2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 100;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
97 int sysctl_tcp_thin_dupack __read_mostly;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100 int sysctl_tcp_early_retrans __read_mostly = 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 /* Adapt the MSS value used to make delayed ack decision to the
127 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
129 struct inet_connection_sock *icsk = inet_csk(sk);
130 const unsigned int lss = icsk->icsk_ack.last_seg_size;
133 icsk->icsk_ack.last_seg_size = 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len = skb_shinfo(skb)->gso_size ? : skb->len;
139 if (len >= icsk->icsk_ack.rcv_mss) {
140 icsk->icsk_ack.rcv_mss = len;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len += skb->data - skb_transport_header(skb);
148 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len -= tcp_sk(sk)->tcp_header_len;
161 icsk->icsk_ack.last_seg_size = len;
163 icsk->icsk_ack.rcv_mss = len;
167 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
173 static void tcp_incr_quickack(struct sock *sk)
175 struct inet_connection_sock *icsk = inet_csk(sk);
176 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
180 if (quickacks > icsk->icsk_ack.quick)
181 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
184 static void tcp_enter_quickack_mode(struct sock *sk)
186 struct inet_connection_sock *icsk = inet_csk(sk);
187 tcp_incr_quickack(sk);
188 icsk->icsk_ack.pingpong = 0;
189 icsk->icsk_ack.ato = TCP_ATO_MIN;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline bool tcp_in_quickack_mode(const struct sock *sk)
198 const struct inet_connection_sock *icsk = inet_csk(sk);
200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
205 if (tp->ecn_flags & TCP_ECN_OK)
206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
222 if (!(tp->ecn_flags & TCP_ECN_OK))
225 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
226 case INET_ECN_NOT_ECT:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp->ecn_flags & TCP_ECN_SEEN)
232 tcp_enter_quickack_mode((struct sock *)tp);
235 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock *)tp);
238 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
242 tp->ecn_flags |= TCP_ECN_SEEN;
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
248 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
249 tp->ecn_flags &= ~TCP_ECN_OK;
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
254 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
255 tp->ecn_flags &= ~TCP_ECN_OK;
258 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
260 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
265 /* Buffer size and advertised window tuning.
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270 static void tcp_fixup_sndbuf(struct sock *sk)
272 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
274 sndmem *= TCP_INIT_CWND;
275 if (sk->sk_sndbuf < sndmem)
276 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
279 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
281 * All tcp_full_space() is split to two parts: "network" buffer, allocated
282 * forward and advertised in receiver window (tp->rcv_wnd) and
283 * "application buffer", required to isolate scheduling/application
284 * latencies from network.
285 * window_clamp is maximal advertised window. It can be less than
286 * tcp_full_space(), in this case tcp_full_space() - window_clamp
287 * is reserved for "application" buffer. The less window_clamp is
288 * the smoother our behaviour from viewpoint of network, but the lower
289 * throughput and the higher sensitivity of the connection to losses. 8)
291 * rcv_ssthresh is more strict window_clamp used at "slow start"
292 * phase to predict further behaviour of this connection.
293 * It is used for two goals:
294 * - to enforce header prediction at sender, even when application
295 * requires some significant "application buffer". It is check #1.
296 * - to prevent pruning of receive queue because of misprediction
297 * of receiver window. Check #2.
299 * The scheme does not work when sender sends good segments opening
300 * window and then starts to feed us spaghetti. But it should work
301 * in common situations. Otherwise, we have to rely on queue collapsing.
304 /* Slow part of check#2. */
305 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
307 struct tcp_sock *tp = tcp_sk(sk);
309 int truesize = tcp_win_from_space(skb->truesize) >> 1;
310 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
312 while (tp->rcv_ssthresh <= window) {
313 if (truesize <= skb->len)
314 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
322 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
324 struct tcp_sock *tp = tcp_sk(sk);
327 if (tp->rcv_ssthresh < tp->window_clamp &&
328 (int)tp->rcv_ssthresh < tcp_space(sk) &&
329 !sk_under_memory_pressure(sk)) {
332 /* Check #2. Increase window, if skb with such overhead
333 * will fit to rcvbuf in future.
335 if (tcp_win_from_space(skb->truesize) <= skb->len)
336 incr = 2 * tp->advmss;
338 incr = __tcp_grow_window(sk, skb);
341 incr = max_t(int, incr, 2 * skb->len);
342 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
344 inet_csk(sk)->icsk_ack.quick |= 1;
349 /* 3. Tuning rcvbuf, when connection enters established state. */
351 static void tcp_fixup_rcvbuf(struct sock *sk)
353 u32 mss = tcp_sk(sk)->advmss;
354 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
357 /* Limit to 10 segments if mss <= 1460,
358 * or 14600/mss segments, with a minimum of two segments.
361 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
363 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
364 while (tcp_win_from_space(rcvmem) < mss)
369 if (sk->sk_rcvbuf < rcvmem)
370 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
373 /* 4. Try to fixup all. It is made immediately after connection enters
376 void tcp_init_buffer_space(struct sock *sk)
378 struct tcp_sock *tp = tcp_sk(sk);
381 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
382 tcp_fixup_rcvbuf(sk);
383 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
384 tcp_fixup_sndbuf(sk);
386 tp->rcvq_space.space = tp->rcv_wnd;
388 maxwin = tcp_full_space(sk);
390 if (tp->window_clamp >= maxwin) {
391 tp->window_clamp = maxwin;
393 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
394 tp->window_clamp = max(maxwin -
395 (maxwin >> sysctl_tcp_app_win),
399 /* Force reservation of one segment. */
400 if (sysctl_tcp_app_win &&
401 tp->window_clamp > 2 * tp->advmss &&
402 tp->window_clamp + tp->advmss > maxwin)
403 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
405 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
406 tp->snd_cwnd_stamp = tcp_time_stamp;
409 /* 5. Recalculate window clamp after socket hit its memory bounds. */
410 static void tcp_clamp_window(struct sock *sk)
412 struct tcp_sock *tp = tcp_sk(sk);
413 struct inet_connection_sock *icsk = inet_csk(sk);
415 icsk->icsk_ack.quick = 0;
417 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
418 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
419 !sk_under_memory_pressure(sk) &&
420 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
421 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
424 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
425 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
428 /* Initialize RCV_MSS value.
429 * RCV_MSS is an our guess about MSS used by the peer.
430 * We haven't any direct information about the MSS.
431 * It's better to underestimate the RCV_MSS rather than overestimate.
432 * Overestimations make us ACKing less frequently than needed.
433 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
435 void tcp_initialize_rcv_mss(struct sock *sk)
437 const struct tcp_sock *tp = tcp_sk(sk);
438 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
440 hint = min(hint, tp->rcv_wnd / 2);
441 hint = min(hint, TCP_MSS_DEFAULT);
442 hint = max(hint, TCP_MIN_MSS);
444 inet_csk(sk)->icsk_ack.rcv_mss = hint;
446 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
448 /* Receiver "autotuning" code.
450 * The algorithm for RTT estimation w/o timestamps is based on
451 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
452 * <http://public.lanl.gov/radiant/pubs.html#DRS>
454 * More detail on this code can be found at
455 * <http://staff.psc.edu/jheffner/>,
456 * though this reference is out of date. A new paper
459 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
461 u32 new_sample = tp->rcv_rtt_est.rtt;
467 if (new_sample != 0) {
468 /* If we sample in larger samples in the non-timestamp
469 * case, we could grossly overestimate the RTT especially
470 * with chatty applications or bulk transfer apps which
471 * are stalled on filesystem I/O.
473 * Also, since we are only going for a minimum in the
474 * non-timestamp case, we do not smooth things out
475 * else with timestamps disabled convergence takes too
479 m -= (new_sample >> 3);
487 /* No previous measure. */
491 if (tp->rcv_rtt_est.rtt != new_sample)
492 tp->rcv_rtt_est.rtt = new_sample;
495 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
497 if (tp->rcv_rtt_est.time == 0)
499 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
501 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
504 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
505 tp->rcv_rtt_est.time = tcp_time_stamp;
508 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
509 const struct sk_buff *skb)
511 struct tcp_sock *tp = tcp_sk(sk);
512 if (tp->rx_opt.rcv_tsecr &&
513 (TCP_SKB_CB(skb)->end_seq -
514 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
515 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
519 * This function should be called every time data is copied to user space.
520 * It calculates the appropriate TCP receive buffer space.
522 void tcp_rcv_space_adjust(struct sock *sk)
524 struct tcp_sock *tp = tcp_sk(sk);
528 if (tp->rcvq_space.time == 0)
531 time = tcp_time_stamp - tp->rcvq_space.time;
532 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
535 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
537 space = max(tp->rcvq_space.space, space);
539 if (tp->rcvq_space.space != space) {
542 tp->rcvq_space.space = space;
544 if (sysctl_tcp_moderate_rcvbuf &&
545 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
546 int new_clamp = space;
548 /* Receive space grows, normalize in order to
549 * take into account packet headers and sk_buff
550 * structure overhead.
555 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
556 while (tcp_win_from_space(rcvmem) < tp->advmss)
559 space = min(space, sysctl_tcp_rmem[2]);
560 if (space > sk->sk_rcvbuf) {
561 sk->sk_rcvbuf = space;
563 /* Make the window clamp follow along. */
564 tp->window_clamp = new_clamp;
570 tp->rcvq_space.seq = tp->copied_seq;
571 tp->rcvq_space.time = tcp_time_stamp;
574 /* There is something which you must keep in mind when you analyze the
575 * behavior of the tp->ato delayed ack timeout interval. When a
576 * connection starts up, we want to ack as quickly as possible. The
577 * problem is that "good" TCP's do slow start at the beginning of data
578 * transmission. The means that until we send the first few ACK's the
579 * sender will sit on his end and only queue most of his data, because
580 * he can only send snd_cwnd unacked packets at any given time. For
581 * each ACK we send, he increments snd_cwnd and transmits more of his
584 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
586 struct tcp_sock *tp = tcp_sk(sk);
587 struct inet_connection_sock *icsk = inet_csk(sk);
590 inet_csk_schedule_ack(sk);
592 tcp_measure_rcv_mss(sk, skb);
594 tcp_rcv_rtt_measure(tp);
596 now = tcp_time_stamp;
598 if (!icsk->icsk_ack.ato) {
599 /* The _first_ data packet received, initialize
600 * delayed ACK engine.
602 tcp_incr_quickack(sk);
603 icsk->icsk_ack.ato = TCP_ATO_MIN;
605 int m = now - icsk->icsk_ack.lrcvtime;
607 if (m <= TCP_ATO_MIN / 2) {
608 /* The fastest case is the first. */
609 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
610 } else if (m < icsk->icsk_ack.ato) {
611 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
612 if (icsk->icsk_ack.ato > icsk->icsk_rto)
613 icsk->icsk_ack.ato = icsk->icsk_rto;
614 } else if (m > icsk->icsk_rto) {
615 /* Too long gap. Apparently sender failed to
616 * restart window, so that we send ACKs quickly.
618 tcp_incr_quickack(sk);
622 icsk->icsk_ack.lrcvtime = now;
624 TCP_ECN_check_ce(tp, skb);
627 tcp_grow_window(sk, skb);
630 /* Called to compute a smoothed rtt estimate. The data fed to this
631 * routine either comes from timestamps, or from segments that were
632 * known _not_ to have been retransmitted [see Karn/Partridge
633 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
634 * piece by Van Jacobson.
635 * NOTE: the next three routines used to be one big routine.
636 * To save cycles in the RFC 1323 implementation it was better to break
637 * it up into three procedures. -- erics
639 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
641 struct tcp_sock *tp = tcp_sk(sk);
642 long m = mrtt; /* RTT */
644 /* The following amusing code comes from Jacobson's
645 * article in SIGCOMM '88. Note that rtt and mdev
646 * are scaled versions of rtt and mean deviation.
647 * This is designed to be as fast as possible
648 * m stands for "measurement".
650 * On a 1990 paper the rto value is changed to:
651 * RTO = rtt + 4 * mdev
653 * Funny. This algorithm seems to be very broken.
654 * These formulae increase RTO, when it should be decreased, increase
655 * too slowly, when it should be increased quickly, decrease too quickly
656 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
657 * does not matter how to _calculate_ it. Seems, it was trap
658 * that VJ failed to avoid. 8)
663 m -= (tp->srtt >> 3); /* m is now error in rtt est */
664 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
666 m = -m; /* m is now abs(error) */
667 m -= (tp->mdev >> 2); /* similar update on mdev */
668 /* This is similar to one of Eifel findings.
669 * Eifel blocks mdev updates when rtt decreases.
670 * This solution is a bit different: we use finer gain
671 * for mdev in this case (alpha*beta).
672 * Like Eifel it also prevents growth of rto,
673 * but also it limits too fast rto decreases,
674 * happening in pure Eifel.
679 m -= (tp->mdev >> 2); /* similar update on mdev */
681 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
682 if (tp->mdev > tp->mdev_max) {
683 tp->mdev_max = tp->mdev;
684 if (tp->mdev_max > tp->rttvar)
685 tp->rttvar = tp->mdev_max;
687 if (after(tp->snd_una, tp->rtt_seq)) {
688 if (tp->mdev_max < tp->rttvar)
689 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
690 tp->rtt_seq = tp->snd_nxt;
691 tp->mdev_max = tcp_rto_min(sk);
694 /* no previous measure. */
695 tp->srtt = m << 3; /* take the measured time to be rtt */
696 tp->mdev = m << 1; /* make sure rto = 3*rtt */
697 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
698 tp->rtt_seq = tp->snd_nxt;
702 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
703 * Note: TCP stack does not yet implement pacing.
704 * FQ packet scheduler can be used to implement cheap but effective
705 * TCP pacing, to smooth the burst on large writes when packets
706 * in flight is significantly lower than cwnd (or rwin)
708 static void tcp_update_pacing_rate(struct sock *sk)
710 const struct tcp_sock *tp = tcp_sk(sk);
713 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
714 rate = (u64)tp->mss_cache * 2 * (HZ << 3);
716 rate *= max(tp->snd_cwnd, tp->packets_out);
718 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
719 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
720 * We probably need usec resolution in the future.
721 * Note: This also takes care of possible srtt=0 case,
722 * when tcp_rtt_estimator() was not yet called.
724 if (tp->srtt > 8 + 2)
725 do_div(rate, tp->srtt);
727 sk->sk_pacing_rate = min_t(u64, rate, ~0U);
730 /* Calculate rto without backoff. This is the second half of Van Jacobson's
731 * routine referred to above.
733 void tcp_set_rto(struct sock *sk)
735 const struct tcp_sock *tp = tcp_sk(sk);
736 /* Old crap is replaced with new one. 8)
739 * 1. If rtt variance happened to be less 50msec, it is hallucination.
740 * It cannot be less due to utterly erratic ACK generation made
741 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
742 * to do with delayed acks, because at cwnd>2 true delack timeout
743 * is invisible. Actually, Linux-2.4 also generates erratic
744 * ACKs in some circumstances.
746 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
748 /* 2. Fixups made earlier cannot be right.
749 * If we do not estimate RTO correctly without them,
750 * all the algo is pure shit and should be replaced
751 * with correct one. It is exactly, which we pretend to do.
754 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
755 * guarantees that rto is higher.
760 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
762 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
765 cwnd = TCP_INIT_CWND;
766 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
770 * Packet counting of FACK is based on in-order assumptions, therefore TCP
771 * disables it when reordering is detected
773 void tcp_disable_fack(struct tcp_sock *tp)
775 /* RFC3517 uses different metric in lost marker => reset on change */
777 tp->lost_skb_hint = NULL;
778 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
781 /* Take a notice that peer is sending D-SACKs */
782 static void tcp_dsack_seen(struct tcp_sock *tp)
784 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
787 static void tcp_update_reordering(struct sock *sk, const int metric,
790 struct tcp_sock *tp = tcp_sk(sk);
791 if (metric > tp->reordering) {
794 tp->reordering = min(TCP_MAX_REORDERING, metric);
796 /* This exciting event is worth to be remembered. 8) */
798 mib_idx = LINUX_MIB_TCPTSREORDER;
799 else if (tcp_is_reno(tp))
800 mib_idx = LINUX_MIB_TCPRENOREORDER;
801 else if (tcp_is_fack(tp))
802 mib_idx = LINUX_MIB_TCPFACKREORDER;
804 mib_idx = LINUX_MIB_TCPSACKREORDER;
806 NET_INC_STATS_BH(sock_net(sk), mib_idx);
807 #if FASTRETRANS_DEBUG > 1
808 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
809 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
813 tp->undo_marker ? tp->undo_retrans : 0);
815 tcp_disable_fack(tp);
819 tcp_disable_early_retrans(tp);
822 /* This must be called before lost_out is incremented */
823 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
825 if ((tp->retransmit_skb_hint == NULL) ||
826 before(TCP_SKB_CB(skb)->seq,
827 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
828 tp->retransmit_skb_hint = skb;
831 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
832 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
835 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
837 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
838 tcp_verify_retransmit_hint(tp, skb);
840 tp->lost_out += tcp_skb_pcount(skb);
841 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
845 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
848 tcp_verify_retransmit_hint(tp, skb);
850 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
851 tp->lost_out += tcp_skb_pcount(skb);
852 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
856 /* This procedure tags the retransmission queue when SACKs arrive.
