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).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_ecn __read_mostly = 2;
85 EXPORT_SYMBOL(sysctl_tcp_ecn);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
91 int sysctl_tcp_stdurg __read_mostly;
92 int sysctl_tcp_rfc1337 __read_mostly;
93 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
94 int sysctl_tcp_frto __read_mostly = 2;
95 int sysctl_tcp_frto_response __read_mostly;
97 int sysctl_tcp_thin_dupack __read_mostly;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100 int sysctl_tcp_abc __read_mostly;
101 int sysctl_tcp_early_retrans __read_mostly = 2;
103 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
104 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
105 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
106 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
107 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
108 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
109 #define FLAG_ECE 0x40 /* ECE in this ACK */
110 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
112 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
115 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
123 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
124 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
126 /* Adapt the MSS value used to make delayed ack decision to the
129 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
131 struct inet_connection_sock *icsk = inet_csk(sk);
132 const unsigned int lss = icsk->icsk_ack.last_seg_size;
135 icsk->icsk_ack.last_seg_size = 0;
137 /* skb->len may jitter because of SACKs, even if peer
138 * sends good full-sized frames.
140 len = skb_shinfo(skb)->gso_size ? : skb->len;
141 if (len >= icsk->icsk_ack.rcv_mss) {
142 icsk->icsk_ack.rcv_mss = len;
144 /* Otherwise, we make more careful check taking into account,
145 * that SACKs block is variable.
147 * "len" is invariant segment length, including TCP header.
149 len += skb->data - skb_transport_header(skb);
150 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
151 /* If PSH is not set, packet should be
152 * full sized, provided peer TCP is not badly broken.
153 * This observation (if it is correct 8)) allows
154 * to handle super-low mtu links fairly.
156 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
157 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
158 /* Subtract also invariant (if peer is RFC compliant),
159 * tcp header plus fixed timestamp option length.
160 * Resulting "len" is MSS free of SACK jitter.
162 len -= tcp_sk(sk)->tcp_header_len;
163 icsk->icsk_ack.last_seg_size = len;
165 icsk->icsk_ack.rcv_mss = len;
169 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
171 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
175 static void tcp_incr_quickack(struct sock *sk)
177 struct inet_connection_sock *icsk = inet_csk(sk);
178 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
182 if (quickacks > icsk->icsk_ack.quick)
183 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
186 static void tcp_enter_quickack_mode(struct sock *sk)
188 struct inet_connection_sock *icsk = inet_csk(sk);
189 tcp_incr_quickack(sk);
190 icsk->icsk_ack.pingpong = 0;
191 icsk->icsk_ack.ato = TCP_ATO_MIN;
194 /* Send ACKs quickly, if "quick" count is not exhausted
195 * and the session is not interactive.
198 static inline bool tcp_in_quickack_mode(const struct sock *sk)
200 const struct inet_connection_sock *icsk = inet_csk(sk);
202 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
205 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
207 if (tp->ecn_flags & TCP_ECN_OK)
208 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
211 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
213 if (tcp_hdr(skb)->cwr)
214 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
217 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
219 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
222 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
224 if (!(tp->ecn_flags & TCP_ECN_OK))
227 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
228 case INET_ECN_NOT_ECT:
229 /* Funny extension: if ECT is not set on a segment,
230 * and we already seen ECT on a previous segment,
231 * it is probably a retransmit.
233 if (tp->ecn_flags & TCP_ECN_SEEN)
234 tcp_enter_quickack_mode((struct sock *)tp);
237 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
240 tp->ecn_flags |= TCP_ECN_SEEN;
244 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
246 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
247 tp->ecn_flags &= ~TCP_ECN_OK;
250 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
252 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
253 tp->ecn_flags &= ~TCP_ECN_OK;
256 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
258 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
263 /* Buffer size and advertised window tuning.
265 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
268 static void tcp_fixup_sndbuf(struct sock *sk)
270 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
272 sndmem *= TCP_INIT_CWND;
273 if (sk->sk_sndbuf < sndmem)
274 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
277 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
279 * All tcp_full_space() is split to two parts: "network" buffer, allocated
280 * forward and advertised in receiver window (tp->rcv_wnd) and
281 * "application buffer", required to isolate scheduling/application
282 * latencies from network.
283 * window_clamp is maximal advertised window. It can be less than
284 * tcp_full_space(), in this case tcp_full_space() - window_clamp
285 * is reserved for "application" buffer. The less window_clamp is
286 * the smoother our behaviour from viewpoint of network, but the lower
287 * throughput and the higher sensitivity of the connection to losses. 8)
289 * rcv_ssthresh is more strict window_clamp used at "slow start"
290 * phase to predict further behaviour of this connection.
291 * It is used for two goals:
292 * - to enforce header prediction at sender, even when application
293 * requires some significant "application buffer". It is check #1.
294 * - to prevent pruning of receive queue because of misprediction
295 * of receiver window. Check #2.
297 * The scheme does not work when sender sends good segments opening
298 * window and then starts to feed us spaghetti. But it should work
299 * in common situations. Otherwise, we have to rely on queue collapsing.
302 /* Slow part of check#2. */
303 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
305 struct tcp_sock *tp = tcp_sk(sk);
307 int truesize = tcp_win_from_space(skb->truesize) >> 1;
308 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
310 while (tp->rcv_ssthresh <= window) {
311 if (truesize <= skb->len)
312 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
320 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
322 struct tcp_sock *tp = tcp_sk(sk);
325 if (tp->rcv_ssthresh < tp->window_clamp &&
326 (int)tp->rcv_ssthresh < tcp_space(sk) &&
327 !sk_under_memory_pressure(sk)) {
330 /* Check #2. Increase window, if skb with such overhead
331 * will fit to rcvbuf in future.
333 if (tcp_win_from_space(skb->truesize) <= skb->len)
334 incr = 2 * tp->advmss;
336 incr = __tcp_grow_window(sk, skb);
339 incr = max_t(int, incr, 2 * skb->len);
340 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
342 inet_csk(sk)->icsk_ack.quick |= 1;
347 /* 3. Tuning rcvbuf, when connection enters established state. */
349 static void tcp_fixup_rcvbuf(struct sock *sk)
351 u32 mss = tcp_sk(sk)->advmss;
352 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
355 /* Limit to 10 segments if mss <= 1460,
356 * or 14600/mss segments, with a minimum of two segments.
359 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
361 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
362 while (tcp_win_from_space(rcvmem) < mss)
367 if (sk->sk_rcvbuf < rcvmem)
368 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
371 /* 4. Try to fixup all. It is made immediately after connection enters
374 static void tcp_init_buffer_space(struct sock *sk)
376 struct tcp_sock *tp = tcp_sk(sk);
379 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
380 tcp_fixup_rcvbuf(sk);
381 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
382 tcp_fixup_sndbuf(sk);
384 tp->rcvq_space.space = tp->rcv_wnd;
386 maxwin = tcp_full_space(sk);
388 if (tp->window_clamp >= maxwin) {
389 tp->window_clamp = maxwin;
391 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
392 tp->window_clamp = max(maxwin -
393 (maxwin >> sysctl_tcp_app_win),
397 /* Force reservation of one segment. */
398 if (sysctl_tcp_app_win &&
399 tp->window_clamp > 2 * tp->advmss &&
400 tp->window_clamp + tp->advmss > maxwin)
401 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
403 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
404 tp->snd_cwnd_stamp = tcp_time_stamp;
407 /* 5. Recalculate window clamp after socket hit its memory bounds. */
408 static void tcp_clamp_window(struct sock *sk)
410 struct tcp_sock *tp = tcp_sk(sk);
411 struct inet_connection_sock *icsk = inet_csk(sk);
413 icsk->icsk_ack.quick = 0;
415 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
416 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
417 !sk_under_memory_pressure(sk) &&
418 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
419 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
422 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
423 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
426 /* Initialize RCV_MSS value.
427 * RCV_MSS is an our guess about MSS used by the peer.
428 * We haven't any direct information about the MSS.
429 * It's better to underestimate the RCV_MSS rather than overestimate.
430 * Overestimations make us ACKing less frequently than needed.
431 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
433 void tcp_initialize_rcv_mss(struct sock *sk)
435 const struct tcp_sock *tp = tcp_sk(sk);
436 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
438 hint = min(hint, tp->rcv_wnd / 2);
439 hint = min(hint, TCP_MSS_DEFAULT);
440 hint = max(hint, TCP_MIN_MSS);
442 inet_csk(sk)->icsk_ack.rcv_mss = hint;
444 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
446 /* Receiver "autotuning" code.
448 * The algorithm for RTT estimation w/o timestamps is based on
449 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
450 * <http://public.lanl.gov/radiant/pubs.html#DRS>
452 * More detail on this code can be found at
453 * <http://staff.psc.edu/jheffner/>,
454 * though this reference is out of date. A new paper
457 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
459 u32 new_sample = tp->rcv_rtt_est.rtt;
465 if (new_sample != 0) {
466 /* If we sample in larger samples in the non-timestamp
467 * case, we could grossly overestimate the RTT especially
468 * with chatty applications or bulk transfer apps which
469 * are stalled on filesystem I/O.
471 * Also, since we are only going for a minimum in the
472 * non-timestamp case, we do not smooth things out
473 * else with timestamps disabled convergence takes too
477 m -= (new_sample >> 3);
485 /* No previous measure. */
489 if (tp->rcv_rtt_est.rtt != new_sample)
490 tp->rcv_rtt_est.rtt = new_sample;
493 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
495 if (tp->rcv_rtt_est.time == 0)
497 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
499 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
502 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
503 tp->rcv_rtt_est.time = tcp_time_stamp;
506 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
507 const struct sk_buff *skb)
509 struct tcp_sock *tp = tcp_sk(sk);
510 if (tp->rx_opt.rcv_tsecr &&
511 (TCP_SKB_CB(skb)->end_seq -
512 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
513 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
517 * This function should be called every time data is copied to user space.
518 * It calculates the appropriate TCP receive buffer space.
520 void tcp_rcv_space_adjust(struct sock *sk)
522 struct tcp_sock *tp = tcp_sk(sk);
526 if (tp->rcvq_space.time == 0)
529 time = tcp_time_stamp - tp->rcvq_space.time;
530 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
533 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
535 space = max(tp->rcvq_space.space, space);
537 if (tp->rcvq_space.space != space) {
540 tp->rcvq_space.space = space;
542 if (sysctl_tcp_moderate_rcvbuf &&
543 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
544 int new_clamp = space;
546 /* Receive space grows, normalize in order to
547 * take into account packet headers and sk_buff
548 * structure overhead.
553 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
554 while (tcp_win_from_space(rcvmem) < tp->advmss)
557 space = min(space, sysctl_tcp_rmem[2]);
558 if (space > sk->sk_rcvbuf) {
559 sk->sk_rcvbuf = space;
561 /* Make the window clamp follow along. */
562 tp->window_clamp = new_clamp;
568 tp->rcvq_space.seq = tp->copied_seq;
569 tp->rcvq_space.time = tcp_time_stamp;
572 /* There is something which you must keep in mind when you analyze the
573 * behavior of the tp->ato delayed ack timeout interval. When a
574 * connection starts up, we want to ack as quickly as possible. The
575 * problem is that "good" TCP's do slow start at the beginning of data
576 * transmission. The means that until we send the first few ACK's the
577 * sender will sit on his end and only queue most of his data, because
578 * he can only send snd_cwnd unacked packets at any given time. For
579 * each ACK we send, he increments snd_cwnd and transmits more of his
582 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
584 struct tcp_sock *tp = tcp_sk(sk);
585 struct inet_connection_sock *icsk = inet_csk(sk);
588 inet_csk_schedule_ack(sk);
590 tcp_measure_rcv_mss(sk, skb);
592 tcp_rcv_rtt_measure(tp);
594 now = tcp_time_stamp;
596 if (!icsk->icsk_ack.ato) {
597 /* The _first_ data packet received, initialize
598 * delayed ACK engine.
600 tcp_incr_quickack(sk);
601 icsk->icsk_ack.ato = TCP_ATO_MIN;
603 int m = now - icsk->icsk_ack.lrcvtime;
605 if (m <= TCP_ATO_MIN / 2) {
606 /* The fastest case is the first. */
607 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
608 } else if (m < icsk->icsk_ack.ato) {
609 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
610 if (icsk->icsk_ack.ato > icsk->icsk_rto)
611 icsk->icsk_ack.ato = icsk->icsk_rto;
612 } else if (m > icsk->icsk_rto) {
613 /* Too long gap. Apparently sender failed to
614 * restart window, so that we send ACKs quickly.
616 tcp_incr_quickack(sk);
620 icsk->icsk_ack.lrcvtime = now;
622 TCP_ECN_check_ce(tp, skb);
625 tcp_grow_window(sk, skb);
628 /* Called to compute a smoothed rtt estimate. The data fed to this
629 * routine either comes from timestamps, or from segments that were
630 * known _not_ to have been retransmitted [see Karn/Partridge
631 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
632 * piece by Van Jacobson.
633 * NOTE: the next three routines used to be one big routine.
634 * To save cycles in the RFC 1323 implementation it was better to break
635 * it up into three procedures. -- erics
637 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
639 struct tcp_sock *tp = tcp_sk(sk);
640 long m = mrtt; /* RTT */
642 /* The following amusing code comes from Jacobson's
643 * article in SIGCOMM '88. Note that rtt and mdev
644 * are scaled versions of rtt and mean deviation.
645 * This is designed to be as fast as possible
646 * m stands for "measurement".
648 * On a 1990 paper the rto value is changed to:
649 * RTO = rtt + 4 * mdev
651 * Funny. This algorithm seems to be very broken.
652 * These formulae increase RTO, when it should be decreased, increase
653 * too slowly, when it should be increased quickly, decrease too quickly
654 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
655 * does not matter how to _calculate_ it. Seems, it was trap
656 * that VJ failed to avoid. 8)
661 m -= (tp->srtt >> 3); /* m is now error in rtt est */
662 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
664 m = -m; /* m is now abs(error) */
665 m -= (tp->mdev >> 2); /* similar update on mdev */
666 /* This is similar to one of Eifel findings.
667 * Eifel blocks mdev updates when rtt decreases.
668 * This solution is a bit different: we use finer gain
669 * for mdev in this case (alpha*beta).
670 * Like Eifel it also prevents growth of rto,
671 * but also it limits too fast rto decreases,
672 * happening in pure Eifel.
677 m -= (tp->mdev >> 2); /* similar update on mdev */
679 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
680 if (tp->mdev > tp->mdev_max) {
681 tp->mdev_max = tp->mdev;
682 if (tp->mdev_max > tp->rttvar)
683 tp->rttvar = tp->mdev_max;
685 if (after(tp->snd_una, tp->rtt_seq)) {
686 if (tp->mdev_max < tp->rttvar)
687 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
688 tp->rtt_seq = tp->snd_nxt;
689 tp->mdev_max = tcp_rto_min(sk);
692 /* no previous measure. */
693 tp->srtt = m << 3; /* take the measured time to be rtt */
694 tp->mdev = m << 1; /* make sure rto = 3*rtt */
695 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
696 tp->rtt_seq = tp->snd_nxt;
700 /* Calculate rto without backoff. This is the second half of Van Jacobson's
701 * routine referred to above.
703 void tcp_set_rto(struct sock *sk)
705 const struct tcp_sock *tp = tcp_sk(sk);
706 /* Old crap is replaced with new one. 8)
709 * 1. If rtt variance happened to be less 50msec, it is hallucination.
710 * It cannot be less due to utterly erratic ACK generation made
711 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
712 * to do with delayed acks, because at cwnd>2 true delack timeout
713 * is invisible. Actually, Linux-2.4 also generates erratic
714 * ACKs in some circumstances.
716 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
718 /* 2. Fixups made earlier cannot be right.
719 * If we do not estimate RTO correctly without them,
720 * all the algo is pure shit and should be replaced
721 * with correct one. It is exactly, which we pretend to do.
724 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
725 * guarantees that rto is higher.
730 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
732 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
735 cwnd = TCP_INIT_CWND;
736 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
739 /* Set slow start threshold and cwnd not falling to slow start */
740 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
742 struct tcp_sock *tp = tcp_sk(sk);
743 const struct inet_connection_sock *icsk = inet_csk(sk);
745 tp->prior_ssthresh = 0;
747 if (icsk->icsk_ca_state < TCP_CA_CWR) {
750 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
751 tp->snd_cwnd = min(tp->snd_cwnd,
752 tcp_packets_in_flight(tp) + 1U);
753 tp->snd_cwnd_cnt = 0;
754 tp->high_seq = tp->snd_nxt;
755 tp->snd_cwnd_stamp = tcp_time_stamp;
756 TCP_ECN_queue_cwr(tp);
758 tcp_set_ca_state(sk, TCP_CA_CWR);
763 * Packet counting of FACK is based on in-order assumptions, therefore TCP
764 * disables it when reordering is detected
766 void tcp_disable_fack(struct tcp_sock *tp)
768 /* RFC3517 uses different metric in lost marker => reset on change */
770 tp->lost_skb_hint = NULL;
771 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
774 /* Take a notice that peer is sending D-SACKs */
775 static void tcp_dsack_seen(struct tcp_sock *tp)
777 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
780 static void tcp_update_reordering(struct sock *sk, const int metric,
783 struct tcp_sock *tp = tcp_sk(sk);
784 if (metric > tp->reordering) {
787 tp->reordering = min(TCP_MAX_REORDERING, metric);
789 /* This exciting event is worth to be remembered. 8) */
791 mib_idx = LINUX_MIB_TCPTSREORDER;
792 else if (tcp_is_reno(tp))
793 mib_idx = LINUX_MIB_TCPRENOREORDER;
794 else if (tcp_is_fack(tp))
795 mib_idx = LINUX_MIB_TCPFACKREORDER;
797 mib_idx = LINUX_MIB_TCPSACKREORDER;
799 NET_INC_STATS_BH(sock_net(sk), mib_idx);
800 #if FASTRETRANS_DEBUG > 1
801 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
802 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
806 tp->undo_marker ? tp->undo_retrans : 0);
808 tcp_disable_fack(tp);
812 tcp_disable_early_retrans(tp);
815 /* This must be called before lost_out is incremented */
816 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
818 if ((tp->retransmit_skb_hint == NULL) ||
819 before(TCP_SKB_CB(skb)->seq,
820 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
821 tp->retransmit_skb_hint = skb;
824 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
825 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
828 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
830 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
831 tcp_verify_retransmit_hint(tp, skb);
833 tp->lost_out += tcp_skb_pcount(skb);
834 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
838 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
841 tcp_verify_retransmit_hint(tp, skb);
843 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
844 tp->lost_out += tcp_skb_pcount(skb);
845 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
849 /* This procedure tags the retransmission queue when SACKs arrive.
