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
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly = 1;
79 int sysctl_tcp_window_scaling __read_mostly = 1;
80 int sysctl_tcp_sack __read_mostly = 1;
81 int sysctl_tcp_fack __read_mostly = 1;
82 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83 EXPORT_SYMBOL(sysctl_tcp_reordering);
84 int sysctl_tcp_ecn __read_mostly = 2;
85 EXPORT_SYMBOL(sysctl_tcp_ecn);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 2;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
91 int sysctl_tcp_stdurg __read_mostly;
92 int sysctl_tcp_rfc1337 __read_mostly;
93 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
94 int sysctl_tcp_frto __read_mostly = 2;
95 int sysctl_tcp_frto_response __read_mostly;
96 int sysctl_tcp_nometrics_save __read_mostly;
98 int sysctl_tcp_thin_dupack __read_mostly;
100 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101 int sysctl_tcp_abc __read_mostly;
103 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
104 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
105 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
106 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
107 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
108 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
109 #define FLAG_ECE 0x40 /* ECE in this ACK */
110 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
111 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
112 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
113 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
114 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
115 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
123 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
124 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
126 /* Adapt the MSS value used to make delayed ack decision to the
129 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
131 struct inet_connection_sock *icsk = inet_csk(sk);
132 const unsigned int lss = icsk->icsk_ack.last_seg_size;
135 icsk->icsk_ack.last_seg_size = 0;
137 /* skb->len may jitter because of SACKs, even if peer
138 * sends good full-sized frames.
140 len = skb_shinfo(skb)->gso_size ? : skb->len;
141 if (len >= icsk->icsk_ack.rcv_mss) {
142 icsk->icsk_ack.rcv_mss = len;
144 /* Otherwise, we make more careful check taking into account,
145 * that SACKs block is variable.
147 * "len" is invariant segment length, including TCP header.
149 len += skb->data - skb_transport_header(skb);
150 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
151 /* If PSH is not set, packet should be
152 * full sized, provided peer TCP is not badly broken.
153 * This observation (if it is correct 8)) allows
154 * to handle super-low mtu links fairly.
156 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
157 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
158 /* Subtract also invariant (if peer is RFC compliant),
159 * tcp header plus fixed timestamp option length.
160 * Resulting "len" is MSS free of SACK jitter.
162 len -= tcp_sk(sk)->tcp_header_len;
163 icsk->icsk_ack.last_seg_size = len;
165 icsk->icsk_ack.rcv_mss = len;
169 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
171 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
175 static void tcp_incr_quickack(struct sock *sk)
177 struct inet_connection_sock *icsk = inet_csk(sk);
178 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
182 if (quickacks > icsk->icsk_ack.quick)
183 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
186 static void tcp_enter_quickack_mode(struct sock *sk)
188 struct inet_connection_sock *icsk = inet_csk(sk);
189 tcp_incr_quickack(sk);
190 icsk->icsk_ack.pingpong = 0;
191 icsk->icsk_ack.ato = TCP_ATO_MIN;
194 /* Send ACKs quickly, if "quick" count is not exhausted
195 * and the session is not interactive.
198 static inline int tcp_in_quickack_mode(const struct sock *sk)
200 const struct inet_connection_sock *icsk = inet_csk(sk);
201 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
204 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
206 if (tp->ecn_flags & TCP_ECN_OK)
207 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
210 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
212 if (tcp_hdr(skb)->cwr)
213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
216 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
218 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
221 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
223 if (!(tp->ecn_flags & TCP_ECN_OK))
226 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
227 case INET_ECN_NOT_ECT:
228 /* Funny extension: if ECT is not set on a segment,
229 * and we already seen ECT on a previous segment,
230 * it is probably a retransmit.
232 if (tp->ecn_flags & TCP_ECN_SEEN)
233 tcp_enter_quickack_mode((struct sock *)tp);
236 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
239 tp->ecn_flags |= TCP_ECN_SEEN;
243 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
245 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
246 tp->ecn_flags &= ~TCP_ECN_OK;
249 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
251 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
252 tp->ecn_flags &= ~TCP_ECN_OK;
255 static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
257 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
262 /* Buffer size and advertised window tuning.
264 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
267 static void tcp_fixup_sndbuf(struct sock *sk)
269 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
271 sndmem *= TCP_INIT_CWND;
272 if (sk->sk_sndbuf < sndmem)
273 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
276 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
278 * All tcp_full_space() is split to two parts: "network" buffer, allocated
279 * forward and advertised in receiver window (tp->rcv_wnd) and
280 * "application buffer", required to isolate scheduling/application
281 * latencies from network.
282 * window_clamp is maximal advertised window. It can be less than
283 * tcp_full_space(), in this case tcp_full_space() - window_clamp
284 * is reserved for "application" buffer. The less window_clamp is
285 * the smoother our behaviour from viewpoint of network, but the lower
286 * throughput and the higher sensitivity of the connection to losses. 8)
288 * rcv_ssthresh is more strict window_clamp used at "slow start"
289 * phase to predict further behaviour of this connection.
290 * It is used for two goals:
291 * - to enforce header prediction at sender, even when application
292 * requires some significant "application buffer". It is check #1.
293 * - to prevent pruning of receive queue because of misprediction
294 * of receiver window. Check #2.
296 * The scheme does not work when sender sends good segments opening
297 * window and then starts to feed us spaghetti. But it should work
298 * in common situations. Otherwise, we have to rely on queue collapsing.
301 /* Slow part of check#2. */
302 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
304 struct tcp_sock *tp = tcp_sk(sk);
306 int truesize = tcp_win_from_space(skb->truesize) >> 1;
307 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
309 while (tp->rcv_ssthresh <= window) {
310 if (truesize <= skb->len)
311 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
319 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
321 struct tcp_sock *tp = tcp_sk(sk);
324 if (tp->rcv_ssthresh < tp->window_clamp &&
325 (int)tp->rcv_ssthresh < tcp_space(sk) &&
326 !sk_under_memory_pressure(sk)) {
329 /* Check #2. Increase window, if skb with such overhead
330 * will fit to rcvbuf in future.
332 if (tcp_win_from_space(skb->truesize) <= skb->len)
333 incr = 2 * tp->advmss;
335 incr = __tcp_grow_window(sk, skb);
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 !sk_under_memory_pressure(sk) &&
416 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 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);
477 } else if (m < new_sample)
480 /* No previous measure. */
484 if (tp->rcv_rtt_est.rtt != new_sample)
485 tp->rcv_rtt_est.rtt = new_sample;
488 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
490 if (tp->rcv_rtt_est.time == 0)
492 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
494 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
497 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
498 tp->rcv_rtt_est.time = tcp_time_stamp;
501 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
502 const struct sk_buff *skb)
504 struct tcp_sock *tp = tcp_sk(sk);
505 if (tp->rx_opt.rcv_tsecr &&
506 (TCP_SKB_CB(skb)->end_seq -
507 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
508 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
512 * This function should be called every time data is copied to user space.
513 * It calculates the appropriate TCP receive buffer space.
515 void tcp_rcv_space_adjust(struct sock *sk)
517 struct tcp_sock *tp = tcp_sk(sk);
521 if (tp->rcvq_space.time == 0)
524 time = tcp_time_stamp - tp->rcvq_space.time;
525 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
528 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
530 space = max(tp->rcvq_space.space, space);
532 if (tp->rcvq_space.space != space) {
535 tp->rcvq_space.space = space;
537 if (sysctl_tcp_moderate_rcvbuf &&
538 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
539 int new_clamp = space;
541 /* Receive space grows, normalize in order to
542 * take into account packet headers and sk_buff
543 * structure overhead.
548 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
549 while (tcp_win_from_space(rcvmem) < tp->advmss)
552 space = min(space, sysctl_tcp_rmem[2]);
553 if (space > sk->sk_rcvbuf) {
554 sk->sk_rcvbuf = space;
556 /* Make the window clamp follow along. */
557 tp->window_clamp = new_clamp;
563 tp->rcvq_space.seq = tp->copied_seq;
564 tp->rcvq_space.time = tcp_time_stamp;
567 /* There is something which you must keep in mind when you analyze the
568 * behavior of the tp->ato delayed ack timeout interval. When a
569 * connection starts up, we want to ack as quickly as possible. The
570 * problem is that "good" TCP's do slow start at the beginning of data
571 * transmission. The means that until we send the first few ACK's the
572 * sender will sit on his end and only queue most of his data, because
573 * he can only send snd_cwnd unacked packets at any given time. For
574 * each ACK we send, he increments snd_cwnd and transmits more of his
577 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
579 struct tcp_sock *tp = tcp_sk(sk);
580 struct inet_connection_sock *icsk = inet_csk(sk);
583 inet_csk_schedule_ack(sk);
585 tcp_measure_rcv_mss(sk, skb);
587 tcp_rcv_rtt_measure(tp);
589 now = tcp_time_stamp;
591 if (!icsk->icsk_ack.ato) {
592 /* The _first_ data packet received, initialize
593 * delayed ACK engine.
595 tcp_incr_quickack(sk);
596 icsk->icsk_ack.ato = TCP_ATO_MIN;
598 int m = now - icsk->icsk_ack.lrcvtime;
600 if (m <= TCP_ATO_MIN / 2) {
601 /* The fastest case is the first. */
602 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
603 } else if (m < icsk->icsk_ack.ato) {
604 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
605 if (icsk->icsk_ack.ato > icsk->icsk_rto)
606 icsk->icsk_ack.ato = icsk->icsk_rto;
607 } else if (m > icsk->icsk_rto) {
608 /* Too long gap. Apparently sender failed to
609 * restart window, so that we send ACKs quickly.
611 tcp_incr_quickack(sk);
615 icsk->icsk_ack.lrcvtime = now;
617 TCP_ECN_check_ce(tp, skb);
620 tcp_grow_window(sk, skb);
623 /* Called to compute a smoothed rtt estimate. The data fed to this
624 * routine either comes from timestamps, or from segments that were
625 * known _not_ to have been retransmitted [see Karn/Partridge
626 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
627 * piece by Van Jacobson.
628 * NOTE: the next three routines used to be one big routine.
629 * To save cycles in the RFC 1323 implementation it was better to break
630 * it up into three procedures. -- erics
632 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
634 struct tcp_sock *tp = tcp_sk(sk);
635 long m = mrtt; /* RTT */
637 /* The following amusing code comes from Jacobson's
638 * article in SIGCOMM '88. Note that rtt and mdev
639 * are scaled versions of rtt and mean deviation.
640 * This is designed to be as fast as possible
641 * m stands for "measurement".
643 * On a 1990 paper the rto value is changed to:
644 * RTO = rtt + 4 * mdev
646 * Funny. This algorithm seems to be very broken.
647 * These formulae increase RTO, when it should be decreased, increase
648 * too slowly, when it should be increased quickly, decrease too quickly
649 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
650 * does not matter how to _calculate_ it. Seems, it was trap
651 * that VJ failed to avoid. 8)
656 m -= (tp->srtt >> 3); /* m is now error in rtt est */
657 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
659 m = -m; /* m is now abs(error) */
660 m -= (tp->mdev >> 2); /* similar update on mdev */
661 /* This is similar to one of Eifel findings.
662 * Eifel blocks mdev updates when rtt decreases.
663 * This solution is a bit different: we use finer gain
664 * for mdev in this case (alpha*beta).
665 * Like Eifel it also prevents growth of rto,
666 * but also it limits too fast rto decreases,
667 * happening in pure Eifel.
672 m -= (tp->mdev >> 2); /* similar update on mdev */
674 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
675 if (tp->mdev > tp->mdev_max) {
676 tp->mdev_max = tp->mdev;
677 if (tp->mdev_max > tp->rttvar)
678 tp->rttvar = tp->mdev_max;
680 if (after(tp->snd_una, tp->rtt_seq)) {
681 if (tp->mdev_max < tp->rttvar)
682 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
683 tp->rtt_seq = tp->snd_nxt;
684 tp->mdev_max = tcp_rto_min(sk);
687 /* no previous measure. */
688 tp->srtt = m << 3; /* take the measured time to be rtt */
689 tp->mdev = m << 1; /* make sure rto = 3*rtt */
690 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
691 tp->rtt_seq = tp->snd_nxt;
695 /* Calculate rto without backoff. This is the second half of Van Jacobson's
696 * routine referred to above.
698 static inline void tcp_set_rto(struct sock *sk)
700 const struct tcp_sock *tp = tcp_sk(sk);
701 /* Old crap is replaced with new one. 8)
704 * 1. If rtt variance happened to be less 50msec, it is hallucination.
705 * It cannot be less due to utterly erratic ACK generation made
706 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
707 * to do with delayed acks, because at cwnd>2 true delack timeout
708 * is invisible. Actually, Linux-2.4 also generates erratic
709 * ACKs in some circumstances.
711 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
713 /* 2. Fixups made earlier cannot be right.
714 * If we do not estimate RTO correctly without them,
715 * all the algo is pure shit and should be replaced
716 * with correct one. It is exactly, which we pretend to do.
719 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
720 * guarantees that rto is higher.
725 /* Save metrics learned by this TCP session.
726 This function is called only, when TCP finishes successfully
727 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
729 void tcp_update_metrics(struct sock *sk)
731 struct tcp_sock *tp = tcp_sk(sk);
732 struct dst_entry *dst = __sk_dst_get(sk);
734 if (sysctl_tcp_nometrics_save)
739 if (dst && (dst->flags & DST_HOST)) {
740 const struct inet_connection_sock *icsk = inet_csk(sk);
744 if (icsk->icsk_backoff || !tp->srtt) {
745 /* This session failed to estimate rtt. Why?
746 * Probably, no packets returned in time.
749 if (!(dst_metric_locked(dst, RTAX_RTT)))
750 dst_metric_set(dst, RTAX_RTT, 0);
754 rtt = dst_metric_rtt(dst, RTAX_RTT);
757 /* If newly calculated rtt larger than stored one,
758 * store new one. Otherwise, use EWMA. Remember,
759 * rtt overestimation is always better than underestimation.
761 if (!(dst_metric_locked(dst, RTAX_RTT))) {
763 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
765 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
768 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
773 /* Scale deviation to rttvar fixed point */
778 var = dst_metric_rtt(dst, RTAX_RTTVAR);
782 var -= (var - m) >> 2;
784 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
787 if (tcp_in_initial_slowstart(tp)) {
788 /* Slow start still did not finish. */
789 if (dst_metric(dst, RTAX_SSTHRESH) &&
790 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
791 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
792 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
793 if (!dst_metric_locked(dst, RTAX_CWND) &&
794 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
795 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
796 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
797 icsk->icsk_ca_state == TCP_CA_Open) {
798 /* Cong. avoidance phase, cwnd is reliable. */
799 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
800 dst_metric_set(dst, RTAX_SSTHRESH,
801 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
802 if (!dst_metric_locked(dst, RTAX_CWND))
803 dst_metric_set(dst, RTAX_CWND,
804 (dst_metric(dst, RTAX_CWND) +
807 /* Else slow start did not finish, cwnd is non-sense,
808 ssthresh may be also invalid.
810 if (!dst_metric_locked(dst, RTAX_CWND))
811 dst_metric_set(dst, RTAX_CWND,
812 (dst_metric(dst, RTAX_CWND) +
813 tp->snd_ssthresh) >> 1);
814 if (dst_metric(dst, RTAX_SSTHRESH) &&
815 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
816 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
817 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
820 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
821 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
822 tp->reordering != sysctl_tcp_reordering)
823 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
828 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
830 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
833 cwnd = TCP_INIT_CWND;
834 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
837 /* Set slow start threshold and cwnd not falling to slow start */
838 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
840 struct tcp_sock *tp = tcp_sk(sk);
841 const struct inet_connection_sock *icsk = inet_csk(sk);
843 tp->prior_ssthresh = 0;
845 if (icsk->icsk_ca_state < TCP_CA_CWR) {
848 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
849 tp->snd_cwnd = min(tp->snd_cwnd,
850 tcp_packets_in_flight(tp) + 1U);
851 tp->snd_cwnd_cnt = 0;
852 tp->high_seq = tp->snd_nxt;
853 tp->snd_cwnd_stamp = tcp_time_stamp;
854 TCP_ECN_queue_cwr(tp);
856 tcp_set_ca_state(sk, TCP_CA_CWR);
861 * Packet counting of FACK is based on in-order assumptions, therefore TCP
862 * disables it when reordering is detected
864 static void tcp_disable_fack(struct tcp_sock *tp)
866 /* RFC3517 uses different metric in lost marker => reset on change */
868 tp->lost_skb_hint = NULL;
869 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
872 /* Take a notice that peer is sending D-SACKs */
873 static void tcp_dsack_seen(struct tcp_sock *tp)
875 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
878 /* Initialize metrics on socket. */
880 static void tcp_init_metrics(struct sock *sk)
882 struct tcp_sock *tp = tcp_sk(sk);
883 struct dst_entry *dst = __sk_dst_get(sk);
890 if (dst_metric_locked(dst, RTAX_CWND))
891 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
892 if (dst_metric(dst, RTAX_SSTHRESH)) {
893 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
894 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
895 tp->snd_ssthresh = tp->snd_cwnd_clamp;
897 /* ssthresh may have been reduced unnecessarily during.
898 * 3WHS. Restore it back to its initial default.
900 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
902 if (dst_metric(dst, RTAX_REORDERING) &&
903 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
904 tcp_disable_fack(tp);
905 tp->reordering = dst_metric(dst, RTAX_REORDERING);
908 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
911 /* Initial rtt is determined from SYN,SYN-ACK.
912 * The segment is small and rtt may appear much
913 * less than real one. Use per-dst memory
914 * to make it more realistic.
