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 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1006 struct sk_buff *skb)
1008 tcp_verify_retransmit_hint(tp, skb);
1010 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1011 tp->lost_out += tcp_skb_pcount(skb);
1012 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1016 /* This procedure tags the retransmission queue when SACKs arrive.
1018 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1019 * Packets in queue with these bits set are counted in variables
1020 * sacked_out, retrans_out and lost_out, correspondingly.
1022 * Valid combinations are:
1023 * Tag InFlight Description
1024 * 0 1 - orig segment is in flight.
1025 * S 0 - nothing flies, orig reached receiver.
1026 * L 0 - nothing flies, orig lost by net.
1027 * R 2 - both orig and retransmit are in flight.
1028 * L|R 1 - orig is lost, retransmit is in flight.
1029 * S|R 1 - orig reached receiver, retrans is still in flight.
1030 * (L|S|R is logically valid, it could occur when L|R is sacked,
1031 * but it is equivalent to plain S and code short-curcuits it to S.
1032 * L|S is logically invalid, it would mean -1 packet in flight 8))
1034 * These 6 states form finite state machine, controlled by the following events:
1035 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1036 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1037 * 3. Loss detection event of one of three flavors:
1038 * A. Scoreboard estimator decided the packet is lost.
1039 * A'. Reno "three dupacks" marks head of queue lost.
1040 * A''. Its FACK modfication, head until snd.fack is lost.
1041 * B. SACK arrives sacking data transmitted after never retransmitted
1042 * hole was sent out.
1043 * C. SACK arrives sacking SND.NXT at the moment, when the
1044 * segment was retransmitted.
1045 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1047 * It is pleasant to note, that state diagram turns out to be commutative,
1048 * so that we are allowed not to be bothered by order of our actions,
1049 * when multiple events arrive simultaneously. (see the function below).
1051 * Reordering detection.
1052 * --------------------
1053 * Reordering metric is maximal distance, which a packet can be displaced
1054 * in packet stream. With SACKs we can estimate it:
1056 * 1. SACK fills old hole and the corresponding segment was not
1057 * ever retransmitted -> reordering. Alas, we cannot use it
1058 * when segment was retransmitted.
1059 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1060 * for retransmitted and already SACKed segment -> reordering..
1061 * Both of these heuristics are not used in Loss state, when we cannot
1062 * account for retransmits accurately.
1064 * SACK block validation.
1065 * ----------------------
1067 * SACK block range validation checks that the received SACK block fits to
1068 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1069 * Note that SND.UNA is not included to the range though being valid because
1070 * it means that the receiver is rather inconsistent with itself reporting
1071 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1072 * perfectly valid, however, in light of RFC2018 which explicitly states
1073 * that "SACK block MUST reflect the newest segment. Even if the newest
1074 * segment is going to be discarded ...", not that it looks very clever
1075 * in case of head skb. Due to potentional receiver driven attacks, we
1076 * choose to avoid immediate execution of a walk in write queue due to
1077 * reneging and defer head skb's loss recovery to standard loss recovery
1078 * procedure that will eventually trigger (nothing forbids us doing this).
1080 * Implements also blockage to start_seq wrap-around. Problem lies in the
1081 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1082 * there's no guarantee that it will be before snd_nxt (n). The problem
1083 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1086 * <- outs wnd -> <- wrapzone ->
1087 * u e n u_w e_w s n_w
1089 * |<------------+------+----- TCP seqno space --------------+---------->|
1090 * ...-- <2^31 ->| |<--------...
1091 * ...---- >2^31 ------>| |<--------...
1093 * Current code wouldn't be vulnerable but it's better still to discard such
1094 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1095 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1096 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1097 * equal to the ideal case (infinite seqno space without wrap caused issues).
1099 * With D-SACK the lower bound is extended to cover sequence space below
1100 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1101 * again, D-SACK block must not to go across snd_una (for the same reason as
1102 * for the normal SACK blocks, explained above). But there all simplicity
1103 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1104 * fully below undo_marker they do not affect behavior in anyway and can
1105 * therefore be safely ignored. In rare cases (which are more or less
1106 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1107 * fragmentation and packet reordering past skb's retransmission. To consider
1108 * them correctly, the acceptable range must be extended even more though
1109 * the exact amount is rather hard to quantify. However, tp->max_window can
1110 * be used as an exaggerated estimate.
1112 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1113 u32 start_seq, u32 end_seq)
1115 /* Too far in future, or reversed (interpretation is ambiguous) */
1116 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1119 /* Nasty start_seq wrap-around check (see comments above) */
1120 if (!before(start_seq, tp->snd_nxt))
1123 /* In outstanding window? ...This is valid exit for D-SACKs too.
1124 * start_seq == snd_una is non-sensical (see comments above)
1126 if (after(start_seq, tp->snd_una))
1129 if (!is_dsack || !tp->undo_marker)
1132 /* ...Then it's D-SACK, and must reside below snd_una completely */
1133 if (!after(end_seq, tp->snd_una))
1136 if (!before(start_seq, tp->undo_marker))
1140 if (!after(end_seq, tp->undo_marker))
1143 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1144 * start_seq < undo_marker and end_seq >= undo_marker.
1146 return !before(start_seq, end_seq - tp->max_window);
1149 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1150 * Event "C". Later note: FACK people cheated me again 8), we have to account
1151 * for reordering! Ugly, but should help.
1153 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1154 * less than what is now known to be received by the other end (derived from
1155 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1156 * retransmitted skbs to avoid some costly processing per ACKs.
1158 static void tcp_mark_lost_retrans(struct sock *sk)
1160 const struct inet_connection_sock *icsk = inet_csk(sk);
1161 struct tcp_sock *tp = tcp_sk(sk);
1162 struct sk_buff *skb;
1164 u32 new_low_seq = tp->snd_nxt;
1165 u32 received_upto = tcp_highest_sack_seq(tp);
1167 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1168 !after(received_upto, tp->lost_retrans_low) ||
1169 icsk->icsk_ca_state != TCP_CA_Recovery)
1172 tcp_for_write_queue(skb, sk) {
1173 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1175 if (skb == tcp_send_head(sk))
1177 if (cnt == tp->retrans_out)
1179 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1182 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1185 if (after(received_upto, ack_seq) &&
1187 !before(received_upto,
1188 ack_seq + tp->reordering * tp->mss_cache))) {
1189 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1190 tp->retrans_out -= tcp_skb_pcount(skb);
1192 tcp_skb_mark_lost_uncond_verify(tp, skb);
1193 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1195 if (before(ack_seq, new_low_seq))
1196 new_low_seq = ack_seq;
1197 cnt += tcp_skb_pcount(skb);
1201 if (tp->retrans_out)
1202 tp->lost_retrans_low = new_low_seq;
1205 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1206 struct tcp_sack_block_wire *sp, int num_sacks,
1209 struct tcp_sock *tp = tcp_sk(sk);
1210 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1211 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1214 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1217 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1218 } else if (num_sacks > 1) {
1219 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1220 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1222 if (!after(end_seq_0, end_seq_1) &&
1223 !before(start_seq_0, start_seq_1)) {
1226 NET_INC_STATS_BH(sock_net(sk),
1227 LINUX_MIB_TCPDSACKOFORECV);
1231 /* D-SACK for already forgotten data... Do dumb counting. */
1233 !after(end_seq_0, prior_snd_una) &&
1234 after(end_seq_0, tp->undo_marker))
1240 struct tcp_sacktag_state {
1246 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1247 * the incoming SACK may not exactly match but we can find smaller MSS
1248 * aligned portion of it that matches. Therefore we might need to fragment
1249 * which may fail and creates some hassle (caller must handle error case
1252 * FIXME: this could be merged to shift decision code
1254 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1255 u32 start_seq, u32 end_seq)
1258 unsigned int pkt_len;
1261 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1262 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1264 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1265 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1266 mss = tcp_skb_mss(skb);
1267 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1270 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1274 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1279 /* Round if necessary so that SACKs cover only full MSSes
1280 * and/or the remaining small portion (if present)
1282 if (pkt_len > mss) {
1283 unsigned int new_len = (pkt_len / mss) * mss;
1284 if (!in_sack && new_len < pkt_len) {
1286 if (new_len > skb->len)
1291 err = tcp_fragment(sk, skb, pkt_len, mss);
1299 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1300 struct tcp_sacktag_state *state,
1301 int dup_sack, int pcount)
1303 struct tcp_sock *tp = tcp_sk(sk);
1304 u8 sacked = TCP_SKB_CB(skb)->sacked;
1305 int fack_count = state->fack_count;
1307 /* Account D-SACK for retransmitted packet. */
1308 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1309 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1311 if (sacked & TCPCB_SACKED_ACKED)
1312 state->reord = min(fack_count, state->reord);
1315 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1316 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1319 if (!(sacked & TCPCB_SACKED_ACKED)) {
1320 if (sacked & TCPCB_SACKED_RETRANS) {
1321 /* If the segment is not tagged as lost,
1322 * we do not clear RETRANS, believing
1323 * that retransmission is still in flight.
1325 if (sacked & TCPCB_LOST) {
1326 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1327 tp->lost_out -= pcount;
1328 tp->retrans_out -= pcount;
1331 if (!(sacked & TCPCB_RETRANS)) {
1332 /* New sack for not retransmitted frame,
1333 * which was in hole. It is reordering.
1335 if (before(TCP_SKB_CB(skb)->seq,
1336 tcp_highest_sack_seq(tp)))
1337 state->reord = min(fack_count,
1340 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1341 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1342 state->flag |= FLAG_ONLY_ORIG_SACKED;
1345 if (sacked & TCPCB_LOST) {
1346 sacked &= ~TCPCB_LOST;
1347 tp->lost_out -= pcount;
1351 sacked |= TCPCB_SACKED_ACKED;
1352 state->flag |= FLAG_DATA_SACKED;
1353 tp->sacked_out += pcount;
1355 fack_count += pcount;
1357 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1358 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1359 before(TCP_SKB_CB(skb)->seq,
1360 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1361 tp->lost_cnt_hint += pcount;
1363 if (fack_count > tp->fackets_out)
1364 tp->fackets_out = fack_count;
1367 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1368 * frames and clear it. undo_retrans is decreased above, L|R frames
1369 * are accounted above as well.
1371 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1372 sacked &= ~TCPCB_SACKED_RETRANS;
1373 tp->retrans_out -= pcount;
1379 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1380 struct tcp_sacktag_state *state,
1381 unsigned int pcount, int shifted, int mss)
1383 struct tcp_sock *tp = tcp_sk(sk);
1384 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1388 /* Tweak before seqno plays */
1389 if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1390 !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1391 tp->lost_cnt_hint += pcount;
1393 TCP_SKB_CB(prev)->end_seq += shifted;
1394 TCP_SKB_CB(skb)->seq += shifted;
1396 skb_shinfo(prev)->gso_segs += pcount;
1397 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1398 skb_shinfo(skb)->gso_segs -= pcount;
1400 /* When we're adding to gso_segs == 1, gso_size will be zero,
1401 * in theory this shouldn't be necessary but as long as DSACK
1402 * code can come after this skb later on it's better to keep
1403 * setting gso_size to something.
1405 if (!skb_shinfo(prev)->gso_size) {
1406 skb_shinfo(prev)->gso_size = mss;
1407 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1410 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1411 if (skb_shinfo(skb)->gso_segs <= 1) {
1412 skb_shinfo(skb)->gso_size = 0;
1413 skb_shinfo(skb)->gso_type = 0;
1416 /* We discard results */
1417 tcp_sacktag_one(skb, sk, state, 0, pcount);
1419 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1420 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1423 BUG_ON(!tcp_skb_pcount(skb));
1424 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1428 /* Whole SKB was eaten :-) */
1430 if (skb == tp->retransmit_skb_hint)
1431 tp->retransmit_skb_hint = prev;
1432 if (skb == tp->scoreboard_skb_hint)
1433 tp->scoreboard_skb_hint = prev;
1434 if (skb == tp->lost_skb_hint) {
1435 tp->lost_skb_hint = prev;
1436 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1439 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1440 if (skb == tcp_highest_sack(sk))
1441 tcp_advance_highest_sack(sk, skb);
1443 tcp_unlink_write_queue(skb, sk);
1444 sk_wmem_free_skb(sk, skb);
1446 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1451 /* I wish gso_size would have a bit more sane initialization than
1452 * something-or-zero which complicates things
1454 static int tcp_skb_seglen(struct sk_buff *skb)
1456 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1459 /* Shifting pages past head area doesn't work */
1460 static int skb_can_shift(struct sk_buff *skb)
1462 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1465 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1468 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1469 struct tcp_sacktag_state *state,
1470 u32 start_seq, u32 end_seq,
1473 struct tcp_sock *tp = tcp_sk(sk);
1474 struct sk_buff *prev;
1480 if (!sk_can_gso(sk))
1483 /* Normally R but no L won't result in plain S */
1485 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1487 if (!skb_can_shift(skb))
1489 /* This frame is about to be dropped (was ACKed). */
1490 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1493 /* Can only happen with delayed DSACK + discard craziness */
1494 if (unlikely(skb == tcp_write_queue_head(sk)))
1496 prev = tcp_write_queue_prev(sk, skb);
1498 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1501 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1502 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1506 pcount = tcp_skb_pcount(skb);
1507 mss = tcp_skb_seglen(skb);
1509 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1510 * drop this restriction as unnecessary
1512 if (mss != tcp_skb_seglen(prev))
1515 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1517 /* CHECKME: This is non-MSS split case only?, this will
1518 * cause skipped skbs due to advancing loop btw, original
1519 * has that feature too
1521 if (tcp_skb_pcount(skb) <= 1)
1524 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1526 /* TODO: head merge to next could be attempted here
1527 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1528 * though it might not be worth of the additional hassle
1530 * ...we can probably just fallback to what was done
1531 * previously. We could try merging non-SACKed ones
1532 * as well but it probably isn't going to buy off
1533 * because later SACKs might again split them, and
1534 * it would make skb timestamp tracking considerably
1540 len = end_seq - TCP_SKB_CB(skb)->seq;
1542 BUG_ON(len > skb->len);
1544 /* MSS boundaries should be honoured or else pcount will
1545 * severely break even though it makes things bit trickier.