858 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
859 * Packets in queue with these bits set are counted in variables
860 * sacked_out, retrans_out and lost_out, correspondingly.
862 * Valid combinations are:
863 * Tag InFlight Description
864 * 0 1 - orig segment is in flight.
865 * S 0 - nothing flies, orig reached receiver.
866 * L 0 - nothing flies, orig lost by net.
867 * R 2 - both orig and retransmit are in flight.
868 * L|R 1 - orig is lost, retransmit is in flight.
869 * S|R 1 - orig reached receiver, retrans is still in flight.
870 * (L|S|R is logically valid, it could occur when L|R is sacked,
871 * but it is equivalent to plain S and code short-curcuits it to S.
872 * L|S is logically invalid, it would mean -1 packet in flight 8))
874 * These 6 states form finite state machine, controlled by the following events:
875 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
876 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
877 * 3. Loss detection event of two flavors:
878 * A. Scoreboard estimator decided the packet is lost.
879 * A'. Reno "three dupacks" marks head of queue lost.
880 * A''. Its FACK modification, head until snd.fack is lost.
881 * B. SACK arrives sacking SND.NXT at the moment, when the
882 * segment was retransmitted.
883 * 4. D-SACK added new rule: D-SACK changes any tag to S.
885 * It is pleasant to note, that state diagram turns out to be commutative,
886 * so that we are allowed not to be bothered by order of our actions,
887 * when multiple events arrive simultaneously. (see the function below).
889 * Reordering detection.
890 * --------------------
891 * Reordering metric is maximal distance, which a packet can be displaced
892 * in packet stream. With SACKs we can estimate it:
894 * 1. SACK fills old hole and the corresponding segment was not
895 * ever retransmitted -> reordering. Alas, we cannot use it
896 * when segment was retransmitted.
897 * 2. The last flaw is solved with D-SACK. D-SACK arrives
898 * for retransmitted and already SACKed segment -> reordering..
899 * Both of these heuristics are not used in Loss state, when we cannot
900 * account for retransmits accurately.
902 * SACK block validation.
903 * ----------------------
905 * SACK block range validation checks that the received SACK block fits to
906 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
907 * Note that SND.UNA is not included to the range though being valid because
908 * it means that the receiver is rather inconsistent with itself reporting
909 * SACK reneging when it should advance SND.UNA. Such SACK block this is
910 * perfectly valid, however, in light of RFC2018 which explicitly states
911 * that "SACK block MUST reflect the newest segment. Even if the newest
912 * segment is going to be discarded ...", not that it looks very clever
913 * in case of head skb. Due to potentional receiver driven attacks, we
914 * choose to avoid immediate execution of a walk in write queue due to
915 * reneging and defer head skb's loss recovery to standard loss recovery
916 * procedure that will eventually trigger (nothing forbids us doing this).
918 * Implements also blockage to start_seq wrap-around. Problem lies in the
919 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
920 * there's no guarantee that it will be before snd_nxt (n). The problem
921 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
924 * <- outs wnd -> <- wrapzone ->
925 * u e n u_w e_w s n_w
927 * |<------------+------+----- TCP seqno space --------------+---------->|
928 * ...-- <2^31 ->| |<--------...
929 * ...---- >2^31 ------>| |<--------...
931 * Current code wouldn't be vulnerable but it's better still to discard such
932 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
933 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
934 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
935 * equal to the ideal case (infinite seqno space without wrap caused issues).
937 * With D-SACK the lower bound is extended to cover sequence space below
938 * SND.UNA down to undo_marker, which is the last point of interest. Yet
939 * again, D-SACK block must not to go across snd_una (for the same reason as
940 * for the normal SACK blocks, explained above). But there all simplicity
941 * ends, TCP might receive valid D-SACKs below that. As long as they reside
942 * fully below undo_marker they do not affect behavior in anyway and can
943 * therefore be safely ignored. In rare cases (which are more or less
944 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
945 * fragmentation and packet reordering past skb's retransmission. To consider
946 * them correctly, the acceptable range must be extended even more though
947 * the exact amount is rather hard to quantify. However, tp->max_window can
948 * be used as an exaggerated estimate.
950 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
951 u32 start_seq, u32 end_seq)
953 /* Too far in future, or reversed (interpretation is ambiguous) */
954 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
957 /* Nasty start_seq wrap-around check (see comments above) */
958 if (!before(start_seq, tp->snd_nxt))
961 /* In outstanding window? ...This is valid exit for D-SACKs too.
962 * start_seq == snd_una is non-sensical (see comments above)
964 if (after(start_seq, tp->snd_una))
967 if (!is_dsack || !tp->undo_marker)
970 /* ...Then it's D-SACK, and must reside below snd_una completely */
971 if (after(end_seq, tp->snd_una))
974 if (!before(start_seq, tp->undo_marker))
978 if (!after(end_seq, tp->undo_marker))
981 /* Undo_marker boundary crossing (overestimates a lot). Known already:
982 * start_seq < undo_marker and end_seq >= undo_marker.
984 return !before(start_seq, end_seq - tp->max_window);
987 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
988 * Event "B". Later note: FACK people cheated me again 8), we have to account
989 * for reordering! Ugly, but should help.
991 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
992 * less than what is now known to be received by the other end (derived from
993 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
994 * retransmitted skbs to avoid some costly processing per ACKs.
996 static void tcp_mark_lost_retrans(struct sock *sk)
998 const struct inet_connection_sock *icsk = inet_csk(sk);
999 struct tcp_sock *tp = tcp_sk(sk);
1000 struct sk_buff *skb;
1002 u32 new_low_seq = tp->snd_nxt;
1003 u32 received_upto = tcp_highest_sack_seq(tp);
1005 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1006 !after(received_upto, tp->lost_retrans_low) ||
1007 icsk->icsk_ca_state != TCP_CA_Recovery)
1010 tcp_for_write_queue(skb, sk) {
1011 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1013 if (skb == tcp_send_head(sk))
1015 if (cnt == tp->retrans_out)
1017 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1020 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1023 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1024 * constraint here (see above) but figuring out that at
1025 * least tp->reordering SACK blocks reside between ack_seq
1026 * and received_upto is not easy task to do cheaply with
1027 * the available datastructures.
1029 * Whether FACK should check here for tp->reordering segs
1030 * in-between one could argue for either way (it would be
1031 * rather simple to implement as we could count fack_count
1032 * during the walk and do tp->fackets_out - fack_count).
1034 if (after(received_upto, ack_seq)) {
1035 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1036 tp->retrans_out -= tcp_skb_pcount(skb);
1038 tcp_skb_mark_lost_uncond_verify(tp, skb);
1039 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1041 if (before(ack_seq, new_low_seq))
1042 new_low_seq = ack_seq;
1043 cnt += tcp_skb_pcount(skb);
1047 if (tp->retrans_out)
1048 tp->lost_retrans_low = new_low_seq;
1051 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1052 struct tcp_sack_block_wire *sp, int num_sacks,
1055 struct tcp_sock *tp = tcp_sk(sk);
1056 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1057 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1058 bool dup_sack = false;
1060 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1063 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1064 } else if (num_sacks > 1) {
1065 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1066 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1068 if (!after(end_seq_0, end_seq_1) &&
1069 !before(start_seq_0, start_seq_1)) {
1072 NET_INC_STATS_BH(sock_net(sk),
1073 LINUX_MIB_TCPDSACKOFORECV);
1077 /* D-SACK for already forgotten data... Do dumb counting. */
1078 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1079 !after(end_seq_0, prior_snd_una) &&
1080 after(end_seq_0, tp->undo_marker))
1086 struct tcp_sacktag_state {
1092 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1093 * the incoming SACK may not exactly match but we can find smaller MSS
1094 * aligned portion of it that matches. Therefore we might need to fragment
1095 * which may fail and creates some hassle (caller must handle error case
1098 * FIXME: this could be merged to shift decision code
1100 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1101 u32 start_seq, u32 end_seq)
1105 unsigned int pkt_len;
1108 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1109 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1111 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1112 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1113 mss = tcp_skb_mss(skb);
1114 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1117 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1121 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1126 /* Round if necessary so that SACKs cover only full MSSes
1127 * and/or the remaining small portion (if present)
1129 if (pkt_len > mss) {
1130 unsigned int new_len = (pkt_len / mss) * mss;
1131 if (!in_sack && new_len < pkt_len) {
1133 if (new_len >= skb->len)
1138 err = tcp_fragment(sk, skb, pkt_len, mss);
1146 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1147 static u8 tcp_sacktag_one(struct sock *sk,
1148 struct tcp_sacktag_state *state, u8 sacked,
1149 u32 start_seq, u32 end_seq,
1150 bool dup_sack, int pcount)
1152 struct tcp_sock *tp = tcp_sk(sk);
1153 int fack_count = state->fack_count;
1155 /* Account D-SACK for retransmitted packet. */
1156 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1157 if (tp->undo_marker && tp->undo_retrans > 0 &&
1158 after(end_seq, tp->undo_marker))
1160 if (sacked & TCPCB_SACKED_ACKED)
1161 state->reord = min(fack_count, state->reord);
1164 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1165 if (!after(end_seq, tp->snd_una))
1168 if (!(sacked & TCPCB_SACKED_ACKED)) {
1169 if (sacked & TCPCB_SACKED_RETRANS) {
1170 /* If the segment is not tagged as lost,
1171 * we do not clear RETRANS, believing
1172 * that retransmission is still in flight.
1174 if (sacked & TCPCB_LOST) {
1175 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1176 tp->lost_out -= pcount;
1177 tp->retrans_out -= pcount;
1180 if (!(sacked & TCPCB_RETRANS)) {
1181 /* New sack for not retransmitted frame,
1182 * which was in hole. It is reordering.
1184 if (before(start_seq,
1185 tcp_highest_sack_seq(tp)))
1186 state->reord = min(fack_count,
1188 if (!after(end_seq, tp->high_seq))
1189 state->flag |= FLAG_ORIG_SACK_ACKED;
1192 if (sacked & TCPCB_LOST) {
1193 sacked &= ~TCPCB_LOST;
1194 tp->lost_out -= pcount;
1198 sacked |= TCPCB_SACKED_ACKED;
1199 state->flag |= FLAG_DATA_SACKED;
1200 tp->sacked_out += pcount;
1202 fack_count += pcount;
1204 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1205 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1206 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1207 tp->lost_cnt_hint += pcount;
1209 if (fack_count > tp->fackets_out)
1210 tp->fackets_out = fack_count;
1213 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1214 * frames and clear it. undo_retrans is decreased above, L|R frames
1215 * are accounted above as well.
1217 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1218 sacked &= ~TCPCB_SACKED_RETRANS;
1219 tp->retrans_out -= pcount;
1225 /* Shift newly-SACKed bytes from this skb to the immediately previous
1226 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1228 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1229 struct tcp_sacktag_state *state,
1230 unsigned int pcount, int shifted, int mss,
1233 struct tcp_sock *tp = tcp_sk(sk);
1234 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1235 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1236 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1240 /* Adjust counters and hints for the newly sacked sequence
1241 * range but discard the return value since prev is already
1242 * marked. We must tag the range first because the seq
1243 * advancement below implicitly advances
1244 * tcp_highest_sack_seq() when skb is highest_sack.
1246 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1247 start_seq, end_seq, dup_sack, pcount);
1249 if (skb == tp->lost_skb_hint)
1250 tp->lost_cnt_hint += pcount;
1252 TCP_SKB_CB(prev)->end_seq += shifted;
1253 TCP_SKB_CB(skb)->seq += shifted;
1255 skb_shinfo(prev)->gso_segs += pcount;
1256 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1257 skb_shinfo(skb)->gso_segs -= pcount;
1259 /* When we're adding to gso_segs == 1, gso_size will be zero,
1260 * in theory this shouldn't be necessary but as long as DSACK
1261 * code can come after this skb later on it's better to keep
1262 * setting gso_size to something.
1264 if (!skb_shinfo(prev)->gso_size) {
1265 skb_shinfo(prev)->gso_size = mss;
1266 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1269 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1270 if (skb_shinfo(skb)->gso_segs <= 1) {
1271 skb_shinfo(skb)->gso_size = 0;
1272 skb_shinfo(skb)->gso_type = 0;
1275 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1276 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1279 BUG_ON(!tcp_skb_pcount(skb));
1280 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1284 /* Whole SKB was eaten :-) */
1286 if (skb == tp->retransmit_skb_hint)
1287 tp->retransmit_skb_hint = prev;
1288 if (skb == tp->scoreboard_skb_hint)
1289 tp->scoreboard_skb_hint = prev;
1290 if (skb == tp->lost_skb_hint) {
1291 tp->lost_skb_hint = prev;
1292 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1295 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1296 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1297 TCP_SKB_CB(prev)->end_seq++;
1299 if (skb == tcp_highest_sack(sk))
1300 tcp_advance_highest_sack(sk, skb);
1302 tcp_unlink_write_queue(skb, sk);
1303 sk_wmem_free_skb(sk, skb);
1305 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1310 /* I wish gso_size would have a bit more sane initialization than
1311 * something-or-zero which complicates things
1313 static int tcp_skb_seglen(const struct sk_buff *skb)
1315 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1318 /* Shifting pages past head area doesn't work */
1319 static int skb_can_shift(const struct sk_buff *skb)
1321 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1324 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1327 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1328 struct tcp_sacktag_state *state,
1329 u32 start_seq, u32 end_seq,
1332 struct tcp_sock *tp = tcp_sk(sk);
1333 struct sk_buff *prev;
1339 if (!sk_can_gso(sk))
1342 /* Normally R but no L won't result in plain S */
1344 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1346 if (!skb_can_shift(skb))
1348 /* This frame is about to be dropped (was ACKed). */
1349 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1352 /* Can only happen with delayed DSACK + discard craziness */
1353 if (unlikely(skb == tcp_write_queue_head(sk)))
1355 prev = tcp_write_queue_prev(sk, skb);
1357 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1360 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1361 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1365 pcount = tcp_skb_pcount(skb);
1366 mss = tcp_skb_seglen(skb);
1368 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1369 * drop this restriction as unnecessary
1371 if (mss != tcp_skb_seglen(prev))
1374 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1376 /* CHECKME: This is non-MSS split case only?, this will
1377 * cause skipped skbs due to advancing loop btw, original
1378 * has that feature too
1380 if (tcp_skb_pcount(skb) <= 1)
1383 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1385 /* TODO: head merge to next could be attempted here
1386 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1387 * though it might not be worth of the additional hassle
1389 * ...we can probably just fallback to what was done
1390 * previously. We could try merging non-SACKed ones
1391 * as well but it probably isn't going to buy off
1392 * because later SACKs might again split them, and
1393 * it would make skb timestamp tracking considerably
1399 len = end_seq - TCP_SKB_CB(skb)->seq;
1401 BUG_ON(len > skb->len);
1403 /* MSS boundaries should be honoured or else pcount will
1404 * severely break even though it makes things bit trickier.
1405 * Optimize common case to avoid most of the divides
1407 mss = tcp_skb_mss(skb);
1409 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1410 * drop this restriction as unnecessary
1412 if (mss != tcp_skb_seglen(prev))
1417 } else if (len < mss) {
1425 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1426 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1429 if (!skb_shift(prev, skb, len))
1431 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1434 /* Hole filled allows collapsing with the next as well, this is very
1435 * useful when hole on every nth skb pattern happens
1437 if (prev == tcp_write_queue_tail(sk))
1439 skb = tcp_write_queue_next(sk, prev);
1441 if (!skb_can_shift(skb) ||
1442 (skb == tcp_send_head(sk)) ||
1443 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1444 (mss != tcp_skb_seglen(skb)))
1448 if (skb_shift(prev, skb, len)) {
1449 pcount += tcp_skb_pcount(skb);
1450 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1454 state->fack_count += pcount;
1461 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1465 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1466 struct tcp_sack_block *next_dup,
1467 struct tcp_sacktag_state *state,
1468 u32 start_seq, u32 end_seq,
1471 struct tcp_sock *tp = tcp_sk(sk);
1472 struct sk_buff *tmp;
1474 tcp_for_write_queue_from(skb, sk) {
1476 bool dup_sack = dup_sack_in;
1478 if (skb == tcp_send_head(sk))
1481 /* queue is in-order => we can short-circuit the walk early */
1482 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1485 if ((next_dup != NULL) &&
1486 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1487 in_sack = tcp_match_skb_to_sack(sk, skb,
1488 next_dup->start_seq,
1494 /* skb reference here is a bit tricky to get right, since
1495 * shifting can eat and free both this skb and the next,
1496 * so not even _safe variant of the loop is enough.
