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/slab.h>
66 #include <linux/module.h>
67 #include <linux/sysctl.h>
68 #include <linux/kernel.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
74 #include <net/netdma.h>
76 int sysctl_tcp_timestamps __read_mostly = 1;
77 int sysctl_tcp_window_scaling __read_mostly = 1;
78 int sysctl_tcp_sack __read_mostly = 1;
79 int sysctl_tcp_fack __read_mostly = 1;
80 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
81 EXPORT_SYMBOL(sysctl_tcp_reordering);
82 int sysctl_tcp_ecn __read_mostly = 2;
83 EXPORT_SYMBOL(sysctl_tcp_ecn);
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 2;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 int sysctl_tcp_stdurg __read_mostly;
90 int sysctl_tcp_rfc1337 __read_mostly;
91 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
92 int sysctl_tcp_frto __read_mostly = 2;
93 int sysctl_tcp_frto_response __read_mostly;
94 int sysctl_tcp_nometrics_save __read_mostly;
96 int sysctl_tcp_thin_dupack __read_mostly;
98 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
99 int sysctl_tcp_abc __read_mostly;
101 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
102 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
103 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
104 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
105 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
106 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
107 #define FLAG_ECE 0x40 /* ECE in this ACK */
108 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
130 struct inet_connection_sock *icsk = inet_csk(sk);
131 const unsigned int lss = icsk->icsk_ack.last_seg_size;
134 icsk->icsk_ack.last_seg_size = 0;
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
139 len = skb_shinfo(skb)->gso_size ? : skb->len;
140 if (len >= icsk->icsk_ack.rcv_mss) {
141 icsk->icsk_ack.rcv_mss = len;
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
146 * "len" is invariant segment length, including TCP header.
148 len += skb->data - skb_transport_header(skb);
149 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
155 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
156 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
161 len -= tcp_sk(sk)->tcp_header_len;
162 icsk->icsk_ack.last_seg_size = len;
164 icsk->icsk_ack.rcv_mss = len;
168 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
174 static void tcp_incr_quickack(struct sock *sk)
176 struct inet_connection_sock *icsk = inet_csk(sk);
177 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
181 if (quickacks > icsk->icsk_ack.quick)
182 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185 static void tcp_enter_quickack_mode(struct sock *sk)
187 struct inet_connection_sock *icsk = inet_csk(sk);
188 tcp_incr_quickack(sk);
189 icsk->icsk_ack.pingpong = 0;
190 icsk->icsk_ack.ato = TCP_ATO_MIN;
193 /* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
197 static inline int tcp_in_quickack_mode(const struct sock *sk)
199 const struct inet_connection_sock *icsk = inet_csk(sk);
200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
205 if (tp->ecn_flags & TCP_ECN_OK)
206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
222 if (!(tp->ecn_flags & TCP_ECN_OK))
225 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
226 case INET_ECN_NOT_ECT:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp->ecn_flags & TCP_ECN_SEEN)
232 tcp_enter_quickack_mode((struct sock *)tp);
235 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
238 tp->ecn_flags |= TCP_ECN_SEEN;
242 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
244 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
245 tp->ecn_flags &= ~TCP_ECN_OK;
248 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
250 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
251 tp->ecn_flags &= ~TCP_ECN_OK;
254 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
256 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
261 /* Buffer size and advertised window tuning.
263 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
266 static void tcp_fixup_sndbuf(struct sock *sk)
268 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
269 sizeof(struct sk_buff);
271 if (sk->sk_sndbuf < 3 * sndmem) {
272 sk->sk_sndbuf = 3 * sndmem;
273 if (sk->sk_sndbuf > sysctl_tcp_wmem[2])
274 sk->sk_sndbuf = sysctl_tcp_wmem[2];
278 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
280 * All tcp_full_space() is split to two parts: "network" buffer, allocated
281 * forward and advertised in receiver window (tp->rcv_wnd) and
282 * "application buffer", required to isolate scheduling/application
283 * latencies from network.
284 * window_clamp is maximal advertised window. It can be less than
285 * tcp_full_space(), in this case tcp_full_space() - window_clamp
286 * is reserved for "application" buffer. The less window_clamp is
287 * the smoother our behaviour from viewpoint of network, but the lower
288 * throughput and the higher sensitivity of the connection to losses. 8)
290 * rcv_ssthresh is more strict window_clamp used at "slow start"
291 * phase to predict further behaviour of this connection.
292 * It is used for two goals:
293 * - to enforce header prediction at sender, even when application
294 * requires some significant "application buffer". It is check #1.
295 * - to prevent pruning of receive queue because of misprediction
296 * of receiver window. Check #2.
298 * The scheme does not work when sender sends good segments opening
299 * window and then starts to feed us spaghetti. But it should work
300 * in common situations. Otherwise, we have to rely on queue collapsing.
303 /* Slow part of check#2. */
304 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
306 struct tcp_sock *tp = tcp_sk(sk);
308 int truesize = tcp_win_from_space(skb->truesize) >> 1;
309 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
311 while (tp->rcv_ssthresh <= window) {
312 if (truesize <= skb->len)
313 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
321 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
323 struct tcp_sock *tp = tcp_sk(sk);
326 if (tp->rcv_ssthresh < tp->window_clamp &&
327 (int)tp->rcv_ssthresh < tcp_space(sk) &&
328 !tcp_memory_pressure) {
331 /* Check #2. Increase window, if skb with such overhead
332 * will fit to rcvbuf in future.
334 if (tcp_win_from_space(skb->truesize) <= skb->len)
335 incr = 2 * tp->advmss;
337 incr = __tcp_grow_window(sk, skb);
340 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
342 inet_csk(sk)->icsk_ack.quick |= 1;
347 /* 3. Tuning rcvbuf, when connection enters established state. */
349 static void tcp_fixup_rcvbuf(struct sock *sk)
351 struct tcp_sock *tp = tcp_sk(sk);
352 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
354 /* Try to select rcvbuf so that 4 mss-sized segments
355 * will fit to window and corresponding skbs will fit to our rcvbuf.
356 * (was 3; 4 is minimum to allow fast retransmit to work.)
358 while (tcp_win_from_space(rcvmem) < tp->advmss)
360 if (sk->sk_rcvbuf < 4 * rcvmem)
361 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
364 /* 4. Try to fixup all. It is made immediately after connection enters
367 static void tcp_init_buffer_space(struct sock *sk)
369 struct tcp_sock *tp = tcp_sk(sk);
372 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
373 tcp_fixup_rcvbuf(sk);
374 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
375 tcp_fixup_sndbuf(sk);
377 tp->rcvq_space.space = tp->rcv_wnd;
379 maxwin = tcp_full_space(sk);
381 if (tp->window_clamp >= maxwin) {
382 tp->window_clamp = maxwin;
384 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
385 tp->window_clamp = max(maxwin -
386 (maxwin >> sysctl_tcp_app_win),
390 /* Force reservation of one segment. */
391 if (sysctl_tcp_app_win &&
392 tp->window_clamp > 2 * tp->advmss &&
393 tp->window_clamp + tp->advmss > maxwin)
394 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
396 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
397 tp->snd_cwnd_stamp = tcp_time_stamp;
400 /* 5. Recalculate window clamp after socket hit its memory bounds. */
401 static void tcp_clamp_window(struct sock *sk)
403 struct tcp_sock *tp = tcp_sk(sk);
404 struct inet_connection_sock *icsk = inet_csk(sk);
406 icsk->icsk_ack.quick = 0;
408 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
409 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
410 !tcp_memory_pressure &&
411 atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
412 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
415 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
416 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
419 /* Initialize RCV_MSS value.
420 * RCV_MSS is an our guess about MSS used by the peer.
421 * We haven't any direct information about the MSS.
422 * It's better to underestimate the RCV_MSS rather than overestimate.
423 * Overestimations make us ACKing less frequently than needed.
424 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
426 void tcp_initialize_rcv_mss(struct sock *sk)
428 struct tcp_sock *tp = tcp_sk(sk);
429 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
431 hint = min(hint, tp->rcv_wnd / 2);
432 hint = min(hint, TCP_MSS_DEFAULT);
433 hint = max(hint, TCP_MIN_MSS);
435 inet_csk(sk)->icsk_ack.rcv_mss = hint;
437 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
439 /* Receiver "autotuning" code.
441 * The algorithm for RTT estimation w/o timestamps is based on
442 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
443 * <http://public.lanl.gov/radiant/pubs.html#DRS>
445 * More detail on this code can be found at
446 * <http://staff.psc.edu/jheffner/>,
447 * though this reference is out of date. A new paper
450 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
452 u32 new_sample = tp->rcv_rtt_est.rtt;
458 if (new_sample != 0) {
459 /* If we sample in larger samples in the non-timestamp
460 * case, we could grossly overestimate the RTT especially
461 * with chatty applications or bulk transfer apps which
462 * are stalled on filesystem I/O.
464 * Also, since we are only going for a minimum in the
465 * non-timestamp case, we do not smooth things out
466 * else with timestamps disabled convergence takes too
470 m -= (new_sample >> 3);
472 } else if (m < new_sample)
475 /* No previous measure. */
479 if (tp->rcv_rtt_est.rtt != new_sample)
480 tp->rcv_rtt_est.rtt = new_sample;
483 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
485 if (tp->rcv_rtt_est.time == 0)
487 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
489 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
492 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
493 tp->rcv_rtt_est.time = tcp_time_stamp;
496 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
497 const struct sk_buff *skb)
499 struct tcp_sock *tp = tcp_sk(sk);
500 if (tp->rx_opt.rcv_tsecr &&
501 (TCP_SKB_CB(skb)->end_seq -
502 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
503 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
507 * This function should be called every time data is copied to user space.
508 * It calculates the appropriate TCP receive buffer space.
510 void tcp_rcv_space_adjust(struct sock *sk)
512 struct tcp_sock *tp = tcp_sk(sk);
516 if (tp->rcvq_space.time == 0)
519 time = tcp_time_stamp - tp->rcvq_space.time;
520 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
523 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
525 space = max(tp->rcvq_space.space, space);
527 if (tp->rcvq_space.space != space) {
530 tp->rcvq_space.space = space;
532 if (sysctl_tcp_moderate_rcvbuf &&
533 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
534 int new_clamp = space;
536 /* Receive space grows, normalize in order to
537 * take into account packet headers and sk_buff
538 * structure overhead.
543 rcvmem = (tp->advmss + MAX_TCP_HEADER +
544 16 + sizeof(struct sk_buff));
545 while (tcp_win_from_space(rcvmem) < tp->advmss)
548 space = min(space, sysctl_tcp_rmem[2]);
549 if (space > sk->sk_rcvbuf) {
550 sk->sk_rcvbuf = space;
552 /* Make the window clamp follow along. */
553 tp->window_clamp = new_clamp;
559 tp->rcvq_space.seq = tp->copied_seq;
560 tp->rcvq_space.time = tcp_time_stamp;
563 /* There is something which you must keep in mind when you analyze the
564 * behavior of the tp->ato delayed ack timeout interval. When a
565 * connection starts up, we want to ack as quickly as possible. The
566 * problem is that "good" TCP's do slow start at the beginning of data
567 * transmission. The means that until we send the first few ACK's the
568 * sender will sit on his end and only queue most of his data, because
569 * he can only send snd_cwnd unacked packets at any given time. For
570 * each ACK we send, he increments snd_cwnd and transmits more of his
573 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
575 struct tcp_sock *tp = tcp_sk(sk);
576 struct inet_connection_sock *icsk = inet_csk(sk);
579 inet_csk_schedule_ack(sk);
581 tcp_measure_rcv_mss(sk, skb);
583 tcp_rcv_rtt_measure(tp);
585 now = tcp_time_stamp;
587 if (!icsk->icsk_ack.ato) {
588 /* The _first_ data packet received, initialize
589 * delayed ACK engine.
591 tcp_incr_quickack(sk);
592 icsk->icsk_ack.ato = TCP_ATO_MIN;
594 int m = now - icsk->icsk_ack.lrcvtime;
596 if (m <= TCP_ATO_MIN / 2) {
597 /* The fastest case is the first. */
598 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
599 } else if (m < icsk->icsk_ack.ato) {
600 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
601 if (icsk->icsk_ack.ato > icsk->icsk_rto)
602 icsk->icsk_ack.ato = icsk->icsk_rto;
603 } else if (m > icsk->icsk_rto) {
604 /* Too long gap. Apparently sender failed to
605 * restart window, so that we send ACKs quickly.
607 tcp_incr_quickack(sk);
611 icsk->icsk_ack.lrcvtime = now;
613 TCP_ECN_check_ce(tp, skb);
616 tcp_grow_window(sk, skb);
619 /* Called to compute a smoothed rtt estimate. The data fed to this
620 * routine either comes from timestamps, or from segments that were
621 * known _not_ to have been retransmitted [see Karn/Partridge
622 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
623 * piece by Van Jacobson.
624 * NOTE: the next three routines used to be one big routine.
625 * To save cycles in the RFC 1323 implementation it was better to break
626 * it up into three procedures. -- erics
628 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
630 struct tcp_sock *tp = tcp_sk(sk);
631 long m = mrtt; /* RTT */
633 /* The following amusing code comes from Jacobson's
634 * article in SIGCOMM '88. Note that rtt and mdev
635 * are scaled versions of rtt and mean deviation.
636 * This is designed to be as fast as possible
637 * m stands for "measurement".
639 * On a 1990 paper the rto value is changed to:
640 * RTO = rtt + 4 * mdev
642 * Funny. This algorithm seems to be very broken.
643 * These formulae increase RTO, when it should be decreased, increase
644 * too slowly, when it should be increased quickly, decrease too quickly
645 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
646 * does not matter how to _calculate_ it. Seems, it was trap
647 * that VJ failed to avoid. 8)
652 m -= (tp->srtt >> 3); /* m is now error in rtt est */
653 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
655 m = -m; /* m is now abs(error) */
656 m -= (tp->mdev >> 2); /* similar update on mdev */
657 /* This is similar to one of Eifel findings.
658 * Eifel blocks mdev updates when rtt decreases.
659 * This solution is a bit different: we use finer gain
660 * for mdev in this case (alpha*beta).
661 * Like Eifel it also prevents growth of rto,
662 * but also it limits too fast rto decreases,
663 * happening in pure Eifel.
668 m -= (tp->mdev >> 2); /* similar update on mdev */
670 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
671 if (tp->mdev > tp->mdev_max) {
672 tp->mdev_max = tp->mdev;
673 if (tp->mdev_max > tp->rttvar)
674 tp->rttvar = tp->mdev_max;
676 if (after(tp->snd_una, tp->rtt_seq)) {
677 if (tp->mdev_max < tp->rttvar)
678 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
679 tp->rtt_seq = tp->snd_nxt;
680 tp->mdev_max = tcp_rto_min(sk);
683 /* no previous measure. */
684 tp->srtt = m << 3; /* take the measured time to be rtt */
685 tp->mdev = m << 1; /* make sure rto = 3*rtt */
686 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
687 tp->rtt_seq = tp->snd_nxt;
691 /* Calculate rto without backoff. This is the second half of Van Jacobson's
692 * routine referred to above.
694 static inline void tcp_set_rto(struct sock *sk)
696 const struct tcp_sock *tp = tcp_sk(sk);
697 /* Old crap is replaced with new one. 8)
700 * 1. If rtt variance happened to be less 50msec, it is hallucination.
701 * It cannot be less due to utterly erratic ACK generation made
702 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
703 * to do with delayed acks, because at cwnd>2 true delack timeout
704 * is invisible. Actually, Linux-2.4 also generates erratic
705 * ACKs in some circumstances.
707 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
709 /* 2. Fixups made earlier cannot be right.
710 * If we do not estimate RTO correctly without them,
711 * all the algo is pure shit and should be replaced
712 * with correct one. It is exactly, which we pretend to do.
715 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
716 * guarantees that rto is higher.
721 /* Save metrics learned by this TCP session.
722 This function is called only, when TCP finishes successfully
723 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
725 void tcp_update_metrics(struct sock *sk)
727 struct tcp_sock *tp = tcp_sk(sk);
728 struct dst_entry *dst = __sk_dst_get(sk);
730 if (sysctl_tcp_nometrics_save)
735 if (dst && (dst->flags & DST_HOST)) {
736 const struct inet_connection_sock *icsk = inet_csk(sk);
740 if (icsk->icsk_backoff || !tp->srtt) {
741 /* This session failed to estimate rtt. Why?
742 * Probably, no packets returned in time.
745 if (!(dst_metric_locked(dst, RTAX_RTT)))
746 dst_metric_set(dst, RTAX_RTT, 0);
750 rtt = dst_metric_rtt(dst, RTAX_RTT);
753 /* If newly calculated rtt larger than stored one,
754 * store new one. Otherwise, use EWMA. Remember,
755 * rtt overestimation is always better than underestimation.
757 if (!(dst_metric_locked(dst, RTAX_RTT))) {
759 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
761 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
764 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
769 /* Scale deviation to rttvar fixed point */
774 var = dst_metric_rtt(dst, RTAX_RTTVAR);
778 var -= (var - m) >> 2;
780 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
783 if (tcp_in_initial_slowstart(tp)) {
784 /* Slow start still did not finish. */
785 if (dst_metric(dst, RTAX_SSTHRESH) &&
786 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
787 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
788 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
789 if (!dst_metric_locked(dst, RTAX_CWND) &&
790 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
791 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
792 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
793 icsk->icsk_ca_state == TCP_CA_Open) {
794 /* Cong. avoidance phase, cwnd is reliable. */
795 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
796 dst_metric_set(dst, RTAX_SSTHRESH,
797 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
798 if (!dst_metric_locked(dst, RTAX_CWND))
799 dst_metric_set(dst, RTAX_CWND,
800 (dst_metric(dst, RTAX_CWND) +
803 /* Else slow start did not finish, cwnd is non-sense,
804 ssthresh may be also invalid.
