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_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 100;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
97 int sysctl_tcp_thin_dupack __read_mostly;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100 int sysctl_tcp_early_retrans __read_mostly = 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
111 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
112 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
115 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
116 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
117 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
119 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
120 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
122 /* Adapt the MSS value used to make delayed ack decision to the
125 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
127 struct inet_connection_sock *icsk = inet_csk(sk);
128 const unsigned int lss = icsk->icsk_ack.last_seg_size;
131 icsk->icsk_ack.last_seg_size = 0;
133 /* skb->len may jitter because of SACKs, even if peer
134 * sends good full-sized frames.
136 len = skb_shinfo(skb)->gso_size ? : skb->len;
137 if (len >= icsk->icsk_ack.rcv_mss) {
138 icsk->icsk_ack.rcv_mss = len;
140 /* Otherwise, we make more careful check taking into account,
141 * that SACKs block is variable.
143 * "len" is invariant segment length, including TCP header.
145 len += skb->data - skb_transport_header(skb);
146 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
147 /* If PSH is not set, packet should be
148 * full sized, provided peer TCP is not badly broken.
149 * This observation (if it is correct 8)) allows
150 * to handle super-low mtu links fairly.
152 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
153 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
154 /* Subtract also invariant (if peer is RFC compliant),
155 * tcp header plus fixed timestamp option length.
156 * Resulting "len" is MSS free of SACK jitter.
158 len -= tcp_sk(sk)->tcp_header_len;
159 icsk->icsk_ack.last_seg_size = len;
161 icsk->icsk_ack.rcv_mss = len;
165 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
166 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
167 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
171 static void tcp_incr_quickack(struct sock *sk)
173 struct inet_connection_sock *icsk = inet_csk(sk);
174 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
178 if (quickacks > icsk->icsk_ack.quick)
179 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
182 static void tcp_enter_quickack_mode(struct sock *sk)
184 struct inet_connection_sock *icsk = inet_csk(sk);
185 tcp_incr_quickack(sk);
186 icsk->icsk_ack.pingpong = 0;
187 icsk->icsk_ack.ato = TCP_ATO_MIN;
190 /* Send ACKs quickly, if "quick" count is not exhausted
191 * and the session is not interactive.
194 static inline bool tcp_in_quickack_mode(const struct sock *sk)
196 const struct inet_connection_sock *icsk = inet_csk(sk);
198 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
201 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
203 if (tp->ecn_flags & TCP_ECN_OK)
204 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
207 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
209 if (tcp_hdr(skb)->cwr)
210 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
213 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
215 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
220 if (!(tp->ecn_flags & TCP_ECN_OK))
223 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
224 case INET_ECN_NOT_ECT:
225 /* Funny extension: if ECT is not set on a segment,
226 * and we already seen ECT on a previous segment,
227 * it is probably a retransmit.
229 if (tp->ecn_flags & TCP_ECN_SEEN)
230 tcp_enter_quickack_mode((struct sock *)tp);
233 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
234 /* Better not delay acks, sender can have a very low cwnd */
235 tcp_enter_quickack_mode((struct sock *)tp);
236 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 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);
740 * Packet counting of FACK is based on in-order assumptions, therefore TCP
741 * disables it when reordering is detected
743 void tcp_disable_fack(struct tcp_sock *tp)
745 /* RFC3517 uses different metric in lost marker => reset on change */
747 tp->lost_skb_hint = NULL;
748 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
751 /* Take a notice that peer is sending D-SACKs */
752 static void tcp_dsack_seen(struct tcp_sock *tp)
754 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
757 static void tcp_update_reordering(struct sock *sk, const int metric,
760 struct tcp_sock *tp = tcp_sk(sk);
761 if (metric > tp->reordering) {
764 tp->reordering = min(TCP_MAX_REORDERING, metric);
766 /* This exciting event is worth to be remembered. 8) */
768 mib_idx = LINUX_MIB_TCPTSREORDER;
769 else if (tcp_is_reno(tp))
770 mib_idx = LINUX_MIB_TCPRENOREORDER;
771 else if (tcp_is_fack(tp))
772 mib_idx = LINUX_MIB_TCPFACKREORDER;
774 mib_idx = LINUX_MIB_TCPSACKREORDER;
776 NET_INC_STATS_BH(sock_net(sk), mib_idx);
777 #if FASTRETRANS_DEBUG > 1
778 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
779 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
783 tp->undo_marker ? tp->undo_retrans : 0);
785 tcp_disable_fack(tp);
789 tcp_disable_early_retrans(tp);
792 /* This must be called before lost_out is incremented */
793 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
795 if ((tp->retransmit_skb_hint == NULL) ||
796 before(TCP_SKB_CB(skb)->seq,
797 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
798 tp->retransmit_skb_hint = skb;
801 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
802 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
805 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
807 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
808 tcp_verify_retransmit_hint(tp, skb);
810 tp->lost_out += tcp_skb_pcount(skb);
811 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
815 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
818 tcp_verify_retransmit_hint(tp, skb);
820 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
821 tp->lost_out += tcp_skb_pcount(skb);
822 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
826 /* This procedure tags the retransmission queue when SACKs arrive.
828 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
829 * Packets in queue with these bits set are counted in variables
830 * sacked_out, retrans_out and lost_out, correspondingly.
832 * Valid combinations are:
833 * Tag InFlight Description
834 * 0 1 - orig segment is in flight.
835 * S 0 - nothing flies, orig reached receiver.
836 * L 0 - nothing flies, orig lost by net.
837 * R 2 - both orig and retransmit are in flight.
838 * L|R 1 - orig is lost, retransmit is in flight.
839 * S|R 1 - orig reached receiver, retrans is still in flight.
840 * (L|S|R is logically valid, it could occur when L|R is sacked,
841 * but it is equivalent to plain S and code short-curcuits it to S.
842 * L|S is logically invalid, it would mean -1 packet in flight 8))
844 * These 6 states form finite state machine, controlled by the following events:
845 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
846 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
847 * 3. Loss detection event of two flavors:
848 * A. Scoreboard estimator decided the packet is lost.
849 * A'. Reno "three dupacks" marks head of queue lost.
850 * A''. Its FACK modification, head until snd.fack is lost.
851 * B. SACK arrives sacking SND.NXT at the moment, when the
852 * segment was retransmitted.
853 * 4. D-SACK added new rule: D-SACK changes any tag to S.
855 * It is pleasant to note, that state diagram turns out to be commutative,
856 * so that we are allowed not to be bothered by order of our actions,
857 * when multiple events arrive simultaneously. (see the function below).
859 * Reordering detection.
860 * --------------------
861 * Reordering metric is maximal distance, which a packet can be displaced
862 * in packet stream. With SACKs we can estimate it:
864 * 1. SACK fills old hole and the corresponding segment was not
865 * ever retransmitted -> reordering. Alas, we cannot use it
866 * when segment was retransmitted.
867 * 2. The last flaw is solved with D-SACK. D-SACK arrives
868 * for retransmitted and already SACKed segment -> reordering..
869 * Both of these heuristics are not used in Loss state, when we cannot
870 * account for retransmits accurately.
872 * SACK block validation.
873 * ----------------------
875 * SACK block range validation checks that the received SACK block fits to
876 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
877 * Note that SND.UNA is not included to the range though being valid because
878 * it means that the receiver is rather inconsistent with itself reporting
879 * SACK reneging when it should advance SND.UNA. Such SACK block this is
880 * perfectly valid, however, in light of RFC2018 which explicitly states
881 * that "SACK block MUST reflect the newest segment. Even if the newest
882 * segment is going to be discarded ...", not that it looks very clever
883 * in case of head skb. Due to potentional receiver driven attacks, we
884 * choose to avoid immediate execution of a walk in write queue due to
885 * reneging and defer head skb's loss recovery to standard loss recovery
886 * procedure that will eventually trigger (nothing forbids us doing this).
888 * Implements also blockage to start_seq wrap-around. Problem lies in the
889 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
890 * there's no guarantee that it will be before snd_nxt (n). The problem
891 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
894 * <- outs wnd -> <- wrapzone ->
895 * u e n u_w e_w s n_w
897 * |<------------+------+----- TCP seqno space --------------+---------->|
898 * ...-- <2^31 ->| |<--------...
899 * ...---- >2^31 ------>| |<--------...
901 * Current code wouldn't be vulnerable but it's better still to discard such
902 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
903 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
904 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
905 * equal to the ideal case (infinite seqno space without wrap caused issues).
907 * With D-SACK the lower bound is extended to cover sequence space below
908 * SND.UNA down to undo_marker, which is the last point of interest. Yet
909 * again, D-SACK block must not to go across snd_una (for the same reason as
910 * for the normal SACK blocks, explained above). But there all simplicity
911 * ends, TCP might receive valid D-SACKs below that. As long as they reside
912 * fully below undo_marker they do not affect behavior in anyway and can
913 * therefore be safely ignored. In rare cases (which are more or less
914 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
915 * fragmentation and packet reordering past skb's retransmission. To consider
916 * them correctly, the acceptable range must be extended even more though
917 * the exact amount is rather hard to quantify. However, tp->max_window can
918 * be used as an exaggerated estimate.
920 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
921 u32 start_seq, u32 end_seq)
923 /* Too far in future, or reversed (interpretation is ambiguous) */
924 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
927 /* Nasty start_seq wrap-around check (see comments above) */
928 if (!before(start_seq, tp->snd_nxt))
931 /* In outstanding window? ...This is valid exit for D-SACKs too.
932 * start_seq == snd_una is non-sensical (see comments above)
934 if (after(start_seq, tp->snd_una))
937 if (!is_dsack || !tp->undo_marker)
940 /* ...Then it's D-SACK, and must reside below snd_una completely */
941 if (after(end_seq, tp->snd_una))
944 if (!before(start_seq, tp->undo_marker))
948 if (!after(end_seq, tp->undo_marker))
951 /* Undo_marker boundary crossing (overestimates a lot). Known already:
952 * start_seq < undo_marker and end_seq >= undo_marker.
954 return !before(start_seq, end_seq - tp->max_window);
957 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
958 * Event "B". Later note: FACK people cheated me again 8), we have to account
959 * for reordering! Ugly, but should help.
961 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
962 * less than what is now known to be received by the other end (derived from
963 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
964 * retransmitted skbs to avoid some costly processing per ACKs.
966 static void tcp_mark_lost_retrans(struct sock *sk)
968 const struct inet_connection_sock *icsk = inet_csk(sk);
969 struct tcp_sock *tp = tcp_sk(sk);
972 u32 new_low_seq = tp->snd_nxt;
973 u32 received_upto = tcp_highest_sack_seq(tp);
975 if (!tcp_is_fack(tp) || !tp->retrans_out ||
976 !after(received_upto, tp->lost_retrans_low) ||
977 icsk->icsk_ca_state != TCP_CA_Recovery)
980 tcp_for_write_queue(skb, sk) {
981 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
983 if (skb == tcp_send_head(sk))
985 if (cnt == tp->retrans_out)
987 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
990 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
993 /* TODO: We would like to get rid of tcp_is_fack(tp) only
994 * constraint here (see above) but figuring out that at
995 * least tp->reordering SACK blocks reside between ack_seq
996 * and received_upto is not easy task to do cheaply with
997 * the available datastructures.
999 * Whether FACK should check here for tp->reordering segs
1000 * in-between one could argue for either way (it would be
1001 * rather simple to implement as we could count fack_count
1002 * during the walk and do tp->fackets_out - fack_count).
1004 if (after(received_upto, ack_seq)) {
1005 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1006 tp->retrans_out -= tcp_skb_pcount(skb);
1008 tcp_skb_mark_lost_uncond_verify(tp, skb);
1009 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1011 if (before(ack_seq, new_low_seq))
1012 new_low_seq = ack_seq;
1013 cnt += tcp_skb_pcount(skb);
1017 if (tp->retrans_out)
1018 tp->lost_retrans_low = new_low_seq;
1021 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1022 struct tcp_sack_block_wire *sp, int num_sacks,
1025 struct tcp_sock *tp = tcp_sk(sk);
1026 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1027 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1028 bool dup_sack = false;
1030 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1033 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1034 } else if (num_sacks > 1) {
1035 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1036 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1038 if (!after(end_seq_0, end_seq_1) &&
1039 !before(start_seq_0, start_seq_1)) {
1042 NET_INC_STATS_BH(sock_net(sk),
1043 LINUX_MIB_TCPDSACKOFORECV);
1047 /* D-SACK for already forgotten data... Do dumb counting. */
1048 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1049 !after(end_seq_0, prior_snd_una) &&
1050 after(end_seq_0, tp->undo_marker))
1056 struct tcp_sacktag_state {
1062 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1063 * the incoming SACK may not exactly match but we can find smaller MSS
1064 * aligned portion of it that matches. Therefore we might need to fragment
1065 * which may fail and creates some hassle (caller must handle error case
1068 * FIXME: this could be merged to shift decision code
1070 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1071 u32 start_seq, u32 end_seq)
1075 unsigned int pkt_len;
1078 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1079 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1081 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1082 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1083 mss = tcp_skb_mss(skb);
1084 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1087 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1091 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1096 /* Round if necessary so that SACKs cover only full MSSes
1097 * and/or the remaining small portion (if present)
1099 if (pkt_len > mss) {
1100 unsigned int new_len = (pkt_len / mss) * mss;
1101 if (!in_sack && new_len < pkt_len) {
1103 if (new_len > skb->len)
1108 err = tcp_fragment(sk, skb, pkt_len, mss);
1116 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1117 static u8 tcp_sacktag_one(struct sock *sk,
1118 struct tcp_sacktag_state *state, u8 sacked,
1119 u32 start_seq, u32 end_seq,
1120 bool dup_sack, int pcount)
1122 struct tcp_sock *tp = tcp_sk(sk);
1123 int fack_count = state->fack_count;
1125 /* Account D-SACK for retransmitted packet. */
1126 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1127 if (tp->undo_marker && tp->undo_retrans &&
1128 after(end_seq, tp->undo_marker))
1130 if (sacked & TCPCB_SACKED_ACKED)
1131 state->reord = min(fack_count, state->reord);
1134 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1135 if (!after(end_seq, tp->snd_una))
1138 if (!(sacked & TCPCB_SACKED_ACKED)) {
1139 if (sacked & TCPCB_SACKED_RETRANS) {
1140 /* If the segment is not tagged as lost,
1141 * we do not clear RETRANS, believing
1142 * that retransmission is still in flight.
1144 if (sacked & TCPCB_LOST) {
1145 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1146 tp->lost_out -= pcount;
1147 tp->retrans_out -= pcount;
1150 if (!(sacked & TCPCB_RETRANS)) {
1151 /* New sack for not retransmitted frame,
1152 * which was in hole. It is reordering.
1154 if (before(start_seq,
1155 tcp_highest_sack_seq(tp)))
1156 state->reord = min(fack_count,
1160 if (sacked & TCPCB_LOST) {
1161 sacked &= ~TCPCB_LOST;
1162 tp->lost_out -= pcount;
1166 sacked |= TCPCB_SACKED_ACKED;
1167 state->flag |= FLAG_DATA_SACKED;
1168 tp->sacked_out += pcount;
1170 fack_count += pcount;
1172 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1173 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1174 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1175 tp->lost_cnt_hint += pcount;
1177 if (fack_count > tp->fackets_out)
1178 tp->fackets_out = fack_count;
1181 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1182 * frames and clear it. undo_retrans is decreased above, L|R frames
1183 * are accounted above as well.