1499 tmp = tcp_shift_skb_data(sk, skb, state,
1500 start_seq, end_seq, dup_sack);
1509 in_sack = tcp_match_skb_to_sack(sk, skb,
1515 if (unlikely(in_sack < 0))
1519 TCP_SKB_CB(skb)->sacked =
1522 TCP_SKB_CB(skb)->sacked,
1523 TCP_SKB_CB(skb)->seq,
1524 TCP_SKB_CB(skb)->end_seq,
1526 tcp_skb_pcount(skb));
1528 if (!before(TCP_SKB_CB(skb)->seq,
1529 tcp_highest_sack_seq(tp)))
1530 tcp_advance_highest_sack(sk, skb);
1533 state->fack_count += tcp_skb_pcount(skb);
1538 /* Avoid all extra work that is being done by sacktag while walking in
1541 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1542 struct tcp_sacktag_state *state,
1545 tcp_for_write_queue_from(skb, sk) {
1546 if (skb == tcp_send_head(sk))
1549 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1552 state->fack_count += tcp_skb_pcount(skb);
1557 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1559 struct tcp_sack_block *next_dup,
1560 struct tcp_sacktag_state *state,
1563 if (next_dup == NULL)
1566 if (before(next_dup->start_seq, skip_to_seq)) {
1567 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1568 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1569 next_dup->start_seq, next_dup->end_seq,
1576 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1578 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1582 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1585 struct tcp_sock *tp = tcp_sk(sk);
1586 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1587 TCP_SKB_CB(ack_skb)->sacked);
1588 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1589 struct tcp_sack_block sp[TCP_NUM_SACKS];
1590 struct tcp_sack_block *cache;
1591 struct tcp_sacktag_state state;
1592 struct sk_buff *skb;
1593 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1595 bool found_dup_sack = false;
1597 int first_sack_index;
1600 state.reord = tp->packets_out;
1602 if (!tp->sacked_out) {
1603 if (WARN_ON(tp->fackets_out))
1604 tp->fackets_out = 0;
1605 tcp_highest_sack_reset(sk);
1608 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1609 num_sacks, prior_snd_una);
1611 state.flag |= FLAG_DSACKING_ACK;
1613 /* Eliminate too old ACKs, but take into
1614 * account more or less fresh ones, they can
1615 * contain valid SACK info.
1617 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1620 if (!tp->packets_out)
1624 first_sack_index = 0;
1625 for (i = 0; i < num_sacks; i++) {
1626 bool dup_sack = !i && found_dup_sack;
1628 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1629 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1631 if (!tcp_is_sackblock_valid(tp, dup_sack,
1632 sp[used_sacks].start_seq,
1633 sp[used_sacks].end_seq)) {
1637 if (!tp->undo_marker)
1638 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1640 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1642 /* Don't count olds caused by ACK reordering */
1643 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1644 !after(sp[used_sacks].end_seq, tp->snd_una))
1646 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1649 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1651 first_sack_index = -1;
1655 /* Ignore very old stuff early */
1656 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1662 /* order SACK blocks to allow in order walk of the retrans queue */
1663 for (i = used_sacks - 1; i > 0; i--) {
1664 for (j = 0; j < i; j++) {
1665 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1666 swap(sp[j], sp[j + 1]);
1668 /* Track where the first SACK block goes to */
1669 if (j == first_sack_index)
1670 first_sack_index = j + 1;
1675 skb = tcp_write_queue_head(sk);
1676 state.fack_count = 0;
1679 if (!tp->sacked_out) {
1680 /* It's already past, so skip checking against it */
1681 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1683 cache = tp->recv_sack_cache;
1684 /* Skip empty blocks in at head of the cache */
1685 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1690 while (i < used_sacks) {
1691 u32 start_seq = sp[i].start_seq;
1692 u32 end_seq = sp[i].end_seq;
1693 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1694 struct tcp_sack_block *next_dup = NULL;
1696 if (found_dup_sack && ((i + 1) == first_sack_index))
1697 next_dup = &sp[i + 1];
1699 /* Skip too early cached blocks */
1700 while (tcp_sack_cache_ok(tp, cache) &&
1701 !before(start_seq, cache->end_seq))
1704 /* Can skip some work by looking recv_sack_cache? */
1705 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1706 after(end_seq, cache->start_seq)) {
1709 if (before(start_seq, cache->start_seq)) {
1710 skb = tcp_sacktag_skip(skb, sk, &state,
1712 skb = tcp_sacktag_walk(skb, sk, next_dup,
1719 /* Rest of the block already fully processed? */
1720 if (!after(end_seq, cache->end_seq))
1723 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1727 /* ...tail remains todo... */
1728 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1729 /* ...but better entrypoint exists! */
1730 skb = tcp_highest_sack(sk);
1733 state.fack_count = tp->fackets_out;
1738 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1739 /* Check overlap against next cached too (past this one already) */
1744 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1745 skb = tcp_highest_sack(sk);
1748 state.fack_count = tp->fackets_out;
1750 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1753 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1754 start_seq, end_seq, dup_sack);
1760 /* Clear the head of the cache sack blocks so we can skip it next time */
1761 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1762 tp->recv_sack_cache[i].start_seq = 0;
1763 tp->recv_sack_cache[i].end_seq = 0;
1765 for (j = 0; j < used_sacks; j++)
1766 tp->recv_sack_cache[i++] = sp[j];
1768 tcp_mark_lost_retrans(sk);
1770 tcp_verify_left_out(tp);
1772 if ((state.reord < tp->fackets_out) &&
1773 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1774 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1778 #if FASTRETRANS_DEBUG > 0
1779 WARN_ON((int)tp->sacked_out < 0);
1780 WARN_ON((int)tp->lost_out < 0);
1781 WARN_ON((int)tp->retrans_out < 0);
1782 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1787 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1788 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1790 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1794 holes = max(tp->lost_out, 1U);
1795 holes = min(holes, tp->packets_out);
1797 if ((tp->sacked_out + holes) > tp->packets_out) {
1798 tp->sacked_out = tp->packets_out - holes;
1804 /* If we receive more dupacks than we expected counting segments
1805 * in assumption of absent reordering, interpret this as reordering.
1806 * The only another reason could be bug in receiver TCP.
1808 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1810 struct tcp_sock *tp = tcp_sk(sk);
1811 if (tcp_limit_reno_sacked(tp))
1812 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1815 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1817 static void tcp_add_reno_sack(struct sock *sk)
1819 struct tcp_sock *tp = tcp_sk(sk);
1821 tcp_check_reno_reordering(sk, 0);
1822 tcp_verify_left_out(tp);
1825 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1827 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1829 struct tcp_sock *tp = tcp_sk(sk);
1832 /* One ACK acked hole. The rest eat duplicate ACKs. */
1833 if (acked - 1 >= tp->sacked_out)
1836 tp->sacked_out -= acked - 1;
1838 tcp_check_reno_reordering(sk, acked);
1839 tcp_verify_left_out(tp);
1842 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1847 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1849 tp->retrans_out = 0;
1852 tp->undo_marker = 0;
1853 tp->undo_retrans = -1;
1856 void tcp_clear_retrans(struct tcp_sock *tp)
1858 tcp_clear_retrans_partial(tp);
1860 tp->fackets_out = 0;
1864 /* Enter Loss state. If "how" is not zero, forget all SACK information
1865 * and reset tags completely, otherwise preserve SACKs. If receiver
1866 * dropped its ofo queue, we will know this due to reneging detection.
1868 void tcp_enter_loss(struct sock *sk, int how)
1870 const struct inet_connection_sock *icsk = inet_csk(sk);
1871 struct tcp_sock *tp = tcp_sk(sk);
1872 struct sk_buff *skb;
1873 bool new_recovery = false;
1875 /* Reduce ssthresh if it has not yet been made inside this window. */
1876 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1877 !after(tp->high_seq, tp->snd_una) ||
1878 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1879 new_recovery = true;
1880 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1881 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1882 tcp_ca_event(sk, CA_EVENT_LOSS);
1885 tp->snd_cwnd_cnt = 0;
1886 tp->snd_cwnd_stamp = tcp_time_stamp;
1888 tcp_clear_retrans_partial(tp);
1890 if (tcp_is_reno(tp))
1891 tcp_reset_reno_sack(tp);
1893 tp->undo_marker = tp->snd_una;
1896 tp->fackets_out = 0;
1898 tcp_clear_all_retrans_hints(tp);
1900 tcp_for_write_queue(skb, sk) {
1901 if (skb == tcp_send_head(sk))
1904 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1905 tp->undo_marker = 0;
1906 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1907 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1908 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1909 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1910 tp->lost_out += tcp_skb_pcount(skb);
1911 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1914 tcp_verify_left_out(tp);
1916 tp->reordering = min_t(unsigned int, tp->reordering,
1917 sysctl_tcp_reordering);
1918 tcp_set_ca_state(sk, TCP_CA_Loss);
1919 tp->high_seq = tp->snd_nxt;
1920 TCP_ECN_queue_cwr(tp);
1922 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1923 * loss recovery is underway except recurring timeout(s) on
1924 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1926 tp->frto = sysctl_tcp_frto &&
1927 (new_recovery || icsk->icsk_retransmits) &&
1928 !inet_csk(sk)->icsk_mtup.probe_size;
1931 /* If ACK arrived pointing to a remembered SACK, it means that our
1932 * remembered SACKs do not reflect real state of receiver i.e.
1933 * receiver _host_ is heavily congested (or buggy).
1935 * Do processing similar to RTO timeout.
1937 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1939 if (flag & FLAG_SACK_RENEGING) {
1940 struct inet_connection_sock *icsk = inet_csk(sk);
1941 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1943 tcp_enter_loss(sk, 1);
1944 icsk->icsk_retransmits++;
1945 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1946 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1947 icsk->icsk_rto, TCP_RTO_MAX);
1953 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1955 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1958 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1959 * counter when SACK is enabled (without SACK, sacked_out is used for
1962 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1963 * segments up to the highest received SACK block so far and holes in
1966 * With reordering, holes may still be in flight, so RFC3517 recovery
1967 * uses pure sacked_out (total number of SACKed segments) even though
1968 * it violates the RFC that uses duplicate ACKs, often these are equal
1969 * but when e.g. out-of-window ACKs or packet duplication occurs,
1970 * they differ. Since neither occurs due to loss, TCP should really
1973 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1975 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1978 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1980 struct tcp_sock *tp = tcp_sk(sk);
1981 unsigned long delay;
1983 /* Delay early retransmit and entering fast recovery for
1984 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1985 * available, or RTO is scheduled to fire first.
1987 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
1988 (flag & FLAG_ECE) || !tp->srtt)
1991 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
1992 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
1995 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2000 static inline int tcp_skb_timedout(const struct sock *sk,
2001 const struct sk_buff *skb)
2003 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2006 static inline int tcp_head_timedout(const struct sock *sk)
2008 const struct tcp_sock *tp = tcp_sk(sk);
2010 return tp->packets_out &&
2011 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2014 /* Linux NewReno/SACK/FACK/ECN state machine.
2015 * --------------------------------------
2017 * "Open" Normal state, no dubious events, fast path.
2018 * "Disorder" In all the respects it is "Open",
2019 * but requires a bit more attention. It is entered when
2020 * we see some SACKs or dupacks. It is split of "Open"
2021 * mainly to move some processing from fast path to slow one.
2022 * "CWR" CWND was reduced due to some Congestion Notification event.
2023 * It can be ECN, ICMP source quench, local device congestion.
2024 * "Recovery" CWND was reduced, we are fast-retransmitting.
2025 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2027 * tcp_fastretrans_alert() is entered:
2028 * - each incoming ACK, if state is not "Open"
2029 * - when arrived ACK is unusual, namely:
2034 * Counting packets in flight is pretty simple.
2036 * in_flight = packets_out - left_out + retrans_out
2038 * packets_out is SND.NXT-SND.UNA counted in packets.
2040 * retrans_out is number of retransmitted segments.
2042 * left_out is number of segments left network, but not ACKed yet.
2044 * left_out = sacked_out + lost_out
2046 * sacked_out: Packets, which arrived to receiver out of order
2047 * and hence not ACKed. With SACKs this number is simply
2048 * amount of SACKed data. Even without SACKs
2049 * it is easy to give pretty reliable estimate of this number,
2050 * counting duplicate ACKs.
2052 * lost_out: Packets lost by network. TCP has no explicit
2053 * "loss notification" feedback from network (for now).
2054 * It means that this number can be only _guessed_.
2055 * Actually, it is the heuristics to predict lossage that
2056 * distinguishes different algorithms.
2058 * F.e. after RTO, when all the queue is considered as lost,
2059 * lost_out = packets_out and in_flight = retrans_out.
2061 * Essentially, we have now two algorithms counting
2064 * FACK: It is the simplest heuristics. As soon as we decided
2065 * that something is lost, we decide that _all_ not SACKed
2066 * packets until the most forward SACK are lost. I.e.
2067 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2068 * It is absolutely correct estimate, if network does not reorder
2069 * packets. And it loses any connection to reality when reordering
2070 * takes place. We use FACK by default until reordering
2071 * is suspected on the path to this destination.
2073 * NewReno: when Recovery is entered, we assume that one segment
2074 * is lost (classic Reno). While we are in Recovery and
2075 * a partial ACK arrives, we assume that one more packet
2076 * is lost (NewReno). This heuristics are the same in NewReno
2079 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2080 * deflation etc. CWND is real congestion window, never inflated, changes
2081 * only according to classic VJ rules.
2083 * Really tricky (and requiring careful tuning) part of algorithm
2084 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2085 * The first determines the moment _when_ we should reduce CWND and,
2086 * hence, slow down forward transmission. In fact, it determines the moment
2087 * when we decide that hole is caused by loss, rather than by a reorder.
2089 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2090 * holes, caused by lost packets.
2092 * And the most logically complicated part of algorithm is undo
2093 * heuristics. We detect false retransmits due to both too early
2094 * fast retransmit (reordering) and underestimated RTO, analyzing
2095 * timestamps and D-SACKs. When we detect that some segments were
2096 * retransmitted by mistake and CWND reduction was wrong, we undo
2097 * window reduction and abort recovery phase. This logic is hidden
2098 * inside several functions named tcp_try_undo_<something>.
2101 /* This function decides, when we should leave Disordered state
2102 * and enter Recovery phase, reducing congestion window.
2104 * Main question: may we further continue forward transmission
2105 * with the same cwnd?
2107 static bool tcp_time_to_recover(struct sock *sk, int flag)
2109 struct tcp_sock *tp = tcp_sk(sk);
2112 /* Trick#1: The loss is proven. */
2116 /* Not-A-Trick#2 : Classic rule... */
2117 if (tcp_dupack_heuristics(tp) > tp->reordering)
2120 /* Trick#3 : when we use RFC2988 timer restart, fast
2121 * retransmit can be triggered by timeout of queue head.
2123 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2126 /* Trick#4: It is still not OK... But will it be useful to delay
2129 packets_out = tp->packets_out;
2130 if (packets_out <= tp->reordering &&
2131 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2132 !tcp_may_send_now(sk)) {
2133 /* We have nothing to send. This connection is limited
2134 * either by receiver window or by application.
2139 /* If a thin stream is detected, retransmit after first
2140 * received dupack. Employ only if SACK is supported in order
2141 * to avoid possible corner-case series of spurious retransmissions
2142 * Use only if there are no unsent data.
2144 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2145 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2146 tcp_is_sack(tp) && !tcp_send_head(sk))
2149 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2150 * retransmissions due to small network reorderings, we implement
2151 * Mitigation A.3 in the RFC and delay the retransmission for a short
2152 * interval if appropriate.
2154 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2155 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2156 !tcp_may_send_now(sk))
2157 return !tcp_pause_early_retransmit(sk, flag);
2162 /* New heuristics: it is possible only after we switched to restart timer
2163 * each time when something is ACKed. Hence, we can detect timed out packets
2164 * during fast retransmit without falling to slow start.
2166 * Usefulness of this as is very questionable, since we should know which of
2167 * the segments is the next to timeout which is relatively expensive to find
2168 * in general case unless we add some data structure just for that. The
2169 * current approach certainly won't find the right one too often and when it
2170 * finally does find _something_ it usually marks large part of the window
2171 * right away (because a retransmission with a larger timestamp blocks the
2172 * loop from advancing). -ij
2174 static void tcp_timeout_skbs(struct sock *sk)
2176 struct tcp_sock *tp = tcp_sk(sk);
2177 struct sk_buff *skb;
2179 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2182 skb = tp->scoreboard_skb_hint;
2183 if (tp->scoreboard_skb_hint == NULL)
2184 skb = tcp_write_queue_head(sk);
2186 tcp_for_write_queue_from(skb, sk) {
2187 if (skb == tcp_send_head(sk))
2189 if (!tcp_skb_timedout(sk, skb))
2192 tcp_skb_mark_lost(tp, skb);
2195 tp->scoreboard_skb_hint = skb;
2197 tcp_verify_left_out(tp);
2200 /* Detect loss in event "A" above by marking head of queue up as lost.
2201 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2202 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2203 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2204 * the maximum SACKed segments to pass before reaching this limit.