806 if (!dst_metric_locked(dst, RTAX_CWND))
807 dst_metric_set(dst, RTAX_CWND,
808 (dst_metric(dst, RTAX_CWND) +
809 tp->snd_ssthresh) >> 1);
810 if (dst_metric(dst, RTAX_SSTHRESH) &&
811 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
812 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
813 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
816 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
817 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
818 tp->reordering != sysctl_tcp_reordering)
819 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
824 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
826 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
829 cwnd = TCP_INIT_CWND;
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 /* ssthresh may have been reduced unnecessarily during.
894 * 3WHS. Restore it back to its initial default.
896 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
898 if (dst_metric(dst, RTAX_REORDERING) &&
899 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
900 tcp_disable_fack(tp);
901 tp->reordering = dst_metric(dst, RTAX_REORDERING);
904 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
907 /* Initial rtt is determined from SYN,SYN-ACK.
908 * The segment is small and rtt may appear much
909 * less than real one. Use per-dst memory
910 * to make it more realistic.
912 * A bit of theory. RTT is time passed after "normal" sized packet
913 * is sent until it is ACKed. In normal circumstances sending small
914 * packets force peer to delay ACKs and calculation is correct too.
915 * The algorithm is adaptive and, provided we follow specs, it
916 * NEVER underestimate RTT. BUT! If peer tries to make some clever
917 * tricks sort of "quick acks" for time long enough to decrease RTT
918 * to low value, and then abruptly stops to do it and starts to delay
919 * ACKs, wait for troubles.
921 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
922 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
923 tp->rtt_seq = tp->snd_nxt;
925 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
926 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
927 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
932 /* RFC2988bis: We've failed to get a valid RTT sample from
933 * 3WHS. This is most likely due to retransmission,
934 * including spurious one. Reset the RTO back to 3secs
935 * from the more aggressive 1sec to avoid more spurious
938 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
939 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
941 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
942 * retransmitted. In light of RFC2988bis' more aggressive 1sec
943 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
944 * retransmission has occurred.
946 if (tp->total_retrans > 1)
949 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
950 tp->snd_cwnd_stamp = tcp_time_stamp;
953 static void tcp_update_reordering(struct sock *sk, const int metric,
956 struct tcp_sock *tp = tcp_sk(sk);
957 if (metric > tp->reordering) {
960 tp->reordering = min(TCP_MAX_REORDERING, metric);
962 /* This exciting event is worth to be remembered. 8) */
964 mib_idx = LINUX_MIB_TCPTSREORDER;
965 else if (tcp_is_reno(tp))
966 mib_idx = LINUX_MIB_TCPRENOREORDER;
967 else if (tcp_is_fack(tp))
968 mib_idx = LINUX_MIB_TCPFACKREORDER;
970 mib_idx = LINUX_MIB_TCPSACKREORDER;
972 NET_INC_STATS_BH(sock_net(sk), mib_idx);
973 #if FASTRETRANS_DEBUG > 1
974 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
975 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
979 tp->undo_marker ? tp->undo_retrans : 0);
981 tcp_disable_fack(tp);
985 /* This must be called before lost_out is incremented */
986 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
988 if ((tp->retransmit_skb_hint == NULL) ||
989 before(TCP_SKB_CB(skb)->seq,
990 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
991 tp->retransmit_skb_hint = skb;
994 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
995 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
998 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1000 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1001 tcp_verify_retransmit_hint(tp, skb);
1003 tp->lost_out += tcp_skb_pcount(skb);
1004 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1008 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1009 struct sk_buff *skb)
1011 tcp_verify_retransmit_hint(tp, skb);
1013 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1014 tp->lost_out += tcp_skb_pcount(skb);
1015 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1019 /* This procedure tags the retransmission queue when SACKs arrive.
1021 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1022 * Packets in queue with these bits set are counted in variables
1023 * sacked_out, retrans_out and lost_out, correspondingly.
1025 * Valid combinations are:
1026 * Tag InFlight Description
1027 * 0 1 - orig segment is in flight.
1028 * S 0 - nothing flies, orig reached receiver.
1029 * L 0 - nothing flies, orig lost by net.
1030 * R 2 - both orig and retransmit are in flight.
1031 * L|R 1 - orig is lost, retransmit is in flight.
1032 * S|R 1 - orig reached receiver, retrans is still in flight.
1033 * (L|S|R is logically valid, it could occur when L|R is sacked,
1034 * but it is equivalent to plain S and code short-curcuits it to S.
1035 * L|S is logically invalid, it would mean -1 packet in flight 8))
1037 * These 6 states form finite state machine, controlled by the following events:
1038 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1039 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1040 * 3. Loss detection event of one of three flavors:
1041 * A. Scoreboard estimator decided the packet is lost.
1042 * A'. Reno "three dupacks" marks head of queue lost.
1043 * A''. Its FACK modfication, head until snd.fack is lost.
1044 * B. SACK arrives sacking data transmitted after never retransmitted
1045 * hole was sent out.
1046 * C. SACK arrives sacking SND.NXT at the moment, when the
1047 * segment was retransmitted.
1048 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1050 * It is pleasant to note, that state diagram turns out to be commutative,
1051 * so that we are allowed not to be bothered by order of our actions,
1052 * when multiple events arrive simultaneously. (see the function below).
1054 * Reordering detection.
1055 * --------------------
1056 * Reordering metric is maximal distance, which a packet can be displaced
1057 * in packet stream. With SACKs we can estimate it:
1059 * 1. SACK fills old hole and the corresponding segment was not
1060 * ever retransmitted -> reordering. Alas, we cannot use it
1061 * when segment was retransmitted.
1062 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1063 * for retransmitted and already SACKed segment -> reordering..
1064 * Both of these heuristics are not used in Loss state, when we cannot
1065 * account for retransmits accurately.
1067 * SACK block validation.
1068 * ----------------------
1070 * SACK block range validation checks that the received SACK block fits to
1071 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1072 * Note that SND.UNA is not included to the range though being valid because
1073 * it means that the receiver is rather inconsistent with itself reporting
1074 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1075 * perfectly valid, however, in light of RFC2018 which explicitly states
1076 * that "SACK block MUST reflect the newest segment. Even if the newest
1077 * segment is going to be discarded ...", not that it looks very clever
1078 * in case of head skb. Due to potentional receiver driven attacks, we
1079 * choose to avoid immediate execution of a walk in write queue due to
1080 * reneging and defer head skb's loss recovery to standard loss recovery
1081 * procedure that will eventually trigger (nothing forbids us doing this).
1083 * Implements also blockage to start_seq wrap-around. Problem lies in the
1084 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1085 * there's no guarantee that it will be before snd_nxt (n). The problem
1086 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1089 * <- outs wnd -> <- wrapzone ->
1090 * u e n u_w e_w s n_w
1092 * |<------------+------+----- TCP seqno space --------------+---------->|
1093 * ...-- <2^31 ->| |<--------...
1094 * ...---- >2^31 ------>| |<--------...
1096 * Current code wouldn't be vulnerable but it's better still to discard such
1097 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1098 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1099 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1100 * equal to the ideal case (infinite seqno space without wrap caused issues).
1102 * With D-SACK the lower bound is extended to cover sequence space below
1103 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1104 * again, D-SACK block must not to go across snd_una (for the same reason as
1105 * for the normal SACK blocks, explained above). But there all simplicity
1106 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1107 * fully below undo_marker they do not affect behavior in anyway and can
1108 * therefore be safely ignored. In rare cases (which are more or less
1109 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1110 * fragmentation and packet reordering past skb's retransmission. To consider
1111 * them correctly, the acceptable range must be extended even more though
1112 * the exact amount is rather hard to quantify. However, tp->max_window can
1113 * be used as an exaggerated estimate.
1115 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1116 u32 start_seq, u32 end_seq)
1118 /* Too far in future, or reversed (interpretation is ambiguous) */
1119 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1122 /* Nasty start_seq wrap-around check (see comments above) */
1123 if (!before(start_seq, tp->snd_nxt))
1126 /* In outstanding window? ...This is valid exit for D-SACKs too.
1127 * start_seq == snd_una is non-sensical (see comments above)
1129 if (after(start_seq, tp->snd_una))
1132 if (!is_dsack || !tp->undo_marker)
1135 /* ...Then it's D-SACK, and must reside below snd_una completely */
1136 if (after(end_seq, tp->snd_una))
1139 if (!before(start_seq, tp->undo_marker))
1143 if (!after(end_seq, tp->undo_marker))
1146 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1147 * start_seq < undo_marker and end_seq >= undo_marker.
1149 return !before(start_seq, end_seq - tp->max_window);
1152 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1153 * Event "C". Later note: FACK people cheated me again 8), we have to account
1154 * for reordering! Ugly, but should help.
1156 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1157 * less than what is now known to be received by the other end (derived from
1158 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1159 * retransmitted skbs to avoid some costly processing per ACKs.
1161 static void tcp_mark_lost_retrans(struct sock *sk)
1163 const struct inet_connection_sock *icsk = inet_csk(sk);
1164 struct tcp_sock *tp = tcp_sk(sk);
1165 struct sk_buff *skb;
1167 u32 new_low_seq = tp->snd_nxt;
1168 u32 received_upto = tcp_highest_sack_seq(tp);
1170 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1171 !after(received_upto, tp->lost_retrans_low) ||
1172 icsk->icsk_ca_state != TCP_CA_Recovery)
1175 tcp_for_write_queue(skb, sk) {
1176 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1178 if (skb == tcp_send_head(sk))
1180 if (cnt == tp->retrans_out)
1182 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1185 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1188 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1189 * constraint here (see above) but figuring out that at
1190 * least tp->reordering SACK blocks reside between ack_seq
1191 * and received_upto is not easy task to do cheaply with
1192 * the available datastructures.
1194 * Whether FACK should check here for tp->reordering segs
1195 * in-between one could argue for either way (it would be
1196 * rather simple to implement as we could count fack_count
1197 * during the walk and do tp->fackets_out - fack_count).
1199 if (after(received_upto, ack_seq)) {
1200 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1201 tp->retrans_out -= tcp_skb_pcount(skb);
1203 tcp_skb_mark_lost_uncond_verify(tp, skb);
1204 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1206 if (before(ack_seq, new_low_seq))
1207 new_low_seq = ack_seq;
1208 cnt += tcp_skb_pcount(skb);
1212 if (tp->retrans_out)
1213 tp->lost_retrans_low = new_low_seq;
1216 static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
1217 struct tcp_sack_block_wire *sp, int num_sacks,
1220 struct tcp_sock *tp = tcp_sk(sk);
1221 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1222 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1225 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1228 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1229 } else if (num_sacks > 1) {
1230 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1231 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1233 if (!after(end_seq_0, end_seq_1) &&
1234 !before(start_seq_0, start_seq_1)) {
1237 NET_INC_STATS_BH(sock_net(sk),
1238 LINUX_MIB_TCPDSACKOFORECV);
1242 /* D-SACK for already forgotten data... Do dumb counting. */
1243 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1244 !after(end_seq_0, prior_snd_una) &&
1245 after(end_seq_0, tp->undo_marker))
1251 struct tcp_sacktag_state {
1257 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1258 * the incoming SACK may not exactly match but we can find smaller MSS
1259 * aligned portion of it that matches. Therefore we might need to fragment
1260 * which may fail and creates some hassle (caller must handle error case
1263 * FIXME: this could be merged to shift decision code
1265 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1266 u32 start_seq, u32 end_seq)
1269 unsigned int pkt_len;
1272 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1273 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1275 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1276 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1277 mss = tcp_skb_mss(skb);
1278 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1281 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1285 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1290 /* Round if necessary so that SACKs cover only full MSSes
1291 * and/or the remaining small portion (if present)
1293 if (pkt_len > mss) {
1294 unsigned int new_len = (pkt_len / mss) * mss;
1295 if (!in_sack && new_len < pkt_len) {
1297 if (new_len > skb->len)
1302 err = tcp_fragment(sk, skb, pkt_len, mss);
1310 static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1311 struct tcp_sacktag_state *state,
1312 int dup_sack, int pcount)
1314 struct tcp_sock *tp = tcp_sk(sk);
1315 u8 sacked = TCP_SKB_CB(skb)->sacked;
1316 int fack_count = state->fack_count;
1318 /* Account D-SACK for retransmitted packet. */
1319 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1320 if (tp->undo_marker && tp->undo_retrans &&
1321 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1323 if (sacked & TCPCB_SACKED_ACKED)
1324 state->reord = min(fack_count, state->reord);
1327 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1328 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1331 if (!(sacked & TCPCB_SACKED_ACKED)) {
1332 if (sacked & TCPCB_SACKED_RETRANS) {
1333 /* If the segment is not tagged as lost,
1334 * we do not clear RETRANS, believing
1335 * that retransmission is still in flight.
1337 if (sacked & TCPCB_LOST) {
1338 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1339 tp->lost_out -= pcount;
1340 tp->retrans_out -= pcount;
1343 if (!(sacked & TCPCB_RETRANS)) {
1344 /* New sack for not retransmitted frame,
1345 * which was in hole. It is reordering.
1347 if (before(TCP_SKB_CB(skb)->seq,
1348 tcp_highest_sack_seq(tp)))
1349 state->reord = min(fack_count,
1352 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1353 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1354 state->flag |= FLAG_ONLY_ORIG_SACKED;
1357 if (sacked & TCPCB_LOST) {
1358 sacked &= ~TCPCB_LOST;
1359 tp->lost_out -= pcount;
1363 sacked |= TCPCB_SACKED_ACKED;
1364 state->flag |= FLAG_DATA_SACKED;
1365 tp->sacked_out += pcount;
1367 fack_count += pcount;
1369 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1370 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1371 before(TCP_SKB_CB(skb)->seq,
1372 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1373 tp->lost_cnt_hint += pcount;
1375 if (fack_count > tp->fackets_out)
1376 tp->fackets_out = fack_count;
1379 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1380 * frames and clear it. undo_retrans is decreased above, L|R frames
1381 * are accounted above as well.
1383 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1384 sacked &= ~TCPCB_SACKED_RETRANS;
1385 tp->retrans_out -= pcount;
1391 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1392 struct tcp_sacktag_state *state,
1393 unsigned int pcount, int shifted, int mss,
1396 struct tcp_sock *tp = tcp_sk(sk);
1397 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1401 /* Tweak before seqno plays */
1402 if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1403 !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1404 tp->lost_cnt_hint += pcount;
1406 TCP_SKB_CB(prev)->end_seq += shifted;
1407 TCP_SKB_CB(skb)->seq += shifted;
1409 skb_shinfo(prev)->gso_segs += pcount;
1410 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1411 skb_shinfo(skb)->gso_segs -= pcount;
1413 /* When we're adding to gso_segs == 1, gso_size will be zero,
1414 * in theory this shouldn't be necessary but as long as DSACK
1415 * code can come after this skb later on it's better to keep
1416 * setting gso_size to something.
1418 if (!skb_shinfo(prev)->gso_size) {
1419 skb_shinfo(prev)->gso_size = mss;
1420 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1423 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1424 if (skb_shinfo(skb)->gso_segs <= 1) {
1425 skb_shinfo(skb)->gso_size = 0;
1426 skb_shinfo(skb)->gso_type = 0;
1429 /* We discard results */
1430 tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
1432 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1433 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1436 BUG_ON(!tcp_skb_pcount(skb));
1437 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1441 /* Whole SKB was eaten :-) */
1443 if (skb == tp->retransmit_skb_hint)
1444 tp->retransmit_skb_hint = prev;
1445 if (skb == tp->scoreboard_skb_hint)
1446 tp->scoreboard_skb_hint = prev;
1447 if (skb == tp->lost_skb_hint) {
1448 tp->lost_skb_hint = prev;
1449 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1452 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1453 if (skb == tcp_highest_sack(sk))
1454 tcp_advance_highest_sack(sk, skb);
1456 tcp_unlink_write_queue(skb, sk);
1457 sk_wmem_free_skb(sk, skb);
1459 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1464 /* I wish gso_size would have a bit more sane initialization than
1465 * something-or-zero which complicates things
1467 static int tcp_skb_seglen(struct sk_buff *skb)
1469 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1472 /* Shifting pages past head area doesn't work */
1473 static int skb_can_shift(struct sk_buff *skb)
1475 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1478 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1481 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1482 struct tcp_sacktag_state *state,
1483 u32 start_seq, u32 end_seq,
1486 struct tcp_sock *tp = tcp_sk(sk);
1487 struct sk_buff *prev;
1493 if (!sk_can_gso(sk))
1496 /* Normally R but no L won't result in plain S */
1498 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1500 if (!skb_can_shift(skb))
1502 /* This frame is about to be dropped (was ACKed). */
1503 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1506 /* Can only happen with delayed DSACK + discard craziness */
1507 if (unlikely(skb == tcp_write_queue_head(sk)))
1509 prev = tcp_write_queue_prev(sk, skb);
1511 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1514 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1515 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1519 pcount = tcp_skb_pcount(skb);
1520 mss = tcp_skb_seglen(skb);
1522 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1523 * drop this restriction as unnecessary
1525 if (mss != tcp_skb_seglen(prev))
1528 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1530 /* CHECKME: This is non-MSS split case only?, this will
1531 * cause skipped skbs due to advancing loop btw, original
1532 * has that feature too
1534 if (tcp_skb_pcount(skb) <= 1)
1537 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1539 /* TODO: head merge to next could be attempted here
1540 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1541 * though it might not be worth of the additional hassle
1543 * ...we can probably just fallback to what was done
1544 * previously. We could try merging non-SACKed ones
1545 * as well but it probably isn't going to buy off
1546 * because later SACKs might again split them, and
1547 * it would make skb timestamp tracking considerably
1553 len = end_seq - TCP_SKB_CB(skb)->seq;
1555 BUG_ON(len > skb->len);
1557 /* MSS boundaries should be honoured or else pcount will
1558 * severely break even though it makes things bit trickier.
