2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #include <linux/slab.h>
66 #include <linux/module.h>
67 #include <linux/sysctl.h>
68 #include <linux/kernel.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
74 #include <net/netdma.h>
76 int sysctl_tcp_timestamps __read_mostly = 1;
77 int sysctl_tcp_window_scaling __read_mostly = 1;
78 int sysctl_tcp_sack __read_mostly = 1;
79 int sysctl_tcp_fack __read_mostly = 1;
80 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
81 EXPORT_SYMBOL(sysctl_tcp_reordering);
82 int sysctl_tcp_ecn __read_mostly = 2;
83 EXPORT_SYMBOL(sysctl_tcp_ecn);
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 1000;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
96 int sysctl_tcp_frto_response __read_mostly;
97 int sysctl_tcp_nometrics_save __read_mostly;
99 int sysctl_tcp_thin_dupack __read_mostly;
101 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102 int sysctl_tcp_abc __read_mostly;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
117 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
118 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
120 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
123 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
126 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
127 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
129 /* Adapt the MSS value used to make delayed ack decision to the
132 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
134 struct inet_connection_sock *icsk = inet_csk(sk);
135 const unsigned int lss = icsk->icsk_ack.last_seg_size;
138 icsk->icsk_ack.last_seg_size = 0;
140 /* skb->len may jitter because of SACKs, even if peer
141 * sends good full-sized frames.
143 len = skb_shinfo(skb)->gso_size ? : skb->len;
144 if (len >= icsk->icsk_ack.rcv_mss) {
145 icsk->icsk_ack.rcv_mss = len;
147 /* Otherwise, we make more careful check taking into account,
148 * that SACKs block is variable.
150 * "len" is invariant segment length, including TCP header.
152 len += skb->data - skb_transport_header(skb);
153 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
154 /* If PSH is not set, packet should be
155 * full sized, provided peer TCP is not badly broken.
156 * This observation (if it is correct 8)) allows
157 * to handle super-low mtu links fairly.
159 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
160 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
161 /* Subtract also invariant (if peer is RFC compliant),
162 * tcp header plus fixed timestamp option length.
163 * Resulting "len" is MSS free of SACK jitter.
165 len -= tcp_sk(sk)->tcp_header_len;
166 icsk->icsk_ack.last_seg_size = len;
168 icsk->icsk_ack.rcv_mss = len;
172 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
173 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
174 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
178 static void tcp_incr_quickack(struct sock *sk)
180 struct inet_connection_sock *icsk = inet_csk(sk);
181 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
185 if (quickacks > icsk->icsk_ack.quick)
186 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
189 static void tcp_enter_quickack_mode(struct sock *sk)
191 struct inet_connection_sock *icsk = inet_csk(sk);
192 tcp_incr_quickack(sk);
193 icsk->icsk_ack.pingpong = 0;
194 icsk->icsk_ack.ato = TCP_ATO_MIN;
197 /* Send ACKs quickly, if "quick" count is not exhausted
198 * and the session is not interactive.
201 static inline int tcp_in_quickack_mode(const struct sock *sk)
203 const struct inet_connection_sock *icsk = inet_csk(sk);
204 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
207 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
209 if (tp->ecn_flags & TCP_ECN_OK)
210 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
213 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
215 if (tcp_hdr(skb)->cwr)
216 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
219 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
221 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
224 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
226 if (!(tp->ecn_flags & TCP_ECN_OK))
229 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
230 case INET_ECN_NOT_ECT:
231 /* Funny extension: if ECT is not set on a segment,
232 * and we already seen ECT on a previous segment,
233 * it is probably a retransmit.
235 if (tp->ecn_flags & TCP_ECN_SEEN)
236 tcp_enter_quickack_mode((struct sock *)tp);
239 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
242 tp->ecn_flags |= TCP_ECN_SEEN;
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
248 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
249 tp->ecn_flags &= ~TCP_ECN_OK;
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
254 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
255 tp->ecn_flags &= ~TCP_ECN_OK;
258 static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
260 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
265 /* Buffer size and advertised window tuning.
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270 static void tcp_fixup_sndbuf(struct sock *sk)
272 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
274 sndmem *= TCP_INIT_CWND;
275 if (sk->sk_sndbuf < sndmem)
276 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
279 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
281 * All tcp_full_space() is split to two parts: "network" buffer, allocated
282 * forward and advertised in receiver window (tp->rcv_wnd) and
283 * "application buffer", required to isolate scheduling/application
284 * latencies from network.
285 * window_clamp is maximal advertised window. It can be less than
286 * tcp_full_space(), in this case tcp_full_space() - window_clamp
287 * is reserved for "application" buffer. The less window_clamp is
288 * the smoother our behaviour from viewpoint of network, but the lower
289 * throughput and the higher sensitivity of the connection to losses. 8)
291 * rcv_ssthresh is more strict window_clamp used at "slow start"
292 * phase to predict further behaviour of this connection.
293 * It is used for two goals:
294 * - to enforce header prediction at sender, even when application
295 * requires some significant "application buffer". It is check #1.
296 * - to prevent pruning of receive queue because of misprediction
297 * of receiver window. Check #2.
299 * The scheme does not work when sender sends good segments opening
300 * window and then starts to feed us spaghetti. But it should work
301 * in common situations. Otherwise, we have to rely on queue collapsing.
304 /* Slow part of check#2. */
305 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
307 struct tcp_sock *tp = tcp_sk(sk);
309 int truesize = tcp_win_from_space(skb->truesize) >> 1;
310 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
312 while (tp->rcv_ssthresh <= window) {
313 if (truesize <= skb->len)
314 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
322 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
324 struct tcp_sock *tp = tcp_sk(sk);
327 if (tp->rcv_ssthresh < tp->window_clamp &&
328 (int)tp->rcv_ssthresh < tcp_space(sk) &&
329 !tcp_memory_pressure) {
332 /* Check #2. Increase window, if skb with such overhead
333 * will fit to rcvbuf in future.
335 if (tcp_win_from_space(skb->truesize) <= skb->len)
336 incr = 2 * tp->advmss;
338 incr = __tcp_grow_window(sk, skb);
341 incr = max_t(int, incr, 2 * skb->len);
342 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
344 inet_csk(sk)->icsk_ack.quick |= 1;
349 /* 3. Tuning rcvbuf, when connection enters established state. */
351 static void tcp_fixup_rcvbuf(struct sock *sk)
353 u32 mss = tcp_sk(sk)->advmss;
354 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
357 /* Limit to 10 segments if mss <= 1460,
358 * or 14600/mss segments, with a minimum of two segments.
361 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
363 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
364 while (tcp_win_from_space(rcvmem) < mss)
369 if (sk->sk_rcvbuf < rcvmem)
370 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
373 /* 4. Try to fixup all. It is made immediately after connection enters
376 static void tcp_init_buffer_space(struct sock *sk)
378 struct tcp_sock *tp = tcp_sk(sk);
381 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
382 tcp_fixup_rcvbuf(sk);
383 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
384 tcp_fixup_sndbuf(sk);
386 tp->rcvq_space.space = tp->rcv_wnd;
388 maxwin = tcp_full_space(sk);
390 if (tp->window_clamp >= maxwin) {
391 tp->window_clamp = maxwin;
393 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
394 tp->window_clamp = max(maxwin -
395 (maxwin >> sysctl_tcp_app_win),
399 /* Force reservation of one segment. */
400 if (sysctl_tcp_app_win &&
401 tp->window_clamp > 2 * tp->advmss &&
402 tp->window_clamp + tp->advmss > maxwin)
403 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
405 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
406 tp->snd_cwnd_stamp = tcp_time_stamp;
409 /* 5. Recalculate window clamp after socket hit its memory bounds. */
410 static void tcp_clamp_window(struct sock *sk)
412 struct tcp_sock *tp = tcp_sk(sk);
413 struct inet_connection_sock *icsk = inet_csk(sk);
415 icsk->icsk_ack.quick = 0;
417 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
418 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
419 !tcp_memory_pressure &&
420 atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
421 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
424 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
425 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
428 /* Initialize RCV_MSS value.
429 * RCV_MSS is an our guess about MSS used by the peer.
430 * We haven't any direct information about the MSS.
431 * It's better to underestimate the RCV_MSS rather than overestimate.
432 * Overestimations make us ACKing less frequently than needed.
433 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
435 void tcp_initialize_rcv_mss(struct sock *sk)
437 const struct tcp_sock *tp = tcp_sk(sk);
438 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
440 hint = min(hint, tp->rcv_wnd / 2);
441 hint = min(hint, TCP_MSS_DEFAULT);
442 hint = max(hint, TCP_MIN_MSS);
444 inet_csk(sk)->icsk_ack.rcv_mss = hint;
446 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
448 /* Receiver "autotuning" code.
450 * The algorithm for RTT estimation w/o timestamps is based on
451 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
452 * <http://public.lanl.gov/radiant/pubs.html#DRS>
454 * More detail on this code can be found at
455 * <http://staff.psc.edu/jheffner/>,
456 * though this reference is out of date. A new paper
459 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
461 u32 new_sample = tp->rcv_rtt_est.rtt;
467 if (new_sample != 0) {
468 /* If we sample in larger samples in the non-timestamp
469 * case, we could grossly overestimate the RTT especially
470 * with chatty applications or bulk transfer apps which
471 * are stalled on filesystem I/O.
473 * Also, since we are only going for a minimum in the
474 * non-timestamp case, we do not smooth things out
475 * else with timestamps disabled convergence takes too
479 m -= (new_sample >> 3);
487 /* No previous measure. */
491 if (tp->rcv_rtt_est.rtt != new_sample)
492 tp->rcv_rtt_est.rtt = new_sample;
495 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
497 if (tp->rcv_rtt_est.time == 0)
499 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
501 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
504 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
505 tp->rcv_rtt_est.time = tcp_time_stamp;
508 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
509 const struct sk_buff *skb)
511 struct tcp_sock *tp = tcp_sk(sk);
512 if (tp->rx_opt.rcv_tsecr &&
513 (TCP_SKB_CB(skb)->end_seq -
514 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
515 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
519 * This function should be called every time data is copied to user space.
520 * It calculates the appropriate TCP receive buffer space.
522 void tcp_rcv_space_adjust(struct sock *sk)
524 struct tcp_sock *tp = tcp_sk(sk);
528 if (tp->rcvq_space.time == 0)
531 time = tcp_time_stamp - tp->rcvq_space.time;
532 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
535 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
537 space = max(tp->rcvq_space.space, space);
539 if (tp->rcvq_space.space != space) {
542 tp->rcvq_space.space = space;
544 if (sysctl_tcp_moderate_rcvbuf &&
545 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
546 int new_clamp = space;
548 /* Receive space grows, normalize in order to
549 * take into account packet headers and sk_buff
550 * structure overhead.
555 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
556 while (tcp_win_from_space(rcvmem) < tp->advmss)
559 space = min(space, sysctl_tcp_rmem[2]);
560 if (space > sk->sk_rcvbuf) {
561 sk->sk_rcvbuf = space;
563 /* Make the window clamp follow along. */
564 tp->window_clamp = new_clamp;
570 tp->rcvq_space.seq = tp->copied_seq;
571 tp->rcvq_space.time = tcp_time_stamp;
574 /* There is something which you must keep in mind when you analyze the
575 * behavior of the tp->ato delayed ack timeout interval. When a
576 * connection starts up, we want to ack as quickly as possible. The
577 * problem is that "good" TCP's do slow start at the beginning of data
578 * transmission. The means that until we send the first few ACK's the
579 * sender will sit on his end and only queue most of his data, because
580 * he can only send snd_cwnd unacked packets at any given time. For
581 * each ACK we send, he increments snd_cwnd and transmits more of his
584 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
586 struct tcp_sock *tp = tcp_sk(sk);
587 struct inet_connection_sock *icsk = inet_csk(sk);
590 inet_csk_schedule_ack(sk);
592 tcp_measure_rcv_mss(sk, skb);
594 tcp_rcv_rtt_measure(tp);
596 now = tcp_time_stamp;
598 if (!icsk->icsk_ack.ato) {
599 /* The _first_ data packet received, initialize
600 * delayed ACK engine.
602 tcp_incr_quickack(sk);
603 icsk->icsk_ack.ato = TCP_ATO_MIN;
605 int m = now - icsk->icsk_ack.lrcvtime;
607 if (m <= TCP_ATO_MIN / 2) {
608 /* The fastest case is the first. */
609 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
610 } else if (m < icsk->icsk_ack.ato) {
611 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
612 if (icsk->icsk_ack.ato > icsk->icsk_rto)
613 icsk->icsk_ack.ato = icsk->icsk_rto;
614 } else if (m > icsk->icsk_rto) {
615 /* Too long gap. Apparently sender failed to
616 * restart window, so that we send ACKs quickly.
618 tcp_incr_quickack(sk);
622 icsk->icsk_ack.lrcvtime = now;
624 TCP_ECN_check_ce(tp, skb);
627 tcp_grow_window(sk, skb);
630 /* Called to compute a smoothed rtt estimate. The data fed to this
631 * routine either comes from timestamps, or from segments that were
632 * known _not_ to have been retransmitted [see Karn/Partridge
633 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
634 * piece by Van Jacobson.
635 * NOTE: the next three routines used to be one big routine.
636 * To save cycles in the RFC 1323 implementation it was better to break
637 * it up into three procedures. -- erics
639 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
641 struct tcp_sock *tp = tcp_sk(sk);
642 long m = mrtt; /* RTT */
644 /* The following amusing code comes from Jacobson's
645 * article in SIGCOMM '88. Note that rtt and mdev
646 * are scaled versions of rtt and mean deviation.
647 * This is designed to be as fast as possible
648 * m stands for "measurement".
650 * On a 1990 paper the rto value is changed to:
651 * RTO = rtt + 4 * mdev
653 * Funny. This algorithm seems to be very broken.
654 * These formulae increase RTO, when it should be decreased, increase
655 * too slowly, when it should be increased quickly, decrease too quickly
656 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
657 * does not matter how to _calculate_ it. Seems, it was trap
658 * that VJ failed to avoid. 8)
663 m -= (tp->srtt >> 3); /* m is now error in rtt est */
664 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
666 m = -m; /* m is now abs(error) */
667 m -= (tp->mdev >> 2); /* similar update on mdev */
668 /* This is similar to one of Eifel findings.
669 * Eifel blocks mdev updates when rtt decreases.
670 * This solution is a bit different: we use finer gain
671 * for mdev in this case (alpha*beta).
672 * Like Eifel it also prevents growth of rto,
673 * but also it limits too fast rto decreases,
674 * happening in pure Eifel.
679 m -= (tp->mdev >> 2); /* similar update on mdev */
681 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
682 if (tp->mdev > tp->mdev_max) {
683 tp->mdev_max = tp->mdev;
684 if (tp->mdev_max > tp->rttvar)
685 tp->rttvar = tp->mdev_max;
687 if (after(tp->snd_una, tp->rtt_seq)) {
688 if (tp->mdev_max < tp->rttvar)
689 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
690 tp->rtt_seq = tp->snd_nxt;
691 tp->mdev_max = tcp_rto_min(sk);
694 /* no previous measure. */
695 tp->srtt = m << 3; /* take the measured time to be rtt */
696 tp->mdev = m << 1; /* make sure rto = 3*rtt */
697 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
698 tp->rtt_seq = tp->snd_nxt;
702 /* Calculate rto without backoff. This is the second half of Van Jacobson's
703 * routine referred to above.
705 static inline void tcp_set_rto(struct sock *sk)
707 const struct tcp_sock *tp = tcp_sk(sk);
708 /* Old crap is replaced with new one. 8)
711 * 1. If rtt variance happened to be less 50msec, it is hallucination.
712 * It cannot be less due to utterly erratic ACK generation made
713 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
714 * to do with delayed acks, because at cwnd>2 true delack timeout
715 * is invisible. Actually, Linux-2.4 also generates erratic
716 * ACKs in some circumstances.
718 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
720 /* 2. Fixups made earlier cannot be right.
721 * If we do not estimate RTO correctly without them,
722 * all the algo is pure shit and should be replaced
723 * with correct one. It is exactly, which we pretend to do.
726 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
727 * guarantees that rto is higher.
732 /* Save metrics learned by this TCP session.
733 This function is called only, when TCP finishes successfully
734 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
736 void tcp_update_metrics(struct sock *sk)
738 struct tcp_sock *tp = tcp_sk(sk);
739 struct dst_entry *dst = __sk_dst_get(sk);
741 if (sysctl_tcp_nometrics_save)
746 if (dst && (dst->flags & DST_HOST)) {
747 const struct inet_connection_sock *icsk = inet_csk(sk);
751 if (icsk->icsk_backoff || !tp->srtt) {
752 /* This session failed to estimate rtt. Why?