1546 * Optimize common case to avoid most of the divides
1548 mss = tcp_skb_mss(skb);
1550 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1551 * drop this restriction as unnecessary
1553 if (mss != tcp_skb_seglen(prev))
1558 } else if (len < mss) {
1566 if (!skb_shift(prev, skb, len))
1568 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss))
1571 /* Hole filled allows collapsing with the next as well, this is very
1572 * useful when hole on every nth skb pattern happens
1574 if (prev == tcp_write_queue_tail(sk))
1576 skb = tcp_write_queue_next(sk, prev);
1578 if (!skb_can_shift(skb) ||
1579 (skb == tcp_send_head(sk)) ||
1580 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1581 (mss != tcp_skb_seglen(skb)))
1585 if (skb_shift(prev, skb, len)) {
1586 pcount += tcp_skb_pcount(skb);
1587 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss);
1591 state->fack_count += pcount;
1598 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1602 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1603 struct tcp_sack_block *next_dup,
1604 struct tcp_sacktag_state *state,
1605 u32 start_seq, u32 end_seq,
1608 struct tcp_sock *tp = tcp_sk(sk);
1609 struct sk_buff *tmp;
1611 tcp_for_write_queue_from(skb, sk) {
1613 int dup_sack = dup_sack_in;
1615 if (skb == tcp_send_head(sk))
1618 /* queue is in-order => we can short-circuit the walk early */
1619 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1622 if ((next_dup != NULL) &&
1623 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1624 in_sack = tcp_match_skb_to_sack(sk, skb,
1625 next_dup->start_seq,
1631 /* skb reference here is a bit tricky to get right, since
1632 * shifting can eat and free both this skb and the next,
1633 * so not even _safe variant of the loop is enough.
1636 tmp = tcp_shift_skb_data(sk, skb, state,
1637 start_seq, end_seq, dup_sack);
1646 in_sack = tcp_match_skb_to_sack(sk, skb,
1652 if (unlikely(in_sack < 0))
1656 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1659 tcp_skb_pcount(skb));
1661 if (!before(TCP_SKB_CB(skb)->seq,
1662 tcp_highest_sack_seq(tp)))
1663 tcp_advance_highest_sack(sk, skb);
1666 state->fack_count += tcp_skb_pcount(skb);
1671 /* Avoid all extra work that is being done by sacktag while walking in
1674 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1675 struct tcp_sacktag_state *state,
1678 tcp_for_write_queue_from(skb, sk) {
1679 if (skb == tcp_send_head(sk))
1682 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1685 state->fack_count += tcp_skb_pcount(skb);
1690 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1692 struct tcp_sack_block *next_dup,
1693 struct tcp_sacktag_state *state,
1696 if (next_dup == NULL)
1699 if (before(next_dup->start_seq, skip_to_seq)) {
1700 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1701 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1702 next_dup->start_seq, next_dup->end_seq,
1709 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1711 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1715 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1718 const struct inet_connection_sock *icsk = inet_csk(sk);
1719 struct tcp_sock *tp = tcp_sk(sk);
1720 unsigned char *ptr = (skb_transport_header(ack_skb) +
1721 TCP_SKB_CB(ack_skb)->sacked);
1722 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1723 struct tcp_sack_block sp[TCP_NUM_SACKS];
1724 struct tcp_sack_block *cache;
1725 struct tcp_sacktag_state state;
1726 struct sk_buff *skb;
1727 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1729 int found_dup_sack = 0;
1731 int first_sack_index;
1734 state.reord = tp->packets_out;
1736 if (!tp->sacked_out) {
1737 if (WARN_ON(tp->fackets_out))
1738 tp->fackets_out = 0;
1739 tcp_highest_sack_reset(sk);
1742 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1743 num_sacks, prior_snd_una);
1745 state.flag |= FLAG_DSACKING_ACK;
1747 /* Eliminate too old ACKs, but take into
1748 * account more or less fresh ones, they can
1749 * contain valid SACK info.
1751 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1754 if (!tp->packets_out)
1758 first_sack_index = 0;
1759 for (i = 0; i < num_sacks; i++) {
1760 int dup_sack = !i && found_dup_sack;
1762 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1763 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1765 if (!tcp_is_sackblock_valid(tp, dup_sack,
1766 sp[used_sacks].start_seq,
1767 sp[used_sacks].end_seq)) {
1771 if (!tp->undo_marker)
1772 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1774 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1776 /* Don't count olds caused by ACK reordering */
1777 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1778 !after(sp[used_sacks].end_seq, tp->snd_una))
1780 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1783 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1785 first_sack_index = -1;
1789 /* Ignore very old stuff early */
1790 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1796 /* order SACK blocks to allow in order walk of the retrans queue */
1797 for (i = used_sacks - 1; i > 0; i--) {
1798 for (j = 0; j < i; j++) {
1799 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1800 struct tcp_sack_block tmp;
1806 /* Track where the first SACK block goes to */
1807 if (j == first_sack_index)
1808 first_sack_index = j + 1;
1813 skb = tcp_write_queue_head(sk);
1814 state.fack_count = 0;
1817 if (!tp->sacked_out) {
1818 /* It's already past, so skip checking against it */
1819 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1821 cache = tp->recv_sack_cache;
1822 /* Skip empty blocks in at head of the cache */
1823 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1828 while (i < used_sacks) {
1829 u32 start_seq = sp[i].start_seq;
1830 u32 end_seq = sp[i].end_seq;
1831 int dup_sack = (found_dup_sack && (i == first_sack_index));
1832 struct tcp_sack_block *next_dup = NULL;
1834 if (found_dup_sack && ((i + 1) == first_sack_index))
1835 next_dup = &sp[i + 1];
1837 /* Event "B" in the comment above. */
1838 if (after(end_seq, tp->high_seq))
1839 state.flag |= FLAG_DATA_LOST;
1841 /* Skip too early cached blocks */
1842 while (tcp_sack_cache_ok(tp, cache) &&
1843 !before(start_seq, cache->end_seq))
1846 /* Can skip some work by looking recv_sack_cache? */
1847 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1848 after(end_seq, cache->start_seq)) {
1851 if (before(start_seq, cache->start_seq)) {
1852 skb = tcp_sacktag_skip(skb, sk, &state,
1854 skb = tcp_sacktag_walk(skb, sk, next_dup,
1861 /* Rest of the block already fully processed? */
1862 if (!after(end_seq, cache->end_seq))
1865 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1869 /* ...tail remains todo... */
1870 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1871 /* ...but better entrypoint exists! */
1872 skb = tcp_highest_sack(sk);
1875 state.fack_count = tp->fackets_out;
1880 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1881 /* Check overlap against next cached too (past this one already) */
1886 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1887 skb = tcp_highest_sack(sk);
1890 state.fack_count = tp->fackets_out;
1892 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1895 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1896 start_seq, end_seq, dup_sack);
1899 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1900 * due to in-order walk
1902 if (after(end_seq, tp->frto_highmark))
1903 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1908 /* Clear the head of the cache sack blocks so we can skip it next time */
1909 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1910 tp->recv_sack_cache[i].start_seq = 0;
1911 tp->recv_sack_cache[i].end_seq = 0;
1913 for (j = 0; j < used_sacks; j++)
1914 tp->recv_sack_cache[i++] = sp[j];
1916 tcp_mark_lost_retrans(sk);
1918 tcp_verify_left_out(tp);
1920 if ((state.reord < tp->fackets_out) &&
1921 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1922 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1923 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1927 #if FASTRETRANS_DEBUG > 0
1928 WARN_ON((int)tp->sacked_out < 0);
1929 WARN_ON((int)tp->lost_out < 0);
1930 WARN_ON((int)tp->retrans_out < 0);
1931 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1936 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1937 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1939 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1943 holes = max(tp->lost_out, 1U);
1944 holes = min(holes, tp->packets_out);
1946 if ((tp->sacked_out + holes) > tp->packets_out) {
1947 tp->sacked_out = tp->packets_out - holes;
1953 /* If we receive more dupacks than we expected counting segments
1954 * in assumption of absent reordering, interpret this as reordering.
1955 * The only another reason could be bug in receiver TCP.
1957 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1959 struct tcp_sock *tp = tcp_sk(sk);
1960 if (tcp_limit_reno_sacked(tp))
1961 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1964 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1966 static void tcp_add_reno_sack(struct sock *sk)
1968 struct tcp_sock *tp = tcp_sk(sk);
1970 tcp_check_reno_reordering(sk, 0);
1971 tcp_verify_left_out(tp);
1974 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1976 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1978 struct tcp_sock *tp = tcp_sk(sk);
1981 /* One ACK acked hole. The rest eat duplicate ACKs. */
1982 if (acked - 1 >= tp->sacked_out)
1985 tp->sacked_out -= acked - 1;
1987 tcp_check_reno_reordering(sk, acked);
1988 tcp_verify_left_out(tp);
1991 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1996 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1998 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2001 /* F-RTO can only be used if TCP has never retransmitted anything other than
2002 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2004 int tcp_use_frto(struct sock *sk)
2006 const struct tcp_sock *tp = tcp_sk(sk);
2007 const struct inet_connection_sock *icsk = inet_csk(sk);
2008 struct sk_buff *skb;
2010 if (!sysctl_tcp_frto)
2013 /* MTU probe and F-RTO won't really play nicely along currently */
2014 if (icsk->icsk_mtup.probe_size)
2017 if (tcp_is_sackfrto(tp))
2020 /* Avoid expensive walking of rexmit queue if possible */
2021 if (tp->retrans_out > 1)
2024 skb = tcp_write_queue_head(sk);
2025 if (tcp_skb_is_last(sk, skb))
2027 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2028 tcp_for_write_queue_from(skb, sk) {
2029 if (skb == tcp_send_head(sk))
2031 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2033 /* Short-circuit when first non-SACKed skb has been checked */
2034 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2040 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2041 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2042 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2043 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2044 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2045 * bits are handled if the Loss state is really to be entered (in
2046 * tcp_enter_frto_loss).
2048 * Do like tcp_enter_loss() would; when RTO expires the second time it
2050 * "Reduce ssthresh if it has not yet been made inside this window."
2052 void tcp_enter_frto(struct sock *sk)
2054 const struct inet_connection_sock *icsk = inet_csk(sk);
2055 struct tcp_sock *tp = tcp_sk(sk);
2056 struct sk_buff *skb;
2058 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2059 tp->snd_una == tp->high_seq ||
2060 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2061 !icsk->icsk_retransmits)) {
2062 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2063 /* Our state is too optimistic in ssthresh() call because cwnd
2064 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2065 * recovery has not yet completed. Pattern would be this: RTO,
2066 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2068 * RFC4138 should be more specific on what to do, even though
2069 * RTO is quite unlikely to occur after the first Cumulative ACK
2070 * due to back-off and complexity of triggering events ...
2072 if (tp->frto_counter) {
2074 stored_cwnd = tp->snd_cwnd;
2076 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2077 tp->snd_cwnd = stored_cwnd;
2079 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2081 /* ... in theory, cong.control module could do "any tricks" in
2082 * ssthresh(), which means that ca_state, lost bits and lost_out
2083 * counter would have to be faked before the call occurs. We
2084 * consider that too expensive, unlikely and hacky, so modules
2085 * using these in ssthresh() must deal these incompatibility
2086 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2088 tcp_ca_event(sk, CA_EVENT_FRTO);
2091 tp->undo_marker = tp->snd_una;
2092 tp->undo_retrans = 0;
2094 skb = tcp_write_queue_head(sk);
2095 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2096 tp->undo_marker = 0;
2097 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2098 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2099 tp->retrans_out -= tcp_skb_pcount(skb);
2101 tcp_verify_left_out(tp);
2103 /* Too bad if TCP was application limited */
2104 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2106 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2107 * The last condition is necessary at least in tp->frto_counter case.
2109 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2110 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2111 after(tp->high_seq, tp->snd_una)) {
2112 tp->frto_highmark = tp->high_seq;
2114 tp->frto_highmark = tp->snd_nxt;
2116 tcp_set_ca_state(sk, TCP_CA_Disorder);
2117 tp->high_seq = tp->snd_nxt;
2118 tp->frto_counter = 1;
2121 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2122 * which indicates that we should follow the traditional RTO recovery,
2123 * i.e. mark everything lost and do go-back-N retransmission.
2125 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2127 struct tcp_sock *tp = tcp_sk(sk);
2128 struct sk_buff *skb;
2131 tp->retrans_out = 0;
2132 if (tcp_is_reno(tp))
2133 tcp_reset_reno_sack(tp);
2135 tcp_for_write_queue(skb, sk) {
2136 if (skb == tcp_send_head(sk))
2139 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2141 * Count the retransmission made on RTO correctly (only when
2142 * waiting for the first ACK and did not get it)...
2144 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2145 /* For some reason this R-bit might get cleared? */
2146 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2147 tp->retrans_out += tcp_skb_pcount(skb);
2148 /* ...enter this if branch just for the first segment */
2149 flag |= FLAG_DATA_ACKED;
2151 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2152 tp->undo_marker = 0;
2153 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2156 /* Marking forward transmissions that were made after RTO lost
2157 * can cause unnecessary retransmissions in some scenarios,
2158 * SACK blocks will mitigate that in some but not in all cases.
2159 * We used to not mark them but it was causing break-ups with
2160 * receivers that do only in-order receival.
2162 * TODO: we could detect presence of such receiver and select
2163 * different behavior per flow.
2165 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2166 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2167 tp->lost_out += tcp_skb_pcount(skb);
2168 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2171 tcp_verify_left_out(tp);
2173 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2174 tp->snd_cwnd_cnt = 0;
2175 tp->snd_cwnd_stamp = tcp_time_stamp;
2176 tp->frto_counter = 0;
2177 tp->bytes_acked = 0;
2179 tp->reordering = min_t(unsigned int, tp->reordering,
2180 sysctl_tcp_reordering);
2181 tcp_set_ca_state(sk, TCP_CA_Loss);
2182 tp->high_seq = tp->snd_nxt;
2183 TCP_ECN_queue_cwr(tp);
2185 tcp_clear_all_retrans_hints(tp);
2188 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2190 tp->retrans_out = 0;
2193 tp->undo_marker = 0;
2194 tp->undo_retrans = 0;
2197 void tcp_clear_retrans(struct tcp_sock *tp)
2199 tcp_clear_retrans_partial(tp);
2201 tp->fackets_out = 0;
2205 /* Enter Loss state. If "how" is not zero, forget all SACK information
2206 * and reset tags completely, otherwise preserve SACKs. If receiver
2207 * dropped its ofo queue, we will know this due to reneging detection.