1559 * Optimize common case to avoid most of the divides
1561 mss = tcp_skb_mss(skb);
1563 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1564 * drop this restriction as unnecessary
1566 if (mss != tcp_skb_seglen(prev))
1571 } else if (len < mss) {
1579 if (!skb_shift(prev, skb, len))
1581 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1584 /* Hole filled allows collapsing with the next as well, this is very
1585 * useful when hole on every nth skb pattern happens
1587 if (prev == tcp_write_queue_tail(sk))
1589 skb = tcp_write_queue_next(sk, prev);
1591 if (!skb_can_shift(skb) ||
1592 (skb == tcp_send_head(sk)) ||
1593 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1594 (mss != tcp_skb_seglen(skb)))
1598 if (skb_shift(prev, skb, len)) {
1599 pcount += tcp_skb_pcount(skb);
1600 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1604 state->fack_count += pcount;
1611 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1615 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1616 struct tcp_sack_block *next_dup,
1617 struct tcp_sacktag_state *state,
1618 u32 start_seq, u32 end_seq,
1621 struct tcp_sock *tp = tcp_sk(sk);
1622 struct sk_buff *tmp;
1624 tcp_for_write_queue_from(skb, sk) {
1626 int dup_sack = dup_sack_in;
1628 if (skb == tcp_send_head(sk))
1631 /* queue is in-order => we can short-circuit the walk early */
1632 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1635 if ((next_dup != NULL) &&
1636 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1637 in_sack = tcp_match_skb_to_sack(sk, skb,
1638 next_dup->start_seq,
1644 /* skb reference here is a bit tricky to get right, since
1645 * shifting can eat and free both this skb and the next,
1646 * so not even _safe variant of the loop is enough.
1649 tmp = tcp_shift_skb_data(sk, skb, state,
1650 start_seq, end_seq, dup_sack);
1659 in_sack = tcp_match_skb_to_sack(sk, skb,
1665 if (unlikely(in_sack < 0))
1669 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1672 tcp_skb_pcount(skb));
1674 if (!before(TCP_SKB_CB(skb)->seq,
1675 tcp_highest_sack_seq(tp)))
1676 tcp_advance_highest_sack(sk, skb);
1679 state->fack_count += tcp_skb_pcount(skb);
1684 /* Avoid all extra work that is being done by sacktag while walking in
1687 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1688 struct tcp_sacktag_state *state,
1691 tcp_for_write_queue_from(skb, sk) {
1692 if (skb == tcp_send_head(sk))
1695 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1698 state->fack_count += tcp_skb_pcount(skb);
1703 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1705 struct tcp_sack_block *next_dup,
1706 struct tcp_sacktag_state *state,
1709 if (next_dup == NULL)
1712 if (before(next_dup->start_seq, skip_to_seq)) {
1713 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1714 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1715 next_dup->start_seq, next_dup->end_seq,
1722 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1724 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1728 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1731 const struct inet_connection_sock *icsk = inet_csk(sk);
1732 struct tcp_sock *tp = tcp_sk(sk);
1733 unsigned char *ptr = (skb_transport_header(ack_skb) +
1734 TCP_SKB_CB(ack_skb)->sacked);
1735 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1736 struct tcp_sack_block sp[TCP_NUM_SACKS];
1737 struct tcp_sack_block *cache;
1738 struct tcp_sacktag_state state;
1739 struct sk_buff *skb;
1740 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1742 int found_dup_sack = 0;
1744 int first_sack_index;
1747 state.reord = tp->packets_out;
1749 if (!tp->sacked_out) {
1750 if (WARN_ON(tp->fackets_out))
1751 tp->fackets_out = 0;
1752 tcp_highest_sack_reset(sk);
1755 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1756 num_sacks, prior_snd_una);
1758 state.flag |= FLAG_DSACKING_ACK;
1760 /* Eliminate too old ACKs, but take into
1761 * account more or less fresh ones, they can
1762 * contain valid SACK info.
1764 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1767 if (!tp->packets_out)
1771 first_sack_index = 0;
1772 for (i = 0; i < num_sacks; i++) {
1773 int dup_sack = !i && found_dup_sack;
1775 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1776 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1778 if (!tcp_is_sackblock_valid(tp, dup_sack,
1779 sp[used_sacks].start_seq,
1780 sp[used_sacks].end_seq)) {
1784 if (!tp->undo_marker)
1785 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1787 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1789 /* Don't count olds caused by ACK reordering */
1790 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1791 !after(sp[used_sacks].end_seq, tp->snd_una))
1793 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1796 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1798 first_sack_index = -1;
1802 /* Ignore very old stuff early */
1803 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1809 /* order SACK blocks to allow in order walk of the retrans queue */
1810 for (i = used_sacks - 1; i > 0; i--) {
1811 for (j = 0; j < i; j++) {
1812 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1813 swap(sp[j], sp[j + 1]);
1815 /* Track where the first SACK block goes to */
1816 if (j == first_sack_index)
1817 first_sack_index = j + 1;
1822 skb = tcp_write_queue_head(sk);
1823 state.fack_count = 0;
1826 if (!tp->sacked_out) {
1827 /* It's already past, so skip checking against it */
1828 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1830 cache = tp->recv_sack_cache;
1831 /* Skip empty blocks in at head of the cache */
1832 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1837 while (i < used_sacks) {
1838 u32 start_seq = sp[i].start_seq;
1839 u32 end_seq = sp[i].end_seq;
1840 int dup_sack = (found_dup_sack && (i == first_sack_index));
1841 struct tcp_sack_block *next_dup = NULL;
1843 if (found_dup_sack && ((i + 1) == first_sack_index))
1844 next_dup = &sp[i + 1];
1846 /* Event "B" in the comment above. */
1847 if (after(end_seq, tp->high_seq))
1848 state.flag |= FLAG_DATA_LOST;
1850 /* Skip too early cached blocks */
1851 while (tcp_sack_cache_ok(tp, cache) &&
1852 !before(start_seq, cache->end_seq))
1855 /* Can skip some work by looking recv_sack_cache? */
1856 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1857 after(end_seq, cache->start_seq)) {
1860 if (before(start_seq, cache->start_seq)) {
1861 skb = tcp_sacktag_skip(skb, sk, &state,
1863 skb = tcp_sacktag_walk(skb, sk, next_dup,
1870 /* Rest of the block already fully processed? */
1871 if (!after(end_seq, cache->end_seq))
1874 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1878 /* ...tail remains todo... */
1879 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1880 /* ...but better entrypoint exists! */
1881 skb = tcp_highest_sack(sk);
1884 state.fack_count = tp->fackets_out;
1889 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1890 /* Check overlap against next cached too (past this one already) */
1895 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1896 skb = tcp_highest_sack(sk);
1899 state.fack_count = tp->fackets_out;
1901 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1904 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1905 start_seq, end_seq, dup_sack);
1908 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1909 * due to in-order walk
1911 if (after(end_seq, tp->frto_highmark))
1912 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1917 /* Clear the head of the cache sack blocks so we can skip it next time */
1918 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1919 tp->recv_sack_cache[i].start_seq = 0;
1920 tp->recv_sack_cache[i].end_seq = 0;
1922 for (j = 0; j < used_sacks; j++)
1923 tp->recv_sack_cache[i++] = sp[j];
1925 tcp_mark_lost_retrans(sk);
1927 tcp_verify_left_out(tp);
1929 if ((state.reord < tp->fackets_out) &&
1930 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1931 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1932 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1936 #if FASTRETRANS_DEBUG > 0
1937 WARN_ON((int)tp->sacked_out < 0);
1938 WARN_ON((int)tp->lost_out < 0);
1939 WARN_ON((int)tp->retrans_out < 0);
1940 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1945 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1946 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1948 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1952 holes = max(tp->lost_out, 1U);
1953 holes = min(holes, tp->packets_out);
1955 if ((tp->sacked_out + holes) > tp->packets_out) {
1956 tp->sacked_out = tp->packets_out - holes;
1962 /* If we receive more dupacks than we expected counting segments
1963 * in assumption of absent reordering, interpret this as reordering.
1964 * The only another reason could be bug in receiver TCP.
1966 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1968 struct tcp_sock *tp = tcp_sk(sk);
1969 if (tcp_limit_reno_sacked(tp))
1970 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1973 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1975 static void tcp_add_reno_sack(struct sock *sk)
1977 struct tcp_sock *tp = tcp_sk(sk);
1979 tcp_check_reno_reordering(sk, 0);
1980 tcp_verify_left_out(tp);
1983 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1985 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1987 struct tcp_sock *tp = tcp_sk(sk);
1990 /* One ACK acked hole. The rest eat duplicate ACKs. */
1991 if (acked - 1 >= tp->sacked_out)
1994 tp->sacked_out -= acked - 1;
1996 tcp_check_reno_reordering(sk, acked);
1997 tcp_verify_left_out(tp);
2000 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2005 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2007 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2010 /* F-RTO can only be used if TCP has never retransmitted anything other than
2011 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2013 int tcp_use_frto(struct sock *sk)
2015 const struct tcp_sock *tp = tcp_sk(sk);
2016 const struct inet_connection_sock *icsk = inet_csk(sk);
2017 struct sk_buff *skb;
2019 if (!sysctl_tcp_frto)
2022 /* MTU probe and F-RTO won't really play nicely along currently */
2023 if (icsk->icsk_mtup.probe_size)
2026 if (tcp_is_sackfrto(tp))
2029 /* Avoid expensive walking of rexmit queue if possible */
2030 if (tp->retrans_out > 1)
2033 skb = tcp_write_queue_head(sk);
2034 if (tcp_skb_is_last(sk, skb))
2036 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2037 tcp_for_write_queue_from(skb, sk) {
2038 if (skb == tcp_send_head(sk))
2040 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2042 /* Short-circuit when first non-SACKed skb has been checked */
2043 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2049 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2050 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2051 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2052 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2053 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2054 * bits are handled if the Loss state is really to be entered (in
2055 * tcp_enter_frto_loss).
2057 * Do like tcp_enter_loss() would; when RTO expires the second time it
2059 * "Reduce ssthresh if it has not yet been made inside this window."
2061 void tcp_enter_frto(struct sock *sk)
2063 const struct inet_connection_sock *icsk = inet_csk(sk);
2064 struct tcp_sock *tp = tcp_sk(sk);
2065 struct sk_buff *skb;
2067 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2068 tp->snd_una == tp->high_seq ||
2069 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2070 !icsk->icsk_retransmits)) {
2071 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2072 /* Our state is too optimistic in ssthresh() call because cwnd
2073 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2074 * recovery has not yet completed. Pattern would be this: RTO,
2075 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2077 * RFC4138 should be more specific on what to do, even though
2078 * RTO is quite unlikely to occur after the first Cumulative ACK
2079 * due to back-off and complexity of triggering events ...
2081 if (tp->frto_counter) {
2083 stored_cwnd = tp->snd_cwnd;
2085 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2086 tp->snd_cwnd = stored_cwnd;
2088 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2090 /* ... in theory, cong.control module could do "any tricks" in
2091 * ssthresh(), which means that ca_state, lost bits and lost_out
2092 * counter would have to be faked before the call occurs. We
2093 * consider that too expensive, unlikely and hacky, so modules
2094 * using these in ssthresh() must deal these incompatibility
2095 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2097 tcp_ca_event(sk, CA_EVENT_FRTO);
2100 tp->undo_marker = tp->snd_una;
2101 tp->undo_retrans = 0;
2103 skb = tcp_write_queue_head(sk);
2104 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2105 tp->undo_marker = 0;
2106 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2107 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2108 tp->retrans_out -= tcp_skb_pcount(skb);
2110 tcp_verify_left_out(tp);
2112 /* Too bad if TCP was application limited */
2113 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2115 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2116 * The last condition is necessary at least in tp->frto_counter case.
2118 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2119 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2120 after(tp->high_seq, tp->snd_una)) {
2121 tp->frto_highmark = tp->high_seq;
2123 tp->frto_highmark = tp->snd_nxt;
2125 tcp_set_ca_state(sk, TCP_CA_Disorder);
2126 tp->high_seq = tp->snd_nxt;
2127 tp->frto_counter = 1;
2130 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2131 * which indicates that we should follow the traditional RTO recovery,
2132 * i.e. mark everything lost and do go-back-N retransmission.
2134 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2136 struct tcp_sock *tp = tcp_sk(sk);
2137 struct sk_buff *skb;
2140 tp->retrans_out = 0;
2141 if (tcp_is_reno(tp))
2142 tcp_reset_reno_sack(tp);
2144 tcp_for_write_queue(skb, sk) {
2145 if (skb == tcp_send_head(sk))
2148 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2150 * Count the retransmission made on RTO correctly (only when
2151 * waiting for the first ACK and did not get it)...
2153 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2154 /* For some reason this R-bit might get cleared? */
2155 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2156 tp->retrans_out += tcp_skb_pcount(skb);
2157 /* ...enter this if branch just for the first segment */
2158 flag |= FLAG_DATA_ACKED;
2160 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2161 tp->undo_marker = 0;
2162 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2165 /* Marking forward transmissions that were made after RTO lost
2166 * can cause unnecessary retransmissions in some scenarios,
2167 * SACK blocks will mitigate that in some but not in all cases.
2168 * We used to not mark them but it was causing break-ups with
2169 * receivers that do only in-order receival.
2171 * TODO: we could detect presence of such receiver and select
2172 * different behavior per flow.
2174 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2175 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2176 tp->lost_out += tcp_skb_pcount(skb);
2177 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2180 tcp_verify_left_out(tp);
2182 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2183 tp->snd_cwnd_cnt = 0;
2184 tp->snd_cwnd_stamp = tcp_time_stamp;
2185 tp->frto_counter = 0;
2186 tp->bytes_acked = 0;
2188 tp->reordering = min_t(unsigned int, tp->reordering,
2189 sysctl_tcp_reordering);
2190 tcp_set_ca_state(sk, TCP_CA_Loss);
2191 tp->high_seq = tp->snd_nxt;
2192 TCP_ECN_queue_cwr(tp);
2194 tcp_clear_all_retrans_hints(tp);
2197 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2199 tp->retrans_out = 0;
2202 tp->undo_marker = 0;
2203 tp->undo_retrans = 0;
2206 void tcp_clear_retrans(struct tcp_sock *tp)
2208 tcp_clear_retrans_partial(tp);
2210 tp->fackets_out = 0;
2214 /* Enter Loss state. If "how" is not zero, forget all SACK information
2215 * and reset tags completely, otherwise preserve SACKs. If receiver
2216 * dropped its ofo queue, we will know this due to reneging detection.
2218 void tcp_enter_loss(struct sock *sk, int how)
2220 const struct inet_connection_sock *icsk = inet_csk(sk);
2221 struct tcp_sock *tp = tcp_sk(sk);
2222 struct sk_buff *skb;
2224 /* Reduce ssthresh if it has not yet been made inside this window. */
2225 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2226 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2227 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2228 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2229 tcp_ca_event(sk, CA_EVENT_LOSS);
2232 tp->snd_cwnd_cnt = 0;
2233 tp->snd_cwnd_stamp = tcp_time_stamp;
2235 tp->bytes_acked = 0;
2236 tcp_clear_retrans_partial(tp);
2238 if (tcp_is_reno(tp))
2239 tcp_reset_reno_sack(tp);
2242 /* Push undo marker, if it was plain RTO and nothing
2243 * was retransmitted. */
2244 tp->undo_marker = tp->snd_una;
2247 tp->fackets_out = 0;
2249 tcp_clear_all_retrans_hints(tp);
2251 tcp_for_write_queue(skb, sk) {
2252 if (skb == tcp_send_head(sk))
2255 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2256 tp->undo_marker = 0;
2257 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2258 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2259 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2260 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2261 tp->lost_out += tcp_skb_pcount(skb);
2262 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2265 tcp_verify_left_out(tp);
2267 tp->reordering = min_t(unsigned int, tp->reordering,
2268 sysctl_tcp_reordering);
2269 tcp_set_ca_state(sk, TCP_CA_Loss);
2270 tp->high_seq = tp->snd_nxt;
2271 TCP_ECN_queue_cwr(tp);
2272 /* Abort F-RTO algorithm if one is in progress */
2273 tp->frto_counter = 0;
2276 /* If ACK arrived pointing to a remembered SACK, it means that our
2277 * remembered SACKs do not reflect real state of receiver i.e.
2278 * receiver _host_ is heavily congested (or buggy).
2280 * Do processing similar to RTO timeout.
2282 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2284 if (flag & FLAG_SACK_RENEGING) {
2285 struct inet_connection_sock *icsk = inet_csk(sk);
2286 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2288 tcp_enter_loss(sk, 1);
2289 icsk->icsk_retransmits++;
2290 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2291 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2292 icsk->icsk_rto, TCP_RTO_MAX);
2298 static inline int tcp_fackets_out(struct tcp_sock *tp)
2300 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2303 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2304 * counter when SACK is enabled (without SACK, sacked_out is used for
2307 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2308 * segments up to the highest received SACK block so far and holes in
2311 * With reordering, holes may still be in flight, so RFC3517 recovery
2312 * uses pure sacked_out (total number of SACKed segments) even though
2313 * it violates the RFC that uses duplicate ACKs, often these are equal
2314 * but when e.g. out-of-window ACKs or packet duplication occurs,
2315 * they differ. Since neither occurs due to loss, TCP should really
2318 static inline int tcp_dupack_heuristics(struct tcp_sock *tp)
2320 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2323 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
2325 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2328 static inline int tcp_head_timedout(struct sock *sk)
2330 struct tcp_sock *tp = tcp_sk(sk);
2332 return tp->packets_out &&
2333 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2336 /* Linux NewReno/SACK/FACK/ECN state machine.
