2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 100;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
97 int sysctl_tcp_thin_dupack __read_mostly;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100 int sysctl_tcp_early_retrans __read_mostly = 3;
101 int sysctl_tcp_default_init_rwnd __read_mostly = TCP_DEFAULT_INIT_RCVWND;
103 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
104 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
105 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
106 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
107 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
108 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
109 #define FLAG_ECE 0x40 /* ECE in this ACK */
110 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
112 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
115 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
130 struct inet_connection_sock *icsk = inet_csk(sk);
131 const unsigned int lss = icsk->icsk_ack.last_seg_size;
134 icsk->icsk_ack.last_seg_size = 0;
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
139 len = skb_shinfo(skb)->gso_size ? : skb->len;
140 if (len >= icsk->icsk_ack.rcv_mss) {
141 icsk->icsk_ack.rcv_mss = len;
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
146 * "len" is invariant segment length, including TCP header.
148 len += skb->data - skb_transport_header(skb);
149 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
155 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
156 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
161 len -= tcp_sk(sk)->tcp_header_len;
162 icsk->icsk_ack.last_seg_size = len;
164 icsk->icsk_ack.rcv_mss = len;
168 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
174 static void tcp_incr_quickack(struct sock *sk)
176 struct inet_connection_sock *icsk = inet_csk(sk);
177 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
181 if (quickacks > icsk->icsk_ack.quick)
182 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185 static void tcp_enter_quickack_mode(struct sock *sk)
187 struct inet_connection_sock *icsk = inet_csk(sk);
188 tcp_incr_quickack(sk);
189 icsk->icsk_ack.pingpong = 0;
190 icsk->icsk_ack.ato = TCP_ATO_MIN;
193 /* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
197 static inline bool tcp_in_quickack_mode(const struct sock *sk)
199 const struct inet_connection_sock *icsk = inet_csk(sk);
201 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
204 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
206 if (tp->ecn_flags & TCP_ECN_OK)
207 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
210 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
212 if (tcp_hdr(skb)->cwr)
213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
216 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
218 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
221 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
223 if (!(tp->ecn_flags & TCP_ECN_OK))
226 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
227 case INET_ECN_NOT_ECT:
228 /* Funny extension: if ECT is not set on a segment,
229 * and we already seen ECT on a previous segment,
230 * it is probably a retransmit.
232 if (tp->ecn_flags & TCP_ECN_SEEN)
233 tcp_enter_quickack_mode((struct sock *)tp);
236 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
237 /* Better not delay acks, sender can have a very low cwnd */
238 tcp_enter_quickack_mode((struct sock *)tp);
239 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
243 tp->ecn_flags |= TCP_ECN_SEEN;
247 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
249 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
250 tp->ecn_flags &= ~TCP_ECN_OK;
253 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
255 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
256 tp->ecn_flags &= ~TCP_ECN_OK;
259 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
261 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
266 /* Buffer size and advertised window tuning.
268 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
271 static void tcp_fixup_sndbuf(struct sock *sk)
273 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
275 sndmem *= TCP_INIT_CWND;
276 if (sk->sk_sndbuf < sndmem)
277 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
280 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
282 * All tcp_full_space() is split to two parts: "network" buffer, allocated
283 * forward and advertised in receiver window (tp->rcv_wnd) and
284 * "application buffer", required to isolate scheduling/application
285 * latencies from network.
286 * window_clamp is maximal advertised window. It can be less than
287 * tcp_full_space(), in this case tcp_full_space() - window_clamp
288 * is reserved for "application" buffer. The less window_clamp is
289 * the smoother our behaviour from viewpoint of network, but the lower
290 * throughput and the higher sensitivity of the connection to losses. 8)
292 * rcv_ssthresh is more strict window_clamp used at "slow start"
293 * phase to predict further behaviour of this connection.
294 * It is used for two goals:
295 * - to enforce header prediction at sender, even when application
296 * requires some significant "application buffer". It is check #1.
297 * - to prevent pruning of receive queue because of misprediction
298 * of receiver window. Check #2.
300 * The scheme does not work when sender sends good segments opening
301 * window and then starts to feed us spaghetti. But it should work
302 * in common situations. Otherwise, we have to rely on queue collapsing.
305 /* Slow part of check#2. */
306 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
308 struct tcp_sock *tp = tcp_sk(sk);
310 int truesize = tcp_win_from_space(skb->truesize) >> 1;
311 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
313 while (tp->rcv_ssthresh <= window) {
314 if (truesize <= skb->len)
315 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
323 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
325 struct tcp_sock *tp = tcp_sk(sk);
328 if (tp->rcv_ssthresh < tp->window_clamp &&
329 (int)tp->rcv_ssthresh < tcp_space(sk) &&
330 !sk_under_memory_pressure(sk)) {
333 /* Check #2. Increase window, if skb with such overhead
334 * will fit to rcvbuf in future.
336 if (tcp_win_from_space(skb->truesize) <= skb->len)
337 incr = 2 * tp->advmss;
339 incr = __tcp_grow_window(sk, skb);
342 incr = max_t(int, incr, 2 * skb->len);
343 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
345 inet_csk(sk)->icsk_ack.quick |= 1;
350 /* 3. Tuning rcvbuf, when connection enters established state. */
352 static void tcp_fixup_rcvbuf(struct sock *sk)
354 u32 mss = tcp_sk(sk)->advmss;
355 u32 icwnd = sysctl_tcp_default_init_rwnd;
358 /* Limit to 10 segments if mss <= 1460,
359 * or 14600/mss segments, with a minimum of two segments.
362 icwnd = max_t(u32, (1460 * icwnd) / mss, 2);
364 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
365 while (tcp_win_from_space(rcvmem) < mss)
370 if (sk->sk_rcvbuf < rcvmem)
371 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
374 /* 4. Try to fixup all. It is made immediately after connection enters
377 void tcp_init_buffer_space(struct sock *sk)
379 struct tcp_sock *tp = tcp_sk(sk);
382 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
383 tcp_fixup_rcvbuf(sk);
384 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
385 tcp_fixup_sndbuf(sk);
387 tp->rcvq_space.space = tp->rcv_wnd;
389 maxwin = tcp_full_space(sk);
391 if (tp->window_clamp >= maxwin) {
392 tp->window_clamp = maxwin;
394 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
395 tp->window_clamp = max(maxwin -
396 (maxwin >> sysctl_tcp_app_win),
400 /* Force reservation of one segment. */
401 if (sysctl_tcp_app_win &&
402 tp->window_clamp > 2 * tp->advmss &&
403 tp->window_clamp + tp->advmss > maxwin)
404 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
406 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
407 tp->snd_cwnd_stamp = tcp_time_stamp;
410 /* 5. Recalculate window clamp after socket hit its memory bounds. */
411 static void tcp_clamp_window(struct sock *sk)
413 struct tcp_sock *tp = tcp_sk(sk);
414 struct inet_connection_sock *icsk = inet_csk(sk);
416 icsk->icsk_ack.quick = 0;
418 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
419 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
420 !sk_under_memory_pressure(sk) &&
421 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
422 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
425 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
426 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
429 /* Initialize RCV_MSS value.
430 * RCV_MSS is an our guess about MSS used by the peer.
431 * We haven't any direct information about the MSS.
432 * It's better to underestimate the RCV_MSS rather than overestimate.
433 * Overestimations make us ACKing less frequently than needed.
434 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
436 void tcp_initialize_rcv_mss(struct sock *sk)
438 const struct tcp_sock *tp = tcp_sk(sk);
439 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
441 hint = min(hint, tp->rcv_wnd / 2);
442 hint = min(hint, TCP_MSS_DEFAULT);
443 hint = max(hint, TCP_MIN_MSS);
445 inet_csk(sk)->icsk_ack.rcv_mss = hint;
447 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
449 /* Receiver "autotuning" code.
451 * The algorithm for RTT estimation w/o timestamps is based on
452 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
453 * <http://public.lanl.gov/radiant/pubs.html#DRS>
455 * More detail on this code can be found at
456 * <http://staff.psc.edu/jheffner/>,
457 * though this reference is out of date. A new paper
460 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
462 u32 new_sample = tp->rcv_rtt_est.rtt;
468 if (new_sample != 0) {
469 /* If we sample in larger samples in the non-timestamp
470 * case, we could grossly overestimate the RTT especially
471 * with chatty applications or bulk transfer apps which
472 * are stalled on filesystem I/O.
474 * Also, since we are only going for a minimum in the
475 * non-timestamp case, we do not smooth things out
476 * else with timestamps disabled convergence takes too
480 m -= (new_sample >> 3);
488 /* No previous measure. */
492 if (tp->rcv_rtt_est.rtt != new_sample)
493 tp->rcv_rtt_est.rtt = new_sample;
496 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
498 if (tp->rcv_rtt_est.time == 0)
500 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
502 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
505 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
506 tp->rcv_rtt_est.time = tcp_time_stamp;
509 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
510 const struct sk_buff *skb)
512 struct tcp_sock *tp = tcp_sk(sk);
513 if (tp->rx_opt.rcv_tsecr &&
514 (TCP_SKB_CB(skb)->end_seq -
515 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
516 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
520 * This function should be called every time data is copied to user space.
521 * It calculates the appropriate TCP receive buffer space.
523 void tcp_rcv_space_adjust(struct sock *sk)
525 struct tcp_sock *tp = tcp_sk(sk);
529 if (tp->rcvq_space.time == 0)
532 time = tcp_time_stamp - tp->rcvq_space.time;
533 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
536 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
538 space = max(tp->rcvq_space.space, space);
540 if (tp->rcvq_space.space != space) {
543 tp->rcvq_space.space = space;
545 if (sysctl_tcp_moderate_rcvbuf &&
546 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
547 int new_clamp = space;
549 /* Receive space grows, normalize in order to
550 * take into account packet headers and sk_buff
551 * structure overhead.
556 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
557 while (tcp_win_from_space(rcvmem) < tp->advmss)
560 space = min(space, sysctl_tcp_rmem[2]);
561 if (space > sk->sk_rcvbuf) {
562 sk->sk_rcvbuf = space;
564 /* Make the window clamp follow along. */
565 tp->window_clamp = new_clamp;
571 tp->rcvq_space.seq = tp->copied_seq;
572 tp->rcvq_space.time = tcp_time_stamp;
575 /* There is something which you must keep in mind when you analyze the
576 * behavior of the tp->ato delayed ack timeout interval. When a
577 * connection starts up, we want to ack as quickly as possible. The
578 * problem is that "good" TCP's do slow start at the beginning of data
579 * transmission. The means that until we send the first few ACK's the
580 * sender will sit on his end and only queue most of his data, because
581 * he can only send snd_cwnd unacked packets at any given time. For
582 * each ACK we send, he increments snd_cwnd and transmits more of his
585 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
587 struct tcp_sock *tp = tcp_sk(sk);
588 struct inet_connection_sock *icsk = inet_csk(sk);
591 inet_csk_schedule_ack(sk);
593 tcp_measure_rcv_mss(sk, skb);
595 tcp_rcv_rtt_measure(tp);
597 now = tcp_time_stamp;
599 if (!icsk->icsk_ack.ato) {
600 /* The _first_ data packet received, initialize
601 * delayed ACK engine.
603 tcp_incr_quickack(sk);
604 icsk->icsk_ack.ato = TCP_ATO_MIN;
606 int m = now - icsk->icsk_ack.lrcvtime;
608 if (m <= TCP_ATO_MIN / 2) {
609 /* The fastest case is the first. */
610 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
611 } else if (m < icsk->icsk_ack.ato) {
612 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
613 if (icsk->icsk_ack.ato > icsk->icsk_rto)
614 icsk->icsk_ack.ato = icsk->icsk_rto;
615 } else if (m > icsk->icsk_rto) {
616 /* Too long gap. Apparently sender failed to
617 * restart window, so that we send ACKs quickly.
619 tcp_incr_quickack(sk);
623 icsk->icsk_ack.lrcvtime = now;
625 TCP_ECN_check_ce(tp, skb);
628 tcp_grow_window(sk, skb);
631 /* Called to compute a smoothed rtt estimate. The data fed to this
632 * routine either comes from timestamps, or from segments that were
633 * known _not_ to have been retransmitted [see Karn/Partridge
634 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
635 * piece by Van Jacobson.
636 * NOTE: the next three routines used to be one big routine.
637 * To save cycles in the RFC 1323 implementation it was better to break
638 * it up into three procedures. -- erics
640 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
642 struct tcp_sock *tp = tcp_sk(sk);
643 long m = mrtt; /* RTT */
645 /* The following amusing code comes from Jacobson's
646 * article in SIGCOMM '88. Note that rtt and mdev
647 * are scaled versions of rtt and mean deviation.
648 * This is designed to be as fast as possible
649 * m stands for "measurement".
651 * On a 1990 paper the rto value is changed to:
652 * RTO = rtt + 4 * mdev
654 * Funny. This algorithm seems to be very broken.
655 * These formulae increase RTO, when it should be decreased, increase
656 * too slowly, when it should be increased quickly, decrease too quickly
657 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
658 * does not matter how to _calculate_ it. Seems, it was trap
659 * that VJ failed to avoid. 8)
664 m -= (tp->srtt >> 3); /* m is now error in rtt est */
665 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
667 m = -m; /* m is now abs(error) */
668 m -= (tp->mdev >> 2); /* similar update on mdev */
669 /* This is similar to one of Eifel findings.
670 * Eifel blocks mdev updates when rtt decreases.
671 * This solution is a bit different: we use finer gain
672 * for mdev in this case (alpha*beta).
673 * Like Eifel it also prevents growth of rto,
674 * but also it limits too fast rto decreases,
675 * happening in pure Eifel.
680 m -= (tp->mdev >> 2); /* similar update on mdev */
682 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
683 if (tp->mdev > tp->mdev_max) {
684 tp->mdev_max = tp->mdev;
685 if (tp->mdev_max > tp->rttvar)
686 tp->rttvar = tp->mdev_max;
688 if (after(tp->snd_una, tp->rtt_seq)) {
689 if (tp->mdev_max < tp->rttvar)
690 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
691 tp->rtt_seq = tp->snd_nxt;
692 tp->mdev_max = tcp_rto_min(sk);
695 /* no previous measure. */
696 tp->srtt = m << 3; /* take the measured time to be rtt */
697 tp->mdev = m << 1; /* make sure rto = 3*rtt */
698 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
699 tp->rtt_seq = tp->snd_nxt;
703 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
704 * Note: TCP stack does not yet implement pacing.
705 * FQ packet scheduler can be used to implement cheap but effective
706 * TCP pacing, to smooth the burst on large writes when packets
707 * in flight is significantly lower than cwnd (or rwin)
709 static void tcp_update_pacing_rate(struct sock *sk)
711 const struct tcp_sock *tp = tcp_sk(sk);
714 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
715 rate = (u64)tp->mss_cache * 2 * (HZ << 3);
717 rate *= max(tp->snd_cwnd, tp->packets_out);
719 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
720 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
721 * We probably need usec resolution in the future.
722 * Note: This also takes care of possible srtt=0 case,
723 * when tcp_rtt_estimator() was not yet called.
725 if (tp->srtt > 8 + 2)
726 do_div(rate, tp->srtt);
728 sk->sk_pacing_rate = min_t(u64, rate, ~0U);
731 /* Calculate rto without backoff. This is the second half of Van Jacobson's
732 * routine referred to above.
734 void tcp_set_rto(struct sock *sk)
736 const struct tcp_sock *tp = tcp_sk(sk);
737 /* Old crap is replaced with new one. 8)
740 * 1. If rtt variance happened to be less 50msec, it is hallucination.
741 * It cannot be less due to utterly erratic ACK generation made
742 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
743 * to do with delayed acks, because at cwnd>2 true delack timeout
744 * is invisible. Actually, Linux-2.4 also generates erratic
745 * ACKs in some circumstances.
747 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
749 /* 2. Fixups made earlier cannot be right.
750 * If we do not estimate RTO correctly without them,
751 * all the algo is pure shit and should be replaced
752 * with correct one. It is exactly, which we pretend to do.
755 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
756 * guarantees that rto is higher.
761 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
763 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
766 cwnd = TCP_INIT_CWND;
767 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
771 * Packet counting of FACK is based on in-order assumptions, therefore TCP
772 * disables it when reordering is detected
774 void tcp_disable_fack(struct tcp_sock *tp)
776 /* RFC3517 uses different metric in lost marker => reset on change */
778 tp->lost_skb_hint = NULL;
779 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
782 /* Take a notice that peer is sending D-SACKs */
783 static void tcp_dsack_seen(struct tcp_sock *tp)
785 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
788 static void tcp_update_reordering(struct sock *sk, const int metric,
791 struct tcp_sock *tp = tcp_sk(sk);
792 if (metric > tp->reordering) {
795 tp->reordering = min(TCP_MAX_REORDERING, metric);
797 /* This exciting event is worth to be remembered. 8) */
799 mib_idx = LINUX_MIB_TCPTSREORDER;
800 else if (tcp_is_reno(tp))
801 mib_idx = LINUX_MIB_TCPRENOREORDER;
802 else if (tcp_is_fack(tp))
803 mib_idx = LINUX_MIB_TCPFACKREORDER;
805 mib_idx = LINUX_MIB_TCPSACKREORDER;
807 NET_INC_STATS_BH(sock_net(sk), mib_idx);
808 #if FASTRETRANS_DEBUG > 1
809 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
810 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
814 tp->undo_marker ? tp->undo_retrans : 0);
816 tcp_disable_fack(tp);
820 tcp_disable_early_retrans(tp);
823 /* This must be called before lost_out is incremented */
824 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
826 if ((tp->retransmit_skb_hint == NULL) ||
827 before(TCP_SKB_CB(skb)->seq,
828 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
829 tp->retransmit_skb_hint = skb;
832 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
833 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
836 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
838 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
839 tcp_verify_retransmit_hint(tp, skb);
841 tp->lost_out += tcp_skb_pcount(skb);
842 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
846 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
849 tcp_verify_retransmit_hint(tp, skb);
851 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
852 tp->lost_out += tcp_skb_pcount(skb);
853 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
857 /* This procedure tags the retransmission queue when SACKs arrive.