1185 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1186 sacked &= ~TCPCB_SACKED_RETRANS;
1187 tp->retrans_out -= pcount;
1193 /* Shift newly-SACKed bytes from this skb to the immediately previous
1194 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1196 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1197 struct tcp_sacktag_state *state,
1198 unsigned int pcount, int shifted, int mss,
1201 struct tcp_sock *tp = tcp_sk(sk);
1202 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1203 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1204 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1208 /* Adjust counters and hints for the newly sacked sequence
1209 * range but discard the return value since prev is already
1210 * marked. We must tag the range first because the seq
1211 * advancement below implicitly advances
1212 * tcp_highest_sack_seq() when skb is highest_sack.
1214 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1215 start_seq, end_seq, dup_sack, pcount);
1217 if (skb == tp->lost_skb_hint)
1218 tp->lost_cnt_hint += pcount;
1220 TCP_SKB_CB(prev)->end_seq += shifted;
1221 TCP_SKB_CB(skb)->seq += shifted;
1223 skb_shinfo(prev)->gso_segs += pcount;
1224 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1225 skb_shinfo(skb)->gso_segs -= pcount;
1227 /* When we're adding to gso_segs == 1, gso_size will be zero,
1228 * in theory this shouldn't be necessary but as long as DSACK
1229 * code can come after this skb later on it's better to keep
1230 * setting gso_size to something.
1232 if (!skb_shinfo(prev)->gso_size) {
1233 skb_shinfo(prev)->gso_size = mss;
1234 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1237 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1238 if (skb_shinfo(skb)->gso_segs <= 1) {
1239 skb_shinfo(skb)->gso_size = 0;
1240 skb_shinfo(skb)->gso_type = 0;
1243 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1244 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1247 BUG_ON(!tcp_skb_pcount(skb));
1248 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1252 /* Whole SKB was eaten :-) */
1254 if (skb == tp->retransmit_skb_hint)
1255 tp->retransmit_skb_hint = prev;
1256 if (skb == tp->scoreboard_skb_hint)
1257 tp->scoreboard_skb_hint = prev;
1258 if (skb == tp->lost_skb_hint) {
1259 tp->lost_skb_hint = prev;
1260 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1263 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1264 if (skb == tcp_highest_sack(sk))
1265 tcp_advance_highest_sack(sk, skb);
1267 tcp_unlink_write_queue(skb, sk);
1268 sk_wmem_free_skb(sk, skb);
1270 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1275 /* I wish gso_size would have a bit more sane initialization than
1276 * something-or-zero which complicates things
1278 static int tcp_skb_seglen(const struct sk_buff *skb)
1280 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1283 /* Shifting pages past head area doesn't work */
1284 static int skb_can_shift(const struct sk_buff *skb)
1286 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1289 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1292 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1293 struct tcp_sacktag_state *state,
1294 u32 start_seq, u32 end_seq,
1297 struct tcp_sock *tp = tcp_sk(sk);
1298 struct sk_buff *prev;
1304 if (!sk_can_gso(sk))
1307 /* Normally R but no L won't result in plain S */
1309 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1311 if (!skb_can_shift(skb))
1313 /* This frame is about to be dropped (was ACKed). */
1314 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1317 /* Can only happen with delayed DSACK + discard craziness */
1318 if (unlikely(skb == tcp_write_queue_head(sk)))
1320 prev = tcp_write_queue_prev(sk, skb);
1322 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1325 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1326 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1330 pcount = tcp_skb_pcount(skb);
1331 mss = tcp_skb_seglen(skb);
1333 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1334 * drop this restriction as unnecessary
1336 if (mss != tcp_skb_seglen(prev))
1339 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1341 /* CHECKME: This is non-MSS split case only?, this will
1342 * cause skipped skbs due to advancing loop btw, original
1343 * has that feature too
1345 if (tcp_skb_pcount(skb) <= 1)
1348 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1350 /* TODO: head merge to next could be attempted here
1351 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1352 * though it might not be worth of the additional hassle
1354 * ...we can probably just fallback to what was done
1355 * previously. We could try merging non-SACKed ones
1356 * as well but it probably isn't going to buy off
1357 * because later SACKs might again split them, and
1358 * it would make skb timestamp tracking considerably
1364 len = end_seq - TCP_SKB_CB(skb)->seq;
1366 BUG_ON(len > skb->len);
1368 /* MSS boundaries should be honoured or else pcount will
1369 * severely break even though it makes things bit trickier.
1370 * Optimize common case to avoid most of the divides
1372 mss = tcp_skb_mss(skb);
1374 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1375 * drop this restriction as unnecessary
1377 if (mss != tcp_skb_seglen(prev))
1382 } else if (len < mss) {
1390 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1391 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1394 if (!skb_shift(prev, skb, len))
1396 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1399 /* Hole filled allows collapsing with the next as well, this is very
1400 * useful when hole on every nth skb pattern happens
1402 if (prev == tcp_write_queue_tail(sk))
1404 skb = tcp_write_queue_next(sk, prev);
1406 if (!skb_can_shift(skb) ||
1407 (skb == tcp_send_head(sk)) ||
1408 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1409 (mss != tcp_skb_seglen(skb)))
1413 if (skb_shift(prev, skb, len)) {
1414 pcount += tcp_skb_pcount(skb);
1415 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1419 state->fack_count += pcount;
1426 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1430 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1431 struct tcp_sack_block *next_dup,
1432 struct tcp_sacktag_state *state,
1433 u32 start_seq, u32 end_seq,
1436 struct tcp_sock *tp = tcp_sk(sk);
1437 struct sk_buff *tmp;
1439 tcp_for_write_queue_from(skb, sk) {
1441 bool dup_sack = dup_sack_in;
1443 if (skb == tcp_send_head(sk))
1446 /* queue is in-order => we can short-circuit the walk early */
1447 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1450 if ((next_dup != NULL) &&
1451 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1452 in_sack = tcp_match_skb_to_sack(sk, skb,
1453 next_dup->start_seq,
1459 /* skb reference here is a bit tricky to get right, since
1460 * shifting can eat and free both this skb and the next,
1461 * so not even _safe variant of the loop is enough.
1464 tmp = tcp_shift_skb_data(sk, skb, state,
1465 start_seq, end_seq, dup_sack);
1474 in_sack = tcp_match_skb_to_sack(sk, skb,
1480 if (unlikely(in_sack < 0))
1484 TCP_SKB_CB(skb)->sacked =
1487 TCP_SKB_CB(skb)->sacked,
1488 TCP_SKB_CB(skb)->seq,
1489 TCP_SKB_CB(skb)->end_seq,
1491 tcp_skb_pcount(skb));
1493 if (!before(TCP_SKB_CB(skb)->seq,
1494 tcp_highest_sack_seq(tp)))
1495 tcp_advance_highest_sack(sk, skb);
1498 state->fack_count += tcp_skb_pcount(skb);
1503 /* Avoid all extra work that is being done by sacktag while walking in
1506 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1507 struct tcp_sacktag_state *state,
1510 tcp_for_write_queue_from(skb, sk) {
1511 if (skb == tcp_send_head(sk))
1514 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1517 state->fack_count += tcp_skb_pcount(skb);
1522 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1524 struct tcp_sack_block *next_dup,
1525 struct tcp_sacktag_state *state,
1528 if (next_dup == NULL)
1531 if (before(next_dup->start_seq, skip_to_seq)) {
1532 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1533 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1534 next_dup->start_seq, next_dup->end_seq,
1541 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1543 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1547 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1550 struct tcp_sock *tp = tcp_sk(sk);
1551 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1552 TCP_SKB_CB(ack_skb)->sacked);
1553 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1554 struct tcp_sack_block sp[TCP_NUM_SACKS];
1555 struct tcp_sack_block *cache;
1556 struct tcp_sacktag_state state;
1557 struct sk_buff *skb;
1558 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1560 bool found_dup_sack = false;
1562 int first_sack_index;
1565 state.reord = tp->packets_out;
1567 if (!tp->sacked_out) {
1568 if (WARN_ON(tp->fackets_out))
1569 tp->fackets_out = 0;
1570 tcp_highest_sack_reset(sk);
1573 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1574 num_sacks, prior_snd_una);
1576 state.flag |= FLAG_DSACKING_ACK;
1578 /* Eliminate too old ACKs, but take into
1579 * account more or less fresh ones, they can
1580 * contain valid SACK info.
1582 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1585 if (!tp->packets_out)
1589 first_sack_index = 0;
1590 for (i = 0; i < num_sacks; i++) {
1591 bool dup_sack = !i && found_dup_sack;
1593 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1594 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1596 if (!tcp_is_sackblock_valid(tp, dup_sack,
1597 sp[used_sacks].start_seq,
1598 sp[used_sacks].end_seq)) {
1602 if (!tp->undo_marker)
1603 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1605 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1607 /* Don't count olds caused by ACK reordering */
1608 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1609 !after(sp[used_sacks].end_seq, tp->snd_una))
1611 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1614 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1616 first_sack_index = -1;
1620 /* Ignore very old stuff early */
1621 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1627 /* order SACK blocks to allow in order walk of the retrans queue */
1628 for (i = used_sacks - 1; i > 0; i--) {
1629 for (j = 0; j < i; j++) {
1630 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1631 swap(sp[j], sp[j + 1]);
1633 /* Track where the first SACK block goes to */
1634 if (j == first_sack_index)
1635 first_sack_index = j + 1;
1640 skb = tcp_write_queue_head(sk);
1641 state.fack_count = 0;
1644 if (!tp->sacked_out) {
1645 /* It's already past, so skip checking against it */
1646 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1648 cache = tp->recv_sack_cache;
1649 /* Skip empty blocks in at head of the cache */
1650 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1655 while (i < used_sacks) {
1656 u32 start_seq = sp[i].start_seq;
1657 u32 end_seq = sp[i].end_seq;
1658 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1659 struct tcp_sack_block *next_dup = NULL;
1661 if (found_dup_sack && ((i + 1) == first_sack_index))
1662 next_dup = &sp[i + 1];
1664 /* Skip too early cached blocks */
1665 while (tcp_sack_cache_ok(tp, cache) &&
1666 !before(start_seq, cache->end_seq))
1669 /* Can skip some work by looking recv_sack_cache? */
1670 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1671 after(end_seq, cache->start_seq)) {
1674 if (before(start_seq, cache->start_seq)) {
1675 skb = tcp_sacktag_skip(skb, sk, &state,
1677 skb = tcp_sacktag_walk(skb, sk, next_dup,
1684 /* Rest of the block already fully processed? */
1685 if (!after(end_seq, cache->end_seq))
1688 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1692 /* ...tail remains todo... */
1693 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1694 /* ...but better entrypoint exists! */
1695 skb = tcp_highest_sack(sk);
1698 state.fack_count = tp->fackets_out;
1703 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1704 /* Check overlap against next cached too (past this one already) */
1709 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1710 skb = tcp_highest_sack(sk);
1713 state.fack_count = tp->fackets_out;
1715 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1718 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1719 start_seq, end_seq, dup_sack);
1725 /* Clear the head of the cache sack blocks so we can skip it next time */
1726 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1727 tp->recv_sack_cache[i].start_seq = 0;
1728 tp->recv_sack_cache[i].end_seq = 0;
1730 for (j = 0; j < used_sacks; j++)
1731 tp->recv_sack_cache[i++] = sp[j];
1733 tcp_mark_lost_retrans(sk);
1735 tcp_verify_left_out(tp);
1737 if ((state.reord < tp->fackets_out) &&
1738 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1739 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1743 #if FASTRETRANS_DEBUG > 0
1744 WARN_ON((int)tp->sacked_out < 0);
1745 WARN_ON((int)tp->lost_out < 0);
1746 WARN_ON((int)tp->retrans_out < 0);
1747 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1752 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1753 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1755 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1759 holes = max(tp->lost_out, 1U);
1760 holes = min(holes, tp->packets_out);
1762 if ((tp->sacked_out + holes) > tp->packets_out) {
1763 tp->sacked_out = tp->packets_out - holes;
1769 /* If we receive more dupacks than we expected counting segments
1770 * in assumption of absent reordering, interpret this as reordering.
1771 * The only another reason could be bug in receiver TCP.
1773 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1775 struct tcp_sock *tp = tcp_sk(sk);
1776 if (tcp_limit_reno_sacked(tp))
1777 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1780 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1782 static void tcp_add_reno_sack(struct sock *sk)
1784 struct tcp_sock *tp = tcp_sk(sk);
1786 tcp_check_reno_reordering(sk, 0);
1787 tcp_verify_left_out(tp);
1790 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1792 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1794 struct tcp_sock *tp = tcp_sk(sk);
1797 /* One ACK acked hole. The rest eat duplicate ACKs. */
1798 if (acked - 1 >= tp->sacked_out)
1801 tp->sacked_out -= acked - 1;
1803 tcp_check_reno_reordering(sk, acked);
1804 tcp_verify_left_out(tp);
1807 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1812 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1814 tp->retrans_out = 0;
1817 tp->undo_marker = 0;
1818 tp->undo_retrans = 0;
1821 void tcp_clear_retrans(struct tcp_sock *tp)
1823 tcp_clear_retrans_partial(tp);
1825 tp->fackets_out = 0;
1829 /* Enter Loss state. If "how" is not zero, forget all SACK information
1830 * and reset tags completely, otherwise preserve SACKs. If receiver
1831 * dropped its ofo queue, we will know this due to reneging detection.
1833 void tcp_enter_loss(struct sock *sk, int how)
1835 const struct inet_connection_sock *icsk = inet_csk(sk);
1836 struct tcp_sock *tp = tcp_sk(sk);
1837 struct sk_buff *skb;
1839 /* Reduce ssthresh if it has not yet been made inside this window. */
1840 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1841 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1842 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1843 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1844 tcp_ca_event(sk, CA_EVENT_LOSS);
1847 tp->snd_cwnd_cnt = 0;
1848 tp->snd_cwnd_stamp = tcp_time_stamp;
1850 tcp_clear_retrans_partial(tp);
1852 if (tcp_is_reno(tp))
1853 tcp_reset_reno_sack(tp);
1856 /* Push undo marker, if it was plain RTO and nothing
1857 * was retransmitted. */
1858 tp->undo_marker = tp->snd_una;
1861 tp->fackets_out = 0;
1863 tcp_clear_all_retrans_hints(tp);
1865 tcp_for_write_queue(skb, sk) {
1866 if (skb == tcp_send_head(sk))
1869 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1870 tp->undo_marker = 0;
1871 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1872 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1873 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1874 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1875 tp->lost_out += tcp_skb_pcount(skb);
1876 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1879 tcp_verify_left_out(tp);
1881 tp->reordering = min_t(unsigned int, tp->reordering,
1882 sysctl_tcp_reordering);
1883 tcp_set_ca_state(sk, TCP_CA_Loss);
1884 tp->high_seq = tp->snd_nxt;
1885 TCP_ECN_queue_cwr(tp);
1888 /* If ACK arrived pointing to a remembered SACK, it means that our
1889 * remembered SACKs do not reflect real state of receiver i.e.
1890 * receiver _host_ is heavily congested (or buggy).
1892 * Do processing similar to RTO timeout.