2337 * --------------------------------------
2339 * "Open" Normal state, no dubious events, fast path.
2340 * "Disorder" In all the respects it is "Open",
2341 * but requires a bit more attention. It is entered when
2342 * we see some SACKs or dupacks. It is split of "Open"
2343 * mainly to move some processing from fast path to slow one.
2344 * "CWR" CWND was reduced due to some Congestion Notification event.
2345 * It can be ECN, ICMP source quench, local device congestion.
2346 * "Recovery" CWND was reduced, we are fast-retransmitting.
2347 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2349 * tcp_fastretrans_alert() is entered:
2350 * - each incoming ACK, if state is not "Open"
2351 * - when arrived ACK is unusual, namely:
2356 * Counting packets in flight is pretty simple.
2358 * in_flight = packets_out - left_out + retrans_out
2360 * packets_out is SND.NXT-SND.UNA counted in packets.
2362 * retrans_out is number of retransmitted segments.
2364 * left_out is number of segments left network, but not ACKed yet.
2366 * left_out = sacked_out + lost_out
2368 * sacked_out: Packets, which arrived to receiver out of order
2369 * and hence not ACKed. With SACKs this number is simply
2370 * amount of SACKed data. Even without SACKs
2371 * it is easy to give pretty reliable estimate of this number,
2372 * counting duplicate ACKs.
2374 * lost_out: Packets lost by network. TCP has no explicit
2375 * "loss notification" feedback from network (for now).
2376 * It means that this number can be only _guessed_.
2377 * Actually, it is the heuristics to predict lossage that
2378 * distinguishes different algorithms.
2380 * F.e. after RTO, when all the queue is considered as lost,
2381 * lost_out = packets_out and in_flight = retrans_out.
2383 * Essentially, we have now two algorithms counting
2386 * FACK: It is the simplest heuristics. As soon as we decided
2387 * that something is lost, we decide that _all_ not SACKed
2388 * packets until the most forward SACK are lost. I.e.
2389 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2390 * It is absolutely correct estimate, if network does not reorder
2391 * packets. And it loses any connection to reality when reordering
2392 * takes place. We use FACK by default until reordering
2393 * is suspected on the path to this destination.
2395 * NewReno: when Recovery is entered, we assume that one segment
2396 * is lost (classic Reno). While we are in Recovery and
2397 * a partial ACK arrives, we assume that one more packet
2398 * is lost (NewReno). This heuristics are the same in NewReno
2401 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2402 * deflation etc. CWND is real congestion window, never inflated, changes
2403 * only according to classic VJ rules.
2405 * Really tricky (and requiring careful tuning) part of algorithm
2406 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2407 * The first determines the moment _when_ we should reduce CWND and,
2408 * hence, slow down forward transmission. In fact, it determines the moment
2409 * when we decide that hole is caused by loss, rather than by a reorder.
2411 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2412 * holes, caused by lost packets.
2414 * And the most logically complicated part of algorithm is undo
2415 * heuristics. We detect false retransmits due to both too early
2416 * fast retransmit (reordering) and underestimated RTO, analyzing
2417 * timestamps and D-SACKs. When we detect that some segments were
2418 * retransmitted by mistake and CWND reduction was wrong, we undo
2419 * window reduction and abort recovery phase. This logic is hidden
2420 * inside several functions named tcp_try_undo_<something>.
2423 /* This function decides, when we should leave Disordered state
2424 * and enter Recovery phase, reducing congestion window.
2426 * Main question: may we further continue forward transmission
2427 * with the same cwnd?
2429 static int tcp_time_to_recover(struct sock *sk)
2431 struct tcp_sock *tp = tcp_sk(sk);
2434 /* Do not perform any recovery during F-RTO algorithm */
2435 if (tp->frto_counter)
2438 /* Trick#1: The loss is proven. */
2442 /* Not-A-Trick#2 : Classic rule... */
2443 if (tcp_dupack_heuristics(tp) > tp->reordering)
2446 /* Trick#3 : when we use RFC2988 timer restart, fast
2447 * retransmit can be triggered by timeout of queue head.
2449 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2452 /* Trick#4: It is still not OK... But will it be useful to delay
2455 packets_out = tp->packets_out;
2456 if (packets_out <= tp->reordering &&
2457 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2458 !tcp_may_send_now(sk)) {
2459 /* We have nothing to send. This connection is limited
2460 * either by receiver window or by application.
2465 /* If a thin stream is detected, retransmit after first
2466 * received dupack. Employ only if SACK is supported in order
2467 * to avoid possible corner-case series of spurious retransmissions
2468 * Use only if there are no unsent data.
2470 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2471 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2472 tcp_is_sack(tp) && !tcp_send_head(sk))
2478 /* New heuristics: it is possible only after we switched to restart timer
2479 * each time when something is ACKed. Hence, we can detect timed out packets
2480 * during fast retransmit without falling to slow start.
2482 * Usefulness of this as is very questionable, since we should know which of
2483 * the segments is the next to timeout which is relatively expensive to find
2484 * in general case unless we add some data structure just for that. The
2485 * current approach certainly won't find the right one too often and when it
2486 * finally does find _something_ it usually marks large part of the window
2487 * right away (because a retransmission with a larger timestamp blocks the
2488 * loop from advancing). -ij
2490 static void tcp_timeout_skbs(struct sock *sk)
2492 struct tcp_sock *tp = tcp_sk(sk);
2493 struct sk_buff *skb;
2495 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2498 skb = tp->scoreboard_skb_hint;
2499 if (tp->scoreboard_skb_hint == NULL)
2500 skb = tcp_write_queue_head(sk);
2502 tcp_for_write_queue_from(skb, sk) {
2503 if (skb == tcp_send_head(sk))
2505 if (!tcp_skb_timedout(sk, skb))
2508 tcp_skb_mark_lost(tp, skb);
2511 tp->scoreboard_skb_hint = skb;
2513 tcp_verify_left_out(tp);
2516 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2517 * is against sacked "cnt", otherwise it's against facked "cnt"
2519 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2521 struct tcp_sock *tp = tcp_sk(sk);
2522 struct sk_buff *skb;
2527 WARN_ON(packets > tp->packets_out);
2528 if (tp->lost_skb_hint) {
2529 skb = tp->lost_skb_hint;
2530 cnt = tp->lost_cnt_hint;
2531 /* Head already handled? */
2532 if (mark_head && skb != tcp_write_queue_head(sk))
2535 skb = tcp_write_queue_head(sk);
2539 tcp_for_write_queue_from(skb, sk) {
2540 if (skb == tcp_send_head(sk))
2542 /* TODO: do this better */
2543 /* this is not the most efficient way to do this... */
2544 tp->lost_skb_hint = skb;
2545 tp->lost_cnt_hint = cnt;
2547 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2551 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2552 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2553 cnt += tcp_skb_pcount(skb);
2555 if (cnt > packets) {
2556 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2557 (oldcnt >= packets))
2560 mss = skb_shinfo(skb)->gso_size;
2561 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2567 tcp_skb_mark_lost(tp, skb);
2572 tcp_verify_left_out(tp);
2575 /* Account newly detected lost packet(s) */
2577 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2579 struct tcp_sock *tp = tcp_sk(sk);
2581 if (tcp_is_reno(tp)) {
2582 tcp_mark_head_lost(sk, 1, 1);
2583 } else if (tcp_is_fack(tp)) {
2584 int lost = tp->fackets_out - tp->reordering;
2587 tcp_mark_head_lost(sk, lost, 0);
2589 int sacked_upto = tp->sacked_out - tp->reordering;
2590 if (sacked_upto >= 0)
2591 tcp_mark_head_lost(sk, sacked_upto, 0);
2592 else if (fast_rexmit)
2593 tcp_mark_head_lost(sk, 1, 1);
2596 tcp_timeout_skbs(sk);
2599 /* CWND moderation, preventing bursts due to too big ACKs
2600 * in dubious situations.
2602 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2604 tp->snd_cwnd = min(tp->snd_cwnd,
2605 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2606 tp->snd_cwnd_stamp = tcp_time_stamp;
2609 /* Lower bound on congestion window is slow start threshold
2610 * unless congestion avoidance choice decides to overide it.
2612 static inline u32 tcp_cwnd_min(const struct sock *sk)
2614 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2616 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2619 /* Decrease cwnd each second ack. */
2620 static void tcp_cwnd_down(struct sock *sk, int flag)
2622 struct tcp_sock *tp = tcp_sk(sk);
2623 int decr = tp->snd_cwnd_cnt + 1;
2625 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2626 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2627 tp->snd_cwnd_cnt = decr & 1;
2630 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2631 tp->snd_cwnd -= decr;
2633 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2634 tp->snd_cwnd_stamp = tcp_time_stamp;
2638 /* Nothing was retransmitted or returned timestamp is less
2639 * than timestamp of the first retransmission.
2641 static inline int tcp_packet_delayed(struct tcp_sock *tp)
2643 return !tp->retrans_stamp ||
2644 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2645 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2648 /* Undo procedures. */
2650 #if FASTRETRANS_DEBUG > 1
2651 static void DBGUNDO(struct sock *sk, const char *msg)
2653 struct tcp_sock *tp = tcp_sk(sk);
2654 struct inet_sock *inet = inet_sk(sk);
2656 if (sk->sk_family == AF_INET) {
2657 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2659 &inet->inet_daddr, ntohs(inet->inet_dport),
2660 tp->snd_cwnd, tcp_left_out(tp),
2661 tp->snd_ssthresh, tp->prior_ssthresh,
2664 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2665 else if (sk->sk_family == AF_INET6) {
2666 struct ipv6_pinfo *np = inet6_sk(sk);
2667 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2669 &np->daddr, ntohs(inet->inet_dport),
2670 tp->snd_cwnd, tcp_left_out(tp),
2671 tp->snd_ssthresh, tp->prior_ssthresh,
2677 #define DBGUNDO(x...) do { } while (0)
2680 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2682 struct tcp_sock *tp = tcp_sk(sk);
2684 if (tp->prior_ssthresh) {
2685 const struct inet_connection_sock *icsk = inet_csk(sk);
2687 if (icsk->icsk_ca_ops->undo_cwnd)
2688 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2690 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2692 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2693 tp->snd_ssthresh = tp->prior_ssthresh;
2694 TCP_ECN_withdraw_cwr(tp);
2697 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2699 tp->snd_cwnd_stamp = tcp_time_stamp;
2702 static inline int tcp_may_undo(struct tcp_sock *tp)
2704 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2707 /* People celebrate: "We love our President!" */
2708 static int tcp_try_undo_recovery(struct sock *sk)
2710 struct tcp_sock *tp = tcp_sk(sk);
2712 if (tcp_may_undo(tp)) {
2715 /* Happy end! We did not retransmit anything
2716 * or our original transmission succeeded.
2718 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2719 tcp_undo_cwr(sk, true);
2720 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2721 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2723 mib_idx = LINUX_MIB_TCPFULLUNDO;
2725 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2726 tp->undo_marker = 0;
2728 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2729 /* Hold old state until something *above* high_seq
2730 * is ACKed. For Reno it is MUST to prevent false
2731 * fast retransmits (RFC2582). SACK TCP is safe. */
2732 tcp_moderate_cwnd(tp);
2735 tcp_set_ca_state(sk, TCP_CA_Open);
2739 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2740 static void tcp_try_undo_dsack(struct sock *sk)
2742 struct tcp_sock *tp = tcp_sk(sk);
2744 if (tp->undo_marker && !tp->undo_retrans) {
2745 DBGUNDO(sk, "D-SACK");
2746 tcp_undo_cwr(sk, true);
2747 tp->undo_marker = 0;
2748 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2752 /* We can clear retrans_stamp when there are no retransmissions in the
2753 * window. It would seem that it is trivially available for us in
2754 * tp->retrans_out, however, that kind of assumptions doesn't consider
2755 * what will happen if errors occur when sending retransmission for the
2756 * second time. ...It could the that such segment has only
2757 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2758 * the head skb is enough except for some reneging corner cases that
2759 * are not worth the effort.
2761 * Main reason for all this complexity is the fact that connection dying
2762 * time now depends on the validity of the retrans_stamp, in particular,
2763 * that successive retransmissions of a segment must not advance
2764 * retrans_stamp under any conditions.
2766 static int tcp_any_retrans_done(struct sock *sk)
2768 struct tcp_sock *tp = tcp_sk(sk);
2769 struct sk_buff *skb;
2771 if (tp->retrans_out)
2774 skb = tcp_write_queue_head(sk);
2775 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2781 /* Undo during fast recovery after partial ACK. */
2783 static int tcp_try_undo_partial(struct sock *sk, int acked)
2785 struct tcp_sock *tp = tcp_sk(sk);
2786 /* Partial ACK arrived. Force Hoe's retransmit. */
2787 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2789 if (tcp_may_undo(tp)) {
2790 /* Plain luck! Hole if filled with delayed
2791 * packet, rather than with a retransmit.
2793 if (!tcp_any_retrans_done(sk))
2794 tp->retrans_stamp = 0;
2796 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2799 tcp_undo_cwr(sk, false);
2800 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2802 /* So... Do not make Hoe's retransmit yet.
2803 * If the first packet was delayed, the rest
2804 * ones are most probably delayed as well.
2811 /* Undo during loss recovery after partial ACK. */
2812 static int tcp_try_undo_loss(struct sock *sk)
2814 struct tcp_sock *tp = tcp_sk(sk);
2816 if (tcp_may_undo(tp)) {
2817 struct sk_buff *skb;
2818 tcp_for_write_queue(skb, sk) {
2819 if (skb == tcp_send_head(sk))
2821 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2824 tcp_clear_all_retrans_hints(tp);
2826 DBGUNDO(sk, "partial loss");
2828 tcp_undo_cwr(sk, true);
2829 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2830 inet_csk(sk)->icsk_retransmits = 0;
2831 tp->undo_marker = 0;
2832 if (tcp_is_sack(tp))
2833 tcp_set_ca_state(sk, TCP_CA_Open);
2839 static inline void tcp_complete_cwr(struct sock *sk)
2841 struct tcp_sock *tp = tcp_sk(sk);
2843 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2844 if (tp->undo_marker) {
2845 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR)
2846 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2848 tp->snd_cwnd = tp->snd_ssthresh;
2849 tp->snd_cwnd_stamp = tcp_time_stamp;
2851 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2854 static void tcp_try_keep_open(struct sock *sk)
2856 struct tcp_sock *tp = tcp_sk(sk);
2857 int state = TCP_CA_Open;
2859 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2860 state = TCP_CA_Disorder;
2862 if (inet_csk(sk)->icsk_ca_state != state) {
2863 tcp_set_ca_state(sk, state);
2864 tp->high_seq = tp->snd_nxt;
2868 static void tcp_try_to_open(struct sock *sk, int flag)
2870 struct tcp_sock *tp = tcp_sk(sk);
2872 tcp_verify_left_out(tp);
2874 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2875 tp->retrans_stamp = 0;
2877 if (flag & FLAG_ECE)
2878 tcp_enter_cwr(sk, 1);
2880 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2881 tcp_try_keep_open(sk);
2882 tcp_moderate_cwnd(tp);
2884 tcp_cwnd_down(sk, flag);
2888 static void tcp_mtup_probe_failed(struct sock *sk)
2890 struct inet_connection_sock *icsk = inet_csk(sk);
2892 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2893 icsk->icsk_mtup.probe_size = 0;
2896 static void tcp_mtup_probe_success(struct sock *sk)
2898 struct tcp_sock *tp = tcp_sk(sk);
2899 struct inet_connection_sock *icsk = inet_csk(sk);
2901 /* FIXME: breaks with very large cwnd */
2902 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2903 tp->snd_cwnd = tp->snd_cwnd *
2904 tcp_mss_to_mtu(sk, tp->mss_cache) /
2905 icsk->icsk_mtup.probe_size;
2906 tp->snd_cwnd_cnt = 0;
2907 tp->snd_cwnd_stamp = tcp_time_stamp;
2908 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2910 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2911 icsk->icsk_mtup.probe_size = 0;
2912 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2915 /* Do a simple retransmit without using the backoff mechanisms in
2916 * tcp_timer. This is used for path mtu discovery.
2917 * The socket is already locked here.
2919 void tcp_simple_retransmit(struct sock *sk)
2921 const struct inet_connection_sock *icsk = inet_csk(sk);
2922 struct tcp_sock *tp = tcp_sk(sk);
2923 struct sk_buff *skb;
2924 unsigned int mss = tcp_current_mss(sk);
2925 u32 prior_lost = tp->lost_out;
2927 tcp_for_write_queue(skb, sk) {
2928 if (skb == tcp_send_head(sk))
2930 if (tcp_skb_seglen(skb) > mss &&
2931 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2932 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2933 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2934 tp->retrans_out -= tcp_skb_pcount(skb);
2936 tcp_skb_mark_lost_uncond_verify(tp, skb);
2940 tcp_clear_retrans_hints_partial(tp);
2942 if (prior_lost == tp->lost_out)
2945 if (tcp_is_reno(tp))
2946 tcp_limit_reno_sacked(tp);
2948 tcp_verify_left_out(tp);
2950 /* Don't muck with the congestion window here.
