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
65 #include <linux/slab.h>
66 #include <linux/module.h>
67 #include <linux/sysctl.h>
68 #include <linux/kernel.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
74 #include <net/netdma.h>
76 int sysctl_tcp_timestamps __read_mostly = 1;
77 int sysctl_tcp_window_scaling __read_mostly = 1;
78 int sysctl_tcp_sack __read_mostly = 1;
79 int sysctl_tcp_fack __read_mostly = 1;
80 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
81 EXPORT_SYMBOL(sysctl_tcp_reordering);
82 int sysctl_tcp_ecn __read_mostly = 2;
83 EXPORT_SYMBOL(sysctl_tcp_ecn);
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 = 2;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 int sysctl_tcp_stdurg __read_mostly;
90 int sysctl_tcp_rfc1337 __read_mostly;
91 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
92 int sysctl_tcp_frto __read_mostly = 2;
93 int sysctl_tcp_frto_response __read_mostly;
94 int sysctl_tcp_nometrics_save __read_mostly;
96 int sysctl_tcp_thin_dupack __read_mostly;
98 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
99 int sysctl_tcp_abc __read_mostly;
101 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
102 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
103 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
104 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
105 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
106 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
107 #define FLAG_ECE 0x40 /* ECE in this ACK */
108 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
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_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
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)
120 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
130 struct inet_connection_sock *icsk = inet_csk(sk);
131 const unsigned int lss = icsk->icsk_ack.last_seg_size;
134 icsk->icsk_ack.last_seg_size = 0;
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
139 len = skb_shinfo(skb)->gso_size ? : skb->len;
140 if (len >= icsk->icsk_ack.rcv_mss) {
141 icsk->icsk_ack.rcv_mss = len;
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
146 * "len" is invariant segment length, including TCP header.
148 len += skb->data - skb_transport_header(skb);
149 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
155 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
156 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
161 len -= tcp_sk(sk)->tcp_header_len;
162 icsk->icsk_ack.last_seg_size = len;
164 icsk->icsk_ack.rcv_mss = len;
168 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
174 static void tcp_incr_quickack(struct sock *sk)
176 struct inet_connection_sock *icsk = inet_csk(sk);
177 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
181 if (quickacks > icsk->icsk_ack.quick)
182 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185 static void tcp_enter_quickack_mode(struct sock *sk)
187 struct inet_connection_sock *icsk = inet_csk(sk);
188 tcp_incr_quickack(sk);
189 icsk->icsk_ack.pingpong = 0;
190 icsk->icsk_ack.ato = TCP_ATO_MIN;
193 /* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
197 static inline int tcp_in_quickack_mode(const struct sock *sk)
199 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 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
238 tp->ecn_flags |= TCP_ECN_SEEN;
242 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
244 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
245 tp->ecn_flags &= ~TCP_ECN_OK;
248 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
250 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
251 tp->ecn_flags &= ~TCP_ECN_OK;
254 static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
256 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
261 /* Buffer size and advertised window tuning.
263 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
266 static void tcp_fixup_sndbuf(struct sock *sk)
268 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
270 sndmem *= TCP_INIT_CWND;
271 if (sk->sk_sndbuf < sndmem)
272 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
275 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
277 * All tcp_full_space() is split to two parts: "network" buffer, allocated
278 * forward and advertised in receiver window (tp->rcv_wnd) and
279 * "application buffer", required to isolate scheduling/application
280 * latencies from network.
281 * window_clamp is maximal advertised window. It can be less than
282 * tcp_full_space(), in this case tcp_full_space() - window_clamp
283 * is reserved for "application" buffer. The less window_clamp is
284 * the smoother our behaviour from viewpoint of network, but the lower
285 * throughput and the higher sensitivity of the connection to losses. 8)
287 * rcv_ssthresh is more strict window_clamp used at "slow start"
288 * phase to predict further behaviour of this connection.
289 * It is used for two goals:
290 * - to enforce header prediction at sender, even when application
291 * requires some significant "application buffer". It is check #1.
292 * - to prevent pruning of receive queue because of misprediction
293 * of receiver window. Check #2.
295 * The scheme does not work when sender sends good segments opening
296 * window and then starts to feed us spaghetti. But it should work
297 * in common situations. Otherwise, we have to rely on queue collapsing.
300 /* Slow part of check#2. */
301 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
303 struct tcp_sock *tp = tcp_sk(sk);
305 int truesize = tcp_win_from_space(skb->truesize) >> 1;
306 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
308 while (tp->rcv_ssthresh <= window) {
309 if (truesize <= skb->len)
310 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
318 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
320 struct tcp_sock *tp = tcp_sk(sk);
323 if (tp->rcv_ssthresh < tp->window_clamp &&
324 (int)tp->rcv_ssthresh < tcp_space(sk) &&
325 !tcp_memory_pressure) {
328 /* Check #2. Increase window, if skb with such overhead
329 * will fit to rcvbuf in future.
331 if (tcp_win_from_space(skb->truesize) <= skb->len)
332 incr = 2 * tp->advmss;
334 incr = __tcp_grow_window(sk, skb);
337 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
339 inet_csk(sk)->icsk_ack.quick |= 1;
344 /* 3. Tuning rcvbuf, when connection enters established state. */
346 static void tcp_fixup_rcvbuf(struct sock *sk)
348 u32 mss = tcp_sk(sk)->advmss;
349 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
352 /* Limit to 10 segments if mss <= 1460,
353 * or 14600/mss segments, with a minimum of two segments.
356 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
358 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
359 while (tcp_win_from_space(rcvmem) < mss)
364 if (sk->sk_rcvbuf < rcvmem)
365 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
368 /* 4. Try to fixup all. It is made immediately after connection enters
371 static void tcp_init_buffer_space(struct sock *sk)
373 struct tcp_sock *tp = tcp_sk(sk);
376 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
377 tcp_fixup_rcvbuf(sk);
378 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
379 tcp_fixup_sndbuf(sk);
381 tp->rcvq_space.space = tp->rcv_wnd;
383 maxwin = tcp_full_space(sk);
385 if (tp->window_clamp >= maxwin) {
386 tp->window_clamp = maxwin;
388 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
389 tp->window_clamp = max(maxwin -
390 (maxwin >> sysctl_tcp_app_win),
394 /* Force reservation of one segment. */
395 if (sysctl_tcp_app_win &&
396 tp->window_clamp > 2 * tp->advmss &&
397 tp->window_clamp + tp->advmss > maxwin)
398 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
400 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
401 tp->snd_cwnd_stamp = tcp_time_stamp;
404 /* 5. Recalculate window clamp after socket hit its memory bounds. */
405 static void tcp_clamp_window(struct sock *sk)
407 struct tcp_sock *tp = tcp_sk(sk);
408 struct inet_connection_sock *icsk = inet_csk(sk);
410 icsk->icsk_ack.quick = 0;
412 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
413 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
414 !tcp_memory_pressure &&
415 atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
416 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
419 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
420 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
423 /* Initialize RCV_MSS value.
424 * RCV_MSS is an our guess about MSS used by the peer.
425 * We haven't any direct information about the MSS.
426 * It's better to underestimate the RCV_MSS rather than overestimate.
427 * Overestimations make us ACKing less frequently than needed.
428 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
430 void tcp_initialize_rcv_mss(struct sock *sk)
432 const struct tcp_sock *tp = tcp_sk(sk);
433 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
435 hint = min(hint, tp->rcv_wnd / 2);
436 hint = min(hint, TCP_MSS_DEFAULT);
437 hint = max(hint, TCP_MIN_MSS);
439 inet_csk(sk)->icsk_ack.rcv_mss = hint;
441 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
443 /* Receiver "autotuning" code.
445 * The algorithm for RTT estimation w/o timestamps is based on
446 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
447 * <http://public.lanl.gov/radiant/pubs.html#DRS>
449 * More detail on this code can be found at
450 * <http://staff.psc.edu/jheffner/>,
451 * though this reference is out of date. A new paper
454 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
456 u32 new_sample = tp->rcv_rtt_est.rtt;
462 if (new_sample != 0) {
463 /* If we sample in larger samples in the non-timestamp
464 * case, we could grossly overestimate the RTT especially
465 * with chatty applications or bulk transfer apps which
466 * are stalled on filesystem I/O.
468 * Also, since we are only going for a minimum in the
469 * non-timestamp case, we do not smooth things out
470 * else with timestamps disabled convergence takes too
474 m -= (new_sample >> 3);
476 } else if (m < new_sample)
479 /* No previous measure. */
483 if (tp->rcv_rtt_est.rtt != new_sample)
484 tp->rcv_rtt_est.rtt = new_sample;
487 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
489 if (tp->rcv_rtt_est.time == 0)
491 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
493 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
496 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
497 tp->rcv_rtt_est.time = tcp_time_stamp;
500 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
501 const struct sk_buff *skb)
503 struct tcp_sock *tp = tcp_sk(sk);
504 if (tp->rx_opt.rcv_tsecr &&
505 (TCP_SKB_CB(skb)->end_seq -
506 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
507 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
511 * This function should be called every time data is copied to user space.
512 * It calculates the appropriate TCP receive buffer space.
514 void tcp_rcv_space_adjust(struct sock *sk)
516 struct tcp_sock *tp = tcp_sk(sk);
520 if (tp->rcvq_space.time == 0)
523 time = tcp_time_stamp - tp->rcvq_space.time;
524 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
527 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
529 space = max(tp->rcvq_space.space, space);
531 if (tp->rcvq_space.space != space) {
534 tp->rcvq_space.space = space;
536 if (sysctl_tcp_moderate_rcvbuf &&
537 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
538 int new_clamp = space;
540 /* Receive space grows, normalize in order to
541 * take into account packet headers and sk_buff
542 * structure overhead.
547 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
548 while (tcp_win_from_space(rcvmem) < tp->advmss)
551 space = min(space, sysctl_tcp_rmem[2]);
552 if (space > sk->sk_rcvbuf) {
553 sk->sk_rcvbuf = space;
555 /* Make the window clamp follow along. */
556 tp->window_clamp = new_clamp;
562 tp->rcvq_space.seq = tp->copied_seq;
563 tp->rcvq_space.time = tcp_time_stamp;
566 /* There is something which you must keep in mind when you analyze the
567 * behavior of the tp->ato delayed ack timeout interval. When a
568 * connection starts up, we want to ack as quickly as possible. The
569 * problem is that "good" TCP's do slow start at the beginning of data
570 * transmission. The means that until we send the first few ACK's the
571 * sender will sit on his end and only queue most of his data, because
572 * he can only send snd_cwnd unacked packets at any given time. For
573 * each ACK we send, he increments snd_cwnd and transmits more of his
576 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
578 struct tcp_sock *tp = tcp_sk(sk);
579 struct inet_connection_sock *icsk = inet_csk(sk);
582 inet_csk_schedule_ack(sk);
584 tcp_measure_rcv_mss(sk, skb);
586 tcp_rcv_rtt_measure(tp);
588 now = tcp_time_stamp;
590 if (!icsk->icsk_ack.ato) {
591 /* The _first_ data packet received, initialize
592 * delayed ACK engine.
594 tcp_incr_quickack(sk);
595 icsk->icsk_ack.ato = TCP_ATO_MIN;
597 int m = now - icsk->icsk_ack.lrcvtime;
599 if (m <= TCP_ATO_MIN / 2) {
600 /* The fastest case is the first. */
601 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
602 } else if (m < icsk->icsk_ack.ato) {
603 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
604 if (icsk->icsk_ack.ato > icsk->icsk_rto)
605 icsk->icsk_ack.ato = icsk->icsk_rto;
606 } else if (m > icsk->icsk_rto) {
607 /* Too long gap. Apparently sender failed to
608 * restart window, so that we send ACKs quickly.
610 tcp_incr_quickack(sk);
614 icsk->icsk_ack.lrcvtime = now;
616 TCP_ECN_check_ce(tp, skb);
619 tcp_grow_window(sk, skb);
622 /* Called to compute a smoothed rtt estimate. The data fed to this
623 * routine either comes from timestamps, or from segments that were
624 * known _not_ to have been retransmitted [see Karn/Partridge
625 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
626 * piece by Van Jacobson.
627 * NOTE: the next three routines used to be one big routine.
628 * To save cycles in the RFC 1323 implementation it was better to break
629 * it up into three procedures. -- erics
631 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
633 struct tcp_sock *tp = tcp_sk(sk);
634 long m = mrtt; /* RTT */
636 /* The following amusing code comes from Jacobson's
637 * article in SIGCOMM '88. Note that rtt and mdev
638 * are scaled versions of rtt and mean deviation.
639 * This is designed to be as fast as possible
640 * m stands for "measurement".
642 * On a 1990 paper the rto value is changed to:
643 * RTO = rtt + 4 * mdev
645 * Funny. This algorithm seems to be very broken.
646 * These formulae increase RTO, when it should be decreased, increase
647 * too slowly, when it should be increased quickly, decrease too quickly
648 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
649 * does not matter how to _calculate_ it. Seems, it was trap
650 * that VJ failed to avoid. 8)
655 m -= (tp->srtt >> 3); /* m is now error in rtt est */
656 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
658 m = -m; /* m is now abs(error) */
659 m -= (tp->mdev >> 2); /* similar update on mdev */
660 /* This is similar to one of Eifel findings.
661 * Eifel blocks mdev updates when rtt decreases.
662 * This solution is a bit different: we use finer gain
663 * for mdev in this case (alpha*beta).
664 * Like Eifel it also prevents growth of rto,
665 * but also it limits too fast rto decreases,
666 * happening in pure Eifel.
671 m -= (tp->mdev >> 2); /* similar update on mdev */
673 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
674 if (tp->mdev > tp->mdev_max) {
675 tp->mdev_max = tp->mdev;
676 if (tp->mdev_max > tp->rttvar)
677 tp->rttvar = tp->mdev_max;
679 if (after(tp->snd_una, tp->rtt_seq)) {
680 if (tp->mdev_max < tp->rttvar)
681 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
682 tp->rtt_seq = tp->snd_nxt;
683 tp->mdev_max = tcp_rto_min(sk);
686 /* no previous measure. */
687 tp->srtt = m << 3; /* take the measured time to be rtt */
688 tp->mdev = m << 1; /* make sure rto = 3*rtt */
689 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
690 tp->rtt_seq = tp->snd_nxt;
694 /* Calculate rto without backoff. This is the second half of Van Jacobson's
695 * routine referred to above.
697 static inline void tcp_set_rto(struct sock *sk)
699 const struct tcp_sock *tp = tcp_sk(sk);
700 /* Old crap is replaced with new one. 8)
703 * 1. If rtt variance happened to be less 50msec, it is hallucination.
704 * It cannot be less due to utterly erratic ACK generation made
705 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
706 * to do with delayed acks, because at cwnd>2 true delack timeout
707 * is invisible. Actually, Linux-2.4 also generates erratic
708 * ACKs in some circumstances.
710 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
712 /* 2. Fixups made earlier cannot be right.
713 * If we do not estimate RTO correctly without them,
714 * all the algo is pure shit and should be replaced
715 * with correct one. It is exactly, which we pretend to do.
718 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
719 * guarantees that rto is higher.
724 /* Save metrics learned by this TCP session.
725 This function is called only, when TCP finishes successfully
726 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
728 void tcp_update_metrics(struct sock *sk)
730 struct tcp_sock *tp = tcp_sk(sk);
731 struct dst_entry *dst = __sk_dst_get(sk);
733 if (sysctl_tcp_nometrics_save)
738 if (dst && (dst->flags & DST_HOST)) {
739 const struct inet_connection_sock *icsk = inet_csk(sk);
743 if (icsk->icsk_backoff || !tp->srtt) {
744 /* This session failed to estimate rtt. Why?
745 * Probably, no packets returned in time.
748 if (!(dst_metric_locked(dst, RTAX_RTT)))
749 dst_metric_set(dst, RTAX_RTT, 0);
753 rtt = dst_metric_rtt(dst, RTAX_RTT);
756 /* If newly calculated rtt larger than stored one,
757 * store new one. Otherwise, use EWMA. Remember,
758 * rtt overestimation is always better than underestimation.
760 if (!(dst_metric_locked(dst, RTAX_RTT))) {
762 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
764 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
767 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
772 /* Scale deviation to rttvar fixed point */
777 var = dst_metric_rtt(dst, RTAX_RTTVAR);
781 var -= (var - m) >> 2;
783 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
786 if (tcp_in_initial_slowstart(tp)) {
787 /* Slow start still did not finish. */
788 if (dst_metric(dst, RTAX_SSTHRESH) &&
789 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
790 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
791 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
792 if (!dst_metric_locked(dst, RTAX_CWND) &&
793 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
794 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
795 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
796 icsk->icsk_ca_state == TCP_CA_Open) {
797 /* Cong. avoidance phase, cwnd is reliable. */
798 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
799 dst_metric_set(dst, RTAX_SSTHRESH,
800 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
801 if (!dst_metric_locked(dst, RTAX_CWND))
802 dst_metric_set(dst, RTAX_CWND,
803 (dst_metric(dst, RTAX_CWND) +
806 /* Else slow start did not finish, cwnd is non-sense,
807 ssthresh may be also invalid.
