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.
705 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
706 * guarantees that rto is higher.
708 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
709 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
712 /* Save metrics learned by this TCP session.
713 This function is called only, when TCP finishes successfully
714 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
716 void tcp_update_metrics(struct sock *sk)
718 struct tcp_sock *tp = tcp_sk(sk);
719 struct dst_entry *dst = __sk_dst_get(sk);
721 if (sysctl_tcp_nometrics_save)
726 if (dst && (dst->flags & DST_HOST)) {
727 const struct inet_connection_sock *icsk = inet_csk(sk);
731 if (icsk->icsk_backoff || !tp->srtt) {
732 /* This session failed to estimate rtt. Why?
733 * Probably, no packets returned in time.
736 if (!(dst_metric_locked(dst, RTAX_RTT)))
737 dst->metrics[RTAX_RTT - 1] = 0;
741 rtt = dst_metric_rtt(dst, RTAX_RTT);
744 /* If newly calculated rtt larger than stored one,
745 * store new one. Otherwise, use EWMA. Remember,
746 * rtt overestimation is always better than underestimation.
748 if (!(dst_metric_locked(dst, RTAX_RTT))) {
750 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
752 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
755 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
760 /* Scale deviation to rttvar fixed point */
765 var = dst_metric_rtt(dst, RTAX_RTTVAR);
769 var -= (var - m) >> 2;
771 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
774 if (tp->snd_ssthresh >= 0xFFFF) {
775 /* Slow start still did not finish. */
776 if (dst_metric(dst, RTAX_SSTHRESH) &&
777 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
778 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
779 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
780 if (!dst_metric_locked(dst, RTAX_CWND) &&
781 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
782 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
783 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
784 icsk->icsk_ca_state == TCP_CA_Open) {
785 /* Cong. avoidance phase, cwnd is reliable. */
786 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
787 dst->metrics[RTAX_SSTHRESH-1] =
788 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
789 if (!dst_metric_locked(dst, RTAX_CWND))
790 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
792 /* Else slow start did not finish, cwnd is non-sense,
793 ssthresh may be also invalid.
795 if (!dst_metric_locked(dst, RTAX_CWND))
796 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
797 if (dst_metric(dst, RTAX_SSTHRESH) &&
798 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
799 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
800 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
803 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
804 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
805 tp->reordering != sysctl_tcp_reordering)
806 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
811 /* Numbers are taken from RFC3390.
813 * John Heffner states:
815 * The RFC specifies a window of no more than 4380 bytes
816 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
817 * is a bit misleading because they use a clamp at 4380 bytes
818 * rather than use a multiplier in the relevant range.
820 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
822 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
825 if (tp->mss_cache > 1460)
828 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
830 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
833 /* Set slow start threshold and cwnd not falling to slow start */
834 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
836 struct tcp_sock *tp = tcp_sk(sk);
837 const struct inet_connection_sock *icsk = inet_csk(sk);
839 tp->prior_ssthresh = 0;
841 if (icsk->icsk_ca_state < TCP_CA_CWR) {
844 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
845 tp->snd_cwnd = min(tp->snd_cwnd,
846 tcp_packets_in_flight(tp) + 1U);
847 tp->snd_cwnd_cnt = 0;
848 tp->high_seq = tp->snd_nxt;
849 tp->snd_cwnd_stamp = tcp_time_stamp;
850 TCP_ECN_queue_cwr(tp);
852 tcp_set_ca_state(sk, TCP_CA_CWR);
857 * Packet counting of FACK is based on in-order assumptions, therefore TCP
858 * disables it when reordering is detected
860 static void tcp_disable_fack(struct tcp_sock *tp)
862 /* RFC3517 uses different metric in lost marker => reset on change */
864 tp->lost_skb_hint = NULL;
865 tp->rx_opt.sack_ok &= ~2;
868 /* Take a notice that peer is sending D-SACKs */
869 static void tcp_dsack_seen(struct tcp_sock *tp)
871 tp->rx_opt.sack_ok |= 4;
874 /* Initialize metrics on socket. */
876 static void tcp_init_metrics(struct sock *sk)
878 struct tcp_sock *tp = tcp_sk(sk);
879 struct dst_entry *dst = __sk_dst_get(sk);
886 if (dst_metric_locked(dst, RTAX_CWND))
887 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
888 if (dst_metric(dst, RTAX_SSTHRESH)) {
889 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
890 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
891 tp->snd_ssthresh = tp->snd_cwnd_clamp;
893 if (dst_metric(dst, RTAX_REORDERING) &&
894 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
895 tcp_disable_fack(tp);
896 tp->reordering = dst_metric(dst, RTAX_REORDERING);
899 if (dst_metric(dst, RTAX_RTT) == 0)
902 if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
905 /* Initial rtt is determined from SYN,SYN-ACK.
906 * The segment is small and rtt may appear much
907 * less than real one. Use per-dst memory
908 * to make it more realistic.
910 * A bit of theory. RTT is time passed after "normal" sized packet
911 * is sent until it is ACKed. In normal circumstances sending small
912 * packets force peer to delay ACKs and calculation is correct too.
913 * The algorithm is adaptive and, provided we follow specs, it
914 * NEVER underestimate RTT. BUT! If peer tries to make some clever
915 * tricks sort of "quick acks" for time long enough to decrease RTT
916 * to low value, and then abruptly stops to do it and starts to delay
917 * ACKs, wait for troubles.
919 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
920 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
921 tp->rtt_seq = tp->snd_nxt;
923 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
924 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
925 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
928 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
930 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
931 tp->snd_cwnd_stamp = tcp_time_stamp;
935 /* Play conservative. If timestamps are not
936 * supported, TCP will fail to recalculate correct
937 * rtt, if initial rto is too small. FORGET ALL AND RESET!
939 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
941 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
942 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
946 static void tcp_update_reordering(struct sock *sk, const int metric,
949 struct tcp_sock *tp = tcp_sk(sk);
950 if (metric > tp->reordering) {
953 tp->reordering = min(TCP_MAX_REORDERING, metric);
955 /* This exciting event is worth to be remembered. 8) */
957 mib_idx = LINUX_MIB_TCPTSREORDER;
958 else if (tcp_is_reno(tp))
959 mib_idx = LINUX_MIB_TCPRENOREORDER;
960 else if (tcp_is_fack(tp))
961 mib_idx = LINUX_MIB_TCPFACKREORDER;
963 mib_idx = LINUX_MIB_TCPSACKREORDER;
965 NET_INC_STATS_BH(sock_net(sk), mib_idx);
966 #if FASTRETRANS_DEBUG > 1
967 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
968 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
972 tp->undo_marker ? tp->undo_retrans : 0);
974 tcp_disable_fack(tp);
978 /* This must be called before lost_out is incremented */
979 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
981 if ((tp->retransmit_skb_hint == NULL) ||
982 before(TCP_SKB_CB(skb)->seq,
983 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
984 tp->retransmit_skb_hint = skb;
987 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
988 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
991 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
993 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
994 tcp_verify_retransmit_hint(tp, skb);
996 tp->lost_out += tcp_skb_pcount(skb);
997 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1001 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1002 struct sk_buff *skb)
1004 tcp_verify_retransmit_hint(tp, skb);
1006 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1007 tp->lost_out += tcp_skb_pcount(skb);
1008 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1012 /* This procedure tags the retransmission queue when SACKs arrive.
1014 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1015 * Packets in queue with these bits set are counted in variables
1016 * sacked_out, retrans_out and lost_out, correspondingly.
1018 * Valid combinations are:
1019 * Tag InFlight Description
1020 * 0 1 - orig segment is in flight.
1021 * S 0 - nothing flies, orig reached receiver.
1022 * L 0 - nothing flies, orig lost by net.
1023 * R 2 - both orig and retransmit are in flight.
1024 * L|R 1 - orig is lost, retransmit is in flight.
1025 * S|R 1 - orig reached receiver, retrans is still in flight.
1026 * (L|S|R is logically valid, it could occur when L|R is sacked,
1027 * but it is equivalent to plain S and code short-curcuits it to S.
1028 * L|S is logically invalid, it would mean -1 packet in flight 8))
1030 * These 6 states form finite state machine, controlled by the following events:
1031 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1032 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1033 * 3. Loss detection event of one of three flavors:
1034 * A. Scoreboard estimator decided the packet is lost.
1035 * A'. Reno "three dupacks" marks head of queue lost.
1036 * A''. Its FACK modfication, head until snd.fack is lost.
1037 * B. SACK arrives sacking data transmitted after never retransmitted
1038 * hole was sent out.
1039 * C. SACK arrives sacking SND.NXT at the moment, when the
1040 * segment was retransmitted.
1041 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1043 * It is pleasant to note, that state diagram turns out to be commutative,
1044 * so that we are allowed not to be bothered by order of our actions,
1045 * when multiple events arrive simultaneously. (see the function below).
1047 * Reordering detection.
1048 * --------------------
1049 * Reordering metric is maximal distance, which a packet can be displaced
1050 * in packet stream. With SACKs we can estimate it:
1052 * 1. SACK fills old hole and the corresponding segment was not
1053 * ever retransmitted -> reordering. Alas, we cannot use it
1054 * when segment was retransmitted.
1055 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1056 * for retransmitted and already SACKed segment -> reordering..
1057 * Both of these heuristics are not used in Loss state, when we cannot
1058 * account for retransmits accurately.
1060 * SACK block validation.
1061 * ----------------------
1063 * SACK block range validation checks that the received SACK block fits to
1064 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1065 * Note that SND.UNA is not included to the range though being valid because
1066 * it means that the receiver is rather inconsistent with itself reporting
1067 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1068 * perfectly valid, however, in light of RFC2018 which explicitly states
1069 * that "SACK block MUST reflect the newest segment. Even if the newest
1070 * segment is going to be discarded ...", not that it looks very clever
1071 * in case of head skb. Due to potentional receiver driven attacks, we
1072 * choose to avoid immediate execution of a walk in write queue due to
1073 * reneging and defer head skb's loss recovery to standard loss recovery
1074 * procedure that will eventually trigger (nothing forbids us doing this).
1076 * Implements also blockage to start_seq wrap-around. Problem lies in the
1077 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1078 * there's no guarantee that it will be before snd_nxt (n). The problem
1079 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1082 * <- outs wnd -> <- wrapzone ->
1083 * u e n u_w e_w s n_w
1085 * |<------------+------+----- TCP seqno space --------------+---------->|
1086 * ...-- <2^31 ->| |<--------...
1087 * ...---- >2^31 ------>| |<--------...
1089 * Current code wouldn't be vulnerable but it's better still to discard such
1090 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1091 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1092 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1093 * equal to the ideal case (infinite seqno space without wrap caused issues).
1095 * With D-SACK the lower bound is extended to cover sequence space below
1096 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1097 * again, D-SACK block must not to go across snd_una (for the same reason as
1098 * for the normal SACK blocks, explained above). But there all simplicity
1099 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1100 * fully below undo_marker they do not affect behavior in anyway and can
1101 * therefore be safely ignored. In rare cases (which are more or less
1102 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1103 * fragmentation and packet reordering past skb's retransmission. To consider
1104 * them correctly, the acceptable range must be extended even more though
1105 * the exact amount is rather hard to quantify. However, tp->max_window can
1106 * be used as an exaggerated estimate.
1108 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1109 u32 start_seq, u32 end_seq)
1111 /* Too far in future, or reversed (interpretation is ambiguous) */
1112 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1115 /* Nasty start_seq wrap-around check (see comments above) */
1116 if (!before(start_seq, tp->snd_nxt))
1119 /* In outstanding window? ...This is valid exit for D-SACKs too.
1120 * start_seq == snd_una is non-sensical (see comments above)
1122 if (after(start_seq, tp->snd_una))
1125 if (!is_dsack || !tp->undo_marker)
1128 /* ...Then it's D-SACK, and must reside below snd_una completely */
1129 if (!after(end_seq, tp->snd_una))
1132 if (!before(start_seq, tp->undo_marker))
1136 if (!after(end_seq, tp->undo_marker))
1139 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1140 * start_seq < undo_marker and end_seq >= undo_marker.
1142 return !before(start_seq, end_seq - tp->max_window);
1145 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1146 * Event "C". Later note: FACK people cheated me again 8), we have to account
1147 * for reordering! Ugly, but should help.
1149 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1150 * less than what is now known to be received by the other end (derived from
1151 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1152 * retransmitted skbs to avoid some costly processing per ACKs.
1154 static void tcp_mark_lost_retrans(struct sock *sk)
1156 const struct inet_connection_sock *icsk = inet_csk(sk);
1157 struct tcp_sock *tp = tcp_sk(sk);
1158 struct sk_buff *skb;
1160 u32 new_low_seq = tp->snd_nxt;
1161 u32 received_upto = tcp_highest_sack_seq(tp);
1163 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1164 !after(received_upto, tp->lost_retrans_low) ||
1165 icsk->icsk_ca_state != TCP_CA_Recovery)
1168 tcp_for_write_queue(skb, sk) {
1169 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1171 if (skb == tcp_send_head(sk))
1173 if (cnt == tp->retrans_out)
1175 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1178 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1181 if (after(received_upto, ack_seq) &&
1183 !before(received_upto,
1184 ack_seq + tp->reordering * tp->mss_cache))) {
1185 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1186 tp->retrans_out -= tcp_skb_pcount(skb);
1188 tcp_skb_mark_lost_uncond_verify(tp, skb);
1189 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1191 if (before(ack_seq, new_low_seq))
1192 new_low_seq = ack_seq;
1193 cnt += tcp_skb_pcount(skb);
1197 if (tp->retrans_out)
1198 tp->lost_retrans_low = new_low_seq;
1201 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1202 struct tcp_sack_block_wire *sp, int num_sacks,
1205 struct tcp_sock *tp = tcp_sk(sk);
1206 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1207 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1210 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1213 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1214 } else if (num_sacks > 1) {
1215 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1216 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1218 if (!after(end_seq_0, end_seq_1) &&
1219 !before(start_seq_0, start_seq_1)) {
1222 NET_INC_STATS_BH(sock_net(sk),
1223 LINUX_MIB_TCPDSACKOFORECV);
1227 /* D-SACK for already forgotten data... Do dumb counting. */
1229 !after(end_seq_0, prior_snd_una) &&
1230 after(end_seq_0, tp->undo_marker))
1236 struct tcp_sacktag_state {
1242 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1243 * the incoming SACK may not exactly match but we can find smaller MSS
1244 * aligned portion of it that matches. Therefore we might need to fragment
1245 * which may fail and creates some hassle (caller must handle error case
1248 * FIXME: this could be merged to shift decision code
1250 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1251 u32 start_seq, u32 end_seq)
1254 unsigned int pkt_len;
1257 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1258 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1260 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1261 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1262 mss = tcp_skb_mss(skb);
1263 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1266 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1270 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1275 /* Round if necessary so that SACKs cover only full MSSes
1276 * and/or the remaining small portion (if present)
1278 if (pkt_len > mss) {
1279 unsigned int new_len = (pkt_len / mss) * mss;
1280 if (!in_sack && new_len < pkt_len) {
1282 if (new_len > skb->len)
1287 err = tcp_fragment(sk, skb, pkt_len, mss);
1295 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1296 struct tcp_sacktag_state *state,
1297 int dup_sack, int pcount)
1299 struct tcp_sock *tp = tcp_sk(sk);
1300 u8 sacked = TCP_SKB_CB(skb)->sacked;
1301 int fack_count = state->fack_count;
1303 /* Account D-SACK for retransmitted packet. */
1304 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1305 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1307 if (sacked & TCPCB_SACKED_ACKED)
1308 state->reord = min(fack_count, state->reord);
1311 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1312 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1315 if (!(sacked & TCPCB_SACKED_ACKED)) {
1316 if (sacked & TCPCB_SACKED_RETRANS) {
1317 /* If the segment is not tagged as lost,
1318 * we do not clear RETRANS, believing
1319 * that retransmission is still in flight.