2951 * Reason is that we do not increase amount of _data_
2952 * in network, but units changed and effective
2953 * cwnd/ssthresh really reduced now.
2955 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2956 tp->high_seq = tp->snd_nxt;
2957 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2958 tp->prior_ssthresh = 0;
2959 tp->undo_marker = 0;
2960 tcp_set_ca_state(sk, TCP_CA_Loss);
2962 tcp_xmit_retransmit_queue(sk);
2964 EXPORT_SYMBOL(tcp_simple_retransmit);
2966 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2967 * (proportional rate reduction with slow start reduction bound) as described in
2968 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2969 * It computes the number of packets to send (sndcnt) based on packets newly
2971 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2972 * cwnd reductions across a full RTT.
2973 * 2) If packets in flight is lower than ssthresh (such as due to excess
2974 * losses and/or application stalls), do not perform any further cwnd
2975 * reductions, but instead slow start up to ssthresh.
2977 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
2978 int fast_rexmit, int flag)
2980 struct tcp_sock *tp = tcp_sk(sk);
2982 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2984 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2985 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2987 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2989 sndcnt = min_t(int, delta,
2990 max_t(int, tp->prr_delivered - tp->prr_out,
2991 newly_acked_sacked) + 1);
2994 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2995 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2998 /* Process an event, which can update packets-in-flight not trivially.
2999 * Main goal of this function is to calculate new estimate for left_out,
3000 * taking into account both packets sitting in receiver's buffer and
3001 * packets lost by network.
3003 * Besides that it does CWND reduction, when packet loss is detected
3004 * and changes state of machine.
3006 * It does _not_ decide what to send, it is made in function
3007 * tcp_xmit_retransmit_queue().
3009 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3010 int newly_acked_sacked, int flag)
3012 struct inet_connection_sock *icsk = inet_csk(sk);
3013 struct tcp_sock *tp = tcp_sk(sk);
3014 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3015 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3016 (tcp_fackets_out(tp) > tp->reordering));
3017 int fast_rexmit = 0, mib_idx;
3019 if (WARN_ON(!tp->packets_out && tp->sacked_out))
3021 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3022 tp->fackets_out = 0;
3024 /* Now state machine starts.
3025 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3026 if (flag & FLAG_ECE)
3027 tp->prior_ssthresh = 0;
3029 /* B. In all the states check for reneging SACKs. */
3030 if (tcp_check_sack_reneging(sk, flag))
3033 /* C. Process data loss notification, provided it is valid. */
3034 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
3035 before(tp->snd_una, tp->high_seq) &&
3036 icsk->icsk_ca_state != TCP_CA_Open &&
3037 tp->fackets_out > tp->reordering) {
3038 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
3039 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
3042 /* D. Check consistency of the current state. */
3043 tcp_verify_left_out(tp);
3045 /* E. Check state exit conditions. State can be terminated
3046 * when high_seq is ACKed. */
3047 if (icsk->icsk_ca_state == TCP_CA_Open) {
3048 WARN_ON(tp->retrans_out != 0);
3049 tp->retrans_stamp = 0;
3050 } else if (!before(tp->snd_una, tp->high_seq)) {
3051 switch (icsk->icsk_ca_state) {
3053 icsk->icsk_retransmits = 0;
3054 if (tcp_try_undo_recovery(sk))
3059 /* CWR is to be held something *above* high_seq
3060 * is ACKed for CWR bit to reach receiver. */
3061 if (tp->snd_una != tp->high_seq) {
3062 tcp_complete_cwr(sk);
3063 tcp_set_ca_state(sk, TCP_CA_Open);
3067 case TCP_CA_Disorder:
3068 tcp_try_undo_dsack(sk);
3069 if (!tp->undo_marker ||
3070 /* For SACK case do not Open to allow to undo
3071 * catching for all duplicate ACKs. */
3072 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3073 tp->undo_marker = 0;
3074 tcp_set_ca_state(sk, TCP_CA_Open);
3078 case TCP_CA_Recovery:
3079 if (tcp_is_reno(tp))
3080 tcp_reset_reno_sack(tp);
3081 if (tcp_try_undo_recovery(sk))
3083 tcp_complete_cwr(sk);
3088 /* F. Process state. */
3089 switch (icsk->icsk_ca_state) {
3090 case TCP_CA_Recovery:
3091 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3092 if (tcp_is_reno(tp) && is_dupack)
3093 tcp_add_reno_sack(sk);
3095 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3098 if (flag & FLAG_DATA_ACKED)
3099 icsk->icsk_retransmits = 0;
3100 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3101 tcp_reset_reno_sack(tp);
3102 if (!tcp_try_undo_loss(sk)) {
3103 tcp_moderate_cwnd(tp);
3104 tcp_xmit_retransmit_queue(sk);
3107 if (icsk->icsk_ca_state != TCP_CA_Open)
3109 /* Loss is undone; fall through to processing in Open state. */
3111 if (tcp_is_reno(tp)) {
3112 if (flag & FLAG_SND_UNA_ADVANCED)
3113 tcp_reset_reno_sack(tp);
3115 tcp_add_reno_sack(sk);
3118 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3119 tcp_try_undo_dsack(sk);
3121 if (!tcp_time_to_recover(sk)) {
3122 tcp_try_to_open(sk, flag);
3126 /* MTU probe failure: don't reduce cwnd */
3127 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3128 icsk->icsk_mtup.probe_size &&
3129 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3130 tcp_mtup_probe_failed(sk);
3131 /* Restores the reduction we did in tcp_mtup_probe() */
3133 tcp_simple_retransmit(sk);
3137 /* Otherwise enter Recovery state */
3139 if (tcp_is_reno(tp))
3140 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3142 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3144 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3146 tp->high_seq = tp->snd_nxt;
3147 tp->prior_ssthresh = 0;
3148 tp->undo_marker = tp->snd_una;
3149 tp->undo_retrans = tp->retrans_out;
3151 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3152 if (!(flag & FLAG_ECE))
3153 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3154 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3155 TCP_ECN_queue_cwr(tp);
3158 tp->bytes_acked = 0;
3159 tp->snd_cwnd_cnt = 0;
3160 tp->prior_cwnd = tp->snd_cwnd;
3161 tp->prr_delivered = 0;
3163 tcp_set_ca_state(sk, TCP_CA_Recovery);
3167 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3168 tcp_update_scoreboard(sk, fast_rexmit);
3169 tp->prr_delivered += newly_acked_sacked;
3170 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3171 tcp_xmit_retransmit_queue(sk);
3174 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3176 tcp_rtt_estimator(sk, seq_rtt);
3178 inet_csk(sk)->icsk_backoff = 0;
3180 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3182 /* Read draft-ietf-tcplw-high-performance before mucking
3183 * with this code. (Supersedes RFC1323)
3185 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3187 /* RTTM Rule: A TSecr value received in a segment is used to
3188 * update the averaged RTT measurement only if the segment
3189 * acknowledges some new data, i.e., only if it advances the
3190 * left edge of the send window.
3192 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3193 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3195 * Changed: reset backoff as soon as we see the first valid sample.
3196 * If we do not, we get strongly overestimated rto. With timestamps
3197 * samples are accepted even from very old segments: f.e., when rtt=1
3198 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3199 * answer arrives rto becomes 120 seconds! If at least one of segments
3200 * in window is lost... Voila. --ANK (010210)
3202 struct tcp_sock *tp = tcp_sk(sk);
3204 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3207 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3209 /* We don't have a timestamp. Can only use
3210 * packets that are not retransmitted to determine
3211 * rtt estimates. Also, we must not reset the
3212 * backoff for rto until we get a non-retransmitted
3213 * packet. This allows us to deal with a situation
3214 * where the network delay has increased suddenly.
3215 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3218 if (flag & FLAG_RETRANS_DATA_ACKED)
3221 tcp_valid_rtt_meas(sk, seq_rtt);
3224 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3227 const struct tcp_sock *tp = tcp_sk(sk);
3228 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3229 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3230 tcp_ack_saw_tstamp(sk, flag);
3231 else if (seq_rtt >= 0)
3232 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3235 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3237 const struct inet_connection_sock *icsk = inet_csk(sk);
3238 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3239 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3242 /* Restart timer after forward progress on connection.
3243 * RFC2988 recommends to restart timer to now+rto.
3245 static void tcp_rearm_rto(struct sock *sk)
3247 struct tcp_sock *tp = tcp_sk(sk);
3249 if (!tp->packets_out) {
3250 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3252 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3253 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3257 /* If we get here, the whole TSO packet has not been acked. */
3258 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3260 struct tcp_sock *tp = tcp_sk(sk);
3263 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3265 packets_acked = tcp_skb_pcount(skb);
3266 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3268 packets_acked -= tcp_skb_pcount(skb);
3270 if (packets_acked) {
3271 BUG_ON(tcp_skb_pcount(skb) == 0);
3272 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3275 return packets_acked;
3278 /* Remove acknowledged frames from the retransmission queue. If our packet
3279 * is before the ack sequence we can discard it as it's confirmed to have
3280 * arrived at the other end.
3282 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3285 struct tcp_sock *tp = tcp_sk(sk);
3286 const struct inet_connection_sock *icsk = inet_csk(sk);
3287 struct sk_buff *skb;
3288 u32 now = tcp_time_stamp;
3289 int fully_acked = 1;
3292 u32 reord = tp->packets_out;
3293 u32 prior_sacked = tp->sacked_out;
3295 s32 ca_seq_rtt = -1;
3296 ktime_t last_ackt = net_invalid_timestamp();
3298 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3299 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3301 u8 sacked = scb->sacked;
3303 /* Determine how many packets and what bytes were acked, tso and else */
3304 if (after(scb->end_seq, tp->snd_una)) {
3305 if (tcp_skb_pcount(skb) == 1 ||
3306 !after(tp->snd_una, scb->seq))
3309 acked_pcount = tcp_tso_acked(sk, skb);
3315 acked_pcount = tcp_skb_pcount(skb);
3318 if (sacked & TCPCB_RETRANS) {
3319 if (sacked & TCPCB_SACKED_RETRANS)
3320 tp->retrans_out -= acked_pcount;
3321 flag |= FLAG_RETRANS_DATA_ACKED;
3324 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3325 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3327 ca_seq_rtt = now - scb->when;
3328 last_ackt = skb->tstamp;
3330 seq_rtt = ca_seq_rtt;
3332 if (!(sacked & TCPCB_SACKED_ACKED))
3333 reord = min(pkts_acked, reord);
3336 if (sacked & TCPCB_SACKED_ACKED)
3337 tp->sacked_out -= acked_pcount;
3338 if (sacked & TCPCB_LOST)
3339 tp->lost_out -= acked_pcount;
3341 tp->packets_out -= acked_pcount;
3342 pkts_acked += acked_pcount;
3344 /* Initial outgoing SYN's get put onto the write_queue
3345 * just like anything else we transmit. It is not
3346 * true data, and if we misinform our callers that
3347 * this ACK acks real data, we will erroneously exit
3348 * connection startup slow start one packet too
3349 * quickly. This is severely frowned upon behavior.
3351 if (!(scb->flags & TCPHDR_SYN)) {
3352 flag |= FLAG_DATA_ACKED;
3354 flag |= FLAG_SYN_ACKED;
3355 tp->retrans_stamp = 0;
3361 tcp_unlink_write_queue(skb, sk);
3362 sk_wmem_free_skb(sk, skb);
3363 tp->scoreboard_skb_hint = NULL;
3364 if (skb == tp->retransmit_skb_hint)
3365 tp->retransmit_skb_hint = NULL;
3366 if (skb == tp->lost_skb_hint)
3367 tp->lost_skb_hint = NULL;
3370 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3371 tp->snd_up = tp->snd_una;
3373 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3374 flag |= FLAG_SACK_RENEGING;
3376 if (flag & FLAG_ACKED) {
3377 const struct tcp_congestion_ops *ca_ops
3378 = inet_csk(sk)->icsk_ca_ops;
3380 if (unlikely(icsk->icsk_mtup.probe_size &&
3381 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3382 tcp_mtup_probe_success(sk);
3385 tcp_ack_update_rtt(sk, flag, seq_rtt);
3388 if (tcp_is_reno(tp)) {
3389 tcp_remove_reno_sacks(sk, pkts_acked);
3393 /* Non-retransmitted hole got filled? That's reordering */
3394 if (reord < prior_fackets)
3395 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3397 delta = tcp_is_fack(tp) ? pkts_acked :
3398 prior_sacked - tp->sacked_out;
3399 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3402 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3404 if (ca_ops->pkts_acked) {
3407 /* Is the ACK triggering packet unambiguous? */
3408 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3409 /* High resolution needed and available? */
3410 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3411 !ktime_equal(last_ackt,
3412 net_invalid_timestamp()))
3413 rtt_us = ktime_us_delta(ktime_get_real(),
3415 else if (ca_seq_rtt >= 0)
3416 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3419 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3423 #if FASTRETRANS_DEBUG > 0
3424 WARN_ON((int)tp->sacked_out < 0);
3425 WARN_ON((int)tp->lost_out < 0);
3426 WARN_ON((int)tp->retrans_out < 0);
3427 if (!tp->packets_out && tcp_is_sack(tp)) {
3428 icsk = inet_csk(sk);
3430 printk(KERN_DEBUG "Leak l=%u %d\n",
3431 tp->lost_out, icsk->icsk_ca_state);
3434 if (tp->sacked_out) {
3435 printk(KERN_DEBUG "Leak s=%u %d\n",
3436 tp->sacked_out, icsk->icsk_ca_state);
3439 if (tp->retrans_out) {
3440 printk(KERN_DEBUG "Leak r=%u %d\n",
3441 tp->retrans_out, icsk->icsk_ca_state);
3442 tp->retrans_out = 0;
3449 static void tcp_ack_probe(struct sock *sk)
3451 const struct tcp_sock *tp = tcp_sk(sk);
3452 struct inet_connection_sock *icsk = inet_csk(sk);
3454 /* Was it a usable window open? */
3456 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3457 icsk->icsk_backoff = 0;
3458 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3459 /* Socket must be waked up by subsequent tcp_data_snd_check().
3460 * This function is not for random using!
3463 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3464 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3469 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3471 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3472 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3475 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3477 const struct tcp_sock *tp = tcp_sk(sk);
3478 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3479 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3482 /* Check that window update is acceptable.
3483 * The function assumes that snd_una<=ack<=snd_next.
3485 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3486 const u32 ack, const u32 ack_seq,
3489 return after(ack, tp->snd_una) ||
3490 after(ack_seq, tp->snd_wl1) ||
3491 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3494 /* Update our send window.
3496 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3497 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3499 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3502 struct tcp_sock *tp = tcp_sk(sk);
3504 u32 nwin = ntohs(tcp_hdr(skb)->window);
3506 if (likely(!tcp_hdr(skb)->syn))
3507 nwin <<= tp->rx_opt.snd_wscale;
3509 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3510 flag |= FLAG_WIN_UPDATE;
3511 tcp_update_wl(tp, ack_seq);
3513 if (tp->snd_wnd != nwin) {
3516 /* Note, it is the only place, where
3517 * fast path is recovered for sending TCP.
3520 tcp_fast_path_check(sk);
3522 if (nwin > tp->max_window) {
3523 tp->max_window = nwin;
3524 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3534 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3535 * continue in congestion avoidance.
3537 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3539 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3540 tp->snd_cwnd_cnt = 0;
3541 tp->bytes_acked = 0;
3542 TCP_ECN_queue_cwr(tp);
3543 tcp_moderate_cwnd(tp);
3546 /* A conservative spurious RTO response algorithm: reduce cwnd using
3547 * rate halving and continue in congestion avoidance.
3549 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3551 tcp_enter_cwr(sk, 0);
3554 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3556 if (flag & FLAG_ECE)
3557 tcp_ratehalving_spur_to_response(sk);
3559 tcp_undo_cwr(sk, true);
3562 /* F-RTO spurious RTO detection algorithm (RFC4138)
3564 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3565 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3566 * window (but not to or beyond highest sequence sent before RTO):
3567 * On First ACK, send two new segments out.
3568 * On Second ACK, RTO was likely spurious. Do spurious response (response
3569 * algorithm is not part of the F-RTO detection algorithm
3570 * given in RFC4138 but can be selected separately).
3571 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3572 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3573 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3574 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3576 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3577 * original window even after we transmit two new data segments.
3580 * on first step, wait until first cumulative ACK arrives, then move to
3581 * the second step. In second step, the next ACK decides.
