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).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presence of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly = 2;
89 int sysctl_tcp_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
101 #define FLAG_ECE 0x40 /* ECE in this ACK */
102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
108 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
110 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
111 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
112 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
113 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
114 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
116 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
118 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
119 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
121 /* Adapt the MSS value used to make delayed ack decision to the
124 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
126 struct inet_connection_sock *icsk = inet_csk(sk);
127 const unsigned int lss = icsk->icsk_ack.last_seg_size;
130 icsk->icsk_ack.last_seg_size = 0;
132 /* skb->len may jitter because of SACKs, even if peer
133 * sends good full-sized frames.
135 len = skb_shinfo(skb)->gso_size ? : skb->len;
136 if (len >= icsk->icsk_ack.rcv_mss) {
137 icsk->icsk_ack.rcv_mss = len;
139 /* Otherwise, we make more careful check taking into account,
140 * that SACKs block is variable.
142 * "len" is invariant segment length, including TCP header.
144 len += skb->data - skb_transport_header(skb);
145 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
146 /* If PSH is not set, packet should be
147 * full sized, provided peer TCP is not badly broken.
148 * This observation (if it is correct 8)) allows
149 * to handle super-low mtu links fairly.
151 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
152 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
153 /* Subtract also invariant (if peer is RFC compliant),
154 * tcp header plus fixed timestamp option length.
155 * Resulting "len" is MSS free of SACK jitter.
157 len -= tcp_sk(sk)->tcp_header_len;
158 icsk->icsk_ack.last_seg_size = len;
160 icsk->icsk_ack.rcv_mss = len;
164 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
166 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
170 static void tcp_incr_quickack(struct sock *sk)
172 struct inet_connection_sock *icsk = inet_csk(sk);
173 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
177 if (quickacks > icsk->icsk_ack.quick)
178 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
181 void tcp_enter_quickack_mode(struct sock *sk)
183 struct inet_connection_sock *icsk = inet_csk(sk);
184 tcp_incr_quickack(sk);
185 icsk->icsk_ack.pingpong = 0;
186 icsk->icsk_ack.ato = TCP_ATO_MIN;
189 /* Send ACKs quickly, if "quick" count is not exhausted
190 * and the session is not interactive.
193 static inline int tcp_in_quickack_mode(const struct sock *sk)
195 const struct inet_connection_sock *icsk = inet_csk(sk);
196 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
201 if (tp->ecn_flags & TCP_ECN_OK)
202 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
207 if (tcp_hdr(skb)->cwr)
208 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
218 if (tp->ecn_flags & TCP_ECN_OK) {
219 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
220 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
221 /* Funny extension: if ECT is not set on a segment,
222 * it is surely retransmit. It is not in ECN RFC,
223 * but Linux follows this rule. */
224 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
225 tcp_enter_quickack_mode((struct sock *)tp);
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
231 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
232 tp->ecn_flags &= ~TCP_ECN_OK;
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
237 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
238 tp->ecn_flags &= ~TCP_ECN_OK;
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
243 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
248 /* Buffer size and advertised window tuning.
250 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
253 static void tcp_fixup_sndbuf(struct sock *sk)
255 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
256 sizeof(struct sk_buff);
258 if (sk->sk_sndbuf < 3 * sndmem)
259 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
264 * All tcp_full_space() is split to two parts: "network" buffer, allocated
265 * forward and advertised in receiver window (tp->rcv_wnd) and
266 * "application buffer", required to isolate scheduling/application
267 * latencies from network.
268 * window_clamp is maximal advertised window. It can be less than
269 * tcp_full_space(), in this case tcp_full_space() - window_clamp
270 * is reserved for "application" buffer. The less window_clamp is
271 * the smoother our behaviour from viewpoint of network, but the lower
272 * throughput and the higher sensitivity of the connection to losses. 8)
274 * rcv_ssthresh is more strict window_clamp used at "slow start"
275 * phase to predict further behaviour of this connection.
276 * It is used for two goals:
277 * - to enforce header prediction at sender, even when application
278 * requires some significant "application buffer". It is check #1.
279 * - to prevent pruning of receive queue because of misprediction
280 * of receiver window. Check #2.
282 * The scheme does not work when sender sends good segments opening
283 * window and then starts to feed us spaghetti. But it should work
284 * in common situations. Otherwise, we have to rely on queue collapsing.
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
290 struct tcp_sock *tp = tcp_sk(sk);
292 int truesize = tcp_win_from_space(skb->truesize) >> 1;
293 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
295 while (tp->rcv_ssthresh <= window) {
296 if (truesize <= skb->len)
297 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
305 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
307 struct tcp_sock *tp = tcp_sk(sk);
310 if (tp->rcv_ssthresh < tp->window_clamp &&
311 (int)tp->rcv_ssthresh < tcp_space(sk) &&
312 !tcp_memory_pressure) {
315 /* Check #2. Increase window, if skb with such overhead
316 * will fit to rcvbuf in future.
318 if (tcp_win_from_space(skb->truesize) <= skb->len)
319 incr = 2 * tp->advmss;
321 incr = __tcp_grow_window(sk, skb);
324 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
326 inet_csk(sk)->icsk_ack.quick |= 1;
331 /* 3. Tuning rcvbuf, when connection enters established state. */
333 static void tcp_fixup_rcvbuf(struct sock *sk)
335 struct tcp_sock *tp = tcp_sk(sk);
336 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
338 /* Try to select rcvbuf so that 4 mss-sized segments
339 * will fit to window and corresponding skbs will fit to our rcvbuf.
340 * (was 3; 4 is minimum to allow fast retransmit to work.)
342 while (tcp_win_from_space(rcvmem) < tp->advmss)
344 if (sk->sk_rcvbuf < 4 * rcvmem)
345 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
348 /* 4. Try to fixup all. It is made immediately after connection enters
351 static void tcp_init_buffer_space(struct sock *sk)
353 struct tcp_sock *tp = tcp_sk(sk);
356 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
357 tcp_fixup_rcvbuf(sk);
358 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
359 tcp_fixup_sndbuf(sk);
361 tp->rcvq_space.space = tp->rcv_wnd;
363 maxwin = tcp_full_space(sk);
365 if (tp->window_clamp >= maxwin) {
366 tp->window_clamp = maxwin;
368 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
369 tp->window_clamp = max(maxwin -
370 (maxwin >> sysctl_tcp_app_win),
374 /* Force reservation of one segment. */
375 if (sysctl_tcp_app_win &&
376 tp->window_clamp > 2 * tp->advmss &&
377 tp->window_clamp + tp->advmss > maxwin)
378 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
380 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
381 tp->snd_cwnd_stamp = tcp_time_stamp;
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock *sk)
387 struct tcp_sock *tp = tcp_sk(sk);
388 struct inet_connection_sock *icsk = inet_csk(sk);
390 icsk->icsk_ack.quick = 0;
392 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
393 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
394 !tcp_memory_pressure &&
395 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
396 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
399 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
400 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
403 /* Initialize RCV_MSS value.
404 * RCV_MSS is an our guess about MSS used by the peer.
405 * We haven't any direct information about the MSS.
406 * It's better to underestimate the RCV_MSS rather than overestimate.
407 * Overestimations make us ACKing less frequently than needed.
408 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
410 void tcp_initialize_rcv_mss(struct sock *sk)
412 struct tcp_sock *tp = tcp_sk(sk);
413 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
415 hint = min(hint, tp->rcv_wnd / 2);
416 hint = min(hint, TCP_MIN_RCVMSS);
417 hint = max(hint, TCP_MIN_MSS);
419 inet_csk(sk)->icsk_ack.rcv_mss = hint;
422 /* Receiver "autotuning" code.
424 * The algorithm for RTT estimation w/o timestamps is based on
425 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
426 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
428 * More detail on this code can be found at
429 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
430 * though this reference is out of date. A new paper
433 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
435 u32 new_sample = tp->rcv_rtt_est.rtt;
441 if (new_sample != 0) {
442 /* If we sample in larger samples in the non-timestamp
443 * case, we could grossly overestimate the RTT especially
444 * with chatty applications or bulk transfer apps which
445 * are stalled on filesystem I/O.
447 * Also, since we are only going for a minimum in the
448 * non-timestamp case, we do not smooth things out
449 * else with timestamps disabled convergence takes too
453 m -= (new_sample >> 3);
455 } else if (m < new_sample)
458 /* No previous measure. */
462 if (tp->rcv_rtt_est.rtt != new_sample)
463 tp->rcv_rtt_est.rtt = new_sample;
466 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
468 if (tp->rcv_rtt_est.time == 0)
470 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
472 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
475 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
476 tp->rcv_rtt_est.time = tcp_time_stamp;
479 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
480 const struct sk_buff *skb)
482 struct tcp_sock *tp = tcp_sk(sk);
483 if (tp->rx_opt.rcv_tsecr &&
484 (TCP_SKB_CB(skb)->end_seq -
485 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
486 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
490 * This function should be called every time data is copied to user space.
491 * It calculates the appropriate TCP receive buffer space.
493 void tcp_rcv_space_adjust(struct sock *sk)
495 struct tcp_sock *tp = tcp_sk(sk);
499 if (tp->rcvq_space.time == 0)
502 time = tcp_time_stamp - tp->rcvq_space.time;
503 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
506 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
508 space = max(tp->rcvq_space.space, space);
510 if (tp->rcvq_space.space != space) {
513 tp->rcvq_space.space = space;
515 if (sysctl_tcp_moderate_rcvbuf &&
516 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
517 int new_clamp = space;
519 /* Receive space grows, normalize in order to
520 * take into account packet headers and sk_buff
521 * structure overhead.
526 rcvmem = (tp->advmss + MAX_TCP_HEADER +
527 16 + sizeof(struct sk_buff));
528 while (tcp_win_from_space(rcvmem) < tp->advmss)
531 space = min(space, sysctl_tcp_rmem[2]);
532 if (space > sk->sk_rcvbuf) {
533 sk->sk_rcvbuf = space;
535 /* Make the window clamp follow along. */
536 tp->window_clamp = new_clamp;
542 tp->rcvq_space.seq = tp->copied_seq;
543 tp->rcvq_space.time = tcp_time_stamp;
546 /* There is something which you must keep in mind when you analyze the
547 * behavior of the tp->ato delayed ack timeout interval. When a
548 * connection starts up, we want to ack as quickly as possible. The
549 * problem is that "good" TCP's do slow start at the beginning of data
550 * transmission. The means that until we send the first few ACK's the
551 * sender will sit on his end and only queue most of his data, because
552 * he can only send snd_cwnd unacked packets at any given time. For
553 * each ACK we send, he increments snd_cwnd and transmits more of his
556 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
558 struct tcp_sock *tp = tcp_sk(sk);
559 struct inet_connection_sock *icsk = inet_csk(sk);
562 inet_csk_schedule_ack(sk);
564 tcp_measure_rcv_mss(sk, skb);
566 tcp_rcv_rtt_measure(tp);
568 now = tcp_time_stamp;
570 if (!icsk->icsk_ack.ato) {
571 /* The _first_ data packet received, initialize
572 * delayed ACK engine.
574 tcp_incr_quickack(sk);
575 icsk->icsk_ack.ato = TCP_ATO_MIN;
577 int m = now - icsk->icsk_ack.lrcvtime;
579 if (m <= TCP_ATO_MIN / 2) {
580 /* The fastest case is the first. */
581 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
582 } else if (m < icsk->icsk_ack.ato) {
583 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
584 if (icsk->icsk_ack.ato > icsk->icsk_rto)
585 icsk->icsk_ack.ato = icsk->icsk_rto;
586 } else if (m > icsk->icsk_rto) {
587 /* Too long gap. Apparently sender failed to
588 * restart window, so that we send ACKs quickly.
590 tcp_incr_quickack(sk);
594 icsk->icsk_ack.lrcvtime = now;
596 TCP_ECN_check_ce(tp, skb);
599 tcp_grow_window(sk, skb);
602 static u32 tcp_rto_min(struct sock *sk)
604 struct dst_entry *dst = __sk_dst_get(sk);
605 u32 rto_min = TCP_RTO_MIN;
607 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
608 rto_min = dst->metrics[RTAX_RTO_MIN - 1];
612 /* Called to compute a smoothed rtt estimate. The data fed to this
613 * routine either comes from timestamps, or from segments that were
614 * known _not_ to have been retransmitted [see Karn/Partridge
615 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
616 * piece by Van Jacobson.
617 * NOTE: the next three routines used to be one big routine.
618 * To save cycles in the RFC 1323 implementation it was better to break
619 * it up into three procedures. -- erics
621 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
623 struct tcp_sock *tp = tcp_sk(sk);
624 long m = mrtt; /* RTT */
626 /* The following amusing code comes from Jacobson's
627 * article in SIGCOMM '88. Note that rtt and mdev
628 * are scaled versions of rtt and mean deviation.
629 * This is designed to be as fast as possible
630 * m stands for "measurement".
632 * On a 1990 paper the rto value is changed to:
633 * RTO = rtt + 4 * mdev
635 * Funny. This algorithm seems to be very broken.
636 * These formulae increase RTO, when it should be decreased, increase
637 * too slowly, when it should be increased quickly, decrease too quickly
638 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
639 * does not matter how to _calculate_ it. Seems, it was trap
640 * that VJ failed to avoid. 8)
645 m -= (tp->srtt >> 3); /* m is now error in rtt est */
646 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
648 m = -m; /* m is now abs(error) */
649 m -= (tp->mdev >> 2); /* similar update on mdev */
650 /* This is similar to one of Eifel findings.
651 * Eifel blocks mdev updates when rtt decreases.
652 * This solution is a bit different: we use finer gain
653 * for mdev in this case (alpha*beta).
654 * Like Eifel it also prevents growth of rto,
655 * but also it limits too fast rto decreases,
656 * happening in pure Eifel.
661 m -= (tp->mdev >> 2); /* similar update on mdev */
663 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
664 if (tp->mdev > tp->mdev_max) {
665 tp->mdev_max = tp->mdev;
666 if (tp->mdev_max > tp->rttvar)
667 tp->rttvar = tp->mdev_max;
669 if (after(tp->snd_una, tp->rtt_seq)) {
670 if (tp->mdev_max < tp->rttvar)
671 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
672 tp->rtt_seq = tp->snd_nxt;
673 tp->mdev_max = tcp_rto_min(sk);
676 /* no previous measure. */
677 tp->srtt = m << 3; /* take the measured time to be rtt */
678 tp->mdev = m << 1; /* make sure rto = 3*rtt */
679 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
680 tp->rtt_seq = tp->snd_nxt;
684 /* Calculate rto without backoff. This is the second half of Van Jacobson's
685 * routine referred to above.
687 static inline void tcp_set_rto(struct sock *sk)
689 const struct tcp_sock *tp = tcp_sk(sk);
690 /* Old crap is replaced with new one. 8)
693 * 1. If rtt variance happened to be less 50msec, it is hallucination.
694 * It cannot be less due to utterly erratic ACK generation made
695 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
696 * to do with delayed acks, because at cwnd>2 true delack timeout
697 * is invisible. Actually, Linux-2.4 also generates erratic
698 * ACKs in some circumstances.
700 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
702 /* 2. Fixups made earlier cannot be right.
703 * If we do not estimate RTO correctly without them,
704 * all the algo is pure shit and should be replaced
705 * with correct one. It is exactly, which we pretend to do.
709 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
710 * guarantees that rto is higher.
712 static inline void tcp_bound_rto(struct sock *sk)
714 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
715 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
718 /* Save metrics learned by this TCP session.
719 This function is called only, when TCP finishes successfully
720 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
722 void tcp_update_metrics(struct sock *sk)
724 struct tcp_sock *tp = tcp_sk(sk);
725 struct dst_entry *dst = __sk_dst_get(sk);
727 if (sysctl_tcp_nometrics_save)
732 if (dst && (dst->flags & DST_HOST)) {
733 const struct inet_connection_sock *icsk = inet_csk(sk);
736 if (icsk->icsk_backoff || !tp->srtt) {
737 /* This session failed to estimate rtt. Why?
738 * Probably, no packets returned in time.
