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);
482 /* No previous measure. */
486 if (tp->rcv_rtt_est.rtt != new_sample)
487 tp->rcv_rtt_est.rtt = new_sample;
490 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
492 if (tp->rcv_rtt_est.time == 0)
494 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
496 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
499 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
500 tp->rcv_rtt_est.time = tcp_time_stamp;
503 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
504 const struct sk_buff *skb)
506 struct tcp_sock *tp = tcp_sk(sk);
507 if (tp->rx_opt.rcv_tsecr &&
508 (TCP_SKB_CB(skb)->end_seq -
509 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
510 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
514 * This function should be called every time data is copied to user space.
515 * It calculates the appropriate TCP receive buffer space.
517 void tcp_rcv_space_adjust(struct sock *sk)
519 struct tcp_sock *tp = tcp_sk(sk);
523 if (tp->rcvq_space.time == 0)
526 time = tcp_time_stamp - tp->rcvq_space.time;
527 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
530 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
532 space = max(tp->rcvq_space.space, space);
534 if (tp->rcvq_space.space != space) {
537 tp->rcvq_space.space = space;
539 if (sysctl_tcp_moderate_rcvbuf &&
540 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
541 int new_clamp = space;
543 /* Receive space grows, normalize in order to
544 * take into account packet headers and sk_buff
545 * structure overhead.
550 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
551 while (tcp_win_from_space(rcvmem) < tp->advmss)
554 space = min(space, sysctl_tcp_rmem[2]);
555 if (space > sk->sk_rcvbuf) {
556 sk->sk_rcvbuf = space;
558 /* Make the window clamp follow along. */
559 tp->window_clamp = new_clamp;
565 tp->rcvq_space.seq = tp->copied_seq;
566 tp->rcvq_space.time = tcp_time_stamp;
569 /* There is something which you must keep in mind when you analyze the
570 * behavior of the tp->ato delayed ack timeout interval. When a
571 * connection starts up, we want to ack as quickly as possible. The
572 * problem is that "good" TCP's do slow start at the beginning of data
573 * transmission. The means that until we send the first few ACK's the
574 * sender will sit on his end and only queue most of his data, because
575 * he can only send snd_cwnd unacked packets at any given time. For
576 * each ACK we send, he increments snd_cwnd and transmits more of his
579 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
581 struct tcp_sock *tp = tcp_sk(sk);
582 struct inet_connection_sock *icsk = inet_csk(sk);
585 inet_csk_schedule_ack(sk);
587 tcp_measure_rcv_mss(sk, skb);
589 tcp_rcv_rtt_measure(tp);
591 now = tcp_time_stamp;
593 if (!icsk->icsk_ack.ato) {
594 /* The _first_ data packet received, initialize
595 * delayed ACK engine.
597 tcp_incr_quickack(sk);
598 icsk->icsk_ack.ato = TCP_ATO_MIN;
600 int m = now - icsk->icsk_ack.lrcvtime;
602 if (m <= TCP_ATO_MIN / 2) {
603 /* The fastest case is the first. */
604 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
605 } else if (m < icsk->icsk_ack.ato) {
606 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
607 if (icsk->icsk_ack.ato > icsk->icsk_rto)
608 icsk->icsk_ack.ato = icsk->icsk_rto;
609 } else if (m > icsk->icsk_rto) {
610 /* Too long gap. Apparently sender failed to
611 * restart window, so that we send ACKs quickly.
613 tcp_incr_quickack(sk);
617 icsk->icsk_ack.lrcvtime = now;
619 TCP_ECN_check_ce(tp, skb);
622 tcp_grow_window(sk, skb);
625 /* Called to compute a smoothed rtt estimate. The data fed to this
626 * routine either comes from timestamps, or from segments that were
627 * known _not_ to have been retransmitted [see Karn/Partridge
628 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
629 * piece by Van Jacobson.
630 * NOTE: the next three routines used to be one big routine.
631 * To save cycles in the RFC 1323 implementation it was better to break
632 * it up into three procedures. -- erics
634 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
636 struct tcp_sock *tp = tcp_sk(sk);
637 long m = mrtt; /* RTT */
639 /* The following amusing code comes from Jacobson's
640 * article in SIGCOMM '88. Note that rtt and mdev
641 * are scaled versions of rtt and mean deviation.
642 * This is designed to be as fast as possible
643 * m stands for "measurement".
645 * On a 1990 paper the rto value is changed to:
646 * RTO = rtt + 4 * mdev
648 * Funny. This algorithm seems to be very broken.
649 * These formulae increase RTO, when it should be decreased, increase
650 * too slowly, when it should be increased quickly, decrease too quickly
651 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
652 * does not matter how to _calculate_ it. Seems, it was trap
653 * that VJ failed to avoid. 8)
658 m -= (tp->srtt >> 3); /* m is now error in rtt est */
659 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
661 m = -m; /* m is now abs(error) */
662 m -= (tp->mdev >> 2); /* similar update on mdev */
663 /* This is similar to one of Eifel findings.
664 * Eifel blocks mdev updates when rtt decreases.
665 * This solution is a bit different: we use finer gain
666 * for mdev in this case (alpha*beta).
667 * Like Eifel it also prevents growth of rto,
668 * but also it limits too fast rto decreases,
669 * happening in pure Eifel.
674 m -= (tp->mdev >> 2); /* similar update on mdev */
676 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
677 if (tp->mdev > tp->mdev_max) {
678 tp->mdev_max = tp->mdev;
679 if (tp->mdev_max > tp->rttvar)
680 tp->rttvar = tp->mdev_max;
682 if (after(tp->snd_una, tp->rtt_seq)) {
683 if (tp->mdev_max < tp->rttvar)
684 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
685 tp->rtt_seq = tp->snd_nxt;
686 tp->mdev_max = tcp_rto_min(sk);
689 /* no previous measure. */
690 tp->srtt = m << 3; /* take the measured time to be rtt */
691 tp->mdev = m << 1; /* make sure rto = 3*rtt */
692 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
693 tp->rtt_seq = tp->snd_nxt;
697 /* Calculate rto without backoff. This is the second half of Van Jacobson's
698 * routine referred to above.
700 static inline void tcp_set_rto(struct sock *sk)
702 const struct tcp_sock *tp = tcp_sk(sk);
703 /* Old crap is replaced with new one. 8)
706 * 1. If rtt variance happened to be less 50msec, it is hallucination.
707 * It cannot be less due to utterly erratic ACK generation made
708 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
709 * to do with delayed acks, because at cwnd>2 true delack timeout
710 * is invisible. Actually, Linux-2.4 also generates erratic
711 * ACKs in some circumstances.
713 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
715 /* 2. Fixups made earlier cannot be right.
716 * If we do not estimate RTO correctly without them,
717 * all the algo is pure shit and should be replaced
718 * with correct one. It is exactly, which we pretend to do.
721 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
722 * guarantees that rto is higher.
727 /* Save metrics learned by this TCP session.
728 This function is called only, when TCP finishes successfully
729 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
731 void tcp_update_metrics(struct sock *sk)
733 struct tcp_sock *tp = tcp_sk(sk);
734 struct dst_entry *dst = __sk_dst_get(sk);
736 if (sysctl_tcp_nometrics_save)
741 if (dst && (dst->flags & DST_HOST)) {
742 const struct inet_connection_sock *icsk = inet_csk(sk);
746 if (icsk->icsk_backoff || !tp->srtt) {
747 /* This session failed to estimate rtt. Why?
748 * Probably, no packets returned in time.
751 if (!(dst_metric_locked(dst, RTAX_RTT)))
752 dst_metric_set(dst, RTAX_RTT, 0);
756 rtt = dst_metric_rtt(dst, RTAX_RTT);
759 /* If newly calculated rtt larger than stored one,
760 * store new one. Otherwise, use EWMA. Remember,
761 * rtt overestimation is always better than underestimation.
763 if (!(dst_metric_locked(dst, RTAX_RTT))) {
765 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
767 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
770 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
775 /* Scale deviation to rttvar fixed point */
780 var = dst_metric_rtt(dst, RTAX_RTTVAR);
784 var -= (var - m) >> 2;
786 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
789 if (tcp_in_initial_slowstart(tp)) {
790 /* Slow start still did not finish. */
791 if (dst_metric(dst, RTAX_SSTHRESH) &&
792 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
793 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
794 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
795 if (!dst_metric_locked(dst, RTAX_CWND) &&
796 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
797 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
798 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
799 icsk->icsk_ca_state == TCP_CA_Open) {
800 /* Cong. avoidance phase, cwnd is reliable. */
801 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
802 dst_metric_set(dst, RTAX_SSTHRESH,
803 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
804 if (!dst_metric_locked(dst, RTAX_CWND))
805 dst_metric_set(dst, RTAX_CWND,
806 (dst_metric(dst, RTAX_CWND) +
809 /* Else slow start did not finish, cwnd is non-sense,
810 ssthresh may be also invalid.