753 * Probably, no packets returned in time.
756 if (!(dst_metric_locked(dst, RTAX_RTT)))
757 dst_metric_set(dst, RTAX_RTT, 0);
761 rtt = dst_metric_rtt(dst, RTAX_RTT);
764 /* If newly calculated rtt larger than stored one,
765 * store new one. Otherwise, use EWMA. Remember,
766 * rtt overestimation is always better than underestimation.
768 if (!(dst_metric_locked(dst, RTAX_RTT))) {
770 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
772 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
775 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
780 /* Scale deviation to rttvar fixed point */
785 var = dst_metric_rtt(dst, RTAX_RTTVAR);
789 var -= (var - m) >> 2;
791 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
794 if (tcp_in_initial_slowstart(tp)) {
795 /* Slow start still did not finish. */
796 if (dst_metric(dst, RTAX_SSTHRESH) &&
797 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
798 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
799 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
800 if (!dst_metric_locked(dst, RTAX_CWND) &&
801 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
802 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
803 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
804 icsk->icsk_ca_state == TCP_CA_Open) {
805 /* Cong. avoidance phase, cwnd is reliable. */
806 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
807 dst_metric_set(dst, RTAX_SSTHRESH,
808 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
809 if (!dst_metric_locked(dst, RTAX_CWND))
810 dst_metric_set(dst, RTAX_CWND,
811 (dst_metric(dst, RTAX_CWND) +
814 /* Else slow start did not finish, cwnd is non-sense,
815 ssthresh may be also invalid.
817 if (!dst_metric_locked(dst, RTAX_CWND))
818 dst_metric_set(dst, RTAX_CWND,
819 (dst_metric(dst, RTAX_CWND) +
820 tp->snd_ssthresh) >> 1);
821 if (dst_metric(dst, RTAX_SSTHRESH) &&
822 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
823 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
824 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
827 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
828 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
829 tp->reordering != sysctl_tcp_reordering)
830 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
835 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
837 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
840 cwnd = TCP_INIT_CWND;
841 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
844 /* Set slow start threshold and cwnd not falling to slow start */
845 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
847 struct tcp_sock *tp = tcp_sk(sk);
848 const struct inet_connection_sock *icsk = inet_csk(sk);
850 tp->prior_ssthresh = 0;
852 if (icsk->icsk_ca_state < TCP_CA_CWR) {
855 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
856 tp->snd_cwnd = min(tp->snd_cwnd,
857 tcp_packets_in_flight(tp) + 1U);
858 tp->snd_cwnd_cnt = 0;
859 tp->high_seq = tp->snd_nxt;
860 tp->snd_cwnd_stamp = tcp_time_stamp;
861 TCP_ECN_queue_cwr(tp);
863 tcp_set_ca_state(sk, TCP_CA_CWR);
868 * Packet counting of FACK is based on in-order assumptions, therefore TCP
869 * disables it when reordering is detected
871 static void tcp_disable_fack(struct tcp_sock *tp)
873 /* RFC3517 uses different metric in lost marker => reset on change */
875 tp->lost_skb_hint = NULL;
876 tp->rx_opt.sack_ok &= ~2;
879 /* Take a notice that peer is sending D-SACKs */
880 static void tcp_dsack_seen(struct tcp_sock *tp)
882 tp->rx_opt.sack_ok |= 4;
885 /* Initialize metrics on socket. */
887 static void tcp_init_metrics(struct sock *sk)
889 struct tcp_sock *tp = tcp_sk(sk);
890 struct dst_entry *dst = __sk_dst_get(sk);
897 if (dst_metric_locked(dst, RTAX_CWND))
898 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
899 if (dst_metric(dst, RTAX_SSTHRESH)) {
900 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
901 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
902 tp->snd_ssthresh = tp->snd_cwnd_clamp;
904 /* ssthresh may have been reduced unnecessarily during.
905 * 3WHS. Restore it back to its initial default.
907 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
909 if (dst_metric(dst, RTAX_REORDERING) &&
910 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
911 tcp_disable_fack(tp);
912 tp->reordering = dst_metric(dst, RTAX_REORDERING);
915 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
918 /* Initial rtt is determined from SYN,SYN-ACK.
919 * The segment is small and rtt may appear much
920 * less than real one. Use per-dst memory
921 * to make it more realistic.
923 * A bit of theory. RTT is time passed after "normal" sized packet
924 * is sent until it is ACKed. In normal circumstances sending small
925 * packets force peer to delay ACKs and calculation is correct too.
926 * The algorithm is adaptive and, provided we follow specs, it
927 * NEVER underestimate RTT. BUT! If peer tries to make some clever
928 * tricks sort of "quick acks" for time long enough to decrease RTT
929 * to low value, and then abruptly stops to do it and starts to delay
930 * ACKs, wait for troubles.
932 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
933 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
934 tp->rtt_seq = tp->snd_nxt;
936 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
937 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
938 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
943 /* RFC2988bis: We've failed to get a valid RTT sample from
944 * 3WHS. This is most likely due to retransmission,
945 * including spurious one. Reset the RTO back to 3secs
946 * from the more aggressive 1sec to avoid more spurious
949 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
950 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
952 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
953 * retransmitted. In light of RFC2988bis' more aggressive 1sec
954 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
955 * retransmission has occurred.
957 if (tp->total_retrans > 1)
960 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
961 tp->snd_cwnd_stamp = tcp_time_stamp;
964 static void tcp_update_reordering(struct sock *sk, const int metric,
967 struct tcp_sock *tp = tcp_sk(sk);
968 if (metric > tp->reordering) {
971 tp->reordering = min(TCP_MAX_REORDERING, metric);
973 /* This exciting event is worth to be remembered. 8) */
975 mib_idx = LINUX_MIB_TCPTSREORDER;
976 else if (tcp_is_reno(tp))
977 mib_idx = LINUX_MIB_TCPRENOREORDER;
978 else if (tcp_is_fack(tp))
979 mib_idx = LINUX_MIB_TCPFACKREORDER;
981 mib_idx = LINUX_MIB_TCPSACKREORDER;
983 NET_INC_STATS_BH(sock_net(sk), mib_idx);
984 #if FASTRETRANS_DEBUG > 1
985 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
986 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
990 tp->undo_marker ? tp->undo_retrans : 0);
992 tcp_disable_fack(tp);
996 /* This must be called before lost_out is incremented */
997 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
999 if ((tp->retransmit_skb_hint == NULL) ||
1000 before(TCP_SKB_CB(skb)->seq,
1001 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1002 tp->retransmit_skb_hint = skb;
1004 if (!tp->lost_out ||
1005 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
1006 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1009 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1011 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1012 tcp_verify_retransmit_hint(tp, skb);
1014 tp->lost_out += tcp_skb_pcount(skb);
1015 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1019 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1020 struct sk_buff *skb)
1022 tcp_verify_retransmit_hint(tp, skb);
1024 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1025 tp->lost_out += tcp_skb_pcount(skb);
1026 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1030 /* This procedure tags the retransmission queue when SACKs arrive.
1032 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1033 * Packets in queue with these bits set are counted in variables
1034 * sacked_out, retrans_out and lost_out, correspondingly.
1036 * Valid combinations are:
1037 * Tag InFlight Description
1038 * 0 1 - orig segment is in flight.
1039 * S 0 - nothing flies, orig reached receiver.
1040 * L 0 - nothing flies, orig lost by net.
1041 * R 2 - both orig and retransmit are in flight.
1042 * L|R 1 - orig is lost, retransmit is in flight.
1043 * S|R 1 - orig reached receiver, retrans is still in flight.
1044 * (L|S|R is logically valid, it could occur when L|R is sacked,
1045 * but it is equivalent to plain S and code short-curcuits it to S.
1046 * L|S is logically invalid, it would mean -1 packet in flight 8))
1048 * These 6 states form finite state machine, controlled by the following events:
1049 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1050 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1051 * 3. Loss detection event of one of three flavors:
1052 * A. Scoreboard estimator decided the packet is lost.
1053 * A'. Reno "three dupacks" marks head of queue lost.
1054 * A''. Its FACK modfication, head until snd.fack is lost.
1055 * B. SACK arrives sacking data transmitted after never retransmitted
1056 * hole was sent out.
1057 * C. SACK arrives sacking SND.NXT at the moment, when the
1058 * segment was retransmitted.
1059 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1061 * It is pleasant to note, that state diagram turns out to be commutative,
1062 * so that we are allowed not to be bothered by order of our actions,
1063 * when multiple events arrive simultaneously. (see the function below).
1065 * Reordering detection.
1066 * --------------------
1067 * Reordering metric is maximal distance, which a packet can be displaced
1068 * in packet stream. With SACKs we can estimate it:
1070 * 1. SACK fills old hole and the corresponding segment was not
1071 * ever retransmitted -> reordering. Alas, we cannot use it
1072 * when segment was retransmitted.
1073 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1074 * for retransmitted and already SACKed segment -> reordering..
1075 * Both of these heuristics are not used in Loss state, when we cannot
1076 * account for retransmits accurately.
1078 * SACK block validation.
1079 * ----------------------
1081 * SACK block range validation checks that the received SACK block fits to
1082 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1083 * Note that SND.UNA is not included to the range though being valid because
1084 * it means that the receiver is rather inconsistent with itself reporting
1085 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1086 * perfectly valid, however, in light of RFC2018 which explicitly states
1087 * that "SACK block MUST reflect the newest segment. Even if the newest
1088 * segment is going to be discarded ...", not that it looks very clever
1089 * in case of head skb. Due to potentional receiver driven attacks, we
1090 * choose to avoid immediate execution of a walk in write queue due to
1091 * reneging and defer head skb's loss recovery to standard loss recovery
1092 * procedure that will eventually trigger (nothing forbids us doing this).
1094 * Implements also blockage to start_seq wrap-around. Problem lies in the
1095 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1096 * there's no guarantee that it will be before snd_nxt (n). The problem
1097 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1100 * <- outs wnd -> <- wrapzone ->
1101 * u e n u_w e_w s n_w
1103 * |<------------+------+----- TCP seqno space --------------+---------->|
1104 * ...-- <2^31 ->| |<--------...
1105 * ...---- >2^31 ------>| |<--------...
1107 * Current code wouldn't be vulnerable but it's better still to discard such
1108 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1109 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1110 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1111 * equal to the ideal case (infinite seqno space without wrap caused issues).
1113 * With D-SACK the lower bound is extended to cover sequence space below
1114 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1115 * again, D-SACK block must not to go across snd_una (for the same reason as
1116 * for the normal SACK blocks, explained above). But there all simplicity
1117 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1118 * fully below undo_marker they do not affect behavior in anyway and can
1119 * therefore be safely ignored. In rare cases (which are more or less
1120 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1121 * fragmentation and packet reordering past skb's retransmission. To consider
1122 * them correctly, the acceptable range must be extended even more though
1123 * the exact amount is rather hard to quantify. However, tp->max_window can
1124 * be used as an exaggerated estimate.
1126 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1127 u32 start_seq, u32 end_seq)
1129 /* Too far in future, or reversed (interpretation is ambiguous) */
1130 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1133 /* Nasty start_seq wrap-around check (see comments above) */
1134 if (!before(start_seq, tp->snd_nxt))
1137 /* In outstanding window? ...This is valid exit for D-SACKs too.
1138 * start_seq == snd_una is non-sensical (see comments above)
1140 if (after(start_seq, tp->snd_una))
1143 if (!is_dsack || !tp->undo_marker)
1146 /* ...Then it's D-SACK, and must reside below snd_una completely */
1147 if (after(end_seq, tp->snd_una))
1150 if (!before(start_seq, tp->undo_marker))
1154 if (!after(end_seq, tp->undo_marker))
1157 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1158 * start_seq < undo_marker and end_seq >= undo_marker.
1160 return !before(start_seq, end_seq - tp->max_window);
1163 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1164 * Event "C". Later note: FACK people cheated me again 8), we have to account
1165 * for reordering! Ugly, but should help.
1167 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1168 * less than what is now known to be received by the other end (derived from
1169 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1170 * retransmitted skbs to avoid some costly processing per ACKs.
1172 static void tcp_mark_lost_retrans(struct sock *sk)
1174 const struct inet_connection_sock *icsk = inet_csk(sk);
1175 struct tcp_sock *tp = tcp_sk(sk);
1176 struct sk_buff *skb;
1178 u32 new_low_seq = tp->snd_nxt;
1179 u32 received_upto = tcp_highest_sack_seq(tp);
1181 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1182 !after(received_upto, tp->lost_retrans_low) ||
1183 icsk->icsk_ca_state != TCP_CA_Recovery)
1186 tcp_for_write_queue(skb, sk) {
1187 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1189 if (skb == tcp_send_head(sk))
1191 if (cnt == tp->retrans_out)
1193 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1196 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1199 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1200 * constraint here (see above) but figuring out that at
1201 * least tp->reordering SACK blocks reside between ack_seq
1202 * and received_upto is not easy task to do cheaply with
1203 * the available datastructures.
1205 * Whether FACK should check here for tp->reordering segs
1206 * in-between one could argue for either way (it would be
1207 * rather simple to implement as we could count fack_count
1208 * during the walk and do tp->fackets_out - fack_count).
1210 if (after(received_upto, ack_seq)) {
1211 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1212 tp->retrans_out -= tcp_skb_pcount(skb);
1214 tcp_skb_mark_lost_uncond_verify(tp, skb);
1215 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1217 if (before(ack_seq, new_low_seq))
1218 new_low_seq = ack_seq;
1219 cnt += tcp_skb_pcount(skb);
1223 if (tp->retrans_out)
1224 tp->lost_retrans_low = new_low_seq;
1227 static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1228 struct tcp_sack_block_wire *sp, int num_sacks,
1231 struct tcp_sock *tp = tcp_sk(sk);
1232 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1233 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1236 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1239 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1240 } else if (num_sacks > 1) {
1241 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1242 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1244 if (!after(end_seq_0, end_seq_1) &&
1245 !before(start_seq_0, start_seq_1)) {
1248 NET_INC_STATS_BH(sock_net(sk),
1249 LINUX_MIB_TCPDSACKOFORECV);
1253 /* D-SACK for already forgotten data... Do dumb counting. */
1254 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1255 !after(end_seq_0, prior_snd_una) &&
1256 after(end_seq_0, tp->undo_marker))
1262 struct tcp_sacktag_state {
1268 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1269 * the incoming SACK may not exactly match but we can find smaller MSS
1270 * aligned portion of it that matches. Therefore we might need to fragment
1271 * which may fail and creates some hassle (caller must handle error case
1274 * FIXME: this could be merged to shift decision code
1276 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1277 u32 start_seq, u32 end_seq)
1280 unsigned int pkt_len;
1283 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1284 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1286 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1287 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1288 mss = tcp_skb_mss(skb);
1289 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1292 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1296 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1301 /* Round if necessary so that SACKs cover only full MSSes
1302 * and/or the remaining small portion (if present)
1304 if (pkt_len > mss) {
1305 unsigned int new_len = (pkt_len / mss) * mss;
1306 if (!in_sack && new_len < pkt_len) {
1308 if (new_len >= skb->len)
1313 err = tcp_fragment(sk, skb, pkt_len, mss);
1321 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1322 static u8 tcp_sacktag_one(struct sock *sk,
1323 struct tcp_sacktag_state *state, u8 sacked,
1324 u32 start_seq, u32 end_seq,
1325 int dup_sack, int pcount)
1327 struct tcp_sock *tp = tcp_sk(sk);
1328 int fack_count = state->fack_count;
1330 /* Account D-SACK for retransmitted packet. */
1331 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1332 if (tp->undo_marker && tp->undo_retrans &&
1333 after(end_seq, tp->undo_marker))
1335 if (sacked & TCPCB_SACKED_ACKED)
1336 state->reord = min(fack_count, state->reord);
1339 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1340 if (!after(end_seq, tp->snd_una))
1343 if (!(sacked & TCPCB_SACKED_ACKED)) {
1344 if (sacked & TCPCB_SACKED_RETRANS) {
1345 /* If the segment is not tagged as lost,
1346 * we do not clear RETRANS, believing
1347 * that retransmission is still in flight.
1349 if (sacked & TCPCB_LOST) {
1350 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1351 tp->lost_out -= pcount;
1352 tp->retrans_out -= pcount;
1355 if (!(sacked & TCPCB_RETRANS)) {
1356 /* New sack for not retransmitted frame,
1357 * which was in hole. It is reordering.