741 if (!(dst_metric_locked(dst, RTAX_RTT)))
742 dst->metrics[RTAX_RTT - 1] = 0;
746 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
748 /* If newly calculated rtt larger than stored one,
749 * store new one. Otherwise, use EWMA. Remember,
750 * rtt overestimation is always better than underestimation.
752 if (!(dst_metric_locked(dst, RTAX_RTT))) {
754 dst->metrics[RTAX_RTT - 1] = tp->srtt;
756 dst->metrics[RTAX_RTT - 1] -= (m >> 3);
759 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
763 /* Scale deviation to rttvar fixed point */
768 if (m >= dst_metric(dst, RTAX_RTTVAR))
769 dst->metrics[RTAX_RTTVAR - 1] = m;
771 dst->metrics[RTAX_RTTVAR-1] -=
772 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
775 if (tp->snd_ssthresh >= 0xFFFF) {
776 /* Slow start still did not finish. */
777 if (dst_metric(dst, RTAX_SSTHRESH) &&
778 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
779 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
780 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
781 if (!dst_metric_locked(dst, RTAX_CWND) &&
782 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
783 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
784 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
785 icsk->icsk_ca_state == TCP_CA_Open) {
786 /* Cong. avoidance phase, cwnd is reliable. */
787 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
788 dst->metrics[RTAX_SSTHRESH-1] =
789 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
790 if (!dst_metric_locked(dst, RTAX_CWND))
791 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
793 /* Else slow start did not finish, cwnd is non-sense,
794 ssthresh may be also invalid.
796 if (!dst_metric_locked(dst, RTAX_CWND))
797 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
798 if (dst->metrics[RTAX_SSTHRESH-1] &&
799 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
800 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
801 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
804 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
805 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
806 tp->reordering != sysctl_tcp_reordering)
807 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
812 /* Numbers are taken from RFC3390.
814 * John Heffner states:
816 * The RFC specifies a window of no more than 4380 bytes
817 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
818 * is a bit misleading because they use a clamp at 4380 bytes
819 * rather than use a multiplier in the relevant range.
821 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
823 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
826 if (tp->mss_cache > 1460)
829 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
831 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
834 /* Set slow start threshold and cwnd not falling to slow start */
835 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
837 struct tcp_sock *tp = tcp_sk(sk);
838 const struct inet_connection_sock *icsk = inet_csk(sk);
840 tp->prior_ssthresh = 0;
842 if (icsk->icsk_ca_state < TCP_CA_CWR) {
845 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
846 tp->snd_cwnd = min(tp->snd_cwnd,
847 tcp_packets_in_flight(tp) + 1U);
848 tp->snd_cwnd_cnt = 0;
849 tp->high_seq = tp->snd_nxt;
850 tp->snd_cwnd_stamp = tcp_time_stamp;
851 TCP_ECN_queue_cwr(tp);
853 tcp_set_ca_state(sk, TCP_CA_CWR);
858 * Packet counting of FACK is based on in-order assumptions, therefore TCP
859 * disables it when reordering is detected
861 static void tcp_disable_fack(struct tcp_sock *tp)
863 /* RFC3517 uses different metric in lost marker => reset on change */
865 tp->lost_skb_hint = NULL;
866 tp->rx_opt.sack_ok &= ~2;
869 /* Take a notice that peer is sending D-SACKs */
870 static void tcp_dsack_seen(struct tcp_sock *tp)
872 tp->rx_opt.sack_ok |= 4;
875 /* Initialize metrics on socket. */
877 static void tcp_init_metrics(struct sock *sk)
879 struct tcp_sock *tp = tcp_sk(sk);
880 struct dst_entry *dst = __sk_dst_get(sk);
887 if (dst_metric_locked(dst, RTAX_CWND))
888 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
889 if (dst_metric(dst, RTAX_SSTHRESH)) {
890 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
891 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
892 tp->snd_ssthresh = tp->snd_cwnd_clamp;
894 if (dst_metric(dst, RTAX_REORDERING) &&
895 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
896 tcp_disable_fack(tp);
897 tp->reordering = dst_metric(dst, RTAX_REORDERING);
900 if (dst_metric(dst, RTAX_RTT) == 0)
903 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
906 /* Initial rtt is determined from SYN,SYN-ACK.
907 * The segment is small and rtt may appear much
908 * less than real one. Use per-dst memory
909 * to make it more realistic.
911 * A bit of theory. RTT is time passed after "normal" sized packet
912 * is sent until it is ACKed. In normal circumstances sending small
913 * packets force peer to delay ACKs and calculation is correct too.
914 * The algorithm is adaptive and, provided we follow specs, it
915 * NEVER underestimate RTT. BUT! If peer tries to make some clever
916 * tricks sort of "quick acks" for time long enough to decrease RTT
917 * to low value, and then abruptly stops to do it and starts to delay
918 * ACKs, wait for troubles.
920 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
921 tp->srtt = dst_metric(dst, RTAX_RTT);
922 tp->rtt_seq = tp->snd_nxt;
924 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
925 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
926 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
930 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
932 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
933 tp->snd_cwnd_stamp = tcp_time_stamp;
937 /* Play conservative. If timestamps are not
938 * supported, TCP will fail to recalculate correct
939 * rtt, if initial rto is too small. FORGET ALL AND RESET!
941 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
943 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
944 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
948 static void tcp_update_reordering(struct sock *sk, const int metric,
951 struct tcp_sock *tp = tcp_sk(sk);
952 if (metric > tp->reordering) {
953 tp->reordering = min(TCP_MAX_REORDERING, metric);
955 /* This exciting event is worth to be remembered. 8) */
957 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
958 else if (tcp_is_reno(tp))
959 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
960 else if (tcp_is_fack(tp))
961 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
963 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
964 #if FASTRETRANS_DEBUG > 1
965 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
966 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
970 tp->undo_marker ? tp->undo_retrans : 0);
972 tcp_disable_fack(tp);
976 /* This procedure tags the retransmission queue when SACKs arrive.
978 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
979 * Packets in queue with these bits set are counted in variables
980 * sacked_out, retrans_out and lost_out, correspondingly.
982 * Valid combinations are:
983 * Tag InFlight Description
984 * 0 1 - orig segment is in flight.
985 * S 0 - nothing flies, orig reached receiver.
986 * L 0 - nothing flies, orig lost by net.
987 * R 2 - both orig and retransmit are in flight.
988 * L|R 1 - orig is lost, retransmit is in flight.
989 * S|R 1 - orig reached receiver, retrans is still in flight.
990 * (L|S|R is logically valid, it could occur when L|R is sacked,
991 * but it is equivalent to plain S and code short-curcuits it to S.
992 * L|S is logically invalid, it would mean -1 packet in flight 8))
994 * These 6 states form finite state machine, controlled by the following events:
995 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
996 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
997 * 3. Loss detection event of one of three flavors:
998 * A. Scoreboard estimator decided the packet is lost.
999 * A'. Reno "three dupacks" marks head of queue lost.
1000 * A''. Its FACK modfication, head until snd.fack is lost.
1001 * B. SACK arrives sacking data transmitted after never retransmitted
1002 * hole was sent out.
1003 * C. SACK arrives sacking SND.NXT at the moment, when the
1004 * segment was retransmitted.
1005 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1007 * It is pleasant to note, that state diagram turns out to be commutative,
1008 * so that we are allowed not to be bothered by order of our actions,
1009 * when multiple events arrive simultaneously. (see the function below).
1011 * Reordering detection.
1012 * --------------------
1013 * Reordering metric is maximal distance, which a packet can be displaced
1014 * in packet stream. With SACKs we can estimate it:
1016 * 1. SACK fills old hole and the corresponding segment was not
1017 * ever retransmitted -> reordering. Alas, we cannot use it
1018 * when segment was retransmitted.
1019 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1020 * for retransmitted and already SACKed segment -> reordering..
1021 * Both of these heuristics are not used in Loss state, when we cannot
1022 * account for retransmits accurately.
1024 * SACK block validation.
1025 * ----------------------
1027 * SACK block range validation checks that the received SACK block fits to
1028 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1029 * Note that SND.UNA is not included to the range though being valid because
1030 * it means that the receiver is rather inconsistent with itself reporting
1031 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1032 * perfectly valid, however, in light of RFC2018 which explicitly states
1033 * that "SACK block MUST reflect the newest segment. Even if the newest
1034 * segment is going to be discarded ...", not that it looks very clever
1035 * in case of head skb. Due to potentional receiver driven attacks, we
1036 * choose to avoid immediate execution of a walk in write queue due to
1037 * reneging and defer head skb's loss recovery to standard loss recovery
1038 * procedure that will eventually trigger (nothing forbids us doing this).
1040 * Implements also blockage to start_seq wrap-around. Problem lies in the
1041 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1042 * there's no guarantee that it will be before snd_nxt (n). The problem
1043 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1046 * <- outs wnd -> <- wrapzone ->
1047 * u e n u_w e_w s n_w
1049 * |<------------+------+----- TCP seqno space --------------+---------->|
1050 * ...-- <2^31 ->| |<--------...
1051 * ...---- >2^31 ------>| |<--------...
1053 * Current code wouldn't be vulnerable but it's better still to discard such
1054 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1055 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1056 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1057 * equal to the ideal case (infinite seqno space without wrap caused issues).
1059 * With D-SACK the lower bound is extended to cover sequence space below
1060 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1061 * again, D-SACK block must not to go across snd_una (for the same reason as
1062 * for the normal SACK blocks, explained above). But there all simplicity
1063 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1064 * fully below undo_marker they do not affect behavior in anyway and can
1065 * therefore be safely ignored. In rare cases (which are more or less
1066 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1067 * fragmentation and packet reordering past skb's retransmission. To consider
1068 * them correctly, the acceptable range must be extended even more though
1069 * the exact amount is rather hard to quantify. However, tp->max_window can
1070 * be used as an exaggerated estimate.
1072 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1073 u32 start_seq, u32 end_seq)
1075 /* Too far in future, or reversed (interpretation is ambiguous) */
1076 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1079 /* Nasty start_seq wrap-around check (see comments above) */
1080 if (!before(start_seq, tp->snd_nxt))
1083 /* In outstanding window? ...This is valid exit for D-SACKs too.
1084 * start_seq == snd_una is non-sensical (see comments above)
1086 if (after(start_seq, tp->snd_una))
1089 if (!is_dsack || !tp->undo_marker)
1092 /* ...Then it's D-SACK, and must reside below snd_una completely */
1093 if (!after(end_seq, tp->snd_una))
1096 if (!before(start_seq, tp->undo_marker))
1100 if (!after(end_seq, tp->undo_marker))
1103 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1104 * start_seq < undo_marker and end_seq >= undo_marker.
1106 return !before(start_seq, end_seq - tp->max_window);
1109 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1110 * Event "C". Later note: FACK people cheated me again 8), we have to account
1111 * for reordering! Ugly, but should help.
1113 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1114 * less than what is now known to be received by the other end (derived from
1115 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1116 * retransmitted skbs to avoid some costly processing per ACKs.
1118 static void tcp_mark_lost_retrans(struct sock *sk)
1120 const struct inet_connection_sock *icsk = inet_csk(sk);
1121 struct tcp_sock *tp = tcp_sk(sk);
1122 struct sk_buff *skb;
1124 u32 new_low_seq = tp->snd_nxt;
1125 u32 received_upto = tcp_highest_sack_seq(tp);
1127 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1128 !after(received_upto, tp->lost_retrans_low) ||
1129 icsk->icsk_ca_state != TCP_CA_Recovery)
1132 tcp_for_write_queue(skb, sk) {
1133 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1135 if (skb == tcp_send_head(sk))
1137 if (cnt == tp->retrans_out)
1139 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1142 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1145 if (after(received_upto, ack_seq) &&
1147 !before(received_upto,
1148 ack_seq + tp->reordering * tp->mss_cache))) {
1149 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1150 tp->retrans_out -= tcp_skb_pcount(skb);
1152 /* clear lost hint */
1153 tp->retransmit_skb_hint = NULL;
1155 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1156 tp->lost_out += tcp_skb_pcount(skb);
1157 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1159 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1161 if (before(ack_seq, new_low_seq))
1162 new_low_seq = ack_seq;
1163 cnt += tcp_skb_pcount(skb);
1167 if (tp->retrans_out)
1168 tp->lost_retrans_low = new_low_seq;
1171 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1172 struct tcp_sack_block_wire *sp, int num_sacks,
1175 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1176 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1179 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1182 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1183 } else if (num_sacks > 1) {
1184 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1185 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1187 if (!after(end_seq_0, end_seq_1) &&
1188 !before(start_seq_0, start_seq_1)) {
1191 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1195 /* D-SACK for already forgotten data... Do dumb counting. */
1197 !after(end_seq_0, prior_snd_una) &&
1198 after(end_seq_0, tp->undo_marker))
1204 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1205 * the incoming SACK may not exactly match but we can find smaller MSS
1206 * aligned portion of it that matches. Therefore we might need to fragment
1207 * which may fail and creates some hassle (caller must handle error case
1210 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1211 u32 start_seq, u32 end_seq)
1214 unsigned int pkt_len;
1216 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1217 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1219 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1220 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1222 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1225 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1227 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1228 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1236 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1237 int *reord, int dup_sack, int fack_count)
1239 struct tcp_sock *tp = tcp_sk(sk);
1240 u8 sacked = TCP_SKB_CB(skb)->sacked;
1243 /* Account D-SACK for retransmitted packet. */
1244 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1245 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1247 if (sacked & TCPCB_SACKED_ACKED)
1248 *reord = min(fack_count, *reord);
1251 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1252 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1255 if (!(sacked & TCPCB_SACKED_ACKED)) {
1256 if (sacked & TCPCB_SACKED_RETRANS) {
1257 /* If the segment is not tagged as lost,
1258 * we do not clear RETRANS, believing
1259 * that retransmission is still in flight.
1261 if (sacked & TCPCB_LOST) {
1262 TCP_SKB_CB(skb)->sacked &=
1263 ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1264 tp->lost_out -= tcp_skb_pcount(skb);
1265 tp->retrans_out -= tcp_skb_pcount(skb);
1267 /* clear lost hint */
1268 tp->retransmit_skb_hint = NULL;
1271 if (!(sacked & TCPCB_RETRANS)) {
1272 /* New sack for not retransmitted frame,
1273 * which was in hole. It is reordering.
1275 if (before(TCP_SKB_CB(skb)->seq,
1276 tcp_highest_sack_seq(tp)))
1277 *reord = min(fack_count, *reord);
1279 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1280 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1281 flag |= FLAG_ONLY_ORIG_SACKED;
1284 if (sacked & TCPCB_LOST) {
1285 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1286 tp->lost_out -= tcp_skb_pcount(skb);
1288 /* clear lost hint */
1289 tp->retransmit_skb_hint = NULL;
1293 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1294 flag |= FLAG_DATA_SACKED;
1295 tp->sacked_out += tcp_skb_pcount(skb);
1297 fack_count += tcp_skb_pcount(skb);
1299 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1300 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1301 before(TCP_SKB_CB(skb)->seq,
1302 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1303 tp->lost_cnt_hint += tcp_skb_pcount(skb);
1305 if (fack_count > tp->fackets_out)
1306 tp->fackets_out = fack_count;
1308 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1309 tcp_advance_highest_sack(sk, skb);
1312 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1313 * frames and clear it. undo_retrans is decreased above, L|R frames
1314 * are accounted above as well.
