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/module.h>
66 #include <linux/sysctl.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71 #include <asm/unaligned.h>
72 #include <net/netdma.h>
74 int sysctl_tcp_timestamps __read_mostly = 1;
75 int sysctl_tcp_window_scaling __read_mostly = 1;
76 int sysctl_tcp_sack __read_mostly = 1;
77 int sysctl_tcp_fack __read_mostly = 1;
78 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
79 int sysctl_tcp_ecn __read_mostly;
80 int sysctl_tcp_dsack __read_mostly = 1;
81 int sysctl_tcp_app_win __read_mostly = 31;
82 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84 int sysctl_tcp_stdurg __read_mostly;
85 int sysctl_tcp_rfc1337 __read_mostly;
86 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
87 int sysctl_tcp_frto __read_mostly = 2;
88 int sysctl_tcp_frto_response __read_mostly;
89 int sysctl_tcp_nometrics_save __read_mostly;
91 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
92 int sysctl_tcp_abc __read_mostly;
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
104 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
105 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
106 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
107 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
118 /* Adapt the MSS value used to make delayed ack decision to the
121 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
123 struct inet_connection_sock *icsk = inet_csk(sk);
124 const unsigned int lss = icsk->icsk_ack.last_seg_size;
127 icsk->icsk_ack.last_seg_size = 0;
129 /* skb->len may jitter because of SACKs, even if peer
130 * sends good full-sized frames.
132 len = skb_shinfo(skb)->gso_size ? : skb->len;
133 if (len >= icsk->icsk_ack.rcv_mss) {
134 icsk->icsk_ack.rcv_mss = len;
136 /* Otherwise, we make more careful check taking into account,
137 * that SACKs block is variable.
139 * "len" is invariant segment length, including TCP header.
141 len += skb->data - skb_transport_header(skb);
142 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
143 /* If PSH is not set, packet should be
144 * full sized, provided peer TCP is not badly broken.
145 * This observation (if it is correct 8)) allows
146 * to handle super-low mtu links fairly.
148 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
149 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
150 /* Subtract also invariant (if peer is RFC compliant),
151 * tcp header plus fixed timestamp option length.
152 * Resulting "len" is MSS free of SACK jitter.
154 len -= tcp_sk(sk)->tcp_header_len;
155 icsk->icsk_ack.last_seg_size = len;
157 icsk->icsk_ack.rcv_mss = len;
161 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
162 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
163 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
167 static void tcp_incr_quickack(struct sock *sk)
169 struct inet_connection_sock *icsk = inet_csk(sk);
170 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
174 if (quickacks > icsk->icsk_ack.quick)
175 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
178 void tcp_enter_quickack_mode(struct sock *sk)
180 struct inet_connection_sock *icsk = inet_csk(sk);
181 tcp_incr_quickack(sk);
182 icsk->icsk_ack.pingpong = 0;
183 icsk->icsk_ack.ato = TCP_ATO_MIN;
186 /* Send ACKs quickly, if "quick" count is not exhausted
187 * and the session is not interactive.
190 static inline int tcp_in_quickack_mode(const struct sock *sk)
192 const struct inet_connection_sock *icsk = inet_csk(sk);
193 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
196 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
198 if (tp->ecn_flags & TCP_ECN_OK)
199 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
202 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
204 if (tcp_hdr(skb)->cwr)
205 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
208 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
210 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
213 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
215 if (tp->ecn_flags & TCP_ECN_OK) {
216 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
217 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
218 /* Funny extension: if ECT is not set on a segment,
219 * it is surely retransmit. It is not in ECN RFC,
220 * but Linux follows this rule. */
221 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
222 tcp_enter_quickack_mode((struct sock *)tp);
226 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
228 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
229 tp->ecn_flags &= ~TCP_ECN_OK;
232 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
234 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
235 tp->ecn_flags &= ~TCP_ECN_OK;
238 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
240 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
245 /* Buffer size and advertised window tuning.
247 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
250 static void tcp_fixup_sndbuf(struct sock *sk)
252 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
253 sizeof(struct sk_buff);
255 if (sk->sk_sndbuf < 3 * sndmem)
256 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
259 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261 * All tcp_full_space() is split to two parts: "network" buffer, allocated
262 * forward and advertised in receiver window (tp->rcv_wnd) and
263 * "application buffer", required to isolate scheduling/application
264 * latencies from network.
265 * window_clamp is maximal advertised window. It can be less than
266 * tcp_full_space(), in this case tcp_full_space() - window_clamp
267 * is reserved for "application" buffer. The less window_clamp is
268 * the smoother our behaviour from viewpoint of network, but the lower
269 * throughput and the higher sensitivity of the connection to losses. 8)
271 * rcv_ssthresh is more strict window_clamp used at "slow start"
272 * phase to predict further behaviour of this connection.
273 * It is used for two goals:
274 * - to enforce header prediction at sender, even when application
275 * requires some significant "application buffer". It is check #1.
276 * - to prevent pruning of receive queue because of misprediction
277 * of receiver window. Check #2.
279 * The scheme does not work when sender sends good segments opening
280 * window and then starts to feed us spaghetti. But it should work
281 * in common situations. Otherwise, we have to rely on queue collapsing.
284 /* Slow part of check#2. */
285 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
287 struct tcp_sock *tp = tcp_sk(sk);
289 int truesize = tcp_win_from_space(skb->truesize) >> 1;
290 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
292 while (tp->rcv_ssthresh <= window) {
293 if (truesize <= skb->len)
294 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
302 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
304 struct tcp_sock *tp = tcp_sk(sk);
307 if (tp->rcv_ssthresh < tp->window_clamp &&
308 (int)tp->rcv_ssthresh < tcp_space(sk) &&
309 !tcp_memory_pressure) {
312 /* Check #2. Increase window, if skb with such overhead
313 * will fit to rcvbuf in future.
315 if (tcp_win_from_space(skb->truesize) <= skb->len)
316 incr = 2 * tp->advmss;
318 incr = __tcp_grow_window(sk, skb);
321 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
323 inet_csk(sk)->icsk_ack.quick |= 1;
328 /* 3. Tuning rcvbuf, when connection enters established state. */
330 static void tcp_fixup_rcvbuf(struct sock *sk)
332 struct tcp_sock *tp = tcp_sk(sk);
333 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
335 /* Try to select rcvbuf so that 4 mss-sized segments
336 * will fit to window and corresponding skbs will fit to our rcvbuf.
337 * (was 3; 4 is minimum to allow fast retransmit to work.)
339 while (tcp_win_from_space(rcvmem) < tp->advmss)
341 if (sk->sk_rcvbuf < 4 * rcvmem)
342 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
345 /* 4. Try to fixup all. It is made immediately after connection enters
348 static void tcp_init_buffer_space(struct sock *sk)
350 struct tcp_sock *tp = tcp_sk(sk);
353 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
354 tcp_fixup_rcvbuf(sk);
355 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
356 tcp_fixup_sndbuf(sk);
358 tp->rcvq_space.space = tp->rcv_wnd;
360 maxwin = tcp_full_space(sk);
362 if (tp->window_clamp >= maxwin) {
363 tp->window_clamp = maxwin;
365 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
366 tp->window_clamp = max(maxwin -
367 (maxwin >> sysctl_tcp_app_win),
371 /* Force reservation of one segment. */
372 if (sysctl_tcp_app_win &&
373 tp->window_clamp > 2 * tp->advmss &&
374 tp->window_clamp + tp->advmss > maxwin)
375 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
377 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
378 tp->snd_cwnd_stamp = tcp_time_stamp;
381 /* 5. Recalculate window clamp after socket hit its memory bounds. */
382 static void tcp_clamp_window(struct sock *sk)
384 struct tcp_sock *tp = tcp_sk(sk);
385 struct inet_connection_sock *icsk = inet_csk(sk);
387 icsk->icsk_ack.quick = 0;
389 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
390 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
391 !tcp_memory_pressure &&
392 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
393 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
396 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
397 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
400 /* Initialize RCV_MSS value.
401 * RCV_MSS is an our guess about MSS used by the peer.
402 * We haven't any direct information about the MSS.
403 * It's better to underestimate the RCV_MSS rather than overestimate.
404 * Overestimations make us ACKing less frequently than needed.
405 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407 void tcp_initialize_rcv_mss(struct sock *sk)
409 struct tcp_sock *tp = tcp_sk(sk);
410 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
412 hint = min(hint, tp->rcv_wnd / 2);
413 hint = min(hint, TCP_MIN_RCVMSS);
414 hint = max(hint, TCP_MIN_MSS);
416 inet_csk(sk)->icsk_ack.rcv_mss = hint;
419 /* Receiver "autotuning" code.
421 * The algorithm for RTT estimation w/o timestamps is based on
422 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
423 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425 * More detail on this code can be found at
426 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
427 * though this reference is out of date. A new paper
430 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
432 u32 new_sample = tp->rcv_rtt_est.rtt;
438 if (new_sample != 0) {
439 /* If we sample in larger samples in the non-timestamp
440 * case, we could grossly overestimate the RTT especially
441 * with chatty applications or bulk transfer apps which
442 * are stalled on filesystem I/O.
444 * Also, since we are only going for a minimum in the
445 * non-timestamp case, we do not smooth things out
446 * else with timestamps disabled convergence takes too
450 m -= (new_sample >> 3);
452 } else if (m < new_sample)
455 /* No previous measure. */
459 if (tp->rcv_rtt_est.rtt != new_sample)
460 tp->rcv_rtt_est.rtt = new_sample;
463 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
465 if (tp->rcv_rtt_est.time == 0)
467 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
469 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
472 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
473 tp->rcv_rtt_est.time = tcp_time_stamp;
476 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
477 const struct sk_buff *skb)
479 struct tcp_sock *tp = tcp_sk(sk);
480 if (tp->rx_opt.rcv_tsecr &&
481 (TCP_SKB_CB(skb)->end_seq -
482 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
483 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
487 * This function should be called every time data is copied to user space.
488 * It calculates the appropriate TCP receive buffer space.
490 void tcp_rcv_space_adjust(struct sock *sk)
492 struct tcp_sock *tp = tcp_sk(sk);
496 if (tp->rcvq_space.time == 0)
499 time = tcp_time_stamp - tp->rcvq_space.time;
500 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
503 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
505 space = max(tp->rcvq_space.space, space);
507 if (tp->rcvq_space.space != space) {
510 tp->rcvq_space.space = space;
512 if (sysctl_tcp_moderate_rcvbuf &&
513 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
514 int new_clamp = space;
516 /* Receive space grows, normalize in order to
517 * take into account packet headers and sk_buff
518 * structure overhead.
523 rcvmem = (tp->advmss + MAX_TCP_HEADER +
524 16 + sizeof(struct sk_buff));
525 while (tcp_win_from_space(rcvmem) < tp->advmss)
528 space = min(space, sysctl_tcp_rmem[2]);
529 if (space > sk->sk_rcvbuf) {
530 sk->sk_rcvbuf = space;
532 /* Make the window clamp follow along. */
533 tp->window_clamp = new_clamp;
539 tp->rcvq_space.seq = tp->copied_seq;
540 tp->rcvq_space.time = tcp_time_stamp;
543 /* There is something which you must keep in mind when you analyze the
544 * behavior of the tp->ato delayed ack timeout interval. When a
545 * connection starts up, we want to ack as quickly as possible. The
546 * problem is that "good" TCP's do slow start at the beginning of data
547 * transmission. The means that until we send the first few ACK's the
548 * sender will sit on his end and only queue most of his data, because
549 * he can only send snd_cwnd unacked packets at any given time. For
550 * each ACK we send, he increments snd_cwnd and transmits more of his
553 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
555 struct tcp_sock *tp = tcp_sk(sk);
556 struct inet_connection_sock *icsk = inet_csk(sk);
559 inet_csk_schedule_ack(sk);
561 tcp_measure_rcv_mss(sk, skb);
563 tcp_rcv_rtt_measure(tp);
565 now = tcp_time_stamp;
567 if (!icsk->icsk_ack.ato) {
568 /* The _first_ data packet received, initialize
569 * delayed ACK engine.
571 tcp_incr_quickack(sk);
572 icsk->icsk_ack.ato = TCP_ATO_MIN;
574 int m = now - icsk->icsk_ack.lrcvtime;
576 if (m <= TCP_ATO_MIN / 2) {
577 /* The fastest case is the first. */
578 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
579 } else if (m < icsk->icsk_ack.ato) {
580 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
581 if (icsk->icsk_ack.ato > icsk->icsk_rto)
582 icsk->icsk_ack.ato = icsk->icsk_rto;
583 } else if (m > icsk->icsk_rto) {
584 /* Too long gap. Apparently sender failed to
585 * restart window, so that we send ACKs quickly.
587 tcp_incr_quickack(sk);
591 icsk->icsk_ack.lrcvtime = now;
593 TCP_ECN_check_ce(tp, skb);
596 tcp_grow_window(sk, skb);
599 static u32 tcp_rto_min(struct sock *sk)
601 struct dst_entry *dst = __sk_dst_get(sk);
602 u32 rto_min = TCP_RTO_MIN;
604 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
605 rto_min = dst_metric(dst, RTAX_RTO_MIN);
609 /* Called to compute a smoothed rtt estimate. The data fed to this
610 * routine either comes from timestamps, or from segments that were
611 * known _not_ to have been retransmitted [see Karn/Partridge
612 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
613 * piece by Van Jacobson.
614 * NOTE: the next three routines used to be one big routine.
615 * To save cycles in the RFC 1323 implementation it was better to break
616 * it up into three procedures. -- erics
618 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
620 struct tcp_sock *tp = tcp_sk(sk);
621 long m = mrtt; /* RTT */
623 /* The following amusing code comes from Jacobson's
624 * article in SIGCOMM '88. Note that rtt and mdev
625 * are scaled versions of rtt and mean deviation.
626 * This is designed to be as fast as possible
627 * m stands for "measurement".
629 * On a 1990 paper the rto value is changed to:
630 * RTO = rtt + 4 * mdev
632 * Funny. This algorithm seems to be very broken.
633 * These formulae increase RTO, when it should be decreased, increase
634 * too slowly, when it should be increased quickly, decrease too quickly
635 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
636 * does not matter how to _calculate_ it. Seems, it was trap
637 * that VJ failed to avoid. 8)
642 m -= (tp->srtt >> 3); /* m is now error in rtt est */
643 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
645 m = -m; /* m is now abs(error) */
646 m -= (tp->mdev >> 2); /* similar update on mdev */
647 /* This is similar to one of Eifel findings.
648 * Eifel blocks mdev updates when rtt decreases.
649 * This solution is a bit different: we use finer gain
650 * for mdev in this case (alpha*beta).
651 * Like Eifel it also prevents growth of rto,
652 * but also it limits too fast rto decreases,
653 * happening in pure Eifel.
658 m -= (tp->mdev >> 2); /* similar update on mdev */
660 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
661 if (tp->mdev > tp->mdev_max) {
662 tp->mdev_max = tp->mdev;
663 if (tp->mdev_max > tp->rttvar)
664 tp->rttvar = tp->mdev_max;
666 if (after(tp->snd_una, tp->rtt_seq)) {
667 if (tp->mdev_max < tp->rttvar)
668 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
669 tp->rtt_seq = tp->snd_nxt;
670 tp->mdev_max = tcp_rto_min(sk);
673 /* no previous measure. */
674 tp->srtt = m << 3; /* take the measured time to be rtt */
675 tp->mdev = m << 1; /* make sure rto = 3*rtt */
676 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
677 tp->rtt_seq = tp->snd_nxt;
681 /* Calculate rto without backoff. This is the second half of Van Jacobson's
682 * routine referred to above.
684 static inline void tcp_set_rto(struct sock *sk)
686 const struct tcp_sock *tp = tcp_sk(sk);
687 /* Old crap is replaced with new one. 8)
690 * 1. If rtt variance happened to be less 50msec, it is hallucination.
691 * It cannot be less due to utterly erratic ACK generation made
692 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
693 * to do with delayed acks, because at cwnd>2 true delack timeout
694 * is invisible. Actually, Linux-2.4 also generates erratic
695 * ACKs in some circumstances.
697 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
699 /* 2. Fixups made earlier cannot be right.
700 * If we do not estimate RTO correctly without them,
701 * all the algo is pure shit and should be replaced
702 * with correct one. It is exactly, which we pretend to do.
706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707 * guarantees that rto is higher.
709 static inline void tcp_bound_rto(struct sock *sk)
711 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
712 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
715 /* Save metrics learned by this TCP session.