2209 void tcp_enter_loss(struct sock *sk, int how)
2211 const struct inet_connection_sock *icsk = inet_csk(sk);
2212 struct tcp_sock *tp = tcp_sk(sk);
2213 struct sk_buff *skb;
2215 /* Reduce ssthresh if it has not yet been made inside this window. */
2216 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2217 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2218 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2219 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2220 tcp_ca_event(sk, CA_EVENT_LOSS);
2223 tp->snd_cwnd_cnt = 0;
2224 tp->snd_cwnd_stamp = tcp_time_stamp;
2226 tp->bytes_acked = 0;
2227 tcp_clear_retrans_partial(tp);
2229 if (tcp_is_reno(tp))
2230 tcp_reset_reno_sack(tp);
2233 /* Push undo marker, if it was plain RTO and nothing
2234 * was retransmitted. */
2235 tp->undo_marker = tp->snd_una;
2238 tp->fackets_out = 0;
2240 tcp_clear_all_retrans_hints(tp);
2242 tcp_for_write_queue(skb, sk) {
2243 if (skb == tcp_send_head(sk))
2246 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2247 tp->undo_marker = 0;
2248 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2249 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2250 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2251 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2252 tp->lost_out += tcp_skb_pcount(skb);
2253 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2256 tcp_verify_left_out(tp);
2258 tp->reordering = min_t(unsigned int, tp->reordering,
2259 sysctl_tcp_reordering);
2260 tcp_set_ca_state(sk, TCP_CA_Loss);
2261 tp->high_seq = tp->snd_nxt;
2262 TCP_ECN_queue_cwr(tp);
2263 /* Abort F-RTO algorithm if one is in progress */
2264 tp->frto_counter = 0;
2267 /* If ACK arrived pointing to a remembered SACK, it means that our
2268 * remembered SACKs do not reflect real state of receiver i.e.
2269 * receiver _host_ is heavily congested (or buggy).
2271 * Do processing similar to RTO timeout.
2273 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2275 if (flag & FLAG_SACK_RENEGING) {
2276 struct inet_connection_sock *icsk = inet_csk(sk);
2277 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2279 tcp_enter_loss(sk, 1);
2280 icsk->icsk_retransmits++;
2281 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2282 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2283 icsk->icsk_rto, TCP_RTO_MAX);
2289 static inline int tcp_fackets_out(struct tcp_sock *tp)
2291 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2294 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2295 * counter when SACK is enabled (without SACK, sacked_out is used for
2298 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2299 * segments up to the highest received SACK block so far and holes in
2302 * With reordering, holes may still be in flight, so RFC3517 recovery
2303 * uses pure sacked_out (total number of SACKed segments) even though
2304 * it violates the RFC that uses duplicate ACKs, often these are equal
2305 * but when e.g. out-of-window ACKs or packet duplication occurs,
2306 * they differ. Since neither occurs due to loss, TCP should really
2309 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2311 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2314 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2316 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2319 static inline int tcp_head_timedout(struct sock *sk)
2321 struct tcp_sock *tp = tcp_sk(sk);
2323 return tp->packets_out &&
2324 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2327 /* Linux NewReno/SACK/FACK/ECN state machine.
2328 * --------------------------------------
2330 * "Open" Normal state, no dubious events, fast path.
2331 * "Disorder" In all the respects it is "Open",
2332 * but requires a bit more attention. It is entered when
2333 * we see some SACKs or dupacks. It is split of "Open"
2334 * mainly to move some processing from fast path to slow one.
2335 * "CWR" CWND was reduced due to some Congestion Notification event.
2336 * It can be ECN, ICMP source quench, local device congestion.
2337 * "Recovery" CWND was reduced, we are fast-retransmitting.
2338 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2340 * tcp_fastretrans_alert() is entered:
2341 * - each incoming ACK, if state is not "Open"
2342 * - when arrived ACK is unusual, namely:
2347 * Counting packets in flight is pretty simple.
2349 * in_flight = packets_out - left_out + retrans_out
2351 * packets_out is SND.NXT-SND.UNA counted in packets.
2353 * retrans_out is number of retransmitted segments.
2355 * left_out is number of segments left network, but not ACKed yet.
2357 * left_out = sacked_out + lost_out
2359 * sacked_out: Packets, which arrived to receiver out of order
2360 * and hence not ACKed. With SACKs this number is simply
2361 * amount of SACKed data. Even without SACKs
2362 * it is easy to give pretty reliable estimate of this number,
2363 * counting duplicate ACKs.
2365 * lost_out: Packets lost by network. TCP has no explicit
2366 * "loss notification" feedback from network (for now).
2367 * It means that this number can be only _guessed_.
2368 * Actually, it is the heuristics to predict lossage that
2369 * distinguishes different algorithms.
2371 * F.e. after RTO, when all the queue is considered as lost,
2372 * lost_out = packets_out and in_flight = retrans_out.
2374 * Essentially, we have now two algorithms counting
2377 * FACK: It is the simplest heuristics. As soon as we decided
2378 * that something is lost, we decide that _all_ not SACKed
2379 * packets until the most forward SACK are lost. I.e.
2380 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2381 * It is absolutely correct estimate, if network does not reorder
2382 * packets. And it loses any connection to reality when reordering
2383 * takes place. We use FACK by default until reordering
2384 * is suspected on the path to this destination.
2386 * NewReno: when Recovery is entered, we assume that one segment
2387 * is lost (classic Reno). While we are in Recovery and
2388 * a partial ACK arrives, we assume that one more packet
2389 * is lost (NewReno). This heuristics are the same in NewReno
2392 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2393 * deflation etc. CWND is real congestion window, never inflated, changes
2394 * only according to classic VJ rules.
2396 * Really tricky (and requiring careful tuning) part of algorithm
2397 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2398 * The first determines the moment _when_ we should reduce CWND and,
2399 * hence, slow down forward transmission. In fact, it determines the moment
2400 * when we decide that hole is caused by loss, rather than by a reorder.
2402 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2403 * holes, caused by lost packets.
2405 * And the most logically complicated part of algorithm is undo
2406 * heuristics. We detect false retransmits due to both too early
2407 * fast retransmit (reordering) and underestimated RTO, analyzing
2408 * timestamps and D-SACKs. When we detect that some segments were
2409 * retransmitted by mistake and CWND reduction was wrong, we undo
2410 * window reduction and abort recovery phase. This logic is hidden
2411 * inside several functions named tcp_try_undo_<something>.
2414 /* This function decides, when we should leave Disordered state
2415 * and enter Recovery phase, reducing congestion window.
2417 * Main question: may we further continue forward transmission
2418 * with the same cwnd?
2420 static int tcp_time_to_recover(struct sock *sk)
2422 struct tcp_sock *tp = tcp_sk(sk);
2425 /* Do not perform any recovery during F-RTO algorithm */
2426 if (tp->frto_counter)
2429 /* Trick#1: The loss is proven. */
2433 /* Not-A-Trick#2 : Classic rule... */
2434 if (tcp_dupack_heurestics(tp) > tp->reordering)
2437 /* Trick#3 : when we use RFC2988 timer restart, fast
2438 * retransmit can be triggered by timeout of queue head.
2440 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2443 /* Trick#4: It is still not OK... But will it be useful to delay
2446 packets_out = tp->packets_out;
2447 if (packets_out <= tp->reordering &&
2448 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2449 !tcp_may_send_now(sk)) {
2450 /* We have nothing to send. This connection is limited
2451 * either by receiver window or by application.
2459 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2460 * is against sacked "cnt", otherwise it's against facked "cnt"
2462 static void tcp_mark_head_lost(struct sock *sk, int packets)
2464 struct tcp_sock *tp = tcp_sk(sk);
2465 struct sk_buff *skb;
2470 WARN_ON(packets > tp->packets_out);
2471 if (tp->lost_skb_hint) {
2472 skb = tp->lost_skb_hint;
2473 cnt = tp->lost_cnt_hint;
2475 skb = tcp_write_queue_head(sk);
2479 tcp_for_write_queue_from(skb, sk) {
2480 if (skb == tcp_send_head(sk))
2482 /* TODO: do this better */
2483 /* this is not the most efficient way to do this... */
2484 tp->lost_skb_hint = skb;
2485 tp->lost_cnt_hint = cnt;
2487 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2491 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2492 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2493 cnt += tcp_skb_pcount(skb);
2495 if (cnt > packets) {
2496 if (tcp_is_sack(tp) || (oldcnt >= packets))
2499 mss = skb_shinfo(skb)->gso_size;
2500 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2506 tcp_skb_mark_lost(tp, skb);
2508 tcp_verify_left_out(tp);
2511 /* Account newly detected lost packet(s) */
2513 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2515 struct tcp_sock *tp = tcp_sk(sk);
2517 if (tcp_is_reno(tp)) {
2518 tcp_mark_head_lost(sk, 1);
2519 } else if (tcp_is_fack(tp)) {
2520 int lost = tp->fackets_out - tp->reordering;
2523 tcp_mark_head_lost(sk, lost);
2525 int sacked_upto = tp->sacked_out - tp->reordering;
2526 if (sacked_upto < fast_rexmit)
2527 sacked_upto = fast_rexmit;
2528 tcp_mark_head_lost(sk, sacked_upto);
2531 /* New heuristics: it is possible only after we switched
2532 * to restart timer each time when something is ACKed.
2533 * Hence, we can detect timed out packets during fast
2534 * retransmit without falling to slow start.
2536 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2537 struct sk_buff *skb;
2539 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2540 : tcp_write_queue_head(sk);
2542 tcp_for_write_queue_from(skb, sk) {
2543 if (skb == tcp_send_head(sk))
2545 if (!tcp_skb_timedout(sk, skb))
2548 tcp_skb_mark_lost(tp, skb);
2551 tp->scoreboard_skb_hint = skb;
2553 tcp_verify_left_out(tp);
2557 /* CWND moderation, preventing bursts due to too big ACKs
2558 * in dubious situations.
2560 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2562 tp->snd_cwnd = min(tp->snd_cwnd,
2563 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2564 tp->snd_cwnd_stamp = tcp_time_stamp;
2567 /* Lower bound on congestion window is slow start threshold
2568 * unless congestion avoidance choice decides to overide it.
2570 static inline u32 tcp_cwnd_min(const struct sock *sk)
2572 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2574 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2577 /* Decrease cwnd each second ack. */
2578 static void tcp_cwnd_down(struct sock *sk, int flag)
2580 struct tcp_sock *tp = tcp_sk(sk);
2581 int decr = tp->snd_cwnd_cnt + 1;
2583 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2584 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2585 tp->snd_cwnd_cnt = decr & 1;
2588 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2589 tp->snd_cwnd -= decr;
2591 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2592 tp->snd_cwnd_stamp = tcp_time_stamp;
2596 /* Nothing was retransmitted or returned timestamp is less
2597 * than timestamp of the first retransmission.
2599 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2601 return !tp->retrans_stamp ||
2602 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2603 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2606 /* Undo procedures. */
2608 #if FASTRETRANS_DEBUG > 1
2609 static void DBGUNDO(struct sock *sk, const char *msg)
2611 struct tcp_sock *tp = tcp_sk(sk);
2612 struct inet_sock *inet = inet_sk(sk);
2614 if (sk->sk_family == AF_INET) {
2615 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2617 &inet->daddr, ntohs(inet->dport),
2618 tp->snd_cwnd, tcp_left_out(tp),
2619 tp->snd_ssthresh, tp->prior_ssthresh,
2622 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2623 else if (sk->sk_family == AF_INET6) {
2624 struct ipv6_pinfo *np = inet6_sk(sk);
2625 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2627 &np->daddr, ntohs(inet->dport),
2628 tp->snd_cwnd, tcp_left_out(tp),
2629 tp->snd_ssthresh, tp->prior_ssthresh,
2635 #define DBGUNDO(x...) do { } while (0)
2638 static void tcp_undo_cwr(struct sock *sk, const int undo)
2640 struct tcp_sock *tp = tcp_sk(sk);
2642 if (tp->prior_ssthresh) {
2643 const struct inet_connection_sock *icsk = inet_csk(sk);
2645 if (icsk->icsk_ca_ops->undo_cwnd)
2646 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2648 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2650 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2651 tp->snd_ssthresh = tp->prior_ssthresh;
2652 TCP_ECN_withdraw_cwr(tp);
2655 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2657 tcp_moderate_cwnd(tp);
2658 tp->snd_cwnd_stamp = tcp_time_stamp;
2661 static inline int tcp_may_undo(struct tcp_sock *tp)
2663 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2666 /* People celebrate: "We love our President!" */
2667 static int tcp_try_undo_recovery(struct sock *sk)
2669 struct tcp_sock *tp = tcp_sk(sk);
2671 if (tcp_may_undo(tp)) {
2674 /* Happy end! We did not retransmit anything
2675 * or our original transmission succeeded.
2677 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2678 tcp_undo_cwr(sk, 1);
2679 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2680 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2682 mib_idx = LINUX_MIB_TCPFULLUNDO;
2684 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2685 tp->undo_marker = 0;
2687 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2688 /* Hold old state until something *above* high_seq
2689 * is ACKed. For Reno it is MUST to prevent false
2690 * fast retransmits (RFC2582). SACK TCP is safe. */
2691 tcp_moderate_cwnd(tp);
2694 tcp_set_ca_state(sk, TCP_CA_Open);
2698 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2699 static void tcp_try_undo_dsack(struct sock *sk)
2701 struct tcp_sock *tp = tcp_sk(sk);
2703 if (tp->undo_marker && !tp->undo_retrans) {
2704 DBGUNDO(sk, "D-SACK");
2705 tcp_undo_cwr(sk, 1);
2706 tp->undo_marker = 0;
2707 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2711 /* Undo during fast recovery after partial ACK. */
2713 static int tcp_try_undo_partial(struct sock *sk, int acked)
2715 struct tcp_sock *tp = tcp_sk(sk);
2716 /* Partial ACK arrived. Force Hoe's retransmit. */
2717 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2719 if (tcp_may_undo(tp)) {
2720 /* Plain luck! Hole if filled with delayed
2721 * packet, rather than with a retransmit.
2723 if (tp->retrans_out == 0)
2724 tp->retrans_stamp = 0;
2726 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2729 tcp_undo_cwr(sk, 0);
2730 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2732 /* So... Do not make Hoe's retransmit yet.
2733 * If the first packet was delayed, the rest
2734 * ones are most probably delayed as well.