2206 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2208 struct tcp_sock *tp = tcp_sk(sk);
2209 struct sk_buff *skb;
2213 /* Use SACK to deduce losses of new sequences sent during recovery */
2214 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2216 WARN_ON(packets > tp->packets_out);
2217 if (tp->lost_skb_hint) {
2218 skb = tp->lost_skb_hint;
2219 cnt = tp->lost_cnt_hint;
2220 /* Head already handled? */
2221 if (mark_head && skb != tcp_write_queue_head(sk))
2224 skb = tcp_write_queue_head(sk);
2228 tcp_for_write_queue_from(skb, sk) {
2229 if (skb == tcp_send_head(sk))
2231 /* TODO: do this better */
2232 /* this is not the most efficient way to do this... */
2233 tp->lost_skb_hint = skb;
2234 tp->lost_cnt_hint = cnt;
2236 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2240 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2241 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2242 cnt += tcp_skb_pcount(skb);
2244 if (cnt > packets) {
2245 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2246 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2247 (oldcnt >= packets))
2250 mss = skb_shinfo(skb)->gso_size;
2251 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2257 tcp_skb_mark_lost(tp, skb);
2262 tcp_verify_left_out(tp);
2265 /* Account newly detected lost packet(s) */
2267 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2269 struct tcp_sock *tp = tcp_sk(sk);
2271 if (tcp_is_reno(tp)) {
2272 tcp_mark_head_lost(sk, 1, 1);
2273 } else if (tcp_is_fack(tp)) {
2274 int lost = tp->fackets_out - tp->reordering;
2277 tcp_mark_head_lost(sk, lost, 0);
2279 int sacked_upto = tp->sacked_out - tp->reordering;
2280 if (sacked_upto >= 0)
2281 tcp_mark_head_lost(sk, sacked_upto, 0);
2282 else if (fast_rexmit)
2283 tcp_mark_head_lost(sk, 1, 1);
2286 tcp_timeout_skbs(sk);
2289 /* CWND moderation, preventing bursts due to too big ACKs
2290 * in dubious situations.
2292 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2294 tp->snd_cwnd = min(tp->snd_cwnd,
2295 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2296 tp->snd_cwnd_stamp = tcp_time_stamp;
2299 /* Nothing was retransmitted or returned timestamp is less
2300 * than timestamp of the first retransmission.
2302 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2304 return !tp->retrans_stamp ||
2305 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2306 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2309 /* Undo procedures. */
2311 #if FASTRETRANS_DEBUG > 1
2312 static void DBGUNDO(struct sock *sk, const char *msg)
2314 struct tcp_sock *tp = tcp_sk(sk);
2315 struct inet_sock *inet = inet_sk(sk);
2317 if (sk->sk_family == AF_INET) {
2318 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2320 &inet->inet_daddr, ntohs(inet->inet_dport),
2321 tp->snd_cwnd, tcp_left_out(tp),
2322 tp->snd_ssthresh, tp->prior_ssthresh,
2325 #if IS_ENABLED(CONFIG_IPV6)
2326 else if (sk->sk_family == AF_INET6) {
2327 struct ipv6_pinfo *np = inet6_sk(sk);
2328 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2330 &np->daddr, ntohs(inet->inet_dport),
2331 tp->snd_cwnd, tcp_left_out(tp),
2332 tp->snd_ssthresh, tp->prior_ssthresh,
2338 #define DBGUNDO(x...) do { } while (0)
2341 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2343 struct tcp_sock *tp = tcp_sk(sk);
2345 if (tp->prior_ssthresh) {
2346 const struct inet_connection_sock *icsk = inet_csk(sk);
2348 if (icsk->icsk_ca_ops->undo_cwnd)
2349 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2351 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2353 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2354 tp->snd_ssthresh = tp->prior_ssthresh;
2355 TCP_ECN_withdraw_cwr(tp);
2358 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2360 tp->snd_cwnd_stamp = tcp_time_stamp;
2363 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2365 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2368 /* People celebrate: "We love our President!" */
2369 static bool tcp_try_undo_recovery(struct sock *sk)
2371 struct tcp_sock *tp = tcp_sk(sk);
2373 if (tcp_may_undo(tp)) {
2376 /* Happy end! We did not retransmit anything
2377 * or our original transmission succeeded.
2379 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2380 tcp_undo_cwr(sk, true);
2381 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2382 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2384 mib_idx = LINUX_MIB_TCPFULLUNDO;
2386 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2387 tp->undo_marker = 0;
2389 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2390 /* Hold old state until something *above* high_seq
2391 * is ACKed. For Reno it is MUST to prevent false
2392 * fast retransmits (RFC2582). SACK TCP is safe. */
2393 tcp_moderate_cwnd(tp);
2396 tcp_set_ca_state(sk, TCP_CA_Open);
2400 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2401 static void tcp_try_undo_dsack(struct sock *sk)
2403 struct tcp_sock *tp = tcp_sk(sk);
2405 if (tp->undo_marker && !tp->undo_retrans) {
2406 DBGUNDO(sk, "D-SACK");
2407 tcp_undo_cwr(sk, true);
2408 tp->undo_marker = 0;
2409 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2413 /* We can clear retrans_stamp when there are no retransmissions in the
2414 * window. It would seem that it is trivially available for us in
2415 * tp->retrans_out, however, that kind of assumptions doesn't consider
2416 * what will happen if errors occur when sending retransmission for the
2417 * second time. ...It could the that such segment has only
2418 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2419 * the head skb is enough except for some reneging corner cases that
2420 * are not worth the effort.
2422 * Main reason for all this complexity is the fact that connection dying
2423 * time now depends on the validity of the retrans_stamp, in particular,
2424 * that successive retransmissions of a segment must not advance
2425 * retrans_stamp under any conditions.
2427 static bool tcp_any_retrans_done(const struct sock *sk)
2429 const struct tcp_sock *tp = tcp_sk(sk);
2430 struct sk_buff *skb;
2432 if (tp->retrans_out)
2435 skb = tcp_write_queue_head(sk);
2436 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2442 /* Undo during fast recovery after partial ACK. */
2444 static int tcp_try_undo_partial(struct sock *sk, int acked)
2446 struct tcp_sock *tp = tcp_sk(sk);
2447 /* Partial ACK arrived. Force Hoe's retransmit. */
2448 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2450 if (tcp_may_undo(tp)) {
2451 /* Plain luck! Hole if filled with delayed
2452 * packet, rather than with a retransmit.
2454 if (!tcp_any_retrans_done(sk))
2455 tp->retrans_stamp = 0;
2457 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2460 tcp_undo_cwr(sk, false);
2461 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2463 /* So... Do not make Hoe's retransmit yet.
2464 * If the first packet was delayed, the rest
2465 * ones are most probably delayed as well.
2472 /* Undo during loss recovery after partial ACK or using F-RTO. */
2473 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2475 struct tcp_sock *tp = tcp_sk(sk);
2477 if (frto_undo || tcp_may_undo(tp)) {
2478 struct sk_buff *skb;
2479 tcp_for_write_queue(skb, sk) {
2480 if (skb == tcp_send_head(sk))
2482 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2485 tcp_clear_all_retrans_hints(tp);
2487 DBGUNDO(sk, "partial loss");
2489 tcp_undo_cwr(sk, true);
2490 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2492 NET_INC_STATS_BH(sock_net(sk),
2493 LINUX_MIB_TCPSPURIOUSRTOS);
2494 inet_csk(sk)->icsk_retransmits = 0;
2495 tp->undo_marker = 0;
2496 if (frto_undo || tcp_is_sack(tp))
2497 tcp_set_ca_state(sk, TCP_CA_Open);
2503 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2504 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2505 * It computes the number of packets to send (sndcnt) based on packets newly
2507 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2508 * cwnd reductions across a full RTT.
2509 * 2) If packets in flight is lower than ssthresh (such as due to excess
2510 * losses and/or application stalls), do not perform any further cwnd
2511 * reductions, but instead slow start up to ssthresh.
2513 static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2515 struct tcp_sock *tp = tcp_sk(sk);
2517 tp->high_seq = tp->snd_nxt;
2518 tp->tlp_high_seq = 0;
2519 tp->snd_cwnd_cnt = 0;
2520 tp->prior_cwnd = tp->snd_cwnd;
2521 tp->prr_delivered = 0;
2524 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2525 TCP_ECN_queue_cwr(tp);
2528 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2531 struct tcp_sock *tp = tcp_sk(sk);
2533 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2535 tp->prr_delivered += newly_acked_sacked;
2536 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2537 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2539 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2541 sndcnt = min_t(int, delta,
2542 max_t(int, tp->prr_delivered - tp->prr_out,
2543 newly_acked_sacked) + 1);
2546 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2547 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2550 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2552 struct tcp_sock *tp = tcp_sk(sk);
2554 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2555 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2556 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2557 tp->snd_cwnd = tp->snd_ssthresh;
2558 tp->snd_cwnd_stamp = tcp_time_stamp;
2560 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2563 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2564 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2566 struct tcp_sock *tp = tcp_sk(sk);
2568 tp->prior_ssthresh = 0;
2569 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2570 tp->undo_marker = 0;
2571 tcp_init_cwnd_reduction(sk, set_ssthresh);
2572 tcp_set_ca_state(sk, TCP_CA_CWR);
2576 static void tcp_try_keep_open(struct sock *sk)
2578 struct tcp_sock *tp = tcp_sk(sk);
2579 int state = TCP_CA_Open;
2581 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2582 state = TCP_CA_Disorder;
2584 if (inet_csk(sk)->icsk_ca_state != state) {
2585 tcp_set_ca_state(sk, state);
2586 tp->high_seq = tp->snd_nxt;
2590 static void tcp_try_to_open(struct sock *sk, int flag, int newly_acked_sacked)
2592 struct tcp_sock *tp = tcp_sk(sk);
2594 tcp_verify_left_out(tp);
2596 if (!tcp_any_retrans_done(sk))
2597 tp->retrans_stamp = 0;
2599 if (flag & FLAG_ECE)
2600 tcp_enter_cwr(sk, 1);
2602 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2603 tcp_try_keep_open(sk);
2604 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2605 tcp_moderate_cwnd(tp);
2607 tcp_cwnd_reduction(sk, newly_acked_sacked, 0);
2611 static void tcp_mtup_probe_failed(struct sock *sk)
2613 struct inet_connection_sock *icsk = inet_csk(sk);
2615 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2616 icsk->icsk_mtup.probe_size = 0;
2619 static void tcp_mtup_probe_success(struct sock *sk)
2621 struct tcp_sock *tp = tcp_sk(sk);
2622 struct inet_connection_sock *icsk = inet_csk(sk);
2624 /* FIXME: breaks with very large cwnd */
2625 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2626 tp->snd_cwnd = tp->snd_cwnd *
2627 tcp_mss_to_mtu(sk, tp->mss_cache) /
2628 icsk->icsk_mtup.probe_size;
2629 tp->snd_cwnd_cnt = 0;
2630 tp->snd_cwnd_stamp = tcp_time_stamp;
2631 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2633 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2634 icsk->icsk_mtup.probe_size = 0;
2635 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2638 /* Do a simple retransmit without using the backoff mechanisms in
2639 * tcp_timer. This is used for path mtu discovery.
2640 * The socket is already locked here.
2642 void tcp_simple_retransmit(struct sock *sk)
2644 const struct inet_connection_sock *icsk = inet_csk(sk);
2645 struct tcp_sock *tp = tcp_sk(sk);
2646 struct sk_buff *skb;
2647 unsigned int mss = tcp_current_mss(sk);
2648 u32 prior_lost = tp->lost_out;
2650 tcp_for_write_queue(skb, sk) {
2651 if (skb == tcp_send_head(sk))
2653 if (tcp_skb_seglen(skb) > mss &&
2654 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2655 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2656 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2657 tp->retrans_out -= tcp_skb_pcount(skb);
2659 tcp_skb_mark_lost_uncond_verify(tp, skb);
2663 tcp_clear_retrans_hints_partial(tp);
2665 if (prior_lost == tp->lost_out)
2668 if (tcp_is_reno(tp))
2669 tcp_limit_reno_sacked(tp);
2671 tcp_verify_left_out(tp);
2673 /* Don't muck with the congestion window here.
2674 * Reason is that we do not increase amount of _data_
2675 * in network, but units changed and effective
2676 * cwnd/ssthresh really reduced now.
2678 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2679 tp->high_seq = tp->snd_nxt;
2680 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2681 tp->prior_ssthresh = 0;
2682 tp->undo_marker = 0;
2683 tcp_set_ca_state(sk, TCP_CA_Loss);
2685 tcp_xmit_retransmit_queue(sk);
2687 EXPORT_SYMBOL(tcp_simple_retransmit);
2689 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2691 struct tcp_sock *tp = tcp_sk(sk);
2694 if (tcp_is_reno(tp))
2695 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2697 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2699 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2701 tp->prior_ssthresh = 0;
2702 tp->undo_marker = tp->snd_una;
2703 tp->undo_retrans = tp->retrans_out ? : -1;
2705 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2707 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2708 tcp_init_cwnd_reduction(sk, true);
2710 tcp_set_ca_state(sk, TCP_CA_Recovery);
2713 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2714 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2716 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2718 struct inet_connection_sock *icsk = inet_csk(sk);
2719 struct tcp_sock *tp = tcp_sk(sk);
2720 bool recovered = !before(tp->snd_una, tp->high_seq);
2722 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2723 /* Step 3.b. A timeout is spurious if not all data are
2724 * lost, i.e., never-retransmitted data are (s)acked.
2726 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2729 if (after(tp->snd_nxt, tp->high_seq) &&
2730 (flag & FLAG_DATA_SACKED || is_dupack)) {
2731 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2732 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2733 tp->high_seq = tp->snd_nxt;
2734 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2736 if (after(tp->snd_nxt, tp->high_seq))
2737 return; /* Step 2.b */
2743 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2744 icsk->icsk_retransmits = 0;
2745 tcp_try_undo_recovery(sk);
2748 if (flag & FLAG_DATA_ACKED)
2749 icsk->icsk_retransmits = 0;
2750 if (tcp_is_reno(tp)) {
2751 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2752 * delivered. Lower inflight to clock out (re)tranmissions.
2754 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2755 tcp_add_reno_sack(sk);
2756 else if (flag & FLAG_SND_UNA_ADVANCED)
2757 tcp_reset_reno_sack(tp);
2759 if (tcp_try_undo_loss(sk, false))
2761 tcp_xmit_retransmit_queue(sk);
2764 /* Process an event, which can update packets-in-flight not trivially.
2765 * Main goal of this function is to calculate new estimate for left_out,
2766 * taking into account both packets sitting in receiver's buffer and
2767 * packets lost by network.
2769 * Besides that it does CWND reduction, when packet loss is detected
2770 * and changes state of machine.
2772 * It does _not_ decide what to send, it is made in function
2773 * tcp_xmit_retransmit_queue().
2775 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
2776 int prior_sacked, int prior_packets,
2777 bool is_dupack, int flag)
2779 struct inet_connection_sock *icsk = inet_csk(sk);
2780 struct tcp_sock *tp = tcp_sk(sk);
2781 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2782 (tcp_fackets_out(tp) > tp->reordering));
2783 int newly_acked_sacked = 0;
2784 int fast_rexmit = 0;
2786 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2788 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2789 tp->fackets_out = 0;
2791 /* Now state machine starts.
2792 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2793 if (flag & FLAG_ECE)
2794 tp->prior_ssthresh = 0;
2796 /* B. In all the states check for reneging SACKs. */
2797 if (tcp_check_sack_reneging(sk, flag))
2800 /* C. Check consistency of the current state. */
2801 tcp_verify_left_out(tp);
2803 /* D. Check state exit conditions. State can be terminated
2804 * when high_seq is ACKed. */
2805 if (icsk->icsk_ca_state == TCP_CA_Open) {
2806 WARN_ON(tp->retrans_out != 0);
2807 tp->retrans_stamp = 0;
2808 } else if (!before(tp->snd_una, tp->high_seq)) {
2809 switch (icsk->icsk_ca_state) {
2811 /* CWR is to be held something *above* high_seq
2812 * is ACKed for CWR bit to reach receiver. */
2813 if (tp->snd_una != tp->high_seq) {
2814 tcp_end_cwnd_reduction(sk);
2815 tcp_set_ca_state(sk, TCP_CA_Open);
2819 case TCP_CA_Recovery:
2820 if (tcp_is_reno(tp))
2821 tcp_reset_reno_sack(tp);
2822 if (tcp_try_undo_recovery(sk))
2824 tcp_end_cwnd_reduction(sk);
2829 /* E. Process state. */
2830 switch (icsk->icsk_ca_state) {
2831 case TCP_CA_Recovery:
2832 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2833 if (tcp_is_reno(tp) && is_dupack)
2834 tcp_add_reno_sack(sk);
2836 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2837 newly_acked_sacked = prior_packets - tp->packets_out +
2838 tp->sacked_out - prior_sacked;
2841 tcp_process_loss(sk, flag, is_dupack);
2842 if (icsk->icsk_ca_state != TCP_CA_Open)
2844 /* Fall through to processing in Open state. */
2846 if (tcp_is_reno(tp)) {
2847 if (flag & FLAG_SND_UNA_ADVANCED)
2848 tcp_reset_reno_sack(tp);
2850 tcp_add_reno_sack(sk);
2852 newly_acked_sacked = prior_packets - tp->packets_out +
2853 tp->sacked_out - prior_sacked;
2855 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2856 tcp_try_undo_dsack(sk);
2858 if (!tcp_time_to_recover(sk, flag)) {
2859 tcp_try_to_open(sk, flag, newly_acked_sacked);
2863 /* MTU probe failure: don't reduce cwnd */
2864 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2865 icsk->icsk_mtup.probe_size &&
2866 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2867 tcp_mtup_probe_failed(sk);
2868 /* Restores the reduction we did in tcp_mtup_probe() */
2870 tcp_simple_retransmit(sk);
2874 /* Otherwise enter Recovery state */
2875 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2879 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2880 tcp_update_scoreboard(sk, fast_rexmit);
2881 tcp_cwnd_reduction(sk, newly_acked_sacked, fast_rexmit);
2882 tcp_xmit_retransmit_queue(sk);
2885 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
2887 tcp_rtt_estimator(sk, seq_rtt);
2889 inet_csk(sk)->icsk_backoff = 0;
2891 EXPORT_SYMBOL(tcp_valid_rtt_meas);
2893 /* Read draft-ietf-tcplw-high-performance before mucking
2894 * with this code. (Supersedes RFC1323)
2896 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2898 /* RTTM Rule: A TSecr value received in a segment is used to
2899 * update the averaged RTT measurement only if the segment
2900 * acknowledges some new data, i.e., only if it advances the
2901 * left edge of the send window.