1321 if (sacked & TCPCB_LOST) {
1322 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1323 tp->lost_out -= pcount;
1324 tp->retrans_out -= pcount;
1327 if (!(sacked & TCPCB_RETRANS)) {
1328 /* New sack for not retransmitted frame,
1329 * which was in hole. It is reordering.
1331 if (before(TCP_SKB_CB(skb)->seq,
1332 tcp_highest_sack_seq(tp)))
1333 state->reord = min(fack_count,
1336 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1337 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1338 state->flag |= FLAG_ONLY_ORIG_SACKED;
1341 if (sacked & TCPCB_LOST) {
1342 sacked &= ~TCPCB_LOST;
1343 tp->lost_out -= pcount;
1347 sacked |= TCPCB_SACKED_ACKED;
1348 state->flag |= FLAG_DATA_SACKED;
1349 tp->sacked_out += pcount;
1351 fack_count += pcount;
1353 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1354 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1355 before(TCP_SKB_CB(skb)->seq,
1356 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1357 tp->lost_cnt_hint += pcount;
1359 if (fack_count > tp->fackets_out)
1360 tp->fackets_out = fack_count;
1363 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1364 * frames and clear it. undo_retrans is decreased above, L|R frames
1365 * are accounted above as well.
1367 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1368 sacked &= ~TCPCB_SACKED_RETRANS;
1369 tp->retrans_out -= pcount;
1375 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1376 struct tcp_sacktag_state *state,
1377 unsigned int pcount, int shifted, int mss,
1380 struct tcp_sock *tp = tcp_sk(sk);
1381 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1385 /* Tweak before seqno plays */
1386 if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1387 !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1388 tp->lost_cnt_hint += pcount;
1390 TCP_SKB_CB(prev)->end_seq += shifted;
1391 TCP_SKB_CB(skb)->seq += shifted;
1393 skb_shinfo(prev)->gso_segs += pcount;
1394 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1395 skb_shinfo(skb)->gso_segs -= pcount;
1397 /* When we're adding to gso_segs == 1, gso_size will be zero,
1398 * in theory this shouldn't be necessary but as long as DSACK
1399 * code can come after this skb later on it's better to keep
1400 * setting gso_size to something.
1402 if (!skb_shinfo(prev)->gso_size) {
1403 skb_shinfo(prev)->gso_size = mss;
1404 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1407 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1408 if (skb_shinfo(skb)->gso_segs <= 1) {
1409 skb_shinfo(skb)->gso_size = 0;
1410 skb_shinfo(skb)->gso_type = 0;
1413 /* We discard results */
1414 tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1416 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1417 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1420 BUG_ON(!tcp_skb_pcount(skb));
1421 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1425 /* Whole SKB was eaten :-) */
1427 if (skb == tp->retransmit_skb_hint)
1428 tp->retransmit_skb_hint = prev;
1429 if (skb == tp->scoreboard_skb_hint)
1430 tp->scoreboard_skb_hint = prev;
1431 if (skb == tp->lost_skb_hint) {
1432 tp->lost_skb_hint = prev;
1433 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1436 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1437 if (skb == tcp_highest_sack(sk))
1438 tcp_advance_highest_sack(sk, skb);
1440 tcp_unlink_write_queue(skb, sk);
1441 sk_wmem_free_skb(sk, skb);
1443 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1448 /* I wish gso_size would have a bit more sane initialization than
1449 * something-or-zero which complicates things
1451 static int tcp_skb_seglen(struct sk_buff *skb)
1453 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1456 /* Shifting pages past head area doesn't work */
1457 static int skb_can_shift(struct sk_buff *skb)
1459 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1462 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1465 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1466 struct tcp_sacktag_state *state,
1467 u32 start_seq, u32 end_seq,
1470 struct tcp_sock *tp = tcp_sk(sk);
1471 struct sk_buff *prev;
1477 if (!sk_can_gso(sk))
1480 /* Normally R but no L won't result in plain S */
1482 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1484 if (!skb_can_shift(skb))
1486 /* This frame is about to be dropped (was ACKed). */
1487 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1490 /* Can only happen with delayed DSACK + discard craziness */
1491 if (unlikely(skb == tcp_write_queue_head(sk)))
1493 prev = tcp_write_queue_prev(sk, skb);
1495 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1498 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1499 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1503 pcount = tcp_skb_pcount(skb);
1504 mss = tcp_skb_seglen(skb);
1506 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1507 * drop this restriction as unnecessary
1509 if (mss != tcp_skb_seglen(prev))
1512 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1514 /* CHECKME: This is non-MSS split case only?, this will
1515 * cause skipped skbs due to advancing loop btw, original
1516 * has that feature too
1518 if (tcp_skb_pcount(skb) <= 1)
1521 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1523 /* TODO: head merge to next could be attempted here
1524 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1525 * though it might not be worth of the additional hassle
1527 * ...we can probably just fallback to what was done
1528 * previously. We could try merging non-SACKed ones
1529 * as well but it probably isn't going to buy off
1530 * because later SACKs might again split them, and
1531 * it would make skb timestamp tracking considerably
1537 len = end_seq - TCP_SKB_CB(skb)->seq;
1539 BUG_ON(len > skb->len);
1541 /* MSS boundaries should be honoured or else pcount will
1542 * severely break even though it makes things bit trickier.
1543 * Optimize common case to avoid most of the divides
1545 mss = tcp_skb_mss(skb);
1547 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1548 * drop this restriction as unnecessary
1550 if (mss != tcp_skb_seglen(prev))
1555 } else if (len < mss) {
1563 if (!skb_shift(prev, skb, len))
1565 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1568 /* Hole filled allows collapsing with the next as well, this is very
1569 * useful when hole on every nth skb pattern happens
1571 if (prev == tcp_write_queue_tail(sk))
1573 skb = tcp_write_queue_next(sk, prev);
1575 if (!skb_can_shift(skb) ||
1576 (skb == tcp_send_head(sk)) ||
1577 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1578 (mss != tcp_skb_seglen(skb)))
1582 if (skb_shift(prev, skb, len)) {
1583 pcount += tcp_skb_pcount(skb);
1584 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1588 state->fack_count += pcount;
1595 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1599 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1600 struct tcp_sack_block *next_dup,
1601 struct tcp_sacktag_state *state,
1602 u32 start_seq, u32 end_seq,
1605 struct tcp_sock *tp = tcp_sk(sk);
1606 struct sk_buff *tmp;
1608 tcp_for_write_queue_from(skb, sk) {
1610 int dup_sack = dup_sack_in;
1612 if (skb == tcp_send_head(sk))
1615 /* queue is in-order => we can short-circuit the walk early */
1616 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1619 if ((next_dup != NULL) &&
1620 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1621 in_sack = tcp_match_skb_to_sack(sk, skb,
1622 next_dup->start_seq,
1628 /* skb reference here is a bit tricky to get right, since
1629 * shifting can eat and free both this skb and the next,
1630 * so not even _safe variant of the loop is enough.
1633 tmp = tcp_shift_skb_data(sk, skb, state,
1634 start_seq, end_seq, dup_sack);
1643 in_sack = tcp_match_skb_to_sack(sk, skb,
1649 if (unlikely(in_sack < 0))
1653 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1656 tcp_skb_pcount(skb));
1658 if (!before(TCP_SKB_CB(skb)->seq,
1659 tcp_highest_sack_seq(tp)))
1660 tcp_advance_highest_sack(sk, skb);
1663 state->fack_count += tcp_skb_pcount(skb);
1668 /* Avoid all extra work that is being done by sacktag while walking in
1671 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1672 struct tcp_sacktag_state *state,
1675 tcp_for_write_queue_from(skb, sk) {
1676 if (skb == tcp_send_head(sk))
1679 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1682 state->fack_count += tcp_skb_pcount(skb);
1687 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1689 struct tcp_sack_block *next_dup,
1690 struct tcp_sacktag_state *state,
1693 if (next_dup == NULL)
1696 if (before(next_dup->start_seq, skip_to_seq)) {
1697 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1698 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1699 next_dup->start_seq, next_dup->end_seq,
1706 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1708 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1712 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1715 const struct inet_connection_sock *icsk = inet_csk(sk);
1716 struct tcp_sock *tp = tcp_sk(sk);
1717 unsigned char *ptr = (skb_transport_header(ack_skb) +
1718 TCP_SKB_CB(ack_skb)->sacked);
1719 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1720 struct tcp_sack_block sp[TCP_NUM_SACKS];
1721 struct tcp_sack_block *cache;
1722 struct tcp_sacktag_state state;
1723 struct sk_buff *skb;
1724 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1726 int found_dup_sack = 0;
1728 int first_sack_index;
1731 state.reord = tp->packets_out;
1733 if (!tp->sacked_out) {
1734 if (WARN_ON(tp->fackets_out))
1735 tp->fackets_out = 0;
1736 tcp_highest_sack_reset(sk);
1739 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1740 num_sacks, prior_snd_una);
1742 state.flag |= FLAG_DSACKING_ACK;
1744 /* Eliminate too old ACKs, but take into
1745 * account more or less fresh ones, they can
1746 * contain valid SACK info.
1748 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1751 if (!tp->packets_out)
1755 first_sack_index = 0;
1756 for (i = 0; i < num_sacks; i++) {
1757 int dup_sack = !i && found_dup_sack;
1759 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1760 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1762 if (!tcp_is_sackblock_valid(tp, dup_sack,
1763 sp[used_sacks].start_seq,
1764 sp[used_sacks].end_seq)) {
1768 if (!tp->undo_marker)
1769 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1771 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1773 /* Don't count olds caused by ACK reordering */
1774 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1775 !after(sp[used_sacks].end_seq, tp->snd_una))
1777 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1780 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1782 first_sack_index = -1;
1786 /* Ignore very old stuff early */
1787 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1793 /* order SACK blocks to allow in order walk of the retrans queue */
1794 for (i = used_sacks - 1; i > 0; i--) {
1795 for (j = 0; j < i; j++) {
1796 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1797 struct tcp_sack_block tmp;
1803 /* Track where the first SACK block goes to */
1804 if (j == first_sack_index)
1805 first_sack_index = j + 1;
1810 skb = tcp_write_queue_head(sk);
1811 state.fack_count = 0;
1814 if (!tp->sacked_out) {
1815 /* It's already past, so skip checking against it */
1816 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1818 cache = tp->recv_sack_cache;
1819 /* Skip empty blocks in at head of the cache */
1820 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1825 while (i < used_sacks) {
1826 u32 start_seq = sp[i].start_seq;
1827 u32 end_seq = sp[i].end_seq;
1828 int dup_sack = (found_dup_sack && (i == first_sack_index));
1829 struct tcp_sack_block *next_dup = NULL;
1831 if (found_dup_sack && ((i + 1) == first_sack_index))
1832 next_dup = &sp[i + 1];
1834 /* Event "B" in the comment above. */
1835 if (after(end_seq, tp->high_seq))
1836 state.flag |= FLAG_DATA_LOST;
1838 /* Skip too early cached blocks */
1839 while (tcp_sack_cache_ok(tp, cache) &&
1840 !before(start_seq, cache->end_seq))
1843 /* Can skip some work by looking recv_sack_cache? */
1844 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1845 after(end_seq, cache->start_seq)) {
1848 if (before(start_seq, cache->start_seq)) {
1849 skb = tcp_sacktag_skip(skb, sk, &state,
1851 skb = tcp_sacktag_walk(skb, sk, next_dup,
1858 /* Rest of the block already fully processed? */
1859 if (!after(end_seq, cache->end_seq))
1862 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1866 /* ...tail remains todo... */
1867 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1868 /* ...but better entrypoint exists! */
1869 skb = tcp_highest_sack(sk);
1872 state.fack_count = tp->fackets_out;
1877 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1878 /* Check overlap against next cached too (past this one already) */
1883 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1884 skb = tcp_highest_sack(sk);
1887 state.fack_count = tp->fackets_out;
1889 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1892 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1893 start_seq, end_seq, dup_sack);
1896 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1897 * due to in-order walk
1899 if (after(end_seq, tp->frto_highmark))
1900 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1905 /* Clear the head of the cache sack blocks so we can skip it next time */
1906 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1907 tp->recv_sack_cache[i].start_seq = 0;
1908 tp->recv_sack_cache[i].end_seq = 0;
1910 for (j = 0; j < used_sacks; j++)
1911 tp->recv_sack_cache[i++] = sp[j];
1913 tcp_mark_lost_retrans(sk);
1915 tcp_verify_left_out(tp);
1917 if ((state.reord < tp->fackets_out) &&
1918 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1919 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1920 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1924 #if FASTRETRANS_DEBUG > 0
1925 WARN_ON((int)tp->sacked_out < 0);
1926 WARN_ON((int)tp->lost_out < 0);
1927 WARN_ON((int)tp->retrans_out < 0);
1928 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1933 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1934 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1936 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1940 holes = max(tp->lost_out, 1U);
1941 holes = min(holes, tp->packets_out);
1943 if ((tp->sacked_out + holes) > tp->packets_out) {
1944 tp->sacked_out = tp->packets_out - holes;
1950 /* If we receive more dupacks than we expected counting segments
1951 * in assumption of absent reordering, interpret this as reordering.
1952 * The only another reason could be bug in receiver TCP.
1954 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1956 struct tcp_sock *tp = tcp_sk(sk);
1957 if (tcp_limit_reno_sacked(tp))
1958 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1961 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1963 static void tcp_add_reno_sack(struct sock *sk)
1965 struct tcp_sock *tp = tcp_sk(sk);
1967 tcp_check_reno_reordering(sk, 0);
1968 tcp_verify_left_out(tp);
1971 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1973 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1975 struct tcp_sock *tp = tcp_sk(sk);
1978 /* One ACK acked hole. The rest eat duplicate ACKs. */
1979 if (acked - 1 >= tp->sacked_out)
1982 tp->sacked_out -= acked - 1;
1984 tcp_check_reno_reordering(sk, acked);
1985 tcp_verify_left_out(tp);
1988 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1993 static int tcp_is_sackfrto(const struct tcp_sock *tp)
1995 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1998 /* F-RTO can only be used if TCP has never retransmitted anything other than
1999 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2001 int tcp_use_frto(struct sock *sk)
2003 const struct tcp_sock *tp = tcp_sk(sk);
2004 const struct inet_connection_sock *icsk = inet_csk(sk);
2005 struct sk_buff *skb;
2007 if (!sysctl_tcp_frto)
2010 /* MTU probe and F-RTO won't really play nicely along currently */
2011 if (icsk->icsk_mtup.probe_size)
2014 if (tcp_is_sackfrto(tp))
2017 /* Avoid expensive walking of rexmit queue if possible */
2018 if (tp->retrans_out > 1)
2021 skb = tcp_write_queue_head(sk);
2022 if (tcp_skb_is_last(sk, skb))
2024 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2025 tcp_for_write_queue_from(skb, sk) {
2026 if (skb == tcp_send_head(sk))
2028 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2030 /* Short-circuit when first non-SACKed skb has been checked */
2031 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2037 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2038 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2039 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2040 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2041 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2042 * bits are handled if the Loss state is really to be entered (in
2043 * tcp_enter_frto_loss).