812 if (!dst_metric_locked(dst, RTAX_CWND))
813 dst_metric_set(dst, RTAX_CWND,
814 (dst_metric(dst, RTAX_CWND) +
815 tp->snd_ssthresh) >> 1);
816 if (dst_metric(dst, RTAX_SSTHRESH) &&
817 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
818 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
819 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
822 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
823 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
824 tp->reordering != sysctl_tcp_reordering)
825 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
830 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
832 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
835 cwnd = TCP_INIT_CWND;
836 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
839 /* Set slow start threshold and cwnd not falling to slow start */
840 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
842 struct tcp_sock *tp = tcp_sk(sk);
843 const struct inet_connection_sock *icsk = inet_csk(sk);
845 tp->prior_ssthresh = 0;
847 if (icsk->icsk_ca_state < TCP_CA_CWR) {
850 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
851 tp->snd_cwnd = min(tp->snd_cwnd,
852 tcp_packets_in_flight(tp) + 1U);
853 tp->snd_cwnd_cnt = 0;
854 tp->high_seq = tp->snd_nxt;
855 tp->snd_cwnd_stamp = tcp_time_stamp;
856 TCP_ECN_queue_cwr(tp);
858 tcp_set_ca_state(sk, TCP_CA_CWR);
863 * Packet counting of FACK is based on in-order assumptions, therefore TCP
864 * disables it when reordering is detected
866 static void tcp_disable_fack(struct tcp_sock *tp)
868 /* RFC3517 uses different metric in lost marker => reset on change */
870 tp->lost_skb_hint = NULL;
871 tp->rx_opt.sack_ok &= ~2;
874 /* Take a notice that peer is sending D-SACKs */
875 static void tcp_dsack_seen(struct tcp_sock *tp)
877 tp->rx_opt.sack_ok |= 4;
880 /* Initialize metrics on socket. */
882 static void tcp_init_metrics(struct sock *sk)
884 struct tcp_sock *tp = tcp_sk(sk);
885 struct dst_entry *dst = __sk_dst_get(sk);
892 if (dst_metric_locked(dst, RTAX_CWND))
893 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
894 if (dst_metric(dst, RTAX_SSTHRESH)) {
895 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
896 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
897 tp->snd_ssthresh = tp->snd_cwnd_clamp;
899 /* ssthresh may have been reduced unnecessarily during.
900 * 3WHS. Restore it back to its initial default.
902 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
904 if (dst_metric(dst, RTAX_REORDERING) &&
905 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
906 tcp_disable_fack(tp);
907 tp->reordering = dst_metric(dst, RTAX_REORDERING);
910 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
913 /* Initial rtt is determined from SYN,SYN-ACK.
914 * The segment is small and rtt may appear much
915 * less than real one. Use per-dst memory
916 * to make it more realistic.
918 * A bit of theory. RTT is time passed after "normal" sized packet
919 * is sent until it is ACKed. In normal circumstances sending small
920 * packets force peer to delay ACKs and calculation is correct too.
921 * The algorithm is adaptive and, provided we follow specs, it
922 * NEVER underestimate RTT. BUT! If peer tries to make some clever
923 * tricks sort of "quick acks" for time long enough to decrease RTT
924 * to low value, and then abruptly stops to do it and starts to delay
925 * ACKs, wait for troubles.
927 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
928 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
929 tp->rtt_seq = tp->snd_nxt;
931 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
932 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
933 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
938 /* RFC2988bis: We've failed to get a valid RTT sample from
939 * 3WHS. This is most likely due to retransmission,
940 * including spurious one. Reset the RTO back to 3secs
941 * from the more aggressive 1sec to avoid more spurious
944 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
945 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
947 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
948 * retransmitted. In light of RFC2988bis' more aggressive 1sec
949 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
950 * retransmission has occurred.
952 if (tp->total_retrans > 1)
955 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
956 tp->snd_cwnd_stamp = tcp_time_stamp;
959 static void tcp_update_reordering(struct sock *sk, const int metric,
962 struct tcp_sock *tp = tcp_sk(sk);
963 if (metric > tp->reordering) {
966 tp->reordering = min(TCP_MAX_REORDERING, metric);
968 /* This exciting event is worth to be remembered. 8) */
970 mib_idx = LINUX_MIB_TCPTSREORDER;
971 else if (tcp_is_reno(tp))
972 mib_idx = LINUX_MIB_TCPRENOREORDER;
973 else if (tcp_is_fack(tp))
974 mib_idx = LINUX_MIB_TCPFACKREORDER;
976 mib_idx = LINUX_MIB_TCPSACKREORDER;
978 NET_INC_STATS_BH(sock_net(sk), mib_idx);
979 #if FASTRETRANS_DEBUG > 1
980 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
981 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
985 tp->undo_marker ? tp->undo_retrans : 0);
987 tcp_disable_fack(tp);
991 /* This must be called before lost_out is incremented */
992 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
994 if ((tp->retransmit_skb_hint == NULL) ||
995 before(TCP_SKB_CB(skb)->seq,
996 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
997 tp->retransmit_skb_hint = skb;
1000 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
1001 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1004 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1006 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1007 tcp_verify_retransmit_hint(tp, skb);
1009 tp->lost_out += tcp_skb_pcount(skb);
1010 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1014 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1015 struct sk_buff *skb)
1017 tcp_verify_retransmit_hint(tp, skb);
1019 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1020 tp->lost_out += tcp_skb_pcount(skb);
1021 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1025 /* This procedure tags the retransmission queue when SACKs arrive.
1027 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1028 * Packets in queue with these bits set are counted in variables
1029 * sacked_out, retrans_out and lost_out, correspondingly.
1031 * Valid combinations are:
1032 * Tag InFlight Description
1033 * 0 1 - orig segment is in flight.
1034 * S 0 - nothing flies, orig reached receiver.
1035 * L 0 - nothing flies, orig lost by net.
1036 * R 2 - both orig and retransmit are in flight.
1037 * L|R 1 - orig is lost, retransmit is in flight.
1038 * S|R 1 - orig reached receiver, retrans is still in flight.
1039 * (L|S|R is logically valid, it could occur when L|R is sacked,
1040 * but it is equivalent to plain S and code short-curcuits it to S.
1041 * L|S is logically invalid, it would mean -1 packet in flight 8))
1043 * These 6 states form finite state machine, controlled by the following events:
1044 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1045 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1046 * 3. Loss detection event of one of three flavors:
1047 * A. Scoreboard estimator decided the packet is lost.
1048 * A'. Reno "three dupacks" marks head of queue lost.
1049 * A''. Its FACK modfication, head until snd.fack is lost.
1050 * B. SACK arrives sacking data transmitted after never retransmitted
1051 * hole was sent out.
1052 * C. SACK arrives sacking SND.NXT at the moment, when the
1053 * segment was retransmitted.
1054 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1056 * It is pleasant to note, that state diagram turns out to be commutative,
1057 * so that we are allowed not to be bothered by order of our actions,
1058 * when multiple events arrive simultaneously. (see the function below).
1060 * Reordering detection.
1061 * --------------------
1062 * Reordering metric is maximal distance, which a packet can be displaced
1063 * in packet stream. With SACKs we can estimate it:
1065 * 1. SACK fills old hole and the corresponding segment was not
1066 * ever retransmitted -> reordering. Alas, we cannot use it
1067 * when segment was retransmitted.
1068 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1069 * for retransmitted and already SACKed segment -> reordering..
1070 * Both of these heuristics are not used in Loss state, when we cannot
1071 * account for retransmits accurately.
1073 * SACK block validation.
1074 * ----------------------
1076 * SACK block range validation checks that the received SACK block fits to
1077 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1078 * Note that SND.UNA is not included to the range though being valid because
1079 * it means that the receiver is rather inconsistent with itself reporting
1080 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1081 * perfectly valid, however, in light of RFC2018 which explicitly states
1082 * that "SACK block MUST reflect the newest segment. Even if the newest
1083 * segment is going to be discarded ...", not that it looks very clever
1084 * in case of head skb. Due to potentional receiver driven attacks, we
1085 * choose to avoid immediate execution of a walk in write queue due to
1086 * reneging and defer head skb's loss recovery to standard loss recovery
1087 * procedure that will eventually trigger (nothing forbids us doing this).
1089 * Implements also blockage to start_seq wrap-around. Problem lies in the
1090 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1091 * there's no guarantee that it will be before snd_nxt (n). The problem
1092 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1095 * <- outs wnd -> <- wrapzone ->
1096 * u e n u_w e_w s n_w
1098 * |<------------+------+----- TCP seqno space --------------+---------->|
1099 * ...-- <2^31 ->| |<--------...
1100 * ...---- >2^31 ------>| |<--------...
1102 * Current code wouldn't be vulnerable but it's better still to discard such
1103 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1104 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1105 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1106 * equal to the ideal case (infinite seqno space without wrap caused issues).
1108 * With D-SACK the lower bound is extended to cover sequence space below
1109 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1110 * again, D-SACK block must not to go across snd_una (for the same reason as
1111 * for the normal SACK blocks, explained above). But there all simplicity
1112 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1113 * fully below undo_marker they do not affect behavior in anyway and can
1114 * therefore be safely ignored. In rare cases (which are more or less
1115 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1116 * fragmentation and packet reordering past skb's retransmission. To consider
1117 * them correctly, the acceptable range must be extended even more though
1118 * the exact amount is rather hard to quantify. However, tp->max_window can
1119 * be used as an exaggerated estimate.
1121 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1122 u32 start_seq, u32 end_seq)
1124 /* Too far in future, or reversed (interpretation is ambiguous) */
1125 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1128 /* Nasty start_seq wrap-around check (see comments above) */
1129 if (!before(start_seq, tp->snd_nxt))
1132 /* In outstanding window? ...This is valid exit for D-SACKs too.
1133 * start_seq == snd_una is non-sensical (see comments above)
1135 if (after(start_seq, tp->snd_una))
1138 if (!is_dsack || !tp->undo_marker)
1141 /* ...Then it's D-SACK, and must reside below snd_una completely */
1142 if (after(end_seq, tp->snd_una))
1145 if (!before(start_seq, tp->undo_marker))
1149 if (!after(end_seq, tp->undo_marker))
1152 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1153 * start_seq < undo_marker and end_seq >= undo_marker.
1155 return !before(start_seq, end_seq - tp->max_window);
1158 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1159 * Event "C". Later note: FACK people cheated me again 8), we have to account
1160 * for reordering! Ugly, but should help.
1162 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1163 * less than what is now known to be received by the other end (derived from
1164 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1165 * retransmitted skbs to avoid some costly processing per ACKs.