1894 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1896 if (flag & FLAG_SACK_RENEGING) {
1897 struct inet_connection_sock *icsk = inet_csk(sk);
1898 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1900 tcp_enter_loss(sk, 1);
1901 icsk->icsk_retransmits++;
1902 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1903 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1904 icsk->icsk_rto, TCP_RTO_MAX);
1910 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1912 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1915 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1916 * counter when SACK is enabled (without SACK, sacked_out is used for
1919 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1920 * segments up to the highest received SACK block so far and holes in
1923 * With reordering, holes may still be in flight, so RFC3517 recovery
1924 * uses pure sacked_out (total number of SACKed segments) even though
1925 * it violates the RFC that uses duplicate ACKs, often these are equal
1926 * but when e.g. out-of-window ACKs or packet duplication occurs,
1927 * they differ. Since neither occurs due to loss, TCP should really
1930 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1932 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1935 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1937 struct tcp_sock *tp = tcp_sk(sk);
1938 unsigned long delay;
1940 /* Delay early retransmit and entering fast recovery for
1941 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1942 * available, or RTO is scheduled to fire first.
1944 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
1945 (flag & FLAG_ECE) || !tp->srtt)
1948 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
1949 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
1952 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
1957 static inline int tcp_skb_timedout(const struct sock *sk,
1958 const struct sk_buff *skb)
1960 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
1963 static inline int tcp_head_timedout(const struct sock *sk)
1965 const struct tcp_sock *tp = tcp_sk(sk);
1967 return tp->packets_out &&
1968 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1971 /* Linux NewReno/SACK/FACK/ECN state machine.
1972 * --------------------------------------
1974 * "Open" Normal state, no dubious events, fast path.
1975 * "Disorder" In all the respects it is "Open",
1976 * but requires a bit more attention. It is entered when
1977 * we see some SACKs or dupacks. It is split of "Open"
1978 * mainly to move some processing from fast path to slow one.
1979 * "CWR" CWND was reduced due to some Congestion Notification event.
1980 * It can be ECN, ICMP source quench, local device congestion.
1981 * "Recovery" CWND was reduced, we are fast-retransmitting.
1982 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1984 * tcp_fastretrans_alert() is entered:
1985 * - each incoming ACK, if state is not "Open"
1986 * - when arrived ACK is unusual, namely:
1991 * Counting packets in flight is pretty simple.
1993 * in_flight = packets_out - left_out + retrans_out
1995 * packets_out is SND.NXT-SND.UNA counted in packets.
1997 * retrans_out is number of retransmitted segments.
1999 * left_out is number of segments left network, but not ACKed yet.
2001 * left_out = sacked_out + lost_out
2003 * sacked_out: Packets, which arrived to receiver out of order
2004 * and hence not ACKed. With SACKs this number is simply
2005 * amount of SACKed data. Even without SACKs
2006 * it is easy to give pretty reliable estimate of this number,
2007 * counting duplicate ACKs.
2009 * lost_out: Packets lost by network. TCP has no explicit
2010 * "loss notification" feedback from network (for now).
2011 * It means that this number can be only _guessed_.
2012 * Actually, it is the heuristics to predict lossage that
2013 * distinguishes different algorithms.
2015 * F.e. after RTO, when all the queue is considered as lost,
2016 * lost_out = packets_out and in_flight = retrans_out.
2018 * Essentially, we have now two algorithms counting
2021 * FACK: It is the simplest heuristics. As soon as we decided
2022 * that something is lost, we decide that _all_ not SACKed
2023 * packets until the most forward SACK are lost. I.e.
2024 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2025 * It is absolutely correct estimate, if network does not reorder
2026 * packets. And it loses any connection to reality when reordering
2027 * takes place. We use FACK by default until reordering
2028 * is suspected on the path to this destination.
2030 * NewReno: when Recovery is entered, we assume that one segment
2031 * is lost (classic Reno). While we are in Recovery and
2032 * a partial ACK arrives, we assume that one more packet
2033 * is lost (NewReno). This heuristics are the same in NewReno
2036 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2037 * deflation etc. CWND is real congestion window, never inflated, changes
2038 * only according to classic VJ rules.
2040 * Really tricky (and requiring careful tuning) part of algorithm
2041 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2042 * The first determines the moment _when_ we should reduce CWND and,
2043 * hence, slow down forward transmission. In fact, it determines the moment
2044 * when we decide that hole is caused by loss, rather than by a reorder.
2046 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2047 * holes, caused by lost packets.
2049 * And the most logically complicated part of algorithm is undo
2050 * heuristics. We detect false retransmits due to both too early
2051 * fast retransmit (reordering) and underestimated RTO, analyzing
2052 * timestamps and D-SACKs. When we detect that some segments were
2053 * retransmitted by mistake and CWND reduction was wrong, we undo
2054 * window reduction and abort recovery phase. This logic is hidden
2055 * inside several functions named tcp_try_undo_<something>.
2058 /* This function decides, when we should leave Disordered state
2059 * and enter Recovery phase, reducing congestion window.
2061 * Main question: may we further continue forward transmission
2062 * with the same cwnd?
2064 static bool tcp_time_to_recover(struct sock *sk, int flag)
2066 struct tcp_sock *tp = tcp_sk(sk);
2069 /* Trick#1: The loss is proven. */
2073 /* Not-A-Trick#2 : Classic rule... */
2074 if (tcp_dupack_heuristics(tp) > tp->reordering)
2077 /* Trick#3 : when we use RFC2988 timer restart, fast
2078 * retransmit can be triggered by timeout of queue head.
2080 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2083 /* Trick#4: It is still not OK... But will it be useful to delay
2086 packets_out = tp->packets_out;
2087 if (packets_out <= tp->reordering &&
2088 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2089 !tcp_may_send_now(sk)) {
2090 /* We have nothing to send. This connection is limited
2091 * either by receiver window or by application.
2096 /* If a thin stream is detected, retransmit after first
2097 * received dupack. Employ only if SACK is supported in order
2098 * to avoid possible corner-case series of spurious retransmissions
2099 * Use only if there are no unsent data.
2101 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2102 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2103 tcp_is_sack(tp) && !tcp_send_head(sk))
2106 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2107 * retransmissions due to small network reorderings, we implement
2108 * Mitigation A.3 in the RFC and delay the retransmission for a short
2109 * interval if appropriate.
2111 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2112 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2113 !tcp_may_send_now(sk))
2114 return !tcp_pause_early_retransmit(sk, flag);
2119 /* New heuristics: it is possible only after we switched to restart timer
2120 * each time when something is ACKed. Hence, we can detect timed out packets
2121 * during fast retransmit without falling to slow start.
2123 * Usefulness of this as is very questionable, since we should know which of
2124 * the segments is the next to timeout which is relatively expensive to find
2125 * in general case unless we add some data structure just for that. The
2126 * current approach certainly won't find the right one too often and when it
2127 * finally does find _something_ it usually marks large part of the window
2128 * right away (because a retransmission with a larger timestamp blocks the
2129 * loop from advancing). -ij
2131 static void tcp_timeout_skbs(struct sock *sk)
2133 struct tcp_sock *tp = tcp_sk(sk);
2134 struct sk_buff *skb;
2136 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2139 skb = tp->scoreboard_skb_hint;
2140 if (tp->scoreboard_skb_hint == NULL)
2141 skb = tcp_write_queue_head(sk);
2143 tcp_for_write_queue_from(skb, sk) {
2144 if (skb == tcp_send_head(sk))
2146 if (!tcp_skb_timedout(sk, skb))
2149 tcp_skb_mark_lost(tp, skb);
2152 tp->scoreboard_skb_hint = skb;
2154 tcp_verify_left_out(tp);
2157 /* Detect loss in event "A" above by marking head of queue up as lost.
2158 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2159 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2160 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2161 * the maximum SACKed segments to pass before reaching this limit.
2163 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2165 struct tcp_sock *tp = tcp_sk(sk);
2166 struct sk_buff *skb;
2170 /* Use SACK to deduce losses of new sequences sent during recovery */
2171 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2173 WARN_ON(packets > tp->packets_out);
2174 if (tp->lost_skb_hint) {
2175 skb = tp->lost_skb_hint;
2176 cnt = tp->lost_cnt_hint;
2177 /* Head already handled? */
2178 if (mark_head && skb != tcp_write_queue_head(sk))
2181 skb = tcp_write_queue_head(sk);
2185 tcp_for_write_queue_from(skb, sk) {
2186 if (skb == tcp_send_head(sk))
2188 /* TODO: do this better */
2189 /* this is not the most efficient way to do this... */
2190 tp->lost_skb_hint = skb;
2191 tp->lost_cnt_hint = cnt;
2193 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2197 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2198 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2199 cnt += tcp_skb_pcount(skb);
2201 if (cnt > packets) {
2202 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2203 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2204 (oldcnt >= packets))
2207 mss = skb_shinfo(skb)->gso_size;
2208 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2214 tcp_skb_mark_lost(tp, skb);
2219 tcp_verify_left_out(tp);
2222 /* Account newly detected lost packet(s) */
2224 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2226 struct tcp_sock *tp = tcp_sk(sk);
2228 if (tcp_is_reno(tp)) {
2229 tcp_mark_head_lost(sk, 1, 1);
2230 } else if (tcp_is_fack(tp)) {
2231 int lost = tp->fackets_out - tp->reordering;
2234 tcp_mark_head_lost(sk, lost, 0);
2236 int sacked_upto = tp->sacked_out - tp->reordering;
2237 if (sacked_upto >= 0)
2238 tcp_mark_head_lost(sk, sacked_upto, 0);
2239 else if (fast_rexmit)
2240 tcp_mark_head_lost(sk, 1, 1);
2243 tcp_timeout_skbs(sk);
2246 /* CWND moderation, preventing bursts due to too big ACKs
2247 * in dubious situations.
2249 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2251 tp->snd_cwnd = min(tp->snd_cwnd,
2252 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2253 tp->snd_cwnd_stamp = tcp_time_stamp;
2256 /* Nothing was retransmitted or returned timestamp is less
2257 * than timestamp of the first retransmission.
2259 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2261 return !tp->retrans_stamp ||
2262 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2263 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2266 /* Undo procedures. */
2268 #if FASTRETRANS_DEBUG > 1
2269 static void DBGUNDO(struct sock *sk, const char *msg)
2271 struct tcp_sock *tp = tcp_sk(sk);
2272 struct inet_sock *inet = inet_sk(sk);
2274 if (sk->sk_family == AF_INET) {
2275 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2277 &inet->inet_daddr, ntohs(inet->inet_dport),
2278 tp->snd_cwnd, tcp_left_out(tp),
2279 tp->snd_ssthresh, tp->prior_ssthresh,
2282 #if IS_ENABLED(CONFIG_IPV6)
2283 else if (sk->sk_family == AF_INET6) {
2284 struct ipv6_pinfo *np = inet6_sk(sk);
2285 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2287 &np->daddr, ntohs(inet->inet_dport),
2288 tp->snd_cwnd, tcp_left_out(tp),
2289 tp->snd_ssthresh, tp->prior_ssthresh,
2295 #define DBGUNDO(x...) do { } while (0)
2298 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2300 struct tcp_sock *tp = tcp_sk(sk);
2302 if (tp->prior_ssthresh) {
2303 const struct inet_connection_sock *icsk = inet_csk(sk);
2305 if (icsk->icsk_ca_ops->undo_cwnd)
2306 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2308 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2310 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2311 tp->snd_ssthresh = tp->prior_ssthresh;
2312 TCP_ECN_withdraw_cwr(tp);
2315 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2317 tp->snd_cwnd_stamp = tcp_time_stamp;
2320 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2322 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2325 /* People celebrate: "We love our President!" */
2326 static bool tcp_try_undo_recovery(struct sock *sk)
2328 struct tcp_sock *tp = tcp_sk(sk);
2330 if (tcp_may_undo(tp)) {
2333 /* Happy end! We did not retransmit anything
2334 * or our original transmission succeeded.
2336 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2337 tcp_undo_cwr(sk, true);
2338 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2339 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2341 mib_idx = LINUX_MIB_TCPFULLUNDO;
2343 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2344 tp->undo_marker = 0;
2346 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2347 /* Hold old state until something *above* high_seq
2348 * is ACKed. For Reno it is MUST to prevent false
2349 * fast retransmits (RFC2582). SACK TCP is safe. */
2350 tcp_moderate_cwnd(tp);
2353 tcp_set_ca_state(sk, TCP_CA_Open);
2357 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2358 static void tcp_try_undo_dsack(struct sock *sk)
2360 struct tcp_sock *tp = tcp_sk(sk);
2362 if (tp->undo_marker && !tp->undo_retrans) {
2363 DBGUNDO(sk, "D-SACK");
2364 tcp_undo_cwr(sk, true);
2365 tp->undo_marker = 0;
2366 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2370 /* We can clear retrans_stamp when there are no retransmissions in the
2371 * window. It would seem that it is trivially available for us in
2372 * tp->retrans_out, however, that kind of assumptions doesn't consider
2373 * what will happen if errors occur when sending retransmission for the
2374 * second time. ...It could the that such segment has only
2375 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2376 * the head skb is enough except for some reneging corner cases that
2377 * are not worth the effort.
2379 * Main reason for all this complexity is the fact that connection dying
2380 * time now depends on the validity of the retrans_stamp, in particular,
2381 * that successive retransmissions of a segment must not advance
2382 * retrans_stamp under any conditions.
2384 static bool tcp_any_retrans_done(const struct sock *sk)
2386 const struct tcp_sock *tp = tcp_sk(sk);
2387 struct sk_buff *skb;
2389 if (tp->retrans_out)
2392 skb = tcp_write_queue_head(sk);
2393 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2399 /* Undo during fast recovery after partial ACK. */
2401 static int tcp_try_undo_partial(struct sock *sk, int acked)
2403 struct tcp_sock *tp = tcp_sk(sk);
2404 /* Partial ACK arrived. Force Hoe's retransmit. */
2405 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2407 if (tcp_may_undo(tp)) {
2408 /* Plain luck! Hole if filled with delayed
2409 * packet, rather than with a retransmit.
2411 if (!tcp_any_retrans_done(sk))
2412 tp->retrans_stamp = 0;
2414 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2417 tcp_undo_cwr(sk, false);
2418 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2420 /* So... Do not make Hoe's retransmit yet.
2421 * If the first packet was delayed, the rest
2422 * ones are most probably delayed as well.
2429 /* Undo during loss recovery after partial ACK. */
2430 static bool tcp_try_undo_loss(struct sock *sk)
2432 struct tcp_sock *tp = tcp_sk(sk);
2434 if (tcp_may_undo(tp)) {
2435 struct sk_buff *skb;
2436 tcp_for_write_queue(skb, sk) {
2437 if (skb == tcp_send_head(sk))
2439 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2442 tcp_clear_all_retrans_hints(tp);
2444 DBGUNDO(sk, "partial loss");
2446 tcp_undo_cwr(sk, true);
2447 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2448 inet_csk(sk)->icsk_retransmits = 0;
2449 tp->undo_marker = 0;
2450 if (tcp_is_sack(tp))
2451 tcp_set_ca_state(sk, TCP_CA_Open);
2457 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2458 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2459 * It computes the number of packets to send (sndcnt) based on packets newly
2461 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2462 * cwnd reductions across a full RTT.
2463 * 2) If packets in flight is lower than ssthresh (such as due to excess
2464 * losses and/or application stalls), do not perform any further cwnd
2465 * reductions, but instead slow start up to ssthresh.