916 * A bit of theory. RTT is time passed after "normal" sized packet
917 * is sent until it is ACKed. In normal circumstances sending small
918 * packets force peer to delay ACKs and calculation is correct too.
919 * The algorithm is adaptive and, provided we follow specs, it
920 * NEVER underestimate RTT. BUT! If peer tries to make some clever
921 * tricks sort of "quick acks" for time long enough to decrease RTT
922 * to low value, and then abruptly stops to do it and starts to delay
923 * ACKs, wait for troubles.
925 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
926 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
927 tp->rtt_seq = tp->snd_nxt;
929 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
930 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
931 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
936 /* RFC2988bis: We've failed to get a valid RTT sample from
937 * 3WHS. This is most likely due to retransmission,
938 * including spurious one. Reset the RTO back to 3secs
939 * from the more aggressive 1sec to avoid more spurious
942 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
943 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
945 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
946 * retransmitted. In light of RFC2988bis' more aggressive 1sec
947 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
948 * retransmission has occurred.
950 if (tp->total_retrans > 1)
953 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
954 tp->snd_cwnd_stamp = tcp_time_stamp;
957 static void tcp_update_reordering(struct sock *sk, const int metric,
960 struct tcp_sock *tp = tcp_sk(sk);
961 if (metric > tp->reordering) {
964 tp->reordering = min(TCP_MAX_REORDERING, metric);
966 /* This exciting event is worth to be remembered. 8) */
968 mib_idx = LINUX_MIB_TCPTSREORDER;
969 else if (tcp_is_reno(tp))
970 mib_idx = LINUX_MIB_TCPRENOREORDER;
971 else if (tcp_is_fack(tp))
972 mib_idx = LINUX_MIB_TCPFACKREORDER;
974 mib_idx = LINUX_MIB_TCPSACKREORDER;
976 NET_INC_STATS_BH(sock_net(sk), mib_idx);
977 #if FASTRETRANS_DEBUG > 1
978 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
979 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
983 tp->undo_marker ? tp->undo_retrans : 0);
985 tcp_disable_fack(tp);
989 /* This must be called before lost_out is incremented */
990 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
992 if ((tp->retransmit_skb_hint == NULL) ||
993 before(TCP_SKB_CB(skb)->seq,
994 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
995 tp->retransmit_skb_hint = skb;
998 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
999 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1002 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1004 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1005 tcp_verify_retransmit_hint(tp, skb);
1007 tp->lost_out += tcp_skb_pcount(skb);
1008 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1012 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1013 struct sk_buff *skb)
1015 tcp_verify_retransmit_hint(tp, skb);
1017 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1018 tp->lost_out += tcp_skb_pcount(skb);
1019 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1023 /* This procedure tags the retransmission queue when SACKs arrive.
1025 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1026 * Packets in queue with these bits set are counted in variables
1027 * sacked_out, retrans_out and lost_out, correspondingly.
1029 * Valid combinations are:
1030 * Tag InFlight Description
1031 * 0 1 - orig segment is in flight.
1032 * S 0 - nothing flies, orig reached receiver.
1033 * L 0 - nothing flies, orig lost by net.
1034 * R 2 - both orig and retransmit are in flight.
1035 * L|R 1 - orig is lost, retransmit is in flight.
1036 * S|R 1 - orig reached receiver, retrans is still in flight.
1037 * (L|S|R is logically valid, it could occur when L|R is sacked,
1038 * but it is equivalent to plain S and code short-curcuits it to S.
1039 * L|S is logically invalid, it would mean -1 packet in flight 8))
1041 * These 6 states form finite state machine, controlled by the following events:
1042 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1043 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1044 * 3. Loss detection event of two flavors:
1045 * A. Scoreboard estimator decided the packet is lost.
1046 * A'. Reno "three dupacks" marks head of queue lost.
1047 * A''. Its FACK modification, head until snd.fack is lost.
1048 * B. SACK arrives sacking SND.NXT at the moment, when the
1049 * segment was retransmitted.
1050 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1052 * It is pleasant to note, that state diagram turns out to be commutative,
1053 * so that we are allowed not to be bothered by order of our actions,
1054 * when multiple events arrive simultaneously. (see the function below).
1056 * Reordering detection.
1057 * --------------------
1058 * Reordering metric is maximal distance, which a packet can be displaced
1059 * in packet stream. With SACKs we can estimate it:
1061 * 1. SACK fills old hole and the corresponding segment was not
1062 * ever retransmitted -> reordering. Alas, we cannot use it
1063 * when segment was retransmitted.
1064 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1065 * for retransmitted and already SACKed segment -> reordering..
1066 * Both of these heuristics are not used in Loss state, when we cannot
1067 * account for retransmits accurately.
1069 * SACK block validation.
1070 * ----------------------
1072 * SACK block range validation checks that the received SACK block fits to
1073 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1074 * Note that SND.UNA is not included to the range though being valid because
1075 * it means that the receiver is rather inconsistent with itself reporting
1076 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1077 * perfectly valid, however, in light of RFC2018 which explicitly states
1078 * that "SACK block MUST reflect the newest segment. Even if the newest
1079 * segment is going to be discarded ...", not that it looks very clever
1080 * in case of head skb. Due to potentional receiver driven attacks, we
1081 * choose to avoid immediate execution of a walk in write queue due to
1082 * reneging and defer head skb's loss recovery to standard loss recovery
1083 * procedure that will eventually trigger (nothing forbids us doing this).
1085 * Implements also blockage to start_seq wrap-around. Problem lies in the
1086 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1087 * there's no guarantee that it will be before snd_nxt (n). The problem
1088 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1091 * <- outs wnd -> <- wrapzone ->
1092 * u e n u_w e_w s n_w
1094 * |<------------+------+----- TCP seqno space --------------+---------->|
1095 * ...-- <2^31 ->| |<--------...
1096 * ...---- >2^31 ------>| |<--------...
1098 * Current code wouldn't be vulnerable but it's better still to discard such
1099 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1100 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1101 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1102 * equal to the ideal case (infinite seqno space without wrap caused issues).
1104 * With D-SACK the lower bound is extended to cover sequence space below
1105 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1106 * again, D-SACK block must not to go across snd_una (for the same reason as
1107 * for the normal SACK blocks, explained above). But there all simplicity
1108 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1109 * fully below undo_marker they do not affect behavior in anyway and can
1110 * therefore be safely ignored. In rare cases (which are more or less
1111 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1112 * fragmentation and packet reordering past skb's retransmission. To consider
1113 * them correctly, the acceptable range must be extended even more though
1114 * the exact amount is rather hard to quantify. However, tp->max_window can
1115 * be used as an exaggerated estimate.
1117 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1118 u32 start_seq, u32 end_seq)
1120 /* Too far in future, or reversed (interpretation is ambiguous) */
1121 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1124 /* Nasty start_seq wrap-around check (see comments above) */
1125 if (!before(start_seq, tp->snd_nxt))
1128 /* In outstanding window? ...This is valid exit for D-SACKs too.
1129 * start_seq == snd_una is non-sensical (see comments above)
1131 if (after(start_seq, tp->snd_una))
1134 if (!is_dsack || !tp->undo_marker)
1137 /* ...Then it's D-SACK, and must reside below snd_una completely */
1138 if (after(end_seq, tp->snd_una))
1141 if (!before(start_seq, tp->undo_marker))
1145 if (!after(end_seq, tp->undo_marker))
1148 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1149 * start_seq < undo_marker and end_seq >= undo_marker.
1151 return !before(start_seq, end_seq - tp->max_window);
1154 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1155 * Event "B". Later note: FACK people cheated me again 8), we have to account
1156 * for reordering! Ugly, but should help.
1158 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1159 * less than what is now known to be received by the other end (derived from
1160 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1161 * retransmitted skbs to avoid some costly processing per ACKs.
1163 static void tcp_mark_lost_retrans(struct sock *sk)
1165 const struct inet_connection_sock *icsk = inet_csk(sk);
1166 struct tcp_sock *tp = tcp_sk(sk);
1167 struct sk_buff *skb;
1169 u32 new_low_seq = tp->snd_nxt;
1170 u32 received_upto = tcp_highest_sack_seq(tp);
1172 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1173 !after(received_upto, tp->lost_retrans_low) ||
1174 icsk->icsk_ca_state != TCP_CA_Recovery)
1177 tcp_for_write_queue(skb, sk) {
1178 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1180 if (skb == tcp_send_head(sk))
1182 if (cnt == tp->retrans_out)
1184 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1187 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1190 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1191 * constraint here (see above) but figuring out that at
1192 * least tp->reordering SACK blocks reside between ack_seq
1193 * and received_upto is not easy task to do cheaply with
1194 * the available datastructures.
1196 * Whether FACK should check here for tp->reordering segs
1197 * in-between one could argue for either way (it would be
1198 * rather simple to implement as we could count fack_count
1199 * during the walk and do tp->fackets_out - fack_count).
1201 if (after(received_upto, ack_seq)) {
1202 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1203 tp->retrans_out -= tcp_skb_pcount(skb);
1205 tcp_skb_mark_lost_uncond_verify(tp, skb);
1206 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1208 if (before(ack_seq, new_low_seq))
1209 new_low_seq = ack_seq;
1210 cnt += tcp_skb_pcount(skb);
1214 if (tp->retrans_out)
1215 tp->lost_retrans_low = new_low_seq;
1218 static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1219 struct tcp_sack_block_wire *sp, int num_sacks,
1222 struct tcp_sock *tp = tcp_sk(sk);
1223 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1224 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1227 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1230 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1231 } else if (num_sacks > 1) {
1232 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1233 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1235 if (!after(end_seq_0, end_seq_1) &&
1236 !before(start_seq_0, start_seq_1)) {
1239 NET_INC_STATS_BH(sock_net(sk),
1240 LINUX_MIB_TCPDSACKOFORECV);
1244 /* D-SACK for already forgotten data... Do dumb counting. */
1245 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1246 !after(end_seq_0, prior_snd_una) &&
1247 after(end_seq_0, tp->undo_marker))
1253 struct tcp_sacktag_state {
1259 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1260 * the incoming SACK may not exactly match but we can find smaller MSS
1261 * aligned portion of it that matches. Therefore we might need to fragment
1262 * which may fail and creates some hassle (caller must handle error case
1265 * FIXME: this could be merged to shift decision code
1267 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1268 u32 start_seq, u32 end_seq)
1271 unsigned int pkt_len;
1274 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1275 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1277 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1278 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1279 mss = tcp_skb_mss(skb);
1280 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1283 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1287 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1292 /* Round if necessary so that SACKs cover only full MSSes
1293 * and/or the remaining small portion (if present)
1295 if (pkt_len > mss) {
1296 unsigned int new_len = (pkt_len / mss) * mss;
1297 if (!in_sack && new_len < pkt_len) {
1299 if (new_len > skb->len)
1304 err = tcp_fragment(sk, skb, pkt_len, mss);
1312 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1313 static u8 tcp_sacktag_one(struct sock *sk,
1314 struct tcp_sacktag_state *state, u8 sacked,
1315 u32 start_seq, u32 end_seq,
1316 int dup_sack, int pcount)
1318 struct tcp_sock *tp = tcp_sk(sk);
1319 int fack_count = state->fack_count;
1321 /* Account D-SACK for retransmitted packet. */
1322 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1323 if (tp->undo_marker && tp->undo_retrans &&
1324 after(end_seq, tp->undo_marker))
1326 if (sacked & TCPCB_SACKED_ACKED)
1327 state->reord = min(fack_count, state->reord);
1330 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1331 if (!after(end_seq, tp->snd_una))
1334 if (!(sacked & TCPCB_SACKED_ACKED)) {
1335 if (sacked & TCPCB_SACKED_RETRANS) {
1336 /* If the segment is not tagged as lost,
1337 * we do not clear RETRANS, believing
1338 * that retransmission is still in flight.
1340 if (sacked & TCPCB_LOST) {
1341 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1342 tp->lost_out -= pcount;
1343 tp->retrans_out -= pcount;
1346 if (!(sacked & TCPCB_RETRANS)) {
1347 /* New sack for not retransmitted frame,
1348 * which was in hole. It is reordering.
1350 if (before(start_seq,
1351 tcp_highest_sack_seq(tp)))
1352 state->reord = min(fack_count,
1355 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1356 if (!after(end_seq, tp->frto_highmark))
1357 state->flag |= FLAG_ONLY_ORIG_SACKED;
1360 if (sacked & TCPCB_LOST) {
1361 sacked &= ~TCPCB_LOST;
1362 tp->lost_out -= pcount;
1366 sacked |= TCPCB_SACKED_ACKED;
1367 state->flag |= FLAG_DATA_SACKED;
1368 tp->sacked_out += pcount;
1370 fack_count += pcount;
1372 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1373 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1374 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1375 tp->lost_cnt_hint += pcount;
1377 if (fack_count > tp->fackets_out)
1378 tp->fackets_out = fack_count;
1381 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1382 * frames and clear it. undo_retrans is decreased above, L|R frames
1383 * are accounted above as well.
1385 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1386 sacked &= ~TCPCB_SACKED_RETRANS;
1387 tp->retrans_out -= pcount;
1393 /* Shift newly-SACKed bytes from this skb to the immediately previous
1394 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1396 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1397 struct tcp_sacktag_state *state,
1398 unsigned int pcount, int shifted, int mss,
1401 struct tcp_sock *tp = tcp_sk(sk);
1402 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1403 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1404 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1408 /* Adjust counters and hints for the newly sacked sequence
1409 * range but discard the return value since prev is already
1410 * marked. We must tag the range first because the seq
1411 * advancement below implicitly advances
1412 * tcp_highest_sack_seq() when skb is highest_sack.
1414 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1415 start_seq, end_seq, dup_sack, pcount);
1417 if (skb == tp->lost_skb_hint)
1418 tp->lost_cnt_hint += pcount;
1420 TCP_SKB_CB(prev)->end_seq += shifted;
1421 TCP_SKB_CB(skb)->seq += shifted;
1423 skb_shinfo(prev)->gso_segs += pcount;
1424 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1425 skb_shinfo(skb)->gso_segs -= pcount;
1427 /* When we're adding to gso_segs == 1, gso_size will be zero,
1428 * in theory this shouldn't be necessary but as long as DSACK
1429 * code can come after this skb later on it's better to keep
1430 * setting gso_size to something.
1432 if (!skb_shinfo(prev)->gso_size) {
1433 skb_shinfo(prev)->gso_size = mss;
1434 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1437 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1438 if (skb_shinfo(skb)->gso_segs <= 1) {
1439 skb_shinfo(skb)->gso_size = 0;
1440 skb_shinfo(skb)->gso_type = 0;
1443 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1444 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1447 BUG_ON(!tcp_skb_pcount(skb));
1448 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1452 /* Whole SKB was eaten :-) */
1454 if (skb == tp->retransmit_skb_hint)
1455 tp->retransmit_skb_hint = prev;
1456 if (skb == tp->scoreboard_skb_hint)
1457 tp->scoreboard_skb_hint = prev;
1458 if (skb == tp->lost_skb_hint) {
1459 tp->lost_skb_hint = prev;
1460 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1463 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1464 if (skb == tcp_highest_sack(sk))
1465 tcp_advance_highest_sack(sk, skb);
1467 tcp_unlink_write_queue(skb, sk);
1468 sk_wmem_free_skb(sk, skb);
1470 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1475 /* I wish gso_size would have a bit more sane initialization than
1476 * something-or-zero which complicates things
1478 static int tcp_skb_seglen(const struct sk_buff *skb)
1480 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1483 /* Shifting pages past head area doesn't work */
1484 static int skb_can_shift(const struct sk_buff *skb)
1486 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1489 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1492 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1493 struct tcp_sacktag_state *state,
1494 u32 start_seq, u32 end_seq,
1497 struct tcp_sock *tp = tcp_sk(sk);
1498 struct sk_buff *prev;
1504 if (!sk_can_gso(sk))
1507 /* Normally R but no L won't result in plain S */
1509 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1511 if (!skb_can_shift(skb))
1513 /* This frame is about to be dropped (was ACKed). */
1514 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1517 /* Can only happen with delayed DSACK + discard craziness */
1518 if (unlikely(skb == tcp_write_queue_head(sk)))
1520 prev = tcp_write_queue_prev(sk, skb);
1522 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1525 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1526 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1530 pcount = tcp_skb_pcount(skb);
1531 mss = tcp_skb_seglen(skb);
1533 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1534 * drop this restriction as unnecessary
1536 if (mss != tcp_skb_seglen(prev))
1539 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1541 /* CHECKME: This is non-MSS split case only?, this will
1542 * cause skipped skbs due to advancing loop btw, original
1543 * has that feature too
1545 if (tcp_skb_pcount(skb) <= 1)
1548 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1550 /* TODO: head merge to next could be attempted here
1551 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1552 * though it might not be worth of the additional hassle
1554 * ...we can probably just fallback to what was done
1555 * previously. We could try merging non-SACKed ones
1556 * as well but it probably isn't going to buy off
1557 * because later SACKs might again split them, and
1558 * it would make skb timestamp tracking considerably
1564 len = end_seq - TCP_SKB_CB(skb)->seq;
1566 BUG_ON(len > skb->len);
1568 /* MSS boundaries should be honoured or else pcount will
1569 * severely break even though it makes things bit trickier.