859 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
860 * Packets in queue with these bits set are counted in variables
861 * sacked_out, retrans_out and lost_out, correspondingly.
863 * Valid combinations are:
864 * Tag InFlight Description
865 * 0 1 - orig segment is in flight.
866 * S 0 - nothing flies, orig reached receiver.
867 * L 0 - nothing flies, orig lost by net.
868 * R 2 - both orig and retransmit are in flight.
869 * L|R 1 - orig is lost, retransmit is in flight.
870 * S|R 1 - orig reached receiver, retrans is still in flight.
871 * (L|S|R is logically valid, it could occur when L|R is sacked,
872 * but it is equivalent to plain S and code short-curcuits it to S.
873 * L|S is logically invalid, it would mean -1 packet in flight 8))
875 * These 6 states form finite state machine, controlled by the following events:
876 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
877 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
878 * 3. Loss detection event of two flavors:
879 * A. Scoreboard estimator decided the packet is lost.
880 * A'. Reno "three dupacks" marks head of queue lost.
881 * A''. Its FACK modification, head until snd.fack is lost.
882 * B. SACK arrives sacking SND.NXT at the moment, when the
883 * segment was retransmitted.
884 * 4. D-SACK added new rule: D-SACK changes any tag to S.
886 * It is pleasant to note, that state diagram turns out to be commutative,
887 * so that we are allowed not to be bothered by order of our actions,
888 * when multiple events arrive simultaneously. (see the function below).
890 * Reordering detection.
891 * --------------------
892 * Reordering metric is maximal distance, which a packet can be displaced
893 * in packet stream. With SACKs we can estimate it:
895 * 1. SACK fills old hole and the corresponding segment was not
896 * ever retransmitted -> reordering. Alas, we cannot use it
897 * when segment was retransmitted.
898 * 2. The last flaw is solved with D-SACK. D-SACK arrives
899 * for retransmitted and already SACKed segment -> reordering..
900 * Both of these heuristics are not used in Loss state, when we cannot
901 * account for retransmits accurately.
903 * SACK block validation.
904 * ----------------------
906 * SACK block range validation checks that the received SACK block fits to
907 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
908 * Note that SND.UNA is not included to the range though being valid because
909 * it means that the receiver is rather inconsistent with itself reporting
910 * SACK reneging when it should advance SND.UNA. Such SACK block this is
911 * perfectly valid, however, in light of RFC2018 which explicitly states
912 * that "SACK block MUST reflect the newest segment. Even if the newest
913 * segment is going to be discarded ...", not that it looks very clever
914 * in case of head skb. Due to potentional receiver driven attacks, we
915 * choose to avoid immediate execution of a walk in write queue due to
916 * reneging and defer head skb's loss recovery to standard loss recovery
917 * procedure that will eventually trigger (nothing forbids us doing this).
919 * Implements also blockage to start_seq wrap-around. Problem lies in the
920 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
921 * there's no guarantee that it will be before snd_nxt (n). The problem
922 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
925 * <- outs wnd -> <- wrapzone ->
926 * u e n u_w e_w s n_w
928 * |<------------+------+----- TCP seqno space --------------+---------->|
929 * ...-- <2^31 ->| |<--------...
930 * ...---- >2^31 ------>| |<--------...
932 * Current code wouldn't be vulnerable but it's better still to discard such
933 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
934 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
935 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
936 * equal to the ideal case (infinite seqno space without wrap caused issues).
938 * With D-SACK the lower bound is extended to cover sequence space below
939 * SND.UNA down to undo_marker, which is the last point of interest. Yet
940 * again, D-SACK block must not to go across snd_una (for the same reason as
941 * for the normal SACK blocks, explained above). But there all simplicity
942 * ends, TCP might receive valid D-SACKs below that. As long as they reside
943 * fully below undo_marker they do not affect behavior in anyway and can
944 * therefore be safely ignored. In rare cases (which are more or less
945 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
946 * fragmentation and packet reordering past skb's retransmission. To consider
947 * them correctly, the acceptable range must be extended even more though
948 * the exact amount is rather hard to quantify. However, tp->max_window can
949 * be used as an exaggerated estimate.
951 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
952 u32 start_seq, u32 end_seq)
954 /* Too far in future, or reversed (interpretation is ambiguous) */
955 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
958 /* Nasty start_seq wrap-around check (see comments above) */
959 if (!before(start_seq, tp->snd_nxt))
962 /* In outstanding window? ...This is valid exit for D-SACKs too.
963 * start_seq == snd_una is non-sensical (see comments above)
965 if (after(start_seq, tp->snd_una))
968 if (!is_dsack || !tp->undo_marker)
971 /* ...Then it's D-SACK, and must reside below snd_una completely */
972 if (after(end_seq, tp->snd_una))
975 if (!before(start_seq, tp->undo_marker))
979 if (!after(end_seq, tp->undo_marker))
982 /* Undo_marker boundary crossing (overestimates a lot). Known already:
983 * start_seq < undo_marker and end_seq >= undo_marker.
985 return !before(start_seq, end_seq - tp->max_window);
988 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
989 * Event "B". Later note: FACK people cheated me again 8), we have to account
990 * for reordering! Ugly, but should help.
992 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
993 * less than what is now known to be received by the other end (derived from
994 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
995 * retransmitted skbs to avoid some costly processing per ACKs.
997 static void tcp_mark_lost_retrans(struct sock *sk)
999 const struct inet_connection_sock *icsk = inet_csk(sk);
1000 struct tcp_sock *tp = tcp_sk(sk);
1001 struct sk_buff *skb;
1003 u32 new_low_seq = tp->snd_nxt;
1004 u32 received_upto = tcp_highest_sack_seq(tp);
1006 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1007 !after(received_upto, tp->lost_retrans_low) ||
1008 icsk->icsk_ca_state != TCP_CA_Recovery)
1011 tcp_for_write_queue(skb, sk) {
1012 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1014 if (skb == tcp_send_head(sk))
1016 if (cnt == tp->retrans_out)
1018 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1021 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1024 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1025 * constraint here (see above) but figuring out that at
1026 * least tp->reordering SACK blocks reside between ack_seq
1027 * and received_upto is not easy task to do cheaply with
1028 * the available datastructures.
1030 * Whether FACK should check here for tp->reordering segs
1031 * in-between one could argue for either way (it would be
1032 * rather simple to implement as we could count fack_count
1033 * during the walk and do tp->fackets_out - fack_count).
1035 if (after(received_upto, ack_seq)) {
1036 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1037 tp->retrans_out -= tcp_skb_pcount(skb);
1039 tcp_skb_mark_lost_uncond_verify(tp, skb);
1040 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1042 if (before(ack_seq, new_low_seq))
1043 new_low_seq = ack_seq;
1044 cnt += tcp_skb_pcount(skb);
1048 if (tp->retrans_out)
1049 tp->lost_retrans_low = new_low_seq;
1052 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1053 struct tcp_sack_block_wire *sp, int num_sacks,
1056 struct tcp_sock *tp = tcp_sk(sk);
1057 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1058 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1059 bool dup_sack = false;
1061 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1064 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1065 } else if (num_sacks > 1) {
1066 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1067 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1069 if (!after(end_seq_0, end_seq_1) &&
1070 !before(start_seq_0, start_seq_1)) {
1073 NET_INC_STATS_BH(sock_net(sk),
1074 LINUX_MIB_TCPDSACKOFORECV);
1078 /* D-SACK for already forgotten data... Do dumb counting. */
1079 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1080 !after(end_seq_0, prior_snd_una) &&
1081 after(end_seq_0, tp->undo_marker))
1087 struct tcp_sacktag_state {
1093 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1094 * the incoming SACK may not exactly match but we can find smaller MSS
1095 * aligned portion of it that matches. Therefore we might need to fragment
1096 * which may fail and creates some hassle (caller must handle error case
1099 * FIXME: this could be merged to shift decision code
1101 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1102 u32 start_seq, u32 end_seq)
1106 unsigned int pkt_len;
1109 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1110 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1112 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1113 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1114 mss = tcp_skb_mss(skb);
1115 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1118 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1122 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1127 /* Round if necessary so that SACKs cover only full MSSes
1128 * and/or the remaining small portion (if present)
1130 if (pkt_len > mss) {
1131 unsigned int new_len = (pkt_len / mss) * mss;
1132 if (!in_sack && new_len < pkt_len) {
1134 if (new_len >= skb->len)
1139 err = tcp_fragment(sk, skb, pkt_len, mss);
1147 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1148 static u8 tcp_sacktag_one(struct sock *sk,
1149 struct tcp_sacktag_state *state, u8 sacked,
1150 u32 start_seq, u32 end_seq,
1151 bool dup_sack, int pcount)
1153 struct tcp_sock *tp = tcp_sk(sk);
1154 int fack_count = state->fack_count;
1156 /* Account D-SACK for retransmitted packet. */
1157 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1158 if (tp->undo_marker && tp->undo_retrans > 0 &&
1159 after(end_seq, tp->undo_marker))
1161 if (sacked & TCPCB_SACKED_ACKED)
1162 state->reord = min(fack_count, state->reord);
1165 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1166 if (!after(end_seq, tp->snd_una))
1169 if (!(sacked & TCPCB_SACKED_ACKED)) {
1170 if (sacked & TCPCB_SACKED_RETRANS) {
1171 /* If the segment is not tagged as lost,
1172 * we do not clear RETRANS, believing
1173 * that retransmission is still in flight.
1175 if (sacked & TCPCB_LOST) {
1176 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1177 tp->lost_out -= pcount;
1178 tp->retrans_out -= pcount;
1181 if (!(sacked & TCPCB_RETRANS)) {
1182 /* New sack for not retransmitted frame,
1183 * which was in hole. It is reordering.
1185 if (before(start_seq,
1186 tcp_highest_sack_seq(tp)))
1187 state->reord = min(fack_count,
1189 if (!after(end_seq, tp->high_seq))
1190 state->flag |= FLAG_ORIG_SACK_ACKED;
1193 if (sacked & TCPCB_LOST) {
1194 sacked &= ~TCPCB_LOST;
1195 tp->lost_out -= pcount;
1199 sacked |= TCPCB_SACKED_ACKED;
1200 state->flag |= FLAG_DATA_SACKED;
1201 tp->sacked_out += pcount;
1203 fack_count += pcount;
1205 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1206 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1207 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1208 tp->lost_cnt_hint += pcount;
1210 if (fack_count > tp->fackets_out)
1211 tp->fackets_out = fack_count;
1214 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1215 * frames and clear it. undo_retrans is decreased above, L|R frames
1216 * are accounted above as well.
1218 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1219 sacked &= ~TCPCB_SACKED_RETRANS;
1220 tp->retrans_out -= pcount;
1226 /* Shift newly-SACKed bytes from this skb to the immediately previous
1227 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1229 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1230 struct tcp_sacktag_state *state,
1231 unsigned int pcount, int shifted, int mss,
1234 struct tcp_sock *tp = tcp_sk(sk);
1235 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1236 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1237 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1241 /* Adjust counters and hints for the newly sacked sequence
1242 * range but discard the return value since prev is already
1243 * marked. We must tag the range first because the seq
1244 * advancement below implicitly advances
1245 * tcp_highest_sack_seq() when skb is highest_sack.
1247 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1248 start_seq, end_seq, dup_sack, pcount);
1250 if (skb == tp->lost_skb_hint)
1251 tp->lost_cnt_hint += pcount;
1253 TCP_SKB_CB(prev)->end_seq += shifted;
1254 TCP_SKB_CB(skb)->seq += shifted;
1256 skb_shinfo(prev)->gso_segs += pcount;
1257 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1258 skb_shinfo(skb)->gso_segs -= pcount;
1260 /* When we're adding to gso_segs == 1, gso_size will be zero,
1261 * in theory this shouldn't be necessary but as long as DSACK
1262 * code can come after this skb later on it's better to keep
1263 * setting gso_size to something.
1265 if (!skb_shinfo(prev)->gso_size) {
1266 skb_shinfo(prev)->gso_size = mss;
1267 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1270 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1271 if (skb_shinfo(skb)->gso_segs <= 1) {
1272 skb_shinfo(skb)->gso_size = 0;
1273 skb_shinfo(skb)->gso_type = 0;
1276 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1277 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1280 BUG_ON(!tcp_skb_pcount(skb));
1281 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1285 /* Whole SKB was eaten :-) */
1287 if (skb == tp->retransmit_skb_hint)
1288 tp->retransmit_skb_hint = prev;
1289 if (skb == tp->scoreboard_skb_hint)
1290 tp->scoreboard_skb_hint = prev;
1291 if (skb == tp->lost_skb_hint) {
1292 tp->lost_skb_hint = prev;
1293 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1296 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1297 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1298 TCP_SKB_CB(prev)->end_seq++;
1300 if (skb == tcp_highest_sack(sk))
1301 tcp_advance_highest_sack(sk, skb);
1303 tcp_unlink_write_queue(skb, sk);
1304 sk_wmem_free_skb(sk, skb);
1306 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1311 /* I wish gso_size would have a bit more sane initialization than
1312 * something-or-zero which complicates things
1314 static int tcp_skb_seglen(const struct sk_buff *skb)
1316 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1319 /* Shifting pages past head area doesn't work */
1320 static int skb_can_shift(const struct sk_buff *skb)
1322 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1325 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1328 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1329 struct tcp_sacktag_state *state,
1330 u32 start_seq, u32 end_seq,
1333 struct tcp_sock *tp = tcp_sk(sk);
1334 struct sk_buff *prev;
1340 if (!sk_can_gso(sk))
1343 /* Normally R but no L won't result in plain S */
1345 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1347 if (!skb_can_shift(skb))
1349 /* This frame is about to be dropped (was ACKed). */
1350 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1353 /* Can only happen with delayed DSACK + discard craziness */
1354 if (unlikely(skb == tcp_write_queue_head(sk)))
1356 prev = tcp_write_queue_prev(sk, skb);
1358 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1361 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1362 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1366 pcount = tcp_skb_pcount(skb);
1367 mss = tcp_skb_seglen(skb);
1369 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1370 * drop this restriction as unnecessary
1372 if (mss != tcp_skb_seglen(prev))
1375 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1377 /* CHECKME: This is non-MSS split case only?, this will
1378 * cause skipped skbs due to advancing loop btw, original
1379 * has that feature too
1381 if (tcp_skb_pcount(skb) <= 1)
1384 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1386 /* TODO: head merge to next could be attempted here
1387 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1388 * though it might not be worth of the additional hassle
1390 * ...we can probably just fallback to what was done
1391 * previously. We could try merging non-SACKed ones
1392 * as well but it probably isn't going to buy off
1393 * because later SACKs might again split them, and
1394 * it would make skb timestamp tracking considerably
1400 len = end_seq - TCP_SKB_CB(skb)->seq;
1402 BUG_ON(len > skb->len);
1404 /* MSS boundaries should be honoured or else pcount will
1405 * severely break even though it makes things bit trickier.
1406 * Optimize common case to avoid most of the divides
1408 mss = tcp_skb_mss(skb);
1410 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1411 * drop this restriction as unnecessary
1413 if (mss != tcp_skb_seglen(prev))
1418 } else if (len < mss) {
1426 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1427 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1430 if (!skb_shift(prev, skb, len))
1432 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1435 /* Hole filled allows collapsing with the next as well, this is very
1436 * useful when hole on every nth skb pattern happens
1438 if (prev == tcp_write_queue_tail(sk))
1440 skb = tcp_write_queue_next(sk, prev);
1442 if (!skb_can_shift(skb) ||
1443 (skb == tcp_send_head(sk)) ||
1444 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1445 (mss != tcp_skb_seglen(skb)))
1449 if (skb_shift(prev, skb, len)) {
1450 pcount += tcp_skb_pcount(skb);
1451 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1455 state->fack_count += pcount;
1462 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1466 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1467 struct tcp_sack_block *next_dup,
1468 struct tcp_sacktag_state *state,
1469 u32 start_seq, u32 end_seq,
1472 struct tcp_sock *tp = tcp_sk(sk);
1473 struct sk_buff *tmp;
1475 tcp_for_write_queue_from(skb, sk) {
1477 bool dup_sack = dup_sack_in;
1479 if (skb == tcp_send_head(sk))
1482 /* queue is in-order => we can short-circuit the walk early */
1483 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1486 if ((next_dup != NULL) &&
1487 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1488 in_sack = tcp_match_skb_to_sack(sk, skb,
1489 next_dup->start_seq,
1495 /* skb reference here is a bit tricky to get right, since
1496 * shifting can eat and free both this skb and the next,
1497 * so not even _safe variant of the loop is enough.