2045 * Do like tcp_enter_loss() would; when RTO expires the second time it
2047 * "Reduce ssthresh if it has not yet been made inside this window."
2049 void tcp_enter_frto(struct sock *sk)
2051 const struct inet_connection_sock *icsk = inet_csk(sk);
2052 struct tcp_sock *tp = tcp_sk(sk);
2053 struct sk_buff *skb;
2055 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2056 tp->snd_una == tp->high_seq ||
2057 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2058 !icsk->icsk_retransmits)) {
2059 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2060 /* Our state is too optimistic in ssthresh() call because cwnd
2061 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2062 * recovery has not yet completed. Pattern would be this: RTO,
2063 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2065 * RFC4138 should be more specific on what to do, even though
2066 * RTO is quite unlikely to occur after the first Cumulative ACK
2067 * due to back-off and complexity of triggering events ...
2069 if (tp->frto_counter) {
2071 stored_cwnd = tp->snd_cwnd;
2073 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2074 tp->snd_cwnd = stored_cwnd;
2076 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2078 /* ... in theory, cong.control module could do "any tricks" in
2079 * ssthresh(), which means that ca_state, lost bits and lost_out
2080 * counter would have to be faked before the call occurs. We
2081 * consider that too expensive, unlikely and hacky, so modules
2082 * using these in ssthresh() must deal these incompatibility
2083 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2085 tcp_ca_event(sk, CA_EVENT_FRTO);
2088 tp->undo_marker = tp->snd_una;
2089 tp->undo_retrans = 0;
2091 skb = tcp_write_queue_head(sk);
2092 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2093 tp->undo_marker = 0;
2094 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2095 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2096 tp->retrans_out -= tcp_skb_pcount(skb);
2098 tcp_verify_left_out(tp);
2100 /* Too bad if TCP was application limited */
2101 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2103 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2104 * The last condition is necessary at least in tp->frto_counter case.
2106 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2107 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2108 after(tp->high_seq, tp->snd_una)) {
2109 tp->frto_highmark = tp->high_seq;
2111 tp->frto_highmark = tp->snd_nxt;
2113 tcp_set_ca_state(sk, TCP_CA_Disorder);
2114 tp->high_seq = tp->snd_nxt;
2115 tp->frto_counter = 1;
2118 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2119 * which indicates that we should follow the traditional RTO recovery,
2120 * i.e. mark everything lost and do go-back-N retransmission.
2122 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2124 struct tcp_sock *tp = tcp_sk(sk);
2125 struct sk_buff *skb;
2128 tp->retrans_out = 0;
2129 if (tcp_is_reno(tp))
2130 tcp_reset_reno_sack(tp);
2132 tcp_for_write_queue(skb, sk) {
2133 if (skb == tcp_send_head(sk))
2136 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2138 * Count the retransmission made on RTO correctly (only when
2139 * waiting for the first ACK and did not get it)...
2141 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2142 /* For some reason this R-bit might get cleared? */
2143 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2144 tp->retrans_out += tcp_skb_pcount(skb);
2145 /* ...enter this if branch just for the first segment */
2146 flag |= FLAG_DATA_ACKED;
2148 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2149 tp->undo_marker = 0;
2150 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2153 /* Marking forward transmissions that were made after RTO lost
2154 * can cause unnecessary retransmissions in some scenarios,
2155 * SACK blocks will mitigate that in some but not in all cases.
2156 * We used to not mark them but it was causing break-ups with
2157 * receivers that do only in-order receival.
2159 * TODO: we could detect presence of such receiver and select
2160 * different behavior per flow.
2162 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2163 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2164 tp->lost_out += tcp_skb_pcount(skb);
2165 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2168 tcp_verify_left_out(tp);
2170 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2171 tp->snd_cwnd_cnt = 0;
2172 tp->snd_cwnd_stamp = tcp_time_stamp;
2173 tp->frto_counter = 0;
2174 tp->bytes_acked = 0;
2176 tp->reordering = min_t(unsigned int, tp->reordering,
2177 sysctl_tcp_reordering);
2178 tcp_set_ca_state(sk, TCP_CA_Loss);
2179 tp->high_seq = tp->snd_nxt;
2180 TCP_ECN_queue_cwr(tp);
2182 tcp_clear_all_retrans_hints(tp);
2185 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2187 tp->retrans_out = 0;
2190 tp->undo_marker = 0;
2191 tp->undo_retrans = 0;
2194 void tcp_clear_retrans(struct tcp_sock *tp)
2196 tcp_clear_retrans_partial(tp);
2198 tp->fackets_out = 0;
2202 /* Enter Loss state. If "how" is not zero, forget all SACK information
2203 * and reset tags completely, otherwise preserve SACKs. If receiver
2204 * dropped its ofo queue, we will know this due to reneging detection.
2206 void tcp_enter_loss(struct sock *sk, int how)
2208 const struct inet_connection_sock *icsk = inet_csk(sk);
2209 struct tcp_sock *tp = tcp_sk(sk);
2210 struct sk_buff *skb;
2212 /* Reduce ssthresh if it has not yet been made inside this window. */
2213 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2214 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2215 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2216 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2217 tcp_ca_event(sk, CA_EVENT_LOSS);
2220 tp->snd_cwnd_cnt = 0;
2221 tp->snd_cwnd_stamp = tcp_time_stamp;
2223 tp->bytes_acked = 0;
2224 tcp_clear_retrans_partial(tp);
2226 if (tcp_is_reno(tp))
2227 tcp_reset_reno_sack(tp);
2230 /* Push undo marker, if it was plain RTO and nothing
2231 * was retransmitted. */
2232 tp->undo_marker = tp->snd_una;
2235 tp->fackets_out = 0;
2237 tcp_clear_all_retrans_hints(tp);
2239 tcp_for_write_queue(skb, sk) {
2240 if (skb == tcp_send_head(sk))
2243 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2244 tp->undo_marker = 0;
2245 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2246 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2247 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2248 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2249 tp->lost_out += tcp_skb_pcount(skb);
2250 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2253 tcp_verify_left_out(tp);
2255 tp->reordering = min_t(unsigned int, tp->reordering,
2256 sysctl_tcp_reordering);
2257 tcp_set_ca_state(sk, TCP_CA_Loss);
2258 tp->high_seq = tp->snd_nxt;
2259 TCP_ECN_queue_cwr(tp);
2260 /* Abort F-RTO algorithm if one is in progress */
2261 tp->frto_counter = 0;
2264 /* If ACK arrived pointing to a remembered SACK, it means that our
2265 * remembered SACKs do not reflect real state of receiver i.e.
2266 * receiver _host_ is heavily congested (or buggy).
2268 * Do processing similar to RTO timeout.
2270 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2272 if (flag & FLAG_SACK_RENEGING) {
2273 struct inet_connection_sock *icsk = inet_csk(sk);
2274 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2276 tcp_enter_loss(sk, 1);
2277 icsk->icsk_retransmits++;
2278 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2279 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2280 icsk->icsk_rto, TCP_RTO_MAX);
2286 static inline int tcp_fackets_out(struct tcp_sock *tp)
2288 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2291 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2292 * counter when SACK is enabled (without SACK, sacked_out is used for
2295 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2296 * segments up to the highest received SACK block so far and holes in
2299 * With reordering, holes may still be in flight, so RFC3517 recovery
2300 * uses pure sacked_out (total number of SACKed segments) even though
2301 * it violates the RFC that uses duplicate ACKs, often these are equal
2302 * but when e.g. out-of-window ACKs or packet duplication occurs,
2303 * they differ. Since neither occurs due to loss, TCP should really
2306 static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2308 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2311 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2313 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
2316 static inline int tcp_head_timedout(struct sock *sk)
2318 struct tcp_sock *tp = tcp_sk(sk);
2320 return tp->packets_out &&
2321 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2324 /* Linux NewReno/SACK/FACK/ECN state machine.
2325 * --------------------------------------
2327 * "Open" Normal state, no dubious events, fast path.
2328 * "Disorder" In all the respects it is "Open",
2329 * but requires a bit more attention. It is entered when
2330 * we see some SACKs or dupacks. It is split of "Open"
2331 * mainly to move some processing from fast path to slow one.
2332 * "CWR" CWND was reduced due to some Congestion Notification event.
2333 * It can be ECN, ICMP source quench, local device congestion.
2334 * "Recovery" CWND was reduced, we are fast-retransmitting.
2335 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2337 * tcp_fastretrans_alert() is entered:
2338 * - each incoming ACK, if state is not "Open"
2339 * - when arrived ACK is unusual, namely:
2344 * Counting packets in flight is pretty simple.
2346 * in_flight = packets_out - left_out + retrans_out
2348 * packets_out is SND.NXT-SND.UNA counted in packets.
2350 * retrans_out is number of retransmitted segments.
2352 * left_out is number of segments left network, but not ACKed yet.
2354 * left_out = sacked_out + lost_out
2356 * sacked_out: Packets, which arrived to receiver out of order
2357 * and hence not ACKed. With SACKs this number is simply
2358 * amount of SACKed data. Even without SACKs
2359 * it is easy to give pretty reliable estimate of this number,
2360 * counting duplicate ACKs.
2362 * lost_out: Packets lost by network. TCP has no explicit
2363 * "loss notification" feedback from network (for now).
2364 * It means that this number can be only _guessed_.
2365 * Actually, it is the heuristics to predict lossage that
2366 * distinguishes different algorithms.
2368 * F.e. after RTO, when all the queue is considered as lost,
2369 * lost_out = packets_out and in_flight = retrans_out.
2371 * Essentially, we have now two algorithms counting
2374 * FACK: It is the simplest heuristics. As soon as we decided
2375 * that something is lost, we decide that _all_ not SACKed
2376 * packets until the most forward SACK are lost. I.e.
2377 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2378 * It is absolutely correct estimate, if network does not reorder
2379 * packets. And it loses any connection to reality when reordering
2380 * takes place. We use FACK by default until reordering
2381 * is suspected on the path to this destination.
2383 * NewReno: when Recovery is entered, we assume that one segment
2384 * is lost (classic Reno). While we are in Recovery and
2385 * a partial ACK arrives, we assume that one more packet
2386 * is lost (NewReno). This heuristics are the same in NewReno
2389 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2390 * deflation etc. CWND is real congestion window, never inflated, changes
2391 * only according to classic VJ rules.
2393 * Really tricky (and requiring careful tuning) part of algorithm
2394 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2395 * The first determines the moment _when_ we should reduce CWND and,
2396 * hence, slow down forward transmission. In fact, it determines the moment
2397 * when we decide that hole is caused by loss, rather than by a reorder.
2399 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2400 * holes, caused by lost packets.
2402 * And the most logically complicated part of algorithm is undo
2403 * heuristics. We detect false retransmits due to both too early
2404 * fast retransmit (reordering) and underestimated RTO, analyzing
2405 * timestamps and D-SACKs. When we detect that some segments were
2406 * retransmitted by mistake and CWND reduction was wrong, we undo
2407 * window reduction and abort recovery phase. This logic is hidden
2408 * inside several functions named tcp_try_undo_<something>.
2411 /* This function decides, when we should leave Disordered state
2412 * and enter Recovery phase, reducing congestion window.
2414 * Main question: may we further continue forward transmission
2415 * with the same cwnd?
2417 static int tcp_time_to_recover(struct sock *sk)
2419 struct tcp_sock *tp = tcp_sk(sk);
2422 /* Do not perform any recovery during F-RTO algorithm */
2423 if (tp->frto_counter)
2426 /* Trick#1: The loss is proven. */
2430 /* Not-A-Trick#2 : Classic rule... */
2431 if (tcp_dupack_heurestics(tp) > tp->reordering)
2434 /* Trick#3 : when we use RFC2988 timer restart, fast
2435 * retransmit can be triggered by timeout of queue head.
2437 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2440 /* Trick#4: It is still not OK... But will it be useful to delay
2443 packets_out = tp->packets_out;
2444 if (packets_out <= tp->reordering &&
2445 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2446 !tcp_may_send_now(sk)) {
2447 /* We have nothing to send. This connection is limited
2448 * either by receiver window or by application.
2456 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2457 * is against sacked "cnt", otherwise it's against facked "cnt"
2459 static void tcp_mark_head_lost(struct sock *sk, int packets)
2461 struct tcp_sock *tp = tcp_sk(sk);
2462 struct sk_buff *skb;
2467 WARN_ON(packets > tp->packets_out);
2468 if (tp->lost_skb_hint) {
2469 skb = tp->lost_skb_hint;
2470 cnt = tp->lost_cnt_hint;
2472 skb = tcp_write_queue_head(sk);
2476 tcp_for_write_queue_from(skb, sk) {
2477 if (skb == tcp_send_head(sk))
2479 /* TODO: do this better */
2480 /* this is not the most efficient way to do this... */
2481 tp->lost_skb_hint = skb;
2482 tp->lost_cnt_hint = cnt;
2484 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2488 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2489 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2490 cnt += tcp_skb_pcount(skb);
2492 if (cnt > packets) {
2493 if (tcp_is_sack(tp) || (oldcnt >= packets))
2496 mss = skb_shinfo(skb)->gso_size;
2497 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2503 tcp_skb_mark_lost(tp, skb);
2505 tcp_verify_left_out(tp);
2508 /* Account newly detected lost packet(s) */
2510 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2512 struct tcp_sock *tp = tcp_sk(sk);
2514 if (tcp_is_reno(tp)) {
2515 tcp_mark_head_lost(sk, 1);
2516 } else if (tcp_is_fack(tp)) {
2517 int lost = tp->fackets_out - tp->reordering;
2520 tcp_mark_head_lost(sk, lost);
2522 int sacked_upto = tp->sacked_out - tp->reordering;
2523 if (sacked_upto < fast_rexmit)
2524 sacked_upto = fast_rexmit;
2525 tcp_mark_head_lost(sk, sacked_upto);
2528 /* New heuristics: it is possible only after we switched
2529 * to restart timer each time when something is ACKed.
2530 * Hence, we can detect timed out packets during fast
2531 * retransmit without falling to slow start.
2533 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
2534 struct sk_buff *skb;
2536 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
2537 : tcp_write_queue_head(sk);
2539 tcp_for_write_queue_from(skb, sk) {
2540 if (skb == tcp_send_head(sk))
2542 if (!tcp_skb_timedout(sk, skb))
2545 tcp_skb_mark_lost(tp, skb);
2548 tp->scoreboard_skb_hint = skb;
2550 tcp_verify_left_out(tp);
2554 /* CWND moderation, preventing bursts due to too big ACKs
2555 * in dubious situations.
2557 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2559 tp->snd_cwnd = min(tp->snd_cwnd,
2560 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2561 tp->snd_cwnd_stamp = tcp_time_stamp;
2564 /* Lower bound on congestion window is slow start threshold
2565 * unless congestion avoidance choice decides to overide it.
2567 static inline u32 tcp_cwnd_min(const struct sock *sk)
2569 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2571 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2574 /* Decrease cwnd each second ack. */
2575 static void tcp_cwnd_down(struct sock *sk, int flag)
2577 struct tcp_sock *tp = tcp_sk(sk);
2578 int decr = tp->snd_cwnd_cnt + 1;
2580 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2581 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2582 tp->snd_cwnd_cnt = decr & 1;
2585 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2586 tp->snd_cwnd -= decr;
2588 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2589 tp->snd_cwnd_stamp = tcp_time_stamp;
2593 /* Nothing was retransmitted or returned timestamp is less
2594 * than timestamp of the first retransmission.