1316 if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) {
1317 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1318 tp->retrans_out -= tcp_skb_pcount(skb);
1319 tp->retransmit_skb_hint = NULL;
1325 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1326 struct tcp_sack_block *next_dup,
1327 u32 start_seq, u32 end_seq,
1328 int dup_sack_in, int *fack_count,
1329 int *reord, int *flag)
1331 tcp_for_write_queue_from(skb, sk) {
1333 int dup_sack = dup_sack_in;
1335 if (skb == tcp_send_head(sk))
1338 /* queue is in-order => we can short-circuit the walk early */
1339 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1342 if ((next_dup != NULL) &&
1343 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1344 in_sack = tcp_match_skb_to_sack(sk, skb,
1345 next_dup->start_seq,
1352 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq,
1354 if (unlikely(in_sack < 0))
1358 *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack,
1361 *fack_count += tcp_skb_pcount(skb);
1366 /* Avoid all extra work that is being done by sacktag while walking in
1369 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1370 u32 skip_to_seq, int *fack_count)
1372 tcp_for_write_queue_from(skb, sk) {
1373 if (skb == tcp_send_head(sk))
1376 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1379 *fack_count += tcp_skb_pcount(skb);
1384 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1386 struct tcp_sack_block *next_dup,
1388 int *fack_count, int *reord,
1391 if (next_dup == NULL)
1394 if (before(next_dup->start_seq, skip_to_seq)) {
1395 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count);
1396 tcp_sacktag_walk(skb, sk, NULL,
1397 next_dup->start_seq, next_dup->end_seq,
1398 1, fack_count, reord, flag);
1404 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1406 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1410 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1413 const struct inet_connection_sock *icsk = inet_csk(sk);
1414 struct tcp_sock *tp = tcp_sk(sk);
1415 unsigned char *ptr = (skb_transport_header(ack_skb) +
1416 TCP_SKB_CB(ack_skb)->sacked);
1417 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1418 struct tcp_sack_block sp[4];
1419 struct tcp_sack_block *cache;
1420 struct sk_buff *skb;
1421 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE) >> 3;
1423 int reord = tp->packets_out;
1425 int found_dup_sack = 0;
1428 int first_sack_index;
1430 if (!tp->sacked_out) {
1431 if (WARN_ON(tp->fackets_out))
1432 tp->fackets_out = 0;
1433 tcp_highest_sack_reset(sk);
1436 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp_wire,
1437 num_sacks, prior_snd_una);
1439 flag |= FLAG_DSACKING_ACK;
1441 /* Eliminate too old ACKs, but take into
1442 * account more or less fresh ones, they can
1443 * contain valid SACK info.
1445 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1448 if (!tp->packets_out)
1452 first_sack_index = 0;
1453 for (i = 0; i < num_sacks; i++) {
1454 int dup_sack = !i && found_dup_sack;
1456 sp[used_sacks].start_seq = ntohl(get_unaligned(&sp_wire[i].start_seq));
1457 sp[used_sacks].end_seq = ntohl(get_unaligned(&sp_wire[i].end_seq));
1459 if (!tcp_is_sackblock_valid(tp, dup_sack,
1460 sp[used_sacks].start_seq,
1461 sp[used_sacks].end_seq)) {
1463 if (!tp->undo_marker)
1464 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1466 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1468 /* Don't count olds caused by ACK reordering */
1469 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1470 !after(sp[used_sacks].end_seq, tp->snd_una))
1472 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1475 first_sack_index = -1;
1479 /* Ignore very old stuff early */
1480 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1486 /* order SACK blocks to allow in order walk of the retrans queue */
1487 for (i = used_sacks - 1; i > 0; i--) {
1488 for (j = 0; j < i; j++) {
1489 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1490 struct tcp_sack_block tmp;
1496 /* Track where the first SACK block goes to */
1497 if (j == first_sack_index)
1498 first_sack_index = j + 1;
1503 skb = tcp_write_queue_head(sk);
1507 if (!tp->sacked_out) {
1508 /* It's already past, so skip checking against it */
1509 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1511 cache = tp->recv_sack_cache;
1512 /* Skip empty blocks in at head of the cache */
1513 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1518 while (i < used_sacks) {
1519 u32 start_seq = sp[i].start_seq;
1520 u32 end_seq = sp[i].end_seq;
1521 int dup_sack = (found_dup_sack && (i == first_sack_index));
1522 struct tcp_sack_block *next_dup = NULL;
1524 if (found_dup_sack && ((i + 1) == first_sack_index))
1525 next_dup = &sp[i + 1];
1527 /* Event "B" in the comment above. */
1528 if (after(end_seq, tp->high_seq))
1529 flag |= FLAG_DATA_LOST;
1531 /* Skip too early cached blocks */
1532 while (tcp_sack_cache_ok(tp, cache) &&
1533 !before(start_seq, cache->end_seq))
1536 /* Can skip some work by looking recv_sack_cache? */
1537 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1538 after(end_seq, cache->start_seq)) {
1541 if (before(start_seq, cache->start_seq)) {
1542 skb = tcp_sacktag_skip(skb, sk, start_seq,
1544 skb = tcp_sacktag_walk(skb, sk, next_dup,
1547 dup_sack, &fack_count,
1551 /* Rest of the block already fully processed? */
1552 if (!after(end_seq, cache->end_seq))
1555 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1557 &fack_count, &reord,
1560 /* ...tail remains todo... */
1561 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1562 /* ...but better entrypoint exists! */
1563 skb = tcp_highest_sack(sk);
1566 fack_count = tp->fackets_out;
1571 skb = tcp_sacktag_skip(skb, sk, cache->end_seq,
1573 /* Check overlap against next cached too (past this one already) */
1578 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1579 skb = tcp_highest_sack(sk);
1582 fack_count = tp->fackets_out;
1584 skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count);
1587 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1588 dup_sack, &fack_count, &reord, &flag);
1591 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1592 * due to in-order walk
1594 if (after(end_seq, tp->frto_highmark))
1595 flag &= ~FLAG_ONLY_ORIG_SACKED;
1600 /* Clear the head of the cache sack blocks so we can skip it next time */
1601 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1602 tp->recv_sack_cache[i].start_seq = 0;
1603 tp->recv_sack_cache[i].end_seq = 0;
1605 for (j = 0; j < used_sacks; j++)
1606 tp->recv_sack_cache[i++] = sp[j];
1608 tcp_mark_lost_retrans(sk);
1610 tcp_verify_left_out(tp);
1612 if ((reord < tp->fackets_out) &&
1613 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1614 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1615 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1619 #if FASTRETRANS_DEBUG > 0
1620 BUG_TRAP((int)tp->sacked_out >= 0);
1621 BUG_TRAP((int)tp->lost_out >= 0);
1622 BUG_TRAP((int)tp->retrans_out >= 0);
1623 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1628 /* If we receive more dupacks than we expected counting segments
1629 * in assumption of absent reordering, interpret this as reordering.
1630 * The only another reason could be bug in receiver TCP.
1632 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1634 struct tcp_sock *tp = tcp_sk(sk);
1637 holes = max(tp->lost_out, 1U);
1638 holes = min(holes, tp->packets_out);
1640 if ((tp->sacked_out + holes) > tp->packets_out) {
1641 tp->sacked_out = tp->packets_out - holes;
1642 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1646 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1648 static void tcp_add_reno_sack(struct sock *sk)
1650 struct tcp_sock *tp = tcp_sk(sk);
1652 tcp_check_reno_reordering(sk, 0);
1653 tcp_verify_left_out(tp);
1656 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1658 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1660 struct tcp_sock *tp = tcp_sk(sk);
1663 /* One ACK acked hole. The rest eat duplicate ACKs. */
1664 if (acked - 1 >= tp->sacked_out)
1667 tp->sacked_out -= acked - 1;
1669 tcp_check_reno_reordering(sk, acked);
1670 tcp_verify_left_out(tp);
1673 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1678 /* F-RTO can only be used if TCP has never retransmitted anything other than
1679 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1681 int tcp_use_frto(struct sock *sk)
1683 const struct tcp_sock *tp = tcp_sk(sk);
1684 struct sk_buff *skb;
1686 if (!sysctl_tcp_frto)
1692 /* Avoid expensive walking of rexmit queue if possible */
1693 if (tp->retrans_out > 1)
1696 skb = tcp_write_queue_head(sk);
1697 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1698 tcp_for_write_queue_from(skb, sk) {
1699 if (skb == tcp_send_head(sk))
1701 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1703 /* Short-circuit when first non-SACKed skb has been checked */
1704 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1710 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1711 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1712 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1713 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1714 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1715 * bits are handled if the Loss state is really to be entered (in
1716 * tcp_enter_frto_loss).
1718 * Do like tcp_enter_loss() would; when RTO expires the second time it
1720 * "Reduce ssthresh if it has not yet been made inside this window."
1722 void tcp_enter_frto(struct sock *sk)
1724 const struct inet_connection_sock *icsk = inet_csk(sk);
1725 struct tcp_sock *tp = tcp_sk(sk);
1726 struct sk_buff *skb;
1728 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1729 tp->snd_una == tp->high_seq ||
1730 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1731 !icsk->icsk_retransmits)) {
1732 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1733 /* Our state is too optimistic in ssthresh() call because cwnd
1734 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1735 * recovery has not yet completed. Pattern would be this: RTO,
1736 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1738 * RFC4138 should be more specific on what to do, even though
1739 * RTO is quite unlikely to occur after the first Cumulative ACK
1740 * due to back-off and complexity of triggering events ...
1742 if (tp->frto_counter) {
1744 stored_cwnd = tp->snd_cwnd;
1746 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1747 tp->snd_cwnd = stored_cwnd;
1749 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1751 /* ... in theory, cong.control module could do "any tricks" in
1752 * ssthresh(), which means that ca_state, lost bits and lost_out
1753 * counter would have to be faked before the call occurs. We
1754 * consider that too expensive, unlikely and hacky, so modules
1755 * using these in ssthresh() must deal these incompatibility
1756 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1758 tcp_ca_event(sk, CA_EVENT_FRTO);
1761 tp->undo_marker = tp->snd_una;
1762 tp->undo_retrans = 0;
1764 skb = tcp_write_queue_head(sk);
1765 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1766 tp->undo_marker = 0;
1767 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1768 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1769 tp->retrans_out -= tcp_skb_pcount(skb);
1771 tcp_verify_left_out(tp);
1773 /* Too bad if TCP was application limited */
1774 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1776 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1777 * The last condition is necessary at least in tp->frto_counter case.
1779 if (IsSackFrto() && (tp->frto_counter ||
1780 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1781 after(tp->high_seq, tp->snd_una)) {
1782 tp->frto_highmark = tp->high_seq;
1784 tp->frto_highmark = tp->snd_nxt;
1786 tcp_set_ca_state(sk, TCP_CA_Disorder);
1787 tp->high_seq = tp->snd_nxt;
1788 tp->frto_counter = 1;
1791 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1792 * which indicates that we should follow the traditional RTO recovery,
1793 * i.e. mark everything lost and do go-back-N retransmission.
1795 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1797 struct tcp_sock *tp = tcp_sk(sk);
1798 struct sk_buff *skb;
1801 tp->retrans_out = 0;
1802 if (tcp_is_reno(tp))
1803 tcp_reset_reno_sack(tp);
1805 tcp_for_write_queue(skb, sk) {
1806 if (skb == tcp_send_head(sk))
1809 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1811 * Count the retransmission made on RTO correctly (only when
1812 * waiting for the first ACK and did not get it)...
1814 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1815 /* For some reason this R-bit might get cleared? */
1816 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1817 tp->retrans_out += tcp_skb_pcount(skb);
1818 /* ...enter this if branch just for the first segment */
1819 flag |= FLAG_DATA_ACKED;
1821 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1822 tp->undo_marker = 0;
1823 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1826 /* Don't lost mark skbs that were fwd transmitted after RTO */
1827 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) &&
1828 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1829 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1830 tp->lost_out += tcp_skb_pcount(skb);
1833 tcp_verify_left_out(tp);
1835 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1836 tp->snd_cwnd_cnt = 0;
1837 tp->snd_cwnd_stamp = tcp_time_stamp;
1838 tp->frto_counter = 0;
1839 tp->bytes_acked = 0;
1841 tp->reordering = min_t(unsigned int, tp->reordering,
1842 sysctl_tcp_reordering);
1843 tcp_set_ca_state(sk, TCP_CA_Loss);
1844 tp->high_seq = tp->frto_highmark;
1845 TCP_ECN_queue_cwr(tp);
1847 tcp_clear_retrans_hints_partial(tp);
1850 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1852 tp->retrans_out = 0;
1855 tp->undo_marker = 0;
1856 tp->undo_retrans = 0;
1859 void tcp_clear_retrans(struct tcp_sock *tp)
1861 tcp_clear_retrans_partial(tp);
1863 tp->fackets_out = 0;
1867 /* Enter Loss state. If "how" is not zero, forget all SACK information
1868 * and reset tags completely, otherwise preserve SACKs. If receiver
1869 * dropped its ofo queue, we will know this due to reneging detection.
1871 void tcp_enter_loss(struct sock *sk, int how)
1873 const struct inet_connection_sock *icsk = inet_csk(sk);
1874 struct tcp_sock *tp = tcp_sk(sk);
1875 struct sk_buff *skb;
1877 /* Reduce ssthresh if it has not yet been made inside this window. */
1878 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1879 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1880 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1881 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1882 tcp_ca_event(sk, CA_EVENT_LOSS);
1885 tp->snd_cwnd_cnt = 0;
1886 tp->snd_cwnd_stamp = tcp_time_stamp;
1888 tp->bytes_acked = 0;
1889 tcp_clear_retrans_partial(tp);
1891 if (tcp_is_reno(tp))
1892 tcp_reset_reno_sack(tp);
1895 /* Push undo marker, if it was plain RTO and nothing
1896 * was retransmitted. */
1897 tp->undo_marker = tp->snd_una;
1898 tcp_clear_retrans_hints_partial(tp);
1901 tp->fackets_out = 0;
1902 tcp_clear_all_retrans_hints(tp);
1905 tcp_for_write_queue(skb, sk) {
1906 if (skb == tcp_send_head(sk))
1909 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1910 tp->undo_marker = 0;
1911 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1912 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1913 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1914 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1915 tp->lost_out += tcp_skb_pcount(skb);
1918 tcp_verify_left_out(tp);
1920 tp->reordering = min_t(unsigned int, tp->reordering,
1921 sysctl_tcp_reordering);
1922 tcp_set_ca_state(sk, TCP_CA_Loss);
1923 tp->high_seq = tp->snd_nxt;
1924 TCP_ECN_queue_cwr(tp);
1925 /* Abort F-RTO algorithm if one is in progress */
1926 tp->frto_counter = 0;
1929 /* If ACK arrived pointing to a remembered SACK, it means that our
1930 * remembered SACKs do not reflect real state of receiver i.e.
1931 * receiver _host_ is heavily congested (or buggy).
1933 * Do processing similar to RTO timeout.
1935 static int tcp_check_sack_reneging(struct sock *sk, int flag)
1937 if (flag & FLAG_SACK_RENEGING) {
1938 struct inet_connection_sock *icsk = inet_csk(sk);
1939 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1941 tcp_enter_loss(sk, 1);
1942 icsk->icsk_retransmits++;
1943 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1944 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1945 icsk->icsk_rto, TCP_RTO_MAX);
1951 static inline int tcp_fackets_out(struct tcp_sock *tp)
1953 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1956 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1957 * counter when SACK is enabled (without SACK, sacked_out is used for
1960 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1961 * segments up to the highest received SACK block so far and holes in
1964 * With reordering, holes may still be in flight, so RFC3517 recovery
1965 * uses pure sacked_out (total number of SACKed segments) even though
1966 * it violates the RFC that uses duplicate ACKs, often these are equal
1967 * but when e.g. out-of-window ACKs or packet duplication occurs,
1968 * they differ. Since neither occurs due to loss, TCP should really
1971 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
1973 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1976 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1978 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1981 static inline int tcp_head_timedout(struct sock *sk)
1983 struct tcp_sock *tp = tcp_sk(sk);
1985 return tp->packets_out &&
1986 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1989 /* Linux NewReno/SACK/FACK/ECN state machine.
1990 * --------------------------------------
1992 * "Open" Normal state, no dubious events, fast path.
1993 * "Disorder" In all the respects it is "Open",
1994 * but requires a bit more attention. It is entered when
1995 * we see some SACKs or dupacks. It is split of "Open"
1996 * mainly to move some processing from fast path to slow one.
1997 * "CWR" CWND was reduced due to some Congestion Notification event.
1998 * It can be ECN, ICMP source quench, local device congestion.
1999 * "Recovery" CWND was reduced, we are fast-retransmitting.
2000 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2002 * tcp_fastretrans_alert() is entered:
2003 * - each incoming ACK, if state is not "Open"
2004 * - when arrived ACK is unusual, namely:
2009 * Counting packets in flight is pretty simple.