1359 if (before(start_seq,
1360 tcp_highest_sack_seq(tp)))
1361 state->reord = min(fack_count,
1364 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1365 if (!after(end_seq, tp->frto_highmark))
1366 state->flag |= FLAG_ONLY_ORIG_SACKED;
1369 if (sacked & TCPCB_LOST) {
1370 sacked &= ~TCPCB_LOST;
1371 tp->lost_out -= pcount;
1375 sacked |= TCPCB_SACKED_ACKED;
1376 state->flag |= FLAG_DATA_SACKED;
1377 tp->sacked_out += pcount;
1379 fack_count += pcount;
1381 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1382 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1383 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1384 tp->lost_cnt_hint += pcount;
1386 if (fack_count > tp->fackets_out)
1387 tp->fackets_out = fack_count;
1390 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1391 * frames and clear it. undo_retrans is decreased above, L|R frames
1392 * are accounted above as well.
1394 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1395 sacked &= ~TCPCB_SACKED_RETRANS;
1396 tp->retrans_out -= pcount;
1402 /* Shift newly-SACKed bytes from this skb to the immediately previous
1403 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1405 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1406 struct tcp_sacktag_state *state,
1407 unsigned int pcount, int shifted, int mss,
1410 struct tcp_sock *tp = tcp_sk(sk);
1411 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1412 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1413 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1417 /* Adjust counters and hints for the newly sacked sequence
1418 * range but discard the return value since prev is already
1419 * marked. We must tag the range first because the seq
1420 * advancement below implicitly advances
1421 * tcp_highest_sack_seq() when skb is highest_sack.
1423 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1424 start_seq, end_seq, dup_sack, pcount);
1426 if (skb == tp->lost_skb_hint)
1427 tp->lost_cnt_hint += pcount;
1429 TCP_SKB_CB(prev)->end_seq += shifted;
1430 TCP_SKB_CB(skb)->seq += shifted;
1432 skb_shinfo(prev)->gso_segs += pcount;
1433 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1434 skb_shinfo(skb)->gso_segs -= pcount;
1436 /* When we're adding to gso_segs == 1, gso_size will be zero,
1437 * in theory this shouldn't be necessary but as long as DSACK
1438 * code can come after this skb later on it's better to keep
1439 * setting gso_size to something.
1441 if (!skb_shinfo(prev)->gso_size) {
1442 skb_shinfo(prev)->gso_size = mss;
1443 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1446 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1447 if (skb_shinfo(skb)->gso_segs <= 1) {
1448 skb_shinfo(skb)->gso_size = 0;
1449 skb_shinfo(skb)->gso_type = 0;
1452 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1453 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1456 BUG_ON(!tcp_skb_pcount(skb));
1457 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1461 /* Whole SKB was eaten :-) */
1463 if (skb == tp->retransmit_skb_hint)
1464 tp->retransmit_skb_hint = prev;
1465 if (skb == tp->scoreboard_skb_hint)
1466 tp->scoreboard_skb_hint = prev;
1467 if (skb == tp->lost_skb_hint) {
1468 tp->lost_skb_hint = prev;
1469 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1472 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1473 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1474 TCP_SKB_CB(prev)->end_seq++;
1476 if (skb == tcp_highest_sack(sk))
1477 tcp_advance_highest_sack(sk, skb);
1479 tcp_unlink_write_queue(skb, sk);
1480 sk_wmem_free_skb(sk, skb);
1482 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1487 /* I wish gso_size would have a bit more sane initialization than
1488 * something-or-zero which complicates things
1490 static int tcp_skb_seglen(const struct sk_buff *skb)
1492 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1495 /* Shifting pages past head area doesn't work */
1496 static int skb_can_shift(const struct sk_buff *skb)
1498 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1501 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1504 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1505 struct tcp_sacktag_state *state,
1506 u32 start_seq, u32 end_seq,
1509 struct tcp_sock *tp = tcp_sk(sk);
1510 struct sk_buff *prev;
1516 if (!sk_can_gso(sk))
1519 /* Normally R but no L won't result in plain S */
1521 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1523 if (!skb_can_shift(skb))
1525 /* This frame is about to be dropped (was ACKed). */
1526 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1529 /* Can only happen with delayed DSACK + discard craziness */
1530 if (unlikely(skb == tcp_write_queue_head(sk)))
1532 prev = tcp_write_queue_prev(sk, skb);
1534 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1537 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1538 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1542 pcount = tcp_skb_pcount(skb);
1543 mss = tcp_skb_seglen(skb);
1545 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1546 * drop this restriction as unnecessary
1548 if (mss != tcp_skb_seglen(prev))
1551 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1553 /* CHECKME: This is non-MSS split case only?, this will
1554 * cause skipped skbs due to advancing loop btw, original
1555 * has that feature too
1557 if (tcp_skb_pcount(skb) <= 1)
1560 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1562 /* TODO: head merge to next could be attempted here
1563 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1564 * though it might not be worth of the additional hassle
1566 * ...we can probably just fallback to what was done
1567 * previously. We could try merging non-SACKed ones
1568 * as well but it probably isn't going to buy off
1569 * because later SACKs might again split them, and
1570 * it would make skb timestamp tracking considerably
1576 len = end_seq - TCP_SKB_CB(skb)->seq;
1578 BUG_ON(len > skb->len);
1580 /* MSS boundaries should be honoured or else pcount will
1581 * severely break even though it makes things bit trickier.
1582 * Optimize common case to avoid most of the divides
1584 mss = tcp_skb_mss(skb);
1586 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1587 * drop this restriction as unnecessary
1589 if (mss != tcp_skb_seglen(prev))
1594 } else if (len < mss) {
1602 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1603 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1606 if (!skb_shift(prev, skb, len))
1608 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1611 /* Hole filled allows collapsing with the next as well, this is very
1612 * useful when hole on every nth skb pattern happens
1614 if (prev == tcp_write_queue_tail(sk))
1616 skb = tcp_write_queue_next(sk, prev);
1618 if (!skb_can_shift(skb) ||
1619 (skb == tcp_send_head(sk)) ||
1620 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1621 (mss != tcp_skb_seglen(skb)))
1625 if (skb_shift(prev, skb, len)) {
1626 pcount += tcp_skb_pcount(skb);
1627 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1631 state->fack_count += pcount;
1638 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1642 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1643 struct tcp_sack_block *next_dup,
1644 struct tcp_sacktag_state *state,
1645 u32 start_seq, u32 end_seq,
1648 struct tcp_sock *tp = tcp_sk(sk);
1649 struct sk_buff *tmp;
1651 tcp_for_write_queue_from(skb, sk) {
1653 int dup_sack = dup_sack_in;
1655 if (skb == tcp_send_head(sk))
1658 /* queue is in-order => we can short-circuit the walk early */
1659 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1662 if ((next_dup != NULL) &&
1663 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1664 in_sack = tcp_match_skb_to_sack(sk, skb,
1665 next_dup->start_seq,
1671 /* skb reference here is a bit tricky to get right, since
1672 * shifting can eat and free both this skb and the next,
1673 * so not even _safe variant of the loop is enough.
1676 tmp = tcp_shift_skb_data(sk, skb, state,
1677 start_seq, end_seq, dup_sack);
1686 in_sack = tcp_match_skb_to_sack(sk, skb,
1692 if (unlikely(in_sack < 0))
1696 TCP_SKB_CB(skb)->sacked =
1699 TCP_SKB_CB(skb)->sacked,
1700 TCP_SKB_CB(skb)->seq,
1701 TCP_SKB_CB(skb)->end_seq,
1703 tcp_skb_pcount(skb));
1705 if (!before(TCP_SKB_CB(skb)->seq,
1706 tcp_highest_sack_seq(tp)))
1707 tcp_advance_highest_sack(sk, skb);
1710 state->fack_count += tcp_skb_pcount(skb);
1715 /* Avoid all extra work that is being done by sacktag while walking in
1718 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1719 struct tcp_sacktag_state *state,
1722 tcp_for_write_queue_from(skb, sk) {
1723 if (skb == tcp_send_head(sk))
1726 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1729 state->fack_count += tcp_skb_pcount(skb);
1734 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1736 struct tcp_sack_block *next_dup,
1737 struct tcp_sacktag_state *state,
1740 if (next_dup == NULL)
1743 if (before(next_dup->start_seq, skip_to_seq)) {
1744 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1745 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1746 next_dup->start_seq, next_dup->end_seq,
1753 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1755 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1759 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1762 const struct inet_connection_sock *icsk = inet_csk(sk);
1763 struct tcp_sock *tp = tcp_sk(sk);
1764 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1765 TCP_SKB_CB(ack_skb)->sacked);
1766 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1767 struct tcp_sack_block sp[TCP_NUM_SACKS];
1768 struct tcp_sack_block *cache;
1769 struct tcp_sacktag_state state;
1770 struct sk_buff *skb;
1771 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1773 int found_dup_sack = 0;
1775 int first_sack_index;
1778 state.reord = tp->packets_out;
1780 if (!tp->sacked_out) {
1781 if (WARN_ON(tp->fackets_out))
1782 tp->fackets_out = 0;
1783 tcp_highest_sack_reset(sk);
1786 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1787 num_sacks, prior_snd_una);
1789 state.flag |= FLAG_DSACKING_ACK;
1791 /* Eliminate too old ACKs, but take into
1792 * account more or less fresh ones, they can
1793 * contain valid SACK info.
1795 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1798 if (!tp->packets_out)
1802 first_sack_index = 0;
1803 for (i = 0; i < num_sacks; i++) {
1804 int dup_sack = !i && found_dup_sack;
1806 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1807 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1809 if (!tcp_is_sackblock_valid(tp, dup_sack,
1810 sp[used_sacks].start_seq,
1811 sp[used_sacks].end_seq)) {
1815 if (!tp->undo_marker)
1816 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1818 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1820 /* Don't count olds caused by ACK reordering */
1821 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1822 !after(sp[used_sacks].end_seq, tp->snd_una))
1824 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1827 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1829 first_sack_index = -1;
1833 /* Ignore very old stuff early */
1834 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1840 /* order SACK blocks to allow in order walk of the retrans queue */
1841 for (i = used_sacks - 1; i > 0; i--) {
1842 for (j = 0; j < i; j++) {
1843 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1844 swap(sp[j], sp[j + 1]);
1846 /* Track where the first SACK block goes to */
1847 if (j == first_sack_index)
1848 first_sack_index = j + 1;
1853 skb = tcp_write_queue_head(sk);
1854 state.fack_count = 0;
1857 if (!tp->sacked_out) {
1858 /* It's already past, so skip checking against it */
1859 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1861 cache = tp->recv_sack_cache;
1862 /* Skip empty blocks in at head of the cache */
1863 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1868 while (i < used_sacks) {
1869 u32 start_seq = sp[i].start_seq;
1870 u32 end_seq = sp[i].end_seq;
1871 int dup_sack = (found_dup_sack && (i == first_sack_index));
1872 struct tcp_sack_block *next_dup = NULL;
1874 if (found_dup_sack && ((i + 1) == first_sack_index))
1875 next_dup = &sp[i + 1];
1877 /* Event "B" in the comment above. */
1878 if (after(end_seq, tp->high_seq))
1879 state.flag |= FLAG_DATA_LOST;
1881 /* Skip too early cached blocks */
1882 while (tcp_sack_cache_ok(tp, cache) &&
1883 !before(start_seq, cache->end_seq))
1886 /* Can skip some work by looking recv_sack_cache? */
1887 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1888 after(end_seq, cache->start_seq)) {
1891 if (before(start_seq, cache->start_seq)) {
1892 skb = tcp_sacktag_skip(skb, sk, &state,
1894 skb = tcp_sacktag_walk(skb, sk, next_dup,
1901 /* Rest of the block already fully processed? */
1902 if (!after(end_seq, cache->end_seq))
1905 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1909 /* ...tail remains todo... */
1910 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1911 /* ...but better entrypoint exists! */
1912 skb = tcp_highest_sack(sk);
1915 state.fack_count = tp->fackets_out;
1920 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1921 /* Check overlap against next cached too (past this one already) */
1926 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1927 skb = tcp_highest_sack(sk);
1930 state.fack_count = tp->fackets_out;
1932 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1935 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1936 start_seq, end_seq, dup_sack);
1939 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1940 * due to in-order walk
1942 if (after(end_seq, tp->frto_highmark))
1943 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1948 /* Clear the head of the cache sack blocks so we can skip it next time */
1949 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1950 tp->recv_sack_cache[i].start_seq = 0;
1951 tp->recv_sack_cache[i].end_seq = 0;
1953 for (j = 0; j < used_sacks; j++)
1954 tp->recv_sack_cache[i++] = sp[j];
1956 tcp_mark_lost_retrans(sk);
1958 tcp_verify_left_out(tp);
1960 if ((state.reord < tp->fackets_out) &&
1961 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1962 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1963 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1967 #if FASTRETRANS_DEBUG > 0
1968 WARN_ON((int)tp->sacked_out < 0);
1969 WARN_ON((int)tp->lost_out < 0);
1970 WARN_ON((int)tp->retrans_out < 0);
1971 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1976 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1977 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1979 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1983 holes = max(tp->lost_out, 1U);
1984 holes = min(holes, tp->packets_out);
1986 if ((tp->sacked_out + holes) > tp->packets_out) {
1987 tp->sacked_out = tp->packets_out - holes;
1993 /* If we receive more dupacks than we expected counting segments
1994 * in assumption of absent reordering, interpret this as reordering.
1995 * The only another reason could be bug in receiver TCP.
1997 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1999 struct tcp_sock *tp = tcp_sk(sk);
2000 if (tcp_limit_reno_sacked(tp))
2001 tcp_update_reordering(sk, tp->packets_out + addend, 0);
2004 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2006 static void tcp_add_reno_sack(struct sock *sk)
2008 struct tcp_sock *tp = tcp_sk(sk);
2010 tcp_check_reno_reordering(sk, 0);
2011 tcp_verify_left_out(tp);
2014 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2016 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
2018 struct tcp_sock *tp = tcp_sk(sk);
2021 /* One ACK acked hole. The rest eat duplicate ACKs. */
2022 if (acked - 1 >= tp->sacked_out)
2025 tp->sacked_out -= acked - 1;
2027 tcp_check_reno_reordering(sk, acked);
2028 tcp_verify_left_out(tp);
2031 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2036 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2038 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2041 /* F-RTO can only be used if TCP has never retransmitted anything other than
2042 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2044 int tcp_use_frto(struct sock *sk)
2046 const struct tcp_sock *tp = tcp_sk(sk);
2047 const struct inet_connection_sock *icsk = inet_csk(sk);
2048 struct sk_buff *skb;
2050 if (!sysctl_tcp_frto)
2053 /* MTU probe and F-RTO won't really play nicely along currently */
2054 if (icsk->icsk_mtup.probe_size)
2057 if (tcp_is_sackfrto(tp))
2060 /* Avoid expensive walking of rexmit queue if possible */
2061 if (tp->retrans_out > 1)
2064 skb = tcp_write_queue_head(sk);
2065 if (tcp_skb_is_last(sk, skb))
2067 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2068 tcp_for_write_queue_from(skb, sk) {
2069 if (skb == tcp_send_head(sk))
2071 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2073 /* Short-circuit when first non-SACKed skb has been checked */
2074 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2080 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2081 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2082 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2083 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2084 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2085 * bits are handled if the Loss state is really to be entered (in
2086 * tcp_enter_frto_loss).
2088 * Do like tcp_enter_loss() would; when RTO expires the second time it
2090 * "Reduce ssthresh if it has not yet been made inside this window."
2092 void tcp_enter_frto(struct sock *sk)
2094 const struct inet_connection_sock *icsk = inet_csk(sk);
2095 struct tcp_sock *tp = tcp_sk(sk);
2096 struct sk_buff *skb;
2098 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2099 tp->snd_una == tp->high_seq ||
2100 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2101 !icsk->icsk_retransmits)) {
2102 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2103 /* Our state is too optimistic in ssthresh() call because cwnd
2104 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2105 * recovery has not yet completed. Pattern would be this: RTO,
2106 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2108 * RFC4138 should be more specific on what to do, even though
2109 * RTO is quite unlikely to occur after the first Cumulative ACK
2110 * due to back-off and complexity of triggering events ...