716 This function is called only, when TCP finishes successfully
717 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
719 void tcp_update_metrics(struct sock *sk)
721 struct tcp_sock *tp = tcp_sk(sk);
722 struct dst_entry *dst = __sk_dst_get(sk);
724 if (sysctl_tcp_nometrics_save)
729 if (dst && (dst->flags & DST_HOST)) {
730 const struct inet_connection_sock *icsk = inet_csk(sk);
733 if (icsk->icsk_backoff || !tp->srtt) {
734 /* This session failed to estimate rtt. Why?
735 * Probably, no packets returned in time.
738 if (!(dst_metric_locked(dst, RTAX_RTT)))
739 dst->metrics[RTAX_RTT - 1] = 0;
743 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
745 /* If newly calculated rtt larger than stored one,
746 * store new one. Otherwise, use EWMA. Remember,
747 * rtt overestimation is always better than underestimation.
749 if (!(dst_metric_locked(dst, RTAX_RTT))) {
751 dst->metrics[RTAX_RTT - 1] = tp->srtt;
753 dst->metrics[RTAX_RTT - 1] -= (m >> 3);
756 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
760 /* Scale deviation to rttvar fixed point */
765 if (m >= dst_metric(dst, RTAX_RTTVAR))
766 dst->metrics[RTAX_RTTVAR - 1] = m;
768 dst->metrics[RTAX_RTTVAR-1] -=
769 (dst_metric(dst, RTAX_RTTVAR) - m)>>2;
772 if (tp->snd_ssthresh >= 0xFFFF) {
773 /* Slow start still did not finish. */
774 if (dst_metric(dst, RTAX_SSTHRESH) &&
775 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
776 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
777 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
778 if (!dst_metric_locked(dst, RTAX_CWND) &&
779 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
780 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
781 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
782 icsk->icsk_ca_state == TCP_CA_Open) {
783 /* Cong. avoidance phase, cwnd is reliable. */
784 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
785 dst->metrics[RTAX_SSTHRESH-1] =
786 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
787 if (!dst_metric_locked(dst, RTAX_CWND))
788 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
790 /* Else slow start did not finish, cwnd is non-sense,
791 ssthresh may be also invalid.
793 if (!dst_metric_locked(dst, RTAX_CWND))
794 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
795 if (dst_metric(dst, RTAX_SSTHRESH) &&
796 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
797 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
798 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
801 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
802 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
803 tp->reordering != sysctl_tcp_reordering)
804 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
809 /* Numbers are taken from RFC3390.
811 * John Heffner states:
813 * The RFC specifies a window of no more than 4380 bytes
814 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
815 * is a bit misleading because they use a clamp at 4380 bytes
816 * rather than use a multiplier in the relevant range.
818 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
820 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
823 if (tp->mss_cache > 1460)
826 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
828 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
831 /* Set slow start threshold and cwnd not falling to slow start */
832 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
834 struct tcp_sock *tp = tcp_sk(sk);
835 const struct inet_connection_sock *icsk = inet_csk(sk);
837 tp->prior_ssthresh = 0;
839 if (icsk->icsk_ca_state < TCP_CA_CWR) {
842 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
843 tp->snd_cwnd = min(tp->snd_cwnd,
844 tcp_packets_in_flight(tp) + 1U);
845 tp->snd_cwnd_cnt = 0;
846 tp->high_seq = tp->snd_nxt;
847 tp->snd_cwnd_stamp = tcp_time_stamp;
848 TCP_ECN_queue_cwr(tp);
850 tcp_set_ca_state(sk, TCP_CA_CWR);
855 * Packet counting of FACK is based on in-order assumptions, therefore TCP
856 * disables it when reordering is detected
858 static void tcp_disable_fack(struct tcp_sock *tp)
860 /* RFC3517 uses different metric in lost marker => reset on change */
862 tp->lost_skb_hint = NULL;
863 tp->rx_opt.sack_ok &= ~2;
866 /* Take a notice that peer is sending D-SACKs */
867 static void tcp_dsack_seen(struct tcp_sock *tp)
869 tp->rx_opt.sack_ok |= 4;
872 /* Initialize metrics on socket. */
874 static void tcp_init_metrics(struct sock *sk)
876 struct tcp_sock *tp = tcp_sk(sk);
877 struct dst_entry *dst = __sk_dst_get(sk);
884 if (dst_metric_locked(dst, RTAX_CWND))
885 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
886 if (dst_metric(dst, RTAX_SSTHRESH)) {
887 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
888 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
889 tp->snd_ssthresh = tp->snd_cwnd_clamp;
891 if (dst_metric(dst, RTAX_REORDERING) &&
892 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
893 tcp_disable_fack(tp);
894 tp->reordering = dst_metric(dst, RTAX_REORDERING);
897 if (dst_metric(dst, RTAX_RTT) == 0)
900 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
903 /* Initial rtt is determined from SYN,SYN-ACK.
904 * The segment is small and rtt may appear much
905 * less than real one. Use per-dst memory
906 * to make it more realistic.
908 * A bit of theory. RTT is time passed after "normal" sized packet
909 * is sent until it is ACKed. In normal circumstances sending small
910 * packets force peer to delay ACKs and calculation is correct too.
911 * The algorithm is adaptive and, provided we follow specs, it
912 * NEVER underestimate RTT. BUT! If peer tries to make some clever
913 * tricks sort of "quick acks" for time long enough to decrease RTT
914 * to low value, and then abruptly stops to do it and starts to delay
915 * ACKs, wait for troubles.
917 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
918 tp->srtt = dst_metric(dst, RTAX_RTT);
919 tp->rtt_seq = tp->snd_nxt;
921 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
922 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
923 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
927 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
929 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
930 tp->snd_cwnd_stamp = tcp_time_stamp;
934 /* Play conservative. If timestamps are not
935 * supported, TCP will fail to recalculate correct
936 * rtt, if initial rto is too small. FORGET ALL AND RESET!
938 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
940 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
941 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
945 static void tcp_update_reordering(struct sock *sk, const int metric,
948 struct tcp_sock *tp = tcp_sk(sk);
949 if (metric > tp->reordering) {
952 tp->reordering = min(TCP_MAX_REORDERING, metric);
954 /* This exciting event is worth to be remembered. 8) */
956 mib_idx = LINUX_MIB_TCPTSREORDER;
957 else if (tcp_is_reno(tp))
958 mib_idx = LINUX_MIB_TCPRENOREORDER;
959 else if (tcp_is_fack(tp))
960 mib_idx = LINUX_MIB_TCPFACKREORDER;
962 mib_idx = LINUX_MIB_TCPSACKREORDER;
964 NET_INC_STATS_BH(mib_idx);
965 #if FASTRETRANS_DEBUG > 1
966 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
967 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
971 tp->undo_marker ? tp->undo_retrans : 0);
973 tcp_disable_fack(tp);
977 /* This procedure tags the retransmission queue when SACKs arrive.
979 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
980 * Packets in queue with these bits set are counted in variables
981 * sacked_out, retrans_out and lost_out, correspondingly.
983 * Valid combinations are:
984 * Tag InFlight Description
985 * 0 1 - orig segment is in flight.
986 * S 0 - nothing flies, orig reached receiver.
987 * L 0 - nothing flies, orig lost by net.
988 * R 2 - both orig and retransmit are in flight.
989 * L|R 1 - orig is lost, retransmit is in flight.
990 * S|R 1 - orig reached receiver, retrans is still in flight.
991 * (L|S|R is logically valid, it could occur when L|R is sacked,
992 * but it is equivalent to plain S and code short-curcuits it to S.
993 * L|S is logically invalid, it would mean -1 packet in flight 8))
995 * These 6 states form finite state machine, controlled by the following events:
996 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
997 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
998 * 3. Loss detection event of one of three flavors:
999 * A. Scoreboard estimator decided the packet is lost.
1000 * A'. Reno "three dupacks" marks head of queue lost.
1001 * A''. Its FACK modfication, head until snd.fack is lost.
1002 * B. SACK arrives sacking data transmitted after never retransmitted
1003 * hole was sent out.
1004 * C. SACK arrives sacking SND.NXT at the moment, when the
1005 * segment was retransmitted.
1006 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1008 * It is pleasant to note, that state diagram turns out to be commutative,
1009 * so that we are allowed not to be bothered by order of our actions,
1010 * when multiple events arrive simultaneously. (see the function below).
1012 * Reordering detection.
1013 * --------------------
1014 * Reordering metric is maximal distance, which a packet can be displaced
1015 * in packet stream. With SACKs we can estimate it:
1017 * 1. SACK fills old hole and the corresponding segment was not
1018 * ever retransmitted -> reordering. Alas, we cannot use it
1019 * when segment was retransmitted.
1020 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1021 * for retransmitted and already SACKed segment -> reordering..
1022 * Both of these heuristics are not used in Loss state, when we cannot
1023 * account for retransmits accurately.
1025 * SACK block validation.
1026 * ----------------------
1028 * SACK block range validation checks that the received SACK block fits to
1029 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1030 * Note that SND.UNA is not included to the range though being valid because
1031 * it means that the receiver is rather inconsistent with itself reporting
1032 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1033 * perfectly valid, however, in light of RFC2018 which explicitly states
1034 * that "SACK block MUST reflect the newest segment. Even if the newest
1035 * segment is going to be discarded ...", not that it looks very clever
1036 * in case of head skb. Due to potentional receiver driven attacks, we
1037 * choose to avoid immediate execution of a walk in write queue due to
1038 * reneging and defer head skb's loss recovery to standard loss recovery
1039 * procedure that will eventually trigger (nothing forbids us doing this).
1041 * Implements also blockage to start_seq wrap-around. Problem lies in the
1042 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1043 * there's no guarantee that it will be before snd_nxt (n). The problem
1044 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1047 * <- outs wnd -> <- wrapzone ->
1048 * u e n u_w e_w s n_w
1050 * |<------------+------+----- TCP seqno space --------------+---------->|
1051 * ...-- <2^31 ->| |<--------...
1052 * ...---- >2^31 ------>| |<--------...
1054 * Current code wouldn't be vulnerable but it's better still to discard such
1055 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1056 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1057 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1058 * equal to the ideal case (infinite seqno space without wrap caused issues).
1060 * With D-SACK the lower bound is extended to cover sequence space below
1061 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1062 * again, D-SACK block must not to go across snd_una (for the same reason as
1063 * for the normal SACK blocks, explained above). But there all simplicity
1064 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1065 * fully below undo_marker they do not affect behavior in anyway and can
1066 * therefore be safely ignored. In rare cases (which are more or less
1067 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1068 * fragmentation and packet reordering past skb's retransmission. To consider
1069 * them correctly, the acceptable range must be extended even more though
1070 * the exact amount is rather hard to quantify. However, tp->max_window can
1071 * be used as an exaggerated estimate.
1073 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1074 u32 start_seq, u32 end_seq)
1076 /* Too far in future, or reversed (interpretation is ambiguous) */
1077 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1080 /* Nasty start_seq wrap-around check (see comments above) */
1081 if (!before(start_seq, tp->snd_nxt))
1084 /* In outstanding window? ...This is valid exit for D-SACKs too.
1085 * start_seq == snd_una is non-sensical (see comments above)
1087 if (after(start_seq, tp->snd_una))
1090 if (!is_dsack || !tp->undo_marker)
1093 /* ...Then it's D-SACK, and must reside below snd_una completely */
1094 if (!after(end_seq, tp->snd_una))
1097 if (!before(start_seq, tp->undo_marker))
1101 if (!after(end_seq, tp->undo_marker))
1104 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1105 * start_seq < undo_marker and end_seq >= undo_marker.
1107 return !before(start_seq, end_seq - tp->max_window);
1110 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1111 * Event "C". Later note: FACK people cheated me again 8), we have to account
1112 * for reordering! Ugly, but should help.
1114 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1115 * less than what is now known to be received by the other end (derived from
1116 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1117 * retransmitted skbs to avoid some costly processing per ACKs.
1119 static void tcp_mark_lost_retrans(struct sock *sk)
1121 const struct inet_connection_sock *icsk = inet_csk(sk);
1122 struct tcp_sock *tp = tcp_sk(sk);
1123 struct sk_buff *skb;
1125 u32 new_low_seq = tp->snd_nxt;
1126 u32 received_upto = tcp_highest_sack_seq(tp);
1128 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1129 !after(received_upto, tp->lost_retrans_low) ||
1130 icsk->icsk_ca_state != TCP_CA_Recovery)
1133 tcp_for_write_queue(skb, sk) {
1134 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1136 if (skb == tcp_send_head(sk))
1138 if (cnt == tp->retrans_out)
1140 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1143 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1146 if (after(received_upto, ack_seq) &&
1148 !before(received_upto,
1149 ack_seq + tp->reordering * tp->mss_cache))) {
1150 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1151 tp->retrans_out -= tcp_skb_pcount(skb);
1153 /* clear lost hint */
1154 tp->retransmit_skb_hint = NULL;
1156 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1157 tp->lost_out += tcp_skb_pcount(skb);
1158 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1160 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1162 if (before(ack_seq, new_low_seq))
1163 new_low_seq = ack_seq;
1164 cnt += tcp_skb_pcount(skb);
1168 if (tp->retrans_out)
1169 tp->lost_retrans_low = new_low_seq;
1172 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1173 struct tcp_sack_block_wire *sp, int num_sacks,
1176 struct tcp_sock *tp = tcp_sk(sk);
1177 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1178 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1181 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1184 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1185 } else if (num_sacks > 1) {
1186 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1187 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1189 if (!after(end_seq_0, end_seq_1) &&
1190 !before(start_seq_0, start_seq_1)) {
1193 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1197 /* D-SACK for already forgotten data... Do dumb counting. */
1199 !after(end_seq_0, prior_snd_una) &&
1200 after(end_seq_0, tp->undo_marker))
1206 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1207 * the incoming SACK may not exactly match but we can find smaller MSS
1208 * aligned portion of it that matches. Therefore we might need to fragment
1209 * which may fail and creates some hassle (caller must handle error case
1212 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1213 u32 start_seq, u32 end_seq)
1216 unsigned int pkt_len;
1218 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1219 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1221 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1222 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1224 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1227 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1229 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1230 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1238 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1239 int *reord, int dup_sack, int fack_count)
1241 struct tcp_sock *tp = tcp_sk(sk);
1242 u8 sacked = TCP_SKB_CB(skb)->sacked;
1245 /* Account D-SACK for retransmitted packet. */
1246 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1247 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1249 if (sacked & TCPCB_SACKED_ACKED)
1250 *reord = min(fack_count, *reord);
1253 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1254 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1257 if (!(sacked & TCPCB_SACKED_ACKED)) {
1258 if (sacked & TCPCB_SACKED_RETRANS) {
1259 /* If the segment is not tagged as lost,
1260 * we do not clear RETRANS, believing
1261 * that retransmission is still in flight.
1263 if (sacked & TCPCB_LOST) {
1264 TCP_SKB_CB(skb)->sacked &=
1265 ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1266 tp->lost_out -= tcp_skb_pcount(skb);
1267 tp->retrans_out -= tcp_skb_pcount(skb);
1269 /* clear lost hint */
1270 tp->retransmit_skb_hint = NULL;
1273 if (!(sacked & TCPCB_RETRANS)) {
1274 /* New sack for not retransmitted frame,
1275 * which was in hole. It is reordering.
1277 if (before(TCP_SKB_CB(skb)->seq,
1278 tcp_highest_sack_seq(tp)))
1279 *reord = min(fack_count, *reord);
1281 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1282 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1283 flag |= FLAG_ONLY_ORIG_SACKED;
1286 if (sacked & TCPCB_LOST) {
1287 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1288 tp->lost_out -= tcp_skb_pcount(skb);
1290 /* clear lost hint */
1291 tp->retransmit_skb_hint = NULL;
1295 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1296 flag |= FLAG_DATA_SACKED;
1297 tp->sacked_out += tcp_skb_pcount(skb);
1299 fack_count += tcp_skb_pcount(skb);
1301 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1302 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1303 before(TCP_SKB_CB(skb)->seq,
1304 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1305 tp->lost_cnt_hint += tcp_skb_pcount(skb);
1307 if (fack_count > tp->fackets_out)
1308 tp->fackets_out = fack_count;
1310 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1311 tcp_advance_highest_sack(sk, skb);
1314 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1315 * frames and clear it. undo_retrans is decreased above, L|R frames
1316 * are accounted above as well.