2741 /* Undo during loss recovery after partial ACK. */
2742 static int tcp_try_undo_loss(struct sock *sk)
2744 struct tcp_sock *tp = tcp_sk(sk);
2746 if (tcp_may_undo(tp)) {
2747 struct sk_buff *skb;
2748 tcp_for_write_queue(skb, sk) {
2749 if (skb == tcp_send_head(sk))
2751 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2754 tcp_clear_all_retrans_hints(tp);
2756 DBGUNDO(sk, "partial loss");
2758 tcp_undo_cwr(sk, 1);
2759 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2760 inet_csk(sk)->icsk_retransmits = 0;
2761 tp->undo_marker = 0;
2762 if (tcp_is_sack(tp))
2763 tcp_set_ca_state(sk, TCP_CA_Open);
2769 static inline void tcp_complete_cwr(struct sock *sk)
2771 struct tcp_sock *tp = tcp_sk(sk);
2772 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2773 tp->snd_cwnd_stamp = tcp_time_stamp;
2774 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2777 static void tcp_try_keep_open(struct sock *sk)
2779 struct tcp_sock *tp = tcp_sk(sk);
2780 int state = TCP_CA_Open;
2782 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2783 state = TCP_CA_Disorder;
2785 if (inet_csk(sk)->icsk_ca_state != state) {
2786 tcp_set_ca_state(sk, state);
2787 tp->high_seq = tp->snd_nxt;
2791 static void tcp_try_to_open(struct sock *sk, int flag)
2793 struct tcp_sock *tp = tcp_sk(sk);
2795 tcp_verify_left_out(tp);
2797 if (!tp->frto_counter && tp->retrans_out == 0)
2798 tp->retrans_stamp = 0;
2800 if (flag & FLAG_ECE)
2801 tcp_enter_cwr(sk, 1);
2803 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2804 tcp_try_keep_open(sk);
2805 tcp_moderate_cwnd(tp);
2807 tcp_cwnd_down(sk, flag);
2811 static void tcp_mtup_probe_failed(struct sock *sk)
2813 struct inet_connection_sock *icsk = inet_csk(sk);
2815 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2816 icsk->icsk_mtup.probe_size = 0;
2819 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2821 struct tcp_sock *tp = tcp_sk(sk);
2822 struct inet_connection_sock *icsk = inet_csk(sk);
2824 /* FIXME: breaks with very large cwnd */
2825 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2826 tp->snd_cwnd = tp->snd_cwnd *
2827 tcp_mss_to_mtu(sk, tp->mss_cache) /
2828 icsk->icsk_mtup.probe_size;
2829 tp->snd_cwnd_cnt = 0;
2830 tp->snd_cwnd_stamp = tcp_time_stamp;
2831 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2833 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2834 icsk->icsk_mtup.probe_size = 0;
2835 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2838 /* Do a simple retransmit without using the backoff mechanisms in
2839 * tcp_timer. This is used for path mtu discovery.
2840 * The socket is already locked here.
2842 void tcp_simple_retransmit(struct sock *sk)
2844 const struct inet_connection_sock *icsk = inet_csk(sk);
2845 struct tcp_sock *tp = tcp_sk(sk);
2846 struct sk_buff *skb;
2847 unsigned int mss = tcp_current_mss(sk, 0);
2848 u32 prior_lost = tp->lost_out;
2850 tcp_for_write_queue(skb, sk) {
2851 if (skb == tcp_send_head(sk))
2853 if (tcp_skb_seglen(skb) > mss &&
2854 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2855 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2856 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2857 tp->retrans_out -= tcp_skb_pcount(skb);
2859 tcp_skb_mark_lost_uncond_verify(tp, skb);
2863 tcp_clear_retrans_hints_partial(tp);
2865 if (prior_lost == tp->lost_out)
2868 if (tcp_is_reno(tp))
2869 tcp_limit_reno_sacked(tp);
2871 tcp_verify_left_out(tp);
2873 /* Don't muck with the congestion window here.
2874 * Reason is that we do not increase amount of _data_
2875 * in network, but units changed and effective
2876 * cwnd/ssthresh really reduced now.
2878 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2879 tp->high_seq = tp->snd_nxt;
2880 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2881 tp->prior_ssthresh = 0;
2882 tp->undo_marker = 0;
2883 tcp_set_ca_state(sk, TCP_CA_Loss);
2885 tcp_xmit_retransmit_queue(sk);
2888 /* Process an event, which can update packets-in-flight not trivially.
2889 * Main goal of this function is to calculate new estimate for left_out,
2890 * taking into account both packets sitting in receiver's buffer and
2891 * packets lost by network.
2893 * Besides that it does CWND reduction, when packet loss is detected
2894 * and changes state of machine.
2896 * It does _not_ decide what to send, it is made in function
2897 * tcp_xmit_retransmit_queue().
2899 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2901 struct inet_connection_sock *icsk = inet_csk(sk);
2902 struct tcp_sock *tp = tcp_sk(sk);
2903 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2904 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2905 (tcp_fackets_out(tp) > tp->reordering));
2906 int fast_rexmit = 0, mib_idx;
2908 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2910 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2911 tp->fackets_out = 0;
2913 /* Now state machine starts.
2914 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2915 if (flag & FLAG_ECE)
2916 tp->prior_ssthresh = 0;
2918 /* B. In all the states check for reneging SACKs. */
2919 if (tcp_check_sack_reneging(sk, flag))
2922 /* C. Process data loss notification, provided it is valid. */
2923 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2924 before(tp->snd_una, tp->high_seq) &&
2925 icsk->icsk_ca_state != TCP_CA_Open &&
2926 tp->fackets_out > tp->reordering) {
2927 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2928 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2931 /* D. Check consistency of the current state. */
2932 tcp_verify_left_out(tp);
2934 /* E. Check state exit conditions. State can be terminated
2935 * when high_seq is ACKed. */
2936 if (icsk->icsk_ca_state == TCP_CA_Open) {
2937 WARN_ON(tp->retrans_out != 0);
2938 tp->retrans_stamp = 0;
2939 } else if (!before(tp->snd_una, tp->high_seq)) {
2940 switch (icsk->icsk_ca_state) {
2942 icsk->icsk_retransmits = 0;
2943 if (tcp_try_undo_recovery(sk))
2948 /* CWR is to be held something *above* high_seq
2949 * is ACKed for CWR bit to reach receiver. */
2950 if (tp->snd_una != tp->high_seq) {
2951 tcp_complete_cwr(sk);
2952 tcp_set_ca_state(sk, TCP_CA_Open);
2956 case TCP_CA_Disorder:
2957 tcp_try_undo_dsack(sk);
2958 if (!tp->undo_marker ||
2959 /* For SACK case do not Open to allow to undo
2960 * catching for all duplicate ACKs. */
2961 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2962 tp->undo_marker = 0;
2963 tcp_set_ca_state(sk, TCP_CA_Open);
2967 case TCP_CA_Recovery:
2968 if (tcp_is_reno(tp))
2969 tcp_reset_reno_sack(tp);
2970 if (tcp_try_undo_recovery(sk))
2972 tcp_complete_cwr(sk);
2977 /* F. Process state. */
2978 switch (icsk->icsk_ca_state) {
2979 case TCP_CA_Recovery:
2980 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2981 if (tcp_is_reno(tp) && is_dupack)
2982 tcp_add_reno_sack(sk);
2984 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2987 if (flag & FLAG_DATA_ACKED)
2988 icsk->icsk_retransmits = 0;
2989 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2990 tcp_reset_reno_sack(tp);
2991 if (!tcp_try_undo_loss(sk)) {
2992 tcp_moderate_cwnd(tp);
2993 tcp_xmit_retransmit_queue(sk);
2996 if (icsk->icsk_ca_state != TCP_CA_Open)
2998 /* Loss is undone; fall through to processing in Open state. */
3000 if (tcp_is_reno(tp)) {
3001 if (flag & FLAG_SND_UNA_ADVANCED)
3002 tcp_reset_reno_sack(tp);
3004 tcp_add_reno_sack(sk);
3007 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3008 tcp_try_undo_dsack(sk);
3010 if (!tcp_time_to_recover(sk)) {
3011 tcp_try_to_open(sk, flag);
3015 /* MTU probe failure: don't reduce cwnd */
3016 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3017 icsk->icsk_mtup.probe_size &&
3018 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3019 tcp_mtup_probe_failed(sk);
3020 /* Restores the reduction we did in tcp_mtup_probe() */
3022 tcp_simple_retransmit(sk);
3026 /* Otherwise enter Recovery state */
3028 if (tcp_is_reno(tp))
3029 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3031 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3033 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3035 tp->high_seq = tp->snd_nxt;
3036 tp->prior_ssthresh = 0;
3037 tp->undo_marker = tp->snd_una;
3038 tp->undo_retrans = tp->retrans_out;
3040 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3041 if (!(flag & FLAG_ECE))
3042 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3043 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3044 TCP_ECN_queue_cwr(tp);
3047 tp->bytes_acked = 0;
3048 tp->snd_cwnd_cnt = 0;
3049 tcp_set_ca_state(sk, TCP_CA_Recovery);
3053 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3054 tcp_update_scoreboard(sk, fast_rexmit);
3055 tcp_cwnd_down(sk, flag);
3056 tcp_xmit_retransmit_queue(sk);
3059 /* Read draft-ietf-tcplw-high-performance before mucking
3060 * with this code. (Supersedes RFC1323)
3062 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3064 /* RTTM Rule: A TSecr value received in a segment is used to
3065 * update the averaged RTT measurement only if the segment
3066 * acknowledges some new data, i.e., only if it advances the
3067 * left edge of the send window.
3069 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3070 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3072 * Changed: reset backoff as soon as we see the first valid sample.
3073 * If we do not, we get strongly overestimated rto. With timestamps
3074 * samples are accepted even from very old segments: f.e., when rtt=1
3075 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3076 * answer arrives rto becomes 120 seconds! If at least one of segments
3077 * in window is lost... Voila. --ANK (010210)
3079 struct tcp_sock *tp = tcp_sk(sk);
3080 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
3081 tcp_rtt_estimator(sk, seq_rtt);
3083 inet_csk(sk)->icsk_backoff = 0;
3087 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3089 /* We don't have a timestamp. Can only use
3090 * packets that are not retransmitted to determine
3091 * rtt estimates. Also, we must not reset the
3092 * backoff for rto until we get a non-retransmitted
3093 * packet. This allows us to deal with a situation
3094 * where the network delay has increased suddenly.
3095 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3098 if (flag & FLAG_RETRANS_DATA_ACKED)
3101 tcp_rtt_estimator(sk, seq_rtt);
3103 inet_csk(sk)->icsk_backoff = 0;
3107 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3110 const struct tcp_sock *tp = tcp_sk(sk);
3111 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3112 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3113 tcp_ack_saw_tstamp(sk, flag);
3114 else if (seq_rtt >= 0)
3115 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3118 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3120 const struct inet_connection_sock *icsk = inet_csk(sk);
3121 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3122 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3125 /* Restart timer after forward progress on connection.
3126 * RFC2988 recommends to restart timer to now+rto.
3128 static void tcp_rearm_rto(struct sock *sk)
3130 struct tcp_sock *tp = tcp_sk(sk);
3132 if (!tp->packets_out) {
3133 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3135 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3136 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3140 /* If we get here, the whole TSO packet has not been acked. */
3141 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3143 struct tcp_sock *tp = tcp_sk(sk);
3146 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3148 packets_acked = tcp_skb_pcount(skb);
3149 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3151 packets_acked -= tcp_skb_pcount(skb);
3153 if (packets_acked) {
3154 BUG_ON(tcp_skb_pcount(skb) == 0);
3155 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3158 return packets_acked;
3161 /* Remove acknowledged frames from the retransmission queue. If our packet
3162 * is before the ack sequence we can discard it as it's confirmed to have
3163 * arrived at the other end.
3165 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3168 struct tcp_sock *tp = tcp_sk(sk);
3169 const struct inet_connection_sock *icsk = inet_csk(sk);
3170 struct sk_buff *skb;
3171 u32 now = tcp_time_stamp;
3172 int fully_acked = 1;
3175 u32 reord = tp->packets_out;
3176 u32 prior_sacked = tp->sacked_out;
3178 s32 ca_seq_rtt = -1;
3179 ktime_t last_ackt = net_invalid_timestamp();
3181 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3182 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3185 u8 sacked = scb->sacked;
3187 /* Determine how many packets and what bytes were acked, tso and else */
3188 if (after(scb->end_seq, tp->snd_una)) {
3189 if (tcp_skb_pcount(skb) == 1 ||
3190 !after(tp->snd_una, scb->seq))
3193 acked_pcount = tcp_tso_acked(sk, skb);
3198 end_seq = tp->snd_una;
3200 acked_pcount = tcp_skb_pcount(skb);
3201 end_seq = scb->end_seq;
3204 /* MTU probing checks */
3205 if (fully_acked && icsk->icsk_mtup.probe_size &&
3206 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
3207 tcp_mtup_probe_success(sk, skb);
3210 if (sacked & TCPCB_RETRANS) {
3211 if (sacked & TCPCB_SACKED_RETRANS)
3212 tp->retrans_out -= acked_pcount;
3213 flag |= FLAG_RETRANS_DATA_ACKED;
3216 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3217 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3219 ca_seq_rtt = now - scb->when;
3220 last_ackt = skb->tstamp;
3222 seq_rtt = ca_seq_rtt;
3224 if (!(sacked & TCPCB_SACKED_ACKED))
3225 reord = min(pkts_acked, reord);
3228 if (sacked & TCPCB_SACKED_ACKED)
3229 tp->sacked_out -= acked_pcount;
3230 if (sacked & TCPCB_LOST)
3231 tp->lost_out -= acked_pcount;
3233 tp->packets_out -= acked_pcount;
3234 pkts_acked += acked_pcount;
3236 /* Initial outgoing SYN's get put onto the write_queue
3237 * just like anything else we transmit. It is not
3238 * true data, and if we misinform our callers that
3239 * this ACK acks real data, we will erroneously exit
3240 * connection startup slow start one packet too
3241 * quickly. This is severely frowned upon behavior.