2903 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2904 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2906 * Changed: reset backoff as soon as we see the first valid sample.
2907 * If we do not, we get strongly overestimated rto. With timestamps
2908 * samples are accepted even from very old segments: f.e., when rtt=1
2909 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2910 * answer arrives rto becomes 120 seconds! If at least one of segments
2911 * in window is lost... Voila. --ANK (010210)
2913 struct tcp_sock *tp = tcp_sk(sk);
2915 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2918 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2920 /* We don't have a timestamp. Can only use
2921 * packets that are not retransmitted to determine
2922 * rtt estimates. Also, we must not reset the
2923 * backoff for rto until we get a non-retransmitted
2924 * packet. This allows us to deal with a situation
2925 * where the network delay has increased suddenly.
2926 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2929 if (flag & FLAG_RETRANS_DATA_ACKED)
2932 tcp_valid_rtt_meas(sk, seq_rtt);
2935 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2938 const struct tcp_sock *tp = tcp_sk(sk);
2939 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2940 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2941 tcp_ack_saw_tstamp(sk, flag);
2942 else if (seq_rtt >= 0)
2943 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2946 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2948 const struct inet_connection_sock *icsk = inet_csk(sk);
2949 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2950 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2953 /* Restart timer after forward progress on connection.
2954 * RFC2988 recommends to restart timer to now+rto.
2956 void tcp_rearm_rto(struct sock *sk)
2958 const struct inet_connection_sock *icsk = inet_csk(sk);
2959 struct tcp_sock *tp = tcp_sk(sk);
2961 /* If the retrans timer is currently being used by Fast Open
2962 * for SYN-ACK retrans purpose, stay put.
2964 if (tp->fastopen_rsk)
2967 if (!tp->packets_out) {
2968 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2970 u32 rto = inet_csk(sk)->icsk_rto;
2971 /* Offset the time elapsed after installing regular RTO */
2972 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2973 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2974 struct sk_buff *skb = tcp_write_queue_head(sk);
2975 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2976 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2977 /* delta may not be positive if the socket is locked
2978 * when the retrans timer fires and is rescheduled.
2983 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2988 /* This function is called when the delayed ER timer fires. TCP enters
2989 * fast recovery and performs fast-retransmit.
2991 void tcp_resume_early_retransmit(struct sock *sk)
2993 struct tcp_sock *tp = tcp_sk(sk);
2997 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2998 if (!tp->do_early_retrans)
3001 tcp_enter_recovery(sk, false);
3002 tcp_update_scoreboard(sk, 1);
3003 tcp_xmit_retransmit_queue(sk);
3006 /* If we get here, the whole TSO packet has not been acked. */
3007 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3009 struct tcp_sock *tp = tcp_sk(sk);
3012 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3014 packets_acked = tcp_skb_pcount(skb);
3015 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3017 packets_acked -= tcp_skb_pcount(skb);
3019 if (packets_acked) {
3020 BUG_ON(tcp_skb_pcount(skb) == 0);
3021 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3024 return packets_acked;
3027 /* Remove acknowledged frames from the retransmission queue. If our packet
3028 * is before the ack sequence we can discard it as it's confirmed to have
3029 * arrived at the other end.
3031 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3034 struct tcp_sock *tp = tcp_sk(sk);
3035 const struct inet_connection_sock *icsk = inet_csk(sk);
3036 struct sk_buff *skb;
3037 u32 now = tcp_time_stamp;
3038 int fully_acked = true;
3041 u32 reord = tp->packets_out;
3042 u32 prior_sacked = tp->sacked_out;
3044 s32 ca_seq_rtt = -1;
3045 ktime_t last_ackt = net_invalid_timestamp();
3047 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3048 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3050 u8 sacked = scb->sacked;
3052 /* Determine how many packets and what bytes were acked, tso and else */
3053 if (after(scb->end_seq, tp->snd_una)) {
3054 if (tcp_skb_pcount(skb) == 1 ||
3055 !after(tp->snd_una, scb->seq))
3058 acked_pcount = tcp_tso_acked(sk, skb);
3062 fully_acked = false;
3064 acked_pcount = tcp_skb_pcount(skb);
3067 if (sacked & TCPCB_RETRANS) {
3068 if (sacked & TCPCB_SACKED_RETRANS)
3069 tp->retrans_out -= acked_pcount;
3070 flag |= FLAG_RETRANS_DATA_ACKED;
3074 ca_seq_rtt = now - scb->when;
3075 last_ackt = skb->tstamp;
3077 seq_rtt = ca_seq_rtt;
3079 if (!(sacked & TCPCB_SACKED_ACKED))
3080 reord = min(pkts_acked, reord);
3081 if (!after(scb->end_seq, tp->high_seq))
3082 flag |= FLAG_ORIG_SACK_ACKED;
3085 if (sacked & TCPCB_SACKED_ACKED)
3086 tp->sacked_out -= acked_pcount;
3087 if (sacked & TCPCB_LOST)
3088 tp->lost_out -= acked_pcount;
3090 tp->packets_out -= acked_pcount;
3091 pkts_acked += acked_pcount;
3093 /* Initial outgoing SYN's get put onto the write_queue
3094 * just like anything else we transmit. It is not
3095 * true data, and if we misinform our callers that
3096 * this ACK acks real data, we will erroneously exit
3097 * connection startup slow start one packet too
3098 * quickly. This is severely frowned upon behavior.
3100 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3101 flag |= FLAG_DATA_ACKED;
3103 flag |= FLAG_SYN_ACKED;
3104 tp->retrans_stamp = 0;
3110 tcp_unlink_write_queue(skb, sk);
3111 sk_wmem_free_skb(sk, skb);
3112 tp->scoreboard_skb_hint = NULL;
3113 if (skb == tp->retransmit_skb_hint)
3114 tp->retransmit_skb_hint = NULL;
3115 if (skb == tp->lost_skb_hint)
3116 tp->lost_skb_hint = NULL;
3119 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3120 tp->snd_up = tp->snd_una;
3122 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3123 flag |= FLAG_SACK_RENEGING;
3125 if (flag & FLAG_ACKED) {
3126 const struct tcp_congestion_ops *ca_ops
3127 = inet_csk(sk)->icsk_ca_ops;
3129 if (unlikely(icsk->icsk_mtup.probe_size &&
3130 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3131 tcp_mtup_probe_success(sk);
3134 tcp_ack_update_rtt(sk, flag, seq_rtt);
3137 if (tcp_is_reno(tp)) {
3138 tcp_remove_reno_sacks(sk, pkts_acked);
3142 /* Non-retransmitted hole got filled? That's reordering */
3143 if (reord < prior_fackets)
3144 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3146 delta = tcp_is_fack(tp) ? pkts_acked :
3147 prior_sacked - tp->sacked_out;
3148 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3151 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3153 if (ca_ops->pkts_acked) {
3156 /* Is the ACK triggering packet unambiguous? */
3157 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3158 /* High resolution needed and available? */
3159 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3160 !ktime_equal(last_ackt,
3161 net_invalid_timestamp()))
3162 rtt_us = ktime_us_delta(ktime_get_real(),
3164 else if (ca_seq_rtt >= 0)
3165 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3168 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3172 #if FASTRETRANS_DEBUG > 0
3173 WARN_ON((int)tp->sacked_out < 0);
3174 WARN_ON((int)tp->lost_out < 0);
3175 WARN_ON((int)tp->retrans_out < 0);
3176 if (!tp->packets_out && tcp_is_sack(tp)) {
3177 icsk = inet_csk(sk);
3179 pr_debug("Leak l=%u %d\n",
3180 tp->lost_out, icsk->icsk_ca_state);
3183 if (tp->sacked_out) {
3184 pr_debug("Leak s=%u %d\n",
3185 tp->sacked_out, icsk->icsk_ca_state);
3188 if (tp->retrans_out) {
3189 pr_debug("Leak r=%u %d\n",
3190 tp->retrans_out, icsk->icsk_ca_state);
3191 tp->retrans_out = 0;
3198 static void tcp_ack_probe(struct sock *sk)
3200 const struct tcp_sock *tp = tcp_sk(sk);
3201 struct inet_connection_sock *icsk = inet_csk(sk);
3203 /* Was it a usable window open? */
3205 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3206 icsk->icsk_backoff = 0;
3207 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3208 /* Socket must be waked up by subsequent tcp_data_snd_check().
3209 * This function is not for random using!
3212 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3213 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3218 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3220 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3221 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3224 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3226 const struct tcp_sock *tp = tcp_sk(sk);
3227 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3228 !tcp_in_cwnd_reduction(sk);
3231 /* Check that window update is acceptable.
3232 * The function assumes that snd_una<=ack<=snd_next.
3234 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3235 const u32 ack, const u32 ack_seq,
3238 return after(ack, tp->snd_una) ||
3239 after(ack_seq, tp->snd_wl1) ||
3240 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3243 /* Update our send window.
3245 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3246 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3248 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3251 struct tcp_sock *tp = tcp_sk(sk);
3253 u32 nwin = ntohs(tcp_hdr(skb)->window);
3255 if (likely(!tcp_hdr(skb)->syn))
3256 nwin <<= tp->rx_opt.snd_wscale;
3258 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3259 flag |= FLAG_WIN_UPDATE;
3260 tcp_update_wl(tp, ack_seq);
3262 if (tp->snd_wnd != nwin) {
3265 /* Note, it is the only place, where
3266 * fast path is recovered for sending TCP.
3269 tcp_fast_path_check(sk);
3271 if (nwin > tp->max_window) {
3272 tp->max_window = nwin;
3273 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3283 /* RFC 5961 7 [ACK Throttling] */
3284 static void tcp_send_challenge_ack(struct sock *sk)
3286 /* unprotected vars, we dont care of overwrites */
3287 static u32 challenge_timestamp;
3288 static unsigned int challenge_count;
3289 u32 now = jiffies / HZ;
3291 if (now != challenge_timestamp) {
3292 challenge_timestamp = now;
3293 challenge_count = 0;
3295 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3296 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3301 static void tcp_store_ts_recent(struct tcp_sock *tp)
3303 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3304 tp->rx_opt.ts_recent_stamp = get_seconds();
3307 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3309 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3310 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3311 * extra check below makes sure this can only happen
3312 * for pure ACK frames. -DaveM
3314 * Not only, also it occurs for expired timestamps.
3317 if (tcp_paws_check(&tp->rx_opt, 0))
3318 tcp_store_ts_recent(tp);
3322 /* This routine deals with acks during a TLP episode.
3323 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3325 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3327 struct tcp_sock *tp = tcp_sk(sk);
3328 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3329 !(flag & (FLAG_SND_UNA_ADVANCED |
3330 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3332 /* Mark the end of TLP episode on receiving TLP dupack or when
3333 * ack is after tlp_high_seq.
3335 if (is_tlp_dupack) {
3336 tp->tlp_high_seq = 0;
3340 if (after(ack, tp->tlp_high_seq)) {
3341 tp->tlp_high_seq = 0;
3342 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3343 if (!(flag & FLAG_DSACKING_ACK)) {
3344 tcp_init_cwnd_reduction(sk, true);
3345 tcp_set_ca_state(sk, TCP_CA_CWR);
3346 tcp_end_cwnd_reduction(sk);
3347 tcp_try_keep_open(sk);
3348 NET_INC_STATS_BH(sock_net(sk),
3349 LINUX_MIB_TCPLOSSPROBERECOVERY);
3354 /* This routine deals with incoming acks, but not outgoing ones. */
3355 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3357 struct inet_connection_sock *icsk = inet_csk(sk);
3358 struct tcp_sock *tp = tcp_sk(sk);
3359 u32 prior_snd_una = tp->snd_una;
3360 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3361 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3362 bool is_dupack = false;
3363 u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
3365 int prior_packets = tp->packets_out;
3366 int prior_sacked = tp->sacked_out;
3368 int previous_packets_out = 0;
3370 /* If the ack is older than previous acks
3371 * then we can probably ignore it.
3373 if (before(ack, prior_snd_una)) {
3374 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3375 if (before(ack, prior_snd_una - tp->max_window)) {
3376 tcp_send_challenge_ack(sk);
3382 /* If the ack includes data we haven't sent yet, discard
3383 * this segment (RFC793 Section 3.9).
3385 if (after(ack, tp->snd_nxt))
3388 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3389 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3392 if (after(ack, prior_snd_una))
3393 flag |= FLAG_SND_UNA_ADVANCED;
3395 prior_fackets = tp->fackets_out;
3396 prior_in_flight = tcp_packets_in_flight(tp);
3398 /* ts_recent update must be made after we are sure that the packet
3401 if (flag & FLAG_UPDATE_TS_RECENT)
3402 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3404 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3405 /* Window is constant, pure forward advance.
3406 * No more checks are required.
3407 * Note, we use the fact that SND.UNA>=SND.WL2.
3409 tcp_update_wl(tp, ack_seq);
3411 flag |= FLAG_WIN_UPDATE;
3413 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3415 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3417 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3420 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3422 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3424 if (TCP_SKB_CB(skb)->sacked)
3425 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3427 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3430 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3433 /* We passed data and got it acked, remove any soft error
3434 * log. Something worked...
3436 sk->sk_err_soft = 0;
3437 icsk->icsk_probes_out = 0;
3438 tp->rcv_tstamp = tcp_time_stamp;
3442 /* See if we can take anything off of the retransmit queue. */
3443 previous_packets_out = tp->packets_out;
3444 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3446 pkts_acked = previous_packets_out - tp->packets_out;
3448 if (tcp_ack_is_dubious(sk, flag)) {
3449 /* Advance CWND, if state allows this. */
3450 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
3451 tcp_cong_avoid(sk, ack, prior_in_flight);
3452 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3453 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3454 prior_packets, is_dupack, flag);
3456 if (flag & FLAG_DATA_ACKED)
3457 tcp_cong_avoid(sk, ack, prior_in_flight);
3460 if (tp->tlp_high_seq)
3461 tcp_process_tlp_ack(sk, ack, flag);
3463 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3464 struct dst_entry *dst = __sk_dst_get(sk);
3469 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3470 tcp_schedule_loss_probe(sk);
3471 if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
3472 tcp_update_pacing_rate(sk);
3476 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3477 if (flag & FLAG_DSACKING_ACK)
3478 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3479 prior_packets, is_dupack, flag);
3480 /* If this ack opens up a zero window, clear backoff. It was
3481 * being used to time the probes, and is probably far higher than
3482 * it needs to be for normal retransmission.
3484 if (tcp_send_head(sk))
3487 if (tp->tlp_high_seq)
3488 tcp_process_tlp_ack(sk, ack, flag);
3492 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3496 /* If data was SACKed, tag it and see if we should send more data.
3497 * If data was DSACKed, see if we can undo a cwnd reduction.
3499 if (TCP_SKB_CB(skb)->sacked) {
3500 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3501 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3502 prior_packets, is_dupack, flag);
3505 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3509 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3510 * But, this can also be called on packets in the established flow when
3511 * the fast version below fails.