1570 * Optimize common case to avoid most of the divides
1572 mss = tcp_skb_mss(skb);
1574 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1575 * drop this restriction as unnecessary
1577 if (mss != tcp_skb_seglen(prev))
1582 } else if (len < mss) {
1590 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1591 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1594 if (!skb_shift(prev, skb, len))
1596 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1599 /* Hole filled allows collapsing with the next as well, this is very
1600 * useful when hole on every nth skb pattern happens
1602 if (prev == tcp_write_queue_tail(sk))
1604 skb = tcp_write_queue_next(sk, prev);
1606 if (!skb_can_shift(skb) ||
1607 (skb == tcp_send_head(sk)) ||
1608 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1609 (mss != tcp_skb_seglen(skb)))
1613 if (skb_shift(prev, skb, len)) {
1614 pcount += tcp_skb_pcount(skb);
1615 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1619 state->fack_count += pcount;
1626 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1630 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1631 struct tcp_sack_block *next_dup,
1632 struct tcp_sacktag_state *state,
1633 u32 start_seq, u32 end_seq,
1636 struct tcp_sock *tp = tcp_sk(sk);
1637 struct sk_buff *tmp;
1639 tcp_for_write_queue_from(skb, sk) {
1641 int dup_sack = dup_sack_in;
1643 if (skb == tcp_send_head(sk))
1646 /* queue is in-order => we can short-circuit the walk early */
1647 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1650 if ((next_dup != NULL) &&
1651 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1652 in_sack = tcp_match_skb_to_sack(sk, skb,
1653 next_dup->start_seq,
1659 /* skb reference here is a bit tricky to get right, since
1660 * shifting can eat and free both this skb and the next,
1661 * so not even _safe variant of the loop is enough.
1664 tmp = tcp_shift_skb_data(sk, skb, state,
1665 start_seq, end_seq, dup_sack);
1674 in_sack = tcp_match_skb_to_sack(sk, skb,
1680 if (unlikely(in_sack < 0))
1684 TCP_SKB_CB(skb)->sacked =
1687 TCP_SKB_CB(skb)->sacked,
1688 TCP_SKB_CB(skb)->seq,
1689 TCP_SKB_CB(skb)->end_seq,
1691 tcp_skb_pcount(skb));
1693 if (!before(TCP_SKB_CB(skb)->seq,
1694 tcp_highest_sack_seq(tp)))
1695 tcp_advance_highest_sack(sk, skb);
1698 state->fack_count += tcp_skb_pcount(skb);
1703 /* Avoid all extra work that is being done by sacktag while walking in
1706 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1707 struct tcp_sacktag_state *state,
1710 tcp_for_write_queue_from(skb, sk) {
1711 if (skb == tcp_send_head(sk))
1714 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1717 state->fack_count += tcp_skb_pcount(skb);
1722 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1724 struct tcp_sack_block *next_dup,
1725 struct tcp_sacktag_state *state,
1728 if (next_dup == NULL)
1731 if (before(next_dup->start_seq, skip_to_seq)) {
1732 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1733 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1734 next_dup->start_seq, next_dup->end_seq,
1741 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1743 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1747 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1750 const struct inet_connection_sock *icsk = inet_csk(sk);
1751 struct tcp_sock *tp = tcp_sk(sk);
1752 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1753 TCP_SKB_CB(ack_skb)->sacked);
1754 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1755 struct tcp_sack_block sp[TCP_NUM_SACKS];
1756 struct tcp_sack_block *cache;
1757 struct tcp_sacktag_state state;
1758 struct sk_buff *skb;
1759 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1761 int found_dup_sack = 0;
1763 int first_sack_index;
1766 state.reord = tp->packets_out;
1768 if (!tp->sacked_out) {
1769 if (WARN_ON(tp->fackets_out))
1770 tp->fackets_out = 0;
1771 tcp_highest_sack_reset(sk);
1774 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1775 num_sacks, prior_snd_una);
1777 state.flag |= FLAG_DSACKING_ACK;
1779 /* Eliminate too old ACKs, but take into
1780 * account more or less fresh ones, they can
1781 * contain valid SACK info.
1783 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1786 if (!tp->packets_out)
1790 first_sack_index = 0;
1791 for (i = 0; i < num_sacks; i++) {
1792 int dup_sack = !i && found_dup_sack;
1794 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1795 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1797 if (!tcp_is_sackblock_valid(tp, dup_sack,
1798 sp[used_sacks].start_seq,
1799 sp[used_sacks].end_seq)) {
1803 if (!tp->undo_marker)
1804 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1806 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1808 /* Don't count olds caused by ACK reordering */
1809 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1810 !after(sp[used_sacks].end_seq, tp->snd_una))
1812 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1815 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1817 first_sack_index = -1;
1821 /* Ignore very old stuff early */
1822 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1828 /* order SACK blocks to allow in order walk of the retrans queue */
1829 for (i = used_sacks - 1; i > 0; i--) {
1830 for (j = 0; j < i; j++) {
1831 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1832 swap(sp[j], sp[j + 1]);
1834 /* Track where the first SACK block goes to */
1835 if (j == first_sack_index)
1836 first_sack_index = j + 1;
1841 skb = tcp_write_queue_head(sk);
1842 state.fack_count = 0;
1845 if (!tp->sacked_out) {
1846 /* It's already past, so skip checking against it */
1847 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1849 cache = tp->recv_sack_cache;
1850 /* Skip empty blocks in at head of the cache */
1851 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1856 while (i < used_sacks) {
1857 u32 start_seq = sp[i].start_seq;
1858 u32 end_seq = sp[i].end_seq;
1859 int dup_sack = (found_dup_sack && (i == first_sack_index));
1860 struct tcp_sack_block *next_dup = NULL;
1862 if (found_dup_sack && ((i + 1) == first_sack_index))
1863 next_dup = &sp[i + 1];
1865 /* Skip too early cached blocks */
1866 while (tcp_sack_cache_ok(tp, cache) &&
1867 !before(start_seq, cache->end_seq))
1870 /* Can skip some work by looking recv_sack_cache? */
1871 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1872 after(end_seq, cache->start_seq)) {
1875 if (before(start_seq, cache->start_seq)) {
1876 skb = tcp_sacktag_skip(skb, sk, &state,
1878 skb = tcp_sacktag_walk(skb, sk, next_dup,
1885 /* Rest of the block already fully processed? */
1886 if (!after(end_seq, cache->end_seq))
1889 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1893 /* ...tail remains todo... */
1894 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1895 /* ...but better entrypoint exists! */
1896 skb = tcp_highest_sack(sk);
1899 state.fack_count = tp->fackets_out;
1904 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1905 /* Check overlap against next cached too (past this one already) */
1910 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1911 skb = tcp_highest_sack(sk);
1914 state.fack_count = tp->fackets_out;
1916 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1919 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1920 start_seq, end_seq, dup_sack);
1923 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1924 * due to in-order walk
1926 if (after(end_seq, tp->frto_highmark))
1927 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1932 /* Clear the head of the cache sack blocks so we can skip it next time */
1933 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1934 tp->recv_sack_cache[i].start_seq = 0;
1935 tp->recv_sack_cache[i].end_seq = 0;
1937 for (j = 0; j < used_sacks; j++)
1938 tp->recv_sack_cache[i++] = sp[j];
1940 tcp_mark_lost_retrans(sk);
1942 tcp_verify_left_out(tp);
1944 if ((state.reord < tp->fackets_out) &&
1945 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1946 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1947 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1951 #if FASTRETRANS_DEBUG > 0
1952 WARN_ON((int)tp->sacked_out < 0);
1953 WARN_ON((int)tp->lost_out < 0);
1954 WARN_ON((int)tp->retrans_out < 0);
1955 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1960 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1961 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1963 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1967 holes = max(tp->lost_out, 1U);
1968 holes = min(holes, tp->packets_out);
1970 if ((tp->sacked_out + holes) > tp->packets_out) {
1971 tp->sacked_out = tp->packets_out - holes;
1977 /* If we receive more dupacks than we expected counting segments
1978 * in assumption of absent reordering, interpret this as reordering.
1979 * The only another reason could be bug in receiver TCP.
1981 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1983 struct tcp_sock *tp = tcp_sk(sk);
1984 if (tcp_limit_reno_sacked(tp))
1985 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1988 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1990 static void tcp_add_reno_sack(struct sock *sk)
1992 struct tcp_sock *tp = tcp_sk(sk);
1994 tcp_check_reno_reordering(sk, 0);
1995 tcp_verify_left_out(tp);
1998 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2000 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
2002 struct tcp_sock *tp = tcp_sk(sk);
2005 /* One ACK acked hole. The rest eat duplicate ACKs. */
2006 if (acked - 1 >= tp->sacked_out)
2009 tp->sacked_out -= acked - 1;
2011 tcp_check_reno_reordering(sk, acked);
2012 tcp_verify_left_out(tp);
2015 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2020 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2022 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2025 /* F-RTO can only be used if TCP has never retransmitted anything other than
2026 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2028 int tcp_use_frto(struct sock *sk)
2030 const struct tcp_sock *tp = tcp_sk(sk);
2031 const struct inet_connection_sock *icsk = inet_csk(sk);
2032 struct sk_buff *skb;
2034 if (!sysctl_tcp_frto)
2037 /* MTU probe and F-RTO won't really play nicely along currently */
2038 if (icsk->icsk_mtup.probe_size)
2041 if (tcp_is_sackfrto(tp))
2044 /* Avoid expensive walking of rexmit queue if possible */
2045 if (tp->retrans_out > 1)
2048 skb = tcp_write_queue_head(sk);
2049 if (tcp_skb_is_last(sk, skb))
2051 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2052 tcp_for_write_queue_from(skb, sk) {
2053 if (skb == tcp_send_head(sk))
2055 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2057 /* Short-circuit when first non-SACKed skb has been checked */
2058 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2064 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2065 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2066 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2067 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2068 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2069 * bits are handled if the Loss state is really to be entered (in
2070 * tcp_enter_frto_loss).
2072 * Do like tcp_enter_loss() would; when RTO expires the second time it
2074 * "Reduce ssthresh if it has not yet been made inside this window."
2076 void tcp_enter_frto(struct sock *sk)
2078 const struct inet_connection_sock *icsk = inet_csk(sk);
2079 struct tcp_sock *tp = tcp_sk(sk);
2080 struct sk_buff *skb;
2082 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2083 tp->snd_una == tp->high_seq ||
2084 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2085 !icsk->icsk_retransmits)) {
2086 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2087 /* Our state is too optimistic in ssthresh() call because cwnd
2088 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2089 * recovery has not yet completed. Pattern would be this: RTO,
2090 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2092 * RFC4138 should be more specific on what to do, even though
2093 * RTO is quite unlikely to occur after the first Cumulative ACK
2094 * due to back-off and complexity of triggering events ...
2096 if (tp->frto_counter) {
2098 stored_cwnd = tp->snd_cwnd;
2100 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2101 tp->snd_cwnd = stored_cwnd;
2103 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2105 /* ... in theory, cong.control module could do "any tricks" in
2106 * ssthresh(), which means that ca_state, lost bits and lost_out
2107 * counter would have to be faked before the call occurs. We
2108 * consider that too expensive, unlikely and hacky, so modules
2109 * using these in ssthresh() must deal these incompatibility
2110 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2112 tcp_ca_event(sk, CA_EVENT_FRTO);
2115 tp->undo_marker = tp->snd_una;
2116 tp->undo_retrans = 0;
2118 skb = tcp_write_queue_head(sk);
2119 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2120 tp->undo_marker = 0;
2121 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2122 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2123 tp->retrans_out -= tcp_skb_pcount(skb);
2125 tcp_verify_left_out(tp);
2127 /* Too bad if TCP was application limited */
2128 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2130 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2131 * The last condition is necessary at least in tp->frto_counter case.
2133 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2134 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2135 after(tp->high_seq, tp->snd_una)) {
2136 tp->frto_highmark = tp->high_seq;
2138 tp->frto_highmark = tp->snd_nxt;
2140 tcp_set_ca_state(sk, TCP_CA_Disorder);
2141 tp->high_seq = tp->snd_nxt;
2142 tp->frto_counter = 1;
2145 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2146 * which indicates that we should follow the traditional RTO recovery,
2147 * i.e. mark everything lost and do go-back-N retransmission.
2149 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2151 struct tcp_sock *tp = tcp_sk(sk);
2152 struct sk_buff *skb;
2155 tp->retrans_out = 0;
2156 if (tcp_is_reno(tp))
2157 tcp_reset_reno_sack(tp);
2159 tcp_for_write_queue(skb, sk) {
2160 if (skb == tcp_send_head(sk))
2163 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2165 * Count the retransmission made on RTO correctly (only when
2166 * waiting for the first ACK and did not get it)...
2168 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2169 /* For some reason this R-bit might get cleared? */
2170 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2171 tp->retrans_out += tcp_skb_pcount(skb);
2172 /* ...enter this if branch just for the first segment */
2173 flag |= FLAG_DATA_ACKED;
2175 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2176 tp->undo_marker = 0;
2177 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2180 /* Marking forward transmissions that were made after RTO lost
2181 * can cause unnecessary retransmissions in some scenarios,
2182 * SACK blocks will mitigate that in some but not in all cases.
2183 * We used to not mark them but it was causing break-ups with
2184 * receivers that do only in-order receival.
2186 * TODO: we could detect presence of such receiver and select
2187 * different behavior per flow.
2189 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2190 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2191 tp->lost_out += tcp_skb_pcount(skb);
2192 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2195 tcp_verify_left_out(tp);
2197 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2198 tp->snd_cwnd_cnt = 0;
2199 tp->snd_cwnd_stamp = tcp_time_stamp;
2200 tp->frto_counter = 0;
2201 tp->bytes_acked = 0;
2203 tp->reordering = min_t(unsigned int, tp->reordering,
2204 sysctl_tcp_reordering);
2205 tcp_set_ca_state(sk, TCP_CA_Loss);
2206 tp->high_seq = tp->snd_nxt;
2207 TCP_ECN_queue_cwr(tp);
2209 tcp_clear_all_retrans_hints(tp);
2212 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2214 tp->retrans_out = 0;
2217 tp->undo_marker = 0;
2218 tp->undo_retrans = 0;
2221 void tcp_clear_retrans(struct tcp_sock *tp)
2223 tcp_clear_retrans_partial(tp);
2225 tp->fackets_out = 0;
2229 /* Enter Loss state. If "how" is not zero, forget all SACK information
2230 * and reset tags completely, otherwise preserve SACKs. If receiver
2231 * dropped its ofo queue, we will know this due to reneging detection.
2233 void tcp_enter_loss(struct sock *sk, int how)
2235 const struct inet_connection_sock *icsk = inet_csk(sk);
2236 struct tcp_sock *tp = tcp_sk(sk);
2237 struct sk_buff *skb;
2239 /* Reduce ssthresh if it has not yet been made inside this window. */
2240 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2241 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2242 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2243 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2244 tcp_ca_event(sk, CA_EVENT_LOSS);
2247 tp->snd_cwnd_cnt = 0;
2248 tp->snd_cwnd_stamp = tcp_time_stamp;
2250 tp->bytes_acked = 0;
2251 tcp_clear_retrans_partial(tp);
2253 if (tcp_is_reno(tp))
2254 tcp_reset_reno_sack(tp);
2257 /* Push undo marker, if it was plain RTO and nothing
2258 * was retransmitted. */
2259 tp->undo_marker = tp->snd_una;
2262 tp->fackets_out = 0;
2264 tcp_clear_all_retrans_hints(tp);
2266 tcp_for_write_queue(skb, sk) {
2267 if (skb == tcp_send_head(sk))
2270 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2271 tp->undo_marker = 0;
2272 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2273 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2274 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2275 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2276 tp->lost_out += tcp_skb_pcount(skb);
2277 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2280 tcp_verify_left_out(tp);
2282 tp->reordering = min_t(unsigned int, tp->reordering,
2283 sysctl_tcp_reordering);
2284 tcp_set_ca_state(sk, TCP_CA_Loss);
2285 tp->high_seq = tp->snd_nxt;
2286 TCP_ECN_queue_cwr(tp);
2287 /* Abort F-RTO algorithm if one is in progress */
2288 tp->frto_counter = 0;
2291 /* If ACK arrived pointing to a remembered SACK, it means that our
2292 * remembered SACKs do not reflect real state of receiver i.e.
2293 * receiver _host_ is heavily congested (or buggy).
2295 * Do processing similar to RTO timeout.
2297 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2299 if (flag & FLAG_SACK_RENEGING) {
2300 struct inet_connection_sock *icsk = inet_csk(sk);
2301 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2303 tcp_enter_loss(sk, 1);
2304 icsk->icsk_retransmits++;
2305 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2306 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2307 icsk->icsk_rto, TCP_RTO_MAX);
2313 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2315 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2318 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2319 * counter when SACK is enabled (without SACK, sacked_out is used for
2322 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2323 * segments up to the highest received SACK block so far and holes in
2326 * With reordering, holes may still be in flight, so RFC3517 recovery
2327 * uses pure sacked_out (total number of SACKed segments) even though
2328 * it violates the RFC that uses duplicate ACKs, often these are equal
2329 * but when e.g. out-of-window ACKs or packet duplication occurs,
2330 * they differ. Since neither occurs due to loss, TCP should really
2333 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2335 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2338 static inline int tcp_skb_timedout(const struct sock *sk,
2339 const struct sk_buff *skb)
2341 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2344 static inline int tcp_head_timedout(const struct sock *sk)
2346 const struct tcp_sock *tp = tcp_sk(sk);
2348 return tp->packets_out &&
2349 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2352 /* Linux NewReno/SACK/FACK/ECN state machine.
2353 * --------------------------------------
2355 * "Open" Normal state, no dubious events, fast path.
2356 * "Disorder" In all the respects it is "Open",
2357 * but requires a bit more attention. It is entered when
2358 * we see some SACKs or dupacks. It is split of "Open"
2359 * mainly to move some processing from fast path to slow one.
2360 * "CWR" CWND was reduced due to some Congestion Notification event.
