2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #include <linux/module.h>
66 #include <linux/sysctl.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71 #include <asm/unaligned.h>
72 #include <net/netdma.h>
74 int sysctl_tcp_timestamps __read_mostly = 1;
75 int sysctl_tcp_window_scaling __read_mostly = 1;
76 int sysctl_tcp_sack __read_mostly = 1;
77 int sysctl_tcp_fack __read_mostly = 1;
78 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
79 int sysctl_tcp_ecn __read_mostly;
80 int sysctl_tcp_dsack __read_mostly = 1;
81 int sysctl_tcp_app_win __read_mostly = 31;
82 int sysctl_tcp_adv_win_scale __read_mostly = 2;
84 int sysctl_tcp_stdurg __read_mostly;
85 int sysctl_tcp_rfc1337 __read_mostly;
86 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
87 int sysctl_tcp_frto __read_mostly = 2;
88 int sysctl_tcp_frto_response __read_mostly;
89 int sysctl_tcp_nometrics_save __read_mostly;
91 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
92 int sysctl_tcp_abc __read_mostly;
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
104 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
105 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
106 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
107 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
118 /* Adapt the MSS value used to make delayed ack decision to the
121 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
123 struct inet_connection_sock *icsk = inet_csk(sk);
124 const unsigned int lss = icsk->icsk_ack.last_seg_size;
127 icsk->icsk_ack.last_seg_size = 0;
129 /* skb->len may jitter because of SACKs, even if peer
130 * sends good full-sized frames.
132 len = skb_shinfo(skb)->gso_size ? : skb->len;
133 if (len >= icsk->icsk_ack.rcv_mss) {
134 icsk->icsk_ack.rcv_mss = len;
136 /* Otherwise, we make more careful check taking into account,
137 * that SACKs block is variable.
139 * "len" is invariant segment length, including TCP header.
141 len += skb->data - skb_transport_header(skb);
142 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
143 /* If PSH is not set, packet should be
144 * full sized, provided peer TCP is not badly broken.
145 * This observation (if it is correct 8)) allows
146 * to handle super-low mtu links fairly.
148 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
149 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
150 /* Subtract also invariant (if peer is RFC compliant),
151 * tcp header plus fixed timestamp option length.
152 * Resulting "len" is MSS free of SACK jitter.
154 len -= tcp_sk(sk)->tcp_header_len;
155 icsk->icsk_ack.last_seg_size = len;
157 icsk->icsk_ack.rcv_mss = len;
161 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
162 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
163 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
167 static void tcp_incr_quickack(struct sock *sk)
169 struct inet_connection_sock *icsk = inet_csk(sk);
170 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
174 if (quickacks > icsk->icsk_ack.quick)
175 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
178 void tcp_enter_quickack_mode(struct sock *sk)
180 struct inet_connection_sock *icsk = inet_csk(sk);
181 tcp_incr_quickack(sk);
182 icsk->icsk_ack.pingpong = 0;
183 icsk->icsk_ack.ato = TCP_ATO_MIN;
186 /* Send ACKs quickly, if "quick" count is not exhausted
187 * and the session is not interactive.
190 static inline int tcp_in_quickack_mode(const struct sock *sk)
192 const struct inet_connection_sock *icsk = inet_csk(sk);
193 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
196 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
198 if (tp->ecn_flags & TCP_ECN_OK)
199 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
202 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
204 if (tcp_hdr(skb)->cwr)
205 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
208 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
210 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
213 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
215 if (tp->ecn_flags & TCP_ECN_OK) {
216 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
217 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
218 /* Funny extension: if ECT is not set on a segment,
219 * it is surely retransmit. It is not in ECN RFC,
220 * but Linux follows this rule. */
221 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
222 tcp_enter_quickack_mode((struct sock *)tp);
226 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
228 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
229 tp->ecn_flags &= ~TCP_ECN_OK;
232 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
234 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
235 tp->ecn_flags &= ~TCP_ECN_OK;
238 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
240 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
245 /* Buffer size and advertised window tuning.
247 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
250 static void tcp_fixup_sndbuf(struct sock *sk)
252 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
253 sizeof(struct sk_buff);
255 if (sk->sk_sndbuf < 3 * sndmem)
256 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
259 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261 * All tcp_full_space() is split to two parts: "network" buffer, allocated
262 * forward and advertised in receiver window (tp->rcv_wnd) and
263 * "application buffer", required to isolate scheduling/application
264 * latencies from network.
265 * window_clamp is maximal advertised window. It can be less than
266 * tcp_full_space(), in this case tcp_full_space() - window_clamp
267 * is reserved for "application" buffer. The less window_clamp is
268 * the smoother our behaviour from viewpoint of network, but the lower
269 * throughput and the higher sensitivity of the connection to losses. 8)
271 * rcv_ssthresh is more strict window_clamp used at "slow start"
272 * phase to predict further behaviour of this connection.
273 * It is used for two goals:
274 * - to enforce header prediction at sender, even when application
275 * requires some significant "application buffer". It is check #1.
276 * - to prevent pruning of receive queue because of misprediction
277 * of receiver window. Check #2.
279 * The scheme does not work when sender sends good segments opening
280 * window and then starts to feed us spaghetti. But it should work
281 * in common situations. Otherwise, we have to rely on queue collapsing.
284 /* Slow part of check#2. */
285 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
287 struct tcp_sock *tp = tcp_sk(sk);
289 int truesize = tcp_win_from_space(skb->truesize) >> 1;
290 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
292 while (tp->rcv_ssthresh <= window) {
293 if (truesize <= skb->len)
294 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
302 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
304 struct tcp_sock *tp = tcp_sk(sk);
307 if (tp->rcv_ssthresh < tp->window_clamp &&
308 (int)tp->rcv_ssthresh < tcp_space(sk) &&
309 !tcp_memory_pressure) {
312 /* Check #2. Increase window, if skb with such overhead
313 * will fit to rcvbuf in future.
315 if (tcp_win_from_space(skb->truesize) <= skb->len)
316 incr = 2 * tp->advmss;
318 incr = __tcp_grow_window(sk, skb);
321 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
323 inet_csk(sk)->icsk_ack.quick |= 1;
328 /* 3. Tuning rcvbuf, when connection enters established state. */
330 static void tcp_fixup_rcvbuf(struct sock *sk)
332 struct tcp_sock *tp = tcp_sk(sk);
333 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
335 /* Try to select rcvbuf so that 4 mss-sized segments
336 * will fit to window and corresponding skbs will fit to our rcvbuf.
337 * (was 3; 4 is minimum to allow fast retransmit to work.)
339 while (tcp_win_from_space(rcvmem) < tp->advmss)
341 if (sk->sk_rcvbuf < 4 * rcvmem)
342 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
345 /* 4. Try to fixup all. It is made immediately after connection enters
348 static void tcp_init_buffer_space(struct sock *sk)
350 struct tcp_sock *tp = tcp_sk(sk);
353 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
354 tcp_fixup_rcvbuf(sk);
355 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
356 tcp_fixup_sndbuf(sk);
358 tp->rcvq_space.space = tp->rcv_wnd;
360 maxwin = tcp_full_space(sk);
362 if (tp->window_clamp >= maxwin) {
363 tp->window_clamp = maxwin;
365 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
366 tp->window_clamp = max(maxwin -
367 (maxwin >> sysctl_tcp_app_win),
371 /* Force reservation of one segment. */
372 if (sysctl_tcp_app_win &&
373 tp->window_clamp > 2 * tp->advmss &&
374 tp->window_clamp + tp->advmss > maxwin)
375 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
377 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
378 tp->snd_cwnd_stamp = tcp_time_stamp;
381 /* 5. Recalculate window clamp after socket hit its memory bounds. */
382 static void tcp_clamp_window(struct sock *sk)
384 struct tcp_sock *tp = tcp_sk(sk);
385 struct inet_connection_sock *icsk = inet_csk(sk);
387 icsk->icsk_ack.quick = 0;
389 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
390 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
391 !tcp_memory_pressure &&
392 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
393 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
396 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
397 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
400 /* Initialize RCV_MSS value.
401 * RCV_MSS is an our guess about MSS used by the peer.
402 * We haven't any direct information about the MSS.
403 * It's better to underestimate the RCV_MSS rather than overestimate.
404 * Overestimations make us ACKing less frequently than needed.
405 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407 void tcp_initialize_rcv_mss(struct sock *sk)
409 struct tcp_sock *tp = tcp_sk(sk);
410 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
412 hint = min(hint, tp->rcv_wnd / 2);
413 hint = min(hint, TCP_MIN_RCVMSS);
414 hint = max(hint, TCP_MIN_MSS);
416 inet_csk(sk)->icsk_ack.rcv_mss = hint;
419 /* Receiver "autotuning" code.
421 * The algorithm for RTT estimation w/o timestamps is based on
422 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
423 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425 * More detail on this code can be found at
426 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
427 * though this reference is out of date. A new paper
430 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
432 u32 new_sample = tp->rcv_rtt_est.rtt;
438 if (new_sample != 0) {
439 /* If we sample in larger samples in the non-timestamp
440 * case, we could grossly overestimate the RTT especially
441 * with chatty applications or bulk transfer apps which
442 * are stalled on filesystem I/O.
444 * Also, since we are only going for a minimum in the
445 * non-timestamp case, we do not smooth things out
446 * else with timestamps disabled convergence takes too
450 m -= (new_sample >> 3);
452 } else if (m < new_sample)
455 /* No previous measure. */
459 if (tp->rcv_rtt_est.rtt != new_sample)
460 tp->rcv_rtt_est.rtt = new_sample;
463 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
465 if (tp->rcv_rtt_est.time == 0)
467 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
469 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
472 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
473 tp->rcv_rtt_est.time = tcp_time_stamp;
476 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
477 const struct sk_buff *skb)
479 struct tcp_sock *tp = tcp_sk(sk);
480 if (tp->rx_opt.rcv_tsecr &&
481 (TCP_SKB_CB(skb)->end_seq -
482 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
483 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
487 * This function should be called every time data is copied to user space.
488 * It calculates the appropriate TCP receive buffer space.
490 void tcp_rcv_space_adjust(struct sock *sk)
492 struct tcp_sock *tp = tcp_sk(sk);
496 if (tp->rcvq_space.time == 0)
499 time = tcp_time_stamp - tp->rcvq_space.time;
500 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
503 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
505 space = max(tp->rcvq_space.space, space);
507 if (tp->rcvq_space.space != space) {
510 tp->rcvq_space.space = space;
512 if (sysctl_tcp_moderate_rcvbuf &&
513 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
514 int new_clamp = space;
516 /* Receive space grows, normalize in order to
517 * take into account packet headers and sk_buff
518 * structure overhead.
523 rcvmem = (tp->advmss + MAX_TCP_HEADER +
524 16 + sizeof(struct sk_buff));
525 while (tcp_win_from_space(rcvmem) < tp->advmss)
528 space = min(space, sysctl_tcp_rmem[2]);
529 if (space > sk->sk_rcvbuf) {
530 sk->sk_rcvbuf = space;
532 /* Make the window clamp follow along. */
533 tp->window_clamp = new_clamp;
539 tp->rcvq_space.seq = tp->copied_seq;
540 tp->rcvq_space.time = tcp_time_stamp;
543 /* There is something which you must keep in mind when you analyze the
544 * behavior of the tp->ato delayed ack timeout interval. When a
545 * connection starts up, we want to ack as quickly as possible. The
546 * problem is that "good" TCP's do slow start at the beginning of data
547 * transmission. The means that until we send the first few ACK's the
548 * sender will sit on his end and only queue most of his data, because
549 * he can only send snd_cwnd unacked packets at any given time. For
550 * each ACK we send, he increments snd_cwnd and transmits more of his
553 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
555 struct tcp_sock *tp = tcp_sk(sk);
556 struct inet_connection_sock *icsk = inet_csk(sk);
559 inet_csk_schedule_ack(sk);
561 tcp_measure_rcv_mss(sk, skb);
563 tcp_rcv_rtt_measure(tp);
565 now = tcp_time_stamp;
567 if (!icsk->icsk_ack.ato) {
568 /* The _first_ data packet received, initialize
569 * delayed ACK engine.
571 tcp_incr_quickack(sk);
572 icsk->icsk_ack.ato = TCP_ATO_MIN;
574 int m = now - icsk->icsk_ack.lrcvtime;
576 if (m <= TCP_ATO_MIN / 2) {
577 /* The fastest case is the first. */
578 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
579 } else if (m < icsk->icsk_ack.ato) {
580 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
581 if (icsk->icsk_ack.ato > icsk->icsk_rto)
582 icsk->icsk_ack.ato = icsk->icsk_rto;
583 } else if (m > icsk->icsk_rto) {
584 /* Too long gap. Apparently sender failed to
585 * restart window, so that we send ACKs quickly.
587 tcp_incr_quickack(sk);
591 icsk->icsk_ack.lrcvtime = now;
593 TCP_ECN_check_ce(tp, skb);
596 tcp_grow_window(sk, skb);
599 static u32 tcp_rto_min(struct sock *sk)
601 struct dst_entry *dst = __sk_dst_get(sk);
602 u32 rto_min = TCP_RTO_MIN;
604 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
605 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
609 /* Called to compute a smoothed rtt estimate. The data fed to this
610 * routine either comes from timestamps, or from segments that were
611 * known _not_ to have been retransmitted [see Karn/Partridge
612 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
613 * piece by Van Jacobson.
614 * NOTE: the next three routines used to be one big routine.
615 * To save cycles in the RFC 1323 implementation it was better to break
616 * it up into three procedures. -- erics
618 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
620 struct tcp_sock *tp = tcp_sk(sk);
621 long m = mrtt; /* RTT */
623 /* The following amusing code comes from Jacobson's
624 * article in SIGCOMM '88. Note that rtt and mdev
625 * are scaled versions of rtt and mean deviation.
626 * This is designed to be as fast as possible
627 * m stands for "measurement".
629 * On a 1990 paper the rto value is changed to:
630 * RTO = rtt + 4 * mdev
632 * Funny. This algorithm seems to be very broken.
633 * These formulae increase RTO, when it should be decreased, increase
634 * too slowly, when it should be increased quickly, decrease too quickly
635 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
636 * does not matter how to _calculate_ it. Seems, it was trap
637 * that VJ failed to avoid. 8)
642 m -= (tp->srtt >> 3); /* m is now error in rtt est */
643 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
645 m = -m; /* m is now abs(error) */
646 m -= (tp->mdev >> 2); /* similar update on mdev */
647 /* This is similar to one of Eifel findings.
648 * Eifel blocks mdev updates when rtt decreases.
649 * This solution is a bit different: we use finer gain
650 * for mdev in this case (alpha*beta).
651 * Like Eifel it also prevents growth of rto,
652 * but also it limits too fast rto decreases,
653 * happening in pure Eifel.
658 m -= (tp->mdev >> 2); /* similar update on mdev */
660 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
661 if (tp->mdev > tp->mdev_max) {
662 tp->mdev_max = tp->mdev;
663 if (tp->mdev_max > tp->rttvar)
664 tp->rttvar = tp->mdev_max;
666 if (after(tp->snd_una, tp->rtt_seq)) {
667 if (tp->mdev_max < tp->rttvar)
668 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
669 tp->rtt_seq = tp->snd_nxt;
670 tp->mdev_max = tcp_rto_min(sk);
673 /* no previous measure. */
674 tp->srtt = m << 3; /* take the measured time to be rtt */
675 tp->mdev = m << 1; /* make sure rto = 3*rtt */
676 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
677 tp->rtt_seq = tp->snd_nxt;
681 /* Calculate rto without backoff. This is the second half of Van Jacobson's
682 * routine referred to above.
684 static inline void tcp_set_rto(struct sock *sk)
686 const struct tcp_sock *tp = tcp_sk(sk);
687 /* Old crap is replaced with new one. 8)
690 * 1. If rtt variance happened to be less 50msec, it is hallucination.
691 * It cannot be less due to utterly erratic ACK generation made
692 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
693 * to do with delayed acks, because at cwnd>2 true delack timeout
694 * is invisible. Actually, Linux-2.4 also generates erratic
695 * ACKs in some circumstances.
697 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
699 /* 2. Fixups made earlier cannot be right.
700 * If we do not estimate RTO correctly without them,
701 * all the algo is pure shit and should be replaced
702 * with correct one. It is exactly, which we pretend to do.
706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707 * guarantees that rto is higher.
709 static inline void tcp_bound_rto(struct sock *sk)
711 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
712 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
715 /* Save metrics learned by this TCP session.