1167 static void tcp_mark_lost_retrans(struct sock *sk)
1169 const struct inet_connection_sock *icsk = inet_csk(sk);
1170 struct tcp_sock *tp = tcp_sk(sk);
1171 struct sk_buff *skb;
1173 u32 new_low_seq = tp->snd_nxt;
1174 u32 received_upto = tcp_highest_sack_seq(tp);
1176 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1177 !after(received_upto, tp->lost_retrans_low) ||
1178 icsk->icsk_ca_state != TCP_CA_Recovery)
1181 tcp_for_write_queue(skb, sk) {
1182 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1184 if (skb == tcp_send_head(sk))
1186 if (cnt == tp->retrans_out)
1188 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1191 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1194 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1195 * constraint here (see above) but figuring out that at
1196 * least tp->reordering SACK blocks reside between ack_seq
1197 * and received_upto is not easy task to do cheaply with
1198 * the available datastructures.
1200 * Whether FACK should check here for tp->reordering segs
1201 * in-between one could argue for either way (it would be
1202 * rather simple to implement as we could count fack_count
1203 * during the walk and do tp->fackets_out - fack_count).
1205 if (after(received_upto, ack_seq)) {
1206 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1207 tp->retrans_out -= tcp_skb_pcount(skb);
1209 tcp_skb_mark_lost_uncond_verify(tp, skb);
1210 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1212 if (before(ack_seq, new_low_seq))
1213 new_low_seq = ack_seq;
1214 cnt += tcp_skb_pcount(skb);
1218 if (tp->retrans_out)
1219 tp->lost_retrans_low = new_low_seq;
1222 static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1223 struct tcp_sack_block_wire *sp, int num_sacks,
1226 struct tcp_sock *tp = tcp_sk(sk);
1227 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1228 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1231 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1234 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1235 } else if (num_sacks > 1) {
1236 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1237 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1239 if (!after(end_seq_0, end_seq_1) &&
1240 !before(start_seq_0, start_seq_1)) {
1243 NET_INC_STATS_BH(sock_net(sk),
1244 LINUX_MIB_TCPDSACKOFORECV);
1248 /* D-SACK for already forgotten data... Do dumb counting. */
1249 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1250 !after(end_seq_0, prior_snd_una) &&
1251 after(end_seq_0, tp->undo_marker))
1257 struct tcp_sacktag_state {
1263 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1264 * the incoming SACK may not exactly match but we can find smaller MSS
1265 * aligned portion of it that matches. Therefore we might need to fragment
1266 * which may fail and creates some hassle (caller must handle error case
1269 * FIXME: this could be merged to shift decision code
1271 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1272 u32 start_seq, u32 end_seq)
1275 unsigned int pkt_len;
1278 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1279 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1281 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1282 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1283 mss = tcp_skb_mss(skb);
1284 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1287 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1291 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1296 /* Round if necessary so that SACKs cover only full MSSes
1297 * and/or the remaining small portion (if present)
1299 if (pkt_len > mss) {
1300 unsigned int new_len = (pkt_len / mss) * mss;
1301 if (!in_sack && new_len < pkt_len) {
1303 if (new_len > skb->len)
1308 err = tcp_fragment(sk, skb, pkt_len, mss);
1316 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1317 static u8 tcp_sacktag_one(struct sock *sk,
1318 struct tcp_sacktag_state *state, u8 sacked,
1319 u32 start_seq, u32 end_seq,
1320 int dup_sack, int pcount)
1322 struct tcp_sock *tp = tcp_sk(sk);
1323 int fack_count = state->fack_count;
1325 /* Account D-SACK for retransmitted packet. */
1326 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1327 if (tp->undo_marker && tp->undo_retrans &&
1328 after(end_seq, tp->undo_marker))
1330 if (sacked & TCPCB_SACKED_ACKED)
1331 state->reord = min(fack_count, state->reord);
1334 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1335 if (!after(end_seq, tp->snd_una))
1338 if (!(sacked & TCPCB_SACKED_ACKED)) {
1339 if (sacked & TCPCB_SACKED_RETRANS) {
1340 /* If the segment is not tagged as lost,
1341 * we do not clear RETRANS, believing
1342 * that retransmission is still in flight.
1344 if (sacked & TCPCB_LOST) {
1345 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1346 tp->lost_out -= pcount;
1347 tp->retrans_out -= pcount;
1350 if (!(sacked & TCPCB_RETRANS)) {
1351 /* New sack for not retransmitted frame,
1352 * which was in hole. It is reordering.
1354 if (before(start_seq,
1355 tcp_highest_sack_seq(tp)))
1356 state->reord = min(fack_count,
1359 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1360 if (!after(end_seq, tp->frto_highmark))
1361 state->flag |= FLAG_ONLY_ORIG_SACKED;
1364 if (sacked & TCPCB_LOST) {
1365 sacked &= ~TCPCB_LOST;
1366 tp->lost_out -= pcount;
1370 sacked |= TCPCB_SACKED_ACKED;
1371 state->flag |= FLAG_DATA_SACKED;
1372 tp->sacked_out += pcount;
1374 fack_count += pcount;
1376 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1377 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1378 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1379 tp->lost_cnt_hint += pcount;
1381 if (fack_count > tp->fackets_out)
1382 tp->fackets_out = fack_count;
1385 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1386 * frames and clear it. undo_retrans is decreased above, L|R frames
1387 * are accounted above as well.
1389 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1390 sacked &= ~TCPCB_SACKED_RETRANS;
1391 tp->retrans_out -= pcount;
1397 /* Shift newly-SACKed bytes from this skb to the immediately previous
1398 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1400 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1401 struct tcp_sacktag_state *state,
1402 unsigned int pcount, int shifted, int mss,
1405 struct tcp_sock *tp = tcp_sk(sk);
1406 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1407 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1408 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1412 /* Adjust counters and hints for the newly sacked sequence
1413 * range but discard the return value since prev is already
1414 * marked. We must tag the range first because the seq
1415 * advancement below implicitly advances
1416 * tcp_highest_sack_seq() when skb is highest_sack.
1418 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1419 start_seq, end_seq, dup_sack, pcount);
1421 if (skb == tp->lost_skb_hint)
1422 tp->lost_cnt_hint += pcount;
1424 TCP_SKB_CB(prev)->end_seq += shifted;
1425 TCP_SKB_CB(skb)->seq += shifted;
1427 skb_shinfo(prev)->gso_segs += pcount;
1428 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1429 skb_shinfo(skb)->gso_segs -= pcount;
1431 /* When we're adding to gso_segs == 1, gso_size will be zero,
1432 * in theory this shouldn't be necessary but as long as DSACK
1433 * code can come after this skb later on it's better to keep
1434 * setting gso_size to something.
1436 if (!skb_shinfo(prev)->gso_size) {
1437 skb_shinfo(prev)->gso_size = mss;
1438 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1441 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1442 if (skb_shinfo(skb)->gso_segs <= 1) {
1443 skb_shinfo(skb)->gso_size = 0;
1444 skb_shinfo(skb)->gso_type = 0;
1447 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1448 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1451 BUG_ON(!tcp_skb_pcount(skb));
1452 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1456 /* Whole SKB was eaten :-) */
1458 if (skb == tp->retransmit_skb_hint)
1459 tp->retransmit_skb_hint = prev;
1460 if (skb == tp->scoreboard_skb_hint)
1461 tp->scoreboard_skb_hint = prev;
1462 if (skb == tp->lost_skb_hint) {
1463 tp->lost_skb_hint = prev;
1464 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1467 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1468 if (skb == tcp_highest_sack(sk))
1469 tcp_advance_highest_sack(sk, skb);
1471 tcp_unlink_write_queue(skb, sk);
1472 sk_wmem_free_skb(sk, skb);
1474 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1479 /* I wish gso_size would have a bit more sane initialization than
1480 * something-or-zero which complicates things
1482 static int tcp_skb_seglen(const struct sk_buff *skb)
1484 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1487 /* Shifting pages past head area doesn't work */
1488 static int skb_can_shift(const struct sk_buff *skb)
1490 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1493 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1496 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1497 struct tcp_sacktag_state *state,
1498 u32 start_seq, u32 end_seq,
1501 struct tcp_sock *tp = tcp_sk(sk);
1502 struct sk_buff *prev;
1508 if (!sk_can_gso(sk))
1511 /* Normally R but no L won't result in plain S */
1513 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1515 if (!skb_can_shift(skb))
1517 /* This frame is about to be dropped (was ACKed). */
1518 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1521 /* Can only happen with delayed DSACK + discard craziness */
1522 if (unlikely(skb == tcp_write_queue_head(sk)))
1524 prev = tcp_write_queue_prev(sk, skb);
1526 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1529 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1530 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1534 pcount = tcp_skb_pcount(skb);
1535 mss = tcp_skb_seglen(skb);
1537 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1538 * drop this restriction as unnecessary
1540 if (mss != tcp_skb_seglen(prev))
1543 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1545 /* CHECKME: This is non-MSS split case only?, this will
1546 * cause skipped skbs due to advancing loop btw, original
1547 * has that feature too
1549 if (tcp_skb_pcount(skb) <= 1)
1552 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1554 /* TODO: head merge to next could be attempted here
1555 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1556 * though it might not be worth of the additional hassle
1558 * ...we can probably just fallback to what was done
1559 * previously. We could try merging non-SACKed ones
1560 * as well but it probably isn't going to buy off
1561 * because later SACKs might again split them, and
1562 * it would make skb timestamp tracking considerably
1568 len = end_seq - TCP_SKB_CB(skb)->seq;
1570 BUG_ON(len > skb->len);
1572 /* MSS boundaries should be honoured or else pcount will
1573 * severely break even though it makes things bit trickier.