3243 if (!(scb->flags & TCPCB_FLAG_SYN)) {
3244 flag |= FLAG_DATA_ACKED;
3246 flag |= FLAG_SYN_ACKED;
3247 tp->retrans_stamp = 0;
3253 tcp_unlink_write_queue(skb, sk);
3254 sk_wmem_free_skb(sk, skb);
3255 tp->scoreboard_skb_hint = NULL;
3256 if (skb == tp->retransmit_skb_hint)
3257 tp->retransmit_skb_hint = NULL;
3258 if (skb == tp->lost_skb_hint)
3259 tp->lost_skb_hint = NULL;
3262 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3263 tp->snd_up = tp->snd_una;
3265 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3266 flag |= FLAG_SACK_RENEGING;
3268 if (flag & FLAG_ACKED) {
3269 const struct tcp_congestion_ops *ca_ops
3270 = inet_csk(sk)->icsk_ca_ops;
3272 tcp_ack_update_rtt(sk, flag, seq_rtt);
3275 if (tcp_is_reno(tp)) {
3276 tcp_remove_reno_sacks(sk, pkts_acked);
3278 /* Non-retransmitted hole got filled? That's reordering */
3279 if (reord < prior_fackets)
3280 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3282 /* No need to care for underflows here because
3283 * the lost_skb_hint gets NULLed if we're past it
3284 * (or something non-trivial happened)
3286 if (tcp_is_fack(tp))
3287 tp->lost_cnt_hint -= pkts_acked;
3289 tp->lost_cnt_hint -= prior_sacked - tp->sacked_out;
3292 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3294 if (ca_ops->pkts_acked) {
3297 /* Is the ACK triggering packet unambiguous? */
3298 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3299 /* High resolution needed and available? */
3300 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3301 !ktime_equal(last_ackt,
3302 net_invalid_timestamp()))
3303 rtt_us = ktime_us_delta(ktime_get_real(),
3305 else if (ca_seq_rtt > 0)
3306 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3309 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3313 #if FASTRETRANS_DEBUG > 0
3314 WARN_ON((int)tp->sacked_out < 0);
3315 WARN_ON((int)tp->lost_out < 0);
3316 WARN_ON((int)tp->retrans_out < 0);
3317 if (!tp->packets_out && tcp_is_sack(tp)) {
3318 icsk = inet_csk(sk);
3320 printk(KERN_DEBUG "Leak l=%u %d\n",
3321 tp->lost_out, icsk->icsk_ca_state);
3324 if (tp->sacked_out) {
3325 printk(KERN_DEBUG "Leak s=%u %d\n",
3326 tp->sacked_out, icsk->icsk_ca_state);
3329 if (tp->retrans_out) {
3330 printk(KERN_DEBUG "Leak r=%u %d\n",
3331 tp->retrans_out, icsk->icsk_ca_state);
3332 tp->retrans_out = 0;
3339 static void tcp_ack_probe(struct sock *sk)
3341 const struct tcp_sock *tp = tcp_sk(sk);
3342 struct inet_connection_sock *icsk = inet_csk(sk);
3344 /* Was it a usable window open? */
3346 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3347 icsk->icsk_backoff = 0;
3348 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3349 /* Socket must be waked up by subsequent tcp_data_snd_check().
3350 * This function is not for random using!
3353 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3354 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3359 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3361 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3362 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3365 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3367 const struct tcp_sock *tp = tcp_sk(sk);
3368 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3369 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3372 /* Check that window update is acceptable.
3373 * The function assumes that snd_una<=ack<=snd_next.
3375 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3376 const u32 ack, const u32 ack_seq,
3379 return (after(ack, tp->snd_una) ||
3380 after(ack_seq, tp->snd_wl1) ||
3381 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3384 /* Update our send window.
3386 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3387 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3389 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3392 struct tcp_sock *tp = tcp_sk(sk);
3394 u32 nwin = ntohs(tcp_hdr(skb)->window);
3396 if (likely(!tcp_hdr(skb)->syn))
3397 nwin <<= tp->rx_opt.snd_wscale;
3399 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3400 flag |= FLAG_WIN_UPDATE;
3401 tcp_update_wl(tp, ack, ack_seq);
3403 if (tp->snd_wnd != nwin) {
3406 /* Note, it is the only place, where
3407 * fast path is recovered for sending TCP.
3410 tcp_fast_path_check(sk);
3412 if (nwin > tp->max_window) {
3413 tp->max_window = nwin;
3414 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3424 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3425 * continue in congestion avoidance.
3427 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3429 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3430 tp->snd_cwnd_cnt = 0;
3431 tp->bytes_acked = 0;
3432 TCP_ECN_queue_cwr(tp);
3433 tcp_moderate_cwnd(tp);
3436 /* A conservative spurious RTO response algorithm: reduce cwnd using
3437 * rate halving and continue in congestion avoidance.
3439 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3441 tcp_enter_cwr(sk, 0);
3444 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3446 if (flag & FLAG_ECE)
3447 tcp_ratehalving_spur_to_response(sk);
3449 tcp_undo_cwr(sk, 1);
3452 /* F-RTO spurious RTO detection algorithm (RFC4138)
3454 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3455 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3456 * window (but not to or beyond highest sequence sent before RTO):
3457 * On First ACK, send two new segments out.
3458 * On Second ACK, RTO was likely spurious. Do spurious response (response
3459 * algorithm is not part of the F-RTO detection algorithm
3460 * given in RFC4138 but can be selected separately).
3461 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3462 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3463 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3464 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3466 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3467 * original window even after we transmit two new data segments.
3470 * on first step, wait until first cumulative ACK arrives, then move to
3471 * the second step. In second step, the next ACK decides.
3473 * F-RTO is implemented (mainly) in four functions:
3474 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3475 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3476 * called when tcp_use_frto() showed green light
3477 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3478 * - tcp_enter_frto_loss() is called if there is not enough evidence
3479 * to prove that the RTO is indeed spurious. It transfers the control
3480 * from F-RTO to the conventional RTO recovery
3482 static int tcp_process_frto(struct sock *sk, int flag)
3484 struct tcp_sock *tp = tcp_sk(sk);
3486 tcp_verify_left_out(tp);
3488 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3489 if (flag & FLAG_DATA_ACKED)
3490 inet_csk(sk)->icsk_retransmits = 0;
3492 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3493 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3494 tp->undo_marker = 0;
3496 if (!before(tp->snd_una, tp->frto_highmark)) {
3497 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3501 if (!tcp_is_sackfrto(tp)) {
3502 /* RFC4138 shortcoming in step 2; should also have case c):
3503 * ACK isn't duplicate nor advances window, e.g., opposite dir
3506 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3509 if (!(flag & FLAG_DATA_ACKED)) {
3510 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3515 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3516 /* Prevent sending of new data. */
3517 tp->snd_cwnd = min(tp->snd_cwnd,
3518 tcp_packets_in_flight(tp));
3522 if ((tp->frto_counter >= 2) &&
3523 (!(flag & FLAG_FORWARD_PROGRESS) ||
3524 ((flag & FLAG_DATA_SACKED) &&
3525 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3526 /* RFC4138 shortcoming (see comment above) */
3527 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3528 (flag & FLAG_NOT_DUP))
3531 tcp_enter_frto_loss(sk, 3, flag);
3536 if (tp->frto_counter == 1) {
3537 /* tcp_may_send_now needs to see updated state */
3538 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3539 tp->frto_counter = 2;
3541 if (!tcp_may_send_now(sk))
3542 tcp_enter_frto_loss(sk, 2, flag);
3546 switch (sysctl_tcp_frto_response) {
3548 tcp_undo_spur_to_response(sk, flag);
3551 tcp_conservative_spur_to_response(tp);
3554 tcp_ratehalving_spur_to_response(sk);
3557 tp->frto_counter = 0;
3558 tp->undo_marker = 0;
3559 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3564 /* This routine deals with incoming acks, but not outgoing ones. */
3565 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3567 struct inet_connection_sock *icsk = inet_csk(sk);
3568 struct tcp_sock *tp = tcp_sk(sk);
3569 u32 prior_snd_una = tp->snd_una;
3570 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3571 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3572 u32 prior_in_flight;
3577 /* If the ack is newer than sent or older than previous acks
3578 * then we can probably ignore it.
3580 if (after(ack, tp->snd_nxt))
3581 goto uninteresting_ack;
3583 if (before(ack, prior_snd_una))
3586 if (after(ack, prior_snd_una))
3587 flag |= FLAG_SND_UNA_ADVANCED;
3589 if (sysctl_tcp_abc) {
3590 if (icsk->icsk_ca_state < TCP_CA_CWR)
3591 tp->bytes_acked += ack - prior_snd_una;
3592 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3593 /* we assume just one segment left network */
3594 tp->bytes_acked += min(ack - prior_snd_una,
3598 prior_fackets = tp->fackets_out;
3599 prior_in_flight = tcp_packets_in_flight(tp);
3601 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3602 /* Window is constant, pure forward advance.
3603 * No more checks are required.
3604 * Note, we use the fact that SND.UNA>=SND.WL2.
3606 tcp_update_wl(tp, ack, ack_seq);
3608 flag |= FLAG_WIN_UPDATE;
3610 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3612 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3614 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3617 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3619 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3621 if (TCP_SKB_CB(skb)->sacked)
3622 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3624 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3627 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3630 /* We passed data and got it acked, remove any soft error
3631 * log. Something worked...
3633 sk->sk_err_soft = 0;
3634 icsk->icsk_probes_out = 0;
3635 tp->rcv_tstamp = tcp_time_stamp;
3636 prior_packets = tp->packets_out;
3640 /* See if we can take anything off of the retransmit queue. */
3641 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3643 if (tp->frto_counter)
3644 frto_cwnd = tcp_process_frto(sk, flag);
3645 /* Guarantee sacktag reordering detection against wrap-arounds */
3646 if (before(tp->frto_highmark, tp->snd_una))
3647 tp->frto_highmark = 0;
3649 if (tcp_ack_is_dubious(sk, flag)) {
3650 /* Advance CWND, if state allows this. */
3651 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3652 tcp_may_raise_cwnd(sk, flag))
3653 tcp_cong_avoid(sk, ack, prior_in_flight);
3654 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3657 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3658 tcp_cong_avoid(sk, ack, prior_in_flight);
3661 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3662 dst_confirm(sk->sk_dst_cache);
3667 /* If this ack opens up a zero window, clear backoff. It was
3668 * being used to time the probes, and is probably far higher than
3669 * it needs to be for normal retransmission.
3671 if (tcp_send_head(sk))
3676 if (TCP_SKB_CB(skb)->sacked) {
3677 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3678 if (icsk->icsk_ca_state == TCP_CA_Open)
3679 tcp_try_keep_open(sk);
3683 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3687 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3688 * But, this can also be called on packets in the established flow when
3689 * the fast version below fails.
3691 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3695 struct tcphdr *th = tcp_hdr(skb);
3696 int length = (th->doff * 4) - sizeof(struct tcphdr);
3698 ptr = (unsigned char *)(th + 1);
3699 opt_rx->saw_tstamp = 0;
3701 while (length > 0) {
3702 int opcode = *ptr++;
3708 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3713 if (opsize < 2) /* "silly options" */
3715 if (opsize > length)
3716 return; /* don't parse partial options */
3719 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3720 u16 in_mss = get_unaligned_be16(ptr);
3722 if (opt_rx->user_mss &&
3723 opt_rx->user_mss < in_mss)
3724 in_mss = opt_rx->user_mss;
3725 opt_rx->mss_clamp = in_mss;
3730 if (opsize == TCPOLEN_WINDOW && th->syn &&
3731 !estab && sysctl_tcp_window_scaling) {
3732 __u8 snd_wscale = *(__u8 *)ptr;
3733 opt_rx->wscale_ok = 1;
3734 if (snd_wscale > 14) {
3735 if (net_ratelimit())
3736 printk(KERN_INFO "tcp_parse_options: Illegal window "
3737 "scaling value %d >14 received.\n",
3741 opt_rx->snd_wscale = snd_wscale;
3744 case TCPOPT_TIMESTAMP:
3745 if ((opsize == TCPOLEN_TIMESTAMP) &&
3746 ((estab && opt_rx->tstamp_ok) ||
3747 (!estab && sysctl_tcp_timestamps))) {
3748 opt_rx->saw_tstamp = 1;
3749 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3750 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3753 case TCPOPT_SACK_PERM:
3754 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3755 !estab && sysctl_tcp_sack) {
3756 opt_rx->sack_ok = 1;
3757 tcp_sack_reset(opt_rx);
3762 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3763 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3765 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3768 #ifdef CONFIG_TCP_MD5SIG
3771 * The MD5 Hash has already been
3772 * checked (see tcp_v{4,6}_do_rcv()).
3784 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3786 __be32 *ptr = (__be32 *)(th + 1);
3788 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3789 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3790 tp->rx_opt.saw_tstamp = 1;
3792 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3794 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3800 /* Fast parse options. This hopes to only see timestamps.
3801 * If it is wrong it falls back on tcp_parse_options().
3803 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3804 struct tcp_sock *tp)
3806 if (th->doff == sizeof(struct tcphdr) >> 2) {
3807 tp->rx_opt.saw_tstamp = 0;
3809 } else if (tp->rx_opt.tstamp_ok &&
3810 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3811 if (tcp_parse_aligned_timestamp(tp, th))
3814 tcp_parse_options(skb, &tp->rx_opt, 1);
3818 #ifdef CONFIG_TCP_MD5SIG
3820 * Parse MD5 Signature option
3822 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3824 int length = (th->doff << 2) - sizeof (*th);
3825 u8 *ptr = (u8*)(th + 1);
3827 /* If the TCP option is too short, we can short cut */
3828 if (length < TCPOLEN_MD5SIG)
3831 while (length > 0) {
3832 int opcode = *ptr++;
3843 if (opsize < 2 || opsize > length)
3845 if (opcode == TCPOPT_MD5SIG)
3855 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3857 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3858 tp->rx_opt.ts_recent_stamp = get_seconds();
3861 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3863 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3864 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3865 * extra check below makes sure this can only happen
3866 * for pure ACK frames. -DaveM
3868 * Not only, also it occurs for expired timestamps.
3871 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3872 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3873 tcp_store_ts_recent(tp);
3877 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3879 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3880 * it can pass through stack. So, the following predicate verifies that
3881 * this segment is not used for anything but congestion avoidance or
3882 * fast retransmit. Moreover, we even are able to eliminate most of such
3883 * second order effects, if we apply some small "replay" window (~RTO)
3884 * to timestamp space.
3886 * All these measures still do not guarantee that we reject wrapped ACKs
3887 * on networks with high bandwidth, when sequence space is recycled fastly,
3888 * but it guarantees that such events will be very rare and do not affect
3889 * connection seriously. This doesn't look nice, but alas, PAWS is really
3892 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3893 * states that events when retransmit arrives after original data are rare.
3894 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3895 * the biggest problem on large power networks even with minor reordering.