1318 if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) {
1319 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1320 tp->retrans_out -= tcp_skb_pcount(skb);
1321 tp->retransmit_skb_hint = NULL;
1327 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1328 struct tcp_sack_block *next_dup,
1329 u32 start_seq, u32 end_seq,
1330 int dup_sack_in, int *fack_count,
1331 int *reord, int *flag)
1333 tcp_for_write_queue_from(skb, sk) {
1335 int dup_sack = dup_sack_in;
1337 if (skb == tcp_send_head(sk))
1340 /* queue is in-order => we can short-circuit the walk early */
1341 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1344 if ((next_dup != NULL) &&
1345 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1346 in_sack = tcp_match_skb_to_sack(sk, skb,
1347 next_dup->start_seq,
1354 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq,
1356 if (unlikely(in_sack < 0))
1360 *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack,
1363 *fack_count += tcp_skb_pcount(skb);
1368 /* Avoid all extra work that is being done by sacktag while walking in
1371 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1372 u32 skip_to_seq, int *fack_count)
1374 tcp_for_write_queue_from(skb, sk) {
1375 if (skb == tcp_send_head(sk))
1378 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1381 *fack_count += tcp_skb_pcount(skb);
1386 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1388 struct tcp_sack_block *next_dup,
1390 int *fack_count, int *reord,
1393 if (next_dup == NULL)
1396 if (before(next_dup->start_seq, skip_to_seq)) {
1397 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count);
1398 skb = tcp_sacktag_walk(skb, sk, NULL,
1399 next_dup->start_seq, next_dup->end_seq,
1400 1, fack_count, reord, flag);
1406 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1408 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1412 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1415 const struct inet_connection_sock *icsk = inet_csk(sk);
1416 struct tcp_sock *tp = tcp_sk(sk);
1417 unsigned char *ptr = (skb_transport_header(ack_skb) +
1418 TCP_SKB_CB(ack_skb)->sacked);
1419 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1420 struct tcp_sack_block sp[4];
1421 struct tcp_sack_block *cache;
1422 struct sk_buff *skb;
1423 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE) >> 3;
1425 int reord = tp->packets_out;
1427 int found_dup_sack = 0;
1430 int first_sack_index;
1432 if (!tp->sacked_out) {
1433 if (WARN_ON(tp->fackets_out))
1434 tp->fackets_out = 0;
1435 tcp_highest_sack_reset(sk);
1438 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1439 num_sacks, prior_snd_una);
1441 flag |= FLAG_DSACKING_ACK;
1443 /* Eliminate too old ACKs, but take into
1444 * account more or less fresh ones, they can
1445 * contain valid SACK info.
1447 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1450 if (!tp->packets_out)
1454 first_sack_index = 0;
1455 for (i = 0; i < num_sacks; i++) {
1456 int dup_sack = !i && found_dup_sack;
1458 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1459 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1461 if (!tcp_is_sackblock_valid(tp, dup_sack,
1462 sp[used_sacks].start_seq,
1463 sp[used_sacks].end_seq)) {
1467 if (!tp->undo_marker)
1468 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1470 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1472 /* Don't count olds caused by ACK reordering */
1473 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1474 !after(sp[used_sacks].end_seq, tp->snd_una))
1476 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1479 NET_INC_STATS_BH(mib_idx);
1481 first_sack_index = -1;
1485 /* Ignore very old stuff early */
1486 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1492 /* order SACK blocks to allow in order walk of the retrans queue */
1493 for (i = used_sacks - 1; i > 0; i--) {
1494 for (j = 0; j < i; j++) {
1495 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1496 struct tcp_sack_block tmp;
1502 /* Track where the first SACK block goes to */
1503 if (j == first_sack_index)
1504 first_sack_index = j + 1;
1509 skb = tcp_write_queue_head(sk);
1513 if (!tp->sacked_out) {
1514 /* It's already past, so skip checking against it */
1515 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1517 cache = tp->recv_sack_cache;
1518 /* Skip empty blocks in at head of the cache */
1519 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1524 while (i < used_sacks) {
1525 u32 start_seq = sp[i].start_seq;
1526 u32 end_seq = sp[i].end_seq;
1527 int dup_sack = (found_dup_sack && (i == first_sack_index));
1528 struct tcp_sack_block *next_dup = NULL;
1530 if (found_dup_sack && ((i + 1) == first_sack_index))
1531 next_dup = &sp[i + 1];
1533 /* Event "B" in the comment above. */
1534 if (after(end_seq, tp->high_seq))
1535 flag |= FLAG_DATA_LOST;
1537 /* Skip too early cached blocks */
1538 while (tcp_sack_cache_ok(tp, cache) &&
1539 !before(start_seq, cache->end_seq))
1542 /* Can skip some work by looking recv_sack_cache? */
1543 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1544 after(end_seq, cache->start_seq)) {
1547 if (before(start_seq, cache->start_seq)) {
1548 skb = tcp_sacktag_skip(skb, sk, start_seq,
1550 skb = tcp_sacktag_walk(skb, sk, next_dup,
1553 dup_sack, &fack_count,
1557 /* Rest of the block already fully processed? */
1558 if (!after(end_seq, cache->end_seq))
1561 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1563 &fack_count, &reord,
1566 /* ...tail remains todo... */
1567 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1568 /* ...but better entrypoint exists! */
1569 skb = tcp_highest_sack(sk);
1572 fack_count = tp->fackets_out;
1577 skb = tcp_sacktag_skip(skb, sk, cache->end_seq,
1579 /* Check overlap against next cached too (past this one already) */
1584 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1585 skb = tcp_highest_sack(sk);
1588 fack_count = tp->fackets_out;
1590 skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count);
1593 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1594 dup_sack, &fack_count, &reord, &flag);
1597 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1598 * due to in-order walk
1600 if (after(end_seq, tp->frto_highmark))
1601 flag &= ~FLAG_ONLY_ORIG_SACKED;
1606 /* Clear the head of the cache sack blocks so we can skip it next time */
1607 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1608 tp->recv_sack_cache[i].start_seq = 0;
1609 tp->recv_sack_cache[i].end_seq = 0;
1611 for (j = 0; j < used_sacks; j++)
1612 tp->recv_sack_cache[i++] = sp[j];
1614 tcp_mark_lost_retrans(sk);
1616 tcp_verify_left_out(tp);
1618 if ((reord < tp->fackets_out) &&
1619 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1620 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1621 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1625 #if FASTRETRANS_DEBUG > 0
1626 BUG_TRAP((int)tp->sacked_out >= 0);
1627 BUG_TRAP((int)tp->lost_out >= 0);
1628 BUG_TRAP((int)tp->retrans_out >= 0);
1629 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1634 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1635 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1637 int tcp_limit_reno_sacked(struct tcp_sock *tp)
1641 holes = max(tp->lost_out, 1U);
1642 holes = min(holes, tp->packets_out);
1644 if ((tp->sacked_out + holes) > tp->packets_out) {
1645 tp->sacked_out = tp->packets_out - holes;
1651 /* If we receive more dupacks than we expected counting segments
1652 * in assumption of absent reordering, interpret this as reordering.
1653 * The only another reason could be bug in receiver TCP.
1655 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1657 struct tcp_sock *tp = tcp_sk(sk);
1658 if (tcp_limit_reno_sacked(tp))
1659 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1662 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1664 static void tcp_add_reno_sack(struct sock *sk)
1666 struct tcp_sock *tp = tcp_sk(sk);
1668 tcp_check_reno_reordering(sk, 0);
1669 tcp_verify_left_out(tp);
1672 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1674 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1676 struct tcp_sock *tp = tcp_sk(sk);
1679 /* One ACK acked hole. The rest eat duplicate ACKs. */
1680 if (acked - 1 >= tp->sacked_out)
1683 tp->sacked_out -= acked - 1;
1685 tcp_check_reno_reordering(sk, acked);
1686 tcp_verify_left_out(tp);
1689 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1694 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1696 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1699 /* F-RTO can only be used if TCP has never retransmitted anything other than
1700 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1702 int tcp_use_frto(struct sock *sk)
1704 const struct tcp_sock *tp = tcp_sk(sk);
1705 const struct inet_connection_sock *icsk = inet_csk(sk);
1706 struct sk_buff *skb;
1708 if (!sysctl_tcp_frto)
1711 /* MTU probe and F-RTO won't really play nicely along currently */
1712 if (icsk->icsk_mtup.probe_size)
1715 if (tcp_is_sackfrto(tp))
1718 /* Avoid expensive walking of rexmit queue if possible */
1719 if (tp->retrans_out > 1)
1722 skb = tcp_write_queue_head(sk);
1723 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1724 tcp_for_write_queue_from(skb, sk) {
1725 if (skb == tcp_send_head(sk))
1727 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1729 /* Short-circuit when first non-SACKed skb has been checked */
1730 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1736 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1737 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1738 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1739 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1740 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1741 * bits are handled if the Loss state is really to be entered (in
1742 * tcp_enter_frto_loss).
1744 * Do like tcp_enter_loss() would; when RTO expires the second time it
1746 * "Reduce ssthresh if it has not yet been made inside this window."
1748 void tcp_enter_frto(struct sock *sk)
1750 const struct inet_connection_sock *icsk = inet_csk(sk);
1751 struct tcp_sock *tp = tcp_sk(sk);
1752 struct sk_buff *skb;
1754 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1755 tp->snd_una == tp->high_seq ||
1756 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1757 !icsk->icsk_retransmits)) {
1758 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1759 /* Our state is too optimistic in ssthresh() call because cwnd
1760 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1761 * recovery has not yet completed. Pattern would be this: RTO,
1762 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1764 * RFC4138 should be more specific on what to do, even though
1765 * RTO is quite unlikely to occur after the first Cumulative ACK
1766 * due to back-off and complexity of triggering events ...
1768 if (tp->frto_counter) {
1770 stored_cwnd = tp->snd_cwnd;
1772 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1773 tp->snd_cwnd = stored_cwnd;
1775 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1777 /* ... in theory, cong.control module could do "any tricks" in
1778 * ssthresh(), which means that ca_state, lost bits and lost_out
1779 * counter would have to be faked before the call occurs. We
1780 * consider that too expensive, unlikely and hacky, so modules
1781 * using these in ssthresh() must deal these incompatibility
1782 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1784 tcp_ca_event(sk, CA_EVENT_FRTO);
1787 tp->undo_marker = tp->snd_una;
1788 tp->undo_retrans = 0;
1790 skb = tcp_write_queue_head(sk);
1791 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1792 tp->undo_marker = 0;
1793 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1794 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1795 tp->retrans_out -= tcp_skb_pcount(skb);
1797 tcp_verify_left_out(tp);
1799 /* Too bad if TCP was application limited */
1800 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1802 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1803 * The last condition is necessary at least in tp->frto_counter case.
1805 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
1806 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1807 after(tp->high_seq, tp->snd_una)) {
1808 tp->frto_highmark = tp->high_seq;
1810 tp->frto_highmark = tp->snd_nxt;
1812 tcp_set_ca_state(sk, TCP_CA_Disorder);
1813 tp->high_seq = tp->snd_nxt;
1814 tp->frto_counter = 1;
1817 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1818 * which indicates that we should follow the traditional RTO recovery,
1819 * i.e. mark everything lost and do go-back-N retransmission.
1821 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1823 struct tcp_sock *tp = tcp_sk(sk);
1824 struct sk_buff *skb;
1827 tp->retrans_out = 0;
1828 if (tcp_is_reno(tp))
1829 tcp_reset_reno_sack(tp);
1831 tcp_for_write_queue(skb, sk) {
1832 if (skb == tcp_send_head(sk))
1835 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1837 * Count the retransmission made on RTO correctly (only when
1838 * waiting for the first ACK and did not get it)...
1840 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1841 /* For some reason this R-bit might get cleared? */
1842 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1843 tp->retrans_out += tcp_skb_pcount(skb);
1844 /* ...enter this if branch just for the first segment */
1845 flag |= FLAG_DATA_ACKED;
1847 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1848 tp->undo_marker = 0;
1849 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1852 /* Marking forward transmissions that were made after RTO lost
1853 * can cause unnecessary retransmissions in some scenarios,
1854 * SACK blocks will mitigate that in some but not in all cases.
1855 * We used to not mark them but it was causing break-ups with
1856 * receivers that do only in-order receival.
1858 * TODO: we could detect presence of such receiver and select
1859 * different behavior per flow.
1861 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1862 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1863 tp->lost_out += tcp_skb_pcount(skb);
1866 tcp_verify_left_out(tp);
1868 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1869 tp->snd_cwnd_cnt = 0;
1870 tp->snd_cwnd_stamp = tcp_time_stamp;
1871 tp->frto_counter = 0;
1872 tp->bytes_acked = 0;
1874 tp->reordering = min_t(unsigned int, tp->reordering,
1875 sysctl_tcp_reordering);
1876 tcp_set_ca_state(sk, TCP_CA_Loss);
1877 tp->high_seq = tp->snd_nxt;
1878 TCP_ECN_queue_cwr(tp);
1880 tcp_clear_retrans_hints_partial(tp);
1883 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1885 tp->retrans_out = 0;
1888 tp->undo_marker = 0;
1889 tp->undo_retrans = 0;
1892 void tcp_clear_retrans(struct tcp_sock *tp)
1894 tcp_clear_retrans_partial(tp);
1896 tp->fackets_out = 0;
1900 /* Enter Loss state. If "how" is not zero, forget all SACK information
1901 * and reset tags completely, otherwise preserve SACKs. If receiver
1902 * dropped its ofo queue, we will know this due to reneging detection.
1904 void tcp_enter_loss(struct sock *sk, int how)
1906 const struct inet_connection_sock *icsk = inet_csk(sk);
1907 struct tcp_sock *tp = tcp_sk(sk);
1908 struct sk_buff *skb;
1910 /* Reduce ssthresh if it has not yet been made inside this window. */
1911 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1912 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1913 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1914 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1915 tcp_ca_event(sk, CA_EVENT_LOSS);
1918 tp->snd_cwnd_cnt = 0;
1919 tp->snd_cwnd_stamp = tcp_time_stamp;
1921 tp->bytes_acked = 0;
1922 tcp_clear_retrans_partial(tp);
1924 if (tcp_is_reno(tp))
1925 tcp_reset_reno_sack(tp);
1928 /* Push undo marker, if it was plain RTO and nothing
1929 * was retransmitted. */
1930 tp->undo_marker = tp->snd_una;
1931 tcp_clear_retrans_hints_partial(tp);
1934 tp->fackets_out = 0;
1935 tcp_clear_all_retrans_hints(tp);
1938 tcp_for_write_queue(skb, sk) {
1939 if (skb == tcp_send_head(sk))
1942 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1943 tp->undo_marker = 0;
1944 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1945 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1946 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1947 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1948 tp->lost_out += tcp_skb_pcount(skb);
1951 tcp_verify_left_out(tp);
1953 tp->reordering = min_t(unsigned int, tp->reordering,
1954 sysctl_tcp_reordering);
1955 tcp_set_ca_state(sk, TCP_CA_Loss);
1956 tp->high_seq = tp->snd_nxt;
1957 TCP_ECN_queue_cwr(tp);
1958 /* Abort F-RTO algorithm if one is in progress */
1959 tp->frto_counter = 0;
1962 /* If ACK arrived pointing to a remembered SACK, it means that our
1963 * remembered SACKs do not reflect real state of receiver i.e.
1964 * receiver _host_ is heavily congested (or buggy).
1966 * Do processing similar to RTO timeout.
1968 static int tcp_check_sack_reneging(struct sock *sk, int flag)
1970 if (flag & FLAG_SACK_RENEGING) {
1971 struct inet_connection_sock *icsk = inet_csk(sk);
1972 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1974 tcp_enter_loss(sk, 1);
1975 icsk->icsk_retransmits++;
1976 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1977 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1978 icsk->icsk_rto, TCP_RTO_MAX);
1984 static inline int tcp_fackets_out(struct tcp_sock *tp)
1986 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1989 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1990 * counter when SACK is enabled (without SACK, sacked_out is used for
1993 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1994 * segments up to the highest received SACK block so far and holes in
1997 * With reordering, holes may still be in flight, so RFC3517 recovery
1998 * uses pure sacked_out (total number of SACKed segments) even though
1999 * it violates the RFC that uses duplicate ACKs, often these are equal
2000 * but when e.g. out-of-window ACKs or packet duplication occurs,
2001 * they differ. Since neither occurs due to loss, TCP should really
2004 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2006 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2009 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2011 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2014 static inline int tcp_head_timedout(struct sock *sk)
2016 struct tcp_sock *tp = tcp_sk(sk);
2018 return tp->packets_out &&
2019 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2022 /* Linux NewReno/SACK/FACK/ECN state machine.
2023 * --------------------------------------
2025 * "Open" Normal state, no dubious events, fast path.
2026 * "Disorder" In all the respects it is "Open",
2027 * but requires a bit more attention. It is entered when
2028 * we see some SACKs or dupacks. It is split of "Open"
2029 * mainly to move some processing from fast path to slow one.
