024 skb_peek(&tp->out_of_order_queue))) {
4025 /* Then ack it now */
4028 /* Else, send delayed ack. */
4029 tcp_send_delayed_ack(sk);
4033 static inline void tcp_ack_snd_check(struct sock *sk)
4035 if (!inet_csk_ack_scheduled(sk)) {
4036 /* We sent a data segment already. */
4039 __tcp_ack_snd_check(sk, 1);
4043 * This routine is only called when we have urgent data
4044 * signaled. Its the 'slow' part of tcp_urg. It could be
4045 * moved inline now as tcp_urg is only called from one
4046 * place. We handle URGent data wrong. We have to - as
4047 * BSD still doesn't use the correction from RFC961.
4048 * For 1003.1g we should support a new option TCP_STDURG to permit
4049 * either form (or just set the sysctl tcp_stdurg).
4052 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4054 struct tcp_sock *tp = tcp_sk(sk);
4055 u32 ptr = ntohs(th->urg_ptr);
4057 if (ptr && !sysctl_tcp_stdurg)
4059 ptr += ntohl(th->seq);
4061 /* Ignore urgent data that we've already seen and read. */
4062 if (after(tp->copied_seq, ptr))
4065 /* Do not replay urg ptr.
4067 * NOTE: interesting situation not covered by specs.
4068 * Misbehaving sender may send urg ptr, pointing to segment,
4069 * which we already have in ofo queue. We are not able to fetch
4070 * such data and will stay in TCP_URG_NOTYET until will be eaten
4071 * by recvmsg(). Seems, we are not obliged to handle such wicked
4072 * situations. But it is worth to think about possibility of some
4073 * DoSes using some hypothetical application level deadlock.
4075 if (before(ptr, tp->rcv_nxt))
4078 /* Do we already have a newer (or duplicate) urgent pointer? */
4079 if (tp->urg_data && !after(ptr, tp->urg_seq))
4082 /* Tell the world about our new urgent pointer. */
4085 /* We may be adding urgent data when the last byte read was
4086 * urgent. To do this requires some care. We cannot just ignore
4087 * tp->copied_seq since we would read the last urgent byte again
4088 * as data, nor can we alter copied_seq until this data arrives
4089 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4091 * NOTE. Double Dutch. Rendering to plain English: author of comment
4092 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4093 * and expect that both A and B disappear from stream. This is _wrong_.
4094 * Though this happens in BSD with high probability, this is occasional.
4095 * Any application relying on this is buggy. Note also, that fix "works"
4096 * only in this artificial test. Insert some normal data between A and B and we will
4097 * decline of BSD again. Verdict: it is better to remove to trap
4100 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4101 !sock_flag(sk, SOCK_URGINLINE) &&
4102 tp->copied_seq != tp->rcv_nxt) {
4103 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4105 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4106 __skb_unlink(skb, &sk->sk_receive_queue);
4111 tp->urg_data = TCP_URG_NOTYET;
4114 /* Disable header prediction. */
4118 /* This is the 'fast' part of urgent handling. */
4119 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4121 struct tcp_sock *tp = tcp_sk(sk);
4123 /* Check if we get a new urgent pointer - normally not. */
4125 tcp_check_urg(sk,th);
4127 /* Do we wait for any urgent data? - normally not... */
4128 if (tp->urg_data == TCP_URG_NOTYET) {
4129 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4132 /* Is the urgent pointer pointing into this packet? */
4133 if (ptr < skb->len) {
4135 if (skb_copy_bits(skb, ptr, &tmp, 1))
4137 tp->urg_data = TCP_URG_VALID | tmp;
4138 if (!sock_flag(sk, SOCK_DEAD))
4139 sk->sk_data_ready(sk, 0);
4144 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4146 struct tcp_sock *tp = tcp_sk(sk);
4147 int chunk = skb->len - hlen;
4151 if (skb_csum_unnecessary(skb))
4152 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4154 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4158 tp->ucopy.len -= chunk;
4159 tp->copied_seq += chunk;
4160 tcp_rcv_space_adjust(sk);
4167 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4171 if (sock_owned_by_user(sk)) {
4173 result = __tcp_checksum_complete(skb);
4176 result = __tcp_checksum_complete(skb);