2467 static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2469 struct tcp_sock *tp = tcp_sk(sk);
2471 tp->high_seq = tp->snd_nxt;
2472 tp->tlp_high_seq = 0;
2473 tp->snd_cwnd_cnt = 0;
2474 tp->prior_cwnd = tp->snd_cwnd;
2475 tp->prr_delivered = 0;
2478 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2479 TCP_ECN_queue_cwr(tp);
2482 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2485 struct tcp_sock *tp = tcp_sk(sk);
2487 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2489 tp->prr_delivered += newly_acked_sacked;
2490 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2491 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2493 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2495 sndcnt = min_t(int, delta,
2496 max_t(int, tp->prr_delivered - tp->prr_out,
2497 newly_acked_sacked) + 1);
2500 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2501 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2504 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2506 struct tcp_sock *tp = tcp_sk(sk);
2508 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2509 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2510 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2511 tp->snd_cwnd = tp->snd_ssthresh;
2512 tp->snd_cwnd_stamp = tcp_time_stamp;
2514 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2517 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2518 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2520 struct tcp_sock *tp = tcp_sk(sk);
2522 tp->prior_ssthresh = 0;
2523 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2524 tp->undo_marker = 0;
2525 tcp_init_cwnd_reduction(sk, set_ssthresh);
2526 tcp_set_ca_state(sk, TCP_CA_CWR);
2530 static void tcp_try_keep_open(struct sock *sk)
2532 struct tcp_sock *tp = tcp_sk(sk);
2533 int state = TCP_CA_Open;
2535 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2536 state = TCP_CA_Disorder;
2538 if (inet_csk(sk)->icsk_ca_state != state) {
2539 tcp_set_ca_state(sk, state);
2540 tp->high_seq = tp->snd_nxt;
2544 static void tcp_try_to_open(struct sock *sk, int flag, int newly_acked_sacked)
2546 struct tcp_sock *tp = tcp_sk(sk);
2548 tcp_verify_left_out(tp);
2550 if (!tcp_any_retrans_done(sk))
2551 tp->retrans_stamp = 0;
2553 if (flag & FLAG_ECE)
2554 tcp_enter_cwr(sk, 1);
2556 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2557 tcp_try_keep_open(sk);
2558 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2559 tcp_moderate_cwnd(tp);
2561 tcp_cwnd_reduction(sk, newly_acked_sacked, 0);
2565 static void tcp_mtup_probe_failed(struct sock *sk)
2567 struct inet_connection_sock *icsk = inet_csk(sk);
2569 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2570 icsk->icsk_mtup.probe_size = 0;
2573 static void tcp_mtup_probe_success(struct sock *sk)
2575 struct tcp_sock *tp = tcp_sk(sk);
2576 struct inet_connection_sock *icsk = inet_csk(sk);
2578 /* FIXME: breaks with very large cwnd */
2579 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2580 tp->snd_cwnd = tp->snd_cwnd *
2581 tcp_mss_to_mtu(sk, tp->mss_cache) /
2582 icsk->icsk_mtup.probe_size;
2583 tp->snd_cwnd_cnt = 0;
2584 tp->snd_cwnd_stamp = tcp_time_stamp;
2585 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2587 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2588 icsk->icsk_mtup.probe_size = 0;
2589 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2592 /* Do a simple retransmit without using the backoff mechanisms in
2593 * tcp_timer. This is used for path mtu discovery.
2594 * The socket is already locked here.
2596 void tcp_simple_retransmit(struct sock *sk)
2598 const struct inet_connection_sock *icsk = inet_csk(sk);
2599 struct tcp_sock *tp = tcp_sk(sk);
2600 struct sk_buff *skb;
2601 unsigned int mss = tcp_current_mss(sk);
2602 u32 prior_lost = tp->lost_out;
2604 tcp_for_write_queue(skb, sk) {
2605 if (skb == tcp_send_head(sk))
2607 if (tcp_skb_seglen(skb) > mss &&
2608 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2609 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2610 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2611 tp->retrans_out -= tcp_skb_pcount(skb);
2613 tcp_skb_mark_lost_uncond_verify(tp, skb);
2617 tcp_clear_retrans_hints_partial(tp);
2619 if (prior_lost == tp->lost_out)
2622 if (tcp_is_reno(tp))
2623 tcp_limit_reno_sacked(tp);
2625 tcp_verify_left_out(tp);
2627 /* Don't muck with the congestion window here.
2628 * Reason is that we do not increase amount of _data_
2629 * in network, but units changed and effective
2630 * cwnd/ssthresh really reduced now.
2632 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2633 tp->high_seq = tp->snd_nxt;
2634 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2635 tp->prior_ssthresh = 0;
2636 tp->undo_marker = 0;
2637 tcp_set_ca_state(sk, TCP_CA_Loss);
2639 tcp_xmit_retransmit_queue(sk);
2641 EXPORT_SYMBOL(tcp_simple_retransmit);
2643 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2645 struct tcp_sock *tp = tcp_sk(sk);
2648 if (tcp_is_reno(tp))
2649 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2651 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2653 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2655 tp->prior_ssthresh = 0;
2656 tp->undo_marker = tp->snd_una;
2657 tp->undo_retrans = tp->retrans_out;
2659 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2661 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2662 tcp_init_cwnd_reduction(sk, true);
2664 tcp_set_ca_state(sk, TCP_CA_Recovery);
2667 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2668 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2670 static void tcp_process_loss(struct sock *sk, int flag)
2672 struct inet_connection_sock *icsk = inet_csk(sk);
2673 struct tcp_sock *tp = tcp_sk(sk);
2675 if (!before(tp->snd_una, tp->high_seq)) {
2676 icsk->icsk_retransmits = 0;
2677 tcp_try_undo_recovery(sk);
2681 if (flag & FLAG_DATA_ACKED)
2682 icsk->icsk_retransmits = 0;
2683 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2684 tcp_reset_reno_sack(tp);
2685 if (tcp_try_undo_loss(sk))
2687 tcp_moderate_cwnd(tp);
2688 tcp_xmit_retransmit_queue(sk);
2691 /* Process an event, which can update packets-in-flight not trivially.
2692 * Main goal of this function is to calculate new estimate for left_out,
2693 * taking into account both packets sitting in receiver's buffer and
2694 * packets lost by network.
2696 * Besides that it does CWND reduction, when packet loss is detected
2697 * and changes state of machine.
2699 * It does _not_ decide what to send, it is made in function
2700 * tcp_xmit_retransmit_queue().
2702 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2703 int prior_sacked, bool is_dupack,
2706 struct inet_connection_sock *icsk = inet_csk(sk);
2707 struct tcp_sock *tp = tcp_sk(sk);
2708 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2709 (tcp_fackets_out(tp) > tp->reordering));
2710 int newly_acked_sacked = 0;
2711 int fast_rexmit = 0;
2713 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2715 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2716 tp->fackets_out = 0;
2718 /* Now state machine starts.
2719 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2720 if (flag & FLAG_ECE)
2721 tp->prior_ssthresh = 0;
2723 /* B. In all the states check for reneging SACKs. */
2724 if (tcp_check_sack_reneging(sk, flag))
2727 /* C. Check consistency of the current state. */
2728 tcp_verify_left_out(tp);
2730 /* D. Check state exit conditions. State can be terminated
2731 * when high_seq is ACKed. */
2732 if (icsk->icsk_ca_state == TCP_CA_Open) {
2733 WARN_ON(tp->retrans_out != 0);
2734 tp->retrans_stamp = 0;
2735 } else if (!before(tp->snd_una, tp->high_seq)) {
2736 switch (icsk->icsk_ca_state) {
2738 /* CWR is to be held something *above* high_seq
2739 * is ACKed for CWR bit to reach receiver. */
2740 if (tp->snd_una != tp->high_seq) {
2741 tcp_end_cwnd_reduction(sk);
2742 tcp_set_ca_state(sk, TCP_CA_Open);
2746 case TCP_CA_Recovery:
2747 if (tcp_is_reno(tp))
2748 tcp_reset_reno_sack(tp);
2749 if (tcp_try_undo_recovery(sk))
2751 tcp_end_cwnd_reduction(sk);
2756 /* E. Process state. */
2757 switch (icsk->icsk_ca_state) {
2758 case TCP_CA_Recovery:
2759 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2760 if (tcp_is_reno(tp) && is_dupack)
2761 tcp_add_reno_sack(sk);
2763 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2764 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
2767 tcp_process_loss(sk, flag);
2768 if (icsk->icsk_ca_state != TCP_CA_Open)
2770 /* Fall through to processing in Open state. */
2772 if (tcp_is_reno(tp)) {
2773 if (flag & FLAG_SND_UNA_ADVANCED)
2774 tcp_reset_reno_sack(tp);
2776 tcp_add_reno_sack(sk);
2778 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
2780 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2781 tcp_try_undo_dsack(sk);
2783 if (!tcp_time_to_recover(sk, flag)) {
2784 tcp_try_to_open(sk, flag, newly_acked_sacked);
2788 /* MTU probe failure: don't reduce cwnd */
2789 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2790 icsk->icsk_mtup.probe_size &&
2791 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2792 tcp_mtup_probe_failed(sk);
2793 /* Restores the reduction we did in tcp_mtup_probe() */
2795 tcp_simple_retransmit(sk);
2799 /* Otherwise enter Recovery state */
2800 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2804 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2805 tcp_update_scoreboard(sk, fast_rexmit);
2806 tcp_cwnd_reduction(sk, newly_acked_sacked, fast_rexmit);
2807 tcp_xmit_retransmit_queue(sk);
2810 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
2812 tcp_rtt_estimator(sk, seq_rtt);
2814 inet_csk(sk)->icsk_backoff = 0;
2816 EXPORT_SYMBOL(tcp_valid_rtt_meas);
2818 /* Read draft-ietf-tcplw-high-performance before mucking
2819 * with this code. (Supersedes RFC1323)
2821 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2823 /* RTTM Rule: A TSecr value received in a segment is used to
2824 * update the averaged RTT measurement only if the segment
2825 * acknowledges some new data, i.e., only if it advances the
2826 * left edge of the send window.
2828 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2829 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2831 * Changed: reset backoff as soon as we see the first valid sample.
2832 * If we do not, we get strongly overestimated rto. With timestamps
2833 * samples are accepted even from very old segments: f.e., when rtt=1
2834 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2835 * answer arrives rto becomes 120 seconds! If at least one of segments
2836 * in window is lost... Voila. --ANK (010210)
2838 struct tcp_sock *tp = tcp_sk(sk);
2840 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2843 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2845 /* We don't have a timestamp. Can only use
2846 * packets that are not retransmitted to determine
2847 * rtt estimates. Also, we must not reset the
2848 * backoff for rto until we get a non-retransmitted
2849 * packet. This allows us to deal with a situation
2850 * where the network delay has increased suddenly.
2851 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2854 if (flag & FLAG_RETRANS_DATA_ACKED)
2857 tcp_valid_rtt_meas(sk, seq_rtt);
2860 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2863 const struct tcp_sock *tp = tcp_sk(sk);
2864 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2865 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2866 tcp_ack_saw_tstamp(sk, flag);
2867 else if (seq_rtt >= 0)
2868 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2871 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2873 const struct inet_connection_sock *icsk = inet_csk(sk);
2874 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2875 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2878 /* Restart timer after forward progress on connection.
2879 * RFC2988 recommends to restart timer to now+rto.
2881 void tcp_rearm_rto(struct sock *sk)
2883 const struct inet_connection_sock *icsk = inet_csk(sk);
2884 struct tcp_sock *tp = tcp_sk(sk);
2886 /* If the retrans timer is currently being used by Fast Open
2887 * for SYN-ACK retrans purpose, stay put.
2889 if (tp->fastopen_rsk)
2892 if (!tp->packets_out) {
2893 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2895 u32 rto = inet_csk(sk)->icsk_rto;
2896 /* Offset the time elapsed after installing regular RTO */
2897 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2898 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2899 struct sk_buff *skb = tcp_write_queue_head(sk);
2900 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2901 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2902 /* delta may not be positive if the socket is locked
2903 * when the retrans timer fires and is rescheduled.
2908 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2913 /* This function is called when the delayed ER timer fires. TCP enters
2914 * fast recovery and performs fast-retransmit.
2916 void tcp_resume_early_retransmit(struct sock *sk)
2918 struct tcp_sock *tp = tcp_sk(sk);
2922 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2923 if (!tp->do_early_retrans)
2926 tcp_enter_recovery(sk, false);
2927 tcp_update_scoreboard(sk, 1);
2928 tcp_xmit_retransmit_queue(sk);
2931 /* If we get here, the whole TSO packet has not been acked. */
2932 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2934 struct tcp_sock *tp = tcp_sk(sk);
2937 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2939 packets_acked = tcp_skb_pcount(skb);
2940 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2942 packets_acked -= tcp_skb_pcount(skb);
2944 if (packets_acked) {
2945 BUG_ON(tcp_skb_pcount(skb) == 0);
2946 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2949 return packets_acked;
2952 /* Remove acknowledged frames from the retransmission queue. If our packet
2953 * is before the ack sequence we can discard it as it's confirmed to have
2954 * arrived at the other end.
2956 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
2959 struct tcp_sock *tp = tcp_sk(sk);
2960 const struct inet_connection_sock *icsk = inet_csk(sk);
2961 struct sk_buff *skb;
2962 u32 now = tcp_time_stamp;
2963 int fully_acked = true;
2966 u32 reord = tp->packets_out;
2967 u32 prior_sacked = tp->sacked_out;
2969 s32 ca_seq_rtt = -1;
2970 ktime_t last_ackt = net_invalid_timestamp();
2972 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2973 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2975 u8 sacked = scb->sacked;
2977 /* Determine how many packets and what bytes were acked, tso and else */
2978 if (after(scb->end_seq, tp->snd_una)) {
2979 if (tcp_skb_pcount(skb) == 1 ||
2980 !after(tp->snd_una, scb->seq))
2983 acked_pcount = tcp_tso_acked(sk, skb);
2987 fully_acked = false;
2989 acked_pcount = tcp_skb_pcount(skb);
2992 if (sacked & TCPCB_RETRANS) {
2993 if (sacked & TCPCB_SACKED_RETRANS)
2994 tp->retrans_out -= acked_pcount;
2995 flag |= FLAG_RETRANS_DATA_ACKED;
2999 ca_seq_rtt = now - scb->when;
3000 last_ackt = skb->tstamp;
3002 seq_rtt = ca_seq_rtt;
3004 if (!(sacked & TCPCB_SACKED_ACKED))
3005 reord = min(pkts_acked, reord);
3008 if (sacked & TCPCB_SACKED_ACKED)
3009 tp->sacked_out -= acked_pcount;
3010 if (sacked & TCPCB_LOST)
3011 tp->lost_out -= acked_pcount;
3013 tp->packets_out -= acked_pcount;
3014 pkts_acked += acked_pcount;
3016 /* Initial outgoing SYN's get put onto the write_queue
3017 * just like anything else we transmit. It is not
3018 * true data, and if we misinform our callers that
3019 * this ACK acks real data, we will erroneously exit
3020 * connection startup slow start one packet too
3021 * quickly. This is severely frowned upon behavior.