3513 void tcp_parse_options(const struct sk_buff *skb,
3514 struct tcp_options_received *opt_rx, int estab,
3515 struct tcp_fastopen_cookie *foc)
3517 const unsigned char *ptr;
3518 const struct tcphdr *th = tcp_hdr(skb);
3519 int length = (th->doff * 4) - sizeof(struct tcphdr);
3521 ptr = (const unsigned char *)(th + 1);
3522 opt_rx->saw_tstamp = 0;
3524 while (length > 0) {
3525 int opcode = *ptr++;
3531 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3536 if (opsize < 2) /* "silly options" */
3538 if (opsize > length)
3539 return; /* don't parse partial options */
3542 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3543 u16 in_mss = get_unaligned_be16(ptr);
3545 if (opt_rx->user_mss &&
3546 opt_rx->user_mss < in_mss)
3547 in_mss = opt_rx->user_mss;
3548 opt_rx->mss_clamp = in_mss;
3553 if (opsize == TCPOLEN_WINDOW && th->syn &&
3554 !estab && sysctl_tcp_window_scaling) {
3555 __u8 snd_wscale = *(__u8 *)ptr;
3556 opt_rx->wscale_ok = 1;
3557 if (snd_wscale > 14) {
3558 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3563 opt_rx->snd_wscale = snd_wscale;
3566 case TCPOPT_TIMESTAMP:
3567 if ((opsize == TCPOLEN_TIMESTAMP) &&
3568 ((estab && opt_rx->tstamp_ok) ||
3569 (!estab && sysctl_tcp_timestamps))) {
3570 opt_rx->saw_tstamp = 1;
3571 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3572 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3575 case TCPOPT_SACK_PERM:
3576 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3577 !estab && sysctl_tcp_sack) {
3578 opt_rx->sack_ok = TCP_SACK_SEEN;
3579 tcp_sack_reset(opt_rx);
3584 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3585 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3587 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3590 #ifdef CONFIG_TCP_MD5SIG
3593 * The MD5 Hash has already been
3594 * checked (see tcp_v{4,6}_do_rcv()).
3599 /* Fast Open option shares code 254 using a
3600 * 16 bits magic number. It's valid only in
3601 * SYN or SYN-ACK with an even size.
3603 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3604 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3605 foc == NULL || !th->syn || (opsize & 1))
3607 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3608 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3609 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3610 memcpy(foc->val, ptr + 2, foc->len);
3611 else if (foc->len != 0)
3621 EXPORT_SYMBOL(tcp_parse_options);
3623 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3625 const __be32 *ptr = (const __be32 *)(th + 1);
3627 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3628 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3629 tp->rx_opt.saw_tstamp = 1;
3631 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3634 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3636 tp->rx_opt.rcv_tsecr = 0;
3642 /* Fast parse options. This hopes to only see timestamps.
3643 * If it is wrong it falls back on tcp_parse_options().
3645 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3646 const struct tcphdr *th, struct tcp_sock *tp)
3648 /* In the spirit of fast parsing, compare doff directly to constant
3649 * values. Because equality is used, short doff can be ignored here.
3651 if (th->doff == (sizeof(*th) / 4)) {
3652 tp->rx_opt.saw_tstamp = 0;
3654 } else if (tp->rx_opt.tstamp_ok &&
3655 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3656 if (tcp_parse_aligned_timestamp(tp, th))
3660 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3661 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3662 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3667 #ifdef CONFIG_TCP_MD5SIG
3669 * Parse MD5 Signature option
3671 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3673 int length = (th->doff << 2) - sizeof(*th);
3674 const u8 *ptr = (const u8 *)(th + 1);
3676 /* If the TCP option is too short, we can short cut */
3677 if (length < TCPOLEN_MD5SIG)
3680 while (length > 0) {
3681 int opcode = *ptr++;
3692 if (opsize < 2 || opsize > length)
3694 if (opcode == TCPOPT_MD5SIG)
3695 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3702 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3705 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3707 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3708 * it can pass through stack. So, the following predicate verifies that
3709 * this segment is not used for anything but congestion avoidance or
3710 * fast retransmit. Moreover, we even are able to eliminate most of such
3711 * second order effects, if we apply some small "replay" window (~RTO)
3712 * to timestamp space.
3714 * All these measures still do not guarantee that we reject wrapped ACKs
3715 * on networks with high bandwidth, when sequence space is recycled fastly,
3716 * but it guarantees that such events will be very rare and do not affect
3717 * connection seriously. This doesn't look nice, but alas, PAWS is really
3720 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3721 * states that events when retransmit arrives after original data are rare.
3722 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3723 * the biggest problem on large power networks even with minor reordering.
3724 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3725 * up to bandwidth of 18Gigabit/sec. 8) ]
3728 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3730 const struct tcp_sock *tp = tcp_sk(sk);
3731 const struct tcphdr *th = tcp_hdr(skb);
3732 u32 seq = TCP_SKB_CB(skb)->seq;
3733 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3735 return (/* 1. Pure ACK with correct sequence number. */
3736 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3738 /* 2. ... and duplicate ACK. */
3739 ack == tp->snd_una &&
3741 /* 3. ... and does not update window. */
3742 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3744 /* 4. ... and sits in replay window. */
3745 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3748 static inline bool tcp_paws_discard(const struct sock *sk,
3749 const struct sk_buff *skb)
3751 const struct tcp_sock *tp = tcp_sk(sk);
3753 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3754 !tcp_disordered_ack(sk, skb);
3757 /* Check segment sequence number for validity.
3759 * Segment controls are considered valid, if the segment
3760 * fits to the window after truncation to the window. Acceptability
3761 * of data (and SYN, FIN, of course) is checked separately.
3762 * See tcp_data_queue(), for example.
3764 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3765 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3766 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3767 * (borrowed from freebsd)
3770 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3772 return !before(end_seq, tp->rcv_wup) &&
3773 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3776 /* When we get a reset we do this. */
3777 void tcp_reset(struct sock *sk)
3779 /* We want the right error as BSD sees it (and indeed as we do). */
3780 switch (sk->sk_state) {
3782 sk->sk_err = ECONNREFUSED;
3784 case TCP_CLOSE_WAIT:
3790 sk->sk_err = ECONNRESET;
3792 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3795 if (!sock_flag(sk, SOCK_DEAD))
3796 sk->sk_error_report(sk);
3802 * Process the FIN bit. This now behaves as it is supposed to work
3803 * and the FIN takes effect when it is validly part of sequence
3804 * space. Not before when we get holes.
3806 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3807 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3810 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3811 * close and we go into CLOSING (and later onto TIME-WAIT)
3813 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3815 static void tcp_fin(struct sock *sk)
3817 struct tcp_sock *tp = tcp_sk(sk);
3819 inet_csk_schedule_ack(sk);
3821 sk->sk_shutdown |= RCV_SHUTDOWN;
3822 sock_set_flag(sk, SOCK_DONE);
3824 switch (sk->sk_state) {
3826 case TCP_ESTABLISHED:
3827 /* Move to CLOSE_WAIT */
3828 tcp_set_state(sk, TCP_CLOSE_WAIT);
3829 inet_csk(sk)->icsk_ack.pingpong = 1;
3832 case TCP_CLOSE_WAIT:
3834 /* Received a retransmission of the FIN, do
3839 /* RFC793: Remain in the LAST-ACK state. */
3843 /* This case occurs when a simultaneous close
3844 * happens, we must ack the received FIN and
3845 * enter the CLOSING state.
3848 tcp_set_state(sk, TCP_CLOSING);
3851 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3853 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3856 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3857 * cases we should never reach this piece of code.
3859 pr_err("%s: Impossible, sk->sk_state=%d\n",
3860 __func__, sk->sk_state);
3864 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3865 * Probably, we should reset in this case. For now drop them.
3867 __skb_queue_purge(&tp->out_of_order_queue);
3868 if (tcp_is_sack(tp))
3869 tcp_sack_reset(&tp->rx_opt);
3872 if (!sock_flag(sk, SOCK_DEAD)) {
3873 sk->sk_state_change(sk);
3875 /* Do not send POLL_HUP for half duplex close. */
3876 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3877 sk->sk_state == TCP_CLOSE)
3878 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3880 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3884 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3887 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3888 if (before(seq, sp->start_seq))
3889 sp->start_seq = seq;
3890 if (after(end_seq, sp->end_seq))
3891 sp->end_seq = end_seq;
3897 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3899 struct tcp_sock *tp = tcp_sk(sk);
3901 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3904 if (before(seq, tp->rcv_nxt))
3905 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3907 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3909 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3911 tp->rx_opt.dsack = 1;
3912 tp->duplicate_sack[0].start_seq = seq;
3913 tp->duplicate_sack[0].end_seq = end_seq;
3917 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3919 struct tcp_sock *tp = tcp_sk(sk);
3921 if (!tp->rx_opt.dsack)
3922 tcp_dsack_set(sk, seq, end_seq);
3924 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3927 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3929 struct tcp_sock *tp = tcp_sk(sk);
3931 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3932 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3933 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3934 tcp_enter_quickack_mode(sk);
3936 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3937 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3939 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3940 end_seq = tp->rcv_nxt;
3941 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3948 /* These routines update the SACK block as out-of-order packets arrive or
3949 * in-order packets close up the sequence space.
3951 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3954 struct tcp_sack_block *sp = &tp->selective_acks[0];
3955 struct tcp_sack_block *swalk = sp + 1;
3957 /* See if the recent change to the first SACK eats into
3958 * or hits the sequence space of other SACK blocks, if so coalesce.
3960 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3961 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3964 /* Zap SWALK, by moving every further SACK up by one slot.
3965 * Decrease num_sacks.
3967 tp->rx_opt.num_sacks--;
3968 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3972 this_sack++, swalk++;
3976 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3978 struct tcp_sock *tp = tcp_sk(sk);
3979 struct tcp_sack_block *sp = &tp->selective_acks[0];
3980 int cur_sacks = tp->rx_opt.num_sacks;
3986 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3987 if (tcp_sack_extend(sp, seq, end_seq)) {
3988 /* Rotate this_sack to the first one. */
3989 for (; this_sack > 0; this_sack--, sp--)
3990 swap(*sp, *(sp - 1));
3992 tcp_sack_maybe_coalesce(tp);
3997 /* Could not find an adjacent existing SACK, build a new one,
3998 * put it at the front, and shift everyone else down. We
3999 * always know there is at least one SACK present already here.
4001 * If the sack array is full, forget about the last one.
4003 if (this_sack >= TCP_NUM_SACKS) {
4005 tp->rx_opt.num_sacks--;
4008 for (; this_sack > 0; this_sack--, sp--)
4012 /* Build the new head SACK, and we're done. */
4013 sp->start_seq = seq;
4014 sp->end_seq = end_seq;
4015 tp->rx_opt.num_sacks++;
4018 /* RCV.NXT advances, some SACKs should be eaten. */
4020 static void tcp_sack_remove(struct tcp_sock *tp)
4022 struct tcp_sack_block *sp = &tp->selective_acks[0];
4023 int num_sacks = tp->rx_opt.num_sacks;
4026 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4027 if (skb_queue_empty(&tp->out_of_order_queue)) {
4028 tp->rx_opt.num_sacks = 0;
4032 for (this_sack = 0; this_sack < num_sacks;) {
4033 /* Check if the start of the sack is covered by RCV.NXT. */
4034 if (!before(tp->rcv_nxt, sp->start_seq)) {
4037 /* RCV.NXT must cover all the block! */
4038 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4040 /* Zap this SACK, by moving forward any other SACKS. */
4041 for (i=this_sack+1; i < num_sacks; i++)
4042 tp->selective_acks[i-1] = tp->selective_acks[i];
4049 tp->rx_opt.num_sacks = num_sacks;
4052 /* This one checks to see if we can put data from the
4053 * out_of_order queue into the receive_queue.
4055 static void tcp_ofo_queue(struct sock *sk)
4057 struct tcp_sock *tp = tcp_sk(sk);
4058 __u32 dsack_high = tp->rcv_nxt;
4059 struct sk_buff *skb;
4061 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4062 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4065 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4066 __u32 dsack = dsack_high;
4067 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4068 dsack_high = TCP_SKB_CB(skb)->end_seq;
4069 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4072 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4073 SOCK_DEBUG(sk, "ofo packet was already received\n");
4074 __skb_unlink(skb, &tp->out_of_order_queue);
4078 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4079 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4080 TCP_SKB_CB(skb)->end_seq);
4082 __skb_unlink(skb, &tp->out_of_order_queue);
4083 __skb_queue_tail(&sk->sk_receive_queue, skb);
4084 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4085 if (tcp_hdr(skb)->fin)
4090 static bool tcp_prune_ofo_queue(struct sock *sk);
4091 static int tcp_prune_queue(struct sock *sk);
4093 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4096 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4097 !sk_rmem_schedule(sk, skb, size)) {
4099 if (tcp_prune_queue(sk) < 0)
4102 if (!sk_rmem_schedule(sk, skb, size)) {
4103 if (!tcp_prune_ofo_queue(sk))
4106 if (!sk_rmem_schedule(sk, skb, size))
4114 * tcp_try_coalesce - try to merge skb to prior one
4117 * @from: buffer to add in queue
4118 * @fragstolen: pointer to boolean
4120 * Before queueing skb @from after @to, try to merge them
4121 * to reduce overall memory use and queue lengths, if cost is small.
4122 * Packets in ofo or receive queues can stay a long time.
4123 * Better try to coalesce them right now to avoid future collapses.
4124 * Returns true if caller should free @from instead of queueing it
4126 static bool tcp_try_coalesce(struct sock *sk,
4128 struct sk_buff *from,
4133 *fragstolen = false;
4135 if (tcp_hdr(from)->fin)
4138 /* Its possible this segment overlaps with prior segment in queue */
4139 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4142 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4145 atomic_add(delta, &sk->sk_rmem_alloc);
4146 sk_mem_charge(sk, delta);
4147 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4148 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4149 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4153 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4155 struct tcp_sock *tp = tcp_sk(sk);
4156 struct sk_buff *skb1;
4159 TCP_ECN_check_ce(tp, skb);
4161 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4162 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4167 /* Disable header prediction. */
4169 inet_csk_schedule_ack(sk);
4171 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4172 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4173 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4175 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4177 /* Initial out of order segment, build 1 SACK. */
4178 if (tcp_is_sack(tp)) {
4179 tp->rx_opt.num_sacks = 1;
4180 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4181 tp->selective_acks[0].end_seq =
4182 TCP_SKB_CB(skb)->end_seq;
4184 __skb_queue_head(&tp->out_of_order_queue, skb);
4188 seq = TCP_SKB_CB(skb)->seq;
4189 end_seq = TCP_SKB_CB(skb)->end_seq;
4191 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4194 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4195 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4197 kfree_skb_partial(skb, fragstolen);
4201 if (!tp->rx_opt.num_sacks ||
4202 tp->selective_acks[0].end_seq != seq)
4205 /* Common case: data arrive in order after hole. */
4206 tp->selective_acks[0].end_seq = end_seq;
4210 /* Find place to insert this segment. */
4212 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4214 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4218 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4221 /* Do skb overlap to previous one? */
4222 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4223 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4224 /* All the bits are present. Drop. */
4225 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4228 tcp_dsack_set(sk, seq, end_seq);
4231 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4232 /* Partial overlap. */
4233 tcp_dsack_set(sk, seq,
4234 TCP_SKB_CB(skb1)->end_seq);
4236 if (skb_queue_is_first(&tp->out_of_order_queue,
4240 skb1 = skb_queue_prev(
4241 &tp->out_of_order_queue,
4246 __skb_queue_head(&tp->out_of_order_queue, skb);
4248 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4250 /* And clean segments covered by new one as whole. */
4251 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4252 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4254 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4256 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4257 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4261 __skb_unlink(skb1, &tp->out_of_order_queue);
4262 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4263 TCP_SKB_CB(skb1)->end_seq);
4264 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4269 if (tcp_is_sack(tp))
4270 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4273 skb_set_owner_r(skb, sk);
4276 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4280 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4282 __skb_pull(skb, hdrlen);
4284 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4285 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4287 __skb_queue_tail(&sk->sk_receive_queue, skb);
4288 skb_set_owner_r(skb, sk);
4293 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4295 struct sk_buff *skb = NULL;
4302 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4306 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4309 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4310 skb_reset_transport_header(skb);
4311 memset(th, 0, sizeof(*th));
4313 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4316 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4317 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4318 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4320 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4321 WARN_ON_ONCE(fragstolen); /* should not happen */
4332 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4334 const struct tcphdr *th = tcp_hdr(skb);
4335 struct tcp_sock *tp = tcp_sk(sk);
4337 bool fragstolen = false;
4339 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4343 __skb_pull(skb, th->doff * 4);
4345 TCP_ECN_accept_cwr(tp, skb);
4347 tp->rx_opt.dsack = 0;
4349 /* Queue data for delivery to the user.
4350 * Packets in sequence go to the receive queue.
4351 * Out of sequence packets to the out_of_order_queue.
4353 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4354 if (tcp_receive_window(tp) == 0)
4357 /* Ok. In sequence. In window. */
4358 if (tp->ucopy.task == current &&
4359 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4360 sock_owned_by_user(sk) && !tp->urg_data) {
4361 int chunk = min_t(unsigned int, skb->len,
4364 __set_current_state(TASK_RUNNING);
4367 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4368 tp->ucopy.len -= chunk;
4369 tp->copied_seq += chunk;
4370 eaten = (chunk == skb->len);
4371 tcp_rcv_space_adjust(sk);
4379 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4382 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4384 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4386 tcp_event_data_recv(sk, skb);
4390 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4393 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4394 * gap in queue is filled.