2596 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2598 return !tp->retrans_stamp ||
2599 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2600 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2603 /* Undo procedures. */
2605 #if FASTRETRANS_DEBUG > 1
2606 static void DBGUNDO(struct sock *sk, const char *msg)
2608 struct tcp_sock *tp = tcp_sk(sk);
2609 struct inet_sock *inet = inet_sk(sk);
2611 if (sk->sk_family == AF_INET) {
2612 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2614 &inet->daddr, ntohs(inet->dport),
2615 tp->snd_cwnd, tcp_left_out(tp),
2616 tp->snd_ssthresh, tp->prior_ssthresh,
2619 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2620 else if (sk->sk_family == AF_INET6) {
2621 struct ipv6_pinfo *np = inet6_sk(sk);
2622 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2624 &np->daddr, ntohs(inet->dport),
2625 tp->snd_cwnd, tcp_left_out(tp),
2626 tp->snd_ssthresh, tp->prior_ssthresh,
2632 #define DBGUNDO(x...) do { } while (0)
2635 static void tcp_undo_cwr(struct sock *sk, const int undo)
2637 struct tcp_sock *tp = tcp_sk(sk);
2639 if (tp->prior_ssthresh) {
2640 const struct inet_connection_sock *icsk = inet_csk(sk);
2642 if (icsk->icsk_ca_ops->undo_cwnd)
2643 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2645 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2647 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2648 tp->snd_ssthresh = tp->prior_ssthresh;
2649 TCP_ECN_withdraw_cwr(tp);
2652 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2654 tcp_moderate_cwnd(tp);
2655 tp->snd_cwnd_stamp = tcp_time_stamp;
2658 static inline int tcp_may_undo(struct tcp_sock *tp)
2660 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2663 /* People celebrate: "We love our President!" */
2664 static int tcp_try_undo_recovery(struct sock *sk)
2666 struct tcp_sock *tp = tcp_sk(sk);
2668 if (tcp_may_undo(tp)) {
2671 /* Happy end! We did not retransmit anything
2672 * or our original transmission succeeded.
2674 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2675 tcp_undo_cwr(sk, 1);
2676 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2677 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2679 mib_idx = LINUX_MIB_TCPFULLUNDO;
2681 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2682 tp->undo_marker = 0;
2684 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2685 /* Hold old state until something *above* high_seq
2686 * is ACKed. For Reno it is MUST to prevent false
2687 * fast retransmits (RFC2582). SACK TCP is safe. */
2688 tcp_moderate_cwnd(tp);
2691 tcp_set_ca_state(sk, TCP_CA_Open);
2695 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2696 static void tcp_try_undo_dsack(struct sock *sk)
2698 struct tcp_sock *tp = tcp_sk(sk);
2700 if (tp->undo_marker && !tp->undo_retrans) {
2701 DBGUNDO(sk, "D-SACK");
2702 tcp_undo_cwr(sk, 1);
2703 tp->undo_marker = 0;
2704 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2708 /* Undo during fast recovery after partial ACK. */
2710 static int tcp_try_undo_partial(struct sock *sk, int acked)
2712 struct tcp_sock *tp = tcp_sk(sk);
2713 /* Partial ACK arrived. Force Hoe's retransmit. */
2714 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2716 if (tcp_may_undo(tp)) {
2717 /* Plain luck! Hole if filled with delayed
2718 * packet, rather than with a retransmit.
2720 if (tp->retrans_out == 0)
2721 tp->retrans_stamp = 0;
2723 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2726 tcp_undo_cwr(sk, 0);
2727 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2729 /* So... Do not make Hoe's retransmit yet.
2730 * If the first packet was delayed, the rest
2731 * ones are most probably delayed as well.
2738 /* Undo during loss recovery after partial ACK. */
2739 static int tcp_try_undo_loss(struct sock *sk)
2741 struct tcp_sock *tp = tcp_sk(sk);
2743 if (tcp_may_undo(tp)) {
2744 struct sk_buff *skb;
2745 tcp_for_write_queue(skb, sk) {
2746 if (skb == tcp_send_head(sk))
2748 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2751 tcp_clear_all_retrans_hints(tp);
2753 DBGUNDO(sk, "partial loss");
2755 tcp_undo_cwr(sk, 1);
2756 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2757 inet_csk(sk)->icsk_retransmits = 0;
2758 tp->undo_marker = 0;
2759 if (tcp_is_sack(tp))
2760 tcp_set_ca_state(sk, TCP_CA_Open);
2766 static inline void tcp_complete_cwr(struct sock *sk)
2768 struct tcp_sock *tp = tcp_sk(sk);
2769 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2770 tp->snd_cwnd_stamp = tcp_time_stamp;
2771 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2774 static void tcp_try_keep_open(struct sock *sk)
2776 struct tcp_sock *tp = tcp_sk(sk);
2777 int state = TCP_CA_Open;
2779 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2780 state = TCP_CA_Disorder;
2782 if (inet_csk(sk)->icsk_ca_state != state) {
2783 tcp_set_ca_state(sk, state);
2784 tp->high_seq = tp->snd_nxt;
2788 static void tcp_try_to_open(struct sock *sk, int flag)
2790 struct tcp_sock *tp = tcp_sk(sk);
2792 tcp_verify_left_out(tp);
2794 if (!tp->frto_counter && tp->retrans_out == 0)
2795 tp->retrans_stamp = 0;
2797 if (flag & FLAG_ECE)
2798 tcp_enter_cwr(sk, 1);
2800 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2801 tcp_try_keep_open(sk);
2802 tcp_moderate_cwnd(tp);
2804 tcp_cwnd_down(sk, flag);
2808 static void tcp_mtup_probe_failed(struct sock *sk)
2810 struct inet_connection_sock *icsk = inet_csk(sk);
2812 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2813 icsk->icsk_mtup.probe_size = 0;
2816 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2818 struct tcp_sock *tp = tcp_sk(sk);
2819 struct inet_connection_sock *icsk = inet_csk(sk);
2821 /* FIXME: breaks with very large cwnd */
2822 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2823 tp->snd_cwnd = tp->snd_cwnd *
2824 tcp_mss_to_mtu(sk, tp->mss_cache) /
2825 icsk->icsk_mtup.probe_size;
2826 tp->snd_cwnd_cnt = 0;
2827 tp->snd_cwnd_stamp = tcp_time_stamp;
2828 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2830 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2831 icsk->icsk_mtup.probe_size = 0;
2832 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2835 /* Do a simple retransmit without using the backoff mechanisms in
2836 * tcp_timer. This is used for path mtu discovery.
2837 * The socket is already locked here.
2839 void tcp_simple_retransmit(struct sock *sk)
2841 const struct inet_connection_sock *icsk = inet_csk(sk);
2842 struct tcp_sock *tp = tcp_sk(sk);
2843 struct sk_buff *skb;
2844 unsigned int mss = tcp_current_mss(sk, 0);
2845 u32 prior_lost = tp->lost_out;
2847 tcp_for_write_queue(skb, sk) {
2848 if (skb == tcp_send_head(sk))
2850 if (tcp_skb_seglen(skb) > mss &&
2851 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2852 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2853 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2854 tp->retrans_out -= tcp_skb_pcount(skb);
2856 tcp_skb_mark_lost_uncond_verify(tp, skb);
2860 tcp_clear_retrans_hints_partial(tp);
2862 if (prior_lost == tp->lost_out)
2865 if (tcp_is_reno(tp))
2866 tcp_limit_reno_sacked(tp);
2868 tcp_verify_left_out(tp);
2870 /* Don't muck with the congestion window here.
2871 * Reason is that we do not increase amount of _data_
2872 * in network, but units changed and effective
2873 * cwnd/ssthresh really reduced now.
2875 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2876 tp->high_seq = tp->snd_nxt;
2877 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2878 tp->prior_ssthresh = 0;
2879 tp->undo_marker = 0;
2880 tcp_set_ca_state(sk, TCP_CA_Loss);
2882 tcp_xmit_retransmit_queue(sk);
2885 /* Process an event, which can update packets-in-flight not trivially.
2886 * Main goal of this function is to calculate new estimate for left_out,
2887 * taking into account both packets sitting in receiver's buffer and
2888 * packets lost by network.
2890 * Besides that it does CWND reduction, when packet loss is detected
2891 * and changes state of machine.
2893 * It does _not_ decide what to send, it is made in function
2894 * tcp_xmit_retransmit_queue().
2896 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
2898 struct inet_connection_sock *icsk = inet_csk(sk);
2899 struct tcp_sock *tp = tcp_sk(sk);
2900 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2901 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2902 (tcp_fackets_out(tp) > tp->reordering));
2903 int fast_rexmit = 0, mib_idx;
2905 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2907 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2908 tp->fackets_out = 0;
2910 /* Now state machine starts.
2911 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2912 if (flag & FLAG_ECE)
2913 tp->prior_ssthresh = 0;
2915 /* B. In all the states check for reneging SACKs. */
2916 if (tcp_check_sack_reneging(sk, flag))
2919 /* C. Process data loss notification, provided it is valid. */
2920 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
2921 before(tp->snd_una, tp->high_seq) &&
2922 icsk->icsk_ca_state != TCP_CA_Open &&
2923 tp->fackets_out > tp->reordering) {
2924 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
2925 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
2928 /* D. Check consistency of the current state. */
2929 tcp_verify_left_out(tp);
2931 /* E. Check state exit conditions. State can be terminated
2932 * when high_seq is ACKed. */
2933 if (icsk->icsk_ca_state == TCP_CA_Open) {
2934 WARN_ON(tp->retrans_out != 0);
2935 tp->retrans_stamp = 0;
2936 } else if (!before(tp->snd_una, tp->high_seq)) {
2937 switch (icsk->icsk_ca_state) {
2939 icsk->icsk_retransmits = 0;
2940 if (tcp_try_undo_recovery(sk))
2945 /* CWR is to be held something *above* high_seq
2946 * is ACKed for CWR bit to reach receiver. */
2947 if (tp->snd_una != tp->high_seq) {
2948 tcp_complete_cwr(sk);
2949 tcp_set_ca_state(sk, TCP_CA_Open);
2953 case TCP_CA_Disorder:
2954 tcp_try_undo_dsack(sk);
2955 if (!tp->undo_marker ||
2956 /* For SACK case do not Open to allow to undo
2957 * catching for all duplicate ACKs. */
2958 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
2959 tp->undo_marker = 0;
2960 tcp_set_ca_state(sk, TCP_CA_Open);
2964 case TCP_CA_Recovery:
2965 if (tcp_is_reno(tp))
2966 tcp_reset_reno_sack(tp);
2967 if (tcp_try_undo_recovery(sk))
2969 tcp_complete_cwr(sk);
2974 /* F. Process state. */
2975 switch (icsk->icsk_ca_state) {
2976 case TCP_CA_Recovery:
2977 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2978 if (tcp_is_reno(tp) && is_dupack)
2979 tcp_add_reno_sack(sk);
2981 do_lost = tcp_try_undo_partial(sk, pkts_acked);
2984 if (flag & FLAG_DATA_ACKED)
2985 icsk->icsk_retransmits = 0;
2986 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2987 tcp_reset_reno_sack(tp);
2988 if (!tcp_try_undo_loss(sk)) {
2989 tcp_moderate_cwnd(tp);
2990 tcp_xmit_retransmit_queue(sk);
2993 if (icsk->icsk_ca_state != TCP_CA_Open)
2995 /* Loss is undone; fall through to processing in Open state. */
2997 if (tcp_is_reno(tp)) {
2998 if (flag & FLAG_SND_UNA_ADVANCED)
2999 tcp_reset_reno_sack(tp);
3001 tcp_add_reno_sack(sk);
3004 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3005 tcp_try_undo_dsack(sk);
3007 if (!tcp_time_to_recover(sk)) {
3008 tcp_try_to_open(sk, flag);
3012 /* MTU probe failure: don't reduce cwnd */
3013 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3014 icsk->icsk_mtup.probe_size &&
3015 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3016 tcp_mtup_probe_failed(sk);
3017 /* Restores the reduction we did in tcp_mtup_probe() */
3019 tcp_simple_retransmit(sk);
3023 /* Otherwise enter Recovery state */
3025 if (tcp_is_reno(tp))
3026 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3028 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3030 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3032 tp->high_seq = tp->snd_nxt;
3033 tp->prior_ssthresh = 0;
3034 tp->undo_marker = tp->snd_una;
3035 tp->undo_retrans = tp->retrans_out;
3037 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3038 if (!(flag & FLAG_ECE))
3039 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3040 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3041 TCP_ECN_queue_cwr(tp);
3044 tp->bytes_acked = 0;
3045 tp->snd_cwnd_cnt = 0;
3046 tcp_set_ca_state(sk, TCP_CA_Recovery);
3050 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3051 tcp_update_scoreboard(sk, fast_rexmit);
3052 tcp_cwnd_down(sk, flag);
3053 tcp_xmit_retransmit_queue(sk);
3056 static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3058 tcp_rtt_estimator(sk, seq_rtt);
3060 inet_csk(sk)->icsk_backoff = 0;
3063 /* Read draft-ietf-tcplw-high-performance before mucking
3064 * with this code. (Supersedes RFC1323)
3066 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3068 /* RTTM Rule: A TSecr value received in a segment is used to
3069 * update the averaged RTT measurement only if the segment
3070 * acknowledges some new data, i.e., only if it advances the
3071 * left edge of the send window.
3073 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3074 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3076 * Changed: reset backoff as soon as we see the first valid sample.
3077 * If we do not, we get strongly overestimated rto. With timestamps
3078 * samples are accepted even from very old segments: f.e., when rtt=1
3079 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3080 * answer arrives rto becomes 120 seconds! If at least one of segments
3081 * in window is lost... Voila. --ANK (010210)
3083 struct tcp_sock *tp = tcp_sk(sk);
3085 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3088 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3090 /* We don't have a timestamp. Can only use
3091 * packets that are not retransmitted to determine
3092 * rtt estimates. Also, we must not reset the
3093 * backoff for rto until we get a non-retransmitted
3094 * packet. This allows us to deal with a situation
3095 * where the network delay has increased suddenly.
3096 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3099 if (flag & FLAG_RETRANS_DATA_ACKED)
3102 tcp_valid_rtt_meas(sk, seq_rtt);
3105 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3108 const struct tcp_sock *tp = tcp_sk(sk);
3109 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3110 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3111 tcp_ack_saw_tstamp(sk, flag);
3112 else if (seq_rtt >= 0)
3113 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3116 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3118 const struct inet_connection_sock *icsk = inet_csk(sk);
3119 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3120 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3123 /* Restart timer after forward progress on connection.
3124 * RFC2988 recommends to restart timer to now+rto.
3126 static void tcp_rearm_rto(struct sock *sk)
3128 struct tcp_sock *tp = tcp_sk(sk);
3130 if (!tp->packets_out) {
3131 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3133 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3134 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3138 /* If we get here, the whole TSO packet has not been acked. */
3139 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3141 struct tcp_sock *tp = tcp_sk(sk);
3144 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3146 packets_acked = tcp_skb_pcount(skb);
3147 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3149 packets_acked -= tcp_skb_pcount(skb);
3151 if (packets_acked) {
3152 BUG_ON(tcp_skb_pcount(skb) == 0);
3153 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3156 return packets_acked;
3159 /* Remove acknowledged frames from the retransmission queue. If our packet
3160 * is before the ack sequence we can discard it as it's confirmed to have
3161 * arrived at the other end.