3583 * F-RTO is implemented (mainly) in four functions:
3584 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3585 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3586 * called when tcp_use_frto() showed green light
3587 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3588 * - tcp_enter_frto_loss() is called if there is not enough evidence
3589 * to prove that the RTO is indeed spurious. It transfers the control
3590 * from F-RTO to the conventional RTO recovery
3592 static int tcp_process_frto(struct sock *sk, int flag)
3594 struct tcp_sock *tp = tcp_sk(sk);
3596 tcp_verify_left_out(tp);
3598 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3599 if (flag & FLAG_DATA_ACKED)
3600 inet_csk(sk)->icsk_retransmits = 0;
3602 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3603 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3604 tp->undo_marker = 0;
3606 if (!before(tp->snd_una, tp->frto_highmark)) {
3607 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3611 if (!tcp_is_sackfrto(tp)) {
3612 /* RFC4138 shortcoming in step 2; should also have case c):
3613 * ACK isn't duplicate nor advances window, e.g., opposite dir
3616 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3619 if (!(flag & FLAG_DATA_ACKED)) {
3620 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3625 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3626 /* Prevent sending of new data. */
3627 tp->snd_cwnd = min(tp->snd_cwnd,
3628 tcp_packets_in_flight(tp));
3632 if ((tp->frto_counter >= 2) &&
3633 (!(flag & FLAG_FORWARD_PROGRESS) ||
3634 ((flag & FLAG_DATA_SACKED) &&
3635 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3636 /* RFC4138 shortcoming (see comment above) */
3637 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3638 (flag & FLAG_NOT_DUP))
3641 tcp_enter_frto_loss(sk, 3, flag);
3646 if (tp->frto_counter == 1) {
3647 /* tcp_may_send_now needs to see updated state */
3648 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3649 tp->frto_counter = 2;
3651 if (!tcp_may_send_now(sk))
3652 tcp_enter_frto_loss(sk, 2, flag);
3656 switch (sysctl_tcp_frto_response) {
3658 tcp_undo_spur_to_response(sk, flag);
3661 tcp_conservative_spur_to_response(tp);
3664 tcp_ratehalving_spur_to_response(sk);
3667 tp->frto_counter = 0;
3668 tp->undo_marker = 0;
3669 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3674 /* This routine deals with incoming acks, but not outgoing ones. */
3675 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3677 struct inet_connection_sock *icsk = inet_csk(sk);
3678 struct tcp_sock *tp = tcp_sk(sk);
3679 u32 prior_snd_una = tp->snd_una;
3680 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3681 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3682 u32 prior_in_flight;
3685 int prior_sacked = tp->sacked_out;
3686 int newly_acked_sacked = 0;
3689 /* If the ack is older than previous acks
3690 * then we can probably ignore it.
3692 if (before(ack, prior_snd_una))
3695 /* If the ack includes data we haven't sent yet, discard
3696 * this segment (RFC793 Section 3.9).
3698 if (after(ack, tp->snd_nxt))
3701 if (after(ack, prior_snd_una))
3702 flag |= FLAG_SND_UNA_ADVANCED;
3704 if (sysctl_tcp_abc) {
3705 if (icsk->icsk_ca_state < TCP_CA_CWR)
3706 tp->bytes_acked += ack - prior_snd_una;
3707 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3708 /* we assume just one segment left network */
3709 tp->bytes_acked += min(ack - prior_snd_una,
3713 prior_fackets = tp->fackets_out;
3714 prior_in_flight = tcp_packets_in_flight(tp);
3716 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3717 /* Window is constant, pure forward advance.
3718 * No more checks are required.
3719 * Note, we use the fact that SND.UNA>=SND.WL2.
3721 tcp_update_wl(tp, ack_seq);
3723 flag |= FLAG_WIN_UPDATE;
3725 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3727 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3729 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3732 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3734 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3736 if (TCP_SKB_CB(skb)->sacked)
3737 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3739 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3742 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3745 /* We passed data and got it acked, remove any soft error
3746 * log. Something worked...
3748 sk->sk_err_soft = 0;
3749 icsk->icsk_probes_out = 0;
3750 tp->rcv_tstamp = tcp_time_stamp;
3751 prior_packets = tp->packets_out;
3755 /* See if we can take anything off of the retransmit queue. */
3756 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3758 newly_acked_sacked = (prior_packets - prior_sacked) -
3759 (tp->packets_out - tp->sacked_out);
3761 if (tp->frto_counter)
3762 frto_cwnd = tcp_process_frto(sk, flag);
3763 /* Guarantee sacktag reordering detection against wrap-arounds */
3764 if (before(tp->frto_highmark, tp->snd_una))
3765 tp->frto_highmark = 0;
3767 if (tcp_ack_is_dubious(sk, flag)) {
3768 /* Advance CWND, if state allows this. */
3769 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3770 tcp_may_raise_cwnd(sk, flag))
3771 tcp_cong_avoid(sk, ack, prior_in_flight);
3772 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3773 newly_acked_sacked, flag);
3775 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3776 tcp_cong_avoid(sk, ack, prior_in_flight);
3779 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3780 dst_confirm(__sk_dst_get(sk));
3785 /* If this ack opens up a zero window, clear backoff. It was
3786 * being used to time the probes, and is probably far higher than
3787 * it needs to be for normal retransmission.
3789 if (tcp_send_head(sk))
3794 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3798 if (TCP_SKB_CB(skb)->sacked) {
3799 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3800 if (icsk->icsk_ca_state == TCP_CA_Open)
3801 tcp_try_keep_open(sk);
3804 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3808 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3809 * But, this can also be called on packets in the established flow when
3810 * the fast version below fails.
3812 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3813 u8 **hvpp, int estab)
3816 struct tcphdr *th = tcp_hdr(skb);
3817 int length = (th->doff * 4) - sizeof(struct tcphdr);
3819 ptr = (unsigned char *)(th + 1);
3820 opt_rx->saw_tstamp = 0;
3822 while (length > 0) {
3823 int opcode = *ptr++;
3829 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3834 if (opsize < 2) /* "silly options" */
3836 if (opsize > length)
3837 return; /* don't parse partial options */
3840 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3841 u16 in_mss = get_unaligned_be16(ptr);
3843 if (opt_rx->user_mss &&
3844 opt_rx->user_mss < in_mss)
3845 in_mss = opt_rx->user_mss;
3846 opt_rx->mss_clamp = in_mss;
3851 if (opsize == TCPOLEN_WINDOW && th->syn &&
3852 !estab && sysctl_tcp_window_scaling) {
3853 __u8 snd_wscale = *(__u8 *)ptr;
3854 opt_rx->wscale_ok = 1;
3855 if (snd_wscale > 14) {
3856 if (net_ratelimit())
3857 printk(KERN_INFO "tcp_parse_options: Illegal window "
3858 "scaling value %d >14 received.\n",
3862 opt_rx->snd_wscale = snd_wscale;
3865 case TCPOPT_TIMESTAMP:
3866 if ((opsize == TCPOLEN_TIMESTAMP) &&
3867 ((estab && opt_rx->tstamp_ok) ||
3868 (!estab && sysctl_tcp_timestamps))) {
3869 opt_rx->saw_tstamp = 1;
3870 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3871 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3874 case TCPOPT_SACK_PERM:
3875 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3876 !estab && sysctl_tcp_sack) {
3877 opt_rx->sack_ok = 1;
3878 tcp_sack_reset(opt_rx);
3883 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3884 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3886 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3889 #ifdef CONFIG_TCP_MD5SIG
3892 * The MD5 Hash has already been
3893 * checked (see tcp_v{4,6}_do_rcv()).
3898 /* This option is variable length.
3901 case TCPOLEN_COOKIE_BASE:
3902 /* not yet implemented */
3904 case TCPOLEN_COOKIE_PAIR:
3905 /* not yet implemented */
3907 case TCPOLEN_COOKIE_MIN+0:
3908 case TCPOLEN_COOKIE_MIN+2:
3909 case TCPOLEN_COOKIE_MIN+4:
3910 case TCPOLEN_COOKIE_MIN+6:
3911 case TCPOLEN_COOKIE_MAX:
3912 /* 16-bit multiple */
3913 opt_rx->cookie_plus = opsize;
3928 EXPORT_SYMBOL(tcp_parse_options);
3930 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3932 __be32 *ptr = (__be32 *)(th + 1);
3934 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3935 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3936 tp->rx_opt.saw_tstamp = 1;
3938 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3940 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3946 /* Fast parse options. This hopes to only see timestamps.
3947 * If it is wrong it falls back on tcp_parse_options().
3949 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3950 struct tcp_sock *tp, u8 **hvpp)
3952 /* In the spirit of fast parsing, compare doff directly to constant
3953 * values. Because equality is used, short doff can be ignored here.
3955 if (th->doff == (sizeof(*th) / 4)) {
3956 tp->rx_opt.saw_tstamp = 0;
3958 } else if (tp->rx_opt.tstamp_ok &&
3959 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3960 if (tcp_parse_aligned_timestamp(tp, th))
3963 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3967 #ifdef CONFIG_TCP_MD5SIG
3969 * Parse MD5 Signature option
3971 u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3973 int length = (th->doff << 2) - sizeof (*th);
3974 u8 *ptr = (u8*)(th + 1);
3976 /* If the TCP option is too short, we can short cut */
3977 if (length < TCPOLEN_MD5SIG)
3980 while (length > 0) {
3981 int opcode = *ptr++;
3992 if (opsize < 2 || opsize > length)
3994 if (opcode == TCPOPT_MD5SIG)
3995 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4002 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4005 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4007 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
4008 tp->rx_opt.ts_recent_stamp = get_seconds();
4011 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4013 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4014 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
4015 * extra check below makes sure this can only happen
4016 * for pure ACK frames. -DaveM
4018 * Not only, also it occurs for expired timestamps.
4021 if (tcp_paws_check(&tp->rx_opt, 0))
4022 tcp_store_ts_recent(tp);
4026 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4028 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4029 * it can pass through stack. So, the following predicate verifies that
4030 * this segment is not used for anything but congestion avoidance or
4031 * fast retransmit. Moreover, we even are able to eliminate most of such
4032 * second order effects, if we apply some small "replay" window (~RTO)
4033 * to timestamp space.
4035 * All these measures still do not guarantee that we reject wrapped ACKs
4036 * on networks with high bandwidth, when sequence space is recycled fastly,
4037 * but it guarantees that such events will be very rare and do not affect
4038 * connection seriously. This doesn't look nice, but alas, PAWS is really
4041 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4042 * states that events when retransmit arrives after original data are rare.
4043 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4044 * the biggest problem on large power networks even with minor reordering.
4045 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4046 * up to bandwidth of 18Gigabit/sec. 8) ]
4049 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4051 struct tcp_sock *tp = tcp_sk(sk);
4052 struct tcphdr *th = tcp_hdr(skb);
4053 u32 seq = TCP_SKB_CB(skb)->seq;
4054 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4056 return (/* 1. Pure ACK with correct sequence number. */
4057 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4059 /* 2. ... and duplicate ACK. */
4060 ack == tp->snd_una &&
4062 /* 3. ... and does not update window. */
4063 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4065 /* 4. ... and sits in replay window. */
4066 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4069 static inline int tcp_paws_discard(const struct sock *sk,
4070 const struct sk_buff *skb)
4072 const struct tcp_sock *tp = tcp_sk(sk);
4074 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4075 !tcp_disordered_ack(sk, skb);
4078 /* Check segment sequence number for validity.
4080 * Segment controls are considered valid, if the segment
4081 * fits to the window after truncation to the window. Acceptability
4082 * of data (and SYN, FIN, of course) is checked separately.
4083 * See tcp_data_queue(), for example.
4085 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4086 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4087 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4088 * (borrowed from freebsd)
4091 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
4093 return !before(end_seq, tp->rcv_wup) &&
4094 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4097 /* When we get a reset we do this. */
4098 static void tcp_reset(struct sock *sk)
4100 /* We want the right error as BSD sees it (and indeed as we do). */
4101 switch (sk->sk_state) {
4103 sk->sk_err = ECONNREFUSED;
4105 case TCP_CLOSE_WAIT:
4111 sk->sk_err = ECONNRESET;
4113 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4116 if (!sock_flag(sk, SOCK_DEAD))
4117 sk->sk_error_report(sk);
4123 * Process the FIN bit. This now behaves as it is supposed to work
4124 * and the FIN takes effect when it is validly part of sequence
4125 * space. Not before when we get holes.
4127 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4128 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4131 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4132 * close and we go into CLOSING (and later onto TIME-WAIT)
4134 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4136 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4138 struct tcp_sock *tp = tcp_sk(sk);
4140 inet_csk_schedule_ack(sk);
4142 sk->sk_shutdown |= RCV_SHUTDOWN;
4143 sock_set_flag(sk, SOCK_DONE);
4145 switch (sk->sk_state) {
4147 case TCP_ESTABLISHED:
4148 /* Move to CLOSE_WAIT */
4149 tcp_set_state(sk, TCP_CLOSE_WAIT);
4150 inet_csk(sk)->icsk_ack.pingpong = 1;
4153 case TCP_CLOSE_WAIT:
4155 /* Received a retransmission of the FIN, do
4160 /* RFC793: Remain in the LAST-ACK state. */
4164 /* This case occurs when a simultaneous close
4165 * happens, we must ack the received FIN and
4166 * enter the CLOSING state.
4169 tcp_set_state(sk, TCP_CLOSING);
4172 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4174 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4177 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4178 * cases we should never reach this piece of code.
4180 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4181 __func__, sk->sk_state);
4185 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4186 * Probably, we should reset in this case. For now drop them.
4188 __skb_queue_purge(&tp->out_of_order_queue);
4189 if (tcp_is_sack(tp))
4190 tcp_sack_reset(&tp->rx_opt);
4193 if (!sock_flag(sk, SOCK_DEAD)) {
4194 sk->sk_state_change(sk);
4196 /* Do not send POLL_HUP for half duplex close. */
4197 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4198 sk->sk_state == TCP_CLOSE)
4199 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4201 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4205 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4208 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4209 if (before(seq, sp->start_seq))
4210 sp->start_seq = seq;
4211 if (after(end_seq, sp->end_seq))
4212 sp->end_seq = end_seq;
4218 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4220 struct tcp_sock *tp = tcp_sk(sk);
4222 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4225 if (before(seq, tp->rcv_nxt))
4226 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4228 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4230 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4232 tp->rx_opt.dsack = 1;
4233 tp->duplicate_sack[0].start_seq = seq;
4234 tp->duplicate_sack[0].end_seq = end_seq;
4238 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4240 struct tcp_sock *tp = tcp_sk(sk);
4242 if (!tp->rx_opt.dsack)
4243 tcp_dsack_set(sk, seq, end_seq);
4245 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4248 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4250 struct tcp_sock *tp = tcp_sk(sk);
4252 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4253 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4254 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4255 tcp_enter_quickack_mode(sk);
4257 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4258 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4260 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4261 end_seq = tp->rcv_nxt;
4262 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4269 /* These routines update the SACK block as out-of-order packets arrive or
4270 * in-order packets close up the sequence space.
4272 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4275 struct tcp_sack_block *sp = &tp->selective_acks[0];
4276 struct tcp_sack_block *swalk = sp + 1;
4278 /* See if the recent change to the first SACK eats into
4279 * or hits the sequence space of other SACK blocks, if so coalesce.
4281 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4282 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4285 /* Zap SWALK, by moving every further SACK up by one slot.
4286 * Decrease num_sacks.
4288 tp->rx_opt.num_sacks--;
4289 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4293 this_sack++, swalk++;
4297 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4299 struct tcp_sock *tp = tcp_sk(sk);
4300 struct tcp_sack_block *sp = &tp->selective_acks[0];
4301 int cur_sacks = tp->rx_opt.num_sacks;
4307 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4308 if (tcp_sack_extend(sp, seq, end_seq)) {
4309 /* Rotate this_sack to the first one. */
4310 for (; this_sack > 0; this_sack--, sp--)
4311 swap(*sp, *(sp - 1));
4313 tcp_sack_maybe_coalesce(tp);
4318 /* Could not find an adjacent existing SACK, build a new one,
4319 * put it at the front, and shift everyone else down. We
4320 * always know there is at least one SACK present already here.
4322 * If the sack array is full, forget about the last one.
4324 if (this_sack >= TCP_NUM_SACKS) {
4326 tp->rx_opt.num_sacks--;
4329 for (; this_sack > 0; this_sack--, sp--)
4333 /* Build the new head SACK, and we're done. */
4334 sp->start_seq = seq;
4335 sp->end_seq = end_seq;
4336 tp->rx_opt.num_sacks++;
4339 /* RCV.NXT advances, some SACKs should be eaten. */
4341 static void tcp_sack_remove(struct tcp_sock *tp)
4343 struct tcp_sack_block *sp = &tp->selective_acks[0];
4344 int num_sacks = tp->rx_opt.num_sacks;
4347 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4348 if (skb_queue_empty(&tp->out_of_order_queue)) {
4349 tp->rx_opt.num_sacks = 0;
4353 for (this_sack = 0; this_sack < num_sacks;) {
4354 /* Check if the start of the sack is covered by RCV.NXT. */
4355 if (!before(tp->rcv_nxt, sp->start_seq)) {
4358 /* RCV.NXT must cover all the block! */
4359 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4361 /* Zap this SACK, by moving forward any other SACKS. */
4362 for (i=this_sack+1; i < num_sacks; i++)
4363 tp->selective_acks[i-1] = tp->selective_acks[i];
4370 tp->rx_opt.num_sacks = num_sacks;
4373 /* This one checks to see if we can put data from the
4374 * out_of_order queue into the receive_queue.
4376 static void tcp_ofo_queue(struct sock *sk)
4378 struct tcp_sock *tp = tcp_sk(sk);
4379 __u32 dsack_high = tp->rcv_nxt;
4380 struct sk_buff *skb;
4382 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4383 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4386 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4387 __u32 dsack = dsack_high;
4388 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4389 dsack_high = TCP_SKB_CB(skb)->end_seq;
4390 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4393 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4394 SOCK_DEBUG(sk, "ofo packet was already received\n");
4395 __skb_unlink(skb, &tp->out_of_order_queue);
4399 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4400 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4401 TCP_SKB_CB(skb)->end_seq);
4403 __skb_unlink(skb, &tp->out_of_order_queue);
4404 __skb_queue_tail(&sk->sk_receive_queue, skb);
4405 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4406 if (tcp_hdr(skb)->fin)
4407 tcp_fin(skb, sk, tcp_hdr(skb));
4411 static int tcp_prune_ofo_queue(struct sock *sk);
4412 static int tcp_prune_queue(struct sock *sk);
4414 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4416 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4417 !sk_rmem_schedule(sk, size)) {
4419 if (tcp_prune_queue(sk) < 0)
4422 if (!sk_rmem_schedule(sk, size)) {
4423 if (!tcp_prune_ofo_queue(sk))
4426 if (!sk_rmem_schedule(sk, size))
4433 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4435 struct tcphdr *th = tcp_hdr(skb);
4436 struct tcp_sock *tp = tcp_sk(sk);
4439 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4443 __skb_pull(skb, th->doff * 4);
4445 TCP_ECN_accept_cwr(tp, skb);
4447 tp->rx_opt.dsack = 0;
4449 /* Queue data for delivery to the user.