851 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
852 * Packets in queue with these bits set are counted in variables
853 * sacked_out, retrans_out and lost_out, correspondingly.
855 * Valid combinations are:
856 * Tag InFlight Description
857 * 0 1 - orig segment is in flight.
858 * S 0 - nothing flies, orig reached receiver.
859 * L 0 - nothing flies, orig lost by net.
860 * R 2 - both orig and retransmit are in flight.
861 * L|R 1 - orig is lost, retransmit is in flight.
862 * S|R 1 - orig reached receiver, retrans is still in flight.
863 * (L|S|R is logically valid, it could occur when L|R is sacked,
864 * but it is equivalent to plain S and code short-curcuits it to S.
865 * L|S is logically invalid, it would mean -1 packet in flight 8))
867 * These 6 states form finite state machine, controlled by the following events:
868 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
869 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
870 * 3. Loss detection event of two flavors:
871 * A. Scoreboard estimator decided the packet is lost.
872 * A'. Reno "three dupacks" marks head of queue lost.
873 * A''. Its FACK modification, head until snd.fack is lost.
874 * B. SACK arrives sacking SND.NXT at the moment, when the
875 * segment was retransmitted.
876 * 4. D-SACK added new rule: D-SACK changes any tag to S.
878 * It is pleasant to note, that state diagram turns out to be commutative,
879 * so that we are allowed not to be bothered by order of our actions,
880 * when multiple events arrive simultaneously. (see the function below).
882 * Reordering detection.
883 * --------------------
884 * Reordering metric is maximal distance, which a packet can be displaced
885 * in packet stream. With SACKs we can estimate it:
887 * 1. SACK fills old hole and the corresponding segment was not
888 * ever retransmitted -> reordering. Alas, we cannot use it
889 * when segment was retransmitted.
890 * 2. The last flaw is solved with D-SACK. D-SACK arrives
891 * for retransmitted and already SACKed segment -> reordering..
892 * Both of these heuristics are not used in Loss state, when we cannot
893 * account for retransmits accurately.
895 * SACK block validation.
896 * ----------------------
898 * SACK block range validation checks that the received SACK block fits to
899 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
900 * Note that SND.UNA is not included to the range though being valid because
901 * it means that the receiver is rather inconsistent with itself reporting
902 * SACK reneging when it should advance SND.UNA. Such SACK block this is
903 * perfectly valid, however, in light of RFC2018 which explicitly states
904 * that "SACK block MUST reflect the newest segment. Even if the newest
905 * segment is going to be discarded ...", not that it looks very clever
906 * in case of head skb. Due to potentional receiver driven attacks, we
907 * choose to avoid immediate execution of a walk in write queue due to
908 * reneging and defer head skb's loss recovery to standard loss recovery
909 * procedure that will eventually trigger (nothing forbids us doing this).
911 * Implements also blockage to start_seq wrap-around. Problem lies in the
912 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
913 * there's no guarantee that it will be before snd_nxt (n). The problem
914 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
917 * <- outs wnd -> <- wrapzone ->
918 * u e n u_w e_w s n_w
920 * |<------------+------+----- TCP seqno space --------------+---------->|
921 * ...-- <2^31 ->| |<--------...
922 * ...---- >2^31 ------>| |<--------...
924 * Current code wouldn't be vulnerable but it's better still to discard such
925 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
926 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
927 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
928 * equal to the ideal case (infinite seqno space without wrap caused issues).
930 * With D-SACK the lower bound is extended to cover sequence space below
931 * SND.UNA down to undo_marker, which is the last point of interest. Yet
932 * again, D-SACK block must not to go across snd_una (for the same reason as
933 * for the normal SACK blocks, explained above). But there all simplicity
934 * ends, TCP might receive valid D-SACKs below that. As long as they reside
935 * fully below undo_marker they do not affect behavior in anyway and can
936 * therefore be safely ignored. In rare cases (which are more or less
937 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
938 * fragmentation and packet reordering past skb's retransmission. To consider
939 * them correctly, the acceptable range must be extended even more though
940 * the exact amount is rather hard to quantify. However, tp->max_window can
941 * be used as an exaggerated estimate.
943 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
944 u32 start_seq, u32 end_seq)
946 /* Too far in future, or reversed (interpretation is ambiguous) */
947 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
950 /* Nasty start_seq wrap-around check (see comments above) */
951 if (!before(start_seq, tp->snd_nxt))
954 /* In outstanding window? ...This is valid exit for D-SACKs too.
955 * start_seq == snd_una is non-sensical (see comments above)
957 if (after(start_seq, tp->snd_una))
960 if (!is_dsack || !tp->undo_marker)
963 /* ...Then it's D-SACK, and must reside below snd_una completely */
964 if (after(end_seq, tp->snd_una))
967 if (!before(start_seq, tp->undo_marker))
971 if (!after(end_seq, tp->undo_marker))
974 /* Undo_marker boundary crossing (overestimates a lot). Known already:
975 * start_seq < undo_marker and end_seq >= undo_marker.
977 return !before(start_seq, end_seq - tp->max_window);
980 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
981 * Event "B". Later note: FACK people cheated me again 8), we have to account
982 * for reordering! Ugly, but should help.
984 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
985 * less than what is now known to be received by the other end (derived from
986 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
987 * retransmitted skbs to avoid some costly processing per ACKs.
989 static void tcp_mark_lost_retrans(struct sock *sk)
991 const struct inet_connection_sock *icsk = inet_csk(sk);
992 struct tcp_sock *tp = tcp_sk(sk);
995 u32 new_low_seq = tp->snd_nxt;
996 u32 received_upto = tcp_highest_sack_seq(tp);
998 if (!tcp_is_fack(tp) || !tp->retrans_out ||
999 !after(received_upto, tp->lost_retrans_low) ||
1000 icsk->icsk_ca_state != TCP_CA_Recovery)
1003 tcp_for_write_queue(skb, sk) {
1004 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1006 if (skb == tcp_send_head(sk))
1008 if (cnt == tp->retrans_out)
1010 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1013 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1016 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1017 * constraint here (see above) but figuring out that at
1018 * least tp->reordering SACK blocks reside between ack_seq
1019 * and received_upto is not easy task to do cheaply with
1020 * the available datastructures.
1022 * Whether FACK should check here for tp->reordering segs
1023 * in-between one could argue for either way (it would be
1024 * rather simple to implement as we could count fack_count
1025 * during the walk and do tp->fackets_out - fack_count).
1027 if (after(received_upto, ack_seq)) {
1028 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1029 tp->retrans_out -= tcp_skb_pcount(skb);
1031 tcp_skb_mark_lost_uncond_verify(tp, skb);
1032 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1034 if (before(ack_seq, new_low_seq))
1035 new_low_seq = ack_seq;
1036 cnt += tcp_skb_pcount(skb);
1040 if (tp->retrans_out)
1041 tp->lost_retrans_low = new_low_seq;
1044 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1045 struct tcp_sack_block_wire *sp, int num_sacks,
1048 struct tcp_sock *tp = tcp_sk(sk);
1049 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1050 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1051 bool dup_sack = false;
1053 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1056 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1057 } else if (num_sacks > 1) {
1058 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1059 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1061 if (!after(end_seq_0, end_seq_1) &&
1062 !before(start_seq_0, start_seq_1)) {
1065 NET_INC_STATS_BH(sock_net(sk),
1066 LINUX_MIB_TCPDSACKOFORECV);
1070 /* D-SACK for already forgotten data... Do dumb counting. */
1071 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1072 !after(end_seq_0, prior_snd_una) &&
1073 after(end_seq_0, tp->undo_marker))
1079 struct tcp_sacktag_state {
1085 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1086 * the incoming SACK may not exactly match but we can find smaller MSS
1087 * aligned portion of it that matches. Therefore we might need to fragment
1088 * which may fail and creates some hassle (caller must handle error case
1091 * FIXME: this could be merged to shift decision code
1093 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1094 u32 start_seq, u32 end_seq)
1098 unsigned int pkt_len;
1101 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1102 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1104 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1105 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1106 mss = tcp_skb_mss(skb);
1107 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1110 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1114 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1119 /* Round if necessary so that SACKs cover only full MSSes
1120 * and/or the remaining small portion (if present)
1122 if (pkt_len > mss) {
1123 unsigned int new_len = (pkt_len / mss) * mss;
1124 if (!in_sack && new_len < pkt_len) {
1126 if (new_len > skb->len)
1131 err = tcp_fragment(sk, skb, pkt_len, mss);
1139 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1140 static u8 tcp_sacktag_one(struct sock *sk,
1141 struct tcp_sacktag_state *state, u8 sacked,
1142 u32 start_seq, u32 end_seq,
1143 bool dup_sack, int pcount)
1145 struct tcp_sock *tp = tcp_sk(sk);
1146 int fack_count = state->fack_count;
1148 /* Account D-SACK for retransmitted packet. */
1149 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1150 if (tp->undo_marker && tp->undo_retrans &&
1151 after(end_seq, tp->undo_marker))
1153 if (sacked & TCPCB_SACKED_ACKED)
1154 state->reord = min(fack_count, state->reord);
1157 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1158 if (!after(end_seq, tp->snd_una))
1161 if (!(sacked & TCPCB_SACKED_ACKED)) {
1162 if (sacked & TCPCB_SACKED_RETRANS) {
1163 /* If the segment is not tagged as lost,
1164 * we do not clear RETRANS, believing
1165 * that retransmission is still in flight.
1167 if (sacked & TCPCB_LOST) {
1168 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1169 tp->lost_out -= pcount;
1170 tp->retrans_out -= pcount;
1173 if (!(sacked & TCPCB_RETRANS)) {
1174 /* New sack for not retransmitted frame,
1175 * which was in hole. It is reordering.
1177 if (before(start_seq,
1178 tcp_highest_sack_seq(tp)))
1179 state->reord = min(fack_count,
1182 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1183 if (!after(end_seq, tp->frto_highmark))
1184 state->flag |= FLAG_ONLY_ORIG_SACKED;
1187 if (sacked & TCPCB_LOST) {
1188 sacked &= ~TCPCB_LOST;
1189 tp->lost_out -= pcount;
1193 sacked |= TCPCB_SACKED_ACKED;
1194 state->flag |= FLAG_DATA_SACKED;
1195 tp->sacked_out += pcount;
1197 fack_count += pcount;
1199 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1200 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1201 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1202 tp->lost_cnt_hint += pcount;
1204 if (fack_count > tp->fackets_out)
1205 tp->fackets_out = fack_count;
1208 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1209 * frames and clear it. undo_retrans is decreased above, L|R frames
1210 * are accounted above as well.
1212 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1213 sacked &= ~TCPCB_SACKED_RETRANS;
1214 tp->retrans_out -= pcount;
1220 /* Shift newly-SACKed bytes from this skb to the immediately previous
1221 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1223 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1224 struct tcp_sacktag_state *state,
1225 unsigned int pcount, int shifted, int mss,
1228 struct tcp_sock *tp = tcp_sk(sk);
1229 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1230 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1231 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1235 /* Adjust counters and hints for the newly sacked sequence
1236 * range but discard the return value since prev is already
1237 * marked. We must tag the range first because the seq
1238 * advancement below implicitly advances
1239 * tcp_highest_sack_seq() when skb is highest_sack.
1241 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1242 start_seq, end_seq, dup_sack, pcount);
1244 if (skb == tp->lost_skb_hint)
1245 tp->lost_cnt_hint += pcount;
1247 TCP_SKB_CB(prev)->end_seq += shifted;
1248 TCP_SKB_CB(skb)->seq += shifted;
1250 skb_shinfo(prev)->gso_segs += pcount;
1251 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1252 skb_shinfo(skb)->gso_segs -= pcount;
1254 /* When we're adding to gso_segs == 1, gso_size will be zero,
1255 * in theory this shouldn't be necessary but as long as DSACK
1256 * code can come after this skb later on it's better to keep
1257 * setting gso_size to something.
1259 if (!skb_shinfo(prev)->gso_size) {
1260 skb_shinfo(prev)->gso_size = mss;
1261 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1264 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1265 if (skb_shinfo(skb)->gso_segs <= 1) {
1266 skb_shinfo(skb)->gso_size = 0;
1267 skb_shinfo(skb)->gso_type = 0;
1270 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1271 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1274 BUG_ON(!tcp_skb_pcount(skb));
1275 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1279 /* Whole SKB was eaten :-) */
1281 if (skb == tp->retransmit_skb_hint)
1282 tp->retransmit_skb_hint = prev;
1283 if (skb == tp->scoreboard_skb_hint)
1284 tp->scoreboard_skb_hint = prev;
1285 if (skb == tp->lost_skb_hint) {
1286 tp->lost_skb_hint = prev;
1287 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1290 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1291 if (skb == tcp_highest_sack(sk))
1292 tcp_advance_highest_sack(sk, skb);
1294 tcp_unlink_write_queue(skb, sk);
1295 sk_wmem_free_skb(sk, skb);
1297 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1302 /* I wish gso_size would have a bit more sane initialization than
1303 * something-or-zero which complicates things
1305 static int tcp_skb_seglen(const struct sk_buff *skb)
1307 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1310 /* Shifting pages past head area doesn't work */
1311 static int skb_can_shift(const struct sk_buff *skb)
1313 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1316 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1319 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1320 struct tcp_sacktag_state *state,
1321 u32 start_seq, u32 end_seq,
1324 struct tcp_sock *tp = tcp_sk(sk);
1325 struct sk_buff *prev;
1331 if (!sk_can_gso(sk))
1334 /* Normally R but no L won't result in plain S */
1336 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1338 if (!skb_can_shift(skb))
1340 /* This frame is about to be dropped (was ACKed). */
1341 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1344 /* Can only happen with delayed DSACK + discard craziness */
1345 if (unlikely(skb == tcp_write_queue_head(sk)))
1347 prev = tcp_write_queue_prev(sk, skb);
1349 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1352 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1353 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1357 pcount = tcp_skb_pcount(skb);
1358 mss = tcp_skb_seglen(skb);
1360 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1361 * drop this restriction as unnecessary
1363 if (mss != tcp_skb_seglen(prev))
1366 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1368 /* CHECKME: This is non-MSS split case only?, this will
1369 * cause skipped skbs due to advancing loop btw, original
1370 * has that feature too
1372 if (tcp_skb_pcount(skb) <= 1)
1375 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1377 /* TODO: head merge to next could be attempted here
1378 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1379 * though it might not be worth of the additional hassle
1381 * ...we can probably just fallback to what was done
1382 * previously. We could try merging non-SACKed ones
1383 * as well but it probably isn't going to buy off
1384 * because later SACKs might again split them, and
1385 * it would make skb timestamp tracking considerably
1391 len = end_seq - TCP_SKB_CB(skb)->seq;
1393 BUG_ON(len > skb->len);
1395 /* MSS boundaries should be honoured or else pcount will
1396 * severely break even though it makes things bit trickier.
1397 * Optimize common case to avoid most of the divides
1399 mss = tcp_skb_mss(skb);
1401 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1402 * drop this restriction as unnecessary
1404 if (mss != tcp_skb_seglen(prev))
1409 } else if (len < mss) {
1417 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1418 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1421 if (!skb_shift(prev, skb, len))
1423 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1426 /* Hole filled allows collapsing with the next as well, this is very
1427 * useful when hole on every nth skb pattern happens
1429 if (prev == tcp_write_queue_tail(sk))
1431 skb = tcp_write_queue_next(sk, prev);
1433 if (!skb_can_shift(skb) ||
1434 (skb == tcp_send_head(sk)) ||
1435 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1436 (mss != tcp_skb_seglen(skb)))
1440 if (skb_shift(prev, skb, len)) {
1441 pcount += tcp_skb_pcount(skb);
1442 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1446 state->fack_count += pcount;
1453 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1457 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1458 struct tcp_sack_block *next_dup,
1459 struct tcp_sacktag_state *state,
1460 u32 start_seq, u32 end_seq,
1463 struct tcp_sock *tp = tcp_sk(sk);
1464 struct sk_buff *tmp;
1466 tcp_for_write_queue_from(skb, sk) {
1468 bool dup_sack = dup_sack_in;
1470 if (skb == tcp_send_head(sk))
1473 /* queue is in-order => we can short-circuit the walk early */
1474 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1477 if ((next_dup != NULL) &&
1478 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1479 in_sack = tcp_match_skb_to_sack(sk, skb,
1480 next_dup->start_seq,
1486 /* skb reference here is a bit tricky to get right, since
1487 * shifting can eat and free both this skb and the next,
1488 * so not even _safe variant of the loop is enough.