3896 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3897 * up to bandwidth of 18Gigabit/sec. 8) ]
3900 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3902 struct tcp_sock *tp = tcp_sk(sk);
3903 struct tcphdr *th = tcp_hdr(skb);
3904 u32 seq = TCP_SKB_CB(skb)->seq;
3905 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3907 return (/* 1. Pure ACK with correct sequence number. */
3908 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3910 /* 2. ... and duplicate ACK. */
3911 ack == tp->snd_una &&
3913 /* 3. ... and does not update window. */
3914 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3916 /* 4. ... and sits in replay window. */
3917 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3920 static inline int tcp_paws_discard(const struct sock *sk,
3921 const struct sk_buff *skb)
3923 const struct tcp_sock *tp = tcp_sk(sk);
3924 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3925 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3926 !tcp_disordered_ack(sk, skb));
3929 /* Check segment sequence number for validity.
3931 * Segment controls are considered valid, if the segment
3932 * fits to the window after truncation to the window. Acceptability
3933 * of data (and SYN, FIN, of course) is checked separately.
3934 * See tcp_data_queue(), for example.
3936 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3937 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3938 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3939 * (borrowed from freebsd)
3942 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3944 return !before(end_seq, tp->rcv_wup) &&
3945 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3948 /* When we get a reset we do this. */
3949 static void tcp_reset(struct sock *sk)
3951 /* We want the right error as BSD sees it (and indeed as we do). */
3952 switch (sk->sk_state) {
3954 sk->sk_err = ECONNREFUSED;
3956 case TCP_CLOSE_WAIT:
3962 sk->sk_err = ECONNRESET;
3965 if (!sock_flag(sk, SOCK_DEAD))
3966 sk->sk_error_report(sk);
3972 * Process the FIN bit. This now behaves as it is supposed to work
3973 * and the FIN takes effect when it is validly part of sequence
3974 * space. Not before when we get holes.
3976 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3977 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3980 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3981 * close and we go into CLOSING (and later onto TIME-WAIT)
3983 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3985 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3987 struct tcp_sock *tp = tcp_sk(sk);
3989 inet_csk_schedule_ack(sk);
3991 sk->sk_shutdown |= RCV_SHUTDOWN;
3992 sock_set_flag(sk, SOCK_DONE);
3994 switch (sk->sk_state) {
3996 case TCP_ESTABLISHED:
3997 /* Move to CLOSE_WAIT */
3998 tcp_set_state(sk, TCP_CLOSE_WAIT);
3999 inet_csk(sk)->icsk_ack.pingpong = 1;
4002 case TCP_CLOSE_WAIT:
4004 /* Received a retransmission of the FIN, do
4009 /* RFC793: Remain in the LAST-ACK state. */
4013 /* This case occurs when a simultaneous close
4014 * happens, we must ack the received FIN and
4015 * enter the CLOSING state.
4018 tcp_set_state(sk, TCP_CLOSING);
4021 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4023 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4026 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4027 * cases we should never reach this piece of code.
4029 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4030 __func__, sk->sk_state);
4034 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4035 * Probably, we should reset in this case. For now drop them.
4037 __skb_queue_purge(&tp->out_of_order_queue);
4038 if (tcp_is_sack(tp))
4039 tcp_sack_reset(&tp->rx_opt);
4042 if (!sock_flag(sk, SOCK_DEAD)) {
4043 sk->sk_state_change(sk);
4045 /* Do not send POLL_HUP for half duplex close. */
4046 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4047 sk->sk_state == TCP_CLOSE)
4048 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4050 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4054 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4057 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4058 if (before(seq, sp->start_seq))
4059 sp->start_seq = seq;
4060 if (after(end_seq, sp->end_seq))
4061 sp->end_seq = end_seq;
4067 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4069 struct tcp_sock *tp = tcp_sk(sk);
4071 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4074 if (before(seq, tp->rcv_nxt))
4075 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4077 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4079 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4081 tp->rx_opt.dsack = 1;
4082 tp->duplicate_sack[0].start_seq = seq;
4083 tp->duplicate_sack[0].end_seq = end_seq;
4084 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 1;
4088 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4090 struct tcp_sock *tp = tcp_sk(sk);
4092 if (!tp->rx_opt.dsack)
4093 tcp_dsack_set(sk, seq, end_seq);
4095 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4098 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4100 struct tcp_sock *tp = tcp_sk(sk);
4102 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4103 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4104 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4105 tcp_enter_quickack_mode(sk);
4107 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4108 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4110 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4111 end_seq = tp->rcv_nxt;
4112 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4119 /* These routines update the SACK block as out-of-order packets arrive or
4120 * in-order packets close up the sequence space.
4122 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4125 struct tcp_sack_block *sp = &tp->selective_acks[0];
4126 struct tcp_sack_block *swalk = sp + 1;
4128 /* See if the recent change to the first SACK eats into
4129 * or hits the sequence space of other SACK blocks, if so coalesce.
4131 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4132 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4135 /* Zap SWALK, by moving every further SACK up by one slot.
4136 * Decrease num_sacks.
4138 tp->rx_opt.num_sacks--;
4139 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
4141 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4145 this_sack++, swalk++;
4149 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
4150 struct tcp_sack_block *sack2)
4154 tmp = sack1->start_seq;
4155 sack1->start_seq = sack2->start_seq;
4156 sack2->start_seq = tmp;
4158 tmp = sack1->end_seq;
4159 sack1->end_seq = sack2->end_seq;
4160 sack2->end_seq = tmp;
4163 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4165 struct tcp_sock *tp = tcp_sk(sk);
4166 struct tcp_sack_block *sp = &tp->selective_acks[0];
4167 int cur_sacks = tp->rx_opt.num_sacks;
4173 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4174 if (tcp_sack_extend(sp, seq, end_seq)) {
4175 /* Rotate this_sack to the first one. */
4176 for (; this_sack > 0; this_sack--, sp--)
4177 tcp_sack_swap(sp, sp - 1);
4179 tcp_sack_maybe_coalesce(tp);
4184 /* Could not find an adjacent existing SACK, build a new one,
4185 * put it at the front, and shift everyone else down. We
4186 * always know there is at least one SACK present already here.
4188 * If the sack array is full, forget about the last one.
4190 if (this_sack >= TCP_NUM_SACKS) {
4192 tp->rx_opt.num_sacks--;
4195 for (; this_sack > 0; this_sack--, sp--)
4199 /* Build the new head SACK, and we're done. */
4200 sp->start_seq = seq;
4201 sp->end_seq = end_seq;
4202 tp->rx_opt.num_sacks++;
4203 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
4206 /* RCV.NXT advances, some SACKs should be eaten. */
4208 static void tcp_sack_remove(struct tcp_sock *tp)
4210 struct tcp_sack_block *sp = &tp->selective_acks[0];
4211 int num_sacks = tp->rx_opt.num_sacks;
4214 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4215 if (skb_queue_empty(&tp->out_of_order_queue)) {
4216 tp->rx_opt.num_sacks = 0;
4217 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
4221 for (this_sack = 0; this_sack < num_sacks;) {
4222 /* Check if the start of the sack is covered by RCV.NXT. */
4223 if (!before(tp->rcv_nxt, sp->start_seq)) {
4226 /* RCV.NXT must cover all the block! */
4227 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4229 /* Zap this SACK, by moving forward any other SACKS. */
4230 for (i=this_sack+1; i < num_sacks; i++)
4231 tp->selective_acks[i-1] = tp->selective_acks[i];
4238 if (num_sacks != tp->rx_opt.num_sacks) {
4239 tp->rx_opt.num_sacks = num_sacks;
4240 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
4245 /* This one checks to see if we can put data from the
4246 * out_of_order queue into the receive_queue.
4248 static void tcp_ofo_queue(struct sock *sk)
4250 struct tcp_sock *tp = tcp_sk(sk);
4251 __u32 dsack_high = tp->rcv_nxt;
4252 struct sk_buff *skb;
4254 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4255 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4258 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4259 __u32 dsack = dsack_high;
4260 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4261 dsack_high = TCP_SKB_CB(skb)->end_seq;
4262 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4265 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4266 SOCK_DEBUG(sk, "ofo packet was already received \n");
4267 __skb_unlink(skb, &tp->out_of_order_queue);
4271 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4272 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4273 TCP_SKB_CB(skb)->end_seq);
4275 __skb_unlink(skb, &tp->out_of_order_queue);
4276 __skb_queue_tail(&sk->sk_receive_queue, skb);
4277 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4278 if (tcp_hdr(skb)->fin)
4279 tcp_fin(skb, sk, tcp_hdr(skb));
4283 static int tcp_prune_ofo_queue(struct sock *sk);
4284 static int tcp_prune_queue(struct sock *sk);
4286 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4288 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4289 !sk_rmem_schedule(sk, size)) {
4291 if (tcp_prune_queue(sk) < 0)
4294 if (!sk_rmem_schedule(sk, size)) {
4295 if (!tcp_prune_ofo_queue(sk))
4298 if (!sk_rmem_schedule(sk, size))
4305 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4307 struct tcphdr *th = tcp_hdr(skb);
4308 struct tcp_sock *tp = tcp_sk(sk);
4311 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4314 __skb_pull(skb, th->doff * 4);
4316 TCP_ECN_accept_cwr(tp, skb);
4318 if (tp->rx_opt.dsack) {
4319 tp->rx_opt.dsack = 0;
4320 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
4323 /* Queue data for delivery to the user.
4324 * Packets in sequence go to the receive queue.
4325 * Out of sequence packets to the out_of_order_queue.
4327 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4328 if (tcp_receive_window(tp) == 0)
4331 /* Ok. In sequence. In window. */
4332 if (tp->ucopy.task == current &&
4333 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4334 sock_owned_by_user(sk) && !tp->urg_data) {
4335 int chunk = min_t(unsigned int, skb->len,
4338 __set_current_state(TASK_RUNNING);
4341 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4342 tp->ucopy.len -= chunk;
4343 tp->copied_seq += chunk;
4344 eaten = (chunk == skb->len && !th->fin);
4345 tcp_rcv_space_adjust(sk);
4353 tcp_try_rmem_schedule(sk, skb->truesize))
4356 skb_set_owner_r(skb, sk);
4357 __skb_queue_tail(&sk->sk_receive_queue, skb);
4359 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4361 tcp_event_data_recv(sk, skb);
4363 tcp_fin(skb, sk, th);
4365 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4368 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4369 * gap in queue is filled.
4371 if (skb_queue_empty(&tp->out_of_order_queue))
4372 inet_csk(sk)->icsk_ack.pingpong = 0;
4375 if (tp->rx_opt.num_sacks)
4376 tcp_sack_remove(tp);
4378 tcp_fast_path_check(sk);
4382 else if (!sock_flag(sk, SOCK_DEAD))
4383 sk->sk_data_ready(sk, 0);
4387 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4388 /* A retransmit, 2nd most common case. Force an immediate ack. */
4389 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4390 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4393 tcp_enter_quickack_mode(sk);
4394 inet_csk_schedule_ack(sk);
4400 /* Out of window. F.e. zero window probe. */
4401 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4404 tcp_enter_quickack_mode(sk);
4406 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4407 /* Partial packet, seq < rcv_next < end_seq */
4408 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4409 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4410 TCP_SKB_CB(skb)->end_seq);
4412 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4414 /* If window is closed, drop tail of packet. But after
4415 * remembering D-SACK for its head made in previous line.
4417 if (!tcp_receive_window(tp))
4422 TCP_ECN_check_ce(tp, skb);
4424 if (tcp_try_rmem_schedule(sk, skb->truesize))
4427 /* Disable header prediction. */
4429 inet_csk_schedule_ack(sk);
4431 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4432 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4434 skb_set_owner_r(skb, sk);
4436 if (!skb_peek(&tp->out_of_order_queue)) {
4437 /* Initial out of order segment, build 1 SACK. */
4438 if (tcp_is_sack(tp)) {
4439 tp->rx_opt.num_sacks = 1;
4440 tp->rx_opt.dsack = 0;
4441 tp->rx_opt.eff_sacks = 1;
4442 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4443 tp->selective_acks[0].end_seq =
4444 TCP_SKB_CB(skb)->end_seq;
4446 __skb_queue_head(&tp->out_of_order_queue, skb);
4448 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4449 u32 seq = TCP_SKB_CB(skb)->seq;
4450 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4452 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4453 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4455 if (!tp->rx_opt.num_sacks ||
4456 tp->selective_acks[0].end_seq != seq)
4459 /* Common case: data arrive in order after hole. */
4460 tp->selective_acks[0].end_seq = end_seq;
4464 /* Find place to insert this segment. */
4466 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4468 } while ((skb1 = skb1->prev) !=
4469 (struct sk_buff *)&tp->out_of_order_queue);
4471 /* Do skb overlap to previous one? */
4472 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4473 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4474 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4475 /* All the bits are present. Drop. */
4477 tcp_dsack_set(sk, seq, end_seq);
4480 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4481 /* Partial overlap. */
4482 tcp_dsack_set(sk, seq,
4483 TCP_SKB_CB(skb1)->end_seq);
4488 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4490 /* And clean segments covered by new one as whole. */
4491 while ((skb1 = skb->next) !=
4492 (struct sk_buff *)&tp->out_of_order_queue &&
4493 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4494 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4495 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4499 __skb_unlink(skb1, &tp->out_of_order_queue);
4500 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4501 TCP_SKB_CB(skb1)->end_seq);
4506 if (tcp_is_sack(tp))
4507 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4511 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4512 struct sk_buff_head *list)
4514 struct sk_buff *next = skb->next;
4516 __skb_unlink(skb, list);
4518 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4523 /* Collapse contiguous sequence of skbs head..tail with
4524 * sequence numbers start..end.
4525 * Segments with FIN/SYN are not collapsed (only because this
4529 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4530 struct sk_buff *head, struct sk_buff *tail,
4533 struct sk_buff *skb;
4535 /* First, check that queue is collapsible and find
4536 * the point where collapsing can be useful. */
4537 for (skb = head; skb != tail;) {
4538 /* No new bits? It is possible on ofo queue. */
4539 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4540 skb = tcp_collapse_one(sk, skb, list);
4544 /* The first skb to collapse is:
4546 * - bloated or contains data before "start" or
4547 * overlaps to the next one.