2011 * in_flight = packets_out - left_out + retrans_out
2013 * packets_out is SND.NXT-SND.UNA counted in packets.
2015 * retrans_out is number of retransmitted segments.
2017 * left_out is number of segments left network, but not ACKed yet.
2019 * left_out = sacked_out + lost_out
2021 * sacked_out: Packets, which arrived to receiver out of order
2022 * and hence not ACKed. With SACKs this number is simply
2023 * amount of SACKed data. Even without SACKs
2024 * it is easy to give pretty reliable estimate of this number,
2025 * counting duplicate ACKs.
2027 * lost_out: Packets lost by network. TCP has no explicit
2028 * "loss notification" feedback from network (for now).
2029 * It means that this number can be only _guessed_.
2030 * Actually, it is the heuristics to predict lossage that
2031 * distinguishes different algorithms.
2033 * F.e. after RTO, when all the queue is considered as lost,
2034 * lost_out = packets_out and in_flight = retrans_out.
2036 * Essentially, we have now two algorithms counting
2039 * FACK: It is the simplest heuristics. As soon as we decided
2040 * that something is lost, we decide that _all_ not SACKed
2041 * packets until the most forward SACK are lost. I.e.
2042 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2043 * It is absolutely correct estimate, if network does not reorder
2044 * packets. And it loses any connection to reality when reordering
2045 * takes place. We use FACK by default until reordering
2046 * is suspected on the path to this destination.
2048 * NewReno: when Recovery is entered, we assume that one segment
2049 * is lost (classic Reno). While we are in Recovery and
2050 * a partial ACK arrives, we assume that one more packet
2051 * is lost (NewReno). This heuristics are the same in NewReno
2054 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2055 * deflation etc. CWND is real congestion window, never inflated, changes
2056 * only according to classic VJ rules.
2058 * Really tricky (and requiring careful tuning) part of algorithm
2059 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2060 * The first determines the moment _when_ we should reduce CWND and,
2061 * hence, slow down forward transmission. In fact, it determines the moment
2062 * when we decide that hole is caused by loss, rather than by a reorder.
2064 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2065 * holes, caused by lost packets.
2067 * And the most logically complicated part of algorithm is undo
2068 * heuristics. We detect false retransmits due to both too early
2069 * fast retransmit (reordering) and underestimated RTO, analyzing
2070 * timestamps and D-SACKs. When we detect that some segments were
2071 * retransmitted by mistake and CWND reduction was wrong, we undo
2072 * window reduction and abort recovery phase. This logic is hidden
2073 * inside several functions named tcp_try_undo_<something>.
2076 /* This function decides, when we should leave Disordered state
2077 * and enter Recovery phase, reducing congestion window.
2079 * Main question: may we further continue forward transmission
2080 * with the same cwnd?
2082 static int tcp_time_to_recover(struct sock *sk)
2084 struct tcp_sock *tp = tcp_sk(sk);
2087 /* Do not perform any recovery during F-RTO algorithm */
2088 if (tp->frto_counter)
2091 /* Trick#1: The loss is proven. */
2095 /* Not-A-Trick#2 : Classic rule... */
2096 if (tcp_dupack_heurestics(tp) > tp->reordering)
2099 /* Trick#3 : when we use RFC2988 timer restart, fast
2100 * retransmit can be triggered by timeout of queue head.
2102 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2105 /* Trick#4: It is still not OK... But will it be useful to delay
2108 packets_out = tp->packets_out;
2109 if (packets_out <= tp->reordering &&
2110 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2111 !tcp_may_send_now(sk)) {
2112 /* We have nothing to send. This connection is limited
2113 * either by receiver window or by application.
2121 /* RFC: This is from the original, I doubt that this is necessary at all:
2122 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2123 * retransmitted past LOST markings in the first place? I'm not fully sure
2124 * about undo and end of connection cases, which can cause R without L?
2126 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
2128 if ((tp->retransmit_skb_hint != NULL) &&
2129 before(TCP_SKB_CB(skb)->seq,
2130 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
2131 tp->retransmit_skb_hint = NULL;
2134 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2135 * is against sacked "cnt", otherwise it's against facked "cnt"
2137 static void tcp_mark_head_lost(struct sock *sk, int packets)
2139 struct tcp_sock *tp = tcp_sk(sk);
2140 struct sk_buff *skb;
2145 BUG_TRAP(packets <= tp->packets_out);
2146 if (tp->lost_skb_hint) {
2147 skb = tp->lost_skb_hint;
2148 cnt = tp->lost_cnt_hint;
2150 skb = tcp_write_queue_head(sk);
2154 tcp_for_write_queue_from(skb, sk) {
2155 if (skb == tcp_send_head(sk))
2157 /* TODO: do this better */
2158 /* this is not the most efficient way to do this... */
2159 tp->lost_skb_hint = skb;
2160 tp->lost_cnt_hint = cnt;
2162 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2166 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2167 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2168 cnt += tcp_skb_pcount(skb);
2170 if (cnt > packets) {
2171 if (tcp_is_sack(tp) || (oldcnt >= packets))
2174 mss = skb_shinfo(skb)->gso_size;
2175 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2181 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2182 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2183 tp->lost_out += tcp_skb_pcount(skb);
2184 tcp_verify_retransmit_hint(tp, skb);
2187 tcp_verify_left_out(tp);
2190 /* Account newly detected lost packet(s) */
2192 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2194 struct tcp_sock *tp = tcp_sk(sk);
2196 if (tcp_is_reno(tp)) {
2197 tcp_mark_head_lost(sk, 1);
2198 } else if (tcp_is_fack(tp)) {
2199 int lost = tp->fackets_out - tp->reordering;
2202 tcp_mark_head_lost(sk, lost);
2204 int sacked_upto = tp->sacked_out - tp->reordering;
2205 if (sacked_upto < fast_rexmit)
2206 sacked_upto = fast_rexmit;
2207 tcp_mark_head_lost(sk, sacked_upto);
2210 /* New heuristics: it is possible only after we switched
2211 * to restart timer each time when something is ACKed.
2212 * Hence, we can detect timed out packets during fast
2213 * retransmit without falling to slow start.
2215 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2216 struct sk_buff *skb;
2218 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2219 : tcp_write_queue_head(sk);
2221 tcp_for_write_queue_from(skb, sk) {
2222 if (skb == tcp_send_head(sk))
2224 if (!tcp_skb_timedout(sk, skb))
2227 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
2228 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2229 tp->lost_out += tcp_skb_pcount(skb);
2230 tcp_verify_retransmit_hint(tp, skb);
2234 tp->scoreboard_skb_hint = skb;
2236 tcp_verify_left_out(tp);
2240 /* CWND moderation, preventing bursts due to too big ACKs
2241 * in dubious situations.
2243 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2245 tp->snd_cwnd = min(tp->snd_cwnd,
2246 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2247 tp->snd_cwnd_stamp = tcp_time_stamp;
2250 /* Lower bound on congestion window is slow start threshold
2251 * unless congestion avoidance choice decides to overide it.
2253 static inline u32 tcp_cwnd_min(const struct sock *sk)
2255 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2257 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2260 /* Decrease cwnd each second ack. */
2261 static void tcp_cwnd_down(struct sock *sk, int flag)
2263 struct tcp_sock *tp = tcp_sk(sk);
2264 int decr = tp->snd_cwnd_cnt + 1;
2266 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2267 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2268 tp->snd_cwnd_cnt = decr & 1;
2271 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2272 tp->snd_cwnd -= decr;
2274 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2275 tp->snd_cwnd_stamp = tcp_time_stamp;
2279 /* Nothing was retransmitted or returned timestamp is less
2280 * than timestamp of the first retransmission.
2282 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2284 return !tp->retrans_stamp ||
2285 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2286 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2289 /* Undo procedures. */
2291 #if FASTRETRANS_DEBUG > 1
2292 static void DBGUNDO(struct sock *sk, const char *msg)
2294 struct tcp_sock *tp = tcp_sk(sk);
2295 struct inet_sock *inet = inet_sk(sk);
2297 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2299 NIPQUAD(inet->daddr), ntohs(inet->dport),
2300 tp->snd_cwnd, tcp_left_out(tp),
2301 tp->snd_ssthresh, tp->prior_ssthresh,
2305 #define DBGUNDO(x...) do { } while (0)
2308 static void tcp_undo_cwr(struct sock *sk, const int undo)
2310 struct tcp_sock *tp = tcp_sk(sk);
2312 if (tp->prior_ssthresh) {
2313 const struct inet_connection_sock *icsk = inet_csk(sk);
2315 if (icsk->icsk_ca_ops->undo_cwnd)
2316 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2318 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2320 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2321 tp->snd_ssthresh = tp->prior_ssthresh;
2322 TCP_ECN_withdraw_cwr(tp);
2325 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2327 tcp_moderate_cwnd(tp);
2328 tp->snd_cwnd_stamp = tcp_time_stamp;
2330 /* There is something screwy going on with the retrans hints after
2332 tcp_clear_all_retrans_hints(tp);
2335 static inline int tcp_may_undo(struct tcp_sock *tp)
2337 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2340 /* People celebrate: "We love our President!" */
2341 static int tcp_try_undo_recovery(struct sock *sk)
2343 struct tcp_sock *tp = tcp_sk(sk);
2345 if (tcp_may_undo(tp)) {
2346 /* Happy end! We did not retransmit anything
2347 * or our original transmission succeeded.
2349 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2350 tcp_undo_cwr(sk, 1);
2351 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2352 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2354 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2355 tp->undo_marker = 0;
2357 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2358 /* Hold old state until something *above* high_seq
2359 * is ACKed. For Reno it is MUST to prevent false
2360 * fast retransmits (RFC2582). SACK TCP is safe. */
2361 tcp_moderate_cwnd(tp);
2364 tcp_set_ca_state(sk, TCP_CA_Open);
2368 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2369 static void tcp_try_undo_dsack(struct sock *sk)
2371 struct tcp_sock *tp = tcp_sk(sk);
2373 if (tp->undo_marker && !tp->undo_retrans) {
2374 DBGUNDO(sk, "D-SACK");
2375 tcp_undo_cwr(sk, 1);
2376 tp->undo_marker = 0;
2377 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2381 /* Undo during fast recovery after partial ACK. */
2383 static int tcp_try_undo_partial(struct sock *sk, int acked)
2385 struct tcp_sock *tp = tcp_sk(sk);
2386 /* Partial ACK arrived. Force Hoe's retransmit. */
2387 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2389 if (tcp_may_undo(tp)) {
2390 /* Plain luck! Hole if filled with delayed
2391 * packet, rather than with a retransmit.
2393 if (tp->retrans_out == 0)
2394 tp->retrans_stamp = 0;
2396 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2399 tcp_undo_cwr(sk, 0);
2400 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2402 /* So... Do not make Hoe's retransmit yet.
2403 * If the first packet was delayed, the rest
2404 * ones are most probably delayed as well.
2411 /* Undo during loss recovery after partial ACK. */
2412 static int tcp_try_undo_loss(struct sock *sk)
2414 struct tcp_sock *tp = tcp_sk(sk);
2416 if (tcp_may_undo(tp)) {
2417 struct sk_buff *skb;
2418 tcp_for_write_queue(skb, sk) {
2419 if (skb == tcp_send_head(sk))
2421 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2424 tcp_clear_all_retrans_hints(tp);
2426 DBGUNDO(sk, "partial loss");
2428 tcp_undo_cwr(sk, 1);
2429 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2430 inet_csk(sk)->icsk_retransmits = 0;
2431 tp->undo_marker = 0;
2432 if (tcp_is_sack(tp))
2433 tcp_set_ca_state(sk, TCP_CA_Open);
2439 static inline void tcp_complete_cwr(struct sock *sk)
2441 struct tcp_sock *tp = tcp_sk(sk);
2442 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2443 tp->snd_cwnd_stamp = tcp_time_stamp;
2444 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2447 static void tcp_try_to_open(struct sock *sk, int flag)
2449 struct tcp_sock *tp = tcp_sk(sk);
2451 tcp_verify_left_out(tp);
2453 if (tp->retrans_out == 0)
2454 tp->retrans_stamp = 0;
2456 if (flag & FLAG_ECE)
2457 tcp_enter_cwr(sk, 1);
2459 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2460 int state = TCP_CA_Open;
2462 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2463 state = TCP_CA_Disorder;
2465 if (inet_csk(sk)->icsk_ca_state != state) {
2466 tcp_set_ca_state(sk, state);
2467 tp->high_seq = tp->snd_nxt;
2469 tcp_moderate_cwnd(tp);
2471 tcp_cwnd_down(sk, flag);
2475 static void tcp_mtup_probe_failed(struct sock *sk)
2477 struct inet_connection_sock *icsk = inet_csk(sk);
2479 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2480 icsk->icsk_mtup.probe_size = 0;
2483 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2485 struct tcp_sock *tp = tcp_sk(sk);
2486 struct inet_connection_sock *icsk = inet_csk(sk);
2488 /* FIXME: breaks with very large cwnd */
2489 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2490 tp->snd_cwnd = tp->snd_cwnd *
2491 tcp_mss_to_mtu(sk, tp->mss_cache) /
2492 icsk->icsk_mtup.probe_size;
2493 tp->snd_cwnd_cnt = 0;
2494 tp->snd_cwnd_stamp = tcp_time_stamp;
2495 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2497 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2498 icsk->icsk_mtup.probe_size = 0;
2499 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2502 /* Process an event, which can update packets-in-flight not trivially.
2503 * Main goal of this function is to calculate new estimate for left_out,
2504 * taking into account both packets sitting in receiver's buffer and
2505 * packets lost by network.
2507 * Besides that it does CWND reduction, when packet loss is detected
2508 * and changes state of machine.
2510 * It does _not_ decide what to send, it is made in function
2511 * tcp_xmit_retransmit_queue().
2513 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2515 struct inet_connection_sock *icsk = inet_csk(sk);
2516 struct tcp_sock *tp = tcp_sk(sk);
2517 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2518 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2519 (tcp_fackets_out(tp) > tp->reordering));
2520 int fast_rexmit = 0;
2522 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2524 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2525 tp->fackets_out = 0;
2527 /* Now state machine starts.
