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, const 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, const 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, const 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(const struct tcp_sock *tp, const 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 = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
270 sndmem *= TCP_INIT_CWND;
271 if (sk->sk_sndbuf < sndmem)
272 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
275 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
277 * All tcp_full_space() is split to two parts: "network" buffer, allocated
278 * forward and advertised in receiver window (tp->rcv_wnd) and
279 * "application buffer", required to isolate scheduling/application
280 * latencies from network.
281 * window_clamp is maximal advertised window. It can be less than
282 * tcp_full_space(), in this case tcp_full_space() - window_clamp
283 * is reserved for "application" buffer. The less window_clamp is
284 * the smoother our behaviour from viewpoint of network, but the lower
285 * throughput and the higher sensitivity of the connection to losses. 8)
287 * rcv_ssthresh is more strict window_clamp used at "slow start"
288 * phase to predict further behaviour of this connection.
289 * It is used for two goals:
290 * - to enforce header prediction at sender, even when application
291 * requires some significant "application buffer". It is check #1.
292 * - to prevent pruning of receive queue because of misprediction
293 * of receiver window. Check #2.
295 * The scheme does not work when sender sends good segments opening
296 * window and then starts to feed us spaghetti. But it should work
297 * in common situations. Otherwise, we have to rely on queue collapsing.
300 /* Slow part of check#2. */
301 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
303 struct tcp_sock *tp = tcp_sk(sk);
305 int truesize = tcp_win_from_space(skb->truesize) >> 1;
306 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
308 while (tp->rcv_ssthresh <= window) {
309 if (truesize <= skb->len)
310 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
318 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
320 struct tcp_sock *tp = tcp_sk(sk);
323 if (tp->rcv_ssthresh < tp->window_clamp &&
324 (int)tp->rcv_ssthresh < tcp_space(sk) &&
325 !tcp_memory_pressure) {
328 /* Check #2. Increase window, if skb with such overhead
329 * will fit to rcvbuf in future.
331 if (tcp_win_from_space(skb->truesize) <= skb->len)
332 incr = 2 * tp->advmss;
334 incr = __tcp_grow_window(sk, skb);
337 incr = max_t(int, incr, 2 * skb->len);
338 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
340 inet_csk(sk)->icsk_ack.quick |= 1;
345 /* 3. Tuning rcvbuf, when connection enters established state. */
347 static void tcp_fixup_rcvbuf(struct sock *sk)
349 u32 mss = tcp_sk(sk)->advmss;
350 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
353 /* Limit to 10 segments if mss <= 1460,
354 * or 14600/mss segments, with a minimum of two segments.
357 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
359 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
360 while (tcp_win_from_space(rcvmem) < mss)
365 if (sk->sk_rcvbuf < rcvmem)
366 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
369 /* 4. Try to fixup all. It is made immediately after connection enters
372 static void tcp_init_buffer_space(struct sock *sk)
374 struct tcp_sock *tp = tcp_sk(sk);
377 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
378 tcp_fixup_rcvbuf(sk);
379 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
380 tcp_fixup_sndbuf(sk);
382 tp->rcvq_space.space = tp->rcv_wnd;
384 maxwin = tcp_full_space(sk);
386 if (tp->window_clamp >= maxwin) {
387 tp->window_clamp = maxwin;
389 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
390 tp->window_clamp = max(maxwin -
391 (maxwin >> sysctl_tcp_app_win),
395 /* Force reservation of one segment. */
396 if (sysctl_tcp_app_win &&
397 tp->window_clamp > 2 * tp->advmss &&
398 tp->window_clamp + tp->advmss > maxwin)
399 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
401 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
402 tp->snd_cwnd_stamp = tcp_time_stamp;
405 /* 5. Recalculate window clamp after socket hit its memory bounds. */
406 static void tcp_clamp_window(struct sock *sk)
408 struct tcp_sock *tp = tcp_sk(sk);
409 struct inet_connection_sock *icsk = inet_csk(sk);
411 icsk->icsk_ack.quick = 0;
413 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
414 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
415 !tcp_memory_pressure &&
416 atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
417 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
420 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
421 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
424 /* Initialize RCV_MSS value.
425 * RCV_MSS is an our guess about MSS used by the peer.
426 * We haven't any direct information about the MSS.
427 * It's better to underestimate the RCV_MSS rather than overestimate.
428 * Overestimations make us ACKing less frequently than needed.
429 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
431 void tcp_initialize_rcv_mss(struct sock *sk)
433 const struct tcp_sock *tp = tcp_sk(sk);
434 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
436 hint = min(hint, tp->rcv_wnd / 2);
437 hint = min(hint, TCP_MSS_DEFAULT);
438 hint = max(hint, TCP_MIN_MSS);
440 inet_csk(sk)->icsk_ack.rcv_mss = hint;
442 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
444 /* Receiver "autotuning" code.
446 * The algorithm for RTT estimation w/o timestamps is based on
447 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
448 * <http://public.lanl.gov/radiant/pubs.html#DRS>
450 * More detail on this code can be found at
451 * <http://staff.psc.edu/jheffner/>,
452 * though this reference is out of date. A new paper
455 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
457 u32 new_sample = tp->rcv_rtt_est.rtt;
463 if (new_sample != 0) {
464 /* If we sample in larger samples in the non-timestamp
465 * case, we could grossly overestimate the RTT especially
466 * with chatty applications or bulk transfer apps which
467 * are stalled on filesystem I/O.
469 * Also, since we are only going for a minimum in the
470 * non-timestamp case, we do not smooth things out
471 * else with timestamps disabled convergence takes too
475 m -= (new_sample >> 3);
483 /* No previous measure. */
487 if (tp->rcv_rtt_est.rtt != new_sample)
488 tp->rcv_rtt_est.rtt = new_sample;
491 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
493 if (tp->rcv_rtt_est.time == 0)
495 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
497 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
500 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
501 tp->rcv_rtt_est.time = tcp_time_stamp;
504 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
505 const struct sk_buff *skb)
507 struct tcp_sock *tp = tcp_sk(sk);
508 if (tp->rx_opt.rcv_tsecr &&
509 (TCP_SKB_CB(skb)->end_seq -
510 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
511 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
515 * This function should be called every time data is copied to user space.
516 * It calculates the appropriate TCP receive buffer space.
518 void tcp_rcv_space_adjust(struct sock *sk)
520 struct tcp_sock *tp = tcp_sk(sk);
524 if (tp->rcvq_space.time == 0)
527 time = tcp_time_stamp - tp->rcvq_space.time;
528 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
531 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
533 space = max(tp->rcvq_space.space, space);
535 if (tp->rcvq_space.space != space) {
538 tp->rcvq_space.space = space;
540 if (sysctl_tcp_moderate_rcvbuf &&
541 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
542 int new_clamp = space;
544 /* Receive space grows, normalize in order to
545 * take into account packet headers and sk_buff
546 * structure overhead.
551 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
552 while (tcp_win_from_space(rcvmem) < tp->advmss)
555 space = min(space, sysctl_tcp_rmem[2]);
556 if (space > sk->sk_rcvbuf) {
557 sk->sk_rcvbuf = space;
559 /* Make the window clamp follow along. */
560 tp->window_clamp = new_clamp;
566 tp->rcvq_space.seq = tp->copied_seq;
567 tp->rcvq_space.time = tcp_time_stamp;
570 /* There is something which you must keep in mind when you analyze the
571 * behavior of the tp->ato delayed ack timeout interval. When a
572 * connection starts up, we want to ack as quickly as possible. The
573 * problem is that "good" TCP's do slow start at the beginning of data
574 * transmission. The means that until we send the first few ACK's the
575 * sender will sit on his end and only queue most of his data, because
576 * he can only send snd_cwnd unacked packets at any given time. For
577 * each ACK we send, he increments snd_cwnd and transmits more of his
580 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
582 struct tcp_sock *tp = tcp_sk(sk);
583 struct inet_connection_sock *icsk = inet_csk(sk);
586 inet_csk_schedule_ack(sk);
588 tcp_measure_rcv_mss(sk, skb);
590 tcp_rcv_rtt_measure(tp);
592 now = tcp_time_stamp;
594 if (!icsk->icsk_ack.ato) {
595 /* The _first_ data packet received, initialize
596 * delayed ACK engine.
598 tcp_incr_quickack(sk);
599 icsk->icsk_ack.ato = TCP_ATO_MIN;
601 int m = now - icsk->icsk_ack.lrcvtime;
603 if (m <= TCP_ATO_MIN / 2) {
604 /* The fastest case is the first. */
605 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
606 } else if (m < icsk->icsk_ack.ato) {
607 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
608 if (icsk->icsk_ack.ato > icsk->icsk_rto)
609 icsk->icsk_ack.ato = icsk->icsk_rto;
610 } else if (m > icsk->icsk_rto) {
611 /* Too long gap. Apparently sender failed to
612 * restart window, so that we send ACKs quickly.
614 tcp_incr_quickack(sk);
618 icsk->icsk_ack.lrcvtime = now;
620 TCP_ECN_check_ce(tp, skb);
623 tcp_grow_window(sk, skb);
626 /* Called to compute a smoothed rtt estimate. The data fed to this
627 * routine either comes from timestamps, or from segments that were
628 * known _not_ to have been retransmitted [see Karn/Partridge
629 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
630 * piece by Van Jacobson.
631 * NOTE: the next three routines used to be one big routine.
632 * To save cycles in the RFC 1323 implementation it was better to break
633 * it up into three procedures. -- erics
635 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
637 struct tcp_sock *tp = tcp_sk(sk);
638 long m = mrtt; /* RTT */
640 /* The following amusing code comes from Jacobson's
641 * article in SIGCOMM '88. Note that rtt and mdev
642 * are scaled versions of rtt and mean deviation.
643 * This is designed to be as fast as possible
644 * m stands for "measurement".
646 * On a 1990 paper the rto value is changed to:
647 * RTO = rtt + 4 * mdev
649 * Funny. This algorithm seems to be very broken.
650 * These formulae increase RTO, when it should be decreased, increase
651 * too slowly, when it should be increased quickly, decrease too quickly
652 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
653 * does not matter how to _calculate_ it. Seems, it was trap
654 * that VJ failed to avoid. 8)
659 m -= (tp->srtt >> 3); /* m is now error in rtt est */
660 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
662 m = -m; /* m is now abs(error) */
663 m -= (tp->mdev >> 2); /* similar update on mdev */
664 /* This is similar to one of Eifel findings.
665 * Eifel blocks mdev updates when rtt decreases.
666 * This solution is a bit different: we use finer gain
667 * for mdev in this case (alpha*beta).
668 * Like Eifel it also prevents growth of rto,
669 * but also it limits too fast rto decreases,
670 * happening in pure Eifel.
675 m -= (tp->mdev >> 2); /* similar update on mdev */
677 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
678 if (tp->mdev > tp->mdev_max) {
679 tp->mdev_max = tp->mdev;
680 if (tp->mdev_max > tp->rttvar)
681 tp->rttvar = tp->mdev_max;
683 if (after(tp->snd_una, tp->rtt_seq)) {
684 if (tp->mdev_max < tp->rttvar)
685 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
686 tp->rtt_seq = tp->snd_nxt;
687 tp->mdev_max = tcp_rto_min(sk);
690 /* no previous measure. */
691 tp->srtt = m << 3; /* take the measured time to be rtt */
692 tp->mdev = m << 1; /* make sure rto = 3*rtt */
693 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
694 tp->rtt_seq = tp->snd_nxt;
698 /* Calculate rto without backoff. This is the second half of Van Jacobson's
699 * routine referred to above.
701 static inline void tcp_set_rto(struct sock *sk)
703 const struct tcp_sock *tp = tcp_sk(sk);
704 /* Old crap is replaced with new one. 8)
707 * 1. If rtt variance happened to be less 50msec, it is hallucination.
708 * It cannot be less due to utterly erratic ACK generation made
709 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
710 * to do with delayed acks, because at cwnd>2 true delack timeout
711 * is invisible. Actually, Linux-2.4 also generates erratic
712 * ACKs in some circumstances.
714 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
716 /* 2. Fixups made earlier cannot be right.
717 * If we do not estimate RTO correctly without them,
718 * all the algo is pure shit and should be replaced
719 * with correct one. It is exactly, which we pretend to do.
722 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
723 * guarantees that rto is higher.
728 /* Save metrics learned by this TCP session.
729 This function is called only, when TCP finishes successfully
730 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
732 void tcp_update_metrics(struct sock *sk)
734 struct tcp_sock *tp = tcp_sk(sk);
735 struct dst_entry *dst = __sk_dst_get(sk);
737 if (sysctl_tcp_nometrics_save)
742 if (dst && (dst->flags & DST_HOST)) {
743 const struct inet_connection_sock *icsk = inet_csk(sk);
747 if (icsk->icsk_backoff || !tp->srtt) {
748 /* This session failed to estimate rtt. Why?
749 * Probably, no packets returned in time.
752 if (!(dst_metric_locked(dst, RTAX_RTT)))
753 dst_metric_set(dst, RTAX_RTT, 0);
757 rtt = dst_metric_rtt(dst, RTAX_RTT);
760 /* If newly calculated rtt larger than stored one,
761 * store new one. Otherwise, use EWMA. Remember,
762 * rtt overestimation is always better than underestimation.
764 if (!(dst_metric_locked(dst, RTAX_RTT))) {
766 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
768 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
771 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
776 /* Scale deviation to rttvar fixed point */
781 var = dst_metric_rtt(dst, RTAX_RTTVAR);
785 var -= (var - m) >> 2;
787 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
790 if (tcp_in_initial_slowstart(tp)) {
791 /* Slow start still did not finish. */
792 if (dst_metric(dst, RTAX_SSTHRESH) &&
793 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
794 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
795 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
796 if (!dst_metric_locked(dst, RTAX_CWND) &&
797 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
798 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
799 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
800 icsk->icsk_ca_state == TCP_CA_Open) {
801 /* Cong. avoidance phase, cwnd is reliable. */
802 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
803 dst_metric_set(dst, RTAX_SSTHRESH,
804 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
805 if (!dst_metric_locked(dst, RTAX_CWND))
806 dst_metric_set(dst, RTAX_CWND,
807 (dst_metric(dst, RTAX_CWND) +
810 /* Else slow start did not finish, cwnd is non-sense,
811 ssthresh may be also invalid.
813 if (!dst_metric_locked(dst, RTAX_CWND))
814 dst_metric_set(dst, RTAX_CWND,
815 (dst_metric(dst, RTAX_CWND) +
816 tp->snd_ssthresh) >> 1);
817 if (dst_metric(dst, RTAX_SSTHRESH) &&
818 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
819 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
820 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
823 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
824 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
825 tp->reordering != sysctl_tcp_reordering)
826 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
831 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
833 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
836 cwnd = TCP_INIT_CWND;
837 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
840 /* Set slow start threshold and cwnd not falling to slow start */
841 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
843 struct tcp_sock *tp = tcp_sk(sk);
844 const struct inet_connection_sock *icsk = inet_csk(sk);
846 tp->prior_ssthresh = 0;
848 if (icsk->icsk_ca_state < TCP_CA_CWR) {
851 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
852 tp->snd_cwnd = min(tp->snd_cwnd,
853 tcp_packets_in_flight(tp) + 1U);
854 tp->snd_cwnd_cnt = 0;
855 tp->high_seq = tp->snd_nxt;
856 tp->snd_cwnd_stamp = tcp_time_stamp;
857 TCP_ECN_queue_cwr(tp);
859 tcp_set_ca_state(sk, TCP_CA_CWR);
864 * Packet counting of FACK is based on in-order assumptions, therefore TCP
865 * disables it when reordering is detected
867 static void tcp_disable_fack(struct tcp_sock *tp)
869 /* RFC3517 uses different metric in lost marker => reset on change */
871 tp->lost_skb_hint = NULL;
872 tp->rx_opt.sack_ok &= ~2;
875 /* Take a notice that peer is sending D-SACKs */
876 static void tcp_dsack_seen(struct tcp_sock *tp)
878 tp->rx_opt.sack_ok |= 4;
881 /* Initialize metrics on socket. */
883 static void tcp_init_metrics(struct sock *sk)
885 struct tcp_sock *tp = tcp_sk(sk);
886 struct dst_entry *dst = __sk_dst_get(sk);
893 if (dst_metric_locked(dst, RTAX_CWND))
894 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
895 if (dst_metric(dst, RTAX_SSTHRESH)) {
896 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
897 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
898 tp->snd_ssthresh = tp->snd_cwnd_clamp;
900 /* ssthresh may have been reduced unnecessarily during.
901 * 3WHS. Restore it back to its initial default.
903 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
905 if (dst_metric(dst, RTAX_REORDERING) &&
906 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
907 tcp_disable_fack(tp);
908 tp->reordering = dst_metric(dst, RTAX_REORDERING);
911 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
914 /* Initial rtt is determined from SYN,SYN-ACK.
915 * The segment is small and rtt may appear much
916 * less than real one. Use per-dst memory
917 * to make it more realistic.
919 * A bit of theory. RTT is time passed after "normal" sized packet
920 * is sent until it is ACKed. In normal circumstances sending small
921 * packets force peer to delay ACKs and calculation is correct too.
922 * The algorithm is adaptive and, provided we follow specs, it
923 * NEVER underestimate RTT. BUT! If peer tries to make some clever
924 * tricks sort of "quick acks" for time long enough to decrease RTT
925 * to low value, and then abruptly stops to do it and starts to delay
926 * ACKs, wait for troubles.