4181 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4183 return !skb_csum_unnecessary(skb) &&
4184 __tcp_checksum_complete_user(sk, skb);
4187 #ifdef CONFIG_NET_DMA
4188 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4190 struct tcp_sock *tp = tcp_sk(sk);
4191 int chunk = skb->len - hlen;
4193 int copied_early = 0;
4195 if (tp->ucopy.wakeup)
4198 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4199 tp->ucopy.dma_chan = get_softnet_dma();
4201 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4203 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4204 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4209 tp->ucopy.dma_cookie = dma_cookie;
4212 tp->ucopy.len -= chunk;
4213 tp->copied_seq += chunk;
4214 tcp_rcv_space_adjust(sk);
4216 if ((tp->ucopy.len == 0) ||
4217 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4218 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4219 tp->ucopy.wakeup = 1;
4220 sk->sk_data_ready(sk, 0);
4222 } else if (chunk > 0) {
4223 tp->ucopy.wakeup = 1;
4224 sk->sk_data_ready(sk, 0);
4227 return copied_early;
4229 #endif /* CONFIG_NET_DMA */
4232 * TCP receive function for the ESTABLISHED state.
4234 * It is split into a fast path and a slow path. The fast path is
4236 * - A zero window was announced from us - zero window probing
4237 * is only handled properly in the slow path.
4238 * - Out of order segments arrived.
4239 * - Urgent data is expected.
4240 * - There is no buffer space left
4241 * - Unexpected TCP flags/window values/header lengths are received
4242 * (detected by checking the TCP header against pred_flags)
4243 * - Data is sent in both directions. Fast path only supports pure senders
4244 * or pure receivers (this means either the sequence number or the ack
4245 * value must stay constant)
4246 * - Unexpected TCP option.
4248 * When these conditions are not satisfied it drops into a standard
4249 * receive procedure patterned after RFC793 to handle all cases.
4250 * The first three cases are guaranteed by proper pred_flags setting,
4251 * the rest is checked inline. Fast processing is turned on in
4252 * tcp_data_queue when everything is OK.
4254 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4255 struct tcphdr *th, unsigned len)
4257 struct tcp_sock *tp = tcp_sk(sk);
4260 * Header prediction.
4261 * The code loosely follows the one in the famous
4262 * "30 instruction TCP receive" Van Jacobson mail.
4264 * Van's trick is to deposit buffers into socket queue
4265 * on a device interrupt, to call tcp_recv function
4266 * on the receive process context and checksum and copy
4267 * the buffer to user space. smart...
4269 * Our current scheme is not silly either but we take the
4270 * extra cost of the net_bh soft interrupt processing...
4271 * We do checksum and copy also but from device to kernel.
4274 tp->rx_opt.saw_tstamp = 0;
4276 /* pred_flags is 0xS?10 << 16 + snd_wnd
4277 * if header_prediction is to be made
4278 * 'S' will always be tp->tcp_header_len >> 2
4279 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4280 * turn it off (when there are holes in the receive
4281 * space for instance)
4282 * PSH flag is ignored.
4285 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4286 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4287 int tcp_header_len = tp->tcp_header_len;
4289 /* Timestamp header prediction: tcp_header_len
4290 * is automatically equal to th->doff*4 due to pred_flags
4294 /* Check timestamp */
4295 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4296 __be32 *ptr = (__be32 *)(th + 1);
4298 /* No? Slow path! */
4299 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4300 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4303 tp->rx_opt.saw_tstamp = 1;
4305 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4307 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4309 /* If PAWS failed, check it more carefully in slow path */
4310 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4313 /* DO NOT update ts_recent here, if checksum fails
4314 * and timestamp was corrupted part, it will result
4315 * in a hung connection since we will drop all
4316 * future packets due to the PAWS test.