716 This function is called only, when TCP finishes successfully
717 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
719 void tcp_update_metrics(struct sock *sk)
721 struct tcp_sock *tp = tcp_sk(sk);
722 struct dst_entry *dst = __sk_dst_get(sk);
724 if (sysctl_tcp_nometrics_save)
729 if (dst && (dst->flags & DST_HOST)) {
730 const struct inet_connection_sock *icsk = inet_csk(sk);
734 if (icsk->icsk_backoff || !tp->srtt) {
735 /* This session failed to estimate rtt. Why?
736 * Probably, no packets returned in time.
739 if (!(dst_metric_locked(dst, RTAX_RTT)))
740 dst->metrics[RTAX_RTT - 1] = 0;
744 rtt = dst_metric_rtt(dst, RTAX_RTT);
747 /* If newly calculated rtt larger than stored one,
748 * store new one. Otherwise, use EWMA. Remember,
749 * rtt overestimation is always better than underestimation.
751 if (!(dst_metric_locked(dst, RTAX_RTT))) {
753 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
755 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
758 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
763 /* Scale deviation to rttvar fixed point */
768 var = dst_metric_rtt(dst, RTAX_RTTVAR);
772 var -= (var - m) >> 2;
774 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
777 if (tp->snd_ssthresh >= 0xFFFF) {
778 /* Slow start still did not finish. */
779 if (dst_metric(dst, RTAX_SSTHRESH) &&
780 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
781 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
782 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
783 if (!dst_metric_locked(dst, RTAX_CWND) &&
784 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
785 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
786 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
787 icsk->icsk_ca_state == TCP_CA_Open) {
788 /* Cong. avoidance phase, cwnd is reliable. */
789 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
790 dst->metrics[RTAX_SSTHRESH-1] =
791 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
792 if (!dst_metric_locked(dst, RTAX_CWND))
793 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
795 /* Else slow start did not finish, cwnd is non-sense,
796 ssthresh may be also invalid.
798 if (!dst_metric_locked(dst, RTAX_CWND))
799 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
800 if (dst_metric(dst, RTAX_SSTHRESH) &&
801 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
802 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
803 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
806 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
807 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
808 tp->reordering != sysctl_tcp_reordering)
809 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
814 /* Numbers are taken from RFC3390.
816 * John Heffner states:
818 * The RFC specifies a window of no more than 4380 bytes
819 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
820 * is a bit misleading because they use a clamp at 4380 bytes
821 * rather than use a multiplier in the relevant range.
823 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
825 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
828 if (tp->mss_cache > 1460)
831 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
833 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
836 /* Set slow start threshold and cwnd not falling to slow start */
837 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
839 struct tcp_sock *tp = tcp_sk(sk);
840 const struct inet_connection_sock *icsk = inet_csk(sk);
842 tp->prior_ssthresh = 0;
844 if (icsk->icsk_ca_state < TCP_CA_CWR) {
847 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
848 tp->snd_cwnd = min(tp->snd_cwnd,
849 tcp_packets_in_flight(tp) + 1U);
850 tp->snd_cwnd_cnt = 0;
851 tp->high_seq = tp->snd_nxt;
852 tp->snd_cwnd_stamp = tcp_time_stamp;
853 TCP_ECN_queue_cwr(tp);
855 tcp_set_ca_state(sk, TCP_CA_CWR);
860 * Packet counting of FACK is based on in-order assumptions, therefore TCP
861 * disables it when reordering is detected
863 static void tcp_disable_fack(struct tcp_sock *tp)
865 /* RFC3517 uses different metric in lost marker => reset on change */
867 tp->lost_skb_hint = NULL;
868 tp->rx_opt.sack_ok &= ~2;
871 /* Take a notice that peer is sending D-SACKs */
872 static void tcp_dsack_seen(struct tcp_sock *tp)
874 tp->rx_opt.sack_ok |= 4;
877 /* Initialize metrics on socket. */
879 static void tcp_init_metrics(struct sock *sk)
881 struct tcp_sock *tp = tcp_sk(sk);
882 struct dst_entry *dst = __sk_dst_get(sk);
889 if (dst_metric_locked(dst, RTAX_CWND))
890 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
891 if (dst_metric(dst, RTAX_SSTHRESH)) {
892 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
893 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
894 tp->snd_ssthresh = tp->snd_cwnd_clamp;
896 if (dst_metric(dst, RTAX_REORDERING) &&
897 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
898 tcp_disable_fack(tp);
899 tp->reordering = dst_metric(dst, RTAX_REORDERING);
902 if (dst_metric(dst, RTAX_RTT) == 0)
905 if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
908 /* Initial rtt is determined from SYN,SYN-ACK.
909 * The segment is small and rtt may appear much
910 * less than real one. Use per-dst memory
911 * to make it more realistic.
913 * A bit of theory. RTT is time passed after "normal" sized packet
914 * is sent until it is ACKed. In normal circumstances sending small
915 * packets force peer to delay ACKs and calculation is correct too.
916 * The algorithm is adaptive and, provided we follow specs, it
917 * NEVER underestimate RTT. BUT! If peer tries to make some clever
918 * tricks sort of "quick acks" for time long enough to decrease RTT
919 * to low value, and then abruptly stops to do it and starts to delay
920 * ACKs, wait for troubles.
922 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
923 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
924 tp->rtt_seq = tp->snd_nxt;
926 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
927 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
928 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
932 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
934 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
935 tp->snd_cwnd_stamp = tcp_time_stamp;
939 /* Play conservative. If timestamps are not
940 * supported, TCP will fail to recalculate correct
941 * rtt, if initial rto is too small. FORGET ALL AND RESET!
943 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
945 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
946 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
950 static void tcp_update_reordering(struct sock *sk, const int metric,
953 struct tcp_sock *tp = tcp_sk(sk);
954 if (metric > tp->reordering) {
957 tp->reordering = min(TCP_MAX_REORDERING, metric);
959 /* This exciting event is worth to be remembered. 8) */
961 mib_idx = LINUX_MIB_TCPTSREORDER;
962 else if (tcp_is_reno(tp))
963 mib_idx = LINUX_MIB_TCPRENOREORDER;
964 else if (tcp_is_fack(tp))
965 mib_idx = LINUX_MIB_TCPFACKREORDER;
967 mib_idx = LINUX_MIB_TCPSACKREORDER;
969 NET_INC_STATS_BH(sock_net(sk), mib_idx);
970 #if FASTRETRANS_DEBUG > 1
971 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
972 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
976 tp->undo_marker ? tp->undo_retrans : 0);
978 tcp_disable_fack(tp);
982 /* This must be called before lost_out is incremented */
983 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
985 if ((tp->retransmit_skb_hint == NULL) ||
986 before(TCP_SKB_CB(skb)->seq,
987 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
988 tp->retransmit_skb_hint = skb;
991 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
992 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
995 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
997 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
998 tcp_verify_retransmit_hint(tp, skb);
1000 tp->lost_out += tcp_skb_pcount(skb);
1001 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1005 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
1007 tcp_verify_retransmit_hint(tp, skb);
1009 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1010 tp->lost_out += tcp_skb_pcount(skb);
1011 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1015 /* This procedure tags the retransmission queue when SACKs arrive.
1017 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1018 * Packets in queue with these bits set are counted in variables
1019 * sacked_out, retrans_out and lost_out, correspondingly.
1021 * Valid combinations are:
1022 * Tag InFlight Description
1023 * 0 1 - orig segment is in flight.
1024 * S 0 - nothing flies, orig reached receiver.
1025 * L 0 - nothing flies, orig lost by net.
1026 * R 2 - both orig and retransmit are in flight.
1027 * L|R 1 - orig is lost, retransmit is in flight.
1028 * S|R 1 - orig reached receiver, retrans is still in flight.
1029 * (L|S|R is logically valid, it could occur when L|R is sacked,
1030 * but it is equivalent to plain S and code short-curcuits it to S.
1031 * L|S is logically invalid, it would mean -1 packet in flight 8))
1033 * These 6 states form finite state machine, controlled by the following events:
1034 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1035 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1036 * 3. Loss detection event of one of three flavors:
1037 * A. Scoreboard estimator decided the packet is lost.
1038 * A'. Reno "three dupacks" marks head of queue lost.
1039 * A''. Its FACK modfication, head until snd.fack is lost.
1040 * B. SACK arrives sacking data transmitted after never retransmitted
1041 * hole was sent out.
1042 * C. SACK arrives sacking SND.NXT at the moment, when the
1043 * segment was retransmitted.
1044 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1046 * It is pleasant to note, that state diagram turns out to be commutative,
1047 * so that we are allowed not to be bothered by order of our actions,
1048 * when multiple events arrive simultaneously. (see the function below).
1050 * Reordering detection.
1051 * --------------------
1052 * Reordering metric is maximal distance, which a packet can be displaced
1053 * in packet stream. With SACKs we can estimate it:
1055 * 1. SACK fills old hole and the corresponding segment was not
1056 * ever retransmitted -> reordering. Alas, we cannot use it
1057 * when segment was retransmitted.
1058 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1059 * for retransmitted and already SACKed segment -> reordering..
1060 * Both of these heuristics are not used in Loss state, when we cannot
1061 * account for retransmits accurately.
1063 * SACK block validation.
1064 * ----------------------
1066 * SACK block range validation checks that the received SACK block fits to
1067 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1068 * Note that SND.UNA is not included to the range though being valid because
1069 * it means that the receiver is rather inconsistent with itself reporting
1070 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1071 * perfectly valid, however, in light of RFC2018 which explicitly states
1072 * that "SACK block MUST reflect the newest segment. Even if the newest
1073 * segment is going to be discarded ...", not that it looks very clever
1074 * in case of head skb. Due to potentional receiver driven attacks, we
1075 * choose to avoid immediate execution of a walk in write queue due to
1076 * reneging and defer head skb's loss recovery to standard loss recovery
1077 * procedure that will eventually trigger (nothing forbids us doing this).
1079 * Implements also blockage to start_seq wrap-around. Problem lies in the
1080 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1081 * there's no guarantee that it will be before snd_nxt (n). The problem
1082 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1085 * <- outs wnd -> <- wrapzone ->
1086 * u e n u_w e_w s n_w
1088 * |<------------+------+----- TCP seqno space --------------+---------->|
1089 * ...-- <2^31 ->| |<--------...
1090 * ...---- >2^31 ------>| |<--------...
1092 * Current code wouldn't be vulnerable but it's better still to discard such
1093 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1094 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1095 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1096 * equal to the ideal case (infinite seqno space without wrap caused issues).
1098 * With D-SACK the lower bound is extended to cover sequence space below
1099 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1100 * again, D-SACK block must not to go across snd_una (for the same reason as
1101 * for the normal SACK blocks, explained above). But there all simplicity
1102 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1103 * fully below undo_marker they do not affect behavior in anyway and can
1104 * therefore be safely ignored. In rare cases (which are more or less
1105 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1106 * fragmentation and packet reordering past skb's retransmission. To consider
1107 * them correctly, the acceptable range must be extended even more though
1108 * the exact amount is rather hard to quantify. However, tp->max_window can
1109 * be used as an exaggerated estimate.
1111 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1112 u32 start_seq, u32 end_seq)
1114 /* Too far in future, or reversed (interpretation is ambiguous) */
1115 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1118 /* Nasty start_seq wrap-around check (see comments above) */
1119 if (!before(start_seq, tp->snd_nxt))
1122 /* In outstanding window? ...This is valid exit for D-SACKs too.
1123 * start_seq == snd_una is non-sensical (see comments above)
1125 if (after(start_seq, tp->snd_una))
1128 if (!is_dsack || !tp->undo_marker)
1131 /* ...Then it's D-SACK, and must reside below snd_una completely */
1132 if (!after(end_seq, tp->snd_una))
1135 if (!before(start_seq, tp->undo_marker))
1139 if (!after(end_seq, tp->undo_marker))
1142 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1143 * start_seq < undo_marker and end_seq >= undo_marker.
1145 return !before(start_seq, end_seq - tp->max_window);
1148 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1149 * Event "C". Later note: FACK people cheated me again 8), we have to account
1150 * for reordering! Ugly, but should help.
1152 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1153 * less than what is now known to be received by the other end (derived from
1154 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1155 * retransmitted skbs to avoid some costly processing per ACKs.
1157 static void tcp_mark_lost_retrans(struct sock *sk)
1159 const struct inet_connection_sock *icsk = inet_csk(sk);
1160 struct tcp_sock *tp = tcp_sk(sk);
1161 struct sk_buff *skb;
1163 u32 new_low_seq = tp->snd_nxt;
1164 u32 received_upto = tcp_highest_sack_seq(tp);
1166 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1167 !after(received_upto, tp->lost_retrans_low) ||
1168 icsk->icsk_ca_state != TCP_CA_Recovery)
1171 tcp_for_write_queue(skb, sk) {
1172 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1174 if (skb == tcp_send_head(sk))
1176 if (cnt == tp->retrans_out)
1178 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1181 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1184 if (after(received_upto, ack_seq) &&
1186 !before(received_upto,
1187 ack_seq + tp->reordering * tp->mss_cache))) {
1188 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1189 tp->retrans_out -= tcp_skb_pcount(skb);
1191 tcp_skb_mark_lost_uncond_verify(tp, skb);
1192 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1194 if (before(ack_seq, new_low_seq))
1195 new_low_seq = ack_seq;
1196 cnt += tcp_skb_pcount(skb);
1200 if (tp->retrans_out)
1201 tp->lost_retrans_low = new_low_seq;
1204 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1205 struct tcp_sack_block_wire *sp, int num_sacks,
1208 struct tcp_sock *tp = tcp_sk(sk);
1209 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1210 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1213 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1216 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1217 } else if (num_sacks > 1) {
1218 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1219 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1221 if (!after(end_seq_0, end_seq_1) &&
1222 !before(start_seq_0, start_seq_1)) {
1225 NET_INC_STATS_BH(sock_net(sk),
1226 LINUX_MIB_TCPDSACKOFORECV);
1230 /* D-SACK for already forgotten data... Do dumb counting. */
1232 !after(end_seq_0, prior_snd_una) &&
1233 after(end_seq_0, tp->undo_marker))
1239 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1240 * the incoming SACK may not exactly match but we can find smaller MSS
1241 * aligned portion of it that matches. Therefore we might need to fragment
1242 * which may fail and creates some hassle (caller must handle error case
1245 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1246 u32 start_seq, u32 end_seq)
1249 unsigned int pkt_len;
1251 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1252 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1254 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1255 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1257 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1260 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1262 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1263 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1271 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1272 int *reord, int dup_sack, int fack_count)
1274 struct tcp_sock *tp = tcp_sk(sk);
1275 u8 sacked = TCP_SKB_CB(skb)->sacked;
1278 /* Account D-SACK for retransmitted packet. */
1279 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1280 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1282 if (sacked & TCPCB_SACKED_ACKED)
1283 *reord = min(fack_count, *reord);
1286 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1287 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1290 if (!(sacked & TCPCB_SACKED_ACKED)) {
1291 if (sacked & TCPCB_SACKED_RETRANS) {
1292 /* If the segment is not tagged as lost,
1293 * we do not clear RETRANS, believing
1294 * that retransmission is still in flight.
1296 if (sacked & TCPCB_LOST) {
1297 TCP_SKB_CB(skb)->sacked &=
1298 ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1299 tp->lost_out -= tcp_skb_pcount(skb);
1300 tp->retrans_out -= tcp_skb_pcount(skb);
1302 /* clear lost hint */
1303 tp->retransmit_skb_hint = NULL;
1306 if (!(sacked & TCPCB_RETRANS)) {
1307 /* New sack for not retransmitted frame,
1308 * which was in hole. It is reordering.
1310 if (before(TCP_SKB_CB(skb)->seq,
1311 tcp_highest_sack_seq(tp)))
1312 *reord = min(fack_count, *reord);
1314 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1315 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1316 flag |= FLAG_ONLY_ORIG_SACKED;
1319 if (sacked & TCPCB_LOST) {
1320 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1321 tp->lost_out -= tcp_skb_pcount(skb);
1323 /* clear lost hint */
1324 tp->retransmit_skb_hint = NULL;
1328 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1329 flag |= FLAG_DATA_SACKED;
1330 tp->sacked_out += tcp_skb_pcount(skb);
1332 fack_count += tcp_skb_pcount(skb);
1334 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1335 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1336 before(TCP_SKB_CB(skb)->seq,
1337 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1338 tp->lost_cnt_hint += tcp_skb_pcount(skb);
1340 if (fack_count > tp->fackets_out)
1341 tp->fackets_out = fack_count;
1343 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1344 tcp_advance_highest_sack(sk, skb);
1347 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1348 * frames and clear it. undo_retrans is decreased above, L|R frames
1349 * are accounted above as well.