2361 * It can be ECN, ICMP source quench, local device congestion.
2362 * "Recovery" CWND was reduced, we are fast-retransmitting.
2363 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2365 * tcp_fastretrans_alert() is entered:
2366 * - each incoming ACK, if state is not "Open"
2367 * - when arrived ACK is unusual, namely:
2372 * Counting packets in flight is pretty simple.
2374 * in_flight = packets_out - left_out + retrans_out
2376 * packets_out is SND.NXT-SND.UNA counted in packets.
2378 * retrans_out is number of retransmitted segments.
2380 * left_out is number of segments left network, but not ACKed yet.
2382 * left_out = sacked_out + lost_out
2384 * sacked_out: Packets, which arrived to receiver out of order
2385 * and hence not ACKed. With SACKs this number is simply
2386 * amount of SACKed data. Even without SACKs
2387 * it is easy to give pretty reliable estimate of this number,
2388 * counting duplicate ACKs.
2390 * lost_out: Packets lost by network. TCP has no explicit
2391 * "loss notification" feedback from network (for now).
2392 * It means that this number can be only _guessed_.
2393 * Actually, it is the heuristics to predict lossage that
2394 * distinguishes different algorithms.
2396 * F.e. after RTO, when all the queue is considered as lost,
2397 * lost_out = packets_out and in_flight = retrans_out.
2399 * Essentially, we have now two algorithms counting
2402 * FACK: It is the simplest heuristics. As soon as we decided
2403 * that something is lost, we decide that _all_ not SACKed
2404 * packets until the most forward SACK are lost. I.e.
2405 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2406 * It is absolutely correct estimate, if network does not reorder
2407 * packets. And it loses any connection to reality when reordering
2408 * takes place. We use FACK by default until reordering
2409 * is suspected on the path to this destination.
2411 * NewReno: when Recovery is entered, we assume that one segment
2412 * is lost (classic Reno). While we are in Recovery and
2413 * a partial ACK arrives, we assume that one more packet
2414 * is lost (NewReno). This heuristics are the same in NewReno
2417 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2418 * deflation etc. CWND is real congestion window, never inflated, changes
2419 * only according to classic VJ rules.
2421 * Really tricky (and requiring careful tuning) part of algorithm
2422 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2423 * The first determines the moment _when_ we should reduce CWND and,
2424 * hence, slow down forward transmission. In fact, it determines the moment
2425 * when we decide that hole is caused by loss, rather than by a reorder.
2427 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2428 * holes, caused by lost packets.
2430 * And the most logically complicated part of algorithm is undo
2431 * heuristics. We detect false retransmits due to both too early
2432 * fast retransmit (reordering) and underestimated RTO, analyzing
2433 * timestamps and D-SACKs. When we detect that some segments were
2434 * retransmitted by mistake and CWND reduction was wrong, we undo
2435 * window reduction and abort recovery phase. This logic is hidden
2436 * inside several functions named tcp_try_undo_<something>.
2439 /* This function decides, when we should leave Disordered state
2440 * and enter Recovery phase, reducing congestion window.
2442 * Main question: may we further continue forward transmission
2443 * with the same cwnd?
2445 static int tcp_time_to_recover(struct sock *sk)
2447 struct tcp_sock *tp = tcp_sk(sk);
2450 /* Do not perform any recovery during F-RTO algorithm */
2451 if (tp->frto_counter)
2454 /* Trick#1: The loss is proven. */
2458 /* Not-A-Trick#2 : Classic rule... */
2459 if (tcp_dupack_heuristics(tp) > tp->reordering)
2462 /* Trick#3 : when we use RFC2988 timer restart, fast
2463 * retransmit can be triggered by timeout of queue head.
2465 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2468 /* Trick#4: It is still not OK... But will it be useful to delay
2471 packets_out = tp->packets_out;
2472 if (packets_out <= tp->reordering &&
2473 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2474 !tcp_may_send_now(sk)) {
2475 /* We have nothing to send. This connection is limited
2476 * either by receiver window or by application.
2481 /* If a thin stream is detected, retransmit after first
2482 * received dupack. Employ only if SACK is supported in order
2483 * to avoid possible corner-case series of spurious retransmissions
2484 * Use only if there are no unsent data.
2486 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2487 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2488 tcp_is_sack(tp) && !tcp_send_head(sk))
2494 /* New heuristics: it is possible only after we switched to restart timer
2495 * each time when something is ACKed. Hence, we can detect timed out packets
2496 * during fast retransmit without falling to slow start.
2498 * Usefulness of this as is very questionable, since we should know which of
2499 * the segments is the next to timeout which is relatively expensive to find
2500 * in general case unless we add some data structure just for that. The
2501 * current approach certainly won't find the right one too often and when it
2502 * finally does find _something_ it usually marks large part of the window
2503 * right away (because a retransmission with a larger timestamp blocks the
2504 * loop from advancing). -ij
2506 static void tcp_timeout_skbs(struct sock *sk)
2508 struct tcp_sock *tp = tcp_sk(sk);
2509 struct sk_buff *skb;
2511 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2514 skb = tp->scoreboard_skb_hint;
2515 if (tp->scoreboard_skb_hint == NULL)
2516 skb = tcp_write_queue_head(sk);
2518 tcp_for_write_queue_from(skb, sk) {
2519 if (skb == tcp_send_head(sk))
2521 if (!tcp_skb_timedout(sk, skb))
2524 tcp_skb_mark_lost(tp, skb);
2527 tp->scoreboard_skb_hint = skb;
2529 tcp_verify_left_out(tp);
2532 /* Detect loss in event "A" above by marking head of queue up as lost.
2533 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2534 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2535 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2536 * the maximum SACKed segments to pass before reaching this limit.
2538 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2540 struct tcp_sock *tp = tcp_sk(sk);
2541 struct sk_buff *skb;
2545 /* Use SACK to deduce losses of new sequences sent during recovery */
2546 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2548 WARN_ON(packets > tp->packets_out);
2549 if (tp->lost_skb_hint) {
2550 skb = tp->lost_skb_hint;
2551 cnt = tp->lost_cnt_hint;
2552 /* Head already handled? */
2553 if (mark_head && skb != tcp_write_queue_head(sk))
2556 skb = tcp_write_queue_head(sk);
2560 tcp_for_write_queue_from(skb, sk) {
2561 if (skb == tcp_send_head(sk))
2563 /* TODO: do this better */
2564 /* this is not the most efficient way to do this... */
2565 tp->lost_skb_hint = skb;
2566 tp->lost_cnt_hint = cnt;
2568 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2572 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2573 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2574 cnt += tcp_skb_pcount(skb);
2576 if (cnt > packets) {
2577 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2578 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2579 (oldcnt >= packets))
2582 mss = skb_shinfo(skb)->gso_size;
2583 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2589 tcp_skb_mark_lost(tp, skb);
2594 tcp_verify_left_out(tp);
2597 /* Account newly detected lost packet(s) */
2599 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2601 struct tcp_sock *tp = tcp_sk(sk);
2603 if (tcp_is_reno(tp)) {
2604 tcp_mark_head_lost(sk, 1, 1);
2605 } else if (tcp_is_fack(tp)) {
2606 int lost = tp->fackets_out - tp->reordering;
2609 tcp_mark_head_lost(sk, lost, 0);
2611 int sacked_upto = tp->sacked_out - tp->reordering;
2612 if (sacked_upto >= 0)
2613 tcp_mark_head_lost(sk, sacked_upto, 0);
2614 else if (fast_rexmit)
2615 tcp_mark_head_lost(sk, 1, 1);
2618 tcp_timeout_skbs(sk);
2621 /* CWND moderation, preventing bursts due to too big ACKs
2622 * in dubious situations.
2624 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2626 tp->snd_cwnd = min(tp->snd_cwnd,
2627 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2628 tp->snd_cwnd_stamp = tcp_time_stamp;
2631 /* Lower bound on congestion window is slow start threshold
2632 * unless congestion avoidance choice decides to overide it.
2634 static inline u32 tcp_cwnd_min(const struct sock *sk)
2636 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2638 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2641 /* Decrease cwnd each second ack. */
2642 static void tcp_cwnd_down(struct sock *sk, int flag)
2644 struct tcp_sock *tp = tcp_sk(sk);
2645 int decr = tp->snd_cwnd_cnt + 1;
2647 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2648 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2649 tp->snd_cwnd_cnt = decr & 1;
2652 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2653 tp->snd_cwnd -= decr;
2655 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2656 tp->snd_cwnd_stamp = tcp_time_stamp;
2660 /* Nothing was retransmitted or returned timestamp is less
2661 * than timestamp of the first retransmission.
2663 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2665 return !tp->retrans_stamp ||
2666 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2667 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2670 /* Undo procedures. */
2672 #if FASTRETRANS_DEBUG > 1
2673 static void DBGUNDO(struct sock *sk, const char *msg)
2675 struct tcp_sock *tp = tcp_sk(sk);
2676 struct inet_sock *inet = inet_sk(sk);
2678 if (sk->sk_family == AF_INET) {
2679 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2681 &inet->inet_daddr, ntohs(inet->inet_dport),
2682 tp->snd_cwnd, tcp_left_out(tp),
2683 tp->snd_ssthresh, tp->prior_ssthresh,
2686 #if IS_ENABLED(CONFIG_IPV6)
2687 else if (sk->sk_family == AF_INET6) {
2688 struct ipv6_pinfo *np = inet6_sk(sk);
2689 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2691 &np->daddr, ntohs(inet->inet_dport),
2692 tp->snd_cwnd, tcp_left_out(tp),
2693 tp->snd_ssthresh, tp->prior_ssthresh,
2699 #define DBGUNDO(x...) do { } while (0)
2702 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2704 struct tcp_sock *tp = tcp_sk(sk);
2706 if (tp->prior_ssthresh) {
2707 const struct inet_connection_sock *icsk = inet_csk(sk);
2709 if (icsk->icsk_ca_ops->undo_cwnd)
2710 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2712 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2714 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2715 tp->snd_ssthresh = tp->prior_ssthresh;
2716 TCP_ECN_withdraw_cwr(tp);
2719 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2721 tp->snd_cwnd_stamp = tcp_time_stamp;
2724 static inline int tcp_may_undo(const struct tcp_sock *tp)
2726 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2729 /* People celebrate: "We love our President!" */
2730 static int tcp_try_undo_recovery(struct sock *sk)
2732 struct tcp_sock *tp = tcp_sk(sk);
2734 if (tcp_may_undo(tp)) {
2737 /* Happy end! We did not retransmit anything
2738 * or our original transmission succeeded.
2740 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2741 tcp_undo_cwr(sk, true);
2742 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2743 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2745 mib_idx = LINUX_MIB_TCPFULLUNDO;
2747 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2748 tp->undo_marker = 0;
2750 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2751 /* Hold old state until something *above* high_seq
2752 * is ACKed. For Reno it is MUST to prevent false
2753 * fast retransmits (RFC2582). SACK TCP is safe. */
2754 tcp_moderate_cwnd(tp);
2757 tcp_set_ca_state(sk, TCP_CA_Open);
2761 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2762 static void tcp_try_undo_dsack(struct sock *sk)
2764 struct tcp_sock *tp = tcp_sk(sk);
2766 if (tp->undo_marker && !tp->undo_retrans) {
2767 DBGUNDO(sk, "D-SACK");
2768 tcp_undo_cwr(sk, true);
2769 tp->undo_marker = 0;
2770 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2774 /* We can clear retrans_stamp when there are no retransmissions in the
2775 * window. It would seem that it is trivially available for us in
2776 * tp->retrans_out, however, that kind of assumptions doesn't consider
2777 * what will happen if errors occur when sending retransmission for the
2778 * second time. ...It could the that such segment has only
2779 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2780 * the head skb is enough except for some reneging corner cases that
2781 * are not worth the effort.
2783 * Main reason for all this complexity is the fact that connection dying
2784 * time now depends on the validity of the retrans_stamp, in particular,
2785 * that successive retransmissions of a segment must not advance
2786 * retrans_stamp under any conditions.
2788 static int tcp_any_retrans_done(const struct sock *sk)
2790 const struct tcp_sock *tp = tcp_sk(sk);
2791 struct sk_buff *skb;
2793 if (tp->retrans_out)
2796 skb = tcp_write_queue_head(sk);
2797 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2803 /* Undo during fast recovery after partial ACK. */
2805 static int tcp_try_undo_partial(struct sock *sk, int acked)
2807 struct tcp_sock *tp = tcp_sk(sk);
2808 /* Partial ACK arrived. Force Hoe's retransmit. */
2809 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2811 if (tcp_may_undo(tp)) {
2812 /* Plain luck! Hole if filled with delayed
2813 * packet, rather than with a retransmit.
2815 if (!tcp_any_retrans_done(sk))
2816 tp->retrans_stamp = 0;
2818 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2821 tcp_undo_cwr(sk, false);
2822 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2824 /* So... Do not make Hoe's retransmit yet.
2825 * If the first packet was delayed, the rest
2826 * ones are most probably delayed as well.
2833 /* Undo during loss recovery after partial ACK. */
2834 static int tcp_try_undo_loss(struct sock *sk)
2836 struct tcp_sock *tp = tcp_sk(sk);
2838 if (tcp_may_undo(tp)) {
2839 struct sk_buff *skb;
2840 tcp_for_write_queue(skb, sk) {
2841 if (skb == tcp_send_head(sk))
2843 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2846 tcp_clear_all_retrans_hints(tp);
2848 DBGUNDO(sk, "partial loss");
2850 tcp_undo_cwr(sk, true);
2851 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2852 inet_csk(sk)->icsk_retransmits = 0;
2853 tp->undo_marker = 0;
2854 if (tcp_is_sack(tp))
2855 tcp_set_ca_state(sk, TCP_CA_Open);
2861 static inline void tcp_complete_cwr(struct sock *sk)
2863 struct tcp_sock *tp = tcp_sk(sk);
2865 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2866 if (tp->undo_marker) {
2867 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR)
2868 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2870 tp->snd_cwnd = tp->snd_ssthresh;
2871 tp->snd_cwnd_stamp = tcp_time_stamp;
2873 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2876 static void tcp_try_keep_open(struct sock *sk)
2878 struct tcp_sock *tp = tcp_sk(sk);
2879 int state = TCP_CA_Open;
2881 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2882 state = TCP_CA_Disorder;
2884 if (inet_csk(sk)->icsk_ca_state != state) {
2885 tcp_set_ca_state(sk, state);
2886 tp->high_seq = tp->snd_nxt;
2890 static void tcp_try_to_open(struct sock *sk, int flag)
2892 struct tcp_sock *tp = tcp_sk(sk);
2894 tcp_verify_left_out(tp);
2896 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2897 tp->retrans_stamp = 0;
2899 if (flag & FLAG_ECE)
2900 tcp_enter_cwr(sk, 1);
2902 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2903 tcp_try_keep_open(sk);
2904 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2905 tcp_moderate_cwnd(tp);
2907 tcp_cwnd_down(sk, flag);
2911 static void tcp_mtup_probe_failed(struct sock *sk)
2913 struct inet_connection_sock *icsk = inet_csk(sk);
2915 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2916 icsk->icsk_mtup.probe_size = 0;
2919 static void tcp_mtup_probe_success(struct sock *sk)
2921 struct tcp_sock *tp = tcp_sk(sk);
2922 struct inet_connection_sock *icsk = inet_csk(sk);
2924 /* FIXME: breaks with very large cwnd */
2925 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2926 tp->snd_cwnd = tp->snd_cwnd *
2927 tcp_mss_to_mtu(sk, tp->mss_cache) /
2928 icsk->icsk_mtup.probe_size;
2929 tp->snd_cwnd_cnt = 0;
2930 tp->snd_cwnd_stamp = tcp_time_stamp;
2931 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2933 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2934 icsk->icsk_mtup.probe_size = 0;
2935 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2938 /* Do a simple retransmit without using the backoff mechanisms in
2939 * tcp_timer. This is used for path mtu discovery.
2940 * The socket is already locked here.
2942 void tcp_simple_retransmit(struct sock *sk)
2944 const struct inet_connection_sock *icsk = inet_csk(sk);
2945 struct tcp_sock *tp = tcp_sk(sk);
2946 struct sk_buff *skb;
2947 unsigned int mss = tcp_current_mss(sk);
2948 u32 prior_lost = tp->lost_out;
2950 tcp_for_write_queue(skb, sk) {
2951 if (skb == tcp_send_head(sk))
2953 if (tcp_skb_seglen(skb) > mss &&
2954 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2955 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2956 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2957 tp->retrans_out -= tcp_skb_pcount(skb);
2959 tcp_skb_mark_lost_uncond_verify(tp, skb);
2963 tcp_clear_retrans_hints_partial(tp);
2965 if (prior_lost == tp->lost_out)
2968 if (tcp_is_reno(tp))
2969 tcp_limit_reno_sacked(tp);
2971 tcp_verify_left_out(tp);
2973 /* Don't muck with the congestion window here.
2974 * Reason is that we do not increase amount of _data_
2975 * in network, but units changed and effective
2976 * cwnd/ssthresh really reduced now.
2978 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2979 tp->high_seq = tp->snd_nxt;
2980 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2981 tp->prior_ssthresh = 0;
2982 tp->undo_marker = 0;
2983 tcp_set_ca_state(sk, TCP_CA_Loss);
2985 tcp_xmit_retransmit_queue(sk);
2987 EXPORT_SYMBOL(tcp_simple_retransmit);
2989 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2990 * (proportional rate reduction with slow start reduction bound) as described in
2991 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2992 * It computes the number of packets to send (sndcnt) based on packets newly
2994 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2995 * cwnd reductions across a full RTT.