4396 if (skb_queue_empty(&tp->out_of_order_queue))
4397 inet_csk(sk)->icsk_ack.pingpong = 0;
4400 if (tp->rx_opt.num_sacks)
4401 tcp_sack_remove(tp);
4403 tcp_fast_path_check(sk);
4406 kfree_skb_partial(skb, fragstolen);
4407 if (!sock_flag(sk, SOCK_DEAD))
4408 sk->sk_data_ready(sk, 0);
4412 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4413 /* A retransmit, 2nd most common case. Force an immediate ack. */
4414 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4415 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4418 tcp_enter_quickack_mode(sk);
4419 inet_csk_schedule_ack(sk);
4425 /* Out of window. F.e. zero window probe. */
4426 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4429 tcp_enter_quickack_mode(sk);
4431 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4432 /* Partial packet, seq < rcv_next < end_seq */
4433 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4434 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4435 TCP_SKB_CB(skb)->end_seq);
4437 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4439 /* If window is closed, drop tail of packet. But after
4440 * remembering D-SACK for its head made in previous line.
4442 if (!tcp_receive_window(tp))
4447 tcp_data_queue_ofo(sk, skb);
4450 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4451 struct sk_buff_head *list)
4453 struct sk_buff *next = NULL;
4455 if (!skb_queue_is_last(list, skb))
4456 next = skb_queue_next(list, skb);
4458 __skb_unlink(skb, list);
4460 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4465 /* Collapse contiguous sequence of skbs head..tail with
4466 * sequence numbers start..end.
4468 * If tail is NULL, this means until the end of the list.
4470 * Segments with FIN/SYN are not collapsed (only because this
4474 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4475 struct sk_buff *head, struct sk_buff *tail,
4478 struct sk_buff *skb, *n;
4481 /* First, check that queue is collapsible and find
4482 * the point where collapsing can be useful. */
4486 skb_queue_walk_from_safe(list, skb, n) {
4489 /* No new bits? It is possible on ofo queue. */
4490 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4491 skb = tcp_collapse_one(sk, skb, list);
4497 /* The first skb to collapse is:
4499 * - bloated or contains data before "start" or
4500 * overlaps to the next one.
4502 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4503 (tcp_win_from_space(skb->truesize) > skb->len ||
4504 before(TCP_SKB_CB(skb)->seq, start))) {
4505 end_of_skbs = false;
4509 if (!skb_queue_is_last(list, skb)) {
4510 struct sk_buff *next = skb_queue_next(list, skb);
4512 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4513 end_of_skbs = false;
4518 /* Decided to skip this, advance start seq. */
4519 start = TCP_SKB_CB(skb)->end_seq;
4521 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4524 while (before(start, end)) {
4525 struct sk_buff *nskb;
4526 unsigned int header = skb_headroom(skb);
4527 int copy = SKB_MAX_ORDER(header, 0);
4529 /* Too big header? This can happen with IPv6. */
4532 if (end - start < copy)
4534 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4538 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4539 skb_set_network_header(nskb, (skb_network_header(skb) -
4541 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4543 skb_reserve(nskb, header);
4544 memcpy(nskb->head, skb->head, header);
4545 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4546 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4547 __skb_queue_before(list, skb, nskb);
4548 skb_set_owner_r(nskb, sk);
4550 /* Copy data, releasing collapsed skbs. */
4552 int offset = start - TCP_SKB_CB(skb)->seq;
4553 int size = TCP_SKB_CB(skb)->end_seq - start;
4557 size = min(copy, size);
4558 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4560 TCP_SKB_CB(nskb)->end_seq += size;
4564 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4565 skb = tcp_collapse_one(sk, skb, list);
4568 tcp_hdr(skb)->syn ||
4576 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4577 * and tcp_collapse() them until all the queue is collapsed.
4579 static void tcp_collapse_ofo_queue(struct sock *sk)
4581 struct tcp_sock *tp = tcp_sk(sk);
4582 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4583 struct sk_buff *head;
4589 start = TCP_SKB_CB(skb)->seq;
4590 end = TCP_SKB_CB(skb)->end_seq;
4594 struct sk_buff *next = NULL;
4596 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4597 next = skb_queue_next(&tp->out_of_order_queue, skb);
4600 /* Segment is terminated when we see gap or when
4601 * we are at the end of all the queue. */
4603 after(TCP_SKB_CB(skb)->seq, end) ||
4604 before(TCP_SKB_CB(skb)->end_seq, start)) {
4605 tcp_collapse(sk, &tp->out_of_order_queue,
4606 head, skb, start, end);
4610 /* Start new segment */
4611 start = TCP_SKB_CB(skb)->seq;
4612 end = TCP_SKB_CB(skb)->end_seq;
4614 if (before(TCP_SKB_CB(skb)->seq, start))
4615 start = TCP_SKB_CB(skb)->seq;
4616 if (after(TCP_SKB_CB(skb)->end_seq, end))
4617 end = TCP_SKB_CB(skb)->end_seq;
4623 * Purge the out-of-order queue.
4624 * Return true if queue was pruned.
4626 static bool tcp_prune_ofo_queue(struct sock *sk)
4628 struct tcp_sock *tp = tcp_sk(sk);
4631 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4632 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4633 __skb_queue_purge(&tp->out_of_order_queue);
4635 /* Reset SACK state. A conforming SACK implementation will
4636 * do the same at a timeout based retransmit. When a connection
4637 * is in a sad state like this, we care only about integrity
4638 * of the connection not performance.
4640 if (tp->rx_opt.sack_ok)
4641 tcp_sack_reset(&tp->rx_opt);
4648 /* Reduce allocated memory if we can, trying to get
4649 * the socket within its memory limits again.
4651 * Return less than zero if we should start dropping frames
4652 * until the socket owning process reads some of the data
4653 * to stabilize the situation.
4655 static int tcp_prune_queue(struct sock *sk)
4657 struct tcp_sock *tp = tcp_sk(sk);
4659 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4661 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4663 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4664 tcp_clamp_window(sk);
4665 else if (sk_under_memory_pressure(sk))
4666 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4668 tcp_collapse_ofo_queue(sk);
4669 if (!skb_queue_empty(&sk->sk_receive_queue))
4670 tcp_collapse(sk, &sk->sk_receive_queue,
4671 skb_peek(&sk->sk_receive_queue),
4673 tp->copied_seq, tp->rcv_nxt);
4676 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4679 /* Collapsing did not help, destructive actions follow.
4680 * This must not ever occur. */
4682 tcp_prune_ofo_queue(sk);
4684 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4687 /* If we are really being abused, tell the caller to silently
4688 * drop receive data on the floor. It will get retransmitted
4689 * and hopefully then we'll have sufficient space.
4691 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4693 /* Massive buffer overcommit. */
4698 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4699 * As additional protections, we do not touch cwnd in retransmission phases,
4700 * and if application hit its sndbuf limit recently.
4702 void tcp_cwnd_application_limited(struct sock *sk)
4704 struct tcp_sock *tp = tcp_sk(sk);
4706 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4707 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4708 /* Limited by application or receiver window. */
4709 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4710 u32 win_used = max(tp->snd_cwnd_used, init_win);
4711 if (win_used < tp->snd_cwnd) {
4712 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4713 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4715 tp->snd_cwnd_used = 0;
4717 tp->snd_cwnd_stamp = tcp_time_stamp;
4720 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4722 const struct tcp_sock *tp = tcp_sk(sk);
4724 /* If the user specified a specific send buffer setting, do
4727 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4730 /* If we are under global TCP memory pressure, do not expand. */
4731 if (sk_under_memory_pressure(sk))
4734 /* If we are under soft global TCP memory pressure, do not expand. */
4735 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4738 /* If we filled the congestion window, do not expand. */
4739 if (tp->packets_out >= tp->snd_cwnd)
4745 /* When incoming ACK allowed to free some skb from write_queue,
4746 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4747 * on the exit from tcp input handler.
4749 * PROBLEM: sndbuf expansion does not work well with largesend.
4751 static void tcp_new_space(struct sock *sk)
4753 struct tcp_sock *tp = tcp_sk(sk);
4755 if (tcp_should_expand_sndbuf(sk)) {
4756 int sndmem = SKB_TRUESIZE(max_t(u32,
4757 tp->rx_opt.mss_clamp,
4760 int demanded = max_t(unsigned int, tp->snd_cwnd,
4761 tp->reordering + 1);
4762 sndmem *= 2 * demanded;
4763 if (sndmem > sk->sk_sndbuf)
4764 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4765 tp->snd_cwnd_stamp = tcp_time_stamp;
4768 sk->sk_write_space(sk);
4771 static void tcp_check_space(struct sock *sk)
4773 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4774 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4775 if (sk->sk_socket &&
4776 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4781 static inline void tcp_data_snd_check(struct sock *sk)
4783 tcp_push_pending_frames(sk);
4784 tcp_check_space(sk);
4788 * Check if sending an ack is needed.
4790 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4792 struct tcp_sock *tp = tcp_sk(sk);
4794 /* More than one full frame received... */
4795 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4796 /* ... and right edge of window advances far enough.
4797 * (tcp_recvmsg() will send ACK otherwise). Or...
4799 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4800 /* We ACK each frame or... */
4801 tcp_in_quickack_mode(sk) ||
4802 /* We have out of order data. */
4803 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4804 /* Then ack it now */
4807 /* Else, send delayed ack. */
4808 tcp_send_delayed_ack(sk);
4812 static inline void tcp_ack_snd_check(struct sock *sk)
4814 if (!inet_csk_ack_scheduled(sk)) {
4815 /* We sent a data segment already. */
4818 __tcp_ack_snd_check(sk, 1);
4822 * This routine is only called when we have urgent data
4823 * signaled. Its the 'slow' part of tcp_urg. It could be
4824 * moved inline now as tcp_urg is only called from one
4825 * place. We handle URGent data wrong. We have to - as
4826 * BSD still doesn't use the correction from RFC961.
4827 * For 1003.1g we should support a new option TCP_STDURG to permit
4828 * either form (or just set the sysctl tcp_stdurg).
4831 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4833 struct tcp_sock *tp = tcp_sk(sk);
4834 u32 ptr = ntohs(th->urg_ptr);
4836 if (ptr && !sysctl_tcp_stdurg)
4838 ptr += ntohl(th->seq);
4840 /* Ignore urgent data that we've already seen and read. */
4841 if (after(tp->copied_seq, ptr))
4844 /* Do not replay urg ptr.
4846 * NOTE: interesting situation not covered by specs.
4847 * Misbehaving sender may send urg ptr, pointing to segment,
4848 * which we already have in ofo queue. We are not able to fetch
4849 * such data and will stay in TCP_URG_NOTYET until will be eaten
4850 * by recvmsg(). Seems, we are not obliged to handle such wicked
4851 * situations. But it is worth to think about possibility of some
4852 * DoSes using some hypothetical application level deadlock.
4854 if (before(ptr, tp->rcv_nxt))
4857 /* Do we already have a newer (or duplicate) urgent pointer? */
4858 if (tp->urg_data && !after(ptr, tp->urg_seq))
4861 /* Tell the world about our new urgent pointer. */
4864 /* We may be adding urgent data when the last byte read was
4865 * urgent. To do this requires some care. We cannot just ignore
4866 * tp->copied_seq since we would read the last urgent byte again
4867 * as data, nor can we alter copied_seq until this data arrives
4868 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4870 * NOTE. Double Dutch. Rendering to plain English: author of comment
4871 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4872 * and expect that both A and B disappear from stream. This is _wrong_.
4873 * Though this happens in BSD with high probability, this is occasional.
4874 * Any application relying on this is buggy. Note also, that fix "works"
4875 * only in this artificial test. Insert some normal data between A and B and we will
4876 * decline of BSD again. Verdict: it is better to remove to trap
4879 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4880 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4881 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4883 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4884 __skb_unlink(skb, &sk->sk_receive_queue);
4889 tp->urg_data = TCP_URG_NOTYET;
4892 /* Disable header prediction. */
4896 /* This is the 'fast' part of urgent handling. */
4897 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4899 struct tcp_sock *tp = tcp_sk(sk);
4901 /* Check if we get a new urgent pointer - normally not. */
4903 tcp_check_urg(sk, th);
4905 /* Do we wait for any urgent data? - normally not... */
4906 if (tp->urg_data == TCP_URG_NOTYET) {
4907 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4910 /* Is the urgent pointer pointing into this packet? */
4911 if (ptr < skb->len) {
4913 if (skb_copy_bits(skb, ptr, &tmp, 1))
4915 tp->urg_data = TCP_URG_VALID | tmp;
4916 if (!sock_flag(sk, SOCK_DEAD))
4917 sk->sk_data_ready(sk, 0);
4922 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4924 struct tcp_sock *tp = tcp_sk(sk);
4925 int chunk = skb->len - hlen;
4929 if (skb_csum_unnecessary(skb))
4930 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4932 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4936 tp->ucopy.len -= chunk;
4937 tp->copied_seq += chunk;
4938 tcp_rcv_space_adjust(sk);
4945 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4946 struct sk_buff *skb)
4950 if (sock_owned_by_user(sk)) {
4952 result = __tcp_checksum_complete(skb);
4955 result = __tcp_checksum_complete(skb);
4960 static inline bool tcp_checksum_complete_user(struct sock *sk,
4961 struct sk_buff *skb)
4963 return !skb_csum_unnecessary(skb) &&
4964 __tcp_checksum_complete_user(sk, skb);
4967 #ifdef CONFIG_NET_DMA
4968 static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4971 struct tcp_sock *tp = tcp_sk(sk);
4972 int chunk = skb->len - hlen;
4974 bool copied_early = false;
4976 if (tp->ucopy.wakeup)
4979 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4980 tp->ucopy.dma_chan = net_dma_find_channel();
4982 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4984 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4986 tp->ucopy.iov, chunk,
4987 tp->ucopy.pinned_list);
4992 tp->ucopy.dma_cookie = dma_cookie;
4993 copied_early = true;
4995 tp->ucopy.len -= chunk;
4996 tp->copied_seq += chunk;
4997 tcp_rcv_space_adjust(sk);
4999 if ((tp->ucopy.len == 0) ||
5000 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5001 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5002 tp->ucopy.wakeup = 1;
5003 sk->sk_data_ready(sk, 0);
5005 } else if (chunk > 0) {
5006 tp->ucopy.wakeup = 1;
5007 sk->sk_data_ready(sk, 0);
5010 return copied_early;
5012 #endif /* CONFIG_NET_DMA */
5014 /* Does PAWS and seqno based validation of an incoming segment, flags will
5015 * play significant role here.
5017 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5018 const struct tcphdr *th, int syn_inerr)
5020 struct tcp_sock *tp = tcp_sk(sk);
5022 /* RFC1323: H1. Apply PAWS check first. */
5023 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5024 tcp_paws_discard(sk, skb)) {
5026 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5027 tcp_send_dupack(sk, skb);
5030 /* Reset is accepted even if it did not pass PAWS. */
5033 /* Step 1: check sequence number */
5034 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5035 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5036 * (RST) segments are validated by checking their SEQ-fields."
5037 * And page 69: "If an incoming segment is not acceptable,
5038 * an acknowledgment should be sent in reply (unless the RST
5039 * bit is set, if so drop the segment and return)".
5044 tcp_send_dupack(sk, skb);
5049 /* Step 2: check RST bit */
5052 * If sequence number exactly matches RCV.NXT, then
5053 * RESET the connection
5055 * Send a challenge ACK
5057 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5060 tcp_send_challenge_ack(sk);
5064 /* step 3: check security and precedence [ignored] */
5066 /* step 4: Check for a SYN
5067 * RFC 5691 4.2 : Send a challenge ack
5072 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5073 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5074 tcp_send_challenge_ack(sk);
5086 * TCP receive function for the ESTABLISHED state.
5088 * It is split into a fast path and a slow path. The fast path is
5090 * - A zero window was announced from us - zero window probing
5091 * is only handled properly in the slow path.
5092 * - Out of order segments arrived.
5093 * - Urgent data is expected.
5094 * - There is no buffer space left
5095 * - Unexpected TCP flags/window values/header lengths are received
5096 * (detected by checking the TCP header against pred_flags)
5097 * - Data is sent in both directions. Fast path only supports pure senders
5098 * or pure receivers (this means either the sequence number or the ack
5099 * value must stay constant)
5100 * - Unexpected TCP option.
5102 * When these conditions are not satisfied it drops into a standard
5103 * receive procedure patterned after RFC793 to handle all cases.
5104 * The first three cases are guaranteed by proper pred_flags setting,
5105 * the rest is checked inline. Fast processing is turned on in
5106 * tcp_data_queue when everything is OK.
5108 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5109 const struct tcphdr *th, unsigned int len)
5111 struct tcp_sock *tp = tcp_sk(sk);
5113 if (unlikely(sk->sk_rx_dst == NULL))
5114 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5116 * Header prediction.
5117 * The code loosely follows the one in the famous
5118 * "30 instruction TCP receive" Van Jacobson mail.
5120 * Van's trick is to deposit buffers into socket queue
5121 * on a device interrupt, to call tcp_recv function
5122 * on the receive process context and checksum and copy
5123 * the buffer to user space. smart...
5125 * Our current scheme is not silly either but we take the
5126 * extra cost of the net_bh soft interrupt processing...
5127 * We do checksum and copy also but from device to kernel.