1491 tmp = tcp_shift_skb_data(sk, skb, state,
1492 start_seq, end_seq, dup_sack);
1501 in_sack = tcp_match_skb_to_sack(sk, skb,
1507 if (unlikely(in_sack < 0))
1511 TCP_SKB_CB(skb)->sacked =
1514 TCP_SKB_CB(skb)->sacked,
1515 TCP_SKB_CB(skb)->seq,
1516 TCP_SKB_CB(skb)->end_seq,
1518 tcp_skb_pcount(skb));
1520 if (!before(TCP_SKB_CB(skb)->seq,
1521 tcp_highest_sack_seq(tp)))
1522 tcp_advance_highest_sack(sk, skb);
1525 state->fack_count += tcp_skb_pcount(skb);
1530 /* Avoid all extra work that is being done by sacktag while walking in
1533 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1534 struct tcp_sacktag_state *state,
1537 tcp_for_write_queue_from(skb, sk) {
1538 if (skb == tcp_send_head(sk))
1541 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1544 state->fack_count += tcp_skb_pcount(skb);
1549 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1551 struct tcp_sack_block *next_dup,
1552 struct tcp_sacktag_state *state,
1555 if (next_dup == NULL)
1558 if (before(next_dup->start_seq, skip_to_seq)) {
1559 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1560 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1561 next_dup->start_seq, next_dup->end_seq,
1568 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1570 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1574 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1577 const struct inet_connection_sock *icsk = inet_csk(sk);
1578 struct tcp_sock *tp = tcp_sk(sk);
1579 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1580 TCP_SKB_CB(ack_skb)->sacked);
1581 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1582 struct tcp_sack_block sp[TCP_NUM_SACKS];
1583 struct tcp_sack_block *cache;
1584 struct tcp_sacktag_state state;
1585 struct sk_buff *skb;
1586 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1588 bool found_dup_sack = false;
1590 int first_sack_index;
1593 state.reord = tp->packets_out;
1595 if (!tp->sacked_out) {
1596 if (WARN_ON(tp->fackets_out))
1597 tp->fackets_out = 0;
1598 tcp_highest_sack_reset(sk);
1601 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1602 num_sacks, prior_snd_una);
1604 state.flag |= FLAG_DSACKING_ACK;
1606 /* Eliminate too old ACKs, but take into
1607 * account more or less fresh ones, they can
1608 * contain valid SACK info.
1610 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1613 if (!tp->packets_out)
1617 first_sack_index = 0;
1618 for (i = 0; i < num_sacks; i++) {
1619 bool dup_sack = !i && found_dup_sack;
1621 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1622 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1624 if (!tcp_is_sackblock_valid(tp, dup_sack,
1625 sp[used_sacks].start_seq,
1626 sp[used_sacks].end_seq)) {
1630 if (!tp->undo_marker)
1631 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1633 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1635 /* Don't count olds caused by ACK reordering */
1636 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1637 !after(sp[used_sacks].end_seq, tp->snd_una))
1639 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1642 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1644 first_sack_index = -1;
1648 /* Ignore very old stuff early */
1649 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1655 /* order SACK blocks to allow in order walk of the retrans queue */
1656 for (i = used_sacks - 1; i > 0; i--) {
1657 for (j = 0; j < i; j++) {
1658 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1659 swap(sp[j], sp[j + 1]);
1661 /* Track where the first SACK block goes to */
1662 if (j == first_sack_index)
1663 first_sack_index = j + 1;
1668 skb = tcp_write_queue_head(sk);
1669 state.fack_count = 0;
1672 if (!tp->sacked_out) {
1673 /* It's already past, so skip checking against it */
1674 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1676 cache = tp->recv_sack_cache;
1677 /* Skip empty blocks in at head of the cache */
1678 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1683 while (i < used_sacks) {
1684 u32 start_seq = sp[i].start_seq;
1685 u32 end_seq = sp[i].end_seq;
1686 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1687 struct tcp_sack_block *next_dup = NULL;
1689 if (found_dup_sack && ((i + 1) == first_sack_index))
1690 next_dup = &sp[i + 1];
1692 /* Skip too early cached blocks */
1693 while (tcp_sack_cache_ok(tp, cache) &&
1694 !before(start_seq, cache->end_seq))
1697 /* Can skip some work by looking recv_sack_cache? */
1698 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1699 after(end_seq, cache->start_seq)) {
1702 if (before(start_seq, cache->start_seq)) {
1703 skb = tcp_sacktag_skip(skb, sk, &state,
1705 skb = tcp_sacktag_walk(skb, sk, next_dup,
1712 /* Rest of the block already fully processed? */
1713 if (!after(end_seq, cache->end_seq))
1716 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1720 /* ...tail remains todo... */
1721 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1722 /* ...but better entrypoint exists! */
1723 skb = tcp_highest_sack(sk);
1726 state.fack_count = tp->fackets_out;
1731 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1732 /* Check overlap against next cached too (past this one already) */
1737 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1738 skb = tcp_highest_sack(sk);
1741 state.fack_count = tp->fackets_out;
1743 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1746 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1747 start_seq, end_seq, dup_sack);
1750 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1751 * due to in-order walk
1753 if (after(end_seq, tp->frto_highmark))
1754 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1759 /* Clear the head of the cache sack blocks so we can skip it next time */
1760 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1761 tp->recv_sack_cache[i].start_seq = 0;
1762 tp->recv_sack_cache[i].end_seq = 0;
1764 for (j = 0; j < used_sacks; j++)
1765 tp->recv_sack_cache[i++] = sp[j];
1767 tcp_mark_lost_retrans(sk);
1769 tcp_verify_left_out(tp);
1771 if ((state.reord < tp->fackets_out) &&
1772 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1773 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1774 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1778 #if FASTRETRANS_DEBUG > 0
1779 WARN_ON((int)tp->sacked_out < 0);
1780 WARN_ON((int)tp->lost_out < 0);
1781 WARN_ON((int)tp->retrans_out < 0);
1782 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1787 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1788 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1790 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1794 holes = max(tp->lost_out, 1U);
1795 holes = min(holes, tp->packets_out);
1797 if ((tp->sacked_out + holes) > tp->packets_out) {
1798 tp->sacked_out = tp->packets_out - holes;
1804 /* If we receive more dupacks than we expected counting segments
1805 * in assumption of absent reordering, interpret this as reordering.
1806 * The only another reason could be bug in receiver TCP.
1808 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1810 struct tcp_sock *tp = tcp_sk(sk);
1811 if (tcp_limit_reno_sacked(tp))
1812 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1815 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1817 static void tcp_add_reno_sack(struct sock *sk)
1819 struct tcp_sock *tp = tcp_sk(sk);
1821 tcp_check_reno_reordering(sk, 0);
1822 tcp_verify_left_out(tp);
1825 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1827 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1829 struct tcp_sock *tp = tcp_sk(sk);
1832 /* One ACK acked hole. The rest eat duplicate ACKs. */
1833 if (acked - 1 >= tp->sacked_out)
1836 tp->sacked_out -= acked - 1;
1838 tcp_check_reno_reordering(sk, acked);
1839 tcp_verify_left_out(tp);
1842 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1847 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1849 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1852 /* F-RTO can only be used if TCP has never retransmitted anything other than
1853 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1855 bool tcp_use_frto(struct sock *sk)
1857 const struct tcp_sock *tp = tcp_sk(sk);
1858 const struct inet_connection_sock *icsk = inet_csk(sk);
1859 struct sk_buff *skb;
1861 if (!sysctl_tcp_frto)
1864 /* MTU probe and F-RTO won't really play nicely along currently */
1865 if (icsk->icsk_mtup.probe_size)
1868 if (tcp_is_sackfrto(tp))
1871 /* Avoid expensive walking of rexmit queue if possible */
1872 if (tp->retrans_out > 1)
1875 skb = tcp_write_queue_head(sk);
1876 if (tcp_skb_is_last(sk, skb))
1878 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1879 tcp_for_write_queue_from(skb, sk) {
1880 if (skb == tcp_send_head(sk))
1882 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1884 /* Short-circuit when first non-SACKed skb has been checked */
1885 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1891 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1892 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1893 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1894 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1895 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1896 * bits are handled if the Loss state is really to be entered (in
1897 * tcp_enter_frto_loss).
1899 * Do like tcp_enter_loss() would; when RTO expires the second time it
1901 * "Reduce ssthresh if it has not yet been made inside this window."
1903 void tcp_enter_frto(struct sock *sk)
1905 const struct inet_connection_sock *icsk = inet_csk(sk);
1906 struct tcp_sock *tp = tcp_sk(sk);
1907 struct sk_buff *skb;
1909 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1910 tp->snd_una == tp->high_seq ||
1911 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1912 !icsk->icsk_retransmits)) {
1913 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1914 /* Our state is too optimistic in ssthresh() call because cwnd
1915 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1916 * recovery has not yet completed. Pattern would be this: RTO,
1917 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1919 * RFC4138 should be more specific on what to do, even though
1920 * RTO is quite unlikely to occur after the first Cumulative ACK
1921 * due to back-off and complexity of triggering events ...
1923 if (tp->frto_counter) {
1925 stored_cwnd = tp->snd_cwnd;
1927 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1928 tp->snd_cwnd = stored_cwnd;
1930 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1932 /* ... in theory, cong.control module could do "any tricks" in
1933 * ssthresh(), which means that ca_state, lost bits and lost_out
1934 * counter would have to be faked before the call occurs. We
1935 * consider that too expensive, unlikely and hacky, so modules
1936 * using these in ssthresh() must deal these incompatibility
1937 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1939 tcp_ca_event(sk, CA_EVENT_FRTO);
1942 tp->undo_marker = tp->snd_una;
1943 tp->undo_retrans = 0;
1945 skb = tcp_write_queue_head(sk);
1946 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1947 tp->undo_marker = 0;
1948 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1949 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1950 tp->retrans_out -= tcp_skb_pcount(skb);
1952 tcp_verify_left_out(tp);
1954 /* Too bad if TCP was application limited */
1955 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1957 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1958 * The last condition is necessary at least in tp->frto_counter case.
1960 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
1961 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1962 after(tp->high_seq, tp->snd_una)) {
1963 tp->frto_highmark = tp->high_seq;
1965 tp->frto_highmark = tp->snd_nxt;
1967 tcp_set_ca_state(sk, TCP_CA_Disorder);
1968 tp->high_seq = tp->snd_nxt;
1969 tp->frto_counter = 1;
1972 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1973 * which indicates that we should follow the traditional RTO recovery,
1974 * i.e. mark everything lost and do go-back-N retransmission.
1976 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1978 struct tcp_sock *tp = tcp_sk(sk);
1979 struct sk_buff *skb;
1982 tp->retrans_out = 0;
1983 if (tcp_is_reno(tp))
1984 tcp_reset_reno_sack(tp);
1986 tcp_for_write_queue(skb, sk) {
1987 if (skb == tcp_send_head(sk))
1990 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1992 * Count the retransmission made on RTO correctly (only when
1993 * waiting for the first ACK and did not get it)...
1995 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1996 /* For some reason this R-bit might get cleared? */
1997 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1998 tp->retrans_out += tcp_skb_pcount(skb);
1999 /* ...enter this if branch just for the first segment */
2000 flag |= FLAG_DATA_ACKED;
2002 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2003 tp->undo_marker = 0;
2004 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2007 /* Marking forward transmissions that were made after RTO lost
2008 * can cause unnecessary retransmissions in some scenarios,
2009 * SACK blocks will mitigate that in some but not in all cases.
2010 * We used to not mark them but it was causing break-ups with
2011 * receivers that do only in-order receival.
2013 * TODO: we could detect presence of such receiver and select
2014 * different behavior per flow.
2016 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2017 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2018 tp->lost_out += tcp_skb_pcount(skb);
2019 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2022 tcp_verify_left_out(tp);
2024 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2025 tp->snd_cwnd_cnt = 0;
2026 tp->snd_cwnd_stamp = tcp_time_stamp;
2027 tp->frto_counter = 0;
2028 tp->bytes_acked = 0;
2030 tp->reordering = min_t(unsigned int, tp->reordering,
2031 sysctl_tcp_reordering);
2032 tcp_set_ca_state(sk, TCP_CA_Loss);
2033 tp->high_seq = tp->snd_nxt;
2034 TCP_ECN_queue_cwr(tp);
2036 tcp_clear_all_retrans_hints(tp);
2039 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2041 tp->retrans_out = 0;
2044 tp->undo_marker = 0;
2045 tp->undo_retrans = 0;
2048 void tcp_clear_retrans(struct tcp_sock *tp)
2050 tcp_clear_retrans_partial(tp);
2052 tp->fackets_out = 0;
2056 /* Enter Loss state. If "how" is not zero, forget all SACK information
2057 * and reset tags completely, otherwise preserve SACKs. If receiver
2058 * dropped its ofo queue, we will know this due to reneging detection.
2060 void tcp_enter_loss(struct sock *sk, int how)
2062 const struct inet_connection_sock *icsk = inet_csk(sk);
2063 struct tcp_sock *tp = tcp_sk(sk);
2064 struct sk_buff *skb;
2066 /* Reduce ssthresh if it has not yet been made inside this window. */
2067 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2068 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2069 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2070 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2071 tcp_ca_event(sk, CA_EVENT_LOSS);
2074 tp->snd_cwnd_cnt = 0;
2075 tp->snd_cwnd_stamp = tcp_time_stamp;
2077 tp->bytes_acked = 0;
2078 tcp_clear_retrans_partial(tp);
2080 if (tcp_is_reno(tp))
2081 tcp_reset_reno_sack(tp);
2084 /* Push undo marker, if it was plain RTO and nothing
2085 * was retransmitted. */
2086 tp->undo_marker = tp->snd_una;
2089 tp->fackets_out = 0;
2091 tcp_clear_all_retrans_hints(tp);
2093 tcp_for_write_queue(skb, sk) {
2094 if (skb == tcp_send_head(sk))
2097 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2098 tp->undo_marker = 0;
2099 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2100 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2101 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2102 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2103 tp->lost_out += tcp_skb_pcount(skb);
2104 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2107 tcp_verify_left_out(tp);
2109 tp->reordering = min_t(unsigned int, tp->reordering,
2110 sysctl_tcp_reordering);
2111 tcp_set_ca_state(sk, TCP_CA_Loss);
2112 tp->high_seq = tp->snd_nxt;
2113 TCP_ECN_queue_cwr(tp);
2114 /* Abort F-RTO algorithm if one is in progress */
2115 tp->frto_counter = 0;
2118 /* If ACK arrived pointing to a remembered SACK, it means that our
2119 * remembered SACKs do not reflect real state of receiver i.e.
2120 * receiver _host_ is heavily congested (or buggy).
2122 * Do processing similar to RTO timeout.
2124 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2126 if (flag & FLAG_SACK_RENEGING) {
2127 struct inet_connection_sock *icsk = inet_csk(sk);
2128 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2130 tcp_enter_loss(sk, 1);
2131 icsk->icsk_retransmits++;
2132 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2133 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2134 icsk->icsk_rto, TCP_RTO_MAX);
2140 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2142 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2145 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2146 * counter when SACK is enabled (without SACK, sacked_out is used for
2149 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2150 * segments up to the highest received SACK block so far and holes in
2153 * With reordering, holes may still be in flight, so RFC3517 recovery
2154 * uses pure sacked_out (total number of SACKed segments) even though
2155 * it violates the RFC that uses duplicate ACKs, often these are equal
2156 * but when e.g. out-of-window ACKs or packet duplication occurs,
2157 * they differ. Since neither occurs due to loss, TCP should really
2160 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2162 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2165 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2167 struct tcp_sock *tp = tcp_sk(sk);
2168 unsigned long delay;
2170 /* Delay early retransmit and entering fast recovery for
2171 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2172 * available, or RTO is scheduled to fire first.
2174 if (sysctl_tcp_early_retrans < 2 || (flag & FLAG_ECE) || !tp->srtt)
2177 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2178 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2181 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, TCP_RTO_MAX);
2182 tp->early_retrans_delayed = 1;
2186 static inline int tcp_skb_timedout(const struct sock *sk,
2187 const struct sk_buff *skb)
2189 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2192 static inline int tcp_head_timedout(const struct sock *sk)
2194 const struct tcp_sock *tp = tcp_sk(sk);
2196 return tp->packets_out &&
2197 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2200 /* Linux NewReno/SACK/FACK/ECN state machine.
2201 * --------------------------------------
2203 * "Open" Normal state, no dubious events, fast path.
2204 * "Disorder" In all the respects it is "Open",
2205 * but requires a bit more attention. It is entered when
2206 * we see some SACKs or dupacks. It is split of "Open"
2207 * mainly to move some processing from fast path to slow one.
2208 * "CWR" CWND was reduced due to some Congestion Notification event.
2209 * It can be ECN, ICMP source quench, local device congestion.
2210 * "Recovery" CWND was reduced, we are fast-retransmitting.
2211 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2213 * tcp_fastretrans_alert() is entered:
2214 * - each incoming ACK, if state is not "Open"
2215 * - when arrived ACK is unusual, namely:
2220 * Counting packets in flight is pretty simple.
2222 * in_flight = packets_out - left_out + retrans_out
2224 * packets_out is SND.NXT-SND.UNA counted in packets.
2226 * retrans_out is number of retransmitted segments.
2228 * left_out is number of segments left network, but not ACKed yet.
2230 * left_out = sacked_out + lost_out
2232 * sacked_out: Packets, which arrived to receiver out of order
2233 * and hence not ACKed. With SACKs this number is simply
2234 * amount of SACKed data. Even without SACKs
2235 * it is easy to give pretty reliable estimate of this number,
2236 * counting duplicate ACKs.
2238 * lost_out: Packets lost by network. TCP has no explicit
2239 * "loss notification" feedback from network (for now).
2240 * It means that this number can be only _guessed_.
2241 * Actually, it is the heuristics to predict lossage that
2242 * distinguishes different algorithms.
2244 * F.e. after RTO, when all the queue is considered as lost,
2245 * lost_out = packets_out and in_flight = retrans_out.
2247 * Essentially, we have now two algorithms counting
2250 * FACK: It is the simplest heuristics. As soon as we decided
2251 * that something is lost, we decide that _all_ not SACKed
2252 * packets until the most forward SACK are lost. I.e.
