2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_ecn __read_mostly = 2;
85 EXPORT_SYMBOL(sysctl_tcp_ecn);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit = 100;
94 int sysctl_tcp_stdurg __read_mostly;
95 int sysctl_tcp_rfc1337 __read_mostly;
96 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
97 int sysctl_tcp_frto __read_mostly = 2;
98 int sysctl_tcp_frto_response __read_mostly;
100 int sysctl_tcp_thin_dupack __read_mostly;
102 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
103 int sysctl_tcp_abc __read_mostly;
104 int sysctl_tcp_early_retrans __read_mostly = 2;
106 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
107 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
108 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
109 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
110 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
111 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
112 #define FLAG_ECE 0x40 /* ECE in this ACK */
113 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
114 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
115 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
116 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
117 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
118 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
120 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
123 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
126 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
127 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
129 /* Adapt the MSS value used to make delayed ack decision to the
132 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
134 struct inet_connection_sock *icsk = inet_csk(sk);
135 const unsigned int lss = icsk->icsk_ack.last_seg_size;
138 icsk->icsk_ack.last_seg_size = 0;
140 /* skb->len may jitter because of SACKs, even if peer
141 * sends good full-sized frames.
143 len = skb_shinfo(skb)->gso_size ? : skb->len;
144 if (len >= icsk->icsk_ack.rcv_mss) {
145 icsk->icsk_ack.rcv_mss = len;
147 /* Otherwise, we make more careful check taking into account,
148 * that SACKs block is variable.
150 * "len" is invariant segment length, including TCP header.
152 len += skb->data - skb_transport_header(skb);
153 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
154 /* If PSH is not set, packet should be
155 * full sized, provided peer TCP is not badly broken.
156 * This observation (if it is correct 8)) allows
157 * to handle super-low mtu links fairly.
159 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
160 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
161 /* Subtract also invariant (if peer is RFC compliant),
162 * tcp header plus fixed timestamp option length.
163 * Resulting "len" is MSS free of SACK jitter.
165 len -= tcp_sk(sk)->tcp_header_len;
166 icsk->icsk_ack.last_seg_size = len;
168 icsk->icsk_ack.rcv_mss = len;
172 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
173 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
174 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
178 static void tcp_incr_quickack(struct sock *sk)
180 struct inet_connection_sock *icsk = inet_csk(sk);
181 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
185 if (quickacks > icsk->icsk_ack.quick)
186 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
189 static void tcp_enter_quickack_mode(struct sock *sk)
191 struct inet_connection_sock *icsk = inet_csk(sk);
192 tcp_incr_quickack(sk);
193 icsk->icsk_ack.pingpong = 0;
194 icsk->icsk_ack.ato = TCP_ATO_MIN;
197 /* Send ACKs quickly, if "quick" count is not exhausted
198 * and the session is not interactive.
201 static inline bool tcp_in_quickack_mode(const struct sock *sk)
203 const struct inet_connection_sock *icsk = inet_csk(sk);
205 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
208 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
210 if (tp->ecn_flags & TCP_ECN_OK)
211 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
214 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
216 if (tcp_hdr(skb)->cwr)
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
222 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
225 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
227 if (!(tp->ecn_flags & TCP_ECN_OK))
230 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
231 case INET_ECN_NOT_ECT:
232 /* Funny extension: if ECT is not set on a segment,
233 * and we already seen ECT on a previous segment,
234 * it is probably a retransmit.
236 if (tp->ecn_flags & TCP_ECN_SEEN)
237 tcp_enter_quickack_mode((struct sock *)tp);
240 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
241 /* Better not delay acks, sender can have a very low cwnd */
242 tcp_enter_quickack_mode((struct sock *)tp);
243 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
247 tp->ecn_flags |= TCP_ECN_SEEN;
251 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
253 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
254 tp->ecn_flags &= ~TCP_ECN_OK;
257 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
259 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
260 tp->ecn_flags &= ~TCP_ECN_OK;
263 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
265 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
270 /* Buffer size and advertised window tuning.
272 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
275 static void tcp_fixup_sndbuf(struct sock *sk)
277 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
279 sndmem *= TCP_INIT_CWND;
280 if (sk->sk_sndbuf < sndmem)
281 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
284 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
286 * All tcp_full_space() is split to two parts: "network" buffer, allocated
287 * forward and advertised in receiver window (tp->rcv_wnd) and
288 * "application buffer", required to isolate scheduling/application
289 * latencies from network.
290 * window_clamp is maximal advertised window. It can be less than
291 * tcp_full_space(), in this case tcp_full_space() - window_clamp
292 * is reserved for "application" buffer. The less window_clamp is
293 * the smoother our behaviour from viewpoint of network, but the lower
294 * throughput and the higher sensitivity of the connection to losses. 8)
296 * rcv_ssthresh is more strict window_clamp used at "slow start"
297 * phase to predict further behaviour of this connection.
298 * It is used for two goals:
299 * - to enforce header prediction at sender, even when application
300 * requires some significant "application buffer". It is check #1.
301 * - to prevent pruning of receive queue because of misprediction
302 * of receiver window. Check #2.
304 * The scheme does not work when sender sends good segments opening
305 * window and then starts to feed us spaghetti. But it should work
306 * in common situations. Otherwise, we have to rely on queue collapsing.
309 /* Slow part of check#2. */
310 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
312 struct tcp_sock *tp = tcp_sk(sk);
314 int truesize = tcp_win_from_space(skb->truesize) >> 1;
315 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
317 while (tp->rcv_ssthresh <= window) {
318 if (truesize <= skb->len)
319 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
327 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
329 struct tcp_sock *tp = tcp_sk(sk);
332 if (tp->rcv_ssthresh < tp->window_clamp &&
333 (int)tp->rcv_ssthresh < tcp_space(sk) &&
334 !sk_under_memory_pressure(sk)) {
337 /* Check #2. Increase window, if skb with such overhead
338 * will fit to rcvbuf in future.
340 if (tcp_win_from_space(skb->truesize) <= skb->len)
341 incr = 2 * tp->advmss;
343 incr = __tcp_grow_window(sk, skb);
346 incr = max_t(int, incr, 2 * skb->len);
347 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
349 inet_csk(sk)->icsk_ack.quick |= 1;
354 /* 3. Tuning rcvbuf, when connection enters established state. */
356 static void tcp_fixup_rcvbuf(struct sock *sk)
358 u32 mss = tcp_sk(sk)->advmss;
359 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
362 /* Limit to 10 segments if mss <= 1460,
363 * or 14600/mss segments, with a minimum of two segments.
366 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
368 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
369 while (tcp_win_from_space(rcvmem) < mss)
374 if (sk->sk_rcvbuf < rcvmem)
375 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
378 /* 4. Try to fixup all. It is made immediately after connection enters
381 static void tcp_init_buffer_space(struct sock *sk)
383 struct tcp_sock *tp = tcp_sk(sk);
386 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
387 tcp_fixup_rcvbuf(sk);
388 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
389 tcp_fixup_sndbuf(sk);
391 tp->rcvq_space.space = tp->rcv_wnd;
393 maxwin = tcp_full_space(sk);
395 if (tp->window_clamp >= maxwin) {
396 tp->window_clamp = maxwin;
398 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
399 tp->window_clamp = max(maxwin -
400 (maxwin >> sysctl_tcp_app_win),
404 /* Force reservation of one segment. */
405 if (sysctl_tcp_app_win &&
406 tp->window_clamp > 2 * tp->advmss &&
407 tp->window_clamp + tp->advmss > maxwin)
408 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
410 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
411 tp->snd_cwnd_stamp = tcp_time_stamp;
414 /* 5. Recalculate window clamp after socket hit its memory bounds. */
415 static void tcp_clamp_window(struct sock *sk)
417 struct tcp_sock *tp = tcp_sk(sk);
418 struct inet_connection_sock *icsk = inet_csk(sk);
420 icsk->icsk_ack.quick = 0;
422 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
423 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
424 !sk_under_memory_pressure(sk) &&
425 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
426 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
429 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
430 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
433 /* Initialize RCV_MSS value.
434 * RCV_MSS is an our guess about MSS used by the peer.
435 * We haven't any direct information about the MSS.
436 * It's better to underestimate the RCV_MSS rather than overestimate.
437 * Overestimations make us ACKing less frequently than needed.
438 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
440 void tcp_initialize_rcv_mss(struct sock *sk)
442 const struct tcp_sock *tp = tcp_sk(sk);
443 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
445 hint = min(hint, tp->rcv_wnd / 2);
446 hint = min(hint, TCP_MSS_DEFAULT);
447 hint = max(hint, TCP_MIN_MSS);
449 inet_csk(sk)->icsk_ack.rcv_mss = hint;
451 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
453 /* Receiver "autotuning" code.
455 * The algorithm for RTT estimation w/o timestamps is based on
456 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
457 * <http://public.lanl.gov/radiant/pubs.html#DRS>
459 * More detail on this code can be found at
460 * <http://staff.psc.edu/jheffner/>,
461 * though this reference is out of date. A new paper
464 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
466 u32 new_sample = tp->rcv_rtt_est.rtt;
472 if (new_sample != 0) {
473 /* If we sample in larger samples in the non-timestamp
474 * case, we could grossly overestimate the RTT especially
475 * with chatty applications or bulk transfer apps which
476 * are stalled on filesystem I/O.
478 * Also, since we are only going for a minimum in the
479 * non-timestamp case, we do not smooth things out
480 * else with timestamps disabled convergence takes too
484 m -= (new_sample >> 3);
492 /* No previous measure. */
496 if (tp->rcv_rtt_est.rtt != new_sample)
497 tp->rcv_rtt_est.rtt = new_sample;
500 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
502 if (tp->rcv_rtt_est.time == 0)
504 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
506 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
509 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
510 tp->rcv_rtt_est.time = tcp_time_stamp;
513 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
514 const struct sk_buff *skb)
516 struct tcp_sock *tp = tcp_sk(sk);
517 if (tp->rx_opt.rcv_tsecr &&
518 (TCP_SKB_CB(skb)->end_seq -
519 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
520 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
524 * This function should be called every time data is copied to user space.
525 * It calculates the appropriate TCP receive buffer space.
527 void tcp_rcv_space_adjust(struct sock *sk)
529 struct tcp_sock *tp = tcp_sk(sk);
533 if (tp->rcvq_space.time == 0)
536 time = tcp_time_stamp - tp->rcvq_space.time;
537 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
540 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
542 space = max(tp->rcvq_space.space, space);
544 if (tp->rcvq_space.space != space) {
547 tp->rcvq_space.space = space;
549 if (sysctl_tcp_moderate_rcvbuf &&
550 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
551 int new_clamp = space;
553 /* Receive space grows, normalize in order to
554 * take into account packet headers and sk_buff
555 * structure overhead.
560 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
561 while (tcp_win_from_space(rcvmem) < tp->advmss)
564 space = min(space, sysctl_tcp_rmem[2]);
565 if (space > sk->sk_rcvbuf) {
566 sk->sk_rcvbuf = space;
568 /* Make the window clamp follow along. */
569 tp->window_clamp = new_clamp;
575 tp->rcvq_space.seq = tp->copied_seq;
576 tp->rcvq_space.time = tcp_time_stamp;
579 /* There is something which you must keep in mind when you analyze the
580 * behavior of the tp->ato delayed ack timeout interval. When a
581 * connection starts up, we want to ack as quickly as possible. The
582 * problem is that "good" TCP's do slow start at the beginning of data
583 * transmission. The means that until we send the first few ACK's the
584 * sender will sit on his end and only queue most of his data, because
585 * he can only send snd_cwnd unacked packets at any given time. For
586 * each ACK we send, he increments snd_cwnd and transmits more of his
589 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
591 struct tcp_sock *tp = tcp_sk(sk);
592 struct inet_connection_sock *icsk = inet_csk(sk);
595 inet_csk_schedule_ack(sk);
597 tcp_measure_rcv_mss(sk, skb);
599 tcp_rcv_rtt_measure(tp);
601 now = tcp_time_stamp;
603 if (!icsk->icsk_ack.ato) {
604 /* The _first_ data packet received, initialize
605 * delayed ACK engine.
607 tcp_incr_quickack(sk);
608 icsk->icsk_ack.ato = TCP_ATO_MIN;
610 int m = now - icsk->icsk_ack.lrcvtime;
612 if (m <= TCP_ATO_MIN / 2) {
613 /* The fastest case is the first. */
614 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
615 } else if (m < icsk->icsk_ack.ato) {
616 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
617 if (icsk->icsk_ack.ato > icsk->icsk_rto)
618 icsk->icsk_ack.ato = icsk->icsk_rto;
619 } else if (m > icsk->icsk_rto) {
620 /* Too long gap. Apparently sender failed to
621 * restart window, so that we send ACKs quickly.
623 tcp_incr_quickack(sk);
627 icsk->icsk_ack.lrcvtime = now;
629 TCP_ECN_check_ce(tp, skb);
632 tcp_grow_window(sk, skb);
635 /* Called to compute a smoothed rtt estimate. The data fed to this
636 * routine either comes from timestamps, or from segments that were
637 * known _not_ to have been retransmitted [see Karn/Partridge
638 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
639 * piece by Van Jacobson.
640 * NOTE: the next three routines used to be one big routine.
641 * To save cycles in the RFC 1323 implementation it was better to break
642 * it up into three procedures. -- erics
644 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
646 struct tcp_sock *tp = tcp_sk(sk);
647 long m = mrtt; /* RTT */
649 /* The following amusing code comes from Jacobson's
650 * article in SIGCOMM '88. Note that rtt and mdev
651 * are scaled versions of rtt and mean deviation.
652 * This is designed to be as fast as possible
653 * m stands for "measurement".
655 * On a 1990 paper the rto value is changed to:
656 * RTO = rtt + 4 * mdev
658 * Funny. This algorithm seems to be very broken.
659 * These formulae increase RTO, when it should be decreased, increase
660 * too slowly, when it should be increased quickly, decrease too quickly
661 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
662 * does not matter how to _calculate_ it. Seems, it was trap
663 * that VJ failed to avoid. 8)
668 m -= (tp->srtt >> 3); /* m is now error in rtt est */
669 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
671 m = -m; /* m is now abs(error) */
672 m -= (tp->mdev >> 2); /* similar update on mdev */
673 /* This is similar to one of Eifel findings.
674 * Eifel blocks mdev updates when rtt decreases.
675 * This solution is a bit different: we use finer gain
676 * for mdev in this case (alpha*beta).
677 * Like Eifel it also prevents growth of rto,
678 * but also it limits too fast rto decreases,
679 * happening in pure Eifel.
684 m -= (tp->mdev >> 2); /* similar update on mdev */
686 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
687 if (tp->mdev > tp->mdev_max) {
688 tp->mdev_max = tp->mdev;
689 if (tp->mdev_max > tp->rttvar)
690 tp->rttvar = tp->mdev_max;
692 if (after(tp->snd_una, tp->rtt_seq)) {
693 if (tp->mdev_max < tp->rttvar)
694 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
695 tp->rtt_seq = tp->snd_nxt;
696 tp->mdev_max = tcp_rto_min(sk);
699 /* no previous measure. */
700 tp->srtt = m << 3; /* take the measured time to be rtt */
701 tp->mdev = m << 1; /* make sure rto = 3*rtt */
702 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
703 tp->rtt_seq = tp->snd_nxt;
707 /* Calculate rto without backoff. This is the second half of Van Jacobson's
708 * routine referred to above.
710 void tcp_set_rto(struct sock *sk)
712 const struct tcp_sock *tp = tcp_sk(sk);
713 /* Old crap is replaced with new one. 8)
716 * 1. If rtt variance happened to be less 50msec, it is hallucination.
717 * It cannot be less due to utterly erratic ACK generation made
718 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
719 * to do with delayed acks, because at cwnd>2 true delack timeout
720 * is invisible. Actually, Linux-2.4 also generates erratic
721 * ACKs in some circumstances.
723 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
725 /* 2. Fixups made earlier cannot be right.
726 * If we do not estimate RTO correctly without them,
727 * all the algo is pure shit and should be replaced
728 * with correct one. It is exactly, which we pretend to do.
731 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
732 * guarantees that rto is higher.
737 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
739 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
742 cwnd = TCP_INIT_CWND;
743 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
746 /* Set slow start threshold and cwnd not falling to slow start */
747 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
749 struct tcp_sock *tp = tcp_sk(sk);
750 const struct inet_connection_sock *icsk = inet_csk(sk);
752 tp->prior_ssthresh = 0;
754 if (icsk->icsk_ca_state < TCP_CA_CWR) {
757 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
758 tp->snd_cwnd = min(tp->snd_cwnd,
759 tcp_packets_in_flight(tp) + 1U);
760 tp->snd_cwnd_cnt = 0;
761 tp->high_seq = tp->snd_nxt;
762 tp->snd_cwnd_stamp = tcp_time_stamp;
763 TCP_ECN_queue_cwr(tp);
765 tcp_set_ca_state(sk, TCP_CA_CWR);
770 * Packet counting of FACK is based on in-order assumptions, therefore TCP
771 * disables it when reordering is detected
773 void tcp_disable_fack(struct tcp_sock *tp)
775 /* RFC3517 uses different metric in lost marker => reset on change */
777 tp->lost_skb_hint = NULL;
778 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
781 /* Take a notice that peer is sending D-SACKs */
782 static void tcp_dsack_seen(struct tcp_sock *tp)
784 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
787 static void tcp_update_reordering(struct sock *sk, const int metric,
790 struct tcp_sock *tp = tcp_sk(sk);
791 if (metric > tp->reordering) {
794 tp->reordering = min(TCP_MAX_REORDERING, metric);
796 /* This exciting event is worth to be remembered. 8) */
798 mib_idx = LINUX_MIB_TCPTSREORDER;
799 else if (tcp_is_reno(tp))
800 mib_idx = LINUX_MIB_TCPRENOREORDER;
801 else if (tcp_is_fack(tp))
802 mib_idx = LINUX_MIB_TCPFACKREORDER;
804 mib_idx = LINUX_MIB_TCPSACKREORDER;
806 NET_INC_STATS_BH(sock_net(sk), mib_idx);
807 #if FASTRETRANS_DEBUG > 1
808 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
809 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
813 tp->undo_marker ? tp->undo_retrans : 0);
815 tcp_disable_fack(tp);
819 tcp_disable_early_retrans(tp);
822 /* This must be called before lost_out is incremented */
823 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
825 if ((tp->retransmit_skb_hint == NULL) ||
826 before(TCP_SKB_CB(skb)->seq,
827 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
828 tp->retransmit_skb_hint = skb;
831 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
832 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
835 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
837 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
838 tcp_verify_retransmit_hint(tp, skb);
840 tp->lost_out += tcp_skb_pcount(skb);
841 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
845 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
848 tcp_verify_retransmit_hint(tp, skb);
850 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
851 tp->lost_out += tcp_skb_pcount(skb);
852 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
856 /* This procedure tags the retransmission queue when SACKs arrive.