4320 if (len <= tcp_header_len) {
4321 /* Bulk data transfer: sender */
4322 if (len == tcp_header_len) {
4323 /* Predicted packet is in window by definition.
4324 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4325 * Hence, check seq<=rcv_wup reduces to:
4327 if (tcp_header_len ==
4328 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4329 tp->rcv_nxt == tp->rcv_wup)
4330 tcp_store_ts_recent(tp);
4332 /* We know that such packets are checksummed
4335 tcp_ack(sk, skb, 0);
4337 tcp_data_snd_check(sk);
4339 } else { /* Header too small */
4340 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4345 int copied_early = 0;
4347 if (tp->copied_seq == tp->rcv_nxt &&
4348 len - tcp_header_len <= tp->ucopy.len) {
4349 #ifdef CONFIG_NET_DMA
4350 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4355 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4356 __set_current_state(TASK_RUNNING);
4358 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4362 /* Predicted packet is in window by definition.
4363 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4364 * Hence, check seq<=rcv_wup reduces to:
4366 if (tcp_header_len ==
4367 (sizeof(struct tcphdr) +
4368 TCPOLEN_TSTAMP_ALIGNED) &&
4369 tp->rcv_nxt == tp->rcv_wup)
4370 tcp_store_ts_recent(tp);
4372 tcp_rcv_rtt_measure_ts(sk, skb);
4374 __skb_pull(skb, tcp_header_len);
4375 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4376 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4379 tcp_cleanup_rbuf(sk, skb->len);
4382 if (tcp_checksum_complete_user(sk, skb))
4385 /* Predicted packet is in window by definition.
4386 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4387 * Hence, check seq<=rcv_wup reduces to:
4389 if (tcp_header_len ==
4390 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4391 tp->rcv_nxt == tp->rcv_wup)
4392 tcp_store_ts_recent(tp);
4394 tcp_rcv_rtt_measure_ts(sk, skb);
4396 if ((int)skb->truesize > sk->sk_forward_alloc)
4399 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4401 /* Bulk data transfer: receiver */
4402 __skb_pull(skb,tcp_header_len);
4403 __skb_queue_tail(&sk->sk_receive_queue, skb);
4404 sk_stream_set_owner_r(skb, sk);
4405 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4408 tcp_event_data_recv(sk, skb);
4410 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4411 /* Well, only one small jumplet in fast path... */
4412 tcp_ack(sk, skb, FLAG_DATA);
4413 tcp_data_snd_check(sk);
4414 if (!inet_csk_ack_scheduled(sk))
4418 __tcp_ack_snd_check(sk, 0);
4420 #ifdef CONFIG_NET_DMA
4422 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4428 sk->sk_data_ready(sk, 0);
4434 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4438 * RFC1323: H1. Apply PAWS check first.
4440 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4441 tcp_paws_discard(sk, skb)) {
4443 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4444 tcp_send_dupack(sk, skb);
4447 /* Resets are accepted even if PAWS failed.
4449 ts_recent update must be made after we are sure
4450 that the packet is in window.
4455 * Standard slow path.
4458 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4459 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4460 * (RST) segments are validated by checking their SEQ-fields."
4461 * And page 69: "If an incoming segment is not acceptable,
4462 * an acknowledgment should be sent in reply (unless the RST bit
4463 * is set, if so drop the segment and return)".