2253 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2254 * It is absolutely correct estimate, if network does not reorder
2255 * packets. And it loses any connection to reality when reordering
2256 * takes place. We use FACK by default until reordering
2257 * is suspected on the path to this destination.
2259 * NewReno: when Recovery is entered, we assume that one segment
2260 * is lost (classic Reno). While we are in Recovery and
2261 * a partial ACK arrives, we assume that one more packet
2262 * is lost (NewReno). This heuristics are the same in NewReno
2265 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2266 * deflation etc. CWND is real congestion window, never inflated, changes
2267 * only according to classic VJ rules.
2269 * Really tricky (and requiring careful tuning) part of algorithm
2270 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2271 * The first determines the moment _when_ we should reduce CWND and,
2272 * hence, slow down forward transmission. In fact, it determines the moment
2273 * when we decide that hole is caused by loss, rather than by a reorder.
2275 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2276 * holes, caused by lost packets.
2278 * And the most logically complicated part of algorithm is undo
2279 * heuristics. We detect false retransmits due to both too early
2280 * fast retransmit (reordering) and underestimated RTO, analyzing
2281 * timestamps and D-SACKs. When we detect that some segments were
2282 * retransmitted by mistake and CWND reduction was wrong, we undo
2283 * window reduction and abort recovery phase. This logic is hidden
2284 * inside several functions named tcp_try_undo_<something>.
2287 /* This function decides, when we should leave Disordered state
2288 * and enter Recovery phase, reducing congestion window.
2290 * Main question: may we further continue forward transmission
2291 * with the same cwnd?
2293 static bool tcp_time_to_recover(struct sock *sk, int flag)
2295 struct tcp_sock *tp = tcp_sk(sk);
2298 /* Do not perform any recovery during F-RTO algorithm */
2299 if (tp->frto_counter)
2302 /* Trick#1: The loss is proven. */
2306 /* Not-A-Trick#2 : Classic rule... */
2307 if (tcp_dupack_heuristics(tp) > tp->reordering)
2310 /* Trick#3 : when we use RFC2988 timer restart, fast
2311 * retransmit can be triggered by timeout of queue head.
2313 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2316 /* Trick#4: It is still not OK... But will it be useful to delay
2319 packets_out = tp->packets_out;
2320 if (packets_out <= tp->reordering &&
2321 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2322 !tcp_may_send_now(sk)) {
2323 /* We have nothing to send. This connection is limited
2324 * either by receiver window or by application.
2329 /* If a thin stream is detected, retransmit after first
2330 * received dupack. Employ only if SACK is supported in order
2331 * to avoid possible corner-case series of spurious retransmissions
2332 * Use only if there are no unsent data.
2334 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2335 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2336 tcp_is_sack(tp) && !tcp_send_head(sk))
2339 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2340 * retransmissions due to small network reorderings, we implement
2341 * Mitigation A.3 in the RFC and delay the retransmission for a short
2342 * interval if appropriate.
2344 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2345 (tp->packets_out == (tp->sacked_out + 1) && tp->packets_out < 4) &&
2346 !tcp_may_send_now(sk))
2347 return !tcp_pause_early_retransmit(sk, flag);
2352 /* New heuristics: it is possible only after we switched to restart timer
2353 * each time when something is ACKed. Hence, we can detect timed out packets
2354 * during fast retransmit without falling to slow start.
2356 * Usefulness of this as is very questionable, since we should know which of
2357 * the segments is the next to timeout which is relatively expensive to find
2358 * in general case unless we add some data structure just for that. The
2359 * current approach certainly won't find the right one too often and when it
2360 * finally does find _something_ it usually marks large part of the window
2361 * right away (because a retransmission with a larger timestamp blocks the
2362 * loop from advancing). -ij
2364 static void tcp_timeout_skbs(struct sock *sk)
2366 struct tcp_sock *tp = tcp_sk(sk);
2367 struct sk_buff *skb;
2369 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2372 skb = tp->scoreboard_skb_hint;
2373 if (tp->scoreboard_skb_hint == NULL)
2374 skb = tcp_write_queue_head(sk);
2376 tcp_for_write_queue_from(skb, sk) {
2377 if (skb == tcp_send_head(sk))
2379 if (!tcp_skb_timedout(sk, skb))
2382 tcp_skb_mark_lost(tp, skb);
2385 tp->scoreboard_skb_hint = skb;
2387 tcp_verify_left_out(tp);
2390 /* Detect loss in event "A" above by marking head of queue up as lost.
2391 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2392 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2393 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2394 * the maximum SACKed segments to pass before reaching this limit.
2396 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2398 struct tcp_sock *tp = tcp_sk(sk);
2399 struct sk_buff *skb;
2403 /* Use SACK to deduce losses of new sequences sent during recovery */
2404 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2406 WARN_ON(packets > tp->packets_out);
2407 if (tp->lost_skb_hint) {
2408 skb = tp->lost_skb_hint;
2409 cnt = tp->lost_cnt_hint;
2410 /* Head already handled? */
2411 if (mark_head && skb != tcp_write_queue_head(sk))
2414 skb = tcp_write_queue_head(sk);
2418 tcp_for_write_queue_from(skb, sk) {
2419 if (skb == tcp_send_head(sk))
2421 /* TODO: do this better */
2422 /* this is not the most efficient way to do this... */
2423 tp->lost_skb_hint = skb;
2424 tp->lost_cnt_hint = cnt;
2426 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2430 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2431 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2432 cnt += tcp_skb_pcount(skb);
2434 if (cnt > packets) {
2435 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2436 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2437 (oldcnt >= packets))
2440 mss = skb_shinfo(skb)->gso_size;
2441 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2447 tcp_skb_mark_lost(tp, skb);
2452 tcp_verify_left_out(tp);
2455 /* Account newly detected lost packet(s) */
2457 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2459 struct tcp_sock *tp = tcp_sk(sk);
2461 if (tcp_is_reno(tp)) {
2462 tcp_mark_head_lost(sk, 1, 1);
2463 } else if (tcp_is_fack(tp)) {
2464 int lost = tp->fackets_out - tp->reordering;
2467 tcp_mark_head_lost(sk, lost, 0);
2469 int sacked_upto = tp->sacked_out - tp->reordering;
2470 if (sacked_upto >= 0)
2471 tcp_mark_head_lost(sk, sacked_upto, 0);
2472 else if (fast_rexmit)
2473 tcp_mark_head_lost(sk, 1, 1);
2476 tcp_timeout_skbs(sk);
2479 /* CWND moderation, preventing bursts due to too big ACKs
2480 * in dubious situations.
2482 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2484 tp->snd_cwnd = min(tp->snd_cwnd,
2485 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2486 tp->snd_cwnd_stamp = tcp_time_stamp;
2489 /* Lower bound on congestion window is slow start threshold
2490 * unless congestion avoidance choice decides to overide it.
2492 static inline u32 tcp_cwnd_min(const struct sock *sk)
2494 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2496 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2499 /* Decrease cwnd each second ack. */
2500 static void tcp_cwnd_down(struct sock *sk, int flag)
2502 struct tcp_sock *tp = tcp_sk(sk);
2503 int decr = tp->snd_cwnd_cnt + 1;
2505 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2506 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2507 tp->snd_cwnd_cnt = decr & 1;
2510 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2511 tp->snd_cwnd -= decr;
2513 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2514 tp->snd_cwnd_stamp = tcp_time_stamp;
2518 /* Nothing was retransmitted or returned timestamp is less
2519 * than timestamp of the first retransmission.
2521 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2523 return !tp->retrans_stamp ||
2524 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2525 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2528 /* Undo procedures. */
2530 #if FASTRETRANS_DEBUG > 1
2531 static void DBGUNDO(struct sock *sk, const char *msg)
2533 struct tcp_sock *tp = tcp_sk(sk);
2534 struct inet_sock *inet = inet_sk(sk);
2536 if (sk->sk_family == AF_INET) {
2537 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2539 &inet->inet_daddr, ntohs(inet->inet_dport),
2540 tp->snd_cwnd, tcp_left_out(tp),
2541 tp->snd_ssthresh, tp->prior_ssthresh,
2544 #if IS_ENABLED(CONFIG_IPV6)
2545 else if (sk->sk_family == AF_INET6) {
2546 struct ipv6_pinfo *np = inet6_sk(sk);
2547 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2549 &np->daddr, ntohs(inet->inet_dport),
2550 tp->snd_cwnd, tcp_left_out(tp),
2551 tp->snd_ssthresh, tp->prior_ssthresh,
2557 #define DBGUNDO(x...) do { } while (0)
2560 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2562 struct tcp_sock *tp = tcp_sk(sk);
2564 if (tp->prior_ssthresh) {
2565 const struct inet_connection_sock *icsk = inet_csk(sk);
2567 if (icsk->icsk_ca_ops->undo_cwnd)
2568 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2570 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2572 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2573 tp->snd_ssthresh = tp->prior_ssthresh;
2574 TCP_ECN_withdraw_cwr(tp);
2577 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2579 tp->snd_cwnd_stamp = tcp_time_stamp;
2582 static inline int tcp_may_undo(const struct tcp_sock *tp)
2584 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2587 /* People celebrate: "We love our President!" */
2588 static bool tcp_try_undo_recovery(struct sock *sk)
2590 struct tcp_sock *tp = tcp_sk(sk);
2592 if (tcp_may_undo(tp)) {
2595 /* Happy end! We did not retransmit anything
2596 * or our original transmission succeeded.
2598 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2599 tcp_undo_cwr(sk, true);
2600 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2601 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2603 mib_idx = LINUX_MIB_TCPFULLUNDO;
2605 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2606 tp->undo_marker = 0;
2608 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2609 /* Hold old state until something *above* high_seq
2610 * is ACKed. For Reno it is MUST to prevent false
2611 * fast retransmits (RFC2582). SACK TCP is safe. */
2612 tcp_moderate_cwnd(tp);
2615 tcp_set_ca_state(sk, TCP_CA_Open);
2619 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2620 static void tcp_try_undo_dsack(struct sock *sk)
2622 struct tcp_sock *tp = tcp_sk(sk);
2624 if (tp->undo_marker && !tp->undo_retrans) {
2625 DBGUNDO(sk, "D-SACK");
2626 tcp_undo_cwr(sk, true);
2627 tp->undo_marker = 0;
2628 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2632 /* We can clear retrans_stamp when there are no retransmissions in the
2633 * window. It would seem that it is trivially available for us in
2634 * tp->retrans_out, however, that kind of assumptions doesn't consider
2635 * what will happen if errors occur when sending retransmission for the
2636 * second time. ...It could the that such segment has only
2637 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2638 * the head skb is enough except for some reneging corner cases that
2639 * are not worth the effort.
2641 * Main reason for all this complexity is the fact that connection dying
2642 * time now depends on the validity of the retrans_stamp, in particular,
2643 * that successive retransmissions of a segment must not advance
2644 * retrans_stamp under any conditions.
2646 static bool tcp_any_retrans_done(const struct sock *sk)
2648 const struct tcp_sock *tp = tcp_sk(sk);
2649 struct sk_buff *skb;
2651 if (tp->retrans_out)
2654 skb = tcp_write_queue_head(sk);
2655 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2661 /* Undo during fast recovery after partial ACK. */
2663 static int tcp_try_undo_partial(struct sock *sk, int acked)
2665 struct tcp_sock *tp = tcp_sk(sk);
2666 /* Partial ACK arrived. Force Hoe's retransmit. */
2667 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2669 if (tcp_may_undo(tp)) {
2670 /* Plain luck! Hole if filled with delayed
2671 * packet, rather than with a retransmit.
2673 if (!tcp_any_retrans_done(sk))
2674 tp->retrans_stamp = 0;
2676 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2679 tcp_undo_cwr(sk, false);
2680 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2682 /* So... Do not make Hoe's retransmit yet.
2683 * If the first packet was delayed, the rest
2684 * ones are most probably delayed as well.
2691 /* Undo during loss recovery after partial ACK. */
2692 static bool tcp_try_undo_loss(struct sock *sk)
2694 struct tcp_sock *tp = tcp_sk(sk);
2696 if (tcp_may_undo(tp)) {
2697 struct sk_buff *skb;
2698 tcp_for_write_queue(skb, sk) {
2699 if (skb == tcp_send_head(sk))
2701 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2704 tcp_clear_all_retrans_hints(tp);
2706 DBGUNDO(sk, "partial loss");
2708 tcp_undo_cwr(sk, true);
2709 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2710 inet_csk(sk)->icsk_retransmits = 0;
2711 tp->undo_marker = 0;
2712 if (tcp_is_sack(tp))
2713 tcp_set_ca_state(sk, TCP_CA_Open);
2719 static inline void tcp_complete_cwr(struct sock *sk)
2721 struct tcp_sock *tp = tcp_sk(sk);
2723 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2724 if (tp->undo_marker) {
2725 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
2726 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2727 tp->snd_cwnd_stamp = tcp_time_stamp;
2728 } else if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH) {
2729 /* PRR algorithm. */
2730 tp->snd_cwnd = tp->snd_ssthresh;
2731 tp->snd_cwnd_stamp = tcp_time_stamp;
2734 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2737 static void tcp_try_keep_open(struct sock *sk)
2739 struct tcp_sock *tp = tcp_sk(sk);
2740 int state = TCP_CA_Open;
2742 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2743 state = TCP_CA_Disorder;
2745 if (inet_csk(sk)->icsk_ca_state != state) {
2746 tcp_set_ca_state(sk, state);
2747 tp->high_seq = tp->snd_nxt;
2751 static void tcp_try_to_open(struct sock *sk, int flag)
2753 struct tcp_sock *tp = tcp_sk(sk);
2755 tcp_verify_left_out(tp);
2757 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2758 tp->retrans_stamp = 0;
2760 if (flag & FLAG_ECE)
2761 tcp_enter_cwr(sk, 1);
2763 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2764 tcp_try_keep_open(sk);
2765 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2766 tcp_moderate_cwnd(tp);
2768 tcp_cwnd_down(sk, flag);
2772 static void tcp_mtup_probe_failed(struct sock *sk)
2774 struct inet_connection_sock *icsk = inet_csk(sk);
2776 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2777 icsk->icsk_mtup.probe_size = 0;
2780 static void tcp_mtup_probe_success(struct sock *sk)
2782 struct tcp_sock *tp = tcp_sk(sk);
2783 struct inet_connection_sock *icsk = inet_csk(sk);
2785 /* FIXME: breaks with very large cwnd */
2786 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2787 tp->snd_cwnd = tp->snd_cwnd *
2788 tcp_mss_to_mtu(sk, tp->mss_cache) /
2789 icsk->icsk_mtup.probe_size;
2790 tp->snd_cwnd_cnt = 0;
2791 tp->snd_cwnd_stamp = tcp_time_stamp;
2792 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2794 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2795 icsk->icsk_mtup.probe_size = 0;
2796 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2799 /* Do a simple retransmit without using the backoff mechanisms in
2800 * tcp_timer. This is used for path mtu discovery.
2801 * The socket is already locked here.
2803 void tcp_simple_retransmit(struct sock *sk)
2805 const struct inet_connection_sock *icsk = inet_csk(sk);
2806 struct tcp_sock *tp = tcp_sk(sk);
2807 struct sk_buff *skb;
2808 unsigned int mss = tcp_current_mss(sk);
2809 u32 prior_lost = tp->lost_out;
2811 tcp_for_write_queue(skb, sk) {
2812 if (skb == tcp_send_head(sk))
2814 if (tcp_skb_seglen(skb) > mss &&
2815 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2816 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2817 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2818 tp->retrans_out -= tcp_skb_pcount(skb);
2820 tcp_skb_mark_lost_uncond_verify(tp, skb);
2824 tcp_clear_retrans_hints_partial(tp);
2826 if (prior_lost == tp->lost_out)
2829 if (tcp_is_reno(tp))
2830 tcp_limit_reno_sacked(tp);
2832 tcp_verify_left_out(tp);
2834 /* Don't muck with the congestion window here.
2835 * Reason is that we do not increase amount of _data_
2836 * in network, but units changed and effective
2837 * cwnd/ssthresh really reduced now.
2839 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2840 tp->high_seq = tp->snd_nxt;
2841 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2842 tp->prior_ssthresh = 0;
2843 tp->undo_marker = 0;
2844 tcp_set_ca_state(sk, TCP_CA_Loss);
2846 tcp_xmit_retransmit_queue(sk);
2848 EXPORT_SYMBOL(tcp_simple_retransmit);
2850 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2851 * (proportional rate reduction with slow start reduction bound) as described in
2852 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2853 * It computes the number of packets to send (sndcnt) based on packets newly
2855 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2856 * cwnd reductions across a full RTT.
2857 * 2) If packets in flight is lower than ssthresh (such as due to excess
2858 * losses and/or application stalls), do not perform any further cwnd
2859 * reductions, but instead slow start up to ssthresh.
2861 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
2862 int fast_rexmit, int flag)
2864 struct tcp_sock *tp = tcp_sk(sk);
2866 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2868 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2869 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2871 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2873 sndcnt = min_t(int, delta,
2874 max_t(int, tp->prr_delivered - tp->prr_out,
2875 newly_acked_sacked) + 1);
2878 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2879 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2882 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2884 struct tcp_sock *tp = tcp_sk(sk);
2887 if (tcp_is_reno(tp))
2888 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2890 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2892 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2894 tp->high_seq = tp->snd_nxt;
2895 tp->prior_ssthresh = 0;
2896 tp->undo_marker = tp->snd_una;
2897 tp->undo_retrans = tp->retrans_out;
2899 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2901 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2902 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2903 TCP_ECN_queue_cwr(tp);
2906 tp->bytes_acked = 0;
2907 tp->snd_cwnd_cnt = 0;
2908 tp->prior_cwnd = tp->snd_cwnd;
2909 tp->prr_delivered = 0;
2911 tcp_set_ca_state(sk, TCP_CA_Recovery);
2914 /* Process an event, which can update packets-in-flight not trivially.