1574 * Optimize common case to avoid most of the divides
1576 mss = tcp_skb_mss(skb);
1578 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1579 * drop this restriction as unnecessary
1581 if (mss != tcp_skb_seglen(prev))
1586 } else if (len < mss) {
1594 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1595 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1598 if (!skb_shift(prev, skb, len))
1600 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1603 /* Hole filled allows collapsing with the next as well, this is very
1604 * useful when hole on every nth skb pattern happens
1606 if (prev == tcp_write_queue_tail(sk))
1608 skb = tcp_write_queue_next(sk, prev);
1610 if (!skb_can_shift(skb) ||
1611 (skb == tcp_send_head(sk)) ||
1612 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1613 (mss != tcp_skb_seglen(skb)))
1617 if (skb_shift(prev, skb, len)) {
1618 pcount += tcp_skb_pcount(skb);
1619 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1623 state->fack_count += pcount;
1630 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1634 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1635 struct tcp_sack_block *next_dup,
1636 struct tcp_sacktag_state *state,
1637 u32 start_seq, u32 end_seq,
1640 struct tcp_sock *tp = tcp_sk(sk);
1641 struct sk_buff *tmp;
1643 tcp_for_write_queue_from(skb, sk) {
1645 int dup_sack = dup_sack_in;
1647 if (skb == tcp_send_head(sk))
1650 /* queue is in-order => we can short-circuit the walk early */
1651 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1654 if ((next_dup != NULL) &&
1655 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1656 in_sack = tcp_match_skb_to_sack(sk, skb,
1657 next_dup->start_seq,
1663 /* skb reference here is a bit tricky to get right, since
1664 * shifting can eat and free both this skb and the next,
1665 * so not even _safe variant of the loop is enough.
1668 tmp = tcp_shift_skb_data(sk, skb, state,
1669 start_seq, end_seq, dup_sack);
1678 in_sack = tcp_match_skb_to_sack(sk, skb,
1684 if (unlikely(in_sack < 0))
1688 TCP_SKB_CB(skb)->sacked =
1691 TCP_SKB_CB(skb)->sacked,
1692 TCP_SKB_CB(skb)->seq,
1693 TCP_SKB_CB(skb)->end_seq,
1695 tcp_skb_pcount(skb));
1697 if (!before(TCP_SKB_CB(skb)->seq,
1698 tcp_highest_sack_seq(tp)))
1699 tcp_advance_highest_sack(sk, skb);
1702 state->fack_count += tcp_skb_pcount(skb);
1707 /* Avoid all extra work that is being done by sacktag while walking in
1710 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1711 struct tcp_sacktag_state *state,
1714 tcp_for_write_queue_from(skb, sk) {
1715 if (skb == tcp_send_head(sk))
1718 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1721 state->fack_count += tcp_skb_pcount(skb);
1726 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1728 struct tcp_sack_block *next_dup,
1729 struct tcp_sacktag_state *state,
1732 if (next_dup == NULL)
1735 if (before(next_dup->start_seq, skip_to_seq)) {
1736 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1737 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1738 next_dup->start_seq, next_dup->end_seq,
1745 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1747 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1751 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1754 const struct inet_connection_sock *icsk = inet_csk(sk);
1755 struct tcp_sock *tp = tcp_sk(sk);
1756 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1757 TCP_SKB_CB(ack_skb)->sacked);
1758 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1759 struct tcp_sack_block sp[TCP_NUM_SACKS];
1760 struct tcp_sack_block *cache;
1761 struct tcp_sacktag_state state;
1762 struct sk_buff *skb;
1763 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1765 int found_dup_sack = 0;
1767 int first_sack_index;
1770 state.reord = tp->packets_out;
1772 if (!tp->sacked_out) {
1773 if (WARN_ON(tp->fackets_out))
1774 tp->fackets_out = 0;
1775 tcp_highest_sack_reset(sk);
1778 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1779 num_sacks, prior_snd_una);
1781 state.flag |= FLAG_DSACKING_ACK;
1783 /* Eliminate too old ACKs, but take into
1784 * account more or less fresh ones, they can
1785 * contain valid SACK info.
1787 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1790 if (!tp->packets_out)
1794 first_sack_index = 0;
1795 for (i = 0; i < num_sacks; i++) {
1796 int dup_sack = !i && found_dup_sack;
1798 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1799 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1801 if (!tcp_is_sackblock_valid(tp, dup_sack,
1802 sp[used_sacks].start_seq,
1803 sp[used_sacks].end_seq)) {
1807 if (!tp->undo_marker)
1808 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1810 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1812 /* Don't count olds caused by ACK reordering */
1813 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1814 !after(sp[used_sacks].end_seq, tp->snd_una))
1816 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1819 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1821 first_sack_index = -1;
1825 /* Ignore very old stuff early */
1826 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1832 /* order SACK blocks to allow in order walk of the retrans queue */
1833 for (i = used_sacks - 1; i > 0; i--) {
1834 for (j = 0; j < i; j++) {
1835 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1836 swap(sp[j], sp[j + 1]);
1838 /* Track where the first SACK block goes to */
1839 if (j == first_sack_index)
1840 first_sack_index = j + 1;
1845 skb = tcp_write_queue_head(sk);
1846 state.fack_count = 0;
1849 if (!tp->sacked_out) {
1850 /* It's already past, so skip checking against it */
1851 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1853 cache = tp->recv_sack_cache;
1854 /* Skip empty blocks in at head of the cache */
1855 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1860 while (i < used_sacks) {
1861 u32 start_seq = sp[i].start_seq;
1862 u32 end_seq = sp[i].end_seq;
1863 int dup_sack = (found_dup_sack && (i == first_sack_index));
1864 struct tcp_sack_block *next_dup = NULL;
1866 if (found_dup_sack && ((i + 1) == first_sack_index))
1867 next_dup = &sp[i + 1];
1869 /* Event "B" in the comment above. */
1870 if (after(end_seq, tp->high_seq))
1871 state.flag |= FLAG_DATA_LOST;
1873 /* Skip too early cached blocks */
1874 while (tcp_sack_cache_ok(tp, cache) &&
1875 !before(start_seq, cache->end_seq))
1878 /* Can skip some work by looking recv_sack_cache? */
1879 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1880 after(end_seq, cache->start_seq)) {
1883 if (before(start_seq, cache->start_seq)) {
1884 skb = tcp_sacktag_skip(skb, sk, &state,
1886 skb = tcp_sacktag_walk(skb, sk, next_dup,
1893 /* Rest of the block already fully processed? */
1894 if (!after(end_seq, cache->end_seq))
1897 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1901 /* ...tail remains todo... */
1902 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1903 /* ...but better entrypoint exists! */
1904 skb = tcp_highest_sack(sk);
1907 state.fack_count = tp->fackets_out;
1912 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1913 /* Check overlap against next cached too (past this one already) */
1918 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1919 skb = tcp_highest_sack(sk);
1922 state.fack_count = tp->fackets_out;
1924 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1927 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1928 start_seq, end_seq, dup_sack);
1931 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1932 * due to in-order walk
1934 if (after(end_seq, tp->frto_highmark))
1935 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1940 /* Clear the head of the cache sack blocks so we can skip it next time */
1941 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1942 tp->recv_sack_cache[i].start_seq = 0;
1943 tp->recv_sack_cache[i].end_seq = 0;
1945 for (j = 0; j < used_sacks; j++)
1946 tp->recv_sack_cache[i++] = sp[j];
1948 tcp_mark_lost_retrans(sk);
1950 tcp_verify_left_out(tp);
1952 if ((state.reord < tp->fackets_out) &&
1953 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1954 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1955 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1959 #if FASTRETRANS_DEBUG > 0
1960 WARN_ON((int)tp->sacked_out < 0);
1961 WARN_ON((int)tp->lost_out < 0);
1962 WARN_ON((int)tp->retrans_out < 0);
1963 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1968 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1969 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1971 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1975 holes = max(tp->lost_out, 1U);
1976 holes = min(holes, tp->packets_out);
1978 if ((tp->sacked_out + holes) > tp->packets_out) {
1979 tp->sacked_out = tp->packets_out - holes;
1985 /* If we receive more dupacks than we expected counting segments
1986 * in assumption of absent reordering, interpret this as reordering.
1987 * The only another reason could be bug in receiver TCP.
1989 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1991 struct tcp_sock *tp = tcp_sk(sk);
1992 if (tcp_limit_reno_sacked(tp))
1993 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1996 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1998 static void tcp_add_reno_sack(struct sock *sk)
2000 struct tcp_sock *tp = tcp_sk(sk);
2002 tcp_check_reno_reordering(sk, 0);
2003 tcp_verify_left_out(tp);
2006 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2008 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
2010 struct tcp_sock *tp = tcp_sk(sk);
2013 /* One ACK acked hole. The rest eat duplicate ACKs. */
2014 if (acked - 1 >= tp->sacked_out)
2017 tp->sacked_out -= acked - 1;
2019 tcp_check_reno_reordering(sk, acked);
2020 tcp_verify_left_out(tp);
2023 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2028 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2030 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2033 /* F-RTO can only be used if TCP has never retransmitted anything other than
2034 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2036 int tcp_use_frto(struct sock *sk)
2038 const struct tcp_sock *tp = tcp_sk(sk);
2039 const struct inet_connection_sock *icsk = inet_csk(sk);
2040 struct sk_buff *skb;
2042 if (!sysctl_tcp_frto)
2045 /* MTU probe and F-RTO won't really play nicely along currently */
2046 if (icsk->icsk_mtup.probe_size)
2049 if (tcp_is_sackfrto(tp))
2052 /* Avoid expensive walking of rexmit queue if possible */
2053 if (tp->retrans_out > 1)
2056 skb = tcp_write_queue_head(sk);
2057 if (tcp_skb_is_last(sk, skb))
2059 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2060 tcp_for_write_queue_from(skb, sk) {
2061 if (skb == tcp_send_head(sk))
2063 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2065 /* Short-circuit when first non-SACKed skb has been checked */
2066 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2072 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2073 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2074 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2075 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2076 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2077 * bits are handled if the Loss state is really to be entered (in
2078 * tcp_enter_frto_loss).
2080 * Do like tcp_enter_loss() would; when RTO expires the second time it
2082 * "Reduce ssthresh if it has not yet been made inside this window."