858 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
859 * Packets in queue with these bits set are counted in variables
860 * sacked_out, retrans_out and lost_out, correspondingly.
862 * Valid combinations are:
863 * Tag InFlight Description
864 * 0 1 - orig segment is in flight.
865 * S 0 - nothing flies, orig reached receiver.
866 * L 0 - nothing flies, orig lost by net.
867 * R 2 - both orig and retransmit are in flight.
868 * L|R 1 - orig is lost, retransmit is in flight.
869 * S|R 1 - orig reached receiver, retrans is still in flight.
870 * (L|S|R is logically valid, it could occur when L|R is sacked,
871 * but it is equivalent to plain S and code short-curcuits it to S.
872 * L|S is logically invalid, it would mean -1 packet in flight 8))
874 * These 6 states form finite state machine, controlled by the following events:
875 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
876 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
877 * 3. Loss detection event of two flavors:
878 * A. Scoreboard estimator decided the packet is lost.
879 * A'. Reno "three dupacks" marks head of queue lost.
880 * A''. Its FACK modification, head until snd.fack is lost.
881 * B. SACK arrives sacking SND.NXT at the moment, when the
882 * segment was retransmitted.
883 * 4. D-SACK added new rule: D-SACK changes any tag to S.
885 * It is pleasant to note, that state diagram turns out to be commutative,
886 * so that we are allowed not to be bothered by order of our actions,
887 * when multiple events arrive simultaneously. (see the function below).
889 * Reordering detection.
890 * --------------------
891 * Reordering metric is maximal distance, which a packet can be displaced
892 * in packet stream. With SACKs we can estimate it:
894 * 1. SACK fills old hole and the corresponding segment was not
895 * ever retransmitted -> reordering. Alas, we cannot use it
896 * when segment was retransmitted.
897 * 2. The last flaw is solved with D-SACK. D-SACK arrives
898 * for retransmitted and already SACKed segment -> reordering..
899 * Both of these heuristics are not used in Loss state, when we cannot
900 * account for retransmits accurately.
902 * SACK block validation.
903 * ----------------------
905 * SACK block range validation checks that the received SACK block fits to
906 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
907 * Note that SND.UNA is not included to the range though being valid because
908 * it means that the receiver is rather inconsistent with itself reporting
909 * SACK reneging when it should advance SND.UNA. Such SACK block this is
910 * perfectly valid, however, in light of RFC2018 which explicitly states
911 * that "SACK block MUST reflect the newest segment. Even if the newest
912 * segment is going to be discarded ...", not that it looks very clever
913 * in case of head skb. Due to potentional receiver driven attacks, we
914 * choose to avoid immediate execution of a walk in write queue due to
915 * reneging and defer head skb's loss recovery to standard loss recovery
916 * procedure that will eventually trigger (nothing forbids us doing this).
918 * Implements also blockage to start_seq wrap-around. Problem lies in the
919 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
920 * there's no guarantee that it will be before snd_nxt (n). The problem
921 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
924 * <- outs wnd -> <- wrapzone ->
925 * u e n u_w e_w s n_w
927 * |<------------+------+----- TCP seqno space --------------+---------->|
928 * ...-- <2^31 ->| |<--------...
929 * ...---- >2^31 ------>| |<--------...
931 * Current code wouldn't be vulnerable but it's better still to discard such
932 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
933 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
934 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
935 * equal to the ideal case (infinite seqno space without wrap caused issues).
937 * With D-SACK the lower bound is extended to cover sequence space below
938 * SND.UNA down to undo_marker, which is the last point of interest. Yet
939 * again, D-SACK block must not to go across snd_una (for the same reason as
940 * for the normal SACK blocks, explained above). But there all simplicity
941 * ends, TCP might receive valid D-SACKs below that. As long as they reside
942 * fully below undo_marker they do not affect behavior in anyway and can
943 * therefore be safely ignored. In rare cases (which are more or less
944 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
945 * fragmentation and packet reordering past skb's retransmission. To consider
946 * them correctly, the acceptable range must be extended even more though
947 * the exact amount is rather hard to quantify. However, tp->max_window can
948 * be used as an exaggerated estimate.
950 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
951 u32 start_seq, u32 end_seq)
953 /* Too far in future, or reversed (interpretation is ambiguous) */
954 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
957 /* Nasty start_seq wrap-around check (see comments above) */
958 if (!before(start_seq, tp->snd_nxt))
961 /* In outstanding window? ...This is valid exit for D-SACKs too.
962 * start_seq == snd_una is non-sensical (see comments above)
964 if (after(start_seq, tp->snd_una))
967 if (!is_dsack || !tp->undo_marker)
970 /* ...Then it's D-SACK, and must reside below snd_una completely */
971 if (after(end_seq, tp->snd_una))
974 if (!before(start_seq, tp->undo_marker))
978 if (!after(end_seq, tp->undo_marker))
981 /* Undo_marker boundary crossing (overestimates a lot). Known already:
982 * start_seq < undo_marker and end_seq >= undo_marker.
984 return !before(start_seq, end_seq - tp->max_window);
987 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
988 * Event "B". Later note: FACK people cheated me again 8), we have to account
989 * for reordering! Ugly, but should help.
991 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
992 * less than what is now known to be received by the other end (derived from
993 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
994 * retransmitted skbs to avoid some costly processing per ACKs.
996 static void tcp_mark_lost_retrans(struct sock *sk)
998 const struct inet_connection_sock *icsk = inet_csk(sk);
999 struct tcp_sock *tp = tcp_sk(sk);
1000 struct sk_buff *skb;
1002 u32 new_low_seq = tp->snd_nxt;
1003 u32 received_upto = tcp_highest_sack_seq(tp);
1005 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1006 !after(received_upto, tp->lost_retrans_low) ||
1007 icsk->icsk_ca_state != TCP_CA_Recovery)
1010 tcp_for_write_queue(skb, sk) {
1011 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1013 if (skb == tcp_send_head(sk))
1015 if (cnt == tp->retrans_out)
1017 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1020 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1023 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1024 * constraint here (see above) but figuring out that at
1025 * least tp->reordering SACK blocks reside between ack_seq
1026 * and received_upto is not easy task to do cheaply with
1027 * the available datastructures.
1029 * Whether FACK should check here for tp->reordering segs
1030 * in-between one could argue for either way (it would be
1031 * rather simple to implement as we could count fack_count
1032 * during the walk and do tp->fackets_out - fack_count).
1034 if (after(received_upto, ack_seq)) {
1035 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1036 tp->retrans_out -= tcp_skb_pcount(skb);
1038 tcp_skb_mark_lost_uncond_verify(tp, skb);
1039 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1041 if (before(ack_seq, new_low_seq))
1042 new_low_seq = ack_seq;
1043 cnt += tcp_skb_pcount(skb);
1047 if (tp->retrans_out)
1048 tp->lost_retrans_low = new_low_seq;
1051 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1052 struct tcp_sack_block_wire *sp, int num_sacks,
1055 struct tcp_sock *tp = tcp_sk(sk);
1056 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1057 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1058 bool dup_sack = false;
1060 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1063 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1064 } else if (num_sacks > 1) {
1065 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1066 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1068 if (!after(end_seq_0, end_seq_1) &&
1069 !before(start_seq_0, start_seq_1)) {
1072 NET_INC_STATS_BH(sock_net(sk),
1073 LINUX_MIB_TCPDSACKOFORECV);
1077 /* D-SACK for already forgotten data... Do dumb counting. */
1078 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1079 !after(end_seq_0, prior_snd_una) &&
1080 after(end_seq_0, tp->undo_marker))
1086 struct tcp_sacktag_state {
1092 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1093 * the incoming SACK may not exactly match but we can find smaller MSS
1094 * aligned portion of it that matches. Therefore we might need to fragment
1095 * which may fail and creates some hassle (caller must handle error case
1098 * FIXME: this could be merged to shift decision code
1100 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1101 u32 start_seq, u32 end_seq)
1105 unsigned int pkt_len;
1108 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1109 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1111 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1112 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1113 mss = tcp_skb_mss(skb);
1114 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1117 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1121 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1126 /* Round if necessary so that SACKs cover only full MSSes
1127 * and/or the remaining small portion (if present)
1129 if (pkt_len > mss) {
1130 unsigned int new_len = (pkt_len / mss) * mss;
1131 if (!in_sack && new_len < pkt_len) {
1133 if (new_len > skb->len)
1138 err = tcp_fragment(sk, skb, pkt_len, mss);
1146 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1147 static u8 tcp_sacktag_one(struct sock *sk,
1148 struct tcp_sacktag_state *state, u8 sacked,
1149 u32 start_seq, u32 end_seq,
1150 bool dup_sack, int pcount)
1152 struct tcp_sock *tp = tcp_sk(sk);
1153 int fack_count = state->fack_count;
1155 /* Account D-SACK for retransmitted packet. */
1156 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1157 if (tp->undo_marker && tp->undo_retrans &&
1158 after(end_seq, tp->undo_marker))
1160 if (sacked & TCPCB_SACKED_ACKED)
1161 state->reord = min(fack_count, state->reord);
1164 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1165 if (!after(end_seq, tp->snd_una))
1168 if (!(sacked & TCPCB_SACKED_ACKED)) {
1169 if (sacked & TCPCB_SACKED_RETRANS) {
1170 /* If the segment is not tagged as lost,
1171 * we do not clear RETRANS, believing
1172 * that retransmission is still in flight.
1174 if (sacked & TCPCB_LOST) {
1175 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1176 tp->lost_out -= pcount;
1177 tp->retrans_out -= pcount;
1180 if (!(sacked & TCPCB_RETRANS)) {
1181 /* New sack for not retransmitted frame,
1182 * which was in hole. It is reordering.
1184 if (before(start_seq,
1185 tcp_highest_sack_seq(tp)))
1186 state->reord = min(fack_count,
1189 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1190 if (!after(end_seq, tp->frto_highmark))
1191 state->flag |= FLAG_ONLY_ORIG_SACKED;
1194 if (sacked & TCPCB_LOST) {
1195 sacked &= ~TCPCB_LOST;
1196 tp->lost_out -= pcount;
1200 sacked |= TCPCB_SACKED_ACKED;
1201 state->flag |= FLAG_DATA_SACKED;
1202 tp->sacked_out += pcount;
1204 fack_count += pcount;
1206 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1207 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1208 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1209 tp->lost_cnt_hint += pcount;
1211 if (fack_count > tp->fackets_out)
1212 tp->fackets_out = fack_count;
1215 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1216 * frames and clear it. undo_retrans is decreased above, L|R frames
1217 * are accounted above as well.
1219 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1220 sacked &= ~TCPCB_SACKED_RETRANS;
1221 tp->retrans_out -= pcount;
1227 /* Shift newly-SACKed bytes from this skb to the immediately previous
1228 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1230 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1231 struct tcp_sacktag_state *state,
1232 unsigned int pcount, int shifted, int mss,
1235 struct tcp_sock *tp = tcp_sk(sk);
1236 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1237 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1238 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1242 /* Adjust counters and hints for the newly sacked sequence
1243 * range but discard the return value since prev is already
1244 * marked. We must tag the range first because the seq
1245 * advancement below implicitly advances
1246 * tcp_highest_sack_seq() when skb is highest_sack.
1248 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1249 start_seq, end_seq, dup_sack, pcount);
1251 if (skb == tp->lost_skb_hint)
1252 tp->lost_cnt_hint += pcount;
1254 TCP_SKB_CB(prev)->end_seq += shifted;
1255 TCP_SKB_CB(skb)->seq += shifted;
1257 skb_shinfo(prev)->gso_segs += pcount;
1258 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1259 skb_shinfo(skb)->gso_segs -= pcount;
1261 /* When we're adding to gso_segs == 1, gso_size will be zero,
1262 * in theory this shouldn't be necessary but as long as DSACK
1263 * code can come after this skb later on it's better to keep
1264 * setting gso_size to something.
1266 if (!skb_shinfo(prev)->gso_size) {
1267 skb_shinfo(prev)->gso_size = mss;
1268 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1271 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1272 if (skb_shinfo(skb)->gso_segs <= 1) {
1273 skb_shinfo(skb)->gso_size = 0;
1274 skb_shinfo(skb)->gso_type = 0;
1277 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1278 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1281 BUG_ON(!tcp_skb_pcount(skb));
1282 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1286 /* Whole SKB was eaten :-) */
1288 if (skb == tp->retransmit_skb_hint)
1289 tp->retransmit_skb_hint = prev;
1290 if (skb == tp->scoreboard_skb_hint)
1291 tp->scoreboard_skb_hint = prev;
1292 if (skb == tp->lost_skb_hint) {
1293 tp->lost_skb_hint = prev;
1294 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1297 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1298 if (skb == tcp_highest_sack(sk))
1299 tcp_advance_highest_sack(sk, skb);
1301 tcp_unlink_write_queue(skb, sk);
1302 sk_wmem_free_skb(sk, skb);
1304 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1309 /* I wish gso_size would have a bit more sane initialization than
1310 * something-or-zero which complicates things
1312 static int tcp_skb_seglen(const struct sk_buff *skb)
1314 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1317 /* Shifting pages past head area doesn't work */
1318 static int skb_can_shift(const struct sk_buff *skb)
1320 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1323 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1326 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1327 struct tcp_sacktag_state *state,
1328 u32 start_seq, u32 end_seq,
1331 struct tcp_sock *tp = tcp_sk(sk);
1332 struct sk_buff *prev;
1338 if (!sk_can_gso(sk))
1341 /* Normally R but no L won't result in plain S */
1343 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1345 if (!skb_can_shift(skb))
1347 /* This frame is about to be dropped (was ACKed). */
1348 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1351 /* Can only happen with delayed DSACK + discard craziness */
1352 if (unlikely(skb == tcp_write_queue_head(sk)))
1354 prev = tcp_write_queue_prev(sk, skb);
1356 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1359 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1360 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1364 pcount = tcp_skb_pcount(skb);
1365 mss = tcp_skb_seglen(skb);
1367 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1368 * drop this restriction as unnecessary
1370 if (mss != tcp_skb_seglen(prev))
1373 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1375 /* CHECKME: This is non-MSS split case only?, this will
1376 * cause skipped skbs due to advancing loop btw, original
1377 * has that feature too
1379 if (tcp_skb_pcount(skb) <= 1)
1382 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1384 /* TODO: head merge to next could be attempted here
1385 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1386 * though it might not be worth of the additional hassle
1388 * ...we can probably just fallback to what was done
1389 * previously. We could try merging non-SACKed ones
1390 * as well but it probably isn't going to buy off
1391 * because later SACKs might again split them, and
1392 * it would make skb timestamp tracking considerably
1398 len = end_seq - TCP_SKB_CB(skb)->seq;
1400 BUG_ON(len > skb->len);
1402 /* MSS boundaries should be honoured or else pcount will
1403 * severely break even though it makes things bit trickier.
1404 * Optimize common case to avoid most of the divides
1406 mss = tcp_skb_mss(skb);
1408 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1409 * drop this restriction as unnecessary
1411 if (mss != tcp_skb_seglen(prev))
1416 } else if (len < mss) {
1424 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1425 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1428 if (!skb_shift(prev, skb, len))
1430 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1433 /* Hole filled allows collapsing with the next as well, this is very
1434 * useful when hole on every nth skb pattern happens
1436 if (prev == tcp_write_queue_tail(sk))
1438 skb = tcp_write_queue_next(sk, prev);
1440 if (!skb_can_shift(skb) ||
1441 (skb == tcp_send_head(sk)) ||
1442 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1443 (mss != tcp_skb_seglen(skb)))
1447 if (skb_shift(prev, skb, len)) {
1448 pcount += tcp_skb_pcount(skb);
1449 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1453 state->fack_count += pcount;
1460 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1464 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1465 struct tcp_sack_block *next_dup,
1466 struct tcp_sacktag_state *state,
1467 u32 start_seq, u32 end_seq,
1470 struct tcp_sock *tp = tcp_sk(sk);
1471 struct sk_buff *tmp;
1473 tcp_for_write_queue_from(skb, sk) {
1475 bool dup_sack = dup_sack_in;
1477 if (skb == tcp_send_head(sk))
1480 /* queue is in-order => we can short-circuit the walk early */
1481 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1484 if ((next_dup != NULL) &&
1485 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1486 in_sack = tcp_match_skb_to_sack(sk, skb,
1487 next_dup->start_seq,
1493 /* skb reference here is a bit tricky to get right, since
1494 * shifting can eat and free both this skb and the next,
1495 * so not even _safe variant of the loop is enough.