2112 if (tp->frto_counter) {
2114 stored_cwnd = tp->snd_cwnd;
2116 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2117 tp->snd_cwnd = stored_cwnd;
2119 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2121 /* ... in theory, cong.control module could do "any tricks" in
2122 * ssthresh(), which means that ca_state, lost bits and lost_out
2123 * counter would have to be faked before the call occurs. We
2124 * consider that too expensive, unlikely and hacky, so modules
2125 * using these in ssthresh() must deal these incompatibility
2126 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2128 tcp_ca_event(sk, CA_EVENT_FRTO);
2131 tp->undo_marker = tp->snd_una;
2132 tp->undo_retrans = 0;
2134 skb = tcp_write_queue_head(sk);
2135 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2136 tp->undo_marker = 0;
2137 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2138 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2139 tp->retrans_out -= tcp_skb_pcount(skb);
2141 tcp_verify_left_out(tp);
2143 /* Too bad if TCP was application limited */
2144 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2146 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2147 * The last condition is necessary at least in tp->frto_counter case.
2149 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2150 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2151 after(tp->high_seq, tp->snd_una)) {
2152 tp->frto_highmark = tp->high_seq;
2154 tp->frto_highmark = tp->snd_nxt;
2156 tcp_set_ca_state(sk, TCP_CA_Disorder);
2157 tp->high_seq = tp->snd_nxt;
2158 tp->frto_counter = 1;
2161 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2162 * which indicates that we should follow the traditional RTO recovery,
2163 * i.e. mark everything lost and do go-back-N retransmission.
2165 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2167 struct tcp_sock *tp = tcp_sk(sk);
2168 struct sk_buff *skb;
2171 tp->retrans_out = 0;
2172 if (tcp_is_reno(tp))
2173 tcp_reset_reno_sack(tp);
2175 tcp_for_write_queue(skb, sk) {
2176 if (skb == tcp_send_head(sk))
2179 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2181 * Count the retransmission made on RTO correctly (only when
2182 * waiting for the first ACK and did not get it)...
2184 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2185 /* For some reason this R-bit might get cleared? */
2186 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2187 tp->retrans_out += tcp_skb_pcount(skb);
2188 /* ...enter this if branch just for the first segment */
2189 flag |= FLAG_DATA_ACKED;
2191 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2192 tp->undo_marker = 0;
2193 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2196 /* Marking forward transmissions that were made after RTO lost
2197 * can cause unnecessary retransmissions in some scenarios,
2198 * SACK blocks will mitigate that in some but not in all cases.
2199 * We used to not mark them but it was causing break-ups with
2200 * receivers that do only in-order receival.
2202 * TODO: we could detect presence of such receiver and select
2203 * different behavior per flow.
2205 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2206 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2207 tp->lost_out += tcp_skb_pcount(skb);
2208 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2211 tcp_verify_left_out(tp);
2213 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2214 tp->snd_cwnd_cnt = 0;
2215 tp->snd_cwnd_stamp = tcp_time_stamp;
2216 tp->frto_counter = 0;
2217 tp->bytes_acked = 0;
2219 tp->reordering = min_t(unsigned int, tp->reordering,
2220 sysctl_tcp_reordering);
2221 tcp_set_ca_state(sk, TCP_CA_Loss);
2222 tp->high_seq = tp->snd_nxt;
2223 TCP_ECN_queue_cwr(tp);
2225 tcp_clear_all_retrans_hints(tp);
2228 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2230 tp->retrans_out = 0;
2233 tp->undo_marker = 0;
2234 tp->undo_retrans = 0;
2237 void tcp_clear_retrans(struct tcp_sock *tp)
2239 tcp_clear_retrans_partial(tp);
2241 tp->fackets_out = 0;
2245 /* Enter Loss state. If "how" is not zero, forget all SACK information
2246 * and reset tags completely, otherwise preserve SACKs. If receiver
2247 * dropped its ofo queue, we will know this due to reneging detection.
2249 void tcp_enter_loss(struct sock *sk, int how)
2251 const struct inet_connection_sock *icsk = inet_csk(sk);
2252 struct tcp_sock *tp = tcp_sk(sk);
2253 struct sk_buff *skb;
2255 /* Reduce ssthresh if it has not yet been made inside this window. */
2256 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2257 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2258 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2259 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2260 tcp_ca_event(sk, CA_EVENT_LOSS);
2263 tp->snd_cwnd_cnt = 0;
2264 tp->snd_cwnd_stamp = tcp_time_stamp;
2266 tp->bytes_acked = 0;
2267 tcp_clear_retrans_partial(tp);
2269 if (tcp_is_reno(tp))
2270 tcp_reset_reno_sack(tp);
2272 tp->undo_marker = tp->snd_una;
2275 tp->fackets_out = 0;
2277 tcp_clear_all_retrans_hints(tp);
2279 tcp_for_write_queue(skb, sk) {
2280 if (skb == tcp_send_head(sk))
2283 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2284 tp->undo_marker = 0;
2285 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2286 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2287 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2288 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2289 tp->lost_out += tcp_skb_pcount(skb);
2290 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2293 tcp_verify_left_out(tp);
2295 tp->reordering = min_t(unsigned int, tp->reordering,
2296 sysctl_tcp_reordering);
2297 tcp_set_ca_state(sk, TCP_CA_Loss);
2298 tp->high_seq = tp->snd_nxt;
2299 TCP_ECN_queue_cwr(tp);
2300 /* Abort F-RTO algorithm if one is in progress */
2301 tp->frto_counter = 0;
2304 /* If ACK arrived pointing to a remembered SACK, it means that our
2305 * remembered SACKs do not reflect real state of receiver i.e.
2306 * receiver _host_ is heavily congested (or buggy).
2308 * Do processing similar to RTO timeout.
2310 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2312 if (flag & FLAG_SACK_RENEGING) {
2313 struct inet_connection_sock *icsk = inet_csk(sk);
2314 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2316 tcp_enter_loss(sk, 1);
2317 icsk->icsk_retransmits++;
2318 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2319 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2320 icsk->icsk_rto, TCP_RTO_MAX);
2326 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2328 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2331 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2332 * counter when SACK is enabled (without SACK, sacked_out is used for
2335 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2336 * segments up to the highest received SACK block so far and holes in
2339 * With reordering, holes may still be in flight, so RFC3517 recovery
2340 * uses pure sacked_out (total number of SACKed segments) even though
2341 * it violates the RFC that uses duplicate ACKs, often these are equal
2342 * but when e.g. out-of-window ACKs or packet duplication occurs,
2343 * they differ. Since neither occurs due to loss, TCP should really
2346 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2348 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2351 static inline int tcp_skb_timedout(const struct sock *sk,
2352 const struct sk_buff *skb)
2354 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2357 static inline int tcp_head_timedout(const struct sock *sk)
2359 const struct tcp_sock *tp = tcp_sk(sk);
2361 return tp->packets_out &&
2362 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2365 /* Linux NewReno/SACK/FACK/ECN state machine.
2366 * --------------------------------------
2368 * "Open" Normal state, no dubious events, fast path.
2369 * "Disorder" In all the respects it is "Open",
2370 * but requires a bit more attention. It is entered when
2371 * we see some SACKs or dupacks. It is split of "Open"
2372 * mainly to move some processing from fast path to slow one.
2373 * "CWR" CWND was reduced due to some Congestion Notification event.
2374 * It can be ECN, ICMP source quench, local device congestion.
2375 * "Recovery" CWND was reduced, we are fast-retransmitting.
2376 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2378 * tcp_fastretrans_alert() is entered:
2379 * - each incoming ACK, if state is not "Open"
2380 * - when arrived ACK is unusual, namely:
2385 * Counting packets in flight is pretty simple.
2387 * in_flight = packets_out - left_out + retrans_out
2389 * packets_out is SND.NXT-SND.UNA counted in packets.
2391 * retrans_out is number of retransmitted segments.
2393 * left_out is number of segments left network, but not ACKed yet.
2395 * left_out = sacked_out + lost_out
2397 * sacked_out: Packets, which arrived to receiver out of order
2398 * and hence not ACKed. With SACKs this number is simply
2399 * amount of SACKed data. Even without SACKs
2400 * it is easy to give pretty reliable estimate of this number,
2401 * counting duplicate ACKs.
2403 * lost_out: Packets lost by network. TCP has no explicit
2404 * "loss notification" feedback from network (for now).
2405 * It means that this number can be only _guessed_.
2406 * Actually, it is the heuristics to predict lossage that
2407 * distinguishes different algorithms.
2409 * F.e. after RTO, when all the queue is considered as lost,
2410 * lost_out = packets_out and in_flight = retrans_out.
2412 * Essentially, we have now two algorithms counting
2415 * FACK: It is the simplest heuristics. As soon as we decided
2416 * that something is lost, we decide that _all_ not SACKed
2417 * packets until the most forward SACK are lost. I.e.
2418 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2419 * It is absolutely correct estimate, if network does not reorder
2420 * packets. And it loses any connection to reality when reordering
2421 * takes place. We use FACK by default until reordering
2422 * is suspected on the path to this destination.
2424 * NewReno: when Recovery is entered, we assume that one segment
2425 * is lost (classic Reno). While we are in Recovery and
2426 * a partial ACK arrives, we assume that one more packet
2427 * is lost (NewReno). This heuristics are the same in NewReno
2430 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2431 * deflation etc. CWND is real congestion window, never inflated, changes
2432 * only according to classic VJ rules.
2434 * Really tricky (and requiring careful tuning) part of algorithm
2435 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2436 * The first determines the moment _when_ we should reduce CWND and,
2437 * hence, slow down forward transmission. In fact, it determines the moment
2438 * when we decide that hole is caused by loss, rather than by a reorder.
2440 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2441 * holes, caused by lost packets.
2443 * And the most logically complicated part of algorithm is undo
2444 * heuristics. We detect false retransmits due to both too early
2445 * fast retransmit (reordering) and underestimated RTO, analyzing
2446 * timestamps and D-SACKs. When we detect that some segments were
2447 * retransmitted by mistake and CWND reduction was wrong, we undo
2448 * window reduction and abort recovery phase. This logic is hidden
2449 * inside several functions named tcp_try_undo_<something>.
2452 /* This function decides, when we should leave Disordered state
2453 * and enter Recovery phase, reducing congestion window.
2455 * Main question: may we further continue forward transmission
2456 * with the same cwnd?
2458 static int tcp_time_to_recover(struct sock *sk)
2460 struct tcp_sock *tp = tcp_sk(sk);
2463 /* Do not perform any recovery during F-RTO algorithm */
2464 if (tp->frto_counter)
2467 /* Trick#1: The loss is proven. */
2471 /* Not-A-Trick#2 : Classic rule... */
2472 if (tcp_dupack_heuristics(tp) > tp->reordering)
2475 /* Trick#3 : when we use RFC2988 timer restart, fast
2476 * retransmit can be triggered by timeout of queue head.
2478 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2481 /* Trick#4: It is still not OK... But will it be useful to delay
2484 packets_out = tp->packets_out;
2485 if (packets_out <= tp->reordering &&
2486 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2487 !tcp_may_send_now(sk)) {
2488 /* We have nothing to send. This connection is limited
2489 * either by receiver window or by application.
2494 /* If a thin stream is detected, retransmit after first
2495 * received dupack. Employ only if SACK is supported in order
2496 * to avoid possible corner-case series of spurious retransmissions
2497 * Use only if there are no unsent data.
2499 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2500 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2501 tcp_is_sack(tp) && !tcp_send_head(sk))
2507 /* New heuristics: it is possible only after we switched to restart timer
2508 * each time when something is ACKed. Hence, we can detect timed out packets
2509 * during fast retransmit without falling to slow start.
2511 * Usefulness of this as is very questionable, since we should know which of
2512 * the segments is the next to timeout which is relatively expensive to find
2513 * in general case unless we add some data structure just for that. The
2514 * current approach certainly won't find the right one too often and when it
2515 * finally does find _something_ it usually marks large part of the window
2516 * right away (because a retransmission with a larger timestamp blocks the
2517 * loop from advancing). -ij
2519 static void tcp_timeout_skbs(struct sock *sk)
2521 struct tcp_sock *tp = tcp_sk(sk);
2522 struct sk_buff *skb;
2524 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2527 skb = tp->scoreboard_skb_hint;
2528 if (tp->scoreboard_skb_hint == NULL)
2529 skb = tcp_write_queue_head(sk);
2531 tcp_for_write_queue_from(skb, sk) {
2532 if (skb == tcp_send_head(sk))
2534 if (!tcp_skb_timedout(sk, skb))
2537 tcp_skb_mark_lost(tp, skb);
2540 tp->scoreboard_skb_hint = skb;
2542 tcp_verify_left_out(tp);
2545 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2546 * is against sacked "cnt", otherwise it's against facked "cnt"
2548 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2550 struct tcp_sock *tp = tcp_sk(sk);
2551 struct sk_buff *skb;
2556 WARN_ON(packets > tp->packets_out);
2557 if (tp->lost_skb_hint) {
2558 skb = tp->lost_skb_hint;
2559 cnt = tp->lost_cnt_hint;
2560 /* Head already handled? */
2561 if (mark_head && skb != tcp_write_queue_head(sk))
2564 skb = tcp_write_queue_head(sk);
2568 tcp_for_write_queue_from(skb, sk) {
2569 if (skb == tcp_send_head(sk))
2571 /* TODO: do this better */
2572 /* this is not the most efficient way to do this... */
2573 tp->lost_skb_hint = skb;
2574 tp->lost_cnt_hint = cnt;
2576 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2580 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2581 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2582 cnt += tcp_skb_pcount(skb);
2584 if (cnt > packets) {
2585 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2586 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2587 (oldcnt >= packets))
2590 mss = skb_shinfo(skb)->gso_size;
2591 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2597 tcp_skb_mark_lost(tp, skb);
2602 tcp_verify_left_out(tp);
2605 /* Account newly detected lost packet(s) */
2607 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2609 struct tcp_sock *tp = tcp_sk(sk);
2611 if (tcp_is_reno(tp)) {
2612 tcp_mark_head_lost(sk, 1, 1);
2613 } else if (tcp_is_fack(tp)) {
2614 int lost = tp->fackets_out - tp->reordering;
2617 tcp_mark_head_lost(sk, lost, 0);
2619 int sacked_upto = tp->sacked_out - tp->reordering;
2620 if (sacked_upto >= 0)
2621 tcp_mark_head_lost(sk, sacked_upto, 0);
2622 else if (fast_rexmit)
2623 tcp_mark_head_lost(sk, 1, 1);
2626 tcp_timeout_skbs(sk);
2629 /* CWND moderation, preventing bursts due to too big ACKs
2630 * in dubious situations.
2632 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2634 tp->snd_cwnd = min(tp->snd_cwnd,
2635 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2636 tp->snd_cwnd_stamp = tcp_time_stamp;
2639 /* Lower bound on congestion window is slow start threshold
2640 * unless congestion avoidance choice decides to overide it.
2642 static inline u32 tcp_cwnd_min(const struct sock *sk)
2644 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2646 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2649 /* Decrease cwnd each second ack. */
2650 static void tcp_cwnd_down(struct sock *sk, int flag)
2652 struct tcp_sock *tp = tcp_sk(sk);
2653 int decr = tp->snd_cwnd_cnt + 1;
2655 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2656 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2657 tp->snd_cwnd_cnt = decr & 1;
2660 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2661 tp->snd_cwnd -= decr;
2663 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2664 tp->snd_cwnd_stamp = tcp_time_stamp;
2668 /* Nothing was retransmitted or returned timestamp is less
2669 * than timestamp of the first retransmission.
2671 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2673 return !tp->retrans_stamp ||
2674 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2675 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2678 /* Undo procedures. */
2680 #if FASTRETRANS_DEBUG > 1
2681 static void DBGUNDO(struct sock *sk, const char *msg)
2683 struct tcp_sock *tp = tcp_sk(sk);
2684 struct inet_sock *inet = inet_sk(sk);
2686 if (sk->sk_family == AF_INET) {
2687 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2689 &inet->inet_daddr, ntohs(inet->inet_dport),
2690 tp->snd_cwnd, tcp_left_out(tp),
2691 tp->snd_ssthresh, tp->prior_ssthresh,
2694 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2695 else if (sk->sk_family == AF_INET6) {
2696 struct ipv6_pinfo *np = inet6_sk(sk);
2697 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2699 &np->daddr, ntohs(inet->inet_dport),
2700 tp->snd_cwnd, tcp_left_out(tp),
2701 tp->snd_ssthresh, tp->prior_ssthresh,
2707 #define DBGUNDO(x...) do { } while (0)
2710 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2712 struct tcp_sock *tp = tcp_sk(sk);
2714 if (tp->prior_ssthresh) {
2715 const struct inet_connection_sock *icsk = inet_csk(sk);
2717 if (icsk->icsk_ca_ops->undo_cwnd)
2718 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2720 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2722 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2723 tp->snd_ssthresh = tp->prior_ssthresh;
2724 TCP_ECN_withdraw_cwr(tp);
2727 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2729 tp->snd_cwnd_stamp = tcp_time_stamp;
2732 static inline int tcp_may_undo(const struct tcp_sock *tp)
2734 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2737 /* People celebrate: "We love our President!" */
2738 static int tcp_try_undo_recovery(struct sock *sk)
2740 struct tcp_sock *tp = tcp_sk(sk);
2742 if (tcp_may_undo(tp)) {
2745 /* Happy end! We did not retransmit anything
2746 * or our original transmission succeeded.