2915 * Main goal of this function is to calculate new estimate for left_out,
2916 * taking into account both packets sitting in receiver's buffer and
2917 * packets lost by network.
2919 * Besides that it does CWND reduction, when packet loss is detected
2920 * and changes state of machine.
2922 * It does _not_ decide what to send, it is made in function
2923 * tcp_xmit_retransmit_queue().
2925 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2926 int newly_acked_sacked, bool is_dupack,
2929 struct inet_connection_sock *icsk = inet_csk(sk);
2930 struct tcp_sock *tp = tcp_sk(sk);
2931 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2932 (tcp_fackets_out(tp) > tp->reordering));
2933 int fast_rexmit = 0;
2935 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2937 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2938 tp->fackets_out = 0;
2940 /* Now state machine starts.
2941 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2942 if (flag & FLAG_ECE)
2943 tp->prior_ssthresh = 0;
2945 /* B. In all the states check for reneging SACKs. */
2946 if (tcp_check_sack_reneging(sk, flag))
2949 /* C. Check consistency of the current state. */
2950 tcp_verify_left_out(tp);
2952 /* D. Check state exit conditions. State can be terminated
2953 * when high_seq is ACKed. */
2954 if (icsk->icsk_ca_state == TCP_CA_Open) {
2955 WARN_ON(tp->retrans_out != 0);
2956 tp->retrans_stamp = 0;
2957 } else if (!before(tp->snd_una, tp->high_seq)) {
2958 switch (icsk->icsk_ca_state) {
2960 icsk->icsk_retransmits = 0;
2961 if (tcp_try_undo_recovery(sk))
2966 /* CWR is to be held something *above* high_seq
2967 * is ACKed for CWR bit to reach receiver. */
2968 if (tp->snd_una != tp->high_seq) {
2969 tcp_complete_cwr(sk);
2970 tcp_set_ca_state(sk, TCP_CA_Open);
2974 case TCP_CA_Recovery:
2975 if (tcp_is_reno(tp))
2976 tcp_reset_reno_sack(tp);
2977 if (tcp_try_undo_recovery(sk))
2979 tcp_complete_cwr(sk);
2984 /* E. Process state. */
2985 switch (icsk->icsk_ca_state) {
2986 case TCP_CA_Recovery:
2987 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2988 if (tcp_is_reno(tp) && is_dupack)
2989 tcp_add_reno_sack(sk);
2991 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2994 if (flag & FLAG_DATA_ACKED)
2995 icsk->icsk_retransmits = 0;
2996 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2997 tcp_reset_reno_sack(tp);
2998 if (!tcp_try_undo_loss(sk)) {
2999 tcp_moderate_cwnd(tp);
3000 tcp_xmit_retransmit_queue(sk);
3003 if (icsk->icsk_ca_state != TCP_CA_Open)
3005 /* Loss is undone; fall through to processing in Open state. */
3007 if (tcp_is_reno(tp)) {
3008 if (flag & FLAG_SND_UNA_ADVANCED)
3009 tcp_reset_reno_sack(tp);
3011 tcp_add_reno_sack(sk);
3014 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3015 tcp_try_undo_dsack(sk);
3017 if (!tcp_time_to_recover(sk, flag)) {
3018 tcp_try_to_open(sk, flag);
3022 /* MTU probe failure: don't reduce cwnd */
3023 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3024 icsk->icsk_mtup.probe_size &&
3025 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3026 tcp_mtup_probe_failed(sk);
3027 /* Restores the reduction we did in tcp_mtup_probe() */
3029 tcp_simple_retransmit(sk);
3033 /* Otherwise enter Recovery state */
3034 tcp_enter_recovery(sk, (flag & FLAG_ECE));
3038 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3039 tcp_update_scoreboard(sk, fast_rexmit);
3040 tp->prr_delivered += newly_acked_sacked;
3041 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3042 tcp_xmit_retransmit_queue(sk);
3045 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3047 tcp_rtt_estimator(sk, seq_rtt);
3049 inet_csk(sk)->icsk_backoff = 0;
3051 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3053 /* Read draft-ietf-tcplw-high-performance before mucking
3054 * with this code. (Supersedes RFC1323)
3056 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3058 /* RTTM Rule: A TSecr value received in a segment is used to
3059 * update the averaged RTT measurement only if the segment
3060 * acknowledges some new data, i.e., only if it advances the
3061 * left edge of the send window.
3063 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3064 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3066 * Changed: reset backoff as soon as we see the first valid sample.
3067 * If we do not, we get strongly overestimated rto. With timestamps
3068 * samples are accepted even from very old segments: f.e., when rtt=1
3069 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3070 * answer arrives rto becomes 120 seconds! If at least one of segments
3071 * in window is lost... Voila. --ANK (010210)
3073 struct tcp_sock *tp = tcp_sk(sk);
3075 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3078 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3080 /* We don't have a timestamp. Can only use
3081 * packets that are not retransmitted to determine
3082 * rtt estimates. Also, we must not reset the
3083 * backoff for rto until we get a non-retransmitted
3084 * packet. This allows us to deal with a situation
3085 * where the network delay has increased suddenly.
3086 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3089 if (flag & FLAG_RETRANS_DATA_ACKED)
3092 tcp_valid_rtt_meas(sk, seq_rtt);
3095 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3098 const struct tcp_sock *tp = tcp_sk(sk);
3099 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3100 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3101 tcp_ack_saw_tstamp(sk, flag);
3102 else if (seq_rtt >= 0)
3103 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3106 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3108 const struct inet_connection_sock *icsk = inet_csk(sk);
3109 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3110 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3113 /* Restart timer after forward progress on connection.
3114 * RFC2988 recommends to restart timer to now+rto.
3116 void tcp_rearm_rto(struct sock *sk)
3118 struct tcp_sock *tp = tcp_sk(sk);
3120 if (!tp->packets_out) {
3121 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3123 u32 rto = inet_csk(sk)->icsk_rto;
3124 /* Offset the time elapsed after installing regular RTO */
3125 if (tp->early_retrans_delayed) {
3126 struct sk_buff *skb = tcp_write_queue_head(sk);
3127 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
3128 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3129 /* delta may not be positive if the socket is locked
3130 * when the delayed ER timer fires and is rescheduled.
3135 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3138 tp->early_retrans_delayed = 0;
3141 /* This function is called when the delayed ER timer fires. TCP enters
3142 * fast recovery and performs fast-retransmit.
3144 void tcp_resume_early_retransmit(struct sock *sk)
3146 struct tcp_sock *tp = tcp_sk(sk);
3150 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3151 if (!tp->do_early_retrans)
3154 tcp_enter_recovery(sk, false);
3155 tcp_update_scoreboard(sk, 1);
3156 tcp_xmit_retransmit_queue(sk);
3159 /* If we get here, the whole TSO packet has not been acked. */
3160 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3162 struct tcp_sock *tp = tcp_sk(sk);
3165 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3167 packets_acked = tcp_skb_pcount(skb);
3168 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3170 packets_acked -= tcp_skb_pcount(skb);
3172 if (packets_acked) {
3173 BUG_ON(tcp_skb_pcount(skb) == 0);
3174 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3177 return packets_acked;
3180 /* Remove acknowledged frames from the retransmission queue. If our packet
3181 * is before the ack sequence we can discard it as it's confirmed to have
3182 * arrived at the other end.
3184 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3187 struct tcp_sock *tp = tcp_sk(sk);
3188 const struct inet_connection_sock *icsk = inet_csk(sk);
3189 struct sk_buff *skb;
3190 u32 now = tcp_time_stamp;
3191 int fully_acked = true;
3194 u32 reord = tp->packets_out;
3195 u32 prior_sacked = tp->sacked_out;
3197 s32 ca_seq_rtt = -1;
3198 ktime_t last_ackt = net_invalid_timestamp();
3200 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3201 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3203 u8 sacked = scb->sacked;
3205 /* Determine how many packets and what bytes were acked, tso and else */
3206 if (after(scb->end_seq, tp->snd_una)) {
3207 if (tcp_skb_pcount(skb) == 1 ||
3208 !after(tp->snd_una, scb->seq))
3211 acked_pcount = tcp_tso_acked(sk, skb);
3215 fully_acked = false;
3217 acked_pcount = tcp_skb_pcount(skb);
3220 if (sacked & TCPCB_RETRANS) {
3221 if (sacked & TCPCB_SACKED_RETRANS)
3222 tp->retrans_out -= acked_pcount;
3223 flag |= FLAG_RETRANS_DATA_ACKED;
3226 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3227 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3229 ca_seq_rtt = now - scb->when;
3230 last_ackt = skb->tstamp;
3232 seq_rtt = ca_seq_rtt;
3234 if (!(sacked & TCPCB_SACKED_ACKED))
3235 reord = min(pkts_acked, reord);
3238 if (sacked & TCPCB_SACKED_ACKED)
3239 tp->sacked_out -= acked_pcount;
3240 if (sacked & TCPCB_LOST)
3241 tp->lost_out -= acked_pcount;
3243 tp->packets_out -= acked_pcount;
3244 pkts_acked += acked_pcount;
3246 /* Initial outgoing SYN's get put onto the write_queue
3247 * just like anything else we transmit. It is not
3248 * true data, and if we misinform our callers that
3249 * this ACK acks real data, we will erroneously exit
3250 * connection startup slow start one packet too
3251 * quickly. This is severely frowned upon behavior.
3253 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3254 flag |= FLAG_DATA_ACKED;
3256 flag |= FLAG_SYN_ACKED;
3257 tp->retrans_stamp = 0;
3263 tcp_unlink_write_queue(skb, sk);
3264 sk_wmem_free_skb(sk, skb);
3265 tp->scoreboard_skb_hint = NULL;
3266 if (skb == tp->retransmit_skb_hint)
3267 tp->retransmit_skb_hint = NULL;
3268 if (skb == tp->lost_skb_hint)
3269 tp->lost_skb_hint = NULL;
3272 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3273 tp->snd_up = tp->snd_una;
3275 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3276 flag |= FLAG_SACK_RENEGING;
3278 if (flag & FLAG_ACKED) {
3279 const struct tcp_congestion_ops *ca_ops
3280 = inet_csk(sk)->icsk_ca_ops;
3282 if (unlikely(icsk->icsk_mtup.probe_size &&
3283 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3284 tcp_mtup_probe_success(sk);
3287 tcp_ack_update_rtt(sk, flag, seq_rtt);
3290 if (tcp_is_reno(tp)) {
3291 tcp_remove_reno_sacks(sk, pkts_acked);
3295 /* Non-retransmitted hole got filled? That's reordering */
3296 if (reord < prior_fackets)
3297 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3299 delta = tcp_is_fack(tp) ? pkts_acked :
3300 prior_sacked - tp->sacked_out;
3301 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3304 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3306 if (ca_ops->pkts_acked) {
3309 /* Is the ACK triggering packet unambiguous? */
3310 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3311 /* High resolution needed and available? */
3312 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3313 !ktime_equal(last_ackt,
3314 net_invalid_timestamp()))
3315 rtt_us = ktime_us_delta(ktime_get_real(),
3317 else if (ca_seq_rtt >= 0)
3318 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3321 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3325 #if FASTRETRANS_DEBUG > 0
3326 WARN_ON((int)tp->sacked_out < 0);
3327 WARN_ON((int)tp->lost_out < 0);
3328 WARN_ON((int)tp->retrans_out < 0);
3329 if (!tp->packets_out && tcp_is_sack(tp)) {
3330 icsk = inet_csk(sk);
3332 pr_debug("Leak l=%u %d\n",
3333 tp->lost_out, icsk->icsk_ca_state);
3336 if (tp->sacked_out) {
3337 pr_debug("Leak s=%u %d\n",
3338 tp->sacked_out, icsk->icsk_ca_state);
3341 if (tp->retrans_out) {
3342 pr_debug("Leak r=%u %d\n",
3343 tp->retrans_out, icsk->icsk_ca_state);
3344 tp->retrans_out = 0;
3351 static void tcp_ack_probe(struct sock *sk)
3353 const struct tcp_sock *tp = tcp_sk(sk);
3354 struct inet_connection_sock *icsk = inet_csk(sk);
3356 /* Was it a usable window open? */
3358 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3359 icsk->icsk_backoff = 0;
3360 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3361 /* Socket must be waked up by subsequent tcp_data_snd_check().
3362 * This function is not for random using!
3365 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3366 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3371 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3373 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3374 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3377 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3379 const struct tcp_sock *tp = tcp_sk(sk);
3380 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3381 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3384 /* Check that window update is acceptable.
3385 * The function assumes that snd_una<=ack<=snd_next.
3387 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3388 const u32 ack, const u32 ack_seq,
3391 return after(ack, tp->snd_una) ||
3392 after(ack_seq, tp->snd_wl1) ||
3393 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3396 /* Update our send window.
3398 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3399 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3401 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3404 struct tcp_sock *tp = tcp_sk(sk);
3406 u32 nwin = ntohs(tcp_hdr(skb)->window);
3408 if (likely(!tcp_hdr(skb)->syn))
3409 nwin <<= tp->rx_opt.snd_wscale;
3411 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3412 flag |= FLAG_WIN_UPDATE;
3413 tcp_update_wl(tp, ack_seq);
3415 if (tp->snd_wnd != nwin) {
3418 /* Note, it is the only place, where
3419 * fast path is recovered for sending TCP.
3422 tcp_fast_path_check(sk);
3424 if (nwin > tp->max_window) {
3425 tp->max_window = nwin;
3426 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3436 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3437 * continue in congestion avoidance.
3439 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3441 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3442 tp->snd_cwnd_cnt = 0;
3443 tp->bytes_acked = 0;
3444 TCP_ECN_queue_cwr(tp);
3445 tcp_moderate_cwnd(tp);
3448 /* A conservative spurious RTO response algorithm: reduce cwnd using
3449 * rate halving and continue in congestion avoidance.
3451 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3453 tcp_enter_cwr(sk, 0);
3456 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3458 if (flag & FLAG_ECE)
3459 tcp_ratehalving_spur_to_response(sk);
3461 tcp_undo_cwr(sk, true);
3464 /* F-RTO spurious RTO detection algorithm (RFC4138)
3466 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3467 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3468 * window (but not to or beyond highest sequence sent before RTO):
3469 * On First ACK, send two new segments out.
3470 * On Second ACK, RTO was likely spurious. Do spurious response (response
3471 * algorithm is not part of the F-RTO detection algorithm
3472 * given in RFC4138 but can be selected separately).
3473 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3474 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3475 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3476 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3478 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3479 * original window even after we transmit two new data segments.
3482 * on first step, wait until first cumulative ACK arrives, then move to
3483 * the second step. In second step, the next ACK decides.
3485 * F-RTO is implemented (mainly) in four functions:
3486 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3487 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3488 * called when tcp_use_frto() showed green light
3489 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3490 * - tcp_enter_frto_loss() is called if there is not enough evidence
3491 * to prove that the RTO is indeed spurious. It transfers the control
3492 * from F-RTO to the conventional RTO recovery
3494 static bool tcp_process_frto(struct sock *sk, int flag)
3496 struct tcp_sock *tp = tcp_sk(sk);
3498 tcp_verify_left_out(tp);
3500 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3501 if (flag & FLAG_DATA_ACKED)
3502 inet_csk(sk)->icsk_retransmits = 0;
3504 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3505 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3506 tp->undo_marker = 0;
3508 if (!before(tp->snd_una, tp->frto_highmark)) {
3509 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3513 if (!tcp_is_sackfrto(tp)) {
3514 /* RFC4138 shortcoming in step 2; should also have case c):
3515 * ACK isn't duplicate nor advances window, e.g., opposite dir
3518 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3521 if (!(flag & FLAG_DATA_ACKED)) {
3522 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3527 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3528 /* Prevent sending of new data. */
3529 tp->snd_cwnd = min(tp->snd_cwnd,
3530 tcp_packets_in_flight(tp));
3534 if ((tp->frto_counter >= 2) &&
3535 (!(flag & FLAG_FORWARD_PROGRESS) ||
3536 ((flag & FLAG_DATA_SACKED) &&
3537 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3538 /* RFC4138 shortcoming (see comment above) */
3539 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3540 (flag & FLAG_NOT_DUP))
3543 tcp_enter_frto_loss(sk, 3, flag);
3548 if (tp->frto_counter == 1) {
3549 /* tcp_may_send_now needs to see updated state */
3550 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3551 tp->frto_counter = 2;
3553 if (!tcp_may_send_now(sk))
3554 tcp_enter_frto_loss(sk, 2, flag);
3558 switch (sysctl_tcp_frto_response) {
3560 tcp_undo_spur_to_response(sk, flag);
3563 tcp_conservative_spur_to_response(tp);
3566 tcp_ratehalving_spur_to_response(sk);
3569 tp->frto_counter = 0;
3570 tp->undo_marker = 0;
3571 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3576 /* This routine deals with incoming acks, but not outgoing ones. */
3577 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3579 struct inet_connection_sock *icsk = inet_csk(sk);
3580 struct tcp_sock *tp = tcp_sk(sk);
3581 u32 prior_snd_una = tp->snd_una;
3582 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3583 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3584 bool is_dupack = false;
3585 u32 prior_in_flight;
3588 int prior_sacked = tp->sacked_out;
3590 int newly_acked_sacked = 0;
3591 bool frto_cwnd = false;
3593 /* If the ack is older than previous acks
3594 * then we can probably ignore it.
3596 if (before(ack, prior_snd_una))
3599 /* If the ack includes data we haven't sent yet, discard
3600 * this segment (RFC793 Section 3.9).
3602 if (after(ack, tp->snd_nxt))
3605 if (tp->early_retrans_delayed)
3608 if (after(ack, prior_snd_una))
3609 flag |= FLAG_SND_UNA_ADVANCED;
3611 if (sysctl_tcp_abc) {
3612 if (icsk->icsk_ca_state < TCP_CA_CWR)
3613 tp->bytes_acked += ack - prior_snd_una;
3614 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3615 /* we assume just one segment left network */
3616 tp->bytes_acked += min(ack - prior_snd_una,
3620 prior_fackets = tp->fackets_out;
3621 prior_in_flight = tcp_packets_in_flight(tp);
3623 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3624 /* Window is constant, pure forward advance.