3163 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3166 struct tcp_sock *tp = tcp_sk(sk);
3167 const struct inet_connection_sock *icsk = inet_csk(sk);
3168 struct sk_buff *skb;
3169 u32 now = tcp_time_stamp;
3170 int fully_acked = 1;
3173 u32 reord = tp->packets_out;
3174 u32 prior_sacked = tp->sacked_out;
3176 s32 ca_seq_rtt = -1;
3177 ktime_t last_ackt = net_invalid_timestamp();
3179 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3180 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3183 u8 sacked = scb->sacked;
3185 /* Determine how many packets and what bytes were acked, tso and else */
3186 if (after(scb->end_seq, tp->snd_una)) {
3187 if (tcp_skb_pcount(skb) == 1 ||
3188 !after(tp->snd_una, scb->seq))
3191 acked_pcount = tcp_tso_acked(sk, skb);
3196 end_seq = tp->snd_una;
3198 acked_pcount = tcp_skb_pcount(skb);
3199 end_seq = scb->end_seq;
3202 /* MTU probing checks */
3203 if (fully_acked && icsk->icsk_mtup.probe_size &&
3204 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
3205 tcp_mtup_probe_success(sk, skb);
3208 if (sacked & TCPCB_RETRANS) {
3209 if (sacked & TCPCB_SACKED_RETRANS)
3210 tp->retrans_out -= acked_pcount;
3211 flag |= FLAG_RETRANS_DATA_ACKED;
3214 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3215 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3217 ca_seq_rtt = now - scb->when;
3218 last_ackt = skb->tstamp;
3220 seq_rtt = ca_seq_rtt;
3222 if (!(sacked & TCPCB_SACKED_ACKED))
3223 reord = min(pkts_acked, reord);
3226 if (sacked & TCPCB_SACKED_ACKED)
3227 tp->sacked_out -= acked_pcount;
3228 if (sacked & TCPCB_LOST)
3229 tp->lost_out -= acked_pcount;
3231 tp->packets_out -= acked_pcount;
3232 pkts_acked += acked_pcount;
3234 /* Initial outgoing SYN's get put onto the write_queue
3235 * just like anything else we transmit. It is not
3236 * true data, and if we misinform our callers that
3237 * this ACK acks real data, we will erroneously exit
3238 * connection startup slow start one packet too
3239 * quickly. This is severely frowned upon behavior.
3241 if (!(scb->flags & TCPCB_FLAG_SYN)) {
3242 flag |= FLAG_DATA_ACKED;
3244 flag |= FLAG_SYN_ACKED;
3245 tp->retrans_stamp = 0;
3251 tcp_unlink_write_queue(skb, sk);
3252 sk_wmem_free_skb(sk, skb);
3253 tp->scoreboard_skb_hint = NULL;
3254 if (skb == tp->retransmit_skb_hint)
3255 tp->retransmit_skb_hint = NULL;
3256 if (skb == tp->lost_skb_hint)
3257 tp->lost_skb_hint = NULL;
3260 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3261 tp->snd_up = tp->snd_una;
3263 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3264 flag |= FLAG_SACK_RENEGING;
3266 if (flag & FLAG_ACKED) {
3267 const struct tcp_congestion_ops *ca_ops
3268 = inet_csk(sk)->icsk_ca_ops;
3270 tcp_ack_update_rtt(sk, flag, seq_rtt);
3273 if (tcp_is_reno(tp)) {
3274 tcp_remove_reno_sacks(sk, pkts_acked);
3276 /* Non-retransmitted hole got filled? That's reordering */
3277 if (reord < prior_fackets)
3278 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3280 /* No need to care for underflows here because
3281 * the lost_skb_hint gets NULLed if we're past it
3282 * (or something non-trivial happened)
3284 if (tcp_is_fack(tp))
3285 tp->lost_cnt_hint -= pkts_acked;
3287 tp->lost_cnt_hint -= prior_sacked - tp->sacked_out;
3290 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3292 if (ca_ops->pkts_acked) {
3295 /* Is the ACK triggering packet unambiguous? */
3296 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3297 /* High resolution needed and available? */
3298 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3299 !ktime_equal(last_ackt,
3300 net_invalid_timestamp()))
3301 rtt_us = ktime_us_delta(ktime_get_real(),
3303 else if (ca_seq_rtt > 0)
3304 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3307 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3311 #if FASTRETRANS_DEBUG > 0
3312 WARN_ON((int)tp->sacked_out < 0);
3313 WARN_ON((int)tp->lost_out < 0);
3314 WARN_ON((int)tp->retrans_out < 0);
3315 if (!tp->packets_out && tcp_is_sack(tp)) {
3316 icsk = inet_csk(sk);
3318 printk(KERN_DEBUG "Leak l=%u %d\n",
3319 tp->lost_out, icsk->icsk_ca_state);
3322 if (tp->sacked_out) {
3323 printk(KERN_DEBUG "Leak s=%u %d\n",
3324 tp->sacked_out, icsk->icsk_ca_state);
3327 if (tp->retrans_out) {
3328 printk(KERN_DEBUG "Leak r=%u %d\n",
3329 tp->retrans_out, icsk->icsk_ca_state);
3330 tp->retrans_out = 0;
3337 static void tcp_ack_probe(struct sock *sk)
3339 const struct tcp_sock *tp = tcp_sk(sk);
3340 struct inet_connection_sock *icsk = inet_csk(sk);
3342 /* Was it a usable window open? */
3344 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3345 icsk->icsk_backoff = 0;
3346 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3347 /* Socket must be waked up by subsequent tcp_data_snd_check().
3348 * This function is not for random using!
3351 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3352 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3357 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3359 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3360 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
3363 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3365 const struct tcp_sock *tp = tcp_sk(sk);
3366 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3367 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3370 /* Check that window update is acceptable.
3371 * The function assumes that snd_una<=ack<=snd_next.
3373 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3374 const u32 ack, const u32 ack_seq,
3377 return (after(ack, tp->snd_una) ||
3378 after(ack_seq, tp->snd_wl1) ||
3379 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3382 /* Update our send window.
3384 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3385 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3387 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3390 struct tcp_sock *tp = tcp_sk(sk);
3392 u32 nwin = ntohs(tcp_hdr(skb)->window);
3394 if (likely(!tcp_hdr(skb)->syn))
3395 nwin <<= tp->rx_opt.snd_wscale;
3397 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3398 flag |= FLAG_WIN_UPDATE;
3399 tcp_update_wl(tp, ack, ack_seq);
3401 if (tp->snd_wnd != nwin) {
3404 /* Note, it is the only place, where
3405 * fast path is recovered for sending TCP.
3408 tcp_fast_path_check(sk);
3410 if (nwin > tp->max_window) {
3411 tp->max_window = nwin;
3412 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3422 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3423 * continue in congestion avoidance.
3425 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3427 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3428 tp->snd_cwnd_cnt = 0;
3429 tp->bytes_acked = 0;
3430 TCP_ECN_queue_cwr(tp);
3431 tcp_moderate_cwnd(tp);
3434 /* A conservative spurious RTO response algorithm: reduce cwnd using
3435 * rate halving and continue in congestion avoidance.
3437 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3439 tcp_enter_cwr(sk, 0);
3442 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3444 if (flag & FLAG_ECE)
3445 tcp_ratehalving_spur_to_response(sk);
3447 tcp_undo_cwr(sk, 1);
3450 /* F-RTO spurious RTO detection algorithm (RFC4138)
3452 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3453 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3454 * window (but not to or beyond highest sequence sent before RTO):
3455 * On First ACK, send two new segments out.
3456 * On Second ACK, RTO was likely spurious. Do spurious response (response
3457 * algorithm is not part of the F-RTO detection algorithm
3458 * given in RFC4138 but can be selected separately).
3459 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3460 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3461 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3462 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3464 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3465 * original window even after we transmit two new data segments.
3468 * on first step, wait until first cumulative ACK arrives, then move to
3469 * the second step. In second step, the next ACK decides.
3471 * F-RTO is implemented (mainly) in four functions:
3472 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3473 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3474 * called when tcp_use_frto() showed green light
3475 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3476 * - tcp_enter_frto_loss() is called if there is not enough evidence
3477 * to prove that the RTO is indeed spurious. It transfers the control
3478 * from F-RTO to the conventional RTO recovery
3480 static int tcp_process_frto(struct sock *sk, int flag)
3482 struct tcp_sock *tp = tcp_sk(sk);
3484 tcp_verify_left_out(tp);
3486 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3487 if (flag & FLAG_DATA_ACKED)
3488 inet_csk(sk)->icsk_retransmits = 0;
3490 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3491 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3492 tp->undo_marker = 0;
3494 if (!before(tp->snd_una, tp->frto_highmark)) {
3495 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3499 if (!tcp_is_sackfrto(tp)) {
3500 /* RFC4138 shortcoming in step 2; should also have case c):
3501 * ACK isn't duplicate nor advances window, e.g., opposite dir
3504 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3507 if (!(flag & FLAG_DATA_ACKED)) {
3508 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3513 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3514 /* Prevent sending of new data. */
3515 tp->snd_cwnd = min(tp->snd_cwnd,
3516 tcp_packets_in_flight(tp));
3520 if ((tp->frto_counter >= 2) &&
3521 (!(flag & FLAG_FORWARD_PROGRESS) ||
3522 ((flag & FLAG_DATA_SACKED) &&
3523 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3524 /* RFC4138 shortcoming (see comment above) */
3525 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3526 (flag & FLAG_NOT_DUP))
3529 tcp_enter_frto_loss(sk, 3, flag);
3534 if (tp->frto_counter == 1) {
3535 /* tcp_may_send_now needs to see updated state */
3536 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3537 tp->frto_counter = 2;
3539 if (!tcp_may_send_now(sk))
3540 tcp_enter_frto_loss(sk, 2, flag);
3544 switch (sysctl_tcp_frto_response) {
3546 tcp_undo_spur_to_response(sk, flag);
3549 tcp_conservative_spur_to_response(tp);
3552 tcp_ratehalving_spur_to_response(sk);
3555 tp->frto_counter = 0;
3556 tp->undo_marker = 0;
3557 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3562 /* This routine deals with incoming acks, but not outgoing ones. */
3563 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3565 struct inet_connection_sock *icsk = inet_csk(sk);
3566 struct tcp_sock *tp = tcp_sk(sk);
3567 u32 prior_snd_una = tp->snd_una;
3568 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3569 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3570 u32 prior_in_flight;
3575 /* If the ack is newer than sent or older than previous acks
3576 * then we can probably ignore it.
3578 if (after(ack, tp->snd_nxt))
3579 goto uninteresting_ack;
3581 if (before(ack, prior_snd_una))
3584 if (after(ack, prior_snd_una))
3585 flag |= FLAG_SND_UNA_ADVANCED;
3587 if (sysctl_tcp_abc) {
3588 if (icsk->icsk_ca_state < TCP_CA_CWR)
3589 tp->bytes_acked += ack - prior_snd_una;
3590 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3591 /* we assume just one segment left network */
3592 tp->bytes_acked += min(ack - prior_snd_una,
3596 prior_fackets = tp->fackets_out;
3597 prior_in_flight = tcp_packets_in_flight(tp);
3599 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3600 /* Window is constant, pure forward advance.
3601 * No more checks are required.
3602 * Note, we use the fact that SND.UNA>=SND.WL2.
3604 tcp_update_wl(tp, ack, ack_seq);
3606 flag |= FLAG_WIN_UPDATE;
3608 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3610 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3612 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3615 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3617 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3619 if (TCP_SKB_CB(skb)->sacked)
3620 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3622 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3625 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3628 /* We passed data and got it acked, remove any soft error
3629 * log. Something worked...
3631 sk->sk_err_soft = 0;
3632 icsk->icsk_probes_out = 0;
3633 tp->rcv_tstamp = tcp_time_stamp;
3634 prior_packets = tp->packets_out;
3638 /* See if we can take anything off of the retransmit queue. */
3639 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3641 if (tp->frto_counter)
3642 frto_cwnd = tcp_process_frto(sk, flag);
3643 /* Guarantee sacktag reordering detection against wrap-arounds */
3644 if (before(tp->frto_highmark, tp->snd_una))
3645 tp->frto_highmark = 0;
3647 if (tcp_ack_is_dubious(sk, flag)) {
3648 /* Advance CWND, if state allows this. */
3649 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3650 tcp_may_raise_cwnd(sk, flag))
3651 tcp_cong_avoid(sk, ack, prior_in_flight);
3652 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3655 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3656 tcp_cong_avoid(sk, ack, prior_in_flight);
3659 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3660 dst_confirm(sk->sk_dst_cache);
3665 /* If this ack opens up a zero window, clear backoff. It was
3666 * being used to time the probes, and is probably far higher than
3667 * it needs to be for normal retransmission.
3669 if (tcp_send_head(sk))
3674 if (TCP_SKB_CB(skb)->sacked) {
3675 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3676 if (icsk->icsk_ca_state == TCP_CA_Open)
3677 tcp_try_keep_open(sk);
3681 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3685 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3686 * But, this can also be called on packets in the established flow when
3687 * the fast version below fails.
3689 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3693 struct tcphdr *th = tcp_hdr(skb);
3694 int length = (th->doff * 4) - sizeof(struct tcphdr);
3696 ptr = (unsigned char *)(th + 1);
3697 opt_rx->saw_tstamp = 0;
3699 while (length > 0) {
3700 int opcode = *ptr++;
3706 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3711 if (opsize < 2) /* "silly options" */
3713 if (opsize > length)
3714 return; /* don't parse partial options */
3717 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3718 u16 in_mss = get_unaligned_be16(ptr);
3720 if (opt_rx->user_mss &&
3721 opt_rx->user_mss < in_mss)
3722 in_mss = opt_rx->user_mss;
3723 opt_rx->mss_clamp = in_mss;
3728 if (opsize == TCPOLEN_WINDOW && th->syn &&
3729 !estab && sysctl_tcp_window_scaling) {
3730 __u8 snd_wscale = *(__u8 *)ptr;
3731 opt_rx->wscale_ok = 1;
3732 if (snd_wscale > 14) {
3733 if (net_ratelimit())
3734 printk(KERN_INFO "tcp_parse_options: Illegal window "
3735 "scaling value %d >14 received.\n",
3739 opt_rx->snd_wscale = snd_wscale;
3742 case TCPOPT_TIMESTAMP:
3743 if ((opsize == TCPOLEN_TIMESTAMP) &&
3744 ((estab && opt_rx->tstamp_ok) ||
3745 (!estab && sysctl_tcp_timestamps))) {
3746 opt_rx->saw_tstamp = 1;
3747 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3748 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3751 case TCPOPT_SACK_PERM:
3752 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3753 !estab && sysctl_tcp_sack) {
3754 opt_rx->sack_ok = 1;
3755 tcp_sack_reset(opt_rx);
3760 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3761 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3763 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3766 #ifdef CONFIG_TCP_MD5SIG
3769 * The MD5 Hash has already been
3770 * checked (see tcp_v{4,6}_do_rcv()).