928 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
929 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
930 tp->rtt_seq = tp->snd_nxt;
932 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
933 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
934 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
939 /* RFC2988bis: We've failed to get a valid RTT sample from
940 * 3WHS. This is most likely due to retransmission,
941 * including spurious one. Reset the RTO back to 3secs
942 * from the more aggressive 1sec to avoid more spurious
945 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
946 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
948 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
949 * retransmitted. In light of RFC2988bis' more aggressive 1sec
950 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
951 * retransmission has occurred.
953 if (tp->total_retrans > 1)
956 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
957 tp->snd_cwnd_stamp = tcp_time_stamp;
960 static void tcp_update_reordering(struct sock *sk, const int metric,
963 struct tcp_sock *tp = tcp_sk(sk);
964 if (metric > tp->reordering) {
967 tp->reordering = min(TCP_MAX_REORDERING, metric);
969 /* This exciting event is worth to be remembered. 8) */
971 mib_idx = LINUX_MIB_TCPTSREORDER;
972 else if (tcp_is_reno(tp))
973 mib_idx = LINUX_MIB_TCPRENOREORDER;
974 else if (tcp_is_fack(tp))
975 mib_idx = LINUX_MIB_TCPFACKREORDER;
977 mib_idx = LINUX_MIB_TCPSACKREORDER;
979 NET_INC_STATS_BH(sock_net(sk), mib_idx);
980 #if FASTRETRANS_DEBUG > 1
981 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
982 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
986 tp->undo_marker ? tp->undo_retrans : 0);
988 tcp_disable_fack(tp);
992 /* This must be called before lost_out is incremented */
993 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
995 if ((tp->retransmit_skb_hint == NULL) ||
996 before(TCP_SKB_CB(skb)->seq,
997 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
998 tp->retransmit_skb_hint = skb;
1000 if (!tp->lost_out ||
1001 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
1002 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1005 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1007 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1008 tcp_verify_retransmit_hint(tp, skb);
1010 tp->lost_out += tcp_skb_pcount(skb);
1011 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1015 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1016 struct sk_buff *skb)
1018 tcp_verify_retransmit_hint(tp, skb);
1020 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1021 tp->lost_out += tcp_skb_pcount(skb);
1022 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1026 /* This procedure tags the retransmission queue when SACKs arrive.
1028 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1029 * Packets in queue with these bits set are counted in variables
1030 * sacked_out, retrans_out and lost_out, correspondingly.
1032 * Valid combinations are:
1033 * Tag InFlight Description
1034 * 0 1 - orig segment is in flight.
1035 * S 0 - nothing flies, orig reached receiver.
1036 * L 0 - nothing flies, orig lost by net.
1037 * R 2 - both orig and retransmit are in flight.
1038 * L|R 1 - orig is lost, retransmit is in flight.
1039 * S|R 1 - orig reached receiver, retrans is still in flight.
1040 * (L|S|R is logically valid, it could occur when L|R is sacked,
1041 * but it is equivalent to plain S and code short-curcuits it to S.
1042 * L|S is logically invalid, it would mean -1 packet in flight 8))
1044 * These 6 states form finite state machine, controlled by the following events:
1045 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1046 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1047 * 3. Loss detection event of one of three flavors:
1048 * A. Scoreboard estimator decided the packet is lost.
1049 * A'. Reno "three dupacks" marks head of queue lost.
1050 * A''. Its FACK modfication, head until snd.fack is lost.
1051 * B. SACK arrives sacking data transmitted after never retransmitted
1052 * hole was sent out.
1053 * C. SACK arrives sacking SND.NXT at the moment, when the
1054 * segment was retransmitted.
1055 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1057 * It is pleasant to note, that state diagram turns out to be commutative,
1058 * so that we are allowed not to be bothered by order of our actions,
1059 * when multiple events arrive simultaneously. (see the function below).
1061 * Reordering detection.
1062 * --------------------
1063 * Reordering metric is maximal distance, which a packet can be displaced
1064 * in packet stream. With SACKs we can estimate it:
1066 * 1. SACK fills old hole and the corresponding segment was not
1067 * ever retransmitted -> reordering. Alas, we cannot use it
1068 * when segment was retransmitted.
1069 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1070 * for retransmitted and already SACKed segment -> reordering..
1071 * Both of these heuristics are not used in Loss state, when we cannot
1072 * account for retransmits accurately.
1074 * SACK block validation.
1075 * ----------------------
1077 * SACK block range validation checks that the received SACK block fits to
1078 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1079 * Note that SND.UNA is not included to the range though being valid because
1080 * it means that the receiver is rather inconsistent with itself reporting
1081 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1082 * perfectly valid, however, in light of RFC2018 which explicitly states
1083 * that "SACK block MUST reflect the newest segment. Even if the newest
1084 * segment is going to be discarded ...", not that it looks very clever
1085 * in case of head skb. Due to potentional receiver driven attacks, we
1086 * choose to avoid immediate execution of a walk in write queue due to
1087 * reneging and defer head skb's loss recovery to standard loss recovery
1088 * procedure that will eventually trigger (nothing forbids us doing this).
1090 * Implements also blockage to start_seq wrap-around. Problem lies in the
1091 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1092 * there's no guarantee that it will be before snd_nxt (n). The problem
1093 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1096 * <- outs wnd -> <- wrapzone ->
1097 * u e n u_w e_w s n_w
1099 * |<------------+------+----- TCP seqno space --------------+---------->|
1100 * ...-- <2^31 ->| |<--------...
1101 * ...---- >2^31 ------>| |<--------...
1103 * Current code wouldn't be vulnerable but it's better still to discard such
1104 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1105 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1106 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1107 * equal to the ideal case (infinite seqno space without wrap caused issues).
1109 * With D-SACK the lower bound is extended to cover sequence space below
1110 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1111 * again, D-SACK block must not to go across snd_una (for the same reason as
1112 * for the normal SACK blocks, explained above). But there all simplicity
1113 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1114 * fully below undo_marker they do not affect behavior in anyway and can
1115 * therefore be safely ignored. In rare cases (which are more or less
1116 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1117 * fragmentation and packet reordering past skb's retransmission. To consider
1118 * them correctly, the acceptable range must be extended even more though
1119 * the exact amount is rather hard to quantify. However, tp->max_window can
1120 * be used as an exaggerated estimate.
1122 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1123 u32 start_seq, u32 end_seq)
1125 /* Too far in future, or reversed (interpretation is ambiguous) */
1126 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1129 /* Nasty start_seq wrap-around check (see comments above) */
1130 if (!before(start_seq, tp->snd_nxt))
1133 /* In outstanding window? ...This is valid exit for D-SACKs too.
1134 * start_seq == snd_una is non-sensical (see comments above)
1136 if (after(start_seq, tp->snd_una))
1139 if (!is_dsack || !tp->undo_marker)
1142 /* ...Then it's D-SACK, and must reside below snd_una completely */
1143 if (after(end_seq, tp->snd_una))
1146 if (!before(start_seq, tp->undo_marker))
1150 if (!after(end_seq, tp->undo_marker))
1153 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1154 * start_seq < undo_marker and end_seq >= undo_marker.
1156 return !before(start_seq, end_seq - tp->max_window);
1159 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1160 * Event "C". Later note: FACK people cheated me again 8), we have to account
1161 * for reordering! Ugly, but should help.
1163 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1164 * less than what is now known to be received by the other end (derived from
1165 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1166 * retransmitted skbs to avoid some costly processing per ACKs.
1168 static void tcp_mark_lost_retrans(struct sock *sk)
1170 const struct inet_connection_sock *icsk = inet_csk(sk);
1171 struct tcp_sock *tp = tcp_sk(sk);
1172 struct sk_buff *skb;
1174 u32 new_low_seq = tp->snd_nxt;
1175 u32 received_upto = tcp_highest_sack_seq(tp);
1177 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1178 !after(received_upto, tp->lost_retrans_low) ||
1179 icsk->icsk_ca_state != TCP_CA_Recovery)
1182 tcp_for_write_queue(skb, sk) {
1183 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1185 if (skb == tcp_send_head(sk))
1187 if (cnt == tp->retrans_out)
1189 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1192 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1195 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1196 * constraint here (see above) but figuring out that at
1197 * least tp->reordering SACK blocks reside between ack_seq
1198 * and received_upto is not easy task to do cheaply with
1199 * the available datastructures.
1201 * Whether FACK should check here for tp->reordering segs
1202 * in-between one could argue for either way (it would be
1203 * rather simple to implement as we could count fack_count
1204 * during the walk and do tp->fackets_out - fack_count).
1206 if (after(received_upto, ack_seq)) {
1207 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1208 tp->retrans_out -= tcp_skb_pcount(skb);
1210 tcp_skb_mark_lost_uncond_verify(tp, skb);
1211 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1213 if (before(ack_seq, new_low_seq))
1214 new_low_seq = ack_seq;
1215 cnt += tcp_skb_pcount(skb);
1219 if (tp->retrans_out)
1220 tp->lost_retrans_low = new_low_seq;
1223 static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1224 struct tcp_sack_block_wire *sp, int num_sacks,
1227 struct tcp_sock *tp = tcp_sk(sk);
1228 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1229 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1232 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1235 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1236 } else if (num_sacks > 1) {
1237 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1238 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1240 if (!after(end_seq_0, end_seq_1) &&
1241 !before(start_seq_0, start_seq_1)) {
1244 NET_INC_STATS_BH(sock_net(sk),
1245 LINUX_MIB_TCPDSACKOFORECV);
1249 /* D-SACK for already forgotten data... Do dumb counting. */
1250 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1251 !after(end_seq_0, prior_snd_una) &&
1252 after(end_seq_0, tp->undo_marker))
1258 struct tcp_sacktag_state {
1264 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1265 * the incoming SACK may not exactly match but we can find smaller MSS
1266 * aligned portion of it that matches. Therefore we might need to fragment
1267 * which may fail and creates some hassle (caller must handle error case
1270 * FIXME: this could be merged to shift decision code
1272 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1273 u32 start_seq, u32 end_seq)
1276 unsigned int pkt_len;
1279 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1280 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1282 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1283 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1284 mss = tcp_skb_mss(skb);
1285 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1288 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1292 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1297 /* Round if necessary so that SACKs cover only full MSSes
1298 * and/or the remaining small portion (if present)
1300 if (pkt_len > mss) {
1301 unsigned int new_len = (pkt_len / mss) * mss;
1302 if (!in_sack && new_len < pkt_len) {
1304 if (new_len > skb->len)
1309 err = tcp_fragment(sk, skb, pkt_len, mss);
1317 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1318 static u8 tcp_sacktag_one(struct sock *sk,
1319 struct tcp_sacktag_state *state, u8 sacked,
1320 u32 start_seq, u32 end_seq,
1321 int dup_sack, int pcount)
1323 struct tcp_sock *tp = tcp_sk(sk);
1324 int fack_count = state->fack_count;
1326 /* Account D-SACK for retransmitted packet. */
1327 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1328 if (tp->undo_marker && tp->undo_retrans &&
1329 after(end_seq, tp->undo_marker))
1331 if (sacked & TCPCB_SACKED_ACKED)
1332 state->reord = min(fack_count, state->reord);
1335 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1336 if (!after(end_seq, tp->snd_una))
1339 if (!(sacked & TCPCB_SACKED_ACKED)) {
1340 if (sacked & TCPCB_SACKED_RETRANS) {
1341 /* If the segment is not tagged as lost,
1342 * we do not clear RETRANS, believing
1343 * that retransmission is still in flight.
1345 if (sacked & TCPCB_LOST) {
1346 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1347 tp->lost_out -= pcount;
1348 tp->retrans_out -= pcount;
1351 if (!(sacked & TCPCB_RETRANS)) {
1352 /* New sack for not retransmitted frame,
1353 * which was in hole. It is reordering.
1355 if (before(start_seq,
1356 tcp_highest_sack_seq(tp)))
1357 state->reord = min(fack_count,
1360 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1361 if (!after(end_seq, tp->frto_highmark))
1362 state->flag |= FLAG_ONLY_ORIG_SACKED;
1365 if (sacked & TCPCB_LOST) {
1366 sacked &= ~TCPCB_LOST;
1367 tp->lost_out -= pcount;
1371 sacked |= TCPCB_SACKED_ACKED;
1372 state->flag |= FLAG_DATA_SACKED;
1373 tp->sacked_out += pcount;
1375 fack_count += pcount;
1377 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1378 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1379 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1380 tp->lost_cnt_hint += pcount;
1382 if (fack_count > tp->fackets_out)
1383 tp->fackets_out = fack_count;
1386 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1387 * frames and clear it. undo_retrans is decreased above, L|R frames
1388 * are accounted above as well.
1390 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1391 sacked &= ~TCPCB_SACKED_RETRANS;
1392 tp->retrans_out -= pcount;
1398 /* Shift newly-SACKed bytes from this skb to the immediately previous
1399 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1401 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1402 struct tcp_sacktag_state *state,
1403 unsigned int pcount, int shifted, int mss,
1406 struct tcp_sock *tp = tcp_sk(sk);
1407 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1408 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1409 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1413 /* Adjust counters and hints for the newly sacked sequence
1414 * range but discard the return value since prev is already
1415 * marked. We must tag the range first because the seq
1416 * advancement below implicitly advances
1417 * tcp_highest_sack_seq() when skb is highest_sack.
1419 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1420 start_seq, end_seq, dup_sack, pcount);
1422 if (skb == tp->lost_skb_hint)
1423 tp->lost_cnt_hint += pcount;
1425 TCP_SKB_CB(prev)->end_seq += shifted;
1426 TCP_SKB_CB(skb)->seq += shifted;
1428 skb_shinfo(prev)->gso_segs += pcount;
1429 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1430 skb_shinfo(skb)->gso_segs -= pcount;
1432 /* When we're adding to gso_segs == 1, gso_size will be zero,
1433 * in theory this shouldn't be necessary but as long as DSACK
1434 * code can come after this skb later on it's better to keep
1435 * setting gso_size to something.
1437 if (!skb_shinfo(prev)->gso_size) {
1438 skb_shinfo(prev)->gso_size = mss;
1439 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1442 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1443 if (skb_shinfo(skb)->gso_segs <= 1) {
1444 skb_shinfo(skb)->gso_size = 0;
1445 skb_shinfo(skb)->gso_type = 0;
1448 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1449 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1452 BUG_ON(!tcp_skb_pcount(skb));
1453 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1457 /* Whole SKB was eaten :-) */
1459 if (skb == tp->retransmit_skb_hint)
1460 tp->retransmit_skb_hint = prev;
1461 if (skb == tp->scoreboard_skb_hint)
1462 tp->scoreboard_skb_hint = prev;
1463 if (skb == tp->lost_skb_hint) {
1464 tp->lost_skb_hint = prev;
1465 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1468 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1469 if (skb == tcp_highest_sack(sk))
1470 tcp_advance_highest_sack(sk, skb);
1472 tcp_unlink_write_queue(skb, sk);
1473 sk_wmem_free_skb(sk, skb);
1475 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1480 /* I wish gso_size would have a bit more sane initialization than
1481 * something-or-zero which complicates things
1483 static int tcp_skb_seglen(const struct sk_buff *skb)
1485 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1488 /* Shifting pages past head area doesn't work */
1489 static int skb_can_shift(const struct sk_buff *skb)
1491 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1494 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1497 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1498 struct tcp_sacktag_state *state,
1499 u32 start_seq, u32 end_seq,
1502 struct tcp_sock *tp = tcp_sk(sk);
1503 struct sk_buff *prev;
1509 if (!sk_can_gso(sk))
1512 /* Normally R but no L won't result in plain S */
1514 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1516 if (!skb_can_shift(skb))
1518 /* This frame is about to be dropped (was ACKed). */
1519 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1522 /* Can only happen with delayed DSACK + discard craziness */
1523 if (unlikely(skb == tcp_write_queue_head(sk)))
1525 prev = tcp_write_queue_prev(sk, skb);
1527 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1530 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1531 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1535 pcount = tcp_skb_pcount(skb);
1536 mss = tcp_skb_seglen(skb);
1538 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1539 * drop this restriction as unnecessary
1541 if (mss != tcp_skb_seglen(prev))
1544 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1546 /* CHECKME: This is non-MSS split case only?, this will
1547 * cause skipped skbs due to advancing loop btw, original
1548 * has that feature too
1550 if (tcp_skb_pcount(skb) <= 1)
1553 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1555 /* TODO: head merge to next could be attempted here
1556 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1557 * though it might not be worth of the additional hassle
1559 * ...we can probably just fallback to what was done
1560 * previously. We could try merging non-SACKed ones
1561 * as well but it probably isn't going to buy off
1562 * because later SACKs might again split them, and
1563 * it would make skb timestamp tracking considerably
1569 len = end_seq - TCP_SKB_CB(skb)->seq;
1571 BUG_ON(len > skb->len);
1573 /* MSS boundaries should be honoured or else pcount will
1574 * severely break even though it makes things bit trickier.