2030 * "CWR" CWND was reduced due to some Congestion Notification event.
2031 * It can be ECN, ICMP source quench, local device congestion.
2032 * "Recovery" CWND was reduced, we are fast-retransmitting.
2033 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2035 * tcp_fastretrans_alert() is entered:
2036 * - each incoming ACK, if state is not "Open"
2037 * - when arrived ACK is unusual, namely:
2042 * Counting packets in flight is pretty simple.
2044 * in_flight = packets_out - left_out + retrans_out
2046 * packets_out is SND.NXT-SND.UNA counted in packets.
2048 * retrans_out is number of retransmitted segments.
2050 * left_out is number of segments left network, but not ACKed yet.
2052 * left_out = sacked_out + lost_out
2054 * sacked_out: Packets, which arrived to receiver out of order
2055 * and hence not ACKed. With SACKs this number is simply
2056 * amount of SACKed data. Even without SACKs
2057 * it is easy to give pretty reliable estimate of this number,
2058 * counting duplicate ACKs.
2060 * lost_out: Packets lost by network. TCP has no explicit
2061 * "loss notification" feedback from network (for now).
2062 * It means that this number can be only _guessed_.
2063 * Actually, it is the heuristics to predict lossage that
2064 * distinguishes different algorithms.
2066 * F.e. after RTO, when all the queue is considered as lost,
2067 * lost_out = packets_out and in_flight = retrans_out.
2069 * Essentially, we have now two algorithms counting
2072 * FACK: It is the simplest heuristics. As soon as we decided
2073 * that something is lost, we decide that _all_ not SACKed
2074 * packets until the most forward SACK are lost. I.e.
2075 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2076 * It is absolutely correct estimate, if network does not reorder
2077 * packets. And it loses any connection to reality when reordering
2078 * takes place. We use FACK by default until reordering
2079 * is suspected on the path to this destination.
2081 * NewReno: when Recovery is entered, we assume that one segment
2082 * is lost (classic Reno). While we are in Recovery and
2083 * a partial ACK arrives, we assume that one more packet
2084 * is lost (NewReno). This heuristics are the same in NewReno
2087 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2088 * deflation etc. CWND is real congestion window, never inflated, changes
2089 * only according to classic VJ rules.
2091 * Really tricky (and requiring careful tuning) part of algorithm
2092 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2093 * The first determines the moment _when_ we should reduce CWND and,
2094 * hence, slow down forward transmission. In fact, it determines the moment
2095 * when we decide that hole is caused by loss, rather than by a reorder.
2097 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2098 * holes, caused by lost packets.
2100 * And the most logically complicated part of algorithm is undo
2101 * heuristics. We detect false retransmits due to both too early
2102 * fast retransmit (reordering) and underestimated RTO, analyzing
2103 * timestamps and D-SACKs. When we detect that some segments were
2104 * retransmitted by mistake and CWND reduction was wrong, we undo
2105 * window reduction and abort recovery phase. This logic is hidden
2106 * inside several functions named tcp_try_undo_<something>.
2109 /* This function decides, when we should leave Disordered state
2110 * and enter Recovery phase, reducing congestion window.
2112 * Main question: may we further continue forward transmission
2113 * with the same cwnd?
2115 static int tcp_time_to_recover(struct sock *sk)
2117 struct tcp_sock *tp = tcp_sk(sk);
2120 /* Do not perform any recovery during F-RTO algorithm */
2121 if (tp->frto_counter)
2124 /* Trick#1: The loss is proven. */
2128 /* Not-A-Trick#2 : Classic rule... */
2129 if (tcp_dupack_heurestics(tp) > tp->reordering)
2132 /* Trick#3 : when we use RFC2988 timer restart, fast
2133 * retransmit can be triggered by timeout of queue head.
2135 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2138 /* Trick#4: It is still not OK... But will it be useful to delay
2141 packets_out = tp->packets_out;
2142 if (packets_out <= tp->reordering &&
2143 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2144 !tcp_may_send_now(sk)) {
2145 /* We have nothing to send. This connection is limited
2146 * either by receiver window or by application.
2154 /* RFC: This is from the original, I doubt that this is necessary at all:
2155 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2156 * retransmitted past LOST markings in the first place? I'm not fully sure
2157 * about undo and end of connection cases, which can cause R without L?
2159 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
2161 if ((tp->retransmit_skb_hint != NULL) &&
2162 before(TCP_SKB_CB(skb)->seq,
2163 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
2164 tp->retransmit_skb_hint = NULL;
2167 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2168 * is against sacked "cnt", otherwise it's against facked "cnt"
2170 static void tcp_mark_head_lost(struct sock *sk, int packets)
2172 struct tcp_sock *tp = tcp_sk(sk);
2173 struct sk_buff *skb;
2178 BUG_TRAP(packets <= tp->packets_out);
2179 if (tp->lost_skb_hint) {
2180 skb = tp->lost_skb_hint;
2181 cnt = tp->lost_cnt_hint;
2183 skb = tcp_write_queue_head(sk);
2187 tcp_for_write_queue_from(skb, sk) {
2188 if (skb == tcp_send_head(sk))
2190 /* TODO: do this better */
2191 /* this is not the most efficient way to do this... */
2192 tp->lost_skb_hint = skb;
2193 tp->lost_cnt_hint = cnt;
2195 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2199 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2200 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2201 cnt += tcp_skb_pcount(skb);
2203 if (cnt > packets) {
2204 if (tcp_is_sack(tp) || (oldcnt >= packets))
2207 mss = skb_shinfo(skb)->gso_size;
2208 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2214 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2215 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2216 tp->lost_out += tcp_skb_pcount(skb);
2217 tcp_verify_retransmit_hint(tp, skb);
2220 tcp_verify_left_out(tp);
2223 /* Account newly detected lost packet(s) */
2225 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2227 struct tcp_sock *tp = tcp_sk(sk);
2229 if (tcp_is_reno(tp)) {
2230 tcp_mark_head_lost(sk, 1);
2231 } else if (tcp_is_fack(tp)) {
2232 int lost = tp->fackets_out - tp->reordering;
2235 tcp_mark_head_lost(sk, lost);
2237 int sacked_upto = tp->sacked_out - tp->reordering;
2238 if (sacked_upto < fast_rexmit)
2239 sacked_upto = fast_rexmit;
2240 tcp_mark_head_lost(sk, sacked_upto);
2243 /* New heuristics: it is possible only after we switched
2244 * to restart timer each time when something is ACKed.
2245 * Hence, we can detect timed out packets during fast
2246 * retransmit without falling to slow start.
2248 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2249 struct sk_buff *skb;
2251 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2252 : tcp_write_queue_head(sk);
2254 tcp_for_write_queue_from(skb, sk) {
2255 if (skb == tcp_send_head(sk))
2257 if (!tcp_skb_timedout(sk, skb))
2260 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2261 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2262 tp->lost_out += tcp_skb_pcount(skb);
2263 tcp_verify_retransmit_hint(tp, skb);
2267 tp->scoreboard_skb_hint = skb;
2269 tcp_verify_left_out(tp);
2273 /* CWND moderation, preventing bursts due to too big ACKs
2274 * in dubious situations.
2276 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2278 tp->snd_cwnd = min(tp->snd_cwnd,
2279 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2280 tp->snd_cwnd_stamp = tcp_time_stamp;
2283 /* Lower bound on congestion window is slow start threshold
2284 * unless congestion avoidance choice decides to overide it.
2286 static inline u32 tcp_cwnd_min(const struct sock *sk)
2288 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2290 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2293 /* Decrease cwnd each second ack. */
2294 static void tcp_cwnd_down(struct sock *sk, int flag)
2296 struct tcp_sock *tp = tcp_sk(sk);
2297 int decr = tp->snd_cwnd_cnt + 1;
2299 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2300 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2301 tp->snd_cwnd_cnt = decr & 1;
2304 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2305 tp->snd_cwnd -= decr;
2307 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2308 tp->snd_cwnd_stamp = tcp_time_stamp;
2312 /* Nothing was retransmitted or returned timestamp is less
2313 * than timestamp of the first retransmission.
2315 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2317 return !tp->retrans_stamp ||
2318 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2319 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2322 /* Undo procedures. */
2324 #if FASTRETRANS_DEBUG > 1
2325 static void DBGUNDO(struct sock *sk, const char *msg)
2327 struct tcp_sock *tp = tcp_sk(sk);
2328 struct inet_sock *inet = inet_sk(sk);
2330 if (sk->sk_family == AF_INET) {
2331 printk(KERN_DEBUG "Undo %s " NIPQUAD_FMT "/%u c%u l%u ss%u/%u p%u\n",
2333 NIPQUAD(inet->daddr), ntohs(inet->dport),
2334 tp->snd_cwnd, tcp_left_out(tp),
2335 tp->snd_ssthresh, tp->prior_ssthresh,
2338 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2339 else if (sk->sk_family == AF_INET6) {
2340 struct ipv6_pinfo *np = inet6_sk(sk);
2341 printk(KERN_DEBUG "Undo %s " NIP6_FMT "/%u c%u l%u ss%u/%u p%u\n",
2343 NIP6(np->daddr), ntohs(inet->dport),
2344 tp->snd_cwnd, tcp_left_out(tp),
2345 tp->snd_ssthresh, tp->prior_ssthresh,
2351 #define DBGUNDO(x...) do { } while (0)
2354 static void tcp_undo_cwr(struct sock *sk, const int undo)
2356 struct tcp_sock *tp = tcp_sk(sk);
2358 if (tp->prior_ssthresh) {
2359 const struct inet_connection_sock *icsk = inet_csk(sk);
2361 if (icsk->icsk_ca_ops->undo_cwnd)
2362 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2364 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2366 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2367 tp->snd_ssthresh = tp->prior_ssthresh;
2368 TCP_ECN_withdraw_cwr(tp);
2371 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2373 tcp_moderate_cwnd(tp);
2374 tp->snd_cwnd_stamp = tcp_time_stamp;
2376 /* There is something screwy going on with the retrans hints after
2378 tcp_clear_all_retrans_hints(tp);
2381 static inline int tcp_may_undo(struct tcp_sock *tp)
2383 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2386 /* People celebrate: "We love our President!" */
2387 static int tcp_try_undo_recovery(struct sock *sk)
2389 struct tcp_sock *tp = tcp_sk(sk);
2391 if (tcp_may_undo(tp)) {
2394 /* Happy end! We did not retransmit anything
2395 * or our original transmission succeeded.
2397 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2398 tcp_undo_cwr(sk, 1);
2399 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2400 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2402 mib_idx = LINUX_MIB_TCPFULLUNDO;
2404 NET_INC_STATS_BH(mib_idx);
2405 tp->undo_marker = 0;
2407 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2408 /* Hold old state until something *above* high_seq
2409 * is ACKed. For Reno it is MUST to prevent false
2410 * fast retransmits (RFC2582). SACK TCP is safe. */
2411 tcp_moderate_cwnd(tp);
2414 tcp_set_ca_state(sk, TCP_CA_Open);
2418 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2419 static void tcp_try_undo_dsack(struct sock *sk)
2421 struct tcp_sock *tp = tcp_sk(sk);
2423 if (tp->undo_marker && !tp->undo_retrans) {
2424 DBGUNDO(sk, "D-SACK");
2425 tcp_undo_cwr(sk, 1);
2426 tp->undo_marker = 0;
2427 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2431 /* Undo during fast recovery after partial ACK. */
2433 static int tcp_try_undo_partial(struct sock *sk, int acked)
2435 struct tcp_sock *tp = tcp_sk(sk);
2436 /* Partial ACK arrived. Force Hoe's retransmit. */
2437 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2439 if (tcp_may_undo(tp)) {
2440 /* Plain luck! Hole if filled with delayed
2441 * packet, rather than with a retransmit.
2443 if (tp->retrans_out == 0)
2444 tp->retrans_stamp = 0;
2446 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2449 tcp_undo_cwr(sk, 0);
2450 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2452 /* So... Do not make Hoe's retransmit yet.
2453 * If the first packet was delayed, the rest
2454 * ones are most probably delayed as well.
2461 /* Undo during loss recovery after partial ACK. */
2462 static int tcp_try_undo_loss(struct sock *sk)
2464 struct tcp_sock *tp = tcp_sk(sk);
2466 if (tcp_may_undo(tp)) {
2467 struct sk_buff *skb;
2468 tcp_for_write_queue(skb, sk) {
2469 if (skb == tcp_send_head(sk))
2471 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2474 tcp_clear_all_retrans_hints(tp);
2476 DBGUNDO(sk, "partial loss");
2478 tcp_undo_cwr(sk, 1);
2479 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2480 inet_csk(sk)->icsk_retransmits = 0;
2481 tp->undo_marker = 0;
2482 if (tcp_is_sack(tp))
2483 tcp_set_ca_state(sk, TCP_CA_Open);
2489 static inline void tcp_complete_cwr(struct sock *sk)
2491 struct tcp_sock *tp = tcp_sk(sk);
2492 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2493 tp->snd_cwnd_stamp = tcp_time_stamp;
2494 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2497 static void tcp_try_keep_open(struct sock *sk)
2499 struct tcp_sock *tp = tcp_sk(sk);
2500 int state = TCP_CA_Open;
2502 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2503 state = TCP_CA_Disorder;
2505 if (inet_csk(sk)->icsk_ca_state != state) {
2506 tcp_set_ca_state(sk, state);
2507 tp->high_seq = tp->snd_nxt;
2511 static void tcp_try_to_open(struct sock *sk, int flag)
2513 struct tcp_sock *tp = tcp_sk(sk);
2515 tcp_verify_left_out(tp);
2517 if (!tp->frto_counter && tp->retrans_out == 0)
2518 tp->retrans_stamp = 0;
2520 if (flag & FLAG_ECE)
2521 tcp_enter_cwr(sk, 1);
2523 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2524 tcp_try_keep_open(sk);
2525 tcp_moderate_cwnd(tp);
2527 tcp_cwnd_down(sk, flag);
2531 static void tcp_mtup_probe_failed(struct sock *sk)
2533 struct inet_connection_sock *icsk = inet_csk(sk);
2535 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2536 icsk->icsk_mtup.probe_size = 0;
2539 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2541 struct tcp_sock *tp = tcp_sk(sk);
2542 struct inet_connection_sock *icsk = inet_csk(sk);
2544 /* FIXME: breaks with very large cwnd */
2545 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2546 tp->snd_cwnd = tp->snd_cwnd *
2547 tcp_mss_to_mtu(sk, tp->mss_cache) /
2548 icsk->icsk_mtup.probe_size;
2549 tp->snd_cwnd_cnt = 0;
2550 tp->snd_cwnd_stamp = tcp_time_stamp;
2551 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2553 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2554 icsk->icsk_mtup.probe_size = 0;
2555 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2558 /* Process an event, which can update packets-in-flight not trivially.
2559 * Main goal of this function is to calculate new estimate for left_out,
2560 * taking into account both packets sitting in receiver's buffer and
2561 * packets lost by network.
2563 * Besides that it does CWND reduction, when packet loss is detected
2564 * and changes state of machine.
2566 * It does _not_ decide what to send, it is made in function
2567 * tcp_xmit_retransmit_queue().
2569 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2571 struct inet_connection_sock *icsk = inet_csk(sk);
2572 struct tcp_sock *tp = tcp_sk(sk);
2573 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2574 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2575 (tcp_fackets_out(tp) > tp->reordering));
2576 int fast_rexmit = 0, mib_idx;
2578 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2580 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2581 tp->fackets_out = 0;
2583 /* Now state machine starts.