4549 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4550 (tcp_win_from_space(skb->truesize) > skb->len ||
4551 before(TCP_SKB_CB(skb)->seq, start) ||
4552 (skb->next != tail &&
4553 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4556 /* Decided to skip this, advance start seq. */
4557 start = TCP_SKB_CB(skb)->end_seq;
4560 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4563 while (before(start, end)) {
4564 struct sk_buff *nskb;
4565 unsigned int header = skb_headroom(skb);
4566 int copy = SKB_MAX_ORDER(header, 0);
4568 /* Too big header? This can happen with IPv6. */
4571 if (end - start < copy)
4573 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4577 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4578 skb_set_network_header(nskb, (skb_network_header(skb) -
4580 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4582 skb_reserve(nskb, header);
4583 memcpy(nskb->head, skb->head, header);
4584 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4585 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4586 __skb_queue_before(list, skb, nskb);
4587 skb_set_owner_r(nskb, sk);
4589 /* Copy data, releasing collapsed skbs. */
4591 int offset = start - TCP_SKB_CB(skb)->seq;
4592 int size = TCP_SKB_CB(skb)->end_seq - start;
4596 size = min(copy, size);
4597 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4599 TCP_SKB_CB(nskb)->end_seq += size;
4603 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4604 skb = tcp_collapse_one(sk, skb, list);
4606 tcp_hdr(skb)->syn ||
4614 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4615 * and tcp_collapse() them until all the queue is collapsed.
4617 static void tcp_collapse_ofo_queue(struct sock *sk)
4619 struct tcp_sock *tp = tcp_sk(sk);
4620 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4621 struct sk_buff *head;
4627 start = TCP_SKB_CB(skb)->seq;
4628 end = TCP_SKB_CB(skb)->end_seq;
4634 /* Segment is terminated when we see gap or when
4635 * we are at the end of all the queue. */
4636 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4637 after(TCP_SKB_CB(skb)->seq, end) ||
4638 before(TCP_SKB_CB(skb)->end_seq, start)) {
4639 tcp_collapse(sk, &tp->out_of_order_queue,
4640 head, skb, start, end);
4642 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4644 /* Start new segment */
4645 start = TCP_SKB_CB(skb)->seq;
4646 end = TCP_SKB_CB(skb)->end_seq;
4648 if (before(TCP_SKB_CB(skb)->seq, start))
4649 start = TCP_SKB_CB(skb)->seq;
4650 if (after(TCP_SKB_CB(skb)->end_seq, end))
4651 end = TCP_SKB_CB(skb)->end_seq;
4657 * Purge the out-of-order queue.
4658 * Return true if queue was pruned.
4660 static int tcp_prune_ofo_queue(struct sock *sk)
4662 struct tcp_sock *tp = tcp_sk(sk);
4665 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4666 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4667 __skb_queue_purge(&tp->out_of_order_queue);
4669 /* Reset SACK state. A conforming SACK implementation will
4670 * do the same at a timeout based retransmit. When a connection
4671 * is in a sad state like this, we care only about integrity
4672 * of the connection not performance.
4674 if (tp->rx_opt.sack_ok)
4675 tcp_sack_reset(&tp->rx_opt);
4682 /* Reduce allocated memory if we can, trying to get
4683 * the socket within its memory limits again.
4685 * Return less than zero if we should start dropping frames
4686 * until the socket owning process reads some of the data
4687 * to stabilize the situation.
4689 static int tcp_prune_queue(struct sock *sk)
4691 struct tcp_sock *tp = tcp_sk(sk);
4693 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4695 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4697 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4698 tcp_clamp_window(sk);
4699 else if (tcp_memory_pressure)
4700 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4702 tcp_collapse_ofo_queue(sk);
4703 tcp_collapse(sk, &sk->sk_receive_queue,
4704 sk->sk_receive_queue.next,
4705 (struct sk_buff *)&sk->sk_receive_queue,
4706 tp->copied_seq, tp->rcv_nxt);
4709 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4712 /* Collapsing did not help, destructive actions follow.
4713 * This must not ever occur. */
4715 tcp_prune_ofo_queue(sk);
4717 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4720 /* If we are really being abused, tell the caller to silently
4721 * drop receive data on the floor. It will get retransmitted
4722 * and hopefully then we'll have sufficient space.
4724 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4726 /* Massive buffer overcommit. */
4731 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4732 * As additional protections, we do not touch cwnd in retransmission phases,
4733 * and if application hit its sndbuf limit recently.
4735 void tcp_cwnd_application_limited(struct sock *sk)
4737 struct tcp_sock *tp = tcp_sk(sk);
4739 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4740 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4741 /* Limited by application or receiver window. */
4742 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4743 u32 win_used = max(tp->snd_cwnd_used, init_win);
4744 if (win_used < tp->snd_cwnd) {
4745 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4746 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4748 tp->snd_cwnd_used = 0;
4750 tp->snd_cwnd_stamp = tcp_time_stamp;
4753 static int tcp_should_expand_sndbuf(struct sock *sk)
4755 struct tcp_sock *tp = tcp_sk(sk);
4757 /* If the user specified a specific send buffer setting, do
4760 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4763 /* If we are under global TCP memory pressure, do not expand. */
4764 if (tcp_memory_pressure)
4767 /* If we are under soft global TCP memory pressure, do not expand. */
4768 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4771 /* If we filled the congestion window, do not expand. */
4772 if (tp->packets_out >= tp->snd_cwnd)
4778 /* When incoming ACK allowed to free some skb from write_queue,
4779 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4780 * on the exit from tcp input handler.
4782 * PROBLEM: sndbuf expansion does not work well with largesend.
4784 static void tcp_new_space(struct sock *sk)
4786 struct tcp_sock *tp = tcp_sk(sk);
4788 if (tcp_should_expand_sndbuf(sk)) {
4789 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4790 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4791 int demanded = max_t(unsigned int, tp->snd_cwnd,
4792 tp->reordering + 1);
4793 sndmem *= 2 * demanded;
4794 if (sndmem > sk->sk_sndbuf)
4795 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4796 tp->snd_cwnd_stamp = tcp_time_stamp;
4799 sk->sk_write_space(sk);
4802 static void tcp_check_space(struct sock *sk)
4804 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4805 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4806 if (sk->sk_socket &&
4807 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4812 static inline void tcp_data_snd_check(struct sock *sk)
4814 tcp_push_pending_frames(sk);
4815 tcp_check_space(sk);
4819 * Check if sending an ack is needed.
4821 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4823 struct tcp_sock *tp = tcp_sk(sk);
4825 /* More than one full frame received... */
4826 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4827 /* ... and right edge of window advances far enough.
4828 * (tcp_recvmsg() will send ACK otherwise). Or...
4830 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4831 /* We ACK each frame or... */
4832 tcp_in_quickack_mode(sk) ||
4833 /* We have out of order data. */
4834 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4835 /* Then ack it now */
4838 /* Else, send delayed ack. */
4839 tcp_send_delayed_ack(sk);
4843 static inline void tcp_ack_snd_check(struct sock *sk)
4845 if (!inet_csk_ack_scheduled(sk)) {
4846 /* We sent a data segment already. */
4849 __tcp_ack_snd_check(sk, 1);
4853 * This routine is only called when we have urgent data
4854 * signaled. Its the 'slow' part of tcp_urg. It could be
4855 * moved inline now as tcp_urg is only called from one
4856 * place. We handle URGent data wrong. We have to - as
4857 * BSD still doesn't use the correction from RFC961.
4858 * For 1003.1g we should support a new option TCP_STDURG to permit
4859 * either form (or just set the sysctl tcp_stdurg).
4862 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4864 struct tcp_sock *tp = tcp_sk(sk);
4865 u32 ptr = ntohs(th->urg_ptr);
4867 if (ptr && !sysctl_tcp_stdurg)
4869 ptr += ntohl(th->seq);
4871 /* Ignore urgent data that we've already seen and read. */
4872 if (after(tp->copied_seq, ptr))
4875 /* Do not replay urg ptr.
4877 * NOTE: interesting situation not covered by specs.
4878 * Misbehaving sender may send urg ptr, pointing to segment,
4879 * which we already have in ofo queue. We are not able to fetch
4880 * such data and will stay in TCP_URG_NOTYET until will be eaten
4881 * by recvmsg(). Seems, we are not obliged to handle such wicked
4882 * situations. But it is worth to think about possibility of some
4883 * DoSes using some hypothetical application level deadlock.
4885 if (before(ptr, tp->rcv_nxt))
4888 /* Do we already have a newer (or duplicate) urgent pointer? */
4889 if (tp->urg_data && !after(ptr, tp->urg_seq))
4892 /* Tell the world about our new urgent pointer. */
4895 /* We may be adding urgent data when the last byte read was
4896 * urgent. To do this requires some care. We cannot just ignore
4897 * tp->copied_seq since we would read the last urgent byte again
4898 * as data, nor can we alter copied_seq until this data arrives
4899 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4901 * NOTE. Double Dutch. Rendering to plain English: author of comment
4902 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4903 * and expect that both A and B disappear from stream. This is _wrong_.
4904 * Though this happens in BSD with high probability, this is occasional.
4905 * Any application relying on this is buggy. Note also, that fix "works"
4906 * only in this artificial test. Insert some normal data between A and B and we will
4907 * decline of BSD again. Verdict: it is better to remove to trap
4910 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4911 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4912 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4914 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4915 __skb_unlink(skb, &sk->sk_receive_queue);
4920 tp->urg_data = TCP_URG_NOTYET;
4923 /* Disable header prediction. */
4927 /* This is the 'fast' part of urgent handling. */
4928 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4930 struct tcp_sock *tp = tcp_sk(sk);
4932 /* Check if we get a new urgent pointer - normally not. */
4934 tcp_check_urg(sk, th);
4936 /* Do we wait for any urgent data? - normally not... */
4937 if (tp->urg_data == TCP_URG_NOTYET) {
4938 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4941 /* Is the urgent pointer pointing into this packet? */
4942 if (ptr < skb->len) {
4944 if (skb_copy_bits(skb, ptr, &tmp, 1))
4946 tp->urg_data = TCP_URG_VALID | tmp;
4947 if (!sock_flag(sk, SOCK_DEAD))
4948 sk->sk_data_ready(sk, 0);
4953 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4955 struct tcp_sock *tp = tcp_sk(sk);
4956 int chunk = skb->len - hlen;
4960 if (skb_csum_unnecessary(skb))
4961 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4963 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4967 tp->ucopy.len -= chunk;
4968 tp->copied_seq += chunk;
4969 tcp_rcv_space_adjust(sk);
4976 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4977 struct sk_buff *skb)
4981 if (sock_owned_by_user(sk)) {
4983 result = __tcp_checksum_complete(skb);
4986 result = __tcp_checksum_complete(skb);
4991 static inline int tcp_checksum_complete_user(struct sock *sk,
4992 struct sk_buff *skb)
4994 return !skb_csum_unnecessary(skb) &&
4995 __tcp_checksum_complete_user(sk, skb);
4998 #ifdef CONFIG_NET_DMA
4999 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5002 struct tcp_sock *tp = tcp_sk(sk);
5003 int chunk = skb->len - hlen;
5005 int copied_early = 0;
5007 if (tp->ucopy.wakeup)
5010 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5011 tp->ucopy.dma_chan = get_softnet_dma();
5013 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5015 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5017 tp->ucopy.iov, chunk,
5018 tp->ucopy.pinned_list);
5023 tp->ucopy.dma_cookie = dma_cookie;
5026 tp->ucopy.len -= chunk;
5027 tp->copied_seq += chunk;
5028 tcp_rcv_space_adjust(sk);
5030 if ((tp->ucopy.len == 0) ||
5031 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5032 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5033 tp->ucopy.wakeup = 1;
5034 sk->sk_data_ready(sk, 0);
5036 } else if (chunk > 0) {
5037 tp->ucopy.wakeup = 1;
5038 sk->sk_data_ready(sk, 0);
5041 return copied_early;
5043 #endif /* CONFIG_NET_DMA */
5045 /* Does PAWS and seqno based validation of an incoming segment, flags will
5046 * play significant role here.
5048 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5049 struct tcphdr *th, int syn_inerr)
5051 struct tcp_sock *tp = tcp_sk(sk);
5053 /* RFC1323: H1. Apply PAWS check first. */
5054 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5055 tcp_paws_discard(sk, skb)) {
5057 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5058 tcp_send_dupack(sk, skb);
5061 /* Reset is accepted even if it did not pass PAWS. */
5064 /* Step 1: check sequence number */
5065 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5066 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5067 * (RST) segments are validated by checking their SEQ-fields."
5068 * And page 69: "If an incoming segment is not acceptable,
5069 * an acknowledgment should be sent in reply (unless the RST
5070 * bit is set, if so drop the segment and return)".
5073 tcp_send_dupack(sk, skb);
5077 /* Step 2: check RST bit */
5083 /* ts_recent update must be made after we are sure that the packet
5086 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5088 /* step 3: check security and precedence [ignored] */
5090 /* step 4: Check for a SYN in window. */
5091 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5093 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5094 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5107 * TCP receive function for the ESTABLISHED state.
5109 * It is split into a fast path and a slow path. The fast path is
5111 * - A zero window was announced from us - zero window probing
5112 * is only handled properly in the slow path.
5113 * - Out of order segments arrived.
5114 * - Urgent data is expected.
5115 * - There is no buffer space left
5116 * - Unexpected TCP flags/window values/header lengths are received
5117 * (detected by checking the TCP header against pred_flags)
5118 * - Data is sent in both directions. Fast path only supports pure senders
5119 * or pure receivers (this means either the sequence number or the ack
5120 * value must stay constant)
5121 * - Unexpected TCP option.
5123 * When these conditions are not satisfied it drops into a standard
5124 * receive procedure patterned after RFC793 to handle all cases.
5125 * The first three cases are guaranteed by proper pred_flags setting,
5126 * the rest is checked inline. Fast processing is turned on in
5127 * tcp_data_queue when everything is OK.
5129 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5130 struct tcphdr *th, unsigned len)
5132 struct tcp_sock *tp = tcp_sk(sk);
5136 * Header prediction.
5137 * The code loosely follows the one in the famous
5138 * "30 instruction TCP receive" Van Jacobson mail.
5140 * Van's trick is to deposit buffers into socket queue
5141 * on a device interrupt, to call tcp_recv function
5142 * on the receive process context and checksum and copy
5143 * the buffer to user space. smart...
5145 * Our current scheme is not silly either but we take the
5146 * extra cost of the net_bh soft interrupt processing...
5147 * We do checksum and copy also but from device to kernel.
5150 tp->rx_opt.saw_tstamp = 0;
5152 /* pred_flags is 0xS?10 << 16 + snd_wnd
5153 * if header_prediction is to be made
5154 * 'S' will always be tp->tcp_header_len >> 2
5155 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5156 * turn it off (when there are holes in the receive
5157 * space for instance)
5158 * PSH flag is ignored.