809 if (!dst_metric_locked(dst, RTAX_CWND))
810 dst_metric_set(dst, RTAX_CWND,
811 (dst_metric(dst, RTAX_CWND) +
812 tp->snd_ssthresh) >> 1);
813 if (dst_metric(dst, RTAX_SSTHRESH) &&
814 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
815 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
816 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
819 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
820 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
821 tp->reordering != sysctl_tcp_reordering)
822 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
827 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
829 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
832 cwnd = TCP_INIT_CWND;
833 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
836 /* Set slow start threshold and cwnd not falling to slow start */
837 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
839 struct tcp_sock *tp = tcp_sk(sk);
840 const struct inet_connection_sock *icsk = inet_csk(sk);
842 tp->prior_ssthresh = 0;
844 if (icsk->icsk_ca_state < TCP_CA_CWR) {
847 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
848 tp->snd_cwnd = min(tp->snd_cwnd,
849 tcp_packets_in_flight(tp) + 1U);
850 tp->snd_cwnd_cnt = 0;
851 tp->high_seq = tp->snd_nxt;
852 tp->snd_cwnd_stamp = tcp_time_stamp;
853 TCP_ECN_queue_cwr(tp);
855 tcp_set_ca_state(sk, TCP_CA_CWR);
860 * Packet counting of FACK is based on in-order assumptions, therefore TCP
861 * disables it when reordering is detected
863 static void tcp_disable_fack(struct tcp_sock *tp)
865 /* RFC3517 uses different metric in lost marker => reset on change */
867 tp->lost_skb_hint = NULL;
868 tp->rx_opt.sack_ok &= ~2;
871 /* Take a notice that peer is sending D-SACKs */
872 static void tcp_dsack_seen(struct tcp_sock *tp)
874 tp->rx_opt.sack_ok |= 4;
877 /* Initialize metrics on socket. */
879 static void tcp_init_metrics(struct sock *sk)
881 struct tcp_sock *tp = tcp_sk(sk);
882 struct dst_entry *dst = __sk_dst_get(sk);
889 if (dst_metric_locked(dst, RTAX_CWND))
890 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
891 if (dst_metric(dst, RTAX_SSTHRESH)) {
892 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
893 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
894 tp->snd_ssthresh = tp->snd_cwnd_clamp;
896 /* ssthresh may have been reduced unnecessarily during.
897 * 3WHS. Restore it back to its initial default.
899 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
901 if (dst_metric(dst, RTAX_REORDERING) &&
902 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
903 tcp_disable_fack(tp);
904 tp->reordering = dst_metric(dst, RTAX_REORDERING);
907 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
910 /* Initial rtt is determined from SYN,SYN-ACK.
911 * The segment is small and rtt may appear much
912 * less than real one. Use per-dst memory
913 * to make it more realistic.
915 * A bit of theory. RTT is time passed after "normal" sized packet
916 * is sent until it is ACKed. In normal circumstances sending small
917 * packets force peer to delay ACKs and calculation is correct too.
918 * The algorithm is adaptive and, provided we follow specs, it
919 * NEVER underestimate RTT. BUT! If peer tries to make some clever
920 * tricks sort of "quick acks" for time long enough to decrease RTT
921 * to low value, and then abruptly stops to do it and starts to delay
922 * ACKs, wait for troubles.
924 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
925 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
926 tp->rtt_seq = tp->snd_nxt;
928 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
929 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
930 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
935 /* RFC2988bis: We've failed to get a valid RTT sample from
936 * 3WHS. This is most likely due to retransmission,
937 * including spurious one. Reset the RTO back to 3secs
938 * from the more aggressive 1sec to avoid more spurious
941 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
942 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
944 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
945 * retransmitted. In light of RFC2988bis' more aggressive 1sec
946 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
947 * retransmission has occurred.
949 if (tp->total_retrans > 1)
952 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
953 tp->snd_cwnd_stamp = tcp_time_stamp;
956 static void tcp_update_reordering(struct sock *sk, const int metric,
959 struct tcp_sock *tp = tcp_sk(sk);
960 if (metric > tp->reordering) {
963 tp->reordering = min(TCP_MAX_REORDERING, metric);
965 /* This exciting event is worth to be remembered. 8) */
967 mib_idx = LINUX_MIB_TCPTSREORDER;
968 else if (tcp_is_reno(tp))
969 mib_idx = LINUX_MIB_TCPRENOREORDER;
970 else if (tcp_is_fack(tp))
971 mib_idx = LINUX_MIB_TCPFACKREORDER;
973 mib_idx = LINUX_MIB_TCPSACKREORDER;
975 NET_INC_STATS_BH(sock_net(sk), mib_idx);
976 #if FASTRETRANS_DEBUG > 1
977 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
978 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
982 tp->undo_marker ? tp->undo_retrans : 0);
984 tcp_disable_fack(tp);
988 /* This must be called before lost_out is incremented */
989 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
991 if ((tp->retransmit_skb_hint == NULL) ||
992 before(TCP_SKB_CB(skb)->seq,
993 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
994 tp->retransmit_skb_hint = skb;
997 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
998 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1001 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1003 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1004 tcp_verify_retransmit_hint(tp, skb);
1006 tp->lost_out += tcp_skb_pcount(skb);
1007 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1011 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1012 struct sk_buff *skb)
1014 tcp_verify_retransmit_hint(tp, skb);
1016 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1017 tp->lost_out += tcp_skb_pcount(skb);
1018 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1022 /* This procedure tags the retransmission queue when SACKs arrive.
1024 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1025 * Packets in queue with these bits set are counted in variables
1026 * sacked_out, retrans_out and lost_out, correspondingly.
1028 * Valid combinations are:
1029 * Tag InFlight Description
1030 * 0 1 - orig segment is in flight.
1031 * S 0 - nothing flies, orig reached receiver.
1032 * L 0 - nothing flies, orig lost by net.
1033 * R 2 - both orig and retransmit are in flight.
1034 * L|R 1 - orig is lost, retransmit is in flight.
1035 * S|R 1 - orig reached receiver, retrans is still in flight.
1036 * (L|S|R is logically valid, it could occur when L|R is sacked,
1037 * but it is equivalent to plain S and code short-curcuits it to S.
1038 * L|S is logically invalid, it would mean -1 packet in flight 8))
1040 * These 6 states form finite state machine, controlled by the following events:
1041 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1042 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1043 * 3. Loss detection event of one of three flavors:
1044 * A. Scoreboard estimator decided the packet is lost.
1045 * A'. Reno "three dupacks" marks head of queue lost.
1046 * A''. Its FACK modfication, head until snd.fack is lost.
1047 * B. SACK arrives sacking data transmitted after never retransmitted
1048 * hole was sent out.
1049 * C. SACK arrives sacking SND.NXT at the moment, when the
1050 * segment was retransmitted.
1051 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1053 * It is pleasant to note, that state diagram turns out to be commutative,
1054 * so that we are allowed not to be bothered by order of our actions,
1055 * when multiple events arrive simultaneously. (see the function below).
1057 * Reordering detection.
1058 * --------------------
1059 * Reordering metric is maximal distance, which a packet can be displaced
1060 * in packet stream. With SACKs we can estimate it:
1062 * 1. SACK fills old hole and the corresponding segment was not
1063 * ever retransmitted -> reordering. Alas, we cannot use it
1064 * when segment was retransmitted.
1065 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1066 * for retransmitted and already SACKed segment -> reordering..
1067 * Both of these heuristics are not used in Loss state, when we cannot
1068 * account for retransmits accurately.
1070 * SACK block validation.
1071 * ----------------------
1073 * SACK block range validation checks that the received SACK block fits to
1074 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1075 * Note that SND.UNA is not included to the range though being valid because
1076 * it means that the receiver is rather inconsistent with itself reporting
1077 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1078 * perfectly valid, however, in light of RFC2018 which explicitly states
1079 * that "SACK block MUST reflect the newest segment. Even if the newest
1080 * segment is going to be discarded ...", not that it looks very clever
1081 * in case of head skb. Due to potentional receiver driven attacks, we
1082 * choose to avoid immediate execution of a walk in write queue due to
1083 * reneging and defer head skb's loss recovery to standard loss recovery
1084 * procedure that will eventually trigger (nothing forbids us doing this).
1086 * Implements also blockage to start_seq wrap-around. Problem lies in the
1087 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1088 * there's no guarantee that it will be before snd_nxt (n). The problem
1089 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1092 * <- outs wnd -> <- wrapzone ->
1093 * u e n u_w e_w s n_w
1095 * |<------------+------+----- TCP seqno space --------------+---------->|
1096 * ...-- <2^31 ->| |<--------...
1097 * ...---- >2^31 ------>| |<--------...
1099 * Current code wouldn't be vulnerable but it's better still to discard such
1100 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1101 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1102 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1103 * equal to the ideal case (infinite seqno space without wrap caused issues).
1105 * With D-SACK the lower bound is extended to cover sequence space below
1106 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1107 * again, D-SACK block must not to go across snd_una (for the same reason as
1108 * for the normal SACK blocks, explained above). But there all simplicity
1109 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1110 * fully below undo_marker they do not affect behavior in anyway and can
1111 * therefore be safely ignored. In rare cases (which are more or less
1112 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1113 * fragmentation and packet reordering past skb's retransmission. To consider
1114 * them correctly, the acceptable range must be extended even more though
1115 * the exact amount is rather hard to quantify. However, tp->max_window can
1116 * be used as an exaggerated estimate.
1118 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1119 u32 start_seq, u32 end_seq)
1121 /* Too far in future, or reversed (interpretation is ambiguous) */
1122 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1125 /* Nasty start_seq wrap-around check (see comments above) */
1126 if (!before(start_seq, tp->snd_nxt))
1129 /* In outstanding window? ...This is valid exit for D-SACKs too.
1130 * start_seq == snd_una is non-sensical (see comments above)
1132 if (after(start_seq, tp->snd_una))
1135 if (!is_dsack || !tp->undo_marker)
1138 /* ...Then it's D-SACK, and must reside below snd_una completely */
1139 if (after(end_seq, tp->snd_una))
1142 if (!before(start_seq, tp->undo_marker))
1146 if (!after(end_seq, tp->undo_marker))
1149 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1150 * start_seq < undo_marker and end_seq >= undo_marker.
1152 return !before(start_seq, end_seq - tp->max_window);
1155 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1156 * Event "C". Later note: FACK people cheated me again 8), we have to account
1157 * for reordering! Ugly, but should help.
1159 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1160 * less than what is now known to be received by the other end (derived from
1161 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1162 * retransmitted skbs to avoid some costly processing per ACKs.
1164 static void tcp_mark_lost_retrans(struct sock *sk)
1166 const struct inet_connection_sock *icsk = inet_csk(sk);
1167 struct tcp_sock *tp = tcp_sk(sk);
1168 struct sk_buff *skb;
1170 u32 new_low_seq = tp->snd_nxt;
1171 u32 received_upto = tcp_highest_sack_seq(tp);
1173 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1174 !after(received_upto, tp->lost_retrans_low) ||
1175 icsk->icsk_ca_state != TCP_CA_Recovery)
1178 tcp_for_write_queue(skb, sk) {
1179 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1181 if (skb == tcp_send_head(sk))
1183 if (cnt == tp->retrans_out)
1185 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1188 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1191 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1192 * constraint here (see above) but figuring out that at
1193 * least tp->reordering SACK blocks reside between ack_seq
1194 * and received_upto is not easy task to do cheaply with
1195 * the available datastructures.
1197 * Whether FACK should check here for tp->reordering segs
1198 * in-between one could argue for either way (it would be
1199 * rather simple to implement as we could count fack_count
1200 * during the walk and do tp->fackets_out - fack_count).
1202 if (after(received_upto, ack_seq)) {
1203 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1204 tp->retrans_out -= tcp_skb_pcount(skb);
1206 tcp_skb_mark_lost_uncond_verify(tp, skb);
1207 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1209 if (before(ack_seq, new_low_seq))
1210 new_low_seq = ack_seq;
1211 cnt += tcp_skb_pcount(skb);
1215 if (tp->retrans_out)
1216 tp->lost_retrans_low = new_low_seq;
1219 static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1220 struct tcp_sack_block_wire *sp, int num_sacks,
1223 struct tcp_sock *tp = tcp_sk(sk);
1224 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1225 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1228 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1231 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1232 } else if (num_sacks > 1) {
1233 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1234 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1236 if (!after(end_seq_0, end_seq_1) &&
1237 !before(start_seq_0, start_seq_1)) {
1240 NET_INC_STATS_BH(sock_net(sk),
1241 LINUX_MIB_TCPDSACKOFORECV);
1245 /* D-SACK for already forgotten data... Do dumb counting. */
1246 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1247 !after(end_seq_0, prior_snd_una) &&
1248 after(end_seq_0, tp->undo_marker))
1254 struct tcp_sacktag_state {
1260 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1261 * the incoming SACK may not exactly match but we can find smaller MSS
1262 * aligned portion of it that matches. Therefore we might need to fragment
1263 * which may fail and creates some hassle (caller must handle error case
1266 * FIXME: this could be merged to shift decision code
1268 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1269 u32 start_seq, u32 end_seq)
1272 unsigned int pkt_len;
1275 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1276 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1278 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1279 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1280 mss = tcp_skb_mss(skb);
1281 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1284 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1288 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1293 /* Round if necessary so that SACKs cover only full MSSes
1294 * and/or the remaining small portion (if present)
1296 if (pkt_len > mss) {
1297 unsigned int new_len = (pkt_len / mss) * mss;
1298 if (!in_sack && new_len < pkt_len) {
1300 if (new_len > skb->len)
1305 err = tcp_fragment(sk, skb, pkt_len, mss);
1313 static u8 tcp_sacktag_one(const struct sk_buff *skb, struct sock *sk,
1314 struct tcp_sacktag_state *state,
1315 int dup_sack, int pcount)
1317 struct tcp_sock *tp = tcp_sk(sk);
1318 u8 sacked = TCP_SKB_CB(skb)->sacked;
1319 int fack_count = state->fack_count;
1321 /* Account D-SACK for retransmitted packet. */
1322 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1323 if (tp->undo_marker && tp->undo_retrans &&
1324 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1326 if (sacked & TCPCB_SACKED_ACKED)
1327 state->reord = min(fack_count, state->reord);
1330 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1331 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1334 if (!(sacked & TCPCB_SACKED_ACKED)) {
1335 if (sacked & TCPCB_SACKED_RETRANS) {
1336 /* If the segment is not tagged as lost,
1337 * we do not clear RETRANS, believing
1338 * that retransmission is still in flight.
1340 if (sacked & TCPCB_LOST) {
1341 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1342 tp->lost_out -= pcount;
1343 tp->retrans_out -= pcount;
1346 if (!(sacked & TCPCB_RETRANS)) {
1347 /* New sack for not retransmitted frame,
1348 * which was in hole. It is reordering.
1350 if (before(TCP_SKB_CB(skb)->seq,
1351 tcp_highest_sack_seq(tp)))
1352 state->reord = min(fack_count,
1355 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1356 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1357 state->flag |= FLAG_ONLY_ORIG_SACKED;
1360 if (sacked & TCPCB_LOST) {
1361 sacked &= ~TCPCB_LOST;
1362 tp->lost_out -= pcount;
1366 sacked |= TCPCB_SACKED_ACKED;
1367 state->flag |= FLAG_DATA_SACKED;
1368 tp->sacked_out += pcount;
1370 fack_count += pcount;
1372 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1373 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1374 before(TCP_SKB_CB(skb)->seq,
1375 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1376 tp->lost_cnt_hint += pcount;
1378 if (fack_count > tp->fackets_out)
1379 tp->fackets_out = fack_count;
1382 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1383 * frames and clear it. undo_retrans is decreased above, L|R frames
1384 * are accounted above as well.
1386 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1387 sacked &= ~TCPCB_SACKED_RETRANS;
1388 tp->retrans_out -= pcount;
1394 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1395 struct tcp_sacktag_state *state,
1396 unsigned int pcount, int shifted, int mss,
1399 struct tcp_sock *tp = tcp_sk(sk);
1400 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1404 if (skb == tp->lost_skb_hint)
1405 tp->lost_cnt_hint += pcount;
1407 TCP_SKB_CB(prev)->end_seq += shifted;
1408 TCP_SKB_CB(skb)->seq += shifted;
1410 skb_shinfo(prev)->gso_segs += pcount;
1411 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1412 skb_shinfo(skb)->gso_segs -= pcount;
1414 /* When we're adding to gso_segs == 1, gso_size will be zero,
1415 * in theory this shouldn't be necessary but as long as DSACK
1416 * code can come after this skb later on it's better to keep
1417 * setting gso_size to something.
1419 if (!skb_shinfo(prev)->gso_size) {
1420 skb_shinfo(prev)->gso_size = mss;
1421 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1424 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1425 if (skb_shinfo(skb)->gso_segs <= 1) {
1426 skb_shinfo(skb)->gso_size = 0;
1427 skb_shinfo(skb)->gso_type = 0;
1430 /* We discard results */
1431 tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1433 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1434 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1437 BUG_ON(!tcp_skb_pcount(skb));
1438 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1442 /* Whole SKB was eaten :-) */
1444 if (skb == tp->retransmit_skb_hint)
1445 tp->retransmit_skb_hint = prev;
1446 if (skb == tp->scoreboard_skb_hint)
1447 tp->scoreboard_skb_hint = prev;
1448 if (skb == tp->lost_skb_hint) {
1449 tp->lost_skb_hint = prev;
1450 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1453 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1454 if (skb == tcp_highest_sack(sk))
1455 tcp_advance_highest_sack(sk, skb);
1457 tcp_unlink_write_queue(skb, sk);
1458 sk_wmem_free_skb(sk, skb);
1460 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1465 /* I wish gso_size would have a bit more sane initialization than
1466 * something-or-zero which complicates things
1468 static int tcp_skb_seglen(const struct sk_buff *skb)
1470 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1473 /* Shifting pages past head area doesn't work */
1474 static int skb_can_shift(const struct sk_buff *skb)
1476 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1479 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1482 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1483 struct tcp_sacktag_state *state,
1484 u32 start_seq, u32 end_seq,
1487 struct tcp_sock *tp = tcp_sk(sk);
1488 struct sk_buff *prev;
1494 if (!sk_can_gso(sk))
1497 /* Normally R but no L won't result in plain S */
1499 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1501 if (!skb_can_shift(skb))
1503 /* This frame is about to be dropped (was ACKed). */
1504 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1507 /* Can only happen with delayed DSACK + discard craziness */
1508 if (unlikely(skb == tcp_write_queue_head(sk)))
1510 prev = tcp_write_queue_prev(sk, skb);
1512 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1515 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1516 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1520 pcount = tcp_skb_pcount(skb);
1521 mss = tcp_skb_seglen(skb);
1523 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1524 * drop this restriction as unnecessary
1526 if (mss != tcp_skb_seglen(prev))
1529 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1531 /* CHECKME: This is non-MSS split case only?, this will
1532 * cause skipped skbs due to advancing loop btw, original
1533 * has that feature too
1535 if (tcp_skb_pcount(skb) <= 1)
1538 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1540 /* TODO: head merge to next could be attempted here
1541 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1542 * though it might not be worth of the additional hassle
1544 * ...we can probably just fallback to what was done
1545 * previously. We could try merging non-SACKed ones
1546 * as well but it probably isn't going to buy off
1547 * because later SACKs might again split them, and
1548 * it would make skb timestamp tracking considerably
1554 len = end_seq - TCP_SKB_CB(skb)->seq;
1556 BUG_ON(len > skb->len);
1558 /* MSS boundaries should be honoured or else pcount will
1559 * severely break even though it makes things bit trickier.