1500 tmp = tcp_shift_skb_data(sk, skb, state,
1501 start_seq, end_seq, dup_sack);
1510 in_sack = tcp_match_skb_to_sack(sk, skb,
1516 if (unlikely(in_sack < 0))
1520 TCP_SKB_CB(skb)->sacked =
1523 TCP_SKB_CB(skb)->sacked,
1524 TCP_SKB_CB(skb)->seq,
1525 TCP_SKB_CB(skb)->end_seq,
1527 tcp_skb_pcount(skb));
1529 if (!before(TCP_SKB_CB(skb)->seq,
1530 tcp_highest_sack_seq(tp)))
1531 tcp_advance_highest_sack(sk, skb);
1534 state->fack_count += tcp_skb_pcount(skb);
1539 /* Avoid all extra work that is being done by sacktag while walking in
1542 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1543 struct tcp_sacktag_state *state,
1546 tcp_for_write_queue_from(skb, sk) {
1547 if (skb == tcp_send_head(sk))
1550 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1553 state->fack_count += tcp_skb_pcount(skb);
1558 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1560 struct tcp_sack_block *next_dup,
1561 struct tcp_sacktag_state *state,
1564 if (next_dup == NULL)
1567 if (before(next_dup->start_seq, skip_to_seq)) {
1568 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1569 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1570 next_dup->start_seq, next_dup->end_seq,
1577 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1579 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1583 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1586 struct tcp_sock *tp = tcp_sk(sk);
1587 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1588 TCP_SKB_CB(ack_skb)->sacked);
1589 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1590 struct tcp_sack_block sp[TCP_NUM_SACKS];
1591 struct tcp_sack_block *cache;
1592 struct tcp_sacktag_state state;
1593 struct sk_buff *skb;
1594 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1596 bool found_dup_sack = false;
1598 int first_sack_index;
1601 state.reord = tp->packets_out;
1603 if (!tp->sacked_out) {
1604 if (WARN_ON(tp->fackets_out))
1605 tp->fackets_out = 0;
1606 tcp_highest_sack_reset(sk);
1609 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1610 num_sacks, prior_snd_una);
1612 state.flag |= FLAG_DSACKING_ACK;
1614 /* Eliminate too old ACKs, but take into
1615 * account more or less fresh ones, they can
1616 * contain valid SACK info.
1618 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1621 if (!tp->packets_out)
1625 first_sack_index = 0;
1626 for (i = 0; i < num_sacks; i++) {
1627 bool dup_sack = !i && found_dup_sack;
1629 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1630 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1632 if (!tcp_is_sackblock_valid(tp, dup_sack,
1633 sp[used_sacks].start_seq,
1634 sp[used_sacks].end_seq)) {
1638 if (!tp->undo_marker)
1639 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1641 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1643 /* Don't count olds caused by ACK reordering */
1644 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1645 !after(sp[used_sacks].end_seq, tp->snd_una))
1647 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1650 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1652 first_sack_index = -1;
1656 /* Ignore very old stuff early */
1657 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1663 /* order SACK blocks to allow in order walk of the retrans queue */
1664 for (i = used_sacks - 1; i > 0; i--) {
1665 for (j = 0; j < i; j++) {
1666 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1667 swap(sp[j], sp[j + 1]);
1669 /* Track where the first SACK block goes to */
1670 if (j == first_sack_index)
1671 first_sack_index = j + 1;
1676 skb = tcp_write_queue_head(sk);
1677 state.fack_count = 0;
1680 if (!tp->sacked_out) {
1681 /* It's already past, so skip checking against it */
1682 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1684 cache = tp->recv_sack_cache;
1685 /* Skip empty blocks in at head of the cache */
1686 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1691 while (i < used_sacks) {
1692 u32 start_seq = sp[i].start_seq;
1693 u32 end_seq = sp[i].end_seq;
1694 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1695 struct tcp_sack_block *next_dup = NULL;
1697 if (found_dup_sack && ((i + 1) == first_sack_index))
1698 next_dup = &sp[i + 1];
1700 /* Skip too early cached blocks */
1701 while (tcp_sack_cache_ok(tp, cache) &&
1702 !before(start_seq, cache->end_seq))
1705 /* Can skip some work by looking recv_sack_cache? */
1706 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1707 after(end_seq, cache->start_seq)) {
1710 if (before(start_seq, cache->start_seq)) {
1711 skb = tcp_sacktag_skip(skb, sk, &state,
1713 skb = tcp_sacktag_walk(skb, sk, next_dup,
1720 /* Rest of the block already fully processed? */
1721 if (!after(end_seq, cache->end_seq))
1724 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1728 /* ...tail remains todo... */
1729 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1730 /* ...but better entrypoint exists! */
1731 skb = tcp_highest_sack(sk);
1734 state.fack_count = tp->fackets_out;
1739 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1740 /* Check overlap against next cached too (past this one already) */
1745 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1746 skb = tcp_highest_sack(sk);
1749 state.fack_count = tp->fackets_out;
1751 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1754 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1755 start_seq, end_seq, dup_sack);
1761 /* Clear the head of the cache sack blocks so we can skip it next time */
1762 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1763 tp->recv_sack_cache[i].start_seq = 0;
1764 tp->recv_sack_cache[i].end_seq = 0;
1766 for (j = 0; j < used_sacks; j++)
1767 tp->recv_sack_cache[i++] = sp[j];
1769 tcp_mark_lost_retrans(sk);
1771 tcp_verify_left_out(tp);
1773 if ((state.reord < tp->fackets_out) &&
1774 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1775 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1779 #if FASTRETRANS_DEBUG > 0
1780 WARN_ON((int)tp->sacked_out < 0);
1781 WARN_ON((int)tp->lost_out < 0);
1782 WARN_ON((int)tp->retrans_out < 0);
1783 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1788 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1789 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1791 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1795 holes = max(tp->lost_out, 1U);
1796 holes = min(holes, tp->packets_out);
1798 if ((tp->sacked_out + holes) > tp->packets_out) {
1799 tp->sacked_out = tp->packets_out - holes;
1805 /* If we receive more dupacks than we expected counting segments
1806 * in assumption of absent reordering, interpret this as reordering.
1807 * The only another reason could be bug in receiver TCP.
1809 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1811 struct tcp_sock *tp = tcp_sk(sk);
1812 if (tcp_limit_reno_sacked(tp))
1813 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1816 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1818 static void tcp_add_reno_sack(struct sock *sk)
1820 struct tcp_sock *tp = tcp_sk(sk);
1822 tcp_check_reno_reordering(sk, 0);
1823 tcp_verify_left_out(tp);
1826 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1828 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1830 struct tcp_sock *tp = tcp_sk(sk);
1833 /* One ACK acked hole. The rest eat duplicate ACKs. */
1834 if (acked - 1 >= tp->sacked_out)
1837 tp->sacked_out -= acked - 1;
1839 tcp_check_reno_reordering(sk, acked);
1840 tcp_verify_left_out(tp);
1843 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1848 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1850 tp->retrans_out = 0;
1853 tp->undo_marker = 0;
1854 tp->undo_retrans = -1;
1857 void tcp_clear_retrans(struct tcp_sock *tp)
1859 tcp_clear_retrans_partial(tp);
1861 tp->fackets_out = 0;
1865 /* Enter Loss state. If "how" is not zero, forget all SACK information
1866 * and reset tags completely, otherwise preserve SACKs. If receiver
1867 * dropped its ofo queue, we will know this due to reneging detection.
1869 void tcp_enter_loss(struct sock *sk, int how)
1871 const struct inet_connection_sock *icsk = inet_csk(sk);
1872 struct tcp_sock *tp = tcp_sk(sk);
1873 struct sk_buff *skb;
1874 bool new_recovery = false;
1876 /* Reduce ssthresh if it has not yet been made inside this window. */
1877 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1878 !after(tp->high_seq, tp->snd_una) ||
1879 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1880 new_recovery = true;
1881 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1882 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1883 tcp_ca_event(sk, CA_EVENT_LOSS);
1886 tp->snd_cwnd_cnt = 0;
1887 tp->snd_cwnd_stamp = tcp_time_stamp;
1889 tcp_clear_retrans_partial(tp);
1891 if (tcp_is_reno(tp))
1892 tcp_reset_reno_sack(tp);
1894 tp->undo_marker = tp->snd_una;
1897 tp->fackets_out = 0;
1899 tcp_clear_all_retrans_hints(tp);
1901 tcp_for_write_queue(skb, sk) {
1902 if (skb == tcp_send_head(sk))
1905 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1906 tp->undo_marker = 0;
1907 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1908 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1909 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1910 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1911 tp->lost_out += tcp_skb_pcount(skb);
1912 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1915 tcp_verify_left_out(tp);
1917 tp->reordering = min_t(unsigned int, tp->reordering,
1918 sysctl_tcp_reordering);
1919 tcp_set_ca_state(sk, TCP_CA_Loss);
1920 tp->high_seq = tp->snd_nxt;
1921 TCP_ECN_queue_cwr(tp);
1923 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1924 * loss recovery is underway except recurring timeout(s) on
1925 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1927 tp->frto = sysctl_tcp_frto &&
1928 (new_recovery || icsk->icsk_retransmits) &&
1929 !inet_csk(sk)->icsk_mtup.probe_size;
1932 /* If ACK arrived pointing to a remembered SACK, it means that our
1933 * remembered SACKs do not reflect real state of receiver i.e.
1934 * receiver _host_ is heavily congested (or buggy).
1936 * Do processing similar to RTO timeout.
1938 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1940 if (flag & FLAG_SACK_RENEGING) {
1941 struct inet_connection_sock *icsk = inet_csk(sk);
1942 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1944 tcp_enter_loss(sk, 1);
1945 icsk->icsk_retransmits++;
1946 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1947 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1948 icsk->icsk_rto, TCP_RTO_MAX);
1954 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1956 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1959 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1960 * counter when SACK is enabled (without SACK, sacked_out is used for
1963 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1964 * segments up to the highest received SACK block so far and holes in
1967 * With reordering, holes may still be in flight, so RFC3517 recovery
1968 * uses pure sacked_out (total number of SACKed segments) even though
1969 * it violates the RFC that uses duplicate ACKs, often these are equal
1970 * but when e.g. out-of-window ACKs or packet duplication occurs,
1971 * they differ. Since neither occurs due to loss, TCP should really
1974 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1976 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1979 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1981 struct tcp_sock *tp = tcp_sk(sk);
1982 unsigned long delay;
1984 /* Delay early retransmit and entering fast recovery for
1985 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1986 * available, or RTO is scheduled to fire first.
1988 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
1989 (flag & FLAG_ECE) || !tp->srtt)
1992 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
1993 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
1996 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2001 static inline int tcp_skb_timedout(const struct sock *sk,
2002 const struct sk_buff *skb)
2004 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2007 static inline int tcp_head_timedout(const struct sock *sk)
2009 const struct tcp_sock *tp = tcp_sk(sk);
2011 return tp->packets_out &&
2012 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2015 /* Linux NewReno/SACK/FACK/ECN state machine.
2016 * --------------------------------------
2018 * "Open" Normal state, no dubious events, fast path.
2019 * "Disorder" In all the respects it is "Open",
2020 * but requires a bit more attention. It is entered when
2021 * we see some SACKs or dupacks. It is split of "Open"
2022 * mainly to move some processing from fast path to slow one.
2023 * "CWR" CWND was reduced due to some Congestion Notification event.
2024 * It can be ECN, ICMP source quench, local device congestion.
2025 * "Recovery" CWND was reduced, we are fast-retransmitting.
2026 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2028 * tcp_fastretrans_alert() is entered:
2029 * - each incoming ACK, if state is not "Open"
2030 * - when arrived ACK is unusual, namely:
2035 * Counting packets in flight is pretty simple.
2037 * in_flight = packets_out - left_out + retrans_out
2039 * packets_out is SND.NXT-SND.UNA counted in packets.
2041 * retrans_out is number of retransmitted segments.
2043 * left_out is number of segments left network, but not ACKed yet.
2045 * left_out = sacked_out + lost_out
2047 * sacked_out: Packets, which arrived to receiver out of order
2048 * and hence not ACKed. With SACKs this number is simply
2049 * amount of SACKed data. Even without SACKs
2050 * it is easy to give pretty reliable estimate of this number,
2051 * counting duplicate ACKs.
2053 * lost_out: Packets lost by network. TCP has no explicit
2054 * "loss notification" feedback from network (for now).
2055 * It means that this number can be only _guessed_.
2056 * Actually, it is the heuristics to predict lossage that
2057 * distinguishes different algorithms.
2059 * F.e. after RTO, when all the queue is considered as lost,
2060 * lost_out = packets_out and in_flight = retrans_out.
2062 * Essentially, we have now two algorithms counting
2065 * FACK: It is the simplest heuristics. As soon as we decided
2066 * that something is lost, we decide that _all_ not SACKed
2067 * packets until the most forward SACK are lost. I.e.
2068 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2069 * It is absolutely correct estimate, if network does not reorder
2070 * packets. And it loses any connection to reality when reordering
2071 * takes place. We use FACK by default until reordering
2072 * is suspected on the path to this destination.
2074 * NewReno: when Recovery is entered, we assume that one segment
2075 * is lost (classic Reno). While we are in Recovery and
2076 * a partial ACK arrives, we assume that one more packet
2077 * is lost (NewReno). This heuristics are the same in NewReno
2080 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2081 * deflation etc. CWND is real congestion window, never inflated, changes
2082 * only according to classic VJ rules.
2084 * Really tricky (and requiring careful tuning) part of algorithm
2085 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2086 * The first determines the moment _when_ we should reduce CWND and,
2087 * hence, slow down forward transmission. In fact, it determines the moment
2088 * when we decide that hole is caused by loss, rather than by a reorder.
2090 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2091 * holes, caused by lost packets.
2093 * And the most logically complicated part of algorithm is undo
2094 * heuristics. We detect false retransmits due to both too early
2095 * fast retransmit (reordering) and underestimated RTO, analyzing
2096 * timestamps and D-SACKs. When we detect that some segments were
2097 * retransmitted by mistake and CWND reduction was wrong, we undo
2098 * window reduction and abort recovery phase. This logic is hidden
2099 * inside several functions named tcp_try_undo_<something>.
2102 /* This function decides, when we should leave Disordered state
2103 * and enter Recovery phase, reducing congestion window.
2105 * Main question: may we further continue forward transmission
2106 * with the same cwnd?
2108 static bool tcp_time_to_recover(struct sock *sk, int flag)
2110 struct tcp_sock *tp = tcp_sk(sk);
2113 /* Trick#1: The loss is proven. */
2117 /* Not-A-Trick#2 : Classic rule... */
2118 if (tcp_dupack_heuristics(tp) > tp->reordering)
2121 /* Trick#3 : when we use RFC2988 timer restart, fast
2122 * retransmit can be triggered by timeout of queue head.
2124 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2127 /* Trick#4: It is still not OK... But will it be useful to delay
2130 packets_out = tp->packets_out;
2131 if (packets_out <= tp->reordering &&
2132 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2133 !tcp_may_send_now(sk)) {
2134 /* We have nothing to send. This connection is limited
2135 * either by receiver window or by application.
2140 /* If a thin stream is detected, retransmit after first
2141 * received dupack. Employ only if SACK is supported in order
2142 * to avoid possible corner-case series of spurious retransmissions
2143 * Use only if there are no unsent data.
2145 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2146 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2147 tcp_is_sack(tp) && !tcp_send_head(sk))
2150 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2151 * retransmissions due to small network reorderings, we implement
2152 * Mitigation A.3 in the RFC and delay the retransmission for a short
2153 * interval if appropriate.
2155 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2156 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2157 !tcp_may_send_now(sk))
2158 return !tcp_pause_early_retransmit(sk, flag);
2163 /* New heuristics: it is possible only after we switched to restart timer
2164 * each time when something is ACKed. Hence, we can detect timed out packets
2165 * during fast retransmit without falling to slow start.
2167 * Usefulness of this as is very questionable, since we should know which of
2168 * the segments is the next to timeout which is relatively expensive to find
2169 * in general case unless we add some data structure just for that. The
2170 * current approach certainly won't find the right one too often and when it
2171 * finally does find _something_ it usually marks large part of the window
2172 * right away (because a retransmission with a larger timestamp blocks the
2173 * loop from advancing). -ij
2175 static void tcp_timeout_skbs(struct sock *sk)
2177 struct tcp_sock *tp = tcp_sk(sk);
2178 struct sk_buff *skb;
2180 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2183 skb = tp->scoreboard_skb_hint;
2184 if (tp->scoreboard_skb_hint == NULL)
2185 skb = tcp_write_queue_head(sk);
2187 tcp_for_write_queue_from(skb, sk) {
2188 if (skb == tcp_send_head(sk))
2190 if (!tcp_skb_timedout(sk, skb))
2193 tcp_skb_mark_lost(tp, skb);
2196 tp->scoreboard_skb_hint = skb;
2198 tcp_verify_left_out(tp);
2201 /* Detect loss in event "A" above by marking head of queue up as lost.