1498 tmp = tcp_shift_skb_data(sk, skb, state,
1499 start_seq, end_seq, dup_sack);
1508 in_sack = tcp_match_skb_to_sack(sk, skb,
1514 if (unlikely(in_sack < 0))
1518 TCP_SKB_CB(skb)->sacked =
1521 TCP_SKB_CB(skb)->sacked,
1522 TCP_SKB_CB(skb)->seq,
1523 TCP_SKB_CB(skb)->end_seq,
1525 tcp_skb_pcount(skb));
1527 if (!before(TCP_SKB_CB(skb)->seq,
1528 tcp_highest_sack_seq(tp)))
1529 tcp_advance_highest_sack(sk, skb);
1532 state->fack_count += tcp_skb_pcount(skb);
1537 /* Avoid all extra work that is being done by sacktag while walking in
1540 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1541 struct tcp_sacktag_state *state,
1544 tcp_for_write_queue_from(skb, sk) {
1545 if (skb == tcp_send_head(sk))
1548 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1551 state->fack_count += tcp_skb_pcount(skb);
1556 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1558 struct tcp_sack_block *next_dup,
1559 struct tcp_sacktag_state *state,
1562 if (next_dup == NULL)
1565 if (before(next_dup->start_seq, skip_to_seq)) {
1566 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1567 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1568 next_dup->start_seq, next_dup->end_seq,
1575 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1577 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1581 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1584 const struct inet_connection_sock *icsk = inet_csk(sk);
1585 struct tcp_sock *tp = tcp_sk(sk);
1586 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1587 TCP_SKB_CB(ack_skb)->sacked);
1588 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1589 struct tcp_sack_block sp[TCP_NUM_SACKS];
1590 struct tcp_sack_block *cache;
1591 struct tcp_sacktag_state state;
1592 struct sk_buff *skb;
1593 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1595 bool found_dup_sack = false;
1597 int first_sack_index;
1600 state.reord = tp->packets_out;
1602 if (!tp->sacked_out) {
1603 if (WARN_ON(tp->fackets_out))
1604 tp->fackets_out = 0;
1605 tcp_highest_sack_reset(sk);
1608 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1609 num_sacks, prior_snd_una);
1611 state.flag |= FLAG_DSACKING_ACK;
1613 /* Eliminate too old ACKs, but take into
1614 * account more or less fresh ones, they can
1615 * contain valid SACK info.
1617 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1620 if (!tp->packets_out)
1624 first_sack_index = 0;
1625 for (i = 0; i < num_sacks; i++) {
1626 bool dup_sack = !i && found_dup_sack;
1628 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1629 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1631 if (!tcp_is_sackblock_valid(tp, dup_sack,
1632 sp[used_sacks].start_seq,
1633 sp[used_sacks].end_seq)) {
1637 if (!tp->undo_marker)
1638 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1640 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1642 /* Don't count olds caused by ACK reordering */
1643 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1644 !after(sp[used_sacks].end_seq, tp->snd_una))
1646 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1649 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1651 first_sack_index = -1;
1655 /* Ignore very old stuff early */
1656 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1662 /* order SACK blocks to allow in order walk of the retrans queue */
1663 for (i = used_sacks - 1; i > 0; i--) {
1664 for (j = 0; j < i; j++) {
1665 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1666 swap(sp[j], sp[j + 1]);
1668 /* Track where the first SACK block goes to */
1669 if (j == first_sack_index)
1670 first_sack_index = j + 1;
1675 skb = tcp_write_queue_head(sk);
1676 state.fack_count = 0;
1679 if (!tp->sacked_out) {
1680 /* It's already past, so skip checking against it */
1681 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1683 cache = tp->recv_sack_cache;
1684 /* Skip empty blocks in at head of the cache */
1685 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1690 while (i < used_sacks) {
1691 u32 start_seq = sp[i].start_seq;
1692 u32 end_seq = sp[i].end_seq;
1693 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1694 struct tcp_sack_block *next_dup = NULL;
1696 if (found_dup_sack && ((i + 1) == first_sack_index))
1697 next_dup = &sp[i + 1];
1699 /* Skip too early cached blocks */
1700 while (tcp_sack_cache_ok(tp, cache) &&
1701 !before(start_seq, cache->end_seq))
1704 /* Can skip some work by looking recv_sack_cache? */
1705 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1706 after(end_seq, cache->start_seq)) {
1709 if (before(start_seq, cache->start_seq)) {
1710 skb = tcp_sacktag_skip(skb, sk, &state,
1712 skb = tcp_sacktag_walk(skb, sk, next_dup,
1719 /* Rest of the block already fully processed? */
1720 if (!after(end_seq, cache->end_seq))
1723 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1727 /* ...tail remains todo... */
1728 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1729 /* ...but better entrypoint exists! */
1730 skb = tcp_highest_sack(sk);
1733 state.fack_count = tp->fackets_out;
1738 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1739 /* Check overlap against next cached too (past this one already) */
1744 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1745 skb = tcp_highest_sack(sk);
1748 state.fack_count = tp->fackets_out;
1750 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1753 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1754 start_seq, end_seq, dup_sack);
1757 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1758 * due to in-order walk
1760 if (after(end_seq, tp->frto_highmark))
1761 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1766 /* Clear the head of the cache sack blocks so we can skip it next time */
1767 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1768 tp->recv_sack_cache[i].start_seq = 0;
1769 tp->recv_sack_cache[i].end_seq = 0;
1771 for (j = 0; j < used_sacks; j++)
1772 tp->recv_sack_cache[i++] = sp[j];
1774 tcp_mark_lost_retrans(sk);
1776 tcp_verify_left_out(tp);
1778 if ((state.reord < tp->fackets_out) &&
1779 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1780 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1781 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1785 #if FASTRETRANS_DEBUG > 0
1786 WARN_ON((int)tp->sacked_out < 0);
1787 WARN_ON((int)tp->lost_out < 0);
1788 WARN_ON((int)tp->retrans_out < 0);
1789 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1794 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1795 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1797 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1801 holes = max(tp->lost_out, 1U);
1802 holes = min(holes, tp->packets_out);
1804 if ((tp->sacked_out + holes) > tp->packets_out) {
1805 tp->sacked_out = tp->packets_out - holes;
1811 /* If we receive more dupacks than we expected counting segments
1812 * in assumption of absent reordering, interpret this as reordering.
1813 * The only another reason could be bug in receiver TCP.
1815 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1817 struct tcp_sock *tp = tcp_sk(sk);
1818 if (tcp_limit_reno_sacked(tp))
1819 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1822 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1824 static void tcp_add_reno_sack(struct sock *sk)
1826 struct tcp_sock *tp = tcp_sk(sk);
1828 tcp_check_reno_reordering(sk, 0);
1829 tcp_verify_left_out(tp);
1832 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1834 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1836 struct tcp_sock *tp = tcp_sk(sk);
1839 /* One ACK acked hole. The rest eat duplicate ACKs. */
1840 if (acked - 1 >= tp->sacked_out)
1843 tp->sacked_out -= acked - 1;
1845 tcp_check_reno_reordering(sk, acked);
1846 tcp_verify_left_out(tp);
1849 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1854 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1856 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1859 /* F-RTO can only be used if TCP has never retransmitted anything other than
1860 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1862 bool tcp_use_frto(struct sock *sk)
1864 const struct tcp_sock *tp = tcp_sk(sk);
1865 const struct inet_connection_sock *icsk = inet_csk(sk);
1866 struct sk_buff *skb;
1868 if (!sysctl_tcp_frto)
1871 /* MTU probe and F-RTO won't really play nicely along currently */
1872 if (icsk->icsk_mtup.probe_size)
1875 if (tcp_is_sackfrto(tp))
1878 /* Avoid expensive walking of rexmit queue if possible */
1879 if (tp->retrans_out > 1)
1882 skb = tcp_write_queue_head(sk);
1883 if (tcp_skb_is_last(sk, skb))
1885 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1886 tcp_for_write_queue_from(skb, sk) {
1887 if (skb == tcp_send_head(sk))
1889 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1891 /* Short-circuit when first non-SACKed skb has been checked */
1892 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1898 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1899 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1900 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1901 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1902 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1903 * bits are handled if the Loss state is really to be entered (in
1904 * tcp_enter_frto_loss).
1906 * Do like tcp_enter_loss() would; when RTO expires the second time it
1908 * "Reduce ssthresh if it has not yet been made inside this window."
1910 void tcp_enter_frto(struct sock *sk)
1912 const struct inet_connection_sock *icsk = inet_csk(sk);
1913 struct tcp_sock *tp = tcp_sk(sk);
1914 struct sk_buff *skb;
1916 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1917 tp->snd_una == tp->high_seq ||
1918 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1919 !icsk->icsk_retransmits)) {
1920 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1921 /* Our state is too optimistic in ssthresh() call because cwnd
1922 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1923 * recovery has not yet completed. Pattern would be this: RTO,
1924 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1926 * RFC4138 should be more specific on what to do, even though
1927 * RTO is quite unlikely to occur after the first Cumulative ACK
1928 * due to back-off and complexity of triggering events ...
1930 if (tp->frto_counter) {
1932 stored_cwnd = tp->snd_cwnd;
1934 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1935 tp->snd_cwnd = stored_cwnd;
1937 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1939 /* ... in theory, cong.control module could do "any tricks" in
1940 * ssthresh(), which means that ca_state, lost bits and lost_out
1941 * counter would have to be faked before the call occurs. We
1942 * consider that too expensive, unlikely and hacky, so modules
1943 * using these in ssthresh() must deal these incompatibility
1944 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1946 tcp_ca_event(sk, CA_EVENT_FRTO);
1949 tp->undo_marker = tp->snd_una;
1950 tp->undo_retrans = 0;
1952 skb = tcp_write_queue_head(sk);
1953 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1954 tp->undo_marker = 0;
1955 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1956 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1957 tp->retrans_out -= tcp_skb_pcount(skb);
1959 tcp_verify_left_out(tp);
1961 /* Too bad if TCP was application limited */
1962 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1964 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1965 * The last condition is necessary at least in tp->frto_counter case.
1967 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
1968 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1969 after(tp->high_seq, tp->snd_una)) {
1970 tp->frto_highmark = tp->high_seq;
1972 tp->frto_highmark = tp->snd_nxt;
1974 tcp_set_ca_state(sk, TCP_CA_Disorder);
1975 tp->high_seq = tp->snd_nxt;
1976 tp->frto_counter = 1;
1979 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1980 * which indicates that we should follow the traditional RTO recovery,
1981 * i.e. mark everything lost and do go-back-N retransmission.
1983 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1985 struct tcp_sock *tp = tcp_sk(sk);
1986 struct sk_buff *skb;
1989 tp->retrans_out = 0;
1990 if (tcp_is_reno(tp))
1991 tcp_reset_reno_sack(tp);
1993 tcp_for_write_queue(skb, sk) {
1994 if (skb == tcp_send_head(sk))
1997 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1999 * Count the retransmission made on RTO correctly (only when
2000 * waiting for the first ACK and did not get it)...
2002 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2003 /* For some reason this R-bit might get cleared? */
2004 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2005 tp->retrans_out += tcp_skb_pcount(skb);
2006 /* ...enter this if branch just for the first segment */
2007 flag |= FLAG_DATA_ACKED;
2009 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2010 tp->undo_marker = 0;
2011 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2014 /* Marking forward transmissions that were made after RTO lost
2015 * can cause unnecessary retransmissions in some scenarios,
2016 * SACK blocks will mitigate that in some but not in all cases.
2017 * We used to not mark them but it was causing break-ups with
2018 * receivers that do only in-order receival.
2020 * TODO: we could detect presence of such receiver and select
2021 * different behavior per flow.
2023 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2024 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2025 tp->lost_out += tcp_skb_pcount(skb);
2026 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2029 tcp_verify_left_out(tp);
2031 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2032 tp->snd_cwnd_cnt = 0;
2033 tp->snd_cwnd_stamp = tcp_time_stamp;
2034 tp->frto_counter = 0;
2035 tp->bytes_acked = 0;
2037 tp->reordering = min_t(unsigned int, tp->reordering,
2038 sysctl_tcp_reordering);
2039 tcp_set_ca_state(sk, TCP_CA_Loss);
2040 tp->high_seq = tp->snd_nxt;
2041 TCP_ECN_queue_cwr(tp);
2043 tcp_clear_all_retrans_hints(tp);
2046 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2048 tp->retrans_out = 0;
2051 tp->undo_marker = 0;
2052 tp->undo_retrans = 0;
2055 void tcp_clear_retrans(struct tcp_sock *tp)
2057 tcp_clear_retrans_partial(tp);
2059 tp->fackets_out = 0;
2063 /* Enter Loss state. If "how" is not zero, forget all SACK information
2064 * and reset tags completely, otherwise preserve SACKs. If receiver
2065 * dropped its ofo queue, we will know this due to reneging detection.
2067 void tcp_enter_loss(struct sock *sk, int how)
2069 const struct inet_connection_sock *icsk = inet_csk(sk);
2070 struct tcp_sock *tp = tcp_sk(sk);
2071 struct sk_buff *skb;
2073 /* Reduce ssthresh if it has not yet been made inside this window. */
2074 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2075 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2076 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2077 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2078 tcp_ca_event(sk, CA_EVENT_LOSS);
2081 tp->snd_cwnd_cnt = 0;
2082 tp->snd_cwnd_stamp = tcp_time_stamp;
2084 tp->bytes_acked = 0;
2085 tcp_clear_retrans_partial(tp);
2087 if (tcp_is_reno(tp))
2088 tcp_reset_reno_sack(tp);
2091 /* Push undo marker, if it was plain RTO and nothing
2092 * was retransmitted. */
2093 tp->undo_marker = tp->snd_una;
2096 tp->fackets_out = 0;
2098 tcp_clear_all_retrans_hints(tp);
2100 tcp_for_write_queue(skb, sk) {
2101 if (skb == tcp_send_head(sk))
2104 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2105 tp->undo_marker = 0;
2106 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2107 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2108 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2109 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2110 tp->lost_out += tcp_skb_pcount(skb);
2111 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2114 tcp_verify_left_out(tp);
2116 tp->reordering = min_t(unsigned int, tp->reordering,
2117 sysctl_tcp_reordering);
2118 tcp_set_ca_state(sk, TCP_CA_Loss);
2119 tp->high_seq = tp->snd_nxt;
2120 TCP_ECN_queue_cwr(tp);
2121 /* Abort F-RTO algorithm if one is in progress */
2122 tp->frto_counter = 0;
2125 /* If ACK arrived pointing to a remembered SACK, it means that our
2126 * remembered SACKs do not reflect real state of receiver i.e.
2127 * receiver _host_ is heavily congested (or buggy).
2129 * Do processing similar to RTO timeout.
2131 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2133 if (flag & FLAG_SACK_RENEGING) {
2134 struct inet_connection_sock *icsk = inet_csk(sk);
2135 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2137 tcp_enter_loss(sk, 1);
2138 icsk->icsk_retransmits++;
2139 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2140 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2141 icsk->icsk_rto, TCP_RTO_MAX);
2147 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2149 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2152 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2153 * counter when SACK is enabled (without SACK, sacked_out is used for
2156 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2157 * segments up to the highest received SACK block so far and holes in
2160 * With reordering, holes may still be in flight, so RFC3517 recovery
2161 * uses pure sacked_out (total number of SACKed segments) even though
2162 * it violates the RFC that uses duplicate ACKs, often these are equal
2163 * but when e.g. out-of-window ACKs or packet duplication occurs,
2164 * they differ. Since neither occurs due to loss, TCP should really
2167 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2169 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2172 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2174 struct tcp_sock *tp = tcp_sk(sk);
2175 unsigned long delay;
2177 /* Delay early retransmit and entering fast recovery for
2178 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2179 * available, or RTO is scheduled to fire first.
2181 if (sysctl_tcp_early_retrans < 2 || (flag & FLAG_ECE) || !tp->srtt)
2184 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2185 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2188 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, TCP_RTO_MAX);
2189 tp->early_retrans_delayed = 1;
2193 static inline int tcp_skb_timedout(const struct sock *sk,
2194 const struct sk_buff *skb)
2196 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2199 static inline int tcp_head_timedout(const struct sock *sk)
2201 const struct tcp_sock *tp = tcp_sk(sk);
2203 return tp->packets_out &&
2204 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2207 /* Linux NewReno/SACK/FACK/ECN state machine.
2208 * --------------------------------------
2210 * "Open" Normal state, no dubious events, fast path.
2211 * "Disorder" In all the respects it is "Open",
2212 * but requires a bit more attention. It is entered when
2213 * we see some SACKs or dupacks. It is split of "Open"
2214 * mainly to move some processing from fast path to slow one.
2215 * "CWR" CWND was reduced due to some Congestion Notification event.
2216 * It can be ECN, ICMP source quench, local device congestion.
2217 * "Recovery" CWND was reduced, we are fast-retransmitting.
2218 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2220 * tcp_fastretrans_alert() is entered:
2221 * - each incoming ACK, if state is not "Open"
2222 * - when arrived ACK is unusual, namely:
2227 * Counting packets in flight is pretty simple.
2229 * in_flight = packets_out - left_out + retrans_out
2231 * packets_out is SND.NXT-SND.UNA counted in packets.
2233 * retrans_out is number of retransmitted segments.
2235 * left_out is number of segments left network, but not ACKed yet.
2237 * left_out = sacked_out + lost_out
2239 * sacked_out: Packets, which arrived to receiver out of order
2240 * and hence not ACKed. With SACKs this number is simply
2241 * amount of SACKed data. Even without SACKs
2242 * it is easy to give pretty reliable estimate of this number,
2243 * counting duplicate ACKs.
2245 * lost_out: Packets lost by network. TCP has no explicit
2246 * "loss notification" feedback from network (for now).
2247 * It means that this number can be only _guessed_.
2248 * Actually, it is the heuristics to predict lossage that
2249 * distinguishes different algorithms.
2251 * F.e. after RTO, when all the queue is considered as lost,
2252 * lost_out = packets_out and in_flight = retrans_out.
2254 * Essentially, we have now two algorithms counting
2257 * FACK: It is the simplest heuristics. As soon as we decided
2258 * that something is lost, we decide that _all_ not SACKed
2259 * packets until the most forward SACK are lost. I.e.
2260 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2261 * It is absolutely correct estimate, if network does not reorder
2262 * packets. And it loses any connection to reality when reordering
2263 * takes place. We use FACK by default until reordering
2264 * is suspected on the path to this destination.
2266 * NewReno: when Recovery is entered, we assume that one segment
2267 * is lost (classic Reno). While we are in Recovery and
2268 * a partial ACK arrives, we assume that one more packet
2269 * is lost (NewReno). This heuristics are the same in NewReno
2272 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2273 * deflation etc. CWND is real congestion window, never inflated, changes
2274 * only according to classic VJ rules.