2584 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2585 if (flag & FLAG_ECE)
2586 tp->prior_ssthresh = 0;
2588 /* B. In all the states check for reneging SACKs. */
2589 if (tcp_check_sack_reneging(sk, flag))
2592 /* C. Process data loss notification, provided it is valid. */
2593 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2594 before(tp->snd_una, tp->high_seq) &&
2595 icsk->icsk_ca_state != TCP_CA_Open &&
2596 tp->fackets_out > tp->reordering) {
2597 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2598 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2601 /* D. Check consistency of the current state. */
2602 tcp_verify_left_out(tp);
2604 /* E. Check state exit conditions. State can be terminated
2605 * when high_seq is ACKed. */
2606 if (icsk->icsk_ca_state == TCP_CA_Open) {
2607 BUG_TRAP(tp->retrans_out == 0);
2608 tp->retrans_stamp = 0;
2609 } else if (!before(tp->snd_una, tp->high_seq)) {
2610 switch (icsk->icsk_ca_state) {
2612 icsk->icsk_retransmits = 0;
2613 if (tcp_try_undo_recovery(sk))
2618 /* CWR is to be held something *above* high_seq
2619 * is ACKed for CWR bit to reach receiver. */
2620 if (tp->snd_una != tp->high_seq) {
2621 tcp_complete_cwr(sk);
2622 tcp_set_ca_state(sk, TCP_CA_Open);
2626 case TCP_CA_Disorder:
2627 tcp_try_undo_dsack(sk);
2628 if (!tp->undo_marker ||
2629 /* For SACK case do not Open to allow to undo
2630 * catching for all duplicate ACKs. */
2631 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2632 tp->undo_marker = 0;
2633 tcp_set_ca_state(sk, TCP_CA_Open);
2637 case TCP_CA_Recovery:
2638 if (tcp_is_reno(tp))
2639 tcp_reset_reno_sack(tp);
2640 if (tcp_try_undo_recovery(sk))
2642 tcp_complete_cwr(sk);
2647 /* F. Process state. */
2648 switch (icsk->icsk_ca_state) {
2649 case TCP_CA_Recovery:
2650 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2651 if (tcp_is_reno(tp) && is_dupack)
2652 tcp_add_reno_sack(sk);
2654 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2657 if (flag & FLAG_DATA_ACKED)
2658 icsk->icsk_retransmits = 0;
2659 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2660 tcp_reset_reno_sack(tp);
2661 if (!tcp_try_undo_loss(sk)) {
2662 tcp_moderate_cwnd(tp);
2663 tcp_xmit_retransmit_queue(sk);
2666 if (icsk->icsk_ca_state != TCP_CA_Open)
2668 /* Loss is undone; fall through to processing in Open state. */
2670 if (tcp_is_reno(tp)) {
2671 if (flag & FLAG_SND_UNA_ADVANCED)
2672 tcp_reset_reno_sack(tp);
2674 tcp_add_reno_sack(sk);
2677 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2678 tcp_try_undo_dsack(sk);
2680 if (!tcp_time_to_recover(sk)) {
2681 tcp_try_to_open(sk, flag);
2685 /* MTU probe failure: don't reduce cwnd */
2686 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2687 icsk->icsk_mtup.probe_size &&
2688 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2689 tcp_mtup_probe_failed(sk);
2690 /* Restores the reduction we did in tcp_mtup_probe() */
2692 tcp_simple_retransmit(sk);
2696 /* Otherwise enter Recovery state */
2698 if (tcp_is_reno(tp))
2699 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2701 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2703 NET_INC_STATS_BH(mib_idx);
2705 tp->high_seq = tp->snd_nxt;
2706 tp->prior_ssthresh = 0;
2707 tp->undo_marker = tp->snd_una;
2708 tp->undo_retrans = tp->retrans_out;
2710 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2711 if (!(flag & FLAG_ECE))
2712 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2713 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2714 TCP_ECN_queue_cwr(tp);
2717 tp->bytes_acked = 0;
2718 tp->snd_cwnd_cnt = 0;
2719 tcp_set_ca_state(sk, TCP_CA_Recovery);
2723 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2724 tcp_update_scoreboard(sk, fast_rexmit);
2725 tcp_cwnd_down(sk, flag);
2726 tcp_xmit_retransmit_queue(sk);
2729 /* Read draft-ietf-tcplw-high-performance before mucking
2730 * with this code. (Supersedes RFC1323)
2732 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2734 /* RTTM Rule: A TSecr value received in a segment is used to
2735 * update the averaged RTT measurement only if the segment
2736 * acknowledges some new data, i.e., only if it advances the
2737 * left edge of the send window.
2739 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2740 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2742 * Changed: reset backoff as soon as we see the first valid sample.
2743 * If we do not, we get strongly overestimated rto. With timestamps
2744 * samples are accepted even from very old segments: f.e., when rtt=1
2745 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2746 * answer arrives rto becomes 120 seconds! If at least one of segments
2747 * in window is lost... Voila. --ANK (010210)
2749 struct tcp_sock *tp = tcp_sk(sk);
2750 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2751 tcp_rtt_estimator(sk, seq_rtt);
2753 inet_csk(sk)->icsk_backoff = 0;
2757 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2759 /* We don't have a timestamp. Can only use
2760 * packets that are not retransmitted to determine
2761 * rtt estimates. Also, we must not reset the
2762 * backoff for rto until we get a non-retransmitted
2763 * packet. This allows us to deal with a situation
2764 * where the network delay has increased suddenly.
2765 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2768 if (flag & FLAG_RETRANS_DATA_ACKED)
2771 tcp_rtt_estimator(sk, seq_rtt);
2773 inet_csk(sk)->icsk_backoff = 0;
2777 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2780 const struct tcp_sock *tp = tcp_sk(sk);
2781 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2782 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2783 tcp_ack_saw_tstamp(sk, flag);
2784 else if (seq_rtt >= 0)
2785 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2788 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2790 const struct inet_connection_sock *icsk = inet_csk(sk);
2791 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2792 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2795 /* Restart timer after forward progress on connection.
2796 * RFC2988 recommends to restart timer to now+rto.
2798 static void tcp_rearm_rto(struct sock *sk)
2800 struct tcp_sock *tp = tcp_sk(sk);
2802 if (!tp->packets_out) {
2803 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2805 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2806 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2810 /* If we get here, the whole TSO packet has not been acked. */
2811 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2813 struct tcp_sock *tp = tcp_sk(sk);
2816 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2818 packets_acked = tcp_skb_pcount(skb);
2819 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2821 packets_acked -= tcp_skb_pcount(skb);
2823 if (packets_acked) {
2824 BUG_ON(tcp_skb_pcount(skb) == 0);
2825 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2828 return packets_acked;
2831 /* Remove acknowledged frames from the retransmission queue. If our packet
2832 * is before the ack sequence we can discard it as it's confirmed to have
2833 * arrived at the other end.
2835 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
2837 struct tcp_sock *tp = tcp_sk(sk);
2838 const struct inet_connection_sock *icsk = inet_csk(sk);
2839 struct sk_buff *skb;
2840 u32 now = tcp_time_stamp;
2841 int fully_acked = 1;
2844 u32 reord = tp->packets_out;
2846 s32 ca_seq_rtt = -1;
2847 ktime_t last_ackt = net_invalid_timestamp();
2849 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2850 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2853 u8 sacked = scb->sacked;
2855 /* Determine how many packets and what bytes were acked, tso and else */
2856 if (after(scb->end_seq, tp->snd_una)) {
2857 if (tcp_skb_pcount(skb) == 1 ||
2858 !after(tp->snd_una, scb->seq))
2861 acked_pcount = tcp_tso_acked(sk, skb);
2866 end_seq = tp->snd_una;
2868 acked_pcount = tcp_skb_pcount(skb);
2869 end_seq = scb->end_seq;
2872 /* MTU probing checks */
2873 if (fully_acked && icsk->icsk_mtup.probe_size &&
2874 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2875 tcp_mtup_probe_success(sk, skb);
2878 if (sacked & TCPCB_RETRANS) {
2879 if (sacked & TCPCB_SACKED_RETRANS)
2880 tp->retrans_out -= acked_pcount;
2881 flag |= FLAG_RETRANS_DATA_ACKED;
2884 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
2885 flag |= FLAG_NONHEAD_RETRANS_ACKED;
2887 ca_seq_rtt = now - scb->when;
2888 last_ackt = skb->tstamp;
2890 seq_rtt = ca_seq_rtt;
2892 if (!(sacked & TCPCB_SACKED_ACKED))
2893 reord = min(pkts_acked, reord);
2896 if (sacked & TCPCB_SACKED_ACKED)
2897 tp->sacked_out -= acked_pcount;
2898 if (sacked & TCPCB_LOST)
2899 tp->lost_out -= acked_pcount;
2901 if (unlikely(tp->urg_mode && !before(end_seq, tp->snd_up)))
2904 tp->packets_out -= acked_pcount;
2905 pkts_acked += acked_pcount;
2907 /* Initial outgoing SYN's get put onto the write_queue
2908 * just like anything else we transmit. It is not
2909 * true data, and if we misinform our callers that
2910 * this ACK acks real data, we will erroneously exit
2911 * connection startup slow start one packet too
2912 * quickly. This is severely frowned upon behavior.
2914 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2915 flag |= FLAG_DATA_ACKED;
2917 flag |= FLAG_SYN_ACKED;
2918 tp->retrans_stamp = 0;
2924 tcp_unlink_write_queue(skb, sk);
2925 sk_wmem_free_skb(sk, skb);
2926 tcp_clear_all_retrans_hints(tp);
2929 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2930 flag |= FLAG_SACK_RENEGING;
2932 if (flag & FLAG_ACKED) {
2933 const struct tcp_congestion_ops *ca_ops
2934 = inet_csk(sk)->icsk_ca_ops;
2936 tcp_ack_update_rtt(sk, flag, seq_rtt);
2939 if (tcp_is_reno(tp)) {
2940 tcp_remove_reno_sacks(sk, pkts_acked);
2942 /* Non-retransmitted hole got filled? That's reordering */
2943 if (reord < prior_fackets)
2944 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2947 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2949 if (ca_ops->pkts_acked) {
2952 /* Is the ACK triggering packet unambiguous? */
2953 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2954 /* High resolution needed and available? */
2955 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2956 !ktime_equal(last_ackt,
2957 net_invalid_timestamp()))
2958 rtt_us = ktime_us_delta(ktime_get_real(),
2960 else if (ca_seq_rtt > 0)
2961 rtt_us = jiffies_to_usecs(ca_seq_rtt);
2964 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2968 #if FASTRETRANS_DEBUG > 0
2969 BUG_TRAP((int)tp->sacked_out >= 0);
2970 BUG_TRAP((int)tp->lost_out >= 0);
2971 BUG_TRAP((int)tp->retrans_out >= 0);
2972 if (!tp->packets_out && tcp_is_sack(tp)) {
2973 icsk = inet_csk(sk);
2975 printk(KERN_DEBUG "Leak l=%u %d\n",
2976 tp->lost_out, icsk->icsk_ca_state);
2979 if (tp->sacked_out) {
2980 printk(KERN_DEBUG "Leak s=%u %d\n",
2981 tp->sacked_out, icsk->icsk_ca_state);
2984 if (tp->retrans_out) {
2985 printk(KERN_DEBUG "Leak r=%u %d\n",
2986 tp->retrans_out, icsk->icsk_ca_state);
2987 tp->retrans_out = 0;
2994 static void tcp_ack_probe(struct sock *sk)
2996 const struct tcp_sock *tp = tcp_sk(sk);
2997 struct inet_connection_sock *icsk = inet_csk(sk);
2999 /* Was it a usable window open? */
3001 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3002 icsk->icsk_backoff = 0;
3003 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3004 /* Socket must be waked up by subsequent tcp_data_snd_check().
3005 * This function is not for random using!
3008 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3009 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3014 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3016 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3017 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3020 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3022 const struct tcp_sock *tp = tcp_sk(sk);
3023 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3024 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3027 /* Check that window update is acceptable.
3028 * The function assumes that snd_una<=ack<=snd_next.
3030 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3031 const u32 ack, const u32 ack_seq,
3034 return (after(ack, tp->snd_una) ||
3035 after(ack_seq, tp->snd_wl1) ||
3036 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3039 /* Update our send window.
3041 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3042 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3044 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3047 struct tcp_sock *tp = tcp_sk(sk);
3049 u32 nwin = ntohs(tcp_hdr(skb)->window);
3051 if (likely(!tcp_hdr(skb)->syn))
3052 nwin <<= tp->rx_opt.snd_wscale;
3054 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3055 flag |= FLAG_WIN_UPDATE;
3056 tcp_update_wl(tp, ack, ack_seq);
3058 if (tp->snd_wnd != nwin) {
3061 /* Note, it is the only place, where
3062 * fast path is recovered for sending TCP.
3065 tcp_fast_path_check(sk);
3067 if (nwin > tp->max_window) {
3068 tp->max_window = nwin;
3069 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3079 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3080 * continue in congestion avoidance.
3082 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3084 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3085 tp->snd_cwnd_cnt = 0;
3086 tp->bytes_acked = 0;
3087 TCP_ECN_queue_cwr(tp);
3088 tcp_moderate_cwnd(tp);
3091 /* A conservative spurious RTO response algorithm: reduce cwnd using
3092 * rate halving and continue in congestion avoidance.
3094 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3096 tcp_enter_cwr(sk, 0);
3099 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3101 if (flag & FLAG_ECE)
3102 tcp_ratehalving_spur_to_response(sk);
3104 tcp_undo_cwr(sk, 1);
3107 /* F-RTO spurious RTO detection algorithm (RFC4138)
3109 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3110 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3111 * window (but not to or beyond highest sequence sent before RTO):
3112 * On First ACK, send two new segments out.
3113 * On Second ACK, RTO was likely spurious. Do spurious response (response
3114 * algorithm is not part of the F-RTO detection algorithm
3115 * given in RFC4138 but can be selected separately).
3116 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3117 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3118 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3119 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3121 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3122 * original window even after we transmit two new data segments.
3125 * on first step, wait until first cumulative ACK arrives, then move to
3126 * the second step. In second step, the next ACK decides.
3128 * F-RTO is implemented (mainly) in four functions:
3129 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3130 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3131 * called when tcp_use_frto() showed green light
3132 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3133 * - tcp_enter_frto_loss() is called if there is not enough evidence
3134 * to prove that the RTO is indeed spurious. It transfers the control
3135 * from F-RTO to the conventional RTO recovery
3137 static int tcp_process_frto(struct sock *sk, int flag)
3139 struct tcp_sock *tp = tcp_sk(sk);
3141 tcp_verify_left_out(tp);
3143 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3144 if (flag & FLAG_DATA_ACKED)
3145 inet_csk(sk)->icsk_retransmits = 0;
3147 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3148 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3149 tp->undo_marker = 0;
3151 if (!before(tp->snd_una, tp->frto_highmark)) {
3152 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3156 if (!tcp_is_sackfrto(tp)) {
3157 /* RFC4138 shortcoming in step 2; should also have case c):
3158 * ACK isn't duplicate nor advances window, e.g., opposite dir
3161 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3164 if (!(flag & FLAG_DATA_ACKED)) {
3165 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3170 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3171 /* Prevent sending of new data. */
3172 tp->snd_cwnd = min(tp->snd_cwnd,
3173 tcp_packets_in_flight(tp));
3177 if ((tp->frto_counter >= 2) &&
3178 (!(flag & FLAG_FORWARD_PROGRESS) ||
3179 ((flag & FLAG_DATA_SACKED) &&
3180 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3181 /* RFC4138 shortcoming (see comment above) */
3182 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3183 (flag & FLAG_NOT_DUP))
3186 tcp_enter_frto_loss(sk, 3, flag);
3191 if (tp->frto_counter == 1) {
3192 /* tcp_may_send_now needs to see updated state */
3193 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3194 tp->frto_counter = 2;
3196 if (!tcp_may_send_now(sk))
3197 tcp_enter_frto_loss(sk, 2, flag);
3201 switch (sysctl_tcp_frto_response) {
3203 tcp_undo_spur_to_response(sk, flag);
3206 tcp_conservative_spur_to_response(tp);
3209 tcp_ratehalving_spur_to_response(sk);
3212 tp->frto_counter = 0;
3213 tp->undo_marker = 0;
3214 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
3219 /* This routine deals with incoming acks, but not outgoing ones. */
3220 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3222 struct inet_connection_sock *icsk = inet_csk(sk);
3223 struct tcp_sock *tp = tcp_sk(sk);
3224 u32 prior_snd_una = tp->snd_una;
3225 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3226 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3227 u32 prior_in_flight;
3232 /* If the ack is newer than sent or older than previous acks
3233 * then we can probably ignore it.
3235 if (after(ack, tp->snd_nxt))
3236 goto uninteresting_ack;
3238 if (before(ack, prior_snd_una))
3241 if (after(ack, prior_snd_una))
3242 flag |= FLAG_SND_UNA_ADVANCED;
3244 if (sysctl_tcp_abc) {
3245 if (icsk->icsk_ca_state < TCP_CA_CWR)
3246 tp->bytes_acked += ack - prior_snd_una;
3247 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3248 /* we assume just one segment left network */
3249 tp->bytes_acked += min(ack - prior_snd_una,
3253 prior_fackets = tp->fackets_out;
3254 prior_in_flight = tcp_packets_in_flight(tp);
3256 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3257 /* Window is constant, pure forward advance.
3258 * No more checks are required.
3259 * Note, we use the fact that SND.UNA>=SND.WL2.
3261 tcp_update_wl(tp, ack, ack_seq);
3263 flag |= FLAG_WIN_UPDATE;
3265 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3267 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3269 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3272 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3274 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3276 if (TCP_SKB_CB(skb)->sacked)
3277 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3279 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3282 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3285 /* We passed data and got it acked, remove any soft error
3286 * log. Something worked...