2996 * 2) If packets in flight is lower than ssthresh (such as due to excess
2997 * losses and/or application stalls), do not perform any further cwnd
2998 * reductions, but instead slow start up to ssthresh.
3000 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
3001 int fast_rexmit, int flag)
3003 struct tcp_sock *tp = tcp_sk(sk);
3005 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
3007 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
3008 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
3010 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
3012 sndcnt = min_t(int, delta,
3013 max_t(int, tp->prr_delivered - tp->prr_out,
3014 newly_acked_sacked) + 1);
3017 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
3018 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3021 /* Process an event, which can update packets-in-flight not trivially.
3022 * Main goal of this function is to calculate new estimate for left_out,
3023 * taking into account both packets sitting in receiver's buffer and
3024 * packets lost by network.
3026 * Besides that it does CWND reduction, when packet loss is detected
3027 * and changes state of machine.
3029 * It does _not_ decide what to send, it is made in function
3030 * tcp_xmit_retransmit_queue().
3032 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3033 int newly_acked_sacked, bool is_dupack,
3036 struct inet_connection_sock *icsk = inet_csk(sk);
3037 struct tcp_sock *tp = tcp_sk(sk);
3038 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3039 (tcp_fackets_out(tp) > tp->reordering));
3040 int fast_rexmit = 0, mib_idx;
3042 if (WARN_ON(!tp->packets_out && tp->sacked_out))
3044 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3045 tp->fackets_out = 0;
3047 /* Now state machine starts.
3048 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3049 if (flag & FLAG_ECE)
3050 tp->prior_ssthresh = 0;
3052 /* B. In all the states check for reneging SACKs. */
3053 if (tcp_check_sack_reneging(sk, flag))
3056 /* C. Check consistency of the current state. */
3057 tcp_verify_left_out(tp);
3059 /* D. Check state exit conditions. State can be terminated
3060 * when high_seq is ACKed. */
3061 if (icsk->icsk_ca_state == TCP_CA_Open) {
3062 WARN_ON(tp->retrans_out != 0);
3063 tp->retrans_stamp = 0;
3064 } else if (!before(tp->snd_una, tp->high_seq)) {
3065 switch (icsk->icsk_ca_state) {
3067 icsk->icsk_retransmits = 0;
3068 if (tcp_try_undo_recovery(sk))
3073 /* CWR is to be held something *above* high_seq
3074 * is ACKed for CWR bit to reach receiver. */
3075 if (tp->snd_una != tp->high_seq) {
3076 tcp_complete_cwr(sk);
3077 tcp_set_ca_state(sk, TCP_CA_Open);
3081 case TCP_CA_Recovery:
3082 if (tcp_is_reno(tp))
3083 tcp_reset_reno_sack(tp);
3084 if (tcp_try_undo_recovery(sk))
3086 tcp_complete_cwr(sk);
3091 /* E. Process state. */
3092 switch (icsk->icsk_ca_state) {
3093 case TCP_CA_Recovery:
3094 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3095 if (tcp_is_reno(tp) && is_dupack)
3096 tcp_add_reno_sack(sk);
3098 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3101 if (flag & FLAG_DATA_ACKED)
3102 icsk->icsk_retransmits = 0;
3103 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3104 tcp_reset_reno_sack(tp);
3105 if (!tcp_try_undo_loss(sk)) {
3106 tcp_moderate_cwnd(tp);
3107 tcp_xmit_retransmit_queue(sk);
3110 if (icsk->icsk_ca_state != TCP_CA_Open)
3112 /* Loss is undone; fall through to processing in Open state. */
3114 if (tcp_is_reno(tp)) {
3115 if (flag & FLAG_SND_UNA_ADVANCED)
3116 tcp_reset_reno_sack(tp);
3118 tcp_add_reno_sack(sk);
3121 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3122 tcp_try_undo_dsack(sk);
3124 if (!tcp_time_to_recover(sk)) {
3125 tcp_try_to_open(sk, flag);
3129 /* MTU probe failure: don't reduce cwnd */
3130 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3131 icsk->icsk_mtup.probe_size &&
3132 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3133 tcp_mtup_probe_failed(sk);
3134 /* Restores the reduction we did in tcp_mtup_probe() */
3136 tcp_simple_retransmit(sk);
3140 /* Otherwise enter Recovery state */
3142 if (tcp_is_reno(tp))
3143 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3145 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3147 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3149 tp->high_seq = tp->snd_nxt;
3150 tp->prior_ssthresh = 0;
3151 tp->undo_marker = tp->snd_una;
3152 tp->undo_retrans = tp->retrans_out;
3154 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3155 if (!(flag & FLAG_ECE))
3156 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3157 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3158 TCP_ECN_queue_cwr(tp);
3161 tp->bytes_acked = 0;
3162 tp->snd_cwnd_cnt = 0;
3163 tp->prior_cwnd = tp->snd_cwnd;
3164 tp->prr_delivered = 0;
3166 tcp_set_ca_state(sk, TCP_CA_Recovery);
3170 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3171 tcp_update_scoreboard(sk, fast_rexmit);
3172 tp->prr_delivered += newly_acked_sacked;
3173 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3174 tcp_xmit_retransmit_queue(sk);
3177 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3179 tcp_rtt_estimator(sk, seq_rtt);
3181 inet_csk(sk)->icsk_backoff = 0;
3183 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3185 /* Read draft-ietf-tcplw-high-performance before mucking
3186 * with this code. (Supersedes RFC1323)
3188 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3190 /* RTTM Rule: A TSecr value received in a segment is used to
3191 * update the averaged RTT measurement only if the segment
3192 * acknowledges some new data, i.e., only if it advances the
3193 * left edge of the send window.
3195 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3196 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3198 * Changed: reset backoff as soon as we see the first valid sample.
3199 * If we do not, we get strongly overestimated rto. With timestamps
3200 * samples are accepted even from very old segments: f.e., when rtt=1
3201 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3202 * answer arrives rto becomes 120 seconds! If at least one of segments
3203 * in window is lost... Voila. --ANK (010210)
3205 struct tcp_sock *tp = tcp_sk(sk);
3207 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3210 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3212 /* We don't have a timestamp. Can only use
3213 * packets that are not retransmitted to determine
3214 * rtt estimates. Also, we must not reset the
3215 * backoff for rto until we get a non-retransmitted
3216 * packet. This allows us to deal with a situation
3217 * where the network delay has increased suddenly.
3218 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3221 if (flag & FLAG_RETRANS_DATA_ACKED)
3224 tcp_valid_rtt_meas(sk, seq_rtt);
3227 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3230 const struct tcp_sock *tp = tcp_sk(sk);
3231 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3232 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3233 tcp_ack_saw_tstamp(sk, flag);
3234 else if (seq_rtt >= 0)
3235 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3238 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3240 const struct inet_connection_sock *icsk = inet_csk(sk);
3241 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3242 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3245 /* Restart timer after forward progress on connection.
3246 * RFC2988 recommends to restart timer to now+rto.
3248 static void tcp_rearm_rto(struct sock *sk)
3250 const struct tcp_sock *tp = tcp_sk(sk);
3252 if (!tp->packets_out) {
3253 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3255 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3256 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3260 /* If we get here, the whole TSO packet has not been acked. */
3261 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3263 struct tcp_sock *tp = tcp_sk(sk);
3266 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3268 packets_acked = tcp_skb_pcount(skb);
3269 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3271 packets_acked -= tcp_skb_pcount(skb);
3273 if (packets_acked) {
3274 BUG_ON(tcp_skb_pcount(skb) == 0);
3275 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3278 return packets_acked;
3281 /* Remove acknowledged frames from the retransmission queue. If our packet
3282 * is before the ack sequence we can discard it as it's confirmed to have
3283 * arrived at the other end.
3285 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3288 struct tcp_sock *tp = tcp_sk(sk);
3289 const struct inet_connection_sock *icsk = inet_csk(sk);
3290 struct sk_buff *skb;
3291 u32 now = tcp_time_stamp;
3292 int fully_acked = 1;
3295 u32 reord = tp->packets_out;
3296 u32 prior_sacked = tp->sacked_out;
3298 s32 ca_seq_rtt = -1;
3299 ktime_t last_ackt = net_invalid_timestamp();
3301 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3302 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3304 u8 sacked = scb->sacked;
3306 /* Determine how many packets and what bytes were acked, tso and else */
3307 if (after(scb->end_seq, tp->snd_una)) {
3308 if (tcp_skb_pcount(skb) == 1 ||
3309 !after(tp->snd_una, scb->seq))
3312 acked_pcount = tcp_tso_acked(sk, skb);
3318 acked_pcount = tcp_skb_pcount(skb);
3321 if (sacked & TCPCB_RETRANS) {
3322 if (sacked & TCPCB_SACKED_RETRANS)
3323 tp->retrans_out -= acked_pcount;
3324 flag |= FLAG_RETRANS_DATA_ACKED;
3327 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3328 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3330 ca_seq_rtt = now - scb->when;
3331 last_ackt = skb->tstamp;
3333 seq_rtt = ca_seq_rtt;
3335 if (!(sacked & TCPCB_SACKED_ACKED))
3336 reord = min(pkts_acked, reord);
3339 if (sacked & TCPCB_SACKED_ACKED)
3340 tp->sacked_out -= acked_pcount;
3341 if (sacked & TCPCB_LOST)
3342 tp->lost_out -= acked_pcount;
3344 tp->packets_out -= acked_pcount;
3345 pkts_acked += acked_pcount;
3347 /* Initial outgoing SYN's get put onto the write_queue
3348 * just like anything else we transmit. It is not
3349 * true data, and if we misinform our callers that
3350 * this ACK acks real data, we will erroneously exit
3351 * connection startup slow start one packet too
3352 * quickly. This is severely frowned upon behavior.
3354 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3355 flag |= FLAG_DATA_ACKED;
3357 flag |= FLAG_SYN_ACKED;
3358 tp->retrans_stamp = 0;
3364 tcp_unlink_write_queue(skb, sk);
3365 sk_wmem_free_skb(sk, skb);
3366 tp->scoreboard_skb_hint = NULL;
3367 if (skb == tp->retransmit_skb_hint)
3368 tp->retransmit_skb_hint = NULL;
3369 if (skb == tp->lost_skb_hint)
3370 tp->lost_skb_hint = NULL;
3373 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3374 tp->snd_up = tp->snd_una;
3376 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3377 flag |= FLAG_SACK_RENEGING;
3379 if (flag & FLAG_ACKED) {
3380 const struct tcp_congestion_ops *ca_ops
3381 = inet_csk(sk)->icsk_ca_ops;
3383 if (unlikely(icsk->icsk_mtup.probe_size &&
3384 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3385 tcp_mtup_probe_success(sk);
3388 tcp_ack_update_rtt(sk, flag, seq_rtt);
3391 if (tcp_is_reno(tp)) {
3392 tcp_remove_reno_sacks(sk, pkts_acked);
3396 /* Non-retransmitted hole got filled? That's reordering */
3397 if (reord < prior_fackets)
3398 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3400 delta = tcp_is_fack(tp) ? pkts_acked :
3401 prior_sacked - tp->sacked_out;
3402 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3405 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3407 if (ca_ops->pkts_acked) {
3410 /* Is the ACK triggering packet unambiguous? */
3411 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3412 /* High resolution needed and available? */
3413 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3414 !ktime_equal(last_ackt,
3415 net_invalid_timestamp()))
3416 rtt_us = ktime_us_delta(ktime_get_real(),
3418 else if (ca_seq_rtt >= 0)
3419 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3422 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3426 #if FASTRETRANS_DEBUG > 0
3427 WARN_ON((int)tp->sacked_out < 0);
3428 WARN_ON((int)tp->lost_out < 0);
3429 WARN_ON((int)tp->retrans_out < 0);
3430 if (!tp->packets_out && tcp_is_sack(tp)) {
3431 icsk = inet_csk(sk);
3433 printk(KERN_DEBUG "Leak l=%u %d\n",
3434 tp->lost_out, icsk->icsk_ca_state);
3437 if (tp->sacked_out) {
3438 printk(KERN_DEBUG "Leak s=%u %d\n",
3439 tp->sacked_out, icsk->icsk_ca_state);
3442 if (tp->retrans_out) {
3443 printk(KERN_DEBUG "Leak r=%u %d\n",
3444 tp->retrans_out, icsk->icsk_ca_state);
3445 tp->retrans_out = 0;
3452 static void tcp_ack_probe(struct sock *sk)
3454 const struct tcp_sock *tp = tcp_sk(sk);
3455 struct inet_connection_sock *icsk = inet_csk(sk);
3457 /* Was it a usable window open? */
3459 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3460 icsk->icsk_backoff = 0;
3461 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3462 /* Socket must be waked up by subsequent tcp_data_snd_check().
3463 * This function is not for random using!
3466 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3467 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3472 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3474 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3475 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3478 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3480 const struct tcp_sock *tp = tcp_sk(sk);
3481 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3482 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3485 /* Check that window update is acceptable.
3486 * The function assumes that snd_una<=ack<=snd_next.
3488 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3489 const u32 ack, const u32 ack_seq,
3492 return after(ack, tp->snd_una) ||
3493 after(ack_seq, tp->snd_wl1) ||
3494 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3497 /* Update our send window.
3499 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3500 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3502 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3505 struct tcp_sock *tp = tcp_sk(sk);
3507 u32 nwin = ntohs(tcp_hdr(skb)->window);
3509 if (likely(!tcp_hdr(skb)->syn))
3510 nwin <<= tp->rx_opt.snd_wscale;
3512 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3513 flag |= FLAG_WIN_UPDATE;
3514 tcp_update_wl(tp, ack_seq);
3516 if (tp->snd_wnd != nwin) {
3519 /* Note, it is the only place, where
3520 * fast path is recovered for sending TCP.
3523 tcp_fast_path_check(sk);
3525 if (nwin > tp->max_window) {
3526 tp->max_window = nwin;
3527 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3537 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3538 * continue in congestion avoidance.
3540 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3542 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3543 tp->snd_cwnd_cnt = 0;
3544 tp->bytes_acked = 0;
3545 TCP_ECN_queue_cwr(tp);
3546 tcp_moderate_cwnd(tp);
3549 /* A conservative spurious RTO response algorithm: reduce cwnd using
3550 * rate halving and continue in congestion avoidance.
3552 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3554 tcp_enter_cwr(sk, 0);
3557 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3559 if (flag & FLAG_ECE)
3560 tcp_ratehalving_spur_to_response(sk);
3562 tcp_undo_cwr(sk, true);
3565 /* F-RTO spurious RTO detection algorithm (RFC4138)
3567 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3568 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3569 * window (but not to or beyond highest sequence sent before RTO):
3570 * On First ACK, send two new segments out.
3571 * On Second ACK, RTO was likely spurious. Do spurious response (response
3572 * algorithm is not part of the F-RTO detection algorithm
3573 * given in RFC4138 but can be selected separately).
3574 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3575 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3576 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3577 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3579 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3580 * original window even after we transmit two new data segments.
3583 * on first step, wait until first cumulative ACK arrives, then move to
3584 * the second step. In second step, the next ACK decides.
3586 * F-RTO is implemented (mainly) in four functions:
3587 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3588 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3589 * called when tcp_use_frto() showed green light
3590 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3591 * - tcp_enter_frto_loss() is called if there is not enough evidence
3592 * to prove that the RTO is indeed spurious. It transfers the control
3593 * from F-RTO to the conventional RTO recovery
3595 static int tcp_process_frto(struct sock *sk, int flag)
3597 struct tcp_sock *tp = tcp_sk(sk);
3599 tcp_verify_left_out(tp);
3601 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3602 if (flag & FLAG_DATA_ACKED)
3603 inet_csk(sk)->icsk_retransmits = 0;
3605 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3606 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3607 tp->undo_marker = 0;
3609 if (!before(tp->snd_una, tp->frto_highmark)) {
3610 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3614 if (!tcp_is_sackfrto(tp)) {
3615 /* RFC4138 shortcoming in step 2; should also have case c):
3616 * ACK isn't duplicate nor advances window, e.g., opposite dir
3619 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3622 if (!(flag & FLAG_DATA_ACKED)) {
3623 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3628 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3629 /* Prevent sending of new data. */
3630 tp->snd_cwnd = min(tp->snd_cwnd,
3631 tcp_packets_in_flight(tp));
3635 if ((tp->frto_counter >= 2) &&
3636 (!(flag & FLAG_FORWARD_PROGRESS) ||
3637 ((flag & FLAG_DATA_SACKED) &&
3638 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3639 /* RFC4138 shortcoming (see comment above) */
3640 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3641 (flag & FLAG_NOT_DUP))
3644 tcp_enter_frto_loss(sk, 3, flag);
3649 if (tp->frto_counter == 1) {
3650 /* tcp_may_send_now needs to see updated state */
3651 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3652 tp->frto_counter = 2;
3654 if (!tcp_may_send_now(sk))
3655 tcp_enter_frto_loss(sk, 2, flag);
3659 switch (sysctl_tcp_frto_response) {
3661 tcp_undo_spur_to_response(sk, flag);
3664 tcp_conservative_spur_to_response(tp);
3667 tcp_ratehalving_spur_to_response(sk);
3670 tp->frto_counter = 0;
3671 tp->undo_marker = 0;
3672 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3677 /* This routine deals with incoming acks, but not outgoing ones. */
3678 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3680 struct inet_connection_sock *icsk = inet_csk(sk);
3681 struct tcp_sock *tp = tcp_sk(sk);
3682 u32 prior_snd_una = tp->snd_una;
3683 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3684 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3685 bool is_dupack = false;
3686 u32 prior_in_flight;
3689 int prior_sacked = tp->sacked_out;
3691 int newly_acked_sacked = 0;
3694 /* If the ack is older than previous acks
3695 * then we can probably ignore it.