2276 * Really tricky (and requiring careful tuning) part of algorithm
2277 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2278 * The first determines the moment _when_ we should reduce CWND and,
2279 * hence, slow down forward transmission. In fact, it determines the moment
2280 * when we decide that hole is caused by loss, rather than by a reorder.
2282 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2283 * holes, caused by lost packets.
2285 * And the most logically complicated part of algorithm is undo
2286 * heuristics. We detect false retransmits due to both too early
2287 * fast retransmit (reordering) and underestimated RTO, analyzing
2288 * timestamps and D-SACKs. When we detect that some segments were
2289 * retransmitted by mistake and CWND reduction was wrong, we undo
2290 * window reduction and abort recovery phase. This logic is hidden
2291 * inside several functions named tcp_try_undo_<something>.
2294 /* This function decides, when we should leave Disordered state
2295 * and enter Recovery phase, reducing congestion window.
2297 * Main question: may we further continue forward transmission
2298 * with the same cwnd?
2300 static bool tcp_time_to_recover(struct sock *sk, int flag)
2302 struct tcp_sock *tp = tcp_sk(sk);
2305 /* Do not perform any recovery during F-RTO algorithm */
2306 if (tp->frto_counter)
2309 /* Trick#1: The loss is proven. */
2313 /* Not-A-Trick#2 : Classic rule... */
2314 if (tcp_dupack_heuristics(tp) > tp->reordering)
2317 /* Trick#3 : when we use RFC2988 timer restart, fast
2318 * retransmit can be triggered by timeout of queue head.
2320 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2323 /* Trick#4: It is still not OK... But will it be useful to delay
2326 packets_out = tp->packets_out;
2327 if (packets_out <= tp->reordering &&
2328 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2329 !tcp_may_send_now(sk)) {
2330 /* We have nothing to send. This connection is limited
2331 * either by receiver window or by application.
2336 /* If a thin stream is detected, retransmit after first
2337 * received dupack. Employ only if SACK is supported in order
2338 * to avoid possible corner-case series of spurious retransmissions
2339 * Use only if there are no unsent data.
2341 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2342 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2343 tcp_is_sack(tp) && !tcp_send_head(sk))
2346 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2347 * retransmissions due to small network reorderings, we implement
2348 * Mitigation A.3 in the RFC and delay the retransmission for a short
2349 * interval if appropriate.
2351 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2352 (tp->packets_out == (tp->sacked_out + 1) && tp->packets_out < 4) &&
2353 !tcp_may_send_now(sk))
2354 return !tcp_pause_early_retransmit(sk, flag);
2359 /* New heuristics: it is possible only after we switched to restart timer
2360 * each time when something is ACKed. Hence, we can detect timed out packets
2361 * during fast retransmit without falling to slow start.
2363 * Usefulness of this as is very questionable, since we should know which of
2364 * the segments is the next to timeout which is relatively expensive to find
2365 * in general case unless we add some data structure just for that. The
2366 * current approach certainly won't find the right one too often and when it
2367 * finally does find _something_ it usually marks large part of the window
2368 * right away (because a retransmission with a larger timestamp blocks the
2369 * loop from advancing). -ij
2371 static void tcp_timeout_skbs(struct sock *sk)
2373 struct tcp_sock *tp = tcp_sk(sk);
2374 struct sk_buff *skb;
2376 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2379 skb = tp->scoreboard_skb_hint;
2380 if (tp->scoreboard_skb_hint == NULL)
2381 skb = tcp_write_queue_head(sk);
2383 tcp_for_write_queue_from(skb, sk) {
2384 if (skb == tcp_send_head(sk))
2386 if (!tcp_skb_timedout(sk, skb))
2389 tcp_skb_mark_lost(tp, skb);
2392 tp->scoreboard_skb_hint = skb;
2394 tcp_verify_left_out(tp);
2397 /* Detect loss in event "A" above by marking head of queue up as lost.
2398 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2399 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2400 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2401 * the maximum SACKed segments to pass before reaching this limit.
2403 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2405 struct tcp_sock *tp = tcp_sk(sk);
2406 struct sk_buff *skb;
2410 /* Use SACK to deduce losses of new sequences sent during recovery */
2411 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2413 WARN_ON(packets > tp->packets_out);
2414 if (tp->lost_skb_hint) {
2415 skb = tp->lost_skb_hint;
2416 cnt = tp->lost_cnt_hint;
2417 /* Head already handled? */
2418 if (mark_head && skb != tcp_write_queue_head(sk))
2421 skb = tcp_write_queue_head(sk);
2425 tcp_for_write_queue_from(skb, sk) {
2426 if (skb == tcp_send_head(sk))
2428 /* TODO: do this better */
2429 /* this is not the most efficient way to do this... */
2430 tp->lost_skb_hint = skb;
2431 tp->lost_cnt_hint = cnt;
2433 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2437 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2438 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2439 cnt += tcp_skb_pcount(skb);
2441 if (cnt > packets) {
2442 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2443 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2444 (oldcnt >= packets))
2447 mss = skb_shinfo(skb)->gso_size;
2448 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2454 tcp_skb_mark_lost(tp, skb);
2459 tcp_verify_left_out(tp);
2462 /* Account newly detected lost packet(s) */
2464 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2466 struct tcp_sock *tp = tcp_sk(sk);
2468 if (tcp_is_reno(tp)) {
2469 tcp_mark_head_lost(sk, 1, 1);
2470 } else if (tcp_is_fack(tp)) {
2471 int lost = tp->fackets_out - tp->reordering;
2474 tcp_mark_head_lost(sk, lost, 0);
2476 int sacked_upto = tp->sacked_out - tp->reordering;
2477 if (sacked_upto >= 0)
2478 tcp_mark_head_lost(sk, sacked_upto, 0);
2479 else if (fast_rexmit)
2480 tcp_mark_head_lost(sk, 1, 1);
2483 tcp_timeout_skbs(sk);
2486 /* CWND moderation, preventing bursts due to too big ACKs
2487 * in dubious situations.
2489 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2491 tp->snd_cwnd = min(tp->snd_cwnd,
2492 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2493 tp->snd_cwnd_stamp = tcp_time_stamp;
2496 /* Lower bound on congestion window is slow start threshold
2497 * unless congestion avoidance choice decides to overide it.
2499 static inline u32 tcp_cwnd_min(const struct sock *sk)
2501 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2503 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2506 /* Decrease cwnd each second ack. */
2507 static void tcp_cwnd_down(struct sock *sk, int flag)
2509 struct tcp_sock *tp = tcp_sk(sk);
2510 int decr = tp->snd_cwnd_cnt + 1;
2512 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2513 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2514 tp->snd_cwnd_cnt = decr & 1;
2517 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2518 tp->snd_cwnd -= decr;
2520 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2521 tp->snd_cwnd_stamp = tcp_time_stamp;
2525 /* Nothing was retransmitted or returned timestamp is less
2526 * than timestamp of the first retransmission.
2528 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2530 return !tp->retrans_stamp ||
2531 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2532 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2535 /* Undo procedures. */
2537 #if FASTRETRANS_DEBUG > 1
2538 static void DBGUNDO(struct sock *sk, const char *msg)
2540 struct tcp_sock *tp = tcp_sk(sk);
2541 struct inet_sock *inet = inet_sk(sk);
2543 if (sk->sk_family == AF_INET) {
2544 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2546 &inet->inet_daddr, ntohs(inet->inet_dport),
2547 tp->snd_cwnd, tcp_left_out(tp),
2548 tp->snd_ssthresh, tp->prior_ssthresh,
2551 #if IS_ENABLED(CONFIG_IPV6)
2552 else if (sk->sk_family == AF_INET6) {
2553 struct ipv6_pinfo *np = inet6_sk(sk);
2554 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2556 &np->daddr, ntohs(inet->inet_dport),
2557 tp->snd_cwnd, tcp_left_out(tp),
2558 tp->snd_ssthresh, tp->prior_ssthresh,
2564 #define DBGUNDO(x...) do { } while (0)
2567 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2569 struct tcp_sock *tp = tcp_sk(sk);
2571 if (tp->prior_ssthresh) {
2572 const struct inet_connection_sock *icsk = inet_csk(sk);
2574 if (icsk->icsk_ca_ops->undo_cwnd)
2575 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2577 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2579 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2580 tp->snd_ssthresh = tp->prior_ssthresh;
2581 TCP_ECN_withdraw_cwr(tp);
2584 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2586 tp->snd_cwnd_stamp = tcp_time_stamp;
2589 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2591 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2594 /* People celebrate: "We love our President!" */
2595 static bool tcp_try_undo_recovery(struct sock *sk)
2597 struct tcp_sock *tp = tcp_sk(sk);
2599 if (tcp_may_undo(tp)) {
2602 /* Happy end! We did not retransmit anything
2603 * or our original transmission succeeded.
2605 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2606 tcp_undo_cwr(sk, true);
2607 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2608 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2610 mib_idx = LINUX_MIB_TCPFULLUNDO;
2612 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2613 tp->undo_marker = 0;
2615 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2616 /* Hold old state until something *above* high_seq
2617 * is ACKed. For Reno it is MUST to prevent false
2618 * fast retransmits (RFC2582). SACK TCP is safe. */
2619 tcp_moderate_cwnd(tp);
2622 tcp_set_ca_state(sk, TCP_CA_Open);
2626 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2627 static void tcp_try_undo_dsack(struct sock *sk)
2629 struct tcp_sock *tp = tcp_sk(sk);
2631 if (tp->undo_marker && !tp->undo_retrans) {
2632 DBGUNDO(sk, "D-SACK");
2633 tcp_undo_cwr(sk, true);
2634 tp->undo_marker = 0;
2635 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2639 /* We can clear retrans_stamp when there are no retransmissions in the
2640 * window. It would seem that it is trivially available for us in
2641 * tp->retrans_out, however, that kind of assumptions doesn't consider
2642 * what will happen if errors occur when sending retransmission for the
2643 * second time. ...It could the that such segment has only
2644 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2645 * the head skb is enough except for some reneging corner cases that
2646 * are not worth the effort.
2648 * Main reason for all this complexity is the fact that connection dying
2649 * time now depends on the validity of the retrans_stamp, in particular,
2650 * that successive retransmissions of a segment must not advance
2651 * retrans_stamp under any conditions.
2653 static bool tcp_any_retrans_done(const struct sock *sk)
2655 const struct tcp_sock *tp = tcp_sk(sk);
2656 struct sk_buff *skb;
2658 if (tp->retrans_out)
2661 skb = tcp_write_queue_head(sk);
2662 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2668 /* Undo during fast recovery after partial ACK. */
2670 static int tcp_try_undo_partial(struct sock *sk, int acked)
2672 struct tcp_sock *tp = tcp_sk(sk);
2673 /* Partial ACK arrived. Force Hoe's retransmit. */
2674 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2676 if (tcp_may_undo(tp)) {
2677 /* Plain luck! Hole if filled with delayed
2678 * packet, rather than with a retransmit.
2680 if (!tcp_any_retrans_done(sk))
2681 tp->retrans_stamp = 0;
2683 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2686 tcp_undo_cwr(sk, false);
2687 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2689 /* So... Do not make Hoe's retransmit yet.
2690 * If the first packet was delayed, the rest
2691 * ones are most probably delayed as well.
2698 /* Undo during loss recovery after partial ACK. */
2699 static bool tcp_try_undo_loss(struct sock *sk)
2701 struct tcp_sock *tp = tcp_sk(sk);
2703 if (tcp_may_undo(tp)) {
2704 struct sk_buff *skb;
2705 tcp_for_write_queue(skb, sk) {
2706 if (skb == tcp_send_head(sk))
2708 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2711 tcp_clear_all_retrans_hints(tp);
2713 DBGUNDO(sk, "partial loss");
2715 tcp_undo_cwr(sk, true);
2716 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2717 inet_csk(sk)->icsk_retransmits = 0;
2718 tp->undo_marker = 0;
2719 if (tcp_is_sack(tp))
2720 tcp_set_ca_state(sk, TCP_CA_Open);
2726 static inline void tcp_complete_cwr(struct sock *sk)
2728 struct tcp_sock *tp = tcp_sk(sk);
2730 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2731 if (tp->undo_marker) {
2732 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
2733 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2734 tp->snd_cwnd_stamp = tcp_time_stamp;
2735 } else if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH) {
2736 /* PRR algorithm. */
2737 tp->snd_cwnd = tp->snd_ssthresh;
2738 tp->snd_cwnd_stamp = tcp_time_stamp;
2741 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2744 static void tcp_try_keep_open(struct sock *sk)
2746 struct tcp_sock *tp = tcp_sk(sk);
2747 int state = TCP_CA_Open;
2749 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2750 state = TCP_CA_Disorder;
2752 if (inet_csk(sk)->icsk_ca_state != state) {
2753 tcp_set_ca_state(sk, state);
2754 tp->high_seq = tp->snd_nxt;
2758 static void tcp_try_to_open(struct sock *sk, int flag)
2760 struct tcp_sock *tp = tcp_sk(sk);
2762 tcp_verify_left_out(tp);
2764 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2765 tp->retrans_stamp = 0;
2767 if (flag & FLAG_ECE)
2768 tcp_enter_cwr(sk, 1);
2770 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2771 tcp_try_keep_open(sk);
2772 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2773 tcp_moderate_cwnd(tp);
2775 tcp_cwnd_down(sk, flag);
2779 static void tcp_mtup_probe_failed(struct sock *sk)
2781 struct inet_connection_sock *icsk = inet_csk(sk);
2783 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2784 icsk->icsk_mtup.probe_size = 0;
2787 static void tcp_mtup_probe_success(struct sock *sk)
2789 struct tcp_sock *tp = tcp_sk(sk);
2790 struct inet_connection_sock *icsk = inet_csk(sk);
2792 /* FIXME: breaks with very large cwnd */
2793 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2794 tp->snd_cwnd = tp->snd_cwnd *
2795 tcp_mss_to_mtu(sk, tp->mss_cache) /
2796 icsk->icsk_mtup.probe_size;
2797 tp->snd_cwnd_cnt = 0;
2798 tp->snd_cwnd_stamp = tcp_time_stamp;
2799 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2801 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2802 icsk->icsk_mtup.probe_size = 0;
2803 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2806 /* Do a simple retransmit without using the backoff mechanisms in
2807 * tcp_timer. This is used for path mtu discovery.
2808 * The socket is already locked here.
2810 void tcp_simple_retransmit(struct sock *sk)
2812 const struct inet_connection_sock *icsk = inet_csk(sk);
2813 struct tcp_sock *tp = tcp_sk(sk);
2814 struct sk_buff *skb;
2815 unsigned int mss = tcp_current_mss(sk);
2816 u32 prior_lost = tp->lost_out;
2818 tcp_for_write_queue(skb, sk) {
2819 if (skb == tcp_send_head(sk))
2821 if (tcp_skb_seglen(skb) > mss &&
2822 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2823 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2824 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2825 tp->retrans_out -= tcp_skb_pcount(skb);
2827 tcp_skb_mark_lost_uncond_verify(tp, skb);
2831 tcp_clear_retrans_hints_partial(tp);
2833 if (prior_lost == tp->lost_out)
2836 if (tcp_is_reno(tp))
2837 tcp_limit_reno_sacked(tp);
2839 tcp_verify_left_out(tp);
2841 /* Don't muck with the congestion window here.
2842 * Reason is that we do not increase amount of _data_
2843 * in network, but units changed and effective
2844 * cwnd/ssthresh really reduced now.
2846 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2847 tp->high_seq = tp->snd_nxt;
2848 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2849 tp->prior_ssthresh = 0;
2850 tp->undo_marker = 0;
2851 tcp_set_ca_state(sk, TCP_CA_Loss);
2853 tcp_xmit_retransmit_queue(sk);
2855 EXPORT_SYMBOL(tcp_simple_retransmit);
2857 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2858 * (proportional rate reduction with slow start reduction bound) as described in
2859 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2860 * It computes the number of packets to send (sndcnt) based on packets newly
2862 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2863 * cwnd reductions across a full RTT.
2864 * 2) If packets in flight is lower than ssthresh (such as due to excess
2865 * losses and/or application stalls), do not perform any further cwnd
2866 * reductions, but instead slow start up to ssthresh.
2868 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
2869 int fast_rexmit, int flag)
2871 struct tcp_sock *tp = tcp_sk(sk);
2873 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2875 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2876 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2878 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2880 sndcnt = min_t(int, delta,
2881 max_t(int, tp->prr_delivered - tp->prr_out,
2882 newly_acked_sacked) + 1);
2885 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2886 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2889 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2891 struct tcp_sock *tp = tcp_sk(sk);
2894 if (tcp_is_reno(tp))
2895 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2897 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2899 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2901 tp->high_seq = tp->snd_nxt;
2902 tp->prior_ssthresh = 0;
2903 tp->undo_marker = tp->snd_una;
2904 tp->undo_retrans = tp->retrans_out;
2906 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2908 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2909 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2910 TCP_ECN_queue_cwr(tp);
2913 tp->bytes_acked = 0;
2914 tp->snd_cwnd_cnt = 0;
2915 tp->prior_cwnd = tp->snd_cwnd;
2916 tp->prr_delivered = 0;
2918 tcp_set_ca_state(sk, TCP_CA_Recovery);
2921 /* Process an event, which can update packets-in-flight not trivially.
2922 * Main goal of this function is to calculate new estimate for left_out,
2923 * taking into account both packets sitting in receiver's buffer and
2924 * packets lost by network.
2926 * Besides that it does CWND reduction, when packet loss is detected
2927 * and changes state of machine.
2929 * It does _not_ decide what to send, it is made in function
2930 * tcp_xmit_retransmit_queue().
2932 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2933 int prior_sacked, bool is_dupack,
2936 struct inet_connection_sock *icsk = inet_csk(sk);
2937 struct tcp_sock *tp = tcp_sk(sk);
2938 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2939 (tcp_fackets_out(tp) > tp->reordering));
2940 int newly_acked_sacked = 0;
2941 int fast_rexmit = 0;
2943 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2945 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2946 tp->fackets_out = 0;
2948 /* Now state machine starts.