3023 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3024 flag |= FLAG_DATA_ACKED;
3026 flag |= FLAG_SYN_ACKED;
3027 tp->retrans_stamp = 0;
3033 tcp_unlink_write_queue(skb, sk);
3034 sk_wmem_free_skb(sk, skb);
3035 tp->scoreboard_skb_hint = NULL;
3036 if (skb == tp->retransmit_skb_hint)
3037 tp->retransmit_skb_hint = NULL;
3038 if (skb == tp->lost_skb_hint)
3039 tp->lost_skb_hint = NULL;
3042 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3043 tp->snd_up = tp->snd_una;
3045 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3046 flag |= FLAG_SACK_RENEGING;
3048 if (flag & FLAG_ACKED) {
3049 const struct tcp_congestion_ops *ca_ops
3050 = inet_csk(sk)->icsk_ca_ops;
3052 if (unlikely(icsk->icsk_mtup.probe_size &&
3053 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3054 tcp_mtup_probe_success(sk);
3057 tcp_ack_update_rtt(sk, flag, seq_rtt);
3060 if (tcp_is_reno(tp)) {
3061 tcp_remove_reno_sacks(sk, pkts_acked);
3065 /* Non-retransmitted hole got filled? That's reordering */
3066 if (reord < prior_fackets)
3067 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3069 delta = tcp_is_fack(tp) ? pkts_acked :
3070 prior_sacked - tp->sacked_out;
3071 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3074 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3076 if (ca_ops->pkts_acked) {
3079 /* Is the ACK triggering packet unambiguous? */
3080 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3081 /* High resolution needed and available? */
3082 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3083 !ktime_equal(last_ackt,
3084 net_invalid_timestamp()))
3085 rtt_us = ktime_us_delta(ktime_get_real(),
3087 else if (ca_seq_rtt >= 0)
3088 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3091 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3095 #if FASTRETRANS_DEBUG > 0
3096 WARN_ON((int)tp->sacked_out < 0);
3097 WARN_ON((int)tp->lost_out < 0);
3098 WARN_ON((int)tp->retrans_out < 0);
3099 if (!tp->packets_out && tcp_is_sack(tp)) {
3100 icsk = inet_csk(sk);
3102 pr_debug("Leak l=%u %d\n",
3103 tp->lost_out, icsk->icsk_ca_state);
3106 if (tp->sacked_out) {
3107 pr_debug("Leak s=%u %d\n",
3108 tp->sacked_out, icsk->icsk_ca_state);
3111 if (tp->retrans_out) {
3112 pr_debug("Leak r=%u %d\n",
3113 tp->retrans_out, icsk->icsk_ca_state);
3114 tp->retrans_out = 0;
3121 static void tcp_ack_probe(struct sock *sk)
3123 const struct tcp_sock *tp = tcp_sk(sk);
3124 struct inet_connection_sock *icsk = inet_csk(sk);
3126 /* Was it a usable window open? */
3128 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3129 icsk->icsk_backoff = 0;
3130 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3131 /* Socket must be waked up by subsequent tcp_data_snd_check().
3132 * This function is not for random using!
3135 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3136 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3141 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3143 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3144 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3147 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3149 const struct tcp_sock *tp = tcp_sk(sk);
3150 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3151 !tcp_in_cwnd_reduction(sk);
3154 /* Check that window update is acceptable.
3155 * The function assumes that snd_una<=ack<=snd_next.
3157 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3158 const u32 ack, const u32 ack_seq,
3161 return after(ack, tp->snd_una) ||
3162 after(ack_seq, tp->snd_wl1) ||
3163 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3166 /* Update our send window.
3168 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3169 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3171 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3174 struct tcp_sock *tp = tcp_sk(sk);
3176 u32 nwin = ntohs(tcp_hdr(skb)->window);
3178 if (likely(!tcp_hdr(skb)->syn))
3179 nwin <<= tp->rx_opt.snd_wscale;
3181 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3182 flag |= FLAG_WIN_UPDATE;
3183 tcp_update_wl(tp, ack_seq);
3185 if (tp->snd_wnd != nwin) {
3188 /* Note, it is the only place, where
3189 * fast path is recovered for sending TCP.
3192 tcp_fast_path_check(sk);
3194 if (nwin > tp->max_window) {
3195 tp->max_window = nwin;
3196 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3206 /* RFC 5961 7 [ACK Throttling] */
3207 static void tcp_send_challenge_ack(struct sock *sk)
3209 /* unprotected vars, we dont care of overwrites */
3210 static u32 challenge_timestamp;
3211 static unsigned int challenge_count;
3212 u32 now = jiffies / HZ;
3214 if (now != challenge_timestamp) {
3215 challenge_timestamp = now;
3216 challenge_count = 0;
3218 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3219 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3224 /* This routine deals with acks during a TLP episode.
3225 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3227 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3229 struct tcp_sock *tp = tcp_sk(sk);
3230 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3231 !(flag & (FLAG_SND_UNA_ADVANCED |
3232 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3234 /* Mark the end of TLP episode on receiving TLP dupack or when
3235 * ack is after tlp_high_seq.
3237 if (is_tlp_dupack) {
3238 tp->tlp_high_seq = 0;
3242 if (after(ack, tp->tlp_high_seq)) {
3243 tp->tlp_high_seq = 0;
3244 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3245 if (!(flag & FLAG_DSACKING_ACK)) {
3246 tcp_init_cwnd_reduction(sk, true);
3247 tcp_set_ca_state(sk, TCP_CA_CWR);
3248 tcp_end_cwnd_reduction(sk);
3249 tcp_set_ca_state(sk, TCP_CA_Open);
3250 NET_INC_STATS_BH(sock_net(sk),
3251 LINUX_MIB_TCPLOSSPROBERECOVERY);
3256 /* This routine deals with incoming acks, but not outgoing ones. */
3257 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3259 struct inet_connection_sock *icsk = inet_csk(sk);
3260 struct tcp_sock *tp = tcp_sk(sk);
3261 u32 prior_snd_una = tp->snd_una;
3262 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3263 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3264 bool is_dupack = false;
3265 u32 prior_in_flight;
3268 int prior_sacked = tp->sacked_out;
3271 /* If the ack is older than previous acks
3272 * then we can probably ignore it.
3274 if (before(ack, prior_snd_una)) {
3275 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3276 if (before(ack, prior_snd_una - tp->max_window)) {
3277 tcp_send_challenge_ack(sk);
3283 /* If the ack includes data we haven't sent yet, discard
3284 * this segment (RFC793 Section 3.9).
3286 if (after(ack, tp->snd_nxt))
3289 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3290 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3293 if (after(ack, prior_snd_una))
3294 flag |= FLAG_SND_UNA_ADVANCED;
3296 prior_fackets = tp->fackets_out;
3297 prior_in_flight = tcp_packets_in_flight(tp);
3299 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3300 /* Window is constant, pure forward advance.
3301 * No more checks are required.
3302 * Note, we use the fact that SND.UNA>=SND.WL2.
3304 tcp_update_wl(tp, ack_seq);
3306 flag |= FLAG_WIN_UPDATE;
3308 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3310 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3312 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3315 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3317 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3319 if (TCP_SKB_CB(skb)->sacked)
3320 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3322 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3325 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3328 /* We passed data and got it acked, remove any soft error
3329 * log. Something worked...
3331 sk->sk_err_soft = 0;
3332 icsk->icsk_probes_out = 0;
3333 tp->rcv_tstamp = tcp_time_stamp;
3334 prior_packets = tp->packets_out;
3338 /* See if we can take anything off of the retransmit queue. */
3339 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3341 pkts_acked = prior_packets - tp->packets_out;
3343 if (tcp_ack_is_dubious(sk, flag)) {
3344 /* Advance CWND, if state allows this. */
3345 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
3346 tcp_cong_avoid(sk, ack, prior_in_flight);
3347 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3348 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3351 if (flag & FLAG_DATA_ACKED)
3352 tcp_cong_avoid(sk, ack, prior_in_flight);
3355 if (tp->tlp_high_seq)
3356 tcp_process_tlp_ack(sk, ack, flag);
3358 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3359 struct dst_entry *dst = __sk_dst_get(sk);
3364 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3365 tcp_schedule_loss_probe(sk);
3369 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3370 if (flag & FLAG_DSACKING_ACK)
3371 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3373 /* If this ack opens up a zero window, clear backoff. It was
3374 * being used to time the probes, and is probably far higher than
3375 * it needs to be for normal retransmission.
3377 if (tcp_send_head(sk))
3380 if (tp->tlp_high_seq)
3381 tcp_process_tlp_ack(sk, ack, flag);
3385 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3389 /* If data was SACKed, tag it and see if we should send more data.
3390 * If data was DSACKed, see if we can undo a cwnd reduction.
3392 if (TCP_SKB_CB(skb)->sacked) {
3393 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3394 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3398 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3402 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3403 * But, this can also be called on packets in the established flow when
3404 * the fast version below fails.
3406 void tcp_parse_options(const struct sk_buff *skb,
3407 struct tcp_options_received *opt_rx, int estab,
3408 struct tcp_fastopen_cookie *foc)
3410 const unsigned char *ptr;
3411 const struct tcphdr *th = tcp_hdr(skb);
3412 int length = (th->doff * 4) - sizeof(struct tcphdr);
3414 ptr = (const unsigned char *)(th + 1);
3415 opt_rx->saw_tstamp = 0;
3417 while (length > 0) {
3418 int opcode = *ptr++;
3424 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3429 if (opsize < 2) /* "silly options" */
3431 if (opsize > length)
3432 return; /* don't parse partial options */
3435 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3436 u16 in_mss = get_unaligned_be16(ptr);
3438 if (opt_rx->user_mss &&
3439 opt_rx->user_mss < in_mss)
3440 in_mss = opt_rx->user_mss;
3441 opt_rx->mss_clamp = in_mss;
3446 if (opsize == TCPOLEN_WINDOW && th->syn &&
3447 !estab && sysctl_tcp_window_scaling) {
3448 __u8 snd_wscale = *(__u8 *)ptr;
3449 opt_rx->wscale_ok = 1;
3450 if (snd_wscale > 14) {
3451 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3456 opt_rx->snd_wscale = snd_wscale;
3459 case TCPOPT_TIMESTAMP:
3460 if ((opsize == TCPOLEN_TIMESTAMP) &&
3461 ((estab && opt_rx->tstamp_ok) ||
3462 (!estab && sysctl_tcp_timestamps))) {
3463 opt_rx->saw_tstamp = 1;
3464 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3465 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3468 case TCPOPT_SACK_PERM:
3469 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3470 !estab && sysctl_tcp_sack) {
3471 opt_rx->sack_ok = TCP_SACK_SEEN;
3472 tcp_sack_reset(opt_rx);
3477 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3478 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3480 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3483 #ifdef CONFIG_TCP_MD5SIG
3486 * The MD5 Hash has already been
3487 * checked (see tcp_v{4,6}_do_rcv()).
3492 /* Fast Open option shares code 254 using a
3493 * 16 bits magic number. It's valid only in
3494 * SYN or SYN-ACK with an even size.
3496 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3497 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3498 foc == NULL || !th->syn || (opsize & 1))
3500 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3501 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3502 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3503 memcpy(foc->val, ptr + 2, foc->len);
3504 else if (foc->len != 0)
3514 EXPORT_SYMBOL(tcp_parse_options);
3516 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3518 const __be32 *ptr = (const __be32 *)(th + 1);
3520 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3521 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3522 tp->rx_opt.saw_tstamp = 1;
3524 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3526 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3532 /* Fast parse options. This hopes to only see timestamps.
3533 * If it is wrong it falls back on tcp_parse_options().
3535 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3536 const struct tcphdr *th, struct tcp_sock *tp)
3538 /* In the spirit of fast parsing, compare doff directly to constant
3539 * values. Because equality is used, short doff can be ignored here.
3541 if (th->doff == (sizeof(*th) / 4)) {
3542 tp->rx_opt.saw_tstamp = 0;
3544 } else if (tp->rx_opt.tstamp_ok &&
3545 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3546 if (tcp_parse_aligned_timestamp(tp, th))
3550 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3551 if (tp->rx_opt.saw_tstamp)
3552 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3557 #ifdef CONFIG_TCP_MD5SIG
3559 * Parse MD5 Signature option
3561 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3563 int length = (th->doff << 2) - sizeof(*th);
3564 const u8 *ptr = (const u8 *)(th + 1);
3566 /* If the TCP option is too short, we can short cut */
3567 if (length < TCPOLEN_MD5SIG)
3570 while (length > 0) {
3571 int opcode = *ptr++;
3582 if (opsize < 2 || opsize > length)
3584 if (opcode == TCPOPT_MD5SIG)
3585 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3592 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3595 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3597 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3598 tp->rx_opt.ts_recent_stamp = get_seconds();
3601 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3603 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3604 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3605 * extra check below makes sure this can only happen
3606 * for pure ACK frames. -DaveM
3608 * Not only, also it occurs for expired timestamps.
3611 if (tcp_paws_check(&tp->rx_opt, 0))
3612 tcp_store_ts_recent(tp);
3616 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3618 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3619 * it can pass through stack. So, the following predicate verifies that
3620 * this segment is not used for anything but congestion avoidance or
3621 * fast retransmit. Moreover, we even are able to eliminate most of such
3622 * second order effects, if we apply some small "replay" window (~RTO)
3623 * to timestamp space.
3625 * All these measures still do not guarantee that we reject wrapped ACKs
3626 * on networks with high bandwidth, when sequence space is recycled fastly,
3627 * but it guarantees that such events will be very rare and do not affect
3628 * connection seriously. This doesn't look nice, but alas, PAWS is really
3631 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3632 * states that events when retransmit arrives after original data are rare.
3633 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3634 * the biggest problem on large power networks even with minor reordering.
3635 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3636 * up to bandwidth of 18Gigabit/sec. 8) ]
3639 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3641 const struct tcp_sock *tp = tcp_sk(sk);
3642 const struct tcphdr *th = tcp_hdr(skb);
3643 u32 seq = TCP_SKB_CB(skb)->seq;
3644 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3646 return (/* 1. Pure ACK with correct sequence number. */
3647 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3649 /* 2. ... and duplicate ACK. */
3650 ack == tp->snd_una &&
3652 /* 3. ... and does not update window. */
3653 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3655 /* 4. ... and sits in replay window. */
3656 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3659 static inline bool tcp_paws_discard(const struct sock *sk,
3660 const struct sk_buff *skb)
3662 const struct tcp_sock *tp = tcp_sk(sk);
3664 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3665 !tcp_disordered_ack(sk, skb);
3668 /* Check segment sequence number for validity.
3670 * Segment controls are considered valid, if the segment
3671 * fits to the window after truncation to the window. Acceptability
3672 * of data (and SYN, FIN, of course) is checked separately.
3673 * See tcp_data_queue(), for example.
3675 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3676 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3677 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3678 * (borrowed from freebsd)
3681 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3683 return !before(end_seq, tp->rcv_wup) &&
3684 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3687 /* When we get a reset we do this. */
3688 void tcp_reset(struct sock *sk)
3690 /* We want the right error as BSD sees it (and indeed as we do). */
3691 switch (sk->sk_state) {
3693 sk->sk_err = ECONNREFUSED;
3695 case TCP_CLOSE_WAIT:
3701 sk->sk_err = ECONNRESET;
3703 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3706 if (!sock_flag(sk, SOCK_DEAD))
3707 sk->sk_error_report(sk);
3713 * Process the FIN bit. This now behaves as it is supposed to work
3714 * and the FIN takes effect when it is validly part of sequence
3715 * space. Not before when we get holes.