1351 if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) {
1352 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1353 tp->retrans_out -= tcp_skb_pcount(skb);
1354 tp->retransmit_skb_hint = NULL;
1360 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1361 struct tcp_sack_block *next_dup,
1362 u32 start_seq, u32 end_seq,
1363 int dup_sack_in, int *fack_count,
1364 int *reord, int *flag)
1366 tcp_for_write_queue_from(skb, sk) {
1368 int dup_sack = dup_sack_in;
1370 if (skb == tcp_send_head(sk))
1373 /* queue is in-order => we can short-circuit the walk early */
1374 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1377 if ((next_dup != NULL) &&
1378 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1379 in_sack = tcp_match_skb_to_sack(sk, skb,
1380 next_dup->start_seq,
1387 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq,
1389 if (unlikely(in_sack < 0))
1393 *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack,
1396 *fack_count += tcp_skb_pcount(skb);
1401 /* Avoid all extra work that is being done by sacktag while walking in
1404 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1405 u32 skip_to_seq, int *fack_count)
1407 tcp_for_write_queue_from(skb, sk) {
1408 if (skb == tcp_send_head(sk))
1411 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1414 *fack_count += tcp_skb_pcount(skb);
1419 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1421 struct tcp_sack_block *next_dup,
1423 int *fack_count, int *reord,
1426 if (next_dup == NULL)
1429 if (before(next_dup->start_seq, skip_to_seq)) {
1430 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count);
1431 skb = tcp_sacktag_walk(skb, sk, NULL,
1432 next_dup->start_seq, next_dup->end_seq,
1433 1, fack_count, reord, flag);
1439 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1441 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1445 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1448 const struct inet_connection_sock *icsk = inet_csk(sk);
1449 struct tcp_sock *tp = tcp_sk(sk);
1450 unsigned char *ptr = (skb_transport_header(ack_skb) +
1451 TCP_SKB_CB(ack_skb)->sacked);
1452 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1453 struct tcp_sack_block sp[TCP_NUM_SACKS];
1454 struct tcp_sack_block *cache;
1455 struct sk_buff *skb;
1456 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1458 int reord = tp->packets_out;
1460 int found_dup_sack = 0;
1463 int first_sack_index;
1465 if (!tp->sacked_out) {
1466 if (WARN_ON(tp->fackets_out))
1467 tp->fackets_out = 0;
1468 tcp_highest_sack_reset(sk);
1471 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1472 num_sacks, prior_snd_una);
1474 flag |= FLAG_DSACKING_ACK;
1476 /* Eliminate too old ACKs, but take into
1477 * account more or less fresh ones, they can
1478 * contain valid SACK info.
1480 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1483 if (!tp->packets_out)
1487 first_sack_index = 0;
1488 for (i = 0; i < num_sacks; i++) {
1489 int dup_sack = !i && found_dup_sack;
1491 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1492 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1494 if (!tcp_is_sackblock_valid(tp, dup_sack,
1495 sp[used_sacks].start_seq,
1496 sp[used_sacks].end_seq)) {
1500 if (!tp->undo_marker)
1501 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1503 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1505 /* Don't count olds caused by ACK reordering */
1506 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1507 !after(sp[used_sacks].end_seq, tp->snd_una))
1509 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1512 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1514 first_sack_index = -1;
1518 /* Ignore very old stuff early */
1519 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1525 /* order SACK blocks to allow in order walk of the retrans queue */
1526 for (i = used_sacks - 1; i > 0; i--) {
1527 for (j = 0; j < i; j++) {
1528 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1529 struct tcp_sack_block tmp;
1535 /* Track where the first SACK block goes to */
1536 if (j == first_sack_index)
1537 first_sack_index = j + 1;
1542 skb = tcp_write_queue_head(sk);
1546 if (!tp->sacked_out) {
1547 /* It's already past, so skip checking against it */
1548 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1550 cache = tp->recv_sack_cache;
1551 /* Skip empty blocks in at head of the cache */
1552 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1557 while (i < used_sacks) {
1558 u32 start_seq = sp[i].start_seq;
1559 u32 end_seq = sp[i].end_seq;
1560 int dup_sack = (found_dup_sack && (i == first_sack_index));
1561 struct tcp_sack_block *next_dup = NULL;
1563 if (found_dup_sack && ((i + 1) == first_sack_index))
1564 next_dup = &sp[i + 1];
1566 /* Event "B" in the comment above. */
1567 if (after(end_seq, tp->high_seq))
1568 flag |= FLAG_DATA_LOST;
1570 /* Skip too early cached blocks */
1571 while (tcp_sack_cache_ok(tp, cache) &&
1572 !before(start_seq, cache->end_seq))
1575 /* Can skip some work by looking recv_sack_cache? */
1576 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1577 after(end_seq, cache->start_seq)) {
1580 if (before(start_seq, cache->start_seq)) {
1581 skb = tcp_sacktag_skip(skb, sk, start_seq,
1583 skb = tcp_sacktag_walk(skb, sk, next_dup,
1586 dup_sack, &fack_count,
1590 /* Rest of the block already fully processed? */
1591 if (!after(end_seq, cache->end_seq))
1594 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1596 &fack_count, &reord,
1599 /* ...tail remains todo... */
1600 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1601 /* ...but better entrypoint exists! */
1602 skb = tcp_highest_sack(sk);
1605 fack_count = tp->fackets_out;
1610 skb = tcp_sacktag_skip(skb, sk, cache->end_seq,
1612 /* Check overlap against next cached too (past this one already) */
1617 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1618 skb = tcp_highest_sack(sk);
1621 fack_count = tp->fackets_out;
1623 skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count);
1626 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1627 dup_sack, &fack_count, &reord, &flag);
1630 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1631 * due to in-order walk
1633 if (after(end_seq, tp->frto_highmark))
1634 flag &= ~FLAG_ONLY_ORIG_SACKED;
1639 /* Clear the head of the cache sack blocks so we can skip it next time */
1640 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1641 tp->recv_sack_cache[i].start_seq = 0;
1642 tp->recv_sack_cache[i].end_seq = 0;
1644 for (j = 0; j < used_sacks; j++)
1645 tp->recv_sack_cache[i++] = sp[j];
1647 tcp_mark_lost_retrans(sk);
1649 tcp_verify_left_out(tp);
1651 if ((reord < tp->fackets_out) &&
1652 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1653 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1654 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
1658 #if FASTRETRANS_DEBUG > 0
1659 WARN_ON((int)tp->sacked_out < 0);
1660 WARN_ON((int)tp->lost_out < 0);
1661 WARN_ON((int)tp->retrans_out < 0);
1662 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1667 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1668 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1670 int tcp_limit_reno_sacked(struct tcp_sock *tp)
1674 holes = max(tp->lost_out, 1U);
1675 holes = min(holes, tp->packets_out);
1677 if ((tp->sacked_out + holes) > tp->packets_out) {
1678 tp->sacked_out = tp->packets_out - holes;
1684 /* If we receive more dupacks than we expected counting segments
1685 * in assumption of absent reordering, interpret this as reordering.
1686 * The only another reason could be bug in receiver TCP.
1688 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1690 struct tcp_sock *tp = tcp_sk(sk);
1691 if (tcp_limit_reno_sacked(tp))
1692 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1695 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1697 static void tcp_add_reno_sack(struct sock *sk)
1699 struct tcp_sock *tp = tcp_sk(sk);
1701 tcp_check_reno_reordering(sk, 0);
1702 tcp_verify_left_out(tp);
1705 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1707 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1709 struct tcp_sock *tp = tcp_sk(sk);
1712 /* One ACK acked hole. The rest eat duplicate ACKs. */
1713 if (acked - 1 >= tp->sacked_out)
1716 tp->sacked_out -= acked - 1;
1718 tcp_check_reno_reordering(sk, acked);
1719 tcp_verify_left_out(tp);
1722 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1727 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1729 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1732 /* F-RTO can only be used if TCP has never retransmitted anything other than
1733 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1735 int tcp_use_frto(struct sock *sk)
1737 const struct tcp_sock *tp = tcp_sk(sk);
1738 const struct inet_connection_sock *icsk = inet_csk(sk);
1739 struct sk_buff *skb;
1741 if (!sysctl_tcp_frto)
1744 /* MTU probe and F-RTO won't really play nicely along currently */
1745 if (icsk->icsk_mtup.probe_size)
1748 if (tcp_is_sackfrto(tp))
1751 /* Avoid expensive walking of rexmit queue if possible */
1752 if (tp->retrans_out > 1)
1755 skb = tcp_write_queue_head(sk);
1756 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1757 tcp_for_write_queue_from(skb, sk) {
1758 if (skb == tcp_send_head(sk))
1760 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1762 /* Short-circuit when first non-SACKed skb has been checked */
1763 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1769 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1770 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1771 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1772 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1773 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1774 * bits are handled if the Loss state is really to be entered (in
1775 * tcp_enter_frto_loss).
1777 * Do like tcp_enter_loss() would; when RTO expires the second time it
1779 * "Reduce ssthresh if it has not yet been made inside this window."
1781 void tcp_enter_frto(struct sock *sk)
1783 const struct inet_connection_sock *icsk = inet_csk(sk);
1784 struct tcp_sock *tp = tcp_sk(sk);
1785 struct sk_buff *skb;
1787 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1788 tp->snd_una == tp->high_seq ||
1789 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1790 !icsk->icsk_retransmits)) {
1791 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1792 /* Our state is too optimistic in ssthresh() call because cwnd
1793 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1794 * recovery has not yet completed. Pattern would be this: RTO,
1795 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1797 * RFC4138 should be more specific on what to do, even though
1798 * RTO is quite unlikely to occur after the first Cumulative ACK
1799 * due to back-off and complexity of triggering events ...
1801 if (tp->frto_counter) {
1803 stored_cwnd = tp->snd_cwnd;
1805 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1806 tp->snd_cwnd = stored_cwnd;
1808 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1810 /* ... in theory, cong.control module could do "any tricks" in
1811 * ssthresh(), which means that ca_state, lost bits and lost_out
1812 * counter would have to be faked before the call occurs. We
1813 * consider that too expensive, unlikely and hacky, so modules
1814 * using these in ssthresh() must deal these incompatibility
1815 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1817 tcp_ca_event(sk, CA_EVENT_FRTO);
1820 tp->undo_marker = tp->snd_una;
1821 tp->undo_retrans = 0;
1823 skb = tcp_write_queue_head(sk);
1824 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1825 tp->undo_marker = 0;
1826 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1827 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1828 tp->retrans_out -= tcp_skb_pcount(skb);
1830 tcp_verify_left_out(tp);
1832 /* Too bad if TCP was application limited */
1833 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1835 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1836 * The last condition is necessary at least in tp->frto_counter case.
1838 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
1839 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1840 after(tp->high_seq, tp->snd_una)) {
1841 tp->frto_highmark = tp->high_seq;
1843 tp->frto_highmark = tp->snd_nxt;
1845 tcp_set_ca_state(sk, TCP_CA_Disorder);
1846 tp->high_seq = tp->snd_nxt;
1847 tp->frto_counter = 1;
1850 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1851 * which indicates that we should follow the traditional RTO recovery,
1852 * i.e. mark everything lost and do go-back-N retransmission.
1854 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1856 struct tcp_sock *tp = tcp_sk(sk);
1857 struct sk_buff *skb;
1860 tp->retrans_out = 0;
1861 if (tcp_is_reno(tp))
1862 tcp_reset_reno_sack(tp);
1864 tcp_for_write_queue(skb, sk) {
1865 if (skb == tcp_send_head(sk))
1868 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1870 * Count the retransmission made on RTO correctly (only when
1871 * waiting for the first ACK and did not get it)...
1873 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
1874 /* For some reason this R-bit might get cleared? */
1875 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1876 tp->retrans_out += tcp_skb_pcount(skb);
1877 /* ...enter this if branch just for the first segment */
1878 flag |= FLAG_DATA_ACKED;
1880 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1881 tp->undo_marker = 0;
1882 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1885 /* Marking forward transmissions that were made after RTO lost
1886 * can cause unnecessary retransmissions in some scenarios,
1887 * SACK blocks will mitigate that in some but not in all cases.
1888 * We used to not mark them but it was causing break-ups with
1889 * receivers that do only in-order receival.
1891 * TODO: we could detect presence of such receiver and select
1892 * different behavior per flow.
1894 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1895 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1896 tp->lost_out += tcp_skb_pcount(skb);
1897 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1900 tcp_verify_left_out(tp);
1902 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1903 tp->snd_cwnd_cnt = 0;
1904 tp->snd_cwnd_stamp = tcp_time_stamp;
1905 tp->frto_counter = 0;
1906 tp->bytes_acked = 0;
1908 tp->reordering = min_t(unsigned int, tp->reordering,
1909 sysctl_tcp_reordering);
1910 tcp_set_ca_state(sk, TCP_CA_Loss);
1911 tp->high_seq = tp->snd_nxt;
1912 TCP_ECN_queue_cwr(tp);
1914 tcp_clear_all_retrans_hints(tp);
1917 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1919 tp->retrans_out = 0;
1922 tp->undo_marker = 0;
1923 tp->undo_retrans = 0;
1926 void tcp_clear_retrans(struct tcp_sock *tp)
1928 tcp_clear_retrans_partial(tp);
1930 tp->fackets_out = 0;
1934 /* Enter Loss state. If "how" is not zero, forget all SACK information
1935 * and reset tags completely, otherwise preserve SACKs. If receiver
1936 * dropped its ofo queue, we will know this due to reneging detection.
1938 void tcp_enter_loss(struct sock *sk, int how)
1940 const struct inet_connection_sock *icsk = inet_csk(sk);
1941 struct tcp_sock *tp = tcp_sk(sk);
1942 struct sk_buff *skb;
1944 /* Reduce ssthresh if it has not yet been made inside this window. */
1945 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1946 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1947 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1948 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1949 tcp_ca_event(sk, CA_EVENT_LOSS);
1952 tp->snd_cwnd_cnt = 0;
1953 tp->snd_cwnd_stamp = tcp_time_stamp;
1955 tp->bytes_acked = 0;
1956 tcp_clear_retrans_partial(tp);
1958 if (tcp_is_reno(tp))
1959 tcp_reset_reno_sack(tp);
1962 /* Push undo marker, if it was plain RTO and nothing
1963 * was retransmitted. */
1964 tp->undo_marker = tp->snd_una;
1967 tp->fackets_out = 0;
1969 tcp_clear_all_retrans_hints(tp);
1971 tcp_for_write_queue(skb, sk) {
1972 if (skb == tcp_send_head(sk))
1975 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1976 tp->undo_marker = 0;
1977 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1978 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1979 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1980 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1981 tp->lost_out += tcp_skb_pcount(skb);
1982 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1985 tcp_verify_left_out(tp);
1987 tp->reordering = min_t(unsigned int, tp->reordering,
1988 sysctl_tcp_reordering);
1989 tcp_set_ca_state(sk, TCP_CA_Loss);
1990 tp->high_seq = tp->snd_nxt;
1991 TCP_ECN_queue_cwr(tp);
1992 /* Abort F-RTO algorithm if one is in progress */
1993 tp->frto_counter = 0;
1996 /* If ACK arrived pointing to a remembered SACK, it means that our
1997 * remembered SACKs do not reflect real state of receiver i.e.
1998 * receiver _host_ is heavily congested (or buggy).
2000 * Do processing similar to RTO timeout.
2002 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2004 if (flag & FLAG_SACK_RENEGING) {
2005 struct inet_connection_sock *icsk = inet_csk(sk);
2006 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2008 tcp_enter_loss(sk, 1);
2009 icsk->icsk_retransmits++;
2010 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2011 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2012 icsk->icsk_rto, TCP_RTO_MAX);
2018 static inline int tcp_fackets_out(struct tcp_sock *tp)
2020 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2023 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2024 * counter when SACK is enabled (without SACK, sacked_out is used for
2027 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2028 * segments up to the highest received SACK block so far and holes in
2031 * With reordering, holes may still be in flight, so RFC3517 recovery
2032 * uses pure sacked_out (total number of SACKed segments) even though
2033 * it violates the RFC that uses duplicate ACKs, often these are equal
2034 * but when e.g. out-of-window ACKs or packet duplication occurs,
2035 * they differ. Since neither occurs due to loss, TCP should really
2038 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2040 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2043 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2045 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2048 static inline int tcp_head_timedout(struct sock *sk)
2050 struct tcp_sock *tp = tcp_sk(sk);
2052 return tp->packets_out &&
2053 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2056 /* Linux NewReno/SACK/FACK/ECN state machine.