2528 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2529 if (flag & FLAG_ECE)
2530 tp->prior_ssthresh = 0;
2532 /* B. In all the states check for reneging SACKs. */
2533 if (tcp_check_sack_reneging(sk, flag))
2536 /* C. Process data loss notification, provided it is valid. */
2537 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2538 before(tp->snd_una, tp->high_seq) &&
2539 icsk->icsk_ca_state != TCP_CA_Open &&
2540 tp->fackets_out > tp->reordering) {
2541 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2542 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2545 /* D. Check consistency of the current state. */
2546 tcp_verify_left_out(tp);
2548 /* E. Check state exit conditions. State can be terminated
2549 * when high_seq is ACKed. */
2550 if (icsk->icsk_ca_state == TCP_CA_Open) {
2551 BUG_TRAP(tp->retrans_out == 0);
2552 tp->retrans_stamp = 0;
2553 } else if (!before(tp->snd_una, tp->high_seq)) {
2554 switch (icsk->icsk_ca_state) {
2556 icsk->icsk_retransmits = 0;
2557 if (tcp_try_undo_recovery(sk))
2562 /* CWR is to be held something *above* high_seq
2563 * is ACKed for CWR bit to reach receiver. */
2564 if (tp->snd_una != tp->high_seq) {
2565 tcp_complete_cwr(sk);
2566 tcp_set_ca_state(sk, TCP_CA_Open);
2570 case TCP_CA_Disorder:
2571 tcp_try_undo_dsack(sk);
2572 if (!tp->undo_marker ||
2573 /* For SACK case do not Open to allow to undo
2574 * catching for all duplicate ACKs. */
2575 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2576 tp->undo_marker = 0;
2577 tcp_set_ca_state(sk, TCP_CA_Open);
2581 case TCP_CA_Recovery:
2582 if (tcp_is_reno(tp))
2583 tcp_reset_reno_sack(tp);
2584 if (tcp_try_undo_recovery(sk))
2586 tcp_complete_cwr(sk);
2591 /* F. Process state. */
2592 switch (icsk->icsk_ca_state) {
2593 case TCP_CA_Recovery:
2594 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2595 if (tcp_is_reno(tp) && is_dupack)
2596 tcp_add_reno_sack(sk);
2598 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2601 if (flag & FLAG_DATA_ACKED)
2602 icsk->icsk_retransmits = 0;
2603 if (!tcp_try_undo_loss(sk)) {
2604 tcp_moderate_cwnd(tp);
2605 tcp_xmit_retransmit_queue(sk);
2608 if (icsk->icsk_ca_state != TCP_CA_Open)
2610 /* Loss is undone; fall through to processing in Open state. */
2612 if (tcp_is_reno(tp)) {
2613 if (flag & FLAG_SND_UNA_ADVANCED)
2614 tcp_reset_reno_sack(tp);
2616 tcp_add_reno_sack(sk);
2619 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2620 tcp_try_undo_dsack(sk);
2622 if (!tcp_time_to_recover(sk)) {
2623 tcp_try_to_open(sk, flag);
2627 /* MTU probe failure: don't reduce cwnd */
2628 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2629 icsk->icsk_mtup.probe_size &&
2630 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2631 tcp_mtup_probe_failed(sk);
2632 /* Restores the reduction we did in tcp_mtup_probe() */
2634 tcp_simple_retransmit(sk);
2638 /* Otherwise enter Recovery state */
2640 if (tcp_is_reno(tp))
2641 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2643 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2645 tp->high_seq = tp->snd_nxt;
2646 tp->prior_ssthresh = 0;
2647 tp->undo_marker = tp->snd_una;
2648 tp->undo_retrans = tp->retrans_out;
2650 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2651 if (!(flag & FLAG_ECE))
2652 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2653 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2654 TCP_ECN_queue_cwr(tp);
2657 tp->bytes_acked = 0;
2658 tp->snd_cwnd_cnt = 0;
2659 tcp_set_ca_state(sk, TCP_CA_Recovery);
2663 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2664 tcp_update_scoreboard(sk, fast_rexmit);
2665 tcp_cwnd_down(sk, flag);
2666 tcp_xmit_retransmit_queue(sk);
2669 /* Read draft-ietf-tcplw-high-performance before mucking
2670 * with this code. (Supersedes RFC1323)
2672 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2674 /* RTTM Rule: A TSecr value received in a segment is used to
2675 * update the averaged RTT measurement only if the segment
2676 * acknowledges some new data, i.e., only if it advances the
2677 * left edge of the send window.
2679 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2680 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2682 * Changed: reset backoff as soon as we see the first valid sample.
2683 * If we do not, we get strongly overestimated rto. With timestamps
2684 * samples are accepted even from very old segments: f.e., when rtt=1
2685 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2686 * answer arrives rto becomes 120 seconds! If at least one of segments
2687 * in window is lost... Voila. --ANK (010210)
2689 struct tcp_sock *tp = tcp_sk(sk);
2690 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2691 tcp_rtt_estimator(sk, seq_rtt);
2693 inet_csk(sk)->icsk_backoff = 0;
2697 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2699 /* We don't have a timestamp. Can only use
2700 * packets that are not retransmitted to determine
2701 * rtt estimates. Also, we must not reset the
2702 * backoff for rto until we get a non-retransmitted
2703 * packet. This allows us to deal with a situation
2704 * where the network delay has increased suddenly.
2705 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2708 if (flag & FLAG_RETRANS_DATA_ACKED)
2711 tcp_rtt_estimator(sk, seq_rtt);
2713 inet_csk(sk)->icsk_backoff = 0;
2717 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2720 const struct tcp_sock *tp = tcp_sk(sk);
2721 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2722 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2723 tcp_ack_saw_tstamp(sk, flag);
2724 else if (seq_rtt >= 0)
2725 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2728 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2730 const struct inet_connection_sock *icsk = inet_csk(sk);
2731 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2732 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2735 /* Restart timer after forward progress on connection.
2736 * RFC2988 recommends to restart timer to now+rto.
2738 static void tcp_rearm_rto(struct sock *sk)
2740 struct tcp_sock *tp = tcp_sk(sk);
2742 if (!tp->packets_out) {
2743 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2745 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2746 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2750 /* If we get here, the whole TSO packet has not been acked. */
2751 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2753 struct tcp_sock *tp = tcp_sk(sk);
2756 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2758 packets_acked = tcp_skb_pcount(skb);
2759 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2761 packets_acked -= tcp_skb_pcount(skb);
2763 if (packets_acked) {
2764 BUG_ON(tcp_skb_pcount(skb) == 0);
2765 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2768 return packets_acked;
2771 /* Remove acknowledged frames from the retransmission queue. If our packet
2772 * is before the ack sequence we can discard it as it's confirmed to have
2773 * arrived at the other end.
2775 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
2777 struct tcp_sock *tp = tcp_sk(sk);
2778 const struct inet_connection_sock *icsk = inet_csk(sk);
2779 struct sk_buff *skb;
2780 u32 now = tcp_time_stamp;
2781 int fully_acked = 1;
2784 u32 reord = tp->packets_out;
2786 s32 ca_seq_rtt = -1;
2787 ktime_t last_ackt = net_invalid_timestamp();
2789 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2790 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2793 u8 sacked = scb->sacked;
2795 /* Determine how many packets and what bytes were acked, tso and else */
2796 if (after(scb->end_seq, tp->snd_una)) {
2797 if (tcp_skb_pcount(skb) == 1 ||
2798 !after(tp->snd_una, scb->seq))
2801 acked_pcount = tcp_tso_acked(sk, skb);
2806 end_seq = tp->snd_una;
2808 acked_pcount = tcp_skb_pcount(skb);
2809 end_seq = scb->end_seq;
2812 /* MTU probing checks */
2813 if (fully_acked && icsk->icsk_mtup.probe_size &&
2814 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2815 tcp_mtup_probe_success(sk, skb);
2818 if (sacked & TCPCB_RETRANS) {
2819 if (sacked & TCPCB_SACKED_RETRANS)
2820 tp->retrans_out -= acked_pcount;
2821 flag |= FLAG_RETRANS_DATA_ACKED;
2824 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
2825 flag |= FLAG_NONHEAD_RETRANS_ACKED;
2827 ca_seq_rtt = now - scb->when;
2828 last_ackt = skb->tstamp;
2830 seq_rtt = ca_seq_rtt;
2832 if (!(sacked & TCPCB_SACKED_ACKED))
2833 reord = min(pkts_acked, reord);
2836 if (sacked & TCPCB_SACKED_ACKED)
2837 tp->sacked_out -= acked_pcount;
2838 if (sacked & TCPCB_LOST)
2839 tp->lost_out -= acked_pcount;
2841 if (unlikely(tp->urg_mode && !before(end_seq, tp->snd_up)))
2844 tp->packets_out -= acked_pcount;
2845 pkts_acked += acked_pcount;
2847 /* Initial outgoing SYN's get put onto the write_queue
2848 * just like anything else we transmit. It is not
2849 * true data, and if we misinform our callers that
2850 * this ACK acks real data, we will erroneously exit
2851 * connection startup slow start one packet too
2852 * quickly. This is severely frowned upon behavior.
2854 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2855 flag |= FLAG_DATA_ACKED;
2857 flag |= FLAG_SYN_ACKED;
2858 tp->retrans_stamp = 0;
2864 tcp_unlink_write_queue(skb, sk);
2865 sk_wmem_free_skb(sk, skb);
2866 tcp_clear_all_retrans_hints(tp);
2869 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2870 flag |= FLAG_SACK_RENEGING;
2872 if (flag & FLAG_ACKED) {
2873 const struct tcp_congestion_ops *ca_ops
2874 = inet_csk(sk)->icsk_ca_ops;
2876 tcp_ack_update_rtt(sk, flag, seq_rtt);
2879 if (tcp_is_reno(tp)) {
2880 tcp_remove_reno_sacks(sk, pkts_acked);
2882 /* Non-retransmitted hole got filled? That's reordering */
2883 if (reord < prior_fackets)
2884 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2887 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2889 if (ca_ops->pkts_acked) {
2892 /* Is the ACK triggering packet unambiguous? */
2893 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2894 /* High resolution needed and available? */
2895 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2896 !ktime_equal(last_ackt,
2897 net_invalid_timestamp()))
2898 rtt_us = ktime_us_delta(ktime_get_real(),
2900 else if (ca_seq_rtt > 0)
2901 rtt_us = jiffies_to_usecs(ca_seq_rtt);
2904 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2908 #if FASTRETRANS_DEBUG > 0
2909 BUG_TRAP((int)tp->sacked_out >= 0);
2910 BUG_TRAP((int)tp->lost_out >= 0);
2911 BUG_TRAP((int)tp->retrans_out >= 0);
2912 if (!tp->packets_out && tcp_is_sack(tp)) {
2913 icsk = inet_csk(sk);
2915 printk(KERN_DEBUG "Leak l=%u %d\n",
2916 tp->lost_out, icsk->icsk_ca_state);
2919 if (tp->sacked_out) {
2920 printk(KERN_DEBUG "Leak s=%u %d\n",
2921 tp->sacked_out, icsk->icsk_ca_state);
2924 if (tp->retrans_out) {
2925 printk(KERN_DEBUG "Leak r=%u %d\n",
2926 tp->retrans_out, icsk->icsk_ca_state);
2927 tp->retrans_out = 0;
2934 static void tcp_ack_probe(struct sock *sk)
2936 const struct tcp_sock *tp = tcp_sk(sk);
2937 struct inet_connection_sock *icsk = inet_csk(sk);
2939 /* Was it a usable window open? */
2941 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
2942 icsk->icsk_backoff = 0;
2943 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2944 /* Socket must be waked up by subsequent tcp_data_snd_check().
2945 * This function is not for random using!
2948 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2949 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2954 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2956 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2957 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2960 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2962 const struct tcp_sock *tp = tcp_sk(sk);
2963 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2964 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2967 /* Check that window update is acceptable.
2968 * The function assumes that snd_una<=ack<=snd_next.
2970 static inline int tcp_may_update_window(const struct tcp_sock *tp,
2971 const u32 ack, const u32 ack_seq,
2974 return (after(ack, tp->snd_una) ||
2975 after(ack_seq, tp->snd_wl1) ||
2976 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2979 /* Update our send window.
2981 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2982 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2984 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2987 struct tcp_sock *tp = tcp_sk(sk);
2989 u32 nwin = ntohs(tcp_hdr(skb)->window);
2991 if (likely(!tcp_hdr(skb)->syn))
2992 nwin <<= tp->rx_opt.snd_wscale;
2994 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2995 flag |= FLAG_WIN_UPDATE;
2996 tcp_update_wl(tp, ack, ack_seq);
2998 if (tp->snd_wnd != nwin) {
3001 /* Note, it is the only place, where
3002 * fast path is recovered for sending TCP.
3005 tcp_fast_path_check(sk);
3007 if (nwin > tp->max_window) {
3008 tp->max_window = nwin;
3009 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3019 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3020 * continue in congestion avoidance.
3022 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3024 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3025 tp->snd_cwnd_cnt = 0;
3026 tp->bytes_acked = 0;
3027 TCP_ECN_queue_cwr(tp);
3028 tcp_moderate_cwnd(tp);
3031 /* A conservative spurious RTO response algorithm: reduce cwnd using
3032 * rate halving and continue in congestion avoidance.
3034 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3036 tcp_enter_cwr(sk, 0);
3039 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3041 if (flag & FLAG_ECE)
3042 tcp_ratehalving_spur_to_response(sk);
3044 tcp_undo_cwr(sk, 1);
3047 /* F-RTO spurious RTO detection algorithm (RFC4138)
3049 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3050 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3051 * window (but not to or beyond highest sequence sent before RTO):
3052 * On First ACK, send two new segments out.
3053 * On Second ACK, RTO was likely spurious. Do spurious response (response
3054 * algorithm is not part of the F-RTO detection algorithm
3055 * given in RFC4138 but can be selected separately).
3056 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3057 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3058 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3059 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3061 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3062 * original window even after we transmit two new data segments.
3065 * on first step, wait until first cumulative ACK arrives, then move to
3066 * the second step. In second step, the next ACK decides.
3068 * F-RTO is implemented (mainly) in four functions:
3069 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3070 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3071 * called when tcp_use_frto() showed green light
3072 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3073 * - tcp_enter_frto_loss() is called if there is not enough evidence
3074 * to prove that the RTO is indeed spurious. It transfers the control
3075 * from F-RTO to the conventional RTO recovery
3077 static int tcp_process_frto(struct sock *sk, int flag)
3079 struct tcp_sock *tp = tcp_sk(sk);
3081 tcp_verify_left_out(tp);
3083 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3084 if (flag & FLAG_DATA_ACKED)
3085 inet_csk(sk)->icsk_retransmits = 0;
3087 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3088 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3089 tp->undo_marker = 0;
3091 if (!before(tp->snd_una, tp->frto_highmark)) {
3092 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3096 if (!IsSackFrto() || tcp_is_reno(tp)) {
3097 /* RFC4138 shortcoming in step 2; should also have case c):
3098 * ACK isn't duplicate nor advances window, e.g., opposite dir
3101 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3104 if (!(flag & FLAG_DATA_ACKED)) {
3105 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3110 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3111 /* Prevent sending of new data. */
3112 tp->snd_cwnd = min(tp->snd_cwnd,
3113 tcp_packets_in_flight(tp));
3117 if ((tp->frto_counter >= 2) &&
3118 (!(flag & FLAG_FORWARD_PROGRESS) ||
3119 ((flag & FLAG_DATA_SACKED) &&
3120 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3121 /* RFC4138 shortcoming (see comment above) */
3122 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3123 (flag & FLAG_NOT_DUP))
3126 tcp_enter_frto_loss(sk, 3, flag);
3131 if (tp->frto_counter == 1) {
3132 /* tcp_may_send_now needs to see updated state */
3133 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3134 tp->frto_counter = 2;
3136 if (!tcp_may_send_now(sk))
3137 tcp_enter_frto_loss(sk, 2, flag);
3141 switch (sysctl_tcp_frto_response) {
3143 tcp_undo_spur_to_response(sk, flag);
3146 tcp_conservative_spur_to_response(tp);
3149 tcp_ratehalving_spur_to_response(sk);
3152 tp->frto_counter = 0;
3153 tp->undo_marker = 0;
3154 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
3159 /* This routine deals with incoming acks, but not outgoing ones. */
3160 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3162 struct inet_connection_sock *icsk = inet_csk(sk);
3163 struct tcp_sock *tp = tcp_sk(sk);
3164 u32 prior_snd_una = tp->snd_una;
3165 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3166 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3167 u32 prior_in_flight;
3172 /* If the ack is newer than sent or older than previous acks
3173 * then we can probably ignore it.
3175 if (after(ack, tp->snd_nxt))
3176 goto uninteresting_ack;
3178 if (before(ack, prior_snd_una))
3181 if (after(ack, prior_snd_una))
3182 flag |= FLAG_SND_UNA_ADVANCED;
3184 if (sysctl_tcp_abc) {
3185 if (icsk->icsk_ca_state < TCP_CA_CWR)
3186 tp->bytes_acked += ack - prior_snd_una;
3187 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3188 /* we assume just one segment left network */
3189 tp->bytes_acked += min(ack - prior_snd_una,
3193 prior_fackets = tp->fackets_out;
3194 prior_in_flight = tcp_packets_in_flight(tp);
3196 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3197 /* Window is constant, pure forward advance.
3198 * No more checks are required.
3199 * Note, we use the fact that SND.UNA>=SND.WL2.
3201 tcp_update_wl(tp, ack, ack_seq);
3203 flag |= FLAG_WIN_UPDATE;
3205 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3207 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3209 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3212 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3214 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3216 if (TCP_SKB_CB(skb)->sacked)
3217 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3219 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3222 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3225 /* We passed data and got it acked, remove any soft error
3226 * log. Something worked...