5161 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5162 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5163 int tcp_header_len = tp->tcp_header_len;
5165 /* Timestamp header prediction: tcp_header_len
5166 * is automatically equal to th->doff*4 due to pred_flags
5170 /* Check timestamp */
5171 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5172 /* No? Slow path! */
5173 if (!tcp_parse_aligned_timestamp(tp, th))
5176 /* If PAWS failed, check it more carefully in slow path */
5177 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5180 /* DO NOT update ts_recent here, if checksum fails
5181 * and timestamp was corrupted part, it will result
5182 * in a hung connection since we will drop all
5183 * future packets due to the PAWS test.
5187 if (len <= tcp_header_len) {
5188 /* Bulk data transfer: sender */
5189 if (len == tcp_header_len) {
5190 /* Predicted packet is in window by definition.
5191 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5192 * Hence, check seq<=rcv_wup reduces to:
5194 if (tcp_header_len ==
5195 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5196 tp->rcv_nxt == tp->rcv_wup)
5197 tcp_store_ts_recent(tp);
5199 /* We know that such packets are checksummed
5202 tcp_ack(sk, skb, 0);
5204 tcp_data_snd_check(sk);
5206 } else { /* Header too small */
5207 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5212 int copied_early = 0;
5214 if (tp->copied_seq == tp->rcv_nxt &&
5215 len - tcp_header_len <= tp->ucopy.len) {
5216 #ifdef CONFIG_NET_DMA
5217 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5222 if (tp->ucopy.task == current &&
5223 sock_owned_by_user(sk) && !copied_early) {
5224 __set_current_state(TASK_RUNNING);
5226 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5230 /* Predicted packet is in window by definition.
5231 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5232 * Hence, check seq<=rcv_wup reduces to:
5234 if (tcp_header_len ==
5235 (sizeof(struct tcphdr) +
5236 TCPOLEN_TSTAMP_ALIGNED) &&
5237 tp->rcv_nxt == tp->rcv_wup)
5238 tcp_store_ts_recent(tp);
5240 tcp_rcv_rtt_measure_ts(sk, skb);
5242 __skb_pull(skb, tcp_header_len);
5243 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5244 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5247 tcp_cleanup_rbuf(sk, skb->len);
5250 if (tcp_checksum_complete_user(sk, skb))
5253 /* Predicted packet is in window by definition.
5254 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5255 * Hence, check seq<=rcv_wup reduces to:
5257 if (tcp_header_len ==
5258 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5259 tp->rcv_nxt == tp->rcv_wup)
5260 tcp_store_ts_recent(tp);
5262 tcp_rcv_rtt_measure_ts(sk, skb);
5264 if ((int)skb->truesize > sk->sk_forward_alloc)
5267 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5269 /* Bulk data transfer: receiver */
5270 __skb_pull(skb, tcp_header_len);
5271 __skb_queue_tail(&sk->sk_receive_queue, skb);
5272 skb_set_owner_r(skb, sk);
5273 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5276 tcp_event_data_recv(sk, skb);
5278 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5279 /* Well, only one small jumplet in fast path... */
5280 tcp_ack(sk, skb, FLAG_DATA);
5281 tcp_data_snd_check(sk);
5282 if (!inet_csk_ack_scheduled(sk))
5286 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5287 __tcp_ack_snd_check(sk, 0);
5289 #ifdef CONFIG_NET_DMA
5291 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5297 sk->sk_data_ready(sk, 0);
5303 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5307 * Standard slow path.
5310 res = tcp_validate_incoming(sk, skb, th, 1);
5316 tcp_ack(sk, skb, FLAG_SLOWPATH);
5318 tcp_rcv_rtt_measure_ts(sk, skb);
5320 /* Process urgent data. */
5321 tcp_urg(sk, skb, th);
5323 /* step 7: process the segment text */
5324 tcp_data_queue(sk, skb);
5326 tcp_data_snd_check(sk);
5327 tcp_ack_snd_check(sk);
5331 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5338 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5339 struct tcphdr *th, unsigned len)
5341 struct tcp_sock *tp = tcp_sk(sk);
5342 struct inet_connection_sock *icsk = inet_csk(sk);
5343 int saved_clamp = tp->rx_opt.mss_clamp;
5345 tcp_parse_options(skb, &tp->rx_opt, 0);
5349 * "If the state is SYN-SENT then
5350 * first check the ACK bit
5351 * If the ACK bit is set
5352 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5353 * a reset (unless the RST bit is set, if so drop
5354 * the segment and return)"
5356 * We do not send data with SYN, so that RFC-correct
5359 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5360 goto reset_and_undo;
5362 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5363 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5365 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5366 goto reset_and_undo;
5369 /* Now ACK is acceptable.
5371 * "If the RST bit is set
5372 * If the ACK was acceptable then signal the user "error:
5373 * connection reset", drop the segment, enter CLOSED state,
5374 * delete TCB, and return."
5383 * "fifth, if neither of the SYN or RST bits is set then
5384 * drop the segment and return."
5390 goto discard_and_undo;
5393 * "If the SYN bit is on ...
5394 * are acceptable then ...
5395 * (our SYN has been ACKed), change the connection
5396 * state to ESTABLISHED..."
5399 TCP_ECN_rcv_synack(tp, th);
5401 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5402 tcp_ack(sk, skb, FLAG_SLOWPATH);
5404 /* Ok.. it's good. Set up sequence numbers and
5405 * move to established.
5407 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5408 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5410 /* RFC1323: The window in SYN & SYN/ACK segments is
5413 tp->snd_wnd = ntohs(th->window);
5414 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
5416 if (!tp->rx_opt.wscale_ok) {
5417 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5418 tp->window_clamp = min(tp->window_clamp, 65535U);
5421 if (tp->rx_opt.saw_tstamp) {
5422 tp->rx_opt.tstamp_ok = 1;
5423 tp->tcp_header_len =
5424 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5425 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5426 tcp_store_ts_recent(tp);
5428 tp->tcp_header_len = sizeof(struct tcphdr);
5431 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5432 tcp_enable_fack(tp);
5435 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5436 tcp_initialize_rcv_mss(sk);
5438 /* Remember, tcp_poll() does not lock socket!
5439 * Change state from SYN-SENT only after copied_seq
5440 * is initialized. */
5441 tp->copied_seq = tp->rcv_nxt;
5443 tcp_set_state(sk, TCP_ESTABLISHED);
5445 security_inet_conn_established(sk, skb);
5447 /* Make sure socket is routed, for correct metrics. */
5448 icsk->icsk_af_ops->rebuild_header(sk);
5450 tcp_init_metrics(sk);
5452 tcp_init_congestion_control(sk);
5454 /* Prevent spurious tcp_cwnd_restart() on first data
5457 tp->lsndtime = tcp_time_stamp;
5459 tcp_init_buffer_space(sk);
5461 if (sock_flag(sk, SOCK_KEEPOPEN))
5462 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5464 if (!tp->rx_opt.snd_wscale)
5465 __tcp_fast_path_on(tp, tp->snd_wnd);
5469 if (!sock_flag(sk, SOCK_DEAD)) {
5470 sk->sk_state_change(sk);
5471 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5474 if (sk->sk_write_pending ||
5475 icsk->icsk_accept_queue.rskq_defer_accept ||
5476 icsk->icsk_ack.pingpong) {
5477 /* Save one ACK. Data will be ready after
5478 * several ticks, if write_pending is set.
5480 * It may be deleted, but with this feature tcpdumps
5481 * look so _wonderfully_ clever, that I was not able
5482 * to stand against the temptation 8) --ANK
5484 inet_csk_schedule_ack(sk);
5485 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5486 icsk->icsk_ack.ato = TCP_ATO_MIN;
5487 tcp_incr_quickack(sk);
5488 tcp_enter_quickack_mode(sk);
5489 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5490 TCP_DELACK_MAX, TCP_RTO_MAX);
5501 /* No ACK in the segment */
5505 * "If the RST bit is set
5507 * Otherwise (no ACK) drop the segment and return."
5510 goto discard_and_undo;
5514 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5515 tcp_paws_check(&tp->rx_opt, 0))
5516 goto discard_and_undo;
5519 /* We see SYN without ACK. It is attempt of
5520 * simultaneous connect with crossed SYNs.
5521 * Particularly, it can be connect to self.
5523 tcp_set_state(sk, TCP_SYN_RECV);
5525 if (tp->rx_opt.saw_tstamp) {
5526 tp->rx_opt.tstamp_ok = 1;
5527 tcp_store_ts_recent(tp);
5528 tp->tcp_header_len =
5529 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5531 tp->tcp_header_len = sizeof(struct tcphdr);
5534 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5535 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5537 /* RFC1323: The window in SYN & SYN/ACK segments is
5540 tp->snd_wnd = ntohs(th->window);
5541 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5542 tp->max_window = tp->snd_wnd;
5544 TCP_ECN_rcv_syn(tp, th);
5547 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5548 tcp_initialize_rcv_mss(sk);
5550 tcp_send_synack(sk);
5552 /* Note, we could accept data and URG from this segment.
5553 * There are no obstacles to make this.
5555 * However, if we ignore data in ACKless segments sometimes,
5556 * we have no reasons to accept it sometimes.
5557 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5558 * is not flawless. So, discard packet for sanity.
5559 * Uncomment this return to process the data.
5566 /* "fifth, if neither of the SYN or RST bits is set then
5567 * drop the segment and return."
5571 tcp_clear_options(&tp->rx_opt);
5572 tp->rx_opt.mss_clamp = saved_clamp;
5576 tcp_clear_options(&tp->rx_opt);
5577 tp->rx_opt.mss_clamp = saved_clamp;
5582 * This function implements the receiving procedure of RFC 793 for
5583 * all states except ESTABLISHED and TIME_WAIT.
5584 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5585 * address independent.
5588 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5589 struct tcphdr *th, unsigned len)
5591 struct tcp_sock *tp = tcp_sk(sk);
5592 struct inet_connection_sock *icsk = inet_csk(sk);
5596 tp->rx_opt.saw_tstamp = 0;
5598 switch (sk->sk_state) {
5610 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5613 /* Now we have several options: In theory there is
5614 * nothing else in the frame. KA9Q has an option to
5615 * send data with the syn, BSD accepts data with the
5616 * syn up to the [to be] advertised window and
5617 * Solaris 2.1 gives you a protocol error. For now
5618 * we just ignore it, that fits the spec precisely
5619 * and avoids incompatibilities. It would be nice in
5620 * future to drop through and process the data.
5622 * Now that TTCP is starting to be used we ought to
5624 * But, this leaves one open to an easy denial of
5625 * service attack, and SYN cookies can't defend
5626 * against this problem. So, we drop the data
5627 * in the interest of security over speed unless
5628 * it's still in use.
5636 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5640 /* Do step6 onward by hand. */
5641 tcp_urg(sk, skb, th);
5643 tcp_data_snd_check(sk);
5647 res = tcp_validate_incoming(sk, skb, th, 0);
5651 /* step 5: check the ACK field */
5653 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5655 switch (sk->sk_state) {
5658 tp->copied_seq = tp->rcv_nxt;
5660 tcp_set_state(sk, TCP_ESTABLISHED);
5661 sk->sk_state_change(sk);
5663 /* Note, that this wakeup is only for marginal
5664 * crossed SYN case. Passively open sockets
5665 * are not waked up, because sk->sk_sleep ==
5666 * NULL and sk->sk_socket == NULL.
5670 SOCK_WAKE_IO, POLL_OUT);
5672 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5673 tp->snd_wnd = ntohs(th->window) <<
5674 tp->rx_opt.snd_wscale;
5675 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5676 TCP_SKB_CB(skb)->seq);
5678 /* tcp_ack considers this ACK as duplicate
5679 * and does not calculate rtt.
5680 * Fix it at least with timestamps.
5682 if (tp->rx_opt.saw_tstamp &&
5683 tp->rx_opt.rcv_tsecr && !tp->srtt)
5684 tcp_ack_saw_tstamp(sk, 0);
5686 if (tp->rx_opt.tstamp_ok)
5687 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5689 /* Make sure socket is routed, for
5692 icsk->icsk_af_ops->rebuild_header(sk);
5694 tcp_init_metrics(sk);
5696 tcp_init_congestion_control(sk);
5698 /* Prevent spurious tcp_cwnd_restart() on
5699 * first data packet.
5701 tp->lsndtime = tcp_time_stamp;
5704 tcp_initialize_rcv_mss(sk);
5705 tcp_init_buffer_space(sk);
5706 tcp_fast_path_on(tp);
5713 if (tp->snd_una == tp->write_seq) {
5714 tcp_set_state(sk, TCP_FIN_WAIT2);
5715 sk->sk_shutdown |= SEND_SHUTDOWN;
5716 dst_confirm(sk->sk_dst_cache);
5718 if (!sock_flag(sk, SOCK_DEAD))
5719 /* Wake up lingering close() */
5720 sk->sk_state_change(sk);
5724 if (tp->linger2 < 0 ||
5725 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5726 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5728 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5732 tmo = tcp_fin_time(sk);
5733 if (tmo > TCP_TIMEWAIT_LEN) {
5734 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5735 } else if (th->fin || sock_owned_by_user(sk)) {
5736 /* Bad case. We could lose such FIN otherwise.
5737 * It is not a big problem, but it looks confusing
5738 * and not so rare event. We still can lose it now,
5739 * if it spins in bh_lock_sock(), but it is really
5742 inet_csk_reset_keepalive_timer(sk, tmo);
5744 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5752 if (tp->snd_una == tp->write_seq) {
5753 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5759 if (tp->snd_una == tp->write_seq) {
5760 tcp_update_metrics(sk);
5769 /* step 6: check the URG bit */
5770 tcp_urg(sk, skb, th);
5772 /* step 7: process the segment text */
5773 switch (sk->sk_state) {
5774 case TCP_CLOSE_WAIT:
5777 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5781 /* RFC 793 says to queue data in these states,
5782 * RFC 1122 says we MUST send a reset.
5783 * BSD 4.4 also does reset.
5785 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5786 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5787 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5788 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5794 case TCP_ESTABLISHED:
5795 tcp_data_queue(sk, skb);
5800 /* tcp_data could move socket to TIME-WAIT */
5801 if (sk->sk_state != TCP_CLOSE) {
5802 tcp_data_snd_check(sk);
5803 tcp_ack_snd_check(sk);
5813 EXPORT_SYMBOL(sysctl_tcp_ecn);
5814 EXPORT_SYMBOL(sysctl_tcp_reordering);
5815 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5816 EXPORT_SYMBOL(tcp_parse_options);
5817 #ifdef CONFIG_TCP_MD5SIG
5818 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5820 EXPORT_SYMBOL(tcp_rcv_established);
5821 EXPORT_SYMBOL(tcp_rcv_state_process);
5822 EXPORT_SYMBOL(tcp_initialize_rcv_mss);