2057 * --------------------------------------
2059 * "Open" Normal state, no dubious events, fast path.
2060 * "Disorder" In all the respects it is "Open",
2061 * but requires a bit more attention. It is entered when
2062 * we see some SACKs or dupacks. It is split of "Open"
2063 * mainly to move some processing from fast path to slow one.
2064 * "CWR" CWND was reduced due to some Congestion Notification event.
2065 * It can be ECN, ICMP source quench, local device congestion.
2066 * "Recovery" CWND was reduced, we are fast-retransmitting.
2067 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2069 * tcp_fastretrans_alert() is entered:
2070 * - each incoming ACK, if state is not "Open"
2071 * - when arrived ACK is unusual, namely:
2076 * Counting packets in flight is pretty simple.
2078 * in_flight = packets_out - left_out + retrans_out
2080 * packets_out is SND.NXT-SND.UNA counted in packets.
2082 * retrans_out is number of retransmitted segments.
2084 * left_out is number of segments left network, but not ACKed yet.
2086 * left_out = sacked_out + lost_out
2088 * sacked_out: Packets, which arrived to receiver out of order
2089 * and hence not ACKed. With SACKs this number is simply
2090 * amount of SACKed data. Even without SACKs
2091 * it is easy to give pretty reliable estimate of this number,
2092 * counting duplicate ACKs.
2094 * lost_out: Packets lost by network. TCP has no explicit
2095 * "loss notification" feedback from network (for now).
2096 * It means that this number can be only _guessed_.
2097 * Actually, it is the heuristics to predict lossage that
2098 * distinguishes different algorithms.
2100 * F.e. after RTO, when all the queue is considered as lost,
2101 * lost_out = packets_out and in_flight = retrans_out.
2103 * Essentially, we have now two algorithms counting
2106 * FACK: It is the simplest heuristics. As soon as we decided
2107 * that something is lost, we decide that _all_ not SACKed
2108 * packets until the most forward SACK are lost. I.e.
2109 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2110 * It is absolutely correct estimate, if network does not reorder
2111 * packets. And it loses any connection to reality when reordering
2112 * takes place. We use FACK by default until reordering
2113 * is suspected on the path to this destination.
2115 * NewReno: when Recovery is entered, we assume that one segment
2116 * is lost (classic Reno). While we are in Recovery and
2117 * a partial ACK arrives, we assume that one more packet
2118 * is lost (NewReno). This heuristics are the same in NewReno
2121 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2122 * deflation etc. CWND is real congestion window, never inflated, changes
2123 * only according to classic VJ rules.
2125 * Really tricky (and requiring careful tuning) part of algorithm
2126 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2127 * The first determines the moment _when_ we should reduce CWND and,
2128 * hence, slow down forward transmission. In fact, it determines the moment
2129 * when we decide that hole is caused by loss, rather than by a reorder.
2131 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2132 * holes, caused by lost packets.
2134 * And the most logically complicated part of algorithm is undo
2135 * heuristics. We detect false retransmits due to both too early
2136 * fast retransmit (reordering) and underestimated RTO, analyzing
2137 * timestamps and D-SACKs. When we detect that some segments were
2138 * retransmitted by mistake and CWND reduction was wrong, we undo
2139 * window reduction and abort recovery phase. This logic is hidden
2140 * inside several functions named tcp_try_undo_<something>.
2143 /* This function decides, when we should leave Disordered state
2144 * and enter Recovery phase, reducing congestion window.
2146 * Main question: may we further continue forward transmission
2147 * with the same cwnd?
2149 static int tcp_time_to_recover(struct sock *sk)
2151 struct tcp_sock *tp = tcp_sk(sk);
2154 /* Do not perform any recovery during F-RTO algorithm */
2155 if (tp->frto_counter)
2158 /* Trick#1: The loss is proven. */
2162 /* Not-A-Trick#2 : Classic rule... */
2163 if (tcp_dupack_heurestics(tp) > tp->reordering)
2166 /* Trick#3 : when we use RFC2988 timer restart, fast
2167 * retransmit can be triggered by timeout of queue head.
2169 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2172 /* Trick#4: It is still not OK... But will it be useful to delay
2175 packets_out = tp->packets_out;
2176 if (packets_out <= tp->reordering &&
2177 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2178 !tcp_may_send_now(sk)) {
2179 /* We have nothing to send. This connection is limited
2180 * either by receiver window or by application.
2188 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2189 * is against sacked "cnt", otherwise it's against facked "cnt"
2191 static void tcp_mark_head_lost(struct sock *sk, int packets)
2193 struct tcp_sock *tp = tcp_sk(sk);
2194 struct sk_buff *skb;
2199 WARN_ON(packets > tp->packets_out);
2200 if (tp->lost_skb_hint) {
2201 skb = tp->lost_skb_hint;
2202 cnt = tp->lost_cnt_hint;
2204 skb = tcp_write_queue_head(sk);
2208 tcp_for_write_queue_from(skb, sk) {
2209 if (skb == tcp_send_head(sk))
2211 /* TODO: do this better */
2212 /* this is not the most efficient way to do this... */
2213 tp->lost_skb_hint = skb;
2214 tp->lost_cnt_hint = cnt;
2216 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2220 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2221 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2222 cnt += tcp_skb_pcount(skb);
2224 if (cnt > packets) {
2225 if (tcp_is_sack(tp) || (oldcnt >= packets))
2228 mss = skb_shinfo(skb)->gso_size;
2229 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2235 tcp_skb_mark_lost(tp, skb);
2237 tcp_verify_left_out(tp);
2240 /* Account newly detected lost packet(s) */
2242 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2244 struct tcp_sock *tp = tcp_sk(sk);
2246 if (tcp_is_reno(tp)) {
2247 tcp_mark_head_lost(sk, 1);
2248 } else if (tcp_is_fack(tp)) {
2249 int lost = tp->fackets_out - tp->reordering;
2252 tcp_mark_head_lost(sk, lost);
2254 int sacked_upto = tp->sacked_out - tp->reordering;
2255 if (sacked_upto < fast_rexmit)
2256 sacked_upto = fast_rexmit;
2257 tcp_mark_head_lost(sk, sacked_upto);
2260 /* New heuristics: it is possible only after we switched
2261 * to restart timer each time when something is ACKed.
2262 * Hence, we can detect timed out packets during fast
2263 * retransmit without falling to slow start.
2265 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2266 struct sk_buff *skb;
2268 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2269 : tcp_write_queue_head(sk);
2271 tcp_for_write_queue_from(skb, sk) {
2272 if (skb == tcp_send_head(sk))
2274 if (!tcp_skb_timedout(sk, skb))
2277 tcp_skb_mark_lost(tp, skb);
2280 tp->scoreboard_skb_hint = skb;
2282 tcp_verify_left_out(tp);
2286 /* CWND moderation, preventing bursts due to too big ACKs
2287 * in dubious situations.
2289 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2291 tp->snd_cwnd = min(tp->snd_cwnd,
2292 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2293 tp->snd_cwnd_stamp = tcp_time_stamp;
2296 /* Lower bound on congestion window is slow start threshold
2297 * unless congestion avoidance choice decides to overide it.
2299 static inline u32 tcp_cwnd_min(const struct sock *sk)
2301 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2303 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2306 /* Decrease cwnd each second ack. */
2307 static void tcp_cwnd_down(struct sock *sk, int flag)
2309 struct tcp_sock *tp = tcp_sk(sk);
2310 int decr = tp->snd_cwnd_cnt + 1;
2312 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2313 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2314 tp->snd_cwnd_cnt = decr & 1;
2317 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2318 tp->snd_cwnd -= decr;
2320 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2321 tp->snd_cwnd_stamp = tcp_time_stamp;
2325 /* Nothing was retransmitted or returned timestamp is less
2326 * than timestamp of the first retransmission.
2328 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2330 return !tp->retrans_stamp ||
2331 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2332 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2335 /* Undo procedures. */
2337 #if FASTRETRANS_DEBUG > 1
2338 static void DBGUNDO(struct sock *sk, const char *msg)
2340 struct tcp_sock *tp = tcp_sk(sk);
2341 struct inet_sock *inet = inet_sk(sk);
2343 if (sk->sk_family == AF_INET) {
2344 printk(KERN_DEBUG "Undo %s " NIPQUAD_FMT "/%u c%u l%u ss%u/%u p%u\n",
2346 NIPQUAD(inet->daddr), ntohs(inet->dport),
2347 tp->snd_cwnd, tcp_left_out(tp),
2348 tp->snd_ssthresh, tp->prior_ssthresh,
2351 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2352 else if (sk->sk_family == AF_INET6) {
2353 struct ipv6_pinfo *np = inet6_sk(sk);
2354 printk(KERN_DEBUG "Undo %s " NIP6_FMT "/%u c%u l%u ss%u/%u p%u\n",
2356 NIP6(np->daddr), ntohs(inet->dport),
2357 tp->snd_cwnd, tcp_left_out(tp),
2358 tp->snd_ssthresh, tp->prior_ssthresh,
2364 #define DBGUNDO(x...) do { } while (0)
2367 static void tcp_undo_cwr(struct sock *sk, const int undo)
2369 struct tcp_sock *tp = tcp_sk(sk);
2371 if (tp->prior_ssthresh) {
2372 const struct inet_connection_sock *icsk = inet_csk(sk);
2374 if (icsk->icsk_ca_ops->undo_cwnd)
2375 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2377 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2379 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2380 tp->snd_ssthresh = tp->prior_ssthresh;
2381 TCP_ECN_withdraw_cwr(tp);
2384 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2386 tcp_moderate_cwnd(tp);
2387 tp->snd_cwnd_stamp = tcp_time_stamp;
2390 static inline int tcp_may_undo(struct tcp_sock *tp)
2392 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2395 /* People celebrate: "We love our President!" */
2396 static int tcp_try_undo_recovery(struct sock *sk)
2398 struct tcp_sock *tp = tcp_sk(sk);
2400 if (tcp_may_undo(tp)) {
2403 /* Happy end! We did not retransmit anything
2404 * or our original transmission succeeded.
2406 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2407 tcp_undo_cwr(sk, 1);
2408 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2409 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2411 mib_idx = LINUX_MIB_TCPFULLUNDO;
2413 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2414 tp->undo_marker = 0;
2416 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2417 /* Hold old state until something *above* high_seq
2418 * is ACKed. For Reno it is MUST to prevent false
2419 * fast retransmits (RFC2582). SACK TCP is safe. */
2420 tcp_moderate_cwnd(tp);
2423 tcp_set_ca_state(sk, TCP_CA_Open);
2427 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2428 static void tcp_try_undo_dsack(struct sock *sk)
2430 struct tcp_sock *tp = tcp_sk(sk);
2432 if (tp->undo_marker && !tp->undo_retrans) {
2433 DBGUNDO(sk, "D-SACK");
2434 tcp_undo_cwr(sk, 1);
2435 tp->undo_marker = 0;
2436 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2440 /* Undo during fast recovery after partial ACK. */
2442 static int tcp_try_undo_partial(struct sock *sk, int acked)
2444 struct tcp_sock *tp = tcp_sk(sk);
2445 /* Partial ACK arrived. Force Hoe's retransmit. */
2446 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2448 if (tcp_may_undo(tp)) {
2449 /* Plain luck! Hole if filled with delayed
2450 * packet, rather than with a retransmit.
2452 if (tp->retrans_out == 0)
2453 tp->retrans_stamp = 0;
2455 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2458 tcp_undo_cwr(sk, 0);
2459 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2461 /* So... Do not make Hoe's retransmit yet.
2462 * If the first packet was delayed, the rest
2463 * ones are most probably delayed as well.
2470 /* Undo during loss recovery after partial ACK. */
2471 static int tcp_try_undo_loss(struct sock *sk)
2473 struct tcp_sock *tp = tcp_sk(sk);
2475 if (tcp_may_undo(tp)) {
2476 struct sk_buff *skb;
2477 tcp_for_write_queue(skb, sk) {
2478 if (skb == tcp_send_head(sk))
2480 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2483 tcp_clear_all_retrans_hints(tp);
2485 DBGUNDO(sk, "partial loss");
2487 tcp_undo_cwr(sk, 1);
2488 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2489 inet_csk(sk)->icsk_retransmits = 0;
2490 tp->undo_marker = 0;
2491 if (tcp_is_sack(tp))
2492 tcp_set_ca_state(sk, TCP_CA_Open);
2498 static inline void tcp_complete_cwr(struct sock *sk)
2500 struct tcp_sock *tp = tcp_sk(sk);
2501 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2502 tp->snd_cwnd_stamp = tcp_time_stamp;
2503 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2506 static void tcp_try_keep_open(struct sock *sk)
2508 struct tcp_sock *tp = tcp_sk(sk);
2509 int state = TCP_CA_Open;
2511 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2512 state = TCP_CA_Disorder;
2514 if (inet_csk(sk)->icsk_ca_state != state) {
2515 tcp_set_ca_state(sk, state);
2516 tp->high_seq = tp->snd_nxt;
2520 static void tcp_try_to_open(struct sock *sk, int flag)
2522 struct tcp_sock *tp = tcp_sk(sk);
2524 tcp_verify_left_out(tp);
2526 if (!tp->frto_counter && tp->retrans_out == 0)
2527 tp->retrans_stamp = 0;
2529 if (flag & FLAG_ECE)
2530 tcp_enter_cwr(sk, 1);
2532 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2533 tcp_try_keep_open(sk);
2534 tcp_moderate_cwnd(tp);
2536 tcp_cwnd_down(sk, flag);
2540 static void tcp_mtup_probe_failed(struct sock *sk)
2542 struct inet_connection_sock *icsk = inet_csk(sk);
2544 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2545 icsk->icsk_mtup.probe_size = 0;
2548 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2550 struct tcp_sock *tp = tcp_sk(sk);
2551 struct inet_connection_sock *icsk = inet_csk(sk);
2553 /* FIXME: breaks with very large cwnd */
2554 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2555 tp->snd_cwnd = tp->snd_cwnd *
2556 tcp_mss_to_mtu(sk, tp->mss_cache) /
2557 icsk->icsk_mtup.probe_size;
2558 tp->snd_cwnd_cnt = 0;
2559 tp->snd_cwnd_stamp = tcp_time_stamp;
2560 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2562 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2563 icsk->icsk_mtup.probe_size = 0;
2564 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2567 /* Process an event, which can update packets-in-flight not trivially.
2568 * Main goal of this function is to calculate new estimate for left_out,
2569 * taking into account both packets sitting in receiver's buffer and
2570 * packets lost by network.
2572 * Besides that it does CWND reduction, when packet loss is detected
2573 * and changes state of machine.
2575 * It does _not_ decide what to send, it is made in function
2576 * tcp_xmit_retransmit_queue().
2578 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2580 struct inet_connection_sock *icsk = inet_csk(sk);
2581 struct tcp_sock *tp = tcp_sk(sk);
2582 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2583 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2584 (tcp_fackets_out(tp) > tp->reordering));
2585 int fast_rexmit = 0, mib_idx;
2587 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2589 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2590 tp->fackets_out = 0;
2592 /* Now state machine starts.