2202 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2203 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2204 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2205 * the maximum SACKed segments to pass before reaching this limit.
2207 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2209 struct tcp_sock *tp = tcp_sk(sk);
2210 struct sk_buff *skb;
2214 /* Use SACK to deduce losses of new sequences sent during recovery */
2215 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2217 WARN_ON(packets > tp->packets_out);
2218 if (tp->lost_skb_hint) {
2219 skb = tp->lost_skb_hint;
2220 cnt = tp->lost_cnt_hint;
2221 /* Head already handled? */
2222 if (mark_head && skb != tcp_write_queue_head(sk))
2225 skb = tcp_write_queue_head(sk);
2229 tcp_for_write_queue_from(skb, sk) {
2230 if (skb == tcp_send_head(sk))
2232 /* TODO: do this better */
2233 /* this is not the most efficient way to do this... */
2234 tp->lost_skb_hint = skb;
2235 tp->lost_cnt_hint = cnt;
2237 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2241 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2242 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2243 cnt += tcp_skb_pcount(skb);
2245 if (cnt > packets) {
2246 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2247 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2248 (oldcnt >= packets))
2251 mss = skb_shinfo(skb)->gso_size;
2252 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2258 tcp_skb_mark_lost(tp, skb);
2263 tcp_verify_left_out(tp);
2266 /* Account newly detected lost packet(s) */
2268 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2270 struct tcp_sock *tp = tcp_sk(sk);
2272 if (tcp_is_reno(tp)) {
2273 tcp_mark_head_lost(sk, 1, 1);
2274 } else if (tcp_is_fack(tp)) {
2275 int lost = tp->fackets_out - tp->reordering;
2278 tcp_mark_head_lost(sk, lost, 0);
2280 int sacked_upto = tp->sacked_out - tp->reordering;
2281 if (sacked_upto >= 0)
2282 tcp_mark_head_lost(sk, sacked_upto, 0);
2283 else if (fast_rexmit)
2284 tcp_mark_head_lost(sk, 1, 1);
2287 tcp_timeout_skbs(sk);
2290 /* CWND moderation, preventing bursts due to too big ACKs
2291 * in dubious situations.
2293 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2295 tp->snd_cwnd = min(tp->snd_cwnd,
2296 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2297 tp->snd_cwnd_stamp = tcp_time_stamp;
2300 /* Nothing was retransmitted or returned timestamp is less
2301 * than timestamp of the first retransmission.
2303 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2305 return !tp->retrans_stamp ||
2306 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2307 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2310 /* Undo procedures. */
2312 #if FASTRETRANS_DEBUG > 1
2313 static void DBGUNDO(struct sock *sk, const char *msg)
2315 struct tcp_sock *tp = tcp_sk(sk);
2316 struct inet_sock *inet = inet_sk(sk);
2318 if (sk->sk_family == AF_INET) {
2319 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2321 &inet->inet_daddr, ntohs(inet->inet_dport),
2322 tp->snd_cwnd, tcp_left_out(tp),
2323 tp->snd_ssthresh, tp->prior_ssthresh,
2326 #if IS_ENABLED(CONFIG_IPV6)
2327 else if (sk->sk_family == AF_INET6) {
2328 struct ipv6_pinfo *np = inet6_sk(sk);
2329 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2331 &np->daddr, ntohs(inet->inet_dport),
2332 tp->snd_cwnd, tcp_left_out(tp),
2333 tp->snd_ssthresh, tp->prior_ssthresh,
2339 #define DBGUNDO(x...) do { } while (0)
2342 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2344 struct tcp_sock *tp = tcp_sk(sk);
2346 if (tp->prior_ssthresh) {
2347 const struct inet_connection_sock *icsk = inet_csk(sk);
2349 if (icsk->icsk_ca_ops->undo_cwnd)
2350 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2352 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2354 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2355 tp->snd_ssthresh = tp->prior_ssthresh;
2356 TCP_ECN_withdraw_cwr(tp);
2359 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2361 tp->snd_cwnd_stamp = tcp_time_stamp;
2364 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2366 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2369 /* People celebrate: "We love our President!" */
2370 static bool tcp_try_undo_recovery(struct sock *sk)
2372 struct tcp_sock *tp = tcp_sk(sk);
2374 if (tcp_may_undo(tp)) {
2377 /* Happy end! We did not retransmit anything
2378 * or our original transmission succeeded.
2380 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2381 tcp_undo_cwr(sk, true);
2382 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2383 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2385 mib_idx = LINUX_MIB_TCPFULLUNDO;
2387 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2388 tp->undo_marker = 0;
2390 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2391 /* Hold old state until something *above* high_seq
2392 * is ACKed. For Reno it is MUST to prevent false
2393 * fast retransmits (RFC2582). SACK TCP is safe. */
2394 tcp_moderate_cwnd(tp);
2397 tcp_set_ca_state(sk, TCP_CA_Open);
2401 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2402 static void tcp_try_undo_dsack(struct sock *sk)
2404 struct tcp_sock *tp = tcp_sk(sk);
2406 if (tp->undo_marker && !tp->undo_retrans) {
2407 DBGUNDO(sk, "D-SACK");
2408 tcp_undo_cwr(sk, true);
2409 tp->undo_marker = 0;
2410 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2414 /* We can clear retrans_stamp when there are no retransmissions in the
2415 * window. It would seem that it is trivially available for us in
2416 * tp->retrans_out, however, that kind of assumptions doesn't consider
2417 * what will happen if errors occur when sending retransmission for the
2418 * second time. ...It could the that such segment has only
2419 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2420 * the head skb is enough except for some reneging corner cases that
2421 * are not worth the effort.
2423 * Main reason for all this complexity is the fact that connection dying
2424 * time now depends on the validity of the retrans_stamp, in particular,
2425 * that successive retransmissions of a segment must not advance
2426 * retrans_stamp under any conditions.
2428 static bool tcp_any_retrans_done(const struct sock *sk)
2430 const struct tcp_sock *tp = tcp_sk(sk);
2431 struct sk_buff *skb;
2433 if (tp->retrans_out)
2436 skb = tcp_write_queue_head(sk);
2437 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2443 /* Undo during fast recovery after partial ACK. */
2445 static int tcp_try_undo_partial(struct sock *sk, int acked)
2447 struct tcp_sock *tp = tcp_sk(sk);
2448 /* Partial ACK arrived. Force Hoe's retransmit. */
2449 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2451 if (tcp_may_undo(tp)) {
2452 /* Plain luck! Hole if filled with delayed
2453 * packet, rather than with a retransmit.
2455 if (!tcp_any_retrans_done(sk))
2456 tp->retrans_stamp = 0;
2458 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2461 tcp_undo_cwr(sk, false);
2462 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2464 /* So... Do not make Hoe's retransmit yet.
2465 * If the first packet was delayed, the rest
2466 * ones are most probably delayed as well.
2473 /* Undo during loss recovery after partial ACK or using F-RTO. */
2474 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2476 struct tcp_sock *tp = tcp_sk(sk);
2478 if (frto_undo || tcp_may_undo(tp)) {
2479 struct sk_buff *skb;
2480 tcp_for_write_queue(skb, sk) {
2481 if (skb == tcp_send_head(sk))
2483 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2486 tcp_clear_all_retrans_hints(tp);
2488 DBGUNDO(sk, "partial loss");
2490 tcp_undo_cwr(sk, true);
2491 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2493 NET_INC_STATS_BH(sock_net(sk),
2494 LINUX_MIB_TCPSPURIOUSRTOS);
2495 inet_csk(sk)->icsk_retransmits = 0;
2496 tp->undo_marker = 0;
2497 if (frto_undo || tcp_is_sack(tp))
2498 tcp_set_ca_state(sk, TCP_CA_Open);
2504 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2505 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2506 * It computes the number of packets to send (sndcnt) based on packets newly
2508 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2509 * cwnd reductions across a full RTT.
2510 * 2) If packets in flight is lower than ssthresh (such as due to excess
2511 * losses and/or application stalls), do not perform any further cwnd
2512 * reductions, but instead slow start up to ssthresh.
2514 static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2516 struct tcp_sock *tp = tcp_sk(sk);
2518 tp->high_seq = tp->snd_nxt;
2519 tp->tlp_high_seq = 0;
2520 tp->snd_cwnd_cnt = 0;
2521 tp->prior_cwnd = tp->snd_cwnd;
2522 tp->prr_delivered = 0;
2525 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2526 TCP_ECN_queue_cwr(tp);
2529 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2532 struct tcp_sock *tp = tcp_sk(sk);
2534 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2536 tp->prr_delivered += newly_acked_sacked;
2537 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2538 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2540 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2542 sndcnt = min_t(int, delta,
2543 max_t(int, tp->prr_delivered - tp->prr_out,
2544 newly_acked_sacked) + 1);
2547 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2548 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2551 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2553 struct tcp_sock *tp = tcp_sk(sk);
2555 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2556 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2557 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2558 tp->snd_cwnd = tp->snd_ssthresh;
2559 tp->snd_cwnd_stamp = tcp_time_stamp;
2561 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2564 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2565 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2567 struct tcp_sock *tp = tcp_sk(sk);
2569 tp->prior_ssthresh = 0;
2570 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2571 tp->undo_marker = 0;
2572 tcp_init_cwnd_reduction(sk, set_ssthresh);
2573 tcp_set_ca_state(sk, TCP_CA_CWR);
2577 static void tcp_try_keep_open(struct sock *sk)
2579 struct tcp_sock *tp = tcp_sk(sk);
2580 int state = TCP_CA_Open;
2582 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2583 state = TCP_CA_Disorder;
2585 if (inet_csk(sk)->icsk_ca_state != state) {
2586 tcp_set_ca_state(sk, state);
2587 tp->high_seq = tp->snd_nxt;
2591 static void tcp_try_to_open(struct sock *sk, int flag, int newly_acked_sacked)
2593 struct tcp_sock *tp = tcp_sk(sk);
2595 tcp_verify_left_out(tp);
2597 if (!tcp_any_retrans_done(sk))
2598 tp->retrans_stamp = 0;
2600 if (flag & FLAG_ECE)
2601 tcp_enter_cwr(sk, 1);
2603 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2604 tcp_try_keep_open(sk);
2605 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2606 tcp_moderate_cwnd(tp);
2608 tcp_cwnd_reduction(sk, newly_acked_sacked, 0);
2612 static void tcp_mtup_probe_failed(struct sock *sk)
2614 struct inet_connection_sock *icsk = inet_csk(sk);
2616 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2617 icsk->icsk_mtup.probe_size = 0;
2620 static void tcp_mtup_probe_success(struct sock *sk)
2622 struct tcp_sock *tp = tcp_sk(sk);
2623 struct inet_connection_sock *icsk = inet_csk(sk);
2625 /* FIXME: breaks with very large cwnd */
2626 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2627 tp->snd_cwnd = tp->snd_cwnd *
2628 tcp_mss_to_mtu(sk, tp->mss_cache) /
2629 icsk->icsk_mtup.probe_size;
2630 tp->snd_cwnd_cnt = 0;
2631 tp->snd_cwnd_stamp = tcp_time_stamp;
2632 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2634 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2635 icsk->icsk_mtup.probe_size = 0;
2636 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2639 /* Do a simple retransmit without using the backoff mechanisms in
2640 * tcp_timer. This is used for path mtu discovery.
2641 * The socket is already locked here.
2643 void tcp_simple_retransmit(struct sock *sk)
2645 const struct inet_connection_sock *icsk = inet_csk(sk);
2646 struct tcp_sock *tp = tcp_sk(sk);
2647 struct sk_buff *skb;
2648 unsigned int mss = tcp_current_mss(sk);
2649 u32 prior_lost = tp->lost_out;
2651 tcp_for_write_queue(skb, sk) {
2652 if (skb == tcp_send_head(sk))
2654 if (tcp_skb_seglen(skb) > mss &&
2655 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2656 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2657 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2658 tp->retrans_out -= tcp_skb_pcount(skb);
2660 tcp_skb_mark_lost_uncond_verify(tp, skb);
2664 tcp_clear_retrans_hints_partial(tp);
2666 if (prior_lost == tp->lost_out)
2669 if (tcp_is_reno(tp))
2670 tcp_limit_reno_sacked(tp);
2672 tcp_verify_left_out(tp);
2674 /* Don't muck with the congestion window here.
2675 * Reason is that we do not increase amount of _data_
2676 * in network, but units changed and effective
2677 * cwnd/ssthresh really reduced now.
2679 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2680 tp->high_seq = tp->snd_nxt;
2681 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2682 tp->prior_ssthresh = 0;
2683 tp->undo_marker = 0;
2684 tcp_set_ca_state(sk, TCP_CA_Loss);
2686 tcp_xmit_retransmit_queue(sk);
2688 EXPORT_SYMBOL(tcp_simple_retransmit);
2690 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2692 struct tcp_sock *tp = tcp_sk(sk);
2695 if (tcp_is_reno(tp))
2696 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2698 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2700 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2702 tp->prior_ssthresh = 0;
2703 tp->undo_marker = tp->snd_una;
2704 tp->undo_retrans = tp->retrans_out ? : -1;
2706 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2708 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2709 tcp_init_cwnd_reduction(sk, true);
2711 tcp_set_ca_state(sk, TCP_CA_Recovery);
2714 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2715 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2717 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2719 struct inet_connection_sock *icsk = inet_csk(sk);
2720 struct tcp_sock *tp = tcp_sk(sk);
2721 bool recovered = !before(tp->snd_una, tp->high_seq);
2723 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2724 /* Step 3.b. A timeout is spurious if not all data are
2725 * lost, i.e., never-retransmitted data are (s)acked.
2727 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2730 if (after(tp->snd_nxt, tp->high_seq) &&
2731 (flag & FLAG_DATA_SACKED || is_dupack)) {
2732 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2733 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2734 tp->high_seq = tp->snd_nxt;
2735 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2737 if (after(tp->snd_nxt, tp->high_seq))
2738 return; /* Step 2.b */
2744 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2745 icsk->icsk_retransmits = 0;
2746 tcp_try_undo_recovery(sk);
2749 if (flag & FLAG_DATA_ACKED)
2750 icsk->icsk_retransmits = 0;
2751 if (tcp_is_reno(tp)) {
2752 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2753 * delivered. Lower inflight to clock out (re)tranmissions.
2755 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2756 tcp_add_reno_sack(sk);
2757 else if (flag & FLAG_SND_UNA_ADVANCED)
2758 tcp_reset_reno_sack(tp);
2760 if (tcp_try_undo_loss(sk, false))
2762 tcp_xmit_retransmit_queue(sk);
2765 /* Process an event, which can update packets-in-flight not trivially.
2766 * Main goal of this function is to calculate new estimate for left_out,
2767 * taking into account both packets sitting in receiver's buffer and
2768 * packets lost by network.
2770 * Besides that it does CWND reduction, when packet loss is detected
2771 * and changes state of machine.
2773 * It does _not_ decide what to send, it is made in function
2774 * tcp_xmit_retransmit_queue().
2776 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2777 int prior_sacked, int prior_packets,
2778 bool is_dupack, int flag)
2780 struct inet_connection_sock *icsk = inet_csk(sk);
2781 struct tcp_sock *tp = tcp_sk(sk);
2782 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2783 (tcp_fackets_out(tp) > tp->reordering));
2784 int newly_acked_sacked = 0;
2785 int fast_rexmit = 0;
2787 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2789 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2790 tp->fackets_out = 0;
2792 /* Now state machine starts.