1575 * Optimize common case to avoid most of the divides
1577 mss = tcp_skb_mss(skb);
1579 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1580 * drop this restriction as unnecessary
1582 if (mss != tcp_skb_seglen(prev))
1587 } else if (len < mss) {
1595 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1596 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1599 if (!skb_shift(prev, skb, len))
1601 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1604 /* Hole filled allows collapsing with the next as well, this is very
1605 * useful when hole on every nth skb pattern happens
1607 if (prev == tcp_write_queue_tail(sk))
1609 skb = tcp_write_queue_next(sk, prev);
1611 if (!skb_can_shift(skb) ||
1612 (skb == tcp_send_head(sk)) ||
1613 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1614 (mss != tcp_skb_seglen(skb)))
1618 if (skb_shift(prev, skb, len)) {
1619 pcount += tcp_skb_pcount(skb);
1620 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1624 state->fack_count += pcount;
1631 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1635 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1636 struct tcp_sack_block *next_dup,
1637 struct tcp_sacktag_state *state,
1638 u32 start_seq, u32 end_seq,
1641 struct tcp_sock *tp = tcp_sk(sk);
1642 struct sk_buff *tmp;
1644 tcp_for_write_queue_from(skb, sk) {
1646 int dup_sack = dup_sack_in;
1648 if (skb == tcp_send_head(sk))
1651 /* queue is in-order => we can short-circuit the walk early */
1652 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1655 if ((next_dup != NULL) &&
1656 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1657 in_sack = tcp_match_skb_to_sack(sk, skb,
1658 next_dup->start_seq,
1664 /* skb reference here is a bit tricky to get right, since
1665 * shifting can eat and free both this skb and the next,
1666 * so not even _safe variant of the loop is enough.
1669 tmp = tcp_shift_skb_data(sk, skb, state,
1670 start_seq, end_seq, dup_sack);
1679 in_sack = tcp_match_skb_to_sack(sk, skb,
1685 if (unlikely(in_sack < 0))
1689 TCP_SKB_CB(skb)->sacked =
1692 TCP_SKB_CB(skb)->sacked,
1693 TCP_SKB_CB(skb)->seq,
1694 TCP_SKB_CB(skb)->end_seq,
1696 tcp_skb_pcount(skb));
1698 if (!before(TCP_SKB_CB(skb)->seq,
1699 tcp_highest_sack_seq(tp)))
1700 tcp_advance_highest_sack(sk, skb);
1703 state->fack_count += tcp_skb_pcount(skb);
1708 /* Avoid all extra work that is being done by sacktag while walking in
1711 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1712 struct tcp_sacktag_state *state,
1715 tcp_for_write_queue_from(skb, sk) {
1716 if (skb == tcp_send_head(sk))
1719 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1722 state->fack_count += tcp_skb_pcount(skb);
1727 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1729 struct tcp_sack_block *next_dup,
1730 struct tcp_sacktag_state *state,
1733 if (next_dup == NULL)
1736 if (before(next_dup->start_seq, skip_to_seq)) {
1737 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1738 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1739 next_dup->start_seq, next_dup->end_seq,
1746 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1748 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1752 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1755 const struct inet_connection_sock *icsk = inet_csk(sk);
1756 struct tcp_sock *tp = tcp_sk(sk);
1757 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1758 TCP_SKB_CB(ack_skb)->sacked);
1759 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1760 struct tcp_sack_block sp[TCP_NUM_SACKS];
1761 struct tcp_sack_block *cache;
1762 struct tcp_sacktag_state state;
1763 struct sk_buff *skb;
1764 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1766 int found_dup_sack = 0;
1768 int first_sack_index;
1771 state.reord = tp->packets_out;
1773 if (!tp->sacked_out) {
1774 if (WARN_ON(tp->fackets_out))
1775 tp->fackets_out = 0;
1776 tcp_highest_sack_reset(sk);
1779 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1780 num_sacks, prior_snd_una);
1782 state.flag |= FLAG_DSACKING_ACK;
1784 /* Eliminate too old ACKs, but take into
1785 * account more or less fresh ones, they can
1786 * contain valid SACK info.
1788 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1791 if (!tp->packets_out)
1795 first_sack_index = 0;
1796 for (i = 0; i < num_sacks; i++) {
1797 int dup_sack = !i && found_dup_sack;
1799 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1800 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1802 if (!tcp_is_sackblock_valid(tp, dup_sack,
1803 sp[used_sacks].start_seq,
1804 sp[used_sacks].end_seq)) {
1808 if (!tp->undo_marker)
1809 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1811 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1813 /* Don't count olds caused by ACK reordering */
1814 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1815 !after(sp[used_sacks].end_seq, tp->snd_una))
1817 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1820 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1822 first_sack_index = -1;
1826 /* Ignore very old stuff early */
1827 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1833 /* order SACK blocks to allow in order walk of the retrans queue */
1834 for (i = used_sacks - 1; i > 0; i--) {
1835 for (j = 0; j < i; j++) {
1836 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1837 swap(sp[j], sp[j + 1]);
1839 /* Track where the first SACK block goes to */
1840 if (j == first_sack_index)
1841 first_sack_index = j + 1;
1846 skb = tcp_write_queue_head(sk);
1847 state.fack_count = 0;
1850 if (!tp->sacked_out) {
1851 /* It's already past, so skip checking against it */
1852 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1854 cache = tp->recv_sack_cache;
1855 /* Skip empty blocks in at head of the cache */
1856 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1861 while (i < used_sacks) {
1862 u32 start_seq = sp[i].start_seq;
1863 u32 end_seq = sp[i].end_seq;
1864 int dup_sack = (found_dup_sack && (i == first_sack_index));
1865 struct tcp_sack_block *next_dup = NULL;
1867 if (found_dup_sack && ((i + 1) == first_sack_index))
1868 next_dup = &sp[i + 1];
1870 /* Event "B" in the comment above. */
1871 if (after(end_seq, tp->high_seq))
1872 state.flag |= FLAG_DATA_LOST;
1874 /* Skip too early cached blocks */
1875 while (tcp_sack_cache_ok(tp, cache) &&
1876 !before(start_seq, cache->end_seq))
1879 /* Can skip some work by looking recv_sack_cache? */
1880 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1881 after(end_seq, cache->start_seq)) {
1884 if (before(start_seq, cache->start_seq)) {
1885 skb = tcp_sacktag_skip(skb, sk, &state,
1887 skb = tcp_sacktag_walk(skb, sk, next_dup,
1894 /* Rest of the block already fully processed? */
1895 if (!after(end_seq, cache->end_seq))
1898 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1902 /* ...tail remains todo... */
1903 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1904 /* ...but better entrypoint exists! */
1905 skb = tcp_highest_sack(sk);
1908 state.fack_count = tp->fackets_out;
1913 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1914 /* Check overlap against next cached too (past this one already) */
1919 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1920 skb = tcp_highest_sack(sk);
1923 state.fack_count = tp->fackets_out;
1925 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1928 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1929 start_seq, end_seq, dup_sack);
1932 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1933 * due to in-order walk
1935 if (after(end_seq, tp->frto_highmark))
1936 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1941 /* Clear the head of the cache sack blocks so we can skip it next time */
1942 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1943 tp->recv_sack_cache[i].start_seq = 0;
1944 tp->recv_sack_cache[i].end_seq = 0;
1946 for (j = 0; j < used_sacks; j++)
1947 tp->recv_sack_cache[i++] = sp[j];
1949 tcp_mark_lost_retrans(sk);
1951 tcp_verify_left_out(tp);
1953 if ((state.reord < tp->fackets_out) &&
1954 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1955 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1956 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1960 #if FASTRETRANS_DEBUG > 0
1961 WARN_ON((int)tp->sacked_out < 0);
1962 WARN_ON((int)tp->lost_out < 0);
1963 WARN_ON((int)tp->retrans_out < 0);
1964 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1969 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1970 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1972 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1976 holes = max(tp->lost_out, 1U);
1977 holes = min(holes, tp->packets_out);
1979 if ((tp->sacked_out + holes) > tp->packets_out) {
1980 tp->sacked_out = tp->packets_out - holes;
1986 /* If we receive more dupacks than we expected counting segments
1987 * in assumption of absent reordering, interpret this as reordering.
1988 * The only another reason could be bug in receiver TCP.
1990 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1992 struct tcp_sock *tp = tcp_sk(sk);
1993 if (tcp_limit_reno_sacked(tp))
1994 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1997 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1999 static void tcp_add_reno_sack(struct sock *sk)
2001 struct tcp_sock *tp = tcp_sk(sk);
2003 tcp_check_reno_reordering(sk, 0);
2004 tcp_verify_left_out(tp);
2007 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2009 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
2011 struct tcp_sock *tp = tcp_sk(sk);
2014 /* One ACK acked hole. The rest eat duplicate ACKs. */
2015 if (acked - 1 >= tp->sacked_out)
2018 tp->sacked_out -= acked - 1;
2020 tcp_check_reno_reordering(sk, acked);
2021 tcp_verify_left_out(tp);
2024 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2029 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2031 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2034 /* F-RTO can only be used if TCP has never retransmitted anything other than
2035 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2037 int tcp_use_frto(struct sock *sk)
2039 const struct tcp_sock *tp = tcp_sk(sk);
2040 const struct inet_connection_sock *icsk = inet_csk(sk);
2041 struct sk_buff *skb;
2043 if (!sysctl_tcp_frto)
2046 /* MTU probe and F-RTO won't really play nicely along currently */
2047 if (icsk->icsk_mtup.probe_size)
2050 if (tcp_is_sackfrto(tp))
2053 /* Avoid expensive walking of rexmit queue if possible */
2054 if (tp->retrans_out > 1)
2057 skb = tcp_write_queue_head(sk);
2058 if (tcp_skb_is_last(sk, skb))
2060 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2061 tcp_for_write_queue_from(skb, sk) {
2062 if (skb == tcp_send_head(sk))
2064 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2066 /* Short-circuit when first non-SACKed skb has been checked */
2067 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2073 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2074 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2075 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2076 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2077 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2078 * bits are handled if the Loss state is really to be entered (in
2079 * tcp_enter_frto_loss).
2081 * Do like tcp_enter_loss() would; when RTO expires the second time it
2083 * "Reduce ssthresh if it has not yet been made inside this window."
2085 void tcp_enter_frto(struct sock *sk)
2087 const struct inet_connection_sock *icsk = inet_csk(sk);
2088 struct tcp_sock *tp = tcp_sk(sk);
2089 struct sk_buff *skb;
2091 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2092 tp->snd_una == tp->high_seq ||
2093 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2094 !icsk->icsk_retransmits)) {
2095 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2096 /* Our state is too optimistic in ssthresh() call because cwnd
2097 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2098 * recovery has not yet completed. Pattern would be this: RTO,
2099 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2101 * RFC4138 should be more specific on what to do, even though
2102 * RTO is quite unlikely to occur after the first Cumulative ACK
2103 * due to back-off and complexity of triggering events ...
2105 if (tp->frto_counter) {
2107 stored_cwnd = tp->snd_cwnd;
2109 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2110 tp->snd_cwnd = stored_cwnd;
2112 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2114 /* ... in theory, cong.control module could do "any tricks" in
2115 * ssthresh(), which means that ca_state, lost bits and lost_out
2116 * counter would have to be faked before the call occurs. We
2117 * consider that too expensive, unlikely and hacky, so modules
2118 * using these in ssthresh() must deal these incompatibility
2119 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2121 tcp_ca_event(sk, CA_EVENT_FRTO);
2124 tp->undo_marker = tp->snd_una;
2125 tp->undo_retrans = 0;
2127 skb = tcp_write_queue_head(sk);
2128 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2129 tp->undo_marker = 0;
2130 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2131 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2132 tp->retrans_out -= tcp_skb_pcount(skb);
2134 tcp_verify_left_out(tp);
2136 /* Too bad if TCP was application limited */
2137 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2139 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2140 * The last condition is necessary at least in tp->frto_counter case.
2142 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2143 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2144 after(tp->high_seq, tp->snd_una)) {
2145 tp->frto_highmark = tp->high_seq;
2147 tp->frto_highmark = tp->snd_nxt;
2149 tcp_set_ca_state(sk, TCP_CA_Disorder);
2150 tp->high_seq = tp->snd_nxt;
2151 tp->frto_counter = 1;
2154 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2155 * which indicates that we should follow the traditional RTO recovery,
2156 * i.e. mark everything lost and do go-back-N retransmission.
2158 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2160 struct tcp_sock *tp = tcp_sk(sk);
2161 struct sk_buff *skb;
2164 tp->retrans_out = 0;
2165 if (tcp_is_reno(tp))
2166 tcp_reset_reno_sack(tp);
2168 tcp_for_write_queue(skb, sk) {
2169 if (skb == tcp_send_head(sk))
2172 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2174 * Count the retransmission made on RTO correctly (only when
2175 * waiting for the first ACK and did not get it)...
2177 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2178 /* For some reason this R-bit might get cleared? */
2179 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2180 tp->retrans_out += tcp_skb_pcount(skb);
2181 /* ...enter this if branch just for the first segment */
2182 flag |= FLAG_DATA_ACKED;
2184 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2185 tp->undo_marker = 0;
2186 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2189 /* Marking forward transmissions that were made after RTO lost
2190 * can cause unnecessary retransmissions in some scenarios,
2191 * SACK blocks will mitigate that in some but not in all cases.
2192 * We used to not mark them but it was causing break-ups with
2193 * receivers that do only in-order receival.
2195 * TODO: we could detect presence of such receiver and select
2196 * different behavior per flow.
2198 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2199 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2200 tp->lost_out += tcp_skb_pcount(skb);
2201 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2204 tcp_verify_left_out(tp);
2206 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2207 tp->snd_cwnd_cnt = 0;
2208 tp->snd_cwnd_stamp = tcp_time_stamp;
2209 tp->frto_counter = 0;
2210 tp->bytes_acked = 0;
2212 tp->reordering = min_t(unsigned int, tp->reordering,
2213 sysctl_tcp_reordering);
2214 tcp_set_ca_state(sk, TCP_CA_Loss);
2215 tp->high_seq = tp->snd_nxt;
2216 TCP_ECN_queue_cwr(tp);
2218 tcp_clear_all_retrans_hints(tp);
2221 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2223 tp->retrans_out = 0;
2226 tp->undo_marker = 0;
2227 tp->undo_retrans = 0;
2230 void tcp_clear_retrans(struct tcp_sock *tp)
2232 tcp_clear_retrans_partial(tp);
2234 tp->fackets_out = 0;
2238 /* Enter Loss state. If "how" is not zero, forget all SACK information
2239 * and reset tags completely, otherwise preserve SACKs. If receiver
2240 * dropped its ofo queue, we will know this due to reneging detection.
2242 void tcp_enter_loss(struct sock *sk, int how)
2244 const struct inet_connection_sock *icsk = inet_csk(sk);
2245 struct tcp_sock *tp = tcp_sk(sk);
2246 struct sk_buff *skb;
2248 /* Reduce ssthresh if it has not yet been made inside this window. */
2249 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2250 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2251 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2252 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2253 tcp_ca_event(sk, CA_EVENT_LOSS);
2256 tp->snd_cwnd_cnt = 0;
2257 tp->snd_cwnd_stamp = tcp_time_stamp;
2259 tp->bytes_acked = 0;
2260 tcp_clear_retrans_partial(tp);
2262 if (tcp_is_reno(tp))
2263 tcp_reset_reno_sack(tp);
2266 /* Push undo marker, if it was plain RTO and nothing
2267 * was retransmitted. */
2268 tp->undo_marker = tp->snd_una;
2271 tp->fackets_out = 0;
2273 tcp_clear_all_retrans_hints(tp);
2275 tcp_for_write_queue(skb, sk) {
2276 if (skb == tcp_send_head(sk))
2279 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2280 tp->undo_marker = 0;
2281 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2282 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2283 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2284 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2285 tp->lost_out += tcp_skb_pcount(skb);
2286 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2289 tcp_verify_left_out(tp);
2291 tp->reordering = min_t(unsigned int, tp->reordering,
2292 sysctl_tcp_reordering);
2293 tcp_set_ca_state(sk, TCP_CA_Loss);
2294 tp->high_seq = tp->snd_nxt;
2295 TCP_ECN_queue_cwr(tp);
2296 /* Abort F-RTO algorithm if one is in progress */
2297 tp->frto_counter = 0;
2300 /* If ACK arrived pointing to a remembered SACK, it means that our
2301 * remembered SACKs do not reflect real state of receiver i.e.
2302 * receiver _host_ is heavily congested (or buggy).
2304 * Do processing similar to RTO timeout.
2306 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2308 if (flag & FLAG_SACK_RENEGING) {
2309 struct inet_connection_sock *icsk = inet_csk(sk);
2310 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2312 tcp_enter_loss(sk, 1);
2313 icsk->icsk_retransmits++;
2314 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2315 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2316 icsk->icsk_rto, TCP_RTO_MAX);
2322 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2324 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2327 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2328 * counter when SACK is enabled (without SACK, sacked_out is used for
2331 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2332 * segments up to the highest received SACK block so far and holes in
2335 * With reordering, holes may still be in flight, so RFC3517 recovery
2336 * uses pure sacked_out (total number of SACKed segments) even though
2337 * it violates the RFC that uses duplicate ACKs, often these are equal
2338 * but when e.g. out-of-window ACKs or packet duplication occurs,
2339 * they differ. Since neither occurs due to loss, TCP should really
2342 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2344 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2347 static inline int tcp_skb_timedout(const struct sock *sk,
2348 const struct sk_buff *skb)
2350 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2353 static inline int tcp_head_timedout(const struct sock *sk)
2355 const struct tcp_sock *tp = tcp_sk(sk);
2357 return tp->packets_out &&
2358 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2361 /* Linux NewReno/SACK/FACK/ECN state machine.
2362 * --------------------------------------
2364 * "Open" Normal state, no dubious events, fast path.
2365 * "Disorder" In all the respects it is "Open",
2366 * but requires a bit more attention. It is entered when
2367 * we see some SACKs or dupacks. It is split of "Open"
2368 * mainly to move some processing from fast path to slow one.
2369 * "CWR" CWND was reduced due to some Congestion Notification event.