3717 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3718 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3721 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3722 * close and we go into CLOSING (and later onto TIME-WAIT)
3724 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3726 static void tcp_fin(struct sock *sk)
3728 struct tcp_sock *tp = tcp_sk(sk);
3730 inet_csk_schedule_ack(sk);
3732 sk->sk_shutdown |= RCV_SHUTDOWN;
3733 sock_set_flag(sk, SOCK_DONE);
3735 switch (sk->sk_state) {
3737 case TCP_ESTABLISHED:
3738 /* Move to CLOSE_WAIT */
3739 tcp_set_state(sk, TCP_CLOSE_WAIT);
3740 inet_csk(sk)->icsk_ack.pingpong = 1;
3743 case TCP_CLOSE_WAIT:
3745 /* Received a retransmission of the FIN, do
3750 /* RFC793: Remain in the LAST-ACK state. */
3754 /* This case occurs when a simultaneous close
3755 * happens, we must ack the received FIN and
3756 * enter the CLOSING state.
3759 tcp_set_state(sk, TCP_CLOSING);
3762 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3764 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3767 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3768 * cases we should never reach this piece of code.
3770 pr_err("%s: Impossible, sk->sk_state=%d\n",
3771 __func__, sk->sk_state);
3775 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3776 * Probably, we should reset in this case. For now drop them.
3778 __skb_queue_purge(&tp->out_of_order_queue);
3779 if (tcp_is_sack(tp))
3780 tcp_sack_reset(&tp->rx_opt);
3783 if (!sock_flag(sk, SOCK_DEAD)) {
3784 sk->sk_state_change(sk);
3786 /* Do not send POLL_HUP for half duplex close. */
3787 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3788 sk->sk_state == TCP_CLOSE)
3789 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3791 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3795 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3798 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3799 if (before(seq, sp->start_seq))
3800 sp->start_seq = seq;
3801 if (after(end_seq, sp->end_seq))
3802 sp->end_seq = end_seq;
3808 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3810 struct tcp_sock *tp = tcp_sk(sk);
3812 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3815 if (before(seq, tp->rcv_nxt))
3816 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3818 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3820 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3822 tp->rx_opt.dsack = 1;
3823 tp->duplicate_sack[0].start_seq = seq;
3824 tp->duplicate_sack[0].end_seq = end_seq;
3828 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3830 struct tcp_sock *tp = tcp_sk(sk);
3832 if (!tp->rx_opt.dsack)
3833 tcp_dsack_set(sk, seq, end_seq);
3835 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3838 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3840 struct tcp_sock *tp = tcp_sk(sk);
3842 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3843 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3844 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3845 tcp_enter_quickack_mode(sk);
3847 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3848 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3850 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3851 end_seq = tp->rcv_nxt;
3852 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3859 /* These routines update the SACK block as out-of-order packets arrive or
3860 * in-order packets close up the sequence space.
3862 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3865 struct tcp_sack_block *sp = &tp->selective_acks[0];
3866 struct tcp_sack_block *swalk = sp + 1;
3868 /* See if the recent change to the first SACK eats into
3869 * or hits the sequence space of other SACK blocks, if so coalesce.
3871 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3872 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3875 /* Zap SWALK, by moving every further SACK up by one slot.
3876 * Decrease num_sacks.
3878 tp->rx_opt.num_sacks--;
3879 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3883 this_sack++, swalk++;
3887 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3889 struct tcp_sock *tp = tcp_sk(sk);
3890 struct tcp_sack_block *sp = &tp->selective_acks[0];
3891 int cur_sacks = tp->rx_opt.num_sacks;
3897 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3898 if (tcp_sack_extend(sp, seq, end_seq)) {
3899 /* Rotate this_sack to the first one. */
3900 for (; this_sack > 0; this_sack--, sp--)
3901 swap(*sp, *(sp - 1));
3903 tcp_sack_maybe_coalesce(tp);
3908 /* Could not find an adjacent existing SACK, build a new one,
3909 * put it at the front, and shift everyone else down. We
3910 * always know there is at least one SACK present already here.
3912 * If the sack array is full, forget about the last one.
3914 if (this_sack >= TCP_NUM_SACKS) {
3916 tp->rx_opt.num_sacks--;
3919 for (; this_sack > 0; this_sack--, sp--)
3923 /* Build the new head SACK, and we're done. */
3924 sp->start_seq = seq;
3925 sp->end_seq = end_seq;
3926 tp->rx_opt.num_sacks++;
3929 /* RCV.NXT advances, some SACKs should be eaten. */
3931 static void tcp_sack_remove(struct tcp_sock *tp)
3933 struct tcp_sack_block *sp = &tp->selective_acks[0];
3934 int num_sacks = tp->rx_opt.num_sacks;
3937 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3938 if (skb_queue_empty(&tp->out_of_order_queue)) {
3939 tp->rx_opt.num_sacks = 0;
3943 for (this_sack = 0; this_sack < num_sacks;) {
3944 /* Check if the start of the sack is covered by RCV.NXT. */
3945 if (!before(tp->rcv_nxt, sp->start_seq)) {
3948 /* RCV.NXT must cover all the block! */
3949 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
3951 /* Zap this SACK, by moving forward any other SACKS. */
3952 for (i=this_sack+1; i < num_sacks; i++)
3953 tp->selective_acks[i-1] = tp->selective_acks[i];
3960 tp->rx_opt.num_sacks = num_sacks;
3963 /* This one checks to see if we can put data from the
3964 * out_of_order queue into the receive_queue.
3966 static void tcp_ofo_queue(struct sock *sk)
3968 struct tcp_sock *tp = tcp_sk(sk);
3969 __u32 dsack_high = tp->rcv_nxt;
3970 struct sk_buff *skb;
3972 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3973 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3976 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3977 __u32 dsack = dsack_high;
3978 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3979 dsack_high = TCP_SKB_CB(skb)->end_seq;
3980 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
3983 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3984 SOCK_DEBUG(sk, "ofo packet was already received\n");
3985 __skb_unlink(skb, &tp->out_of_order_queue);
3989 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3990 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3991 TCP_SKB_CB(skb)->end_seq);
3993 __skb_unlink(skb, &tp->out_of_order_queue);
3994 __skb_queue_tail(&sk->sk_receive_queue, skb);
3995 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3996 if (tcp_hdr(skb)->fin)
4001 static bool tcp_prune_ofo_queue(struct sock *sk);
4002 static int tcp_prune_queue(struct sock *sk);
4004 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4007 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4008 !sk_rmem_schedule(sk, skb, size)) {
4010 if (tcp_prune_queue(sk) < 0)
4013 if (!sk_rmem_schedule(sk, skb, size)) {
4014 if (!tcp_prune_ofo_queue(sk))
4017 if (!sk_rmem_schedule(sk, skb, size))
4025 * tcp_try_coalesce - try to merge skb to prior one
4028 * @from: buffer to add in queue
4029 * @fragstolen: pointer to boolean
4031 * Before queueing skb @from after @to, try to merge them
4032 * to reduce overall memory use and queue lengths, if cost is small.
4033 * Packets in ofo or receive queues can stay a long time.
4034 * Better try to coalesce them right now to avoid future collapses.
4035 * Returns true if caller should free @from instead of queueing it
4037 static bool tcp_try_coalesce(struct sock *sk,
4039 struct sk_buff *from,
4044 *fragstolen = false;
4046 if (tcp_hdr(from)->fin)
4049 /* Its possible this segment overlaps with prior segment in queue */
4050 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4053 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4056 atomic_add(delta, &sk->sk_rmem_alloc);
4057 sk_mem_charge(sk, delta);
4058 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4059 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4060 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4064 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4066 struct tcp_sock *tp = tcp_sk(sk);
4067 struct sk_buff *skb1;
4070 TCP_ECN_check_ce(tp, skb);
4072 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4073 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4078 /* Disable header prediction. */
4080 inet_csk_schedule_ack(sk);
4082 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4083 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4084 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4086 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4088 /* Initial out of order segment, build 1 SACK. */
4089 if (tcp_is_sack(tp)) {
4090 tp->rx_opt.num_sacks = 1;
4091 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4092 tp->selective_acks[0].end_seq =
4093 TCP_SKB_CB(skb)->end_seq;
4095 __skb_queue_head(&tp->out_of_order_queue, skb);
4099 seq = TCP_SKB_CB(skb)->seq;
4100 end_seq = TCP_SKB_CB(skb)->end_seq;
4102 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4105 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4106 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4108 kfree_skb_partial(skb, fragstolen);
4112 if (!tp->rx_opt.num_sacks ||
4113 tp->selective_acks[0].end_seq != seq)
4116 /* Common case: data arrive in order after hole. */
4117 tp->selective_acks[0].end_seq = end_seq;
4121 /* Find place to insert this segment. */
4123 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4125 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4129 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4132 /* Do skb overlap to previous one? */
4133 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4134 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4135 /* All the bits are present. Drop. */
4136 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4139 tcp_dsack_set(sk, seq, end_seq);
4142 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4143 /* Partial overlap. */
4144 tcp_dsack_set(sk, seq,
4145 TCP_SKB_CB(skb1)->end_seq);
4147 if (skb_queue_is_first(&tp->out_of_order_queue,
4151 skb1 = skb_queue_prev(
4152 &tp->out_of_order_queue,
4157 __skb_queue_head(&tp->out_of_order_queue, skb);
4159 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4161 /* And clean segments covered by new one as whole. */
4162 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4163 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4165 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4167 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4168 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4172 __skb_unlink(skb1, &tp->out_of_order_queue);
4173 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4174 TCP_SKB_CB(skb1)->end_seq);
4175 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4180 if (tcp_is_sack(tp))
4181 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4184 skb_set_owner_r(skb, sk);
4187 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4191 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4193 __skb_pull(skb, hdrlen);
4195 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4196 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4198 __skb_queue_tail(&sk->sk_receive_queue, skb);
4199 skb_set_owner_r(skb, sk);
4204 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4206 struct sk_buff *skb = NULL;
4213 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4217 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4220 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4221 skb_reset_transport_header(skb);
4222 memset(th, 0, sizeof(*th));
4224 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4227 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4228 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4229 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4231 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4232 WARN_ON_ONCE(fragstolen); /* should not happen */
4243 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4245 const struct tcphdr *th = tcp_hdr(skb);
4246 struct tcp_sock *tp = tcp_sk(sk);
4248 bool fragstolen = false;
4250 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4254 __skb_pull(skb, th->doff * 4);
4256 TCP_ECN_accept_cwr(tp, skb);
4258 tp->rx_opt.dsack = 0;
4260 /* Queue data for delivery to the user.
4261 * Packets in sequence go to the receive queue.
4262 * Out of sequence packets to the out_of_order_queue.
4264 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4265 if (tcp_receive_window(tp) == 0)
4268 /* Ok. In sequence. In window. */
4269 if (tp->ucopy.task == current &&
4270 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4271 sock_owned_by_user(sk) && !tp->urg_data) {
4272 int chunk = min_t(unsigned int, skb->len,
4275 __set_current_state(TASK_RUNNING);
4278 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4279 tp->ucopy.len -= chunk;
4280 tp->copied_seq += chunk;
4281 eaten = (chunk == skb->len);
4282 tcp_rcv_space_adjust(sk);
4290 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4293 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4295 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4297 tcp_event_data_recv(sk, skb);
4301 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4304 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4305 * gap in queue is filled.
4307 if (skb_queue_empty(&tp->out_of_order_queue))
4308 inet_csk(sk)->icsk_ack.pingpong = 0;
4311 if (tp->rx_opt.num_sacks)
4312 tcp_sack_remove(tp);
4314 tcp_fast_path_check(sk);
4317 kfree_skb_partial(skb, fragstolen);
4318 if (!sock_flag(sk, SOCK_DEAD))
4319 sk->sk_data_ready(sk, 0);
4323 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4324 /* A retransmit, 2nd most common case. Force an immediate ack. */
4325 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4326 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4329 tcp_enter_quickack_mode(sk);
4330 inet_csk_schedule_ack(sk);
4336 /* Out of window. F.e. zero window probe. */
4337 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4340 tcp_enter_quickack_mode(sk);
4342 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4343 /* Partial packet, seq < rcv_next < end_seq */
4344 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4345 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4346 TCP_SKB_CB(skb)->end_seq);
4348 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4350 /* If window is closed, drop tail of packet. But after
4351 * remembering D-SACK for its head made in previous line.
4353 if (!tcp_receive_window(tp))
4358 tcp_data_queue_ofo(sk, skb);
4361 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4362 struct sk_buff_head *list)
4364 struct sk_buff *next = NULL;
4366 if (!skb_queue_is_last(list, skb))
4367 next = skb_queue_next(list, skb);
4369 __skb_unlink(skb, list);
4371 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4376 /* Collapse contiguous sequence of skbs head..tail with
4377 * sequence numbers start..end.
4379 * If tail is NULL, this means until the end of the list.
4381 * Segments with FIN/SYN are not collapsed (only because this
4385 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4386 struct sk_buff *head, struct sk_buff *tail,
4389 struct sk_buff *skb, *n;
4392 /* First, check that queue is collapsible and find
4393 * the point where collapsing can be useful. */
4397 skb_queue_walk_from_safe(list, skb, n) {
4400 /* No new bits? It is possible on ofo queue. */
4401 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4402 skb = tcp_collapse_one(sk, skb, list);
4408 /* The first skb to collapse is:
4410 * - bloated or contains data before "start" or
4411 * overlaps to the next one.
4413 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4414 (tcp_win_from_space(skb->truesize) > skb->len ||
4415 before(TCP_SKB_CB(skb)->seq, start))) {
4416 end_of_skbs = false;
4420 if (!skb_queue_is_last(list, skb)) {
4421 struct sk_buff *next = skb_queue_next(list, skb);
4423 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4424 end_of_skbs = false;
4429 /* Decided to skip this, advance start seq. */
4430 start = TCP_SKB_CB(skb)->end_seq;
4432 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4435 while (before(start, end)) {
4436 struct sk_buff *nskb;
4437 unsigned int header = skb_headroom(skb);
4438 int copy = SKB_MAX_ORDER(header, 0);
4440 /* Too big header? This can happen with IPv6. */
4443 if (end - start < copy)
4445 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4449 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4450 skb_set_network_header(nskb, (skb_network_header(skb) -
4452 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4454 skb_reserve(nskb, header);
4455 memcpy(nskb->head, skb->head, header);
4456 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4457 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4458 __skb_queue_before(list, skb, nskb);
4459 skb_set_owner_r(nskb, sk);
4461 /* Copy data, releasing collapsed skbs. */
4463 int offset = start - TCP_SKB_CB(skb)->seq;
4464 int size = TCP_SKB_CB(skb)->end_seq - start;
4468 size = min(copy, size);
4469 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4471 TCP_SKB_CB(nskb)->end_seq += size;
4475 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4476 skb = tcp_collapse_one(sk, skb, list);
4479 tcp_hdr(skb)->syn ||
4487 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4488 * and tcp_collapse() them until all the queue is collapsed.