3228 sk->sk_err_soft = 0;
3229 tp->rcv_tstamp = tcp_time_stamp;
3230 prior_packets = tp->packets_out;
3234 /* See if we can take anything off of the retransmit queue. */
3235 flag |= tcp_clean_rtx_queue(sk, prior_fackets);
3237 if (tp->frto_counter)
3238 frto_cwnd = tcp_process_frto(sk, flag);
3239 /* Guarantee sacktag reordering detection against wrap-arounds */
3240 if (before(tp->frto_highmark, tp->snd_una))
3241 tp->frto_highmark = 0;
3243 if (tcp_ack_is_dubious(sk, flag)) {
3244 /* Advance CWND, if state allows this. */
3245 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3246 tcp_may_raise_cwnd(sk, flag))
3247 tcp_cong_avoid(sk, ack, prior_in_flight);
3248 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3251 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3252 tcp_cong_avoid(sk, ack, prior_in_flight);
3255 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3256 dst_confirm(sk->sk_dst_cache);
3261 icsk->icsk_probes_out = 0;
3263 /* If this ack opens up a zero window, clear backoff. It was
3264 * being used to time the probes, and is probably far higher than
3265 * it needs to be for normal retransmission.
3267 if (tcp_send_head(sk))
3272 if (TCP_SKB_CB(skb)->sacked)
3273 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3276 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3280 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3281 * But, this can also be called on packets in the established flow when
3282 * the fast version below fails.
3284 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3288 struct tcphdr *th = tcp_hdr(skb);
3289 int length = (th->doff * 4) - sizeof(struct tcphdr);
3291 ptr = (unsigned char *)(th + 1);
3292 opt_rx->saw_tstamp = 0;
3294 while (length > 0) {
3295 int opcode = *ptr++;
3301 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3306 if (opsize < 2) /* "silly options" */
3308 if (opsize > length)
3309 return; /* don't parse partial options */
3312 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3313 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3315 if (opt_rx->user_mss &&
3316 opt_rx->user_mss < in_mss)
3317 in_mss = opt_rx->user_mss;
3318 opt_rx->mss_clamp = in_mss;
3323 if (opsize == TCPOLEN_WINDOW && th->syn &&
3324 !estab && sysctl_tcp_window_scaling) {
3325 __u8 snd_wscale = *(__u8 *)ptr;
3326 opt_rx->wscale_ok = 1;
3327 if (snd_wscale > 14) {
3328 if (net_ratelimit())
3329 printk(KERN_INFO "tcp_parse_options: Illegal window "
3330 "scaling value %d >14 received.\n",
3334 opt_rx->snd_wscale = snd_wscale;
3337 case TCPOPT_TIMESTAMP:
3338 if ((opsize == TCPOLEN_TIMESTAMP) &&
3339 ((estab && opt_rx->tstamp_ok) ||
3340 (!estab && sysctl_tcp_timestamps))) {
3341 opt_rx->saw_tstamp = 1;
3342 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3343 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3346 case TCPOPT_SACK_PERM:
3347 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3348 !estab && sysctl_tcp_sack) {
3349 opt_rx->sack_ok = 1;
3350 tcp_sack_reset(opt_rx);
3355 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3356 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3358 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3361 #ifdef CONFIG_TCP_MD5SIG
3364 * The MD5 Hash has already been
3365 * checked (see tcp_v{4,6}_do_rcv()).
3377 /* Fast parse options. This hopes to only see timestamps.
3378 * If it is wrong it falls back on tcp_parse_options().
3380 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3381 struct tcp_sock *tp)
3383 if (th->doff == sizeof(struct tcphdr) >> 2) {
3384 tp->rx_opt.saw_tstamp = 0;
3386 } else if (tp->rx_opt.tstamp_ok &&
3387 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3388 __be32 *ptr = (__be32 *)(th + 1);
3389 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3390 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3391 tp->rx_opt.saw_tstamp = 1;
3393 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3395 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3399 tcp_parse_options(skb, &tp->rx_opt, 1);
3403 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3405 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3406 tp->rx_opt.ts_recent_stamp = get_seconds();
3409 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3411 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3412 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3413 * extra check below makes sure this can only happen
3414 * for pure ACK frames. -DaveM
3416 * Not only, also it occurs for expired timestamps.
3419 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3420 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3421 tcp_store_ts_recent(tp);
3425 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3427 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3428 * it can pass through stack. So, the following predicate verifies that
3429 * this segment is not used for anything but congestion avoidance or
3430 * fast retransmit. Moreover, we even are able to eliminate most of such
3431 * second order effects, if we apply some small "replay" window (~RTO)
3432 * to timestamp space.
3434 * All these measures still do not guarantee that we reject wrapped ACKs
3435 * on networks with high bandwidth, when sequence space is recycled fastly,
3436 * but it guarantees that such events will be very rare and do not affect
3437 * connection seriously. This doesn't look nice, but alas, PAWS is really
3440 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3441 * states that events when retransmit arrives after original data are rare.
3442 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3443 * the biggest problem on large power networks even with minor reordering.
3444 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3445 * up to bandwidth of 18Gigabit/sec. 8) ]
3448 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3450 struct tcp_sock *tp = tcp_sk(sk);
3451 struct tcphdr *th = tcp_hdr(skb);
3452 u32 seq = TCP_SKB_CB(skb)->seq;
3453 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3455 return (/* 1. Pure ACK with correct sequence number. */
3456 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3458 /* 2. ... and duplicate ACK. */
3459 ack == tp->snd_una &&
3461 /* 3. ... and does not update window. */
3462 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3464 /* 4. ... and sits in replay window. */
3465 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3468 static inline int tcp_paws_discard(const struct sock *sk,
3469 const struct sk_buff *skb)
3471 const struct tcp_sock *tp = tcp_sk(sk);
3472 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3473 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3474 !tcp_disordered_ack(sk, skb));
3477 /* Check segment sequence number for validity.
3479 * Segment controls are considered valid, if the segment
3480 * fits to the window after truncation to the window. Acceptability
3481 * of data (and SYN, FIN, of course) is checked separately.
3482 * See tcp_data_queue(), for example.
3484 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3485 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3486 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3487 * (borrowed from freebsd)
3490 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3492 return !before(end_seq, tp->rcv_wup) &&
3493 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3496 /* When we get a reset we do this. */
3497 static void tcp_reset(struct sock *sk)
3499 /* We want the right error as BSD sees it (and indeed as we do). */
3500 switch (sk->sk_state) {
3502 sk->sk_err = ECONNREFUSED;
3504 case TCP_CLOSE_WAIT:
3510 sk->sk_err = ECONNRESET;
3513 if (!sock_flag(sk, SOCK_DEAD))
3514 sk->sk_error_report(sk);
3520 * Process the FIN bit. This now behaves as it is supposed to work
3521 * and the FIN takes effect when it is validly part of sequence
3522 * space. Not before when we get holes.
3524 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3525 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3528 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3529 * close and we go into CLOSING (and later onto TIME-WAIT)
3531 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3533 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3535 struct tcp_sock *tp = tcp_sk(sk);
3537 inet_csk_schedule_ack(sk);
3539 sk->sk_shutdown |= RCV_SHUTDOWN;
3540 sock_set_flag(sk, SOCK_DONE);
3542 switch (sk->sk_state) {
3544 case TCP_ESTABLISHED:
3545 /* Move to CLOSE_WAIT */
3546 tcp_set_state(sk, TCP_CLOSE_WAIT);
3547 inet_csk(sk)->icsk_ack.pingpong = 1;
3550 case TCP_CLOSE_WAIT:
3552 /* Received a retransmission of the FIN, do
3557 /* RFC793: Remain in the LAST-ACK state. */
3561 /* This case occurs when a simultaneous close
3562 * happens, we must ack the received FIN and
3563 * enter the CLOSING state.
3566 tcp_set_state(sk, TCP_CLOSING);
3569 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3571 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3574 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3575 * cases we should never reach this piece of code.
3577 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3578 __FUNCTION__, sk->sk_state);
3582 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3583 * Probably, we should reset in this case. For now drop them.
3585 __skb_queue_purge(&tp->out_of_order_queue);
3586 if (tcp_is_sack(tp))
3587 tcp_sack_reset(&tp->rx_opt);
3590 if (!sock_flag(sk, SOCK_DEAD)) {
3591 sk->sk_state_change(sk);
3593 /* Do not send POLL_HUP for half duplex close. */
3594 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3595 sk->sk_state == TCP_CLOSE)
3596 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3598 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3602 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3605 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3606 if (before(seq, sp->start_seq))
3607 sp->start_seq = seq;
3608 if (after(end_seq, sp->end_seq))
3609 sp->end_seq = end_seq;
3615 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3617 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3618 if (before(seq, tp->rcv_nxt))
3619 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3621 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3623 tp->rx_opt.dsack = 1;
3624 tp->duplicate_sack[0].start_seq = seq;
3625 tp->duplicate_sack[0].end_seq = end_seq;
3626 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1,
3627 4 - tp->rx_opt.tstamp_ok);
3631 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3633 if (!tp->rx_opt.dsack)
3634 tcp_dsack_set(tp, seq, end_seq);
3636 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3639 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3641 struct tcp_sock *tp = tcp_sk(sk);
3643 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3644 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3645 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3646 tcp_enter_quickack_mode(sk);
3648 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3649 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3651 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3652 end_seq = tp->rcv_nxt;
3653 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3660 /* These routines update the SACK block as out-of-order packets arrive or
3661 * in-order packets close up the sequence space.
3663 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3666 struct tcp_sack_block *sp = &tp->selective_acks[0];
3667 struct tcp_sack_block *swalk = sp + 1;
3669 /* See if the recent change to the first SACK eats into
3670 * or hits the sequence space of other SACK blocks, if so coalesce.
3672 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3673 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3676 /* Zap SWALK, by moving every further SACK up by one slot.
3677 * Decrease num_sacks.
3679 tp->rx_opt.num_sacks--;
3680 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks +
3682 4 - tp->rx_opt.tstamp_ok);
3683 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3687 this_sack++, swalk++;
3691 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
3692 struct tcp_sack_block *sack2)
3696 tmp = sack1->start_seq;
3697 sack1->start_seq = sack2->start_seq;
3698 sack2->start_seq = tmp;
3700 tmp = sack1->end_seq;
3701 sack1->end_seq = sack2->end_seq;
3702 sack2->end_seq = tmp;
3705 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3707 struct tcp_sock *tp = tcp_sk(sk);
3708 struct tcp_sack_block *sp = &tp->selective_acks[0];
3709 int cur_sacks = tp->rx_opt.num_sacks;
3715 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3716 if (tcp_sack_extend(sp, seq, end_seq)) {
3717 /* Rotate this_sack to the first one. */
3718 for (; this_sack > 0; this_sack--, sp--)
3719 tcp_sack_swap(sp, sp - 1);
3721 tcp_sack_maybe_coalesce(tp);
3726 /* Could not find an adjacent existing SACK, build a new one,
3727 * put it at the front, and shift everyone else down. We
3728 * always know there is at least one SACK present already here.
3730 * If the sack array is full, forget about the last one.
3732 if (this_sack >= 4) {
3734 tp->rx_opt.num_sacks--;
3737 for (; this_sack > 0; this_sack--, sp--)
3741 /* Build the new head SACK, and we're done. */
3742 sp->start_seq = seq;
3743 sp->end_seq = end_seq;
3744 tp->rx_opt.num_sacks++;
3745 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack,
3746 4 - tp->rx_opt.tstamp_ok);
3749 /* RCV.NXT advances, some SACKs should be eaten. */
3751 static void tcp_sack_remove(struct tcp_sock *tp)
3753 struct tcp_sack_block *sp = &tp->selective_acks[0];
3754 int num_sacks = tp->rx_opt.num_sacks;
3757 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3758 if (skb_queue_empty(&tp->out_of_order_queue)) {
3759 tp->rx_opt.num_sacks = 0;
3760 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3764 for (this_sack = 0; this_sack < num_sacks;) {
3765 /* Check if the start of the sack is covered by RCV.NXT. */
3766 if (!before(tp->rcv_nxt, sp->start_seq)) {
3769 /* RCV.NXT must cover all the block! */
3770 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3772 /* Zap this SACK, by moving forward any other SACKS. */
3773 for (i=this_sack+1; i < num_sacks; i++)
3774 tp->selective_acks[i-1] = tp->selective_acks[i];
3781 if (num_sacks != tp->rx_opt.num_sacks) {
3782 tp->rx_opt.num_sacks = num_sacks;
3783 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks +
3785 4 - tp->rx_opt.tstamp_ok);
3789 /* This one checks to see if we can put data from the
3790 * out_of_order queue into the receive_queue.
3792 static void tcp_ofo_queue(struct sock *sk)
3794 struct tcp_sock *tp = tcp_sk(sk);
3795 __u32 dsack_high = tp->rcv_nxt;
3796 struct sk_buff *skb;
3798 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3799 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3802 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3803 __u32 dsack = dsack_high;
3804 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3805 dsack_high = TCP_SKB_CB(skb)->end_seq;
3806 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3809 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3810 SOCK_DEBUG(sk, "ofo packet was already received \n");
3811 __skb_unlink(skb, &tp->out_of_order_queue);
3815 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3816 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3817 TCP_SKB_CB(skb)->end_seq);
3819 __skb_unlink(skb, &tp->out_of_order_queue);
3820 __skb_queue_tail(&sk->sk_receive_queue, skb);
3821 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3822 if (tcp_hdr(skb)->fin)
3823 tcp_fin(skb, sk, tcp_hdr(skb));
3827 static int tcp_prune_queue(struct sock *sk);
3829 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3831 struct tcphdr *th = tcp_hdr(skb);
3832 struct tcp_sock *tp = tcp_sk(sk);
3835 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3838 __skb_pull(skb, th->doff * 4);
3840 TCP_ECN_accept_cwr(tp, skb);
3842 if (tp->rx_opt.dsack) {
3843 tp->rx_opt.dsack = 0;
3844 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3845 4 - tp->rx_opt.tstamp_ok);
3848 /* Queue data for delivery to the user.
3849 * Packets in sequence go to the receive queue.
3850 * Out of sequence packets to the out_of_order_queue.
3852 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3853 if (tcp_receive_window(tp) == 0)
3856 /* Ok. In sequence. In window. */
3857 if (tp->ucopy.task == current &&
3858 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3859 sock_owned_by_user(sk) && !tp->urg_data) {
3860 int chunk = min_t(unsigned int, skb->len,
3863 __set_current_state(TASK_RUNNING);
3866 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3867 tp->ucopy.len -= chunk;
3868 tp->copied_seq += chunk;
3869 eaten = (chunk == skb->len && !th->fin);
3870 tcp_rcv_space_adjust(sk);
3878 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3879 !sk_rmem_schedule(sk, skb->truesize))) {
3880 if (tcp_prune_queue(sk) < 0 ||
3881 !sk_rmem_schedule(sk, skb->truesize))
3884 skb_set_owner_r(skb, sk);
3885 __skb_queue_tail(&sk->sk_receive_queue, skb);
3887 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3889 tcp_event_data_recv(sk, skb);
3891 tcp_fin(skb, sk, th);
3893 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3896 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3897 * gap in queue is filled.
3899 if (skb_queue_empty(&tp->out_of_order_queue))
3900 inet_csk(sk)->icsk_ack.pingpong = 0;
3903 if (tp->rx_opt.num_sacks)
3904 tcp_sack_remove(tp);
3906 tcp_fast_path_check(sk);
3910 else if (!sock_flag(sk, SOCK_DEAD))
3911 sk->sk_data_ready(sk, 0);
3915 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3916 /* A retransmit, 2nd most common case. Force an immediate ack. */
3917 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3918 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3921 tcp_enter_quickack_mode(sk);
3922 inet_csk_schedule_ack(sk);
3928 /* Out of window. F.e. zero window probe. */
3929 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3932 tcp_enter_quickack_mode(sk);
3934 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3935 /* Partial packet, seq < rcv_next < end_seq */
3936 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3937 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3938 TCP_SKB_CB(skb)->end_seq);
3940 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3942 /* If window is closed, drop tail of packet. But after
3943 * remembering D-SACK for its head made in previous line.