2748 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2749 tcp_undo_cwr(sk, true);
2750 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2751 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2753 mib_idx = LINUX_MIB_TCPFULLUNDO;
2755 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2756 tp->undo_marker = 0;
2758 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2759 /* Hold old state until something *above* high_seq
2760 * is ACKed. For Reno it is MUST to prevent false
2761 * fast retransmits (RFC2582). SACK TCP is safe. */
2762 tcp_moderate_cwnd(tp);
2765 tcp_set_ca_state(sk, TCP_CA_Open);
2769 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2770 static void tcp_try_undo_dsack(struct sock *sk)
2772 struct tcp_sock *tp = tcp_sk(sk);
2774 if (tp->undo_marker && !tp->undo_retrans) {
2775 DBGUNDO(sk, "D-SACK");
2776 tcp_undo_cwr(sk, true);
2777 tp->undo_marker = 0;
2778 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2782 /* We can clear retrans_stamp when there are no retransmissions in the
2783 * window. It would seem that it is trivially available for us in
2784 * tp->retrans_out, however, that kind of assumptions doesn't consider
2785 * what will happen if errors occur when sending retransmission for the
2786 * second time. ...It could the that such segment has only
2787 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2788 * the head skb is enough except for some reneging corner cases that
2789 * are not worth the effort.
2791 * Main reason for all this complexity is the fact that connection dying
2792 * time now depends on the validity of the retrans_stamp, in particular,
2793 * that successive retransmissions of a segment must not advance
2794 * retrans_stamp under any conditions.
2796 static int tcp_any_retrans_done(const struct sock *sk)
2798 const struct tcp_sock *tp = tcp_sk(sk);
2799 struct sk_buff *skb;
2801 if (tp->retrans_out)
2804 skb = tcp_write_queue_head(sk);
2805 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2811 /* Undo during fast recovery after partial ACK. */
2813 static int tcp_try_undo_partial(struct sock *sk, int acked)
2815 struct tcp_sock *tp = tcp_sk(sk);
2816 /* Partial ACK arrived. Force Hoe's retransmit. */
2817 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2819 if (tcp_may_undo(tp)) {
2820 /* Plain luck! Hole if filled with delayed
2821 * packet, rather than with a retransmit.
2823 if (!tcp_any_retrans_done(sk))
2824 tp->retrans_stamp = 0;
2826 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2829 tcp_undo_cwr(sk, false);
2830 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2832 /* So... Do not make Hoe's retransmit yet.
2833 * If the first packet was delayed, the rest
2834 * ones are most probably delayed as well.
2841 /* Undo during loss recovery after partial ACK. */
2842 static int tcp_try_undo_loss(struct sock *sk)
2844 struct tcp_sock *tp = tcp_sk(sk);
2846 if (tcp_may_undo(tp)) {
2847 struct sk_buff *skb;
2848 tcp_for_write_queue(skb, sk) {
2849 if (skb == tcp_send_head(sk))
2851 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2854 tcp_clear_all_retrans_hints(tp);
2856 DBGUNDO(sk, "partial loss");
2858 tcp_undo_cwr(sk, true);
2859 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2860 inet_csk(sk)->icsk_retransmits = 0;
2861 tp->undo_marker = 0;
2862 if (tcp_is_sack(tp))
2863 tcp_set_ca_state(sk, TCP_CA_Open);
2869 static inline void tcp_complete_cwr(struct sock *sk)
2871 struct tcp_sock *tp = tcp_sk(sk);
2873 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2874 if (tp->undo_marker) {
2875 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
2876 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2877 tp->snd_cwnd_stamp = tcp_time_stamp;
2878 } else if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH) {
2879 /* PRR algorithm. */
2880 tp->snd_cwnd = tp->snd_ssthresh;
2881 tp->snd_cwnd_stamp = tcp_time_stamp;
2884 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2887 static void tcp_try_keep_open(struct sock *sk)
2889 struct tcp_sock *tp = tcp_sk(sk);
2890 int state = TCP_CA_Open;
2892 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2893 state = TCP_CA_Disorder;
2895 if (inet_csk(sk)->icsk_ca_state != state) {
2896 tcp_set_ca_state(sk, state);
2897 tp->high_seq = tp->snd_nxt;
2901 static void tcp_try_to_open(struct sock *sk, int flag)
2903 struct tcp_sock *tp = tcp_sk(sk);
2905 tcp_verify_left_out(tp);
2907 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2908 tp->retrans_stamp = 0;
2910 if (flag & FLAG_ECE)
2911 tcp_enter_cwr(sk, 1);
2913 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2914 tcp_try_keep_open(sk);
2915 tcp_moderate_cwnd(tp);
2917 tcp_cwnd_down(sk, flag);
2921 static void tcp_mtup_probe_failed(struct sock *sk)
2923 struct inet_connection_sock *icsk = inet_csk(sk);
2925 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2926 icsk->icsk_mtup.probe_size = 0;
2929 static void tcp_mtup_probe_success(struct sock *sk)
2931 struct tcp_sock *tp = tcp_sk(sk);
2932 struct inet_connection_sock *icsk = inet_csk(sk);
2934 /* FIXME: breaks with very large cwnd */
2935 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2936 tp->snd_cwnd = tp->snd_cwnd *
2937 tcp_mss_to_mtu(sk, tp->mss_cache) /
2938 icsk->icsk_mtup.probe_size;
2939 tp->snd_cwnd_cnt = 0;
2940 tp->snd_cwnd_stamp = tcp_time_stamp;
2941 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2943 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2944 icsk->icsk_mtup.probe_size = 0;
2945 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2948 /* Do a simple retransmit without using the backoff mechanisms in
2949 * tcp_timer. This is used for path mtu discovery.
2950 * The socket is already locked here.
2952 void tcp_simple_retransmit(struct sock *sk)
2954 const struct inet_connection_sock *icsk = inet_csk(sk);
2955 struct tcp_sock *tp = tcp_sk(sk);
2956 struct sk_buff *skb;
2957 unsigned int mss = tcp_current_mss(sk);
2958 u32 prior_lost = tp->lost_out;
2960 tcp_for_write_queue(skb, sk) {
2961 if (skb == tcp_send_head(sk))
2963 if (tcp_skb_seglen(skb) > mss &&
2964 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2965 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2966 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2967 tp->retrans_out -= tcp_skb_pcount(skb);
2969 tcp_skb_mark_lost_uncond_verify(tp, skb);
2973 tcp_clear_retrans_hints_partial(tp);
2975 if (prior_lost == tp->lost_out)
2978 if (tcp_is_reno(tp))
2979 tcp_limit_reno_sacked(tp);
2981 tcp_verify_left_out(tp);
2983 /* Don't muck with the congestion window here.
2984 * Reason is that we do not increase amount of _data_
2985 * in network, but units changed and effective
2986 * cwnd/ssthresh really reduced now.
2988 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2989 tp->high_seq = tp->snd_nxt;
2990 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2991 tp->prior_ssthresh = 0;
2992 tp->undo_marker = 0;
2993 tcp_set_ca_state(sk, TCP_CA_Loss);
2995 tcp_xmit_retransmit_queue(sk);
2997 EXPORT_SYMBOL(tcp_simple_retransmit);
2999 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
3000 * (proportional rate reduction with slow start reduction bound) as described in
3001 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
3002 * It computes the number of packets to send (sndcnt) based on packets newly
3004 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
3005 * cwnd reductions across a full RTT.
3006 * 2) If packets in flight is lower than ssthresh (such as due to excess
3007 * losses and/or application stalls), do not perform any further cwnd
3008 * reductions, but instead slow start up to ssthresh.
3010 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
3011 int fast_rexmit, int flag)
3013 struct tcp_sock *tp = tcp_sk(sk);
3015 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
3017 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
3018 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
3020 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
3022 sndcnt = min_t(int, delta,
3023 max_t(int, tp->prr_delivered - tp->prr_out,
3024 newly_acked_sacked) + 1);
3027 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
3028 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3031 /* Process an event, which can update packets-in-flight not trivially.
3032 * Main goal of this function is to calculate new estimate for left_out,
3033 * taking into account both packets sitting in receiver's buffer and
3034 * packets lost by network.
3036 * Besides that it does CWND reduction, when packet loss is detected
3037 * and changes state of machine.
3039 * It does _not_ decide what to send, it is made in function
3040 * tcp_xmit_retransmit_queue().
3042 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3043 int newly_acked_sacked, int flag)
3045 struct inet_connection_sock *icsk = inet_csk(sk);
3046 struct tcp_sock *tp = tcp_sk(sk);
3047 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3048 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3049 (tcp_fackets_out(tp) > tp->reordering));
3050 int fast_rexmit = 0, mib_idx;
3052 if (WARN_ON(!tp->packets_out && tp->sacked_out))
3054 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3055 tp->fackets_out = 0;
3057 /* Now state machine starts.
3058 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3059 if (flag & FLAG_ECE)
3060 tp->prior_ssthresh = 0;
3062 /* B. In all the states check for reneging SACKs. */
3063 if (tcp_check_sack_reneging(sk, flag))
3066 /* C. Process data loss notification, provided it is valid. */
3067 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
3068 before(tp->snd_una, tp->high_seq) &&
3069 icsk->icsk_ca_state != TCP_CA_Open &&
3070 tp->fackets_out > tp->reordering) {
3071 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
3072 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
3075 /* D. Check consistency of the current state. */
3076 tcp_verify_left_out(tp);
3078 /* E. Check state exit conditions. State can be terminated
3079 * when high_seq is ACKed. */
3080 if (icsk->icsk_ca_state == TCP_CA_Open) {
3081 WARN_ON(tp->retrans_out != 0);
3082 tp->retrans_stamp = 0;
3083 } else if (!before(tp->snd_una, tp->high_seq)) {
3084 switch (icsk->icsk_ca_state) {
3086 icsk->icsk_retransmits = 0;
3087 if (tcp_try_undo_recovery(sk))
3092 /* CWR is to be held something *above* high_seq
3093 * is ACKed for CWR bit to reach receiver. */
3094 if (tp->snd_una != tp->high_seq) {
3095 tcp_complete_cwr(sk);
3096 tcp_set_ca_state(sk, TCP_CA_Open);
3100 case TCP_CA_Disorder:
3101 tcp_try_undo_dsack(sk);
3102 if (!tp->undo_marker ||
3103 /* For SACK case do not Open to allow to undo
3104 * catching for all duplicate ACKs. */
3105 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3106 tp->undo_marker = 0;
3107 tcp_set_ca_state(sk, TCP_CA_Open);
3111 case TCP_CA_Recovery:
3112 if (tcp_is_reno(tp))
3113 tcp_reset_reno_sack(tp);
3114 if (tcp_try_undo_recovery(sk))
3116 tcp_complete_cwr(sk);
3121 /* F. Process state. */
3122 switch (icsk->icsk_ca_state) {
3123 case TCP_CA_Recovery:
3124 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3125 if (tcp_is_reno(tp) && is_dupack)
3126 tcp_add_reno_sack(sk);
3128 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3131 if (flag & FLAG_DATA_ACKED)
3132 icsk->icsk_retransmits = 0;
3133 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3134 tcp_reset_reno_sack(tp);
3135 if (!tcp_try_undo_loss(sk)) {
3136 tcp_moderate_cwnd(tp);
3137 tcp_xmit_retransmit_queue(sk);
3140 if (icsk->icsk_ca_state != TCP_CA_Open)
3142 /* Loss is undone; fall through to processing in Open state. */
3144 if (tcp_is_reno(tp)) {
3145 if (flag & FLAG_SND_UNA_ADVANCED)
3146 tcp_reset_reno_sack(tp);
3148 tcp_add_reno_sack(sk);
3151 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3152 tcp_try_undo_dsack(sk);
3154 if (!tcp_time_to_recover(sk)) {
3155 tcp_try_to_open(sk, flag);
3159 /* MTU probe failure: don't reduce cwnd */
3160 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3161 icsk->icsk_mtup.probe_size &&
3162 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3163 tcp_mtup_probe_failed(sk);
3164 /* Restores the reduction we did in tcp_mtup_probe() */
3166 tcp_simple_retransmit(sk);
3170 /* Otherwise enter Recovery state */
3172 if (tcp_is_reno(tp))
3173 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3175 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3177 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3179 tp->high_seq = tp->snd_nxt;
3180 tp->prior_ssthresh = 0;
3181 tp->undo_marker = tp->snd_una;
3182 tp->undo_retrans = tp->retrans_out;
3184 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3185 if (!(flag & FLAG_ECE))
3186 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3187 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3188 TCP_ECN_queue_cwr(tp);
3191 tp->bytes_acked = 0;
3192 tp->snd_cwnd_cnt = 0;
3193 tp->prior_cwnd = tp->snd_cwnd;
3194 tp->prr_delivered = 0;
3196 tcp_set_ca_state(sk, TCP_CA_Recovery);
3200 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3201 tcp_update_scoreboard(sk, fast_rexmit);
3202 tp->prr_delivered += newly_acked_sacked;
3203 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3204 tcp_xmit_retransmit_queue(sk);
3207 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3209 tcp_rtt_estimator(sk, seq_rtt);
3211 inet_csk(sk)->icsk_backoff = 0;
3213 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3215 /* Read draft-ietf-tcplw-high-performance before mucking
3216 * with this code. (Supersedes RFC1323)
3218 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3220 /* RTTM Rule: A TSecr value received in a segment is used to
3221 * update the averaged RTT measurement only if the segment
3222 * acknowledges some new data, i.e., only if it advances the
3223 * left edge of the send window.
3225 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3226 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3228 * Changed: reset backoff as soon as we see the first valid sample.
3229 * If we do not, we get strongly overestimated rto. With timestamps
3230 * samples are accepted even from very old segments: f.e., when rtt=1
3231 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3232 * answer arrives rto becomes 120 seconds! If at least one of segments
3233 * in window is lost... Voila. --ANK (010210)
3235 struct tcp_sock *tp = tcp_sk(sk);
3237 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3240 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3242 /* We don't have a timestamp. Can only use
3243 * packets that are not retransmitted to determine
3244 * rtt estimates. Also, we must not reset the
3245 * backoff for rto until we get a non-retransmitted
3246 * packet. This allows us to deal with a situation
3247 * where the network delay has increased suddenly.
3248 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3251 if (flag & FLAG_RETRANS_DATA_ACKED)
3254 tcp_valid_rtt_meas(sk, seq_rtt);
3257 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3260 const struct tcp_sock *tp = tcp_sk(sk);
3261 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3262 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3263 tcp_ack_saw_tstamp(sk, flag);
3264 else if (seq_rtt >= 0)
3265 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3268 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3270 const struct inet_connection_sock *icsk = inet_csk(sk);
3271 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3272 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3275 /* Restart timer after forward progress on connection.
3276 * RFC2988 recommends to restart timer to now+rto.
3278 static void tcp_rearm_rto(struct sock *sk)
3280 const struct tcp_sock *tp = tcp_sk(sk);
3282 if (!tp->packets_out) {
3283 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3285 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3286 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3290 /* If we get here, the whole TSO packet has not been acked. */
3291 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3293 struct tcp_sock *tp = tcp_sk(sk);
3296 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3298 packets_acked = tcp_skb_pcount(skb);
3299 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3301 packets_acked -= tcp_skb_pcount(skb);
3303 if (packets_acked) {
3304 BUG_ON(tcp_skb_pcount(skb) == 0);
3305 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3308 return packets_acked;
3311 /* Remove acknowledged frames from the retransmission queue. If our packet
3312 * is before the ack sequence we can discard it as it's confirmed to have
3313 * arrived at the other end.