2084 void tcp_enter_frto(struct sock *sk)
2086 const struct inet_connection_sock *icsk = inet_csk(sk);
2087 struct tcp_sock *tp = tcp_sk(sk);
2088 struct sk_buff *skb;
2090 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2091 tp->snd_una == tp->high_seq ||
2092 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2093 !icsk->icsk_retransmits)) {
2094 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2095 /* Our state is too optimistic in ssthresh() call because cwnd
2096 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2097 * recovery has not yet completed. Pattern would be this: RTO,
2098 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2100 * RFC4138 should be more specific on what to do, even though
2101 * RTO is quite unlikely to occur after the first Cumulative ACK
2102 * due to back-off and complexity of triggering events ...
2104 if (tp->frto_counter) {
2106 stored_cwnd = tp->snd_cwnd;
2108 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2109 tp->snd_cwnd = stored_cwnd;
2111 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2113 /* ... in theory, cong.control module could do "any tricks" in
2114 * ssthresh(), which means that ca_state, lost bits and lost_out
2115 * counter would have to be faked before the call occurs. We
2116 * consider that too expensive, unlikely and hacky, so modules
2117 * using these in ssthresh() must deal these incompatibility
2118 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2120 tcp_ca_event(sk, CA_EVENT_FRTO);
2123 tp->undo_marker = tp->snd_una;
2124 tp->undo_retrans = 0;
2126 skb = tcp_write_queue_head(sk);
2127 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2128 tp->undo_marker = 0;
2129 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2130 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2131 tp->retrans_out -= tcp_skb_pcount(skb);
2133 tcp_verify_left_out(tp);
2135 /* Too bad if TCP was application limited */
2136 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2138 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2139 * The last condition is necessary at least in tp->frto_counter case.
2141 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2142 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2143 after(tp->high_seq, tp->snd_una)) {
2144 tp->frto_highmark = tp->high_seq;
2146 tp->frto_highmark = tp->snd_nxt;
2148 tcp_set_ca_state(sk, TCP_CA_Disorder);
2149 tp->high_seq = tp->snd_nxt;
2150 tp->frto_counter = 1;
2153 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2154 * which indicates that we should follow the traditional RTO recovery,
2155 * i.e. mark everything lost and do go-back-N retransmission.
2157 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2159 struct tcp_sock *tp = tcp_sk(sk);
2160 struct sk_buff *skb;
2163 tp->retrans_out = 0;
2164 if (tcp_is_reno(tp))
2165 tcp_reset_reno_sack(tp);
2167 tcp_for_write_queue(skb, sk) {
2168 if (skb == tcp_send_head(sk))
2171 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2173 * Count the retransmission made on RTO correctly (only when
2174 * waiting for the first ACK and did not get it)...
2176 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2177 /* For some reason this R-bit might get cleared? */
2178 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2179 tp->retrans_out += tcp_skb_pcount(skb);
2180 /* ...enter this if branch just for the first segment */
2181 flag |= FLAG_DATA_ACKED;
2183 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2184 tp->undo_marker = 0;
2185 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2188 /* Marking forward transmissions that were made after RTO lost
2189 * can cause unnecessary retransmissions in some scenarios,
2190 * SACK blocks will mitigate that in some but not in all cases.
2191 * We used to not mark them but it was causing break-ups with
2192 * receivers that do only in-order receival.
2194 * TODO: we could detect presence of such receiver and select
2195 * different behavior per flow.
2197 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2198 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2199 tp->lost_out += tcp_skb_pcount(skb);
2200 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2203 tcp_verify_left_out(tp);
2205 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2206 tp->snd_cwnd_cnt = 0;
2207 tp->snd_cwnd_stamp = tcp_time_stamp;
2208 tp->frto_counter = 0;
2209 tp->bytes_acked = 0;
2211 tp->reordering = min_t(unsigned int, tp->reordering,
2212 sysctl_tcp_reordering);
2213 tcp_set_ca_state(sk, TCP_CA_Loss);
2214 tp->high_seq = tp->snd_nxt;
2215 TCP_ECN_queue_cwr(tp);
2217 tcp_clear_all_retrans_hints(tp);
2220 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2222 tp->retrans_out = 0;
2225 tp->undo_marker = 0;
2226 tp->undo_retrans = 0;
2229 void tcp_clear_retrans(struct tcp_sock *tp)
2231 tcp_clear_retrans_partial(tp);
2233 tp->fackets_out = 0;
2237 /* Enter Loss state. If "how" is not zero, forget all SACK information
2238 * and reset tags completely, otherwise preserve SACKs. If receiver
2239 * dropped its ofo queue, we will know this due to reneging detection.
2241 void tcp_enter_loss(struct sock *sk, int how)
2243 const struct inet_connection_sock *icsk = inet_csk(sk);
2244 struct tcp_sock *tp = tcp_sk(sk);
2245 struct sk_buff *skb;
2247 /* Reduce ssthresh if it has not yet been made inside this window. */
2248 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2249 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2250 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2251 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2252 tcp_ca_event(sk, CA_EVENT_LOSS);
2255 tp->snd_cwnd_cnt = 0;
2256 tp->snd_cwnd_stamp = tcp_time_stamp;
2258 tp->bytes_acked = 0;
2259 tcp_clear_retrans_partial(tp);
2261 if (tcp_is_reno(tp))
2262 tcp_reset_reno_sack(tp);
2265 /* Push undo marker, if it was plain RTO and nothing
2266 * was retransmitted. */
2267 tp->undo_marker = tp->snd_una;
2270 tp->fackets_out = 0;
2272 tcp_clear_all_retrans_hints(tp);
2274 tcp_for_write_queue(skb, sk) {
2275 if (skb == tcp_send_head(sk))
2278 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2279 tp->undo_marker = 0;
2280 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2281 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2282 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2283 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2284 tp->lost_out += tcp_skb_pcount(skb);
2285 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2288 tcp_verify_left_out(tp);
2290 tp->reordering = min_t(unsigned int, tp->reordering,
2291 sysctl_tcp_reordering);
2292 tcp_set_ca_state(sk, TCP_CA_Loss);
2293 tp->high_seq = tp->snd_nxt;
2294 TCP_ECN_queue_cwr(tp);
2295 /* Abort F-RTO algorithm if one is in progress */
2296 tp->frto_counter = 0;
2299 /* If ACK arrived pointing to a remembered SACK, it means that our
2300 * remembered SACKs do not reflect real state of receiver i.e.
2301 * receiver _host_ is heavily congested (or buggy).
2303 * Do processing similar to RTO timeout.
2305 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2307 if (flag & FLAG_SACK_RENEGING) {
2308 struct inet_connection_sock *icsk = inet_csk(sk);
2309 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2311 tcp_enter_loss(sk, 1);
2312 icsk->icsk_retransmits++;
2313 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2314 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2315 icsk->icsk_rto, TCP_RTO_MAX);
2321 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2323 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2326 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2327 * counter when SACK is enabled (without SACK, sacked_out is used for
2330 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2331 * segments up to the highest received SACK block so far and holes in
2334 * With reordering, holes may still be in flight, so RFC3517 recovery
2335 * uses pure sacked_out (total number of SACKed segments) even though
2336 * it violates the RFC that uses duplicate ACKs, often these are equal
2337 * but when e.g. out-of-window ACKs or packet duplication occurs,
2338 * they differ. Since neither occurs due to loss, TCP should really
2341 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2343 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2346 static inline int tcp_skb_timedout(const struct sock *sk,
2347 const struct sk_buff *skb)
2349 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2352 static inline int tcp_head_timedout(const struct sock *sk)
2354 const struct tcp_sock *tp = tcp_sk(sk);
2356 return tp->packets_out &&
2357 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2360 /* Linux NewReno/SACK/FACK/ECN state machine.
2361 * --------------------------------------
2363 * "Open" Normal state, no dubious events, fast path.
2364 * "Disorder" In all the respects it is "Open",
2365 * but requires a bit more attention. It is entered when
2366 * we see some SACKs or dupacks. It is split of "Open"
2367 * mainly to move some processing from fast path to slow one.
2368 * "CWR" CWND was reduced due to some Congestion Notification event.
2369 * It can be ECN, ICMP source quench, local device congestion.
2370 * "Recovery" CWND was reduced, we are fast-retransmitting.
2371 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2373 * tcp_fastretrans_alert() is entered:
2374 * - each incoming ACK, if state is not "Open"
2375 * - when arrived ACK is unusual, namely:
2380 * Counting packets in flight is pretty simple.
2382 * in_flight = packets_out - left_out + retrans_out
2384 * packets_out is SND.NXT-SND.UNA counted in packets.
2386 * retrans_out is number of retransmitted segments.
2388 * left_out is number of segments left network, but not ACKed yet.
2390 * left_out = sacked_out + lost_out
2392 * sacked_out: Packets, which arrived to receiver out of order
2393 * and hence not ACKed. With SACKs this number is simply
2394 * amount of SACKed data. Even without SACKs
2395 * it is easy to give pretty reliable estimate of this number,
2396 * counting duplicate ACKs.
2398 * lost_out: Packets lost by network. TCP has no explicit
2399 * "loss notification" feedback from network (for now).
2400 * It means that this number can be only _guessed_.
2401 * Actually, it is the heuristics to predict lossage that
2402 * distinguishes different algorithms.
2404 * F.e. after RTO, when all the queue is considered as lost,
2405 * lost_out = packets_out and in_flight = retrans_out.
2407 * Essentially, we have now two algorithms counting
2410 * FACK: It is the simplest heuristics. As soon as we decided
2411 * that something is lost, we decide that _all_ not SACKed
2412 * packets until the most forward SACK are lost. I.e.
2413 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2414 * It is absolutely correct estimate, if network does not reorder
2415 * packets. And it loses any connection to reality when reordering
2416 * takes place. We use FACK by default until reordering
2417 * is suspected on the path to this destination.