2793 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2794 if (flag & FLAG_ECE)
2795 tp->prior_ssthresh = 0;
2797 /* B. In all the states check for reneging SACKs. */
2798 if (tcp_check_sack_reneging(sk, flag))
2801 /* C. Check consistency of the current state. */
2802 tcp_verify_left_out(tp);
2804 /* D. Check state exit conditions. State can be terminated
2805 * when high_seq is ACKed. */
2806 if (icsk->icsk_ca_state == TCP_CA_Open) {
2807 WARN_ON(tp->retrans_out != 0);
2808 tp->retrans_stamp = 0;
2809 } else if (!before(tp->snd_una, tp->high_seq)) {
2810 switch (icsk->icsk_ca_state) {
2812 /* CWR is to be held something *above* high_seq
2813 * is ACKed for CWR bit to reach receiver. */
2814 if (tp->snd_una != tp->high_seq) {
2815 tcp_end_cwnd_reduction(sk);
2816 tcp_set_ca_state(sk, TCP_CA_Open);
2820 case TCP_CA_Recovery:
2821 if (tcp_is_reno(tp))
2822 tcp_reset_reno_sack(tp);
2823 if (tcp_try_undo_recovery(sk))
2825 tcp_end_cwnd_reduction(sk);
2830 /* E. Process state. */
2831 switch (icsk->icsk_ca_state) {
2832 case TCP_CA_Recovery:
2833 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2834 if (tcp_is_reno(tp) && is_dupack)
2835 tcp_add_reno_sack(sk);
2837 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2838 newly_acked_sacked = prior_packets - tp->packets_out +
2839 tp->sacked_out - prior_sacked;
2842 tcp_process_loss(sk, flag, is_dupack);
2843 if (icsk->icsk_ca_state != TCP_CA_Open)
2845 /* Fall through to processing in Open state. */
2847 if (tcp_is_reno(tp)) {
2848 if (flag & FLAG_SND_UNA_ADVANCED)
2849 tcp_reset_reno_sack(tp);
2851 tcp_add_reno_sack(sk);
2853 newly_acked_sacked = prior_packets - tp->packets_out +
2854 tp->sacked_out - prior_sacked;
2856 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2857 tcp_try_undo_dsack(sk);
2859 if (!tcp_time_to_recover(sk, flag)) {
2860 tcp_try_to_open(sk, flag, newly_acked_sacked);
2864 /* MTU probe failure: don't reduce cwnd */
2865 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2866 icsk->icsk_mtup.probe_size &&
2867 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2868 tcp_mtup_probe_failed(sk);
2869 /* Restores the reduction we did in tcp_mtup_probe() */
2871 tcp_simple_retransmit(sk);
2875 /* Otherwise enter Recovery state */
2876 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2880 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2881 tcp_update_scoreboard(sk, fast_rexmit);
2882 tcp_cwnd_reduction(sk, newly_acked_sacked, fast_rexmit);
2883 tcp_xmit_retransmit_queue(sk);
2886 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
2888 tcp_rtt_estimator(sk, seq_rtt);
2890 inet_csk(sk)->icsk_backoff = 0;
2892 EXPORT_SYMBOL(tcp_valid_rtt_meas);
2894 /* Read draft-ietf-tcplw-high-performance before mucking
2895 * with this code. (Supersedes RFC1323)
2897 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2899 /* RTTM Rule: A TSecr value received in a segment is used to
2900 * update the averaged RTT measurement only if the segment
2901 * acknowledges some new data, i.e., only if it advances the
2902 * left edge of the send window.
2904 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2905 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2907 * Changed: reset backoff as soon as we see the first valid sample.
2908 * If we do not, we get strongly overestimated rto. With timestamps
2909 * samples are accepted even from very old segments: f.e., when rtt=1
2910 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2911 * answer arrives rto becomes 120 seconds! If at least one of segments
2912 * in window is lost... Voila. --ANK (010210)
2914 struct tcp_sock *tp = tcp_sk(sk);
2916 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2919 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2921 /* We don't have a timestamp. Can only use
2922 * packets that are not retransmitted to determine
2923 * rtt estimates. Also, we must not reset the
2924 * backoff for rto until we get a non-retransmitted
2925 * packet. This allows us to deal with a situation
2926 * where the network delay has increased suddenly.
2927 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2930 if (flag & FLAG_RETRANS_DATA_ACKED)
2933 tcp_valid_rtt_meas(sk, seq_rtt);
2936 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2939 const struct tcp_sock *tp = tcp_sk(sk);
2940 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2941 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2942 tcp_ack_saw_tstamp(sk, flag);
2943 else if (seq_rtt >= 0)
2944 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2947 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2949 const struct inet_connection_sock *icsk = inet_csk(sk);
2950 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2951 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2954 /* Restart timer after forward progress on connection.
2955 * RFC2988 recommends to restart timer to now+rto.
2957 void tcp_rearm_rto(struct sock *sk)
2959 const struct inet_connection_sock *icsk = inet_csk(sk);
2960 struct tcp_sock *tp = tcp_sk(sk);
2962 /* If the retrans timer is currently being used by Fast Open
2963 * for SYN-ACK retrans purpose, stay put.
2965 if (tp->fastopen_rsk)
2968 if (!tp->packets_out) {
2969 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2971 u32 rto = inet_csk(sk)->icsk_rto;
2972 /* Offset the time elapsed after installing regular RTO */
2973 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2974 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2975 struct sk_buff *skb = tcp_write_queue_head(sk);
2976 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2977 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2978 /* delta may not be positive if the socket is locked
2979 * when the retrans timer fires and is rescheduled.
2984 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2989 /* This function is called when the delayed ER timer fires. TCP enters
2990 * fast recovery and performs fast-retransmit.
2992 void tcp_resume_early_retransmit(struct sock *sk)
2994 struct tcp_sock *tp = tcp_sk(sk);
2998 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2999 if (!tp->do_early_retrans)
3002 tcp_enter_recovery(sk, false);
3003 tcp_update_scoreboard(sk, 1);
3004 tcp_xmit_retransmit_queue(sk);
3007 /* If we get here, the whole TSO packet has not been acked. */
3008 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3010 struct tcp_sock *tp = tcp_sk(sk);
3013 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3015 packets_acked = tcp_skb_pcount(skb);
3016 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3018 packets_acked -= tcp_skb_pcount(skb);
3020 if (packets_acked) {
3021 BUG_ON(tcp_skb_pcount(skb) == 0);
3022 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3025 return packets_acked;
3028 /* Remove acknowledged frames from the retransmission queue. If our packet
3029 * is before the ack sequence we can discard it as it's confirmed to have
3030 * arrived at the other end.
3032 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3035 struct tcp_sock *tp = tcp_sk(sk);
3036 const struct inet_connection_sock *icsk = inet_csk(sk);
3037 struct sk_buff *skb;
3038 u32 now = tcp_time_stamp;
3039 int fully_acked = true;
3042 u32 reord = tp->packets_out;
3043 u32 prior_sacked = tp->sacked_out;
3045 s32 ca_seq_rtt = -1;
3046 ktime_t last_ackt = net_invalid_timestamp();
3048 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3049 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3051 u8 sacked = scb->sacked;
3053 /* Determine how many packets and what bytes were acked, tso and else */
3054 if (after(scb->end_seq, tp->snd_una)) {
3055 if (tcp_skb_pcount(skb) == 1 ||
3056 !after(tp->snd_una, scb->seq))
3059 acked_pcount = tcp_tso_acked(sk, skb);
3063 fully_acked = false;
3065 acked_pcount = tcp_skb_pcount(skb);
3068 if (sacked & TCPCB_RETRANS) {
3069 if (sacked & TCPCB_SACKED_RETRANS)
3070 tp->retrans_out -= acked_pcount;
3071 flag |= FLAG_RETRANS_DATA_ACKED;
3075 ca_seq_rtt = now - scb->when;
3076 last_ackt = skb->tstamp;
3078 seq_rtt = ca_seq_rtt;
3080 if (!(sacked & TCPCB_SACKED_ACKED))
3081 reord = min(pkts_acked, reord);
3082 if (!after(scb->end_seq, tp->high_seq))
3083 flag |= FLAG_ORIG_SACK_ACKED;
3086 if (sacked & TCPCB_SACKED_ACKED)
3087 tp->sacked_out -= acked_pcount;
3088 if (sacked & TCPCB_LOST)
3089 tp->lost_out -= acked_pcount;
3091 tp->packets_out -= acked_pcount;
3092 pkts_acked += acked_pcount;
3094 /* Initial outgoing SYN's get put onto the write_queue
3095 * just like anything else we transmit. It is not
3096 * true data, and if we misinform our callers that
3097 * this ACK acks real data, we will erroneously exit
3098 * connection startup slow start one packet too
3099 * quickly. This is severely frowned upon behavior.
3101 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3102 flag |= FLAG_DATA_ACKED;
3104 flag |= FLAG_SYN_ACKED;
3105 tp->retrans_stamp = 0;
3111 tcp_unlink_write_queue(skb, sk);
3112 sk_wmem_free_skb(sk, skb);
3113 tp->scoreboard_skb_hint = NULL;
3114 if (skb == tp->retransmit_skb_hint)
3115 tp->retransmit_skb_hint = NULL;
3116 if (skb == tp->lost_skb_hint)
3117 tp->lost_skb_hint = NULL;
3120 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3121 tp->snd_up = tp->snd_una;
3123 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3124 flag |= FLAG_SACK_RENEGING;
3126 if (flag & FLAG_ACKED) {
3127 const struct tcp_congestion_ops *ca_ops
3128 = inet_csk(sk)->icsk_ca_ops;
3130 if (unlikely(icsk->icsk_mtup.probe_size &&
3131 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3132 tcp_mtup_probe_success(sk);
3135 tcp_ack_update_rtt(sk, flag, seq_rtt);
3138 if (tcp_is_reno(tp)) {
3139 tcp_remove_reno_sacks(sk, pkts_acked);
3143 /* Non-retransmitted hole got filled? That's reordering */
3144 if (reord < prior_fackets)
3145 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3147 delta = tcp_is_fack(tp) ? pkts_acked :
3148 prior_sacked - tp->sacked_out;
3149 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3152 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3154 if (ca_ops->pkts_acked) {
3157 /* Is the ACK triggering packet unambiguous? */
3158 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3159 /* High resolution needed and available? */
3160 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3161 !ktime_equal(last_ackt,
3162 net_invalid_timestamp()))
3163 rtt_us = ktime_us_delta(ktime_get_real(),
3165 else if (ca_seq_rtt >= 0)
3166 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3169 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3173 #if FASTRETRANS_DEBUG > 0
3174 WARN_ON((int)tp->sacked_out < 0);
3175 WARN_ON((int)tp->lost_out < 0);
3176 WARN_ON((int)tp->retrans_out < 0);
3177 if (!tp->packets_out && tcp_is_sack(tp)) {
3178 icsk = inet_csk(sk);
3180 pr_debug("Leak l=%u %d\n",
3181 tp->lost_out, icsk->icsk_ca_state);
3184 if (tp->sacked_out) {
3185 pr_debug("Leak s=%u %d\n",
3186 tp->sacked_out, icsk->icsk_ca_state);
3189 if (tp->retrans_out) {
3190 pr_debug("Leak r=%u %d\n",
3191 tp->retrans_out, icsk->icsk_ca_state);
3192 tp->retrans_out = 0;
3199 static void tcp_ack_probe(struct sock *sk)
3201 const struct tcp_sock *tp = tcp_sk(sk);
3202 struct inet_connection_sock *icsk = inet_csk(sk);
3204 /* Was it a usable window open? */
3206 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3207 icsk->icsk_backoff = 0;
3208 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3209 /* Socket must be waked up by subsequent tcp_data_snd_check().
3210 * This function is not for random using!
3213 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3214 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3219 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3221 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3222 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3225 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3227 const struct tcp_sock *tp = tcp_sk(sk);
3228 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3229 !tcp_in_cwnd_reduction(sk);
3232 /* Check that window update is acceptable.
3233 * The function assumes that snd_una<=ack<=snd_next.
3235 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3236 const u32 ack, const u32 ack_seq,
3239 return after(ack, tp->snd_una) ||
3240 after(ack_seq, tp->snd_wl1) ||
3241 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3244 /* Update our send window.
3246 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3247 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3249 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3252 struct tcp_sock *tp = tcp_sk(sk);
3254 u32 nwin = ntohs(tcp_hdr(skb)->window);
3256 if (likely(!tcp_hdr(skb)->syn))
3257 nwin <<= tp->rx_opt.snd_wscale;
3259 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3260 flag |= FLAG_WIN_UPDATE;
3261 tcp_update_wl(tp, ack_seq);
3263 if (tp->snd_wnd != nwin) {
3266 /* Note, it is the only place, where
3267 * fast path is recovered for sending TCP.
3270 tcp_fast_path_check(sk);
3272 if (nwin > tp->max_window) {
3273 tp->max_window = nwin;
3274 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3284 /* RFC 5961 7 [ACK Throttling] */
3285 static void tcp_send_challenge_ack(struct sock *sk)
3287 /* unprotected vars, we dont care of overwrites */
3288 static u32 challenge_timestamp;
3289 static unsigned int challenge_count;
3290 u32 now = jiffies / HZ;
3292 if (now != challenge_timestamp) {
3293 challenge_timestamp = now;
3294 challenge_count = 0;
3296 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3297 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3302 static void tcp_store_ts_recent(struct tcp_sock *tp)
3304 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3305 tp->rx_opt.ts_recent_stamp = get_seconds();
3308 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3310 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3311 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3312 * extra check below makes sure this can only happen
3313 * for pure ACK frames. -DaveM
3315 * Not only, also it occurs for expired timestamps.
3318 if (tcp_paws_check(&tp->rx_opt, 0))
3319 tcp_store_ts_recent(tp);
3323 /* This routine deals with acks during a TLP episode.
3324 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3326 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3328 struct tcp_sock *tp = tcp_sk(sk);
3329 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3330 !(flag & (FLAG_SND_UNA_ADVANCED |
3331 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3333 /* Mark the end of TLP episode on receiving TLP dupack or when
3334 * ack is after tlp_high_seq.
3336 if (is_tlp_dupack) {
3337 tp->tlp_high_seq = 0;
3341 if (after(ack, tp->tlp_high_seq)) {
3342 tp->tlp_high_seq = 0;
3343 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3344 if (!(flag & FLAG_DSACKING_ACK)) {
3345 tcp_init_cwnd_reduction(sk, true);
3346 tcp_set_ca_state(sk, TCP_CA_CWR);
3347 tcp_end_cwnd_reduction(sk);
3348 tcp_try_keep_open(sk);
3349 NET_INC_STATS_BH(sock_net(sk),
3350 LINUX_MIB_TCPLOSSPROBERECOVERY);
3355 /* This routine deals with incoming acks, but not outgoing ones. */
3356 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3358 struct inet_connection_sock *icsk = inet_csk(sk);
3359 struct tcp_sock *tp = tcp_sk(sk);
3360 u32 prior_snd_una = tp->snd_una;
3361 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3362 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3363 bool is_dupack = false;
3364 u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
3366 int prior_packets = tp->packets_out;
3367 int prior_sacked = tp->sacked_out;
3369 int previous_packets_out = 0;
3371 /* If the ack is older than previous acks
3372 * then we can probably ignore it.
3374 if (before(ack, prior_snd_una)) {
3375 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3376 if (before(ack, prior_snd_una - tp->max_window)) {
3377 tcp_send_challenge_ack(sk);
3383 /* If the ack includes data we haven't sent yet, discard
3384 * this segment (RFC793 Section 3.9).
3386 if (after(ack, tp->snd_nxt))
3389 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3390 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3393 if (after(ack, prior_snd_una))
3394 flag |= FLAG_SND_UNA_ADVANCED;
3396 prior_fackets = tp->fackets_out;
3397 prior_in_flight = tcp_packets_in_flight(tp);
3399 /* ts_recent update must be made after we are sure that the packet
3402 if (flag & FLAG_UPDATE_TS_RECENT)
3403 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3405 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3406 /* Window is constant, pure forward advance.
3407 * No more checks are required.
3408 * Note, we use the fact that SND.UNA>=SND.WL2.
3410 tcp_update_wl(tp, ack_seq);
3412 flag |= FLAG_WIN_UPDATE;
3414 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3416 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3418 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3421 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3423 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3425 if (TCP_SKB_CB(skb)->sacked)
3426 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3428 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3431 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3434 /* We passed data and got it acked, remove any soft error
3435 * log. Something worked...
3437 sk->sk_err_soft = 0;
3438 icsk->icsk_probes_out = 0;
3439 tp->rcv_tstamp = tcp_time_stamp;
3443 /* See if we can take anything off of the retransmit queue. */
3444 previous_packets_out = tp->packets_out;
3445 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3447 pkts_acked = previous_packets_out - tp->packets_out;
3449 if (tcp_ack_is_dubious(sk, flag)) {
3450 /* Advance CWND, if state allows this. */
3451 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
3452 tcp_cong_avoid(sk, ack, prior_in_flight);
3453 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3454 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3455 prior_packets, is_dupack, flag);
3457 if (flag & FLAG_DATA_ACKED)
3458 tcp_cong_avoid(sk, ack, prior_in_flight);
3461 if (tp->tlp_high_seq)
3462 tcp_process_tlp_ack(sk, ack, flag);
3464 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3465 struct dst_entry *dst = __sk_dst_get(sk);
3470 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3471 tcp_schedule_loss_probe(sk);
3472 if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
3473 tcp_update_pacing_rate(sk);
3477 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3478 if (flag & FLAG_DSACKING_ACK)
3479 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3480 prior_packets, is_dupack, flag);
3481 /* If this ack opens up a zero window, clear backoff. It was
3482 * being used to time the probes, and is probably far higher than
3483 * it needs to be for normal retransmission.
3485 if (tcp_send_head(sk))
3488 if (tp->tlp_high_seq)
3489 tcp_process_tlp_ack(sk, ack, flag);
3493 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3497 /* If data was SACKed, tag it and see if we should send more data.
3498 * If data was DSACKed, see if we can undo a cwnd reduction.
3500 if (TCP_SKB_CB(skb)->sacked) {
3501 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3502 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3503 prior_packets, is_dupack, flag);
3506 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3510 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3511 * But, this can also be called on packets in the established flow when
3512 * the fast version below fails.