3945 if (!tcp_receive_window(tp))
3950 TCP_ECN_check_ce(tp, skb);
3952 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3953 !sk_rmem_schedule(sk, skb->truesize)) {
3954 if (tcp_prune_queue(sk) < 0 ||
3955 !sk_rmem_schedule(sk, skb->truesize))
3959 /* Disable header prediction. */
3961 inet_csk_schedule_ack(sk);
3963 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3964 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3966 skb_set_owner_r(skb, sk);
3968 if (!skb_peek(&tp->out_of_order_queue)) {
3969 /* Initial out of order segment, build 1 SACK. */
3970 if (tcp_is_sack(tp)) {
3971 tp->rx_opt.num_sacks = 1;
3972 tp->rx_opt.dsack = 0;
3973 tp->rx_opt.eff_sacks = 1;
3974 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3975 tp->selective_acks[0].end_seq =
3976 TCP_SKB_CB(skb)->end_seq;
3978 __skb_queue_head(&tp->out_of_order_queue, skb);
3980 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3981 u32 seq = TCP_SKB_CB(skb)->seq;
3982 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3984 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3985 __skb_append(skb1, skb, &tp->out_of_order_queue);
3987 if (!tp->rx_opt.num_sacks ||
3988 tp->selective_acks[0].end_seq != seq)
3991 /* Common case: data arrive in order after hole. */
3992 tp->selective_acks[0].end_seq = end_seq;
3996 /* Find place to insert this segment. */
3998 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4000 } while ((skb1 = skb1->prev) !=
4001 (struct sk_buff *)&tp->out_of_order_queue);
4003 /* Do skb overlap to previous one? */
4004 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4005 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4006 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4007 /* All the bits are present. Drop. */
4009 tcp_dsack_set(tp, seq, end_seq);
4012 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4013 /* Partial overlap. */
4014 tcp_dsack_set(tp, seq,
4015 TCP_SKB_CB(skb1)->end_seq);
4020 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
4022 /* And clean segments covered by new one as whole. */
4023 while ((skb1 = skb->next) !=
4024 (struct sk_buff *)&tp->out_of_order_queue &&
4025 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4026 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4027 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq,
4031 __skb_unlink(skb1, &tp->out_of_order_queue);
4032 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq,
4033 TCP_SKB_CB(skb1)->end_seq);
4038 if (tcp_is_sack(tp))
4039 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4043 /* Collapse contiguous sequence of skbs head..tail with
4044 * sequence numbers start..end.
4045 * Segments with FIN/SYN are not collapsed (only because this
4049 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4050 struct sk_buff *head, struct sk_buff *tail,
4053 struct sk_buff *skb;
4055 /* First, check that queue is collapsible and find
4056 * the point where collapsing can be useful. */
4057 for (skb = head; skb != tail;) {
4058 /* No new bits? It is possible on ofo queue. */
4059 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4060 struct sk_buff *next = skb->next;
4061 __skb_unlink(skb, list);
4063 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4068 /* The first skb to collapse is:
4070 * - bloated or contains data before "start" or
4071 * overlaps to the next one.
4073 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4074 (tcp_win_from_space(skb->truesize) > skb->len ||
4075 before(TCP_SKB_CB(skb)->seq, start) ||
4076 (skb->next != tail &&
4077 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4080 /* Decided to skip this, advance start seq. */
4081 start = TCP_SKB_CB(skb)->end_seq;
4084 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4087 while (before(start, end)) {
4088 struct sk_buff *nskb;
4089 unsigned int header = skb_headroom(skb);
4090 int copy = SKB_MAX_ORDER(header, 0);
4092 /* Too big header? This can happen with IPv6. */
4095 if (end - start < copy)
4097 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4101 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4102 skb_set_network_header(nskb, (skb_network_header(skb) -
4104 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4106 skb_reserve(nskb, header);
4107 memcpy(nskb->head, skb->head, header);
4108 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4109 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4110 __skb_insert(nskb, skb->prev, skb, list);
4111 skb_set_owner_r(nskb, sk);
4113 /* Copy data, releasing collapsed skbs. */
4115 int offset = start - TCP_SKB_CB(skb)->seq;
4116 int size = TCP_SKB_CB(skb)->end_seq - start;
4120 size = min(copy, size);
4121 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4123 TCP_SKB_CB(nskb)->end_seq += size;
4127 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4128 struct sk_buff *next = skb->next;
4129 __skb_unlink(skb, list);
4131 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4134 tcp_hdr(skb)->syn ||
4142 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4143 * and tcp_collapse() them until all the queue is collapsed.
4145 static void tcp_collapse_ofo_queue(struct sock *sk)
4147 struct tcp_sock *tp = tcp_sk(sk);
4148 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4149 struct sk_buff *head;
4155 start = TCP_SKB_CB(skb)->seq;
4156 end = TCP_SKB_CB(skb)->end_seq;
4162 /* Segment is terminated when we see gap or when
4163 * we are at the end of all the queue. */
4164 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4165 after(TCP_SKB_CB(skb)->seq, end) ||
4166 before(TCP_SKB_CB(skb)->end_seq, start)) {
4167 tcp_collapse(sk, &tp->out_of_order_queue,
4168 head, skb, start, end);
4170 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4172 /* Start new segment */
4173 start = TCP_SKB_CB(skb)->seq;
4174 end = TCP_SKB_CB(skb)->end_seq;
4176 if (before(TCP_SKB_CB(skb)->seq, start))
4177 start = TCP_SKB_CB(skb)->seq;
4178 if (after(TCP_SKB_CB(skb)->end_seq, end))
4179 end = TCP_SKB_CB(skb)->end_seq;
4184 /* Reduce allocated memory if we can, trying to get
4185 * the socket within its memory limits again.
4187 * Return less than zero if we should start dropping frames
4188 * until the socket owning process reads some of the data
4189 * to stabilize the situation.
4191 static int tcp_prune_queue(struct sock *sk)
4193 struct tcp_sock *tp = tcp_sk(sk);
4195 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4197 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
4199 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4200 tcp_clamp_window(sk);
4201 else if (tcp_memory_pressure)
4202 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4204 tcp_collapse_ofo_queue(sk);
4205 tcp_collapse(sk, &sk->sk_receive_queue,
4206 sk->sk_receive_queue.next,
4207 (struct sk_buff *)&sk->sk_receive_queue,
4208 tp->copied_seq, tp->rcv_nxt);
4211 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4214 /* Collapsing did not help, destructive actions follow.
4215 * This must not ever occur. */
4217 /* First, purge the out_of_order queue. */
4218 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4219 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
4220 __skb_queue_purge(&tp->out_of_order_queue);
4222 /* Reset SACK state. A conforming SACK implementation will
4223 * do the same at a timeout based retransmit. When a connection
4224 * is in a sad state like this, we care only about integrity
4225 * of the connection not performance.
4227 if (tcp_is_sack(tp))
4228 tcp_sack_reset(&tp->rx_opt);
4232 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4235 /* If we are really being abused, tell the caller to silently
4236 * drop receive data on the floor. It will get retransmitted
4237 * and hopefully then we'll have sufficient space.
4239 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4241 /* Massive buffer overcommit. */
4246 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4247 * As additional protections, we do not touch cwnd in retransmission phases,
4248 * and if application hit its sndbuf limit recently.
4250 void tcp_cwnd_application_limited(struct sock *sk)
4252 struct tcp_sock *tp = tcp_sk(sk);
4254 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4255 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4256 /* Limited by application or receiver window. */
4257 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4258 u32 win_used = max(tp->snd_cwnd_used, init_win);
4259 if (win_used < tp->snd_cwnd) {
4260 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4261 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4263 tp->snd_cwnd_used = 0;
4265 tp->snd_cwnd_stamp = tcp_time_stamp;
4268 static int tcp_should_expand_sndbuf(struct sock *sk)
4270 struct tcp_sock *tp = tcp_sk(sk);
4272 /* If the user specified a specific send buffer setting, do
4275 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4278 /* If we are under global TCP memory pressure, do not expand. */
4279 if (tcp_memory_pressure)
4282 /* If we are under soft global TCP memory pressure, do not expand. */
4283 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4286 /* If we filled the congestion window, do not expand. */
4287 if (tp->packets_out >= tp->snd_cwnd)
4293 /* When incoming ACK allowed to free some skb from write_queue,
4294 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4295 * on the exit from tcp input handler.
4297 * PROBLEM: sndbuf expansion does not work well with largesend.
4299 static void tcp_new_space(struct sock *sk)
4301 struct tcp_sock *tp = tcp_sk(sk);
4303 if (tcp_should_expand_sndbuf(sk)) {
4304 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4305 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4306 demanded = max_t(unsigned int, tp->snd_cwnd,
4307 tp->reordering + 1);
4308 sndmem *= 2 * demanded;
4309 if (sndmem > sk->sk_sndbuf)
4310 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4311 tp->snd_cwnd_stamp = tcp_time_stamp;
4314 sk->sk_write_space(sk);
4317 static void tcp_check_space(struct sock *sk)
4319 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4320 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4321 if (sk->sk_socket &&
4322 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4327 static inline void tcp_data_snd_check(struct sock *sk)
4329 tcp_push_pending_frames(sk);
4330 tcp_check_space(sk);
4334 * Check if sending an ack is needed.
4336 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4338 struct tcp_sock *tp = tcp_sk(sk);
4340 /* More than one full frame received... */
4341 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4342 /* ... and right edge of window advances far enough.
4343 * (tcp_recvmsg() will send ACK otherwise). Or...
4345 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4346 /* We ACK each frame or... */
4347 tcp_in_quickack_mode(sk) ||
4348 /* We have out of order data. */
4349 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4350 /* Then ack it now */
4353 /* Else, send delayed ack. */
4354 tcp_send_delayed_ack(sk);
4358 static inline void tcp_ack_snd_check(struct sock *sk)
4360 if (!inet_csk_ack_scheduled(sk)) {
4361 /* We sent a data segment already. */
4364 __tcp_ack_snd_check(sk, 1);
4368 * This routine is only called when we have urgent data
4369 * signaled. Its the 'slow' part of tcp_urg. It could be
4370 * moved inline now as tcp_urg is only called from one
4371 * place. We handle URGent data wrong. We have to - as
4372 * BSD still doesn't use the correction from RFC961.
4373 * For 1003.1g we should support a new option TCP_STDURG to permit
4374 * either form (or just set the sysctl tcp_stdurg).
4377 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4379 struct tcp_sock *tp = tcp_sk(sk);
4380 u32 ptr = ntohs(th->urg_ptr);
4382 if (ptr && !sysctl_tcp_stdurg)
4384 ptr += ntohl(th->seq);
4386 /* Ignore urgent data that we've already seen and read. */
4387 if (after(tp->copied_seq, ptr))
4390 /* Do not replay urg ptr.
4392 * NOTE: interesting situation not covered by specs.
4393 * Misbehaving sender may send urg ptr, pointing to segment,
4394 * which we already have in ofo queue. We are not able to fetch
4395 * such data and will stay in TCP_URG_NOTYET until will be eaten
4396 * by recvmsg(). Seems, we are not obliged to handle such wicked
4397 * situations. But it is worth to think about possibility of some
4398 * DoSes using some hypothetical application level deadlock.
4400 if (before(ptr, tp->rcv_nxt))
4403 /* Do we already have a newer (or duplicate) urgent pointer? */
4404 if (tp->urg_data && !after(ptr, tp->urg_seq))
4407 /* Tell the world about our new urgent pointer. */
4410 /* We may be adding urgent data when the last byte read was
4411 * urgent. To do this requires some care. We cannot just ignore
4412 * tp->copied_seq since we would read the last urgent byte again
4413 * as data, nor can we alter copied_seq until this data arrives
4414 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4416 * NOTE. Double Dutch. Rendering to plain English: author of comment
4417 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4418 * and expect that both A and B disappear from stream. This is _wrong_.
4419 * Though this happens in BSD with high probability, this is occasional.
4420 * Any application relying on this is buggy. Note also, that fix "works"
4421 * only in this artificial test. Insert some normal data between A and B and we will
4422 * decline of BSD again. Verdict: it is better to remove to trap
4425 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4426 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4427 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4429 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4430 __skb_unlink(skb, &sk->sk_receive_queue);
4435 tp->urg_data = TCP_URG_NOTYET;
4438 /* Disable header prediction. */
4442 /* This is the 'fast' part of urgent handling. */
4443 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4445 struct tcp_sock *tp = tcp_sk(sk);
4447 /* Check if we get a new urgent pointer - normally not. */
4449 tcp_check_urg(sk, th);
4451 /* Do we wait for any urgent data? - normally not... */
4452 if (tp->urg_data == TCP_URG_NOTYET) {
4453 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4456 /* Is the urgent pointer pointing into this packet? */
4457 if (ptr < skb->len) {
4459 if (skb_copy_bits(skb, ptr, &tmp, 1))
4461 tp->urg_data = TCP_URG_VALID | tmp;
4462 if (!sock_flag(sk, SOCK_DEAD))
4463 sk->sk_data_ready(sk, 0);
4468 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4470 struct tcp_sock *tp = tcp_sk(sk);
4471 int chunk = skb->len - hlen;
4475 if (skb_csum_unnecessary(skb))
4476 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4478 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4482 tp->ucopy.len -= chunk;
4483 tp->copied_seq += chunk;
4484 tcp_rcv_space_adjust(sk);
4491 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4492 struct sk_buff *skb)
4496 if (sock_owned_by_user(sk)) {
4498 result = __tcp_checksum_complete(skb);
4501 result = __tcp_checksum_complete(skb);
4506 static inline int tcp_checksum_complete_user(struct sock *sk,
4507 struct sk_buff *skb)
4509 return !skb_csum_unnecessary(skb) &&
4510 __tcp_checksum_complete_user(sk, skb);
4513 #ifdef CONFIG_NET_DMA
4514 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4517 struct tcp_sock *tp = tcp_sk(sk);
4518 int chunk = skb->len - hlen;
4520 int copied_early = 0;
4522 if (tp->ucopy.wakeup)
4525 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4526 tp->ucopy.dma_chan = get_softnet_dma();
4528 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4530 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4532 tp->ucopy.iov, chunk,
4533 tp->ucopy.pinned_list);
4538 tp->ucopy.dma_cookie = dma_cookie;
4541 tp->ucopy.len -= chunk;
4542 tp->copied_seq += chunk;
4543 tcp_rcv_space_adjust(sk);
4545 if ((tp->ucopy.len == 0) ||
4546 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4547 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4548 tp->ucopy.wakeup = 1;
4549 sk->sk_data_ready(sk, 0);
4551 } else if (chunk > 0) {
4552 tp->ucopy.wakeup = 1;
4553 sk->sk_data_ready(sk, 0);
4556 return copied_early;
4558 #endif /* CONFIG_NET_DMA */
4561 * TCP receive function for the ESTABLISHED state.
4563 * It is split into a fast path and a slow path. The fast path is
4565 * - A zero window was announced from us - zero window probing
4566 * is only handled properly in the slow path.
4567 * - Out of order segments arrived.
4568 * - Urgent data is expected.
4569 * - There is no buffer space left
4570 * - Unexpected TCP flags/window values/header lengths are received
4571 * (detected by checking the TCP header against pred_flags)
4572 * - Data is sent in both directions. Fast path only supports pure senders
4573 * or pure receivers (this means either the sequence number or the ack
4574 * value must stay constant)
4575 * - Unexpected TCP option.
4577 * When these conditions are not satisfied it drops into a standard
4578 * receive procedure patterned after RFC793 to handle all cases.
4579 * The first three cases are guaranteed by proper pred_flags setting,
4580 * the rest is checked inline. Fast processing is turned on in
4581 * tcp_data_queue when everything is OK.
4583 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4584 struct tcphdr *th, unsigned len)
4586 struct tcp_sock *tp = tcp_sk(sk);
4589 * Header prediction.
4590 * The code loosely follows the one in the famous
4591 * "30 instruction TCP receive" Van Jacobson mail.
4593 * Van's trick is to deposit buffers into socket queue
4594 * on a device interrupt, to call tcp_recv function
4595 * on the receive process context and checksum and copy
4596 * the buffer to user space. smart...
4598 * Our current scheme is not silly either but we take the
4599 * extra cost of the net_bh soft interrupt processing...
4600 * We do checksum and copy also but from device to kernel.