4490 static void tcp_collapse_ofo_queue(struct sock *sk)
4492 struct tcp_sock *tp = tcp_sk(sk);
4493 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4494 struct sk_buff *head;
4500 start = TCP_SKB_CB(skb)->seq;
4501 end = TCP_SKB_CB(skb)->end_seq;
4505 struct sk_buff *next = NULL;
4507 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4508 next = skb_queue_next(&tp->out_of_order_queue, skb);
4511 /* Segment is terminated when we see gap or when
4512 * we are at the end of all the queue. */
4514 after(TCP_SKB_CB(skb)->seq, end) ||
4515 before(TCP_SKB_CB(skb)->end_seq, start)) {
4516 tcp_collapse(sk, &tp->out_of_order_queue,
4517 head, skb, start, end);
4521 /* Start new segment */
4522 start = TCP_SKB_CB(skb)->seq;
4523 end = TCP_SKB_CB(skb)->end_seq;
4525 if (before(TCP_SKB_CB(skb)->seq, start))
4526 start = TCP_SKB_CB(skb)->seq;
4527 if (after(TCP_SKB_CB(skb)->end_seq, end))
4528 end = TCP_SKB_CB(skb)->end_seq;
4534 * Purge the out-of-order queue.
4535 * Return true if queue was pruned.
4537 static bool tcp_prune_ofo_queue(struct sock *sk)
4539 struct tcp_sock *tp = tcp_sk(sk);
4542 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4543 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4544 __skb_queue_purge(&tp->out_of_order_queue);
4546 /* Reset SACK state. A conforming SACK implementation will
4547 * do the same at a timeout based retransmit. When a connection
4548 * is in a sad state like this, we care only about integrity
4549 * of the connection not performance.
4551 if (tp->rx_opt.sack_ok)
4552 tcp_sack_reset(&tp->rx_opt);
4559 /* Reduce allocated memory if we can, trying to get
4560 * the socket within its memory limits again.
4562 * Return less than zero if we should start dropping frames
4563 * until the socket owning process reads some of the data
4564 * to stabilize the situation.
4566 static int tcp_prune_queue(struct sock *sk)
4568 struct tcp_sock *tp = tcp_sk(sk);
4570 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4572 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4574 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4575 tcp_clamp_window(sk);
4576 else if (sk_under_memory_pressure(sk))
4577 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4579 tcp_collapse_ofo_queue(sk);
4580 if (!skb_queue_empty(&sk->sk_receive_queue))
4581 tcp_collapse(sk, &sk->sk_receive_queue,
4582 skb_peek(&sk->sk_receive_queue),
4584 tp->copied_seq, tp->rcv_nxt);
4587 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4590 /* Collapsing did not help, destructive actions follow.
4591 * This must not ever occur. */
4593 tcp_prune_ofo_queue(sk);
4595 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4598 /* If we are really being abused, tell the caller to silently
4599 * drop receive data on the floor. It will get retransmitted
4600 * and hopefully then we'll have sufficient space.
4602 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4604 /* Massive buffer overcommit. */
4609 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4610 * As additional protections, we do not touch cwnd in retransmission phases,
4611 * and if application hit its sndbuf limit recently.
4613 void tcp_cwnd_application_limited(struct sock *sk)
4615 struct tcp_sock *tp = tcp_sk(sk);
4617 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4618 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4619 /* Limited by application or receiver window. */
4620 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4621 u32 win_used = max(tp->snd_cwnd_used, init_win);
4622 if (win_used < tp->snd_cwnd) {
4623 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4624 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4626 tp->snd_cwnd_used = 0;
4628 tp->snd_cwnd_stamp = tcp_time_stamp;
4631 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4633 const struct tcp_sock *tp = tcp_sk(sk);
4635 /* If the user specified a specific send buffer setting, do
4638 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4641 /* If we are under global TCP memory pressure, do not expand. */
4642 if (sk_under_memory_pressure(sk))
4645 /* If we are under soft global TCP memory pressure, do not expand. */
4646 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4649 /* If we filled the congestion window, do not expand. */
4650 if (tp->packets_out >= tp->snd_cwnd)
4656 /* When incoming ACK allowed to free some skb from write_queue,
4657 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4658 * on the exit from tcp input handler.
4660 * PROBLEM: sndbuf expansion does not work well with largesend.
4662 static void tcp_new_space(struct sock *sk)
4664 struct tcp_sock *tp = tcp_sk(sk);
4666 if (tcp_should_expand_sndbuf(sk)) {
4667 int sndmem = SKB_TRUESIZE(max_t(u32,
4668 tp->rx_opt.mss_clamp,
4671 int demanded = max_t(unsigned int, tp->snd_cwnd,
4672 tp->reordering + 1);
4673 sndmem *= 2 * demanded;
4674 if (sndmem > sk->sk_sndbuf)
4675 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4676 tp->snd_cwnd_stamp = tcp_time_stamp;
4679 sk->sk_write_space(sk);
4682 static void tcp_check_space(struct sock *sk)
4684 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4685 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4686 if (sk->sk_socket &&
4687 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4692 static inline void tcp_data_snd_check(struct sock *sk)
4694 tcp_push_pending_frames(sk);
4695 tcp_check_space(sk);
4699 * Check if sending an ack is needed.
4701 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4703 struct tcp_sock *tp = tcp_sk(sk);
4705 /* More than one full frame received... */
4706 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4707 /* ... and right edge of window advances far enough.
4708 * (tcp_recvmsg() will send ACK otherwise). Or...
4710 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4711 /* We ACK each frame or... */
4712 tcp_in_quickack_mode(sk) ||
4713 /* We have out of order data. */
4714 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4715 /* Then ack it now */
4718 /* Else, send delayed ack. */
4719 tcp_send_delayed_ack(sk);
4723 static inline void tcp_ack_snd_check(struct sock *sk)
4725 if (!inet_csk_ack_scheduled(sk)) {
4726 /* We sent a data segment already. */
4729 __tcp_ack_snd_check(sk, 1);
4733 * This routine is only called when we have urgent data
4734 * signaled. Its the 'slow' part of tcp_urg. It could be
4735 * moved inline now as tcp_urg is only called from one
4736 * place. We handle URGent data wrong. We have to - as
4737 * BSD still doesn't use the correction from RFC961.
4738 * For 1003.1g we should support a new option TCP_STDURG to permit
4739 * either form (or just set the sysctl tcp_stdurg).
4742 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4744 struct tcp_sock *tp = tcp_sk(sk);
4745 u32 ptr = ntohs(th->urg_ptr);
4747 if (ptr && !sysctl_tcp_stdurg)
4749 ptr += ntohl(th->seq);
4751 /* Ignore urgent data that we've already seen and read. */
4752 if (after(tp->copied_seq, ptr))
4755 /* Do not replay urg ptr.
4757 * NOTE: interesting situation not covered by specs.
4758 * Misbehaving sender may send urg ptr, pointing to segment,
4759 * which we already have in ofo queue. We are not able to fetch
4760 * such data and will stay in TCP_URG_NOTYET until will be eaten
4761 * by recvmsg(). Seems, we are not obliged to handle such wicked
4762 * situations. But it is worth to think about possibility of some
4763 * DoSes using some hypothetical application level deadlock.
4765 if (before(ptr, tp->rcv_nxt))
4768 /* Do we already have a newer (or duplicate) urgent pointer? */
4769 if (tp->urg_data && !after(ptr, tp->urg_seq))
4772 /* Tell the world about our new urgent pointer. */
4775 /* We may be adding urgent data when the last byte read was
4776 * urgent. To do this requires some care. We cannot just ignore
4777 * tp->copied_seq since we would read the last urgent byte again
4778 * as data, nor can we alter copied_seq until this data arrives
4779 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4781 * NOTE. Double Dutch. Rendering to plain English: author of comment
4782 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4783 * and expect that both A and B disappear from stream. This is _wrong_.
4784 * Though this happens in BSD with high probability, this is occasional.
4785 * Any application relying on this is buggy. Note also, that fix "works"
4786 * only in this artificial test. Insert some normal data between A and B and we will
4787 * decline of BSD again. Verdict: it is better to remove to trap
4790 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4791 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4792 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4794 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4795 __skb_unlink(skb, &sk->sk_receive_queue);
4800 tp->urg_data = TCP_URG_NOTYET;
4803 /* Disable header prediction. */
4807 /* This is the 'fast' part of urgent handling. */
4808 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4810 struct tcp_sock *tp = tcp_sk(sk);
4812 /* Check if we get a new urgent pointer - normally not. */
4814 tcp_check_urg(sk, th);
4816 /* Do we wait for any urgent data? - normally not... */
4817 if (tp->urg_data == TCP_URG_NOTYET) {
4818 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4821 /* Is the urgent pointer pointing into this packet? */
4822 if (ptr < skb->len) {
4824 if (skb_copy_bits(skb, ptr, &tmp, 1))
4826 tp->urg_data = TCP_URG_VALID | tmp;
4827 if (!sock_flag(sk, SOCK_DEAD))
4828 sk->sk_data_ready(sk, 0);
4833 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4835 struct tcp_sock *tp = tcp_sk(sk);
4836 int chunk = skb->len - hlen;
4840 if (skb_csum_unnecessary(skb))
4841 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4843 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4847 tp->ucopy.len -= chunk;
4848 tp->copied_seq += chunk;
4849 tcp_rcv_space_adjust(sk);
4856 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4857 struct sk_buff *skb)
4861 if (sock_owned_by_user(sk)) {
4863 result = __tcp_checksum_complete(skb);
4866 result = __tcp_checksum_complete(skb);
4871 static inline bool tcp_checksum_complete_user(struct sock *sk,
4872 struct sk_buff *skb)
4874 return !skb_csum_unnecessary(skb) &&
4875 __tcp_checksum_complete_user(sk, skb);
4878 #ifdef CONFIG_NET_DMA
4879 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4882 struct tcp_sock *tp = tcp_sk(sk);
4883 int chunk = skb->len - hlen;
4885 bool copied_early = false;
4887 if (tp->ucopy.wakeup)
4890 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4891 tp->ucopy.dma_chan = net_dma_find_channel();
4893 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4895 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4897 tp->ucopy.iov, chunk,
4898 tp->ucopy.pinned_list);
4903 tp->ucopy.dma_cookie = dma_cookie;
4904 copied_early = true;
4906 tp->ucopy.len -= chunk;
4907 tp->copied_seq += chunk;
4908 tcp_rcv_space_adjust(sk);
4910 if ((tp->ucopy.len == 0) ||
4911 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4912 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4913 tp->ucopy.wakeup = 1;
4914 sk->sk_data_ready(sk, 0);
4916 } else if (chunk > 0) {
4917 tp->ucopy.wakeup = 1;
4918 sk->sk_data_ready(sk, 0);
4921 return copied_early;
4923 #endif /* CONFIG_NET_DMA */
4925 /* Does PAWS and seqno based validation of an incoming segment, flags will
4926 * play significant role here.
4928 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4929 const struct tcphdr *th, int syn_inerr)
4931 struct tcp_sock *tp = tcp_sk(sk);
4933 /* RFC1323: H1. Apply PAWS check first. */
4934 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4935 tcp_paws_discard(sk, skb)) {
4937 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4938 tcp_send_dupack(sk, skb);
4941 /* Reset is accepted even if it did not pass PAWS. */
4944 /* Step 1: check sequence number */
4945 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4946 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4947 * (RST) segments are validated by checking their SEQ-fields."
4948 * And page 69: "If an incoming segment is not acceptable,
4949 * an acknowledgment should be sent in reply (unless the RST
4950 * bit is set, if so drop the segment and return)".
4955 tcp_send_dupack(sk, skb);
4960 /* Step 2: check RST bit */
4963 * If sequence number exactly matches RCV.NXT, then
4964 * RESET the connection
4966 * Send a challenge ACK
4968 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
4971 tcp_send_challenge_ack(sk);
4975 /* step 3: check security and precedence [ignored] */
4977 /* step 4: Check for a SYN
4978 * RFC 5691 4.2 : Send a challenge ack
4983 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4984 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
4985 tcp_send_challenge_ack(sk);
4997 * TCP receive function for the ESTABLISHED state.
4999 * It is split into a fast path and a slow path. The fast path is
5001 * - A zero window was announced from us - zero window probing
5002 * is only handled properly in the slow path.
5003 * - Out of order segments arrived.
5004 * - Urgent data is expected.
5005 * - There is no buffer space left
5006 * - Unexpected TCP flags/window values/header lengths are received
5007 * (detected by checking the TCP header against pred_flags)
5008 * - Data is sent in both directions. Fast path only supports pure senders
5009 * or pure receivers (this means either the sequence number or the ack
5010 * value must stay constant)
5011 * - Unexpected TCP option.
5013 * When these conditions are not satisfied it drops into a standard
5014 * receive procedure patterned after RFC793 to handle all cases.
5015 * The first three cases are guaranteed by proper pred_flags setting,
5016 * the rest is checked inline. Fast processing is turned on in
5017 * tcp_data_queue when everything is OK.
5019 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5020 const struct tcphdr *th, unsigned int len)
5022 struct tcp_sock *tp = tcp_sk(sk);
5024 if (unlikely(sk->sk_rx_dst == NULL))
5025 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5027 * Header prediction.
5028 * The code loosely follows the one in the famous
5029 * "30 instruction TCP receive" Van Jacobson mail.
5031 * Van's trick is to deposit buffers into socket queue
5032 * on a device interrupt, to call tcp_recv function
5033 * on the receive process context and checksum and copy
5034 * the buffer to user space. smart...
5036 * Our current scheme is not silly either but we take the
5037 * extra cost of the net_bh soft interrupt processing...
5038 * We do checksum and copy also but from device to kernel.
5041 tp->rx_opt.saw_tstamp = 0;
5043 /* pred_flags is 0xS?10 << 16 + snd_wnd
5044 * if header_prediction is to be made
5045 * 'S' will always be tp->tcp_header_len >> 2
5046 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5047 * turn it off (when there are holes in the receive
5048 * space for instance)
5049 * PSH flag is ignored.
5052 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5053 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5054 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5055 int tcp_header_len = tp->tcp_header_len;
5057 /* Timestamp header prediction: tcp_header_len
5058 * is automatically equal to th->doff*4 due to pred_flags
5062 /* Check timestamp */
5063 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5064 /* No? Slow path! */
5065 if (!tcp_parse_aligned_timestamp(tp, th))
5068 /* If PAWS failed, check it more carefully in slow path */
5069 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5072 /* DO NOT update ts_recent here, if checksum fails
5073 * and timestamp was corrupted part, it will result
5074 * in a hung connection since we will drop all
5075 * future packets due to the PAWS test.
5079 if (len <= tcp_header_len) {
5080 /* Bulk data transfer: sender */
5081 if (len == tcp_header_len) {
5082 /* Predicted packet is in window by definition.
5083 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5084 * Hence, check seq<=rcv_wup reduces to:
5086 if (tcp_header_len ==
5087 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5088 tp->rcv_nxt == tp->rcv_wup)
5089 tcp_store_ts_recent(tp);
5091 /* We know that such packets are checksummed
5094 tcp_ack(sk, skb, 0);
5096 tcp_data_snd_check(sk);
5098 } else { /* Header too small */
5099 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5104 int copied_early = 0;
5105 bool fragstolen = false;
5107 if (tp->copied_seq == tp->rcv_nxt &&
5108 len - tcp_header_len <= tp->ucopy.len) {
5109 #ifdef CONFIG_NET_DMA
5110 if (tp->ucopy.task == current &&
5111 sock_owned_by_user(sk) &&
5112 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5117 if (tp->ucopy.task == current &&
5118 sock_owned_by_user(sk) && !copied_early) {
5119 __set_current_state(TASK_RUNNING);
5121 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5125 /* Predicted packet is in window by definition.