3288 sk->sk_err_soft = 0;
3289 tp->rcv_tstamp = tcp_time_stamp;
3290 prior_packets = tp->packets_out;
3294 /* See if we can take anything off of the retransmit queue. */
3295 flag |= tcp_clean_rtx_queue(sk, prior_fackets);
3297 if (tp->frto_counter)
3298 frto_cwnd = tcp_process_frto(sk, flag);
3299 /* Guarantee sacktag reordering detection against wrap-arounds */
3300 if (before(tp->frto_highmark, tp->snd_una))
3301 tp->frto_highmark = 0;
3303 if (tcp_ack_is_dubious(sk, flag)) {
3304 /* Advance CWND, if state allows this. */
3305 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3306 tcp_may_raise_cwnd(sk, flag))
3307 tcp_cong_avoid(sk, ack, prior_in_flight);
3308 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3311 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3312 tcp_cong_avoid(sk, ack, prior_in_flight);
3315 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3316 dst_confirm(sk->sk_dst_cache);
3321 icsk->icsk_probes_out = 0;
3323 /* If this ack opens up a zero window, clear backoff. It was
3324 * being used to time the probes, and is probably far higher than
3325 * it needs to be for normal retransmission.
3327 if (tcp_send_head(sk))
3332 if (TCP_SKB_CB(skb)->sacked) {
3333 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3334 if (icsk->icsk_ca_state == TCP_CA_Open)
3335 tcp_try_keep_open(sk);
3339 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3343 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3344 * But, this can also be called on packets in the established flow when
3345 * the fast version below fails.
3347 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3351 struct tcphdr *th = tcp_hdr(skb);
3352 int length = (th->doff * 4) - sizeof(struct tcphdr);
3354 ptr = (unsigned char *)(th + 1);
3355 opt_rx->saw_tstamp = 0;
3357 while (length > 0) {
3358 int opcode = *ptr++;
3364 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3369 if (opsize < 2) /* "silly options" */
3371 if (opsize > length)
3372 return; /* don't parse partial options */
3375 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3376 u16 in_mss = get_unaligned_be16(ptr);
3378 if (opt_rx->user_mss &&
3379 opt_rx->user_mss < in_mss)
3380 in_mss = opt_rx->user_mss;
3381 opt_rx->mss_clamp = in_mss;
3386 if (opsize == TCPOLEN_WINDOW && th->syn &&
3387 !estab && sysctl_tcp_window_scaling) {
3388 __u8 snd_wscale = *(__u8 *)ptr;
3389 opt_rx->wscale_ok = 1;
3390 if (snd_wscale > 14) {
3391 if (net_ratelimit())
3392 printk(KERN_INFO "tcp_parse_options: Illegal window "
3393 "scaling value %d >14 received.\n",
3397 opt_rx->snd_wscale = snd_wscale;
3400 case TCPOPT_TIMESTAMP:
3401 if ((opsize == TCPOLEN_TIMESTAMP) &&
3402 ((estab && opt_rx->tstamp_ok) ||
3403 (!estab && sysctl_tcp_timestamps))) {
3404 opt_rx->saw_tstamp = 1;
3405 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3406 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3409 case TCPOPT_SACK_PERM:
3410 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3411 !estab && sysctl_tcp_sack) {
3412 opt_rx->sack_ok = 1;
3413 tcp_sack_reset(opt_rx);
3418 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3419 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3421 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3424 #ifdef CONFIG_TCP_MD5SIG
3427 * The MD5 Hash has already been
3428 * checked (see tcp_v{4,6}_do_rcv()).
3440 /* Fast parse options. This hopes to only see timestamps.
3441 * If it is wrong it falls back on tcp_parse_options().
3443 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3444 struct tcp_sock *tp)
3446 if (th->doff == sizeof(struct tcphdr) >> 2) {
3447 tp->rx_opt.saw_tstamp = 0;
3449 } else if (tp->rx_opt.tstamp_ok &&
3450 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3451 __be32 *ptr = (__be32 *)(th + 1);
3452 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3453 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3454 tp->rx_opt.saw_tstamp = 1;
3456 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3458 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3462 tcp_parse_options(skb, &tp->rx_opt, 1);
3466 #ifdef CONFIG_TCP_MD5SIG
3468 * Parse MD5 Signature option
3470 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3472 int length = (th->doff << 2) - sizeof (*th);
3473 u8 *ptr = (u8*)(th + 1);
3475 /* If the TCP option is too short, we can short cut */
3476 if (length < TCPOLEN_MD5SIG)
3479 while (length > 0) {
3480 int opcode = *ptr++;
3491 if (opsize < 2 || opsize > length)
3493 if (opcode == TCPOPT_MD5SIG)
3503 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3505 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3506 tp->rx_opt.ts_recent_stamp = get_seconds();
3509 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3511 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3512 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3513 * extra check below makes sure this can only happen
3514 * for pure ACK frames. -DaveM
3516 * Not only, also it occurs for expired timestamps.
3519 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3520 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3521 tcp_store_ts_recent(tp);
3525 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3527 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3528 * it can pass through stack. So, the following predicate verifies that
3529 * this segment is not used for anything but congestion avoidance or
3530 * fast retransmit. Moreover, we even are able to eliminate most of such
3531 * second order effects, if we apply some small "replay" window (~RTO)
3532 * to timestamp space.
3534 * All these measures still do not guarantee that we reject wrapped ACKs
3535 * on networks with high bandwidth, when sequence space is recycled fastly,
3536 * but it guarantees that such events will be very rare and do not affect
3537 * connection seriously. This doesn't look nice, but alas, PAWS is really
3540 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3541 * states that events when retransmit arrives after original data are rare.
3542 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3543 * the biggest problem on large power networks even with minor reordering.
3544 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3545 * up to bandwidth of 18Gigabit/sec. 8) ]
3548 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3550 struct tcp_sock *tp = tcp_sk(sk);
3551 struct tcphdr *th = tcp_hdr(skb);
3552 u32 seq = TCP_SKB_CB(skb)->seq;
3553 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3555 return (/* 1. Pure ACK with correct sequence number. */
3556 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3558 /* 2. ... and duplicate ACK. */
3559 ack == tp->snd_una &&
3561 /* 3. ... and does not update window. */
3562 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3564 /* 4. ... and sits in replay window. */
3565 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3568 static inline int tcp_paws_discard(const struct sock *sk,
3569 const struct sk_buff *skb)
3571 const struct tcp_sock *tp = tcp_sk(sk);
3572 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3573 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3574 !tcp_disordered_ack(sk, skb));
3577 /* Check segment sequence number for validity.
3579 * Segment controls are considered valid, if the segment
3580 * fits to the window after truncation to the window. Acceptability
3581 * of data (and SYN, FIN, of course) is checked separately.
3582 * See tcp_data_queue(), for example.
3584 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3585 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3586 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3587 * (borrowed from freebsd)
3590 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3592 return !before(end_seq, tp->rcv_wup) &&
3593 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3596 /* When we get a reset we do this. */
3597 static void tcp_reset(struct sock *sk)
3599 /* We want the right error as BSD sees it (and indeed as we do). */
3600 switch (sk->sk_state) {
3602 sk->sk_err = ECONNREFUSED;
3604 case TCP_CLOSE_WAIT:
3610 sk->sk_err = ECONNRESET;
3613 if (!sock_flag(sk, SOCK_DEAD))
3614 sk->sk_error_report(sk);
3620 * Process the FIN bit. This now behaves as it is supposed to work
3621 * and the FIN takes effect when it is validly part of sequence
3622 * space. Not before when we get holes.
3624 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3625 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3628 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3629 * close and we go into CLOSING (and later onto TIME-WAIT)
3631 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3633 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3635 struct tcp_sock *tp = tcp_sk(sk);
3637 inet_csk_schedule_ack(sk);
3639 sk->sk_shutdown |= RCV_SHUTDOWN;
3640 sock_set_flag(sk, SOCK_DONE);
3642 switch (sk->sk_state) {
3644 case TCP_ESTABLISHED:
3645 /* Move to CLOSE_WAIT */
3646 tcp_set_state(sk, TCP_CLOSE_WAIT);
3647 inet_csk(sk)->icsk_ack.pingpong = 1;
3650 case TCP_CLOSE_WAIT:
3652 /* Received a retransmission of the FIN, do
3657 /* RFC793: Remain in the LAST-ACK state. */
3661 /* This case occurs when a simultaneous close
3662 * happens, we must ack the received FIN and
3663 * enter the CLOSING state.
3666 tcp_set_state(sk, TCP_CLOSING);
3669 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3671 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3674 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3675 * cases we should never reach this piece of code.
3677 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3678 __func__, sk->sk_state);
3682 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3683 * Probably, we should reset in this case. For now drop them.
3685 __skb_queue_purge(&tp->out_of_order_queue);
3686 if (tcp_is_sack(tp))
3687 tcp_sack_reset(&tp->rx_opt);
3690 if (!sock_flag(sk, SOCK_DEAD)) {
3691 sk->sk_state_change(sk);
3693 /* Do not send POLL_HUP for half duplex close. */
3694 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3695 sk->sk_state == TCP_CLOSE)
3696 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3698 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3702 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3705 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3706 if (before(seq, sp->start_seq))
3707 sp->start_seq = seq;
3708 if (after(end_seq, sp->end_seq))
3709 sp->end_seq = end_seq;
3715 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3717 struct tcp_sock *tp = tcp_sk(sk);
3719 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3722 if (before(seq, tp->rcv_nxt))
3723 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3725 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3727 NET_INC_STATS_BH(mib_idx);
3729 tp->rx_opt.dsack = 1;
3730 tp->duplicate_sack[0].start_seq = seq;
3731 tp->duplicate_sack[0].end_seq = end_seq;
3732 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1,
3733 4 - tp->rx_opt.tstamp_ok);
3737 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3739 struct tcp_sock *tp = tcp_sk(sk);
3741 if (!tp->rx_opt.dsack)
3742 tcp_dsack_set(sk, seq, end_seq);
3744 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3747 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3749 struct tcp_sock *tp = tcp_sk(sk);
3751 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3752 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3753 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3754 tcp_enter_quickack_mode(sk);
3756 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3757 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3759 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3760 end_seq = tp->rcv_nxt;
3761 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3768 /* These routines update the SACK block as out-of-order packets arrive or
3769 * in-order packets close up the sequence space.
3771 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3774 struct tcp_sack_block *sp = &tp->selective_acks[0];
3775 struct tcp_sack_block *swalk = sp + 1;
3777 /* See if the recent change to the first SACK eats into
3778 * or hits the sequence space of other SACK blocks, if so coalesce.
3780 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3781 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3784 /* Zap SWALK, by moving every further SACK up by one slot.
3785 * Decrease num_sacks.
3787 tp->rx_opt.num_sacks--;
3788 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks +
3790 4 - tp->rx_opt.tstamp_ok);
3791 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3795 this_sack++, swalk++;
3799 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
3800 struct tcp_sack_block *sack2)
3804 tmp = sack1->start_seq;
3805 sack1->start_seq = sack2->start_seq;
3806 sack2->start_seq = tmp;
3808 tmp = sack1->end_seq;
3809 sack1->end_seq = sack2->end_seq;
3810 sack2->end_seq = tmp;
3813 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3815 struct tcp_sock *tp = tcp_sk(sk);
3816 struct tcp_sack_block *sp = &tp->selective_acks[0];
3817 int cur_sacks = tp->rx_opt.num_sacks;
3823 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3824 if (tcp_sack_extend(sp, seq, end_seq)) {
3825 /* Rotate this_sack to the first one. */
3826 for (; this_sack > 0; this_sack--, sp--)
3827 tcp_sack_swap(sp, sp - 1);
3829 tcp_sack_maybe_coalesce(tp);
3834 /* Could not find an adjacent existing SACK, build a new one,
3835 * put it at the front, and shift everyone else down. We
3836 * always know there is at least one SACK present already here.
3838 * If the sack array is full, forget about the last one.
3840 if (this_sack >= 4) {
3842 tp->rx_opt.num_sacks--;
3845 for (; this_sack > 0; this_sack--, sp--)
3849 /* Build the new head SACK, and we're done. */
3850 sp->start_seq = seq;
3851 sp->end_seq = end_seq;
3852 tp->rx_opt.num_sacks++;
3853 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack,
3854 4 - tp->rx_opt.tstamp_ok);
3857 /* RCV.NXT advances, some SACKs should be eaten. */
3859 static void tcp_sack_remove(struct tcp_sock *tp)
3861 struct tcp_sack_block *sp = &tp->selective_acks[0];
3862 int num_sacks = tp->rx_opt.num_sacks;
3865 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3866 if (skb_queue_empty(&tp->out_of_order_queue)) {
3867 tp->rx_opt.num_sacks = 0;
3868 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3872 for (this_sack = 0; this_sack < num_sacks;) {
3873 /* Check if the start of the sack is covered by RCV.NXT. */
3874 if (!before(tp->rcv_nxt, sp->start_seq)) {
3877 /* RCV.NXT must cover all the block! */
3878 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3880 /* Zap this SACK, by moving forward any other SACKS. */
3881 for (i=this_sack+1; i < num_sacks; i++)
3882 tp->selective_acks[i-1] = tp->selective_acks[i];
3889 if (num_sacks != tp->rx_opt.num_sacks) {
3890 tp->rx_opt.num_sacks = num_sacks;
3891 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks +
3893 4 - tp->rx_opt.tstamp_ok);
3897 /* This one checks to see if we can put data from the
3898 * out_of_order queue into the receive_queue.
3900 static void tcp_ofo_queue(struct sock *sk)
3902 struct tcp_sock *tp = tcp_sk(sk);
3903 __u32 dsack_high = tp->rcv_nxt;
3904 struct sk_buff *skb;
3906 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3907 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3910 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3911 __u32 dsack = dsack_high;
3912 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3913 dsack_high = TCP_SKB_CB(skb)->end_seq;
3914 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
3917 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3918 SOCK_DEBUG(sk, "ofo packet was already received \n");
3919 __skb_unlink(skb, &tp->out_of_order_queue);
3923 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3924 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3925 TCP_SKB_CB(skb)->end_seq);
3927 __skb_unlink(skb, &tp->out_of_order_queue);
3928 __skb_queue_tail(&sk->sk_receive_queue, skb);
3929 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3930 if (tcp_hdr(skb)->fin)
3931 tcp_fin(skb, sk, tcp_hdr(skb));
3935 static int tcp_prune_ofo_queue(struct sock *sk);
3936 static int tcp_prune_queue(struct sock *sk);
3938 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
3940 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3941 !sk_rmem_schedule(sk, size)) {
3943 if (tcp_prune_queue(sk) < 0)
3946 if (!sk_rmem_schedule(sk, size)) {
3947 if (!tcp_prune_ofo_queue(sk))
3950 if (!sk_rmem_schedule(sk, size))
3957 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3959 struct tcphdr *th = tcp_hdr(skb);
3960 struct tcp_sock *tp = tcp_sk(sk);
3963 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3966 __skb_pull(skb, th->doff * 4);
3968 TCP_ECN_accept_cwr(tp, skb);
3970 if (tp->rx_opt.dsack) {
3971 tp->rx_opt.dsack = 0;
3972 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3973 4 - tp->rx_opt.tstamp_ok);
3976 /* Queue data for delivery to the user.
3977 * Packets in sequence go to the receive queue.
3978 * Out of sequence packets to the out_of_order_queue.
3980 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3981 if (tcp_receive_window(tp) == 0)
3984 /* Ok. In sequence. In window. */
3985 if (tp->ucopy.task == current &&
3986 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3987 sock_owned_by_user(sk) && !tp->urg_data) {
3988 int chunk = min_t(unsigned int, skb->len,
3991 __set_current_state(TASK_RUNNING);
3994 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3995 tp->ucopy.len -= chunk;
3996 tp->copied_seq += chunk;
3997 eaten = (chunk == skb->len && !th->fin);
3998 tcp_rcv_space_adjust(sk);
4006 tcp_try_rmem_schedule(sk, skb->truesize))
4009 skb_set_owner_r(skb, sk);
4010 __skb_queue_tail(&sk->sk_receive_queue, skb);
4012 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4014 tcp_event_data_recv(sk, skb);
4016 tcp_fin(skb, sk, th);
4018 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4021 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4022 * gap in queue is filled.