2593 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2594 if (flag & FLAG_ECE)
2595 tp->prior_ssthresh = 0;
2597 /* B. In all the states check for reneging SACKs. */
2598 if (tcp_check_sack_reneging(sk, flag))
2601 /* C. Process data loss notification, provided it is valid. */
2602 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2603 before(tp->snd_una, tp->high_seq) &&
2604 icsk->icsk_ca_state != TCP_CA_Open &&
2605 tp->fackets_out > tp->reordering) {
2606 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2607 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2610 /* D. Check consistency of the current state. */
2611 tcp_verify_left_out(tp);
2613 /* E. Check state exit conditions. State can be terminated
2614 * when high_seq is ACKed. */
2615 if (icsk->icsk_ca_state == TCP_CA_Open) {
2616 WARN_ON(tp->retrans_out != 0);
2617 tp->retrans_stamp = 0;
2618 } else if (!before(tp->snd_una, tp->high_seq)) {
2619 switch (icsk->icsk_ca_state) {
2621 icsk->icsk_retransmits = 0;
2622 if (tcp_try_undo_recovery(sk))
2627 /* CWR is to be held something *above* high_seq
2628 * is ACKed for CWR bit to reach receiver. */
2629 if (tp->snd_una != tp->high_seq) {
2630 tcp_complete_cwr(sk);
2631 tcp_set_ca_state(sk, TCP_CA_Open);
2635 case TCP_CA_Disorder:
2636 tcp_try_undo_dsack(sk);
2637 if (!tp->undo_marker ||
2638 /* For SACK case do not Open to allow to undo
2639 * catching for all duplicate ACKs. */
2640 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2641 tp->undo_marker = 0;
2642 tcp_set_ca_state(sk, TCP_CA_Open);
2646 case TCP_CA_Recovery:
2647 if (tcp_is_reno(tp))
2648 tcp_reset_reno_sack(tp);
2649 if (tcp_try_undo_recovery(sk))
2651 tcp_complete_cwr(sk);
2656 /* F. Process state. */
2657 switch (icsk->icsk_ca_state) {
2658 case TCP_CA_Recovery:
2659 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2660 if (tcp_is_reno(tp) && is_dupack)
2661 tcp_add_reno_sack(sk);
2663 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2666 if (flag & FLAG_DATA_ACKED)
2667 icsk->icsk_retransmits = 0;
2668 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2669 tcp_reset_reno_sack(tp);
2670 if (!tcp_try_undo_loss(sk)) {
2671 tcp_moderate_cwnd(tp);
2672 tcp_xmit_retransmit_queue(sk);
2675 if (icsk->icsk_ca_state != TCP_CA_Open)
2677 /* Loss is undone; fall through to processing in Open state. */
2679 if (tcp_is_reno(tp)) {
2680 if (flag & FLAG_SND_UNA_ADVANCED)
2681 tcp_reset_reno_sack(tp);
2683 tcp_add_reno_sack(sk);
2686 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2687 tcp_try_undo_dsack(sk);
2689 if (!tcp_time_to_recover(sk)) {
2690 tcp_try_to_open(sk, flag);
2694 /* MTU probe failure: don't reduce cwnd */
2695 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2696 icsk->icsk_mtup.probe_size &&
2697 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2698 tcp_mtup_probe_failed(sk);
2699 /* Restores the reduction we did in tcp_mtup_probe() */
2701 tcp_simple_retransmit(sk);
2705 /* Otherwise enter Recovery state */
2707 if (tcp_is_reno(tp))
2708 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2710 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2712 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2714 tp->high_seq = tp->snd_nxt;
2715 tp->prior_ssthresh = 0;
2716 tp->undo_marker = tp->snd_una;
2717 tp->undo_retrans = tp->retrans_out;
2719 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2720 if (!(flag & FLAG_ECE))
2721 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2722 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2723 TCP_ECN_queue_cwr(tp);
2726 tp->bytes_acked = 0;
2727 tp->snd_cwnd_cnt = 0;
2728 tcp_set_ca_state(sk, TCP_CA_Recovery);
2732 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2733 tcp_update_scoreboard(sk, fast_rexmit);
2734 tcp_cwnd_down(sk, flag);
2735 tcp_xmit_retransmit_queue(sk);
2738 /* Read draft-ietf-tcplw-high-performance before mucking
2739 * with this code. (Supersedes RFC1323)
2741 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2743 /* RTTM Rule: A TSecr value received in a segment is used to
2744 * update the averaged RTT measurement only if the segment
2745 * acknowledges some new data, i.e., only if it advances the
2746 * left edge of the send window.
2748 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2749 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2751 * Changed: reset backoff as soon as we see the first valid sample.
2752 * If we do not, we get strongly overestimated rto. With timestamps
2753 * samples are accepted even from very old segments: f.e., when rtt=1
2754 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2755 * answer arrives rto becomes 120 seconds! If at least one of segments
2756 * in window is lost... Voila. --ANK (010210)
2758 struct tcp_sock *tp = tcp_sk(sk);
2759 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2760 tcp_rtt_estimator(sk, seq_rtt);
2762 inet_csk(sk)->icsk_backoff = 0;
2766 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2768 /* We don't have a timestamp. Can only use
2769 * packets that are not retransmitted to determine
2770 * rtt estimates. Also, we must not reset the
2771 * backoff for rto until we get a non-retransmitted
2772 * packet. This allows us to deal with a situation
2773 * where the network delay has increased suddenly.
2774 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2777 if (flag & FLAG_RETRANS_DATA_ACKED)
2780 tcp_rtt_estimator(sk, seq_rtt);
2782 inet_csk(sk)->icsk_backoff = 0;
2786 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2789 const struct tcp_sock *tp = tcp_sk(sk);
2790 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2791 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2792 tcp_ack_saw_tstamp(sk, flag);
2793 else if (seq_rtt >= 0)
2794 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2797 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
2799 const struct inet_connection_sock *icsk = inet_csk(sk);
2800 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
2801 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2804 /* Restart timer after forward progress on connection.
2805 * RFC2988 recommends to restart timer to now+rto.
2807 static void tcp_rearm_rto(struct sock *sk)
2809 struct tcp_sock *tp = tcp_sk(sk);
2811 if (!tp->packets_out) {
2812 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2814 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2815 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2819 /* If we get here, the whole TSO packet has not been acked. */
2820 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2822 struct tcp_sock *tp = tcp_sk(sk);
2825 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2827 packets_acked = tcp_skb_pcount(skb);
2828 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2830 packets_acked -= tcp_skb_pcount(skb);
2832 if (packets_acked) {
2833 BUG_ON(tcp_skb_pcount(skb) == 0);
2834 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2837 return packets_acked;
2840 /* Remove acknowledged frames from the retransmission queue. If our packet
2841 * is before the ack sequence we can discard it as it's confirmed to have
2842 * arrived at the other end.
2844 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
2846 struct tcp_sock *tp = tcp_sk(sk);
2847 const struct inet_connection_sock *icsk = inet_csk(sk);
2848 struct sk_buff *skb;
2849 u32 now = tcp_time_stamp;
2850 int fully_acked = 1;
2853 u32 reord = tp->packets_out;
2855 s32 ca_seq_rtt = -1;
2856 ktime_t last_ackt = net_invalid_timestamp();
2858 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
2859 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2862 u8 sacked = scb->sacked;
2864 /* Determine how many packets and what bytes were acked, tso and else */
2865 if (after(scb->end_seq, tp->snd_una)) {
2866 if (tcp_skb_pcount(skb) == 1 ||
2867 !after(tp->snd_una, scb->seq))
2870 acked_pcount = tcp_tso_acked(sk, skb);
2875 end_seq = tp->snd_una;
2877 acked_pcount = tcp_skb_pcount(skb);
2878 end_seq = scb->end_seq;
2881 /* MTU probing checks */
2882 if (fully_acked && icsk->icsk_mtup.probe_size &&
2883 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2884 tcp_mtup_probe_success(sk, skb);
2887 if (sacked & TCPCB_RETRANS) {
2888 if (sacked & TCPCB_SACKED_RETRANS)
2889 tp->retrans_out -= acked_pcount;
2890 flag |= FLAG_RETRANS_DATA_ACKED;
2893 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
2894 flag |= FLAG_NONHEAD_RETRANS_ACKED;
2896 ca_seq_rtt = now - scb->when;
2897 last_ackt = skb->tstamp;
2899 seq_rtt = ca_seq_rtt;
2901 if (!(sacked & TCPCB_SACKED_ACKED))
2902 reord = min(pkts_acked, reord);
2905 if (sacked & TCPCB_SACKED_ACKED)
2906 tp->sacked_out -= acked_pcount;
2907 if (sacked & TCPCB_LOST)
2908 tp->lost_out -= acked_pcount;
2910 if (unlikely(tp->urg_mode && !before(end_seq, tp->snd_up)))
2913 tp->packets_out -= acked_pcount;
2914 pkts_acked += acked_pcount;
2916 /* Initial outgoing SYN's get put onto the write_queue
2917 * just like anything else we transmit. It is not
2918 * true data, and if we misinform our callers that
2919 * this ACK acks real data, we will erroneously exit
2920 * connection startup slow start one packet too
2921 * quickly. This is severely frowned upon behavior.
2923 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2924 flag |= FLAG_DATA_ACKED;
2926 flag |= FLAG_SYN_ACKED;
2927 tp->retrans_stamp = 0;
2933 tcp_unlink_write_queue(skb, sk);
2934 sk_wmem_free_skb(sk, skb);
2935 tcp_clear_all_retrans_hints(tp);
2938 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2939 flag |= FLAG_SACK_RENEGING;
2941 if (flag & FLAG_ACKED) {
2942 const struct tcp_congestion_ops *ca_ops
2943 = inet_csk(sk)->icsk_ca_ops;
2945 tcp_ack_update_rtt(sk, flag, seq_rtt);
2948 if (tcp_is_reno(tp)) {
2949 tcp_remove_reno_sacks(sk, pkts_acked);
2951 /* Non-retransmitted hole got filled? That's reordering */
2952 if (reord < prior_fackets)
2953 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2956 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
2958 if (ca_ops->pkts_acked) {
2961 /* Is the ACK triggering packet unambiguous? */
2962 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
2963 /* High resolution needed and available? */
2964 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2965 !ktime_equal(last_ackt,
2966 net_invalid_timestamp()))
2967 rtt_us = ktime_us_delta(ktime_get_real(),
2969 else if (ca_seq_rtt > 0)
2970 rtt_us = jiffies_to_usecs(ca_seq_rtt);
2973 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2977 #if FASTRETRANS_DEBUG > 0
2978 WARN_ON((int)tp->sacked_out < 0);
2979 WARN_ON((int)tp->lost_out < 0);
2980 WARN_ON((int)tp->retrans_out < 0);
2981 if (!tp->packets_out && tcp_is_sack(tp)) {
2982 icsk = inet_csk(sk);
2984 printk(KERN_DEBUG "Leak l=%u %d\n",
2985 tp->lost_out, icsk->icsk_ca_state);
2988 if (tp->sacked_out) {
2989 printk(KERN_DEBUG "Leak s=%u %d\n",
2990 tp->sacked_out, icsk->icsk_ca_state);
2993 if (tp->retrans_out) {
2994 printk(KERN_DEBUG "Leak r=%u %d\n",
2995 tp->retrans_out, icsk->icsk_ca_state);
2996 tp->retrans_out = 0;
3003 static void tcp_ack_probe(struct sock *sk)
3005 const struct tcp_sock *tp = tcp_sk(sk);
3006 struct inet_connection_sock *icsk = inet_csk(sk);
3008 /* Was it a usable window open? */
3010 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3011 icsk->icsk_backoff = 0;
3012 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3013 /* Socket must be waked up by subsequent tcp_data_snd_check().
3014 * This function is not for random using!
3017 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3018 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3023 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3025 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3026 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3029 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3031 const struct tcp_sock *tp = tcp_sk(sk);
3032 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3033 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3036 /* Check that window update is acceptable.
3037 * The function assumes that snd_una<=ack<=snd_next.
3039 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3040 const u32 ack, const u32 ack_seq,
3043 return (after(ack, tp->snd_una) ||
3044 after(ack_seq, tp->snd_wl1) ||
3045 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3048 /* Update our send window.
3050 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3051 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3053 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3056 struct tcp_sock *tp = tcp_sk(sk);
3058 u32 nwin = ntohs(tcp_hdr(skb)->window);
3060 if (likely(!tcp_hdr(skb)->syn))
3061 nwin <<= tp->rx_opt.snd_wscale;
3063 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3064 flag |= FLAG_WIN_UPDATE;
3065 tcp_update_wl(tp, ack, ack_seq);
3067 if (tp->snd_wnd != nwin) {
3070 /* Note, it is the only place, where
3071 * fast path is recovered for sending TCP.
3074 tcp_fast_path_check(sk);
3076 if (nwin > tp->max_window) {
3077 tp->max_window = nwin;
3078 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3088 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3089 * continue in congestion avoidance.
3091 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3093 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3094 tp->snd_cwnd_cnt = 0;
3095 tp->bytes_acked = 0;
3096 TCP_ECN_queue_cwr(tp);
3097 tcp_moderate_cwnd(tp);
3100 /* A conservative spurious RTO response algorithm: reduce cwnd using
3101 * rate halving and continue in congestion avoidance.
3103 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3105 tcp_enter_cwr(sk, 0);
3108 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3110 if (flag & FLAG_ECE)
3111 tcp_ratehalving_spur_to_response(sk);
3113 tcp_undo_cwr(sk, 1);
3116 /* F-RTO spurious RTO detection algorithm (RFC4138)
3118 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3119 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3120 * window (but not to or beyond highest sequence sent before RTO):
3121 * On First ACK, send two new segments out.
3122 * On Second ACK, RTO was likely spurious. Do spurious response (response
3123 * algorithm is not part of the F-RTO detection algorithm
3124 * given in RFC4138 but can be selected separately).
3125 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3126 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3127 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3128 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3130 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3131 * original window even after we transmit two new data segments.
3134 * on first step, wait until first cumulative ACK arrives, then move to
3135 * the second step. In second step, the next ACK decides.
3137 * F-RTO is implemented (mainly) in four functions:
3138 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3139 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3140 * called when tcp_use_frto() showed green light
3141 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3142 * - tcp_enter_frto_loss() is called if there is not enough evidence
3143 * to prove that the RTO is indeed spurious. It transfers the control
3144 * from F-RTO to the conventional RTO recovery
3146 static int tcp_process_frto(struct sock *sk, int flag)
3148 struct tcp_sock *tp = tcp_sk(sk);
3150 tcp_verify_left_out(tp);
3152 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3153 if (flag & FLAG_DATA_ACKED)
3154 inet_csk(sk)->icsk_retransmits = 0;
3156 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3157 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3158 tp->undo_marker = 0;
3160 if (!before(tp->snd_una, tp->frto_highmark)) {
3161 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3165 if (!tcp_is_sackfrto(tp)) {
3166 /* RFC4138 shortcoming in step 2; should also have case c):
3167 * ACK isn't duplicate nor advances window, e.g., opposite dir
3170 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3173 if (!(flag & FLAG_DATA_ACKED)) {
3174 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3179 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3180 /* Prevent sending of new data. */
3181 tp->snd_cwnd = min(tp->snd_cwnd,
3182 tcp_packets_in_flight(tp));
3186 if ((tp->frto_counter >= 2) &&
3187 (!(flag & FLAG_FORWARD_PROGRESS) ||
3188 ((flag & FLAG_DATA_SACKED) &&
3189 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3190 /* RFC4138 shortcoming (see comment above) */
3191 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3192 (flag & FLAG_NOT_DUP))
3195 tcp_enter_frto_loss(sk, 3, flag);
3200 if (tp->frto_counter == 1) {
3201 /* tcp_may_send_now needs to see updated state */
3202 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3203 tp->frto_counter = 2;
3205 if (!tcp_may_send_now(sk))
3206 tcp_enter_frto_loss(sk, 2, flag);
3210 switch (sysctl_tcp_frto_response) {
3212 tcp_undo_spur_to_response(sk, flag);
3215 tcp_conservative_spur_to_response(tp);
3218 tcp_ratehalving_spur_to_response(sk);
3221 tp->frto_counter = 0;
3222 tp->undo_marker = 0;
3223 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3228 /* This routine deals with incoming acks, but not outgoing ones. */
3229 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3231 struct inet_connection_sock *icsk = inet_csk(sk);
3232 struct tcp_sock *tp = tcp_sk(sk);
3233 u32 prior_snd_una = tp->snd_una;
3234 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3235 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3236 u32 prior_in_flight;
3241 /* If the ack is newer than sent or older than previous acks
3242 * then we can probably ignore it.
3244 if (after(ack, tp->snd_nxt))
3245 goto uninteresting_ack;
3247 if (before(ack, prior_snd_una))
3250 if (after(ack, prior_snd_una))
3251 flag |= FLAG_SND_UNA_ADVANCED;
3253 if (sysctl_tcp_abc) {
3254 if (icsk->icsk_ca_state < TCP_CA_CWR)
3255 tp->bytes_acked += ack - prior_snd_una;
3256 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3257 /* we assume just one segment left network */
3258 tp->bytes_acked += min(ack - prior_snd_una,
3262 prior_fackets = tp->fackets_out;
3263 prior_in_flight = tcp_packets_in_flight(tp);
3265 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3266 /* Window is constant, pure forward advance.
3267 * No more checks are required.
3268 * Note, we use the fact that SND.UNA>=SND.WL2.
3270 tcp_update_wl(tp, ack, ack_seq);
3272 flag |= FLAG_WIN_UPDATE;
3274 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3276 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3278 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3281 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3283 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3285 if (TCP_SKB_CB(skb)->sacked)
3286 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3288 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3291 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3294 /* We passed data and got it acked, remove any soft error
3295 * log. Something worked...