2370 * It can be ECN, ICMP source quench, local device congestion.
2371 * "Recovery" CWND was reduced, we are fast-retransmitting.
2372 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2374 * tcp_fastretrans_alert() is entered:
2375 * - each incoming ACK, if state is not "Open"
2376 * - when arrived ACK is unusual, namely:
2381 * Counting packets in flight is pretty simple.
2383 * in_flight = packets_out - left_out + retrans_out
2385 * packets_out is SND.NXT-SND.UNA counted in packets.
2387 * retrans_out is number of retransmitted segments.
2389 * left_out is number of segments left network, but not ACKed yet.
2391 * left_out = sacked_out + lost_out
2393 * sacked_out: Packets, which arrived to receiver out of order
2394 * and hence not ACKed. With SACKs this number is simply
2395 * amount of SACKed data. Even without SACKs
2396 * it is easy to give pretty reliable estimate of this number,
2397 * counting duplicate ACKs.
2399 * lost_out: Packets lost by network. TCP has no explicit
2400 * "loss notification" feedback from network (for now).
2401 * It means that this number can be only _guessed_.
2402 * Actually, it is the heuristics to predict lossage that
2403 * distinguishes different algorithms.
2405 * F.e. after RTO, when all the queue is considered as lost,
2406 * lost_out = packets_out and in_flight = retrans_out.
2408 * Essentially, we have now two algorithms counting
2411 * FACK: It is the simplest heuristics. As soon as we decided
2412 * that something is lost, we decide that _all_ not SACKed
2413 * packets until the most forward SACK are lost. I.e.
2414 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2415 * It is absolutely correct estimate, if network does not reorder
2416 * packets. And it loses any connection to reality when reordering
2417 * takes place. We use FACK by default until reordering
2418 * is suspected on the path to this destination.
2420 * NewReno: when Recovery is entered, we assume that one segment
2421 * is lost (classic Reno). While we are in Recovery and
2422 * a partial ACK arrives, we assume that one more packet
2423 * is lost (NewReno). This heuristics are the same in NewReno
2426 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2427 * deflation etc. CWND is real congestion window, never inflated, changes
2428 * only according to classic VJ rules.
2430 * Really tricky (and requiring careful tuning) part of algorithm
2431 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2432 * The first determines the moment _when_ we should reduce CWND and,
2433 * hence, slow down forward transmission. In fact, it determines the moment
2434 * when we decide that hole is caused by loss, rather than by a reorder.
2436 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2437 * holes, caused by lost packets.
2439 * And the most logically complicated part of algorithm is undo
2440 * heuristics. We detect false retransmits due to both too early
2441 * fast retransmit (reordering) and underestimated RTO, analyzing
2442 * timestamps and D-SACKs. When we detect that some segments were
2443 * retransmitted by mistake and CWND reduction was wrong, we undo
2444 * window reduction and abort recovery phase. This logic is hidden
2445 * inside several functions named tcp_try_undo_<something>.
2448 /* This function decides, when we should leave Disordered state
2449 * and enter Recovery phase, reducing congestion window.
2451 * Main question: may we further continue forward transmission
2452 * with the same cwnd?
2454 static int tcp_time_to_recover(struct sock *sk)
2456 struct tcp_sock *tp = tcp_sk(sk);
2459 /* Do not perform any recovery during F-RTO algorithm */
2460 if (tp->frto_counter)
2463 /* Trick#1: The loss is proven. */
2467 /* Not-A-Trick#2 : Classic rule... */
2468 if (tcp_dupack_heuristics(tp) > tp->reordering)
2471 /* Trick#3 : when we use RFC2988 timer restart, fast
2472 * retransmit can be triggered by timeout of queue head.
2474 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2477 /* Trick#4: It is still not OK... But will it be useful to delay
2480 packets_out = tp->packets_out;
2481 if (packets_out <= tp->reordering &&
2482 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2483 !tcp_may_send_now(sk)) {
2484 /* We have nothing to send. This connection is limited
2485 * either by receiver window or by application.
2490 /* If a thin stream is detected, retransmit after first
2491 * received dupack. Employ only if SACK is supported in order
2492 * to avoid possible corner-case series of spurious retransmissions
2493 * Use only if there are no unsent data.
2495 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2496 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2497 tcp_is_sack(tp) && !tcp_send_head(sk))
2503 /* New heuristics: it is possible only after we switched to restart timer
2504 * each time when something is ACKed. Hence, we can detect timed out packets
2505 * during fast retransmit without falling to slow start.
2507 * Usefulness of this as is very questionable, since we should know which of
2508 * the segments is the next to timeout which is relatively expensive to find
2509 * in general case unless we add some data structure just for that. The
2510 * current approach certainly won't find the right one too often and when it
2511 * finally does find _something_ it usually marks large part of the window
2512 * right away (because a retransmission with a larger timestamp blocks the
2513 * loop from advancing). -ij
2515 static void tcp_timeout_skbs(struct sock *sk)
2517 struct tcp_sock *tp = tcp_sk(sk);
2518 struct sk_buff *skb;
2520 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2523 skb = tp->scoreboard_skb_hint;
2524 if (tp->scoreboard_skb_hint == NULL)
2525 skb = tcp_write_queue_head(sk);
2527 tcp_for_write_queue_from(skb, sk) {
2528 if (skb == tcp_send_head(sk))
2530 if (!tcp_skb_timedout(sk, skb))
2533 tcp_skb_mark_lost(tp, skb);
2536 tp->scoreboard_skb_hint = skb;
2538 tcp_verify_left_out(tp);
2541 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2542 * is against sacked "cnt", otherwise it's against facked "cnt"
2544 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2546 struct tcp_sock *tp = tcp_sk(sk);
2547 struct sk_buff *skb;
2552 WARN_ON(packets > tp->packets_out);
2553 if (tp->lost_skb_hint) {
2554 skb = tp->lost_skb_hint;
2555 cnt = tp->lost_cnt_hint;
2556 /* Head already handled? */
2557 if (mark_head && skb != tcp_write_queue_head(sk))
2560 skb = tcp_write_queue_head(sk);
2564 tcp_for_write_queue_from(skb, sk) {
2565 if (skb == tcp_send_head(sk))
2567 /* TODO: do this better */
2568 /* this is not the most efficient way to do this... */
2569 tp->lost_skb_hint = skb;
2570 tp->lost_cnt_hint = cnt;
2572 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2576 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2577 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2578 cnt += tcp_skb_pcount(skb);
2580 if (cnt > packets) {
2581 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2582 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2583 (oldcnt >= packets))
2586 mss = skb_shinfo(skb)->gso_size;
2587 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2593 tcp_skb_mark_lost(tp, skb);
2598 tcp_verify_left_out(tp);
2601 /* Account newly detected lost packet(s) */
2603 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2605 struct tcp_sock *tp = tcp_sk(sk);
2607 if (tcp_is_reno(tp)) {
2608 tcp_mark_head_lost(sk, 1, 1);
2609 } else if (tcp_is_fack(tp)) {
2610 int lost = tp->fackets_out - tp->reordering;
2613 tcp_mark_head_lost(sk, lost, 0);
2615 int sacked_upto = tp->sacked_out - tp->reordering;
2616 if (sacked_upto >= 0)
2617 tcp_mark_head_lost(sk, sacked_upto, 0);
2618 else if (fast_rexmit)
2619 tcp_mark_head_lost(sk, 1, 1);
2622 tcp_timeout_skbs(sk);
2625 /* CWND moderation, preventing bursts due to too big ACKs
2626 * in dubious situations.
2628 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2630 tp->snd_cwnd = min(tp->snd_cwnd,
2631 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2632 tp->snd_cwnd_stamp = tcp_time_stamp;
2635 /* Lower bound on congestion window is slow start threshold
2636 * unless congestion avoidance choice decides to overide it.
2638 static inline u32 tcp_cwnd_min(const struct sock *sk)
2640 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2642 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2645 /* Decrease cwnd each second ack. */
2646 static void tcp_cwnd_down(struct sock *sk, int flag)
2648 struct tcp_sock *tp = tcp_sk(sk);
2649 int decr = tp->snd_cwnd_cnt + 1;
2651 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2652 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2653 tp->snd_cwnd_cnt = decr & 1;
2656 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2657 tp->snd_cwnd -= decr;
2659 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2660 tp->snd_cwnd_stamp = tcp_time_stamp;
2664 /* Nothing was retransmitted or returned timestamp is less
2665 * than timestamp of the first retransmission.
2667 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2669 return !tp->retrans_stamp ||
2670 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2671 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2674 /* Undo procedures. */
2676 #if FASTRETRANS_DEBUG > 1
2677 static void DBGUNDO(struct sock *sk, const char *msg)
2679 struct tcp_sock *tp = tcp_sk(sk);
2680 struct inet_sock *inet = inet_sk(sk);
2682 if (sk->sk_family == AF_INET) {
2683 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2685 &inet->inet_daddr, ntohs(inet->inet_dport),
2686 tp->snd_cwnd, tcp_left_out(tp),
2687 tp->snd_ssthresh, tp->prior_ssthresh,
2690 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2691 else if (sk->sk_family == AF_INET6) {
2692 struct ipv6_pinfo *np = inet6_sk(sk);
2693 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2695 &np->daddr, ntohs(inet->inet_dport),
2696 tp->snd_cwnd, tcp_left_out(tp),
2697 tp->snd_ssthresh, tp->prior_ssthresh,
2703 #define DBGUNDO(x...) do { } while (0)
2706 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2708 struct tcp_sock *tp = tcp_sk(sk);
2710 if (tp->prior_ssthresh) {
2711 const struct inet_connection_sock *icsk = inet_csk(sk);
2713 if (icsk->icsk_ca_ops->undo_cwnd)
2714 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2716 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2718 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2719 tp->snd_ssthresh = tp->prior_ssthresh;
2720 TCP_ECN_withdraw_cwr(tp);
2723 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2725 tp->snd_cwnd_stamp = tcp_time_stamp;
2728 static inline int tcp_may_undo(const struct tcp_sock *tp)
2730 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2733 /* People celebrate: "We love our President!" */
2734 static int tcp_try_undo_recovery(struct sock *sk)
2736 struct tcp_sock *tp = tcp_sk(sk);
2738 if (tcp_may_undo(tp)) {
2741 /* Happy end! We did not retransmit anything
2742 * or our original transmission succeeded.
2744 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2745 tcp_undo_cwr(sk, true);
2746 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2747 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2749 mib_idx = LINUX_MIB_TCPFULLUNDO;
2751 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2752 tp->undo_marker = 0;
2754 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2755 /* Hold old state until something *above* high_seq
2756 * is ACKed. For Reno it is MUST to prevent false
2757 * fast retransmits (RFC2582). SACK TCP is safe. */
2758 tcp_moderate_cwnd(tp);
2761 tcp_set_ca_state(sk, TCP_CA_Open);
2765 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2766 static void tcp_try_undo_dsack(struct sock *sk)
2768 struct tcp_sock *tp = tcp_sk(sk);
2770 if (tp->undo_marker && !tp->undo_retrans) {
2771 DBGUNDO(sk, "D-SACK");
2772 tcp_undo_cwr(sk, true);
2773 tp->undo_marker = 0;
2774 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2778 /* We can clear retrans_stamp when there are no retransmissions in the
2779 * window. It would seem that it is trivially available for us in
2780 * tp->retrans_out, however, that kind of assumptions doesn't consider
2781 * what will happen if errors occur when sending retransmission for the
2782 * second time. ...It could the that such segment has only
2783 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2784 * the head skb is enough except for some reneging corner cases that
2785 * are not worth the effort.
2787 * Main reason for all this complexity is the fact that connection dying
2788 * time now depends on the validity of the retrans_stamp, in particular,
2789 * that successive retransmissions of a segment must not advance
2790 * retrans_stamp under any conditions.
2792 static int tcp_any_retrans_done(const struct sock *sk)
2794 const struct tcp_sock *tp = tcp_sk(sk);
2795 struct sk_buff *skb;
2797 if (tp->retrans_out)
2800 skb = tcp_write_queue_head(sk);
2801 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2807 /* Undo during fast recovery after partial ACK. */
2809 static int tcp_try_undo_partial(struct sock *sk, int acked)
2811 struct tcp_sock *tp = tcp_sk(sk);
2812 /* Partial ACK arrived. Force Hoe's retransmit. */
2813 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2815 if (tcp_may_undo(tp)) {
2816 /* Plain luck! Hole if filled with delayed
2817 * packet, rather than with a retransmit.
2819 if (!tcp_any_retrans_done(sk))
2820 tp->retrans_stamp = 0;
2822 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2825 tcp_undo_cwr(sk, false);
2826 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2828 /* So... Do not make Hoe's retransmit yet.
2829 * If the first packet was delayed, the rest
2830 * ones are most probably delayed as well.
2837 /* Undo during loss recovery after partial ACK. */
2838 static int tcp_try_undo_loss(struct sock *sk)
2840 struct tcp_sock *tp = tcp_sk(sk);
2842 if (tcp_may_undo(tp)) {
2843 struct sk_buff *skb;
2844 tcp_for_write_queue(skb, sk) {
2845 if (skb == tcp_send_head(sk))
2847 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2850 tcp_clear_all_retrans_hints(tp);
2852 DBGUNDO(sk, "partial loss");
2854 tcp_undo_cwr(sk, true);
2855 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2856 inet_csk(sk)->icsk_retransmits = 0;
2857 tp->undo_marker = 0;
2858 if (tcp_is_sack(tp))
2859 tcp_set_ca_state(sk, TCP_CA_Open);
2865 static inline void tcp_complete_cwr(struct sock *sk)
2867 struct tcp_sock *tp = tcp_sk(sk);
2869 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2870 if (tp->undo_marker) {
2871 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR)
2872 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2874 tp->snd_cwnd = tp->snd_ssthresh;
2875 tp->snd_cwnd_stamp = tcp_time_stamp;
2877 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2880 static void tcp_try_keep_open(struct sock *sk)
2882 struct tcp_sock *tp = tcp_sk(sk);
2883 int state = TCP_CA_Open;
2885 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2886 state = TCP_CA_Disorder;
2888 if (inet_csk(sk)->icsk_ca_state != state) {
2889 tcp_set_ca_state(sk, state);
2890 tp->high_seq = tp->snd_nxt;
2894 static void tcp_try_to_open(struct sock *sk, int flag)
2896 struct tcp_sock *tp = tcp_sk(sk);
2898 tcp_verify_left_out(tp);
2900 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2901 tp->retrans_stamp = 0;
2903 if (flag & FLAG_ECE)
2904 tcp_enter_cwr(sk, 1);
2906 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2907 tcp_try_keep_open(sk);
2908 tcp_moderate_cwnd(tp);
2910 tcp_cwnd_down(sk, flag);
2914 static void tcp_mtup_probe_failed(struct sock *sk)
2916 struct inet_connection_sock *icsk = inet_csk(sk);
2918 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2919 icsk->icsk_mtup.probe_size = 0;
2922 static void tcp_mtup_probe_success(struct sock *sk)
2924 struct tcp_sock *tp = tcp_sk(sk);
2925 struct inet_connection_sock *icsk = inet_csk(sk);
2927 /* FIXME: breaks with very large cwnd */
2928 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2929 tp->snd_cwnd = tp->snd_cwnd *
2930 tcp_mss_to_mtu(sk, tp->mss_cache) /
2931 icsk->icsk_mtup.probe_size;
2932 tp->snd_cwnd_cnt = 0;
2933 tp->snd_cwnd_stamp = tcp_time_stamp;
2934 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2936 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2937 icsk->icsk_mtup.probe_size = 0;
2938 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2941 /* Do a simple retransmit without using the backoff mechanisms in
2942 * tcp_timer. This is used for path mtu discovery.
2943 * The socket is already locked here.
2945 void tcp_simple_retransmit(struct sock *sk)
2947 const struct inet_connection_sock *icsk = inet_csk(sk);
2948 struct tcp_sock *tp = tcp_sk(sk);
2949 struct sk_buff *skb;
2950 unsigned int mss = tcp_current_mss(sk);
2951 u32 prior_lost = tp->lost_out;
2953 tcp_for_write_queue(skb, sk) {
2954 if (skb == tcp_send_head(sk))
2956 if (tcp_skb_seglen(skb) > mss &&
2957 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2958 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2959 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2960 tp->retrans_out -= tcp_skb_pcount(skb);
2962 tcp_skb_mark_lost_uncond_verify(tp, skb);
2966 tcp_clear_retrans_hints_partial(tp);
2968 if (prior_lost == tp->lost_out)
2971 if (tcp_is_reno(tp))
2972 tcp_limit_reno_sacked(tp);
2974 tcp_verify_left_out(tp);
2976 /* Don't muck with the congestion window here.
2977 * Reason is that we do not increase amount of _data_
2978 * in network, but units changed and effective
2979 * cwnd/ssthresh really reduced now.
2981 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2982 tp->high_seq = tp->snd_nxt;
2983 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2984 tp->prior_ssthresh = 0;
2985 tp->undo_marker = 0;
2986 tcp_set_ca_state(sk, TCP_CA_Loss);
2988 tcp_xmit_retransmit_queue(sk);
2990 EXPORT_SYMBOL(tcp_simple_retransmit);
2992 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2993 * (proportional rate reduction with slow start reduction bound) as described in
2994 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2995 * It computes the number of packets to send (sndcnt) based on packets newly
2997 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2998 * cwnd reductions across a full RTT.
2999 * 2) If packets in flight is lower than ssthresh (such as due to excess
3000 * losses and/or application stalls), do not perform any further cwnd
3001 * reductions, but instead slow start up to ssthresh.