1560 * Optimize common case to avoid most of the divides
1562 mss = tcp_skb_mss(skb);
1564 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1565 * drop this restriction as unnecessary
1567 if (mss != tcp_skb_seglen(prev))
1572 } else if (len < mss) {
1580 if (!skb_shift(prev, skb, len))
1582 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1585 /* Hole filled allows collapsing with the next as well, this is very
1586 * useful when hole on every nth skb pattern happens
1588 if (prev == tcp_write_queue_tail(sk))
1590 skb = tcp_write_queue_next(sk, prev);
1592 if (!skb_can_shift(skb) ||
1593 (skb == tcp_send_head(sk)) ||
1594 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1595 (mss != tcp_skb_seglen(skb)))
1599 if (skb_shift(prev, skb, len)) {
1600 pcount += tcp_skb_pcount(skb);
1601 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1605 state->fack_count += pcount;
1612 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1616 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1617 struct tcp_sack_block *next_dup,
1618 struct tcp_sacktag_state *state,
1619 u32 start_seq, u32 end_seq,
1622 struct tcp_sock *tp = tcp_sk(sk);
1623 struct sk_buff *tmp;
1625 tcp_for_write_queue_from(skb, sk) {
1627 int dup_sack = dup_sack_in;
1629 if (skb == tcp_send_head(sk))
1632 /* queue is in-order => we can short-circuit the walk early */
1633 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1636 if ((next_dup != NULL) &&
1637 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1638 in_sack = tcp_match_skb_to_sack(sk, skb,
1639 next_dup->start_seq,
1645 /* skb reference here is a bit tricky to get right, since
1646 * shifting can eat and free both this skb and the next,
1647 * so not even _safe variant of the loop is enough.
1650 tmp = tcp_shift_skb_data(sk, skb, state,
1651 start_seq, end_seq, dup_sack);
1660 in_sack = tcp_match_skb_to_sack(sk, skb,
1666 if (unlikely(in_sack < 0))
1670 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1673 tcp_skb_pcount(skb));
1675 if (!before(TCP_SKB_CB(skb)->seq,
1676 tcp_highest_sack_seq(tp)))
1677 tcp_advance_highest_sack(sk, skb);
1680 state->fack_count += tcp_skb_pcount(skb);
1685 /* Avoid all extra work that is being done by sacktag while walking in
1688 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1689 struct tcp_sacktag_state *state,
1692 tcp_for_write_queue_from(skb, sk) {
1693 if (skb == tcp_send_head(sk))
1696 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1699 state->fack_count += tcp_skb_pcount(skb);
1704 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1706 struct tcp_sack_block *next_dup,
1707 struct tcp_sacktag_state *state,
1710 if (next_dup == NULL)
1713 if (before(next_dup->start_seq, skip_to_seq)) {
1714 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1715 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1716 next_dup->start_seq, next_dup->end_seq,
1723 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1725 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1729 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1732 const struct inet_connection_sock *icsk = inet_csk(sk);
1733 struct tcp_sock *tp = tcp_sk(sk);
1734 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1735 TCP_SKB_CB(ack_skb)->sacked);
1736 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1737 struct tcp_sack_block sp[TCP_NUM_SACKS];
1738 struct tcp_sack_block *cache;
1739 struct tcp_sacktag_state state;
1740 struct sk_buff *skb;
1741 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1743 int found_dup_sack = 0;
1745 int first_sack_index;
1748 state.reord = tp->packets_out;
1750 if (!tp->sacked_out) {
1751 if (WARN_ON(tp->fackets_out))
1752 tp->fackets_out = 0;
1753 tcp_highest_sack_reset(sk);
1756 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1757 num_sacks, prior_snd_una);
1759 state.flag |= FLAG_DSACKING_ACK;
1761 /* Eliminate too old ACKs, but take into
1762 * account more or less fresh ones, they can
1763 * contain valid SACK info.
1765 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1768 if (!tp->packets_out)
1772 first_sack_index = 0;
1773 for (i = 0; i < num_sacks; i++) {
1774 int dup_sack = !i && found_dup_sack;
1776 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1777 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1779 if (!tcp_is_sackblock_valid(tp, dup_sack,
1780 sp[used_sacks].start_seq,
1781 sp[used_sacks].end_seq)) {
1785 if (!tp->undo_marker)
1786 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1788 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1790 /* Don't count olds caused by ACK reordering */
1791 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1792 !after(sp[used_sacks].end_seq, tp->snd_una))
1794 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1797 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1799 first_sack_index = -1;
1803 /* Ignore very old stuff early */
1804 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1810 /* order SACK blocks to allow in order walk of the retrans queue */
1811 for (i = used_sacks - 1; i > 0; i--) {
1812 for (j = 0; j < i; j++) {
1813 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1814 swap(sp[j], sp[j + 1]);
1816 /* Track where the first SACK block goes to */
1817 if (j == first_sack_index)
1818 first_sack_index = j + 1;
1823 skb = tcp_write_queue_head(sk);
1824 state.fack_count = 0;
1827 if (!tp->sacked_out) {
1828 /* It's already past, so skip checking against it */
1829 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1831 cache = tp->recv_sack_cache;
1832 /* Skip empty blocks in at head of the cache */
1833 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1838 while (i < used_sacks) {
1839 u32 start_seq = sp[i].start_seq;
1840 u32 end_seq = sp[i].end_seq;
1841 int dup_sack = (found_dup_sack && (i == first_sack_index));
1842 struct tcp_sack_block *next_dup = NULL;
1844 if (found_dup_sack && ((i + 1) == first_sack_index))
1845 next_dup = &sp[i + 1];
1847 /* Event "B" in the comment above. */
1848 if (after(end_seq, tp->high_seq))
1849 state.flag |= FLAG_DATA_LOST;
1851 /* Skip too early cached blocks */
1852 while (tcp_sack_cache_ok(tp, cache) &&
1853 !before(start_seq, cache->end_seq))
1856 /* Can skip some work by looking recv_sack_cache? */
1857 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1858 after(end_seq, cache->start_seq)) {
1861 if (before(start_seq, cache->start_seq)) {
1862 skb = tcp_sacktag_skip(skb, sk, &state,
1864 skb = tcp_sacktag_walk(skb, sk, next_dup,
1871 /* Rest of the block already fully processed? */
1872 if (!after(end_seq, cache->end_seq))
1875 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1879 /* ...tail remains todo... */
1880 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1881 /* ...but better entrypoint exists! */
1882 skb = tcp_highest_sack(sk);
1885 state.fack_count = tp->fackets_out;
1890 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1891 /* Check overlap against next cached too (past this one already) */
1896 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1897 skb = tcp_highest_sack(sk);
1900 state.fack_count = tp->fackets_out;
1902 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1905 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1906 start_seq, end_seq, dup_sack);
1909 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1910 * due to in-order walk
1912 if (after(end_seq, tp->frto_highmark))
1913 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1918 /* Clear the head of the cache sack blocks so we can skip it next time */
1919 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1920 tp->recv_sack_cache[i].start_seq = 0;
1921 tp->recv_sack_cache[i].end_seq = 0;
1923 for (j = 0; j < used_sacks; j++)
1924 tp->recv_sack_cache[i++] = sp[j];
1926 tcp_mark_lost_retrans(sk);
1928 tcp_verify_left_out(tp);
1930 if ((state.reord < tp->fackets_out) &&
1931 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1932 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1933 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1937 #if FASTRETRANS_DEBUG > 0
1938 WARN_ON((int)tp->sacked_out < 0);
1939 WARN_ON((int)tp->lost_out < 0);
1940 WARN_ON((int)tp->retrans_out < 0);
1941 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1946 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1947 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1949 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1953 holes = max(tp->lost_out, 1U);
1954 holes = min(holes, tp->packets_out);
1956 if ((tp->sacked_out + holes) > tp->packets_out) {
1957 tp->sacked_out = tp->packets_out - holes;
1963 /* If we receive more dupacks than we expected counting segments
1964 * in assumption of absent reordering, interpret this as reordering.
1965 * The only another reason could be bug in receiver TCP.
1967 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1969 struct tcp_sock *tp = tcp_sk(sk);
1970 if (tcp_limit_reno_sacked(tp))
1971 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1974 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1976 static void tcp_add_reno_sack(struct sock *sk)
1978 struct tcp_sock *tp = tcp_sk(sk);
1980 tcp_check_reno_reordering(sk, 0);
1981 tcp_verify_left_out(tp);
1984 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1986 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1988 struct tcp_sock *tp = tcp_sk(sk);
1991 /* One ACK acked hole. The rest eat duplicate ACKs. */
1992 if (acked - 1 >= tp->sacked_out)
1995 tp->sacked_out -= acked - 1;
1997 tcp_check_reno_reordering(sk, acked);
1998 tcp_verify_left_out(tp);
2001 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2006 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2008 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2011 /* F-RTO can only be used if TCP has never retransmitted anything other than
2012 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2014 int tcp_use_frto(struct sock *sk)
2016 const struct tcp_sock *tp = tcp_sk(sk);
2017 const struct inet_connection_sock *icsk = inet_csk(sk);
2018 struct sk_buff *skb;
2020 if (!sysctl_tcp_frto)
2023 /* MTU probe and F-RTO won't really play nicely along currently */
2024 if (icsk->icsk_mtup.probe_size)
2027 if (tcp_is_sackfrto(tp))
2030 /* Avoid expensive walking of rexmit queue if possible */
2031 if (tp->retrans_out > 1)
2034 skb = tcp_write_queue_head(sk);
2035 if (tcp_skb_is_last(sk, skb))
2037 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2038 tcp_for_write_queue_from(skb, sk) {
2039 if (skb == tcp_send_head(sk))
2041 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2043 /* Short-circuit when first non-SACKed skb has been checked */
2044 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2050 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2051 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2052 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2053 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2054 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2055 * bits are handled if the Loss state is really to be entered (in
2056 * tcp_enter_frto_loss).
2058 * Do like tcp_enter_loss() would; when RTO expires the second time it
2060 * "Reduce ssthresh if it has not yet been made inside this window."
2062 void tcp_enter_frto(struct sock *sk)
2064 const struct inet_connection_sock *icsk = inet_csk(sk);
2065 struct tcp_sock *tp = tcp_sk(sk);
2066 struct sk_buff *skb;
2068 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2069 tp->snd_una == tp->high_seq ||
2070 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2071 !icsk->icsk_retransmits)) {
2072 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2073 /* Our state is too optimistic in ssthresh() call because cwnd
2074 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2075 * recovery has not yet completed. Pattern would be this: RTO,
2076 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2078 * RFC4138 should be more specific on what to do, even though
2079 * RTO is quite unlikely to occur after the first Cumulative ACK
2080 * due to back-off and complexity of triggering events ...
2082 if (tp->frto_counter) {
2084 stored_cwnd = tp->snd_cwnd;
2086 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2087 tp->snd_cwnd = stored_cwnd;
2089 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2091 /* ... in theory, cong.control module could do "any tricks" in
2092 * ssthresh(), which means that ca_state, lost bits and lost_out
2093 * counter would have to be faked before the call occurs. We
2094 * consider that too expensive, unlikely and hacky, so modules
2095 * using these in ssthresh() must deal these incompatibility
2096 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2098 tcp_ca_event(sk, CA_EVENT_FRTO);
2101 tp->undo_marker = tp->snd_una;
2102 tp->undo_retrans = 0;
2104 skb = tcp_write_queue_head(sk);
2105 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2106 tp->undo_marker = 0;
2107 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2108 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2109 tp->retrans_out -= tcp_skb_pcount(skb);
2111 tcp_verify_left_out(tp);
2113 /* Too bad if TCP was application limited */
2114 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2116 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2117 * The last condition is necessary at least in tp->frto_counter case.
2119 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2120 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2121 after(tp->high_seq, tp->snd_una)) {
2122 tp->frto_highmark = tp->high_seq;
2124 tp->frto_highmark = tp->snd_nxt;
2126 tcp_set_ca_state(sk, TCP_CA_Disorder);
2127 tp->high_seq = tp->snd_nxt;
2128 tp->frto_counter = 1;
2131 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2132 * which indicates that we should follow the traditional RTO recovery,
2133 * i.e. mark everything lost and do go-back-N retransmission.
2135 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2137 struct tcp_sock *tp = tcp_sk(sk);
2138 struct sk_buff *skb;
2141 tp->retrans_out = 0;
2142 if (tcp_is_reno(tp))
2143 tcp_reset_reno_sack(tp);
2145 tcp_for_write_queue(skb, sk) {
2146 if (skb == tcp_send_head(sk))
2149 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2151 * Count the retransmission made on RTO correctly (only when
2152 * waiting for the first ACK and did not get it)...
2154 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2155 /* For some reason this R-bit might get cleared? */
2156 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2157 tp->retrans_out += tcp_skb_pcount(skb);
2158 /* ...enter this if branch just for the first segment */
2159 flag |= FLAG_DATA_ACKED;
2161 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2162 tp->undo_marker = 0;
2163 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2166 /* Marking forward transmissions that were made after RTO lost
2167 * can cause unnecessary retransmissions in some scenarios,
2168 * SACK blocks will mitigate that in some but not in all cases.
2169 * We used to not mark them but it was causing break-ups with
2170 * receivers that do only in-order receival.
2172 * TODO: we could detect presence of such receiver and select
2173 * different behavior per flow.
2175 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2176 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2177 tp->lost_out += tcp_skb_pcount(skb);
2178 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2181 tcp_verify_left_out(tp);
2183 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2184 tp->snd_cwnd_cnt = 0;
2185 tp->snd_cwnd_stamp = tcp_time_stamp;
2186 tp->frto_counter = 0;
2187 tp->bytes_acked = 0;
2189 tp->reordering = min_t(unsigned int, tp->reordering,
2190 sysctl_tcp_reordering);
2191 tcp_set_ca_state(sk, TCP_CA_Loss);
2192 tp->high_seq = tp->snd_nxt;
2193 TCP_ECN_queue_cwr(tp);
2195 tcp_clear_all_retrans_hints(tp);
2198 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2200 tp->retrans_out = 0;
2203 tp->undo_marker = 0;
2204 tp->undo_retrans = 0;
2207 void tcp_clear_retrans(struct tcp_sock *tp)
2209 tcp_clear_retrans_partial(tp);
2211 tp->fackets_out = 0;
2215 /* Enter Loss state. If "how" is not zero, forget all SACK information
2216 * and reset tags completely, otherwise preserve SACKs. If receiver
2217 * dropped its ofo queue, we will know this due to reneging detection.
2219 void tcp_enter_loss(struct sock *sk, int how)
2221 const struct inet_connection_sock *icsk = inet_csk(sk);
2222 struct tcp_sock *tp = tcp_sk(sk);
2223 struct sk_buff *skb;
2225 /* Reduce ssthresh if it has not yet been made inside this window. */
2226 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2227 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2228 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2229 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2230 tcp_ca_event(sk, CA_EVENT_LOSS);
2233 tp->snd_cwnd_cnt = 0;
2234 tp->snd_cwnd_stamp = tcp_time_stamp;
2236 tp->bytes_acked = 0;
2237 tcp_clear_retrans_partial(tp);
2239 if (tcp_is_reno(tp))
2240 tcp_reset_reno_sack(tp);
2243 /* Push undo marker, if it was plain RTO and nothing
2244 * was retransmitted. */
2245 tp->undo_marker = tp->snd_una;
2248 tp->fackets_out = 0;
2250 tcp_clear_all_retrans_hints(tp);
2252 tcp_for_write_queue(skb, sk) {
2253 if (skb == tcp_send_head(sk))
2256 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2257 tp->undo_marker = 0;
2258 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2259 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2260 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2261 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2262 tp->lost_out += tcp_skb_pcount(skb);
2263 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2266 tcp_verify_left_out(tp);
2268 tp->reordering = min_t(unsigned int, tp->reordering,
2269 sysctl_tcp_reordering);
2270 tcp_set_ca_state(sk, TCP_CA_Loss);
2271 tp->high_seq = tp->snd_nxt;
2272 TCP_ECN_queue_cwr(tp);
2273 /* Abort F-RTO algorithm if one is in progress */
2274 tp->frto_counter = 0;
2277 /* If ACK arrived pointing to a remembered SACK, it means that our
2278 * remembered SACKs do not reflect real state of receiver i.e.
2279 * receiver _host_ is heavily congested (or buggy).
2281 * Do processing similar to RTO timeout.