3297 sk->sk_err_soft = 0;
3298 icsk->icsk_probes_out = 0;
3299 tp->rcv_tstamp = tcp_time_stamp;
3300 prior_packets = tp->packets_out;
3304 /* See if we can take anything off of the retransmit queue. */
3305 flag |= tcp_clean_rtx_queue(sk, prior_fackets);
3307 if (tp->frto_counter)
3308 frto_cwnd = tcp_process_frto(sk, flag);
3309 /* Guarantee sacktag reordering detection against wrap-arounds */
3310 if (before(tp->frto_highmark, tp->snd_una))
3311 tp->frto_highmark = 0;
3313 if (tcp_ack_is_dubious(sk, flag)) {
3314 /* Advance CWND, if state allows this. */
3315 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3316 tcp_may_raise_cwnd(sk, flag))
3317 tcp_cong_avoid(sk, ack, prior_in_flight);
3318 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3321 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3322 tcp_cong_avoid(sk, ack, prior_in_flight);
3325 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3326 dst_confirm(sk->sk_dst_cache);
3331 /* If this ack opens up a zero window, clear backoff. It was
3332 * being used to time the probes, and is probably far higher than
3333 * it needs to be for normal retransmission.
3335 if (tcp_send_head(sk))
3340 if (TCP_SKB_CB(skb)->sacked) {
3341 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3342 if (icsk->icsk_ca_state == TCP_CA_Open)
3343 tcp_try_keep_open(sk);
3347 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3351 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3352 * But, this can also be called on packets in the established flow when
3353 * the fast version below fails.
3355 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3359 struct tcphdr *th = tcp_hdr(skb);
3360 int length = (th->doff * 4) - sizeof(struct tcphdr);
3362 ptr = (unsigned char *)(th + 1);
3363 opt_rx->saw_tstamp = 0;
3365 while (length > 0) {
3366 int opcode = *ptr++;
3372 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3377 if (opsize < 2) /* "silly options" */
3379 if (opsize > length)
3380 return; /* don't parse partial options */
3383 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3384 u16 in_mss = get_unaligned_be16(ptr);
3386 if (opt_rx->user_mss &&
3387 opt_rx->user_mss < in_mss)
3388 in_mss = opt_rx->user_mss;
3389 opt_rx->mss_clamp = in_mss;
3394 if (opsize == TCPOLEN_WINDOW && th->syn &&
3395 !estab && sysctl_tcp_window_scaling) {
3396 __u8 snd_wscale = *(__u8 *)ptr;
3397 opt_rx->wscale_ok = 1;
3398 if (snd_wscale > 14) {
3399 if (net_ratelimit())
3400 printk(KERN_INFO "tcp_parse_options: Illegal window "
3401 "scaling value %d >14 received.\n",
3405 opt_rx->snd_wscale = snd_wscale;
3408 case TCPOPT_TIMESTAMP:
3409 if ((opsize == TCPOLEN_TIMESTAMP) &&
3410 ((estab && opt_rx->tstamp_ok) ||
3411 (!estab && sysctl_tcp_timestamps))) {
3412 opt_rx->saw_tstamp = 1;
3413 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3414 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3417 case TCPOPT_SACK_PERM:
3418 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3419 !estab && sysctl_tcp_sack) {
3420 opt_rx->sack_ok = 1;
3421 tcp_sack_reset(opt_rx);
3426 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3427 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3429 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3432 #ifdef CONFIG_TCP_MD5SIG
3435 * The MD5 Hash has already been
3436 * checked (see tcp_v{4,6}_do_rcv()).
3448 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3450 __be32 *ptr = (__be32 *)(th + 1);
3452 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3453 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3454 tp->rx_opt.saw_tstamp = 1;
3456 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3458 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3464 /* Fast parse options. This hopes to only see timestamps.
3465 * If it is wrong it falls back on tcp_parse_options().
3467 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3468 struct tcp_sock *tp)
3470 if (th->doff == sizeof(struct tcphdr) >> 2) {
3471 tp->rx_opt.saw_tstamp = 0;
3473 } else if (tp->rx_opt.tstamp_ok &&
3474 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3475 if (tcp_parse_aligned_timestamp(tp, th))
3478 tcp_parse_options(skb, &tp->rx_opt, 1);
3482 #ifdef CONFIG_TCP_MD5SIG
3484 * Parse MD5 Signature option
3486 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3488 int length = (th->doff << 2) - sizeof (*th);
3489 u8 *ptr = (u8*)(th + 1);
3491 /* If the TCP option is too short, we can short cut */
3492 if (length < TCPOLEN_MD5SIG)
3495 while (length > 0) {
3496 int opcode = *ptr++;
3507 if (opsize < 2 || opsize > length)
3509 if (opcode == TCPOPT_MD5SIG)
3519 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3521 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3522 tp->rx_opt.ts_recent_stamp = get_seconds();
3525 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3527 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3528 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3529 * extra check below makes sure this can only happen
3530 * for pure ACK frames. -DaveM
3532 * Not only, also it occurs for expired timestamps.
3535 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3536 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3537 tcp_store_ts_recent(tp);
3541 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3543 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3544 * it can pass through stack. So, the following predicate verifies that
3545 * this segment is not used for anything but congestion avoidance or
3546 * fast retransmit. Moreover, we even are able to eliminate most of such
3547 * second order effects, if we apply some small "replay" window (~RTO)
3548 * to timestamp space.
3550 * All these measures still do not guarantee that we reject wrapped ACKs
3551 * on networks with high bandwidth, when sequence space is recycled fastly,
3552 * but it guarantees that such events will be very rare and do not affect
3553 * connection seriously. This doesn't look nice, but alas, PAWS is really
3556 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3557 * states that events when retransmit arrives after original data are rare.
3558 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3559 * the biggest problem on large power networks even with minor reordering.
3560 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3561 * up to bandwidth of 18Gigabit/sec. 8) ]
3564 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3566 struct tcp_sock *tp = tcp_sk(sk);
3567 struct tcphdr *th = tcp_hdr(skb);
3568 u32 seq = TCP_SKB_CB(skb)->seq;
3569 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3571 return (/* 1. Pure ACK with correct sequence number. */
3572 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3574 /* 2. ... and duplicate ACK. */
3575 ack == tp->snd_una &&
3577 /* 3. ... and does not update window. */
3578 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3580 /* 4. ... and sits in replay window. */
3581 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3584 static inline int tcp_paws_discard(const struct sock *sk,
3585 const struct sk_buff *skb)
3587 const struct tcp_sock *tp = tcp_sk(sk);
3588 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3589 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3590 !tcp_disordered_ack(sk, skb));
3593 /* Check segment sequence number for validity.
3595 * Segment controls are considered valid, if the segment
3596 * fits to the window after truncation to the window. Acceptability
3597 * of data (and SYN, FIN, of course) is checked separately.
3598 * See tcp_data_queue(), for example.
3600 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3601 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3602 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3603 * (borrowed from freebsd)
3606 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3608 return !before(end_seq, tp->rcv_wup) &&
3609 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3612 /* When we get a reset we do this. */
3613 static void tcp_reset(struct sock *sk)
3615 /* We want the right error as BSD sees it (and indeed as we do). */
3616 switch (sk->sk_state) {
3618 sk->sk_err = ECONNREFUSED;
3620 case TCP_CLOSE_WAIT:
3626 sk->sk_err = ECONNRESET;
3629 if (!sock_flag(sk, SOCK_DEAD))
3630 sk->sk_error_report(sk);
3636 * Process the FIN bit. This now behaves as it is supposed to work
3637 * and the FIN takes effect when it is validly part of sequence
3638 * space. Not before when we get holes.
3640 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3641 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3644 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3645 * close and we go into CLOSING (and later onto TIME-WAIT)
3647 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3649 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3651 struct tcp_sock *tp = tcp_sk(sk);
3653 inet_csk_schedule_ack(sk);
3655 sk->sk_shutdown |= RCV_SHUTDOWN;
3656 sock_set_flag(sk, SOCK_DONE);
3658 switch (sk->sk_state) {
3660 case TCP_ESTABLISHED:
3661 /* Move to CLOSE_WAIT */
3662 tcp_set_state(sk, TCP_CLOSE_WAIT);
3663 inet_csk(sk)->icsk_ack.pingpong = 1;
3666 case TCP_CLOSE_WAIT:
3668 /* Received a retransmission of the FIN, do
3673 /* RFC793: Remain in the LAST-ACK state. */
3677 /* This case occurs when a simultaneous close
3678 * happens, we must ack the received FIN and
3679 * enter the CLOSING state.
3682 tcp_set_state(sk, TCP_CLOSING);
3685 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3687 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3690 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3691 * cases we should never reach this piece of code.
3693 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3694 __func__, sk->sk_state);
3698 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3699 * Probably, we should reset in this case. For now drop them.
3701 __skb_queue_purge(&tp->out_of_order_queue);
3702 if (tcp_is_sack(tp))
3703 tcp_sack_reset(&tp->rx_opt);
3706 if (!sock_flag(sk, SOCK_DEAD)) {
3707 sk->sk_state_change(sk);
3709 /* Do not send POLL_HUP for half duplex close. */
3710 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3711 sk->sk_state == TCP_CLOSE)
3712 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3714 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3718 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3721 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3722 if (before(seq, sp->start_seq))
3723 sp->start_seq = seq;
3724 if (after(end_seq, sp->end_seq))
3725 sp->end_seq = end_seq;
3731 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3733 struct tcp_sock *tp = tcp_sk(sk);
3735 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3738 if (before(seq, tp->rcv_nxt))
3739 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3741 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3743 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3745 tp->rx_opt.dsack = 1;
3746 tp->duplicate_sack[0].start_seq = seq;
3747 tp->duplicate_sack[0].end_seq = end_seq;
3748 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 1;
3752 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3754 struct tcp_sock *tp = tcp_sk(sk);
3756 if (!tp->rx_opt.dsack)
3757 tcp_dsack_set(sk, seq, end_seq);
3759 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3762 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3764 struct tcp_sock *tp = tcp_sk(sk);
3766 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3767 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3768 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3769 tcp_enter_quickack_mode(sk);
3771 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3772 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3774 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3775 end_seq = tp->rcv_nxt;
3776 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3783 /* These routines update the SACK block as out-of-order packets arrive or
3784 * in-order packets close up the sequence space.
3786 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3789 struct tcp_sack_block *sp = &tp->selective_acks[0];
3790 struct tcp_sack_block *swalk = sp + 1;
3792 /* See if the recent change to the first SACK eats into
3793 * or hits the sequence space of other SACK blocks, if so coalesce.
3795 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3796 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3799 /* Zap SWALK, by moving every further SACK up by one slot.
3800 * Decrease num_sacks.
3802 tp->rx_opt.num_sacks--;
3803 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
3805 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3809 this_sack++, swalk++;
3813 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
3814 struct tcp_sack_block *sack2)
3818 tmp = sack1->start_seq;
3819 sack1->start_seq = sack2->start_seq;
3820 sack2->start_seq = tmp;
3822 tmp = sack1->end_seq;
3823 sack1->end_seq = sack2->end_seq;
3824 sack2->end_seq = tmp;
3827 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3829 struct tcp_sock *tp = tcp_sk(sk);
3830 struct tcp_sack_block *sp = &tp->selective_acks[0];
3831 int cur_sacks = tp->rx_opt.num_sacks;
3837 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3838 if (tcp_sack_extend(sp, seq, end_seq)) {
3839 /* Rotate this_sack to the first one. */
3840 for (; this_sack > 0; this_sack--, sp--)
3841 tcp_sack_swap(sp, sp - 1);
3843 tcp_sack_maybe_coalesce(tp);
3848 /* Could not find an adjacent existing SACK, build a new one,
3849 * put it at the front, and shift everyone else down. We
3850 * always know there is at least one SACK present already here.
3852 * If the sack array is full, forget about the last one.
3854 if (this_sack >= TCP_NUM_SACKS) {
3856 tp->rx_opt.num_sacks--;
3859 for (; this_sack > 0; this_sack--, sp--)
3863 /* Build the new head SACK, and we're done. */
3864 sp->start_seq = seq;
3865 sp->end_seq = end_seq;
3866 tp->rx_opt.num_sacks++;
3867 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
3870 /* RCV.NXT advances, some SACKs should be eaten. */
3872 static void tcp_sack_remove(struct tcp_sock *tp)
3874 struct tcp_sack_block *sp = &tp->selective_acks[0];
3875 int num_sacks = tp->rx_opt.num_sacks;
3878 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3879 if (skb_queue_empty(&tp->out_of_order_queue)) {
3880 tp->rx_opt.num_sacks = 0;
3881 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3885 for (this_sack = 0; this_sack < num_sacks;) {
3886 /* Check if the start of the sack is covered by RCV.NXT. */
3887 if (!before(tp->rcv_nxt, sp->start_seq)) {
3890 /* RCV.NXT must cover all the block! */
3891 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
3893 /* Zap this SACK, by moving forward any other SACKS. */
3894 for (i=this_sack+1; i < num_sacks; i++)
3895 tp->selective_acks[i-1] = tp->selective_acks[i];
3902 if (num_sacks != tp->rx_opt.num_sacks) {
3903 tp->rx_opt.num_sacks = num_sacks;
3904 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
3909 /* This one checks to see if we can put data from the
3910 * out_of_order queue into the receive_queue.
3912 static void tcp_ofo_queue(struct sock *sk)
3914 struct tcp_sock *tp = tcp_sk(sk);
3915 __u32 dsack_high = tp->rcv_nxt;
3916 struct sk_buff *skb;
3918 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3919 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3922 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3923 __u32 dsack = dsack_high;
3924 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3925 dsack_high = TCP_SKB_CB(skb)->end_seq;
3926 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
3929 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3930 SOCK_DEBUG(sk, "ofo packet was already received \n");
3931 __skb_unlink(skb, &tp->out_of_order_queue);
3935 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3936 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3937 TCP_SKB_CB(skb)->end_seq);
3939 __skb_unlink(skb, &tp->out_of_order_queue);
3940 __skb_queue_tail(&sk->sk_receive_queue, skb);
3941 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3942 if (tcp_hdr(skb)->fin)
3943 tcp_fin(skb, sk, tcp_hdr(skb));
3947 static int tcp_prune_ofo_queue(struct sock *sk);
3948 static int tcp_prune_queue(struct sock *sk);
3950 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
3952 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3953 !sk_rmem_schedule(sk, size)) {
3955 if (tcp_prune_queue(sk) < 0)
3958 if (!sk_rmem_schedule(sk, size)) {
3959 if (!tcp_prune_ofo_queue(sk))
3962 if (!sk_rmem_schedule(sk, size))
3969 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3971 struct tcphdr *th = tcp_hdr(skb);
3972 struct tcp_sock *tp = tcp_sk(sk);
3975 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3978 __skb_pull(skb, th->doff * 4);
3980 TCP_ECN_accept_cwr(tp, skb);
3982 if (tp->rx_opt.dsack) {
3983 tp->rx_opt.dsack = 0;
3984 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
3987 /* Queue data for delivery to the user.
3988 * Packets in sequence go to the receive queue.
3989 * Out of sequence packets to the out_of_order_queue.
3991 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3992 if (tcp_receive_window(tp) == 0)
3995 /* Ok. In sequence. In window. */
3996 if (tp->ucopy.task == current &&
3997 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3998 sock_owned_by_user(sk) && !tp->urg_data) {
3999 int chunk = min_t(unsigned int, skb->len,
4002 __set_current_state(TASK_RUNNING);
4005 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4006 tp->ucopy.len -= chunk;
4007 tp->copied_seq += chunk;
4008 eaten = (chunk == skb->len && !th->fin);
4009 tcp_rcv_space_adjust(sk);
4017 tcp_try_rmem_schedule(sk, skb->truesize))
4020 skb_set_owner_r(skb, sk);
4021 __skb_queue_tail(&sk->sk_receive_queue, skb);
4023 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4025 tcp_event_data_recv(sk, skb);
4027 tcp_fin(skb, sk, th);
4029 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4032 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4033 * gap in queue is filled.
4035 if (skb_queue_empty(&tp->out_of_order_queue))
4036 inet_csk(sk)->icsk_ack.pingpong = 0;
4039 if (tp->rx_opt.num_sacks)
4040 tcp_sack_remove(tp);
4042 tcp_fast_path_check(sk);
4046 else if (!sock_flag(sk, SOCK_DEAD))
4047 sk->sk_data_ready(sk, 0);
4051 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4052 /* A retransmit, 2nd most common case. Force an immediate ack. */
4053 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4054 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4057 tcp_enter_quickack_mode(sk);
4058 inet_csk_schedule_ack(sk);
4064 /* Out of window. F.e. zero window probe. */
4065 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4068 tcp_enter_quickack_mode(sk);
4070 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4071 /* Partial packet, seq < rcv_next < end_seq */
4072 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4073 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4074 TCP_SKB_CB(skb)->end_seq);
4076 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4078 /* If window is closed, drop tail of packet. But after
4079 * remembering D-SACK for its head made in previous line.