3514 void tcp_parse_options(const struct sk_buff *skb,
3515 struct tcp_options_received *opt_rx, int estab,
3516 struct tcp_fastopen_cookie *foc)
3518 const unsigned char *ptr;
3519 const struct tcphdr *th = tcp_hdr(skb);
3520 int length = (th->doff * 4) - sizeof(struct tcphdr);
3522 ptr = (const unsigned char *)(th + 1);
3523 opt_rx->saw_tstamp = 0;
3525 while (length > 0) {
3526 int opcode = *ptr++;
3532 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3537 if (opsize < 2) /* "silly options" */
3539 if (opsize > length)
3540 return; /* don't parse partial options */
3543 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3544 u16 in_mss = get_unaligned_be16(ptr);
3546 if (opt_rx->user_mss &&
3547 opt_rx->user_mss < in_mss)
3548 in_mss = opt_rx->user_mss;
3549 opt_rx->mss_clamp = in_mss;
3554 if (opsize == TCPOLEN_WINDOW && th->syn &&
3555 !estab && sysctl_tcp_window_scaling) {
3556 __u8 snd_wscale = *(__u8 *)ptr;
3557 opt_rx->wscale_ok = 1;
3558 if (snd_wscale > 14) {
3559 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3564 opt_rx->snd_wscale = snd_wscale;
3567 case TCPOPT_TIMESTAMP:
3568 if ((opsize == TCPOLEN_TIMESTAMP) &&
3569 ((estab && opt_rx->tstamp_ok) ||
3570 (!estab && sysctl_tcp_timestamps))) {
3571 opt_rx->saw_tstamp = 1;
3572 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3573 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3576 case TCPOPT_SACK_PERM:
3577 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3578 !estab && sysctl_tcp_sack) {
3579 opt_rx->sack_ok = TCP_SACK_SEEN;
3580 tcp_sack_reset(opt_rx);
3585 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3586 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3588 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3591 #ifdef CONFIG_TCP_MD5SIG
3594 * The MD5 Hash has already been
3595 * checked (see tcp_v{4,6}_do_rcv()).
3600 /* Fast Open option shares code 254 using a
3601 * 16 bits magic number. It's valid only in
3602 * SYN or SYN-ACK with an even size.
3604 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3605 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3606 foc == NULL || !th->syn || (opsize & 1))
3608 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3609 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3610 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3611 memcpy(foc->val, ptr + 2, foc->len);
3612 else if (foc->len != 0)
3622 EXPORT_SYMBOL(tcp_parse_options);
3624 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3626 const __be32 *ptr = (const __be32 *)(th + 1);
3628 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3629 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3630 tp->rx_opt.saw_tstamp = 1;
3632 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3635 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3637 tp->rx_opt.rcv_tsecr = 0;
3643 /* Fast parse options. This hopes to only see timestamps.
3644 * If it is wrong it falls back on tcp_parse_options().
3646 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3647 const struct tcphdr *th, struct tcp_sock *tp)
3649 /* In the spirit of fast parsing, compare doff directly to constant
3650 * values. Because equality is used, short doff can be ignored here.
3652 if (th->doff == (sizeof(*th) / 4)) {
3653 tp->rx_opt.saw_tstamp = 0;
3655 } else if (tp->rx_opt.tstamp_ok &&
3656 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3657 if (tcp_parse_aligned_timestamp(tp, th))
3661 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3662 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3663 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3668 #ifdef CONFIG_TCP_MD5SIG
3670 * Parse MD5 Signature option
3672 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3674 int length = (th->doff << 2) - sizeof(*th);
3675 const u8 *ptr = (const u8 *)(th + 1);
3677 /* If the TCP option is too short, we can short cut */
3678 if (length < TCPOLEN_MD5SIG)
3681 while (length > 0) {
3682 int opcode = *ptr++;
3693 if (opsize < 2 || opsize > length)
3695 if (opcode == TCPOPT_MD5SIG)
3696 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3703 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3706 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3708 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3709 * it can pass through stack. So, the following predicate verifies that
3710 * this segment is not used for anything but congestion avoidance or
3711 * fast retransmit. Moreover, we even are able to eliminate most of such
3712 * second order effects, if we apply some small "replay" window (~RTO)
3713 * to timestamp space.
3715 * All these measures still do not guarantee that we reject wrapped ACKs
3716 * on networks with high bandwidth, when sequence space is recycled fastly,
3717 * but it guarantees that such events will be very rare and do not affect
3718 * connection seriously. This doesn't look nice, but alas, PAWS is really
3721 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3722 * states that events when retransmit arrives after original data are rare.
3723 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3724 * the biggest problem on large power networks even with minor reordering.
3725 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3726 * up to bandwidth of 18Gigabit/sec. 8) ]
3729 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3731 const struct tcp_sock *tp = tcp_sk(sk);
3732 const struct tcphdr *th = tcp_hdr(skb);
3733 u32 seq = TCP_SKB_CB(skb)->seq;
3734 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3736 return (/* 1. Pure ACK with correct sequence number. */
3737 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3739 /* 2. ... and duplicate ACK. */
3740 ack == tp->snd_una &&
3742 /* 3. ... and does not update window. */
3743 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3745 /* 4. ... and sits in replay window. */
3746 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3749 static inline bool tcp_paws_discard(const struct sock *sk,
3750 const struct sk_buff *skb)
3752 const struct tcp_sock *tp = tcp_sk(sk);
3754 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3755 !tcp_disordered_ack(sk, skb);
3758 /* Check segment sequence number for validity.
3760 * Segment controls are considered valid, if the segment
3761 * fits to the window after truncation to the window. Acceptability
3762 * of data (and SYN, FIN, of course) is checked separately.
3763 * See tcp_data_queue(), for example.
3765 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3766 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3767 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3768 * (borrowed from freebsd)
3771 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3773 return !before(end_seq, tp->rcv_wup) &&
3774 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3777 /* When we get a reset we do this. */
3778 void tcp_reset(struct sock *sk)
3780 /* We want the right error as BSD sees it (and indeed as we do). */
3781 switch (sk->sk_state) {
3783 sk->sk_err = ECONNREFUSED;
3785 case TCP_CLOSE_WAIT:
3791 sk->sk_err = ECONNRESET;
3793 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3796 if (!sock_flag(sk, SOCK_DEAD))
3797 sk->sk_error_report(sk);
3803 * Process the FIN bit. This now behaves as it is supposed to work
3804 * and the FIN takes effect when it is validly part of sequence
3805 * space. Not before when we get holes.
3807 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3808 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3811 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3812 * close and we go into CLOSING (and later onto TIME-WAIT)
3814 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3816 static void tcp_fin(struct sock *sk)
3818 struct tcp_sock *tp = tcp_sk(sk);
3820 inet_csk_schedule_ack(sk);
3822 sk->sk_shutdown |= RCV_SHUTDOWN;
3823 sock_set_flag(sk, SOCK_DONE);
3825 switch (sk->sk_state) {
3827 case TCP_ESTABLISHED:
3828 /* Move to CLOSE_WAIT */
3829 tcp_set_state(sk, TCP_CLOSE_WAIT);
3830 inet_csk(sk)->icsk_ack.pingpong = 1;
3833 case TCP_CLOSE_WAIT:
3835 /* Received a retransmission of the FIN, do
3840 /* RFC793: Remain in the LAST-ACK state. */
3844 /* This case occurs when a simultaneous close
3845 * happens, we must ack the received FIN and
3846 * enter the CLOSING state.
3849 tcp_set_state(sk, TCP_CLOSING);
3852 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3854 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3857 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3858 * cases we should never reach this piece of code.
3860 pr_err("%s: Impossible, sk->sk_state=%d\n",
3861 __func__, sk->sk_state);
3865 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3866 * Probably, we should reset in this case. For now drop them.
3868 __skb_queue_purge(&tp->out_of_order_queue);
3869 if (tcp_is_sack(tp))
3870 tcp_sack_reset(&tp->rx_opt);
3873 if (!sock_flag(sk, SOCK_DEAD)) {
3874 sk->sk_state_change(sk);
3876 /* Do not send POLL_HUP for half duplex close. */
3877 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3878 sk->sk_state == TCP_CLOSE)
3879 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3881 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3885 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3888 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3889 if (before(seq, sp->start_seq))
3890 sp->start_seq = seq;
3891 if (after(end_seq, sp->end_seq))
3892 sp->end_seq = end_seq;
3898 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3900 struct tcp_sock *tp = tcp_sk(sk);
3902 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3905 if (before(seq, tp->rcv_nxt))
3906 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3908 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3910 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3912 tp->rx_opt.dsack = 1;
3913 tp->duplicate_sack[0].start_seq = seq;
3914 tp->duplicate_sack[0].end_seq = end_seq;
3918 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3920 struct tcp_sock *tp = tcp_sk(sk);
3922 if (!tp->rx_opt.dsack)
3923 tcp_dsack_set(sk, seq, end_seq);
3925 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3928 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3930 struct tcp_sock *tp = tcp_sk(sk);
3932 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3933 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3934 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3935 tcp_enter_quickack_mode(sk);
3937 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3938 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3940 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3941 end_seq = tp->rcv_nxt;
3942 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3949 /* These routines update the SACK block as out-of-order packets arrive or
3950 * in-order packets close up the sequence space.
3952 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3955 struct tcp_sack_block *sp = &tp->selective_acks[0];
3956 struct tcp_sack_block *swalk = sp + 1;
3958 /* See if the recent change to the first SACK eats into
3959 * or hits the sequence space of other SACK blocks, if so coalesce.
3961 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3962 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3965 /* Zap SWALK, by moving every further SACK up by one slot.
3966 * Decrease num_sacks.
3968 tp->rx_opt.num_sacks--;
3969 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3973 this_sack++, swalk++;
3977 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3979 struct tcp_sock *tp = tcp_sk(sk);
3980 struct tcp_sack_block *sp = &tp->selective_acks[0];
3981 int cur_sacks = tp->rx_opt.num_sacks;
3987 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3988 if (tcp_sack_extend(sp, seq, end_seq)) {
3989 /* Rotate this_sack to the first one. */
3990 for (; this_sack > 0; this_sack--, sp--)
3991 swap(*sp, *(sp - 1));
3993 tcp_sack_maybe_coalesce(tp);
3998 /* Could not find an adjacent existing SACK, build a new one,
3999 * put it at the front, and shift everyone else down. We
4000 * always know there is at least one SACK present already here.
4002 * If the sack array is full, forget about the last one.
4004 if (this_sack >= TCP_NUM_SACKS) {
4006 tp->rx_opt.num_sacks--;
4009 for (; this_sack > 0; this_sack--, sp--)
4013 /* Build the new head SACK, and we're done. */
4014 sp->start_seq = seq;
4015 sp->end_seq = end_seq;
4016 tp->rx_opt.num_sacks++;
4019 /* RCV.NXT advances, some SACKs should be eaten. */
4021 static void tcp_sack_remove(struct tcp_sock *tp)
4023 struct tcp_sack_block *sp = &tp->selective_acks[0];
4024 int num_sacks = tp->rx_opt.num_sacks;
4027 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4028 if (skb_queue_empty(&tp->out_of_order_queue)) {
4029 tp->rx_opt.num_sacks = 0;
4033 for (this_sack = 0; this_sack < num_sacks;) {
4034 /* Check if the start of the sack is covered by RCV.NXT. */
4035 if (!before(tp->rcv_nxt, sp->start_seq)) {
4038 /* RCV.NXT must cover all the block! */
4039 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4041 /* Zap this SACK, by moving forward any other SACKS. */
4042 for (i=this_sack+1; i < num_sacks; i++)
4043 tp->selective_acks[i-1] = tp->selective_acks[i];
4050 tp->rx_opt.num_sacks = num_sacks;
4053 /* This one checks to see if we can put data from the
4054 * out_of_order queue into the receive_queue.
4056 static void tcp_ofo_queue(struct sock *sk)
4058 struct tcp_sock *tp = tcp_sk(sk);
4059 __u32 dsack_high = tp->rcv_nxt;
4060 struct sk_buff *skb;
4062 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4063 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4066 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4067 __u32 dsack = dsack_high;
4068 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4069 dsack_high = TCP_SKB_CB(skb)->end_seq;
4070 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4073 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4074 SOCK_DEBUG(sk, "ofo packet was already received\n");
4075 __skb_unlink(skb, &tp->out_of_order_queue);
4079 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4080 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4081 TCP_SKB_CB(skb)->end_seq);
4083 __skb_unlink(skb, &tp->out_of_order_queue);
4084 __skb_queue_tail(&sk->sk_receive_queue, skb);
4085 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4086 if (tcp_hdr(skb)->fin)
4091 static bool tcp_prune_ofo_queue(struct sock *sk);
4092 static int tcp_prune_queue(struct sock *sk);
4094 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4097 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4098 !sk_rmem_schedule(sk, skb, size)) {
4100 if (tcp_prune_queue(sk) < 0)
4103 if (!sk_rmem_schedule(sk, skb, size)) {
4104 if (!tcp_prune_ofo_queue(sk))
4107 if (!sk_rmem_schedule(sk, skb, size))
4115 * tcp_try_coalesce - try to merge skb to prior one
4118 * @from: buffer to add in queue
4119 * @fragstolen: pointer to boolean
4121 * Before queueing skb @from after @to, try to merge them
4122 * to reduce overall memory use and queue lengths, if cost is small.
4123 * Packets in ofo or receive queues can stay a long time.
4124 * Better try to coalesce them right now to avoid future collapses.
4125 * Returns true if caller should free @from instead of queueing it
4127 static bool tcp_try_coalesce(struct sock *sk,
4129 struct sk_buff *from,
4134 *fragstolen = false;
4136 if (tcp_hdr(from)->fin)
4139 /* Its possible this segment overlaps with prior segment in queue */
4140 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4143 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4146 atomic_add(delta, &sk->sk_rmem_alloc);
4147 sk_mem_charge(sk, delta);
4148 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4149 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4150 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4154 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4156 struct tcp_sock *tp = tcp_sk(sk);
4157 struct sk_buff *skb1;
4160 TCP_ECN_check_ce(tp, skb);
4162 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4163 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4168 /* Disable header prediction. */
4170 inet_csk_schedule_ack(sk);
4172 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4173 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4174 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4176 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4178 /* Initial out of order segment, build 1 SACK. */
4179 if (tcp_is_sack(tp)) {
4180 tp->rx_opt.num_sacks = 1;
4181 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4182 tp->selective_acks[0].end_seq =
4183 TCP_SKB_CB(skb)->end_seq;
4185 __skb_queue_head(&tp->out_of_order_queue, skb);
4189 seq = TCP_SKB_CB(skb)->seq;
4190 end_seq = TCP_SKB_CB(skb)->end_seq;
4192 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4195 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4196 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4198 kfree_skb_partial(skb, fragstolen);
4202 if (!tp->rx_opt.num_sacks ||
4203 tp->selective_acks[0].end_seq != seq)
4206 /* Common case: data arrive in order after hole. */
4207 tp->selective_acks[0].end_seq = end_seq;
4211 /* Find place to insert this segment. */
4213 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4215 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4219 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4222 /* Do skb overlap to previous one? */
4223 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4224 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4225 /* All the bits are present. Drop. */
4226 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4229 tcp_dsack_set(sk, seq, end_seq);
4232 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4233 /* Partial overlap. */
4234 tcp_dsack_set(sk, seq,
4235 TCP_SKB_CB(skb1)->end_seq);
4237 if (skb_queue_is_first(&tp->out_of_order_queue,
4241 skb1 = skb_queue_prev(
4242 &tp->out_of_order_queue,
4247 __skb_queue_head(&tp->out_of_order_queue, skb);
4249 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4251 /* And clean segments covered by new one as whole. */
4252 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4253 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4255 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4257 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4258 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4262 __skb_unlink(skb1, &tp->out_of_order_queue);
4263 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4264 TCP_SKB_CB(skb1)->end_seq);
4265 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4270 if (tcp_is_sack(tp))
4271 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4274 skb_set_owner_r(skb, sk);
4277 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4281 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4283 __skb_pull(skb, hdrlen);
4285 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4286 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4288 __skb_queue_tail(&sk->sk_receive_queue, skb);
4289 skb_set_owner_r(skb, sk);
4294 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4296 struct sk_buff *skb = NULL;
4303 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4307 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4310 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4311 skb_reset_transport_header(skb);
4312 memset(th, 0, sizeof(*th));
4314 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4317 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4318 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4319 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4321 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4322 WARN_ON_ONCE(fragstolen); /* should not happen */
4333 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4335 const struct tcphdr *th = tcp_hdr(skb);
4336 struct tcp_sock *tp = tcp_sk(sk);
4338 bool fragstolen = false;
4340 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4344 __skb_pull(skb, th->doff * 4);
4346 TCP_ECN_accept_cwr(tp, skb);
4348 tp->rx_opt.dsack = 0;
4350 /* Queue data for delivery to the user.
4351 * Packets in sequence go to the receive queue.
4352 * Out of sequence packets to the out_of_order_queue.
4354 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4355 if (tcp_receive_window(tp) == 0)
4358 /* Ok. In sequence. In window. */
4359 if (tp->ucopy.task == current &&
4360 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4361 sock_owned_by_user(sk) && !tp->urg_data) {
4362 int chunk = min_t(unsigned int, skb->len,
4365 __set_current_state(TASK_RUNNING);
4368 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4369 tp->ucopy.len -= chunk;
4370 tp->copied_seq += chunk;
4371 eaten = (chunk == skb->len);
4372 tcp_rcv_space_adjust(sk);
4380 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4383 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4385 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4387 tcp_event_data_recv(sk, skb);
4391 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4394 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4395 * gap in queue is filled.