2283 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2285 if (flag & FLAG_SACK_RENEGING) {
2286 struct inet_connection_sock *icsk = inet_csk(sk);
2287 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2289 tcp_enter_loss(sk, 1);
2290 icsk->icsk_retransmits++;
2291 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2292 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2293 icsk->icsk_rto, TCP_RTO_MAX);
2299 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2301 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2304 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2305 * counter when SACK is enabled (without SACK, sacked_out is used for
2308 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2309 * segments up to the highest received SACK block so far and holes in
2312 * With reordering, holes may still be in flight, so RFC3517 recovery
2313 * uses pure sacked_out (total number of SACKed segments) even though
2314 * it violates the RFC that uses duplicate ACKs, often these are equal
2315 * but when e.g. out-of-window ACKs or packet duplication occurs,
2316 * they differ. Since neither occurs due to loss, TCP should really
2319 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2321 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2324 static inline int tcp_skb_timedout(const struct sock *sk,
2325 const struct sk_buff *skb)
2327 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2330 static inline int tcp_head_timedout(const struct sock *sk)
2332 const struct tcp_sock *tp = tcp_sk(sk);
2334 return tp->packets_out &&
2335 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2338 /* Linux NewReno/SACK/FACK/ECN state machine.
2339 * --------------------------------------
2341 * "Open" Normal state, no dubious events, fast path.
2342 * "Disorder" In all the respects it is "Open",
2343 * but requires a bit more attention. It is entered when
2344 * we see some SACKs or dupacks. It is split of "Open"
2345 * mainly to move some processing from fast path to slow one.
2346 * "CWR" CWND was reduced due to some Congestion Notification event.
2347 * It can be ECN, ICMP source quench, local device congestion.
2348 * "Recovery" CWND was reduced, we are fast-retransmitting.
2349 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2351 * tcp_fastretrans_alert() is entered:
2352 * - each incoming ACK, if state is not "Open"
2353 * - when arrived ACK is unusual, namely:
2358 * Counting packets in flight is pretty simple.
2360 * in_flight = packets_out - left_out + retrans_out
2362 * packets_out is SND.NXT-SND.UNA counted in packets.
2364 * retrans_out is number of retransmitted segments.
2366 * left_out is number of segments left network, but not ACKed yet.
2368 * left_out = sacked_out + lost_out
2370 * sacked_out: Packets, which arrived to receiver out of order
2371 * and hence not ACKed. With SACKs this number is simply
2372 * amount of SACKed data. Even without SACKs
2373 * it is easy to give pretty reliable estimate of this number,
2374 * counting duplicate ACKs.
2376 * lost_out: Packets lost by network. TCP has no explicit
2377 * "loss notification" feedback from network (for now).
2378 * It means that this number can be only _guessed_.
2379 * Actually, it is the heuristics to predict lossage that
2380 * distinguishes different algorithms.
2382 * F.e. after RTO, when all the queue is considered as lost,
2383 * lost_out = packets_out and in_flight = retrans_out.
2385 * Essentially, we have now two algorithms counting
2388 * FACK: It is the simplest heuristics. As soon as we decided
2389 * that something is lost, we decide that _all_ not SACKed
2390 * packets until the most forward SACK are lost. I.e.
2391 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2392 * It is absolutely correct estimate, if network does not reorder
2393 * packets. And it loses any connection to reality when reordering
2394 * takes place. We use FACK by default until reordering
2395 * is suspected on the path to this destination.
2397 * NewReno: when Recovery is entered, we assume that one segment
2398 * is lost (classic Reno). While we are in Recovery and
2399 * a partial ACK arrives, we assume that one more packet
2400 * is lost (NewReno). This heuristics are the same in NewReno
2403 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2404 * deflation etc. CWND is real congestion window, never inflated, changes
2405 * only according to classic VJ rules.
2407 * Really tricky (and requiring careful tuning) part of algorithm
2408 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2409 * The first determines the moment _when_ we should reduce CWND and,
2410 * hence, slow down forward transmission. In fact, it determines the moment
2411 * when we decide that hole is caused by loss, rather than by a reorder.
2413 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2414 * holes, caused by lost packets.
2416 * And the most logically complicated part of algorithm is undo
2417 * heuristics. We detect false retransmits due to both too early
2418 * fast retransmit (reordering) and underestimated RTO, analyzing
2419 * timestamps and D-SACKs. When we detect that some segments were
2420 * retransmitted by mistake and CWND reduction was wrong, we undo
2421 * window reduction and abort recovery phase. This logic is hidden
2422 * inside several functions named tcp_try_undo_<something>.
2425 /* This function decides, when we should leave Disordered state
2426 * and enter Recovery phase, reducing congestion window.
2428 * Main question: may we further continue forward transmission
2429 * with the same cwnd?
2431 static int tcp_time_to_recover(struct sock *sk)
2433 struct tcp_sock *tp = tcp_sk(sk);
2436 /* Do not perform any recovery during F-RTO algorithm */
2437 if (tp->frto_counter)
2440 /* Trick#1: The loss is proven. */
2444 /* Not-A-Trick#2 : Classic rule... */
2445 if (tcp_dupack_heuristics(tp) > tp->reordering)
2448 /* Trick#3 : when we use RFC2988 timer restart, fast
2449 * retransmit can be triggered by timeout of queue head.
2451 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2454 /* Trick#4: It is still not OK... But will it be useful to delay
2457 packets_out = tp->packets_out;
2458 if (packets_out <= tp->reordering &&
2459 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2460 !tcp_may_send_now(sk)) {
2461 /* We have nothing to send. This connection is limited
2462 * either by receiver window or by application.
2467 /* If a thin stream is detected, retransmit after first
2468 * received dupack. Employ only if SACK is supported in order
2469 * to avoid possible corner-case series of spurious retransmissions
2470 * Use only if there are no unsent data.
2472 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2473 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2474 tcp_is_sack(tp) && !tcp_send_head(sk))
2480 /* New heuristics: it is possible only after we switched to restart timer
2481 * each time when something is ACKed. Hence, we can detect timed out packets
2482 * during fast retransmit without falling to slow start.
2484 * Usefulness of this as is very questionable, since we should know which of
2485 * the segments is the next to timeout which is relatively expensive to find
2486 * in general case unless we add some data structure just for that. The
2487 * current approach certainly won't find the right one too often and when it
2488 * finally does find _something_ it usually marks large part of the window
2489 * right away (because a retransmission with a larger timestamp blocks the
2490 * loop from advancing). -ij
2492 static void tcp_timeout_skbs(struct sock *sk)
2494 struct tcp_sock *tp = tcp_sk(sk);
2495 struct sk_buff *skb;
2497 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2500 skb = tp->scoreboard_skb_hint;
2501 if (tp->scoreboard_skb_hint == NULL)
2502 skb = tcp_write_queue_head(sk);
2504 tcp_for_write_queue_from(skb, sk) {
2505 if (skb == tcp_send_head(sk))
2507 if (!tcp_skb_timedout(sk, skb))
2510 tcp_skb_mark_lost(tp, skb);
2513 tp->scoreboard_skb_hint = skb;
2515 tcp_verify_left_out(tp);
2518 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2519 * is against sacked "cnt", otherwise it's against facked "cnt"
2521 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2523 struct tcp_sock *tp = tcp_sk(sk);
2524 struct sk_buff *skb;
2529 WARN_ON(packets > tp->packets_out);
2530 if (tp->lost_skb_hint) {
2531 skb = tp->lost_skb_hint;
2532 cnt = tp->lost_cnt_hint;
2533 /* Head already handled? */
2534 if (mark_head && skb != tcp_write_queue_head(sk))
2537 skb = tcp_write_queue_head(sk);
2541 tcp_for_write_queue_from(skb, sk) {
2542 if (skb == tcp_send_head(sk))
2544 /* TODO: do this better */
2545 /* this is not the most efficient way to do this... */
2546 tp->lost_skb_hint = skb;
2547 tp->lost_cnt_hint = cnt;
2549 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2553 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2554 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2555 cnt += tcp_skb_pcount(skb);
2557 if (cnt > packets) {
2558 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2559 (oldcnt >= packets))
2562 mss = skb_shinfo(skb)->gso_size;
2563 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2569 tcp_skb_mark_lost(tp, skb);
2574 tcp_verify_left_out(tp);
2577 /* Account newly detected lost packet(s) */
2579 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2581 struct tcp_sock *tp = tcp_sk(sk);
2583 if (tcp_is_reno(tp)) {
2584 tcp_mark_head_lost(sk, 1, 1);
2585 } else if (tcp_is_fack(tp)) {
2586 int lost = tp->fackets_out - tp->reordering;
2589 tcp_mark_head_lost(sk, lost, 0);
2591 int sacked_upto = tp->sacked_out - tp->reordering;
2592 if (sacked_upto >= 0)
2593 tcp_mark_head_lost(sk, sacked_upto, 0);
2594 else if (fast_rexmit)
2595 tcp_mark_head_lost(sk, 1, 1);
2598 tcp_timeout_skbs(sk);
2601 /* CWND moderation, preventing bursts due to too big ACKs
2602 * in dubious situations.
2604 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2606 tp->snd_cwnd = min(tp->snd_cwnd,
2607 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2608 tp->snd_cwnd_stamp = tcp_time_stamp;
2611 /* Lower bound on congestion window is slow start threshold
2612 * unless congestion avoidance choice decides to overide it.
2614 static inline u32 tcp_cwnd_min(const struct sock *sk)
2616 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2618 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2621 /* Decrease cwnd each second ack. */
2622 static void tcp_cwnd_down(struct sock *sk, int flag)
2624 struct tcp_sock *tp = tcp_sk(sk);
2625 int decr = tp->snd_cwnd_cnt + 1;
2627 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2628 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2629 tp->snd_cwnd_cnt = decr & 1;
2632 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2633 tp->snd_cwnd -= decr;
2635 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2636 tp->snd_cwnd_stamp = tcp_time_stamp;
2640 /* Nothing was retransmitted or returned timestamp is less
2641 * than timestamp of the first retransmission.
2643 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2645 return !tp->retrans_stamp ||
2646 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2647 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2650 /* Undo procedures. */
2652 #if FASTRETRANS_DEBUG > 1
2653 static void DBGUNDO(struct sock *sk, const char *msg)
2655 struct tcp_sock *tp = tcp_sk(sk);
2656 struct inet_sock *inet = inet_sk(sk);
2658 if (sk->sk_family == AF_INET) {
2659 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2661 &inet->inet_daddr, ntohs(inet->inet_dport),
2662 tp->snd_cwnd, tcp_left_out(tp),
2663 tp->snd_ssthresh, tp->prior_ssthresh,
2666 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2667 else if (sk->sk_family == AF_INET6) {
2668 struct ipv6_pinfo *np = inet6_sk(sk);
2669 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2671 &np->daddr, ntohs(inet->inet_dport),
2672 tp->snd_cwnd, tcp_left_out(tp),
2673 tp->snd_ssthresh, tp->prior_ssthresh,
2679 #define DBGUNDO(x...) do { } while (0)
2682 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2684 struct tcp_sock *tp = tcp_sk(sk);
2686 if (tp->prior_ssthresh) {
2687 const struct inet_connection_sock *icsk = inet_csk(sk);
2689 if (icsk->icsk_ca_ops->undo_cwnd)
2690 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2692 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2694 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2695 tp->snd_ssthresh = tp->prior_ssthresh;
2696 TCP_ECN_withdraw_cwr(tp);
2699 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2701 tp->snd_cwnd_stamp = tcp_time_stamp;
2704 static inline int tcp_may_undo(const struct tcp_sock *tp)
2706 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2709 /* People celebrate: "We love our President!" */
2710 static int tcp_try_undo_recovery(struct sock *sk)
2712 struct tcp_sock *tp = tcp_sk(sk);
2714 if (tcp_may_undo(tp)) {
2717 /* Happy end! We did not retransmit anything
2718 * or our original transmission succeeded.
2720 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2721 tcp_undo_cwr(sk, true);
2722 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2723 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2725 mib_idx = LINUX_MIB_TCPFULLUNDO;
2727 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2728 tp->undo_marker = 0;
2730 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2731 /* Hold old state until something *above* high_seq
2732 * is ACKed. For Reno it is MUST to prevent false
2733 * fast retransmits (RFC2582). SACK TCP is safe. */
2734 tcp_moderate_cwnd(tp);
2737 tcp_set_ca_state(sk, TCP_CA_Open);
2741 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2742 static void tcp_try_undo_dsack(struct sock *sk)
2744 struct tcp_sock *tp = tcp_sk(sk);
2746 if (tp->undo_marker && !tp->undo_retrans) {
2747 DBGUNDO(sk, "D-SACK");
2748 tcp_undo_cwr(sk, true);
2749 tp->undo_marker = 0;
2750 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2754 /* We can clear retrans_stamp when there are no retransmissions in the
2755 * window. It would seem that it is trivially available for us in
2756 * tp->retrans_out, however, that kind of assumptions doesn't consider
2757 * what will happen if errors occur when sending retransmission for the
2758 * second time. ...It could the that such segment has only
2759 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2760 * the head skb is enough except for some reneging corner cases that
2761 * are not worth the effort.
2763 * Main reason for all this complexity is the fact that connection dying
2764 * time now depends on the validity of the retrans_stamp, in particular,
2765 * that successive retransmissions of a segment must not advance
2766 * retrans_stamp under any conditions.
2768 static int tcp_any_retrans_done(const struct sock *sk)
2770 const struct tcp_sock *tp = tcp_sk(sk);
2771 struct sk_buff *skb;
2773 if (tp->retrans_out)
2776 skb = tcp_write_queue_head(sk);
2777 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2783 /* Undo during fast recovery after partial ACK. */
2785 static int tcp_try_undo_partial(struct sock *sk, int acked)
2787 struct tcp_sock *tp = tcp_sk(sk);
2788 /* Partial ACK arrived. Force Hoe's retransmit. */
2789 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2791 if (tcp_may_undo(tp)) {
2792 /* Plain luck! Hole if filled with delayed
2793 * packet, rather than with a retransmit.
2795 if (!tcp_any_retrans_done(sk))
2796 tp->retrans_stamp = 0;
2798 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2801 tcp_undo_cwr(sk, false);
2802 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2804 /* So... Do not make Hoe's retransmit yet.
2805 * If the first packet was delayed, the rest
2806 * ones are most probably delayed as well.
2813 /* Undo during loss recovery after partial ACK. */
2814 static int tcp_try_undo_loss(struct sock *sk)
2816 struct tcp_sock *tp = tcp_sk(sk);
2818 if (tcp_may_undo(tp)) {
2819 struct sk_buff *skb;
2820 tcp_for_write_queue(skb, sk) {
2821 if (skb == tcp_send_head(sk))
2823 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2826 tcp_clear_all_retrans_hints(tp);
2828 DBGUNDO(sk, "partial loss");
2830 tcp_undo_cwr(sk, true);
2831 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2832 inet_csk(sk)->icsk_retransmits = 0;
2833 tp->undo_marker = 0;
2834 if (tcp_is_sack(tp))
2835 tcp_set_ca_state(sk, TCP_CA_Open);
2841 static inline void tcp_complete_cwr(struct sock *sk)
2843 struct tcp_sock *tp = tcp_sk(sk);
2845 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2846 if (tp->undo_marker) {
2847 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR)
2848 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2850 tp->snd_cwnd = tp->snd_ssthresh;
2851 tp->snd_cwnd_stamp = tcp_time_stamp;
2853 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2856 static void tcp_try_keep_open(struct sock *sk)
2858 struct tcp_sock *tp = tcp_sk(sk);
2859 int state = TCP_CA_Open;
2861 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2862 state = TCP_CA_Disorder;
2864 if (inet_csk(sk)->icsk_ca_state != state) {
2865 tcp_set_ca_state(sk, state);
2866 tp->high_seq = tp->snd_nxt;
2870 static void tcp_try_to_open(struct sock *sk, int flag)
2872 struct tcp_sock *tp = tcp_sk(sk);
2874 tcp_verify_left_out(tp);
2876 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2877 tp->retrans_stamp = 0;
2879 if (flag & FLAG_ECE)
2880 tcp_enter_cwr(sk, 1);
2882 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2883 tcp_try_keep_open(sk);
2884 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2885 tcp_moderate_cwnd(tp);
2887 tcp_cwnd_down(sk, flag);
2891 static void tcp_mtup_probe_failed(struct sock *sk)
2893 struct inet_connection_sock *icsk = inet_csk(sk);
2895 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2896 icsk->icsk_mtup.probe_size = 0;
2899 static void tcp_mtup_probe_success(struct sock *sk)
2901 struct tcp_sock *tp = tcp_sk(sk);
2902 struct inet_connection_sock *icsk = inet_csk(sk);
2904 /* FIXME: breaks with very large cwnd */
2905 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2906 tp->snd_cwnd = tp->snd_cwnd *
2907 tcp_mss_to_mtu(sk, tp->mss_cache) /
2908 icsk->icsk_mtup.probe_size;
2909 tp->snd_cwnd_cnt = 0;
2910 tp->snd_cwnd_stamp = tcp_time_stamp;
2911 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2913 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2914 icsk->icsk_mtup.probe_size = 0;
2915 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2918 /* Do a simple retransmit without using the backoff mechanisms in
2919 * tcp_timer. This is used for path mtu discovery.
2920 * The socket is already locked here.
2922 void tcp_simple_retransmit(struct sock *sk)
2924 const struct inet_connection_sock *icsk = inet_csk(sk);
2925 struct tcp_sock *tp = tcp_sk(sk);
2926 struct sk_buff *skb;
2927 unsigned int mss = tcp_current_mss(sk);
2928 u32 prior_lost = tp->lost_out;
2930 tcp_for_write_queue(skb, sk) {
2931 if (skb == tcp_send_head(sk))
2933 if (tcp_skb_seglen(skb) > mss &&
2934 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2935 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2936 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2937 tp->retrans_out -= tcp_skb_pcount(skb);
2939 tcp_skb_mark_lost_uncond_verify(tp, skb);
2943 tcp_clear_retrans_hints_partial(tp);
2945 if (prior_lost == tp->lost_out)
2948 if (tcp_is_reno(tp))
2949 tcp_limit_reno_sacked(tp);
2951 tcp_verify_left_out(tp);
2953 /* Don't muck with the congestion window here.