4081 if (!tcp_receive_window(tp))
4086 TCP_ECN_check_ce(tp, skb);
4088 if (tcp_try_rmem_schedule(sk, skb->truesize))
4091 /* Disable header prediction. */
4093 inet_csk_schedule_ack(sk);
4095 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4096 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4098 skb_set_owner_r(skb, sk);
4100 if (!skb_peek(&tp->out_of_order_queue)) {
4101 /* Initial out of order segment, build 1 SACK. */
4102 if (tcp_is_sack(tp)) {
4103 tp->rx_opt.num_sacks = 1;
4104 tp->rx_opt.dsack = 0;
4105 tp->rx_opt.eff_sacks = 1;
4106 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4107 tp->selective_acks[0].end_seq =
4108 TCP_SKB_CB(skb)->end_seq;
4110 __skb_queue_head(&tp->out_of_order_queue, skb);
4112 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4113 u32 seq = TCP_SKB_CB(skb)->seq;
4114 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4116 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4117 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4119 if (!tp->rx_opt.num_sacks ||
4120 tp->selective_acks[0].end_seq != seq)
4123 /* Common case: data arrive in order after hole. */
4124 tp->selective_acks[0].end_seq = end_seq;
4128 /* Find place to insert this segment. */
4130 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4132 } while ((skb1 = skb1->prev) !=
4133 (struct sk_buff *)&tp->out_of_order_queue);
4135 /* Do skb overlap to previous one? */
4136 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4137 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4138 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4139 /* All the bits are present. Drop. */
4141 tcp_dsack_set(sk, seq, end_seq);
4144 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4145 /* Partial overlap. */
4146 tcp_dsack_set(sk, seq,
4147 TCP_SKB_CB(skb1)->end_seq);
4152 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
4154 /* And clean segments covered by new one as whole. */
4155 while ((skb1 = skb->next) !=
4156 (struct sk_buff *)&tp->out_of_order_queue &&
4157 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4158 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4159 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4163 __skb_unlink(skb1, &tp->out_of_order_queue);
4164 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4165 TCP_SKB_CB(skb1)->end_seq);
4170 if (tcp_is_sack(tp))
4171 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4175 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4176 struct sk_buff_head *list)
4178 struct sk_buff *next = skb->next;
4180 __skb_unlink(skb, list);
4182 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4187 /* Collapse contiguous sequence of skbs head..tail with
4188 * sequence numbers start..end.
4189 * Segments with FIN/SYN are not collapsed (only because this
4193 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4194 struct sk_buff *head, struct sk_buff *tail,
4197 struct sk_buff *skb;
4199 /* First, check that queue is collapsible and find
4200 * the point where collapsing can be useful. */
4201 for (skb = head; skb != tail;) {
4202 /* No new bits? It is possible on ofo queue. */
4203 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4204 skb = tcp_collapse_one(sk, skb, list);
4208 /* The first skb to collapse is:
4210 * - bloated or contains data before "start" or
4211 * overlaps to the next one.
4213 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4214 (tcp_win_from_space(skb->truesize) > skb->len ||
4215 before(TCP_SKB_CB(skb)->seq, start) ||
4216 (skb->next != tail &&
4217 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4220 /* Decided to skip this, advance start seq. */
4221 start = TCP_SKB_CB(skb)->end_seq;
4224 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4227 while (before(start, end)) {
4228 struct sk_buff *nskb;
4229 unsigned int header = skb_headroom(skb);
4230 int copy = SKB_MAX_ORDER(header, 0);
4232 /* Too big header? This can happen with IPv6. */
4235 if (end - start < copy)
4237 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4241 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4242 skb_set_network_header(nskb, (skb_network_header(skb) -
4244 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4246 skb_reserve(nskb, header);
4247 memcpy(nskb->head, skb->head, header);
4248 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4249 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4250 __skb_insert(nskb, skb->prev, skb, list);
4251 skb_set_owner_r(nskb, sk);
4253 /* Copy data, releasing collapsed skbs. */
4255 int offset = start - TCP_SKB_CB(skb)->seq;
4256 int size = TCP_SKB_CB(skb)->end_seq - start;
4260 size = min(copy, size);
4261 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4263 TCP_SKB_CB(nskb)->end_seq += size;
4267 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4268 skb = tcp_collapse_one(sk, skb, list);
4270 tcp_hdr(skb)->syn ||
4278 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4279 * and tcp_collapse() them until all the queue is collapsed.
4281 static void tcp_collapse_ofo_queue(struct sock *sk)
4283 struct tcp_sock *tp = tcp_sk(sk);
4284 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4285 struct sk_buff *head;
4291 start = TCP_SKB_CB(skb)->seq;
4292 end = TCP_SKB_CB(skb)->end_seq;
4298 /* Segment is terminated when we see gap or when
4299 * we are at the end of all the queue. */
4300 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4301 after(TCP_SKB_CB(skb)->seq, end) ||
4302 before(TCP_SKB_CB(skb)->end_seq, start)) {
4303 tcp_collapse(sk, &tp->out_of_order_queue,
4304 head, skb, start, end);
4306 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4308 /* Start new segment */
4309 start = TCP_SKB_CB(skb)->seq;
4310 end = TCP_SKB_CB(skb)->end_seq;
4312 if (before(TCP_SKB_CB(skb)->seq, start))
4313 start = TCP_SKB_CB(skb)->seq;
4314 if (after(TCP_SKB_CB(skb)->end_seq, end))
4315 end = TCP_SKB_CB(skb)->end_seq;
4321 * Purge the out-of-order queue.
4322 * Return true if queue was pruned.
4324 static int tcp_prune_ofo_queue(struct sock *sk)
4326 struct tcp_sock *tp = tcp_sk(sk);
4329 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4330 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4331 __skb_queue_purge(&tp->out_of_order_queue);
4333 /* Reset SACK state. A conforming SACK implementation will
4334 * do the same at a timeout based retransmit. When a connection
4335 * is in a sad state like this, we care only about integrity
4336 * of the connection not performance.
4338 if (tp->rx_opt.sack_ok)
4339 tcp_sack_reset(&tp->rx_opt);
4346 /* Reduce allocated memory if we can, trying to get
4347 * the socket within its memory limits again.
4349 * Return less than zero if we should start dropping frames
4350 * until the socket owning process reads some of the data
4351 * to stabilize the situation.
4353 static int tcp_prune_queue(struct sock *sk)
4355 struct tcp_sock *tp = tcp_sk(sk);
4357 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4359 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4361 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4362 tcp_clamp_window(sk);
4363 else if (tcp_memory_pressure)
4364 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4366 tcp_collapse_ofo_queue(sk);
4367 tcp_collapse(sk, &sk->sk_receive_queue,
4368 sk->sk_receive_queue.next,
4369 (struct sk_buff *)&sk->sk_receive_queue,
4370 tp->copied_seq, tp->rcv_nxt);
4373 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4376 /* Collapsing did not help, destructive actions follow.
4377 * This must not ever occur. */
4379 tcp_prune_ofo_queue(sk);
4381 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4384 /* If we are really being abused, tell the caller to silently
4385 * drop receive data on the floor. It will get retransmitted
4386 * and hopefully then we'll have sufficient space.
4388 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4390 /* Massive buffer overcommit. */
4395 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4396 * As additional protections, we do not touch cwnd in retransmission phases,
4397 * and if application hit its sndbuf limit recently.
4399 void tcp_cwnd_application_limited(struct sock *sk)
4401 struct tcp_sock *tp = tcp_sk(sk);
4403 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4404 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4405 /* Limited by application or receiver window. */
4406 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4407 u32 win_used = max(tp->snd_cwnd_used, init_win);
4408 if (win_used < tp->snd_cwnd) {
4409 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4410 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4412 tp->snd_cwnd_used = 0;
4414 tp->snd_cwnd_stamp = tcp_time_stamp;
4417 static int tcp_should_expand_sndbuf(struct sock *sk)
4419 struct tcp_sock *tp = tcp_sk(sk);
4421 /* If the user specified a specific send buffer setting, do
4424 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4427 /* If we are under global TCP memory pressure, do not expand. */
4428 if (tcp_memory_pressure)
4431 /* If we are under soft global TCP memory pressure, do not expand. */
4432 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4435 /* If we filled the congestion window, do not expand. */
4436 if (tp->packets_out >= tp->snd_cwnd)
4442 /* When incoming ACK allowed to free some skb from write_queue,
4443 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4444 * on the exit from tcp input handler.
4446 * PROBLEM: sndbuf expansion does not work well with largesend.
4448 static void tcp_new_space(struct sock *sk)
4450 struct tcp_sock *tp = tcp_sk(sk);
4452 if (tcp_should_expand_sndbuf(sk)) {
4453 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4454 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4455 demanded = max_t(unsigned int, tp->snd_cwnd,
4456 tp->reordering + 1);
4457 sndmem *= 2 * demanded;
4458 if (sndmem > sk->sk_sndbuf)
4459 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4460 tp->snd_cwnd_stamp = tcp_time_stamp;
4463 sk->sk_write_space(sk);
4466 static void tcp_check_space(struct sock *sk)
4468 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4469 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4470 if (sk->sk_socket &&
4471 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4476 static inline void tcp_data_snd_check(struct sock *sk)
4478 tcp_push_pending_frames(sk);
4479 tcp_check_space(sk);
4483 * Check if sending an ack is needed.
4485 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4487 struct tcp_sock *tp = tcp_sk(sk);
4489 /* More than one full frame received... */
4490 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4491 /* ... and right edge of window advances far enough.
4492 * (tcp_recvmsg() will send ACK otherwise). Or...
4494 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4495 /* We ACK each frame or... */
4496 tcp_in_quickack_mode(sk) ||
4497 /* We have out of order data. */
4498 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4499 /* Then ack it now */
4502 /* Else, send delayed ack. */
4503 tcp_send_delayed_ack(sk);
4507 static inline void tcp_ack_snd_check(struct sock *sk)
4509 if (!inet_csk_ack_scheduled(sk)) {
4510 /* We sent a data segment already. */
4513 __tcp_ack_snd_check(sk, 1);
4517 * This routine is only called when we have urgent data
4518 * signaled. Its the 'slow' part of tcp_urg. It could be
4519 * moved inline now as tcp_urg is only called from one
4520 * place. We handle URGent data wrong. We have to - as
4521 * BSD still doesn't use the correction from RFC961.
4522 * For 1003.1g we should support a new option TCP_STDURG to permit
4523 * either form (or just set the sysctl tcp_stdurg).
4526 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4528 struct tcp_sock *tp = tcp_sk(sk);
4529 u32 ptr = ntohs(th->urg_ptr);
4531 if (ptr && !sysctl_tcp_stdurg)
4533 ptr += ntohl(th->seq);
4535 /* Ignore urgent data that we've already seen and read. */
4536 if (after(tp->copied_seq, ptr))
4539 /* Do not replay urg ptr.
4541 * NOTE: interesting situation not covered by specs.
4542 * Misbehaving sender may send urg ptr, pointing to segment,
4543 * which we already have in ofo queue. We are not able to fetch
4544 * such data and will stay in TCP_URG_NOTYET until will be eaten
4545 * by recvmsg(). Seems, we are not obliged to handle such wicked
4546 * situations. But it is worth to think about possibility of some
4547 * DoSes using some hypothetical application level deadlock.
4549 if (before(ptr, tp->rcv_nxt))
4552 /* Do we already have a newer (or duplicate) urgent pointer? */
4553 if (tp->urg_data && !after(ptr, tp->urg_seq))
4556 /* Tell the world about our new urgent pointer. */
4559 /* We may be adding urgent data when the last byte read was
4560 * urgent. To do this requires some care. We cannot just ignore
4561 * tp->copied_seq since we would read the last urgent byte again
4562 * as data, nor can we alter copied_seq until this data arrives
4563 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4565 * NOTE. Double Dutch. Rendering to plain English: author of comment
4566 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4567 * and expect that both A and B disappear from stream. This is _wrong_.
4568 * Though this happens in BSD with high probability, this is occasional.
4569 * Any application relying on this is buggy. Note also, that fix "works"
4570 * only in this artificial test. Insert some normal data between A and B and we will
4571 * decline of BSD again. Verdict: it is better to remove to trap
4574 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4575 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4576 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4578 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4579 __skb_unlink(skb, &sk->sk_receive_queue);
4584 tp->urg_data = TCP_URG_NOTYET;
4587 /* Disable header prediction. */
4591 /* This is the 'fast' part of urgent handling. */
4592 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4594 struct tcp_sock *tp = tcp_sk(sk);
4596 /* Check if we get a new urgent pointer - normally not. */
4598 tcp_check_urg(sk, th);
4600 /* Do we wait for any urgent data? - normally not... */
4601 if (tp->urg_data == TCP_URG_NOTYET) {
4602 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4605 /* Is the urgent pointer pointing into this packet? */
4606 if (ptr < skb->len) {
4608 if (skb_copy_bits(skb, ptr, &tmp, 1))
4610 tp->urg_data = TCP_URG_VALID | tmp;
4611 if (!sock_flag(sk, SOCK_DEAD))
4612 sk->sk_data_ready(sk, 0);
4617 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4619 struct tcp_sock *tp = tcp_sk(sk);
4620 int chunk = skb->len - hlen;
4624 if (skb_csum_unnecessary(skb))
4625 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4627 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4631 tp->ucopy.len -= chunk;
4632 tp->copied_seq += chunk;
4633 tcp_rcv_space_adjust(sk);
4640 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4641 struct sk_buff *skb)
4645 if (sock_owned_by_user(sk)) {
4647 result = __tcp_checksum_complete(skb);
4650 result = __tcp_checksum_complete(skb);
4655 static inline int tcp_checksum_complete_user(struct sock *sk,
4656 struct sk_buff *skb)
4658 return !skb_csum_unnecessary(skb) &&
4659 __tcp_checksum_complete_user(sk, skb);
4662 #ifdef CONFIG_NET_DMA
4663 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4666 struct tcp_sock *tp = tcp_sk(sk);
4667 int chunk = skb->len - hlen;
4669 int copied_early = 0;
4671 if (tp->ucopy.wakeup)
4674 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4675 tp->ucopy.dma_chan = get_softnet_dma();
4677 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4679 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4681 tp->ucopy.iov, chunk,
4682 tp->ucopy.pinned_list);
4687 tp->ucopy.dma_cookie = dma_cookie;
4690 tp->ucopy.len -= chunk;
4691 tp->copied_seq += chunk;
4692 tcp_rcv_space_adjust(sk);
4694 if ((tp->ucopy.len == 0) ||
4695 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4696 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4697 tp->ucopy.wakeup = 1;
4698 sk->sk_data_ready(sk, 0);
4700 } else if (chunk > 0) {
4701 tp->ucopy.wakeup = 1;
4702 sk->sk_data_ready(sk, 0);
4705 return copied_early;
4707 #endif /* CONFIG_NET_DMA */
4709 /* Does PAWS and seqno based validation of an incoming segment, flags will
4710 * play significant role here.
4712 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
4713 struct tcphdr *th, int syn_inerr)
4715 struct tcp_sock *tp = tcp_sk(sk);
4717 /* RFC1323: H1. Apply PAWS check first. */
4718 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4719 tcp_paws_discard(sk, skb)) {
4721 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
4722 tcp_send_dupack(sk, skb);
4725 /* Reset is accepted even if it did not pass PAWS. */
4728 /* Step 1: check sequence number */
4729 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4730 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4731 * (RST) segments are validated by checking their SEQ-fields."
4732 * And page 69: "If an incoming segment is not acceptable,
4733 * an acknowledgment should be sent in reply (unless the RST
4734 * bit is set, if so drop the segment and return)".
4737 tcp_send_dupack(sk, skb);
4741 /* Step 2: check RST bit */
4747 /* ts_recent update must be made after we are sure that the packet
4750 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4752 /* step 3: check security and precedence [ignored] */
4754 /* step 4: Check for a SYN in window. */
4755 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4757 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4758 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
4771 * TCP receive function for the ESTABLISHED state.
4773 * It is split into a fast path and a slow path. The fast path is
4775 * - A zero window was announced from us - zero window probing
4776 * is only handled properly in the slow path.