4397 if (skb_queue_empty(&tp->out_of_order_queue))
4398 inet_csk(sk)->icsk_ack.pingpong = 0;
4401 if (tp->rx_opt.num_sacks)
4402 tcp_sack_remove(tp);
4404 tcp_fast_path_check(sk);
4407 kfree_skb_partial(skb, fragstolen);
4408 if (!sock_flag(sk, SOCK_DEAD))
4409 sk->sk_data_ready(sk, 0);
4413 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4414 /* A retransmit, 2nd most common case. Force an immediate ack. */
4415 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4416 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4419 tcp_enter_quickack_mode(sk);
4420 inet_csk_schedule_ack(sk);
4426 /* Out of window. F.e. zero window probe. */
4427 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4430 tcp_enter_quickack_mode(sk);
4432 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4433 /* Partial packet, seq < rcv_next < end_seq */
4434 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4435 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4436 TCP_SKB_CB(skb)->end_seq);
4438 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4440 /* If window is closed, drop tail of packet. But after
4441 * remembering D-SACK for its head made in previous line.
4443 if (!tcp_receive_window(tp))
4448 tcp_data_queue_ofo(sk, skb);
4451 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4452 struct sk_buff_head *list)
4454 struct sk_buff *next = NULL;
4456 if (!skb_queue_is_last(list, skb))
4457 next = skb_queue_next(list, skb);
4459 __skb_unlink(skb, list);
4461 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4466 /* Collapse contiguous sequence of skbs head..tail with
4467 * sequence numbers start..end.
4469 * If tail is NULL, this means until the end of the list.
4471 * Segments with FIN/SYN are not collapsed (only because this
4475 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4476 struct sk_buff *head, struct sk_buff *tail,
4479 struct sk_buff *skb, *n;
4482 /* First, check that queue is collapsible and find
4483 * the point where collapsing can be useful. */
4487 skb_queue_walk_from_safe(list, skb, n) {
4490 /* No new bits? It is possible on ofo queue. */
4491 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4492 skb = tcp_collapse_one(sk, skb, list);
4498 /* The first skb to collapse is:
4500 * - bloated or contains data before "start" or
4501 * overlaps to the next one.
4503 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4504 (tcp_win_from_space(skb->truesize) > skb->len ||
4505 before(TCP_SKB_CB(skb)->seq, start))) {
4506 end_of_skbs = false;
4510 if (!skb_queue_is_last(list, skb)) {
4511 struct sk_buff *next = skb_queue_next(list, skb);
4513 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4514 end_of_skbs = false;
4519 /* Decided to skip this, advance start seq. */
4520 start = TCP_SKB_CB(skb)->end_seq;
4522 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4525 while (before(start, end)) {
4526 struct sk_buff *nskb;
4527 unsigned int header = skb_headroom(skb);
4528 int copy = SKB_MAX_ORDER(header, 0);
4530 /* Too big header? This can happen with IPv6. */
4533 if (end - start < copy)
4535 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4539 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4540 skb_set_network_header(nskb, (skb_network_header(skb) -
4542 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4544 skb_reserve(nskb, header);
4545 memcpy(nskb->head, skb->head, header);
4546 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4547 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4548 __skb_queue_before(list, skb, nskb);
4549 skb_set_owner_r(nskb, sk);
4551 /* Copy data, releasing collapsed skbs. */
4553 int offset = start - TCP_SKB_CB(skb)->seq;
4554 int size = TCP_SKB_CB(skb)->end_seq - start;
4558 size = min(copy, size);
4559 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4561 TCP_SKB_CB(nskb)->end_seq += size;
4565 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4566 skb = tcp_collapse_one(sk, skb, list);
4569 tcp_hdr(skb)->syn ||
4577 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4578 * and tcp_collapse() them until all the queue is collapsed.
4580 static void tcp_collapse_ofo_queue(struct sock *sk)
4582 struct tcp_sock *tp = tcp_sk(sk);
4583 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4584 struct sk_buff *head;
4590 start = TCP_SKB_CB(skb)->seq;
4591 end = TCP_SKB_CB(skb)->end_seq;
4595 struct sk_buff *next = NULL;
4597 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4598 next = skb_queue_next(&tp->out_of_order_queue, skb);
4601 /* Segment is terminated when we see gap or when
4602 * we are at the end of all the queue. */
4604 after(TCP_SKB_CB(skb)->seq, end) ||
4605 before(TCP_SKB_CB(skb)->end_seq, start)) {
4606 tcp_collapse(sk, &tp->out_of_order_queue,
4607 head, skb, start, end);
4611 /* Start new segment */
4612 start = TCP_SKB_CB(skb)->seq;
4613 end = TCP_SKB_CB(skb)->end_seq;
4615 if (before(TCP_SKB_CB(skb)->seq, start))
4616 start = TCP_SKB_CB(skb)->seq;
4617 if (after(TCP_SKB_CB(skb)->end_seq, end))
4618 end = TCP_SKB_CB(skb)->end_seq;
4624 * Purge the out-of-order queue.
4625 * Return true if queue was pruned.
4627 static bool tcp_prune_ofo_queue(struct sock *sk)
4629 struct tcp_sock *tp = tcp_sk(sk);
4632 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4633 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4634 __skb_queue_purge(&tp->out_of_order_queue);
4636 /* Reset SACK state. A conforming SACK implementation will
4637 * do the same at a timeout based retransmit. When a connection
4638 * is in a sad state like this, we care only about integrity
4639 * of the connection not performance.
4641 if (tp->rx_opt.sack_ok)
4642 tcp_sack_reset(&tp->rx_opt);
4649 /* Reduce allocated memory if we can, trying to get
4650 * the socket within its memory limits again.
4652 * Return less than zero if we should start dropping frames
4653 * until the socket owning process reads some of the data
4654 * to stabilize the situation.
4656 static int tcp_prune_queue(struct sock *sk)
4658 struct tcp_sock *tp = tcp_sk(sk);
4660 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4662 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4664 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4665 tcp_clamp_window(sk);
4666 else if (sk_under_memory_pressure(sk))
4667 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4669 tcp_collapse_ofo_queue(sk);
4670 if (!skb_queue_empty(&sk->sk_receive_queue))
4671 tcp_collapse(sk, &sk->sk_receive_queue,
4672 skb_peek(&sk->sk_receive_queue),
4674 tp->copied_seq, tp->rcv_nxt);
4677 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4680 /* Collapsing did not help, destructive actions follow.
4681 * This must not ever occur. */
4683 tcp_prune_ofo_queue(sk);
4685 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4688 /* If we are really being abused, tell the caller to silently
4689 * drop receive data on the floor. It will get retransmitted
4690 * and hopefully then we'll have sufficient space.
4692 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4694 /* Massive buffer overcommit. */
4699 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4700 * As additional protections, we do not touch cwnd in retransmission phases,
4701 * and if application hit its sndbuf limit recently.
4703 void tcp_cwnd_application_limited(struct sock *sk)
4705 struct tcp_sock *tp = tcp_sk(sk);
4707 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4708 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4709 /* Limited by application or receiver window. */
4710 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4711 u32 win_used = max(tp->snd_cwnd_used, init_win);
4712 if (win_used < tp->snd_cwnd) {
4713 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4714 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4716 tp->snd_cwnd_used = 0;
4718 tp->snd_cwnd_stamp = tcp_time_stamp;
4721 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4723 const struct tcp_sock *tp = tcp_sk(sk);
4725 /* If the user specified a specific send buffer setting, do
4728 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4731 /* If we are under global TCP memory pressure, do not expand. */
4732 if (sk_under_memory_pressure(sk))
4735 /* If we are under soft global TCP memory pressure, do not expand. */
4736 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4739 /* If we filled the congestion window, do not expand. */
4740 if (tp->packets_out >= tp->snd_cwnd)
4746 /* When incoming ACK allowed to free some skb from write_queue,
4747 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4748 * on the exit from tcp input handler.
4750 * PROBLEM: sndbuf expansion does not work well with largesend.
4752 static void tcp_new_space(struct sock *sk)
4754 struct tcp_sock *tp = tcp_sk(sk);
4756 if (tcp_should_expand_sndbuf(sk)) {
4757 int sndmem = SKB_TRUESIZE(max_t(u32,
4758 tp->rx_opt.mss_clamp,
4761 int demanded = max_t(unsigned int, tp->snd_cwnd,
4762 tp->reordering + 1);
4763 sndmem *= 2 * demanded;
4764 if (sndmem > sk->sk_sndbuf)
4765 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4766 tp->snd_cwnd_stamp = tcp_time_stamp;
4769 sk->sk_write_space(sk);
4772 static void tcp_check_space(struct sock *sk)
4774 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4775 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4776 if (sk->sk_socket &&
4777 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4782 static inline void tcp_data_snd_check(struct sock *sk)
4784 tcp_push_pending_frames(sk);
4785 tcp_check_space(sk);
4789 * Check if sending an ack is needed.
4791 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4793 struct tcp_sock *tp = tcp_sk(sk);
4795 /* More than one full frame received... */
4796 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4797 /* ... and right edge of window advances far enough.
4798 * (tcp_recvmsg() will send ACK otherwise). Or...
4800 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4801 /* We ACK each frame or... */
4802 tcp_in_quickack_mode(sk) ||
4803 /* We have out of order data. */
4804 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4805 /* Then ack it now */
4808 /* Else, send delayed ack. */
4809 tcp_send_delayed_ack(sk);
4813 static inline void tcp_ack_snd_check(struct sock *sk)
4815 if (!inet_csk_ack_scheduled(sk)) {
4816 /* We sent a data segment already. */
4819 __tcp_ack_snd_check(sk, 1);
4823 * This routine is only called when we have urgent data
4824 * signaled. Its the 'slow' part of tcp_urg. It could be
4825 * moved inline now as tcp_urg is only called from one
4826 * place. We handle URGent data wrong. We have to - as
4827 * BSD still doesn't use the correction from RFC961.
4828 * For 1003.1g we should support a new option TCP_STDURG to permit
4829 * either form (or just set the sysctl tcp_stdurg).
4832 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4834 struct tcp_sock *tp = tcp_sk(sk);
4835 u32 ptr = ntohs(th->urg_ptr);
4837 if (ptr && !sysctl_tcp_stdurg)
4839 ptr += ntohl(th->seq);
4841 /* Ignore urgent data that we've already seen and read. */
4842 if (after(tp->copied_seq, ptr))
4845 /* Do not replay urg ptr.
4847 * NOTE: interesting situation not covered by specs.
4848 * Misbehaving sender may send urg ptr, pointing to segment,
4849 * which we already have in ofo queue. We are not able to fetch
4850 * such data and will stay in TCP_URG_NOTYET until will be eaten
4851 * by recvmsg(). Seems, we are not obliged to handle such wicked
4852 * situations. But it is worth to think about possibility of some
4853 * DoSes using some hypothetical application level deadlock.
4855 if (before(ptr, tp->rcv_nxt))
4858 /* Do we already have a newer (or duplicate) urgent pointer? */
4859 if (tp->urg_data && !after(ptr, tp->urg_seq))
4862 /* Tell the world about our new urgent pointer. */
4865 /* We may be adding urgent data when the last byte read was
4866 * urgent. To do this requires some care. We cannot just ignore
4867 * tp->copied_seq since we would read the last urgent byte again
4868 * as data, nor can we alter copied_seq until this data arrives
4869 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4871 * NOTE. Double Dutch. Rendering to plain English: author of comment
4872 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4873 * and expect that both A and B disappear from stream. This is _wrong_.
4874 * Though this happens in BSD with high probability, this is occasional.
4875 * Any application relying on this is buggy. Note also, that fix "works"
4876 * only in this artificial test. Insert some normal data between A and B and we will
4877 * decline of BSD again. Verdict: it is better to remove to trap
4880 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4881 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4882 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4884 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4885 __skb_unlink(skb, &sk->sk_receive_queue);
4890 tp->urg_data = TCP_URG_NOTYET;
4893 /* Disable header prediction. */
4897 /* This is the 'fast' part of urgent handling. */
4898 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4900 struct tcp_sock *tp = tcp_sk(sk);
4902 /* Check if we get a new urgent pointer - normally not. */
4904 tcp_check_urg(sk, th);
4906 /* Do we wait for any urgent data? - normally not... */
4907 if (tp->urg_data == TCP_URG_NOTYET) {
4908 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4911 /* Is the urgent pointer pointing into this packet? */
4912 if (ptr < skb->len) {
4914 if (skb_copy_bits(skb, ptr, &tmp, 1))
4916 tp->urg_data = TCP_URG_VALID | tmp;
4917 if (!sock_flag(sk, SOCK_DEAD))
4918 sk->sk_data_ready(sk, 0);
4923 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4925 struct tcp_sock *tp = tcp_sk(sk);
4926 int chunk = skb->len - hlen;
4930 if (skb_csum_unnecessary(skb))
4931 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4933 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4937 tp->ucopy.len -= chunk;
4938 tp->copied_seq += chunk;
4939 tcp_rcv_space_adjust(sk);
4946 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4947 struct sk_buff *skb)
4951 if (sock_owned_by_user(sk)) {
4953 result = __tcp_checksum_complete(skb);
4956 result = __tcp_checksum_complete(skb);
4961 static inline bool tcp_checksum_complete_user(struct sock *sk,
4962 struct sk_buff *skb)
4964 return !skb_csum_unnecessary(skb) &&
4965 __tcp_checksum_complete_user(sk, skb);
4968 #ifdef CONFIG_NET_DMA
4969 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4972 struct tcp_sock *tp = tcp_sk(sk);
4973 int chunk = skb->len - hlen;
4975 bool copied_early = false;
4977 if (tp->ucopy.wakeup)
4980 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4981 tp->ucopy.dma_chan = net_dma_find_channel();
4983 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4985 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4987 tp->ucopy.iov, chunk,
4988 tp->ucopy.pinned_list);
4993 tp->ucopy.dma_cookie = dma_cookie;
4994 copied_early = true;
4996 tp->ucopy.len -= chunk;
4997 tp->copied_seq += chunk;
4998 tcp_rcv_space_adjust(sk);
5000 if ((tp->ucopy.len == 0) ||
5001 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5002 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5003 tp->ucopy.wakeup = 1;
5004 sk->sk_data_ready(sk, 0);
5006 } else if (chunk > 0) {
5007 tp->ucopy.wakeup = 1;
5008 sk->sk_data_ready(sk, 0);
5011 return copied_early;
5013 #endif /* CONFIG_NET_DMA */
5015 /* Does PAWS and seqno based validation of an incoming segment, flags will
5016 * play significant role here.
5018 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5019 const struct tcphdr *th, int syn_inerr)
5021 struct tcp_sock *tp = tcp_sk(sk);
5023 /* RFC1323: H1. Apply PAWS check first. */
5024 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5025 tcp_paws_discard(sk, skb)) {
5027 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5028 tcp_send_dupack(sk, skb);
5031 /* Reset is accepted even if it did not pass PAWS. */
5034 /* Step 1: check sequence number */
5035 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5036 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5037 * (RST) segments are validated by checking their SEQ-fields."
5038 * And page 69: "If an incoming segment is not acceptable,
5039 * an acknowledgment should be sent in reply (unless the RST
5040 * bit is set, if so drop the segment and return)".
5045 tcp_send_dupack(sk, skb);
5050 /* Step 2: check RST bit */
5053 * If sequence number exactly matches RCV.NXT, then
5054 * RESET the connection
5056 * Send a challenge ACK
5058 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5061 tcp_send_challenge_ack(sk);
5065 /* step 3: check security and precedence [ignored] */
5067 /* step 4: Check for a SYN
5068 * RFC 5691 4.2 : Send a challenge ack
5073 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5074 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5075 tcp_send_challenge_ack(sk);
5087 * TCP receive function for the ESTABLISHED state.
5089 * It is split into a fast path and a slow path. The fast path is
5091 * - A zero window was announced from us - zero window probing
5092 * is only handled properly in the slow path.
5093 * - Out of order segments arrived.
5094 * - Urgent data is expected.
5095 * - There is no buffer space left
5096 * - Unexpected TCP flags/window values/header lengths are received
5097 * (detected by checking the TCP header against pred_flags)
5098 * - Data is sent in both directions. Fast path only supports pure senders
5099 * or pure receivers (this means either the sequence number or the ack
5100 * value must stay constant)
5101 * - Unexpected TCP option.
5103 * When these conditions are not satisfied it drops into a standard
5104 * receive procedure patterned after RFC793 to handle all cases.
5105 * The first three cases are guaranteed by proper pred_flags setting,
5106 * the rest is checked inline. Fast processing is turned on in
5107 * tcp_data_queue when everything is OK.
5109 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5110 const struct tcphdr *th, unsigned int len)
5112 struct tcp_sock *tp = tcp_sk(sk);
5114 if (unlikely(sk->sk_rx_dst == NULL))
5115 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5117 * Header prediction.
5118 * The code loosely follows the one in the famous
5119 * "30 instruction TCP receive" Van Jacobson mail.
5121 * Van's trick is to deposit buffers into socket queue
5122 * on a device interrupt, to call tcp_recv function
5123 * on the receive process context and checksum and copy
5124 * the buffer to user space. smart...