2954 * Reason is that we do not increase amount of _data_
2955 * in network, but units changed and effective
2956 * cwnd/ssthresh really reduced now.
2958 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2959 tp->high_seq = tp->snd_nxt;
2960 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2961 tp->prior_ssthresh = 0;
2962 tp->undo_marker = 0;
2963 tcp_set_ca_state(sk, TCP_CA_Loss);
2965 tcp_xmit_retransmit_queue(sk);
2967 EXPORT_SYMBOL(tcp_simple_retransmit);
2969 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2970 * (proportional rate reduction with slow start reduction bound) as described in
2971 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2972 * It computes the number of packets to send (sndcnt) based on packets newly
2974 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2975 * cwnd reductions across a full RTT.
2976 * 2) If packets in flight is lower than ssthresh (such as due to excess
2977 * losses and/or application stalls), do not perform any further cwnd
2978 * reductions, but instead slow start up to ssthresh.
2980 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
2981 int fast_rexmit, int flag)
2983 struct tcp_sock *tp = tcp_sk(sk);
2985 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2987 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2988 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2990 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2992 sndcnt = min_t(int, delta,
2993 max_t(int, tp->prr_delivered - tp->prr_out,
2994 newly_acked_sacked) + 1);
2997 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2998 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3001 /* Process an event, which can update packets-in-flight not trivially.
3002 * Main goal of this function is to calculate new estimate for left_out,
3003 * taking into account both packets sitting in receiver's buffer and
3004 * packets lost by network.
3006 * Besides that it does CWND reduction, when packet loss is detected
3007 * and changes state of machine.
3009 * It does _not_ decide what to send, it is made in function
3010 * tcp_xmit_retransmit_queue().
3012 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3013 int newly_acked_sacked, bool is_dupack,
3016 struct inet_connection_sock *icsk = inet_csk(sk);
3017 struct tcp_sock *tp = tcp_sk(sk);
3018 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3019 (tcp_fackets_out(tp) > tp->reordering));
3020 int fast_rexmit = 0, mib_idx;
3022 if (WARN_ON(!tp->packets_out && tp->sacked_out))
3024 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3025 tp->fackets_out = 0;
3027 /* Now state machine starts.
3028 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3029 if (flag & FLAG_ECE)
3030 tp->prior_ssthresh = 0;
3032 /* B. In all the states check for reneging SACKs. */
3033 if (tcp_check_sack_reneging(sk, flag))
3036 /* C. Process data loss notification, provided it is valid. */
3037 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
3038 before(tp->snd_una, tp->high_seq) &&
3039 icsk->icsk_ca_state != TCP_CA_Open &&
3040 tp->fackets_out > tp->reordering) {
3041 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
3042 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
3045 /* D. Check consistency of the current state. */
3046 tcp_verify_left_out(tp);
3048 /* E. Check state exit conditions. State can be terminated
3049 * when high_seq is ACKed. */
3050 if (icsk->icsk_ca_state == TCP_CA_Open) {
3051 WARN_ON(tp->retrans_out != 0);
3052 tp->retrans_stamp = 0;
3053 } else if (!before(tp->snd_una, tp->high_seq)) {
3054 switch (icsk->icsk_ca_state) {
3056 icsk->icsk_retransmits = 0;
3057 if (tcp_try_undo_recovery(sk))
3062 /* CWR is to be held something *above* high_seq
3063 * is ACKed for CWR bit to reach receiver. */
3064 if (tp->snd_una != tp->high_seq) {
3065 tcp_complete_cwr(sk);
3066 tcp_set_ca_state(sk, TCP_CA_Open);
3070 case TCP_CA_Recovery:
3071 if (tcp_is_reno(tp))
3072 tcp_reset_reno_sack(tp);
3073 if (tcp_try_undo_recovery(sk))
3075 tcp_complete_cwr(sk);
3080 /* F. Process state. */
3081 switch (icsk->icsk_ca_state) {
3082 case TCP_CA_Recovery:
3083 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3084 if (tcp_is_reno(tp) && is_dupack)
3085 tcp_add_reno_sack(sk);
3087 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3090 if (flag & FLAG_DATA_ACKED)
3091 icsk->icsk_retransmits = 0;
3092 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3093 tcp_reset_reno_sack(tp);
3094 if (!tcp_try_undo_loss(sk)) {
3095 tcp_moderate_cwnd(tp);
3096 tcp_xmit_retransmit_queue(sk);
3099 if (icsk->icsk_ca_state != TCP_CA_Open)
3101 /* Loss is undone; fall through to processing in Open state. */
3103 if (tcp_is_reno(tp)) {
3104 if (flag & FLAG_SND_UNA_ADVANCED)
3105 tcp_reset_reno_sack(tp);
3107 tcp_add_reno_sack(sk);
3110 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3111 tcp_try_undo_dsack(sk);
3113 if (!tcp_time_to_recover(sk)) {
3114 tcp_try_to_open(sk, flag);
3118 /* MTU probe failure: don't reduce cwnd */
3119 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3120 icsk->icsk_mtup.probe_size &&
3121 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3122 tcp_mtup_probe_failed(sk);
3123 /* Restores the reduction we did in tcp_mtup_probe() */
3125 tcp_simple_retransmit(sk);
3129 /* Otherwise enter Recovery state */
3131 if (tcp_is_reno(tp))
3132 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3134 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3136 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3138 tp->high_seq = tp->snd_nxt;
3139 tp->prior_ssthresh = 0;
3140 tp->undo_marker = tp->snd_una;
3141 tp->undo_retrans = tp->retrans_out;
3143 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3144 if (!(flag & FLAG_ECE))
3145 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3146 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3147 TCP_ECN_queue_cwr(tp);
3150 tp->bytes_acked = 0;
3151 tp->snd_cwnd_cnt = 0;
3152 tp->prior_cwnd = tp->snd_cwnd;
3153 tp->prr_delivered = 0;
3155 tcp_set_ca_state(sk, TCP_CA_Recovery);
3159 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3160 tcp_update_scoreboard(sk, fast_rexmit);
3161 tp->prr_delivered += newly_acked_sacked;
3162 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3163 tcp_xmit_retransmit_queue(sk);
3166 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3168 tcp_rtt_estimator(sk, seq_rtt);
3170 inet_csk(sk)->icsk_backoff = 0;
3172 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3174 /* Read draft-ietf-tcplw-high-performance before mucking
3175 * with this code. (Supersedes RFC1323)
3177 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3179 /* RTTM Rule: A TSecr value received in a segment is used to
3180 * update the averaged RTT measurement only if the segment
3181 * acknowledges some new data, i.e., only if it advances the
3182 * left edge of the send window.
3184 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3185 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3187 * Changed: reset backoff as soon as we see the first valid sample.
3188 * If we do not, we get strongly overestimated rto. With timestamps
3189 * samples are accepted even from very old segments: f.e., when rtt=1
3190 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3191 * answer arrives rto becomes 120 seconds! If at least one of segments
3192 * in window is lost... Voila. --ANK (010210)
3194 struct tcp_sock *tp = tcp_sk(sk);
3196 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3199 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3201 /* We don't have a timestamp. Can only use
3202 * packets that are not retransmitted to determine
3203 * rtt estimates. Also, we must not reset the
3204 * backoff for rto until we get a non-retransmitted
3205 * packet. This allows us to deal with a situation
3206 * where the network delay has increased suddenly.
3207 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3210 if (flag & FLAG_RETRANS_DATA_ACKED)
3213 tcp_valid_rtt_meas(sk, seq_rtt);
3216 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3219 const struct tcp_sock *tp = tcp_sk(sk);
3220 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3221 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3222 tcp_ack_saw_tstamp(sk, flag);
3223 else if (seq_rtt >= 0)
3224 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3227 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3229 const struct inet_connection_sock *icsk = inet_csk(sk);
3230 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3231 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3234 /* Restart timer after forward progress on connection.
3235 * RFC2988 recommends to restart timer to now+rto.
3237 static void tcp_rearm_rto(struct sock *sk)
3239 const struct tcp_sock *tp = tcp_sk(sk);
3241 if (!tp->packets_out) {
3242 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3244 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3245 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3249 /* If we get here, the whole TSO packet has not been acked. */
3250 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3252 struct tcp_sock *tp = tcp_sk(sk);
3255 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3257 packets_acked = tcp_skb_pcount(skb);
3258 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3260 packets_acked -= tcp_skb_pcount(skb);
3262 if (packets_acked) {
3263 BUG_ON(tcp_skb_pcount(skb) == 0);
3264 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3267 return packets_acked;
3270 /* Remove acknowledged frames from the retransmission queue. If our packet
3271 * is before the ack sequence we can discard it as it's confirmed to have
3272 * arrived at the other end.
3274 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3277 struct tcp_sock *tp = tcp_sk(sk);
3278 const struct inet_connection_sock *icsk = inet_csk(sk);
3279 struct sk_buff *skb;
3280 u32 now = tcp_time_stamp;
3281 int fully_acked = 1;
3284 u32 reord = tp->packets_out;
3285 u32 prior_sacked = tp->sacked_out;
3287 s32 ca_seq_rtt = -1;
3288 ktime_t last_ackt = net_invalid_timestamp();
3290 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3291 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3293 u8 sacked = scb->sacked;
3295 /* Determine how many packets and what bytes were acked, tso and else */
3296 if (after(scb->end_seq, tp->snd_una)) {
3297 if (tcp_skb_pcount(skb) == 1 ||
3298 !after(tp->snd_una, scb->seq))
3301 acked_pcount = tcp_tso_acked(sk, skb);
3307 acked_pcount = tcp_skb_pcount(skb);
3310 if (sacked & TCPCB_RETRANS) {
3311 if (sacked & TCPCB_SACKED_RETRANS)
3312 tp->retrans_out -= acked_pcount;
3313 flag |= FLAG_RETRANS_DATA_ACKED;
3316 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3317 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3319 ca_seq_rtt = now - scb->when;
3320 last_ackt = skb->tstamp;
3322 seq_rtt = ca_seq_rtt;
3324 if (!(sacked & TCPCB_SACKED_ACKED))
3325 reord = min(pkts_acked, reord);
3328 if (sacked & TCPCB_SACKED_ACKED)
3329 tp->sacked_out -= acked_pcount;
3330 if (sacked & TCPCB_LOST)
3331 tp->lost_out -= acked_pcount;
3333 tp->packets_out -= acked_pcount;
3334 pkts_acked += acked_pcount;
3336 /* Initial outgoing SYN's get put onto the write_queue
3337 * just like anything else we transmit. It is not
3338 * true data, and if we misinform our callers that
3339 * this ACK acks real data, we will erroneously exit
3340 * connection startup slow start one packet too
3341 * quickly. This is severely frowned upon behavior.
3343 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3344 flag |= FLAG_DATA_ACKED;
3346 flag |= FLAG_SYN_ACKED;
3347 tp->retrans_stamp = 0;
3353 tcp_unlink_write_queue(skb, sk);
3354 sk_wmem_free_skb(sk, skb);
3355 tp->scoreboard_skb_hint = NULL;
3356 if (skb == tp->retransmit_skb_hint)
3357 tp->retransmit_skb_hint = NULL;
3358 if (skb == tp->lost_skb_hint)
3359 tp->lost_skb_hint = NULL;
3362 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3363 tp->snd_up = tp->snd_una;
3365 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3366 flag |= FLAG_SACK_RENEGING;
3368 if (flag & FLAG_ACKED) {
3369 const struct tcp_congestion_ops *ca_ops
3370 = inet_csk(sk)->icsk_ca_ops;
3372 if (unlikely(icsk->icsk_mtup.probe_size &&
3373 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3374 tcp_mtup_probe_success(sk);
3377 tcp_ack_update_rtt(sk, flag, seq_rtt);
3380 if (tcp_is_reno(tp)) {
3381 tcp_remove_reno_sacks(sk, pkts_acked);
3385 /* Non-retransmitted hole got filled? That's reordering */
3386 if (reord < prior_fackets)
3387 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3389 delta = tcp_is_fack(tp) ? pkts_acked :
3390 prior_sacked - tp->sacked_out;
3391 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3394 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3396 if (ca_ops->pkts_acked) {
3399 /* Is the ACK triggering packet unambiguous? */
3400 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3401 /* High resolution needed and available? */
3402 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3403 !ktime_equal(last_ackt,
3404 net_invalid_timestamp()))
3405 rtt_us = ktime_us_delta(ktime_get_real(),
3407 else if (ca_seq_rtt >= 0)
3408 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3411 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3415 #if FASTRETRANS_DEBUG > 0
3416 WARN_ON((int)tp->sacked_out < 0);
3417 WARN_ON((int)tp->lost_out < 0);
3418 WARN_ON((int)tp->retrans_out < 0);
3419 if (!tp->packets_out && tcp_is_sack(tp)) {
3420 icsk = inet_csk(sk);
3422 printk(KERN_DEBUG "Leak l=%u %d\n",
3423 tp->lost_out, icsk->icsk_ca_state);
3426 if (tp->sacked_out) {
3427 printk(KERN_DEBUG "Leak s=%u %d\n",
3428 tp->sacked_out, icsk->icsk_ca_state);
3431 if (tp->retrans_out) {
3432 printk(KERN_DEBUG "Leak r=%u %d\n",
3433 tp->retrans_out, icsk->icsk_ca_state);
3434 tp->retrans_out = 0;
3441 static void tcp_ack_probe(struct sock *sk)
3443 const struct tcp_sock *tp = tcp_sk(sk);
3444 struct inet_connection_sock *icsk = inet_csk(sk);
3446 /* Was it a usable window open? */
3448 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3449 icsk->icsk_backoff = 0;
3450 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3451 /* Socket must be waked up by subsequent tcp_data_snd_check().
3452 * This function is not for random using!
3455 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3456 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3461 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3463 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3464 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3467 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3469 const struct tcp_sock *tp = tcp_sk(sk);
3470 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3471 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3474 /* Check that window update is acceptable.
3475 * The function assumes that snd_una<=ack<=snd_next.
3477 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3478 const u32 ack, const u32 ack_seq,
3481 return after(ack, tp->snd_una) ||
3482 after(ack_seq, tp->snd_wl1) ||
3483 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3486 /* Update our send window.
3488 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3489 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3491 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3494 struct tcp_sock *tp = tcp_sk(sk);
3496 u32 nwin = ntohs(tcp_hdr(skb)->window);
3498 if (likely(!tcp_hdr(skb)->syn))
3499 nwin <<= tp->rx_opt.snd_wscale;
3501 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3502 flag |= FLAG_WIN_UPDATE;
3503 tcp_update_wl(tp, ack_seq);
3505 if (tp->snd_wnd != nwin) {
3508 /* Note, it is the only place, where
3509 * fast path is recovered for sending TCP.
3512 tcp_fast_path_check(sk);
3514 if (nwin > tp->max_window) {
3515 tp->max_window = nwin;
3516 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3526 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3527 * continue in congestion avoidance.
3529 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3531 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3532 tp->snd_cwnd_cnt = 0;
3533 tp->bytes_acked = 0;
3534 TCP_ECN_queue_cwr(tp);
3535 tcp_moderate_cwnd(tp);
3538 /* A conservative spurious RTO response algorithm: reduce cwnd using
3539 * rate halving and continue in congestion avoidance.
3541 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3543 tcp_enter_cwr(sk, 0);
3546 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3548 if (flag & FLAG_ECE)
3549 tcp_ratehalving_spur_to_response(sk);
3551 tcp_undo_cwr(sk, true);
3554 /* F-RTO spurious RTO detection algorithm (RFC4138)
3556 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3557 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3558 * window (but not to or beyond highest sequence sent before RTO):
3559 * On First ACK, send two new segments out.
3560 * On Second ACK, RTO was likely spurious. Do spurious response (response
3561 * algorithm is not part of the F-RTO detection algorithm
3562 * given in RFC4138 but can be selected separately).
3563 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3564 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3565 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3566 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3568 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3569 * original window even after we transmit two new data segments.
3572 * on first step, wait until first cumulative ACK arrives, then move to
3573 * the second step. In second step, the next ACK decides.
3575 * F-RTO is implemented (mainly) in four functions:
3576 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3577 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3578 * called when tcp_use_frto() showed green light
3579 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3580 * - tcp_enter_frto_loss() is called if there is not enough evidence
3581 * to prove that the RTO is indeed spurious. It transfers the control
3582 * from F-RTO to the conventional RTO recovery
3584 static int tcp_process_frto(struct sock *sk, int flag)
3586 struct tcp_sock *tp = tcp_sk(sk);
3588 tcp_verify_left_out(tp);
3590 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3591 if (flag & FLAG_DATA_ACKED)
3592 inet_csk(sk)->icsk_retransmits = 0;
3594 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3595 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3596 tp->undo_marker = 0;
3598 if (!before(tp->snd_una, tp->frto_highmark)) {
3599 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3603 if (!tcp_is_sackfrto(tp)) {
3604 /* RFC4138 shortcoming in step 2; should also have case c):
3605 * ACK isn't duplicate nor advances window, e.g., opposite dir
3608 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3611 if (!(flag & FLAG_DATA_ACKED)) {
3612 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3617 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3618 /* Prevent sending of new data. */
3619 tp->snd_cwnd = min(tp->snd_cwnd,
3620 tcp_packets_in_flight(tp));
3624 if ((tp->frto_counter >= 2) &&
3625 (!(flag & FLAG_FORWARD_PROGRESS) ||
3626 ((flag & FLAG_DATA_SACKED) &&
3627 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3628 /* RFC4138 shortcoming (see comment above) */
3629 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3630 (flag & FLAG_NOT_DUP))
3633 tcp_enter_frto_loss(sk, 3, flag);
3638 if (tp->frto_counter == 1) {
3639 /* tcp_may_send_now needs to see updated state */
3640 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3641 tp->frto_counter = 2;
3643 if (!tcp_may_send_now(sk))
3644 tcp_enter_frto_loss(sk, 2, flag);
3648 switch (sysctl_tcp_frto_response) {
3650 tcp_undo_spur_to_response(sk, flag);
3653 tcp_conservative_spur_to_response(tp);
3656 tcp_ratehalving_spur_to_response(sk);
3659 tp->frto_counter = 0;
3660 tp->undo_marker = 0;
3661 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3666 /* This routine deals with incoming acks, but not outgoing ones. */
3667 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3669 struct inet_connection_sock *icsk = inet_csk(sk);
3670 struct tcp_sock *tp = tcp_sk(sk);
3671 u32 prior_snd_una = tp->snd_una;
3672 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3673 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3674 bool is_dupack = false;
3675 u32 prior_in_flight;
3678 int prior_sacked = tp->sacked_out;
3680 int newly_acked_sacked = 0;
3683 /* If the ack is older than previous acks
3684 * then we can probably ignore it.