4450 * Packets in sequence go to the receive queue.
4451 * Out of sequence packets to the out_of_order_queue.
4453 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4454 if (tcp_receive_window(tp) == 0)
4457 /* Ok. In sequence. In window. */
4458 if (tp->ucopy.task == current &&
4459 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4460 sock_owned_by_user(sk) && !tp->urg_data) {
4461 int chunk = min_t(unsigned int, skb->len,
4464 __set_current_state(TASK_RUNNING);
4467 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4468 tp->ucopy.len -= chunk;
4469 tp->copied_seq += chunk;
4470 eaten = (chunk == skb->len);
4471 tcp_rcv_space_adjust(sk);
4479 tcp_try_rmem_schedule(sk, skb->truesize))
4482 skb_set_owner_r(skb, sk);
4483 __skb_queue_tail(&sk->sk_receive_queue, skb);
4485 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4487 tcp_event_data_recv(sk, skb);
4489 tcp_fin(skb, sk, th);
4491 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4494 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4495 * gap in queue is filled.
4497 if (skb_queue_empty(&tp->out_of_order_queue))
4498 inet_csk(sk)->icsk_ack.pingpong = 0;
4501 if (tp->rx_opt.num_sacks)
4502 tcp_sack_remove(tp);
4504 tcp_fast_path_check(sk);
4508 else if (!sock_flag(sk, SOCK_DEAD))
4509 sk->sk_data_ready(sk, 0);
4513 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4514 /* A retransmit, 2nd most common case. Force an immediate ack. */
4515 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4516 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4519 tcp_enter_quickack_mode(sk);
4520 inet_csk_schedule_ack(sk);
4526 /* Out of window. F.e. zero window probe. */
4527 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4530 tcp_enter_quickack_mode(sk);
4532 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4533 /* Partial packet, seq < rcv_next < end_seq */
4534 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4535 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4536 TCP_SKB_CB(skb)->end_seq);
4538 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4540 /* If window is closed, drop tail of packet. But after
4541 * remembering D-SACK for its head made in previous line.
4543 if (!tcp_receive_window(tp))
4548 TCP_ECN_check_ce(tp, skb);
4550 if (tcp_try_rmem_schedule(sk, skb->truesize))
4553 /* Disable header prediction. */
4555 inet_csk_schedule_ack(sk);
4557 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4558 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4560 skb_set_owner_r(skb, sk);
4562 if (!skb_peek(&tp->out_of_order_queue)) {
4563 /* Initial out of order segment, build 1 SACK. */
4564 if (tcp_is_sack(tp)) {
4565 tp->rx_opt.num_sacks = 1;
4566 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4567 tp->selective_acks[0].end_seq =
4568 TCP_SKB_CB(skb)->end_seq;
4570 __skb_queue_head(&tp->out_of_order_queue, skb);
4572 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4573 u32 seq = TCP_SKB_CB(skb)->seq;
4574 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4576 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4577 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4579 if (!tp->rx_opt.num_sacks ||
4580 tp->selective_acks[0].end_seq != seq)
4583 /* Common case: data arrive in order after hole. */
4584 tp->selective_acks[0].end_seq = end_seq;
4588 /* Find place to insert this segment. */
4590 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4592 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4596 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4599 /* Do skb overlap to previous one? */
4600 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4601 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4602 /* All the bits are present. Drop. */
4604 tcp_dsack_set(sk, seq, end_seq);
4607 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4608 /* Partial overlap. */
4609 tcp_dsack_set(sk, seq,
4610 TCP_SKB_CB(skb1)->end_seq);
4612 if (skb_queue_is_first(&tp->out_of_order_queue,
4616 skb1 = skb_queue_prev(
4617 &tp->out_of_order_queue,
4622 __skb_queue_head(&tp->out_of_order_queue, skb);
4624 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4626 /* And clean segments covered by new one as whole. */
4627 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4628 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4630 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4632 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4633 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4637 __skb_unlink(skb1, &tp->out_of_order_queue);
4638 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4639 TCP_SKB_CB(skb1)->end_seq);
4644 if (tcp_is_sack(tp))
4645 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4649 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4650 struct sk_buff_head *list)
4652 struct sk_buff *next = NULL;
4654 if (!skb_queue_is_last(list, skb))
4655 next = skb_queue_next(list, skb);
4657 __skb_unlink(skb, list);
4659 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4664 /* Collapse contiguous sequence of skbs head..tail with
4665 * sequence numbers start..end.
4667 * If tail is NULL, this means until the end of the list.
4669 * Segments with FIN/SYN are not collapsed (only because this
4673 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4674 struct sk_buff *head, struct sk_buff *tail,
4677 struct sk_buff *skb, *n;
4680 /* First, check that queue is collapsible and find
4681 * the point where collapsing can be useful. */
4685 skb_queue_walk_from_safe(list, skb, n) {
4688 /* No new bits? It is possible on ofo queue. */
4689 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4690 skb = tcp_collapse_one(sk, skb, list);
4696 /* The first skb to collapse is:
4698 * - bloated or contains data before "start" or
4699 * overlaps to the next one.
4701 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4702 (tcp_win_from_space(skb->truesize) > skb->len ||
4703 before(TCP_SKB_CB(skb)->seq, start))) {
4704 end_of_skbs = false;
4708 if (!skb_queue_is_last(list, skb)) {
4709 struct sk_buff *next = skb_queue_next(list, skb);
4711 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4712 end_of_skbs = false;
4717 /* Decided to skip this, advance start seq. */
4718 start = TCP_SKB_CB(skb)->end_seq;
4720 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4723 while (before(start, end)) {
4724 struct sk_buff *nskb;
4725 unsigned int header = skb_headroom(skb);
4726 int copy = SKB_MAX_ORDER(header, 0);
4728 /* Too big header? This can happen with IPv6. */
4731 if (end - start < copy)
4733 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4737 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4738 skb_set_network_header(nskb, (skb_network_header(skb) -
4740 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4742 skb_reserve(nskb, header);
4743 memcpy(nskb->head, skb->head, header);
4744 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4745 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4746 __skb_queue_before(list, skb, nskb);
4747 skb_set_owner_r(nskb, sk);
4749 /* Copy data, releasing collapsed skbs. */
4751 int offset = start - TCP_SKB_CB(skb)->seq;
4752 int size = TCP_SKB_CB(skb)->end_seq - start;
4756 size = min(copy, size);
4757 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4759 TCP_SKB_CB(nskb)->end_seq += size;
4763 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4764 skb = tcp_collapse_one(sk, skb, list);
4767 tcp_hdr(skb)->syn ||
4775 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4776 * and tcp_collapse() them until all the queue is collapsed.
4778 static void tcp_collapse_ofo_queue(struct sock *sk)
4780 struct tcp_sock *tp = tcp_sk(sk);
4781 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4782 struct sk_buff *head;
4788 start = TCP_SKB_CB(skb)->seq;
4789 end = TCP_SKB_CB(skb)->end_seq;
4793 struct sk_buff *next = NULL;
4795 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4796 next = skb_queue_next(&tp->out_of_order_queue, skb);
4799 /* Segment is terminated when we see gap or when
4800 * we are at the end of all the queue. */
4802 after(TCP_SKB_CB(skb)->seq, end) ||
4803 before(TCP_SKB_CB(skb)->end_seq, start)) {
4804 tcp_collapse(sk, &tp->out_of_order_queue,
4805 head, skb, start, end);
4809 /* Start new segment */
4810 start = TCP_SKB_CB(skb)->seq;
4811 end = TCP_SKB_CB(skb)->end_seq;
4813 if (before(TCP_SKB_CB(skb)->seq, start))
4814 start = TCP_SKB_CB(skb)->seq;
4815 if (after(TCP_SKB_CB(skb)->end_seq, end))
4816 end = TCP_SKB_CB(skb)->end_seq;
4822 * Purge the out-of-order queue.
4823 * Return true if queue was pruned.
4825 static int tcp_prune_ofo_queue(struct sock *sk)
4827 struct tcp_sock *tp = tcp_sk(sk);
4830 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4831 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4832 __skb_queue_purge(&tp->out_of_order_queue);
4834 /* Reset SACK state. A conforming SACK implementation will
4835 * do the same at a timeout based retransmit. When a connection
4836 * is in a sad state like this, we care only about integrity
4837 * of the connection not performance.
4839 if (tp->rx_opt.sack_ok)
4840 tcp_sack_reset(&tp->rx_opt);
4847 /* Reduce allocated memory if we can, trying to get
4848 * the socket within its memory limits again.
4850 * Return less than zero if we should start dropping frames
4851 * until the socket owning process reads some of the data
4852 * to stabilize the situation.
4854 static int tcp_prune_queue(struct sock *sk)
4856 struct tcp_sock *tp = tcp_sk(sk);
4858 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4860 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4862 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4863 tcp_clamp_window(sk);
4864 else if (tcp_memory_pressure)
4865 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4867 tcp_collapse_ofo_queue(sk);
4868 if (!skb_queue_empty(&sk->sk_receive_queue))
4869 tcp_collapse(sk, &sk->sk_receive_queue,
4870 skb_peek(&sk->sk_receive_queue),
4872 tp->copied_seq, tp->rcv_nxt);
4875 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4878 /* Collapsing did not help, destructive actions follow.
4879 * This must not ever occur. */
4881 tcp_prune_ofo_queue(sk);
4883 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4886 /* If we are really being abused, tell the caller to silently
4887 * drop receive data on the floor. It will get retransmitted
4888 * and hopefully then we'll have sufficient space.
4890 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4892 /* Massive buffer overcommit. */
4897 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4898 * As additional protections, we do not touch cwnd in retransmission phases,
4899 * and if application hit its sndbuf limit recently.
4901 void tcp_cwnd_application_limited(struct sock *sk)
4903 struct tcp_sock *tp = tcp_sk(sk);
4905 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4906 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4907 /* Limited by application or receiver window. */
4908 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4909 u32 win_used = max(tp->snd_cwnd_used, init_win);
4910 if (win_used < tp->snd_cwnd) {
4911 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4912 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4914 tp->snd_cwnd_used = 0;
4916 tp->snd_cwnd_stamp = tcp_time_stamp;
4919 static int tcp_should_expand_sndbuf(struct sock *sk)
4921 struct tcp_sock *tp = tcp_sk(sk);
4923 /* If the user specified a specific send buffer setting, do
4926 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4929 /* If we are under global TCP memory pressure, do not expand. */
4930 if (tcp_memory_pressure)
4933 /* If we are under soft global TCP memory pressure, do not expand. */
4934 if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4937 /* If we filled the congestion window, do not expand. */
4938 if (tp->packets_out >= tp->snd_cwnd)
4944 /* When incoming ACK allowed to free some skb from write_queue,
4945 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4946 * on the exit from tcp input handler.
4948 * PROBLEM: sndbuf expansion does not work well with largesend.
4950 static void tcp_new_space(struct sock *sk)
4952 struct tcp_sock *tp = tcp_sk(sk);
4954 if (tcp_should_expand_sndbuf(sk)) {
4955 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
4956 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4957 int demanded = max_t(unsigned int, tp->snd_cwnd,
4958 tp->reordering + 1);
4959 sndmem *= 2 * demanded;
4960 if (sndmem > sk->sk_sndbuf)
4961 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4962 tp->snd_cwnd_stamp = tcp_time_stamp;
4965 sk->sk_write_space(sk);
4968 static void tcp_check_space(struct sock *sk)
4970 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4971 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4972 if (sk->sk_socket &&
4973 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4978 static inline void tcp_data_snd_check(struct sock *sk)
4980 tcp_push_pending_frames(sk);
4981 tcp_check_space(sk);
4985 * Check if sending an ack is needed.
4987 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4989 struct tcp_sock *tp = tcp_sk(sk);
4991 /* More than one full frame received... */
4992 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4993 /* ... and right edge of window advances far enough.
4994 * (tcp_recvmsg() will send ACK otherwise). Or...
4996 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4997 /* We ACK each frame or... */
4998 tcp_in_quickack_mode(sk) ||
4999 /* We have out of order data. */
5000 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5001 /* Then ack it now */
5004 /* Else, send delayed ack. */
5005 tcp_send_delayed_ack(sk);
5009 static inline void tcp_ack_snd_check(struct sock *sk)
5011 if (!inet_csk_ack_scheduled(sk)) {
5012 /* We sent a data segment already. */
5015 __tcp_ack_snd_check(sk, 1);
5019 * This routine is only called when we have urgent data
5020 * signaled. Its the 'slow' part of tcp_urg. It could be
5021 * moved inline now as tcp_urg is only called from one
5022 * place. We handle URGent data wrong. We have to - as
5023 * BSD still doesn't use the correction from RFC961.
5024 * For 1003.1g we should support a new option TCP_STDURG to permit
5025 * either form (or just set the sysctl tcp_stdurg).
5028 static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
5030 struct tcp_sock *tp = tcp_sk(sk);
5031 u32 ptr = ntohs(th->urg_ptr);
5033 if (ptr && !sysctl_tcp_stdurg)
5035 ptr += ntohl(th->seq);
5037 /* Ignore urgent data that we've already seen and read. */
5038 if (after(tp->copied_seq, ptr))
5041 /* Do not replay urg ptr.
5043 * NOTE: interesting situation not covered by specs.
5044 * Misbehaving sender may send urg ptr, pointing to segment,
5045 * which we already have in ofo queue. We are not able to fetch
5046 * such data and will stay in TCP_URG_NOTYET until will be eaten
5047 * by recvmsg(). Seems, we are not obliged to handle such wicked
5048 * situations. But it is worth to think about possibility of some
5049 * DoSes using some hypothetical application level deadlock.
5051 if (before(ptr, tp->rcv_nxt))
5054 /* Do we already have a newer (or duplicate) urgent pointer? */
5055 if (tp->urg_data && !after(ptr, tp->urg_seq))
5058 /* Tell the world about our new urgent pointer. */
5061 /* We may be adding urgent data when the last byte read was
5062 * urgent. To do this requires some care. We cannot just ignore
5063 * tp->copied_seq since we would read the last urgent byte again
5064 * as data, nor can we alter copied_seq until this data arrives
5065 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5067 * NOTE. Double Dutch. Rendering to plain English: author of comment
5068 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5069 * and expect that both A and B disappear from stream. This is _wrong_.
5070 * Though this happens in BSD with high probability, this is occasional.
5071 * Any application relying on this is buggy. Note also, that fix "works"
5072 * only in this artificial test. Insert some normal data between A and B and we will
5073 * decline of BSD again. Verdict: it is better to remove to trap
5076 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5077 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5078 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5080 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5081 __skb_unlink(skb, &sk->sk_receive_queue);
5086 tp->urg_data = TCP_URG_NOTYET;
5089 /* Disable header prediction. */
5093 /* This is the 'fast' part of urgent handling. */
5094 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
5096 struct tcp_sock *tp = tcp_sk(sk);
5098 /* Check if we get a new urgent pointer - normally not. */
5100 tcp_check_urg(sk, th);
5102 /* Do we wait for any urgent data? - normally not... */
5103 if (tp->urg_data == TCP_URG_NOTYET) {
5104 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5107 /* Is the urgent pointer pointing into this packet? */
5108 if (ptr < skb->len) {
5110 if (skb_copy_bits(skb, ptr, &tmp, 1))
5112 tp->urg_data = TCP_URG_VALID | tmp;
5113 if (!sock_flag(sk, SOCK_DEAD))
5114 sk->sk_data_ready(sk, 0);
5119 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5121 struct tcp_sock *tp = tcp_sk(sk);
5122 int chunk = skb->len - hlen;
5126 if (skb_csum_unnecessary(skb))
5127 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5129 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5133 tp->ucopy.len -= chunk;
5134 tp->copied_seq += chunk;
5135 tcp_rcv_space_adjust(sk);
5142 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5143 struct sk_buff *skb)
5147 if (sock_owned_by_user(sk)) {
5149 result = __tcp_checksum_complete(skb);
5152 result = __tcp_checksum_complete(skb);
5157 static inline int tcp_checksum_complete_user(struct sock *sk,
5158 struct sk_buff *skb)
5160 return !skb_csum_unnecessary(skb) &&
5161 __tcp_checksum_complete_user(sk, skb);
5164 #ifdef CONFIG_NET_DMA
5165 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5168 struct tcp_sock *tp = tcp_sk(sk);
5169 int chunk = skb->len - hlen;
5171 int copied_early = 0;
5173 if (tp->ucopy.wakeup)
5176 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5177 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5179 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5181 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5183 tp->ucopy.iov, chunk,
5184 tp->ucopy.pinned_list);
5189 tp->ucopy.dma_cookie = dma_cookie;
5192 tp->ucopy.len -= chunk;
5193 tp->copied_seq += chunk;
5194 tcp_rcv_space_adjust(sk);
5196 if ((tp->ucopy.len == 0) ||
5197 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5198 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5199 tp->ucopy.wakeup = 1;
5200 sk->sk_data_ready(sk, 0);
5202 } else if (chunk > 0) {
5203 tp->ucopy.wakeup = 1;
5204 sk->sk_data_ready(sk, 0);
5207 return copied_early;
5209 #endif /* CONFIG_NET_DMA */
5211 /* Does PAWS and seqno based validation of an incoming segment, flags will
5212 * play significant role here.