5130 tp->rx_opt.saw_tstamp = 0;
5132 /* pred_flags is 0xS?10 << 16 + snd_wnd
5133 * if header_prediction is to be made
5134 * 'S' will always be tp->tcp_header_len >> 2
5135 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5136 * turn it off (when there are holes in the receive
5137 * space for instance)
5138 * PSH flag is ignored.
5141 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5142 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5143 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5144 int tcp_header_len = tp->tcp_header_len;
5146 /* Timestamp header prediction: tcp_header_len
5147 * is automatically equal to th->doff*4 due to pred_flags
5151 /* Check timestamp */
5152 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5153 /* No? Slow path! */
5154 if (!tcp_parse_aligned_timestamp(tp, th))
5157 /* If PAWS failed, check it more carefully in slow path */
5158 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5161 /* DO NOT update ts_recent here, if checksum fails
5162 * and timestamp was corrupted part, it will result
5163 * in a hung connection since we will drop all
5164 * future packets due to the PAWS test.
5168 if (len <= tcp_header_len) {
5169 /* Bulk data transfer: sender */
5170 if (len == tcp_header_len) {
5171 /* Predicted packet is in window by definition.
5172 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5173 * Hence, check seq<=rcv_wup reduces to:
5175 if (tcp_header_len ==
5176 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5177 tp->rcv_nxt == tp->rcv_wup)
5178 tcp_store_ts_recent(tp);
5180 /* We know that such packets are checksummed
5183 tcp_ack(sk, skb, 0);
5185 tcp_data_snd_check(sk);
5187 } else { /* Header too small */
5188 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5193 int copied_early = 0;
5194 bool fragstolen = false;
5196 if (tp->copied_seq == tp->rcv_nxt &&
5197 len - tcp_header_len <= tp->ucopy.len) {
5198 #ifdef CONFIG_NET_DMA
5199 if (tp->ucopy.task == current &&
5200 sock_owned_by_user(sk) &&
5201 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5206 if (tp->ucopy.task == current &&
5207 sock_owned_by_user(sk) && !copied_early) {
5208 __set_current_state(TASK_RUNNING);
5210 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5214 /* Predicted packet is in window by definition.
5215 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5216 * Hence, check seq<=rcv_wup reduces to:
5218 if (tcp_header_len ==
5219 (sizeof(struct tcphdr) +
5220 TCPOLEN_TSTAMP_ALIGNED) &&
5221 tp->rcv_nxt == tp->rcv_wup)
5222 tcp_store_ts_recent(tp);
5224 tcp_rcv_rtt_measure_ts(sk, skb);
5226 __skb_pull(skb, tcp_header_len);
5227 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5228 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5231 tcp_cleanup_rbuf(sk, skb->len);
5234 if (tcp_checksum_complete_user(sk, skb))
5237 if ((int)skb->truesize > sk->sk_forward_alloc)
5240 /* Predicted packet is in window by definition.
5241 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5242 * Hence, check seq<=rcv_wup reduces to:
5244 if (tcp_header_len ==
5245 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5246 tp->rcv_nxt == tp->rcv_wup)
5247 tcp_store_ts_recent(tp);
5249 tcp_rcv_rtt_measure_ts(sk, skb);
5251 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5253 /* Bulk data transfer: receiver */
5254 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5258 tcp_event_data_recv(sk, skb);
5260 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5261 /* Well, only one small jumplet in fast path... */
5262 tcp_ack(sk, skb, FLAG_DATA);
5263 tcp_data_snd_check(sk);
5264 if (!inet_csk_ack_scheduled(sk))
5268 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5269 __tcp_ack_snd_check(sk, 0);
5271 #ifdef CONFIG_NET_DMA
5273 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5277 kfree_skb_partial(skb, fragstolen);
5278 sk->sk_data_ready(sk, 0);
5284 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5287 if (!th->ack && !th->rst)
5291 * Standard slow path.
5294 if (!tcp_validate_incoming(sk, skb, th, 1))
5298 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5301 tcp_rcv_rtt_measure_ts(sk, skb);
5303 /* Process urgent data. */
5304 tcp_urg(sk, skb, th);
5306 /* step 7: process the segment text */
5307 tcp_data_queue(sk, skb);
5309 tcp_data_snd_check(sk);
5310 tcp_ack_snd_check(sk);
5314 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5315 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5321 EXPORT_SYMBOL(tcp_rcv_established);
5323 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5325 struct tcp_sock *tp = tcp_sk(sk);
5326 struct inet_connection_sock *icsk = inet_csk(sk);
5328 tcp_set_state(sk, TCP_ESTABLISHED);
5331 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5332 security_inet_conn_established(sk, skb);
5335 /* Make sure socket is routed, for correct metrics. */
5336 icsk->icsk_af_ops->rebuild_header(sk);
5338 tcp_init_metrics(sk);
5340 tcp_init_congestion_control(sk);
5342 /* Prevent spurious tcp_cwnd_restart() on first data
5345 tp->lsndtime = tcp_time_stamp;
5347 tcp_init_buffer_space(sk);
5349 if (sock_flag(sk, SOCK_KEEPOPEN))
5350 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5352 if (!tp->rx_opt.snd_wscale)
5353 __tcp_fast_path_on(tp, tp->snd_wnd);
5357 if (!sock_flag(sk, SOCK_DEAD)) {
5358 sk->sk_state_change(sk);
5359 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5363 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5364 struct tcp_fastopen_cookie *cookie)
5366 struct tcp_sock *tp = tcp_sk(sk);
5367 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5368 u16 mss = tp->rx_opt.mss_clamp;
5371 if (mss == tp->rx_opt.user_mss) {
5372 struct tcp_options_received opt;
5374 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5375 tcp_clear_options(&opt);
5376 opt.user_mss = opt.mss_clamp = 0;
5377 tcp_parse_options(synack, &opt, 0, NULL);
5378 mss = opt.mss_clamp;
5381 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5384 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5385 * the remote receives only the retransmitted (regular) SYNs: either
5386 * the original SYN-data or the corresponding SYN-ACK is lost.
5388 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5390 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5392 if (data) { /* Retransmit unacked data in SYN */
5393 tcp_for_write_queue_from(data, sk) {
5394 if (data == tcp_send_head(sk) ||
5395 __tcp_retransmit_skb(sk, data))
5401 tp->syn_data_acked = tp->syn_data;
5405 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5406 const struct tcphdr *th, unsigned int len)
5408 struct inet_connection_sock *icsk = inet_csk(sk);
5409 struct tcp_sock *tp = tcp_sk(sk);
5410 struct tcp_fastopen_cookie foc = { .len = -1 };
5411 int saved_clamp = tp->rx_opt.mss_clamp;
5413 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5414 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5415 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5419 * "If the state is SYN-SENT then
5420 * first check the ACK bit
5421 * If the ACK bit is set
5422 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5423 * a reset (unless the RST bit is set, if so drop
5424 * the segment and return)"
5426 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5427 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5428 goto reset_and_undo;
5430 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5431 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5433 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5434 goto reset_and_undo;
5437 /* Now ACK is acceptable.
5439 * "If the RST bit is set
5440 * If the ACK was acceptable then signal the user "error:
5441 * connection reset", drop the segment, enter CLOSED state,
5442 * delete TCB, and return."
5451 * "fifth, if neither of the SYN or RST bits is set then
5452 * drop the segment and return."
5458 goto discard_and_undo;
5461 * "If the SYN bit is on ...
5462 * are acceptable then ...
5463 * (our SYN has been ACKed), change the connection
5464 * state to ESTABLISHED..."
5467 TCP_ECN_rcv_synack(tp, th);
5469 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5470 tcp_ack(sk, skb, FLAG_SLOWPATH);
5472 /* Ok.. it's good. Set up sequence numbers and
5473 * move to established.
5475 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5476 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5478 /* RFC1323: The window in SYN & SYN/ACK segments is
5481 tp->snd_wnd = ntohs(th->window);
5483 if (!tp->rx_opt.wscale_ok) {
5484 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5485 tp->window_clamp = min(tp->window_clamp, 65535U);
5488 if (tp->rx_opt.saw_tstamp) {
5489 tp->rx_opt.tstamp_ok = 1;
5490 tp->tcp_header_len =
5491 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5492 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5493 tcp_store_ts_recent(tp);
5495 tp->tcp_header_len = sizeof(struct tcphdr);
5498 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5499 tcp_enable_fack(tp);
5502 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5503 tcp_initialize_rcv_mss(sk);
5505 /* Remember, tcp_poll() does not lock socket!
5506 * Change state from SYN-SENT only after copied_seq
5507 * is initialized. */
5508 tp->copied_seq = tp->rcv_nxt;
5512 tcp_finish_connect(sk, skb);
5514 if ((tp->syn_fastopen || tp->syn_data) &&
5515 tcp_rcv_fastopen_synack(sk, skb, &foc))
5518 if (sk->sk_write_pending ||
5519 icsk->icsk_accept_queue.rskq_defer_accept ||
5520 icsk->icsk_ack.pingpong) {
5521 /* Save one ACK. Data will be ready after
5522 * several ticks, if write_pending is set.
5524 * It may be deleted, but with this feature tcpdumps
5525 * look so _wonderfully_ clever, that I was not able
5526 * to stand against the temptation 8) --ANK
5528 inet_csk_schedule_ack(sk);
5529 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5530 tcp_enter_quickack_mode(sk);
5531 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5532 TCP_DELACK_MAX, TCP_RTO_MAX);
5543 /* No ACK in the segment */
5547 * "If the RST bit is set
5549 * Otherwise (no ACK) drop the segment and return."
5552 goto discard_and_undo;
5556 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5557 tcp_paws_reject(&tp->rx_opt, 0))
5558 goto discard_and_undo;
5561 /* We see SYN without ACK. It is attempt of
5562 * simultaneous connect with crossed SYNs.
5563 * Particularly, it can be connect to self.
5565 tcp_set_state(sk, TCP_SYN_RECV);
5567 if (tp->rx_opt.saw_tstamp) {
5568 tp->rx_opt.tstamp_ok = 1;
5569 tcp_store_ts_recent(tp);
5570 tp->tcp_header_len =
5571 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5573 tp->tcp_header_len = sizeof(struct tcphdr);
5576 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5577 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5579 /* RFC1323: The window in SYN & SYN/ACK segments is
5582 tp->snd_wnd = ntohs(th->window);
5583 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5584 tp->max_window = tp->snd_wnd;
5586 TCP_ECN_rcv_syn(tp, th);
5589 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5590 tcp_initialize_rcv_mss(sk);
5592 tcp_send_synack(sk);
5594 /* Note, we could accept data and URG from this segment.
5595 * There are no obstacles to make this (except that we must
5596 * either change tcp_recvmsg() to prevent it from returning data
5597 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5599 * However, if we ignore data in ACKless segments sometimes,
5600 * we have no reasons to accept it sometimes.
5601 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5602 * is not flawless. So, discard packet for sanity.
5603 * Uncomment this return to process the data.
5610 /* "fifth, if neither of the SYN or RST bits is set then
5611 * drop the segment and return."
5615 tcp_clear_options(&tp->rx_opt);
5616 tp->rx_opt.mss_clamp = saved_clamp;
5620 tcp_clear_options(&tp->rx_opt);
5621 tp->rx_opt.mss_clamp = saved_clamp;
5626 * This function implements the receiving procedure of RFC 793 for
5627 * all states except ESTABLISHED and TIME_WAIT.
5628 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5629 * address independent.
5632 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5633 const struct tcphdr *th, unsigned int len)
5635 struct tcp_sock *tp = tcp_sk(sk);
5636 struct inet_connection_sock *icsk = inet_csk(sk);
5637 struct request_sock *req;
5640 tp->rx_opt.saw_tstamp = 0;
5642 switch (sk->sk_state) {
5656 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5659 /* Now we have several options: In theory there is
5660 * nothing else in the frame. KA9Q has an option to
5661 * send data with the syn, BSD accepts data with the
5662 * syn up to the [to be] advertised window and
5663 * Solaris 2.1 gives you a protocol error. For now
5664 * we just ignore it, that fits the spec precisely
5665 * and avoids incompatibilities. It would be nice in
5666 * future to drop through and process the data.
5668 * Now that TTCP is starting to be used we ought to
5670 * But, this leaves one open to an easy denial of
5671 * service attack, and SYN cookies can't defend
5672 * against this problem. So, we drop the data
5673 * in the interest of security over speed unless
5674 * it's still in use.
5682 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5686 /* Do step6 onward by hand. */
5687 tcp_urg(sk, skb, th);
5689 tcp_data_snd_check(sk);
5693 req = tp->fastopen_rsk;
5695 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5696 sk->sk_state != TCP_FIN_WAIT1);
5698 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5702 if (!th->ack && !th->rst)
5705 if (!tcp_validate_incoming(sk, skb, th, 0))
5708 /* step 5: check the ACK field */
5710 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5711 FLAG_UPDATE_TS_RECENT) > 0;
5713 switch (sk->sk_state) {
5716 /* Once we leave TCP_SYN_RECV, we no longer
5717 * need req so release it.
5720 tcp_synack_rtt_meas(sk, req);
5721 tp->total_retrans = req->num_retrans;
5723 reqsk_fastopen_remove(sk, req, false);
5725 /* Make sure socket is routed, for
5728 icsk->icsk_af_ops->rebuild_header(sk);
5729 tcp_init_congestion_control(sk);
5732 tcp_init_buffer_space(sk);
5733 tp->copied_seq = tp->rcv_nxt;
5736 tcp_set_state(sk, TCP_ESTABLISHED);
5737 sk->sk_state_change(sk);
5739 /* Note, that this wakeup is only for marginal
5740 * crossed SYN case. Passively open sockets
5741 * are not waked up, because sk->sk_sleep ==
5742 * NULL and sk->sk_socket == NULL.
5746 SOCK_WAKE_IO, POLL_OUT);
5748 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5749 tp->snd_wnd = ntohs(th->window) <<
5750 tp->rx_opt.snd_wscale;
5751 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5753 if (tp->rx_opt.tstamp_ok)
5754 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5757 /* Re-arm the timer because data may
5758 * have been sent out. This is similar
5759 * to the regular data transmission case
5760 * when new data has just been ack'ed.
5762 * (TFO) - we could try to be more
5763 * aggressive and retranmitting any data
5764 * sooner based on when they were sent
5769 tcp_init_metrics(sk);
5771 tcp_update_pacing_rate(sk);
5773 /* Prevent spurious tcp_cwnd_restart() on
5774 * first data packet.
5776 tp->lsndtime = tcp_time_stamp;
5778 tcp_initialize_rcv_mss(sk);
5779 tcp_fast_path_on(tp);
5786 /* If we enter the TCP_FIN_WAIT1 state and we are a
5787 * Fast Open socket and this is the first acceptable
5788 * ACK we have received, this would have acknowledged
5789 * our SYNACK so stop the SYNACK timer.
5792 /* Return RST if ack_seq is invalid.
5793 * Note that RFC793 only says to generate a
5794 * DUPACK for it but for TCP Fast Open it seems
5795 * better to treat this case like TCP_SYN_RECV
5800 /* We no longer need the request sock. */
5801 reqsk_fastopen_remove(sk, req, false);
5804 if (tp->snd_una == tp->write_seq) {
5805 struct dst_entry *dst;
5807 tcp_set_state(sk, TCP_FIN_WAIT2);
5808 sk->sk_shutdown |= SEND_SHUTDOWN;
5810 dst = __sk_dst_get(sk);
5814 if (!sock_flag(sk, SOCK_DEAD))
5815 /* Wake up lingering close() */
5816 sk->sk_state_change(sk);
5820 if (tp->linger2 < 0 ||
5821 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5822 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5824 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5828 tmo = tcp_fin_time(sk);
5829 if (tmo > TCP_TIMEWAIT_LEN) {
5830 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5831 } else if (th->fin || sock_owned_by_user(sk)) {
5832 /* Bad case. We could lose such FIN otherwise.
5833 * It is not a big problem, but it looks confusing
5834 * and not so rare event. We still can lose it now,
5835 * if it spins in bh_lock_sock(), but it is really
5838 inet_csk_reset_keepalive_timer(sk, tmo);
5840 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5848 if (tp->snd_una == tp->write_seq) {
5849 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5855 if (tp->snd_una == tp->write_seq) {
5856 tcp_update_metrics(sk);
5864 /* step 6: check the URG bit */
5865 tcp_urg(sk, skb, th);
5867 /* step 7: process the segment text */
5868 switch (sk->sk_state) {
5869 case TCP_CLOSE_WAIT:
5872 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5876 /* RFC 793 says to queue data in these states,
5877 * RFC 1122 says we MUST send a reset.
5878 * BSD 4.4 also does reset.
5880 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5881 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5882 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5883 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5889 case TCP_ESTABLISHED:
5890 tcp_data_queue(sk, skb);
5895 /* tcp_data could move socket to TIME-WAIT */
5896 if (sk->sk_state != TCP_CLOSE) {
5897 tcp_data_snd_check(sk);
5898 tcp_ack_snd_check(sk);
5907 EXPORT_SYMBOL(tcp_rcv_state_process);