2949 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2950 if (flag & FLAG_ECE)
2951 tp->prior_ssthresh = 0;
2953 /* B. In all the states check for reneging SACKs. */
2954 if (tcp_check_sack_reneging(sk, flag))
2957 /* C. Check consistency of the current state. */
2958 tcp_verify_left_out(tp);
2960 /* D. Check state exit conditions. State can be terminated
2961 * when high_seq is ACKed. */
2962 if (icsk->icsk_ca_state == TCP_CA_Open) {
2963 WARN_ON(tp->retrans_out != 0);
2964 tp->retrans_stamp = 0;
2965 } else if (!before(tp->snd_una, tp->high_seq)) {
2966 switch (icsk->icsk_ca_state) {
2968 icsk->icsk_retransmits = 0;
2969 if (tcp_try_undo_recovery(sk))
2974 /* CWR is to be held something *above* high_seq
2975 * is ACKed for CWR bit to reach receiver. */
2976 if (tp->snd_una != tp->high_seq) {
2977 tcp_complete_cwr(sk);
2978 tcp_set_ca_state(sk, TCP_CA_Open);
2982 case TCP_CA_Recovery:
2983 if (tcp_is_reno(tp))
2984 tcp_reset_reno_sack(tp);
2985 if (tcp_try_undo_recovery(sk))
2987 tcp_complete_cwr(sk);
2992 /* E. Process state. */
2993 switch (icsk->icsk_ca_state) {
2994 case TCP_CA_Recovery:
2995 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2996 if (tcp_is_reno(tp) && is_dupack)
2997 tcp_add_reno_sack(sk);
2999 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3000 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
3003 if (flag & FLAG_DATA_ACKED)
3004 icsk->icsk_retransmits = 0;
3005 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3006 tcp_reset_reno_sack(tp);
3007 if (!tcp_try_undo_loss(sk)) {
3008 tcp_moderate_cwnd(tp);
3009 tcp_xmit_retransmit_queue(sk);
3012 if (icsk->icsk_ca_state != TCP_CA_Open)
3014 /* Loss is undone; fall through to processing in Open state. */
3016 if (tcp_is_reno(tp)) {
3017 if (flag & FLAG_SND_UNA_ADVANCED)
3018 tcp_reset_reno_sack(tp);
3020 tcp_add_reno_sack(sk);
3022 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
3024 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3025 tcp_try_undo_dsack(sk);
3027 if (!tcp_time_to_recover(sk, flag)) {
3028 tcp_try_to_open(sk, flag);
3032 /* MTU probe failure: don't reduce cwnd */
3033 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3034 icsk->icsk_mtup.probe_size &&
3035 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3036 tcp_mtup_probe_failed(sk);
3037 /* Restores the reduction we did in tcp_mtup_probe() */
3039 tcp_simple_retransmit(sk);
3043 /* Otherwise enter Recovery state */
3044 tcp_enter_recovery(sk, (flag & FLAG_ECE));
3048 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3049 tcp_update_scoreboard(sk, fast_rexmit);
3050 tp->prr_delivered += newly_acked_sacked;
3051 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3052 tcp_xmit_retransmit_queue(sk);
3055 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3057 tcp_rtt_estimator(sk, seq_rtt);
3059 inet_csk(sk)->icsk_backoff = 0;
3061 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3063 /* Read draft-ietf-tcplw-high-performance before mucking
3064 * with this code. (Supersedes RFC1323)
3066 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3068 /* RTTM Rule: A TSecr value received in a segment is used to
3069 * update the averaged RTT measurement only if the segment
3070 * acknowledges some new data, i.e., only if it advances the
3071 * left edge of the send window.
3073 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3074 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3076 * Changed: reset backoff as soon as we see the first valid sample.
3077 * If we do not, we get strongly overestimated rto. With timestamps
3078 * samples are accepted even from very old segments: f.e., when rtt=1
3079 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3080 * answer arrives rto becomes 120 seconds! If at least one of segments
3081 * in window is lost... Voila. --ANK (010210)
3083 struct tcp_sock *tp = tcp_sk(sk);
3085 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3088 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3090 /* We don't have a timestamp. Can only use
3091 * packets that are not retransmitted to determine
3092 * rtt estimates. Also, we must not reset the
3093 * backoff for rto until we get a non-retransmitted
3094 * packet. This allows us to deal with a situation
3095 * where the network delay has increased suddenly.
3096 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3099 if (flag & FLAG_RETRANS_DATA_ACKED)
3102 tcp_valid_rtt_meas(sk, seq_rtt);
3105 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3108 const struct tcp_sock *tp = tcp_sk(sk);
3109 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3110 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3111 tcp_ack_saw_tstamp(sk, flag);
3112 else if (seq_rtt >= 0)
3113 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3116 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3118 const struct inet_connection_sock *icsk = inet_csk(sk);
3119 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3120 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3123 /* Restart timer after forward progress on connection.
3124 * RFC2988 recommends to restart timer to now+rto.
3126 void tcp_rearm_rto(struct sock *sk)
3128 struct tcp_sock *tp = tcp_sk(sk);
3130 if (!tp->packets_out) {
3131 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3133 u32 rto = inet_csk(sk)->icsk_rto;
3134 /* Offset the time elapsed after installing regular RTO */
3135 if (tp->early_retrans_delayed) {
3136 struct sk_buff *skb = tcp_write_queue_head(sk);
3137 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
3138 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3139 /* delta may not be positive if the socket is locked
3140 * when the delayed ER timer fires and is rescheduled.
3145 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3148 tp->early_retrans_delayed = 0;
3151 /* This function is called when the delayed ER timer fires. TCP enters
3152 * fast recovery and performs fast-retransmit.
3154 void tcp_resume_early_retransmit(struct sock *sk)
3156 struct tcp_sock *tp = tcp_sk(sk);
3160 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3161 if (!tp->do_early_retrans)
3164 tcp_enter_recovery(sk, false);
3165 tcp_update_scoreboard(sk, 1);
3166 tcp_xmit_retransmit_queue(sk);
3169 /* If we get here, the whole TSO packet has not been acked. */
3170 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3172 struct tcp_sock *tp = tcp_sk(sk);
3175 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3177 packets_acked = tcp_skb_pcount(skb);
3178 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3180 packets_acked -= tcp_skb_pcount(skb);
3182 if (packets_acked) {
3183 BUG_ON(tcp_skb_pcount(skb) == 0);
3184 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3187 return packets_acked;
3190 /* Remove acknowledged frames from the retransmission queue. If our packet
3191 * is before the ack sequence we can discard it as it's confirmed to have
3192 * arrived at the other end.
3194 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3197 struct tcp_sock *tp = tcp_sk(sk);
3198 const struct inet_connection_sock *icsk = inet_csk(sk);
3199 struct sk_buff *skb;
3200 u32 now = tcp_time_stamp;
3201 int fully_acked = true;
3204 u32 reord = tp->packets_out;
3205 u32 prior_sacked = tp->sacked_out;
3207 s32 ca_seq_rtt = -1;
3208 ktime_t last_ackt = net_invalid_timestamp();
3210 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3211 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3213 u8 sacked = scb->sacked;
3215 /* Determine how many packets and what bytes were acked, tso and else */
3216 if (after(scb->end_seq, tp->snd_una)) {
3217 if (tcp_skb_pcount(skb) == 1 ||
3218 !after(tp->snd_una, scb->seq))
3221 acked_pcount = tcp_tso_acked(sk, skb);
3225 fully_acked = false;
3227 acked_pcount = tcp_skb_pcount(skb);
3230 if (sacked & TCPCB_RETRANS) {
3231 if (sacked & TCPCB_SACKED_RETRANS)
3232 tp->retrans_out -= acked_pcount;
3233 flag |= FLAG_RETRANS_DATA_ACKED;
3236 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3237 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3239 ca_seq_rtt = now - scb->when;
3240 last_ackt = skb->tstamp;
3242 seq_rtt = ca_seq_rtt;
3244 if (!(sacked & TCPCB_SACKED_ACKED))
3245 reord = min(pkts_acked, reord);
3248 if (sacked & TCPCB_SACKED_ACKED)
3249 tp->sacked_out -= acked_pcount;
3250 if (sacked & TCPCB_LOST)
3251 tp->lost_out -= acked_pcount;
3253 tp->packets_out -= acked_pcount;
3254 pkts_acked += acked_pcount;
3256 /* Initial outgoing SYN's get put onto the write_queue
3257 * just like anything else we transmit. It is not
3258 * true data, and if we misinform our callers that
3259 * this ACK acks real data, we will erroneously exit
3260 * connection startup slow start one packet too
3261 * quickly. This is severely frowned upon behavior.
3263 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3264 flag |= FLAG_DATA_ACKED;
3266 flag |= FLAG_SYN_ACKED;
3267 tp->retrans_stamp = 0;
3273 tcp_unlink_write_queue(skb, sk);
3274 sk_wmem_free_skb(sk, skb);
3275 tp->scoreboard_skb_hint = NULL;
3276 if (skb == tp->retransmit_skb_hint)
3277 tp->retransmit_skb_hint = NULL;
3278 if (skb == tp->lost_skb_hint)
3279 tp->lost_skb_hint = NULL;
3282 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3283 tp->snd_up = tp->snd_una;
3285 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3286 flag |= FLAG_SACK_RENEGING;
3288 if (flag & FLAG_ACKED) {
3289 const struct tcp_congestion_ops *ca_ops
3290 = inet_csk(sk)->icsk_ca_ops;
3292 if (unlikely(icsk->icsk_mtup.probe_size &&
3293 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3294 tcp_mtup_probe_success(sk);
3297 tcp_ack_update_rtt(sk, flag, seq_rtt);
3300 if (tcp_is_reno(tp)) {
3301 tcp_remove_reno_sacks(sk, pkts_acked);
3305 /* Non-retransmitted hole got filled? That's reordering */
3306 if (reord < prior_fackets)
3307 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3309 delta = tcp_is_fack(tp) ? pkts_acked :
3310 prior_sacked - tp->sacked_out;
3311 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3314 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3316 if (ca_ops->pkts_acked) {
3319 /* Is the ACK triggering packet unambiguous? */
3320 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3321 /* High resolution needed and available? */
3322 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3323 !ktime_equal(last_ackt,
3324 net_invalid_timestamp()))
3325 rtt_us = ktime_us_delta(ktime_get_real(),
3327 else if (ca_seq_rtt >= 0)
3328 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3331 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3335 #if FASTRETRANS_DEBUG > 0
3336 WARN_ON((int)tp->sacked_out < 0);
3337 WARN_ON((int)tp->lost_out < 0);
3338 WARN_ON((int)tp->retrans_out < 0);
3339 if (!tp->packets_out && tcp_is_sack(tp)) {
3340 icsk = inet_csk(sk);
3342 pr_debug("Leak l=%u %d\n",
3343 tp->lost_out, icsk->icsk_ca_state);
3346 if (tp->sacked_out) {
3347 pr_debug("Leak s=%u %d\n",
3348 tp->sacked_out, icsk->icsk_ca_state);
3351 if (tp->retrans_out) {
3352 pr_debug("Leak r=%u %d\n",
3353 tp->retrans_out, icsk->icsk_ca_state);
3354 tp->retrans_out = 0;
3361 static void tcp_ack_probe(struct sock *sk)
3363 const struct tcp_sock *tp = tcp_sk(sk);
3364 struct inet_connection_sock *icsk = inet_csk(sk);
3366 /* Was it a usable window open? */
3368 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3369 icsk->icsk_backoff = 0;
3370 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3371 /* Socket must be waked up by subsequent tcp_data_snd_check().
3372 * This function is not for random using!
3375 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3376 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3381 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3383 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3384 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3387 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3389 const struct tcp_sock *tp = tcp_sk(sk);
3390 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3391 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3394 /* Check that window update is acceptable.
3395 * The function assumes that snd_una<=ack<=snd_next.
3397 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3398 const u32 ack, const u32 ack_seq,
3401 return after(ack, tp->snd_una) ||
3402 after(ack_seq, tp->snd_wl1) ||
3403 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3406 /* Update our send window.
3408 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3409 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3411 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3414 struct tcp_sock *tp = tcp_sk(sk);
3416 u32 nwin = ntohs(tcp_hdr(skb)->window);
3418 if (likely(!tcp_hdr(skb)->syn))
3419 nwin <<= tp->rx_opt.snd_wscale;
3421 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3422 flag |= FLAG_WIN_UPDATE;
3423 tcp_update_wl(tp, ack_seq);
3425 if (tp->snd_wnd != nwin) {
3428 /* Note, it is the only place, where
3429 * fast path is recovered for sending TCP.
3432 tcp_fast_path_check(sk);
3434 if (nwin > tp->max_window) {
3435 tp->max_window = nwin;
3436 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3446 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3447 * continue in congestion avoidance.
3449 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3451 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3452 tp->snd_cwnd_cnt = 0;
3453 tp->bytes_acked = 0;
3454 TCP_ECN_queue_cwr(tp);
3455 tcp_moderate_cwnd(tp);
3458 /* A conservative spurious RTO response algorithm: reduce cwnd using
3459 * rate halving and continue in congestion avoidance.
3461 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3463 tcp_enter_cwr(sk, 0);
3466 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3468 if (flag & FLAG_ECE)
3469 tcp_ratehalving_spur_to_response(sk);
3471 tcp_undo_cwr(sk, true);
3474 /* F-RTO spurious RTO detection algorithm (RFC4138)
3476 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3477 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3478 * window (but not to or beyond highest sequence sent before RTO):
3479 * On First ACK, send two new segments out.
3480 * On Second ACK, RTO was likely spurious. Do spurious response (response
3481 * algorithm is not part of the F-RTO detection algorithm
3482 * given in RFC4138 but can be selected separately).
3483 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3484 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3485 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3486 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3488 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3489 * original window even after we transmit two new data segments.
3492 * on first step, wait until first cumulative ACK arrives, then move to
3493 * the second step. In second step, the next ACK decides.
3495 * F-RTO is implemented (mainly) in four functions:
3496 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3497 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3498 * called when tcp_use_frto() showed green light
3499 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3500 * - tcp_enter_frto_loss() is called if there is not enough evidence
3501 * to prove that the RTO is indeed spurious. It transfers the control
3502 * from F-RTO to the conventional RTO recovery
3504 static bool tcp_process_frto(struct sock *sk, int flag)
3506 struct tcp_sock *tp = tcp_sk(sk);
3508 tcp_verify_left_out(tp);
3510 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3511 if (flag & FLAG_DATA_ACKED)
3512 inet_csk(sk)->icsk_retransmits = 0;
3514 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3515 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3516 tp->undo_marker = 0;
3518 if (!before(tp->snd_una, tp->frto_highmark)) {
3519 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3523 if (!tcp_is_sackfrto(tp)) {
3524 /* RFC4138 shortcoming in step 2; should also have case c):
3525 * ACK isn't duplicate nor advances window, e.g., opposite dir
3528 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3531 if (!(flag & FLAG_DATA_ACKED)) {
3532 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3537 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3538 /* Prevent sending of new data. */
3539 tp->snd_cwnd = min(tp->snd_cwnd,
3540 tcp_packets_in_flight(tp));
3544 if ((tp->frto_counter >= 2) &&
3545 (!(flag & FLAG_FORWARD_PROGRESS) ||
3546 ((flag & FLAG_DATA_SACKED) &&
3547 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3548 /* RFC4138 shortcoming (see comment above) */
3549 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3550 (flag & FLAG_NOT_DUP))
3553 tcp_enter_frto_loss(sk, 3, flag);
3558 if (tp->frto_counter == 1) {
3559 /* tcp_may_send_now needs to see updated state */
3560 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3561 tp->frto_counter = 2;
3563 if (!tcp_may_send_now(sk))
3564 tcp_enter_frto_loss(sk, 2, flag);
3568 switch (sysctl_tcp_frto_response) {
3570 tcp_undo_spur_to_response(sk, flag);
3573 tcp_conservative_spur_to_response(tp);
3576 tcp_ratehalving_spur_to_response(sk);
3579 tp->frto_counter = 0;
3580 tp->undo_marker = 0;
3581 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3586 /* This routine deals with incoming acks, but not outgoing ones. */
3587 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3589 struct inet_connection_sock *icsk = inet_csk(sk);
3590 struct tcp_sock *tp = tcp_sk(sk);
3591 u32 prior_snd_una = tp->snd_una;
3592 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3593 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3594 bool is_dupack = false;
3595 u32 prior_in_flight;
3598 int prior_sacked = tp->sacked_out;
3600 bool frto_cwnd = false;
3602 /* If the ack is older than previous acks
3603 * then we can probably ignore it.
3605 if (before(ack, prior_snd_una))
3608 /* If the ack includes data we haven't sent yet, discard
3609 * this segment (RFC793 Section 3.9).
3611 if (after(ack, tp->snd_nxt))
3614 if (tp->early_retrans_delayed)
3617 if (after(ack, prior_snd_una))
3618 flag |= FLAG_SND_UNA_ADVANCED;
3620 if (sysctl_tcp_abc) {
3621 if (icsk->icsk_ca_state < TCP_CA_CWR)
3622 tp->bytes_acked += ack - prior_snd_una;
3623 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3624 /* we assume just one segment left network */
3625 tp->bytes_acked += min(ack - prior_snd_una,
3629 prior_fackets = tp->fackets_out;
3630 prior_in_flight = tcp_packets_in_flight(tp);
3632 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3633 /* Window is constant, pure forward advance.
3634 * No more checks are required.
3635 * Note, we use the fact that SND.UNA>=SND.WL2.
3637 tcp_update_wl(tp, ack_seq);
3639 flag |= FLAG_WIN_UPDATE;
3641 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3643 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3645 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3648 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3650 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3652 if (TCP_SKB_CB(skb)->sacked)
3653 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3655 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3658 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3661 /* We passed data and got it acked, remove any soft error
3662 * log. Something worked...