4466 tcp_send_dupack(sk, skb);
4475 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4477 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4478 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4479 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4486 tcp_ack(sk, skb, FLAG_SLOWPATH);
4488 tcp_rcv_rtt_measure_ts(sk, skb);
4490 /* Process urgent data. */
4491 tcp_urg(sk, skb, th);
4493 /* step 7: process the segment text */
4494 tcp_data_queue(sk, skb);
4496 tcp_data_snd_check(sk);
4497 tcp_ack_snd_check(sk);
4501 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4508 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4509 struct tcphdr *th, unsigned len)
4511 struct tcp_sock *tp = tcp_sk(sk);
4512 struct inet_connection_sock *icsk = inet_csk(sk);
4513 int saved_clamp = tp->rx_opt.mss_clamp;
4515 tcp_parse_options(skb, &tp->rx_opt, 0);
4519 * "If the state is SYN-SENT then
4520 * first check the ACK bit
4521 * If the ACK bit is set
4522 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4523 * a reset (unless the RST bit is set, if so drop
4524 * the segment and return)"
4526 * We do not send data with SYN, so that RFC-correct
4529 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4530 goto reset_and_undo;
4532 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4533 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4535 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4536 goto reset_and_undo;
4539 /* Now ACK is acceptable.
4541 * "If the RST bit is set
4542 * If the ACK was acceptable then signal the user "error:
4543 * connection reset", drop the segment, enter CLOSED state,
4544 * delete TCB, and return."
4553 * "fifth, if neither of the SYN or RST bits is set then
4554 * drop the segment and return."
4560 goto discard_and_undo;
4563 * "If the SYN bit is on ...
4564 * are acceptable then ...
4565 * (our SYN has been ACKed), change the connection
4566 * state to ESTABLISHED..."
4569 TCP_ECN_rcv_synack(tp, th);
4571 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4572 tcp_ack(sk, skb, FLAG_SLOWPATH);
4574 /* Ok.. it's good. Set up sequence numbers and
4575 * move to established.
4577 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4578 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4580 /* RFC1323: The window in SYN & SYN/ACK segments is
4583 tp->snd_wnd = ntohs(th->window);
4584 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4586 if (!tp->rx_opt.wscale_ok) {
4587 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4588 tp->window_clamp = min(tp->window_clamp, 65535U);
4591 if (tp->rx_opt.saw_tstamp) {
4592 tp->rx_opt.tstamp_ok = 1;
4593 tp->tcp_header_len =
4594 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4595 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4596 tcp_store_ts_recent(tp);
4598 tp->tcp_header_len = sizeof(struct tcphdr);
4601 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4602 tp->rx_opt.sack_ok |= 2;
4605 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4606 tcp_initialize_rcv_mss(sk);
4608 /* Remember, tcp_poll() does not lock socket!
4609 * Change state from SYN-SENT only after copied_seq
4610 * is initialized. */
4611 tp->copied_seq = tp->rcv_nxt;
4613 tcp_set_state(sk, TCP_ESTABLISHED);
4615 security_inet_conn_established(sk, skb);
4617 /* Make sure socket is routed, for correct metrics. */
4618 icsk->icsk_af_ops->rebuild_header(sk);
4620 tcp_init_metrics(sk);
4622 tcp_init_congestion_control(sk);
4624 /* Prevent spurious tcp_cwnd_restart() on first data
4627 tp->lsndtime = tcp_time_stamp;
4629 tcp_init_buffer_space(sk);
4631 if (sock_flag(sk, SOCK_KEEPOPEN))
4632 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4634 if (!tp->rx_opt.snd_wscale)
4635 __tcp_fast_path_on(tp, tp->snd_wnd);
4639 if (!sock_flag(sk, SOCK_DEAD)) {
4640 sk->sk_state_change(sk);
4641 sk_wake_async(sk, 0, POLL_OUT);
4644 if (sk->sk_write_pending ||
4645 icsk->icsk_accept_queue.rskq_defer_accept ||
4646 icsk->icsk_ack.pingpong) {
4647 /* Save one ACK. Data will be ready after
4648 * several ticks, if write_pending is set.