3315 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3318 struct tcp_sock *tp = tcp_sk(sk);
3319 const struct inet_connection_sock *icsk = inet_csk(sk);
3320 struct sk_buff *skb;
3321 u32 now = tcp_time_stamp;
3322 int fully_acked = 1;
3325 u32 reord = tp->packets_out;
3326 u32 prior_sacked = tp->sacked_out;
3328 s32 ca_seq_rtt = -1;
3329 ktime_t last_ackt = net_invalid_timestamp();
3331 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3332 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3334 u8 sacked = scb->sacked;
3336 /* Determine how many packets and what bytes were acked, tso and else */
3337 if (after(scb->end_seq, tp->snd_una)) {
3338 if (tcp_skb_pcount(skb) == 1 ||
3339 !after(tp->snd_una, scb->seq))
3342 acked_pcount = tcp_tso_acked(sk, skb);
3348 acked_pcount = tcp_skb_pcount(skb);
3351 if (sacked & TCPCB_RETRANS) {
3352 if (sacked & TCPCB_SACKED_RETRANS)
3353 tp->retrans_out -= acked_pcount;
3354 flag |= FLAG_RETRANS_DATA_ACKED;
3357 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3358 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3360 ca_seq_rtt = now - scb->when;
3361 last_ackt = skb->tstamp;
3363 seq_rtt = ca_seq_rtt;
3365 if (!(sacked & TCPCB_SACKED_ACKED))
3366 reord = min(pkts_acked, reord);
3369 if (sacked & TCPCB_SACKED_ACKED)
3370 tp->sacked_out -= acked_pcount;
3371 if (sacked & TCPCB_LOST)
3372 tp->lost_out -= acked_pcount;
3374 tp->packets_out -= acked_pcount;
3375 pkts_acked += acked_pcount;
3377 /* Initial outgoing SYN's get put onto the write_queue
3378 * just like anything else we transmit. It is not
3379 * true data, and if we misinform our callers that
3380 * this ACK acks real data, we will erroneously exit
3381 * connection startup slow start one packet too
3382 * quickly. This is severely frowned upon behavior.
3384 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3385 flag |= FLAG_DATA_ACKED;
3387 flag |= FLAG_SYN_ACKED;
3388 tp->retrans_stamp = 0;
3394 tcp_unlink_write_queue(skb, sk);
3395 sk_wmem_free_skb(sk, skb);
3396 tp->scoreboard_skb_hint = NULL;
3397 if (skb == tp->retransmit_skb_hint)
3398 tp->retransmit_skb_hint = NULL;
3399 if (skb == tp->lost_skb_hint)
3400 tp->lost_skb_hint = NULL;
3403 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3404 tp->snd_up = tp->snd_una;
3406 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3407 flag |= FLAG_SACK_RENEGING;
3409 if (flag & FLAG_ACKED) {
3410 const struct tcp_congestion_ops *ca_ops
3411 = inet_csk(sk)->icsk_ca_ops;
3413 if (unlikely(icsk->icsk_mtup.probe_size &&
3414 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3415 tcp_mtup_probe_success(sk);
3418 tcp_ack_update_rtt(sk, flag, seq_rtt);
3421 if (tcp_is_reno(tp)) {
3422 tcp_remove_reno_sacks(sk, pkts_acked);
3426 /* Non-retransmitted hole got filled? That's reordering */
3427 if (reord < prior_fackets)
3428 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3430 delta = tcp_is_fack(tp) ? pkts_acked :
3431 prior_sacked - tp->sacked_out;
3432 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3435 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3437 if (ca_ops->pkts_acked) {
3440 /* Is the ACK triggering packet unambiguous? */
3441 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3442 /* High resolution needed and available? */
3443 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3444 !ktime_equal(last_ackt,
3445 net_invalid_timestamp()))
3446 rtt_us = ktime_us_delta(ktime_get_real(),
3448 else if (ca_seq_rtt >= 0)
3449 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3452 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3456 #if FASTRETRANS_DEBUG > 0
3457 WARN_ON((int)tp->sacked_out < 0);
3458 WARN_ON((int)tp->lost_out < 0);
3459 WARN_ON((int)tp->retrans_out < 0);
3460 if (!tp->packets_out && tcp_is_sack(tp)) {
3461 icsk = inet_csk(sk);
3463 printk(KERN_DEBUG "Leak l=%u %d\n",
3464 tp->lost_out, icsk->icsk_ca_state);
3467 if (tp->sacked_out) {
3468 printk(KERN_DEBUG "Leak s=%u %d\n",
3469 tp->sacked_out, icsk->icsk_ca_state);
3472 if (tp->retrans_out) {
3473 printk(KERN_DEBUG "Leak r=%u %d\n",
3474 tp->retrans_out, icsk->icsk_ca_state);
3475 tp->retrans_out = 0;
3482 static void tcp_ack_probe(struct sock *sk)
3484 const struct tcp_sock *tp = tcp_sk(sk);
3485 struct inet_connection_sock *icsk = inet_csk(sk);
3487 /* Was it a usable window open? */
3489 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3490 icsk->icsk_backoff = 0;
3491 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3492 /* Socket must be waked up by subsequent tcp_data_snd_check().
3493 * This function is not for random using!
3496 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3497 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3502 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3504 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3505 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3508 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3510 const struct tcp_sock *tp = tcp_sk(sk);
3511 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3512 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3515 /* Check that window update is acceptable.
3516 * The function assumes that snd_una<=ack<=snd_next.
3518 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3519 const u32 ack, const u32 ack_seq,
3522 return after(ack, tp->snd_una) ||
3523 after(ack_seq, tp->snd_wl1) ||
3524 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3527 /* Update our send window.
3529 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3530 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3532 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3535 struct tcp_sock *tp = tcp_sk(sk);
3537 u32 nwin = ntohs(tcp_hdr(skb)->window);
3539 if (likely(!tcp_hdr(skb)->syn))
3540 nwin <<= tp->rx_opt.snd_wscale;
3542 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3543 flag |= FLAG_WIN_UPDATE;
3544 tcp_update_wl(tp, ack_seq);
3546 if (tp->snd_wnd != nwin) {
3549 /* Note, it is the only place, where
3550 * fast path is recovered for sending TCP.
3553 tcp_fast_path_check(sk);
3555 if (nwin > tp->max_window) {
3556 tp->max_window = nwin;
3557 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3567 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3568 * continue in congestion avoidance.
3570 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3572 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3573 tp->snd_cwnd_cnt = 0;
3574 tp->bytes_acked = 0;
3575 TCP_ECN_queue_cwr(tp);
3576 tcp_moderate_cwnd(tp);
3579 /* A conservative spurious RTO response algorithm: reduce cwnd using
3580 * rate halving and continue in congestion avoidance.
3582 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3584 tcp_enter_cwr(sk, 0);
3587 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3589 if (flag & FLAG_ECE)
3590 tcp_ratehalving_spur_to_response(sk);
3592 tcp_undo_cwr(sk, true);
3595 /* F-RTO spurious RTO detection algorithm (RFC4138)
3597 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3598 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3599 * window (but not to or beyond highest sequence sent before RTO):
3600 * On First ACK, send two new segments out.
3601 * On Second ACK, RTO was likely spurious. Do spurious response (response
3602 * algorithm is not part of the F-RTO detection algorithm
3603 * given in RFC4138 but can be selected separately).
3604 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3605 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3606 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3607 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3609 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3610 * original window even after we transmit two new data segments.
3613 * on first step, wait until first cumulative ACK arrives, then move to
3614 * the second step. In second step, the next ACK decides.
3616 * F-RTO is implemented (mainly) in four functions:
3617 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3618 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3619 * called when tcp_use_frto() showed green light
3620 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3621 * - tcp_enter_frto_loss() is called if there is not enough evidence
3622 * to prove that the RTO is indeed spurious. It transfers the control
3623 * from F-RTO to the conventional RTO recovery
3625 static int tcp_process_frto(struct sock *sk, int flag)
3627 struct tcp_sock *tp = tcp_sk(sk);
3629 tcp_verify_left_out(tp);
3631 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3632 if (flag & FLAG_DATA_ACKED)
3633 inet_csk(sk)->icsk_retransmits = 0;
3635 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3636 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3637 tp->undo_marker = 0;
3639 if (!before(tp->snd_una, tp->frto_highmark)) {
3640 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3644 if (!tcp_is_sackfrto(tp)) {
3645 /* RFC4138 shortcoming in step 2; should also have case c):
3646 * ACK isn't duplicate nor advances window, e.g., opposite dir
3649 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3652 if (!(flag & FLAG_DATA_ACKED)) {
3653 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3658 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3659 if (!tcp_packets_in_flight(tp)) {
3660 tcp_enter_frto_loss(sk, 2, flag);
3664 /* Prevent sending of new data. */
3665 tp->snd_cwnd = min(tp->snd_cwnd,
3666 tcp_packets_in_flight(tp));
3670 if ((tp->frto_counter >= 2) &&
3671 (!(flag & FLAG_FORWARD_PROGRESS) ||
3672 ((flag & FLAG_DATA_SACKED) &&
3673 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3674 /* RFC4138 shortcoming (see comment above) */
3675 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3676 (flag & FLAG_NOT_DUP))
3679 tcp_enter_frto_loss(sk, 3, flag);
3684 if (tp->frto_counter == 1) {
3685 /* tcp_may_send_now needs to see updated state */
3686 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3687 tp->frto_counter = 2;
3689 if (!tcp_may_send_now(sk))
3690 tcp_enter_frto_loss(sk, 2, flag);
3694 switch (sysctl_tcp_frto_response) {
3696 tcp_undo_spur_to_response(sk, flag);
3699 tcp_conservative_spur_to_response(tp);
3702 tcp_ratehalving_spur_to_response(sk);
3705 tp->frto_counter = 0;
3706 tp->undo_marker = 0;
3707 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3712 /* RFC 5961 7 [ACK Throttling] */
3713 static void tcp_send_challenge_ack(struct sock *sk)
3715 /* unprotected vars, we dont care of overwrites */
3716 static u32 challenge_timestamp;
3717 static unsigned int challenge_count;
3718 u32 count, now = jiffies / HZ;
3720 if (now != challenge_timestamp) {
3721 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3723 challenge_timestamp = now;
3724 ACCESS_ONCE(challenge_count) =
3726 ((u64) random32() * sysctl_tcp_challenge_ack_limit)
3729 count = ACCESS_ONCE(challenge_count);
3731 ACCESS_ONCE(challenge_count) = count - 1;
3732 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3737 static void tcp_store_ts_recent(struct tcp_sock *tp)
3739 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3740 tp->rx_opt.ts_recent_stamp = get_seconds();
3743 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3745 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3746 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3747 * extra check below makes sure this can only happen
3748 * for pure ACK frames. -DaveM
3750 * Not only, also it occurs for expired timestamps.
3753 if (tcp_paws_check(&tp->rx_opt, 0))
3754 tcp_store_ts_recent(tp);
3758 /* This routine deals with incoming acks, but not outgoing ones. */
3759 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3761 struct inet_connection_sock *icsk = inet_csk(sk);
3762 struct tcp_sock *tp = tcp_sk(sk);
3763 u32 prior_snd_una = tp->snd_una;
3764 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3765 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3766 u32 prior_in_flight;
3769 int prior_sacked = tp->sacked_out;
3770 int newly_acked_sacked = 0;
3773 /* If the ack is older than previous acks
3774 * then we can probably ignore it.
3776 if (before(ack, prior_snd_una)) {
3777 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3778 if (before(ack, prior_snd_una - tp->max_window)) {
3779 tcp_send_challenge_ack(sk);
3785 /* If the ack includes data we haven't sent yet, discard
3786 * this segment (RFC793 Section 3.9).
3788 if (after(ack, tp->snd_nxt))
3791 if (after(ack, prior_snd_una))
3792 flag |= FLAG_SND_UNA_ADVANCED;
3794 if (sysctl_tcp_abc) {
3795 if (icsk->icsk_ca_state < TCP_CA_CWR)
3796 tp->bytes_acked += ack - prior_snd_una;
3797 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3798 /* we assume just one segment left network */
3799 tp->bytes_acked += min(ack - prior_snd_una,
3803 prior_fackets = tp->fackets_out;
3804 prior_in_flight = tcp_packets_in_flight(tp);
3806 /* ts_recent update must be made after we are sure that the packet
3809 if (flag & FLAG_UPDATE_TS_RECENT)
3810 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3812 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3813 /* Window is constant, pure forward advance.
3814 * No more checks are required.
3815 * Note, we use the fact that SND.UNA>=SND.WL2.
3817 tcp_update_wl(tp, ack_seq);
3819 flag |= FLAG_WIN_UPDATE;
3821 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3823 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3825 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3828 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3830 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3832 if (TCP_SKB_CB(skb)->sacked)
3833 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3835 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3838 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3841 /* We passed data and got it acked, remove any soft error
3842 * log. Something worked...
3844 sk->sk_err_soft = 0;
3845 icsk->icsk_probes_out = 0;
3846 tp->rcv_tstamp = tcp_time_stamp;
3847 prior_packets = tp->packets_out;
3851 /* See if we can take anything off of the retransmit queue. */
3852 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3854 newly_acked_sacked = (prior_packets - prior_sacked) -
3855 (tp->packets_out - tp->sacked_out);
3857 if (tp->frto_counter)
3858 frto_cwnd = tcp_process_frto(sk, flag);
3859 /* Guarantee sacktag reordering detection against wrap-arounds */
3860 if (before(tp->frto_highmark, tp->snd_una))
3861 tp->frto_highmark = 0;
3863 if (tcp_ack_is_dubious(sk, flag)) {
3864 /* Advance CWND, if state allows this. */
3865 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3866 tcp_may_raise_cwnd(sk, flag))
3867 tcp_cong_avoid(sk, ack, prior_in_flight);
3868 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3869 newly_acked_sacked, flag);
3871 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3872 tcp_cong_avoid(sk, ack, prior_in_flight);
3875 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3876 dst_confirm(__sk_dst_get(sk));
3881 /* If this ack opens up a zero window, clear backoff. It was
3882 * being used to time the probes, and is probably far higher than
3883 * it needs to be for normal retransmission.
3885 if (tcp_send_head(sk))
3890 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3894 if (TCP_SKB_CB(skb)->sacked) {
3895 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3896 if (icsk->icsk_ca_state == TCP_CA_Open)
3897 tcp_try_keep_open(sk);
3900 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3904 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3905 * But, this can also be called on packets in the established flow when
3906 * the fast version below fails.
3908 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3909 const u8 **hvpp, int estab)
3911 const unsigned char *ptr;
3912 const struct tcphdr *th = tcp_hdr(skb);
3913 int length = (th->doff * 4) - sizeof(struct tcphdr);
3915 ptr = (const unsigned char *)(th + 1);
3916 opt_rx->saw_tstamp = 0;
3918 while (length > 0) {
3919 int opcode = *ptr++;
3925 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3930 if (opsize < 2) /* "silly options" */
3932 if (opsize > length)
3933 return; /* don't parse partial options */
3936 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3937 u16 in_mss = get_unaligned_be16(ptr);
3939 if (opt_rx->user_mss &&
3940 opt_rx->user_mss < in_mss)
3941 in_mss = opt_rx->user_mss;
3942 opt_rx->mss_clamp = in_mss;
3947 if (opsize == TCPOLEN_WINDOW && th->syn &&
3948 !estab && sysctl_tcp_window_scaling) {
3949 __u8 snd_wscale = *(__u8 *)ptr;
3950 opt_rx->wscale_ok = 1;
3951 if (snd_wscale > 14) {
3952 if (net_ratelimit())
3953 printk(KERN_INFO "tcp_parse_options: Illegal window "
3954 "scaling value %d >14 received.\n",
3958 opt_rx->snd_wscale = snd_wscale;
3961 case TCPOPT_TIMESTAMP:
3962 if ((opsize == TCPOLEN_TIMESTAMP) &&
3963 ((estab && opt_rx->tstamp_ok) ||
3964 (!estab && sysctl_tcp_timestamps))) {
3965 opt_rx->saw_tstamp = 1;
3966 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3967 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3970 case TCPOPT_SACK_PERM:
3971 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3972 !estab && sysctl_tcp_sack) {
3973 opt_rx->sack_ok = 1;
3974 tcp_sack_reset(opt_rx);
3979 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3980 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3982 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3985 #ifdef CONFIG_TCP_MD5SIG
3988 * The MD5 Hash has already been
3989 * checked (see tcp_v{4,6}_do_rcv()).
3994 /* This option is variable length.
3997 case TCPOLEN_COOKIE_BASE:
3998 /* not yet implemented */
4000 case TCPOLEN_COOKIE_PAIR:
4001 /* not yet implemented */
4003 case TCPOLEN_COOKIE_MIN+0:
4004 case TCPOLEN_COOKIE_MIN+2:
4005 case TCPOLEN_COOKIE_MIN+4:
4006 case TCPOLEN_COOKIE_MIN+6:
4007 case TCPOLEN_COOKIE_MAX:
4008 /* 16-bit multiple */
4009 opt_rx->cookie_plus = opsize;
4024 EXPORT_SYMBOL(tcp_parse_options);
4026 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4028 const __be32 *ptr = (const __be32 *)(th + 1);
4030 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4031 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4032 tp->rx_opt.saw_tstamp = 1;
4034 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4036 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4042 /* Fast parse options. This hopes to only see timestamps.