3697 if (before(ack, prior_snd_una))
3700 /* If the ack includes data we haven't sent yet, discard
3701 * this segment (RFC793 Section 3.9).
3703 if (after(ack, tp->snd_nxt))
3706 if (after(ack, prior_snd_una))
3707 flag |= FLAG_SND_UNA_ADVANCED;
3709 if (sysctl_tcp_abc) {
3710 if (icsk->icsk_ca_state < TCP_CA_CWR)
3711 tp->bytes_acked += ack - prior_snd_una;
3712 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3713 /* we assume just one segment left network */
3714 tp->bytes_acked += min(ack - prior_snd_una,
3718 prior_fackets = tp->fackets_out;
3719 prior_in_flight = tcp_packets_in_flight(tp);
3721 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3722 /* Window is constant, pure forward advance.
3723 * No more checks are required.
3724 * Note, we use the fact that SND.UNA>=SND.WL2.
3726 tcp_update_wl(tp, ack_seq);
3728 flag |= FLAG_WIN_UPDATE;
3730 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3732 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3734 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3737 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3739 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3741 if (TCP_SKB_CB(skb)->sacked)
3742 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3744 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3747 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3750 /* We passed data and got it acked, remove any soft error
3751 * log. Something worked...
3753 sk->sk_err_soft = 0;
3754 icsk->icsk_probes_out = 0;
3755 tp->rcv_tstamp = tcp_time_stamp;
3756 prior_packets = tp->packets_out;
3760 /* See if we can take anything off of the retransmit queue. */
3761 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3763 pkts_acked = prior_packets - tp->packets_out;
3764 newly_acked_sacked = (prior_packets - prior_sacked) -
3765 (tp->packets_out - tp->sacked_out);
3767 if (tp->frto_counter)
3768 frto_cwnd = tcp_process_frto(sk, flag);
3769 /* Guarantee sacktag reordering detection against wrap-arounds */
3770 if (before(tp->frto_highmark, tp->snd_una))
3771 tp->frto_highmark = 0;
3773 if (tcp_ack_is_dubious(sk, flag)) {
3774 /* Advance CWND, if state allows this. */
3775 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3776 tcp_may_raise_cwnd(sk, flag))
3777 tcp_cong_avoid(sk, ack, prior_in_flight);
3778 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3779 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3782 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3783 tcp_cong_avoid(sk, ack, prior_in_flight);
3786 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3787 dst_confirm(__sk_dst_get(sk));
3792 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3793 if (flag & FLAG_DSACKING_ACK)
3794 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3796 /* If this ack opens up a zero window, clear backoff. It was
3797 * being used to time the probes, and is probably far higher than
3798 * it needs to be for normal retransmission.
3800 if (tcp_send_head(sk))
3805 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3809 /* If data was SACKed, tag it and see if we should send more data.
3810 * If data was DSACKed, see if we can undo a cwnd reduction.
3812 if (TCP_SKB_CB(skb)->sacked) {
3813 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3814 newly_acked_sacked = tp->sacked_out - prior_sacked;
3815 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked,
3819 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3823 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3824 * But, this can also be called on packets in the established flow when
3825 * the fast version below fails.
3827 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3828 const u8 **hvpp, int estab)
3830 const unsigned char *ptr;
3831 const struct tcphdr *th = tcp_hdr(skb);
3832 int length = (th->doff * 4) - sizeof(struct tcphdr);
3834 ptr = (const unsigned char *)(th + 1);
3835 opt_rx->saw_tstamp = 0;
3837 while (length > 0) {
3838 int opcode = *ptr++;
3844 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3849 if (opsize < 2) /* "silly options" */
3851 if (opsize > length)
3852 return; /* don't parse partial options */
3855 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3856 u16 in_mss = get_unaligned_be16(ptr);
3858 if (opt_rx->user_mss &&
3859 opt_rx->user_mss < in_mss)
3860 in_mss = opt_rx->user_mss;
3861 opt_rx->mss_clamp = in_mss;
3866 if (opsize == TCPOLEN_WINDOW && th->syn &&
3867 !estab && sysctl_tcp_window_scaling) {
3868 __u8 snd_wscale = *(__u8 *)ptr;
3869 opt_rx->wscale_ok = 1;
3870 if (snd_wscale > 14) {
3871 if (net_ratelimit())
3872 pr_info("%s: Illegal window scaling value %d >14 received\n",
3877 opt_rx->snd_wscale = snd_wscale;
3880 case TCPOPT_TIMESTAMP:
3881 if ((opsize == TCPOLEN_TIMESTAMP) &&
3882 ((estab && opt_rx->tstamp_ok) ||
3883 (!estab && sysctl_tcp_timestamps))) {
3884 opt_rx->saw_tstamp = 1;
3885 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3886 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3889 case TCPOPT_SACK_PERM:
3890 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3891 !estab && sysctl_tcp_sack) {
3892 opt_rx->sack_ok = TCP_SACK_SEEN;
3893 tcp_sack_reset(opt_rx);
3898 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3899 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3901 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3904 #ifdef CONFIG_TCP_MD5SIG
3907 * The MD5 Hash has already been
3908 * checked (see tcp_v{4,6}_do_rcv()).
3913 /* This option is variable length.
3916 case TCPOLEN_COOKIE_BASE:
3917 /* not yet implemented */
3919 case TCPOLEN_COOKIE_PAIR:
3920 /* not yet implemented */
3922 case TCPOLEN_COOKIE_MIN+0:
3923 case TCPOLEN_COOKIE_MIN+2:
3924 case TCPOLEN_COOKIE_MIN+4:
3925 case TCPOLEN_COOKIE_MIN+6:
3926 case TCPOLEN_COOKIE_MAX:
3927 /* 16-bit multiple */
3928 opt_rx->cookie_plus = opsize;
3943 EXPORT_SYMBOL(tcp_parse_options);
3945 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3947 const __be32 *ptr = (const __be32 *)(th + 1);
3949 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3950 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3951 tp->rx_opt.saw_tstamp = 1;
3953 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3955 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3961 /* Fast parse options. This hopes to only see timestamps.
3962 * If it is wrong it falls back on tcp_parse_options().
3964 static int tcp_fast_parse_options(const struct sk_buff *skb,
3965 const struct tcphdr *th,
3966 struct tcp_sock *tp, const u8 **hvpp)
3968 /* In the spirit of fast parsing, compare doff directly to constant
3969 * values. Because equality is used, short doff can be ignored here.
3971 if (th->doff == (sizeof(*th) / 4)) {
3972 tp->rx_opt.saw_tstamp = 0;
3974 } else if (tp->rx_opt.tstamp_ok &&
3975 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3976 if (tcp_parse_aligned_timestamp(tp, th))
3979 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
3983 #ifdef CONFIG_TCP_MD5SIG
3985 * Parse MD5 Signature option
3987 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3989 int length = (th->doff << 2) - sizeof(*th);
3990 const u8 *ptr = (const u8 *)(th + 1);
3992 /* If the TCP option is too short, we can short cut */
3993 if (length < TCPOLEN_MD5SIG)
3996 while (length > 0) {
3997 int opcode = *ptr++;
4008 if (opsize < 2 || opsize > length)
4010 if (opcode == TCPOPT_MD5SIG)
4011 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4018 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4021 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4023 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
4024 tp->rx_opt.ts_recent_stamp = get_seconds();
4027 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4029 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4030 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
4031 * extra check below makes sure this can only happen
4032 * for pure ACK frames. -DaveM
4034 * Not only, also it occurs for expired timestamps.
4037 if (tcp_paws_check(&tp->rx_opt, 0))
4038 tcp_store_ts_recent(tp);
4042 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4044 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4045 * it can pass through stack. So, the following predicate verifies that
4046 * this segment is not used for anything but congestion avoidance or
4047 * fast retransmit. Moreover, we even are able to eliminate most of such
4048 * second order effects, if we apply some small "replay" window (~RTO)
4049 * to timestamp space.
4051 * All these measures still do not guarantee that we reject wrapped ACKs
4052 * on networks with high bandwidth, when sequence space is recycled fastly,
4053 * but it guarantees that such events will be very rare and do not affect
4054 * connection seriously. This doesn't look nice, but alas, PAWS is really
4057 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4058 * states that events when retransmit arrives after original data are rare.
4059 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4060 * the biggest problem on large power networks even with minor reordering.
4061 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4062 * up to bandwidth of 18Gigabit/sec. 8) ]
4065 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4067 const struct tcp_sock *tp = tcp_sk(sk);
4068 const struct tcphdr *th = tcp_hdr(skb);
4069 u32 seq = TCP_SKB_CB(skb)->seq;
4070 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4072 return (/* 1. Pure ACK with correct sequence number. */
4073 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4075 /* 2. ... and duplicate ACK. */
4076 ack == tp->snd_una &&
4078 /* 3. ... and does not update window. */
4079 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4081 /* 4. ... and sits in replay window. */
4082 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4085 static inline int tcp_paws_discard(const struct sock *sk,
4086 const struct sk_buff *skb)
4088 const struct tcp_sock *tp = tcp_sk(sk);
4090 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4091 !tcp_disordered_ack(sk, skb);
4094 /* Check segment sequence number for validity.
4096 * Segment controls are considered valid, if the segment
4097 * fits to the window after truncation to the window. Acceptability
4098 * of data (and SYN, FIN, of course) is checked separately.
4099 * See tcp_data_queue(), for example.
4101 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4102 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4103 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4104 * (borrowed from freebsd)
4107 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4109 return !before(end_seq, tp->rcv_wup) &&
4110 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4113 /* When we get a reset we do this. */
4114 static void tcp_reset(struct sock *sk)
4116 /* We want the right error as BSD sees it (and indeed as we do). */
4117 switch (sk->sk_state) {
4119 sk->sk_err = ECONNREFUSED;
4121 case TCP_CLOSE_WAIT:
4127 sk->sk_err = ECONNRESET;
4129 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4132 if (!sock_flag(sk, SOCK_DEAD))
4133 sk->sk_error_report(sk);
4139 * Process the FIN bit. This now behaves as it is supposed to work
4140 * and the FIN takes effect when it is validly part of sequence
4141 * space. Not before when we get holes.
4143 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4144 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4147 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4148 * close and we go into CLOSING (and later onto TIME-WAIT)
4150 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4152 static void tcp_fin(struct sock *sk)
4154 struct tcp_sock *tp = tcp_sk(sk);
4156 inet_csk_schedule_ack(sk);
4158 sk->sk_shutdown |= RCV_SHUTDOWN;
4159 sock_set_flag(sk, SOCK_DONE);
4161 switch (sk->sk_state) {
4163 case TCP_ESTABLISHED:
4164 /* Move to CLOSE_WAIT */
4165 tcp_set_state(sk, TCP_CLOSE_WAIT);
4166 inet_csk(sk)->icsk_ack.pingpong = 1;
4169 case TCP_CLOSE_WAIT:
4171 /* Received a retransmission of the FIN, do
4176 /* RFC793: Remain in the LAST-ACK state. */
4180 /* This case occurs when a simultaneous close
4181 * happens, we must ack the received FIN and
4182 * enter the CLOSING state.
4185 tcp_set_state(sk, TCP_CLOSING);
4188 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4190 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4193 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4194 * cases we should never reach this piece of code.
4196 pr_err("%s: Impossible, sk->sk_state=%d\n",
4197 __func__, sk->sk_state);
4201 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4202 * Probably, we should reset in this case. For now drop them.
4204 __skb_queue_purge(&tp->out_of_order_queue);
4205 if (tcp_is_sack(tp))
4206 tcp_sack_reset(&tp->rx_opt);
4209 if (!sock_flag(sk, SOCK_DEAD)) {
4210 sk->sk_state_change(sk);
4212 /* Do not send POLL_HUP for half duplex close. */
4213 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4214 sk->sk_state == TCP_CLOSE)
4215 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4217 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4221 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4224 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4225 if (before(seq, sp->start_seq))
4226 sp->start_seq = seq;
4227 if (after(end_seq, sp->end_seq))
4228 sp->end_seq = end_seq;
4234 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4236 struct tcp_sock *tp = tcp_sk(sk);
4238 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4241 if (before(seq, tp->rcv_nxt))
4242 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4244 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4246 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4248 tp->rx_opt.dsack = 1;
4249 tp->duplicate_sack[0].start_seq = seq;
4250 tp->duplicate_sack[0].end_seq = end_seq;
4254 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4256 struct tcp_sock *tp = tcp_sk(sk);
4258 if (!tp->rx_opt.dsack)
4259 tcp_dsack_set(sk, seq, end_seq);
4261 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4264 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4266 struct tcp_sock *tp = tcp_sk(sk);
4268 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4269 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4270 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4271 tcp_enter_quickack_mode(sk);
4273 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4274 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4276 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4277 end_seq = tp->rcv_nxt;
4278 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4285 /* These routines update the SACK block as out-of-order packets arrive or
4286 * in-order packets close up the sequence space.
4288 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4291 struct tcp_sack_block *sp = &tp->selective_acks[0];
4292 struct tcp_sack_block *swalk = sp + 1;
4294 /* See if the recent change to the first SACK eats into
4295 * or hits the sequence space of other SACK blocks, if so coalesce.
4297 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4298 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4301 /* Zap SWALK, by moving every further SACK up by one slot.
4302 * Decrease num_sacks.
4304 tp->rx_opt.num_sacks--;
4305 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4309 this_sack++, swalk++;
4313 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4315 struct tcp_sock *tp = tcp_sk(sk);
4316 struct tcp_sack_block *sp = &tp->selective_acks[0];
4317 int cur_sacks = tp->rx_opt.num_sacks;
4323 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4324 if (tcp_sack_extend(sp, seq, end_seq)) {
4325 /* Rotate this_sack to the first one. */
4326 for (; this_sack > 0; this_sack--, sp--)
4327 swap(*sp, *(sp - 1));
4329 tcp_sack_maybe_coalesce(tp);
4334 /* Could not find an adjacent existing SACK, build a new one,
4335 * put it at the front, and shift everyone else down. We
4336 * always know there is at least one SACK present already here.
4338 * If the sack array is full, forget about the last one.
4340 if (this_sack >= TCP_NUM_SACKS) {
4342 tp->rx_opt.num_sacks--;
4345 for (; this_sack > 0; this_sack--, sp--)
4349 /* Build the new head SACK, and we're done. */
4350 sp->start_seq = seq;
4351 sp->end_seq = end_seq;
4352 tp->rx_opt.num_sacks++;
4355 /* RCV.NXT advances, some SACKs should be eaten. */
4357 static void tcp_sack_remove(struct tcp_sock *tp)
4359 struct tcp_sack_block *sp = &tp->selective_acks[0];
4360 int num_sacks = tp->rx_opt.num_sacks;
4363 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4364 if (skb_queue_empty(&tp->out_of_order_queue)) {
4365 tp->rx_opt.num_sacks = 0;
4369 for (this_sack = 0; this_sack < num_sacks;) {
4370 /* Check if the start of the sack is covered by RCV.NXT. */
4371 if (!before(tp->rcv_nxt, sp->start_seq)) {
4374 /* RCV.NXT must cover all the block! */
4375 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4377 /* Zap this SACK, by moving forward any other SACKS. */
4378 for (i=this_sack+1; i < num_sacks; i++)
4379 tp->selective_acks[i-1] = tp->selective_acks[i];
4386 tp->rx_opt.num_sacks = num_sacks;
4389 /* This one checks to see if we can put data from the
4390 * out_of_order queue into the receive_queue.
4392 static void tcp_ofo_queue(struct sock *sk)
4394 struct tcp_sock *tp = tcp_sk(sk);
4395 __u32 dsack_high = tp->rcv_nxt;
4396 struct sk_buff *skb;
4398 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4399 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4402 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4403 __u32 dsack = dsack_high;
4404 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4405 dsack_high = TCP_SKB_CB(skb)->end_seq;
4406 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4409 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4410 SOCK_DEBUG(sk, "ofo packet was already received\n");
4411 __skb_unlink(skb, &tp->out_of_order_queue);
4415 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4416 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4417 TCP_SKB_CB(skb)->end_seq);
4419 __skb_unlink(skb, &tp->out_of_order_queue);
4420 __skb_queue_tail(&sk->sk_receive_queue, skb);
4421 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4422 if (tcp_hdr(skb)->fin)
4427 static int tcp_prune_ofo_queue(struct sock *sk);
4428 static int tcp_prune_queue(struct sock *sk);
4430 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4432 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4433 !sk_rmem_schedule(sk, size)) {
4435 if (tcp_prune_queue(sk) < 0)
4438 if (!sk_rmem_schedule(sk, size)) {
4439 if (!tcp_prune_ofo_queue(sk))
4442 if (!sk_rmem_schedule(sk, size))
4449 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4451 struct tcp_sock *tp = tcp_sk(sk);
4452 struct sk_buff *skb1;
4455 TCP_ECN_check_ce(tp, skb);
4457 if (tcp_try_rmem_schedule(sk, skb->truesize)) {
4458 /* TODO: should increment a counter */
4463 /* Disable header prediction. */
4465 inet_csk_schedule_ack(sk);
4467 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4468 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4470 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4472 /* Initial out of order segment, build 1 SACK. */
4473 if (tcp_is_sack(tp)) {
4474 tp->rx_opt.num_sacks = 1;
4475 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4476 tp->selective_acks[0].end_seq =
4477 TCP_SKB_CB(skb)->end_seq;
4479 __skb_queue_head(&tp->out_of_order_queue, skb);
4483 seq = TCP_SKB_CB(skb)->seq;
4484 end_seq = TCP_SKB_CB(skb)->end_seq;
4486 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4487 /* Packets in ofo can stay in queue a long time.