3003 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
3004 int fast_rexmit, int flag)
3006 struct tcp_sock *tp = tcp_sk(sk);
3008 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
3010 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
3011 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
3013 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
3015 sndcnt = min_t(int, delta,
3016 max_t(int, tp->prr_delivered - tp->prr_out,
3017 newly_acked_sacked) + 1);
3020 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
3021 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3024 /* Process an event, which can update packets-in-flight not trivially.
3025 * Main goal of this function is to calculate new estimate for left_out,
3026 * taking into account both packets sitting in receiver's buffer and
3027 * packets lost by network.
3029 * Besides that it does CWND reduction, when packet loss is detected
3030 * and changes state of machine.
3032 * It does _not_ decide what to send, it is made in function
3033 * tcp_xmit_retransmit_queue().
3035 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3036 int newly_acked_sacked, int flag)
3038 struct inet_connection_sock *icsk = inet_csk(sk);
3039 struct tcp_sock *tp = tcp_sk(sk);
3040 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3041 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3042 (tcp_fackets_out(tp) > tp->reordering));
3043 int fast_rexmit = 0, mib_idx;
3045 if (WARN_ON(!tp->packets_out && tp->sacked_out))
3047 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3048 tp->fackets_out = 0;
3050 /* Now state machine starts.
3051 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3052 if (flag & FLAG_ECE)
3053 tp->prior_ssthresh = 0;
3055 /* B. In all the states check for reneging SACKs. */
3056 if (tcp_check_sack_reneging(sk, flag))
3059 /* C. Process data loss notification, provided it is valid. */
3060 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
3061 before(tp->snd_una, tp->high_seq) &&
3062 icsk->icsk_ca_state != TCP_CA_Open &&
3063 tp->fackets_out > tp->reordering) {
3064 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
3065 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
3068 /* D. Check consistency of the current state. */
3069 tcp_verify_left_out(tp);
3071 /* E. Check state exit conditions. State can be terminated
3072 * when high_seq is ACKed. */
3073 if (icsk->icsk_ca_state == TCP_CA_Open) {
3074 WARN_ON(tp->retrans_out != 0);
3075 tp->retrans_stamp = 0;
3076 } else if (!before(tp->snd_una, tp->high_seq)) {
3077 switch (icsk->icsk_ca_state) {
3079 icsk->icsk_retransmits = 0;
3080 if (tcp_try_undo_recovery(sk))
3085 /* CWR is to be held something *above* high_seq
3086 * is ACKed for CWR bit to reach receiver. */
3087 if (tp->snd_una != tp->high_seq) {
3088 tcp_complete_cwr(sk);
3089 tcp_set_ca_state(sk, TCP_CA_Open);
3093 case TCP_CA_Disorder:
3094 tcp_try_undo_dsack(sk);
3095 if (!tp->undo_marker ||
3096 /* For SACK case do not Open to allow to undo
3097 * catching for all duplicate ACKs. */
3098 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3099 tp->undo_marker = 0;
3100 tcp_set_ca_state(sk, TCP_CA_Open);
3104 case TCP_CA_Recovery:
3105 if (tcp_is_reno(tp))
3106 tcp_reset_reno_sack(tp);
3107 if (tcp_try_undo_recovery(sk))
3109 tcp_complete_cwr(sk);
3114 /* F. Process state. */
3115 switch (icsk->icsk_ca_state) {
3116 case TCP_CA_Recovery:
3117 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3118 if (tcp_is_reno(tp) && is_dupack)
3119 tcp_add_reno_sack(sk);
3121 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3124 if (flag & FLAG_DATA_ACKED)
3125 icsk->icsk_retransmits = 0;
3126 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3127 tcp_reset_reno_sack(tp);
3128 if (!tcp_try_undo_loss(sk)) {
3129 tcp_moderate_cwnd(tp);
3130 tcp_xmit_retransmit_queue(sk);
3133 if (icsk->icsk_ca_state != TCP_CA_Open)
3135 /* Loss is undone; fall through to processing in Open state. */
3137 if (tcp_is_reno(tp)) {
3138 if (flag & FLAG_SND_UNA_ADVANCED)
3139 tcp_reset_reno_sack(tp);
3141 tcp_add_reno_sack(sk);
3144 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3145 tcp_try_undo_dsack(sk);
3147 if (!tcp_time_to_recover(sk)) {
3148 tcp_try_to_open(sk, flag);
3152 /* MTU probe failure: don't reduce cwnd */
3153 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3154 icsk->icsk_mtup.probe_size &&
3155 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3156 tcp_mtup_probe_failed(sk);
3157 /* Restores the reduction we did in tcp_mtup_probe() */
3159 tcp_simple_retransmit(sk);
3163 /* Otherwise enter Recovery state */
3165 if (tcp_is_reno(tp))
3166 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3168 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3170 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3172 tp->high_seq = tp->snd_nxt;
3173 tp->prior_ssthresh = 0;
3174 tp->undo_marker = tp->snd_una;
3175 tp->undo_retrans = tp->retrans_out;
3177 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3178 if (!(flag & FLAG_ECE))
3179 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3180 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3181 TCP_ECN_queue_cwr(tp);
3184 tp->bytes_acked = 0;
3185 tp->snd_cwnd_cnt = 0;
3186 tp->prior_cwnd = tp->snd_cwnd;
3187 tp->prr_delivered = 0;
3189 tcp_set_ca_state(sk, TCP_CA_Recovery);
3193 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3194 tcp_update_scoreboard(sk, fast_rexmit);
3195 tp->prr_delivered += newly_acked_sacked;
3196 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3197 tcp_xmit_retransmit_queue(sk);
3200 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3202 tcp_rtt_estimator(sk, seq_rtt);
3204 inet_csk(sk)->icsk_backoff = 0;
3206 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3208 /* Read draft-ietf-tcplw-high-performance before mucking
3209 * with this code. (Supersedes RFC1323)
3211 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3213 /* RTTM Rule: A TSecr value received in a segment is used to
3214 * update the averaged RTT measurement only if the segment
3215 * acknowledges some new data, i.e., only if it advances the
3216 * left edge of the send window.
3218 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3219 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3221 * Changed: reset backoff as soon as we see the first valid sample.
3222 * If we do not, we get strongly overestimated rto. With timestamps
3223 * samples are accepted even from very old segments: f.e., when rtt=1
3224 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3225 * answer arrives rto becomes 120 seconds! If at least one of segments
3226 * in window is lost... Voila. --ANK (010210)
3228 struct tcp_sock *tp = tcp_sk(sk);
3230 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3233 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3235 /* We don't have a timestamp. Can only use
3236 * packets that are not retransmitted to determine
3237 * rtt estimates. Also, we must not reset the
3238 * backoff for rto until we get a non-retransmitted
3239 * packet. This allows us to deal with a situation
3240 * where the network delay has increased suddenly.
3241 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3244 if (flag & FLAG_RETRANS_DATA_ACKED)
3247 tcp_valid_rtt_meas(sk, seq_rtt);
3250 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3253 const struct tcp_sock *tp = tcp_sk(sk);
3254 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3255 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3256 tcp_ack_saw_tstamp(sk, flag);
3257 else if (seq_rtt >= 0)
3258 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3261 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3263 const struct inet_connection_sock *icsk = inet_csk(sk);
3264 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3265 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3268 /* Restart timer after forward progress on connection.
3269 * RFC2988 recommends to restart timer to now+rto.
3271 static void tcp_rearm_rto(struct sock *sk)
3273 const struct tcp_sock *tp = tcp_sk(sk);
3275 if (!tp->packets_out) {
3276 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3278 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3279 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3283 /* If we get here, the whole TSO packet has not been acked. */
3284 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3286 struct tcp_sock *tp = tcp_sk(sk);
3289 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3291 packets_acked = tcp_skb_pcount(skb);
3292 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3294 packets_acked -= tcp_skb_pcount(skb);
3296 if (packets_acked) {
3297 BUG_ON(tcp_skb_pcount(skb) == 0);
3298 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3301 return packets_acked;
3304 /* Remove acknowledged frames from the retransmission queue. If our packet
3305 * is before the ack sequence we can discard it as it's confirmed to have
3306 * arrived at the other end.
3308 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3311 struct tcp_sock *tp = tcp_sk(sk);
3312 const struct inet_connection_sock *icsk = inet_csk(sk);
3313 struct sk_buff *skb;
3314 u32 now = tcp_time_stamp;
3315 int fully_acked = 1;
3318 u32 reord = tp->packets_out;
3319 u32 prior_sacked = tp->sacked_out;
3321 s32 ca_seq_rtt = -1;
3322 ktime_t last_ackt = net_invalid_timestamp();
3324 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3325 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3327 u8 sacked = scb->sacked;
3329 /* Determine how many packets and what bytes were acked, tso and else */
3330 if (after(scb->end_seq, tp->snd_una)) {
3331 if (tcp_skb_pcount(skb) == 1 ||
3332 !after(tp->snd_una, scb->seq))
3335 acked_pcount = tcp_tso_acked(sk, skb);
3341 acked_pcount = tcp_skb_pcount(skb);
3344 if (sacked & TCPCB_RETRANS) {
3345 if (sacked & TCPCB_SACKED_RETRANS)
3346 tp->retrans_out -= acked_pcount;
3347 flag |= FLAG_RETRANS_DATA_ACKED;
3350 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3351 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3353 ca_seq_rtt = now - scb->when;
3354 last_ackt = skb->tstamp;
3356 seq_rtt = ca_seq_rtt;
3358 if (!(sacked & TCPCB_SACKED_ACKED))
3359 reord = min(pkts_acked, reord);
3362 if (sacked & TCPCB_SACKED_ACKED)
3363 tp->sacked_out -= acked_pcount;
3364 if (sacked & TCPCB_LOST)
3365 tp->lost_out -= acked_pcount;
3367 tp->packets_out -= acked_pcount;
3368 pkts_acked += acked_pcount;
3370 /* Initial outgoing SYN's get put onto the write_queue
3371 * just like anything else we transmit. It is not
3372 * true data, and if we misinform our callers that
3373 * this ACK acks real data, we will erroneously exit
3374 * connection startup slow start one packet too
3375 * quickly. This is severely frowned upon behavior.
3377 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3378 flag |= FLAG_DATA_ACKED;
3380 flag |= FLAG_SYN_ACKED;
3381 tp->retrans_stamp = 0;
3387 tcp_unlink_write_queue(skb, sk);
3388 sk_wmem_free_skb(sk, skb);
3389 tp->scoreboard_skb_hint = NULL;
3390 if (skb == tp->retransmit_skb_hint)
3391 tp->retransmit_skb_hint = NULL;
3392 if (skb == tp->lost_skb_hint)
3393 tp->lost_skb_hint = NULL;
3396 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3397 tp->snd_up = tp->snd_una;
3399 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3400 flag |= FLAG_SACK_RENEGING;
3402 if (flag & FLAG_ACKED) {
3403 const struct tcp_congestion_ops *ca_ops
3404 = inet_csk(sk)->icsk_ca_ops;
3406 if (unlikely(icsk->icsk_mtup.probe_size &&
3407 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3408 tcp_mtup_probe_success(sk);
3411 tcp_ack_update_rtt(sk, flag, seq_rtt);
3414 if (tcp_is_reno(tp)) {
3415 tcp_remove_reno_sacks(sk, pkts_acked);
3419 /* Non-retransmitted hole got filled? That's reordering */
3420 if (reord < prior_fackets)
3421 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3423 delta = tcp_is_fack(tp) ? pkts_acked :
3424 prior_sacked - tp->sacked_out;
3425 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3428 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3430 if (ca_ops->pkts_acked) {
3433 /* Is the ACK triggering packet unambiguous? */
3434 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3435 /* High resolution needed and available? */
3436 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3437 !ktime_equal(last_ackt,
3438 net_invalid_timestamp()))
3439 rtt_us = ktime_us_delta(ktime_get_real(),
3441 else if (ca_seq_rtt >= 0)
3442 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3445 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3449 #if FASTRETRANS_DEBUG > 0
3450 WARN_ON((int)tp->sacked_out < 0);
3451 WARN_ON((int)tp->lost_out < 0);
3452 WARN_ON((int)tp->retrans_out < 0);
3453 if (!tp->packets_out && tcp_is_sack(tp)) {
3454 icsk = inet_csk(sk);
3456 printk(KERN_DEBUG "Leak l=%u %d\n",
3457 tp->lost_out, icsk->icsk_ca_state);
3460 if (tp->sacked_out) {
3461 printk(KERN_DEBUG "Leak s=%u %d\n",
3462 tp->sacked_out, icsk->icsk_ca_state);
3465 if (tp->retrans_out) {
3466 printk(KERN_DEBUG "Leak r=%u %d\n",
3467 tp->retrans_out, icsk->icsk_ca_state);
3468 tp->retrans_out = 0;
3475 static void tcp_ack_probe(struct sock *sk)
3477 const struct tcp_sock *tp = tcp_sk(sk);
3478 struct inet_connection_sock *icsk = inet_csk(sk);
3480 /* Was it a usable window open? */
3482 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3483 icsk->icsk_backoff = 0;
3484 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3485 /* Socket must be waked up by subsequent tcp_data_snd_check().
3486 * This function is not for random using!
3489 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3490 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3495 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3497 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3498 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3501 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3503 const struct tcp_sock *tp = tcp_sk(sk);
3504 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3505 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3508 /* Check that window update is acceptable.
3509 * The function assumes that snd_una<=ack<=snd_next.
3511 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3512 const u32 ack, const u32 ack_seq,
3515 return after(ack, tp->snd_una) ||
3516 after(ack_seq, tp->snd_wl1) ||
3517 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3520 /* Update our send window.
3522 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3523 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3525 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3528 struct tcp_sock *tp = tcp_sk(sk);
3530 u32 nwin = ntohs(tcp_hdr(skb)->window);
3532 if (likely(!tcp_hdr(skb)->syn))
3533 nwin <<= tp->rx_opt.snd_wscale;
3535 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3536 flag |= FLAG_WIN_UPDATE;
3537 tcp_update_wl(tp, ack_seq);
3539 if (tp->snd_wnd != nwin) {
3542 /* Note, it is the only place, where
3543 * fast path is recovered for sending TCP.
3546 tcp_fast_path_check(sk);
3548 if (nwin > tp->max_window) {
3549 tp->max_window = nwin;
3550 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3560 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3561 * continue in congestion avoidance.
3563 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3565 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3566 tp->snd_cwnd_cnt = 0;
3567 tp->bytes_acked = 0;
3568 TCP_ECN_queue_cwr(tp);
3569 tcp_moderate_cwnd(tp);
3572 /* A conservative spurious RTO response algorithm: reduce cwnd using
3573 * rate halving and continue in congestion avoidance.
3575 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3577 tcp_enter_cwr(sk, 0);
3580 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3582 if (flag & FLAG_ECE)
3583 tcp_ratehalving_spur_to_response(sk);
3585 tcp_undo_cwr(sk, true);
3588 /* F-RTO spurious RTO detection algorithm (RFC4138)
3590 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3591 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3592 * window (but not to or beyond highest sequence sent before RTO):
3593 * On First ACK, send two new segments out.
3594 * On Second ACK, RTO was likely spurious. Do spurious response (response
3595 * algorithm is not part of the F-RTO detection algorithm
3596 * given in RFC4138 but can be selected separately).
3597 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3598 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3599 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3600 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3602 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3603 * original window even after we transmit two new data segments.
3606 * on first step, wait until first cumulative ACK arrives, then move to
3607 * the second step. In second step, the next ACK decides.
3609 * F-RTO is implemented (mainly) in four functions:
3610 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3611 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3612 * called when tcp_use_frto() showed green light
3613 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3614 * - tcp_enter_frto_loss() is called if there is not enough evidence
3615 * to prove that the RTO is indeed spurious. It transfers the control
3616 * from F-RTO to the conventional RTO recovery
3618 static int tcp_process_frto(struct sock *sk, int flag)
3620 struct tcp_sock *tp = tcp_sk(sk);
3622 tcp_verify_left_out(tp);
3624 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3625 if (flag & FLAG_DATA_ACKED)
3626 inet_csk(sk)->icsk_retransmits = 0;
3628 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3629 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3630 tp->undo_marker = 0;
3632 if (!before(tp->snd_una, tp->frto_highmark)) {
3633 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3637 if (!tcp_is_sackfrto(tp)) {
3638 /* RFC4138 shortcoming in step 2; should also have case c):
3639 * ACK isn't duplicate nor advances window, e.g., opposite dir
3642 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3645 if (!(flag & FLAG_DATA_ACKED)) {
3646 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3651 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3652 /* Prevent sending of new data. */
3653 tp->snd_cwnd = min(tp->snd_cwnd,
3654 tcp_packets_in_flight(tp));
3658 if ((tp->frto_counter >= 2) &&
3659 (!(flag & FLAG_FORWARD_PROGRESS) ||
3660 ((flag & FLAG_DATA_SACKED) &&
3661 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3662 /* RFC4138 shortcoming (see comment above) */
3663 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3664 (flag & FLAG_NOT_DUP))
3667 tcp_enter_frto_loss(sk, 3, flag);
3672 if (tp->frto_counter == 1) {
3673 /* tcp_may_send_now needs to see updated state */
3674 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3675 tp->frto_counter = 2;
3677 if (!tcp_may_send_now(sk))
3678 tcp_enter_frto_loss(sk, 2, flag);
3682 switch (sysctl_tcp_frto_response) {
3684 tcp_undo_spur_to_response(sk, flag);
3687 tcp_conservative_spur_to_response(tp);
3690 tcp_ratehalving_spur_to_response(sk);
3693 tp->frto_counter = 0;
3694 tp->undo_marker = 0;
3695 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3700 /* This routine deals with incoming acks, but not outgoing ones. */
3701 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3703 struct inet_connection_sock *icsk = inet_csk(sk);
3704 struct tcp_sock *tp = tcp_sk(sk);
3705 u32 prior_snd_una = tp->snd_una;
3706 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3707 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3708 u32 prior_in_flight;
3711 int prior_sacked = tp->sacked_out;
3712 int newly_acked_sacked = 0;
3715 /* If the ack is older than previous acks
3716 * then we can probably ignore it.