3664 sk->sk_err_soft = 0;
3665 icsk->icsk_probes_out = 0;
3666 tp->rcv_tstamp = tcp_time_stamp;
3667 prior_packets = tp->packets_out;
3671 /* See if we can take anything off of the retransmit queue. */
3672 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3674 pkts_acked = prior_packets - tp->packets_out;
3676 if (tp->frto_counter)
3677 frto_cwnd = tcp_process_frto(sk, flag);
3678 /* Guarantee sacktag reordering detection against wrap-arounds */
3679 if (before(tp->frto_highmark, tp->snd_una))
3680 tp->frto_highmark = 0;
3682 if (tcp_ack_is_dubious(sk, flag)) {
3683 /* Advance CWND, if state allows this. */
3684 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3685 tcp_may_raise_cwnd(sk, flag))
3686 tcp_cong_avoid(sk, ack, prior_in_flight);
3687 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3688 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3691 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3692 tcp_cong_avoid(sk, ack, prior_in_flight);
3695 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3696 struct dst_entry *dst = __sk_dst_get(sk);
3703 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3704 if (flag & FLAG_DSACKING_ACK)
3705 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3707 /* If this ack opens up a zero window, clear backoff. It was
3708 * being used to time the probes, and is probably far higher than
3709 * it needs to be for normal retransmission.
3711 if (tcp_send_head(sk))
3716 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3720 /* If data was SACKed, tag it and see if we should send more data.
3721 * If data was DSACKed, see if we can undo a cwnd reduction.
3723 if (TCP_SKB_CB(skb)->sacked) {
3724 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3725 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3729 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3733 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3734 * But, this can also be called on packets in the established flow when
3735 * the fast version below fails.
3737 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3738 const u8 **hvpp, int estab,
3739 struct tcp_fastopen_cookie *foc)
3741 const unsigned char *ptr;
3742 const struct tcphdr *th = tcp_hdr(skb);
3743 int length = (th->doff * 4) - sizeof(struct tcphdr);
3745 ptr = (const unsigned char *)(th + 1);
3746 opt_rx->saw_tstamp = 0;
3748 while (length > 0) {
3749 int opcode = *ptr++;
3755 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3760 if (opsize < 2) /* "silly options" */
3762 if (opsize > length)
3763 return; /* don't parse partial options */
3766 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3767 u16 in_mss = get_unaligned_be16(ptr);
3769 if (opt_rx->user_mss &&
3770 opt_rx->user_mss < in_mss)
3771 in_mss = opt_rx->user_mss;
3772 opt_rx->mss_clamp = in_mss;
3777 if (opsize == TCPOLEN_WINDOW && th->syn &&
3778 !estab && sysctl_tcp_window_scaling) {
3779 __u8 snd_wscale = *(__u8 *)ptr;
3780 opt_rx->wscale_ok = 1;
3781 if (snd_wscale > 14) {
3782 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3787 opt_rx->snd_wscale = snd_wscale;
3790 case TCPOPT_TIMESTAMP:
3791 if ((opsize == TCPOLEN_TIMESTAMP) &&
3792 ((estab && opt_rx->tstamp_ok) ||
3793 (!estab && sysctl_tcp_timestamps))) {
3794 opt_rx->saw_tstamp = 1;
3795 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3796 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3799 case TCPOPT_SACK_PERM:
3800 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3801 !estab && sysctl_tcp_sack) {
3802 opt_rx->sack_ok = TCP_SACK_SEEN;
3803 tcp_sack_reset(opt_rx);
3808 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3809 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3811 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3814 #ifdef CONFIG_TCP_MD5SIG
3817 * The MD5 Hash has already been
3818 * checked (see tcp_v{4,6}_do_rcv()).
3823 /* This option is variable length.
3826 case TCPOLEN_COOKIE_BASE:
3827 /* not yet implemented */
3829 case TCPOLEN_COOKIE_PAIR:
3830 /* not yet implemented */
3832 case TCPOLEN_COOKIE_MIN+0:
3833 case TCPOLEN_COOKIE_MIN+2:
3834 case TCPOLEN_COOKIE_MIN+4:
3835 case TCPOLEN_COOKIE_MIN+6:
3836 case TCPOLEN_COOKIE_MAX:
3837 /* 16-bit multiple */
3838 opt_rx->cookie_plus = opsize;
3848 /* Fast Open option shares code 254 using a
3849 * 16 bits magic number. It's valid only in
3850 * SYN or SYN-ACK with an even size.
3852 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3853 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3854 foc == NULL || !th->syn || (opsize & 1))
3856 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3857 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3858 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3859 memcpy(foc->val, ptr + 2, foc->len);
3860 else if (foc->len != 0)
3870 EXPORT_SYMBOL(tcp_parse_options);
3872 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3874 const __be32 *ptr = (const __be32 *)(th + 1);
3876 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3877 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3878 tp->rx_opt.saw_tstamp = 1;
3880 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3882 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3888 /* Fast parse options. This hopes to only see timestamps.
3889 * If it is wrong it falls back on tcp_parse_options().
3891 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3892 const struct tcphdr *th,
3893 struct tcp_sock *tp, const u8 **hvpp)
3895 /* In the spirit of fast parsing, compare doff directly to constant
3896 * values. Because equality is used, short doff can be ignored here.
3898 if (th->doff == (sizeof(*th) / 4)) {
3899 tp->rx_opt.saw_tstamp = 0;
3901 } else if (tp->rx_opt.tstamp_ok &&
3902 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3903 if (tcp_parse_aligned_timestamp(tp, th))
3906 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1, NULL);
3910 #ifdef CONFIG_TCP_MD5SIG
3912 * Parse MD5 Signature option
3914 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3916 int length = (th->doff << 2) - sizeof(*th);
3917 const u8 *ptr = (const u8 *)(th + 1);
3919 /* If the TCP option is too short, we can short cut */
3920 if (length < TCPOLEN_MD5SIG)
3923 while (length > 0) {
3924 int opcode = *ptr++;
3935 if (opsize < 2 || opsize > length)
3937 if (opcode == TCPOPT_MD5SIG)
3938 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3945 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3948 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3950 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3951 tp->rx_opt.ts_recent_stamp = get_seconds();
3954 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3956 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3957 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3958 * extra check below makes sure this can only happen
3959 * for pure ACK frames. -DaveM
3961 * Not only, also it occurs for expired timestamps.
3964 if (tcp_paws_check(&tp->rx_opt, 0))
3965 tcp_store_ts_recent(tp);
3969 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3971 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3972 * it can pass through stack. So, the following predicate verifies that
3973 * this segment is not used for anything but congestion avoidance or
3974 * fast retransmit. Moreover, we even are able to eliminate most of such
3975 * second order effects, if we apply some small "replay" window (~RTO)
3976 * to timestamp space.
3978 * All these measures still do not guarantee that we reject wrapped ACKs
3979 * on networks with high bandwidth, when sequence space is recycled fastly,
3980 * but it guarantees that such events will be very rare and do not affect
3981 * connection seriously. This doesn't look nice, but alas, PAWS is really
3984 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3985 * states that events when retransmit arrives after original data are rare.
3986 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3987 * the biggest problem on large power networks even with minor reordering.
3988 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3989 * up to bandwidth of 18Gigabit/sec. 8) ]
3992 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3994 const struct tcp_sock *tp = tcp_sk(sk);
3995 const struct tcphdr *th = tcp_hdr(skb);
3996 u32 seq = TCP_SKB_CB(skb)->seq;
3997 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3999 return (/* 1. Pure ACK with correct sequence number. */
4000 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4002 /* 2. ... and duplicate ACK. */
4003 ack == tp->snd_una &&
4005 /* 3. ... and does not update window. */
4006 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4008 /* 4. ... and sits in replay window. */
4009 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4012 static inline bool tcp_paws_discard(const struct sock *sk,
4013 const struct sk_buff *skb)
4015 const struct tcp_sock *tp = tcp_sk(sk);
4017 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4018 !tcp_disordered_ack(sk, skb);
4021 /* Check segment sequence number for validity.
4023 * Segment controls are considered valid, if the segment
4024 * fits to the window after truncation to the window. Acceptability
4025 * of data (and SYN, FIN, of course) is checked separately.
4026 * See tcp_data_queue(), for example.
4028 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4029 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4030 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4031 * (borrowed from freebsd)
4034 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4036 return !before(end_seq, tp->rcv_wup) &&
4037 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4040 /* When we get a reset we do this. */
4041 static void tcp_reset(struct sock *sk)
4043 /* We want the right error as BSD sees it (and indeed as we do). */
4044 switch (sk->sk_state) {
4046 sk->sk_err = ECONNREFUSED;
4048 case TCP_CLOSE_WAIT:
4054 sk->sk_err = ECONNRESET;
4056 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4059 if (!sock_flag(sk, SOCK_DEAD))
4060 sk->sk_error_report(sk);
4066 * Process the FIN bit. This now behaves as it is supposed to work
4067 * and the FIN takes effect when it is validly part of sequence
4068 * space. Not before when we get holes.
4070 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4071 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4074 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4075 * close and we go into CLOSING (and later onto TIME-WAIT)
4077 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4079 static void tcp_fin(struct sock *sk)
4081 struct tcp_sock *tp = tcp_sk(sk);
4083 inet_csk_schedule_ack(sk);
4085 sk->sk_shutdown |= RCV_SHUTDOWN;
4086 sock_set_flag(sk, SOCK_DONE);
4088 switch (sk->sk_state) {
4090 case TCP_ESTABLISHED:
4091 /* Move to CLOSE_WAIT */
4092 tcp_set_state(sk, TCP_CLOSE_WAIT);
4093 inet_csk(sk)->icsk_ack.pingpong = 1;
4096 case TCP_CLOSE_WAIT:
4098 /* Received a retransmission of the FIN, do
4103 /* RFC793: Remain in the LAST-ACK state. */
4107 /* This case occurs when a simultaneous close
4108 * happens, we must ack the received FIN and
4109 * enter the CLOSING state.
4112 tcp_set_state(sk, TCP_CLOSING);
4115 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4117 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4120 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4121 * cases we should never reach this piece of code.
4123 pr_err("%s: Impossible, sk->sk_state=%d\n",
4124 __func__, sk->sk_state);
4128 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4129 * Probably, we should reset in this case. For now drop them.
4131 __skb_queue_purge(&tp->out_of_order_queue);
4132 if (tcp_is_sack(tp))
4133 tcp_sack_reset(&tp->rx_opt);
4136 if (!sock_flag(sk, SOCK_DEAD)) {
4137 sk->sk_state_change(sk);
4139 /* Do not send POLL_HUP for half duplex close. */
4140 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4141 sk->sk_state == TCP_CLOSE)
4142 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4144 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4148 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4151 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4152 if (before(seq, sp->start_seq))
4153 sp->start_seq = seq;
4154 if (after(end_seq, sp->end_seq))
4155 sp->end_seq = end_seq;
4161 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4163 struct tcp_sock *tp = tcp_sk(sk);
4165 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4168 if (before(seq, tp->rcv_nxt))
4169 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4171 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4173 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4175 tp->rx_opt.dsack = 1;
4176 tp->duplicate_sack[0].start_seq = seq;
4177 tp->duplicate_sack[0].end_seq = end_seq;
4181 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4183 struct tcp_sock *tp = tcp_sk(sk);
4185 if (!tp->rx_opt.dsack)
4186 tcp_dsack_set(sk, seq, end_seq);
4188 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4191 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4193 struct tcp_sock *tp = tcp_sk(sk);
4195 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4196 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4197 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4198 tcp_enter_quickack_mode(sk);
4200 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4201 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4203 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4204 end_seq = tp->rcv_nxt;
4205 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4212 /* These routines update the SACK block as out-of-order packets arrive or
4213 * in-order packets close up the sequence space.
4215 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4218 struct tcp_sack_block *sp = &tp->selective_acks[0];
4219 struct tcp_sack_block *swalk = sp + 1;
4221 /* See if the recent change to the first SACK eats into
4222 * or hits the sequence space of other SACK blocks, if so coalesce.
4224 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4225 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4228 /* Zap SWALK, by moving every further SACK up by one slot.
4229 * Decrease num_sacks.
4231 tp->rx_opt.num_sacks--;
4232 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4236 this_sack++, swalk++;
4240 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4242 struct tcp_sock *tp = tcp_sk(sk);
4243 struct tcp_sack_block *sp = &tp->selective_acks[0];
4244 int cur_sacks = tp->rx_opt.num_sacks;
4250 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4251 if (tcp_sack_extend(sp, seq, end_seq)) {
4252 /* Rotate this_sack to the first one. */
4253 for (; this_sack > 0; this_sack--, sp--)
4254 swap(*sp, *(sp - 1));
4256 tcp_sack_maybe_coalesce(tp);
4261 /* Could not find an adjacent existing SACK, build a new one,
4262 * put it at the front, and shift everyone else down. We
4263 * always know there is at least one SACK present already here.
4265 * If the sack array is full, forget about the last one.
4267 if (this_sack >= TCP_NUM_SACKS) {
4269 tp->rx_opt.num_sacks--;
4272 for (; this_sack > 0; this_sack--, sp--)
4276 /* Build the new head SACK, and we're done. */
4277 sp->start_seq = seq;
4278 sp->end_seq = end_seq;
4279 tp->rx_opt.num_sacks++;
4282 /* RCV.NXT advances, some SACKs should be eaten. */
4284 static void tcp_sack_remove(struct tcp_sock *tp)
4286 struct tcp_sack_block *sp = &tp->selective_acks[0];
4287 int num_sacks = tp->rx_opt.num_sacks;
4290 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4291 if (skb_queue_empty(&tp->out_of_order_queue)) {
4292 tp->rx_opt.num_sacks = 0;
4296 for (this_sack = 0; this_sack < num_sacks;) {
4297 /* Check if the start of the sack is covered by RCV.NXT. */
4298 if (!before(tp->rcv_nxt, sp->start_seq)) {
4301 /* RCV.NXT must cover all the block! */
4302 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4304 /* Zap this SACK, by moving forward any other SACKS. */
4305 for (i=this_sack+1; i < num_sacks; i++)
4306 tp->selective_acks[i-1] = tp->selective_acks[i];
4313 tp->rx_opt.num_sacks = num_sacks;
4316 /* This one checks to see if we can put data from the
4317 * out_of_order queue into the receive_queue.
4319 static void tcp_ofo_queue(struct sock *sk)
4321 struct tcp_sock *tp = tcp_sk(sk);
4322 __u32 dsack_high = tp->rcv_nxt;
4323 struct sk_buff *skb;
4325 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4326 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4329 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4330 __u32 dsack = dsack_high;
4331 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4332 dsack_high = TCP_SKB_CB(skb)->end_seq;
4333 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4336 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4337 SOCK_DEBUG(sk, "ofo packet was already received\n");
4338 __skb_unlink(skb, &tp->out_of_order_queue);
4342 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4343 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4344 TCP_SKB_CB(skb)->end_seq);
4346 __skb_unlink(skb, &tp->out_of_order_queue);
4347 __skb_queue_tail(&sk->sk_receive_queue, skb);
4348 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4349 if (tcp_hdr(skb)->fin)
4354 static bool tcp_prune_ofo_queue(struct sock *sk);
4355 static int tcp_prune_queue(struct sock *sk);
4357 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4360 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4361 !sk_rmem_schedule(sk, skb, size)) {
4363 if (tcp_prune_queue(sk) < 0)
4366 if (!sk_rmem_schedule(sk, skb, size)) {
4367 if (!tcp_prune_ofo_queue(sk))
4370 if (!sk_rmem_schedule(sk, skb, size))
4378 * tcp_try_coalesce - try to merge skb to prior one
4381 * @from: buffer to add in queue
4382 * @fragstolen: pointer to boolean
4384 * Before queueing skb @from after @to, try to merge them
4385 * to reduce overall memory use and queue lengths, if cost is small.
4386 * Packets in ofo or receive queues can stay a long time.
4387 * Better try to coalesce them right now to avoid future collapses.
4388 * Returns true if caller should free @from instead of queueing it
4390 static bool tcp_try_coalesce(struct sock *sk,
4392 struct sk_buff *from,
4397 *fragstolen = false;
4399 if (tcp_hdr(from)->fin)
4402 /* Its possible this segment overlaps with prior segment in queue */
4403 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4406 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4409 atomic_add(delta, &sk->sk_rmem_alloc);
4410 sk_mem_charge(sk, delta);
4411 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4412 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4413 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4417 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4419 struct tcp_sock *tp = tcp_sk(sk);
4420 struct sk_buff *skb1;
4423 TCP_ECN_check_ce(tp, skb);
4425 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4426 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4431 /* Disable header prediction. */
4433 inet_csk_schedule_ack(sk);
4435 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4436 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4437 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4439 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4441 /* Initial out of order segment, build 1 SACK. */
4442 if (tcp_is_sack(tp)) {
4443 tp->rx_opt.num_sacks = 1;
4444 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4445 tp->selective_acks[0].end_seq =
4446 TCP_SKB_CB(skb)->end_seq;
4448 __skb_queue_head(&tp->out_of_order_queue, skb);
4452 seq = TCP_SKB_CB(skb)->seq;
4453 end_seq = TCP_SKB_CB(skb)->end_seq;
4455 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4458 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4459 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4461 kfree_skb_partial(skb, fragstolen);
4465 if (!tp->rx_opt.num_sacks ||
4466 tp->selective_acks[0].end_seq != seq)
4469 /* Common case: data arrive in order after hole. */
4470 tp->selective_acks[0].end_seq = end_seq;
4474 /* Find place to insert this segment. */
4476 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4478 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4482 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4485 /* Do skb overlap to previous one? */
4486 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4487 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4488 /* All the bits are present. Drop. */
4489 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4492 tcp_dsack_set(sk, seq, end_seq);
4495 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4496 /* Partial overlap. */
4497 tcp_dsack_set(sk, seq,
4498 TCP_SKB_CB(skb1)->end_seq);
4500 if (skb_queue_is_first(&tp->out_of_order_queue,
4504 skb1 = skb_queue_prev(
4505 &tp->out_of_order_queue,
4510 __skb_queue_head(&tp->out_of_order_queue, skb);
4512 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4514 /* And clean segments covered by new one as whole. */
4515 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4516 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4518 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4520 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4521 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4525 __skb_unlink(skb1, &tp->out_of_order_queue);
4526 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4527 TCP_SKB_CB(skb1)->end_seq);
4528 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4533 if (tcp_is_sack(tp))
4534 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4537 skb_set_owner_r(skb, sk);
4540 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4544 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4546 __skb_pull(skb, hdrlen);
4548 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4549 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4551 __skb_queue_tail(&sk->sk_receive_queue, skb);
4552 skb_set_owner_r(skb, sk);
4557 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4559 struct sk_buff *skb = NULL;
4563 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4567 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4570 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4571 skb_reset_transport_header(skb);
4572 memset(th, 0, sizeof(*th));
4574 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4577 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4578 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4579 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4581 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4582 WARN_ON_ONCE(fragstolen); /* should not happen */
4593 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4595 const struct tcphdr *th = tcp_hdr(skb);
4596 struct tcp_sock *tp = tcp_sk(sk);
4598 bool fragstolen = false;
4600 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4604 __skb_pull(skb, th->doff * 4);
4606 TCP_ECN_accept_cwr(tp, skb);
4608 tp->rx_opt.dsack = 0;
4610 /* Queue data for delivery to the user.