4777 * - Out of order segments arrived.
4778 * - Urgent data is expected.
4779 * - There is no buffer space left
4780 * - Unexpected TCP flags/window values/header lengths are received
4781 * (detected by checking the TCP header against pred_flags)
4782 * - Data is sent in both directions. Fast path only supports pure senders
4783 * or pure receivers (this means either the sequence number or the ack
4784 * value must stay constant)
4785 * - Unexpected TCP option.
4787 * When these conditions are not satisfied it drops into a standard
4788 * receive procedure patterned after RFC793 to handle all cases.
4789 * The first three cases are guaranteed by proper pred_flags setting,
4790 * the rest is checked inline. Fast processing is turned on in
4791 * tcp_data_queue when everything is OK.
4793 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4794 struct tcphdr *th, unsigned len)
4796 struct tcp_sock *tp = tcp_sk(sk);
4800 * Header prediction.
4801 * The code loosely follows the one in the famous
4802 * "30 instruction TCP receive" Van Jacobson mail.
4804 * Van's trick is to deposit buffers into socket queue
4805 * on a device interrupt, to call tcp_recv function
4806 * on the receive process context and checksum and copy
4807 * the buffer to user space. smart...
4809 * Our current scheme is not silly either but we take the
4810 * extra cost of the net_bh soft interrupt processing...
4811 * We do checksum and copy also but from device to kernel.
4814 tp->rx_opt.saw_tstamp = 0;
4816 /* pred_flags is 0xS?10 << 16 + snd_wnd
4817 * if header_prediction is to be made
4818 * 'S' will always be tp->tcp_header_len >> 2
4819 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4820 * turn it off (when there are holes in the receive
4821 * space for instance)
4822 * PSH flag is ignored.
4825 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4826 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4827 int tcp_header_len = tp->tcp_header_len;
4829 /* Timestamp header prediction: tcp_header_len
4830 * is automatically equal to th->doff*4 due to pred_flags
4834 /* Check timestamp */
4835 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4836 /* No? Slow path! */
4837 if (!tcp_parse_aligned_timestamp(tp, th))
4840 /* If PAWS failed, check it more carefully in slow path */
4841 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4844 /* DO NOT update ts_recent here, if checksum fails
4845 * and timestamp was corrupted part, it will result
4846 * in a hung connection since we will drop all
4847 * future packets due to the PAWS test.
4851 if (len <= tcp_header_len) {
4852 /* Bulk data transfer: sender */
4853 if (len == tcp_header_len) {
4854 /* Predicted packet is in window by definition.
4855 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4856 * Hence, check seq<=rcv_wup reduces to:
4858 if (tcp_header_len ==
4859 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4860 tp->rcv_nxt == tp->rcv_wup)
4861 tcp_store_ts_recent(tp);
4863 /* We know that such packets are checksummed
4866 tcp_ack(sk, skb, 0);
4868 tcp_data_snd_check(sk);
4870 } else { /* Header too small */
4871 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
4876 int copied_early = 0;
4878 if (tp->copied_seq == tp->rcv_nxt &&
4879 len - tcp_header_len <= tp->ucopy.len) {
4880 #ifdef CONFIG_NET_DMA
4881 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4886 if (tp->ucopy.task == current &&
4887 sock_owned_by_user(sk) && !copied_early) {
4888 __set_current_state(TASK_RUNNING);
4890 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4894 /* Predicted packet is in window by definition.
4895 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4896 * Hence, check seq<=rcv_wup reduces to:
4898 if (tcp_header_len ==
4899 (sizeof(struct tcphdr) +
4900 TCPOLEN_TSTAMP_ALIGNED) &&
4901 tp->rcv_nxt == tp->rcv_wup)
4902 tcp_store_ts_recent(tp);
4904 tcp_rcv_rtt_measure_ts(sk, skb);
4906 __skb_pull(skb, tcp_header_len);
4907 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4908 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
4911 tcp_cleanup_rbuf(sk, skb->len);
4914 if (tcp_checksum_complete_user(sk, skb))
4917 /* Predicted packet is in window by definition.
4918 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4919 * Hence, check seq<=rcv_wup reduces to:
4921 if (tcp_header_len ==
4922 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4923 tp->rcv_nxt == tp->rcv_wup)
4924 tcp_store_ts_recent(tp);
4926 tcp_rcv_rtt_measure_ts(sk, skb);
4928 if ((int)skb->truesize > sk->sk_forward_alloc)
4931 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
4933 /* Bulk data transfer: receiver */
4934 __skb_pull(skb, tcp_header_len);
4935 __skb_queue_tail(&sk->sk_receive_queue, skb);
4936 skb_set_owner_r(skb, sk);
4937 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4940 tcp_event_data_recv(sk, skb);
4942 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4943 /* Well, only one small jumplet in fast path... */
4944 tcp_ack(sk, skb, FLAG_DATA);
4945 tcp_data_snd_check(sk);
4946 if (!inet_csk_ack_scheduled(sk))
4950 __tcp_ack_snd_check(sk, 0);
4952 #ifdef CONFIG_NET_DMA
4954 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4960 sk->sk_data_ready(sk, 0);
4966 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
4970 * Standard slow path.
4973 res = tcp_validate_incoming(sk, skb, th, 1);
4979 tcp_ack(sk, skb, FLAG_SLOWPATH);
4981 tcp_rcv_rtt_measure_ts(sk, skb);
4983 /* Process urgent data. */
4984 tcp_urg(sk, skb, th);
4986 /* step 7: process the segment text */
4987 tcp_data_queue(sk, skb);
4989 tcp_data_snd_check(sk);
4990 tcp_ack_snd_check(sk);
4994 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5001 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5002 struct tcphdr *th, unsigned len)
5004 struct tcp_sock *tp = tcp_sk(sk);
5005 struct inet_connection_sock *icsk = inet_csk(sk);
5006 int saved_clamp = tp->rx_opt.mss_clamp;
5008 tcp_parse_options(skb, &tp->rx_opt, 0);
5012 * "If the state is SYN-SENT then
5013 * first check the ACK bit
5014 * If the ACK bit is set
5015 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5016 * a reset (unless the RST bit is set, if so drop
5017 * the segment and return)"
5019 * We do not send data with SYN, so that RFC-correct
5022 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5023 goto reset_and_undo;
5025 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5026 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5028 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5029 goto reset_and_undo;
5032 /* Now ACK is acceptable.
5034 * "If the RST bit is set
5035 * If the ACK was acceptable then signal the user "error:
5036 * connection reset", drop the segment, enter CLOSED state,
5037 * delete TCB, and return."
5046 * "fifth, if neither of the SYN or RST bits is set then
5047 * drop the segment and return."
5053 goto discard_and_undo;
5056 * "If the SYN bit is on ...
5057 * are acceptable then ...
5058 * (our SYN has been ACKed), change the connection
5059 * state to ESTABLISHED..."
5062 TCP_ECN_rcv_synack(tp, th);
5064 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5065 tcp_ack(sk, skb, FLAG_SLOWPATH);
5067 /* Ok.. it's good. Set up sequence numbers and
5068 * move to established.
5070 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5071 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5073 /* RFC1323: The window in SYN & SYN/ACK segments is
5076 tp->snd_wnd = ntohs(th->window);
5077 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
5079 if (!tp->rx_opt.wscale_ok) {
5080 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5081 tp->window_clamp = min(tp->window_clamp, 65535U);
5084 if (tp->rx_opt.saw_tstamp) {
5085 tp->rx_opt.tstamp_ok = 1;
5086 tp->tcp_header_len =
5087 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5088 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5089 tcp_store_ts_recent(tp);
5091 tp->tcp_header_len = sizeof(struct tcphdr);
5094 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5095 tcp_enable_fack(tp);
5098 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5099 tcp_initialize_rcv_mss(sk);
5101 /* Remember, tcp_poll() does not lock socket!
5102 * Change state from SYN-SENT only after copied_seq
5103 * is initialized. */
5104 tp->copied_seq = tp->rcv_nxt;
5106 tcp_set_state(sk, TCP_ESTABLISHED);
5108 security_inet_conn_established(sk, skb);
5110 /* Make sure socket is routed, for correct metrics. */
5111 icsk->icsk_af_ops->rebuild_header(sk);
5113 tcp_init_metrics(sk);
5115 tcp_init_congestion_control(sk);
5117 /* Prevent spurious tcp_cwnd_restart() on first data
5120 tp->lsndtime = tcp_time_stamp;
5122 tcp_init_buffer_space(sk);
5124 if (sock_flag(sk, SOCK_KEEPOPEN))
5125 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5127 if (!tp->rx_opt.snd_wscale)
5128 __tcp_fast_path_on(tp, tp->snd_wnd);
5132 if (!sock_flag(sk, SOCK_DEAD)) {
5133 sk->sk_state_change(sk);
5134 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5137 if (sk->sk_write_pending ||
5138 icsk->icsk_accept_queue.rskq_defer_accept ||
5139 icsk->icsk_ack.pingpong) {
5140 /* Save one ACK. Data will be ready after
5141 * several ticks, if write_pending is set.
5143 * It may be deleted, but with this feature tcpdumps
5144 * look so _wonderfully_ clever, that I was not able
5145 * to stand against the temptation 8) --ANK
5147 inet_csk_schedule_ack(sk);
5148 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5149 icsk->icsk_ack.ato = TCP_ATO_MIN;
5150 tcp_incr_quickack(sk);
5151 tcp_enter_quickack_mode(sk);
5152 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5153 TCP_DELACK_MAX, TCP_RTO_MAX);
5164 /* No ACK in the segment */
5168 * "If the RST bit is set
5170 * Otherwise (no ACK) drop the segment and return."
5173 goto discard_and_undo;
5177 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5178 tcp_paws_check(&tp->rx_opt, 0))
5179 goto discard_and_undo;
5182 /* We see SYN without ACK. It is attempt of
5183 * simultaneous connect with crossed SYNs.
5184 * Particularly, it can be connect to self.
5186 tcp_set_state(sk, TCP_SYN_RECV);
5188 if (tp->rx_opt.saw_tstamp) {
5189 tp->rx_opt.tstamp_ok = 1;
5190 tcp_store_ts_recent(tp);
5191 tp->tcp_header_len =
5192 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5194 tp->tcp_header_len = sizeof(struct tcphdr);
5197 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5198 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5200 /* RFC1323: The window in SYN & SYN/ACK segments is
5203 tp->snd_wnd = ntohs(th->window);
5204 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5205 tp->max_window = tp->snd_wnd;
5207 TCP_ECN_rcv_syn(tp, th);
5210 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5211 tcp_initialize_rcv_mss(sk);
5213 tcp_send_synack(sk);
5215 /* Note, we could accept data and URG from this segment.
5216 * There are no obstacles to make this.
5218 * However, if we ignore data in ACKless segments sometimes,
5219 * we have no reasons to accept it sometimes.
5220 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5221 * is not flawless. So, discard packet for sanity.
5222 * Uncomment this return to process the data.
5229 /* "fifth, if neither of the SYN or RST bits is set then
5230 * drop the segment and return."
5234 tcp_clear_options(&tp->rx_opt);
5235 tp->rx_opt.mss_clamp = saved_clamp;
5239 tcp_clear_options(&tp->rx_opt);
5240 tp->rx_opt.mss_clamp = saved_clamp;
5245 * This function implements the receiving procedure of RFC 793 for
5246 * all states except ESTABLISHED and TIME_WAIT.
5247 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5248 * address independent.
5251 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5252 struct tcphdr *th, unsigned len)
5254 struct tcp_sock *tp = tcp_sk(sk);
5255 struct inet_connection_sock *icsk = inet_csk(sk);
5259 tp->rx_opt.saw_tstamp = 0;
5261 switch (sk->sk_state) {
5273 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5276 /* Now we have several options: In theory there is
5277 * nothing else in the frame. KA9Q has an option to
5278 * send data with the syn, BSD accepts data with the
5279 * syn up to the [to be] advertised window and
5280 * Solaris 2.1 gives you a protocol error. For now
5281 * we just ignore it, that fits the spec precisely
5282 * and avoids incompatibilities. It would be nice in
5283 * future to drop through and process the data.
5285 * Now that TTCP is starting to be used we ought to
5287 * But, this leaves one open to an easy denial of
5288 * service attack, and SYN cookies can't defend
5289 * against this problem. So, we drop the data
5290 * in the interest of security over speed unless
5291 * it's still in use.
5299 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5303 /* Do step6 onward by hand. */
5304 tcp_urg(sk, skb, th);
5306 tcp_data_snd_check(sk);
5310 res = tcp_validate_incoming(sk, skb, th, 0);
5314 /* step 5: check the ACK field */
5316 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5318 switch (sk->sk_state) {
5321 tp->copied_seq = tp->rcv_nxt;
5323 tcp_set_state(sk, TCP_ESTABLISHED);
5324 sk->sk_state_change(sk);
5326 /* Note, that this wakeup is only for marginal
5327 * crossed SYN case. Passively open sockets
5328 * are not waked up, because sk->sk_sleep ==
5329 * NULL and sk->sk_socket == NULL.
5333 SOCK_WAKE_IO, POLL_OUT);
5335 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5336 tp->snd_wnd = ntohs(th->window) <<
5337 tp->rx_opt.snd_wscale;
5338 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5339 TCP_SKB_CB(skb)->seq);
5341 /* tcp_ack considers this ACK as duplicate
5342 * and does not calculate rtt.
5343 * Fix it at least with timestamps.
5345 if (tp->rx_opt.saw_tstamp &&
5346 tp->rx_opt.rcv_tsecr && !tp->srtt)
5347 tcp_ack_saw_tstamp(sk, 0);
5349 if (tp->rx_opt.tstamp_ok)
5350 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5352 /* Make sure socket is routed, for
5355 icsk->icsk_af_ops->rebuild_header(sk);
5357 tcp_init_metrics(sk);
5359 tcp_init_congestion_control(sk);
5361 /* Prevent spurious tcp_cwnd_restart() on
5362 * first data packet.
5364 tp->lsndtime = tcp_time_stamp;
5367 tcp_initialize_rcv_mss(sk);
5368 tcp_init_buffer_space(sk);
5369 tcp_fast_path_on(tp);
5376 if (tp->snd_una == tp->write_seq) {
5377 tcp_set_state(sk, TCP_FIN_WAIT2);
5378 sk->sk_shutdown |= SEND_SHUTDOWN;
5379 dst_confirm(sk->sk_dst_cache);
5381 if (!sock_flag(sk, SOCK_DEAD))
5382 /* Wake up lingering close() */
5383 sk->sk_state_change(sk);
5387 if (tp->linger2 < 0 ||
5388 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5389 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5391 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5395 tmo = tcp_fin_time(sk);
5396 if (tmo > TCP_TIMEWAIT_LEN) {
5397 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5398 } else if (th->fin || sock_owned_by_user(sk)) {
5399 /* Bad case. We could lose such FIN otherwise.
5400 * It is not a big problem, but it looks confusing
5401 * and not so rare event. We still can lose it now,
5402 * if it spins in bh_lock_sock(), but it is really
5405 inet_csk_reset_keepalive_timer(sk, tmo);
5407 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5415 if (tp->snd_una == tp->write_seq) {
5416 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5422 if (tp->snd_una == tp->write_seq) {
5423 tcp_update_metrics(sk);
5432 /* step 6: check the URG bit */
5433 tcp_urg(sk, skb, th);
5435 /* step 7: process the segment text */
5436 switch (sk->sk_state) {
5437 case TCP_CLOSE_WAIT:
5440 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5444 /* RFC 793 says to queue data in these states,
5445 * RFC 1122 says we MUST send a reset.
5446 * BSD 4.4 also does reset.
5448 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5449 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5450 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5451 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5457 case TCP_ESTABLISHED:
5458 tcp_data_queue(sk, skb);
5463 /* tcp_data could move socket to TIME-WAIT */
5464 if (sk->sk_state != TCP_CLOSE) {
5465 tcp_data_snd_check(sk);
5466 tcp_ack_snd_check(sk);
5476 EXPORT_SYMBOL(sysctl_tcp_ecn);
5477 EXPORT_SYMBOL(sysctl_tcp_reordering);
5478 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5479 EXPORT_SYMBOL(tcp_parse_options);
5480 #ifdef CONFIG_TCP_MD5SIG
5481 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5483 EXPORT_SYMBOL(tcp_rcv_established);
5484 EXPORT_SYMBOL(tcp_rcv_state_process);
5485 EXPORT_SYMBOL(tcp_initialize_rcv_mss);