2419 * NewReno: when Recovery is entered, we assume that one segment
2420 * is lost (classic Reno). While we are in Recovery and
2421 * a partial ACK arrives, we assume that one more packet
2422 * is lost (NewReno). This heuristics are the same in NewReno
2425 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2426 * deflation etc. CWND is real congestion window, never inflated, changes
2427 * only according to classic VJ rules.
2429 * Really tricky (and requiring careful tuning) part of algorithm
2430 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2431 * The first determines the moment _when_ we should reduce CWND and,
2432 * hence, slow down forward transmission. In fact, it determines the moment
2433 * when we decide that hole is caused by loss, rather than by a reorder.
2435 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2436 * holes, caused by lost packets.
2438 * And the most logically complicated part of algorithm is undo
2439 * heuristics. We detect false retransmits due to both too early
2440 * fast retransmit (reordering) and underestimated RTO, analyzing
2441 * timestamps and D-SACKs. When we detect that some segments were
2442 * retransmitted by mistake and CWND reduction was wrong, we undo
2443 * window reduction and abort recovery phase. This logic is hidden
2444 * inside several functions named tcp_try_undo_<something>.
2447 /* This function decides, when we should leave Disordered state
2448 * and enter Recovery phase, reducing congestion window.
2450 * Main question: may we further continue forward transmission
2451 * with the same cwnd?
2453 static int tcp_time_to_recover(struct sock *sk)
2455 struct tcp_sock *tp = tcp_sk(sk);
2458 /* Do not perform any recovery during F-RTO algorithm */
2459 if (tp->frto_counter)
2462 /* Trick#1: The loss is proven. */
2466 /* Not-A-Trick#2 : Classic rule... */
2467 if (tcp_dupack_heuristics(tp) > tp->reordering)
2470 /* Trick#3 : when we use RFC2988 timer restart, fast
2471 * retransmit can be triggered by timeout of queue head.
2473 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2476 /* Trick#4: It is still not OK... But will it be useful to delay
2479 packets_out = tp->packets_out;
2480 if (packets_out <= tp->reordering &&
2481 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2482 !tcp_may_send_now(sk)) {
2483 /* We have nothing to send. This connection is limited
2484 * either by receiver window or by application.
2489 /* If a thin stream is detected, retransmit after first
2490 * received dupack. Employ only if SACK is supported in order
2491 * to avoid possible corner-case series of spurious retransmissions
2492 * Use only if there are no unsent data.
2494 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2495 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2496 tcp_is_sack(tp) && !tcp_send_head(sk))
2502 /* New heuristics: it is possible only after we switched to restart timer
2503 * each time when something is ACKed. Hence, we can detect timed out packets
2504 * during fast retransmit without falling to slow start.
2506 * Usefulness of this as is very questionable, since we should know which of
2507 * the segments is the next to timeout which is relatively expensive to find
2508 * in general case unless we add some data structure just for that. The
2509 * current approach certainly won't find the right one too often and when it
2510 * finally does find _something_ it usually marks large part of the window
2511 * right away (because a retransmission with a larger timestamp blocks the
2512 * loop from advancing). -ij
2514 static void tcp_timeout_skbs(struct sock *sk)
2516 struct tcp_sock *tp = tcp_sk(sk);
2517 struct sk_buff *skb;
2519 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2522 skb = tp->scoreboard_skb_hint;
2523 if (tp->scoreboard_skb_hint == NULL)
2524 skb = tcp_write_queue_head(sk);
2526 tcp_for_write_queue_from(skb, sk) {
2527 if (skb == tcp_send_head(sk))
2529 if (!tcp_skb_timedout(sk, skb))
2532 tcp_skb_mark_lost(tp, skb);
2535 tp->scoreboard_skb_hint = skb;
2537 tcp_verify_left_out(tp);
2540 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2541 * is against sacked "cnt", otherwise it's against facked "cnt"
2543 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2545 struct tcp_sock *tp = tcp_sk(sk);
2546 struct sk_buff *skb;
2551 WARN_ON(packets > tp->packets_out);
2552 if (tp->lost_skb_hint) {
2553 skb = tp->lost_skb_hint;
2554 cnt = tp->lost_cnt_hint;
2555 /* Head already handled? */
2556 if (mark_head && skb != tcp_write_queue_head(sk))
2559 skb = tcp_write_queue_head(sk);
2563 tcp_for_write_queue_from(skb, sk) {
2564 if (skb == tcp_send_head(sk))
2566 /* TODO: do this better */
2567 /* this is not the most efficient way to do this... */
2568 tp->lost_skb_hint = skb;
2569 tp->lost_cnt_hint = cnt;
2571 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2575 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2576 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2577 cnt += tcp_skb_pcount(skb);
2579 if (cnt > packets) {
2580 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2581 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2582 (oldcnt >= packets))
2585 mss = skb_shinfo(skb)->gso_size;
2586 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2592 tcp_skb_mark_lost(tp, skb);
2597 tcp_verify_left_out(tp);
2600 /* Account newly detected lost packet(s) */
2602 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2604 struct tcp_sock *tp = tcp_sk(sk);
2606 if (tcp_is_reno(tp)) {
2607 tcp_mark_head_lost(sk, 1, 1);
2608 } else if (tcp_is_fack(tp)) {
2609 int lost = tp->fackets_out - tp->reordering;
2612 tcp_mark_head_lost(sk, lost, 0);
2614 int sacked_upto = tp->sacked_out - tp->reordering;
2615 if (sacked_upto >= 0)
2616 tcp_mark_head_lost(sk, sacked_upto, 0);
2617 else if (fast_rexmit)
2618 tcp_mark_head_lost(sk, 1, 1);
2621 tcp_timeout_skbs(sk);
2624 /* CWND moderation, preventing bursts due to too big ACKs
2625 * in dubious situations.
2627 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2629 tp->snd_cwnd = min(tp->snd_cwnd,
2630 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2631 tp->snd_cwnd_stamp = tcp_time_stamp;
2634 /* Lower bound on congestion window is slow start threshold
2635 * unless congestion avoidance choice decides to overide it.
2637 static inline u32 tcp_cwnd_min(const struct sock *sk)
2639 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2641 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2644 /* Decrease cwnd each second ack. */
2645 static void tcp_cwnd_down(struct sock *sk, int flag)
2647 struct tcp_sock *tp = tcp_sk(sk);
2648 int decr = tp->snd_cwnd_cnt + 1;
2650 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2651 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2652 tp->snd_cwnd_cnt = decr & 1;
2655 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2656 tp->snd_cwnd -= decr;
2658 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2659 tp->snd_cwnd_stamp = tcp_time_stamp;
2663 /* Nothing was retransmitted or returned timestamp is less
2664 * than timestamp of the first retransmission.
2666 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2668 return !tp->retrans_stamp ||
2669 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2670 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2673 /* Undo procedures. */
2675 #if FASTRETRANS_DEBUG > 1
2676 static void DBGUNDO(struct sock *sk, const char *msg)
2678 struct tcp_sock *tp = tcp_sk(sk);
2679 struct inet_sock *inet = inet_sk(sk);
2681 if (sk->sk_family == AF_INET) {
2682 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2684 &inet->inet_daddr, ntohs(inet->inet_dport),
2685 tp->snd_cwnd, tcp_left_out(tp),
2686 tp->snd_ssthresh, tp->prior_ssthresh,
2689 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2690 else if (sk->sk_family == AF_INET6) {
2691 struct ipv6_pinfo *np = inet6_sk(sk);
2692 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2694 &np->daddr, ntohs(inet->inet_dport),
2695 tp->snd_cwnd, tcp_left_out(tp),
2696 tp->snd_ssthresh, tp->prior_ssthresh,
2702 #define DBGUNDO(x...) do { } while (0)
2705 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2707 struct tcp_sock *tp = tcp_sk(sk);
2709 if (tp->prior_ssthresh) {
2710 const struct inet_connection_sock *icsk = inet_csk(sk);
2712 if (icsk->icsk_ca_ops->undo_cwnd)
2713 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2715 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2717 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2718 tp->snd_ssthresh = tp->prior_ssthresh;
2719 TCP_ECN_withdraw_cwr(tp);
2722 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2724 tp->snd_cwnd_stamp = tcp_time_stamp;
2727 static inline int tcp_may_undo(const struct tcp_sock *tp)
2729 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2732 /* People celebrate: "We love our President!" */
2733 static int tcp_try_undo_recovery(struct sock *sk)
2735 struct tcp_sock *tp = tcp_sk(sk);
2737 if (tcp_may_undo(tp)) {
2740 /* Happy end! We did not retransmit anything
2741 * or our original transmission succeeded.
2743 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2744 tcp_undo_cwr(sk, true);
2745 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2746 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2748 mib_idx = LINUX_MIB_TCPFULLUNDO;
2750 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2751 tp->undo_marker = 0;
2753 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2754 /* Hold old state until something *above* high_seq
2755 * is ACKed. For Reno it is MUST to prevent false
2756 * fast retransmits (RFC2582). SACK TCP is safe. */
2757 tcp_moderate_cwnd(tp);
2760 tcp_set_ca_state(sk, TCP_CA_Open);
2764 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2765 static void tcp_try_undo_dsack(struct sock *sk)
2767 struct tcp_sock *tp = tcp_sk(sk);
2769 if (tp->undo_marker && !tp->undo_retrans) {
2770 DBGUNDO(sk, "D-SACK");
2771 tcp_undo_cwr(sk, true);
2772 tp->undo_marker = 0;
2773 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2777 /* We can clear retrans_stamp when there are no retransmissions in the
2778 * window. It would seem that it is trivially available for us in
2779 * tp->retrans_out, however, that kind of assumptions doesn't consider
2780 * what will happen if errors occur when sending retransmission for the
2781 * second time. ...It could the that such segment has only
2782 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2783 * the head skb is enough except for some reneging corner cases that
2784 * are not worth the effort.