4043 * If it is wrong it falls back on tcp_parse_options().
4045 static int tcp_fast_parse_options(const struct sk_buff *skb,
4046 const struct tcphdr *th,
4047 struct tcp_sock *tp, const u8 **hvpp)
4049 /* In the spirit of fast parsing, compare doff directly to constant
4050 * values. Because equality is used, short doff can be ignored here.
4052 if (th->doff == (sizeof(*th) / 4)) {
4053 tp->rx_opt.saw_tstamp = 0;
4055 } else if (tp->rx_opt.tstamp_ok &&
4056 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4057 if (tcp_parse_aligned_timestamp(tp, th))
4060 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
4064 #ifdef CONFIG_TCP_MD5SIG
4066 * Parse MD5 Signature option
4068 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4070 int length = (th->doff << 2) - sizeof(*th);
4071 const u8 *ptr = (const u8 *)(th + 1);
4073 /* If the TCP option is too short, we can short cut */
4074 if (length < TCPOLEN_MD5SIG)
4077 while (length > 0) {
4078 int opcode = *ptr++;
4089 if (opsize < 2 || opsize > length)
4091 if (opcode == TCPOPT_MD5SIG)
4092 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4099 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4102 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4104 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4105 * it can pass through stack. So, the following predicate verifies that
4106 * this segment is not used for anything but congestion avoidance or
4107 * fast retransmit. Moreover, we even are able to eliminate most of such
4108 * second order effects, if we apply some small "replay" window (~RTO)
4109 * to timestamp space.
4111 * All these measures still do not guarantee that we reject wrapped ACKs
4112 * on networks with high bandwidth, when sequence space is recycled fastly,
4113 * but it guarantees that such events will be very rare and do not affect
4114 * connection seriously. This doesn't look nice, but alas, PAWS is really
4117 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4118 * states that events when retransmit arrives after original data are rare.
4119 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4120 * the biggest problem on large power networks even with minor reordering.
4121 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4122 * up to bandwidth of 18Gigabit/sec. 8) ]
4125 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4127 const struct tcp_sock *tp = tcp_sk(sk);
4128 const struct tcphdr *th = tcp_hdr(skb);
4129 u32 seq = TCP_SKB_CB(skb)->seq;
4130 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4132 return (/* 1. Pure ACK with correct sequence number. */
4133 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4135 /* 2. ... and duplicate ACK. */
4136 ack == tp->snd_una &&
4138 /* 3. ... and does not update window. */
4139 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4141 /* 4. ... and sits in replay window. */
4142 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4145 static inline int tcp_paws_discard(const struct sock *sk,
4146 const struct sk_buff *skb)
4148 const struct tcp_sock *tp = tcp_sk(sk);
4150 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4151 !tcp_disordered_ack(sk, skb);
4154 /* Check segment sequence number for validity.
4156 * Segment controls are considered valid, if the segment
4157 * fits to the window after truncation to the window. Acceptability
4158 * of data (and SYN, FIN, of course) is checked separately.
4159 * See tcp_data_queue(), for example.
4161 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4162 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4163 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4164 * (borrowed from freebsd)
4167 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4169 return !before(end_seq, tp->rcv_wup) &&
4170 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4173 /* When we get a reset we do this. */
4174 static void tcp_reset(struct sock *sk)
4176 /* We want the right error as BSD sees it (and indeed as we do). */
4177 switch (sk->sk_state) {
4179 sk->sk_err = ECONNREFUSED;
4181 case TCP_CLOSE_WAIT:
4187 sk->sk_err = ECONNRESET;
4189 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4192 if (!sock_flag(sk, SOCK_DEAD))
4193 sk->sk_error_report(sk);
4199 * Process the FIN bit. This now behaves as it is supposed to work
4200 * and the FIN takes effect when it is validly part of sequence
4201 * space. Not before when we get holes.
4203 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4204 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4207 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4208 * close and we go into CLOSING (and later onto TIME-WAIT)
4210 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4212 static void tcp_fin(struct sock *sk)
4214 struct tcp_sock *tp = tcp_sk(sk);
4216 inet_csk_schedule_ack(sk);
4218 sk->sk_shutdown |= RCV_SHUTDOWN;
4219 sock_set_flag(sk, SOCK_DONE);
4221 switch (sk->sk_state) {
4223 case TCP_ESTABLISHED:
4224 /* Move to CLOSE_WAIT */
4225 tcp_set_state(sk, TCP_CLOSE_WAIT);
4226 inet_csk(sk)->icsk_ack.pingpong = 1;
4229 case TCP_CLOSE_WAIT:
4231 /* Received a retransmission of the FIN, do
4236 /* RFC793: Remain in the LAST-ACK state. */
4240 /* This case occurs when a simultaneous close
4241 * happens, we must ack the received FIN and
4242 * enter the CLOSING state.
4245 tcp_set_state(sk, TCP_CLOSING);
4248 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4250 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4253 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4254 * cases we should never reach this piece of code.
4256 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4257 __func__, sk->sk_state);
4261 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4262 * Probably, we should reset in this case. For now drop them.
4264 __skb_queue_purge(&tp->out_of_order_queue);
4265 if (tcp_is_sack(tp))
4266 tcp_sack_reset(&tp->rx_opt);
4269 if (!sock_flag(sk, SOCK_DEAD)) {
4270 sk->sk_state_change(sk);
4272 /* Do not send POLL_HUP for half duplex close. */
4273 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4274 sk->sk_state == TCP_CLOSE)
4275 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4277 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4281 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4284 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4285 if (before(seq, sp->start_seq))
4286 sp->start_seq = seq;
4287 if (after(end_seq, sp->end_seq))
4288 sp->end_seq = end_seq;
4294 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4296 struct tcp_sock *tp = tcp_sk(sk);
4298 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4301 if (before(seq, tp->rcv_nxt))
4302 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4304 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4306 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4308 tp->rx_opt.dsack = 1;
4309 tp->duplicate_sack[0].start_seq = seq;
4310 tp->duplicate_sack[0].end_seq = end_seq;
4314 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4316 struct tcp_sock *tp = tcp_sk(sk);
4318 if (!tp->rx_opt.dsack)
4319 tcp_dsack_set(sk, seq, end_seq);
4321 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4324 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4326 struct tcp_sock *tp = tcp_sk(sk);
4328 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4329 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4330 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4331 tcp_enter_quickack_mode(sk);
4333 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4334 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4336 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4337 end_seq = tp->rcv_nxt;
4338 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4345 /* These routines update the SACK block as out-of-order packets arrive or
4346 * in-order packets close up the sequence space.
4348 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4351 struct tcp_sack_block *sp = &tp->selective_acks[0];
4352 struct tcp_sack_block *swalk = sp + 1;
4354 /* See if the recent change to the first SACK eats into
4355 * or hits the sequence space of other SACK blocks, if so coalesce.
4357 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4358 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4361 /* Zap SWALK, by moving every further SACK up by one slot.
4362 * Decrease num_sacks.
4364 tp->rx_opt.num_sacks--;
4365 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4369 this_sack++, swalk++;
4373 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4375 struct tcp_sock *tp = tcp_sk(sk);
4376 struct tcp_sack_block *sp = &tp->selective_acks[0];
4377 int cur_sacks = tp->rx_opt.num_sacks;
4383 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4384 if (tcp_sack_extend(sp, seq, end_seq)) {
4385 /* Rotate this_sack to the first one. */
4386 for (; this_sack > 0; this_sack--, sp--)
4387 swap(*sp, *(sp - 1));
4389 tcp_sack_maybe_coalesce(tp);
4394 /* Could not find an adjacent existing SACK, build a new one,
4395 * put it at the front, and shift everyone else down. We
4396 * always know there is at least one SACK present already here.
4398 * If the sack array is full, forget about the last one.
4400 if (this_sack >= TCP_NUM_SACKS) {
4402 tp->rx_opt.num_sacks--;
4405 for (; this_sack > 0; this_sack--, sp--)
4409 /* Build the new head SACK, and we're done. */
4410 sp->start_seq = seq;
4411 sp->end_seq = end_seq;
4412 tp->rx_opt.num_sacks++;
4415 /* RCV.NXT advances, some SACKs should be eaten. */
4417 static void tcp_sack_remove(struct tcp_sock *tp)
4419 struct tcp_sack_block *sp = &tp->selective_acks[0];
4420 int num_sacks = tp->rx_opt.num_sacks;
4423 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4424 if (skb_queue_empty(&tp->out_of_order_queue)) {
4425 tp->rx_opt.num_sacks = 0;
4429 for (this_sack = 0; this_sack < num_sacks;) {
4430 /* Check if the start of the sack is covered by RCV.NXT. */
4431 if (!before(tp->rcv_nxt, sp->start_seq)) {
4434 /* RCV.NXT must cover all the block! */
4435 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4437 /* Zap this SACK, by moving forward any other SACKS. */
4438 for (i=this_sack+1; i < num_sacks; i++)
4439 tp->selective_acks[i-1] = tp->selective_acks[i];
4446 tp->rx_opt.num_sacks = num_sacks;
4449 /* This one checks to see if we can put data from the
4450 * out_of_order queue into the receive_queue.
4452 static void tcp_ofo_queue(struct sock *sk)
4454 struct tcp_sock *tp = tcp_sk(sk);
4455 __u32 dsack_high = tp->rcv_nxt;
4456 struct sk_buff *skb;
4458 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4459 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4462 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4463 __u32 dsack = dsack_high;
4464 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4465 dsack_high = TCP_SKB_CB(skb)->end_seq;
4466 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4469 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4470 SOCK_DEBUG(sk, "ofo packet was already received\n");
4471 __skb_unlink(skb, &tp->out_of_order_queue);
4475 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4476 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4477 TCP_SKB_CB(skb)->end_seq);
4479 __skb_unlink(skb, &tp->out_of_order_queue);
4480 __skb_queue_tail(&sk->sk_receive_queue, skb);
4481 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4482 if (tcp_hdr(skb)->fin)
4487 static int tcp_prune_ofo_queue(struct sock *sk);
4488 static int tcp_prune_queue(struct sock *sk);
4490 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4492 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4493 !sk_rmem_schedule(sk, size)) {
4495 if (tcp_prune_queue(sk) < 0)
4498 if (!sk_rmem_schedule(sk, size)) {
4499 if (!tcp_prune_ofo_queue(sk))
4502 if (!sk_rmem_schedule(sk, size))
4509 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4511 const struct tcphdr *th = tcp_hdr(skb);
4512 struct tcp_sock *tp = tcp_sk(sk);
4515 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4519 __skb_pull(skb, th->doff * 4);
4521 TCP_ECN_accept_cwr(tp, skb);
4523 tp->rx_opt.dsack = 0;
4525 /* Queue data for delivery to the user.
4526 * Packets in sequence go to the receive queue.
4527 * Out of sequence packets to the out_of_order_queue.
4529 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4530 if (tcp_receive_window(tp) == 0)
4533 /* Ok. In sequence. In window. */
4534 if (tp->ucopy.task == current &&
4535 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4536 sock_owned_by_user(sk) && !tp->urg_data) {
4537 int chunk = min_t(unsigned int, skb->len,
4540 __set_current_state(TASK_RUNNING);
4543 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4544 tp->ucopy.len -= chunk;
4545 tp->copied_seq += chunk;
4546 eaten = (chunk == skb->len);
4547 tcp_rcv_space_adjust(sk);
4555 tcp_try_rmem_schedule(sk, skb->truesize))
4558 skb_set_owner_r(skb, sk);
4559 __skb_queue_tail(&sk->sk_receive_queue, skb);
4561 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4563 tcp_event_data_recv(sk, skb);
4567 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4570 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4571 * gap in queue is filled.
4573 if (skb_queue_empty(&tp->out_of_order_queue))
4574 inet_csk(sk)->icsk_ack.pingpong = 0;
4577 if (tp->rx_opt.num_sacks)
4578 tcp_sack_remove(tp);
4580 tcp_fast_path_check(sk);
4584 else if (!sock_flag(sk, SOCK_DEAD))
4585 sk->sk_data_ready(sk, 0);
4589 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4590 /* A retransmit, 2nd most common case. Force an immediate ack. */
4591 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4592 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4595 tcp_enter_quickack_mode(sk);
4596 inet_csk_schedule_ack(sk);
4602 /* Out of window. F.e. zero window probe. */
4603 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4606 tcp_enter_quickack_mode(sk);
4608 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4609 /* Partial packet, seq < rcv_next < end_seq */
4610 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4611 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4612 TCP_SKB_CB(skb)->end_seq);
4614 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4616 /* If window is closed, drop tail of packet. But after
4617 * remembering D-SACK for its head made in previous line.
4619 if (!tcp_receive_window(tp))
4624 TCP_ECN_check_ce(tp, skb);
4626 if (tcp_try_rmem_schedule(sk, skb->truesize))
4629 /* Disable header prediction. */
4631 inet_csk_schedule_ack(sk);
4633 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4634 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4636 skb_set_owner_r(skb, sk);
4638 if (!skb_peek(&tp->out_of_order_queue)) {
4639 /* Initial out of order segment, build 1 SACK. */
4640 if (tcp_is_sack(tp)) {
4641 tp->rx_opt.num_sacks = 1;
4642 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4643 tp->selective_acks[0].end_seq =
4644 TCP_SKB_CB(skb)->end_seq;
4646 __skb_queue_head(&tp->out_of_order_queue, skb);
4648 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4649 u32 seq = TCP_SKB_CB(skb)->seq;
4650 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4652 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4653 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4655 if (!tp->rx_opt.num_sacks ||
4656 tp->selective_acks[0].end_seq != seq)
4659 /* Common case: data arrive in order after hole. */
4660 tp->selective_acks[0].end_seq = end_seq;
4664 /* Find place to insert this segment. */
4666 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4668 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4672 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4675 /* Do skb overlap to previous one? */
4676 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4677 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4678 /* All the bits are present. Drop. */
4680 tcp_dsack_set(sk, seq, end_seq);
4683 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4684 /* Partial overlap. */
4685 tcp_dsack_set(sk, seq,
4686 TCP_SKB_CB(skb1)->end_seq);
4688 if (skb_queue_is_first(&tp->out_of_order_queue,
4692 skb1 = skb_queue_prev(
4693 &tp->out_of_order_queue,
4698 __skb_queue_head(&tp->out_of_order_queue, skb);
4700 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4702 /* And clean segments covered by new one as whole. */
4703 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4704 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4706 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4708 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4709 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4713 __skb_unlink(skb1, &tp->out_of_order_queue);
4714 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4715 TCP_SKB_CB(skb1)->end_seq);
4720 if (tcp_is_sack(tp))
4721 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4725 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4726 struct sk_buff_head *list)
4728 struct sk_buff *next = NULL;
4730 if (!skb_queue_is_last(list, skb))
4731 next = skb_queue_next(list, skb);
4733 __skb_unlink(skb, list);
4735 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4740 /* Collapse contiguous sequence of skbs head..tail with
4741 * sequence numbers start..end.
4743 * If tail is NULL, this means until the end of the list.
4745 * Segments with FIN/SYN are not collapsed (only because this
4749 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4750 struct sk_buff *head, struct sk_buff *tail,
4753 struct sk_buff *skb, *n;
4756 /* First, check that queue is collapsible and find
4757 * the point where collapsing can be useful. */
4761 skb_queue_walk_from_safe(list, skb, n) {
4764 /* No new bits? It is possible on ofo queue. */
4765 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4766 skb = tcp_collapse_one(sk, skb, list);
4772 /* The first skb to collapse is:
4774 * - bloated or contains data before "start" or
4775 * overlaps to the next one.
4777 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4778 (tcp_win_from_space(skb->truesize) > skb->len ||
4779 before(TCP_SKB_CB(skb)->seq, start))) {
4780 end_of_skbs = false;
4784 if (!skb_queue_is_last(list, skb)) {
4785 struct sk_buff *next = skb_queue_next(list, skb);
4787 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4788 end_of_skbs = false;
4793 /* Decided to skip this, advance start seq. */
4794 start = TCP_SKB_CB(skb)->end_seq;
4796 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4799 while (before(start, end)) {
4800 struct sk_buff *nskb;
4801 unsigned int header = skb_headroom(skb);
4802 int copy = SKB_MAX_ORDER(header, 0);
4804 /* Too big header? This can happen with IPv6. */
4807 if (end - start < copy)
4809 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4813 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4814 skb_set_network_header(nskb, (skb_network_header(skb) -
4816 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4818 skb_reserve(nskb, header);
4819 memcpy(nskb->head, skb->head, header);
4820 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4821 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4822 __skb_queue_before(list, skb, nskb);
4823 skb_set_owner_r(nskb, sk);
4825 /* Copy data, releasing collapsed skbs. */
4827 int offset = start - TCP_SKB_CB(skb)->seq;
4828 int size = TCP_SKB_CB(skb)->end_seq - start;
4832 size = min(copy, size);
4833 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4835 TCP_SKB_CB(nskb)->end_seq += size;
4839 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4840 skb = tcp_collapse_one(sk, skb, list);
4843 tcp_hdr(skb)->syn ||
4851 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4852 * and tcp_collapse() them until all the queue is collapsed.