4611 * Packets in sequence go to the receive queue.
4612 * Out of sequence packets to the out_of_order_queue.
4614 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4615 if (tcp_receive_window(tp) == 0)
4618 /* Ok. In sequence. In window. */
4619 if (tp->ucopy.task == current &&
4620 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4621 sock_owned_by_user(sk) && !tp->urg_data) {
4622 int chunk = min_t(unsigned int, skb->len,
4625 __set_current_state(TASK_RUNNING);
4628 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4629 tp->ucopy.len -= chunk;
4630 tp->copied_seq += chunk;
4631 eaten = (chunk == skb->len);
4632 tcp_rcv_space_adjust(sk);
4640 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4643 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4645 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4647 tcp_event_data_recv(sk, skb);
4651 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4654 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4655 * gap in queue is filled.
4657 if (skb_queue_empty(&tp->out_of_order_queue))
4658 inet_csk(sk)->icsk_ack.pingpong = 0;
4661 if (tp->rx_opt.num_sacks)
4662 tcp_sack_remove(tp);
4664 tcp_fast_path_check(sk);
4667 kfree_skb_partial(skb, fragstolen);
4668 else if (!sock_flag(sk, SOCK_DEAD))
4669 sk->sk_data_ready(sk, 0);
4673 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4674 /* A retransmit, 2nd most common case. Force an immediate ack. */
4675 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4676 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4679 tcp_enter_quickack_mode(sk);
4680 inet_csk_schedule_ack(sk);
4686 /* Out of window. F.e. zero window probe. */
4687 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4690 tcp_enter_quickack_mode(sk);
4692 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4693 /* Partial packet, seq < rcv_next < end_seq */
4694 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4695 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4696 TCP_SKB_CB(skb)->end_seq);
4698 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4700 /* If window is closed, drop tail of packet. But after
4701 * remembering D-SACK for its head made in previous line.
4703 if (!tcp_receive_window(tp))
4708 tcp_data_queue_ofo(sk, skb);
4711 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4712 struct sk_buff_head *list)
4714 struct sk_buff *next = NULL;
4716 if (!skb_queue_is_last(list, skb))
4717 next = skb_queue_next(list, skb);
4719 __skb_unlink(skb, list);
4721 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4726 /* Collapse contiguous sequence of skbs head..tail with
4727 * sequence numbers start..end.
4729 * If tail is NULL, this means until the end of the list.
4731 * Segments with FIN/SYN are not collapsed (only because this
4735 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4736 struct sk_buff *head, struct sk_buff *tail,
4739 struct sk_buff *skb, *n;
4742 /* First, check that queue is collapsible and find
4743 * the point where collapsing can be useful. */
4747 skb_queue_walk_from_safe(list, skb, n) {
4750 /* No new bits? It is possible on ofo queue. */
4751 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4752 skb = tcp_collapse_one(sk, skb, list);
4758 /* The first skb to collapse is:
4760 * - bloated or contains data before "start" or
4761 * overlaps to the next one.
4763 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4764 (tcp_win_from_space(skb->truesize) > skb->len ||
4765 before(TCP_SKB_CB(skb)->seq, start))) {
4766 end_of_skbs = false;
4770 if (!skb_queue_is_last(list, skb)) {
4771 struct sk_buff *next = skb_queue_next(list, skb);
4773 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4774 end_of_skbs = false;
4779 /* Decided to skip this, advance start seq. */
4780 start = TCP_SKB_CB(skb)->end_seq;
4782 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4785 while (before(start, end)) {
4786 struct sk_buff *nskb;
4787 unsigned int header = skb_headroom(skb);
4788 int copy = SKB_MAX_ORDER(header, 0);
4790 /* Too big header? This can happen with IPv6. */
4793 if (end - start < copy)
4795 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4799 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4800 skb_set_network_header(nskb, (skb_network_header(skb) -
4802 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4804 skb_reserve(nskb, header);
4805 memcpy(nskb->head, skb->head, header);
4806 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4807 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4808 __skb_queue_before(list, skb, nskb);
4809 skb_set_owner_r(nskb, sk);
4811 /* Copy data, releasing collapsed skbs. */
4813 int offset = start - TCP_SKB_CB(skb)->seq;
4814 int size = TCP_SKB_CB(skb)->end_seq - start;
4818 size = min(copy, size);
4819 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4821 TCP_SKB_CB(nskb)->end_seq += size;
4825 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4826 skb = tcp_collapse_one(sk, skb, list);
4829 tcp_hdr(skb)->syn ||
4837 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4838 * and tcp_collapse() them until all the queue is collapsed.
4840 static void tcp_collapse_ofo_queue(struct sock *sk)
4842 struct tcp_sock *tp = tcp_sk(sk);
4843 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4844 struct sk_buff *head;
4850 start = TCP_SKB_CB(skb)->seq;
4851 end = TCP_SKB_CB(skb)->end_seq;
4855 struct sk_buff *next = NULL;
4857 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4858 next = skb_queue_next(&tp->out_of_order_queue, skb);
4861 /* Segment is terminated when we see gap or when
4862 * we are at the end of all the queue. */
4864 after(TCP_SKB_CB(skb)->seq, end) ||
4865 before(TCP_SKB_CB(skb)->end_seq, start)) {
4866 tcp_collapse(sk, &tp->out_of_order_queue,
4867 head, skb, start, end);
4871 /* Start new segment */
4872 start = TCP_SKB_CB(skb)->seq;
4873 end = TCP_SKB_CB(skb)->end_seq;
4875 if (before(TCP_SKB_CB(skb)->seq, start))
4876 start = TCP_SKB_CB(skb)->seq;
4877 if (after(TCP_SKB_CB(skb)->end_seq, end))
4878 end = TCP_SKB_CB(skb)->end_seq;
4884 * Purge the out-of-order queue.
4885 * Return true if queue was pruned.
4887 static bool tcp_prune_ofo_queue(struct sock *sk)
4889 struct tcp_sock *tp = tcp_sk(sk);
4892 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4893 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4894 __skb_queue_purge(&tp->out_of_order_queue);
4896 /* Reset SACK state. A conforming SACK implementation will
4897 * do the same at a timeout based retransmit. When a connection
4898 * is in a sad state like this, we care only about integrity
4899 * of the connection not performance.
4901 if (tp->rx_opt.sack_ok)
4902 tcp_sack_reset(&tp->rx_opt);
4909 /* Reduce allocated memory if we can, trying to get
4910 * the socket within its memory limits again.
4912 * Return less than zero if we should start dropping frames
4913 * until the socket owning process reads some of the data
4914 * to stabilize the situation.
4916 static int tcp_prune_queue(struct sock *sk)
4918 struct tcp_sock *tp = tcp_sk(sk);
4920 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4922 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4924 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4925 tcp_clamp_window(sk);
4926 else if (sk_under_memory_pressure(sk))
4927 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4929 tcp_collapse_ofo_queue(sk);
4930 if (!skb_queue_empty(&sk->sk_receive_queue))
4931 tcp_collapse(sk, &sk->sk_receive_queue,
4932 skb_peek(&sk->sk_receive_queue),
4934 tp->copied_seq, tp->rcv_nxt);
4937 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4940 /* Collapsing did not help, destructive actions follow.
4941 * This must not ever occur. */
4943 tcp_prune_ofo_queue(sk);
4945 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4948 /* If we are really being abused, tell the caller to silently
4949 * drop receive data on the floor. It will get retransmitted
4950 * and hopefully then we'll have sufficient space.
4952 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4954 /* Massive buffer overcommit. */
4959 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4960 * As additional protections, we do not touch cwnd in retransmission phases,
4961 * and if application hit its sndbuf limit recently.
4963 void tcp_cwnd_application_limited(struct sock *sk)
4965 struct tcp_sock *tp = tcp_sk(sk);
4967 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4968 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4969 /* Limited by application or receiver window. */
4970 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4971 u32 win_used = max(tp->snd_cwnd_used, init_win);
4972 if (win_used < tp->snd_cwnd) {
4973 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4974 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4976 tp->snd_cwnd_used = 0;
4978 tp->snd_cwnd_stamp = tcp_time_stamp;
4981 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4983 const struct tcp_sock *tp = tcp_sk(sk);
4985 /* If the user specified a specific send buffer setting, do
4988 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4991 /* If we are under global TCP memory pressure, do not expand. */
4992 if (sk_under_memory_pressure(sk))
4995 /* If we are under soft global TCP memory pressure, do not expand. */
4996 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4999 /* If we filled the congestion window, do not expand. */
5000 if (tp->packets_out >= tp->snd_cwnd)
5006 /* When incoming ACK allowed to free some skb from write_queue,
5007 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5008 * on the exit from tcp input handler.
5010 * PROBLEM: sndbuf expansion does not work well with largesend.
5012 static void tcp_new_space(struct sock *sk)
5014 struct tcp_sock *tp = tcp_sk(sk);
5016 if (tcp_should_expand_sndbuf(sk)) {
5017 int sndmem = SKB_TRUESIZE(max_t(u32,
5018 tp->rx_opt.mss_clamp,
5021 int demanded = max_t(unsigned int, tp->snd_cwnd,
5022 tp->reordering + 1);
5023 sndmem *= 2 * demanded;
5024 if (sndmem > sk->sk_sndbuf)
5025 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5026 tp->snd_cwnd_stamp = tcp_time_stamp;
5029 sk->sk_write_space(sk);
5032 static void tcp_check_space(struct sock *sk)
5034 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5035 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5036 if (sk->sk_socket &&
5037 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5042 static inline void tcp_data_snd_check(struct sock *sk)
5044 tcp_push_pending_frames(sk);
5045 tcp_check_space(sk);
5049 * Check if sending an ack is needed.
5051 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5053 struct tcp_sock *tp = tcp_sk(sk);
5055 /* More than one full frame received... */
5056 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5057 /* ... and right edge of window advances far enough.
5058 * (tcp_recvmsg() will send ACK otherwise). Or...
5060 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5061 /* We ACK each frame or... */
5062 tcp_in_quickack_mode(sk) ||
5063 /* We have out of order data. */
5064 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5065 /* Then ack it now */
5068 /* Else, send delayed ack. */
5069 tcp_send_delayed_ack(sk);
5073 static inline void tcp_ack_snd_check(struct sock *sk)
5075 if (!inet_csk_ack_scheduled(sk)) {
5076 /* We sent a data segment already. */
5079 __tcp_ack_snd_check(sk, 1);
5083 * This routine is only called when we have urgent data
5084 * signaled. Its the 'slow' part of tcp_urg. It could be
5085 * moved inline now as tcp_urg is only called from one
5086 * place. We handle URGent data wrong. We have to - as
5087 * BSD still doesn't use the correction from RFC961.
5088 * For 1003.1g we should support a new option TCP_STDURG to permit
5089 * either form (or just set the sysctl tcp_stdurg).
5092 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5094 struct tcp_sock *tp = tcp_sk(sk);
5095 u32 ptr = ntohs(th->urg_ptr);
5097 if (ptr && !sysctl_tcp_stdurg)
5099 ptr += ntohl(th->seq);
5101 /* Ignore urgent data that we've already seen and read. */
5102 if (after(tp->copied_seq, ptr))
5105 /* Do not replay urg ptr.
5107 * NOTE: interesting situation not covered by specs.
5108 * Misbehaving sender may send urg ptr, pointing to segment,
5109 * which we already have in ofo queue. We are not able to fetch
5110 * such data and will stay in TCP_URG_NOTYET until will be eaten
5111 * by recvmsg(). Seems, we are not obliged to handle such wicked
5112 * situations. But it is worth to think about possibility of some
5113 * DoSes using some hypothetical application level deadlock.
5115 if (before(ptr, tp->rcv_nxt))
5118 /* Do we already have a newer (or duplicate) urgent pointer? */
5119 if (tp->urg_data && !after(ptr, tp->urg_seq))
5122 /* Tell the world about our new urgent pointer. */
5125 /* We may be adding urgent data when the last byte read was
5126 * urgent. To do this requires some care. We cannot just ignore
5127 * tp->copied_seq since we would read the last urgent byte again
5128 * as data, nor can we alter copied_seq until this data arrives
5129 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5131 * NOTE. Double Dutch. Rendering to plain English: author of comment
5132 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5133 * and expect that both A and B disappear from stream. This is _wrong_.
5134 * Though this happens in BSD with high probability, this is occasional.
5135 * Any application relying on this is buggy. Note also, that fix "works"
5136 * only in this artificial test. Insert some normal data between A and B and we will
5137 * decline of BSD again. Verdict: it is better to remove to trap
5140 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5141 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5142 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5144 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5145 __skb_unlink(skb, &sk->sk_receive_queue);
5150 tp->urg_data = TCP_URG_NOTYET;
5153 /* Disable header prediction. */
5157 /* This is the 'fast' part of urgent handling. */
5158 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5160 struct tcp_sock *tp = tcp_sk(sk);
5162 /* Check if we get a new urgent pointer - normally not. */
5164 tcp_check_urg(sk, th);
5166 /* Do we wait for any urgent data? - normally not... */
5167 if (tp->urg_data == TCP_URG_NOTYET) {
5168 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5171 /* Is the urgent pointer pointing into this packet? */
5172 if (ptr < skb->len) {
5174 if (skb_copy_bits(skb, ptr, &tmp, 1))
5176 tp->urg_data = TCP_URG_VALID | tmp;
5177 if (!sock_flag(sk, SOCK_DEAD))
5178 sk->sk_data_ready(sk, 0);
5183 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5185 struct tcp_sock *tp = tcp_sk(sk);
5186 int chunk = skb->len - hlen;
5190 if (skb_csum_unnecessary(skb))
5191 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5193 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5197 tp->ucopy.len -= chunk;
5198 tp->copied_seq += chunk;
5199 tcp_rcv_space_adjust(sk);
5206 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5207 struct sk_buff *skb)
5211 if (sock_owned_by_user(sk)) {
5213 result = __tcp_checksum_complete(skb);
5216 result = __tcp_checksum_complete(skb);
5221 static inline bool tcp_checksum_complete_user(struct sock *sk,
5222 struct sk_buff *skb)
5224 return !skb_csum_unnecessary(skb) &&
5225 __tcp_checksum_complete_user(sk, skb);
5228 #ifdef CONFIG_NET_DMA
5229 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5232 struct tcp_sock *tp = tcp_sk(sk);
5233 int chunk = skb->len - hlen;
5235 bool copied_early = false;
5237 if (tp->ucopy.wakeup)
5240 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5241 tp->ucopy.dma_chan = net_dma_find_channel();
5243 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5245 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5247 tp->ucopy.iov, chunk,
5248 tp->ucopy.pinned_list);
5253 tp->ucopy.dma_cookie = dma_cookie;
5254 copied_early = true;
5256 tp->ucopy.len -= chunk;
5257 tp->copied_seq += chunk;
5258 tcp_rcv_space_adjust(sk);
5260 if ((tp->ucopy.len == 0) ||
5261 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5262 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5263 tp->ucopy.wakeup = 1;
5264 sk->sk_data_ready(sk, 0);
5266 } else if (chunk > 0) {
5267 tp->ucopy.wakeup = 1;
5268 sk->sk_data_ready(sk, 0);
5271 return copied_early;
5273 #endif /* CONFIG_NET_DMA */
5275 static void tcp_send_challenge_ack(struct sock *sk)
5277 /* unprotected vars, we dont care of overwrites */
5278 static u32 challenge_timestamp;
5279 static unsigned int challenge_count;
5280 u32 now = jiffies / HZ;
5282 if (now != challenge_timestamp) {
5283 challenge_timestamp = now;
5284 challenge_count = 0;
5286 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
5287 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
5292 /* Does PAWS and seqno based validation of an incoming segment, flags will
5293 * play significant role here.
5295 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5296 const struct tcphdr *th, int syn_inerr)
5298 const u8 *hash_location;
5299 struct tcp_sock *tp = tcp_sk(sk);
5301 /* RFC1323: H1. Apply PAWS check first. */
5302 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5303 tp->rx_opt.saw_tstamp &&
5304 tcp_paws_discard(sk, skb)) {
5306 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5307 tcp_send_dupack(sk, skb);
5310 /* Reset is accepted even if it did not pass PAWS. */
5313 /* Step 1: check sequence number */
5314 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5315 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5316 * (RST) segments are validated by checking their SEQ-fields."
5317 * And page 69: "If an incoming segment is not acceptable,
5318 * an acknowledgment should be sent in reply (unless the RST
5319 * bit is set, if so drop the segment and return)".
5324 tcp_send_dupack(sk, skb);
5329 /* Step 2: check RST bit */
5332 * If sequence number exactly matches RCV.NXT, then
5333 * RESET the connection
5335 * Send a challenge ACK
5337 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5340 tcp_send_challenge_ack(sk);
5344 /* ts_recent update must be made after we are sure that the packet
5347 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5349 /* step 3: check security and precedence [ignored] */
5351 /* step 4: Check for a SYN
5352 * RFC 5691 4.2 : Send a challenge ack
5357 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5358 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5359 tcp_send_challenge_ack(sk);
5371 * TCP receive function for the ESTABLISHED state.