3686 if (before(ack, prior_snd_una))
3689 /* If the ack includes data we haven't sent yet, discard
3690 * this segment (RFC793 Section 3.9).
3692 if (after(ack, tp->snd_nxt))
3695 if (after(ack, prior_snd_una))
3696 flag |= FLAG_SND_UNA_ADVANCED;
3698 if (sysctl_tcp_abc) {
3699 if (icsk->icsk_ca_state < TCP_CA_CWR)
3700 tp->bytes_acked += ack - prior_snd_una;
3701 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3702 /* we assume just one segment left network */
3703 tp->bytes_acked += min(ack - prior_snd_una,
3707 prior_fackets = tp->fackets_out;
3708 prior_in_flight = tcp_packets_in_flight(tp);
3710 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3711 /* Window is constant, pure forward advance.
3712 * No more checks are required.
3713 * Note, we use the fact that SND.UNA>=SND.WL2.
3715 tcp_update_wl(tp, ack_seq);
3717 flag |= FLAG_WIN_UPDATE;
3719 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3721 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3723 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3726 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3728 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3730 if (TCP_SKB_CB(skb)->sacked)
3731 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3733 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3736 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3739 /* We passed data and got it acked, remove any soft error
3740 * log. Something worked...
3742 sk->sk_err_soft = 0;
3743 icsk->icsk_probes_out = 0;
3744 tp->rcv_tstamp = tcp_time_stamp;
3745 prior_packets = tp->packets_out;
3749 /* See if we can take anything off of the retransmit queue. */
3750 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3752 pkts_acked = prior_packets - tp->packets_out;
3753 newly_acked_sacked = (prior_packets - prior_sacked) -
3754 (tp->packets_out - tp->sacked_out);
3756 if (tp->frto_counter)
3757 frto_cwnd = tcp_process_frto(sk, flag);
3758 /* Guarantee sacktag reordering detection against wrap-arounds */
3759 if (before(tp->frto_highmark, tp->snd_una))
3760 tp->frto_highmark = 0;
3762 if (tcp_ack_is_dubious(sk, flag)) {
3763 /* Advance CWND, if state allows this. */
3764 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3765 tcp_may_raise_cwnd(sk, flag))
3766 tcp_cong_avoid(sk, ack, prior_in_flight);
3767 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3768 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3771 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3772 tcp_cong_avoid(sk, ack, prior_in_flight);
3775 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3776 dst_confirm(__sk_dst_get(sk));
3781 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3782 if (flag & FLAG_DSACKING_ACK)
3783 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3785 /* If this ack opens up a zero window, clear backoff. It was
3786 * being used to time the probes, and is probably far higher than
3787 * it needs to be for normal retransmission.
3789 if (tcp_send_head(sk))
3794 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3798 /* If data was SACKed, tag it and see if we should send more data.
3799 * If data was DSACKed, see if we can undo a cwnd reduction.
3801 if (TCP_SKB_CB(skb)->sacked) {
3802 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3803 newly_acked_sacked = tp->sacked_out - prior_sacked;
3804 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3808 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3812 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3813 * But, this can also be called on packets in the established flow when
3814 * the fast version below fails.
3816 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3817 const u8 **hvpp, int estab)
3819 const unsigned char *ptr;
3820 const struct tcphdr *th = tcp_hdr(skb);
3821 int length = (th->doff * 4) - sizeof(struct tcphdr);
3823 ptr = (const unsigned char *)(th + 1);
3824 opt_rx->saw_tstamp = 0;
3826 while (length > 0) {
3827 int opcode = *ptr++;
3833 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3838 if (opsize < 2) /* "silly options" */
3840 if (opsize > length)
3841 return; /* don't parse partial options */
3844 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3845 u16 in_mss = get_unaligned_be16(ptr);
3847 if (opt_rx->user_mss &&
3848 opt_rx->user_mss < in_mss)
3849 in_mss = opt_rx->user_mss;
3850 opt_rx->mss_clamp = in_mss;
3855 if (opsize == TCPOLEN_WINDOW && th->syn &&
3856 !estab && sysctl_tcp_window_scaling) {
3857 __u8 snd_wscale = *(__u8 *)ptr;
3858 opt_rx->wscale_ok = 1;
3859 if (snd_wscale > 14) {
3860 if (net_ratelimit())
3861 printk(KERN_INFO "tcp_parse_options: Illegal window "
3862 "scaling value %d >14 received.\n",
3866 opt_rx->snd_wscale = snd_wscale;
3869 case TCPOPT_TIMESTAMP:
3870 if ((opsize == TCPOLEN_TIMESTAMP) &&
3871 ((estab && opt_rx->tstamp_ok) ||
3872 (!estab && sysctl_tcp_timestamps))) {
3873 opt_rx->saw_tstamp = 1;
3874 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3875 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3878 case TCPOPT_SACK_PERM:
3879 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3880 !estab && sysctl_tcp_sack) {
3881 opt_rx->sack_ok = 1;
3882 tcp_sack_reset(opt_rx);
3887 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3888 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3890 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3893 #ifdef CONFIG_TCP_MD5SIG
3896 * The MD5 Hash has already been
3897 * checked (see tcp_v{4,6}_do_rcv()).
3902 /* This option is variable length.
3905 case TCPOLEN_COOKIE_BASE:
3906 /* not yet implemented */
3908 case TCPOLEN_COOKIE_PAIR:
3909 /* not yet implemented */
3911 case TCPOLEN_COOKIE_MIN+0:
3912 case TCPOLEN_COOKIE_MIN+2:
3913 case TCPOLEN_COOKIE_MIN+4:
3914 case TCPOLEN_COOKIE_MIN+6:
3915 case TCPOLEN_COOKIE_MAX:
3916 /* 16-bit multiple */
3917 opt_rx->cookie_plus = opsize;
3932 EXPORT_SYMBOL(tcp_parse_options);
3934 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3936 const __be32 *ptr = (const __be32 *)(th + 1);
3938 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3939 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3940 tp->rx_opt.saw_tstamp = 1;
3942 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3944 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3950 /* Fast parse options. This hopes to only see timestamps.
3951 * If it is wrong it falls back on tcp_parse_options().
3953 static int tcp_fast_parse_options(const struct sk_buff *skb,
3954 const struct tcphdr *th,
3955 struct tcp_sock *tp, const u8 **hvpp)
3957 /* In the spirit of fast parsing, compare doff directly to constant
3958 * values. Because equality is used, short doff can be ignored here.
3960 if (th->doff == (sizeof(*th) / 4)) {
3961 tp->rx_opt.saw_tstamp = 0;
3963 } else if (tp->rx_opt.tstamp_ok &&
3964 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3965 if (tcp_parse_aligned_timestamp(tp, th))
3968 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3972 #ifdef CONFIG_TCP_MD5SIG
3974 * Parse MD5 Signature option
3976 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3978 int length = (th->doff << 2) - sizeof(*th);
3979 const u8 *ptr = (const u8 *)(th + 1);
3981 /* If the TCP option is too short, we can short cut */
3982 if (length < TCPOLEN_MD5SIG)
3985 while (length > 0) {
3986 int opcode = *ptr++;
3997 if (opsize < 2 || opsize > length)
3999 if (opcode == TCPOPT_MD5SIG)
4000 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4007 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4010 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4012 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
4013 tp->rx_opt.ts_recent_stamp = get_seconds();
4016 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4018 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4019 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
4020 * extra check below makes sure this can only happen
4021 * for pure ACK frames. -DaveM
4023 * Not only, also it occurs for expired timestamps.
4026 if (tcp_paws_check(&tp->rx_opt, 0))
4027 tcp_store_ts_recent(tp);
4031 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4033 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4034 * it can pass through stack. So, the following predicate verifies that
4035 * this segment is not used for anything but congestion avoidance or
4036 * fast retransmit. Moreover, we even are able to eliminate most of such
4037 * second order effects, if we apply some small "replay" window (~RTO)
4038 * to timestamp space.
4040 * All these measures still do not guarantee that we reject wrapped ACKs
4041 * on networks with high bandwidth, when sequence space is recycled fastly,
4042 * but it guarantees that such events will be very rare and do not affect
4043 * connection seriously. This doesn't look nice, but alas, PAWS is really
4046 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4047 * states that events when retransmit arrives after original data are rare.
4048 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4049 * the biggest problem on large power networks even with minor reordering.
4050 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4051 * up to bandwidth of 18Gigabit/sec. 8) ]
4054 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4056 const struct tcp_sock *tp = tcp_sk(sk);
4057 const struct tcphdr *th = tcp_hdr(skb);
4058 u32 seq = TCP_SKB_CB(skb)->seq;
4059 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4061 return (/* 1. Pure ACK with correct sequence number. */
4062 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4064 /* 2. ... and duplicate ACK. */
4065 ack == tp->snd_una &&
4067 /* 3. ... and does not update window. */
4068 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4070 /* 4. ... and sits in replay window. */
4071 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4074 static inline int tcp_paws_discard(const struct sock *sk,
4075 const struct sk_buff *skb)
4077 const struct tcp_sock *tp = tcp_sk(sk);
4079 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4080 !tcp_disordered_ack(sk, skb);
4083 /* Check segment sequence number for validity.
4085 * Segment controls are considered valid, if the segment
4086 * fits to the window after truncation to the window. Acceptability
4087 * of data (and SYN, FIN, of course) is checked separately.
4088 * See tcp_data_queue(), for example.
4090 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4091 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4092 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4093 * (borrowed from freebsd)
4096 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4098 return !before(end_seq, tp->rcv_wup) &&
4099 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4102 /* When we get a reset we do this. */
4103 static void tcp_reset(struct sock *sk)
4105 /* We want the right error as BSD sees it (and indeed as we do). */
4106 switch (sk->sk_state) {
4108 sk->sk_err = ECONNREFUSED;
4110 case TCP_CLOSE_WAIT:
4116 sk->sk_err = ECONNRESET;
4118 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4121 if (!sock_flag(sk, SOCK_DEAD))
4122 sk->sk_error_report(sk);
4128 * Process the FIN bit. This now behaves as it is supposed to work
4129 * and the FIN takes effect when it is validly part of sequence
4130 * space. Not before when we get holes.
4132 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4133 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4136 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4137 * close and we go into CLOSING (and later onto TIME-WAIT)
4139 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4141 static void tcp_fin(struct sock *sk)
4143 struct tcp_sock *tp = tcp_sk(sk);
4145 inet_csk_schedule_ack(sk);
4147 sk->sk_shutdown |= RCV_SHUTDOWN;
4148 sock_set_flag(sk, SOCK_DONE);
4150 switch (sk->sk_state) {
4152 case TCP_ESTABLISHED:
4153 /* Move to CLOSE_WAIT */
4154 tcp_set_state(sk, TCP_CLOSE_WAIT);
4155 inet_csk(sk)->icsk_ack.pingpong = 1;
4158 case TCP_CLOSE_WAIT:
4160 /* Received a retransmission of the FIN, do
4165 /* RFC793: Remain in the LAST-ACK state. */
4169 /* This case occurs when a simultaneous close
4170 * happens, we must ack the received FIN and
4171 * enter the CLOSING state.
4174 tcp_set_state(sk, TCP_CLOSING);
4177 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4179 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4182 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4183 * cases we should never reach this piece of code.
4185 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4186 __func__, sk->sk_state);
4190 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4191 * Probably, we should reset in this case. For now drop them.
4193 __skb_queue_purge(&tp->out_of_order_queue);
4194 if (tcp_is_sack(tp))
4195 tcp_sack_reset(&tp->rx_opt);
4198 if (!sock_flag(sk, SOCK_DEAD)) {
4199 sk->sk_state_change(sk);
4201 /* Do not send POLL_HUP for half duplex close. */
4202 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4203 sk->sk_state == TCP_CLOSE)
4204 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4206 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4210 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4213 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4214 if (before(seq, sp->start_seq))
4215 sp->start_seq = seq;
4216 if (after(end_seq, sp->end_seq))
4217 sp->end_seq = end_seq;
4223 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4225 struct tcp_sock *tp = tcp_sk(sk);
4227 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4230 if (before(seq, tp->rcv_nxt))
4231 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4233 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4235 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4237 tp->rx_opt.dsack = 1;
4238 tp->duplicate_sack[0].start_seq = seq;
4239 tp->duplicate_sack[0].end_seq = end_seq;
4243 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4245 struct tcp_sock *tp = tcp_sk(sk);
4247 if (!tp->rx_opt.dsack)
4248 tcp_dsack_set(sk, seq, end_seq);
4250 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4253 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4255 struct tcp_sock *tp = tcp_sk(sk);
4257 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4258 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4259 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4260 tcp_enter_quickack_mode(sk);
4262 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4263 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4265 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4266 end_seq = tp->rcv_nxt;
4267 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4274 /* These routines update the SACK block as out-of-order packets arrive or
4275 * in-order packets close up the sequence space.
4277 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4280 struct tcp_sack_block *sp = &tp->selective_acks[0];
4281 struct tcp_sack_block *swalk = sp + 1;
4283 /* See if the recent change to the first SACK eats into
4284 * or hits the sequence space of other SACK blocks, if so coalesce.
4286 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4287 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4290 /* Zap SWALK, by moving every further SACK up by one slot.
4291 * Decrease num_sacks.
4293 tp->rx_opt.num_sacks--;
4294 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4298 this_sack++, swalk++;
4302 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4304 struct tcp_sock *tp = tcp_sk(sk);
4305 struct tcp_sack_block *sp = &tp->selective_acks[0];
4306 int cur_sacks = tp->rx_opt.num_sacks;
4312 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4313 if (tcp_sack_extend(sp, seq, end_seq)) {
4314 /* Rotate this_sack to the first one. */
4315 for (; this_sack > 0; this_sack--, sp--)
4316 swap(*sp, *(sp - 1));
4318 tcp_sack_maybe_coalesce(tp);
4323 /* Could not find an adjacent existing SACK, build a new one,
4324 * put it at the front, and shift everyone else down. We
4325 * always know there is at least one SACK present already here.
4327 * If the sack array is full, forget about the last one.
4329 if (this_sack >= TCP_NUM_SACKS) {
4331 tp->rx_opt.num_sacks--;
4334 for (; this_sack > 0; this_sack--, sp--)
4338 /* Build the new head SACK, and we're done. */
4339 sp->start_seq = seq;
4340 sp->end_seq = end_seq;
4341 tp->rx_opt.num_sacks++;
4344 /* RCV.NXT advances, some SACKs should be eaten. */
4346 static void tcp_sack_remove(struct tcp_sock *tp)
4348 struct tcp_sack_block *sp = &tp->selective_acks[0];
4349 int num_sacks = tp->rx_opt.num_sacks;
4352 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4353 if (skb_queue_empty(&tp->out_of_order_queue)) {
4354 tp->rx_opt.num_sacks = 0;
4358 for (this_sack = 0; this_sack < num_sacks;) {
4359 /* Check if the start of the sack is covered by RCV.NXT. */
4360 if (!before(tp->rcv_nxt, sp->start_seq)) {
4363 /* RCV.NXT must cover all the block! */
4364 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4366 /* Zap this SACK, by moving forward any other SACKS. */
4367 for (i=this_sack+1; i < num_sacks; i++)
4368 tp->selective_acks[i-1] = tp->selective_acks[i];
4375 tp->rx_opt.num_sacks = num_sacks;
4378 /* This one checks to see if we can put data from the
4379 * out_of_order queue into the receive_queue.
4381 static void tcp_ofo_queue(struct sock *sk)
4383 struct tcp_sock *tp = tcp_sk(sk);
4384 __u32 dsack_high = tp->rcv_nxt;
4385 struct sk_buff *skb;
4387 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4388 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4391 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4392 __u32 dsack = dsack_high;
4393 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4394 dsack_high = TCP_SKB_CB(skb)->end_seq;
4395 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4398 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4399 SOCK_DEBUG(sk, "ofo packet was already received\n");
4400 __skb_unlink(skb, &tp->out_of_order_queue);
4404 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4405 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4406 TCP_SKB_CB(skb)->end_seq);
4408 __skb_unlink(skb, &tp->out_of_order_queue);
4409 __skb_queue_tail(&sk->sk_receive_queue, skb);
4410 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4411 if (tcp_hdr(skb)->fin)
4416 static int tcp_prune_ofo_queue(struct sock *sk);
4417 static int tcp_prune_queue(struct sock *sk);
4419 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4421 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4422 !sk_rmem_schedule(sk, size)) {
4424 if (tcp_prune_queue(sk) < 0)
4427 if (!sk_rmem_schedule(sk, size)) {
4428 if (!tcp_prune_ofo_queue(sk))
4431 if (!sk_rmem_schedule(sk, size))
4438 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4440 const struct tcphdr *th = tcp_hdr(skb);
4441 struct tcp_sock *tp = tcp_sk(sk);
4444 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4448 __skb_pull(skb, th->doff * 4);
4450 TCP_ECN_accept_cwr(tp, skb);
4452 tp->rx_opt.dsack = 0;
4454 /* Queue data for delivery to the user.