2786 * Main reason for all this complexity is the fact that connection dying
2787 * time now depends on the validity of the retrans_stamp, in particular,
2788 * that successive retransmissions of a segment must not advance
2789 * retrans_stamp under any conditions.
2791 static int tcp_any_retrans_done(const struct sock *sk)
2793 const struct tcp_sock *tp = tcp_sk(sk);
2794 struct sk_buff *skb;
2796 if (tp->retrans_out)
2799 skb = tcp_write_queue_head(sk);
2800 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2806 /* Undo during fast recovery after partial ACK. */
2808 static int tcp_try_undo_partial(struct sock *sk, int acked)
2810 struct tcp_sock *tp = tcp_sk(sk);
2811 /* Partial ACK arrived. Force Hoe's retransmit. */
2812 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2814 if (tcp_may_undo(tp)) {
2815 /* Plain luck! Hole if filled with delayed
2816 * packet, rather than with a retransmit.
2818 if (!tcp_any_retrans_done(sk))
2819 tp->retrans_stamp = 0;
2821 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2824 tcp_undo_cwr(sk, false);
2825 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2827 /* So... Do not make Hoe's retransmit yet.
2828 * If the first packet was delayed, the rest
2829 * ones are most probably delayed as well.
2836 /* Undo during loss recovery after partial ACK. */
2837 static int tcp_try_undo_loss(struct sock *sk)
2839 struct tcp_sock *tp = tcp_sk(sk);
2841 if (tcp_may_undo(tp)) {
2842 struct sk_buff *skb;
2843 tcp_for_write_queue(skb, sk) {
2844 if (skb == tcp_send_head(sk))
2846 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2849 tcp_clear_all_retrans_hints(tp);
2851 DBGUNDO(sk, "partial loss");
2853 tcp_undo_cwr(sk, true);
2854 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2855 inet_csk(sk)->icsk_retransmits = 0;
2856 tp->undo_marker = 0;
2857 if (tcp_is_sack(tp))
2858 tcp_set_ca_state(sk, TCP_CA_Open);
2864 static inline void tcp_complete_cwr(struct sock *sk)
2866 struct tcp_sock *tp = tcp_sk(sk);
2868 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2869 if (tp->undo_marker) {
2870 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR)
2871 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2873 tp->snd_cwnd = tp->snd_ssthresh;
2874 tp->snd_cwnd_stamp = tcp_time_stamp;
2876 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2879 static void tcp_try_keep_open(struct sock *sk)
2881 struct tcp_sock *tp = tcp_sk(sk);
2882 int state = TCP_CA_Open;
2884 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2885 state = TCP_CA_Disorder;
2887 if (inet_csk(sk)->icsk_ca_state != state) {
2888 tcp_set_ca_state(sk, state);
2889 tp->high_seq = tp->snd_nxt;
2893 static void tcp_try_to_open(struct sock *sk, int flag)
2895 struct tcp_sock *tp = tcp_sk(sk);
2897 tcp_verify_left_out(tp);
2899 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2900 tp->retrans_stamp = 0;
2902 if (flag & FLAG_ECE)
2903 tcp_enter_cwr(sk, 1);
2905 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2906 tcp_try_keep_open(sk);
2907 tcp_moderate_cwnd(tp);
2909 tcp_cwnd_down(sk, flag);
2913 static void tcp_mtup_probe_failed(struct sock *sk)
2915 struct inet_connection_sock *icsk = inet_csk(sk);
2917 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2918 icsk->icsk_mtup.probe_size = 0;
2921 static void tcp_mtup_probe_success(struct sock *sk)
2923 struct tcp_sock *tp = tcp_sk(sk);
2924 struct inet_connection_sock *icsk = inet_csk(sk);
2926 /* FIXME: breaks with very large cwnd */
2927 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2928 tp->snd_cwnd = tp->snd_cwnd *
2929 tcp_mss_to_mtu(sk, tp->mss_cache) /
2930 icsk->icsk_mtup.probe_size;
2931 tp->snd_cwnd_cnt = 0;
2932 tp->snd_cwnd_stamp = tcp_time_stamp;
2933 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2935 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2936 icsk->icsk_mtup.probe_size = 0;
2937 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2940 /* Do a simple retransmit without using the backoff mechanisms in
2941 * tcp_timer. This is used for path mtu discovery.
2942 * The socket is already locked here.
2944 void tcp_simple_retransmit(struct sock *sk)
2946 const struct inet_connection_sock *icsk = inet_csk(sk);
2947 struct tcp_sock *tp = tcp_sk(sk);
2948 struct sk_buff *skb;
2949 unsigned int mss = tcp_current_mss(sk);
2950 u32 prior_lost = tp->lost_out;
2952 tcp_for_write_queue(skb, sk) {
2953 if (skb == tcp_send_head(sk))
2955 if (tcp_skb_seglen(skb) > mss &&
2956 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2957 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2958 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2959 tp->retrans_out -= tcp_skb_pcount(skb);
2961 tcp_skb_mark_lost_uncond_verify(tp, skb);
2965 tcp_clear_retrans_hints_partial(tp);
2967 if (prior_lost == tp->lost_out)
2970 if (tcp_is_reno(tp))
2971 tcp_limit_reno_sacked(tp);
2973 tcp_verify_left_out(tp);
2975 /* Don't muck with the congestion window here.
2976 * Reason is that we do not increase amount of _data_
2977 * in network, but units changed and effective
2978 * cwnd/ssthresh really reduced now.
2980 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2981 tp->high_seq = tp->snd_nxt;
2982 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2983 tp->prior_ssthresh = 0;
2984 tp->undo_marker = 0;
2985 tcp_set_ca_state(sk, TCP_CA_Loss);
2987 tcp_xmit_retransmit_queue(sk);
2989 EXPORT_SYMBOL(tcp_simple_retransmit);
2991 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2992 * (proportional rate reduction with slow start reduction bound) as described in
2993 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2994 * It computes the number of packets to send (sndcnt) based on packets newly
2996 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2997 * cwnd reductions across a full RTT.
2998 * 2) If packets in flight is lower than ssthresh (such as due to excess
2999 * losses and/or application stalls), do not perform any further cwnd
3000 * reductions, but instead slow start up to ssthresh.
3002 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
3003 int fast_rexmit, int flag)
3005 struct tcp_sock *tp = tcp_sk(sk);
3007 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
3009 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
3010 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
3012 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
3014 sndcnt = min_t(int, delta,
3015 max_t(int, tp->prr_delivered - tp->prr_out,
3016 newly_acked_sacked) + 1);
3019 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
3020 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3023 /* Process an event, which can update packets-in-flight not trivially.
3024 * Main goal of this function is to calculate new estimate for left_out,
3025 * taking into account both packets sitting in receiver's buffer and
3026 * packets lost by network.
3028 * Besides that it does CWND reduction, when packet loss is detected
3029 * and changes state of machine.
3031 * It does _not_ decide what to send, it is made in function
3032 * tcp_xmit_retransmit_queue().
3034 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3035 int newly_acked_sacked, int flag)
3037 struct inet_connection_sock *icsk = inet_csk(sk);
3038 struct tcp_sock *tp = tcp_sk(sk);
3039 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3040 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3041 (tcp_fackets_out(tp) > tp->reordering));
3042 int fast_rexmit = 0, mib_idx;
3044 if (WARN_ON(!tp->packets_out && tp->sacked_out))
3046 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3047 tp->fackets_out = 0;
3049 /* Now state machine starts.
3050 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3051 if (flag & FLAG_ECE)
3052 tp->prior_ssthresh = 0;
3054 /* B. In all the states check for reneging SACKs. */
3055 if (tcp_check_sack_reneging(sk, flag))
3058 /* C. Process data loss notification, provided it is valid. */
3059 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
3060 before(tp->snd_una, tp->high_seq) &&
3061 icsk->icsk_ca_state != TCP_CA_Open &&
3062 tp->fackets_out > tp->reordering) {
3063 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
3064 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
3067 /* D. Check consistency of the current state. */
3068 tcp_verify_left_out(tp);
3070 /* E. Check state exit conditions. State can be terminated
3071 * when high_seq is ACKed. */
3072 if (icsk->icsk_ca_state == TCP_CA_Open) {
3073 WARN_ON(tp->retrans_out != 0);
3074 tp->retrans_stamp = 0;
3075 } else if (!before(tp->snd_una, tp->high_seq)) {
3076 switch (icsk->icsk_ca_state) {
3078 icsk->icsk_retransmits = 0;
3079 if (tcp_try_undo_recovery(sk))
3084 /* CWR is to be held something *above* high_seq
3085 * is ACKed for CWR bit to reach receiver. */
3086 if (tp->snd_una != tp->high_seq) {
3087 tcp_complete_cwr(sk);
3088 tcp_set_ca_state(sk, TCP_CA_Open);
3092 case TCP_CA_Disorder:
3093 tcp_try_undo_dsack(sk);
3094 if (!tp->undo_marker ||
3095 /* For SACK case do not Open to allow to undo
3096 * catching for all duplicate ACKs. */
3097 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3098 tp->undo_marker = 0;
3099 tcp_set_ca_state(sk, TCP_CA_Open);
3103 case TCP_CA_Recovery:
3104 if (tcp_is_reno(tp))
3105 tcp_reset_reno_sack(tp);
3106 if (tcp_try_undo_recovery(sk))
3108 tcp_complete_cwr(sk);
3113 /* F. Process state. */
3114 switch (icsk->icsk_ca_state) {
3115 case TCP_CA_Recovery:
3116 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3117 if (tcp_is_reno(tp) && is_dupack)
3118 tcp_add_reno_sack(sk);
3120 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3123 if (flag & FLAG_DATA_ACKED)
3124 icsk->icsk_retransmits = 0;
3125 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3126 tcp_reset_reno_sack(tp);
3127 if (!tcp_try_undo_loss(sk)) {
3128 tcp_moderate_cwnd(tp);
3129 tcp_xmit_retransmit_queue(sk);
3132 if (icsk->icsk_ca_state != TCP_CA_Open)
3134 /* Loss is undone; fall through to processing in Open state. */
3136 if (tcp_is_reno(tp)) {
3137 if (flag & FLAG_SND_UNA_ADVANCED)
3138 tcp_reset_reno_sack(tp);
3140 tcp_add_reno_sack(sk);
3143 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3144 tcp_try_undo_dsack(sk);
3146 if (!tcp_time_to_recover(sk)) {
3147 tcp_try_to_open(sk, flag);
3151 /* MTU probe failure: don't reduce cwnd */
3152 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3153 icsk->icsk_mtup.probe_size &&
3154 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3155 tcp_mtup_probe_failed(sk);
3156 /* Restores the reduction we did in tcp_mtup_probe() */
3158 tcp_simple_retransmit(sk);
3162 /* Otherwise enter Recovery state */
3164 if (tcp_is_reno(tp))
3165 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3167 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3169 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3171 tp->high_seq = tp->snd_nxt;
3172 tp->prior_ssthresh = 0;
3173 tp->undo_marker = tp->snd_una;
3174 tp->undo_retrans = tp->retrans_out;
3176 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3177 if (!(flag & FLAG_ECE))
3178 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3179 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3180 TCP_ECN_queue_cwr(tp);
3183 tp->bytes_acked = 0;
3184 tp->snd_cwnd_cnt = 0;
3185 tp->prior_cwnd = tp->snd_cwnd;
3186 tp->prr_delivered = 0;
3188 tcp_set_ca_state(sk, TCP_CA_Recovery);
3192 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3193 tcp_update_scoreboard(sk, fast_rexmit);
3194 tp->prr_delivered += newly_acked_sacked;
3195 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3196 tcp_xmit_retransmit_queue(sk);
3199 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3201 tcp_rtt_estimator(sk, seq_rtt);
3203 inet_csk(sk)->icsk_backoff = 0;
3205 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3207 /* Read draft-ietf-tcplw-high-performance before mucking
3208 * with this code. (Supersedes RFC1323)
3210 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3212 /* RTTM Rule: A TSecr value received in a segment is used to
3213 * update the averaged RTT measurement only if the segment
3214 * acknowledges some new data, i.e., only if it advances the
3215 * left edge of the send window.