5214 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5215 struct tcphdr *th, int syn_inerr)
5218 struct tcp_sock *tp = tcp_sk(sk);
5220 /* RFC1323: H1. Apply PAWS check first. */
5221 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5222 tp->rx_opt.saw_tstamp &&
5223 tcp_paws_discard(sk, skb)) {
5225 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5226 tcp_send_dupack(sk, skb);
5229 /* Reset is accepted even if it did not pass PAWS. */
5232 /* Step 1: check sequence number */
5233 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5234 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5235 * (RST) segments are validated by checking their SEQ-fields."
5236 * And page 69: "If an incoming segment is not acceptable,
5237 * an acknowledgment should be sent in reply (unless the RST
5238 * bit is set, if so drop the segment and return)".
5241 tcp_send_dupack(sk, skb);
5245 /* Step 2: check RST bit */
5251 /* ts_recent update must be made after we are sure that the packet
5254 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5256 /* step 3: check security and precedence [ignored] */
5258 /* step 4: Check for a SYN in window. */
5259 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5261 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5262 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5275 * TCP receive function for the ESTABLISHED state.
5277 * It is split into a fast path and a slow path. The fast path is
5279 * - A zero window was announced from us - zero window probing
5280 * is only handled properly in the slow path.
5281 * - Out of order segments arrived.
5282 * - Urgent data is expected.
5283 * - There is no buffer space left
5284 * - Unexpected TCP flags/window values/header lengths are received
5285 * (detected by checking the TCP header against pred_flags)
5286 * - Data is sent in both directions. Fast path only supports pure senders
5287 * or pure receivers (this means either the sequence number or the ack
5288 * value must stay constant)
5289 * - Unexpected TCP option.
5291 * When these conditions are not satisfied it drops into a standard
5292 * receive procedure patterned after RFC793 to handle all cases.
5293 * The first three cases are guaranteed by proper pred_flags setting,
5294 * the rest is checked inline. Fast processing is turned on in
5295 * tcp_data_queue when everything is OK.
5297 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5298 struct tcphdr *th, unsigned len)
5300 struct tcp_sock *tp = tcp_sk(sk);
5304 * Header prediction.
5305 * The code loosely follows the one in the famous
5306 * "30 instruction TCP receive" Van Jacobson mail.
5308 * Van's trick is to deposit buffers into socket queue
5309 * on a device interrupt, to call tcp_recv function
5310 * on the receive process context and checksum and copy
5311 * the buffer to user space. smart...
5313 * Our current scheme is not silly either but we take the
5314 * extra cost of the net_bh soft interrupt processing...
5315 * We do checksum and copy also but from device to kernel.
5318 tp->rx_opt.saw_tstamp = 0;
5320 /* pred_flags is 0xS?10 << 16 + snd_wnd
5321 * if header_prediction is to be made
5322 * 'S' will always be tp->tcp_header_len >> 2
5323 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5324 * turn it off (when there are holes in the receive
5325 * space for instance)
5326 * PSH flag is ignored.
5329 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5330 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5331 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5332 int tcp_header_len = tp->tcp_header_len;
5334 /* Timestamp header prediction: tcp_header_len
5335 * is automatically equal to th->doff*4 due to pred_flags
5339 /* Check timestamp */
5340 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5341 /* No? Slow path! */
5342 if (!tcp_parse_aligned_timestamp(tp, th))
5345 /* If PAWS failed, check it more carefully in slow path */
5346 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5349 /* DO NOT update ts_recent here, if checksum fails
5350 * and timestamp was corrupted part, it will result
5351 * in a hung connection since we will drop all
5352 * future packets due to the PAWS test.
5356 if (len <= tcp_header_len) {
5357 /* Bulk data transfer: sender */
5358 if (len == tcp_header_len) {
5359 /* Predicted packet is in window by definition.
5360 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5361 * Hence, check seq<=rcv_wup reduces to:
5363 if (tcp_header_len ==
5364 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5365 tp->rcv_nxt == tp->rcv_wup)
5366 tcp_store_ts_recent(tp);
5368 /* We know that such packets are checksummed
5371 tcp_ack(sk, skb, 0);
5373 tcp_data_snd_check(sk);
5375 } else { /* Header too small */
5376 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5381 int copied_early = 0;
5383 if (tp->copied_seq == tp->rcv_nxt &&
5384 len - tcp_header_len <= tp->ucopy.len) {
5385 #ifdef CONFIG_NET_DMA
5386 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5391 if (tp->ucopy.task == current &&
5392 sock_owned_by_user(sk) && !copied_early) {
5393 __set_current_state(TASK_RUNNING);
5395 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5399 /* Predicted packet is in window by definition.
5400 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5401 * Hence, check seq<=rcv_wup reduces to:
5403 if (tcp_header_len ==
5404 (sizeof(struct tcphdr) +
5405 TCPOLEN_TSTAMP_ALIGNED) &&
5406 tp->rcv_nxt == tp->rcv_wup)
5407 tcp_store_ts_recent(tp);
5409 tcp_rcv_rtt_measure_ts(sk, skb);
5411 __skb_pull(skb, tcp_header_len);
5412 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5413 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5416 tcp_cleanup_rbuf(sk, skb->len);
5419 if (tcp_checksum_complete_user(sk, skb))
5422 /* Predicted packet is in window by definition.
5423 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5424 * Hence, check seq<=rcv_wup reduces to:
5426 if (tcp_header_len ==
5427 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5428 tp->rcv_nxt == tp->rcv_wup)
5429 tcp_store_ts_recent(tp);
5431 tcp_rcv_rtt_measure_ts(sk, skb);
5433 if ((int)skb->truesize > sk->sk_forward_alloc)
5436 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5438 /* Bulk data transfer: receiver */
5439 __skb_pull(skb, tcp_header_len);
5440 __skb_queue_tail(&sk->sk_receive_queue, skb);
5441 skb_set_owner_r(skb, sk);
5442 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5445 tcp_event_data_recv(sk, skb);
5447 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5448 /* Well, only one small jumplet in fast path... */
5449 tcp_ack(sk, skb, FLAG_DATA);
5450 tcp_data_snd_check(sk);
5451 if (!inet_csk_ack_scheduled(sk))
5455 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5456 __tcp_ack_snd_check(sk, 0);
5458 #ifdef CONFIG_NET_DMA
5460 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5466 sk->sk_data_ready(sk, 0);
5472 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5476 * Standard slow path.
5479 res = tcp_validate_incoming(sk, skb, th, 1);
5484 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5487 tcp_rcv_rtt_measure_ts(sk, skb);
5489 /* Process urgent data. */
5490 tcp_urg(sk, skb, th);
5492 /* step 7: process the segment text */
5493 tcp_data_queue(sk, skb);
5495 tcp_data_snd_check(sk);
5496 tcp_ack_snd_check(sk);
5500 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5506 EXPORT_SYMBOL(tcp_rcv_established);
5508 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5509 struct tcphdr *th, unsigned len)
5512 struct inet_connection_sock *icsk = inet_csk(sk);
5513 struct tcp_sock *tp = tcp_sk(sk);
5514 struct tcp_cookie_values *cvp = tp->cookie_values;
5515 int saved_clamp = tp->rx_opt.mss_clamp;
5517 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5521 * "If the state is SYN-SENT then
5522 * first check the ACK bit
5523 * If the ACK bit is set
5524 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5525 * a reset (unless the RST bit is set, if so drop
5526 * the segment and return)"
5528 * We do not send data with SYN, so that RFC-correct
5531 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5532 goto reset_and_undo;
5534 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5535 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5537 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5538 goto reset_and_undo;
5541 /* Now ACK is acceptable.
5543 * "If the RST bit is set
5544 * If the ACK was acceptable then signal the user "error:
5545 * connection reset", drop the segment, enter CLOSED state,
5546 * delete TCB, and return."
5555 * "fifth, if neither of the SYN or RST bits is set then
5556 * drop the segment and return."
5562 goto discard_and_undo;
5565 * "If the SYN bit is on ...
5566 * are acceptable then ...
5567 * (our SYN has been ACKed), change the connection
5568 * state to ESTABLISHED..."
5571 TCP_ECN_rcv_synack(tp, th);
5573 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5574 tcp_ack(sk, skb, FLAG_SLOWPATH);
5576 /* Ok.. it's good. Set up sequence numbers and
5577 * move to established.
5579 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5580 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5582 /* RFC1323: The window in SYN & SYN/ACK segments is
5585 tp->snd_wnd = ntohs(th->window);
5586 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5588 if (!tp->rx_opt.wscale_ok) {
5589 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5590 tp->window_clamp = min(tp->window_clamp, 65535U);
5593 if (tp->rx_opt.saw_tstamp) {
5594 tp->rx_opt.tstamp_ok = 1;
5595 tp->tcp_header_len =
5596 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5597 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5598 tcp_store_ts_recent(tp);
5600 tp->tcp_header_len = sizeof(struct tcphdr);
5603 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5604 tcp_enable_fack(tp);
5607 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5608 tcp_initialize_rcv_mss(sk);
5610 /* Remember, tcp_poll() does not lock socket!
5611 * Change state from SYN-SENT only after copied_seq
5612 * is initialized. */
5613 tp->copied_seq = tp->rcv_nxt;
5616 cvp->cookie_pair_size > 0 &&
5617 tp->rx_opt.cookie_plus > 0) {
5618 int cookie_size = tp->rx_opt.cookie_plus
5619 - TCPOLEN_COOKIE_BASE;
5620 int cookie_pair_size = cookie_size
5621 + cvp->cookie_desired;
5623 /* A cookie extension option was sent and returned.
5624 * Note that each incoming SYNACK replaces the
5625 * Responder cookie. The initial exchange is most
5626 * fragile, as protection against spoofing relies
5627 * entirely upon the sequence and timestamp (above).
5628 * This replacement strategy allows the correct pair to
5629 * pass through, while any others will be filtered via
5630 * Responder verification later.
5632 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5633 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5634 hash_location, cookie_size);
5635 cvp->cookie_pair_size = cookie_pair_size;
5640 tcp_set_state(sk, TCP_ESTABLISHED);
5642 security_inet_conn_established(sk, skb);
5644 /* Make sure socket is routed, for correct metrics. */
5645 icsk->icsk_af_ops->rebuild_header(sk);
5647 tcp_init_metrics(sk);
5649 tcp_init_congestion_control(sk);
5651 /* Prevent spurious tcp_cwnd_restart() on first data
5654 tp->lsndtime = tcp_time_stamp;
5656 tcp_init_buffer_space(sk);
5658 if (sock_flag(sk, SOCK_KEEPOPEN))
5659 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5661 if (!tp->rx_opt.snd_wscale)
5662 __tcp_fast_path_on(tp, tp->snd_wnd);
5666 if (!sock_flag(sk, SOCK_DEAD)) {
5667 sk->sk_state_change(sk);
5668 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5671 if (sk->sk_write_pending ||
5672 icsk->icsk_accept_queue.rskq_defer_accept ||
5673 icsk->icsk_ack.pingpong) {
5674 /* Save one ACK. Data will be ready after
5675 * several ticks, if write_pending is set.
5677 * It may be deleted, but with this feature tcpdumps
5678 * look so _wonderfully_ clever, that I was not able
5679 * to stand against the temptation 8) --ANK
5681 inet_csk_schedule_ack(sk);
5682 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5683 icsk->icsk_ack.ato = TCP_ATO_MIN;
5684 tcp_incr_quickack(sk);
5685 tcp_enter_quickack_mode(sk);
5686 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5687 TCP_DELACK_MAX, TCP_RTO_MAX);
5698 /* No ACK in the segment */
5702 * "If the RST bit is set
5704 * Otherwise (no ACK) drop the segment and return."
5707 goto discard_and_undo;
5711 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5712 tcp_paws_reject(&tp->rx_opt, 0))
5713 goto discard_and_undo;
5716 /* We see SYN without ACK. It is attempt of
5717 * simultaneous connect with crossed SYNs.
5718 * Particularly, it can be connect to self.
5720 tcp_set_state(sk, TCP_SYN_RECV);
5722 if (tp->rx_opt.saw_tstamp) {
5723 tp->rx_opt.tstamp_ok = 1;
5724 tcp_store_ts_recent(tp);
5725 tp->tcp_header_len =
5726 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5728 tp->tcp_header_len = sizeof(struct tcphdr);
5731 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5732 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5734 /* RFC1323: The window in SYN & SYN/ACK segments is
5737 tp->snd_wnd = ntohs(th->window);
5738 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5739 tp->max_window = tp->snd_wnd;
5741 TCP_ECN_rcv_syn(tp, th);
5744 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5745 tcp_initialize_rcv_mss(sk);
5747 tcp_send_synack(sk);
5749 /* Note, we could accept data and URG from this segment.
5750 * There are no obstacles to make this.
5752 * However, if we ignore data in ACKless segments sometimes,
5753 * we have no reasons to accept it sometimes.
5754 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5755 * is not flawless. So, discard packet for sanity.
5756 * Uncomment this return to process the data.
5763 /* "fifth, if neither of the SYN or RST bits is set then
5764 * drop the segment and return."
5768 tcp_clear_options(&tp->rx_opt);
5769 tp->rx_opt.mss_clamp = saved_clamp;
5773 tcp_clear_options(&tp->rx_opt);
5774 tp->rx_opt.mss_clamp = saved_clamp;
5779 * This function implements the receiving procedure of RFC 793 for
5780 * all states except ESTABLISHED and TIME_WAIT.
5781 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5782 * address independent.
5785 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5786 struct tcphdr *th, unsigned len)
5788 struct tcp_sock *tp = tcp_sk(sk);
5789 struct inet_connection_sock *icsk = inet_csk(sk);
5793 tp->rx_opt.saw_tstamp = 0;
5795 switch (sk->sk_state) {
5807 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5810 /* Now we have several options: In theory there is
5811 * nothing else in the frame. KA9Q has an option to
5812 * send data with the syn, BSD accepts data with the
5813 * syn up to the [to be] advertised window and
5814 * Solaris 2.1 gives you a protocol error. For now
5815 * we just ignore it, that fits the spec precisely
5816 * and avoids incompatibilities. It would be nice in
5817 * future to drop through and process the data.
5819 * Now that TTCP is starting to be used we ought to
5821 * But, this leaves one open to an easy denial of
5822 * service attack, and SYN cookies can't defend
5823 * against this problem. So, we drop the data
5824 * in the interest of security over speed unless
5825 * it's still in use.
5833 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5837 /* Do step6 onward by hand. */
5838 tcp_urg(sk, skb, th);
5840 tcp_data_snd_check(sk);
5844 res = tcp_validate_incoming(sk, skb, th, 0);
5848 /* step 5: check the ACK field */
5850 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5852 switch (sk->sk_state) {
5855 tp->copied_seq = tp->rcv_nxt;
5857 tcp_set_state(sk, TCP_ESTABLISHED);
5858 sk->sk_state_change(sk);
5860 /* Note, that this wakeup is only for marginal
5861 * crossed SYN case. Passively open sockets
5862 * are not waked up, because sk->sk_sleep ==
5863 * NULL and sk->sk_socket == NULL.
5867 SOCK_WAKE_IO, POLL_OUT);
5869 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5870 tp->snd_wnd = ntohs(th->window) <<
5871 tp->rx_opt.snd_wscale;
5872 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5874 if (tp->rx_opt.tstamp_ok)
5875 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5877 /* Make sure socket is routed, for
5880 icsk->icsk_af_ops->rebuild_header(sk);
5882 tcp_init_metrics(sk);
5884 tcp_init_congestion_control(sk);
5886 /* Prevent spurious tcp_cwnd_restart() on
5887 * first data packet.
5889 tp->lsndtime = tcp_time_stamp;
5892 tcp_initialize_rcv_mss(sk);
5893 tcp_init_buffer_space(sk);
5894 tcp_fast_path_on(tp);
5901 if (tp->snd_una == tp->write_seq) {
5902 tcp_set_state(sk, TCP_FIN_WAIT2);
5903 sk->sk_shutdown |= SEND_SHUTDOWN;
5904 dst_confirm(__sk_dst_get(sk));
5906 if (!sock_flag(sk, SOCK_DEAD))
5907 /* Wake up lingering close() */
5908 sk->sk_state_change(sk);
5912 if (tp->linger2 < 0 ||
5913 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5914 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5916 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5920 tmo = tcp_fin_time(sk);
5921 if (tmo > TCP_TIMEWAIT_LEN) {
5922 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5923 } else if (th->fin || sock_owned_by_user(sk)) {
5924 /* Bad case. We could lose such FIN otherwise.
5925 * It is not a big problem, but it looks confusing
5926 * and not so rare event. We still can lose it now,
5927 * if it spins in bh_lock_sock(), but it is really
5930 inet_csk_reset_keepalive_timer(sk, tmo);
5932 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5940 if (tp->snd_una == tp->write_seq) {
5941 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5947 if (tp->snd_una == tp->write_seq) {
5948 tcp_update_metrics(sk);
5957 /* step 6: check the URG bit */
5958 tcp_urg(sk, skb, th);
5960 /* step 7: process the segment text */
5961 switch (sk->sk_state) {
5962 case TCP_CLOSE_WAIT:
5965 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5969 /* RFC 793 says to queue data in these states,
5970 * RFC 1122 says we MUST send a reset.
5971 * BSD 4.4 also does reset.
5973 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5974 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5975 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5976 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5982 case TCP_ESTABLISHED:
5983 tcp_data_queue(sk, skb);
5988 /* tcp_data could move socket to TIME-WAIT */
5989 if (sk->sk_state != TCP_CLOSE) {
5990 tcp_data_snd_check(sk);
5991 tcp_ack_snd_check(sk);
6000 EXPORT_SYMBOL(tcp_rcv_state_process);