5126 * Our current scheme is not silly either but we take the
5127 * extra cost of the net_bh soft interrupt processing...
5128 * We do checksum and copy also but from device to kernel.
5131 tp->rx_opt.saw_tstamp = 0;
5133 /* pred_flags is 0xS?10 << 16 + snd_wnd
5134 * if header_prediction is to be made
5135 * 'S' will always be tp->tcp_header_len >> 2
5136 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5137 * turn it off (when there are holes in the receive
5138 * space for instance)
5139 * PSH flag is ignored.
5142 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5143 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5144 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5145 int tcp_header_len = tp->tcp_header_len;
5147 /* Timestamp header prediction: tcp_header_len
5148 * is automatically equal to th->doff*4 due to pred_flags
5152 /* Check timestamp */
5153 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5154 /* No? Slow path! */
5155 if (!tcp_parse_aligned_timestamp(tp, th))
5158 /* If PAWS failed, check it more carefully in slow path */
5159 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5162 /* DO NOT update ts_recent here, if checksum fails
5163 * and timestamp was corrupted part, it will result
5164 * in a hung connection since we will drop all
5165 * future packets due to the PAWS test.
5169 if (len <= tcp_header_len) {
5170 /* Bulk data transfer: sender */
5171 if (len == tcp_header_len) {
5172 /* Predicted packet is in window by definition.
5173 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5174 * Hence, check seq<=rcv_wup reduces to:
5176 if (tcp_header_len ==
5177 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5178 tp->rcv_nxt == tp->rcv_wup)
5179 tcp_store_ts_recent(tp);
5181 /* We know that such packets are checksummed
5184 tcp_ack(sk, skb, 0);
5186 tcp_data_snd_check(sk);
5188 } else { /* Header too small */
5189 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5194 int copied_early = 0;
5195 bool fragstolen = false;
5197 if (tp->copied_seq == tp->rcv_nxt &&
5198 len - tcp_header_len <= tp->ucopy.len) {
5199 #ifdef CONFIG_NET_DMA
5200 if (tp->ucopy.task == current &&
5201 sock_owned_by_user(sk) &&
5202 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5207 if (tp->ucopy.task == current &&
5208 sock_owned_by_user(sk) && !copied_early) {
5209 __set_current_state(TASK_RUNNING);
5211 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5215 /* Predicted packet is in window by definition.
5216 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5217 * Hence, check seq<=rcv_wup reduces to:
5219 if (tcp_header_len ==
5220 (sizeof(struct tcphdr) +
5221 TCPOLEN_TSTAMP_ALIGNED) &&
5222 tp->rcv_nxt == tp->rcv_wup)
5223 tcp_store_ts_recent(tp);
5225 tcp_rcv_rtt_measure_ts(sk, skb);
5227 __skb_pull(skb, tcp_header_len);
5228 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5229 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5232 tcp_cleanup_rbuf(sk, skb->len);
5235 if (tcp_checksum_complete_user(sk, skb))
5238 if ((int)skb->truesize > sk->sk_forward_alloc)
5241 /* Predicted packet is in window by definition.
5242 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5243 * Hence, check seq<=rcv_wup reduces to:
5245 if (tcp_header_len ==
5246 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5247 tp->rcv_nxt == tp->rcv_wup)
5248 tcp_store_ts_recent(tp);
5250 tcp_rcv_rtt_measure_ts(sk, skb);
5252 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5254 /* Bulk data transfer: receiver */
5255 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5259 tcp_event_data_recv(sk, skb);
5261 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5262 /* Well, only one small jumplet in fast path... */
5263 tcp_ack(sk, skb, FLAG_DATA);
5264 tcp_data_snd_check(sk);
5265 if (!inet_csk_ack_scheduled(sk))
5269 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5270 __tcp_ack_snd_check(sk, 0);
5272 #ifdef CONFIG_NET_DMA
5274 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5278 kfree_skb_partial(skb, fragstolen);
5279 sk->sk_data_ready(sk, 0);
5285 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5288 if (!th->ack && !th->rst)
5292 * Standard slow path.
5295 if (!tcp_validate_incoming(sk, skb, th, 1))
5299 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5302 tcp_rcv_rtt_measure_ts(sk, skb);
5304 /* Process urgent data. */
5305 tcp_urg(sk, skb, th);
5307 /* step 7: process the segment text */
5308 tcp_data_queue(sk, skb);
5310 tcp_data_snd_check(sk);
5311 tcp_ack_snd_check(sk);
5315 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5316 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5322 EXPORT_SYMBOL(tcp_rcv_established);
5324 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5326 struct tcp_sock *tp = tcp_sk(sk);
5327 struct inet_connection_sock *icsk = inet_csk(sk);
5329 tcp_set_state(sk, TCP_ESTABLISHED);
5332 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5333 security_inet_conn_established(sk, skb);
5336 /* Make sure socket is routed, for correct metrics. */
5337 icsk->icsk_af_ops->rebuild_header(sk);
5339 tcp_init_metrics(sk);
5341 tcp_init_congestion_control(sk);
5343 /* Prevent spurious tcp_cwnd_restart() on first data
5346 tp->lsndtime = tcp_time_stamp;
5348 tcp_init_buffer_space(sk);
5350 if (sock_flag(sk, SOCK_KEEPOPEN))
5351 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5353 if (!tp->rx_opt.snd_wscale)
5354 __tcp_fast_path_on(tp, tp->snd_wnd);
5358 if (!sock_flag(sk, SOCK_DEAD)) {
5359 sk->sk_state_change(sk);
5360 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5364 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5365 struct tcp_fastopen_cookie *cookie)
5367 struct tcp_sock *tp = tcp_sk(sk);
5368 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5369 u16 mss = tp->rx_opt.mss_clamp;
5372 if (mss == tp->rx_opt.user_mss) {
5373 struct tcp_options_received opt;
5375 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5376 tcp_clear_options(&opt);
5377 opt.user_mss = opt.mss_clamp = 0;
5378 tcp_parse_options(synack, &opt, 0, NULL);
5379 mss = opt.mss_clamp;
5382 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5385 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5386 * the remote receives only the retransmitted (regular) SYNs: either
5387 * the original SYN-data or the corresponding SYN-ACK is lost.
5389 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5391 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5393 if (data) { /* Retransmit unacked data in SYN */
5394 tcp_for_write_queue_from(data, sk) {
5395 if (data == tcp_send_head(sk) ||
5396 __tcp_retransmit_skb(sk, data))
5402 tp->syn_data_acked = tp->syn_data;
5406 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5407 const struct tcphdr *th, unsigned int len)
5409 struct inet_connection_sock *icsk = inet_csk(sk);
5410 struct tcp_sock *tp = tcp_sk(sk);
5411 struct tcp_fastopen_cookie foc = { .len = -1 };
5412 int saved_clamp = tp->rx_opt.mss_clamp;
5414 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5415 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5416 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5420 * "If the state is SYN-SENT then
5421 * first check the ACK bit
5422 * If the ACK bit is set
5423 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5424 * a reset (unless the RST bit is set, if so drop
5425 * the segment and return)"
5427 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5428 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5429 goto reset_and_undo;
5431 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5432 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5434 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5435 goto reset_and_undo;
5438 /* Now ACK is acceptable.
5440 * "If the RST bit is set
5441 * If the ACK was acceptable then signal the user "error:
5442 * connection reset", drop the segment, enter CLOSED state,
5443 * delete TCB, and return."
5452 * "fifth, if neither of the SYN or RST bits is set then
5453 * drop the segment and return."
5459 goto discard_and_undo;
5462 * "If the SYN bit is on ...
5463 * are acceptable then ...
5464 * (our SYN has been ACKed), change the connection
5465 * state to ESTABLISHED..."
5468 TCP_ECN_rcv_synack(tp, th);
5470 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5471 tcp_ack(sk, skb, FLAG_SLOWPATH);
5473 /* Ok.. it's good. Set up sequence numbers and
5474 * move to established.
5476 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5477 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5479 /* RFC1323: The window in SYN & SYN/ACK segments is
5482 tp->snd_wnd = ntohs(th->window);
5484 if (!tp->rx_opt.wscale_ok) {
5485 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5486 tp->window_clamp = min(tp->window_clamp, 65535U);
5489 if (tp->rx_opt.saw_tstamp) {
5490 tp->rx_opt.tstamp_ok = 1;
5491 tp->tcp_header_len =
5492 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5493 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5494 tcp_store_ts_recent(tp);
5496 tp->tcp_header_len = sizeof(struct tcphdr);
5499 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5500 tcp_enable_fack(tp);
5503 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5504 tcp_initialize_rcv_mss(sk);
5506 /* Remember, tcp_poll() does not lock socket!
5507 * Change state from SYN-SENT only after copied_seq
5508 * is initialized. */
5509 tp->copied_seq = tp->rcv_nxt;
5513 tcp_finish_connect(sk, skb);
5515 if ((tp->syn_fastopen || tp->syn_data) &&
5516 tcp_rcv_fastopen_synack(sk, skb, &foc))
5519 if (sk->sk_write_pending ||
5520 icsk->icsk_accept_queue.rskq_defer_accept ||
5521 icsk->icsk_ack.pingpong) {
5522 /* Save one ACK. Data will be ready after
5523 * several ticks, if write_pending is set.
5525 * It may be deleted, but with this feature tcpdumps
5526 * look so _wonderfully_ clever, that I was not able
5527 * to stand against the temptation 8) --ANK
5529 inet_csk_schedule_ack(sk);
5530 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5531 tcp_enter_quickack_mode(sk);
5532 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5533 TCP_DELACK_MAX, TCP_RTO_MAX);
5544 /* No ACK in the segment */
5548 * "If the RST bit is set
5550 * Otherwise (no ACK) drop the segment and return."
5553 goto discard_and_undo;
5557 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5558 tcp_paws_reject(&tp->rx_opt, 0))
5559 goto discard_and_undo;
5562 /* We see SYN without ACK. It is attempt of
5563 * simultaneous connect with crossed SYNs.
5564 * Particularly, it can be connect to self.
5566 tcp_set_state(sk, TCP_SYN_RECV);
5568 if (tp->rx_opt.saw_tstamp) {
5569 tp->rx_opt.tstamp_ok = 1;
5570 tcp_store_ts_recent(tp);
5571 tp->tcp_header_len =
5572 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5574 tp->tcp_header_len = sizeof(struct tcphdr);
5577 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5578 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5580 /* RFC1323: The window in SYN & SYN/ACK segments is
5583 tp->snd_wnd = ntohs(th->window);
5584 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5585 tp->max_window = tp->snd_wnd;
5587 TCP_ECN_rcv_syn(tp, th);
5590 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5591 tcp_initialize_rcv_mss(sk);
5593 tcp_send_synack(sk);
5595 /* Note, we could accept data and URG from this segment.
5596 * There are no obstacles to make this (except that we must
5597 * either change tcp_recvmsg() to prevent it from returning data
5598 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5600 * However, if we ignore data in ACKless segments sometimes,
5601 * we have no reasons to accept it sometimes.
5602 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5603 * is not flawless. So, discard packet for sanity.
5604 * Uncomment this return to process the data.
5611 /* "fifth, if neither of the SYN or RST bits is set then
5612 * drop the segment and return."
5616 tcp_clear_options(&tp->rx_opt);
5617 tp->rx_opt.mss_clamp = saved_clamp;
5621 tcp_clear_options(&tp->rx_opt);
5622 tp->rx_opt.mss_clamp = saved_clamp;
5627 * This function implements the receiving procedure of RFC 793 for
5628 * all states except ESTABLISHED and TIME_WAIT.
5629 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5630 * address independent.
5633 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5634 const struct tcphdr *th, unsigned int len)
5636 struct tcp_sock *tp = tcp_sk(sk);
5637 struct inet_connection_sock *icsk = inet_csk(sk);
5638 struct request_sock *req;
5641 tp->rx_opt.saw_tstamp = 0;
5643 switch (sk->sk_state) {
5657 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5660 /* Now we have several options: In theory there is
5661 * nothing else in the frame. KA9Q has an option to
5662 * send data with the syn, BSD accepts data with the
5663 * syn up to the [to be] advertised window and
5664 * Solaris 2.1 gives you a protocol error. For now
5665 * we just ignore it, that fits the spec precisely
5666 * and avoids incompatibilities. It would be nice in
5667 * future to drop through and process the data.
5669 * Now that TTCP is starting to be used we ought to
5671 * But, this leaves one open to an easy denial of
5672 * service attack, and SYN cookies can't defend
5673 * against this problem. So, we drop the data
5674 * in the interest of security over speed unless
5675 * it's still in use.
5683 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5687 /* Do step6 onward by hand. */
5688 tcp_urg(sk, skb, th);
5690 tcp_data_snd_check(sk);
5694 req = tp->fastopen_rsk;
5696 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5697 sk->sk_state != TCP_FIN_WAIT1);
5699 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5703 if (!th->ack && !th->rst)
5706 if (!tcp_validate_incoming(sk, skb, th, 0))
5709 /* step 5: check the ACK field */
5711 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5712 FLAG_UPDATE_TS_RECENT) > 0;
5714 switch (sk->sk_state) {
5717 /* Once we leave TCP_SYN_RECV, we no longer
5718 * need req so release it.
5721 tcp_synack_rtt_meas(sk, req);
5722 tp->total_retrans = req->num_retrans;
5724 reqsk_fastopen_remove(sk, req, false);
5726 /* Make sure socket is routed, for
5729 icsk->icsk_af_ops->rebuild_header(sk);
5730 tcp_init_congestion_control(sk);
5733 tcp_init_buffer_space(sk);
5734 tp->copied_seq = tp->rcv_nxt;
5737 tcp_set_state(sk, TCP_ESTABLISHED);
5738 sk->sk_state_change(sk);
5740 /* Note, that this wakeup is only for marginal
5741 * crossed SYN case. Passively open sockets
5742 * are not waked up, because sk->sk_sleep ==
5743 * NULL and sk->sk_socket == NULL.
5747 SOCK_WAKE_IO, POLL_OUT);
5749 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5750 tp->snd_wnd = ntohs(th->window) <<
5751 tp->rx_opt.snd_wscale;
5752 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5754 if (tp->rx_opt.tstamp_ok)
5755 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5758 /* Re-arm the timer because data may
5759 * have been sent out. This is similar
5760 * to the regular data transmission case
5761 * when new data has just been ack'ed.
5763 * (TFO) - we could try to be more
5764 * aggressive and retranmitting any data
5765 * sooner based on when they were sent
5770 tcp_init_metrics(sk);
5772 tcp_update_pacing_rate(sk);
5774 /* Prevent spurious tcp_cwnd_restart() on
5775 * first data packet.
5777 tp->lsndtime = tcp_time_stamp;
5779 tcp_initialize_rcv_mss(sk);
5780 tcp_fast_path_on(tp);
5787 /* If we enter the TCP_FIN_WAIT1 state and we are a
5788 * Fast Open socket and this is the first acceptable
5789 * ACK we have received, this would have acknowledged
5790 * our SYNACK so stop the SYNACK timer.
5793 /* Return RST if ack_seq is invalid.
5794 * Note that RFC793 only says to generate a
5795 * DUPACK for it but for TCP Fast Open it seems
5796 * better to treat this case like TCP_SYN_RECV
5801 /* We no longer need the request sock. */
5802 reqsk_fastopen_remove(sk, req, false);
5805 if (tp->snd_una == tp->write_seq) {
5806 struct dst_entry *dst;
5808 tcp_set_state(sk, TCP_FIN_WAIT2);
5809 sk->sk_shutdown |= SEND_SHUTDOWN;
5811 dst = __sk_dst_get(sk);
5815 if (!sock_flag(sk, SOCK_DEAD))
5816 /* Wake up lingering close() */
5817 sk->sk_state_change(sk);
5821 if (tp->linger2 < 0 ||
5822 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5823 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5825 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5829 tmo = tcp_fin_time(sk);
5830 if (tmo > TCP_TIMEWAIT_LEN) {
5831 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5832 } else if (th->fin || sock_owned_by_user(sk)) {
5833 /* Bad case. We could lose such FIN otherwise.
5834 * It is not a big problem, but it looks confusing
5835 * and not so rare event. We still can lose it now,
5836 * if it spins in bh_lock_sock(), but it is really
5839 inet_csk_reset_keepalive_timer(sk, tmo);
5841 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5849 if (tp->snd_una == tp->write_seq) {
5850 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5856 if (tp->snd_una == tp->write_seq) {
5857 tcp_update_metrics(sk);
5865 /* step 6: check the URG bit */
5866 tcp_urg(sk, skb, th);
5868 /* step 7: process the segment text */
5869 switch (sk->sk_state) {
5870 case TCP_CLOSE_WAIT:
5873 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5877 /* RFC 793 says to queue data in these states,
5878 * RFC 1122 says we MUST send a reset.
5879 * BSD 4.4 also does reset.
5881 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5882 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5883 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5884 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5890 case TCP_ESTABLISHED:
5891 tcp_data_queue(sk, skb);
5896 /* tcp_data could move socket to TIME-WAIT */
5897 if (sk->sk_state != TCP_CLOSE) {
5898 tcp_data_snd_check(sk);
5899 tcp_ack_snd_check(sk);
5908 EXPORT_SYMBOL(tcp_rcv_state_process);