5126 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5127 * Hence, check seq<=rcv_wup reduces to:
5129 if (tcp_header_len ==
5130 (sizeof(struct tcphdr) +
5131 TCPOLEN_TSTAMP_ALIGNED) &&
5132 tp->rcv_nxt == tp->rcv_wup)
5133 tcp_store_ts_recent(tp);
5135 tcp_rcv_rtt_measure_ts(sk, skb);
5137 __skb_pull(skb, tcp_header_len);
5138 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5139 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5142 tcp_cleanup_rbuf(sk, skb->len);
5145 if (tcp_checksum_complete_user(sk, skb))
5148 if ((int)skb->truesize > sk->sk_forward_alloc)
5151 /* Predicted packet is in window by definition.
5152 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5153 * Hence, check seq<=rcv_wup reduces to:
5155 if (tcp_header_len ==
5156 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5157 tp->rcv_nxt == tp->rcv_wup)
5158 tcp_store_ts_recent(tp);
5160 tcp_rcv_rtt_measure_ts(sk, skb);
5162 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5164 /* Bulk data transfer: receiver */
5165 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5169 tcp_event_data_recv(sk, skb);
5171 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5172 /* Well, only one small jumplet in fast path... */
5173 tcp_ack(sk, skb, FLAG_DATA);
5174 tcp_data_snd_check(sk);
5175 if (!inet_csk_ack_scheduled(sk))
5179 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5180 __tcp_ack_snd_check(sk, 0);
5182 #ifdef CONFIG_NET_DMA
5184 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5188 kfree_skb_partial(skb, fragstolen);
5189 sk->sk_data_ready(sk, 0);
5195 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5198 if (!th->ack && !th->rst)
5202 * Standard slow path.
5205 if (!tcp_validate_incoming(sk, skb, th, 1))
5209 if (tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5212 /* ts_recent update must be made after we are sure that the packet
5215 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5217 tcp_rcv_rtt_measure_ts(sk, skb);
5219 /* Process urgent data. */
5220 tcp_urg(sk, skb, th);
5222 /* step 7: process the segment text */
5223 tcp_data_queue(sk, skb);
5225 tcp_data_snd_check(sk);
5226 tcp_ack_snd_check(sk);
5230 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5236 EXPORT_SYMBOL(tcp_rcv_established);
5238 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5240 struct tcp_sock *tp = tcp_sk(sk);
5241 struct inet_connection_sock *icsk = inet_csk(sk);
5243 tcp_set_state(sk, TCP_ESTABLISHED);
5246 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5247 security_inet_conn_established(sk, skb);
5250 /* Make sure socket is routed, for correct metrics. */
5251 icsk->icsk_af_ops->rebuild_header(sk);
5253 tcp_init_metrics(sk);
5255 tcp_init_congestion_control(sk);
5257 /* Prevent spurious tcp_cwnd_restart() on first data
5260 tp->lsndtime = tcp_time_stamp;
5262 tcp_init_buffer_space(sk);
5264 if (sock_flag(sk, SOCK_KEEPOPEN))
5265 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5267 if (!tp->rx_opt.snd_wscale)
5268 __tcp_fast_path_on(tp, tp->snd_wnd);
5272 if (!sock_flag(sk, SOCK_DEAD)) {
5273 sk->sk_state_change(sk);
5274 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5278 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5279 struct tcp_fastopen_cookie *cookie)
5281 struct tcp_sock *tp = tcp_sk(sk);
5282 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5283 u16 mss = tp->rx_opt.mss_clamp;
5286 if (mss == tp->rx_opt.user_mss) {
5287 struct tcp_options_received opt;
5289 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5290 tcp_clear_options(&opt);
5291 opt.user_mss = opt.mss_clamp = 0;
5292 tcp_parse_options(synack, &opt, 0, NULL);
5293 mss = opt.mss_clamp;
5296 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5299 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5300 * the remote receives only the retransmitted (regular) SYNs: either
5301 * the original SYN-data or the corresponding SYN-ACK is lost.
5303 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5305 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5307 if (data) { /* Retransmit unacked data in SYN */
5308 tcp_for_write_queue_from(data, sk) {
5309 if (data == tcp_send_head(sk) ||
5310 __tcp_retransmit_skb(sk, data))
5316 tp->syn_data_acked = tp->syn_data;
5320 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5321 const struct tcphdr *th, unsigned int len)
5323 struct inet_connection_sock *icsk = inet_csk(sk);
5324 struct tcp_sock *tp = tcp_sk(sk);
5325 struct tcp_fastopen_cookie foc = { .len = -1 };
5326 int saved_clamp = tp->rx_opt.mss_clamp;
5328 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5329 if (tp->rx_opt.saw_tstamp)
5330 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5334 * "If the state is SYN-SENT then
5335 * first check the ACK bit
5336 * If the ACK bit is set
5337 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5338 * a reset (unless the RST bit is set, if so drop
5339 * the segment and return)"
5341 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5342 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5343 goto reset_and_undo;
5345 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5346 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5348 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5349 goto reset_and_undo;
5352 /* Now ACK is acceptable.
5354 * "If the RST bit is set
5355 * If the ACK was acceptable then signal the user "error:
5356 * connection reset", drop the segment, enter CLOSED state,
5357 * delete TCB, and return."
5366 * "fifth, if neither of the SYN or RST bits is set then
5367 * drop the segment and return."
5373 goto discard_and_undo;
5376 * "If the SYN bit is on ...
5377 * are acceptable then ...
5378 * (our SYN has been ACKed), change the connection
5379 * state to ESTABLISHED..."
5382 TCP_ECN_rcv_synack(tp, th);
5384 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5385 tcp_ack(sk, skb, FLAG_SLOWPATH);
5387 /* Ok.. it's good. Set up sequence numbers and
5388 * move to established.
5390 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5391 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5393 /* RFC1323: The window in SYN & SYN/ACK segments is
5396 tp->snd_wnd = ntohs(th->window);
5398 if (!tp->rx_opt.wscale_ok) {
5399 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5400 tp->window_clamp = min(tp->window_clamp, 65535U);
5403 if (tp->rx_opt.saw_tstamp) {
5404 tp->rx_opt.tstamp_ok = 1;
5405 tp->tcp_header_len =
5406 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5407 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5408 tcp_store_ts_recent(tp);
5410 tp->tcp_header_len = sizeof(struct tcphdr);
5413 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5414 tcp_enable_fack(tp);
5417 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5418 tcp_initialize_rcv_mss(sk);
5420 /* Remember, tcp_poll() does not lock socket!
5421 * Change state from SYN-SENT only after copied_seq
5422 * is initialized. */
5423 tp->copied_seq = tp->rcv_nxt;
5427 tcp_finish_connect(sk, skb);
5429 if ((tp->syn_fastopen || tp->syn_data) &&
5430 tcp_rcv_fastopen_synack(sk, skb, &foc))
5433 if (sk->sk_write_pending ||
5434 icsk->icsk_accept_queue.rskq_defer_accept ||
5435 icsk->icsk_ack.pingpong) {
5436 /* Save one ACK. Data will be ready after
5437 * several ticks, if write_pending is set.
5439 * It may be deleted, but with this feature tcpdumps
5440 * look so _wonderfully_ clever, that I was not able
5441 * to stand against the temptation 8) --ANK
5443 inet_csk_schedule_ack(sk);
5444 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5445 tcp_enter_quickack_mode(sk);
5446 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5447 TCP_DELACK_MAX, TCP_RTO_MAX);
5458 /* No ACK in the segment */
5462 * "If the RST bit is set
5464 * Otherwise (no ACK) drop the segment and return."
5467 goto discard_and_undo;
5471 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5472 tcp_paws_reject(&tp->rx_opt, 0))
5473 goto discard_and_undo;
5476 /* We see SYN without ACK. It is attempt of
5477 * simultaneous connect with crossed SYNs.
5478 * Particularly, it can be connect to self.
5480 tcp_set_state(sk, TCP_SYN_RECV);
5482 if (tp->rx_opt.saw_tstamp) {
5483 tp->rx_opt.tstamp_ok = 1;
5484 tcp_store_ts_recent(tp);
5485 tp->tcp_header_len =
5486 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5488 tp->tcp_header_len = sizeof(struct tcphdr);
5491 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5492 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5494 /* RFC1323: The window in SYN & SYN/ACK segments is
5497 tp->snd_wnd = ntohs(th->window);
5498 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5499 tp->max_window = tp->snd_wnd;
5501 TCP_ECN_rcv_syn(tp, th);
5504 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5505 tcp_initialize_rcv_mss(sk);
5507 tcp_send_synack(sk);
5509 /* Note, we could accept data and URG from this segment.
5510 * There are no obstacles to make this (except that we must
5511 * either change tcp_recvmsg() to prevent it from returning data
5512 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5514 * However, if we ignore data in ACKless segments sometimes,
5515 * we have no reasons to accept it sometimes.
5516 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5517 * is not flawless. So, discard packet for sanity.
5518 * Uncomment this return to process the data.
5525 /* "fifth, if neither of the SYN or RST bits is set then
5526 * drop the segment and return."
5530 tcp_clear_options(&tp->rx_opt);
5531 tp->rx_opt.mss_clamp = saved_clamp;
5535 tcp_clear_options(&tp->rx_opt);
5536 tp->rx_opt.mss_clamp = saved_clamp;
5541 * This function implements the receiving procedure of RFC 793 for
5542 * all states except ESTABLISHED and TIME_WAIT.
5543 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5544 * address independent.
5547 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5548 const struct tcphdr *th, unsigned int len)
5550 struct tcp_sock *tp = tcp_sk(sk);
5551 struct inet_connection_sock *icsk = inet_csk(sk);
5552 struct request_sock *req;
5555 tp->rx_opt.saw_tstamp = 0;
5557 switch (sk->sk_state) {
5571 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5574 /* Now we have several options: In theory there is
5575 * nothing else in the frame. KA9Q has an option to
5576 * send data with the syn, BSD accepts data with the
5577 * syn up to the [to be] advertised window and
5578 * Solaris 2.1 gives you a protocol error. For now
5579 * we just ignore it, that fits the spec precisely
5580 * and avoids incompatibilities. It would be nice in
5581 * future to drop through and process the data.
5583 * Now that TTCP is starting to be used we ought to
5585 * But, this leaves one open to an easy denial of
5586 * service attack, and SYN cookies can't defend
5587 * against this problem. So, we drop the data
5588 * in the interest of security over speed unless
5589 * it's still in use.
5597 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5601 /* Do step6 onward by hand. */
5602 tcp_urg(sk, skb, th);
5604 tcp_data_snd_check(sk);
5608 req = tp->fastopen_rsk;
5610 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5611 sk->sk_state != TCP_FIN_WAIT1);
5613 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5617 if (!th->ack && !th->rst)
5620 if (!tcp_validate_incoming(sk, skb, th, 0))
5623 /* step 5: check the ACK field */
5625 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5627 switch (sk->sk_state) {
5630 /* Once we leave TCP_SYN_RECV, we no longer
5631 * need req so release it.
5634 tcp_synack_rtt_meas(sk, req);
5635 tp->total_retrans = req->num_retrans;
5637 reqsk_fastopen_remove(sk, req, false);
5639 /* Make sure socket is routed, for
5642 icsk->icsk_af_ops->rebuild_header(sk);
5643 tcp_init_congestion_control(sk);
5646 tcp_init_buffer_space(sk);
5647 tp->copied_seq = tp->rcv_nxt;
5650 tcp_set_state(sk, TCP_ESTABLISHED);
5651 sk->sk_state_change(sk);
5653 /* Note, that this wakeup is only for marginal
5654 * crossed SYN case. Passively open sockets
5655 * are not waked up, because sk->sk_sleep ==
5656 * NULL and sk->sk_socket == NULL.
5660 SOCK_WAKE_IO, POLL_OUT);
5662 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5663 tp->snd_wnd = ntohs(th->window) <<
5664 tp->rx_opt.snd_wscale;
5665 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5667 if (tp->rx_opt.tstamp_ok)
5668 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5671 /* Re-arm the timer because data may
5672 * have been sent out. This is similar
5673 * to the regular data transmission case
5674 * when new data has just been ack'ed.
5676 * (TFO) - we could try to be more
5677 * aggressive and retranmitting any data
5678 * sooner based on when they were sent
5683 tcp_init_metrics(sk);
5685 /* Prevent spurious tcp_cwnd_restart() on
5686 * first data packet.
5688 tp->lsndtime = tcp_time_stamp;
5690 tcp_initialize_rcv_mss(sk);
5691 tcp_fast_path_on(tp);
5698 /* If we enter the TCP_FIN_WAIT1 state and we are a
5699 * Fast Open socket and this is the first acceptable
5700 * ACK we have received, this would have acknowledged
5701 * our SYNACK so stop the SYNACK timer.
5704 /* Return RST if ack_seq is invalid.
5705 * Note that RFC793 only says to generate a
5706 * DUPACK for it but for TCP Fast Open it seems
5707 * better to treat this case like TCP_SYN_RECV
5712 /* We no longer need the request sock. */
5713 reqsk_fastopen_remove(sk, req, false);
5716 if (tp->snd_una == tp->write_seq) {
5717 struct dst_entry *dst;
5719 tcp_set_state(sk, TCP_FIN_WAIT2);
5720 sk->sk_shutdown |= SEND_SHUTDOWN;
5722 dst = __sk_dst_get(sk);
5726 if (!sock_flag(sk, SOCK_DEAD))
5727 /* Wake up lingering close() */
5728 sk->sk_state_change(sk);
5732 if (tp->linger2 < 0 ||
5733 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5734 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5736 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5740 tmo = tcp_fin_time(sk);
5741 if (tmo > TCP_TIMEWAIT_LEN) {
5742 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5743 } else if (th->fin || sock_owned_by_user(sk)) {
5744 /* Bad case. We could lose such FIN otherwise.
5745 * It is not a big problem, but it looks confusing
5746 * and not so rare event. We still can lose it now,
5747 * if it spins in bh_lock_sock(), but it is really
5750 inet_csk_reset_keepalive_timer(sk, tmo);
5752 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5760 if (tp->snd_una == tp->write_seq) {
5761 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5767 if (tp->snd_una == tp->write_seq) {
5768 tcp_update_metrics(sk);
5776 /* ts_recent update must be made after we are sure that the packet
5779 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5781 /* step 6: check the URG bit */
5782 tcp_urg(sk, skb, th);
5784 /* step 7: process the segment text */
5785 switch (sk->sk_state) {
5786 case TCP_CLOSE_WAIT:
5789 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5793 /* RFC 793 says to queue data in these states,
5794 * RFC 1122 says we MUST send a reset.
5795 * BSD 4.4 also does reset.
5797 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5798 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5799 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5800 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5806 case TCP_ESTABLISHED:
5807 tcp_data_queue(sk, skb);
5812 /* tcp_data could move socket to TIME-WAIT */
5813 if (sk->sk_state != TCP_CLOSE) {
5814 tcp_data_snd_check(sk);
5815 tcp_ack_snd_check(sk);
5824 EXPORT_SYMBOL(tcp_rcv_state_process);