3625 * No more checks are required.
3626 * Note, we use the fact that SND.UNA>=SND.WL2.
3628 tcp_update_wl(tp, ack_seq);
3630 flag |= FLAG_WIN_UPDATE;
3632 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3634 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3636 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3639 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3641 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3643 if (TCP_SKB_CB(skb)->sacked)
3644 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3646 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3649 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3652 /* We passed data and got it acked, remove any soft error
3653 * log. Something worked...
3655 sk->sk_err_soft = 0;
3656 icsk->icsk_probes_out = 0;
3657 tp->rcv_tstamp = tcp_time_stamp;
3658 prior_packets = tp->packets_out;
3662 /* See if we can take anything off of the retransmit queue. */
3663 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3665 pkts_acked = prior_packets - tp->packets_out;
3666 newly_acked_sacked = (prior_packets - prior_sacked) -
3667 (tp->packets_out - tp->sacked_out);
3669 if (tp->frto_counter)
3670 frto_cwnd = tcp_process_frto(sk, flag);
3671 /* Guarantee sacktag reordering detection against wrap-arounds */
3672 if (before(tp->frto_highmark, tp->snd_una))
3673 tp->frto_highmark = 0;
3675 if (tcp_ack_is_dubious(sk, flag)) {
3676 /* Advance CWND, if state allows this. */
3677 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3678 tcp_may_raise_cwnd(sk, flag))
3679 tcp_cong_avoid(sk, ack, prior_in_flight);
3680 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3681 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3684 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3685 tcp_cong_avoid(sk, ack, prior_in_flight);
3688 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3689 struct dst_entry *dst = __sk_dst_get(sk);
3696 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3697 if (flag & FLAG_DSACKING_ACK)
3698 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3700 /* If this ack opens up a zero window, clear backoff. It was
3701 * being used to time the probes, and is probably far higher than
3702 * it needs to be for normal retransmission.
3704 if (tcp_send_head(sk))
3709 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3713 /* If data was SACKed, tag it and see if we should send more data.
3714 * If data was DSACKed, see if we can undo a cwnd reduction.
3716 if (TCP_SKB_CB(skb)->sacked) {
3717 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3718 newly_acked_sacked = tp->sacked_out - prior_sacked;
3719 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3723 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3727 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3728 * But, this can also be called on packets in the established flow when
3729 * the fast version below fails.
3731 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3732 const u8 **hvpp, int estab)
3734 const unsigned char *ptr;
3735 const struct tcphdr *th = tcp_hdr(skb);
3736 int length = (th->doff * 4) - sizeof(struct tcphdr);
3738 ptr = (const unsigned char *)(th + 1);
3739 opt_rx->saw_tstamp = 0;
3741 while (length > 0) {
3742 int opcode = *ptr++;
3748 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3753 if (opsize < 2) /* "silly options" */
3755 if (opsize > length)
3756 return; /* don't parse partial options */
3759 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3760 u16 in_mss = get_unaligned_be16(ptr);
3762 if (opt_rx->user_mss &&
3763 opt_rx->user_mss < in_mss)
3764 in_mss = opt_rx->user_mss;
3765 opt_rx->mss_clamp = in_mss;
3770 if (opsize == TCPOLEN_WINDOW && th->syn &&
3771 !estab && sysctl_tcp_window_scaling) {
3772 __u8 snd_wscale = *(__u8 *)ptr;
3773 opt_rx->wscale_ok = 1;
3774 if (snd_wscale > 14) {
3775 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3780 opt_rx->snd_wscale = snd_wscale;
3783 case TCPOPT_TIMESTAMP:
3784 if ((opsize == TCPOLEN_TIMESTAMP) &&
3785 ((estab && opt_rx->tstamp_ok) ||
3786 (!estab && sysctl_tcp_timestamps))) {
3787 opt_rx->saw_tstamp = 1;
3788 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3789 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3792 case TCPOPT_SACK_PERM:
3793 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3794 !estab && sysctl_tcp_sack) {
3795 opt_rx->sack_ok = TCP_SACK_SEEN;
3796 tcp_sack_reset(opt_rx);
3801 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3802 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3804 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3807 #ifdef CONFIG_TCP_MD5SIG
3810 * The MD5 Hash has already been
3811 * checked (see tcp_v{4,6}_do_rcv()).
3816 /* This option is variable length.
3819 case TCPOLEN_COOKIE_BASE:
3820 /* not yet implemented */
3822 case TCPOLEN_COOKIE_PAIR:
3823 /* not yet implemented */
3825 case TCPOLEN_COOKIE_MIN+0:
3826 case TCPOLEN_COOKIE_MIN+2:
3827 case TCPOLEN_COOKIE_MIN+4:
3828 case TCPOLEN_COOKIE_MIN+6:
3829 case TCPOLEN_COOKIE_MAX:
3830 /* 16-bit multiple */
3831 opt_rx->cookie_plus = opsize;
3846 EXPORT_SYMBOL(tcp_parse_options);
3848 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3850 const __be32 *ptr = (const __be32 *)(th + 1);
3852 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3853 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3854 tp->rx_opt.saw_tstamp = 1;
3856 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3858 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3864 /* Fast parse options. This hopes to only see timestamps.
3865 * If it is wrong it falls back on tcp_parse_options().
3867 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3868 const struct tcphdr *th,
3869 struct tcp_sock *tp, const u8 **hvpp)
3871 /* In the spirit of fast parsing, compare doff directly to constant
3872 * values. Because equality is used, short doff can be ignored here.
3874 if (th->doff == (sizeof(*th) / 4)) {
3875 tp->rx_opt.saw_tstamp = 0;
3877 } else if (tp->rx_opt.tstamp_ok &&
3878 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3879 if (tcp_parse_aligned_timestamp(tp, th))
3882 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3886 #ifdef CONFIG_TCP_MD5SIG
3888 * Parse MD5 Signature option
3890 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3892 int length = (th->doff << 2) - sizeof(*th);
3893 const u8 *ptr = (const u8 *)(th + 1);
3895 /* If the TCP option is too short, we can short cut */
3896 if (length < TCPOLEN_MD5SIG)
3899 while (length > 0) {
3900 int opcode = *ptr++;
3911 if (opsize < 2 || opsize > length)
3913 if (opcode == TCPOPT_MD5SIG)
3914 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3921 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3924 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3926 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3927 tp->rx_opt.ts_recent_stamp = get_seconds();
3930 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3932 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3933 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3934 * extra check below makes sure this can only happen
3935 * for pure ACK frames. -DaveM
3937 * Not only, also it occurs for expired timestamps.
3940 if (tcp_paws_check(&tp->rx_opt, 0))
3941 tcp_store_ts_recent(tp);
3945 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3947 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3948 * it can pass through stack. So, the following predicate verifies that
3949 * this segment is not used for anything but congestion avoidance or
3950 * fast retransmit. Moreover, we even are able to eliminate most of such
3951 * second order effects, if we apply some small "replay" window (~RTO)
3952 * to timestamp space.
3954 * All these measures still do not guarantee that we reject wrapped ACKs
3955 * on networks with high bandwidth, when sequence space is recycled fastly,
3956 * but it guarantees that such events will be very rare and do not affect
3957 * connection seriously. This doesn't look nice, but alas, PAWS is really
3960 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3961 * states that events when retransmit arrives after original data are rare.
3962 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3963 * the biggest problem on large power networks even with minor reordering.
3964 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3965 * up to bandwidth of 18Gigabit/sec. 8) ]
3968 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3970 const struct tcp_sock *tp = tcp_sk(sk);
3971 const struct tcphdr *th = tcp_hdr(skb);
3972 u32 seq = TCP_SKB_CB(skb)->seq;
3973 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3975 return (/* 1. Pure ACK with correct sequence number. */
3976 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3978 /* 2. ... and duplicate ACK. */
3979 ack == tp->snd_una &&
3981 /* 3. ... and does not update window. */
3982 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3984 /* 4. ... and sits in replay window. */
3985 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3988 static inline int tcp_paws_discard(const struct sock *sk,
3989 const struct sk_buff *skb)
3991 const struct tcp_sock *tp = tcp_sk(sk);
3993 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3994 !tcp_disordered_ack(sk, skb);
3997 /* Check segment sequence number for validity.
3999 * Segment controls are considered valid, if the segment
4000 * fits to the window after truncation to the window. Acceptability
4001 * of data (and SYN, FIN, of course) is checked separately.
4002 * See tcp_data_queue(), for example.
4004 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4005 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4006 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4007 * (borrowed from freebsd)
4010 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4012 return !before(end_seq, tp->rcv_wup) &&
4013 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4016 /* When we get a reset we do this. */
4017 static void tcp_reset(struct sock *sk)
4019 /* We want the right error as BSD sees it (and indeed as we do). */
4020 switch (sk->sk_state) {
4022 sk->sk_err = ECONNREFUSED;
4024 case TCP_CLOSE_WAIT:
4030 sk->sk_err = ECONNRESET;
4032 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4035 if (!sock_flag(sk, SOCK_DEAD))
4036 sk->sk_error_report(sk);
4042 * Process the FIN bit. This now behaves as it is supposed to work
4043 * and the FIN takes effect when it is validly part of sequence
4044 * space. Not before when we get holes.
4046 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4047 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4050 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4051 * close and we go into CLOSING (and later onto TIME-WAIT)
4053 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4055 static void tcp_fin(struct sock *sk)
4057 struct tcp_sock *tp = tcp_sk(sk);
4059 inet_csk_schedule_ack(sk);
4061 sk->sk_shutdown |= RCV_SHUTDOWN;
4062 sock_set_flag(sk, SOCK_DONE);
4064 switch (sk->sk_state) {
4066 case TCP_ESTABLISHED:
4067 /* Move to CLOSE_WAIT */
4068 tcp_set_state(sk, TCP_CLOSE_WAIT);
4069 inet_csk(sk)->icsk_ack.pingpong = 1;
4072 case TCP_CLOSE_WAIT:
4074 /* Received a retransmission of the FIN, do
4079 /* RFC793: Remain in the LAST-ACK state. */
4083 /* This case occurs when a simultaneous close
4084 * happens, we must ack the received FIN and
4085 * enter the CLOSING state.
4088 tcp_set_state(sk, TCP_CLOSING);
4091 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4093 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4096 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4097 * cases we should never reach this piece of code.
4099 pr_err("%s: Impossible, sk->sk_state=%d\n",
4100 __func__, sk->sk_state);
4104 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4105 * Probably, we should reset in this case. For now drop them.
4107 __skb_queue_purge(&tp->out_of_order_queue);
4108 if (tcp_is_sack(tp))
4109 tcp_sack_reset(&tp->rx_opt);
4112 if (!sock_flag(sk, SOCK_DEAD)) {
4113 sk->sk_state_change(sk);
4115 /* Do not send POLL_HUP for half duplex close. */
4116 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4117 sk->sk_state == TCP_CLOSE)
4118 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4120 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4124 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4127 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4128 if (before(seq, sp->start_seq))
4129 sp->start_seq = seq;
4130 if (after(end_seq, sp->end_seq))
4131 sp->end_seq = end_seq;
4137 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4139 struct tcp_sock *tp = tcp_sk(sk);
4141 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4144 if (before(seq, tp->rcv_nxt))
4145 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4147 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4149 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4151 tp->rx_opt.dsack = 1;
4152 tp->duplicate_sack[0].start_seq = seq;
4153 tp->duplicate_sack[0].end_seq = end_seq;
4157 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4159 struct tcp_sock *tp = tcp_sk(sk);
4161 if (!tp->rx_opt.dsack)
4162 tcp_dsack_set(sk, seq, end_seq);
4164 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4167 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4169 struct tcp_sock *tp = tcp_sk(sk);
4171 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4172 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4173 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4174 tcp_enter_quickack_mode(sk);
4176 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4177 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4179 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4180 end_seq = tp->rcv_nxt;
4181 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4188 /* These routines update the SACK block as out-of-order packets arrive or
4189 * in-order packets close up the sequence space.
4191 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4194 struct tcp_sack_block *sp = &tp->selective_acks[0];
4195 struct tcp_sack_block *swalk = sp + 1;
4197 /* See if the recent change to the first SACK eats into
4198 * or hits the sequence space of other SACK blocks, if so coalesce.
4200 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4201 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4204 /* Zap SWALK, by moving every further SACK up by one slot.
4205 * Decrease num_sacks.
4207 tp->rx_opt.num_sacks--;
4208 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4212 this_sack++, swalk++;
4216 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4218 struct tcp_sock *tp = tcp_sk(sk);
4219 struct tcp_sack_block *sp = &tp->selective_acks[0];
4220 int cur_sacks = tp->rx_opt.num_sacks;
4226 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4227 if (tcp_sack_extend(sp, seq, end_seq)) {
4228 /* Rotate this_sack to the first one. */
4229 for (; this_sack > 0; this_sack--, sp--)
4230 swap(*sp, *(sp - 1));
4232 tcp_sack_maybe_coalesce(tp);
4237 /* Could not find an adjacent existing SACK, build a new one,
4238 * put it at the front, and shift everyone else down. We
4239 * always know there is at least one SACK present already here.
4241 * If the sack array is full, forget about the last one.
4243 if (this_sack >= TCP_NUM_SACKS) {
4245 tp->rx_opt.num_sacks--;
4248 for (; this_sack > 0; this_sack--, sp--)
4252 /* Build the new head SACK, and we're done. */
4253 sp->start_seq = seq;
4254 sp->end_seq = end_seq;
4255 tp->rx_opt.num_sacks++;
4258 /* RCV.NXT advances, some SACKs should be eaten. */
4260 static void tcp_sack_remove(struct tcp_sock *tp)
4262 struct tcp_sack_block *sp = &tp->selective_acks[0];
4263 int num_sacks = tp->rx_opt.num_sacks;
4266 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4267 if (skb_queue_empty(&tp->out_of_order_queue)) {
4268 tp->rx_opt.num_sacks = 0;
4272 for (this_sack = 0; this_sack < num_sacks;) {
4273 /* Check if the start of the sack is covered by RCV.NXT. */
4274 if (!before(tp->rcv_nxt, sp->start_seq)) {
4277 /* RCV.NXT must cover all the block! */
4278 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4280 /* Zap this SACK, by moving forward any other SACKS. */
4281 for (i=this_sack+1; i < num_sacks; i++)
4282 tp->selective_acks[i-1] = tp->selective_acks[i];
4289 tp->rx_opt.num_sacks = num_sacks;
4292 /* This one checks to see if we can put data from the
4293 * out_of_order queue into the receive_queue.
4295 static void tcp_ofo_queue(struct sock *sk)
4297 struct tcp_sock *tp = tcp_sk(sk);
4298 __u32 dsack_high = tp->rcv_nxt;
4299 struct sk_buff *skb;
4301 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4302 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4305 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4306 __u32 dsack = dsack_high;
4307 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4308 dsack_high = TCP_SKB_CB(skb)->end_seq;
4309 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4312 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4313 SOCK_DEBUG(sk, "ofo packet was already received\n");
4314 __skb_unlink(skb, &tp->out_of_order_queue);
4318 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4319 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4320 TCP_SKB_CB(skb)->end_seq);
4322 __skb_unlink(skb, &tp->out_of_order_queue);
4323 __skb_queue_tail(&sk->sk_receive_queue, skb);
4324 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4325 if (tcp_hdr(skb)->fin)
4330 static bool tcp_prune_ofo_queue(struct sock *sk);
4331 static int tcp_prune_queue(struct sock *sk);
4333 static int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4335 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4336 !sk_rmem_schedule(sk, size)) {
4338 if (tcp_prune_queue(sk) < 0)
4341 if (!sk_rmem_schedule(sk, size)) {
4342 if (!tcp_prune_ofo_queue(sk))
4345 if (!sk_rmem_schedule(sk, size))
4353 * tcp_try_coalesce - try to merge skb to prior one
4356 * @from: buffer to add in queue
4357 * @fragstolen: pointer to boolean
4359 * Before queueing skb @from after @to, try to merge them
4360 * to reduce overall memory use and queue lengths, if cost is small.
4361 * Packets in ofo or receive queues can stay a long time.
4362 * Better try to coalesce them right now to avoid future collapses.