4455 * Packets in sequence go to the receive queue.
4456 * Out of sequence packets to the out_of_order_queue.
4458 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4459 if (tcp_receive_window(tp) == 0)
4462 /* Ok. In sequence. In window. */
4463 if (tp->ucopy.task == current &&
4464 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4465 sock_owned_by_user(sk) && !tp->urg_data) {
4466 int chunk = min_t(unsigned int, skb->len,
4469 __set_current_state(TASK_RUNNING);
4472 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4473 tp->ucopy.len -= chunk;
4474 tp->copied_seq += chunk;
4475 eaten = (chunk == skb->len);
4476 tcp_rcv_space_adjust(sk);
4484 tcp_try_rmem_schedule(sk, skb->truesize))
4487 skb_set_owner_r(skb, sk);
4488 __skb_queue_tail(&sk->sk_receive_queue, skb);
4490 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4492 tcp_event_data_recv(sk, skb);
4496 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4499 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4500 * gap in queue is filled.
4502 if (skb_queue_empty(&tp->out_of_order_queue))
4503 inet_csk(sk)->icsk_ack.pingpong = 0;
4506 if (tp->rx_opt.num_sacks)
4507 tcp_sack_remove(tp);
4509 tcp_fast_path_check(sk);
4513 else if (!sock_flag(sk, SOCK_DEAD))
4514 sk->sk_data_ready(sk, 0);
4518 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4519 /* A retransmit, 2nd most common case. Force an immediate ack. */
4520 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4521 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4524 tcp_enter_quickack_mode(sk);
4525 inet_csk_schedule_ack(sk);
4531 /* Out of window. F.e. zero window probe. */
4532 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4535 tcp_enter_quickack_mode(sk);
4537 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4538 /* Partial packet, seq < rcv_next < end_seq */
4539 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4540 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4541 TCP_SKB_CB(skb)->end_seq);
4543 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4545 /* If window is closed, drop tail of packet. But after
4546 * remembering D-SACK for its head made in previous line.
4548 if (!tcp_receive_window(tp))
4553 TCP_ECN_check_ce(tp, skb);
4555 if (tcp_try_rmem_schedule(sk, skb->truesize))
4558 /* Disable header prediction. */
4560 inet_csk_schedule_ack(sk);
4562 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4563 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4565 skb_set_owner_r(skb, sk);
4567 if (!skb_peek(&tp->out_of_order_queue)) {
4568 /* Initial out of order segment, build 1 SACK. */
4569 if (tcp_is_sack(tp)) {
4570 tp->rx_opt.num_sacks = 1;
4571 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4572 tp->selective_acks[0].end_seq =
4573 TCP_SKB_CB(skb)->end_seq;
4575 __skb_queue_head(&tp->out_of_order_queue, skb);
4577 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4578 u32 seq = TCP_SKB_CB(skb)->seq;
4579 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4581 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4582 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4584 if (!tp->rx_opt.num_sacks ||
4585 tp->selective_acks[0].end_seq != seq)
4588 /* Common case: data arrive in order after hole. */
4589 tp->selective_acks[0].end_seq = end_seq;
4593 /* Find place to insert this segment. */
4595 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4597 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4601 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4604 /* Do skb overlap to previous one? */
4605 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4606 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4607 /* All the bits are present. Drop. */
4609 tcp_dsack_set(sk, seq, end_seq);
4612 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4613 /* Partial overlap. */
4614 tcp_dsack_set(sk, seq,
4615 TCP_SKB_CB(skb1)->end_seq);
4617 if (skb_queue_is_first(&tp->out_of_order_queue,
4621 skb1 = skb_queue_prev(
4622 &tp->out_of_order_queue,
4627 __skb_queue_head(&tp->out_of_order_queue, skb);
4629 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4631 /* And clean segments covered by new one as whole. */
4632 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4633 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4635 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4637 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4638 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4642 __skb_unlink(skb1, &tp->out_of_order_queue);
4643 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4644 TCP_SKB_CB(skb1)->end_seq);
4649 if (tcp_is_sack(tp))
4650 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4654 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4655 struct sk_buff_head *list)
4657 struct sk_buff *next = NULL;
4659 if (!skb_queue_is_last(list, skb))
4660 next = skb_queue_next(list, skb);
4662 __skb_unlink(skb, list);
4664 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4669 /* Collapse contiguous sequence of skbs head..tail with
4670 * sequence numbers start..end.
4672 * If tail is NULL, this means until the end of the list.
4674 * Segments with FIN/SYN are not collapsed (only because this
4678 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4679 struct sk_buff *head, struct sk_buff *tail,
4682 struct sk_buff *skb, *n;
4685 /* First, check that queue is collapsible and find
4686 * the point where collapsing can be useful. */
4690 skb_queue_walk_from_safe(list, skb, n) {
4693 /* No new bits? It is possible on ofo queue. */
4694 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4695 skb = tcp_collapse_one(sk, skb, list);
4701 /* The first skb to collapse is:
4703 * - bloated or contains data before "start" or
4704 * overlaps to the next one.
4706 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4707 (tcp_win_from_space(skb->truesize) > skb->len ||
4708 before(TCP_SKB_CB(skb)->seq, start))) {
4709 end_of_skbs = false;
4713 if (!skb_queue_is_last(list, skb)) {
4714 struct sk_buff *next = skb_queue_next(list, skb);
4716 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4717 end_of_skbs = false;
4722 /* Decided to skip this, advance start seq. */
4723 start = TCP_SKB_CB(skb)->end_seq;
4725 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4728 while (before(start, end)) {
4729 struct sk_buff *nskb;
4730 unsigned int header = skb_headroom(skb);
4731 int copy = SKB_MAX_ORDER(header, 0);
4733 /* Too big header? This can happen with IPv6. */
4736 if (end - start < copy)
4738 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4742 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4743 skb_set_network_header(nskb, (skb_network_header(skb) -
4745 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4747 skb_reserve(nskb, header);
4748 memcpy(nskb->head, skb->head, header);
4749 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4750 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4751 __skb_queue_before(list, skb, nskb);
4752 skb_set_owner_r(nskb, sk);
4754 /* Copy data, releasing collapsed skbs. */
4756 int offset = start - TCP_SKB_CB(skb)->seq;
4757 int size = TCP_SKB_CB(skb)->end_seq - start;
4761 size = min(copy, size);
4762 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4764 TCP_SKB_CB(nskb)->end_seq += size;
4768 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4769 skb = tcp_collapse_one(sk, skb, list);
4772 tcp_hdr(skb)->syn ||
4780 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4781 * and tcp_collapse() them until all the queue is collapsed.
4783 static void tcp_collapse_ofo_queue(struct sock *sk)
4785 struct tcp_sock *tp = tcp_sk(sk);
4786 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4787 struct sk_buff *head;
4793 start = TCP_SKB_CB(skb)->seq;
4794 end = TCP_SKB_CB(skb)->end_seq;
4798 struct sk_buff *next = NULL;
4800 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4801 next = skb_queue_next(&tp->out_of_order_queue, skb);
4804 /* Segment is terminated when we see gap or when
4805 * we are at the end of all the queue. */
4807 after(TCP_SKB_CB(skb)->seq, end) ||
4808 before(TCP_SKB_CB(skb)->end_seq, start)) {
4809 tcp_collapse(sk, &tp->out_of_order_queue,
4810 head, skb, start, end);
4814 /* Start new segment */
4815 start = TCP_SKB_CB(skb)->seq;
4816 end = TCP_SKB_CB(skb)->end_seq;
4818 if (before(TCP_SKB_CB(skb)->seq, start))
4819 start = TCP_SKB_CB(skb)->seq;
4820 if (after(TCP_SKB_CB(skb)->end_seq, end))
4821 end = TCP_SKB_CB(skb)->end_seq;
4827 * Purge the out-of-order queue.
4828 * Return true if queue was pruned.
4830 static int tcp_prune_ofo_queue(struct sock *sk)
4832 struct tcp_sock *tp = tcp_sk(sk);
4835 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4836 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4837 __skb_queue_purge(&tp->out_of_order_queue);
4839 /* Reset SACK state. A conforming SACK implementation will
4840 * do the same at a timeout based retransmit. When a connection
4841 * is in a sad state like this, we care only about integrity
4842 * of the connection not performance.
4844 if (tp->rx_opt.sack_ok)
4845 tcp_sack_reset(&tp->rx_opt);
4852 /* Reduce allocated memory if we can, trying to get
4853 * the socket within its memory limits again.
4855 * Return less than zero if we should start dropping frames
4856 * until the socket owning process reads some of the data
4857 * to stabilize the situation.
4859 static int tcp_prune_queue(struct sock *sk)
4861 struct tcp_sock *tp = tcp_sk(sk);
4863 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4865 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4867 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4868 tcp_clamp_window(sk);
4869 else if (tcp_memory_pressure)
4870 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4872 tcp_collapse_ofo_queue(sk);
4873 if (!skb_queue_empty(&sk->sk_receive_queue))
4874 tcp_collapse(sk, &sk->sk_receive_queue,
4875 skb_peek(&sk->sk_receive_queue),
4877 tp->copied_seq, tp->rcv_nxt);
4880 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4883 /* Collapsing did not help, destructive actions follow.
4884 * This must not ever occur. */
4886 tcp_prune_ofo_queue(sk);
4888 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4891 /* If we are really being abused, tell the caller to silently
4892 * drop receive data on the floor. It will get retransmitted
4893 * and hopefully then we'll have sufficient space.
4895 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4897 /* Massive buffer overcommit. */
4902 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4903 * As additional protections, we do not touch cwnd in retransmission phases,
4904 * and if application hit its sndbuf limit recently.
4906 void tcp_cwnd_application_limited(struct sock *sk)
4908 struct tcp_sock *tp = tcp_sk(sk);
4910 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4911 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4912 /* Limited by application or receiver window. */
4913 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4914 u32 win_used = max(tp->snd_cwnd_used, init_win);
4915 if (win_used < tp->snd_cwnd) {
4916 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4917 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4919 tp->snd_cwnd_used = 0;
4921 tp->snd_cwnd_stamp = tcp_time_stamp;
4924 static int tcp_should_expand_sndbuf(const struct sock *sk)
4926 const struct tcp_sock *tp = tcp_sk(sk);
4928 /* If the user specified a specific send buffer setting, do
4931 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4934 /* If we are under global TCP memory pressure, do not expand. */
4935 if (tcp_memory_pressure)
4938 /* If we are under soft global TCP memory pressure, do not expand. */
4939 if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4942 /* If we filled the congestion window, do not expand. */
4943 if (tp->packets_out >= tp->snd_cwnd)
4949 /* When incoming ACK allowed to free some skb from write_queue,
4950 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4951 * on the exit from tcp input handler.
4953 * PROBLEM: sndbuf expansion does not work well with largesend.
4955 static void tcp_new_space(struct sock *sk)
4957 struct tcp_sock *tp = tcp_sk(sk);
4959 if (tcp_should_expand_sndbuf(sk)) {
4960 int sndmem = SKB_TRUESIZE(max_t(u32,
4961 tp->rx_opt.mss_clamp,
4964 int demanded = max_t(unsigned int, tp->snd_cwnd,
4965 tp->reordering + 1);
4966 sndmem *= 2 * demanded;
4967 if (sndmem > sk->sk_sndbuf)
4968 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4969 tp->snd_cwnd_stamp = tcp_time_stamp;
4972 sk->sk_write_space(sk);
4975 static void tcp_check_space(struct sock *sk)
4977 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4978 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4979 if (sk->sk_socket &&
4980 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4985 static inline void tcp_data_snd_check(struct sock *sk)
4987 tcp_push_pending_frames(sk);
4988 tcp_check_space(sk);
4992 * Check if sending an ack is needed.
4994 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4996 struct tcp_sock *tp = tcp_sk(sk);
4998 /* More than one full frame received... */
4999 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5000 /* ... and right edge of window advances far enough.
5001 * (tcp_recvmsg() will send ACK otherwise). Or...
5003 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5004 /* We ACK each frame or... */
5005 tcp_in_quickack_mode(sk) ||
5006 /* We have out of order data. */
5007 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5008 /* Then ack it now */
5011 /* Else, send delayed ack. */
5012 tcp_send_delayed_ack(sk);
5016 static inline void tcp_ack_snd_check(struct sock *sk)
5018 if (!inet_csk_ack_scheduled(sk)) {
5019 /* We sent a data segment already. */
5022 __tcp_ack_snd_check(sk, 1);
5026 * This routine is only called when we have urgent data
5027 * signaled. Its the 'slow' part of tcp_urg. It could be
5028 * moved inline now as tcp_urg is only called from one
5029 * place. We handle URGent data wrong. We have to - as
5030 * BSD still doesn't use the correction from RFC961.
5031 * For 1003.1g we should support a new option TCP_STDURG to permit
5032 * either form (or just set the sysctl tcp_stdurg).
5035 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5037 struct tcp_sock *tp = tcp_sk(sk);
5038 u32 ptr = ntohs(th->urg_ptr);
5040 if (ptr && !sysctl_tcp_stdurg)
5042 ptr += ntohl(th->seq);
5044 /* Ignore urgent data that we've already seen and read. */
5045 if (after(tp->copied_seq, ptr))
5048 /* Do not replay urg ptr.
5050 * NOTE: interesting situation not covered by specs.
5051 * Misbehaving sender may send urg ptr, pointing to segment,
5052 * which we already have in ofo queue. We are not able to fetch
5053 * such data and will stay in TCP_URG_NOTYET until will be eaten
5054 * by recvmsg(). Seems, we are not obliged to handle such wicked
5055 * situations. But it is worth to think about possibility of some
5056 * DoSes using some hypothetical application level deadlock.
5058 if (before(ptr, tp->rcv_nxt))
5061 /* Do we already have a newer (or duplicate) urgent pointer? */
5062 if (tp->urg_data && !after(ptr, tp->urg_seq))
5065 /* Tell the world about our new urgent pointer. */
5068 /* We may be adding urgent data when the last byte read was
5069 * urgent. To do this requires some care. We cannot just ignore
5070 * tp->copied_seq since we would read the last urgent byte again
5071 * as data, nor can we alter copied_seq until this data arrives
5072 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5074 * NOTE. Double Dutch. Rendering to plain English: author of comment
5075 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5076 * and expect that both A and B disappear from stream. This is _wrong_.
5077 * Though this happens in BSD with high probability, this is occasional.
5078 * Any application relying on this is buggy. Note also, that fix "works"
5079 * only in this artificial test. Insert some normal data between A and B and we will
5080 * decline of BSD again. Verdict: it is better to remove to trap
5083 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5084 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5085 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5087 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5088 __skb_unlink(skb, &sk->sk_receive_queue);
5093 tp->urg_data = TCP_URG_NOTYET;
5096 /* Disable header prediction. */
5100 /* This is the 'fast' part of urgent handling. */
5101 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5103 struct tcp_sock *tp = tcp_sk(sk);
5105 /* Check if we get a new urgent pointer - normally not. */
5107 tcp_check_urg(sk, th);
5109 /* Do we wait for any urgent data? - normally not... */
5110 if (tp->urg_data == TCP_URG_NOTYET) {
5111 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5114 /* Is the urgent pointer pointing into this packet? */
5115 if (ptr < skb->len) {
5117 if (skb_copy_bits(skb, ptr, &tmp, 1))
5119 tp->urg_data = TCP_URG_VALID | tmp;
5120 if (!sock_flag(sk, SOCK_DEAD))
5121 sk->sk_data_ready(sk, 0);
5126 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5128 struct tcp_sock *tp = tcp_sk(sk);
5129 int chunk = skb->len - hlen;
5133 if (skb_csum_unnecessary(skb))
5134 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5136 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5140 tp->ucopy.len -= chunk;
5141 tp->copied_seq += chunk;
5142 tcp_rcv_space_adjust(sk);
5149 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5150 struct sk_buff *skb)
5154 if (sock_owned_by_user(sk)) {
5156 result = __tcp_checksum_complete(skb);
5159 result = __tcp_checksum_complete(skb);
5164 static inline int tcp_checksum_complete_user(struct sock *sk,
5165 struct sk_buff *skb)
5167 return !skb_csum_unnecessary(skb) &&
5168 __tcp_checksum_complete_user(sk, skb);
5171 #ifdef CONFIG_NET_DMA
5172 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5175 struct tcp_sock *tp = tcp_sk(sk);
5176 int chunk = skb->len - hlen;
5178 int copied_early = 0;
5180 if (tp->ucopy.wakeup)
5183 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5184 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5186 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5188 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5190 tp->ucopy.iov, chunk,
5191 tp->ucopy.pinned_list);
5196 tp->ucopy.dma_cookie = dma_cookie;
5199 tp->ucopy.len -= chunk;
5200 tp->copied_seq += chunk;
5201 tcp_rcv_space_adjust(sk);
5203 if ((tp->ucopy.len == 0) ||
5204 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5205 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5206 tp->ucopy.wakeup = 1;
5207 sk->sk_data_ready(sk, 0);
5209 } else if (chunk > 0) {
5210 tp->ucopy.wakeup = 1;
5211 sk->sk_data_ready(sk, 0);
5214 return copied_early;
5216 #endif /* CONFIG_NET_DMA */
5218 /* Does PAWS and seqno based validation of an incoming segment, flags will
5219 * play significant role here.