4650 * It may be deleted, but with this feature tcpdumps
4651 * look so _wonderfully_ clever, that I was not able
4652 * to stand against the temptation 8) --ANK
4654 inet_csk_schedule_ack(sk);
4655 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4656 icsk->icsk_ack.ato = TCP_ATO_MIN;
4657 tcp_incr_quickack(sk);
4658 tcp_enter_quickack_mode(sk);
4659 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4660 TCP_DELACK_MAX, TCP_RTO_MAX);
4671 /* No ACK in the segment */
4675 * "If the RST bit is set
4677 * Otherwise (no ACK) drop the segment and return."
4680 goto discard_and_undo;
4684 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4685 goto discard_and_undo;
4688 /* We see SYN without ACK. It is attempt of
4689 * simultaneous connect with crossed SYNs.
4690 * Particularly, it can be connect to self.
4692 tcp_set_state(sk, TCP_SYN_RECV);
4694 if (tp->rx_opt.saw_tstamp) {
4695 tp->rx_opt.tstamp_ok = 1;
4696 tcp_store_ts_recent(tp);
4697 tp->tcp_header_len =
4698 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4700 tp->tcp_header_len = sizeof(struct tcphdr);
4703 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4704 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4706 /* RFC1323: The window in SYN & SYN/ACK segments is
4709 tp->snd_wnd = ntohs(th->window);
4710 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4711 tp->max_window = tp->snd_wnd;
4713 TCP_ECN_rcv_syn(tp, th);
4716 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4717 tcp_initialize_rcv_mss(sk);
4720 tcp_send_synack(sk);
4722 /* Note, we could accept data and URG from this segment.
4723 * There are no obstacles to make this.
4725 * However, if we ignore data in ACKless segments sometimes,
4726 * we have no reasons to accept it sometimes.
4727 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4728 * is not flawless. So, discard packet for sanity.
4729 * Uncomment this return to process the data.
4736 /* "fifth, if neither of the SYN or RST bits is set then
4737 * drop the segment and return."
4741 tcp_clear_options(&tp->rx_opt);
4742 tp->rx_opt.mss_clamp = saved_clamp;
4746 tcp_clear_options(&tp->rx_opt);
4747 tp->rx_opt.mss_clamp = saved_clamp;
4753 * This function implements the receiving procedure of RFC 793 for
4754 * all states except ESTABLISHED and TIME_WAIT.
4755 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4756 * address independent.
4759 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4760 struct tcphdr *th, unsigned len)
4762 struct tcp_sock *tp = tcp_sk(sk);
4763 struct inet_connection_sock *icsk = inet_csk(sk);
4766 tp->rx_opt.saw_tstamp = 0;
4768 switch (sk->sk_state) {
4780 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4783 /* Now we have several options: In theory there is
4784 * nothing else in the frame. KA9Q has an option to
4785 * send data with the syn, BSD accepts data with the
4786 * syn up to the [to be] advertised window and
4787 * Solaris 2.1 gives you a protocol error. For now
4788 * we just ignore it, that fits the spec precisely
4789 * and avoids incompatibilities. It would be nice in
4790 * future to drop through and process the data.
4792 * Now that TTCP is starting to be used we ought to
4794 * But, this leaves one open to an easy denial of
4795 * service attack, and SYN cookies can't defend
4796 * against this problem. So, we drop the data
4797 * in the interest of security over speed unless
4798 * it's still in use.
4806 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4810 /* Do step6 onward by hand. */
4811 tcp_urg(sk, skb, th);
4813 tcp_data_snd_check(sk);
4817 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4818 tcp_paws_discard(sk, skb)) {
4820 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4821 tcp_send_dupack(sk, skb);
4824 /* Reset is accepted even if it did not pass PAWS. */
4827 /* step 1: check sequence number */
4828 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4830 tcp_send_dupack(sk, skb);
4834 /* step 2: check RST bit */
4840 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4842 /* step 3: check security and precedence [ignored] */
4846 * Check for a SYN in window.