3782 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3784 __be32 *ptr = (__be32 *)(th + 1);
3786 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3787 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3788 tp->rx_opt.saw_tstamp = 1;
3790 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3792 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3798 /* Fast parse options. This hopes to only see timestamps.
3799 * If it is wrong it falls back on tcp_parse_options().
3801 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3802 struct tcp_sock *tp)
3804 if (th->doff == sizeof(struct tcphdr) >> 2) {
3805 tp->rx_opt.saw_tstamp = 0;
3807 } else if (tp->rx_opt.tstamp_ok &&
3808 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3809 if (tcp_parse_aligned_timestamp(tp, th))
3812 tcp_parse_options(skb, &tp->rx_opt, 1);
3816 #ifdef CONFIG_TCP_MD5SIG
3818 * Parse MD5 Signature option
3820 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3822 int length = (th->doff << 2) - sizeof (*th);
3823 u8 *ptr = (u8*)(th + 1);
3825 /* If the TCP option is too short, we can short cut */
3826 if (length < TCPOLEN_MD5SIG)
3829 while (length > 0) {
3830 int opcode = *ptr++;
3841 if (opsize < 2 || opsize > length)
3843 if (opcode == TCPOPT_MD5SIG)
3853 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3855 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3856 tp->rx_opt.ts_recent_stamp = get_seconds();
3859 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3861 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3862 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3863 * extra check below makes sure this can only happen
3864 * for pure ACK frames. -DaveM
3866 * Not only, also it occurs for expired timestamps.
3869 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3870 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3871 tcp_store_ts_recent(tp);
3875 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3877 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3878 * it can pass through stack. So, the following predicate verifies that
3879 * this segment is not used for anything but congestion avoidance or
3880 * fast retransmit. Moreover, we even are able to eliminate most of such
3881 * second order effects, if we apply some small "replay" window (~RTO)
3882 * to timestamp space.
3884 * All these measures still do not guarantee that we reject wrapped ACKs
3885 * on networks with high bandwidth, when sequence space is recycled fastly,
3886 * but it guarantees that such events will be very rare and do not affect
3887 * connection seriously. This doesn't look nice, but alas, PAWS is really
3890 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3891 * states that events when retransmit arrives after original data are rare.
3892 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3893 * the biggest problem on large power networks even with minor reordering.
3894 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3895 * up to bandwidth of 18Gigabit/sec. 8) ]
3898 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3900 struct tcp_sock *tp = tcp_sk(sk);
3901 struct tcphdr *th = tcp_hdr(skb);
3902 u32 seq = TCP_SKB_CB(skb)->seq;
3903 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3905 return (/* 1. Pure ACK with correct sequence number. */
3906 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3908 /* 2. ... and duplicate ACK. */
3909 ack == tp->snd_una &&
3911 /* 3. ... and does not update window. */
3912 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3914 /* 4. ... and sits in replay window. */
3915 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3918 static inline int tcp_paws_discard(const struct sock *sk,
3919 const struct sk_buff *skb)
3921 const struct tcp_sock *tp = tcp_sk(sk);
3922 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3923 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3924 !tcp_disordered_ack(sk, skb));
3927 /* Check segment sequence number for validity.
3929 * Segment controls are considered valid, if the segment
3930 * fits to the window after truncation to the window. Acceptability
3931 * of data (and SYN, FIN, of course) is checked separately.
3932 * See tcp_data_queue(), for example.
3934 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3935 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3936 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3937 * (borrowed from freebsd)
3940 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3942 return !before(end_seq, tp->rcv_wup) &&
3943 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3946 /* When we get a reset we do this. */
3947 static void tcp_reset(struct sock *sk)
3949 /* We want the right error as BSD sees it (and indeed as we do). */
3950 switch (sk->sk_state) {
3952 sk->sk_err = ECONNREFUSED;
3954 case TCP_CLOSE_WAIT:
3960 sk->sk_err = ECONNRESET;
3963 if (!sock_flag(sk, SOCK_DEAD))
3964 sk->sk_error_report(sk);
3970 * Process the FIN bit. This now behaves as it is supposed to work
3971 * and the FIN takes effect when it is validly part of sequence
3972 * space. Not before when we get holes.
3974 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3975 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3978 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3979 * close and we go into CLOSING (and later onto TIME-WAIT)
3981 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3983 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3985 struct tcp_sock *tp = tcp_sk(sk);
3987 inet_csk_schedule_ack(sk);
3989 sk->sk_shutdown |= RCV_SHUTDOWN;
3990 sock_set_flag(sk, SOCK_DONE);
3992 switch (sk->sk_state) {
3994 case TCP_ESTABLISHED:
3995 /* Move to CLOSE_WAIT */
3996 tcp_set_state(sk, TCP_CLOSE_WAIT);
3997 inet_csk(sk)->icsk_ack.pingpong = 1;
4000 case TCP_CLOSE_WAIT:
4002 /* Received a retransmission of the FIN, do
4007 /* RFC793: Remain in the LAST-ACK state. */
4011 /* This case occurs when a simultaneous close
4012 * happens, we must ack the received FIN and
4013 * enter the CLOSING state.
4016 tcp_set_state(sk, TCP_CLOSING);
4019 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4021 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4024 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4025 * cases we should never reach this piece of code.
4027 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4028 __func__, sk->sk_state);
4032 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4033 * Probably, we should reset in this case. For now drop them.
4035 __skb_queue_purge(&tp->out_of_order_queue);
4036 if (tcp_is_sack(tp))
4037 tcp_sack_reset(&tp->rx_opt);
4040 if (!sock_flag(sk, SOCK_DEAD)) {
4041 sk->sk_state_change(sk);
4043 /* Do not send POLL_HUP for half duplex close. */
4044 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4045 sk->sk_state == TCP_CLOSE)
4046 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4048 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4052 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4055 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4056 if (before(seq, sp->start_seq))
4057 sp->start_seq = seq;
4058 if (after(end_seq, sp->end_seq))
4059 sp->end_seq = end_seq;
4065 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4067 struct tcp_sock *tp = tcp_sk(sk);
4069 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4072 if (before(seq, tp->rcv_nxt))
4073 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4075 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4077 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4079 tp->rx_opt.dsack = 1;
4080 tp->duplicate_sack[0].start_seq = seq;
4081 tp->duplicate_sack[0].end_seq = end_seq;
4082 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 1;
4086 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4088 struct tcp_sock *tp = tcp_sk(sk);
4090 if (!tp->rx_opt.dsack)
4091 tcp_dsack_set(sk, seq, end_seq);
4093 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4096 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4098 struct tcp_sock *tp = tcp_sk(sk);
4100 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4101 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4102 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4103 tcp_enter_quickack_mode(sk);
4105 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4106 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4108 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4109 end_seq = tp->rcv_nxt;
4110 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4117 /* These routines update the SACK block as out-of-order packets arrive or
4118 * in-order packets close up the sequence space.
4120 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4123 struct tcp_sack_block *sp = &tp->selective_acks[0];
4124 struct tcp_sack_block *swalk = sp + 1;
4126 /* See if the recent change to the first SACK eats into
4127 * or hits the sequence space of other SACK blocks, if so coalesce.
4129 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4130 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4133 /* Zap SWALK, by moving every further SACK up by one slot.
4134 * Decrease num_sacks.
4136 tp->rx_opt.num_sacks--;
4137 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
4139 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4143 this_sack++, swalk++;
4147 static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
4148 struct tcp_sack_block *sack2)
4152 tmp = sack1->start_seq;
4153 sack1->start_seq = sack2->start_seq;
4154 sack2->start_seq = tmp;
4156 tmp = sack1->end_seq;
4157 sack1->end_seq = sack2->end_seq;
4158 sack2->end_seq = tmp;
4161 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4163 struct tcp_sock *tp = tcp_sk(sk);
4164 struct tcp_sack_block *sp = &tp->selective_acks[0];
4165 int cur_sacks = tp->rx_opt.num_sacks;
4171 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4172 if (tcp_sack_extend(sp, seq, end_seq)) {
4173 /* Rotate this_sack to the first one. */
4174 for (; this_sack > 0; this_sack--, sp--)
4175 tcp_sack_swap(sp, sp - 1);
4177 tcp_sack_maybe_coalesce(tp);
4182 /* Could not find an adjacent existing SACK, build a new one,
4183 * put it at the front, and shift everyone else down. We
4184 * always know there is at least one SACK present already here.
4186 * If the sack array is full, forget about the last one.
4188 if (this_sack >= TCP_NUM_SACKS) {
4190 tp->rx_opt.num_sacks--;
4193 for (; this_sack > 0; this_sack--, sp--)
4197 /* Build the new head SACK, and we're done. */
4198 sp->start_seq = seq;
4199 sp->end_seq = end_seq;
4200 tp->rx_opt.num_sacks++;
4201 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
4204 /* RCV.NXT advances, some SACKs should be eaten. */
4206 static void tcp_sack_remove(struct tcp_sock *tp)
4208 struct tcp_sack_block *sp = &tp->selective_acks[0];
4209 int num_sacks = tp->rx_opt.num_sacks;
4212 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4213 if (skb_queue_empty(&tp->out_of_order_queue)) {
4214 tp->rx_opt.num_sacks = 0;
4215 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
4219 for (this_sack = 0; this_sack < num_sacks;) {
4220 /* Check if the start of the sack is covered by RCV.NXT. */
4221 if (!before(tp->rcv_nxt, sp->start_seq)) {
4224 /* RCV.NXT must cover all the block! */
4225 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4227 /* Zap this SACK, by moving forward any other SACKS. */
4228 for (i=this_sack+1; i < num_sacks; i++)
4229 tp->selective_acks[i-1] = tp->selective_acks[i];
4236 if (num_sacks != tp->rx_opt.num_sacks) {
4237 tp->rx_opt.num_sacks = num_sacks;
4238 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks +
4243 /* This one checks to see if we can put data from the
4244 * out_of_order queue into the receive_queue.
4246 static void tcp_ofo_queue(struct sock *sk)
4248 struct tcp_sock *tp = tcp_sk(sk);
4249 __u32 dsack_high = tp->rcv_nxt;
4250 struct sk_buff *skb;
4252 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4253 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4256 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4257 __u32 dsack = dsack_high;
4258 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4259 dsack_high = TCP_SKB_CB(skb)->end_seq;
4260 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4263 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4264 SOCK_DEBUG(sk, "ofo packet was already received \n");
4265 __skb_unlink(skb, &tp->out_of_order_queue);
4269 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4270 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4271 TCP_SKB_CB(skb)->end_seq);
4273 __skb_unlink(skb, &tp->out_of_order_queue);
4274 __skb_queue_tail(&sk->sk_receive_queue, skb);
4275 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4276 if (tcp_hdr(skb)->fin)
4277 tcp_fin(skb, sk, tcp_hdr(skb));
4281 static int tcp_prune_ofo_queue(struct sock *sk);
4282 static int tcp_prune_queue(struct sock *sk);
4284 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4286 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4287 !sk_rmem_schedule(sk, size)) {
4289 if (tcp_prune_queue(sk) < 0)
4292 if (!sk_rmem_schedule(sk, size)) {
4293 if (!tcp_prune_ofo_queue(sk))
4296 if (!sk_rmem_schedule(sk, size))
4303 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4305 struct tcphdr *th = tcp_hdr(skb);
4306 struct tcp_sock *tp = tcp_sk(sk);
4309 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4312 __skb_pull(skb, th->doff * 4);
4314 TCP_ECN_accept_cwr(tp, skb);
4316 if (tp->rx_opt.dsack) {
4317 tp->rx_opt.dsack = 0;
4318 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
4321 /* Queue data for delivery to the user.
4322 * Packets in sequence go to the receive queue.
4323 * Out of sequence packets to the out_of_order_queue.
4325 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4326 if (tcp_receive_window(tp) == 0)
4329 /* Ok. In sequence. In window. */
4330 if (tp->ucopy.task == current &&
4331 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4332 sock_owned_by_user(sk) && !tp->urg_data) {
4333 int chunk = min_t(unsigned int, skb->len,
4336 __set_current_state(TASK_RUNNING);
4339 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4340 tp->ucopy.len -= chunk;
4341 tp->copied_seq += chunk;
4342 eaten = (chunk == skb->len && !th->fin);
4343 tcp_rcv_space_adjust(sk);
4351 tcp_try_rmem_schedule(sk, skb->truesize))
4354 skb_set_owner_r(skb, sk);
4355 __skb_queue_tail(&sk->sk_receive_queue, skb);
4357 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4359 tcp_event_data_recv(sk, skb);
4361 tcp_fin(skb, sk, th);
4363 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4366 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4367 * gap in queue is filled.
4369 if (skb_queue_empty(&tp->out_of_order_queue))
4370 inet_csk(sk)->icsk_ack.pingpong = 0;
4373 if (tp->rx_opt.num_sacks)
4374 tcp_sack_remove(tp);
4376 tcp_fast_path_check(sk);
4380 else if (!sock_flag(sk, SOCK_DEAD))
4381 sk->sk_data_ready(sk, 0);
4385 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4386 /* A retransmit, 2nd most common case. Force an immediate ack. */
4387 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4388 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4391 tcp_enter_quickack_mode(sk);
4392 inet_csk_schedule_ack(sk);
4398 /* Out of window. F.e. zero window probe. */
4399 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4402 tcp_enter_quickack_mode(sk);
4404 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4405 /* Partial packet, seq < rcv_next < end_seq */
4406 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4407 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4408 TCP_SKB_CB(skb)->end_seq);
4410 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4412 /* If window is closed, drop tail of packet. But after
4413 * remembering D-SACK for its head made in previous line.
4415 if (!tcp_receive_window(tp))
4420 TCP_ECN_check_ce(tp, skb);
4422 if (tcp_try_rmem_schedule(sk, skb->truesize))
4425 /* Disable header prediction. */
4427 inet_csk_schedule_ack(sk);
4429 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4430 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4432 skb_set_owner_r(skb, sk);
4434 if (!skb_peek(&tp->out_of_order_queue)) {
4435 /* Initial out of order segment, build 1 SACK. */
4436 if (tcp_is_sack(tp)) {
4437 tp->rx_opt.num_sacks = 1;
4438 tp->rx_opt.dsack = 0;
4439 tp->rx_opt.eff_sacks = 1;
4440 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4441 tp->selective_acks[0].end_seq =
4442 TCP_SKB_CB(skb)->end_seq;
4444 __skb_queue_head(&tp->out_of_order_queue, skb);
4446 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4447 u32 seq = TCP_SKB_CB(skb)->seq;
4448 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4450 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4451 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4453 if (!tp->rx_opt.num_sacks ||
4454 tp->selective_acks[0].end_seq != seq)
4457 /* Common case: data arrive in order after hole. */
4458 tp->selective_acks[0].end_seq = end_seq;
4462 /* Find place to insert this segment. */
4464 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4466 } while ((skb1 = skb1->prev) !=
4467 (struct sk_buff *)&tp->out_of_order_queue);
4469 /* Do skb overlap to previous one? */
4470 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
4471 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4472 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4473 /* All the bits are present. Drop. */
4475 tcp_dsack_set(sk, seq, end_seq);
4478 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4479 /* Partial overlap. */
4480 tcp_dsack_set(sk, seq,
4481 TCP_SKB_CB(skb1)->end_seq);
4486 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4488 /* And clean segments covered by new one as whole. */
4489 while ((skb1 = skb->next) !=
4490 (struct sk_buff *)&tp->out_of_order_queue &&
4491 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
4492 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4493 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4497 __skb_unlink(skb1, &tp->out_of_order_queue);
4498 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4499 TCP_SKB_CB(skb1)->end_seq);
4504 if (tcp_is_sack(tp))
4505 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4509 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4510 struct sk_buff_head *list)
4512 struct sk_buff *next = skb->next;
4514 __skb_unlink(skb, list);
4516 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4521 /* Collapse contiguous sequence of skbs head..tail with
4522 * sequence numbers start..end.