4024 if (skb_queue_empty(&tp->out_of_order_queue))
4025 inet_csk(sk)->icsk_ack.pingpong = 0;
4028 if (tp->rx_opt.num_sacks)
4029 tcp_sack_remove(tp);
4031 tcp_fast_path_check(sk);
4035 else if (!sock_flag(sk, SOCK_DEAD))
4036 sk->sk_data_ready(sk, 0);
4040 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4041 /* A retransmit, 2nd most common case. Force an immediate ack. */
4042 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
4043 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4046 tcp_enter_quickack_mode(sk);
4047 inet_csk_schedule_ack(sk);
4053 /* Out of window. F.e. zero window probe. */
4054 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4057 tcp_enter_quickack_mode(sk);
4059 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4060 /* Partial packet, seq < rcv_next < end_seq */
4061 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4062 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4063 TCP_SKB_CB(skb)->end_seq);
4065 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4067 /* If window is closed, drop tail of packet. But after
4068 * remembering D-SACK for its head made in previous line.
4070 if (!tcp_receive_window(tp))
4075 TCP_ECN_check_ce(tp, skb);
4077 if (tcp_try_rmem_schedule(sk, skb->truesize))
4080 /* Disable header prediction. */
4082 inet_csk_schedule_ack(sk);
4084 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4085 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4087 skb_set_owner_r(skb, sk);
4089 if (!skb_peek(&tp->out_of_order_queue)) {
4090 /* Initial out of order segment, build 1 SACK. */
4091 if (tcp_is_sack(tp)) {
4092 tp->rx_opt.num_sacks = 1;
4093 tp->rx_opt.dsack = 0;
4094 tp->rx_opt.eff_sacks = 1;
4095 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4096 tp->selective_acks[0].end_seq =
4097 TCP_SKB_CB(skb)->end_seq;
4099 __skb_queue_head(&tp->out_of_order_queue, skb);
4101 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4102 u32 seq = TCP_SKB_CB(skb)->seq;
4103 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4105 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4106 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4108 if (!tp->rx_opt.num_sacks ||
4109 tp->selective_acks[0].end_seq != seq)
4112 /* Common case: data arrive in order after hole. */
4113 tp->selective_acks[0].end_seq = end_seq;
4117 /* Find place to insert this segment. */
4119 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4121 } while ((skb1 = skb1->prev) !=
4122 (struct sk_buff *)&tp->out_of_order_queue);
4124 /* Do skb overlap to previous one? */
4125 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4126 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4127 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4128 /* All the bits are present. Drop. */
4130 tcp_dsack_set(sk, seq, end_seq);
4133 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4134 /* Partial overlap. */
4135 tcp_dsack_set(sk, seq,
4136 TCP_SKB_CB(skb1)->end_seq);
4141 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
4143 /* And clean segments covered by new one as whole. */
4144 while ((skb1 = skb->next) !=
4145 (struct sk_buff *)&tp->out_of_order_queue &&
4146 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4147 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4148 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4152 __skb_unlink(skb1, &tp->out_of_order_queue);
4153 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4154 TCP_SKB_CB(skb1)->end_seq);
4159 if (tcp_is_sack(tp))
4160 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4164 /* Collapse contiguous sequence of skbs head..tail with
4165 * sequence numbers start..end.
4166 * Segments with FIN/SYN are not collapsed (only because this
4170 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4171 struct sk_buff *head, struct sk_buff *tail,
4174 struct sk_buff *skb;
4176 /* First, check that queue is collapsible and find
4177 * the point where collapsing can be useful. */
4178 for (skb = head; skb != tail;) {
4179 /* No new bits? It is possible on ofo queue. */
4180 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4181 struct sk_buff *next = skb->next;
4182 __skb_unlink(skb, list);
4184 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4189 /* The first skb to collapse is:
4191 * - bloated or contains data before "start" or
4192 * overlaps to the next one.
4194 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4195 (tcp_win_from_space(skb->truesize) > skb->len ||
4196 before(TCP_SKB_CB(skb)->seq, start) ||
4197 (skb->next != tail &&
4198 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4201 /* Decided to skip this, advance start seq. */
4202 start = TCP_SKB_CB(skb)->end_seq;
4205 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4208 while (before(start, end)) {
4209 struct sk_buff *nskb;
4210 unsigned int header = skb_headroom(skb);
4211 int copy = SKB_MAX_ORDER(header, 0);
4213 /* Too big header? This can happen with IPv6. */
4216 if (end - start < copy)
4218 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4222 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4223 skb_set_network_header(nskb, (skb_network_header(skb) -
4225 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4227 skb_reserve(nskb, header);
4228 memcpy(nskb->head, skb->head, header);
4229 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4230 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4231 __skb_insert(nskb, skb->prev, skb, list);
4232 skb_set_owner_r(nskb, sk);
4234 /* Copy data, releasing collapsed skbs. */
4236 int offset = start - TCP_SKB_CB(skb)->seq;
4237 int size = TCP_SKB_CB(skb)->end_seq - start;
4241 size = min(copy, size);
4242 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4244 TCP_SKB_CB(nskb)->end_seq += size;
4248 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4249 struct sk_buff *next = skb->next;
4250 __skb_unlink(skb, list);
4252 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4255 tcp_hdr(skb)->syn ||
4263 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4264 * and tcp_collapse() them until all the queue is collapsed.
4266 static void tcp_collapse_ofo_queue(struct sock *sk)
4268 struct tcp_sock *tp = tcp_sk(sk);
4269 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4270 struct sk_buff *head;
4276 start = TCP_SKB_CB(skb)->seq;
4277 end = TCP_SKB_CB(skb)->end_seq;
4283 /* Segment is terminated when we see gap or when
4284 * we are at the end of all the queue. */
4285 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4286 after(TCP_SKB_CB(skb)->seq, end) ||
4287 before(TCP_SKB_CB(skb)->end_seq, start)) {
4288 tcp_collapse(sk, &tp->out_of_order_queue,
4289 head, skb, start, end);
4291 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4293 /* Start new segment */
4294 start = TCP_SKB_CB(skb)->seq;
4295 end = TCP_SKB_CB(skb)->end_seq;
4297 if (before(TCP_SKB_CB(skb)->seq, start))
4298 start = TCP_SKB_CB(skb)->seq;
4299 if (after(TCP_SKB_CB(skb)->end_seq, end))
4300 end = TCP_SKB_CB(skb)->end_seq;
4306 * Purge the out-of-order queue.
4307 * Return true if queue was pruned.
4309 static int tcp_prune_ofo_queue(struct sock *sk)
4311 struct tcp_sock *tp = tcp_sk(sk);
4314 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4315 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
4316 __skb_queue_purge(&tp->out_of_order_queue);
4318 /* Reset SACK state. A conforming SACK implementation will
4319 * do the same at a timeout based retransmit. When a connection
4320 * is in a sad state like this, we care only about integrity
4321 * of the connection not performance.
4323 if (tp->rx_opt.sack_ok)
4324 tcp_sack_reset(&tp->rx_opt);
4331 /* Reduce allocated memory if we can, trying to get
4332 * the socket within its memory limits again.
4334 * Return less than zero if we should start dropping frames
4335 * until the socket owning process reads some of the data
4336 * to stabilize the situation.
4338 static int tcp_prune_queue(struct sock *sk)
4340 struct tcp_sock *tp = tcp_sk(sk);
4342 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4344 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
4346 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4347 tcp_clamp_window(sk);
4348 else if (tcp_memory_pressure)
4349 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4351 tcp_collapse_ofo_queue(sk);
4352 tcp_collapse(sk, &sk->sk_receive_queue,
4353 sk->sk_receive_queue.next,
4354 (struct sk_buff *)&sk->sk_receive_queue,
4355 tp->copied_seq, tp->rcv_nxt);
4358 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4361 /* Collapsing did not help, destructive actions follow.
4362 * This must not ever occur. */
4364 tcp_prune_ofo_queue(sk);
4366 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4369 /* If we are really being abused, tell the caller to silently
4370 * drop receive data on the floor. It will get retransmitted
4371 * and hopefully then we'll have sufficient space.
4373 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4375 /* Massive buffer overcommit. */
4380 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4381 * As additional protections, we do not touch cwnd in retransmission phases,
4382 * and if application hit its sndbuf limit recently.
4384 void tcp_cwnd_application_limited(struct sock *sk)
4386 struct tcp_sock *tp = tcp_sk(sk);
4388 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4389 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4390 /* Limited by application or receiver window. */
4391 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4392 u32 win_used = max(tp->snd_cwnd_used, init_win);
4393 if (win_used < tp->snd_cwnd) {
4394 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4395 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4397 tp->snd_cwnd_used = 0;
4399 tp->snd_cwnd_stamp = tcp_time_stamp;
4402 static int tcp_should_expand_sndbuf(struct sock *sk)
4404 struct tcp_sock *tp = tcp_sk(sk);
4406 /* If the user specified a specific send buffer setting, do
4409 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4412 /* If we are under global TCP memory pressure, do not expand. */
4413 if (tcp_memory_pressure)
4416 /* If we are under soft global TCP memory pressure, do not expand. */
4417 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4420 /* If we filled the congestion window, do not expand. */
4421 if (tp->packets_out >= tp->snd_cwnd)
4427 /* When incoming ACK allowed to free some skb from write_queue,
4428 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4429 * on the exit from tcp input handler.
4431 * PROBLEM: sndbuf expansion does not work well with largesend.
4433 static void tcp_new_space(struct sock *sk)
4435 struct tcp_sock *tp = tcp_sk(sk);
4437 if (tcp_should_expand_sndbuf(sk)) {
4438 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4439 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4440 demanded = max_t(unsigned int, tp->snd_cwnd,
4441 tp->reordering + 1);
4442 sndmem *= 2 * demanded;
4443 if (sndmem > sk->sk_sndbuf)
4444 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4445 tp->snd_cwnd_stamp = tcp_time_stamp;
4448 sk->sk_write_space(sk);
4451 static void tcp_check_space(struct sock *sk)
4453 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4454 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4455 if (sk->sk_socket &&
4456 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4461 static inline void tcp_data_snd_check(struct sock *sk)
4463 tcp_push_pending_frames(sk);
4464 tcp_check_space(sk);
4468 * Check if sending an ack is needed.
4470 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4472 struct tcp_sock *tp = tcp_sk(sk);
4474 /* More than one full frame received... */
4475 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4476 /* ... and right edge of window advances far enough.
4477 * (tcp_recvmsg() will send ACK otherwise). Or...
4479 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4480 /* We ACK each frame or... */
4481 tcp_in_quickack_mode(sk) ||
4482 /* We have out of order data. */
4483 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4484 /* Then ack it now */
4487 /* Else, send delayed ack. */
4488 tcp_send_delayed_ack(sk);
4492 static inline void tcp_ack_snd_check(struct sock *sk)
4494 if (!inet_csk_ack_scheduled(sk)) {
4495 /* We sent a data segment already. */
4498 __tcp_ack_snd_check(sk, 1);
4502 * This routine is only called when we have urgent data
4503 * signaled. Its the 'slow' part of tcp_urg. It could be
4504 * moved inline now as tcp_urg is only called from one
4505 * place. We handle URGent data wrong. We have to - as
4506 * BSD still doesn't use the correction from RFC961.
4507 * For 1003.1g we should support a new option TCP_STDURG to permit
4508 * either form (or just set the sysctl tcp_stdurg).
4511 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4513 struct tcp_sock *tp = tcp_sk(sk);
4514 u32 ptr = ntohs(th->urg_ptr);
4516 if (ptr && !sysctl_tcp_stdurg)
4518 ptr += ntohl(th->seq);
4520 /* Ignore urgent data that we've already seen and read. */
4521 if (after(tp->copied_seq, ptr))
4524 /* Do not replay urg ptr.
4526 * NOTE: interesting situation not covered by specs.
4527 * Misbehaving sender may send urg ptr, pointing to segment,
4528 * which we already have in ofo queue. We are not able to fetch
4529 * such data and will stay in TCP_URG_NOTYET until will be eaten
4530 * by recvmsg(). Seems, we are not obliged to handle such wicked
4531 * situations. But it is worth to think about possibility of some
4532 * DoSes using some hypothetical application level deadlock.
4534 if (before(ptr, tp->rcv_nxt))
4537 /* Do we already have a newer (or duplicate) urgent pointer? */
4538 if (tp->urg_data && !after(ptr, tp->urg_seq))
4541 /* Tell the world about our new urgent pointer. */
4544 /* We may be adding urgent data when the last byte read was
4545 * urgent. To do this requires some care. We cannot just ignore
4546 * tp->copied_seq since we would read the last urgent byte again
4547 * as data, nor can we alter copied_seq until this data arrives
4548 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4550 * NOTE. Double Dutch. Rendering to plain English: author of comment
4551 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4552 * and expect that both A and B disappear from stream. This is _wrong_.
4553 * Though this happens in BSD with high probability, this is occasional.
4554 * Any application relying on this is buggy. Note also, that fix "works"
4555 * only in this artificial test. Insert some normal data between A and B and we will
4556 * decline of BSD again. Verdict: it is better to remove to trap
4559 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4560 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4561 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4563 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4564 __skb_unlink(skb, &sk->sk_receive_queue);
4569 tp->urg_data = TCP_URG_NOTYET;
4572 /* Disable header prediction. */
4576 /* This is the 'fast' part of urgent handling. */
4577 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4579 struct tcp_sock *tp = tcp_sk(sk);
4581 /* Check if we get a new urgent pointer - normally not. */
4583 tcp_check_urg(sk, th);
4585 /* Do we wait for any urgent data? - normally not... */
4586 if (tp->urg_data == TCP_URG_NOTYET) {
4587 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4590 /* Is the urgent pointer pointing into this packet? */
4591 if (ptr < skb->len) {
4593 if (skb_copy_bits(skb, ptr, &tmp, 1))
4595 tp->urg_data = TCP_URG_VALID | tmp;
4596 if (!sock_flag(sk, SOCK_DEAD))
4597 sk->sk_data_ready(sk, 0);
4602 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4604 struct tcp_sock *tp = tcp_sk(sk);
4605 int chunk = skb->len - hlen;
4609 if (skb_csum_unnecessary(skb))
4610 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4612 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4616 tp->ucopy.len -= chunk;
4617 tp->copied_seq += chunk;
4618 tcp_rcv_space_adjust(sk);
4625 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4626 struct sk_buff *skb)
4630 if (sock_owned_by_user(sk)) {
4632 result = __tcp_checksum_complete(skb);
4635 result = __tcp_checksum_complete(skb);
4640 static inline int tcp_checksum_complete_user(struct sock *sk,
4641 struct sk_buff *skb)
4643 return !skb_csum_unnecessary(skb) &&
4644 __tcp_checksum_complete_user(sk, skb);
4647 #ifdef CONFIG_NET_DMA
4648 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4651 struct tcp_sock *tp = tcp_sk(sk);
4652 int chunk = skb->len - hlen;
4654 int copied_early = 0;
4656 if (tp->ucopy.wakeup)
4659 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4660 tp->ucopy.dma_chan = get_softnet_dma();
4662 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4664 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4666 tp->ucopy.iov, chunk,
4667 tp->ucopy.pinned_list);
4672 tp->ucopy.dma_cookie = dma_cookie;
4675 tp->ucopy.len -= chunk;
4676 tp->copied_seq += chunk;
4677 tcp_rcv_space_adjust(sk);
4679 if ((tp->ucopy.len == 0) ||
4680 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4681 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4682 tp->ucopy.wakeup = 1;
4683 sk->sk_data_ready(sk, 0);
4685 } else if (chunk > 0) {
4686 tp->ucopy.wakeup = 1;
4687 sk->sk_data_ready(sk, 0);
4690 return copied_early;
4692 #endif /* CONFIG_NET_DMA */
4695 * TCP receive function for the ESTABLISHED state.
4697 * It is split into a fast path and a slow path. The fast path is
4699 * - A zero window was announced from us - zero window probing
4700 * is only handled properly in the slow path.
4701 * - Out of order segments arrived.
4702 * - Urgent data is expected.
4703 * - There is no buffer space left
4704 * - Unexpected TCP flags/window values/header lengths are received
4705 * (detected by checking the TCP header against pred_flags)
4706 * - Data is sent in both directions. Fast path only supports pure senders
4707 * or pure receivers (this means either the sequence number or the ack
4708 * value must stay constant)
4709 * - Unexpected TCP option.
4711 * When these conditions are not satisfied it drops into a standard
4712 * receive procedure patterned after RFC793 to handle all cases.
4713 * The first three cases are guaranteed by proper pred_flags setting,
4714 * the rest is checked inline. Fast processing is turned on in
4715 * tcp_data_queue when everything is OK.
4717 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4718 struct tcphdr *th, unsigned len)
4720 struct tcp_sock *tp = tcp_sk(sk);
4723 * Header prediction.