4488 * Better try to coalesce them right now
4489 * to avoid future tcp_collapse_ofo_queue(),
4490 * probably the most expensive function in tcp stack.
4492 if (skb->len <= skb_tailroom(skb1) && !tcp_hdr(skb)->fin) {
4493 NET_INC_STATS_BH(sock_net(sk),
4494 LINUX_MIB_TCPRCVCOALESCE);
4495 BUG_ON(skb_copy_bits(skb, 0,
4496 skb_put(skb1, skb->len),
4498 TCP_SKB_CB(skb1)->end_seq = end_seq;
4499 TCP_SKB_CB(skb1)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
4503 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4506 if (!tp->rx_opt.num_sacks ||
4507 tp->selective_acks[0].end_seq != seq)
4510 /* Common case: data arrive in order after hole. */
4511 tp->selective_acks[0].end_seq = end_seq;
4515 /* Find place to insert this segment. */
4517 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4519 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4523 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4526 /* Do skb overlap to previous one? */
4527 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4528 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4529 /* All the bits are present. Drop. */
4532 tcp_dsack_set(sk, seq, end_seq);
4535 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4536 /* Partial overlap. */
4537 tcp_dsack_set(sk, seq,
4538 TCP_SKB_CB(skb1)->end_seq);
4540 if (skb_queue_is_first(&tp->out_of_order_queue,
4544 skb1 = skb_queue_prev(
4545 &tp->out_of_order_queue,
4550 __skb_queue_head(&tp->out_of_order_queue, skb);
4552 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4554 /* And clean segments covered by new one as whole. */
4555 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4556 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4558 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4560 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4561 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4565 __skb_unlink(skb1, &tp->out_of_order_queue);
4566 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4567 TCP_SKB_CB(skb1)->end_seq);
4572 if (tcp_is_sack(tp))
4573 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4576 skb_set_owner_r(skb, sk);
4580 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4582 const struct tcphdr *th = tcp_hdr(skb);
4583 struct tcp_sock *tp = tcp_sk(sk);
4586 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4590 __skb_pull(skb, th->doff * 4);
4592 TCP_ECN_accept_cwr(tp, skb);
4594 tp->rx_opt.dsack = 0;
4596 /* Queue data for delivery to the user.
4597 * Packets in sequence go to the receive queue.
4598 * Out of sequence packets to the out_of_order_queue.
4600 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4601 if (tcp_receive_window(tp) == 0)
4604 /* Ok. In sequence. In window. */
4605 if (tp->ucopy.task == current &&
4606 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4607 sock_owned_by_user(sk) && !tp->urg_data) {
4608 int chunk = min_t(unsigned int, skb->len,
4611 __set_current_state(TASK_RUNNING);
4614 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4615 tp->ucopy.len -= chunk;
4616 tp->copied_seq += chunk;
4617 eaten = (chunk == skb->len);
4618 tcp_rcv_space_adjust(sk);
4626 tcp_try_rmem_schedule(sk, skb->truesize))
4629 skb_set_owner_r(skb, sk);
4630 __skb_queue_tail(&sk->sk_receive_queue, skb);
4632 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4634 tcp_event_data_recv(sk, skb);
4638 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4641 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4642 * gap in queue is filled.
4644 if (skb_queue_empty(&tp->out_of_order_queue))
4645 inet_csk(sk)->icsk_ack.pingpong = 0;
4648 if (tp->rx_opt.num_sacks)
4649 tcp_sack_remove(tp);
4651 tcp_fast_path_check(sk);
4655 else if (!sock_flag(sk, SOCK_DEAD))
4656 sk->sk_data_ready(sk, 0);
4660 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4661 /* A retransmit, 2nd most common case. Force an immediate ack. */
4662 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4663 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4666 tcp_enter_quickack_mode(sk);
4667 inet_csk_schedule_ack(sk);
4673 /* Out of window. F.e. zero window probe. */
4674 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4677 tcp_enter_quickack_mode(sk);
4679 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4680 /* Partial packet, seq < rcv_next < end_seq */
4681 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4682 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4683 TCP_SKB_CB(skb)->end_seq);
4685 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4687 /* If window is closed, drop tail of packet. But after
4688 * remembering D-SACK for its head made in previous line.
4690 if (!tcp_receive_window(tp))
4695 tcp_data_queue_ofo(sk, skb);
4698 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4699 struct sk_buff_head *list)
4701 struct sk_buff *next = NULL;
4703 if (!skb_queue_is_last(list, skb))
4704 next = skb_queue_next(list, skb);
4706 __skb_unlink(skb, list);
4708 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4713 /* Collapse contiguous sequence of skbs head..tail with
4714 * sequence numbers start..end.
4716 * If tail is NULL, this means until the end of the list.
4718 * Segments with FIN/SYN are not collapsed (only because this
4722 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4723 struct sk_buff *head, struct sk_buff *tail,
4726 struct sk_buff *skb, *n;
4729 /* First, check that queue is collapsible and find
4730 * the point where collapsing can be useful. */
4734 skb_queue_walk_from_safe(list, skb, n) {
4737 /* No new bits? It is possible on ofo queue. */
4738 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4739 skb = tcp_collapse_one(sk, skb, list);
4745 /* The first skb to collapse is:
4747 * - bloated or contains data before "start" or
4748 * overlaps to the next one.
4750 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4751 (tcp_win_from_space(skb->truesize) > skb->len ||
4752 before(TCP_SKB_CB(skb)->seq, start))) {
4753 end_of_skbs = false;
4757 if (!skb_queue_is_last(list, skb)) {
4758 struct sk_buff *next = skb_queue_next(list, skb);
4760 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4761 end_of_skbs = false;
4766 /* Decided to skip this, advance start seq. */
4767 start = TCP_SKB_CB(skb)->end_seq;
4769 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4772 while (before(start, end)) {
4773 struct sk_buff *nskb;
4774 unsigned int header = skb_headroom(skb);
4775 int copy = SKB_MAX_ORDER(header, 0);
4777 /* Too big header? This can happen with IPv6. */
4780 if (end - start < copy)
4782 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4786 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4787 skb_set_network_header(nskb, (skb_network_header(skb) -
4789 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4791 skb_reserve(nskb, header);
4792 memcpy(nskb->head, skb->head, header);
4793 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4794 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4795 __skb_queue_before(list, skb, nskb);
4796 skb_set_owner_r(nskb, sk);
4798 /* Copy data, releasing collapsed skbs. */
4800 int offset = start - TCP_SKB_CB(skb)->seq;
4801 int size = TCP_SKB_CB(skb)->end_seq - start;
4805 size = min(copy, size);
4806 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4808 TCP_SKB_CB(nskb)->end_seq += size;
4812 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4813 skb = tcp_collapse_one(sk, skb, list);
4816 tcp_hdr(skb)->syn ||
4824 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4825 * and tcp_collapse() them until all the queue is collapsed.
4827 static void tcp_collapse_ofo_queue(struct sock *sk)
4829 struct tcp_sock *tp = tcp_sk(sk);
4830 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4831 struct sk_buff *head;
4837 start = TCP_SKB_CB(skb)->seq;
4838 end = TCP_SKB_CB(skb)->end_seq;
4842 struct sk_buff *next = NULL;
4844 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4845 next = skb_queue_next(&tp->out_of_order_queue, skb);
4848 /* Segment is terminated when we see gap or when
4849 * we are at the end of all the queue. */
4851 after(TCP_SKB_CB(skb)->seq, end) ||
4852 before(TCP_SKB_CB(skb)->end_seq, start)) {
4853 tcp_collapse(sk, &tp->out_of_order_queue,
4854 head, skb, start, end);
4858 /* Start new segment */
4859 start = TCP_SKB_CB(skb)->seq;
4860 end = TCP_SKB_CB(skb)->end_seq;
4862 if (before(TCP_SKB_CB(skb)->seq, start))
4863 start = TCP_SKB_CB(skb)->seq;
4864 if (after(TCP_SKB_CB(skb)->end_seq, end))
4865 end = TCP_SKB_CB(skb)->end_seq;
4871 * Purge the out-of-order queue.
4872 * Return true if queue was pruned.
4874 static int tcp_prune_ofo_queue(struct sock *sk)
4876 struct tcp_sock *tp = tcp_sk(sk);
4879 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4880 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4881 __skb_queue_purge(&tp->out_of_order_queue);
4883 /* Reset SACK state. A conforming SACK implementation will
4884 * do the same at a timeout based retransmit. When a connection
4885 * is in a sad state like this, we care only about integrity
4886 * of the connection not performance.
4888 if (tp->rx_opt.sack_ok)
4889 tcp_sack_reset(&tp->rx_opt);
4896 /* Reduce allocated memory if we can, trying to get
4897 * the socket within its memory limits again.
4899 * Return less than zero if we should start dropping frames
4900 * until the socket owning process reads some of the data
4901 * to stabilize the situation.
4903 static int tcp_prune_queue(struct sock *sk)
4905 struct tcp_sock *tp = tcp_sk(sk);
4907 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4909 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4911 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4912 tcp_clamp_window(sk);
4913 else if (sk_under_memory_pressure(sk))
4914 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4916 tcp_collapse_ofo_queue(sk);
4917 if (!skb_queue_empty(&sk->sk_receive_queue))
4918 tcp_collapse(sk, &sk->sk_receive_queue,
4919 skb_peek(&sk->sk_receive_queue),
4921 tp->copied_seq, tp->rcv_nxt);
4924 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4927 /* Collapsing did not help, destructive actions follow.
4928 * This must not ever occur. */
4930 tcp_prune_ofo_queue(sk);
4932 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4935 /* If we are really being abused, tell the caller to silently
4936 * drop receive data on the floor. It will get retransmitted
4937 * and hopefully then we'll have sufficient space.
4939 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4941 /* Massive buffer overcommit. */
4946 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4947 * As additional protections, we do not touch cwnd in retransmission phases,
4948 * and if application hit its sndbuf limit recently.
4950 void tcp_cwnd_application_limited(struct sock *sk)
4952 struct tcp_sock *tp = tcp_sk(sk);
4954 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4955 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4956 /* Limited by application or receiver window. */
4957 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4958 u32 win_used = max(tp->snd_cwnd_used, init_win);
4959 if (win_used < tp->snd_cwnd) {
4960 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4961 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4963 tp->snd_cwnd_used = 0;
4965 tp->snd_cwnd_stamp = tcp_time_stamp;
4968 static int tcp_should_expand_sndbuf(const struct sock *sk)
4970 const struct tcp_sock *tp = tcp_sk(sk);
4972 /* If the user specified a specific send buffer setting, do
4975 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4978 /* If we are under global TCP memory pressure, do not expand. */
4979 if (sk_under_memory_pressure(sk))
4982 /* If we are under soft global TCP memory pressure, do not expand. */
4983 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4986 /* If we filled the congestion window, do not expand. */
4987 if (tp->packets_out >= tp->snd_cwnd)
4993 /* When incoming ACK allowed to free some skb from write_queue,
4994 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4995 * on the exit from tcp input handler.
4997 * PROBLEM: sndbuf expansion does not work well with largesend.
4999 static void tcp_new_space(struct sock *sk)
5001 struct tcp_sock *tp = tcp_sk(sk);
5003 if (tcp_should_expand_sndbuf(sk)) {
5004 int sndmem = SKB_TRUESIZE(max_t(u32,
5005 tp->rx_opt.mss_clamp,
5008 int demanded = max_t(unsigned int, tp->snd_cwnd,
5009 tp->reordering + 1);
5010 sndmem *= 2 * demanded;
5011 if (sndmem > sk->sk_sndbuf)
5012 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5013 tp->snd_cwnd_stamp = tcp_time_stamp;
5016 sk->sk_write_space(sk);
5019 static void tcp_check_space(struct sock *sk)
5021 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5022 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5023 if (sk->sk_socket &&
5024 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5029 static inline void tcp_data_snd_check(struct sock *sk)
5031 tcp_push_pending_frames(sk);
5032 tcp_check_space(sk);
5036 * Check if sending an ack is needed.
5038 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5040 struct tcp_sock *tp = tcp_sk(sk);
5042 /* More than one full frame received... */
5043 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5044 /* ... and right edge of window advances far enough.
5045 * (tcp_recvmsg() will send ACK otherwise). Or...
5047 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5048 /* We ACK each frame or... */
5049 tcp_in_quickack_mode(sk) ||
5050 /* We have out of order data. */
5051 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5052 /* Then ack it now */
5055 /* Else, send delayed ack. */
5056 tcp_send_delayed_ack(sk);
5060 static inline void tcp_ack_snd_check(struct sock *sk)
5062 if (!inet_csk_ack_scheduled(sk)) {
5063 /* We sent a data segment already. */
5066 __tcp_ack_snd_check(sk, 1);
5070 * This routine is only called when we have urgent data
5071 * signaled. Its the 'slow' part of tcp_urg. It could be
5072 * moved inline now as tcp_urg is only called from one
5073 * place. We handle URGent data wrong. We have to - as
5074 * BSD still doesn't use the correction from RFC961.
5075 * For 1003.1g we should support a new option TCP_STDURG to permit
5076 * either form (or just set the sysctl tcp_stdurg).
5079 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5081 struct tcp_sock *tp = tcp_sk(sk);
5082 u32 ptr = ntohs(th->urg_ptr);
5084 if (ptr && !sysctl_tcp_stdurg)
5086 ptr += ntohl(th->seq);
5088 /* Ignore urgent data that we've already seen and read. */
5089 if (after(tp->copied_seq, ptr))
5092 /* Do not replay urg ptr.
5094 * NOTE: interesting situation not covered by specs.
5095 * Misbehaving sender may send urg ptr, pointing to segment,
5096 * which we already have in ofo queue. We are not able to fetch
5097 * such data and will stay in TCP_URG_NOTYET until will be eaten
5098 * by recvmsg(). Seems, we are not obliged to handle such wicked
5099 * situations. But it is worth to think about possibility of some
5100 * DoSes using some hypothetical application level deadlock.
5102 if (before(ptr, tp->rcv_nxt))
5105 /* Do we already have a newer (or duplicate) urgent pointer? */
5106 if (tp->urg_data && !after(ptr, tp->urg_seq))
5109 /* Tell the world about our new urgent pointer. */
5112 /* We may be adding urgent data when the last byte read was
5113 * urgent. To do this requires some care. We cannot just ignore
5114 * tp->copied_seq since we would read the last urgent byte again
5115 * as data, nor can we alter copied_seq until this data arrives
5116 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5118 * NOTE. Double Dutch. Rendering to plain English: author of comment
5119 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5120 * and expect that both A and B disappear from stream. This is _wrong_.
5121 * Though this happens in BSD with high probability, this is occasional.
5122 * Any application relying on this is buggy. Note also, that fix "works"
5123 * only in this artificial test. Insert some normal data between A and B and we will
5124 * decline of BSD again. Verdict: it is better to remove to trap
5127 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5128 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5129 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5131 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5132 __skb_unlink(skb, &sk->sk_receive_queue);
5137 tp->urg_data = TCP_URG_NOTYET;
5140 /* Disable header prediction. */
5144 /* This is the 'fast' part of urgent handling. */
5145 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5147 struct tcp_sock *tp = tcp_sk(sk);
5149 /* Check if we get a new urgent pointer - normally not. */
5151 tcp_check_urg(sk, th);
5153 /* Do we wait for any urgent data? - normally not... */
5154 if (tp->urg_data == TCP_URG_NOTYET) {
5155 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5158 /* Is the urgent pointer pointing into this packet? */
5159 if (ptr < skb->len) {
5161 if (skb_copy_bits(skb, ptr, &tmp, 1))
5163 tp->urg_data = TCP_URG_VALID | tmp;
5164 if (!sock_flag(sk, SOCK_DEAD))
5165 sk->sk_data_ready(sk, 0);
5170 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5172 struct tcp_sock *tp = tcp_sk(sk);
5173 int chunk = skb->len - hlen;
5177 if (skb_csum_unnecessary(skb))
5178 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5180 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5184 tp->ucopy.len -= chunk;
5185 tp->copied_seq += chunk;
5186 tcp_rcv_space_adjust(sk);
5193 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5194 struct sk_buff *skb)
5198 if (sock_owned_by_user(sk)) {
5200 result = __tcp_checksum_complete(skb);
5203 result = __tcp_checksum_complete(skb);
5208 static inline int tcp_checksum_complete_user(struct sock *sk,
5209 struct sk_buff *skb)
5211 return !skb_csum_unnecessary(skb) &&
5212 __tcp_checksum_complete_user(sk, skb);
5215 #ifdef CONFIG_NET_DMA
5216 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5219 struct tcp_sock *tp = tcp_sk(sk);
5220 int chunk = skb->len - hlen;
5222 int copied_early = 0;
5224 if (tp->ucopy.wakeup)
5227 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5228 tp->ucopy.dma_chan = net_dma_find_channel();
5230 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5232 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5234 tp->ucopy.iov, chunk,
5235 tp->ucopy.pinned_list);
5240 tp->ucopy.dma_cookie = dma_cookie;
5243 tp->ucopy.len -= chunk;
5244 tp->copied_seq += chunk;
5245 tcp_rcv_space_adjust(sk);
5247 if ((tp->ucopy.len == 0) ||
5248 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5249 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5250 tp->ucopy.wakeup = 1;
5251 sk->sk_data_ready(sk, 0);
5253 } else if (chunk > 0) {
5254 tp->ucopy.wakeup = 1;
5255 sk->sk_data_ready(sk, 0);
5258 return copied_early;
5260 #endif /* CONFIG_NET_DMA */
5262 /* Does PAWS and seqno based validation of an incoming segment, flags will
5263 * play significant role here.