5373 * It is split into a fast path and a slow path. The fast path is
5375 * - A zero window was announced from us - zero window probing
5376 * is only handled properly in the slow path.
5377 * - Out of order segments arrived.
5378 * - Urgent data is expected.
5379 * - There is no buffer space left
5380 * - Unexpected TCP flags/window values/header lengths are received
5381 * (detected by checking the TCP header against pred_flags)
5382 * - Data is sent in both directions. Fast path only supports pure senders
5383 * or pure receivers (this means either the sequence number or the ack
5384 * value must stay constant)
5385 * - Unexpected TCP option.
5387 * When these conditions are not satisfied it drops into a standard
5388 * receive procedure patterned after RFC793 to handle all cases.
5389 * The first three cases are guaranteed by proper pred_flags setting,
5390 * the rest is checked inline. Fast processing is turned on in
5391 * tcp_data_queue when everything is OK.
5393 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5394 const struct tcphdr *th, unsigned int len)
5396 struct tcp_sock *tp = tcp_sk(sk);
5398 if (unlikely(sk->sk_rx_dst == NULL))
5399 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5401 * Header prediction.
5402 * The code loosely follows the one in the famous
5403 * "30 instruction TCP receive" Van Jacobson mail.
5405 * Van's trick is to deposit buffers into socket queue
5406 * on a device interrupt, to call tcp_recv function
5407 * on the receive process context and checksum and copy
5408 * the buffer to user space. smart...
5410 * Our current scheme is not silly either but we take the
5411 * extra cost of the net_bh soft interrupt processing...
5412 * We do checksum and copy also but from device to kernel.
5415 tp->rx_opt.saw_tstamp = 0;
5417 /* pred_flags is 0xS?10 << 16 + snd_wnd
5418 * if header_prediction is to be made
5419 * 'S' will always be tp->tcp_header_len >> 2
5420 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5421 * turn it off (when there are holes in the receive
5422 * space for instance)
5423 * PSH flag is ignored.
5426 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5427 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5428 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5429 int tcp_header_len = tp->tcp_header_len;
5431 /* Timestamp header prediction: tcp_header_len
5432 * is automatically equal to th->doff*4 due to pred_flags
5436 /* Check timestamp */
5437 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5438 /* No? Slow path! */
5439 if (!tcp_parse_aligned_timestamp(tp, th))
5442 /* If PAWS failed, check it more carefully in slow path */
5443 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5446 /* DO NOT update ts_recent here, if checksum fails
5447 * and timestamp was corrupted part, it will result
5448 * in a hung connection since we will drop all
5449 * future packets due to the PAWS test.
5453 if (len <= tcp_header_len) {
5454 /* Bulk data transfer: sender */
5455 if (len == tcp_header_len) {
5456 /* Predicted packet is in window by definition.
5457 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5458 * Hence, check seq<=rcv_wup reduces to:
5460 if (tcp_header_len ==
5461 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5462 tp->rcv_nxt == tp->rcv_wup)
5463 tcp_store_ts_recent(tp);
5465 /* We know that such packets are checksummed
5468 tcp_ack(sk, skb, 0);
5470 tcp_data_snd_check(sk);
5472 } else { /* Header too small */
5473 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5478 int copied_early = 0;
5479 bool fragstolen = false;
5481 if (tp->copied_seq == tp->rcv_nxt &&
5482 len - tcp_header_len <= tp->ucopy.len) {
5483 #ifdef CONFIG_NET_DMA
5484 if (tp->ucopy.task == current &&
5485 sock_owned_by_user(sk) &&
5486 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5491 if (tp->ucopy.task == current &&
5492 sock_owned_by_user(sk) && !copied_early) {
5493 __set_current_state(TASK_RUNNING);
5495 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5499 /* Predicted packet is in window by definition.
5500 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5501 * Hence, check seq<=rcv_wup reduces to:
5503 if (tcp_header_len ==
5504 (sizeof(struct tcphdr) +
5505 TCPOLEN_TSTAMP_ALIGNED) &&
5506 tp->rcv_nxt == tp->rcv_wup)
5507 tcp_store_ts_recent(tp);
5509 tcp_rcv_rtt_measure_ts(sk, skb);
5511 __skb_pull(skb, tcp_header_len);
5512 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5513 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5516 tcp_cleanup_rbuf(sk, skb->len);
5519 if (tcp_checksum_complete_user(sk, skb))
5522 /* Predicted packet is in window by definition.
5523 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5524 * Hence, check seq<=rcv_wup reduces to:
5526 if (tcp_header_len ==
5527 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5528 tp->rcv_nxt == tp->rcv_wup)
5529 tcp_store_ts_recent(tp);
5531 tcp_rcv_rtt_measure_ts(sk, skb);
5533 if ((int)skb->truesize > sk->sk_forward_alloc)
5536 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5538 /* Bulk data transfer: receiver */
5539 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5543 tcp_event_data_recv(sk, skb);
5545 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5546 /* Well, only one small jumplet in fast path... */
5547 tcp_ack(sk, skb, FLAG_DATA);
5548 tcp_data_snd_check(sk);
5549 if (!inet_csk_ack_scheduled(sk))
5553 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5554 __tcp_ack_snd_check(sk, 0);
5556 #ifdef CONFIG_NET_DMA
5558 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5562 kfree_skb_partial(skb, fragstolen);
5564 sk->sk_data_ready(sk, 0);
5570 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5574 * Standard slow path.
5577 if (!tcp_validate_incoming(sk, skb, th, 1))
5581 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5584 tcp_rcv_rtt_measure_ts(sk, skb);
5586 /* Process urgent data. */
5587 tcp_urg(sk, skb, th);
5589 /* step 7: process the segment text */
5590 tcp_data_queue(sk, skb);
5592 tcp_data_snd_check(sk);
5593 tcp_ack_snd_check(sk);
5597 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5603 EXPORT_SYMBOL(tcp_rcv_established);
5605 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5607 struct tcp_sock *tp = tcp_sk(sk);
5608 struct inet_connection_sock *icsk = inet_csk(sk);
5610 tcp_set_state(sk, TCP_ESTABLISHED);
5613 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5614 security_inet_conn_established(sk, skb);
5617 /* Make sure socket is routed, for correct metrics. */
5618 icsk->icsk_af_ops->rebuild_header(sk);
5620 tcp_init_metrics(sk);
5622 tcp_init_congestion_control(sk);
5624 /* Prevent spurious tcp_cwnd_restart() on first data
5627 tp->lsndtime = tcp_time_stamp;
5629 tcp_init_buffer_space(sk);
5631 if (sock_flag(sk, SOCK_KEEPOPEN))
5632 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5634 if (!tp->rx_opt.snd_wscale)
5635 __tcp_fast_path_on(tp, tp->snd_wnd);
5639 if (!sock_flag(sk, SOCK_DEAD)) {
5640 sk->sk_state_change(sk);
5641 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5645 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5646 struct tcp_fastopen_cookie *cookie)
5648 struct tcp_sock *tp = tcp_sk(sk);
5649 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5650 u16 mss = tp->rx_opt.mss_clamp;
5653 if (mss == tp->rx_opt.user_mss) {
5654 struct tcp_options_received opt;
5655 const u8 *hash_location;
5657 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5658 tcp_clear_options(&opt);
5659 opt.user_mss = opt.mss_clamp = 0;
5660 tcp_parse_options(synack, &opt, &hash_location, 0, NULL);
5661 mss = opt.mss_clamp;
5664 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5667 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5668 * the remote receives only the retransmitted (regular) SYNs: either
5669 * the original SYN-data or the corresponding SYN-ACK is lost.
5671 syn_drop = (cookie->len <= 0 && data &&
5672 inet_csk(sk)->icsk_retransmits);
5674 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5676 if (data) { /* Retransmit unacked data in SYN */
5677 tcp_retransmit_skb(sk, data);
5684 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5685 const struct tcphdr *th, unsigned int len)
5687 const u8 *hash_location;
5688 struct inet_connection_sock *icsk = inet_csk(sk);
5689 struct tcp_sock *tp = tcp_sk(sk);
5690 struct tcp_cookie_values *cvp = tp->cookie_values;
5691 struct tcp_fastopen_cookie foc = { .len = -1 };
5692 int saved_clamp = tp->rx_opt.mss_clamp;
5694 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0, &foc);
5698 * "If the state is SYN-SENT then
5699 * first check the ACK bit
5700 * If the ACK bit is set
5701 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5702 * a reset (unless the RST bit is set, if so drop
5703 * the segment and return)"
5705 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5706 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5707 goto reset_and_undo;
5709 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5710 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5712 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5713 goto reset_and_undo;
5716 /* Now ACK is acceptable.
5718 * "If the RST bit is set
5719 * If the ACK was acceptable then signal the user "error:
5720 * connection reset", drop the segment, enter CLOSED state,
5721 * delete TCB, and return."
5730 * "fifth, if neither of the SYN or RST bits is set then
5731 * drop the segment and return."
5737 goto discard_and_undo;
5740 * "If the SYN bit is on ...
5741 * are acceptable then ...
5742 * (our SYN has been ACKed), change the connection
5743 * state to ESTABLISHED..."
5746 TCP_ECN_rcv_synack(tp, th);
5748 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5749 tcp_ack(sk, skb, FLAG_SLOWPATH);
5751 /* Ok.. it's good. Set up sequence numbers and
5752 * move to established.
5754 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5755 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5757 /* RFC1323: The window in SYN & SYN/ACK segments is
5760 tp->snd_wnd = ntohs(th->window);
5762 if (!tp->rx_opt.wscale_ok) {
5763 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5764 tp->window_clamp = min(tp->window_clamp, 65535U);
5767 if (tp->rx_opt.saw_tstamp) {
5768 tp->rx_opt.tstamp_ok = 1;
5769 tp->tcp_header_len =
5770 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5771 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5772 tcp_store_ts_recent(tp);
5774 tp->tcp_header_len = sizeof(struct tcphdr);
5777 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5778 tcp_enable_fack(tp);
5781 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5782 tcp_initialize_rcv_mss(sk);
5784 /* Remember, tcp_poll() does not lock socket!
5785 * Change state from SYN-SENT only after copied_seq
5786 * is initialized. */
5787 tp->copied_seq = tp->rcv_nxt;
5790 cvp->cookie_pair_size > 0 &&
5791 tp->rx_opt.cookie_plus > 0) {
5792 int cookie_size = tp->rx_opt.cookie_plus
5793 - TCPOLEN_COOKIE_BASE;
5794 int cookie_pair_size = cookie_size
5795 + cvp->cookie_desired;
5797 /* A cookie extension option was sent and returned.
5798 * Note that each incoming SYNACK replaces the
5799 * Responder cookie. The initial exchange is most
5800 * fragile, as protection against spoofing relies
5801 * entirely upon the sequence and timestamp (above).
5802 * This replacement strategy allows the correct pair to
5803 * pass through, while any others will be filtered via
5804 * Responder verification later.
5806 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5807 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5808 hash_location, cookie_size);
5809 cvp->cookie_pair_size = cookie_pair_size;
5815 tcp_finish_connect(sk, skb);
5817 if ((tp->syn_fastopen || tp->syn_data) &&
5818 tcp_rcv_fastopen_synack(sk, skb, &foc))
5821 if (sk->sk_write_pending ||
5822 icsk->icsk_accept_queue.rskq_defer_accept ||
5823 icsk->icsk_ack.pingpong) {
5824 /* Save one ACK. Data will be ready after
5825 * several ticks, if write_pending is set.
5827 * It may be deleted, but with this feature tcpdumps
5828 * look so _wonderfully_ clever, that I was not able
5829 * to stand against the temptation 8) --ANK
5831 inet_csk_schedule_ack(sk);
5832 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5833 tcp_enter_quickack_mode(sk);
5834 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5835 TCP_DELACK_MAX, TCP_RTO_MAX);
5846 /* No ACK in the segment */
5850 * "If the RST bit is set
5852 * Otherwise (no ACK) drop the segment and return."
5855 goto discard_and_undo;
5859 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5860 tcp_paws_reject(&tp->rx_opt, 0))
5861 goto discard_and_undo;
5864 /* We see SYN without ACK. It is attempt of
5865 * simultaneous connect with crossed SYNs.
5866 * Particularly, it can be connect to self.
5868 tcp_set_state(sk, TCP_SYN_RECV);
5870 if (tp->rx_opt.saw_tstamp) {
5871 tp->rx_opt.tstamp_ok = 1;
5872 tcp_store_ts_recent(tp);
5873 tp->tcp_header_len =
5874 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5876 tp->tcp_header_len = sizeof(struct tcphdr);
5879 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5880 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5882 /* RFC1323: The window in SYN & SYN/ACK segments is
5885 tp->snd_wnd = ntohs(th->window);
5886 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5887 tp->max_window = tp->snd_wnd;
5889 TCP_ECN_rcv_syn(tp, th);
5892 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5893 tcp_initialize_rcv_mss(sk);
5895 tcp_send_synack(sk);
5897 /* Note, we could accept data and URG from this segment.
5898 * There are no obstacles to make this.
5900 * However, if we ignore data in ACKless segments sometimes,
5901 * we have no reasons to accept it sometimes.
5902 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5903 * is not flawless. So, discard packet for sanity.
5904 * Uncomment this return to process the data.
5911 /* "fifth, if neither of the SYN or RST bits is set then
5912 * drop the segment and return."
5916 tcp_clear_options(&tp->rx_opt);
5917 tp->rx_opt.mss_clamp = saved_clamp;
5921 tcp_clear_options(&tp->rx_opt);
5922 tp->rx_opt.mss_clamp = saved_clamp;
5927 * This function implements the receiving procedure of RFC 793 for
5928 * all states except ESTABLISHED and TIME_WAIT.
5929 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5930 * address independent.
5933 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5934 const struct tcphdr *th, unsigned int len)
5936 struct tcp_sock *tp = tcp_sk(sk);
5937 struct inet_connection_sock *icsk = inet_csk(sk);
5940 tp->rx_opt.saw_tstamp = 0;
5942 switch (sk->sk_state) {
5956 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5959 /* Now we have several options: In theory there is
5960 * nothing else in the frame. KA9Q has an option to
5961 * send data with the syn, BSD accepts data with the
5962 * syn up to the [to be] advertised window and
5963 * Solaris 2.1 gives you a protocol error. For now
5964 * we just ignore it, that fits the spec precisely
5965 * and avoids incompatibilities. It would be nice in
5966 * future to drop through and process the data.
5968 * Now that TTCP is starting to be used we ought to
5970 * But, this leaves one open to an easy denial of
5971 * service attack, and SYN cookies can't defend
5972 * against this problem. So, we drop the data
5973 * in the interest of security over speed unless
5974 * it's still in use.
5982 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5986 /* Do step6 onward by hand. */
5987 tcp_urg(sk, skb, th);
5989 tcp_data_snd_check(sk);
5993 if (!tcp_validate_incoming(sk, skb, th, 0))
5996 /* step 5: check the ACK field */
5998 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
6000 switch (sk->sk_state) {
6003 tp->copied_seq = tp->rcv_nxt;
6005 tcp_set_state(sk, TCP_ESTABLISHED);
6006 sk->sk_state_change(sk);
6008 /* Note, that this wakeup is only for marginal
6009 * crossed SYN case. Passively open sockets
6010 * are not waked up, because sk->sk_sleep ==
6011 * NULL and sk->sk_socket == NULL.
6015 SOCK_WAKE_IO, POLL_OUT);
6017 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6018 tp->snd_wnd = ntohs(th->window) <<
6019 tp->rx_opt.snd_wscale;
6020 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6022 if (tp->rx_opt.tstamp_ok)
6023 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6025 /* Make sure socket is routed, for
6028 icsk->icsk_af_ops->rebuild_header(sk);
6030 tcp_init_metrics(sk);
6032 tcp_init_congestion_control(sk);
6034 /* Prevent spurious tcp_cwnd_restart() on
6035 * first data packet.
6037 tp->lsndtime = tcp_time_stamp;
6040 tcp_initialize_rcv_mss(sk);
6041 tcp_init_buffer_space(sk);
6042 tcp_fast_path_on(tp);
6049 if (tp->snd_una == tp->write_seq) {
6050 struct dst_entry *dst;
6052 tcp_set_state(sk, TCP_FIN_WAIT2);
6053 sk->sk_shutdown |= SEND_SHUTDOWN;
6055 dst = __sk_dst_get(sk);
6059 if (!sock_flag(sk, SOCK_DEAD))
6060 /* Wake up lingering close() */
6061 sk->sk_state_change(sk);
6065 if (tp->linger2 < 0 ||
6066 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6067 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6069 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6073 tmo = tcp_fin_time(sk);
6074 if (tmo > TCP_TIMEWAIT_LEN) {
6075 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6076 } else if (th->fin || sock_owned_by_user(sk)) {
6077 /* Bad case. We could lose such FIN otherwise.
6078 * It is not a big problem, but it looks confusing
6079 * and not so rare event. We still can lose it now,
6080 * if it spins in bh_lock_sock(), but it is really
6083 inet_csk_reset_keepalive_timer(sk, tmo);
6085 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6093 if (tp->snd_una == tp->write_seq) {
6094 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6100 if (tp->snd_una == tp->write_seq) {
6101 tcp_update_metrics(sk);
6110 /* step 6: check the URG bit */
6111 tcp_urg(sk, skb, th);
6113 /* step 7: process the segment text */
6114 switch (sk->sk_state) {
6115 case TCP_CLOSE_WAIT:
6118 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6122 /* RFC 793 says to queue data in these states,
6123 * RFC 1122 says we MUST send a reset.
6124 * BSD 4.4 also does reset.
6126 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6127 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6128 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6129 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6135 case TCP_ESTABLISHED:
6136 tcp_data_queue(sk, skb);
6141 /* tcp_data could move socket to TIME-WAIT */
6142 if (sk->sk_state != TCP_CLOSE) {
6143 tcp_data_snd_check(sk);
6144 tcp_ack_snd_check(sk);
6153 EXPORT_SYMBOL(tcp_rcv_state_process);