4523 * Segments with FIN/SYN are not collapsed (only because this
4527 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4528 struct sk_buff *head, struct sk_buff *tail,
4531 struct sk_buff *skb;
4533 /* First, check that queue is collapsible and find
4534 * the point where collapsing can be useful. */
4535 for (skb = head; skb != tail;) {
4536 /* No new bits? It is possible on ofo queue. */
4537 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4538 skb = tcp_collapse_one(sk, skb, list);
4542 /* The first skb to collapse is:
4544 * - bloated or contains data before "start" or
4545 * overlaps to the next one.
4547 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4548 (tcp_win_from_space(skb->truesize) > skb->len ||
4549 before(TCP_SKB_CB(skb)->seq, start) ||
4550 (skb->next != tail &&
4551 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4554 /* Decided to skip this, advance start seq. */
4555 start = TCP_SKB_CB(skb)->end_seq;
4558 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4561 while (before(start, end)) {
4562 struct sk_buff *nskb;
4563 unsigned int header = skb_headroom(skb);
4564 int copy = SKB_MAX_ORDER(header, 0);
4566 /* Too big header? This can happen with IPv6. */
4569 if (end - start < copy)
4571 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4575 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4576 skb_set_network_header(nskb, (skb_network_header(skb) -
4578 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4580 skb_reserve(nskb, header);
4581 memcpy(nskb->head, skb->head, header);
4582 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4583 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4584 __skb_queue_before(list, skb, nskb);
4585 skb_set_owner_r(nskb, sk);
4587 /* Copy data, releasing collapsed skbs. */
4589 int offset = start - TCP_SKB_CB(skb)->seq;
4590 int size = TCP_SKB_CB(skb)->end_seq - start;
4594 size = min(copy, size);
4595 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4597 TCP_SKB_CB(nskb)->end_seq += size;
4601 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4602 skb = tcp_collapse_one(sk, skb, list);
4604 tcp_hdr(skb)->syn ||
4612 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4613 * and tcp_collapse() them until all the queue is collapsed.
4615 static void tcp_collapse_ofo_queue(struct sock *sk)
4617 struct tcp_sock *tp = tcp_sk(sk);
4618 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4619 struct sk_buff *head;
4625 start = TCP_SKB_CB(skb)->seq;
4626 end = TCP_SKB_CB(skb)->end_seq;
4632 /* Segment is terminated when we see gap or when
4633 * we are at the end of all the queue. */
4634 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4635 after(TCP_SKB_CB(skb)->seq, end) ||
4636 before(TCP_SKB_CB(skb)->end_seq, start)) {
4637 tcp_collapse(sk, &tp->out_of_order_queue,
4638 head, skb, start, end);
4640 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4642 /* Start new segment */
4643 start = TCP_SKB_CB(skb)->seq;
4644 end = TCP_SKB_CB(skb)->end_seq;
4646 if (before(TCP_SKB_CB(skb)->seq, start))
4647 start = TCP_SKB_CB(skb)->seq;
4648 if (after(TCP_SKB_CB(skb)->end_seq, end))
4649 end = TCP_SKB_CB(skb)->end_seq;
4655 * Purge the out-of-order queue.
4656 * Return true if queue was pruned.
4658 static int tcp_prune_ofo_queue(struct sock *sk)
4660 struct tcp_sock *tp = tcp_sk(sk);
4663 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4664 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4665 __skb_queue_purge(&tp->out_of_order_queue);
4667 /* Reset SACK state. A conforming SACK implementation will
4668 * do the same at a timeout based retransmit. When a connection
4669 * is in a sad state like this, we care only about integrity
4670 * of the connection not performance.
4672 if (tp->rx_opt.sack_ok)
4673 tcp_sack_reset(&tp->rx_opt);
4680 /* Reduce allocated memory if we can, trying to get
4681 * the socket within its memory limits again.
4683 * Return less than zero if we should start dropping frames
4684 * until the socket owning process reads some of the data
4685 * to stabilize the situation.
4687 static int tcp_prune_queue(struct sock *sk)
4689 struct tcp_sock *tp = tcp_sk(sk);
4691 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4693 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4695 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4696 tcp_clamp_window(sk);
4697 else if (tcp_memory_pressure)
4698 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4700 tcp_collapse_ofo_queue(sk);
4701 tcp_collapse(sk, &sk->sk_receive_queue,
4702 sk->sk_receive_queue.next,
4703 (struct sk_buff *)&sk->sk_receive_queue,
4704 tp->copied_seq, tp->rcv_nxt);
4707 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4710 /* Collapsing did not help, destructive actions follow.
4711 * This must not ever occur. */
4713 tcp_prune_ofo_queue(sk);
4715 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4718 /* If we are really being abused, tell the caller to silently
4719 * drop receive data on the floor. It will get retransmitted
4720 * and hopefully then we'll have sufficient space.
4722 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4724 /* Massive buffer overcommit. */
4729 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4730 * As additional protections, we do not touch cwnd in retransmission phases,
4731 * and if application hit its sndbuf limit recently.
4733 void tcp_cwnd_application_limited(struct sock *sk)
4735 struct tcp_sock *tp = tcp_sk(sk);
4737 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4738 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4739 /* Limited by application or receiver window. */
4740 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4741 u32 win_used = max(tp->snd_cwnd_used, init_win);
4742 if (win_used < tp->snd_cwnd) {
4743 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4744 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4746 tp->snd_cwnd_used = 0;
4748 tp->snd_cwnd_stamp = tcp_time_stamp;
4751 static int tcp_should_expand_sndbuf(struct sock *sk)
4753 struct tcp_sock *tp = tcp_sk(sk);
4755 /* If the user specified a specific send buffer setting, do
4758 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4761 /* If we are under global TCP memory pressure, do not expand. */
4762 if (tcp_memory_pressure)
4765 /* If we are under soft global TCP memory pressure, do not expand. */
4766 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4769 /* If we filled the congestion window, do not expand. */
4770 if (tp->packets_out >= tp->snd_cwnd)
4776 /* When incoming ACK allowed to free some skb from write_queue,
4777 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4778 * on the exit from tcp input handler.
4780 * PROBLEM: sndbuf expansion does not work well with largesend.
4782 static void tcp_new_space(struct sock *sk)
4784 struct tcp_sock *tp = tcp_sk(sk);
4786 if (tcp_should_expand_sndbuf(sk)) {
4787 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4788 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4789 int demanded = max_t(unsigned int, tp->snd_cwnd,
4790 tp->reordering + 1);
4791 sndmem *= 2 * demanded;
4792 if (sndmem > sk->sk_sndbuf)
4793 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4794 tp->snd_cwnd_stamp = tcp_time_stamp;
4797 sk->sk_write_space(sk);
4800 static void tcp_check_space(struct sock *sk)
4802 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4803 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4804 if (sk->sk_socket &&
4805 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4810 static inline void tcp_data_snd_check(struct sock *sk)
4812 tcp_push_pending_frames(sk);
4813 tcp_check_space(sk);
4817 * Check if sending an ack is needed.
4819 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4821 struct tcp_sock *tp = tcp_sk(sk);
4823 /* More than one full frame received... */
4824 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
4825 /* ... and right edge of window advances far enough.
4826 * (tcp_recvmsg() will send ACK otherwise). Or...
4828 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4829 /* We ACK each frame or... */
4830 tcp_in_quickack_mode(sk) ||
4831 /* We have out of order data. */
4832 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4833 /* Then ack it now */
4836 /* Else, send delayed ack. */
4837 tcp_send_delayed_ack(sk);
4841 static inline void tcp_ack_snd_check(struct sock *sk)
4843 if (!inet_csk_ack_scheduled(sk)) {
4844 /* We sent a data segment already. */
4847 __tcp_ack_snd_check(sk, 1);
4851 * This routine is only called when we have urgent data
4852 * signaled. Its the 'slow' part of tcp_urg. It could be
4853 * moved inline now as tcp_urg is only called from one
4854 * place. We handle URGent data wrong. We have to - as
4855 * BSD still doesn't use the correction from RFC961.
4856 * For 1003.1g we should support a new option TCP_STDURG to permit
4857 * either form (or just set the sysctl tcp_stdurg).
4860 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
4862 struct tcp_sock *tp = tcp_sk(sk);
4863 u32 ptr = ntohs(th->urg_ptr);
4865 if (ptr && !sysctl_tcp_stdurg)
4867 ptr += ntohl(th->seq);
4869 /* Ignore urgent data that we've already seen and read. */
4870 if (after(tp->copied_seq, ptr))
4873 /* Do not replay urg ptr.
4875 * NOTE: interesting situation not covered by specs.
4876 * Misbehaving sender may send urg ptr, pointing to segment,
4877 * which we already have in ofo queue. We are not able to fetch
4878 * such data and will stay in TCP_URG_NOTYET until will be eaten
4879 * by recvmsg(). Seems, we are not obliged to handle such wicked
4880 * situations. But it is worth to think about possibility of some
4881 * DoSes using some hypothetical application level deadlock.
4883 if (before(ptr, tp->rcv_nxt))
4886 /* Do we already have a newer (or duplicate) urgent pointer? */
4887 if (tp->urg_data && !after(ptr, tp->urg_seq))
4890 /* Tell the world about our new urgent pointer. */
4893 /* We may be adding urgent data when the last byte read was
4894 * urgent. To do this requires some care. We cannot just ignore
4895 * tp->copied_seq since we would read the last urgent byte again
4896 * as data, nor can we alter copied_seq until this data arrives
4897 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4899 * NOTE. Double Dutch. Rendering to plain English: author of comment
4900 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4901 * and expect that both A and B disappear from stream. This is _wrong_.
4902 * Though this happens in BSD with high probability, this is occasional.
4903 * Any application relying on this is buggy. Note also, that fix "works"
4904 * only in this artificial test. Insert some normal data between A and B and we will
4905 * decline of BSD again. Verdict: it is better to remove to trap
4908 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4909 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4910 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4912 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4913 __skb_unlink(skb, &sk->sk_receive_queue);
4918 tp->urg_data = TCP_URG_NOTYET;
4921 /* Disable header prediction. */
4925 /* This is the 'fast' part of urgent handling. */
4926 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4928 struct tcp_sock *tp = tcp_sk(sk);
4930 /* Check if we get a new urgent pointer - normally not. */
4932 tcp_check_urg(sk, th);
4934 /* Do we wait for any urgent data? - normally not... */
4935 if (tp->urg_data == TCP_URG_NOTYET) {
4936 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4939 /* Is the urgent pointer pointing into this packet? */
4940 if (ptr < skb->len) {
4942 if (skb_copy_bits(skb, ptr, &tmp, 1))
4944 tp->urg_data = TCP_URG_VALID | tmp;
4945 if (!sock_flag(sk, SOCK_DEAD))
4946 sk->sk_data_ready(sk, 0);
4951 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4953 struct tcp_sock *tp = tcp_sk(sk);
4954 int chunk = skb->len - hlen;
4958 if (skb_csum_unnecessary(skb))
4959 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4961 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4965 tp->ucopy.len -= chunk;
4966 tp->copied_seq += chunk;
4967 tcp_rcv_space_adjust(sk);
4974 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4975 struct sk_buff *skb)
4979 if (sock_owned_by_user(sk)) {
4981 result = __tcp_checksum_complete(skb);
4984 result = __tcp_checksum_complete(skb);
4989 static inline int tcp_checksum_complete_user(struct sock *sk,
4990 struct sk_buff *skb)
4992 return !skb_csum_unnecessary(skb) &&
4993 __tcp_checksum_complete_user(sk, skb);
4996 #ifdef CONFIG_NET_DMA
4997 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5000 struct tcp_sock *tp = tcp_sk(sk);
5001 int chunk = skb->len - hlen;
5003 int copied_early = 0;
5005 if (tp->ucopy.wakeup)
5008 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5009 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5011 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5013 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5015 tp->ucopy.iov, chunk,
5016 tp->ucopy.pinned_list);
5021 tp->ucopy.dma_cookie = dma_cookie;
5024 tp->ucopy.len -= chunk;
5025 tp->copied_seq += chunk;
5026 tcp_rcv_space_adjust(sk);
5028 if ((tp->ucopy.len == 0) ||
5029 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5030 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5031 tp->ucopy.wakeup = 1;
5032 sk->sk_data_ready(sk, 0);
5034 } else if (chunk > 0) {
5035 tp->ucopy.wakeup = 1;
5036 sk->sk_data_ready(sk, 0);
5039 return copied_early;
5041 #endif /* CONFIG_NET_DMA */
5043 /* Does PAWS and seqno based validation of an incoming segment, flags will
5044 * play significant role here.
5046 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5047 struct tcphdr *th, int syn_inerr)
5049 struct tcp_sock *tp = tcp_sk(sk);
5051 /* RFC1323: H1. Apply PAWS check first. */
5052 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5053 tcp_paws_discard(sk, skb)) {
5055 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5056 tcp_send_dupack(sk, skb);
5059 /* Reset is accepted even if it did not pass PAWS. */
5062 /* Step 1: check sequence number */
5063 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5064 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5065 * (RST) segments are validated by checking their SEQ-fields."
5066 * And page 69: "If an incoming segment is not acceptable,
5067 * an acknowledgment should be sent in reply (unless the RST
5068 * bit is set, if so drop the segment and return)".
5071 tcp_send_dupack(sk, skb);
5075 /* Step 2: check RST bit */
5081 /* ts_recent update must be made after we are sure that the packet
5084 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5086 /* step 3: check security and precedence [ignored] */
5088 /* step 4: Check for a SYN in window. */
5089 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5091 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5092 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5105 * TCP receive function for the ESTABLISHED state.
5107 * It is split into a fast path and a slow path. The fast path is
5109 * - A zero window was announced from us - zero window probing
5110 * is only handled properly in the slow path.
5111 * - Out of order segments arrived.
5112 * - Urgent data is expected.
5113 * - There is no buffer space left
5114 * - Unexpected TCP flags/window values/header lengths are received
5115 * (detected by checking the TCP header against pred_flags)
5116 * - Data is sent in both directions. Fast path only supports pure senders
5117 * or pure receivers (this means either the sequence number or the ack
5118 * value must stay constant)
5119 * - Unexpected TCP option.
5121 * When these conditions are not satisfied it drops into a standard
5122 * receive procedure patterned after RFC793 to handle all cases.
5123 * The first three cases are guaranteed by proper pred_flags setting,
5124 * the rest is checked inline. Fast processing is turned on in
5125 * tcp_data_queue when everything is OK.
5127 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5128 struct tcphdr *th, unsigned len)
5130 struct tcp_sock *tp = tcp_sk(sk);
5134 * Header prediction.
5135 * The code loosely follows the one in the famous
5136 * "30 instruction TCP receive" Van Jacobson mail.
5138 * Van's trick is to deposit buffers into socket queue
5139 * on a device interrupt, to call tcp_recv function
5140 * on the receive process context and checksum and copy
5141 * the buffer to user space. smart...