3718 if (before(ack, prior_snd_una))
3721 /* If the ack includes data we haven't sent yet, discard
3722 * this segment (RFC793 Section 3.9).
3724 if (after(ack, tp->snd_nxt))
3727 if (after(ack, prior_snd_una))
3728 flag |= FLAG_SND_UNA_ADVANCED;
3730 if (sysctl_tcp_abc) {
3731 if (icsk->icsk_ca_state < TCP_CA_CWR)
3732 tp->bytes_acked += ack - prior_snd_una;
3733 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3734 /* we assume just one segment left network */
3735 tp->bytes_acked += min(ack - prior_snd_una,
3739 prior_fackets = tp->fackets_out;
3740 prior_in_flight = tcp_packets_in_flight(tp);
3742 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3743 /* Window is constant, pure forward advance.
3744 * No more checks are required.
3745 * Note, we use the fact that SND.UNA>=SND.WL2.
3747 tcp_update_wl(tp, ack_seq);
3749 flag |= FLAG_WIN_UPDATE;
3751 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3753 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3755 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3758 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3760 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3762 if (TCP_SKB_CB(skb)->sacked)
3763 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3765 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3768 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3771 /* We passed data and got it acked, remove any soft error
3772 * log. Something worked...
3774 sk->sk_err_soft = 0;
3775 icsk->icsk_probes_out = 0;
3776 tp->rcv_tstamp = tcp_time_stamp;
3777 prior_packets = tp->packets_out;
3781 /* See if we can take anything off of the retransmit queue. */
3782 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3784 newly_acked_sacked = (prior_packets - prior_sacked) -
3785 (tp->packets_out - tp->sacked_out);
3787 if (tp->frto_counter)
3788 frto_cwnd = tcp_process_frto(sk, flag);
3789 /* Guarantee sacktag reordering detection against wrap-arounds */
3790 if (before(tp->frto_highmark, tp->snd_una))
3791 tp->frto_highmark = 0;
3793 if (tcp_ack_is_dubious(sk, flag)) {
3794 /* Advance CWND, if state allows this. */
3795 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3796 tcp_may_raise_cwnd(sk, flag))
3797 tcp_cong_avoid(sk, ack, prior_in_flight);
3798 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3799 newly_acked_sacked, flag);
3801 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3802 tcp_cong_avoid(sk, ack, prior_in_flight);
3805 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3806 dst_confirm(__sk_dst_get(sk));
3811 /* If this ack opens up a zero window, clear backoff. It was
3812 * being used to time the probes, and is probably far higher than
3813 * it needs to be for normal retransmission.
3815 if (tcp_send_head(sk))
3820 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3824 if (TCP_SKB_CB(skb)->sacked) {
3825 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3826 if (icsk->icsk_ca_state == TCP_CA_Open)
3827 tcp_try_keep_open(sk);
3830 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3834 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3835 * But, this can also be called on packets in the established flow when
3836 * the fast version below fails.
3838 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3839 const u8 **hvpp, int estab)
3841 const unsigned char *ptr;
3842 const struct tcphdr *th = tcp_hdr(skb);
3843 int length = (th->doff * 4) - sizeof(struct tcphdr);
3845 ptr = (const unsigned char *)(th + 1);
3846 opt_rx->saw_tstamp = 0;
3848 while (length > 0) {
3849 int opcode = *ptr++;
3855 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3860 if (opsize < 2) /* "silly options" */
3862 if (opsize > length)
3863 return; /* don't parse partial options */
3866 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3867 u16 in_mss = get_unaligned_be16(ptr);
3869 if (opt_rx->user_mss &&
3870 opt_rx->user_mss < in_mss)
3871 in_mss = opt_rx->user_mss;
3872 opt_rx->mss_clamp = in_mss;
3877 if (opsize == TCPOLEN_WINDOW && th->syn &&
3878 !estab && sysctl_tcp_window_scaling) {
3879 __u8 snd_wscale = *(__u8 *)ptr;
3880 opt_rx->wscale_ok = 1;
3881 if (snd_wscale > 14) {
3882 if (net_ratelimit())
3883 printk(KERN_INFO "tcp_parse_options: Illegal window "
3884 "scaling value %d >14 received.\n",
3888 opt_rx->snd_wscale = snd_wscale;
3891 case TCPOPT_TIMESTAMP:
3892 if ((opsize == TCPOLEN_TIMESTAMP) &&
3893 ((estab && opt_rx->tstamp_ok) ||
3894 (!estab && sysctl_tcp_timestamps))) {
3895 opt_rx->saw_tstamp = 1;
3896 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3897 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3900 case TCPOPT_SACK_PERM:
3901 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3902 !estab && sysctl_tcp_sack) {
3903 opt_rx->sack_ok = 1;
3904 tcp_sack_reset(opt_rx);
3909 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3910 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3912 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3915 #ifdef CONFIG_TCP_MD5SIG
3918 * The MD5 Hash has already been
3919 * checked (see tcp_v{4,6}_do_rcv()).
3924 /* This option is variable length.
3927 case TCPOLEN_COOKIE_BASE:
3928 /* not yet implemented */
3930 case TCPOLEN_COOKIE_PAIR:
3931 /* not yet implemented */
3933 case TCPOLEN_COOKIE_MIN+0:
3934 case TCPOLEN_COOKIE_MIN+2:
3935 case TCPOLEN_COOKIE_MIN+4:
3936 case TCPOLEN_COOKIE_MIN+6:
3937 case TCPOLEN_COOKIE_MAX:
3938 /* 16-bit multiple */
3939 opt_rx->cookie_plus = opsize;
3954 EXPORT_SYMBOL(tcp_parse_options);
3956 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3958 const __be32 *ptr = (const __be32 *)(th + 1);
3960 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3961 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3962 tp->rx_opt.saw_tstamp = 1;
3964 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3966 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3972 /* Fast parse options. This hopes to only see timestamps.
3973 * If it is wrong it falls back on tcp_parse_options().
3975 static int tcp_fast_parse_options(const struct sk_buff *skb,
3976 const struct tcphdr *th,
3977 struct tcp_sock *tp, const u8 **hvpp)
3979 /* In the spirit of fast parsing, compare doff directly to constant
3980 * values. Because equality is used, short doff can be ignored here.
3982 if (th->doff == (sizeof(*th) / 4)) {
3983 tp->rx_opt.saw_tstamp = 0;
3985 } else if (tp->rx_opt.tstamp_ok &&
3986 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3987 if (tcp_parse_aligned_timestamp(tp, th))
3990 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3994 #ifdef CONFIG_TCP_MD5SIG
3996 * Parse MD5 Signature option
3998 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4000 int length = (th->doff << 2) - sizeof(*th);
4001 const u8 *ptr = (const u8 *)(th + 1);
4003 /* If the TCP option is too short, we can short cut */
4004 if (length < TCPOLEN_MD5SIG)
4007 while (length > 0) {
4008 int opcode = *ptr++;
4019 if (opsize < 2 || opsize > length)
4021 if (opcode == TCPOPT_MD5SIG)
4022 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4029 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4032 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4034 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
4035 tp->rx_opt.ts_recent_stamp = get_seconds();
4038 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4040 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4041 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
4042 * extra check below makes sure this can only happen
4043 * for pure ACK frames. -DaveM
4045 * Not only, also it occurs for expired timestamps.
4048 if (tcp_paws_check(&tp->rx_opt, 0))
4049 tcp_store_ts_recent(tp);
4053 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4055 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4056 * it can pass through stack. So, the following predicate verifies that
4057 * this segment is not used for anything but congestion avoidance or
4058 * fast retransmit. Moreover, we even are able to eliminate most of such
4059 * second order effects, if we apply some small "replay" window (~RTO)
4060 * to timestamp space.
4062 * All these measures still do not guarantee that we reject wrapped ACKs
4063 * on networks with high bandwidth, when sequence space is recycled fastly,
4064 * but it guarantees that such events will be very rare and do not affect
4065 * connection seriously. This doesn't look nice, but alas, PAWS is really
4068 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4069 * states that events when retransmit arrives after original data are rare.
4070 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4071 * the biggest problem on large power networks even with minor reordering.
4072 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4073 * up to bandwidth of 18Gigabit/sec. 8) ]
4076 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4078 const struct tcp_sock *tp = tcp_sk(sk);
4079 const struct tcphdr *th = tcp_hdr(skb);
4080 u32 seq = TCP_SKB_CB(skb)->seq;
4081 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4083 return (/* 1. Pure ACK with correct sequence number. */
4084 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4086 /* 2. ... and duplicate ACK. */
4087 ack == tp->snd_una &&
4089 /* 3. ... and does not update window. */
4090 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4092 /* 4. ... and sits in replay window. */
4093 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4096 static inline int tcp_paws_discard(const struct sock *sk,
4097 const struct sk_buff *skb)
4099 const struct tcp_sock *tp = tcp_sk(sk);
4101 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4102 !tcp_disordered_ack(sk, skb);
4105 /* Check segment sequence number for validity.
4107 * Segment controls are considered valid, if the segment
4108 * fits to the window after truncation to the window. Acceptability
4109 * of data (and SYN, FIN, of course) is checked separately.
4110 * See tcp_data_queue(), for example.
4112 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4113 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4114 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4115 * (borrowed from freebsd)
4118 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4120 return !before(end_seq, tp->rcv_wup) &&
4121 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4124 /* When we get a reset we do this. */
4125 static void tcp_reset(struct sock *sk)
4127 /* We want the right error as BSD sees it (and indeed as we do). */
4128 switch (sk->sk_state) {
4130 sk->sk_err = ECONNREFUSED;
4132 case TCP_CLOSE_WAIT:
4138 sk->sk_err = ECONNRESET;
4140 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4143 if (!sock_flag(sk, SOCK_DEAD))
4144 sk->sk_error_report(sk);
4150 * Process the FIN bit. This now behaves as it is supposed to work
4151 * and the FIN takes effect when it is validly part of sequence
4152 * space. Not before when we get holes.
4154 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4155 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4158 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4159 * close and we go into CLOSING (and later onto TIME-WAIT)
4161 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4163 static void tcp_fin(struct sock *sk)
4165 struct tcp_sock *tp = tcp_sk(sk);
4167 inet_csk_schedule_ack(sk);
4169 sk->sk_shutdown |= RCV_SHUTDOWN;
4170 sock_set_flag(sk, SOCK_DONE);
4172 switch (sk->sk_state) {
4174 case TCP_ESTABLISHED:
4175 /* Move to CLOSE_WAIT */
4176 tcp_set_state(sk, TCP_CLOSE_WAIT);
4177 inet_csk(sk)->icsk_ack.pingpong = 1;
4180 case TCP_CLOSE_WAIT:
4182 /* Received a retransmission of the FIN, do
4187 /* RFC793: Remain in the LAST-ACK state. */
4191 /* This case occurs when a simultaneous close
4192 * happens, we must ack the received FIN and
4193 * enter the CLOSING state.
4196 tcp_set_state(sk, TCP_CLOSING);
4199 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4201 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4204 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4205 * cases we should never reach this piece of code.
4207 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4208 __func__, sk->sk_state);
4212 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4213 * Probably, we should reset in this case. For now drop them.
4215 __skb_queue_purge(&tp->out_of_order_queue);
4216 if (tcp_is_sack(tp))
4217 tcp_sack_reset(&tp->rx_opt);
4220 if (!sock_flag(sk, SOCK_DEAD)) {
4221 sk->sk_state_change(sk);
4223 /* Do not send POLL_HUP for half duplex close. */
4224 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4225 sk->sk_state == TCP_CLOSE)
4226 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4228 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4232 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4235 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4236 if (before(seq, sp->start_seq))
4237 sp->start_seq = seq;
4238 if (after(end_seq, sp->end_seq))
4239 sp->end_seq = end_seq;
4245 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4247 struct tcp_sock *tp = tcp_sk(sk);
4249 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4252 if (before(seq, tp->rcv_nxt))
4253 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4255 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4257 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4259 tp->rx_opt.dsack = 1;
4260 tp->duplicate_sack[0].start_seq = seq;
4261 tp->duplicate_sack[0].end_seq = end_seq;
4265 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4267 struct tcp_sock *tp = tcp_sk(sk);
4269 if (!tp->rx_opt.dsack)
4270 tcp_dsack_set(sk, seq, end_seq);
4272 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4275 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4277 struct tcp_sock *tp = tcp_sk(sk);
4279 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4280 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4281 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4282 tcp_enter_quickack_mode(sk);
4284 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4285 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4287 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4288 end_seq = tp->rcv_nxt;
4289 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4296 /* These routines update the SACK block as out-of-order packets arrive or
4297 * in-order packets close up the sequence space.
4299 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4302 struct tcp_sack_block *sp = &tp->selective_acks[0];
4303 struct tcp_sack_block *swalk = sp + 1;
4305 /* See if the recent change to the first SACK eats into
4306 * or hits the sequence space of other SACK blocks, if so coalesce.
4308 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4309 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4312 /* Zap SWALK, by moving every further SACK up by one slot.
4313 * Decrease num_sacks.
4315 tp->rx_opt.num_sacks--;
4316 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4320 this_sack++, swalk++;
4324 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4326 struct tcp_sock *tp = tcp_sk(sk);
4327 struct tcp_sack_block *sp = &tp->selective_acks[0];
4328 int cur_sacks = tp->rx_opt.num_sacks;
4334 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4335 if (tcp_sack_extend(sp, seq, end_seq)) {
4336 /* Rotate this_sack to the first one. */
4337 for (; this_sack > 0; this_sack--, sp--)
4338 swap(*sp, *(sp - 1));
4340 tcp_sack_maybe_coalesce(tp);
4345 /* Could not find an adjacent existing SACK, build a new one,
4346 * put it at the front, and shift everyone else down. We
4347 * always know there is at least one SACK present already here.
4349 * If the sack array is full, forget about the last one.
4351 if (this_sack >= TCP_NUM_SACKS) {
4353 tp->rx_opt.num_sacks--;
4356 for (; this_sack > 0; this_sack--, sp--)
4360 /* Build the new head SACK, and we're done. */
4361 sp->start_seq = seq;
4362 sp->end_seq = end_seq;
4363 tp->rx_opt.num_sacks++;
4366 /* RCV.NXT advances, some SACKs should be eaten. */
4368 static void tcp_sack_remove(struct tcp_sock *tp)
4370 struct tcp_sack_block *sp = &tp->selective_acks[0];
4371 int num_sacks = tp->rx_opt.num_sacks;
4374 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4375 if (skb_queue_empty(&tp->out_of_order_queue)) {
4376 tp->rx_opt.num_sacks = 0;
4380 for (this_sack = 0; this_sack < num_sacks;) {
4381 /* Check if the start of the sack is covered by RCV.NXT. */
4382 if (!before(tp->rcv_nxt, sp->start_seq)) {
4385 /* RCV.NXT must cover all the block! */
4386 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4388 /* Zap this SACK, by moving forward any other SACKS. */
4389 for (i=this_sack+1; i < num_sacks; i++)
4390 tp->selective_acks[i-1] = tp->selective_acks[i];
4397 tp->rx_opt.num_sacks = num_sacks;
4400 /* This one checks to see if we can put data from the
4401 * out_of_order queue into the receive_queue.
4403 static void tcp_ofo_queue(struct sock *sk)
4405 struct tcp_sock *tp = tcp_sk(sk);
4406 __u32 dsack_high = tp->rcv_nxt;
4407 struct sk_buff *skb;
4409 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4410 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4413 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4414 __u32 dsack = dsack_high;
4415 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4416 dsack_high = TCP_SKB_CB(skb)->end_seq;
4417 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4420 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4421 SOCK_DEBUG(sk, "ofo packet was already received\n");
4422 __skb_unlink(skb, &tp->out_of_order_queue);
4426 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4427 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4428 TCP_SKB_CB(skb)->end_seq);
4430 __skb_unlink(skb, &tp->out_of_order_queue);
4431 __skb_queue_tail(&sk->sk_receive_queue, skb);
4432 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4433 if (tcp_hdr(skb)->fin)
4438 static int tcp_prune_ofo_queue(struct sock *sk);
4439 static int tcp_prune_queue(struct sock *sk);
4441 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4443 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4444 !sk_rmem_schedule(sk, size)) {
4446 if (tcp_prune_queue(sk) < 0)
4449 if (!sk_rmem_schedule(sk, size)) {
4450 if (!tcp_prune_ofo_queue(sk))
4453 if (!sk_rmem_schedule(sk, size))
4460 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4462 const struct tcphdr *th = tcp_hdr(skb);
4463 struct tcp_sock *tp = tcp_sk(sk);
4466 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4470 __skb_pull(skb, th->doff * 4);
4472 TCP_ECN_accept_cwr(tp, skb);
4474 tp->rx_opt.dsack = 0;
4476 /* Queue data for delivery to the user.