3217 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3218 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3220 * Changed: reset backoff as soon as we see the first valid sample.
3221 * If we do not, we get strongly overestimated rto. With timestamps
3222 * samples are accepted even from very old segments: f.e., when rtt=1
3223 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3224 * answer arrives rto becomes 120 seconds! If at least one of segments
3225 * in window is lost... Voila. --ANK (010210)
3227 struct tcp_sock *tp = tcp_sk(sk);
3229 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3232 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3234 /* We don't have a timestamp. Can only use
3235 * packets that are not retransmitted to determine
3236 * rtt estimates. Also, we must not reset the
3237 * backoff for rto until we get a non-retransmitted
3238 * packet. This allows us to deal with a situation
3239 * where the network delay has increased suddenly.
3240 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3243 if (flag & FLAG_RETRANS_DATA_ACKED)
3246 tcp_valid_rtt_meas(sk, seq_rtt);
3249 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3252 const struct tcp_sock *tp = tcp_sk(sk);
3253 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3254 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3255 tcp_ack_saw_tstamp(sk, flag);
3256 else if (seq_rtt >= 0)
3257 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3260 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3262 const struct inet_connection_sock *icsk = inet_csk(sk);
3263 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3264 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3267 /* Restart timer after forward progress on connection.
3268 * RFC2988 recommends to restart timer to now+rto.
3270 static void tcp_rearm_rto(struct sock *sk)
3272 const struct tcp_sock *tp = tcp_sk(sk);
3274 if (!tp->packets_out) {
3275 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3277 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3278 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3282 /* If we get here, the whole TSO packet has not been acked. */
3283 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3285 struct tcp_sock *tp = tcp_sk(sk);
3288 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3290 packets_acked = tcp_skb_pcount(skb);
3291 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3293 packets_acked -= tcp_skb_pcount(skb);
3295 if (packets_acked) {
3296 BUG_ON(tcp_skb_pcount(skb) == 0);
3297 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3300 return packets_acked;
3303 /* Remove acknowledged frames from the retransmission queue. If our packet
3304 * is before the ack sequence we can discard it as it's confirmed to have
3305 * arrived at the other end.
3307 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3310 struct tcp_sock *tp = tcp_sk(sk);
3311 const struct inet_connection_sock *icsk = inet_csk(sk);
3312 struct sk_buff *skb;
3313 u32 now = tcp_time_stamp;
3314 int fully_acked = 1;
3317 u32 reord = tp->packets_out;
3318 u32 prior_sacked = tp->sacked_out;
3320 s32 ca_seq_rtt = -1;
3321 ktime_t last_ackt = net_invalid_timestamp();
3323 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3324 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3326 u8 sacked = scb->sacked;
3328 /* Determine how many packets and what bytes were acked, tso and else */
3329 if (after(scb->end_seq, tp->snd_una)) {
3330 if (tcp_skb_pcount(skb) == 1 ||
3331 !after(tp->snd_una, scb->seq))
3334 acked_pcount = tcp_tso_acked(sk, skb);
3340 acked_pcount = tcp_skb_pcount(skb);
3343 if (sacked & TCPCB_RETRANS) {
3344 if (sacked & TCPCB_SACKED_RETRANS)
3345 tp->retrans_out -= acked_pcount;
3346 flag |= FLAG_RETRANS_DATA_ACKED;
3349 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3350 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3352 ca_seq_rtt = now - scb->when;
3353 last_ackt = skb->tstamp;
3355 seq_rtt = ca_seq_rtt;
3357 if (!(sacked & TCPCB_SACKED_ACKED))
3358 reord = min(pkts_acked, reord);
3361 if (sacked & TCPCB_SACKED_ACKED)
3362 tp->sacked_out -= acked_pcount;
3363 if (sacked & TCPCB_LOST)
3364 tp->lost_out -= acked_pcount;
3366 tp->packets_out -= acked_pcount;
3367 pkts_acked += acked_pcount;
3369 /* Initial outgoing SYN's get put onto the write_queue
3370 * just like anything else we transmit. It is not
3371 * true data, and if we misinform our callers that
3372 * this ACK acks real data, we will erroneously exit
3373 * connection startup slow start one packet too
3374 * quickly. This is severely frowned upon behavior.
3376 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3377 flag |= FLAG_DATA_ACKED;
3379 flag |= FLAG_SYN_ACKED;
3380 tp->retrans_stamp = 0;
3386 tcp_unlink_write_queue(skb, sk);
3387 sk_wmem_free_skb(sk, skb);
3388 tp->scoreboard_skb_hint = NULL;
3389 if (skb == tp->retransmit_skb_hint)
3390 tp->retransmit_skb_hint = NULL;
3391 if (skb == tp->lost_skb_hint)
3392 tp->lost_skb_hint = NULL;
3395 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3396 tp->snd_up = tp->snd_una;
3398 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3399 flag |= FLAG_SACK_RENEGING;
3401 if (flag & FLAG_ACKED) {
3402 const struct tcp_congestion_ops *ca_ops
3403 = inet_csk(sk)->icsk_ca_ops;
3405 if (unlikely(icsk->icsk_mtup.probe_size &&
3406 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3407 tcp_mtup_probe_success(sk);
3410 tcp_ack_update_rtt(sk, flag, seq_rtt);
3413 if (tcp_is_reno(tp)) {
3414 tcp_remove_reno_sacks(sk, pkts_acked);
3418 /* Non-retransmitted hole got filled? That's reordering */
3419 if (reord < prior_fackets)
3420 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3422 delta = tcp_is_fack(tp) ? pkts_acked :
3423 prior_sacked - tp->sacked_out;
3424 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3427 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3429 if (ca_ops->pkts_acked) {
3432 /* Is the ACK triggering packet unambiguous? */
3433 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3434 /* High resolution needed and available? */
3435 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3436 !ktime_equal(last_ackt,
3437 net_invalid_timestamp()))
3438 rtt_us = ktime_us_delta(ktime_get_real(),
3440 else if (ca_seq_rtt >= 0)
3441 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3444 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3448 #if FASTRETRANS_DEBUG > 0
3449 WARN_ON((int)tp->sacked_out < 0);
3450 WARN_ON((int)tp->lost_out < 0);
3451 WARN_ON((int)tp->retrans_out < 0);
3452 if (!tp->packets_out && tcp_is_sack(tp)) {
3453 icsk = inet_csk(sk);
3455 printk(KERN_DEBUG "Leak l=%u %d\n",
3456 tp->lost_out, icsk->icsk_ca_state);
3459 if (tp->sacked_out) {
3460 printk(KERN_DEBUG "Leak s=%u %d\n",
3461 tp->sacked_out, icsk->icsk_ca_state);
3464 if (tp->retrans_out) {
3465 printk(KERN_DEBUG "Leak r=%u %d\n",
3466 tp->retrans_out, icsk->icsk_ca_state);
3467 tp->retrans_out = 0;
3474 static void tcp_ack_probe(struct sock *sk)
3476 const struct tcp_sock *tp = tcp_sk(sk);
3477 struct inet_connection_sock *icsk = inet_csk(sk);
3479 /* Was it a usable window open? */
3481 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3482 icsk->icsk_backoff = 0;
3483 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3484 /* Socket must be waked up by subsequent tcp_data_snd_check().
3485 * This function is not for random using!
3488 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3489 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3494 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3496 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3497 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3500 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3502 const struct tcp_sock *tp = tcp_sk(sk);
3503 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3504 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3507 /* Check that window update is acceptable.
3508 * The function assumes that snd_una<=ack<=snd_next.