5143 * Our current scheme is not silly either but we take the
5144 * extra cost of the net_bh soft interrupt processing...
5145 * We do checksum and copy also but from device to kernel.
5148 tp->rx_opt.saw_tstamp = 0;
5150 /* pred_flags is 0xS?10 << 16 + snd_wnd
5151 * if header_prediction is to be made
5152 * 'S' will always be tp->tcp_header_len >> 2
5153 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5154 * turn it off (when there are holes in the receive
5155 * space for instance)
5156 * PSH flag is ignored.
5159 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5160 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5161 int tcp_header_len = tp->tcp_header_len;
5163 /* Timestamp header prediction: tcp_header_len
5164 * is automatically equal to th->doff*4 due to pred_flags
5168 /* Check timestamp */
5169 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5170 /* No? Slow path! */
5171 if (!tcp_parse_aligned_timestamp(tp, th))
5174 /* If PAWS failed, check it more carefully in slow path */
5175 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5178 /* DO NOT update ts_recent here, if checksum fails
5179 * and timestamp was corrupted part, it will result
5180 * in a hung connection since we will drop all
5181 * future packets due to the PAWS test.
5185 if (len <= tcp_header_len) {
5186 /* Bulk data transfer: sender */
5187 if (len == tcp_header_len) {
5188 /* Predicted packet is in window by definition.
5189 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5190 * Hence, check seq<=rcv_wup reduces to:
5192 if (tcp_header_len ==
5193 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5194 tp->rcv_nxt == tp->rcv_wup)
5195 tcp_store_ts_recent(tp);
5197 /* We know that such packets are checksummed
5200 tcp_ack(sk, skb, 0);
5202 tcp_data_snd_check(sk);
5204 } else { /* Header too small */
5205 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5210 int copied_early = 0;
5212 if (tp->copied_seq == tp->rcv_nxt &&
5213 len - tcp_header_len <= tp->ucopy.len) {
5214 #ifdef CONFIG_NET_DMA
5215 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5220 if (tp->ucopy.task == current &&
5221 sock_owned_by_user(sk) && !copied_early) {
5222 __set_current_state(TASK_RUNNING);
5224 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5228 /* Predicted packet is in window by definition.
5229 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5230 * Hence, check seq<=rcv_wup reduces to:
5232 if (tcp_header_len ==
5233 (sizeof(struct tcphdr) +
5234 TCPOLEN_TSTAMP_ALIGNED) &&
5235 tp->rcv_nxt == tp->rcv_wup)
5236 tcp_store_ts_recent(tp);
5238 tcp_rcv_rtt_measure_ts(sk, skb);
5240 __skb_pull(skb, tcp_header_len);
5241 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5242 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5245 tcp_cleanup_rbuf(sk, skb->len);
5248 if (tcp_checksum_complete_user(sk, skb))
5251 /* Predicted packet is in window by definition.
5252 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5253 * Hence, check seq<=rcv_wup reduces to:
5255 if (tcp_header_len ==
5256 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5257 tp->rcv_nxt == tp->rcv_wup)
5258 tcp_store_ts_recent(tp);
5260 tcp_rcv_rtt_measure_ts(sk, skb);
5262 if ((int)skb->truesize > sk->sk_forward_alloc)
5265 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5267 /* Bulk data transfer: receiver */
5268 __skb_pull(skb, tcp_header_len);
5269 __skb_queue_tail(&sk->sk_receive_queue, skb);
5270 skb_set_owner_r(skb, sk);
5271 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5274 tcp_event_data_recv(sk, skb);
5276 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5277 /* Well, only one small jumplet in fast path... */
5278 tcp_ack(sk, skb, FLAG_DATA);
5279 tcp_data_snd_check(sk);
5280 if (!inet_csk_ack_scheduled(sk))
5284 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5285 __tcp_ack_snd_check(sk, 0);
5287 #ifdef CONFIG_NET_DMA
5289 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5295 sk->sk_data_ready(sk, 0);
5301 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5305 * Standard slow path.
5308 res = tcp_validate_incoming(sk, skb, th, 1);
5314 tcp_ack(sk, skb, FLAG_SLOWPATH);
5316 tcp_rcv_rtt_measure_ts(sk, skb);
5318 /* Process urgent data. */
5319 tcp_urg(sk, skb, th);
5321 /* step 7: process the segment text */
5322 tcp_data_queue(sk, skb);
5324 tcp_data_snd_check(sk);
5325 tcp_ack_snd_check(sk);
5329 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5336 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5337 struct tcphdr *th, unsigned len)
5339 struct tcp_sock *tp = tcp_sk(sk);
5340 struct inet_connection_sock *icsk = inet_csk(sk);
5341 int saved_clamp = tp->rx_opt.mss_clamp;
5343 tcp_parse_options(skb, &tp->rx_opt, 0);
5347 * "If the state is SYN-SENT then
5348 * first check the ACK bit
5349 * If the ACK bit is set
5350 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5351 * a reset (unless the RST bit is set, if so drop
5352 * the segment and return)"
5354 * We do not send data with SYN, so that RFC-correct
5357 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5358 goto reset_and_undo;
5360 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5361 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5363 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5364 goto reset_and_undo;
5367 /* Now ACK is acceptable.
5369 * "If the RST bit is set
5370 * If the ACK was acceptable then signal the user "error:
5371 * connection reset", drop the segment, enter CLOSED state,
5372 * delete TCB, and return."
5381 * "fifth, if neither of the SYN or RST bits is set then
5382 * drop the segment and return."
5388 goto discard_and_undo;
5391 * "If the SYN bit is on ...
5392 * are acceptable then ...
5393 * (our SYN has been ACKed), change the connection
5394 * state to ESTABLISHED..."
5397 TCP_ECN_rcv_synack(tp, th);
5399 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5400 tcp_ack(sk, skb, FLAG_SLOWPATH);
5402 /* Ok.. it's good. Set up sequence numbers and
5403 * move to established.
5405 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5406 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5408 /* RFC1323: The window in SYN & SYN/ACK segments is
5411 tp->snd_wnd = ntohs(th->window);
5412 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
5414 if (!tp->rx_opt.wscale_ok) {
5415 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5416 tp->window_clamp = min(tp->window_clamp, 65535U);
5419 if (tp->rx_opt.saw_tstamp) {
5420 tp->rx_opt.tstamp_ok = 1;
5421 tp->tcp_header_len =
5422 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5423 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5424 tcp_store_ts_recent(tp);
5426 tp->tcp_header_len = sizeof(struct tcphdr);
5429 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5430 tcp_enable_fack(tp);
5433 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5434 tcp_initialize_rcv_mss(sk);
5436 /* Remember, tcp_poll() does not lock socket!
5437 * Change state from SYN-SENT only after copied_seq
5438 * is initialized. */
5439 tp->copied_seq = tp->rcv_nxt;
5441 tcp_set_state(sk, TCP_ESTABLISHED);
5443 security_inet_conn_established(sk, skb);
5445 /* Make sure socket is routed, for correct metrics. */
5446 icsk->icsk_af_ops->rebuild_header(sk);
5448 tcp_init_metrics(sk);
5450 tcp_init_congestion_control(sk);
5452 /* Prevent spurious tcp_cwnd_restart() on first data
5455 tp->lsndtime = tcp_time_stamp;
5457 tcp_init_buffer_space(sk);
5459 if (sock_flag(sk, SOCK_KEEPOPEN))
5460 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5462 if (!tp->rx_opt.snd_wscale)
5463 __tcp_fast_path_on(tp, tp->snd_wnd);
5467 if (!sock_flag(sk, SOCK_DEAD)) {
5468 sk->sk_state_change(sk);
5469 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5472 if (sk->sk_write_pending ||
5473 icsk->icsk_accept_queue.rskq_defer_accept ||
5474 icsk->icsk_ack.pingpong) {
5475 /* Save one ACK. Data will be ready after
5476 * several ticks, if write_pending is set.
5478 * It may be deleted, but with this feature tcpdumps
5479 * look so _wonderfully_ clever, that I was not able
5480 * to stand against the temptation 8) --ANK
5482 inet_csk_schedule_ack(sk);
5483 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5484 icsk->icsk_ack.ato = TCP_ATO_MIN;
5485 tcp_incr_quickack(sk);
5486 tcp_enter_quickack_mode(sk);
5487 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5488 TCP_DELACK_MAX, TCP_RTO_MAX);
5499 /* No ACK in the segment */
5503 * "If the RST bit is set
5505 * Otherwise (no ACK) drop the segment and return."
5508 goto discard_and_undo;
5512 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5513 tcp_paws_check(&tp->rx_opt, 0))
5514 goto discard_and_undo;
5517 /* We see SYN without ACK. It is attempt of
5518 * simultaneous connect with crossed SYNs.
5519 * Particularly, it can be connect to self.
5521 tcp_set_state(sk, TCP_SYN_RECV);
5523 if (tp->rx_opt.saw_tstamp) {
5524 tp->rx_opt.tstamp_ok = 1;
5525 tcp_store_ts_recent(tp);
5526 tp->tcp_header_len =
5527 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5529 tp->tcp_header_len = sizeof(struct tcphdr);
5532 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5533 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5535 /* RFC1323: The window in SYN & SYN/ACK segments is
5538 tp->snd_wnd = ntohs(th->window);
5539 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5540 tp->max_window = tp->snd_wnd;
5542 TCP_ECN_rcv_syn(tp, th);
5545 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5546 tcp_initialize_rcv_mss(sk);
5548 tcp_send_synack(sk);
5550 /* Note, we could accept data and URG from this segment.
5551 * There are no obstacles to make this.
5553 * However, if we ignore data in ACKless segments sometimes,
5554 * we have no reasons to accept it sometimes.
5555 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5556 * is not flawless. So, discard packet for sanity.
5557 * Uncomment this return to process the data.
5564 /* "fifth, if neither of the SYN or RST bits is set then
5565 * drop the segment and return."
5569 tcp_clear_options(&tp->rx_opt);
5570 tp->rx_opt.mss_clamp = saved_clamp;
5574 tcp_clear_options(&tp->rx_opt);
5575 tp->rx_opt.mss_clamp = saved_clamp;
5580 * This function implements the receiving procedure of RFC 793 for
5581 * all states except ESTABLISHED and TIME_WAIT.
5582 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5583 * address independent.
5586 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5587 struct tcphdr *th, unsigned len)
5589 struct tcp_sock *tp = tcp_sk(sk);
5590 struct inet_connection_sock *icsk = inet_csk(sk);
5594 tp->rx_opt.saw_tstamp = 0;
5596 switch (sk->sk_state) {
5608 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5611 /* Now we have several options: In theory there is
5612 * nothing else in the frame. KA9Q has an option to
5613 * send data with the syn, BSD accepts data with the
5614 * syn up to the [to be] advertised window and
5615 * Solaris 2.1 gives you a protocol error. For now
5616 * we just ignore it, that fits the spec precisely
5617 * and avoids incompatibilities. It would be nice in
5618 * future to drop through and process the data.
5620 * Now that TTCP is starting to be used we ought to
5622 * But, this leaves one open to an easy denial of
5623 * service attack, and SYN cookies can't defend
5624 * against this problem. So, we drop the data
5625 * in the interest of security over speed unless
5626 * it's still in use.
5634 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5638 /* Do step6 onward by hand. */
5639 tcp_urg(sk, skb, th);
5641 tcp_data_snd_check(sk);
5645 res = tcp_validate_incoming(sk, skb, th, 0);
5649 /* step 5: check the ACK field */
5651 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5653 switch (sk->sk_state) {
5656 tp->copied_seq = tp->rcv_nxt;
5658 tcp_set_state(sk, TCP_ESTABLISHED);
5659 sk->sk_state_change(sk);
5661 /* Note, that this wakeup is only for marginal
5662 * crossed SYN case. Passively open sockets
5663 * are not waked up, because sk->sk_sleep ==
5664 * NULL and sk->sk_socket == NULL.
5668 SOCK_WAKE_IO, POLL_OUT);
5670 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5671 tp->snd_wnd = ntohs(th->window) <<
5672 tp->rx_opt.snd_wscale;
5673 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5674 TCP_SKB_CB(skb)->seq);
5676 /* tcp_ack considers this ACK as duplicate
5677 * and does not calculate rtt.
5678 * Fix it at least with timestamps.
5680 if (tp->rx_opt.saw_tstamp &&
5681 tp->rx_opt.rcv_tsecr && !tp->srtt)
5682 tcp_ack_saw_tstamp(sk, 0);
5684 if (tp->rx_opt.tstamp_ok)
5685 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5687 /* Make sure socket is routed, for
5690 icsk->icsk_af_ops->rebuild_header(sk);
5692 tcp_init_metrics(sk);
5694 tcp_init_congestion_control(sk);
5696 /* Prevent spurious tcp_cwnd_restart() on
5697 * first data packet.
5699 tp->lsndtime = tcp_time_stamp;
5702 tcp_initialize_rcv_mss(sk);
5703 tcp_init_buffer_space(sk);
5704 tcp_fast_path_on(tp);
5711 if (tp->snd_una == tp->write_seq) {
5712 tcp_set_state(sk, TCP_FIN_WAIT2);
5713 sk->sk_shutdown |= SEND_SHUTDOWN;
5714 dst_confirm(sk->sk_dst_cache);
5716 if (!sock_flag(sk, SOCK_DEAD))
5717 /* Wake up lingering close() */
5718 sk->sk_state_change(sk);
5722 if (tp->linger2 < 0 ||
5723 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5724 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5726 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5730 tmo = tcp_fin_time(sk);
5731 if (tmo > TCP_TIMEWAIT_LEN) {
5732 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5733 } else if (th->fin || sock_owned_by_user(sk)) {
5734 /* Bad case. We could lose such FIN otherwise.
5735 * It is not a big problem, but it looks confusing
5736 * and not so rare event. We still can lose it now,
5737 * if it spins in bh_lock_sock(), but it is really
5740 inet_csk_reset_keepalive_timer(sk, tmo);
5742 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5750 if (tp->snd_una == tp->write_seq) {
5751 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5757 if (tp->snd_una == tp->write_seq) {
5758 tcp_update_metrics(sk);
5767 /* step 6: check the URG bit */
5768 tcp_urg(sk, skb, th);
5770 /* step 7: process the segment text */
5771 switch (sk->sk_state) {
5772 case TCP_CLOSE_WAIT:
5775 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5779 /* RFC 793 says to queue data in these states,
5780 * RFC 1122 says we MUST send a reset.
5781 * BSD 4.4 also does reset.
5783 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5784 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5785 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5786 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5792 case TCP_ESTABLISHED:
5793 tcp_data_queue(sk, skb);
5798 /* tcp_data could move socket to TIME-WAIT */
5799 if (sk->sk_state != TCP_CLOSE) {
5800 tcp_data_snd_check(sk);
5801 tcp_ack_snd_check(sk);
5811 EXPORT_SYMBOL(sysctl_tcp_ecn);
5812 EXPORT_SYMBOL(sysctl_tcp_reordering);
5813 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
5814 EXPORT_SYMBOL(tcp_parse_options);
5815 #ifdef CONFIG_TCP_MD5SIG
5816 EXPORT_SYMBOL(tcp_parse_md5sig_option);
5818 EXPORT_SYMBOL(tcp_rcv_established);
5819 EXPORT_SYMBOL(tcp_rcv_state_process);
5820 EXPORT_SYMBOL(tcp_initialize_rcv_mss);