4603 tp->rx_opt.saw_tstamp = 0;
4605 /* pred_flags is 0xS?10 << 16 + snd_wnd
4606 * if header_prediction is to be made
4607 * 'S' will always be tp->tcp_header_len >> 2
4608 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4609 * turn it off (when there are holes in the receive
4610 * space for instance)
4611 * PSH flag is ignored.
4614 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4615 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4616 int tcp_header_len = tp->tcp_header_len;
4618 /* Timestamp header prediction: tcp_header_len
4619 * is automatically equal to th->doff*4 due to pred_flags
4623 /* Check timestamp */
4624 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4625 __be32 *ptr = (__be32 *)(th + 1);
4627 /* No? Slow path! */
4628 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4629 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4632 tp->rx_opt.saw_tstamp = 1;
4634 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4636 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4638 /* If PAWS failed, check it more carefully in slow path */
4639 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4642 /* DO NOT update ts_recent here, if checksum fails
4643 * and timestamp was corrupted part, it will result
4644 * in a hung connection since we will drop all
4645 * future packets due to the PAWS test.
4649 if (len <= tcp_header_len) {
4650 /* Bulk data transfer: sender */
4651 if (len == tcp_header_len) {
4652 /* Predicted packet is in window by definition.
4653 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4654 * Hence, check seq<=rcv_wup reduces to:
4656 if (tcp_header_len ==
4657 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4658 tp->rcv_nxt == tp->rcv_wup)
4659 tcp_store_ts_recent(tp);
4661 /* We know that such packets are checksummed
4664 tcp_ack(sk, skb, 0);
4666 tcp_data_snd_check(sk);
4668 } else { /* Header too small */
4669 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4674 int copied_early = 0;
4676 if (tp->copied_seq == tp->rcv_nxt &&
4677 len - tcp_header_len <= tp->ucopy.len) {
4678 #ifdef CONFIG_NET_DMA
4679 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4684 if (tp->ucopy.task == current &&
4685 sock_owned_by_user(sk) && !copied_early) {
4686 __set_current_state(TASK_RUNNING);
4688 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4692 /* Predicted packet is in window by definition.
4693 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4694 * Hence, check seq<=rcv_wup reduces to:
4696 if (tcp_header_len ==
4697 (sizeof(struct tcphdr) +
4698 TCPOLEN_TSTAMP_ALIGNED) &&
4699 tp->rcv_nxt == tp->rcv_wup)
4700 tcp_store_ts_recent(tp);
4702 tcp_rcv_rtt_measure_ts(sk, skb);
4704 __skb_pull(skb, tcp_header_len);
4705 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4706 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4709 tcp_cleanup_rbuf(sk, skb->len);
4712 if (tcp_checksum_complete_user(sk, skb))
4715 /* Predicted packet is in window by definition.
4716 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4717 * Hence, check seq<=rcv_wup reduces to:
4719 if (tcp_header_len ==
4720 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4721 tp->rcv_nxt == tp->rcv_wup)
4722 tcp_store_ts_recent(tp);
4724 tcp_rcv_rtt_measure_ts(sk, skb);
4726 if ((int)skb->truesize > sk->sk_forward_alloc)
4729 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4731 /* Bulk data transfer: receiver */
4732 __skb_pull(skb, tcp_header_len);
4733 __skb_queue_tail(&sk->sk_receive_queue, skb);
4734 skb_set_owner_r(skb, sk);
4735 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4738 tcp_event_data_recv(sk, skb);
4740 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4741 /* Well, only one small jumplet in fast path... */
4742 tcp_ack(sk, skb, FLAG_DATA);
4743 tcp_data_snd_check(sk);
4744 if (!inet_csk_ack_scheduled(sk))
4748 __tcp_ack_snd_check(sk, 0);
4750 #ifdef CONFIG_NET_DMA
4752 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4758 sk->sk_data_ready(sk, 0);
4764 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
4768 * RFC1323: H1. Apply PAWS check first.
4770 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4771 tcp_paws_discard(sk, skb)) {
4773 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4774 tcp_send_dupack(sk, skb);
4777 /* Resets are accepted even if PAWS failed.
4779 ts_recent update must be made after we are sure
4780 that the packet is in window.
4785 * Standard slow path.
4788 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4789 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4790 * (RST) segments are validated by checking their SEQ-fields."
4791 * And page 69: "If an incoming segment is not acceptable,
4792 * an acknowledgment should be sent in reply (unless the RST bit
4793 * is set, if so drop the segment and return)".
4796 tcp_send_dupack(sk, skb);
4805 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4807 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4808 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4809 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4816 tcp_ack(sk, skb, FLAG_SLOWPATH);
4818 tcp_rcv_rtt_measure_ts(sk, skb);
4820 /* Process urgent data. */
4821 tcp_urg(sk, skb, th);
4823 /* step 7: process the segment text */
4824 tcp_data_queue(sk, skb);
4826 tcp_data_snd_check(sk);
4827 tcp_ack_snd_check(sk);
4831 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4838 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4839 struct tcphdr *th, unsigned len)
4841 struct tcp_sock *tp = tcp_sk(sk);
4842 struct inet_connection_sock *icsk = inet_csk(sk);
4843 int saved_clamp = tp->rx_opt.mss_clamp;
4845 tcp_parse_options(skb, &tp->rx_opt, 0);
4849 * "If the state is SYN-SENT then
4850 * first check the ACK bit
4851 * If the ACK bit is set
4852 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4853 * a reset (unless the RST bit is set, if so drop
4854 * the segment and return)"
4856 * We do not send data with SYN, so that RFC-correct
4859 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4860 goto reset_and_undo;
4862 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4863 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4865 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4866 goto reset_and_undo;
4869 /* Now ACK is acceptable.
4871 * "If the RST bit is set
4872 * If the ACK was acceptable then signal the user "error:
4873 * connection reset", drop the segment, enter CLOSED state,
4874 * delete TCB, and return."
4883 * "fifth, if neither of the SYN or RST bits is set then
4884 * drop the segment and return."
4890 goto discard_and_undo;
4893 * "If the SYN bit is on ...
4894 * are acceptable then ...
4895 * (our SYN has been ACKed), change the connection
4896 * state to ESTABLISHED..."
4899 TCP_ECN_rcv_synack(tp, th);
4901 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4902 tcp_ack(sk, skb, FLAG_SLOWPATH);
4904 /* Ok.. it's good. Set up sequence numbers and
4905 * move to established.
4907 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4908 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4910 /* RFC1323: The window in SYN & SYN/ACK segments is
4913 tp->snd_wnd = ntohs(th->window);
4914 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4916 if (!tp->rx_opt.wscale_ok) {
4917 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4918 tp->window_clamp = min(tp->window_clamp, 65535U);
4921 if (tp->rx_opt.saw_tstamp) {
4922 tp->rx_opt.tstamp_ok = 1;
4923 tp->tcp_header_len =
4924 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4925 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4926 tcp_store_ts_recent(tp);
4928 tp->tcp_header_len = sizeof(struct tcphdr);
4931 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4932 tcp_enable_fack(tp);
4935 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4936 tcp_initialize_rcv_mss(sk);
4938 /* Remember, tcp_poll() does not lock socket!
4939 * Change state from SYN-SENT only after copied_seq
4940 * is initialized. */
4941 tp->copied_seq = tp->rcv_nxt;
4943 tcp_set_state(sk, TCP_ESTABLISHED);
4945 security_inet_conn_established(sk, skb);
4947 /* Make sure socket is routed, for correct metrics. */
4948 icsk->icsk_af_ops->rebuild_header(sk);
4950 tcp_init_metrics(sk);
4952 tcp_init_congestion_control(sk);
4954 /* Prevent spurious tcp_cwnd_restart() on first data
4957 tp->lsndtime = tcp_time_stamp;
4959 tcp_init_buffer_space(sk);
4961 if (sock_flag(sk, SOCK_KEEPOPEN))
4962 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4964 if (!tp->rx_opt.snd_wscale)
4965 __tcp_fast_path_on(tp, tp->snd_wnd);
4969 if (!sock_flag(sk, SOCK_DEAD)) {
4970 sk->sk_state_change(sk);
4971 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
4974 if (sk->sk_write_pending ||
4975 icsk->icsk_accept_queue.rskq_defer_accept ||
4976 icsk->icsk_ack.pingpong) {
4977 /* Save one ACK. Data will be ready after
4978 * several ticks, if write_pending is set.
4980 * It may be deleted, but with this feature tcpdumps
4981 * look so _wonderfully_ clever, that I was not able
4982 * to stand against the temptation 8) --ANK
4984 inet_csk_schedule_ack(sk);
4985 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4986 icsk->icsk_ack.ato = TCP_ATO_MIN;
4987 tcp_incr_quickack(sk);
4988 tcp_enter_quickack_mode(sk);
4989 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4990 TCP_DELACK_MAX, TCP_RTO_MAX);
5001 /* No ACK in the segment */
5005 * "If the RST bit is set
5007 * Otherwise (no ACK) drop the segment and return."
5010 goto discard_and_undo;
5014 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5015 tcp_paws_check(&tp->rx_opt, 0))
5016 goto discard_and_undo;
5019 /* We see SYN without ACK. It is attempt of
5020 * simultaneous connect with crossed SYNs.
5021 * Particularly, it can be connect to self.
5023 tcp_set_state(sk, TCP_SYN_RECV);
5025 if (tp->rx_opt.saw_tstamp) {
5026 tp->rx_opt.tstamp_ok = 1;
5027 tcp_store_ts_recent(tp);
5028 tp->tcp_header_len =
5029 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5031 tp->tcp_header_len = sizeof(struct tcphdr);
5034 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5035 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5037 /* RFC1323: The window in SYN & SYN/ACK segments is
5040 tp->snd_wnd = ntohs(th->window);
5041 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5042 tp->max_window = tp->snd_wnd;
5044 TCP_ECN_rcv_syn(tp, th);
5047 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5048 tcp_initialize_rcv_mss(sk);
5050 tcp_send_synack(sk);
5052 /* Note, we could accept data and URG from this segment.
5053 * There are no obstacles to make this.
5055 * However, if we ignore data in ACKless segments sometimes,
5056 * we have no reasons to accept it sometimes.
5057 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5058 * is not flawless. So, discard packet for sanity.
5059 * Uncomment this return to process the data.
5066 /* "fifth, if neither of the SYN or RST bits is set then
5067 * drop the segment and return."
5071 tcp_clear_options(&tp->rx_opt);
5072 tp->rx_opt.mss_clamp = saved_clamp;
5076 tcp_clear_options(&tp->rx_opt);
5077 tp->rx_opt.mss_clamp = saved_clamp;
5082 * This function implements the receiving procedure of RFC 793 for
5083 * all states except ESTABLISHED and TIME_WAIT.
5084 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5085 * address independent.
5088 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5089 struct tcphdr *th, unsigned len)
5091 struct tcp_sock *tp = tcp_sk(sk);
5092 struct inet_connection_sock *icsk = inet_csk(sk);
5095 tp->rx_opt.saw_tstamp = 0;
5097 switch (sk->sk_state) {
5109 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5112 /* Now we have several options: In theory there is
5113 * nothing else in the frame. KA9Q has an option to
5114 * send data with the syn, BSD accepts data with the
5115 * syn up to the [to be] advertised window and
5116 * Solaris 2.1 gives you a protocol error. For now
5117 * we just ignore it, that fits the spec precisely
5118 * and avoids incompatibilities. It would be nice in
5119 * future to drop through and process the data.
5121 * Now that TTCP is starting to be used we ought to
5123 * But, this leaves one open to an easy denial of
5124 * service attack, and SYN cookies can't defend
5125 * against this problem. So, we drop the data
5126 * in the interest of security over speed unless
5127 * it's still in use.
5135 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5139 /* Do step6 onward by hand. */
5140 tcp_urg(sk, skb, th);
5142 tcp_data_snd_check(sk);
5146 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5147 tcp_paws_discard(sk, skb)) {
5149 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
5150 tcp_send_dupack(sk, skb);
5153 /* Reset is accepted even if it did not pass PAWS. */
5156 /* step 1: check sequence number */
5157 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5159 tcp_send_dupack(sk, skb);
5163 /* step 2: check RST bit */
5169 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5171 /* step 3: check security and precedence [ignored] */
5175 * Check for a SYN in window.
5177 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5178 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
5183 /* step 5: check the ACK field */
5185 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5187 switch (sk->sk_state) {
5190 tp->copied_seq = tp->rcv_nxt;
5192 tcp_set_state(sk, TCP_ESTABLISHED);
5193 sk->sk_state_change(sk);
5195 /* Note, that this wakeup is only for marginal
5196 * crossed SYN case. Passively open sockets
5197 * are not waked up, because sk->sk_sleep ==
5198 * NULL and sk->sk_socket == NULL.
5202 SOCK_WAKE_IO, POLL_OUT);
5204 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5205 tp->snd_wnd = ntohs(th->window) <<
5206 tp->rx_opt.snd_wscale;
5207 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5208 TCP_SKB_CB(skb)->seq);
5210 /* tcp_ack considers this ACK as duplicate
5211 * and does not calculate rtt.
5212 * Fix it at least with timestamps.
5214 if (tp->rx_opt.saw_tstamp &&
5215 tp->rx_opt.rcv_tsecr && !tp->srtt)
5216 tcp_ack_saw_tstamp(sk, 0);
5218 if (tp->rx_opt.tstamp_ok)
5219 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5221 /* Make sure socket is routed, for
5224 icsk->icsk_af_ops->rebuild_header(sk);
5226 tcp_init_metrics(sk);
5228 tcp_init_congestion_control(sk);
5230 /* Prevent spurious tcp_cwnd_restart() on
5231 * first data packet.
5233 tp->lsndtime = tcp_time_stamp;
5236 tcp_initialize_rcv_mss(sk);
5237 tcp_init_buffer_space(sk);
5238 tcp_fast_path_on(tp);
5245 if (tp->snd_una == tp->write_seq) {
5246 tcp_set_state(sk, TCP_FIN_WAIT2);
5247 sk->sk_shutdown |= SEND_SHUTDOWN;
5248 dst_confirm(sk->sk_dst_cache);
5250 if (!sock_flag(sk, SOCK_DEAD))
5251 /* Wake up lingering close() */
5252 sk->sk_state_change(sk);
5256 if (tp->linger2 < 0 ||
5257 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5258 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5260 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5264 tmo = tcp_fin_time(sk);
5265 if (tmo > TCP_TIMEWAIT_LEN) {
5266 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5267 } else if (th->fin || sock_owned_by_user(sk)) {
5268 /* Bad case. We could lose such FIN otherwise.
5269 * It is not a big problem, but it looks confusing
5270 * and not so rare event. We still can lose it now,
5271 * if it spins in bh_lock_sock(), but it is really
5274 inet_csk_reset_keepalive_timer(sk, tmo);
5276 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5284 if (tp->snd_una == tp->write_seq) {
5285 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5291 if (tp->snd_una == tp->write_seq) {
5292 tcp_update_metrics(sk);
5301 /* step 6: check the URG bit */
5302 tcp_urg(sk, skb, th);
5304 /* step 7: process the segment text */
5305 switch (sk->sk_state) {
5306 case TCP_CLOSE_WAIT:
5309 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5313 /* RFC 793 says to queue data in these states,
5314 * RFC 1122 says we MUST send a reset.
5315 * BSD 4.4 also does reset.
5317 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5318 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5319 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5320 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5326 case TCP_ESTABLISHED:
5327 tcp_data_queue(sk, skb);
5332 /* tcp_data could move socket to TIME-WAIT */
5333 if (sk->sk_state != TCP_CLOSE) {
5334 tcp_data_snd_check(sk);
5335 tcp_ack_snd_check(sk);
5345 EXPORT_SYMBOL(sysctl_tcp_ecn);
5346 EXPORT_SYMBOL(sysctl_tcp_reordering);
5347 EXPORT_SYMBOL(tcp_parse_options);
5348 EXPORT_SYMBOL(tcp_rcv_established);
5349 EXPORT_SYMBOL(tcp_rcv_state_process);
5350 EXPORT_SYMBOL(tcp_initialize_rcv_mss);