4724 * The code loosely follows the one in the famous
4725 * "30 instruction TCP receive" Van Jacobson mail.
4727 * Van's trick is to deposit buffers into socket queue
4728 * on a device interrupt, to call tcp_recv function
4729 * on the receive process context and checksum and copy
4730 * the buffer to user space. smart...
4732 * Our current scheme is not silly either but we take the
4733 * extra cost of the net_bh soft interrupt processing...
4734 * We do checksum and copy also but from device to kernel.
4737 tp->rx_opt.saw_tstamp = 0;
4739 /* pred_flags is 0xS?10 << 16 + snd_wnd
4740 * if header_prediction is to be made
4741 * 'S' will always be tp->tcp_header_len >> 2
4742 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4743 * turn it off (when there are holes in the receive
4744 * space for instance)
4745 * PSH flag is ignored.
4748 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4749 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4750 int tcp_header_len = tp->tcp_header_len;
4752 /* Timestamp header prediction: tcp_header_len
4753 * is automatically equal to th->doff*4 due to pred_flags
4757 /* Check timestamp */
4758 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4759 __be32 *ptr = (__be32 *)(th + 1);
4761 /* No? Slow path! */
4762 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4763 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4766 tp->rx_opt.saw_tstamp = 1;
4768 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4770 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4772 /* If PAWS failed, check it more carefully in slow path */
4773 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4776 /* DO NOT update ts_recent here, if checksum fails
4777 * and timestamp was corrupted part, it will result
4778 * in a hung connection since we will drop all
4779 * future packets due to the PAWS test.
4783 if (len <= tcp_header_len) {
4784 /* Bulk data transfer: sender */
4785 if (len == tcp_header_len) {
4786 /* Predicted packet is in window by definition.
4787 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4788 * Hence, check seq<=rcv_wup reduces to:
4790 if (tcp_header_len ==
4791 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4792 tp->rcv_nxt == tp->rcv_wup)
4793 tcp_store_ts_recent(tp);
4795 /* We know that such packets are checksummed
4798 tcp_ack(sk, skb, 0);
4800 tcp_data_snd_check(sk);
4802 } else { /* Header too small */
4803 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4808 int copied_early = 0;
4810 if (tp->copied_seq == tp->rcv_nxt &&
4811 len - tcp_header_len <= tp->ucopy.len) {
4812 #ifdef CONFIG_NET_DMA
4813 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4818 if (tp->ucopy.task == current &&
4819 sock_owned_by_user(sk) && !copied_early) {
4820 __set_current_state(TASK_RUNNING);
4822 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4826 /* Predicted packet is in window by definition.
4827 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4828 * Hence, check seq<=rcv_wup reduces to:
4830 if (tcp_header_len ==
4831 (sizeof(struct tcphdr) +
4832 TCPOLEN_TSTAMP_ALIGNED) &&
4833 tp->rcv_nxt == tp->rcv_wup)
4834 tcp_store_ts_recent(tp);
4836 tcp_rcv_rtt_measure_ts(sk, skb);
4838 __skb_pull(skb, tcp_header_len);
4839 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4840 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4843 tcp_cleanup_rbuf(sk, skb->len);
4846 if (tcp_checksum_complete_user(sk, skb))
4849 /* Predicted packet is in window by definition.
4850 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4851 * Hence, check seq<=rcv_wup reduces to:
4853 if (tcp_header_len ==
4854 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4855 tp->rcv_nxt == tp->rcv_wup)
4856 tcp_store_ts_recent(tp);
4858 tcp_rcv_rtt_measure_ts(sk, skb);
4860 if ((int)skb->truesize > sk->sk_forward_alloc)
4863 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4865 /* Bulk data transfer: receiver */
4866 __skb_pull(skb, tcp_header_len);
4867 __skb_queue_tail(&sk->sk_receive_queue, skb);
4868 skb_set_owner_r(skb, sk);
4869 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4872 tcp_event_data_recv(sk, skb);
4874 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4875 /* Well, only one small jumplet in fast path... */
4876 tcp_ack(sk, skb, FLAG_DATA);
4877 tcp_data_snd_check(sk);
4878 if (!inet_csk_ack_scheduled(sk))
4882 __tcp_ack_snd_check(sk, 0);
4884 #ifdef CONFIG_NET_DMA
4886 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4892 sk->sk_data_ready(sk, 0);
4898 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
4902 * RFC1323: H1. Apply PAWS check first.
4904 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4905 tcp_paws_discard(sk, skb)) {
4907 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4908 tcp_send_dupack(sk, skb);
4911 /* Resets are accepted even if PAWS failed.
4913 ts_recent update must be made after we are sure
4914 that the packet is in window.
4919 * Standard slow path.
4922 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4923 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4924 * (RST) segments are validated by checking their SEQ-fields."
4925 * And page 69: "If an incoming segment is not acceptable,
4926 * an acknowledgment should be sent in reply (unless the RST bit
4927 * is set, if so drop the segment and return)".
4930 tcp_send_dupack(sk, skb);
4939 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4941 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4942 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4943 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4950 tcp_ack(sk, skb, FLAG_SLOWPATH);
4952 tcp_rcv_rtt_measure_ts(sk, skb);
4954 /* Process urgent data. */
4955 tcp_urg(sk, skb, th);
4957 /* step 7: process the segment text */
4958 tcp_data_queue(sk, skb);
4960 tcp_data_snd_check(sk);
4961 tcp_ack_snd_check(sk);
4965 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4972 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4973 struct tcphdr *th, unsigned len)
4975 struct tcp_sock *tp = tcp_sk(sk);
4976 struct inet_connection_sock *icsk = inet_csk(sk);
4977 int saved_clamp = tp->rx_opt.mss_clamp;
4979 tcp_parse_options(skb, &tp->rx_opt, 0);
4983 * "If the state is SYN-SENT then
4984 * first check the ACK bit
4985 * If the ACK bit is set
4986 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4987 * a reset (unless the RST bit is set, if so drop
4988 * the segment and return)"
4990 * We do not send data with SYN, so that RFC-correct
4993 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4994 goto reset_and_undo;
4996 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4997 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4999 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
5000 goto reset_and_undo;
5003 /* Now ACK is acceptable.
5005 * "If the RST bit is set
5006 * If the ACK was acceptable then signal the user "error:
5007 * connection reset", drop the segment, enter CLOSED state,
5008 * delete TCB, and return."
5017 * "fifth, if neither of the SYN or RST bits is set then
5018 * drop the segment and return."
5024 goto discard_and_undo;
5027 * "If the SYN bit is on ...
5028 * are acceptable then ...
5029 * (our SYN has been ACKed), change the connection
5030 * state to ESTABLISHED..."
5033 TCP_ECN_rcv_synack(tp, th);
5035 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5036 tcp_ack(sk, skb, FLAG_SLOWPATH);
5038 /* Ok.. it's good. Set up sequence numbers and
5039 * move to established.
5041 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5042 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5044 /* RFC1323: The window in SYN & SYN/ACK segments is
5047 tp->snd_wnd = ntohs(th->window);
5048 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
5050 if (!tp->rx_opt.wscale_ok) {
5051 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5052 tp->window_clamp = min(tp->window_clamp, 65535U);
5055 if (tp->rx_opt.saw_tstamp) {
5056 tp->rx_opt.tstamp_ok = 1;
5057 tp->tcp_header_len =
5058 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5059 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5060 tcp_store_ts_recent(tp);
5062 tp->tcp_header_len = sizeof(struct tcphdr);
5065 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5066 tcp_enable_fack(tp);
5069 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5070 tcp_initialize_rcv_mss(sk);
5072 /* Remember, tcp_poll() does not lock socket!
5073 * Change state from SYN-SENT only after copied_seq
5074 * is initialized. */
5075 tp->copied_seq = tp->rcv_nxt;
5077 tcp_set_state(sk, TCP_ESTABLISHED);
5079 security_inet_conn_established(sk, skb);
5081 /* Make sure socket is routed, for correct metrics. */
5082 icsk->icsk_af_ops->rebuild_header(sk);
5084 tcp_init_metrics(sk);
5086 tcp_init_congestion_control(sk);
5088 /* Prevent spurious tcp_cwnd_restart() on first data
5091 tp->lsndtime = tcp_time_stamp;
5093 tcp_init_buffer_space(sk);
5095 if (sock_flag(sk, SOCK_KEEPOPEN))
5096 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5098 if (!tp->rx_opt.snd_wscale)
5099 __tcp_fast_path_on(tp, tp->snd_wnd);
5103 if (!sock_flag(sk, SOCK_DEAD)) {
5104 sk->sk_state_change(sk);
5105 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5108 if (sk->sk_write_pending ||
5109 icsk->icsk_accept_queue.rskq_defer_accept ||
5110 icsk->icsk_ack.pingpong) {
5111 /* Save one ACK. Data will be ready after
5112 * several ticks, if write_pending is set.
5114 * It may be deleted, but with this feature tcpdumps
5115 * look so _wonderfully_ clever, that I was not able
5116 * to stand against the temptation 8) --ANK
5118 inet_csk_schedule_ack(sk);
5119 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5120 icsk->icsk_ack.ato = TCP_ATO_MIN;
5121 tcp_incr_quickack(sk);
5122 tcp_enter_quickack_mode(sk);
5123 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5124 TCP_DELACK_MAX, TCP_RTO_MAX);
5135 /* No ACK in the segment */
5139 * "If the RST bit is set
5141 * Otherwise (no ACK) drop the segment and return."
5144 goto discard_and_undo;
5148 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5149 tcp_paws_check(&tp->rx_opt, 0))
5150 goto discard_and_undo;
5153 /* We see SYN without ACK. It is attempt of
5154 * simultaneous connect with crossed SYNs.
5155 * Particularly, it can be connect to self.
5157 tcp_set_state(sk, TCP_SYN_RECV);
5159 if (tp->rx_opt.saw_tstamp) {
5160 tp->rx_opt.tstamp_ok = 1;
5161 tcp_store_ts_recent(tp);
5162 tp->tcp_header_len =
5163 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5165 tp->tcp_header_len = sizeof(struct tcphdr);
5168 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5169 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5171 /* RFC1323: The window in SYN & SYN/ACK segments is
5174 tp->snd_wnd = ntohs(th->window);
5175 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5176 tp->max_window = tp->snd_wnd;
5178 TCP_ECN_rcv_syn(tp, th);
5181 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5182 tcp_initialize_rcv_mss(sk);
5184 tcp_send_synack(sk);
5186 /* Note, we could accept data and URG from this segment.
5187 * There are no obstacles to make this.
5189 * However, if we ignore data in ACKless segments sometimes,
5190 * we have no reasons to accept it sometimes.
5191 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5192 * is not flawless. So, discard packet for sanity.
5193 * Uncomment this return to process the data.
5200 /* "fifth, if neither of the SYN or RST bits is set then
5201 * drop the segment and return."
5205 tcp_clear_options(&tp->rx_opt);
5206 tp->rx_opt.mss_clamp = saved_clamp;
5210 tcp_clear_options(&tp->rx_opt);
5211 tp->rx_opt.mss_clamp = saved_clamp;
5216 * This function implements the receiving procedure of RFC 793 for
5217 * all states except ESTABLISHED and TIME_WAIT.
5218 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5219 * address independent.
5222 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5223 struct tcphdr *th, unsigned len)
5225 struct tcp_sock *tp = tcp_sk(sk);
5226 struct inet_connection_sock *icsk = inet_csk(sk);
5229 tp->rx_opt.saw_tstamp = 0;
5231 switch (sk->sk_state) {
5243 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5246 /* Now we have several options: In theory there is
5247 * nothing else in the frame. KA9Q has an option to
5248 * send data with the syn, BSD accepts data with the
5249 * syn up to the [to be] advertised window and
5250 * Solaris 2.1 gives you a protocol error. For now
5251 * we just ignore it, that fits the spec precisely
5252 * and avoids incompatibilities. It would be nice in
5253 * future to drop through and process the data.
5255 * Now that TTCP is starting to be used we ought to
5257 * But, this leaves one open to an easy denial of
5258 * service attack, and SYN cookies can't defend
5259 * against this problem. So, we drop the data
5260 * in the interest of security over speed unless
5261 * it's still in use.
5269 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5273 /* Do step6 onward by hand. */
5274 tcp_urg(sk, skb, th);
5276 tcp_data_snd_check(sk);
5280 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5281 tcp_paws_discard(sk, skb)) {
5283 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
5284 tcp_send_dupack(sk, skb);
5287 /* Reset is accepted even if it did not pass PAWS. */
5290 /* step 1: check sequence number */
5291 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5293 tcp_send_dupack(sk, skb);
5297 /* step 2: check RST bit */
5303 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5305 /* step 3: check security and precedence [ignored] */
5309 * Check for a SYN in window.
5311 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5312 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
5317 /* step 5: check the ACK field */
5319 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5321 switch (sk->sk_state) {
5324 tp->copied_seq = tp->rcv_nxt;
5326 tcp_set_state(sk, TCP_ESTABLISHED);
5327 sk->sk_state_change(sk);
5329 /* Note, that this wakeup is only for marginal
5330 * crossed SYN case. Passively open sockets
5331 * are not waked up, because sk->sk_sleep ==
5332 * NULL and sk->sk_socket == NULL.
5336 SOCK_WAKE_IO, POLL_OUT);
5338 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5339 tp->snd_wnd = ntohs(th->window) <<
5340 tp->rx_opt.snd_wscale;
5341 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5342 TCP_SKB_CB(skb)->seq);
5344 /* tcp_ack considers this ACK as duplicate
5345 * and does not calculate rtt.
5346 * Fix it at least with timestamps.
5348 if (tp->rx_opt.saw_tstamp &&
5349 tp->rx_opt.rcv_tsecr && !tp->srtt)
5350 tcp_ack_saw_tstamp(sk, 0);
5352 if (tp->rx_opt.tstamp_ok)
5353 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5355 /* Make sure socket is routed, for
5358 icsk->icsk_af_ops->rebuild_header(sk);
5360 tcp_init_metrics(sk);
5362 tcp_init_congestion_control(sk);
5364 /* Prevent spurious tcp_cwnd_restart() on
5365 * first data packet.
5367 tp->lsndtime = tcp_time_stamp;
5370 tcp_initialize_rcv_mss(sk);
5371 tcp_init_buffer_space(sk);
5372 tcp_fast_path_on(tp);
5379 if (tp->snd_una == tp->write_seq) {
5380 tcp_set_state(sk, TCP_FIN_WAIT2);
5381 sk->sk_shutdown |= SEND_SHUTDOWN;
5382 dst_confirm(sk->sk_dst_cache);
5384 if (!sock_flag(sk, SOCK_DEAD))
5385 /* Wake up lingering close() */
5386 sk->sk_state_change(sk);
5390 if (tp->linger2 < 0 ||
5391 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5392 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5394 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5398 tmo = tcp_fin_time(sk);
5399 if (tmo > TCP_TIMEWAIT_LEN) {
5400 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5401 } else if (th->fin || sock_owned_by_user(sk)) {
5402 /* Bad case. We could lose such FIN otherwise.
5403 * It is not a big problem, but it looks confusing
5404 * and not so rare event. We still can lose it now,
5405 * if it spins in bh_lock_sock(), but it is really
5408 inet_csk_reset_keepalive_timer(sk, tmo);
5410 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5418 if (tp->snd_una == tp->write_seq) {
5419 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5425 if (tp->snd_una == tp->write_seq) {
5426 tcp_update_metrics(sk);
5435 /* step 6: check the URG bit */
5436 tcp_urg(sk, skb, th);
5438 /* step 7: process the segment text */
5439 switch (sk->sk_state) {
5440 case TCP_CLOSE_WAIT:
5443 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5447 /* RFC 793 says to queue data in these states,
5448 * RFC 1122 says we MUST send a reset.
5449 * BSD 4.4 also does reset.
5451 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5452 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5453 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5454 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5460 case TCP_ESTABLISHED:
5461 tcp_data_queue(sk, skb);
5466 /* tcp_data could move socket to TIME-WAIT */
5467 if (sk->sk_state != TCP_CLOSE) {
5468 tcp_data_snd_check(sk);
5469 tcp_ack_snd_check(sk);
5479 EXPORT_SYMBOL(sysctl_tcp_ecn);
5480 EXPORT_SYMBOL(sysctl_tcp_reordering);
5481 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5482 EXPORT_SYMBOL(tcp_parse_options);
5483 #ifdef CONFIG_TCP_MD5SIG
5484 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5486 EXPORT_SYMBOL(tcp_rcv_established);
5487 EXPORT_SYMBOL(tcp_rcv_state_process);
5488 EXPORT_SYMBOL(tcp_initialize_rcv_mss);