4363 * Returns true if caller should free @from instead of queueing it
4365 static bool tcp_try_coalesce(struct sock *sk,
4367 struct sk_buff *from,
4372 *fragstolen = false;
4374 if (tcp_hdr(from)->fin)
4377 /* Its possible this segment overlaps with prior segment in queue */
4378 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4381 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4384 atomic_add(delta, &sk->sk_rmem_alloc);
4385 sk_mem_charge(sk, delta);
4386 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4387 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4388 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4392 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4394 struct tcp_sock *tp = tcp_sk(sk);
4395 struct sk_buff *skb1;
4398 TCP_ECN_check_ce(tp, skb);
4400 if (unlikely(tcp_try_rmem_schedule(sk, skb->truesize))) {
4401 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4406 /* Disable header prediction. */
4408 inet_csk_schedule_ack(sk);
4410 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4411 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4412 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4414 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4416 /* Initial out of order segment, build 1 SACK. */
4417 if (tcp_is_sack(tp)) {
4418 tp->rx_opt.num_sacks = 1;
4419 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4420 tp->selective_acks[0].end_seq =
4421 TCP_SKB_CB(skb)->end_seq;
4423 __skb_queue_head(&tp->out_of_order_queue, skb);
4427 seq = TCP_SKB_CB(skb)->seq;
4428 end_seq = TCP_SKB_CB(skb)->end_seq;
4430 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4433 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4434 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4436 kfree_skb_partial(skb, fragstolen);
4440 if (!tp->rx_opt.num_sacks ||
4441 tp->selective_acks[0].end_seq != seq)
4444 /* Common case: data arrive in order after hole. */
4445 tp->selective_acks[0].end_seq = end_seq;
4449 /* Find place to insert this segment. */
4451 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4453 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4457 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4460 /* Do skb overlap to previous one? */
4461 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4462 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4463 /* All the bits are present. Drop. */
4464 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4467 tcp_dsack_set(sk, seq, end_seq);
4470 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4471 /* Partial overlap. */
4472 tcp_dsack_set(sk, seq,
4473 TCP_SKB_CB(skb1)->end_seq);
4475 if (skb_queue_is_first(&tp->out_of_order_queue,
4479 skb1 = skb_queue_prev(
4480 &tp->out_of_order_queue,
4485 __skb_queue_head(&tp->out_of_order_queue, skb);
4487 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4489 /* And clean segments covered by new one as whole. */
4490 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4491 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4493 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4495 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4496 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4500 __skb_unlink(skb1, &tp->out_of_order_queue);
4501 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4502 TCP_SKB_CB(skb1)->end_seq);
4503 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4508 if (tcp_is_sack(tp))
4509 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4512 skb_set_owner_r(skb, sk);
4515 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4519 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4521 __skb_pull(skb, hdrlen);
4523 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4524 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4526 __skb_queue_tail(&sk->sk_receive_queue, skb);
4527 skb_set_owner_r(skb, sk);
4532 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4534 struct sk_buff *skb;
4538 if (tcp_try_rmem_schedule(sk, size + sizeof(*th)))
4541 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4545 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4546 skb_reset_transport_header(skb);
4547 memset(th, 0, sizeof(*th));
4549 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4552 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4553 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4554 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4556 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4557 WARN_ON_ONCE(fragstolen); /* should not happen */
4568 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4570 const struct tcphdr *th = tcp_hdr(skb);
4571 struct tcp_sock *tp = tcp_sk(sk);
4573 bool fragstolen = false;
4575 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4579 __skb_pull(skb, th->doff * 4);
4581 TCP_ECN_accept_cwr(tp, skb);
4583 tp->rx_opt.dsack = 0;
4585 /* Queue data for delivery to the user.
4586 * Packets in sequence go to the receive queue.
4587 * Out of sequence packets to the out_of_order_queue.
4589 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4590 if (tcp_receive_window(tp) == 0)
4593 /* Ok. In sequence. In window. */
4594 if (tp->ucopy.task == current &&
4595 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4596 sock_owned_by_user(sk) && !tp->urg_data) {
4597 int chunk = min_t(unsigned int, skb->len,
4600 __set_current_state(TASK_RUNNING);
4603 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4604 tp->ucopy.len -= chunk;
4605 tp->copied_seq += chunk;
4606 eaten = (chunk == skb->len);
4607 tcp_rcv_space_adjust(sk);
4615 tcp_try_rmem_schedule(sk, skb->truesize))
4618 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4620 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4622 tcp_event_data_recv(sk, skb);
4626 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4629 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4630 * gap in queue is filled.
4632 if (skb_queue_empty(&tp->out_of_order_queue))
4633 inet_csk(sk)->icsk_ack.pingpong = 0;
4636 if (tp->rx_opt.num_sacks)
4637 tcp_sack_remove(tp);
4639 tcp_fast_path_check(sk);
4642 kfree_skb_partial(skb, fragstolen);
4643 else if (!sock_flag(sk, SOCK_DEAD))
4644 sk->sk_data_ready(sk, 0);
4648 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4649 /* A retransmit, 2nd most common case. Force an immediate ack. */
4650 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4651 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4654 tcp_enter_quickack_mode(sk);
4655 inet_csk_schedule_ack(sk);
4661 /* Out of window. F.e. zero window probe. */
4662 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4665 tcp_enter_quickack_mode(sk);
4667 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4668 /* Partial packet, seq < rcv_next < end_seq */
4669 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4670 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4671 TCP_SKB_CB(skb)->end_seq);
4673 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4675 /* If window is closed, drop tail of packet. But after
4676 * remembering D-SACK for its head made in previous line.
4678 if (!tcp_receive_window(tp))
4683 tcp_data_queue_ofo(sk, skb);
4686 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4687 struct sk_buff_head *list)
4689 struct sk_buff *next = NULL;
4691 if (!skb_queue_is_last(list, skb))
4692 next = skb_queue_next(list, skb);
4694 __skb_unlink(skb, list);
4696 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4701 /* Collapse contiguous sequence of skbs head..tail with
4702 * sequence numbers start..end.
4704 * If tail is NULL, this means until the end of the list.
4706 * Segments with FIN/SYN are not collapsed (only because this
4710 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4711 struct sk_buff *head, struct sk_buff *tail,
4714 struct sk_buff *skb, *n;
4717 /* First, check that queue is collapsible and find
4718 * the point where collapsing can be useful. */
4722 skb_queue_walk_from_safe(list, skb, n) {
4725 /* No new bits? It is possible on ofo queue. */
4726 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4727 skb = tcp_collapse_one(sk, skb, list);
4733 /* The first skb to collapse is:
4735 * - bloated or contains data before "start" or
4736 * overlaps to the next one.
4738 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4739 (tcp_win_from_space(skb->truesize) > skb->len ||
4740 before(TCP_SKB_CB(skb)->seq, start))) {
4741 end_of_skbs = false;
4745 if (!skb_queue_is_last(list, skb)) {
4746 struct sk_buff *next = skb_queue_next(list, skb);
4748 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4749 end_of_skbs = false;
4754 /* Decided to skip this, advance start seq. */
4755 start = TCP_SKB_CB(skb)->end_seq;
4757 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4760 while (before(start, end)) {
4761 struct sk_buff *nskb;
4762 unsigned int header = skb_headroom(skb);
4763 int copy = SKB_MAX_ORDER(header, 0);
4765 /* Too big header? This can happen with IPv6. */
4768 if (end - start < copy)
4770 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4774 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4775 skb_set_network_header(nskb, (skb_network_header(skb) -
4777 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4779 skb_reserve(nskb, header);
4780 memcpy(nskb->head, skb->head, header);
4781 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4782 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4783 __skb_queue_before(list, skb, nskb);
4784 skb_set_owner_r(nskb, sk);
4786 /* Copy data, releasing collapsed skbs. */
4788 int offset = start - TCP_SKB_CB(skb)->seq;
4789 int size = TCP_SKB_CB(skb)->end_seq - start;
4793 size = min(copy, size);
4794 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4796 TCP_SKB_CB(nskb)->end_seq += size;
4800 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4801 skb = tcp_collapse_one(sk, skb, list);
4804 tcp_hdr(skb)->syn ||
4812 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4813 * and tcp_collapse() them until all the queue is collapsed.
4815 static void tcp_collapse_ofo_queue(struct sock *sk)
4817 struct tcp_sock *tp = tcp_sk(sk);
4818 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4819 struct sk_buff *head;
4825 start = TCP_SKB_CB(skb)->seq;
4826 end = TCP_SKB_CB(skb)->end_seq;
4830 struct sk_buff *next = NULL;
4832 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4833 next = skb_queue_next(&tp->out_of_order_queue, skb);
4836 /* Segment is terminated when we see gap or when
4837 * we are at the end of all the queue. */
4839 after(TCP_SKB_CB(skb)->seq, end) ||
4840 before(TCP_SKB_CB(skb)->end_seq, start)) {
4841 tcp_collapse(sk, &tp->out_of_order_queue,
4842 head, skb, start, end);
4846 /* Start new segment */
4847 start = TCP_SKB_CB(skb)->seq;
4848 end = TCP_SKB_CB(skb)->end_seq;
4850 if (before(TCP_SKB_CB(skb)->seq, start))
4851 start = TCP_SKB_CB(skb)->seq;
4852 if (after(TCP_SKB_CB(skb)->end_seq, end))
4853 end = TCP_SKB_CB(skb)->end_seq;
4859 * Purge the out-of-order queue.
4860 * Return true if queue was pruned.
4862 static bool tcp_prune_ofo_queue(struct sock *sk)
4864 struct tcp_sock *tp = tcp_sk(sk);
4867 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4868 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4869 __skb_queue_purge(&tp->out_of_order_queue);
4871 /* Reset SACK state. A conforming SACK implementation will
4872 * do the same at a timeout based retransmit. When a connection
4873 * is in a sad state like this, we care only about integrity
4874 * of the connection not performance.
4876 if (tp->rx_opt.sack_ok)
4877 tcp_sack_reset(&tp->rx_opt);
4884 /* Reduce allocated memory if we can, trying to get
4885 * the socket within its memory limits again.
4887 * Return less than zero if we should start dropping frames
4888 * until the socket owning process reads some of the data
4889 * to stabilize the situation.
4891 static int tcp_prune_queue(struct sock *sk)
4893 struct tcp_sock *tp = tcp_sk(sk);
4895 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4897 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4899 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4900 tcp_clamp_window(sk);
4901 else if (sk_under_memory_pressure(sk))
4902 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4904 tcp_collapse_ofo_queue(sk);
4905 if (!skb_queue_empty(&sk->sk_receive_queue))
4906 tcp_collapse(sk, &sk->sk_receive_queue,
4907 skb_peek(&sk->sk_receive_queue),
4909 tp->copied_seq, tp->rcv_nxt);
4912 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4915 /* Collapsing did not help, destructive actions follow.
4916 * This must not ever occur. */
4918 tcp_prune_ofo_queue(sk);
4920 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4923 /* If we are really being abused, tell the caller to silently
4924 * drop receive data on the floor. It will get retransmitted
4925 * and hopefully then we'll have sufficient space.
4927 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4929 /* Massive buffer overcommit. */
4934 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4935 * As additional protections, we do not touch cwnd in retransmission phases,
4936 * and if application hit its sndbuf limit recently.
4938 void tcp_cwnd_application_limited(struct sock *sk)
4940 struct tcp_sock *tp = tcp_sk(sk);
4942 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4943 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4944 /* Limited by application or receiver window. */
4945 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4946 u32 win_used = max(tp->snd_cwnd_used, init_win);
4947 if (win_used < tp->snd_cwnd) {
4948 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4949 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4951 tp->snd_cwnd_used = 0;
4953 tp->snd_cwnd_stamp = tcp_time_stamp;
4956 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4958 const struct tcp_sock *tp = tcp_sk(sk);
4960 /* If the user specified a specific send buffer setting, do
4963 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4966 /* If we are under global TCP memory pressure, do not expand. */
4967 if (sk_under_memory_pressure(sk))
4970 /* If we are under soft global TCP memory pressure, do not expand. */
4971 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4974 /* If we filled the congestion window, do not expand. */
4975 if (tp->packets_out >= tp->snd_cwnd)
4981 /* When incoming ACK allowed to free some skb from write_queue,
4982 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4983 * on the exit from tcp input handler.
4985 * PROBLEM: sndbuf expansion does not work well with largesend.
4987 static void tcp_new_space(struct sock *sk)
4989 struct tcp_sock *tp = tcp_sk(sk);
4991 if (tcp_should_expand_sndbuf(sk)) {
4992 int sndmem = SKB_TRUESIZE(max_t(u32,
4993 tp->rx_opt.mss_clamp,
4996 int demanded = max_t(unsigned int, tp->snd_cwnd,
4997 tp->reordering + 1);
4998 sndmem *= 2 * demanded;
4999 if (sndmem > sk->sk_sndbuf)
5000 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5001 tp->snd_cwnd_stamp = tcp_time_stamp;
5004 sk->sk_write_space(sk);
5007 static void tcp_check_space(struct sock *sk)
5009 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5010 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5011 if (sk->sk_socket &&
5012 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5017 static inline void tcp_data_snd_check(struct sock *sk)
5019 tcp_push_pending_frames(sk);
5020 tcp_check_space(sk);
5024 * Check if sending an ack is needed.
5026 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5028 struct tcp_sock *tp = tcp_sk(sk);
5030 /* More than one full frame received... */
5031 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5032 /* ... and right edge of window advances far enough.
5033 * (tcp_recvmsg() will send ACK otherwise). Or...
5035 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5036 /* We ACK each frame or... */
5037 tcp_in_quickack_mode(sk) ||
5038 /* We have out of order data. */
5039 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5040 /* Then ack it now */
5043 /* Else, send delayed ack. */
5044 tcp_send_delayed_ack(sk);
5048 static inline void tcp_ack_snd_check(struct sock *sk)
5050 if (!inet_csk_ack_scheduled(sk)) {
5051 /* We sent a data segment already. */
5054 __tcp_ack_snd_check(sk, 1);
5058 * This routine is only called when we have urgent data
5059 * signaled. Its the 'slow' part of tcp_urg. It could be
5060 * moved inline now as tcp_urg is only called from one
5061 * place. We handle URGent data wrong. We have to - as
5062 * BSD still doesn't use the correction from RFC961.
5063 * For 1003.1g we should support a new option TCP_STDURG to permit
5064 * either form (or just set the sysctl tcp_stdurg).
5067 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5069 struct tcp_sock *tp = tcp_sk(sk);
5070 u32 ptr = ntohs(th->urg_ptr);
5072 if (ptr && !sysctl_tcp_stdurg)
5074 ptr += ntohl(th->seq);
5076 /* Ignore urgent data that we've already seen and read. */
5077 if (after(tp->copied_seq, ptr))
5080 /* Do not replay urg ptr.
5082 * NOTE: interesting situation not covered by specs.
5083 * Misbehaving sender may send urg ptr, pointing to segment,
5084 * which we already have in ofo queue. We are not able to fetch
5085 * such data and will stay in TCP_URG_NOTYET until will be eaten
5086 * by recvmsg(). Seems, we are not obliged to handle such wicked
5087 * situations. But it is worth to think about possibility of some
5088 * DoSes using some hypothetical application level deadlock.
5090 if (before(ptr, tp->rcv_nxt))
5093 /* Do we already have a newer (or duplicate) urgent pointer? */
5094 if (tp->urg_data && !after(ptr, tp->urg_seq))
5097 /* Tell the world about our new urgent pointer. */
5100 /* We may be adding urgent data when the last byte read was
5101 * urgent. To do this requires some care. We cannot just ignore
5102 * tp->copied_seq since we would read the last urgent byte again
5103 * as data, nor can we alter copied_seq until this data arrives
5104 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5106 * NOTE. Double Dutch. Rendering to plain English: author of comment
5107 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5108 * and expect that both A and B disappear from stream. This is _wrong_.
5109 * Though this happens in BSD with high probability, this is occasional.
5110 * Any application relying on this is buggy. Note also, that fix "works"
5111 * only in this artificial test. Insert some normal data between A and B and we will
5112 * decline of BSD again. Verdict: it is better to remove to trap
5115 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5116 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5117 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5119 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5120 __skb_unlink(skb, &sk->sk_receive_queue);
5125 tp->urg_data = TCP_URG_NOTYET;
5128 /* Disable header prediction. */
5132 /* This is the 'fast' part of urgent handling. */
5133 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5135 struct tcp_sock *tp = tcp_sk(sk);
5137 /* Check if we get a new urgent pointer - normally not. */
5139 tcp_check_urg(sk, th);
5141 /* Do we wait for any urgent data? - normally not... */
5142 if (tp->urg_data == TCP_URG_NOTYET) {
5143 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5146 /* Is the urgent pointer pointing into this packet? */
5147 if (ptr < skb->len) {
5149 if (skb_copy_bits(skb, ptr, &tmp, 1))
5151 tp->urg_data = TCP_URG_VALID | tmp;
5152 if (!sock_flag(sk, SOCK_DEAD))
5153 sk->sk_data_ready(sk, 0);
5158 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5160 struct tcp_sock *tp = tcp_sk(sk);
5161 int chunk = skb->len - hlen;
5165 if (skb_csum_unnecessary(skb))
5166 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5168 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5172 tp->ucopy.len -= chunk;
5173 tp->copied_seq += chunk;
5174 tcp_rcv_space_adjust(sk);
5181 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5182 struct sk_buff *skb)
5186 if (sock_owned_by_user(sk)) {
5188 result = __tcp_checksum_complete(skb);
5191 result = __tcp_checksum_complete(skb);
5196 static inline int tcp_checksum_complete_user(struct sock *sk,
5197 struct sk_buff *skb)
5199 return !skb_csum_unnecessary(skb) &&
5200 __tcp_checksum_complete_user(sk, skb);
5203 #ifdef CONFIG_NET_DMA
5204 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5207 struct tcp_sock *tp = tcp_sk(sk);
5208 int chunk = skb->len - hlen;
5210 bool copied_early = false;
5212 if (tp->ucopy.wakeup)
5215 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5216 tp->ucopy.dma_chan = net_dma_find_channel();
5218 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5220 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5222 tp->ucopy.iov, chunk,
5223 tp->ucopy.pinned_list);
5228 tp->ucopy.dma_cookie = dma_cookie;
5229 copied_early = true;
5231 tp->ucopy.len -= chunk;
5232 tp->copied_seq += chunk;
5233 tcp_rcv_space_adjust(sk);
5235 if ((tp->ucopy.len == 0) ||
5236 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5237 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5238 tp->ucopy.wakeup = 1;
5239 sk->sk_data_ready(sk, 0);
5241 } else if (chunk > 0) {
5242 tp->ucopy.wakeup = 1;
5243 sk->sk_data_ready(sk, 0);
5246 return copied_early;
5248 #endif /* CONFIG_NET_DMA */
5250 /* Does PAWS and seqno based validation of an incoming segment, flags will
5251 * play significant role here.
5253 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5254 const struct tcphdr *th, int syn_inerr)
5256 const u8 *hash_location;
5257 struct tcp_sock *tp = tcp_sk(sk);
5259 /* RFC1323: H1. Apply PAWS check first. */
5260 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5261 tp->rx_opt.saw_tstamp &&
5262 tcp_paws_discard(sk, skb)) {
5264 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5265 tcp_send_dupack(sk, skb);
5268 /* Reset is accepted even if it did not pass PAWS. */
5271 /* Step 1: check sequence number */
5272 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5273 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5274 * (RST) segments are validated by checking their SEQ-fields."