4477 * Packets in sequence go to the receive queue.
4478 * Out of sequence packets to the out_of_order_queue.
4480 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4481 if (tcp_receive_window(tp) == 0)
4484 /* Ok. In sequence. In window. */
4485 if (tp->ucopy.task == current &&
4486 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4487 sock_owned_by_user(sk) && !tp->urg_data) {
4488 int chunk = min_t(unsigned int, skb->len,
4491 __set_current_state(TASK_RUNNING);
4494 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4495 tp->ucopy.len -= chunk;
4496 tp->copied_seq += chunk;
4497 eaten = (chunk == skb->len);
4498 tcp_rcv_space_adjust(sk);
4506 tcp_try_rmem_schedule(sk, skb->truesize))
4509 skb_set_owner_r(skb, sk);
4510 __skb_queue_tail(&sk->sk_receive_queue, skb);
4512 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4514 tcp_event_data_recv(sk, skb);
4518 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4521 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4522 * gap in queue is filled.
4524 if (skb_queue_empty(&tp->out_of_order_queue))
4525 inet_csk(sk)->icsk_ack.pingpong = 0;
4528 if (tp->rx_opt.num_sacks)
4529 tcp_sack_remove(tp);
4531 tcp_fast_path_check(sk);
4535 else if (!sock_flag(sk, SOCK_DEAD))
4536 sk->sk_data_ready(sk, 0);
4540 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4541 /* A retransmit, 2nd most common case. Force an immediate ack. */
4542 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4543 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4546 tcp_enter_quickack_mode(sk);
4547 inet_csk_schedule_ack(sk);
4553 /* Out of window. F.e. zero window probe. */
4554 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4557 tcp_enter_quickack_mode(sk);
4559 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4560 /* Partial packet, seq < rcv_next < end_seq */
4561 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4562 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4563 TCP_SKB_CB(skb)->end_seq);
4565 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4567 /* If window is closed, drop tail of packet. But after
4568 * remembering D-SACK for its head made in previous line.
4570 if (!tcp_receive_window(tp))
4575 TCP_ECN_check_ce(tp, skb);
4577 if (tcp_try_rmem_schedule(sk, skb->truesize))
4580 /* Disable header prediction. */
4582 inet_csk_schedule_ack(sk);
4584 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4585 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4587 skb_set_owner_r(skb, sk);
4589 if (!skb_peek(&tp->out_of_order_queue)) {
4590 /* Initial out of order segment, build 1 SACK. */
4591 if (tcp_is_sack(tp)) {
4592 tp->rx_opt.num_sacks = 1;
4593 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4594 tp->selective_acks[0].end_seq =
4595 TCP_SKB_CB(skb)->end_seq;
4597 __skb_queue_head(&tp->out_of_order_queue, skb);
4599 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4600 u32 seq = TCP_SKB_CB(skb)->seq;
4601 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4603 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4604 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4606 if (!tp->rx_opt.num_sacks ||
4607 tp->selective_acks[0].end_seq != seq)
4610 /* Common case: data arrive in order after hole. */
4611 tp->selective_acks[0].end_seq = end_seq;
4615 /* Find place to insert this segment. */
4617 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4619 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4623 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4626 /* Do skb overlap to previous one? */
4627 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4628 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4629 /* All the bits are present. Drop. */
4631 tcp_dsack_set(sk, seq, end_seq);
4634 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4635 /* Partial overlap. */
4636 tcp_dsack_set(sk, seq,
4637 TCP_SKB_CB(skb1)->end_seq);
4639 if (skb_queue_is_first(&tp->out_of_order_queue,
4643 skb1 = skb_queue_prev(
4644 &tp->out_of_order_queue,
4649 __skb_queue_head(&tp->out_of_order_queue, skb);
4651 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4653 /* And clean segments covered by new one as whole. */
4654 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4655 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4657 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4659 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4660 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4664 __skb_unlink(skb1, &tp->out_of_order_queue);
4665 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4666 TCP_SKB_CB(skb1)->end_seq);
4671 if (tcp_is_sack(tp))
4672 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4676 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4677 struct sk_buff_head *list)
4679 struct sk_buff *next = NULL;
4681 if (!skb_queue_is_last(list, skb))
4682 next = skb_queue_next(list, skb);
4684 __skb_unlink(skb, list);
4686 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4691 /* Collapse contiguous sequence of skbs head..tail with
4692 * sequence numbers start..end.
4694 * If tail is NULL, this means until the end of the list.
4696 * Segments with FIN/SYN are not collapsed (only because this
4700 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4701 struct sk_buff *head, struct sk_buff *tail,
4704 struct sk_buff *skb, *n;
4707 /* First, check that queue is collapsible and find
4708 * the point where collapsing can be useful. */
4712 skb_queue_walk_from_safe(list, skb, n) {
4715 /* No new bits? It is possible on ofo queue. */
4716 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4717 skb = tcp_collapse_one(sk, skb, list);
4723 /* The first skb to collapse is:
4725 * - bloated or contains data before "start" or
4726 * overlaps to the next one.
4728 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4729 (tcp_win_from_space(skb->truesize) > skb->len ||
4730 before(TCP_SKB_CB(skb)->seq, start))) {
4731 end_of_skbs = false;
4735 if (!skb_queue_is_last(list, skb)) {
4736 struct sk_buff *next = skb_queue_next(list, skb);
4738 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4739 end_of_skbs = false;
4744 /* Decided to skip this, advance start seq. */
4745 start = TCP_SKB_CB(skb)->end_seq;
4747 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4750 while (before(start, end)) {
4751 struct sk_buff *nskb;
4752 unsigned int header = skb_headroom(skb);
4753 int copy = SKB_MAX_ORDER(header, 0);
4755 /* Too big header? This can happen with IPv6. */
4758 if (end - start < copy)
4760 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4764 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4765 skb_set_network_header(nskb, (skb_network_header(skb) -
4767 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4769 skb_reserve(nskb, header);
4770 memcpy(nskb->head, skb->head, header);
4771 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4772 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4773 __skb_queue_before(list, skb, nskb);
4774 skb_set_owner_r(nskb, sk);
4776 /* Copy data, releasing collapsed skbs. */
4778 int offset = start - TCP_SKB_CB(skb)->seq;
4779 int size = TCP_SKB_CB(skb)->end_seq - start;
4783 size = min(copy, size);
4784 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4786 TCP_SKB_CB(nskb)->end_seq += size;
4790 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4791 skb = tcp_collapse_one(sk, skb, list);
4794 tcp_hdr(skb)->syn ||
4802 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4803 * and tcp_collapse() them until all the queue is collapsed.
4805 static void tcp_collapse_ofo_queue(struct sock *sk)
4807 struct tcp_sock *tp = tcp_sk(sk);
4808 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4809 struct sk_buff *head;
4815 start = TCP_SKB_CB(skb)->seq;
4816 end = TCP_SKB_CB(skb)->end_seq;
4820 struct sk_buff *next = NULL;
4822 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4823 next = skb_queue_next(&tp->out_of_order_queue, skb);
4826 /* Segment is terminated when we see gap or when
4827 * we are at the end of all the queue. */
4829 after(TCP_SKB_CB(skb)->seq, end) ||
4830 before(TCP_SKB_CB(skb)->end_seq, start)) {
4831 tcp_collapse(sk, &tp->out_of_order_queue,
4832 head, skb, start, end);
4836 /* Start new segment */
4837 start = TCP_SKB_CB(skb)->seq;
4838 end = TCP_SKB_CB(skb)->end_seq;
4840 if (before(TCP_SKB_CB(skb)->seq, start))
4841 start = TCP_SKB_CB(skb)->seq;
4842 if (after(TCP_SKB_CB(skb)->end_seq, end))
4843 end = TCP_SKB_CB(skb)->end_seq;
4849 * Purge the out-of-order queue.
4850 * Return true if queue was pruned.
4852 static int tcp_prune_ofo_queue(struct sock *sk)
4854 struct tcp_sock *tp = tcp_sk(sk);
4857 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4858 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4859 __skb_queue_purge(&tp->out_of_order_queue);
4861 /* Reset SACK state. A conforming SACK implementation will
4862 * do the same at a timeout based retransmit. When a connection
4863 * is in a sad state like this, we care only about integrity
4864 * of the connection not performance.
4866 if (tp->rx_opt.sack_ok)
4867 tcp_sack_reset(&tp->rx_opt);
4874 /* Reduce allocated memory if we can, trying to get
4875 * the socket within its memory limits again.
4877 * Return less than zero if we should start dropping frames
4878 * until the socket owning process reads some of the data
4879 * to stabilize the situation.
4881 static int tcp_prune_queue(struct sock *sk)
4883 struct tcp_sock *tp = tcp_sk(sk);
4885 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4887 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4889 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4890 tcp_clamp_window(sk);
4891 else if (tcp_memory_pressure)
4892 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4894 tcp_collapse_ofo_queue(sk);
4895 if (!skb_queue_empty(&sk->sk_receive_queue))
4896 tcp_collapse(sk, &sk->sk_receive_queue,
4897 skb_peek(&sk->sk_receive_queue),
4899 tp->copied_seq, tp->rcv_nxt);
4902 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4905 /* Collapsing did not help, destructive actions follow.
4906 * This must not ever occur. */
4908 tcp_prune_ofo_queue(sk);
4910 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4913 /* If we are really being abused, tell the caller to silently
4914 * drop receive data on the floor. It will get retransmitted
4915 * and hopefully then we'll have sufficient space.
4917 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4919 /* Massive buffer overcommit. */
4924 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4925 * As additional protections, we do not touch cwnd in retransmission phases,
4926 * and if application hit its sndbuf limit recently.
4928 void tcp_cwnd_application_limited(struct sock *sk)
4930 struct tcp_sock *tp = tcp_sk(sk);
4932 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4933 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4934 /* Limited by application or receiver window. */
4935 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4936 u32 win_used = max(tp->snd_cwnd_used, init_win);
4937 if (win_used < tp->snd_cwnd) {
4938 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4939 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4941 tp->snd_cwnd_used = 0;
4943 tp->snd_cwnd_stamp = tcp_time_stamp;
4946 static int tcp_should_expand_sndbuf(const struct sock *sk)
4948 const struct tcp_sock *tp = tcp_sk(sk);
4950 /* If the user specified a specific send buffer setting, do
4953 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4956 /* If we are under global TCP memory pressure, do not expand. */
4957 if (tcp_memory_pressure)
4960 /* If we are under soft global TCP memory pressure, do not expand. */
4961 if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4964 /* If we filled the congestion window, do not expand. */
4965 if (tp->packets_out >= tp->snd_cwnd)
4971 /* When incoming ACK allowed to free some skb from write_queue,
4972 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4973 * on the exit from tcp input handler.
4975 * PROBLEM: sndbuf expansion does not work well with largesend.
4977 static void tcp_new_space(struct sock *sk)
4979 struct tcp_sock *tp = tcp_sk(sk);
4981 if (tcp_should_expand_sndbuf(sk)) {
4982 int sndmem = SKB_TRUESIZE(max_t(u32,
4983 tp->rx_opt.mss_clamp,
4986 int demanded = max_t(unsigned int, tp->snd_cwnd,
4987 tp->reordering + 1);
4988 sndmem *= 2 * demanded;
4989 if (sndmem > sk->sk_sndbuf)
4990 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4991 tp->snd_cwnd_stamp = tcp_time_stamp;
4994 sk->sk_write_space(sk);
4997 static void tcp_check_space(struct sock *sk)
4999 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5000 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5001 if (sk->sk_socket &&
5002 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5007 static inline void tcp_data_snd_check(struct sock *sk)
5009 tcp_push_pending_frames(sk);
5010 tcp_check_space(sk);
5014 * Check if sending an ack is needed.
5016 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5018 struct tcp_sock *tp = tcp_sk(sk);
5020 /* More than one full frame received... */
5021 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5022 /* ... and right edge of window advances far enough.
5023 * (tcp_recvmsg() will send ACK otherwise). Or...
5025 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5026 /* We ACK each frame or... */
5027 tcp_in_quickack_mode(sk) ||
5028 /* We have out of order data. */
5029 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5030 /* Then ack it now */
5033 /* Else, send delayed ack. */
5034 tcp_send_delayed_ack(sk);
5038 static inline void tcp_ack_snd_check(struct sock *sk)
5040 if (!inet_csk_ack_scheduled(sk)) {
5041 /* We sent a data segment already. */
5044 __tcp_ack_snd_check(sk, 1);
5048 * This routine is only called when we have urgent data
5049 * signaled. Its the 'slow' part of tcp_urg. It could be
5050 * moved inline now as tcp_urg is only called from one
5051 * place. We handle URGent data wrong. We have to - as
5052 * BSD still doesn't use the correction from RFC961.
5053 * For 1003.1g we should support a new option TCP_STDURG to permit
5054 * either form (or just set the sysctl tcp_stdurg).
5057 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5059 struct tcp_sock *tp = tcp_sk(sk);
5060 u32 ptr = ntohs(th->urg_ptr);
5062 if (ptr && !sysctl_tcp_stdurg)
5064 ptr += ntohl(th->seq);
5066 /* Ignore urgent data that we've already seen and read. */
5067 if (after(tp->copied_seq, ptr))
5070 /* Do not replay urg ptr.
5072 * NOTE: interesting situation not covered by specs.
5073 * Misbehaving sender may send urg ptr, pointing to segment,
5074 * which we already have in ofo queue. We are not able to fetch
5075 * such data and will stay in TCP_URG_NOTYET until will be eaten
5076 * by recvmsg(). Seems, we are not obliged to handle such wicked
5077 * situations. But it is worth to think about possibility of some
5078 * DoSes using some hypothetical application level deadlock.
5080 if (before(ptr, tp->rcv_nxt))
5083 /* Do we already have a newer (or duplicate) urgent pointer? */
5084 if (tp->urg_data && !after(ptr, tp->urg_seq))
5087 /* Tell the world about our new urgent pointer. */
5090 /* We may be adding urgent data when the last byte read was
5091 * urgent. To do this requires some care. We cannot just ignore
5092 * tp->copied_seq since we would read the last urgent byte again
5093 * as data, nor can we alter copied_seq until this data arrives
5094 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5096 * NOTE. Double Dutch. Rendering to plain English: author of comment
5097 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5098 * and expect that both A and B disappear from stream. This is _wrong_.
5099 * Though this happens in BSD with high probability, this is occasional.
5100 * Any application relying on this is buggy. Note also, that fix "works"
5101 * only in this artificial test. Insert some normal data between A and B and we will
5102 * decline of BSD again. Verdict: it is better to remove to trap
5105 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5106 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5107 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5109 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5110 __skb_unlink(skb, &sk->sk_receive_queue);
5115 tp->urg_data = TCP_URG_NOTYET;
5118 /* Disable header prediction. */
5122 /* This is the 'fast' part of urgent handling. */
5123 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5125 struct tcp_sock *tp = tcp_sk(sk);
5127 /* Check if we get a new urgent pointer - normally not. */
5129 tcp_check_urg(sk, th);
5131 /* Do we wait for any urgent data? - normally not... */
5132 if (tp->urg_data == TCP_URG_NOTYET) {
5133 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5136 /* Is the urgent pointer pointing into this packet? */
5137 if (ptr < skb->len) {
5139 if (skb_copy_bits(skb, ptr, &tmp, 1))
5141 tp->urg_data = TCP_URG_VALID | tmp;
5142 if (!sock_flag(sk, SOCK_DEAD))
5143 sk->sk_data_ready(sk, 0);
5148 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5150 struct tcp_sock *tp = tcp_sk(sk);
5151 int chunk = skb->len - hlen;
5155 if (skb_csum_unnecessary(skb))
5156 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5158 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5162 tp->ucopy.len -= chunk;
5163 tp->copied_seq += chunk;
5164 tcp_rcv_space_adjust(sk);
5171 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5172 struct sk_buff *skb)
5176 if (sock_owned_by_user(sk)) {
5178 result = __tcp_checksum_complete(skb);
5181 result = __tcp_checksum_complete(skb);
5186 static inline int tcp_checksum_complete_user(struct sock *sk,
5187 struct sk_buff *skb)
5189 return !skb_csum_unnecessary(skb) &&
5190 __tcp_checksum_complete_user(sk, skb);
5193 #ifdef CONFIG_NET_DMA
5194 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5197 struct tcp_sock *tp = tcp_sk(sk);
5198 int chunk = skb->len - hlen;
5200 int copied_early = 0;
5202 if (tp->ucopy.wakeup)
5205 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5206 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5208 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5210 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5212 tp->ucopy.iov, chunk,
5213 tp->ucopy.pinned_list);
5218 tp->ucopy.dma_cookie = dma_cookie;
5221 tp->ucopy.len -= chunk;
5222 tp->copied_seq += chunk;
5223 tcp_rcv_space_adjust(sk);
5225 if ((tp->ucopy.len == 0) ||
5226 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5227 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5228 tp->ucopy.wakeup = 1;
5229 sk->sk_data_ready(sk, 0);
5231 } else if (chunk > 0) {
5232 tp->ucopy.wakeup = 1;
5233 sk->sk_data_ready(sk, 0);
5236 return copied_early;
5238 #endif /* CONFIG_NET_DMA */
5240 /* Does PAWS and seqno based validation of an incoming segment, flags will
5241 * play significant role here.