5265 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5266 const struct tcphdr *th, int syn_inerr)
5268 const u8 *hash_location;
5269 struct tcp_sock *tp = tcp_sk(sk);
5271 /* RFC1323: H1. Apply PAWS check first. */
5272 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5273 tp->rx_opt.saw_tstamp &&
5274 tcp_paws_discard(sk, skb)) {
5276 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5277 tcp_send_dupack(sk, skb);
5280 /* Reset is accepted even if it did not pass PAWS. */
5283 /* Step 1: check sequence number */
5284 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5285 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5286 * (RST) segments are validated by checking their SEQ-fields."
5287 * And page 69: "If an incoming segment is not acceptable,
5288 * an acknowledgment should be sent in reply (unless the RST
5289 * bit is set, if so drop the segment and return)".
5292 tcp_send_dupack(sk, skb);
5296 /* Step 2: check RST bit */
5302 /* ts_recent update must be made after we are sure that the packet
5305 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5307 /* step 3: check security and precedence [ignored] */
5309 /* step 4: Check for a SYN in window. */
5310 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5312 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5313 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5326 * TCP receive function for the ESTABLISHED state.
5328 * It is split into a fast path and a slow path. The fast path is
5330 * - A zero window was announced from us - zero window probing
5331 * is only handled properly in the slow path.
5332 * - Out of order segments arrived.
5333 * - Urgent data is expected.
5334 * - There is no buffer space left
5335 * - Unexpected TCP flags/window values/header lengths are received
5336 * (detected by checking the TCP header against pred_flags)
5337 * - Data is sent in both directions. Fast path only supports pure senders
5338 * or pure receivers (this means either the sequence number or the ack
5339 * value must stay constant)
5340 * - Unexpected TCP option.
5342 * When these conditions are not satisfied it drops into a standard
5343 * receive procedure patterned after RFC793 to handle all cases.
5344 * The first three cases are guaranteed by proper pred_flags setting,
5345 * the rest is checked inline. Fast processing is turned on in
5346 * tcp_data_queue when everything is OK.
5348 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5349 const struct tcphdr *th, unsigned int len)
5351 struct tcp_sock *tp = tcp_sk(sk);
5355 * Header prediction.
5356 * The code loosely follows the one in the famous
5357 * "30 instruction TCP receive" Van Jacobson mail.
5359 * Van's trick is to deposit buffers into socket queue
5360 * on a device interrupt, to call tcp_recv function
5361 * on the receive process context and checksum and copy
5362 * the buffer to user space. smart...
5364 * Our current scheme is not silly either but we take the
5365 * extra cost of the net_bh soft interrupt processing...
5366 * We do checksum and copy also but from device to kernel.
5369 tp->rx_opt.saw_tstamp = 0;
5371 /* pred_flags is 0xS?10 << 16 + snd_wnd
5372 * if header_prediction is to be made
5373 * 'S' will always be tp->tcp_header_len >> 2
5374 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5375 * turn it off (when there are holes in the receive
5376 * space for instance)
5377 * PSH flag is ignored.
5380 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5381 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5382 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5383 int tcp_header_len = tp->tcp_header_len;
5385 /* Timestamp header prediction: tcp_header_len
5386 * is automatically equal to th->doff*4 due to pred_flags
5390 /* Check timestamp */
5391 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5392 /* No? Slow path! */
5393 if (!tcp_parse_aligned_timestamp(tp, th))
5396 /* If PAWS failed, check it more carefully in slow path */
5397 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5400 /* DO NOT update ts_recent here, if checksum fails
5401 * and timestamp was corrupted part, it will result
5402 * in a hung connection since we will drop all
5403 * future packets due to the PAWS test.
5407 if (len <= tcp_header_len) {
5408 /* Bulk data transfer: sender */
5409 if (len == tcp_header_len) {
5410 /* Predicted packet is in window by definition.
5411 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5412 * Hence, check seq<=rcv_wup reduces to:
5414 if (tcp_header_len ==
5415 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5416 tp->rcv_nxt == tp->rcv_wup)
5417 tcp_store_ts_recent(tp);
5419 /* We know that such packets are checksummed
5422 tcp_ack(sk, skb, 0);
5424 tcp_data_snd_check(sk);
5426 } else { /* Header too small */
5427 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5432 int copied_early = 0;
5434 if (tp->copied_seq == tp->rcv_nxt &&
5435 len - tcp_header_len <= tp->ucopy.len) {
5436 #ifdef CONFIG_NET_DMA
5437 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5442 if (tp->ucopy.task == current &&
5443 sock_owned_by_user(sk) && !copied_early) {
5444 __set_current_state(TASK_RUNNING);
5446 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5450 /* Predicted packet is in window by definition.
5451 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5452 * Hence, check seq<=rcv_wup reduces to:
5454 if (tcp_header_len ==
5455 (sizeof(struct tcphdr) +
5456 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 __skb_pull(skb, tcp_header_len);
5463 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5464 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5467 tcp_cleanup_rbuf(sk, skb->len);
5470 if (tcp_checksum_complete_user(sk, skb))
5473 /* Predicted packet is in window by definition.
5474 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5475 * Hence, check seq<=rcv_wup reduces to:
5477 if (tcp_header_len ==
5478 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5479 tp->rcv_nxt == tp->rcv_wup)
5480 tcp_store_ts_recent(tp);
5482 tcp_rcv_rtt_measure_ts(sk, skb);
5484 if ((int)skb->truesize > sk->sk_forward_alloc)
5487 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5489 /* Bulk data transfer: receiver */
5490 __skb_pull(skb, tcp_header_len);
5491 __skb_queue_tail(&sk->sk_receive_queue, skb);
5492 skb_set_owner_r(skb, sk);
5493 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5496 tcp_event_data_recv(sk, skb);
5498 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5499 /* Well, only one small jumplet in fast path... */
5500 tcp_ack(sk, skb, FLAG_DATA);
5501 tcp_data_snd_check(sk);
5502 if (!inet_csk_ack_scheduled(sk))
5506 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5507 __tcp_ack_snd_check(sk, 0);
5509 #ifdef CONFIG_NET_DMA
5511 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5517 sk->sk_data_ready(sk, 0);
5523 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5527 * Standard slow path.
5530 res = tcp_validate_incoming(sk, skb, th, 1);
5535 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5538 tcp_rcv_rtt_measure_ts(sk, skb);
5540 /* Process urgent data. */
5541 tcp_urg(sk, skb, th);
5543 /* step 7: process the segment text */
5544 tcp_data_queue(sk, skb);
5546 tcp_data_snd_check(sk);
5547 tcp_ack_snd_check(sk);
5551 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5557 EXPORT_SYMBOL(tcp_rcv_established);
5559 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5560 const struct tcphdr *th, unsigned int len)
5562 const u8 *hash_location;
5563 struct inet_connection_sock *icsk = inet_csk(sk);
5564 struct tcp_sock *tp = tcp_sk(sk);
5565 struct tcp_cookie_values *cvp = tp->cookie_values;
5566 int saved_clamp = tp->rx_opt.mss_clamp;
5568 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5572 * "If the state is SYN-SENT then
5573 * first check the ACK bit
5574 * If the ACK bit is set
5575 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5576 * a reset (unless the RST bit is set, if so drop
5577 * the segment and return)"
5579 * We do not send data with SYN, so that RFC-correct
5582 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5583 goto reset_and_undo;
5585 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5586 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5588 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5589 goto reset_and_undo;
5592 /* Now ACK is acceptable.
5594 * "If the RST bit is set
5595 * If the ACK was acceptable then signal the user "error:
5596 * connection reset", drop the segment, enter CLOSED state,
5597 * delete TCB, and return."
5606 * "fifth, if neither of the SYN or RST bits is set then
5607 * drop the segment and return."
5613 goto discard_and_undo;
5616 * "If the SYN bit is on ...
5617 * are acceptable then ...
5618 * (our SYN has been ACKed), change the connection
5619 * state to ESTABLISHED..."
5622 TCP_ECN_rcv_synack(tp, th);
5624 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5625 tcp_ack(sk, skb, FLAG_SLOWPATH);
5627 /* Ok.. it's good. Set up sequence numbers and
5628 * move to established.
5630 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5631 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5633 /* RFC1323: The window in SYN & SYN/ACK segments is
5636 tp->snd_wnd = ntohs(th->window);
5637 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5639 if (!tp->rx_opt.wscale_ok) {
5640 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5641 tp->window_clamp = min(tp->window_clamp, 65535U);
5644 if (tp->rx_opt.saw_tstamp) {
5645 tp->rx_opt.tstamp_ok = 1;
5646 tp->tcp_header_len =
5647 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5648 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5649 tcp_store_ts_recent(tp);
5651 tp->tcp_header_len = sizeof(struct tcphdr);
5654 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5655 tcp_enable_fack(tp);
5658 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5659 tcp_initialize_rcv_mss(sk);
5661 /* Remember, tcp_poll() does not lock socket!
5662 * Change state from SYN-SENT only after copied_seq
5663 * is initialized. */
5664 tp->copied_seq = tp->rcv_nxt;
5667 cvp->cookie_pair_size > 0 &&
5668 tp->rx_opt.cookie_plus > 0) {
5669 int cookie_size = tp->rx_opt.cookie_plus
5670 - TCPOLEN_COOKIE_BASE;
5671 int cookie_pair_size = cookie_size
5672 + cvp->cookie_desired;
5674 /* A cookie extension option was sent and returned.
5675 * Note that each incoming SYNACK replaces the
5676 * Responder cookie. The initial exchange is most
5677 * fragile, as protection against spoofing relies
5678 * entirely upon the sequence and timestamp (above).
5679 * This replacement strategy allows the correct pair to
5680 * pass through, while any others will be filtered via
5681 * Responder verification later.
5683 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5684 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5685 hash_location, cookie_size);
5686 cvp->cookie_pair_size = cookie_pair_size;
5691 tcp_set_state(sk, TCP_ESTABLISHED);
5693 security_inet_conn_established(sk, skb);
5695 /* Make sure socket is routed, for correct metrics. */
5696 icsk->icsk_af_ops->rebuild_header(sk);
5698 tcp_init_metrics(sk);
5700 tcp_init_congestion_control(sk);
5702 /* Prevent spurious tcp_cwnd_restart() on first data
5705 tp->lsndtime = tcp_time_stamp;
5707 tcp_init_buffer_space(sk);
5709 if (sock_flag(sk, SOCK_KEEPOPEN))
5710 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5712 if (!tp->rx_opt.snd_wscale)
5713 __tcp_fast_path_on(tp, tp->snd_wnd);
5717 if (!sock_flag(sk, SOCK_DEAD)) {
5718 sk->sk_state_change(sk);
5719 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5722 if (sk->sk_write_pending ||
5723 icsk->icsk_accept_queue.rskq_defer_accept ||
5724 icsk->icsk_ack.pingpong) {
5725 /* Save one ACK. Data will be ready after
5726 * several ticks, if write_pending is set.
5728 * It may be deleted, but with this feature tcpdumps
5729 * look so _wonderfully_ clever, that I was not able
5730 * to stand against the temptation 8) --ANK
5732 inet_csk_schedule_ack(sk);
5733 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5734 icsk->icsk_ack.ato = TCP_ATO_MIN;
5735 tcp_incr_quickack(sk);
5736 tcp_enter_quickack_mode(sk);
5737 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5738 TCP_DELACK_MAX, TCP_RTO_MAX);
5749 /* No ACK in the segment */
5753 * "If the RST bit is set
5755 * Otherwise (no ACK) drop the segment and return."
5758 goto discard_and_undo;
5762 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5763 tcp_paws_reject(&tp->rx_opt, 0))
5764 goto discard_and_undo;
5767 /* We see SYN without ACK. It is attempt of
5768 * simultaneous connect with crossed SYNs.
5769 * Particularly, it can be connect to self.
5771 tcp_set_state(sk, TCP_SYN_RECV);
5773 if (tp->rx_opt.saw_tstamp) {
5774 tp->rx_opt.tstamp_ok = 1;
5775 tcp_store_ts_recent(tp);
5776 tp->tcp_header_len =
5777 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5779 tp->tcp_header_len = sizeof(struct tcphdr);
5782 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5783 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5785 /* RFC1323: The window in SYN & SYN/ACK segments is
5788 tp->snd_wnd = ntohs(th->window);
5789 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5790 tp->max_window = tp->snd_wnd;
5792 TCP_ECN_rcv_syn(tp, th);
5795 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5796 tcp_initialize_rcv_mss(sk);
5798 tcp_send_synack(sk);
5800 /* Note, we could accept data and URG from this segment.
5801 * There are no obstacles to make this.
5803 * However, if we ignore data in ACKless segments sometimes,
5804 * we have no reasons to accept it sometimes.
5805 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5806 * is not flawless. So, discard packet for sanity.
5807 * Uncomment this return to process the data.
5814 /* "fifth, if neither of the SYN or RST bits is set then
5815 * drop the segment and return."
5819 tcp_clear_options(&tp->rx_opt);
5820 tp->rx_opt.mss_clamp = saved_clamp;
5824 tcp_clear_options(&tp->rx_opt);
5825 tp->rx_opt.mss_clamp = saved_clamp;
5830 * This function implements the receiving procedure of RFC 793 for
5831 * all states except ESTABLISHED and TIME_WAIT.
5832 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5833 * address independent.
5836 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5837 const struct tcphdr *th, unsigned int len)
5839 struct tcp_sock *tp = tcp_sk(sk);
5840 struct inet_connection_sock *icsk = inet_csk(sk);
5844 tp->rx_opt.saw_tstamp = 0;
5846 switch (sk->sk_state) {
5860 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5863 /* Now we have several options: In theory there is
5864 * nothing else in the frame. KA9Q has an option to
5865 * send data with the syn, BSD accepts data with the
5866 * syn up to the [to be] advertised window and
5867 * Solaris 2.1 gives you a protocol error. For now
5868 * we just ignore it, that fits the spec precisely
5869 * and avoids incompatibilities. It would be nice in
5870 * future to drop through and process the data.
5872 * Now that TTCP is starting to be used we ought to
5874 * But, this leaves one open to an easy denial of
5875 * service attack, and SYN cookies can't defend
5876 * against this problem. So, we drop the data
5877 * in the interest of security over speed unless
5878 * it's still in use.
5886 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5890 /* Do step6 onward by hand. */
5891 tcp_urg(sk, skb, th);
5893 tcp_data_snd_check(sk);
5897 res = tcp_validate_incoming(sk, skb, th, 0);
5901 /* step 5: check the ACK field */
5903 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
5905 switch (sk->sk_state) {
5908 tp->copied_seq = tp->rcv_nxt;
5910 tcp_set_state(sk, TCP_ESTABLISHED);
5911 sk->sk_state_change(sk);
5913 /* Note, that this wakeup is only for marginal
5914 * crossed SYN case. Passively open sockets
5915 * are not waked up, because sk->sk_sleep ==
5916 * NULL and sk->sk_socket == NULL.
5920 SOCK_WAKE_IO, POLL_OUT);
5922 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5923 tp->snd_wnd = ntohs(th->window) <<
5924 tp->rx_opt.snd_wscale;
5925 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5927 if (tp->rx_opt.tstamp_ok)
5928 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5930 /* Make sure socket is routed, for
5933 icsk->icsk_af_ops->rebuild_header(sk);
5935 tcp_init_metrics(sk);
5937 tcp_init_congestion_control(sk);
5939 /* Prevent spurious tcp_cwnd_restart() on
5940 * first data packet.
5942 tp->lsndtime = tcp_time_stamp;
5945 tcp_initialize_rcv_mss(sk);
5946 tcp_init_buffer_space(sk);
5947 tcp_fast_path_on(tp);
5954 if (tp->snd_una == tp->write_seq) {
5955 tcp_set_state(sk, TCP_FIN_WAIT2);
5956 sk->sk_shutdown |= SEND_SHUTDOWN;
5957 dst_confirm(__sk_dst_get(sk));
5959 if (!sock_flag(sk, SOCK_DEAD))
5960 /* Wake up lingering close() */
5961 sk->sk_state_change(sk);
5965 if (tp->linger2 < 0 ||
5966 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5967 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5969 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5973 tmo = tcp_fin_time(sk);
5974 if (tmo > TCP_TIMEWAIT_LEN) {
5975 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5976 } else if (th->fin || sock_owned_by_user(sk)) {
5977 /* Bad case. We could lose such FIN otherwise.
5978 * It is not a big problem, but it looks confusing
5979 * and not so rare event. We still can lose it now,
5980 * if it spins in bh_lock_sock(), but it is really
5983 inet_csk_reset_keepalive_timer(sk, tmo);
5985 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5993 if (tp->snd_una == tp->write_seq) {
5994 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6000 if (tp->snd_una == tp->write_seq) {
6001 tcp_update_metrics(sk);
6010 /* step 6: check the URG bit */
6011 tcp_urg(sk, skb, th);
6013 /* step 7: process the segment text */
6014 switch (sk->sk_state) {
6015 case TCP_CLOSE_WAIT:
6018 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6022 /* RFC 793 says to queue data in these states,
6023 * RFC 1122 says we MUST send a reset.
6024 * BSD 4.4 also does reset.
6026 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6027 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6028 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6029 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6035 case TCP_ESTABLISHED:
6036 tcp_data_queue(sk, skb);
6041 /* tcp_data could move socket to TIME-WAIT */
6042 if (sk->sk_state != TCP_CLOSE) {
6043 tcp_data_snd_check(sk);
6044 tcp_ack_snd_check(sk);
6053 EXPORT_SYMBOL(tcp_rcv_state_process);