5275 * And page 69: "If an incoming segment is not acceptable,
5276 * an acknowledgment should be sent in reply (unless the RST
5277 * bit is set, if so drop the segment and return)".
5280 tcp_send_dupack(sk, skb);
5284 /* Step 2: check RST bit */
5290 /* ts_recent update must be made after we are sure that the packet
5293 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5295 /* step 3: check security and precedence [ignored] */
5297 /* step 4: Check for a SYN in window. */
5298 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5300 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5301 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5314 * TCP receive function for the ESTABLISHED state.
5316 * It is split into a fast path and a slow path. The fast path is
5318 * - A zero window was announced from us - zero window probing
5319 * is only handled properly in the slow path.
5320 * - Out of order segments arrived.
5321 * - Urgent data is expected.
5322 * - There is no buffer space left
5323 * - Unexpected TCP flags/window values/header lengths are received
5324 * (detected by checking the TCP header against pred_flags)
5325 * - Data is sent in both directions. Fast path only supports pure senders
5326 * or pure receivers (this means either the sequence number or the ack
5327 * value must stay constant)
5328 * - Unexpected TCP option.
5330 * When these conditions are not satisfied it drops into a standard
5331 * receive procedure patterned after RFC793 to handle all cases.
5332 * The first three cases are guaranteed by proper pred_flags setting,
5333 * the rest is checked inline. Fast processing is turned on in
5334 * tcp_data_queue when everything is OK.
5336 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5337 const struct tcphdr *th, unsigned int len)
5339 struct tcp_sock *tp = tcp_sk(sk);
5342 if (sk->sk_rx_dst) {
5343 struct dst_entry *dst = sk->sk_rx_dst;
5344 if (unlikely(dst->obsolete)) {
5345 if (dst->ops->check(dst, 0) == NULL) {
5347 sk->sk_rx_dst = NULL;
5351 if (unlikely(sk->sk_rx_dst == NULL))
5352 sk->sk_rx_dst = dst_clone(skb_dst(skb));
5355 * Header prediction.
5356 * The code loosely follows the one in the famous
5357 * "30 instruction TCP receive" Van Jacobson mail.
5359 * Van's trick is to deposit buffers into socket queue
5360 * on a device interrupt, to call tcp_recv function
5361 * on the receive process context and checksum and copy
5362 * the buffer to user space. smart...
5364 * Our current scheme is not silly either but we take the
5365 * extra cost of the net_bh soft interrupt processing...
5366 * We do checksum and copy also but from device to kernel.
5369 tp->rx_opt.saw_tstamp = 0;
5371 /* pred_flags is 0xS?10 << 16 + snd_wnd
5372 * if header_prediction is to be made
5373 * 'S' will always be tp->tcp_header_len >> 2
5374 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5375 * turn it off (when there are holes in the receive
5376 * space for instance)
5377 * PSH flag is ignored.
5380 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5381 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5382 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5383 int tcp_header_len = tp->tcp_header_len;
5385 /* Timestamp header prediction: tcp_header_len
5386 * is automatically equal to th->doff*4 due to pred_flags
5390 /* Check timestamp */
5391 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5392 /* No? Slow path! */
5393 if (!tcp_parse_aligned_timestamp(tp, th))
5396 /* If PAWS failed, check it more carefully in slow path */
5397 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5400 /* DO NOT update ts_recent here, if checksum fails
5401 * and timestamp was corrupted part, it will result
5402 * in a hung connection since we will drop all
5403 * future packets due to the PAWS test.
5407 if (len <= tcp_header_len) {
5408 /* Bulk data transfer: sender */
5409 if (len == tcp_header_len) {
5410 /* Predicted packet is in window by definition.
5411 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5412 * Hence, check seq<=rcv_wup reduces to:
5414 if (tcp_header_len ==
5415 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5416 tp->rcv_nxt == tp->rcv_wup)
5417 tcp_store_ts_recent(tp);
5419 /* We know that such packets are checksummed
5422 tcp_ack(sk, skb, 0);
5424 tcp_data_snd_check(sk);
5426 } else { /* Header too small */
5427 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5432 int copied_early = 0;
5433 bool fragstolen = false;
5435 if (tp->copied_seq == tp->rcv_nxt &&
5436 len - tcp_header_len <= tp->ucopy.len) {
5437 #ifdef CONFIG_NET_DMA
5438 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5443 if (tp->ucopy.task == current &&
5444 sock_owned_by_user(sk) && !copied_early) {
5445 __set_current_state(TASK_RUNNING);
5447 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5451 /* Predicted packet is in window by definition.
5452 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5453 * Hence, check seq<=rcv_wup reduces to:
5455 if (tcp_header_len ==
5456 (sizeof(struct tcphdr) +
5457 TCPOLEN_TSTAMP_ALIGNED) &&
5458 tp->rcv_nxt == tp->rcv_wup)
5459 tcp_store_ts_recent(tp);
5461 tcp_rcv_rtt_measure_ts(sk, skb);
5463 __skb_pull(skb, tcp_header_len);
5464 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5465 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5468 tcp_cleanup_rbuf(sk, skb->len);
5471 if (tcp_checksum_complete_user(sk, skb))
5474 /* Predicted packet is in window by definition.
5475 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5476 * Hence, check seq<=rcv_wup reduces to:
5478 if (tcp_header_len ==
5479 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5480 tp->rcv_nxt == tp->rcv_wup)
5481 tcp_store_ts_recent(tp);
5483 tcp_rcv_rtt_measure_ts(sk, skb);
5485 if ((int)skb->truesize > sk->sk_forward_alloc)
5488 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5490 /* Bulk data transfer: receiver */
5491 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5495 tcp_event_data_recv(sk, skb);
5497 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5498 /* Well, only one small jumplet in fast path... */
5499 tcp_ack(sk, skb, FLAG_DATA);
5500 tcp_data_snd_check(sk);
5501 if (!inet_csk_ack_scheduled(sk))
5505 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5506 __tcp_ack_snd_check(sk, 0);
5508 #ifdef CONFIG_NET_DMA
5510 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5514 kfree_skb_partial(skb, fragstolen);
5516 sk->sk_data_ready(sk, 0);
5522 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5526 * Standard slow path.
5529 res = tcp_validate_incoming(sk, skb, th, 1);
5534 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5537 tcp_rcv_rtt_measure_ts(sk, skb);
5539 /* Process urgent data. */
5540 tcp_urg(sk, skb, th);
5542 /* step 7: process the segment text */
5543 tcp_data_queue(sk, skb);
5545 tcp_data_snd_check(sk);
5546 tcp_ack_snd_check(sk);
5550 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5556 EXPORT_SYMBOL(tcp_rcv_established);
5558 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5560 struct tcp_sock *tp = tcp_sk(sk);
5561 struct inet_connection_sock *icsk = inet_csk(sk);
5563 tcp_set_state(sk, TCP_ESTABLISHED);
5566 sk->sk_rx_dst = dst_clone(skb_dst(skb));
5567 security_inet_conn_established(sk, skb);
5570 /* Make sure socket is routed, for correct metrics. */
5571 icsk->icsk_af_ops->rebuild_header(sk);
5573 tcp_init_metrics(sk);
5575 tcp_init_congestion_control(sk);
5577 /* Prevent spurious tcp_cwnd_restart() on first data
5580 tp->lsndtime = tcp_time_stamp;
5582 tcp_init_buffer_space(sk);
5584 if (sock_flag(sk, SOCK_KEEPOPEN))
5585 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5587 if (!tp->rx_opt.snd_wscale)
5588 __tcp_fast_path_on(tp, tp->snd_wnd);
5592 if (!sock_flag(sk, SOCK_DEAD)) {
5593 sk->sk_state_change(sk);
5594 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5598 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5599 const struct tcphdr *th, unsigned int len)
5601 const u8 *hash_location;
5602 struct inet_connection_sock *icsk = inet_csk(sk);
5603 struct tcp_sock *tp = tcp_sk(sk);
5604 struct tcp_cookie_values *cvp = tp->cookie_values;
5605 int saved_clamp = tp->rx_opt.mss_clamp;
5607 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5611 * "If the state is SYN-SENT then
5612 * first check the ACK bit
5613 * If the ACK bit is set
5614 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5615 * a reset (unless the RST bit is set, if so drop
5616 * the segment and return)"
5618 * We do not send data with SYN, so that RFC-correct
5621 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5622 goto reset_and_undo;
5624 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5625 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5627 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5628 goto reset_and_undo;
5631 /* Now ACK is acceptable.
5633 * "If the RST bit is set
5634 * If the ACK was acceptable then signal the user "error:
5635 * connection reset", drop the segment, enter CLOSED state,
5636 * delete TCB, and return."
5645 * "fifth, if neither of the SYN or RST bits is set then
5646 * drop the segment and return."
5652 goto discard_and_undo;
5655 * "If the SYN bit is on ...
5656 * are acceptable then ...
5657 * (our SYN has been ACKed), change the connection
5658 * state to ESTABLISHED..."
5661 TCP_ECN_rcv_synack(tp, th);
5663 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5664 tcp_ack(sk, skb, FLAG_SLOWPATH);
5666 /* Ok.. it's good. Set up sequence numbers and
5667 * move to established.
5669 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5670 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5672 /* RFC1323: The window in SYN & SYN/ACK segments is
5675 tp->snd_wnd = ntohs(th->window);
5676 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5678 if (!tp->rx_opt.wscale_ok) {
5679 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5680 tp->window_clamp = min(tp->window_clamp, 65535U);
5683 if (tp->rx_opt.saw_tstamp) {
5684 tp->rx_opt.tstamp_ok = 1;
5685 tp->tcp_header_len =
5686 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5687 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5688 tcp_store_ts_recent(tp);
5690 tp->tcp_header_len = sizeof(struct tcphdr);
5693 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5694 tcp_enable_fack(tp);
5697 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5698 tcp_initialize_rcv_mss(sk);
5700 /* Remember, tcp_poll() does not lock socket!
5701 * Change state from SYN-SENT only after copied_seq
5702 * is initialized. */
5703 tp->copied_seq = tp->rcv_nxt;
5706 cvp->cookie_pair_size > 0 &&
5707 tp->rx_opt.cookie_plus > 0) {
5708 int cookie_size = tp->rx_opt.cookie_plus
5709 - TCPOLEN_COOKIE_BASE;
5710 int cookie_pair_size = cookie_size
5711 + cvp->cookie_desired;
5713 /* A cookie extension option was sent and returned.
5714 * Note that each incoming SYNACK replaces the
5715 * Responder cookie. The initial exchange is most
5716 * fragile, as protection against spoofing relies
5717 * entirely upon the sequence and timestamp (above).
5718 * This replacement strategy allows the correct pair to
5719 * pass through, while any others will be filtered via
5720 * Responder verification later.
5722 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5723 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5724 hash_location, cookie_size);
5725 cvp->cookie_pair_size = cookie_pair_size;
5731 tcp_finish_connect(sk, skb);
5733 if (sk->sk_write_pending ||
5734 icsk->icsk_accept_queue.rskq_defer_accept ||
5735 icsk->icsk_ack.pingpong) {
5736 /* Save one ACK. Data will be ready after
5737 * several ticks, if write_pending is set.
5739 * It may be deleted, but with this feature tcpdumps
5740 * look so _wonderfully_ clever, that I was not able
5741 * to stand against the temptation 8) --ANK
5743 inet_csk_schedule_ack(sk);
5744 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5745 tcp_enter_quickack_mode(sk);
5746 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5747 TCP_DELACK_MAX, TCP_RTO_MAX);
5758 /* No ACK in the segment */
5762 * "If the RST bit is set
5764 * Otherwise (no ACK) drop the segment and return."
5767 goto discard_and_undo;
5771 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5772 tcp_paws_reject(&tp->rx_opt, 0))
5773 goto discard_and_undo;
5776 /* We see SYN without ACK. It is attempt of
5777 * simultaneous connect with crossed SYNs.
5778 * Particularly, it can be connect to self.
5780 tcp_set_state(sk, TCP_SYN_RECV);
5782 if (tp->rx_opt.saw_tstamp) {
5783 tp->rx_opt.tstamp_ok = 1;
5784 tcp_store_ts_recent(tp);
5785 tp->tcp_header_len =
5786 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5788 tp->tcp_header_len = sizeof(struct tcphdr);
5791 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5792 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5794 /* RFC1323: The window in SYN & SYN/ACK segments is
5797 tp->snd_wnd = ntohs(th->window);
5798 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5799 tp->max_window = tp->snd_wnd;
5801 TCP_ECN_rcv_syn(tp, th);
5804 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5805 tcp_initialize_rcv_mss(sk);
5807 tcp_send_synack(sk);
5809 /* Note, we could accept data and URG from this segment.
5810 * There are no obstacles to make this.
5812 * However, if we ignore data in ACKless segments sometimes,
5813 * we have no reasons to accept it sometimes.
5814 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5815 * is not flawless. So, discard packet for sanity.
5816 * Uncomment this return to process the data.
5823 /* "fifth, if neither of the SYN or RST bits is set then
5824 * drop the segment and return."
5828 tcp_clear_options(&tp->rx_opt);
5829 tp->rx_opt.mss_clamp = saved_clamp;
5833 tcp_clear_options(&tp->rx_opt);
5834 tp->rx_opt.mss_clamp = saved_clamp;
5839 * This function implements the receiving procedure of RFC 793 for
5840 * all states except ESTABLISHED and TIME_WAIT.
5841 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5842 * address independent.
5845 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5846 const struct tcphdr *th, unsigned int len)
5848 struct tcp_sock *tp = tcp_sk(sk);
5849 struct inet_connection_sock *icsk = inet_csk(sk);
5853 tp->rx_opt.saw_tstamp = 0;
5855 switch (sk->sk_state) {
5869 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5872 /* Now we have several options: In theory there is
5873 * nothing else in the frame. KA9Q has an option to
5874 * send data with the syn, BSD accepts data with the
5875 * syn up to the [to be] advertised window and
5876 * Solaris 2.1 gives you a protocol error. For now
5877 * we just ignore it, that fits the spec precisely
5878 * and avoids incompatibilities. It would be nice in
5879 * future to drop through and process the data.
5881 * Now that TTCP is starting to be used we ought to
5883 * But, this leaves one open to an easy denial of
5884 * service attack, and SYN cookies can't defend
5885 * against this problem. So, we drop the data
5886 * in the interest of security over speed unless
5887 * it's still in use.
5895 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5899 /* Do step6 onward by hand. */
5900 tcp_urg(sk, skb, th);
5902 tcp_data_snd_check(sk);
5906 res = tcp_validate_incoming(sk, skb, th, 0);
5910 /* step 5: check the ACK field */
5912 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5914 switch (sk->sk_state) {
5917 tp->copied_seq = tp->rcv_nxt;
5919 tcp_set_state(sk, TCP_ESTABLISHED);
5920 sk->sk_state_change(sk);
5922 /* Note, that this wakeup is only for marginal
5923 * crossed SYN case. Passively open sockets
5924 * are not waked up, because sk->sk_sleep ==
5925 * NULL and sk->sk_socket == NULL.
5929 SOCK_WAKE_IO, POLL_OUT);
5931 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5932 tp->snd_wnd = ntohs(th->window) <<
5933 tp->rx_opt.snd_wscale;
5934 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5936 if (tp->rx_opt.tstamp_ok)
5937 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5939 /* Make sure socket is routed, for
5942 icsk->icsk_af_ops->rebuild_header(sk);
5944 tcp_init_metrics(sk);
5946 tcp_init_congestion_control(sk);
5948 /* Prevent spurious tcp_cwnd_restart() on
5949 * first data packet.
5951 tp->lsndtime = tcp_time_stamp;
5954 tcp_initialize_rcv_mss(sk);
5955 tcp_init_buffer_space(sk);
5956 tcp_fast_path_on(tp);
5963 if (tp->snd_una == tp->write_seq) {
5964 struct dst_entry *dst;
5966 tcp_set_state(sk, TCP_FIN_WAIT2);
5967 sk->sk_shutdown |= SEND_SHUTDOWN;
5969 dst = __sk_dst_get(sk);
5973 if (!sock_flag(sk, SOCK_DEAD))
5974 /* Wake up lingering close() */
5975 sk->sk_state_change(sk);
5979 if (tp->linger2 < 0 ||
5980 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5981 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5983 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5987 tmo = tcp_fin_time(sk);
5988 if (tmo > TCP_TIMEWAIT_LEN) {
5989 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5990 } else if (th->fin || sock_owned_by_user(sk)) {
5991 /* Bad case. We could lose such FIN otherwise.
5992 * It is not a big problem, but it looks confusing
5993 * and not so rare event. We still can lose it now,
5994 * if it spins in bh_lock_sock(), but it is really
5997 inet_csk_reset_keepalive_timer(sk, tmo);
5999 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6007 if (tp->snd_una == tp->write_seq) {
6008 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6014 if (tp->snd_una == tp->write_seq) {
6015 tcp_update_metrics(sk);
6024 /* step 6: check the URG bit */
6025 tcp_urg(sk, skb, th);
6027 /* step 7: process the segment text */
6028 switch (sk->sk_state) {
6029 case TCP_CLOSE_WAIT:
6032 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6036 /* RFC 793 says to queue data in these states,
6037 * RFC 1122 says we MUST send a reset.
6038 * BSD 4.4 also does reset.
6040 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6041 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6042 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6043 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6049 case TCP_ESTABLISHED:
6050 tcp_data_queue(sk, skb);
6055 /* tcp_data could move socket to TIME-WAIT */
6056 if (sk->sk_state != TCP_CLOSE) {
6057 tcp_data_snd_check(sk);
6058 tcp_ack_snd_check(sk);
6067 EXPORT_SYMBOL(tcp_rcv_state_process);