5243 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5244 const struct tcphdr *th, int syn_inerr)
5246 const u8 *hash_location;
5247 struct tcp_sock *tp = tcp_sk(sk);
5249 /* RFC1323: H1. Apply PAWS check first. */
5250 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5251 tp->rx_opt.saw_tstamp &&
5252 tcp_paws_discard(sk, skb)) {
5254 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5255 tcp_send_dupack(sk, skb);
5258 /* Reset is accepted even if it did not pass PAWS. */
5261 /* Step 1: check sequence number */
5262 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5263 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5264 * (RST) segments are validated by checking their SEQ-fields."
5265 * And page 69: "If an incoming segment is not acceptable,
5266 * an acknowledgment should be sent in reply (unless the RST
5267 * bit is set, if so drop the segment and return)".
5270 tcp_send_dupack(sk, skb);
5274 /* Step 2: check RST bit */
5280 /* ts_recent update must be made after we are sure that the packet
5283 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5285 /* step 3: check security and precedence [ignored] */
5287 /* step 4: Check for a SYN in window. */
5288 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5290 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5291 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5304 * TCP receive function for the ESTABLISHED state.
5306 * It is split into a fast path and a slow path. The fast path is
5308 * - A zero window was announced from us - zero window probing
5309 * is only handled properly in the slow path.
5310 * - Out of order segments arrived.
5311 * - Urgent data is expected.
5312 * - There is no buffer space left
5313 * - Unexpected TCP flags/window values/header lengths are received
5314 * (detected by checking the TCP header against pred_flags)
5315 * - Data is sent in both directions. Fast path only supports pure senders
5316 * or pure receivers (this means either the sequence number or the ack
5317 * value must stay constant)
5318 * - Unexpected TCP option.
5320 * When these conditions are not satisfied it drops into a standard
5321 * receive procedure patterned after RFC793 to handle all cases.
5322 * The first three cases are guaranteed by proper pred_flags setting,
5323 * the rest is checked inline. Fast processing is turned on in
5324 * tcp_data_queue when everything is OK.
5326 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5327 const struct tcphdr *th, unsigned int len)
5329 struct tcp_sock *tp = tcp_sk(sk);
5333 * Header prediction.
5334 * The code loosely follows the one in the famous
5335 * "30 instruction TCP receive" Van Jacobson mail.
5337 * Van's trick is to deposit buffers into socket queue
5338 * on a device interrupt, to call tcp_recv function
5339 * on the receive process context and checksum and copy
5340 * the buffer to user space. smart...
5342 * Our current scheme is not silly either but we take the
5343 * extra cost of the net_bh soft interrupt processing...
5344 * We do checksum and copy also but from device to kernel.
5347 tp->rx_opt.saw_tstamp = 0;
5349 /* pred_flags is 0xS?10 << 16 + snd_wnd
5350 * if header_prediction is to be made
5351 * 'S' will always be tp->tcp_header_len >> 2
5352 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5353 * turn it off (when there are holes in the receive
5354 * space for instance)
5355 * PSH flag is ignored.
5358 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5359 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5360 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5361 int tcp_header_len = tp->tcp_header_len;
5363 /* Timestamp header prediction: tcp_header_len
5364 * is automatically equal to th->doff*4 due to pred_flags
5368 /* Check timestamp */
5369 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5370 /* No? Slow path! */
5371 if (!tcp_parse_aligned_timestamp(tp, th))
5374 /* If PAWS failed, check it more carefully in slow path */
5375 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5378 /* DO NOT update ts_recent here, if checksum fails
5379 * and timestamp was corrupted part, it will result
5380 * in a hung connection since we will drop all
5381 * future packets due to the PAWS test.
5385 if (len <= tcp_header_len) {
5386 /* Bulk data transfer: sender */
5387 if (len == tcp_header_len) {
5388 /* Predicted packet is in window by definition.
5389 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5390 * Hence, check seq<=rcv_wup reduces to:
5392 if (tcp_header_len ==
5393 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5394 tp->rcv_nxt == tp->rcv_wup)
5395 tcp_store_ts_recent(tp);
5397 /* We know that such packets are checksummed
5400 tcp_ack(sk, skb, 0);
5402 tcp_data_snd_check(sk);
5404 } else { /* Header too small */
5405 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5410 int copied_early = 0;
5412 if (tp->copied_seq == tp->rcv_nxt &&
5413 len - tcp_header_len <= tp->ucopy.len) {
5414 #ifdef CONFIG_NET_DMA
5415 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5420 if (tp->ucopy.task == current &&
5421 sock_owned_by_user(sk) && !copied_early) {
5422 __set_current_state(TASK_RUNNING);
5424 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5428 /* Predicted packet is in window by definition.
5429 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5430 * Hence, check seq<=rcv_wup reduces to:
5432 if (tcp_header_len ==
5433 (sizeof(struct tcphdr) +
5434 TCPOLEN_TSTAMP_ALIGNED) &&
5435 tp->rcv_nxt == tp->rcv_wup)
5436 tcp_store_ts_recent(tp);
5438 tcp_rcv_rtt_measure_ts(sk, skb);
5440 __skb_pull(skb, tcp_header_len);
5441 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5442 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5445 tcp_cleanup_rbuf(sk, skb->len);
5448 if (tcp_checksum_complete_user(sk, skb))
5451 /* Predicted packet is in window by definition.
5452 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5453 * Hence, check seq<=rcv_wup reduces to:
5455 if (tcp_header_len ==
5456 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5457 tp->rcv_nxt == tp->rcv_wup)
5458 tcp_store_ts_recent(tp);
5460 tcp_rcv_rtt_measure_ts(sk, skb);
5462 if ((int)skb->truesize > sk->sk_forward_alloc)
5465 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5467 /* Bulk data transfer: receiver */
5468 __skb_pull(skb, tcp_header_len);
5469 __skb_queue_tail(&sk->sk_receive_queue, skb);
5470 skb_set_owner_r(skb, sk);
5471 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5474 tcp_event_data_recv(sk, skb);
5476 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5477 /* Well, only one small jumplet in fast path... */
5478 tcp_ack(sk, skb, FLAG_DATA);
5479 tcp_data_snd_check(sk);
5480 if (!inet_csk_ack_scheduled(sk))
5484 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5485 __tcp_ack_snd_check(sk, 0);
5487 #ifdef CONFIG_NET_DMA
5489 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5495 sk->sk_data_ready(sk, 0);
5501 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5505 * Standard slow path.
5508 res = tcp_validate_incoming(sk, skb, th, 1);
5513 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5516 tcp_rcv_rtt_measure_ts(sk, skb);
5518 /* Process urgent data. */
5519 tcp_urg(sk, skb, th);
5521 /* step 7: process the segment text */
5522 tcp_data_queue(sk, skb);
5524 tcp_data_snd_check(sk);
5525 tcp_ack_snd_check(sk);
5529 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5535 EXPORT_SYMBOL(tcp_rcv_established);
5537 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5538 const struct tcphdr *th, unsigned int len)
5540 const u8 *hash_location;
5541 struct inet_connection_sock *icsk = inet_csk(sk);
5542 struct tcp_sock *tp = tcp_sk(sk);
5543 struct tcp_cookie_values *cvp = tp->cookie_values;
5544 int saved_clamp = tp->rx_opt.mss_clamp;
5546 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5550 * "If the state is SYN-SENT then
5551 * first check the ACK bit
5552 * If the ACK bit is set
5553 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5554 * a reset (unless the RST bit is set, if so drop
5555 * the segment and return)"
5557 * We do not send data with SYN, so that RFC-correct
5560 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5561 goto reset_and_undo;
5563 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5564 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5566 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5567 goto reset_and_undo;
5570 /* Now ACK is acceptable.
5572 * "If the RST bit is set
5573 * If the ACK was acceptable then signal the user "error:
5574 * connection reset", drop the segment, enter CLOSED state,
5575 * delete TCB, and return."
5584 * "fifth, if neither of the SYN or RST bits is set then
5585 * drop the segment and return."
5591 goto discard_and_undo;
5594 * "If the SYN bit is on ...
5595 * are acceptable then ...
5596 * (our SYN has been ACKed), change the connection
5597 * state to ESTABLISHED..."
5600 TCP_ECN_rcv_synack(tp, th);
5602 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5603 tcp_ack(sk, skb, FLAG_SLOWPATH);
5605 /* Ok.. it's good. Set up sequence numbers and
5606 * move to established.
5608 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5609 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5611 /* RFC1323: The window in SYN & SYN/ACK segments is
5614 tp->snd_wnd = ntohs(th->window);
5615 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5617 if (!tp->rx_opt.wscale_ok) {
5618 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5619 tp->window_clamp = min(tp->window_clamp, 65535U);
5622 if (tp->rx_opt.saw_tstamp) {
5623 tp->rx_opt.tstamp_ok = 1;
5624 tp->tcp_header_len =
5625 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5626 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5627 tcp_store_ts_recent(tp);
5629 tp->tcp_header_len = sizeof(struct tcphdr);
5632 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5633 tcp_enable_fack(tp);
5636 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5637 tcp_initialize_rcv_mss(sk);
5639 /* Remember, tcp_poll() does not lock socket!
5640 * Change state from SYN-SENT only after copied_seq
5641 * is initialized. */
5642 tp->copied_seq = tp->rcv_nxt;
5645 cvp->cookie_pair_size > 0 &&
5646 tp->rx_opt.cookie_plus > 0) {
5647 int cookie_size = tp->rx_opt.cookie_plus
5648 - TCPOLEN_COOKIE_BASE;
5649 int cookie_pair_size = cookie_size
5650 + cvp->cookie_desired;
5652 /* A cookie extension option was sent and returned.
5653 * Note that each incoming SYNACK replaces the
5654 * Responder cookie. The initial exchange is most
5655 * fragile, as protection against spoofing relies
5656 * entirely upon the sequence and timestamp (above).
5657 * This replacement strategy allows the correct pair to
5658 * pass through, while any others will be filtered via
5659 * Responder verification later.
5661 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5662 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5663 hash_location, cookie_size);
5664 cvp->cookie_pair_size = cookie_pair_size;
5669 tcp_set_state(sk, TCP_ESTABLISHED);
5671 security_inet_conn_established(sk, skb);
5673 /* Make sure socket is routed, for correct metrics. */
5674 icsk->icsk_af_ops->rebuild_header(sk);
5676 tcp_init_metrics(sk);
5678 tcp_init_congestion_control(sk);
5680 /* Prevent spurious tcp_cwnd_restart() on first data
5683 tp->lsndtime = tcp_time_stamp;
5685 tcp_init_buffer_space(sk);
5687 if (sock_flag(sk, SOCK_KEEPOPEN))
5688 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5690 if (!tp->rx_opt.snd_wscale)
5691 __tcp_fast_path_on(tp, tp->snd_wnd);
5695 if (!sock_flag(sk, SOCK_DEAD)) {
5696 sk->sk_state_change(sk);
5697 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5700 if (sk->sk_write_pending ||
5701 icsk->icsk_accept_queue.rskq_defer_accept ||
5702 icsk->icsk_ack.pingpong) {
5703 /* Save one ACK. Data will be ready after
5704 * several ticks, if write_pending is set.
5706 * It may be deleted, but with this feature tcpdumps
5707 * look so _wonderfully_ clever, that I was not able
5708 * to stand against the temptation 8) --ANK
5710 inet_csk_schedule_ack(sk);
5711 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5712 icsk->icsk_ack.ato = TCP_ATO_MIN;
5713 tcp_incr_quickack(sk);
5714 tcp_enter_quickack_mode(sk);
5715 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5716 TCP_DELACK_MAX, TCP_RTO_MAX);
5727 /* No ACK in the segment */
5731 * "If the RST bit is set
5733 * Otherwise (no ACK) drop the segment and return."
5736 goto discard_and_undo;
5740 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5741 tcp_paws_reject(&tp->rx_opt, 0))
5742 goto discard_and_undo;
5745 /* We see SYN without ACK. It is attempt of
5746 * simultaneous connect with crossed SYNs.
5747 * Particularly, it can be connect to self.
5749 tcp_set_state(sk, TCP_SYN_RECV);
5751 if (tp->rx_opt.saw_tstamp) {
5752 tp->rx_opt.tstamp_ok = 1;
5753 tcp_store_ts_recent(tp);
5754 tp->tcp_header_len =
5755 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5757 tp->tcp_header_len = sizeof(struct tcphdr);
5760 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5761 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5763 /* RFC1323: The window in SYN & SYN/ACK segments is
5766 tp->snd_wnd = ntohs(th->window);
5767 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5768 tp->max_window = tp->snd_wnd;
5770 TCP_ECN_rcv_syn(tp, th);
5773 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5774 tcp_initialize_rcv_mss(sk);
5776 tcp_send_synack(sk);
5778 /* Note, we could accept data and URG from this segment.
5779 * There are no obstacles to make this.
5781 * However, if we ignore data in ACKless segments sometimes,
5782 * we have no reasons to accept it sometimes.
5783 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5784 * is not flawless. So, discard packet for sanity.
5785 * Uncomment this return to process the data.
5792 /* "fifth, if neither of the SYN or RST bits is set then
5793 * drop the segment and return."
5797 tcp_clear_options(&tp->rx_opt);
5798 tp->rx_opt.mss_clamp = saved_clamp;
5802 tcp_clear_options(&tp->rx_opt);
5803 tp->rx_opt.mss_clamp = saved_clamp;
5808 * This function implements the receiving procedure of RFC 793 for
5809 * all states except ESTABLISHED and TIME_WAIT.
5810 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5811 * address independent.
5814 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5815 const struct tcphdr *th, unsigned int len)
5817 struct tcp_sock *tp = tcp_sk(sk);
5818 struct inet_connection_sock *icsk = inet_csk(sk);
5822 tp->rx_opt.saw_tstamp = 0;
5824 switch (sk->sk_state) {
5836 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5839 /* Now we have several options: In theory there is
5840 * nothing else in the frame. KA9Q has an option to
5841 * send data with the syn, BSD accepts data with the
5842 * syn up to the [to be] advertised window and
5843 * Solaris 2.1 gives you a protocol error. For now
5844 * we just ignore it, that fits the spec precisely
5845 * and avoids incompatibilities. It would be nice in
5846 * future to drop through and process the data.
5848 * Now that TTCP is starting to be used we ought to
5850 * But, this leaves one open to an easy denial of
5851 * service attack, and SYN cookies can't defend
5852 * against this problem. So, we drop the data
5853 * in the interest of security over speed unless
5854 * it's still in use.
5862 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5866 /* Do step6 onward by hand. */
5867 tcp_urg(sk, skb, th);
5869 tcp_data_snd_check(sk);
5873 res = tcp_validate_incoming(sk, skb, th, 0);
5877 /* step 5: check the ACK field */
5879 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5881 switch (sk->sk_state) {
5884 tp->copied_seq = tp->rcv_nxt;
5886 tcp_set_state(sk, TCP_ESTABLISHED);
5887 sk->sk_state_change(sk);
5889 /* Note, that this wakeup is only for marginal
5890 * crossed SYN case. Passively open sockets
5891 * are not waked up, because sk->sk_sleep ==
5892 * NULL and sk->sk_socket == NULL.
5896 SOCK_WAKE_IO, POLL_OUT);
5898 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5899 tp->snd_wnd = ntohs(th->window) <<
5900 tp->rx_opt.snd_wscale;
5901 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5903 if (tp->rx_opt.tstamp_ok)
5904 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5906 /* Make sure socket is routed, for
5909 icsk->icsk_af_ops->rebuild_header(sk);
5911 tcp_init_metrics(sk);
5913 tcp_init_congestion_control(sk);
5915 /* Prevent spurious tcp_cwnd_restart() on
5916 * first data packet.
5918 tp->lsndtime = tcp_time_stamp;
5921 tcp_initialize_rcv_mss(sk);
5922 tcp_init_buffer_space(sk);
5923 tcp_fast_path_on(tp);
5930 if (tp->snd_una == tp->write_seq) {
5931 tcp_set_state(sk, TCP_FIN_WAIT2);
5932 sk->sk_shutdown |= SEND_SHUTDOWN;
5933 dst_confirm(__sk_dst_get(sk));
5935 if (!sock_flag(sk, SOCK_DEAD))
5936 /* Wake up lingering close() */
5937 sk->sk_state_change(sk);
5941 if (tp->linger2 < 0 ||
5942 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5943 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5945 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5949 tmo = tcp_fin_time(sk);
5950 if (tmo > TCP_TIMEWAIT_LEN) {
5951 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5952 } else if (th->fin || sock_owned_by_user(sk)) {
5953 /* Bad case. We could lose such FIN otherwise.
5954 * It is not a big problem, but it looks confusing
5955 * and not so rare event. We still can lose it now,
5956 * if it spins in bh_lock_sock(), but it is really
5959 inet_csk_reset_keepalive_timer(sk, tmo);
5961 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5969 if (tp->snd_una == tp->write_seq) {
5970 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5976 if (tp->snd_una == tp->write_seq) {
5977 tcp_update_metrics(sk);
5986 /* step 6: check the URG bit */
5987 tcp_urg(sk, skb, th);
5989 /* step 7: process the segment text */
5990 switch (sk->sk_state) {
5991 case TCP_CLOSE_WAIT:
5994 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5998 /* RFC 793 says to queue data in these states,
5999 * RFC 1122 says we MUST send a reset.
6000 * BSD 4.4 also does reset.
6002 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6003 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6004 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6005 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6011 case TCP_ESTABLISHED:
6012 tcp_data_queue(sk, skb);
6017 /* tcp_data could move socket to TIME-WAIT */
6018 if (sk->sk_state != TCP_CLOSE) {
6019 tcp_data_snd_check(sk);
6020 tcp_ack_snd_check(sk);
6029 EXPORT_SYMBOL(tcp_rcv_state_process);