5221 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5222 const struct tcphdr *th, int syn_inerr)
5224 const u8 *hash_location;
5225 struct tcp_sock *tp = tcp_sk(sk);
5227 /* RFC1323: H1. Apply PAWS check first. */
5228 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5229 tp->rx_opt.saw_tstamp &&
5230 tcp_paws_discard(sk, skb)) {
5232 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5233 tcp_send_dupack(sk, skb);
5236 /* Reset is accepted even if it did not pass PAWS. */
5239 /* Step 1: check sequence number */
5240 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5241 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5242 * (RST) segments are validated by checking their SEQ-fields."
5243 * And page 69: "If an incoming segment is not acceptable,
5244 * an acknowledgment should be sent in reply (unless the RST
5245 * bit is set, if so drop the segment and return)".
5248 tcp_send_dupack(sk, skb);
5252 /* Step 2: check RST bit */
5258 /* ts_recent update must be made after we are sure that the packet
5261 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5263 /* step 3: check security and precedence [ignored] */
5265 /* step 4: Check for a SYN in window. */
5266 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5268 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5269 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5282 * TCP receive function for the ESTABLISHED state.
5284 * It is split into a fast path and a slow path. The fast path is
5286 * - A zero window was announced from us - zero window probing
5287 * is only handled properly in the slow path.
5288 * - Out of order segments arrived.
5289 * - Urgent data is expected.
5290 * - There is no buffer space left
5291 * - Unexpected TCP flags/window values/header lengths are received
5292 * (detected by checking the TCP header against pred_flags)
5293 * - Data is sent in both directions. Fast path only supports pure senders
5294 * or pure receivers (this means either the sequence number or the ack
5295 * value must stay constant)
5296 * - Unexpected TCP option.
5298 * When these conditions are not satisfied it drops into a standard
5299 * receive procedure patterned after RFC793 to handle all cases.
5300 * The first three cases are guaranteed by proper pred_flags setting,
5301 * the rest is checked inline. Fast processing is turned on in
5302 * tcp_data_queue when everything is OK.
5304 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5305 const struct tcphdr *th, unsigned int len)
5307 struct tcp_sock *tp = tcp_sk(sk);
5311 * Header prediction.
5312 * The code loosely follows the one in the famous
5313 * "30 instruction TCP receive" Van Jacobson mail.
5315 * Van's trick is to deposit buffers into socket queue
5316 * on a device interrupt, to call tcp_recv function
5317 * on the receive process context and checksum and copy
5318 * the buffer to user space. smart...
5320 * Our current scheme is not silly either but we take the
5321 * extra cost of the net_bh soft interrupt processing...
5322 * We do checksum and copy also but from device to kernel.
5325 tp->rx_opt.saw_tstamp = 0;
5327 /* pred_flags is 0xS?10 << 16 + snd_wnd
5328 * if header_prediction is to be made
5329 * 'S' will always be tp->tcp_header_len >> 2
5330 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5331 * turn it off (when there are holes in the receive
5332 * space for instance)
5333 * PSH flag is ignored.
5336 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5337 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5338 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5339 int tcp_header_len = tp->tcp_header_len;
5341 /* Timestamp header prediction: tcp_header_len
5342 * is automatically equal to th->doff*4 due to pred_flags
5346 /* Check timestamp */
5347 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5348 /* No? Slow path! */
5349 if (!tcp_parse_aligned_timestamp(tp, th))
5352 /* If PAWS failed, check it more carefully in slow path */
5353 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5356 /* DO NOT update ts_recent here, if checksum fails
5357 * and timestamp was corrupted part, it will result
5358 * in a hung connection since we will drop all
5359 * future packets due to the PAWS test.
5363 if (len <= tcp_header_len) {
5364 /* Bulk data transfer: sender */
5365 if (len == tcp_header_len) {
5366 /* Predicted packet is in window by definition.
5367 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5368 * Hence, check seq<=rcv_wup reduces to:
5370 if (tcp_header_len ==
5371 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5372 tp->rcv_nxt == tp->rcv_wup)
5373 tcp_store_ts_recent(tp);
5375 /* We know that such packets are checksummed
5378 tcp_ack(sk, skb, 0);
5380 tcp_data_snd_check(sk);
5382 } else { /* Header too small */
5383 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5388 int copied_early = 0;
5390 if (tp->copied_seq == tp->rcv_nxt &&
5391 len - tcp_header_len <= tp->ucopy.len) {
5392 #ifdef CONFIG_NET_DMA
5393 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5398 if (tp->ucopy.task == current &&
5399 sock_owned_by_user(sk) && !copied_early) {
5400 __set_current_state(TASK_RUNNING);
5402 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5406 /* Predicted packet is in window by definition.
5407 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5408 * Hence, check seq<=rcv_wup reduces to:
5410 if (tcp_header_len ==
5411 (sizeof(struct tcphdr) +
5412 TCPOLEN_TSTAMP_ALIGNED) &&
5413 tp->rcv_nxt == tp->rcv_wup)
5414 tcp_store_ts_recent(tp);
5416 tcp_rcv_rtt_measure_ts(sk, skb);
5418 __skb_pull(skb, tcp_header_len);
5419 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5420 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5423 tcp_cleanup_rbuf(sk, skb->len);
5426 if (tcp_checksum_complete_user(sk, skb))
5429 /* Predicted packet is in window by definition.
5430 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5431 * Hence, check seq<=rcv_wup reduces to:
5433 if (tcp_header_len ==
5434 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5435 tp->rcv_nxt == tp->rcv_wup)
5436 tcp_store_ts_recent(tp);
5438 tcp_rcv_rtt_measure_ts(sk, skb);
5440 if ((int)skb->truesize > sk->sk_forward_alloc)
5443 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5445 /* Bulk data transfer: receiver */
5446 __skb_pull(skb, tcp_header_len);
5447 __skb_queue_tail(&sk->sk_receive_queue, skb);
5448 skb_set_owner_r(skb, sk);
5449 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5452 tcp_event_data_recv(sk, skb);
5454 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5455 /* Well, only one small jumplet in fast path... */
5456 tcp_ack(sk, skb, FLAG_DATA);
5457 tcp_data_snd_check(sk);
5458 if (!inet_csk_ack_scheduled(sk))
5462 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5463 __tcp_ack_snd_check(sk, 0);
5465 #ifdef CONFIG_NET_DMA
5467 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5473 sk->sk_data_ready(sk, 0);
5479 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5483 * Standard slow path.
5486 res = tcp_validate_incoming(sk, skb, th, 1);
5491 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5494 tcp_rcv_rtt_measure_ts(sk, skb);
5496 /* Process urgent data. */
5497 tcp_urg(sk, skb, th);
5499 /* step 7: process the segment text */
5500 tcp_data_queue(sk, skb);
5502 tcp_data_snd_check(sk);
5503 tcp_ack_snd_check(sk);
5507 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5513 EXPORT_SYMBOL(tcp_rcv_established);
5515 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5516 const struct tcphdr *th, unsigned int len)
5518 const u8 *hash_location;
5519 struct inet_connection_sock *icsk = inet_csk(sk);
5520 struct tcp_sock *tp = tcp_sk(sk);
5521 struct tcp_cookie_values *cvp = tp->cookie_values;
5522 int saved_clamp = tp->rx_opt.mss_clamp;
5524 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5528 * "If the state is SYN-SENT then
5529 * first check the ACK bit
5530 * If the ACK bit is set
5531 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5532 * a reset (unless the RST bit is set, if so drop
5533 * the segment and return)"
5535 * We do not send data with SYN, so that RFC-correct
5538 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5539 goto reset_and_undo;
5541 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5542 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5544 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5545 goto reset_and_undo;
5548 /* Now ACK is acceptable.
5550 * "If the RST bit is set
5551 * If the ACK was acceptable then signal the user "error:
5552 * connection reset", drop the segment, enter CLOSED state,
5553 * delete TCB, and return."
5562 * "fifth, if neither of the SYN or RST bits is set then
5563 * drop the segment and return."
5569 goto discard_and_undo;
5572 * "If the SYN bit is on ...
5573 * are acceptable then ...
5574 * (our SYN has been ACKed), change the connection
5575 * state to ESTABLISHED..."
5578 TCP_ECN_rcv_synack(tp, th);
5580 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5581 tcp_ack(sk, skb, FLAG_SLOWPATH);
5583 /* Ok.. it's good. Set up sequence numbers and
5584 * move to established.
5586 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5587 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5589 /* RFC1323: The window in SYN & SYN/ACK segments is
5592 tp->snd_wnd = ntohs(th->window);
5593 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5595 if (!tp->rx_opt.wscale_ok) {
5596 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5597 tp->window_clamp = min(tp->window_clamp, 65535U);
5600 if (tp->rx_opt.saw_tstamp) {
5601 tp->rx_opt.tstamp_ok = 1;
5602 tp->tcp_header_len =
5603 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5604 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5605 tcp_store_ts_recent(tp);
5607 tp->tcp_header_len = sizeof(struct tcphdr);
5610 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5611 tcp_enable_fack(tp);
5614 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5615 tcp_initialize_rcv_mss(sk);
5617 /* Remember, tcp_poll() does not lock socket!
5618 * Change state from SYN-SENT only after copied_seq
5619 * is initialized. */
5620 tp->copied_seq = tp->rcv_nxt;
5623 cvp->cookie_pair_size > 0 &&
5624 tp->rx_opt.cookie_plus > 0) {
5625 int cookie_size = tp->rx_opt.cookie_plus
5626 - TCPOLEN_COOKIE_BASE;
5627 int cookie_pair_size = cookie_size
5628 + cvp->cookie_desired;
5630 /* A cookie extension option was sent and returned.
5631 * Note that each incoming SYNACK replaces the
5632 * Responder cookie. The initial exchange is most
5633 * fragile, as protection against spoofing relies
5634 * entirely upon the sequence and timestamp (above).
5635 * This replacement strategy allows the correct pair to
5636 * pass through, while any others will be filtered via
5637 * Responder verification later.
5639 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5640 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5641 hash_location, cookie_size);
5642 cvp->cookie_pair_size = cookie_pair_size;
5647 tcp_set_state(sk, TCP_ESTABLISHED);
5649 security_inet_conn_established(sk, skb);
5651 /* Make sure socket is routed, for correct metrics. */
5652 icsk->icsk_af_ops->rebuild_header(sk);
5654 tcp_init_metrics(sk);
5656 tcp_init_congestion_control(sk);
5658 /* Prevent spurious tcp_cwnd_restart() on first data
5661 tp->lsndtime = tcp_time_stamp;
5663 tcp_init_buffer_space(sk);
5665 if (sock_flag(sk, SOCK_KEEPOPEN))
5666 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5668 if (!tp->rx_opt.snd_wscale)
5669 __tcp_fast_path_on(tp, tp->snd_wnd);
5673 if (!sock_flag(sk, SOCK_DEAD)) {
5674 sk->sk_state_change(sk);
5675 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5678 if (sk->sk_write_pending ||
5679 icsk->icsk_accept_queue.rskq_defer_accept ||
5680 icsk->icsk_ack.pingpong) {
5681 /* Save one ACK. Data will be ready after
5682 * several ticks, if write_pending is set.
5684 * It may be deleted, but with this feature tcpdumps
5685 * look so _wonderfully_ clever, that I was not able
5686 * to stand against the temptation 8) --ANK
5688 inet_csk_schedule_ack(sk);
5689 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5690 icsk->icsk_ack.ato = TCP_ATO_MIN;
5691 tcp_incr_quickack(sk);
5692 tcp_enter_quickack_mode(sk);
5693 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5694 TCP_DELACK_MAX, TCP_RTO_MAX);
5705 /* No ACK in the segment */
5709 * "If the RST bit is set
5711 * Otherwise (no ACK) drop the segment and return."
5714 goto discard_and_undo;
5718 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5719 tcp_paws_reject(&tp->rx_opt, 0))
5720 goto discard_and_undo;
5723 /* We see SYN without ACK. It is attempt of
5724 * simultaneous connect with crossed SYNs.
5725 * Particularly, it can be connect to self.
5727 tcp_set_state(sk, TCP_SYN_RECV);
5729 if (tp->rx_opt.saw_tstamp) {
5730 tp->rx_opt.tstamp_ok = 1;
5731 tcp_store_ts_recent(tp);
5732 tp->tcp_header_len =
5733 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5735 tp->tcp_header_len = sizeof(struct tcphdr);
5738 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5739 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5741 /* RFC1323: The window in SYN & SYN/ACK segments is
5744 tp->snd_wnd = ntohs(th->window);
5745 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5746 tp->max_window = tp->snd_wnd;
5748 TCP_ECN_rcv_syn(tp, th);
5751 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5752 tcp_initialize_rcv_mss(sk);
5754 tcp_send_synack(sk);
5756 /* Note, we could accept data and URG from this segment.
5757 * There are no obstacles to make this.
5759 * However, if we ignore data in ACKless segments sometimes,
5760 * we have no reasons to accept it sometimes.
5761 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5762 * is not flawless. So, discard packet for sanity.
5763 * Uncomment this return to process the data.
5770 /* "fifth, if neither of the SYN or RST bits is set then
5771 * drop the segment and return."
5775 tcp_clear_options(&tp->rx_opt);
5776 tp->rx_opt.mss_clamp = saved_clamp;
5780 tcp_clear_options(&tp->rx_opt);
5781 tp->rx_opt.mss_clamp = saved_clamp;
5786 * This function implements the receiving procedure of RFC 793 for
5787 * all states except ESTABLISHED and TIME_WAIT.
5788 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5789 * address independent.
5792 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5793 const struct tcphdr *th, unsigned int len)
5795 struct tcp_sock *tp = tcp_sk(sk);
5796 struct inet_connection_sock *icsk = inet_csk(sk);
5800 tp->rx_opt.saw_tstamp = 0;
5802 switch (sk->sk_state) {
5816 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5819 /* Now we have several options: In theory there is
5820 * nothing else in the frame. KA9Q has an option to
5821 * send data with the syn, BSD accepts data with the
5822 * syn up to the [to be] advertised window and
5823 * Solaris 2.1 gives you a protocol error. For now
5824 * we just ignore it, that fits the spec precisely
5825 * and avoids incompatibilities. It would be nice in
5826 * future to drop through and process the data.
5828 * Now that TTCP is starting to be used we ought to
5830 * But, this leaves one open to an easy denial of
5831 * service attack, and SYN cookies can't defend
5832 * against this problem. So, we drop the data
5833 * in the interest of security over speed unless
5834 * it's still in use.
5842 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5846 /* Do step6 onward by hand. */
5847 tcp_urg(sk, skb, th);
5849 tcp_data_snd_check(sk);
5853 res = tcp_validate_incoming(sk, skb, th, 0);
5857 /* step 5: check the ACK field */
5859 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5861 switch (sk->sk_state) {
5864 tp->copied_seq = tp->rcv_nxt;
5866 tcp_set_state(sk, TCP_ESTABLISHED);
5867 sk->sk_state_change(sk);
5869 /* Note, that this wakeup is only for marginal
5870 * crossed SYN case. Passively open sockets
5871 * are not waked up, because sk->sk_sleep ==
5872 * NULL and sk->sk_socket == NULL.
5876 SOCK_WAKE_IO, POLL_OUT);
5878 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5879 tp->snd_wnd = ntohs(th->window) <<
5880 tp->rx_opt.snd_wscale;
5881 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5883 if (tp->rx_opt.tstamp_ok)
5884 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5886 /* Make sure socket is routed, for
5889 icsk->icsk_af_ops->rebuild_header(sk);
5891 tcp_init_metrics(sk);
5893 tcp_init_congestion_control(sk);
5895 /* Prevent spurious tcp_cwnd_restart() on
5896 * first data packet.
5898 tp->lsndtime = tcp_time_stamp;
5901 tcp_initialize_rcv_mss(sk);
5902 tcp_init_buffer_space(sk);
5903 tcp_fast_path_on(tp);
5910 if (tp->snd_una == tp->write_seq) {
5911 tcp_set_state(sk, TCP_FIN_WAIT2);
5912 sk->sk_shutdown |= SEND_SHUTDOWN;
5913 dst_confirm(__sk_dst_get(sk));
5915 if (!sock_flag(sk, SOCK_DEAD))
5916 /* Wake up lingering close() */
5917 sk->sk_state_change(sk);
5921 if (tp->linger2 < 0 ||
5922 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5923 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5925 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5929 tmo = tcp_fin_time(sk);
5930 if (tmo > TCP_TIMEWAIT_LEN) {
5931 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5932 } else if (th->fin || sock_owned_by_user(sk)) {
5933 /* Bad case. We could lose such FIN otherwise.
5934 * It is not a big problem, but it looks confusing
5935 * and not so rare event. We still can lose it now,
5936 * if it spins in bh_lock_sock(), but it is really
5939 inet_csk_reset_keepalive_timer(sk, tmo);
5941 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5949 if (tp->snd_una == tp->write_seq) {
5950 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5956 if (tp->snd_una == tp->write_seq) {
5957 tcp_update_metrics(sk);
5966 /* step 6: check the URG bit */
5967 tcp_urg(sk, skb, th);
5969 /* step 7: process the segment text */
5970 switch (sk->sk_state) {
5971 case TCP_CLOSE_WAIT:
5974 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5978 /* RFC 793 says to queue data in these states,
5979 * RFC 1122 says we MUST send a reset.
5980 * BSD 4.4 also does reset.
5982 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5983 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5984 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5985 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5991 case TCP_ESTABLISHED:
5992 tcp_data_queue(sk, skb);
5997 /* tcp_data could move socket to TIME-WAIT */
5998 if (sk->sk_state != TCP_CLOSE) {
5999 tcp_data_snd_check(sk);
6000 tcp_ack_snd_check(sk);
6009 EXPORT_SYMBOL(tcp_rcv_state_process);