4848 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4849 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4854 /* step 5: check the ACK field */
4856 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4858 switch (sk->sk_state) {
4861 tp->copied_seq = tp->rcv_nxt;
4863 tcp_set_state(sk, TCP_ESTABLISHED);
4864 sk->sk_state_change(sk);
4866 /* Note, that this wakeup is only for marginal
4867 * crossed SYN case. Passively open sockets
4868 * are not waked up, because sk->sk_sleep ==
4869 * NULL and sk->sk_socket == NULL.
4871 if (sk->sk_socket) {
4872 sk_wake_async(sk,0,POLL_OUT);
4875 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4876 tp->snd_wnd = ntohs(th->window) <<
4877 tp->rx_opt.snd_wscale;
4878 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4879 TCP_SKB_CB(skb)->seq);
4881 /* tcp_ack considers this ACK as duplicate
4882 * and does not calculate rtt.
4883 * Fix it at least with timestamps.
4885 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4887 tcp_ack_saw_tstamp(sk, 0);
4889 if (tp->rx_opt.tstamp_ok)
4890 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4892 /* Make sure socket is routed, for
4895 icsk->icsk_af_ops->rebuild_header(sk);
4897 tcp_init_metrics(sk);
4899 tcp_init_congestion_control(sk);
4901 /* Prevent spurious tcp_cwnd_restart() on
4902 * first data packet.
4904 tp->lsndtime = tcp_time_stamp;
4907 tcp_initialize_rcv_mss(sk);
4908 tcp_init_buffer_space(sk);
4909 tcp_fast_path_on(tp);
4916 if (tp->snd_una == tp->write_seq) {
4917 tcp_set_state(sk, TCP_FIN_WAIT2);
4918 sk->sk_shutdown |= SEND_SHUTDOWN;
4919 dst_confirm(sk->sk_dst_cache);
4921 if (!sock_flag(sk, SOCK_DEAD))
4922 /* Wake up lingering close() */
4923 sk->sk_state_change(sk);
4927 if (tp->linger2 < 0 ||
4928 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4929 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4931 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4935 tmo = tcp_fin_time(sk);
4936 if (tmo > TCP_TIMEWAIT_LEN) {
4937 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4938 } else if (th->fin || sock_owned_by_user(sk)) {
4939 /* Bad case. We could lose such FIN otherwise.
4940 * It is not a big problem, but it looks confusing
4941 * and not so rare event. We still can lose it now,
4942 * if it spins in bh_lock_sock(), but it is really
4945 inet_csk_reset_keepalive_timer(sk, tmo);
4947 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4955 if (tp->snd_una == tp->write_seq) {
4956 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4962 if (tp->snd_una == tp->write_seq) {
4963 tcp_update_metrics(sk);
4972 /* step 6: check the URG bit */
4973 tcp_urg(sk, skb, th);
4975 /* step 7: process the segment text */
4976 switch (sk->sk_state) {
4977 case TCP_CLOSE_WAIT:
4980 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4984 /* RFC 793 says to queue data in these states,
4985 * RFC 1122 says we MUST send a reset.
4986 * BSD 4.4 also does reset.
4988 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4989 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4990 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4991 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4997 case TCP_ESTABLISHED:
4998 tcp_data_queue(sk, skb);
5003 /* tcp_data could move socket to TIME-WAIT */
5004 if (sk->sk_state != TCP_CLOSE) {
5005 tcp_data_snd_check(sk);
5006 tcp_ack_snd_check(sk);
5016 EXPORT_SYMBOL(sysctl_tcp_ecn);
5017 EXPORT_SYMBOL(sysctl_tcp_reordering);
5018 EXPORT_SYMBOL(tcp_parse_options);
5019 EXPORT_SYMBOL(tcp_rcv_established);
5020 EXPORT_SYMBOL(tcp_rcv_state_process);
5021 EXPORT_SYMBOL(tcp_initialize_rcv_mss);