4854 static void tcp_collapse_ofo_queue(struct sock *sk)
4856 struct tcp_sock *tp = tcp_sk(sk);
4857 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4858 struct sk_buff *head;
4864 start = TCP_SKB_CB(skb)->seq;
4865 end = TCP_SKB_CB(skb)->end_seq;
4869 struct sk_buff *next = NULL;
4871 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4872 next = skb_queue_next(&tp->out_of_order_queue, skb);
4875 /* Segment is terminated when we see gap or when
4876 * we are at the end of all the queue. */
4878 after(TCP_SKB_CB(skb)->seq, end) ||
4879 before(TCP_SKB_CB(skb)->end_seq, start)) {
4880 tcp_collapse(sk, &tp->out_of_order_queue,
4881 head, skb, start, end);
4885 /* Start new segment */
4886 start = TCP_SKB_CB(skb)->seq;
4887 end = TCP_SKB_CB(skb)->end_seq;
4889 if (before(TCP_SKB_CB(skb)->seq, start))
4890 start = TCP_SKB_CB(skb)->seq;
4891 if (after(TCP_SKB_CB(skb)->end_seq, end))
4892 end = TCP_SKB_CB(skb)->end_seq;
4898 * Purge the out-of-order queue.
4899 * Return true if queue was pruned.
4901 static int tcp_prune_ofo_queue(struct sock *sk)
4903 struct tcp_sock *tp = tcp_sk(sk);
4906 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4907 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4908 __skb_queue_purge(&tp->out_of_order_queue);
4910 /* Reset SACK state. A conforming SACK implementation will
4911 * do the same at a timeout based retransmit. When a connection
4912 * is in a sad state like this, we care only about integrity
4913 * of the connection not performance.
4915 if (tp->rx_opt.sack_ok)
4916 tcp_sack_reset(&tp->rx_opt);
4923 /* Reduce allocated memory if we can, trying to get
4924 * the socket within its memory limits again.
4926 * Return less than zero if we should start dropping frames
4927 * until the socket owning process reads some of the data
4928 * to stabilize the situation.
4930 static int tcp_prune_queue(struct sock *sk)
4932 struct tcp_sock *tp = tcp_sk(sk);
4934 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4936 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4938 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4939 tcp_clamp_window(sk);
4940 else if (tcp_memory_pressure)
4941 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4943 tcp_collapse_ofo_queue(sk);
4944 if (!skb_queue_empty(&sk->sk_receive_queue))
4945 tcp_collapse(sk, &sk->sk_receive_queue,
4946 skb_peek(&sk->sk_receive_queue),
4948 tp->copied_seq, tp->rcv_nxt);
4951 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4954 /* Collapsing did not help, destructive actions follow.
4955 * This must not ever occur. */
4957 tcp_prune_ofo_queue(sk);
4959 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4962 /* If we are really being abused, tell the caller to silently
4963 * drop receive data on the floor. It will get retransmitted
4964 * and hopefully then we'll have sufficient space.
4966 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4968 /* Massive buffer overcommit. */
4973 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4974 * As additional protections, we do not touch cwnd in retransmission phases,
4975 * and if application hit its sndbuf limit recently.
4977 void tcp_cwnd_application_limited(struct sock *sk)
4979 struct tcp_sock *tp = tcp_sk(sk);
4981 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4982 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4983 /* Limited by application or receiver window. */
4984 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4985 u32 win_used = max(tp->snd_cwnd_used, init_win);
4986 if (win_used < tp->snd_cwnd) {
4987 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4988 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4990 tp->snd_cwnd_used = 0;
4992 tp->snd_cwnd_stamp = tcp_time_stamp;
4995 static int tcp_should_expand_sndbuf(const struct sock *sk)
4997 const struct tcp_sock *tp = tcp_sk(sk);
4999 /* If the user specified a specific send buffer setting, do
5002 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5005 /* If we are under global TCP memory pressure, do not expand. */
5006 if (tcp_memory_pressure)
5009 /* If we are under soft global TCP memory pressure, do not expand. */
5010 if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
5013 /* If we filled the congestion window, do not expand. */
5014 if (tp->packets_out >= tp->snd_cwnd)
5020 /* When incoming ACK allowed to free some skb from write_queue,
5021 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5022 * on the exit from tcp input handler.
5024 * PROBLEM: sndbuf expansion does not work well with largesend.
5026 static void tcp_new_space(struct sock *sk)
5028 struct tcp_sock *tp = tcp_sk(sk);
5030 if (tcp_should_expand_sndbuf(sk)) {
5031 int sndmem = SKB_TRUESIZE(max_t(u32,
5032 tp->rx_opt.mss_clamp,
5035 int demanded = max_t(unsigned int, tp->snd_cwnd,
5036 tp->reordering + 1);
5037 sndmem *= 2 * demanded;
5038 if (sndmem > sk->sk_sndbuf)
5039 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5040 tp->snd_cwnd_stamp = tcp_time_stamp;
5043 sk->sk_write_space(sk);
5046 static void tcp_check_space(struct sock *sk)
5048 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5049 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5050 if (sk->sk_socket &&
5051 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5056 static inline void tcp_data_snd_check(struct sock *sk)
5058 tcp_push_pending_frames(sk);
5059 tcp_check_space(sk);
5063 * Check if sending an ack is needed.
5065 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5067 struct tcp_sock *tp = tcp_sk(sk);
5069 /* More than one full frame received... */
5070 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5071 /* ... and right edge of window advances far enough.
5072 * (tcp_recvmsg() will send ACK otherwise). Or...
5074 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5075 /* We ACK each frame or... */
5076 tcp_in_quickack_mode(sk) ||
5077 /* We have out of order data. */
5078 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5079 /* Then ack it now */
5082 /* Else, send delayed ack. */
5083 tcp_send_delayed_ack(sk);
5087 static inline void tcp_ack_snd_check(struct sock *sk)
5089 if (!inet_csk_ack_scheduled(sk)) {
5090 /* We sent a data segment already. */
5093 __tcp_ack_snd_check(sk, 1);
5097 * This routine is only called when we have urgent data
5098 * signaled. Its the 'slow' part of tcp_urg. It could be
5099 * moved inline now as tcp_urg is only called from one
5100 * place. We handle URGent data wrong. We have to - as
5101 * BSD still doesn't use the correction from RFC961.
5102 * For 1003.1g we should support a new option TCP_STDURG to permit
5103 * either form (or just set the sysctl tcp_stdurg).
5106 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5108 struct tcp_sock *tp = tcp_sk(sk);
5109 u32 ptr = ntohs(th->urg_ptr);
5111 if (ptr && !sysctl_tcp_stdurg)
5113 ptr += ntohl(th->seq);
5115 /* Ignore urgent data that we've already seen and read. */
5116 if (after(tp->copied_seq, ptr))
5119 /* Do not replay urg ptr.
5121 * NOTE: interesting situation not covered by specs.
5122 * Misbehaving sender may send urg ptr, pointing to segment,
5123 * which we already have in ofo queue. We are not able to fetch
5124 * such data and will stay in TCP_URG_NOTYET until will be eaten
5125 * by recvmsg(). Seems, we are not obliged to handle such wicked
5126 * situations. But it is worth to think about possibility of some
5127 * DoSes using some hypothetical application level deadlock.
5129 if (before(ptr, tp->rcv_nxt))
5132 /* Do we already have a newer (or duplicate) urgent pointer? */
5133 if (tp->urg_data && !after(ptr, tp->urg_seq))
5136 /* Tell the world about our new urgent pointer. */
5139 /* We may be adding urgent data when the last byte read was
5140 * urgent. To do this requires some care. We cannot just ignore
5141 * tp->copied_seq since we would read the last urgent byte again
5142 * as data, nor can we alter copied_seq until this data arrives
5143 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5145 * NOTE. Double Dutch. Rendering to plain English: author of comment
5146 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5147 * and expect that both A and B disappear from stream. This is _wrong_.
5148 * Though this happens in BSD with high probability, this is occasional.
5149 * Any application relying on this is buggy. Note also, that fix "works"
5150 * only in this artificial test. Insert some normal data between A and B and we will
5151 * decline of BSD again. Verdict: it is better to remove to trap
5154 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5155 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5156 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5158 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5159 __skb_unlink(skb, &sk->sk_receive_queue);
5164 tp->urg_data = TCP_URG_NOTYET;
5167 /* Disable header prediction. */
5171 /* This is the 'fast' part of urgent handling. */
5172 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5174 struct tcp_sock *tp = tcp_sk(sk);
5176 /* Check if we get a new urgent pointer - normally not. */
5178 tcp_check_urg(sk, th);
5180 /* Do we wait for any urgent data? - normally not... */
5181 if (tp->urg_data == TCP_URG_NOTYET) {
5182 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5185 /* Is the urgent pointer pointing into this packet? */
5186 if (ptr < skb->len) {
5188 if (skb_copy_bits(skb, ptr, &tmp, 1))
5190 tp->urg_data = TCP_URG_VALID | tmp;
5191 if (!sock_flag(sk, SOCK_DEAD))
5192 sk->sk_data_ready(sk, 0);
5197 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5199 struct tcp_sock *tp = tcp_sk(sk);
5200 int chunk = skb->len - hlen;
5204 if (skb_csum_unnecessary(skb))
5205 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5207 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5211 tp->ucopy.len -= chunk;
5212 tp->copied_seq += chunk;
5213 tcp_rcv_space_adjust(sk);
5220 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5221 struct sk_buff *skb)
5225 if (sock_owned_by_user(sk)) {
5227 result = __tcp_checksum_complete(skb);
5230 result = __tcp_checksum_complete(skb);
5235 static inline int tcp_checksum_complete_user(struct sock *sk,
5236 struct sk_buff *skb)
5238 return !skb_csum_unnecessary(skb) &&
5239 __tcp_checksum_complete_user(sk, skb);
5242 #ifdef CONFIG_NET_DMA
5243 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5246 struct tcp_sock *tp = tcp_sk(sk);
5247 int chunk = skb->len - hlen;
5249 int copied_early = 0;
5251 if (tp->ucopy.wakeup)
5254 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5255 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5257 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5259 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5261 tp->ucopy.iov, chunk,
5262 tp->ucopy.pinned_list);
5267 tp->ucopy.dma_cookie = dma_cookie;
5270 tp->ucopy.len -= chunk;
5271 tp->copied_seq += chunk;
5272 tcp_rcv_space_adjust(sk);
5274 if ((tp->ucopy.len == 0) ||
5275 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5276 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5277 tp->ucopy.wakeup = 1;
5278 sk->sk_data_ready(sk, 0);
5280 } else if (chunk > 0) {
5281 tp->ucopy.wakeup = 1;
5282 sk->sk_data_ready(sk, 0);
5285 return copied_early;
5287 #endif /* CONFIG_NET_DMA */
5289 /* Does PAWS and seqno based validation of an incoming segment, flags will
5290 * play significant role here.
5292 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5293 const struct tcphdr *th, int syn_inerr)
5295 const u8 *hash_location;
5296 struct tcp_sock *tp = tcp_sk(sk);
5298 /* RFC1323: H1. Apply PAWS check first. */
5299 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5300 tp->rx_opt.saw_tstamp &&
5301 tcp_paws_discard(sk, skb)) {
5303 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5304 tcp_send_dupack(sk, skb);
5307 /* Reset is accepted even if it did not pass PAWS. */
5310 /* Step 1: check sequence number */
5311 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5312 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5313 * (RST) segments are validated by checking their SEQ-fields."
5314 * And page 69: "If an incoming segment is not acceptable,
5315 * an acknowledgment should be sent in reply (unless the RST
5316 * bit is set, if so drop the segment and return)".
5321 tcp_send_dupack(sk, skb);
5326 /* Step 2: check RST bit */
5329 * If sequence number exactly matches RCV.NXT, then
5330 * RESET the connection
5332 * Send a challenge ACK
5334 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5337 tcp_send_challenge_ack(sk);
5341 /* step 3: check security and precedence [ignored] */
5343 /* step 4: Check for a SYN
5344 * RFC 5691 4.2 : Send a challenge ack
5349 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5350 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5351 tcp_send_challenge_ack(sk);
5363 * TCP receive function for the ESTABLISHED state.
5365 * It is split into a fast path and a slow path. The fast path is
5367 * - A zero window was announced from us - zero window probing
5368 * is only handled properly in the slow path.
5369 * - Out of order segments arrived.
5370 * - Urgent data is expected.
5371 * - There is no buffer space left
5372 * - Unexpected TCP flags/window values/header lengths are received
5373 * (detected by checking the TCP header against pred_flags)
5374 * - Data is sent in both directions. Fast path only supports pure senders
5375 * or pure receivers (this means either the sequence number or the ack
5376 * value must stay constant)
5377 * - Unexpected TCP option.
5379 * When these conditions are not satisfied it drops into a standard
5380 * receive procedure patterned after RFC793 to handle all cases.
5381 * The first three cases are guaranteed by proper pred_flags setting,
5382 * the rest is checked inline. Fast processing is turned on in
5383 * tcp_data_queue when everything is OK.
5385 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5386 const struct tcphdr *th, unsigned int len)
5388 struct tcp_sock *tp = tcp_sk(sk);
5391 * Header prediction.
5392 * The code loosely follows the one in the famous
5393 * "30 instruction TCP receive" Van Jacobson mail.
5395 * Van's trick is to deposit buffers into socket queue
5396 * on a device interrupt, to call tcp_recv function
5397 * on the receive process context and checksum and copy
5398 * the buffer to user space. smart...
5400 * Our current scheme is not silly either but we take the
5401 * extra cost of the net_bh soft interrupt processing...
5402 * We do checksum and copy also but from device to kernel.
5405 tp->rx_opt.saw_tstamp = 0;
5407 /* pred_flags is 0xS?10 << 16 + snd_wnd
5408 * if header_prediction is to be made
5409 * 'S' will always be tp->tcp_header_len >> 2
5410 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5411 * turn it off (when there are holes in the receive
5412 * space for instance)
5413 * PSH flag is ignored.
5416 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5417 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5418 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5419 int tcp_header_len = tp->tcp_header_len;
5421 /* Timestamp header prediction: tcp_header_len
5422 * is automatically equal to th->doff*4 due to pred_flags
5426 /* Check timestamp */
5427 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5428 /* No? Slow path! */
5429 if (!tcp_parse_aligned_timestamp(tp, th))
5432 /* If PAWS failed, check it more carefully in slow path */
5433 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5436 /* DO NOT update ts_recent here, if checksum fails
5437 * and timestamp was corrupted part, it will result
5438 * in a hung connection since we will drop all
5439 * future packets due to the PAWS test.
5443 if (len <= tcp_header_len) {
5444 /* Bulk data transfer: sender */
5445 if (len == tcp_header_len) {
5446 /* Predicted packet is in window by definition.
5447 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5448 * Hence, check seq<=rcv_wup reduces to:
5450 if (tcp_header_len ==
5451 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5452 tp->rcv_nxt == tp->rcv_wup)
5453 tcp_store_ts_recent(tp);
5455 /* We know that such packets are checksummed
5458 tcp_ack(sk, skb, 0);
5460 tcp_data_snd_check(sk);
5462 } else { /* Header too small */
5463 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5468 int copied_early = 0;
5470 if (tp->copied_seq == tp->rcv_nxt &&
5471 len - tcp_header_len <= tp->ucopy.len) {
5472 #ifdef CONFIG_NET_DMA
5473 if (tp->ucopy.task == current &&
5474 sock_owned_by_user(sk) &&
5475 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5480 if (tp->ucopy.task == current &&
5481 sock_owned_by_user(sk) && !copied_early) {
5482 __set_current_state(TASK_RUNNING);
5484 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5488 /* Predicted packet is in window by definition.
5489 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5490 * Hence, check seq<=rcv_wup reduces to:
5492 if (tcp_header_len ==
5493 (sizeof(struct tcphdr) +
5494 TCPOLEN_TSTAMP_ALIGNED) &&
5495 tp->rcv_nxt == tp->rcv_wup)
5496 tcp_store_ts_recent(tp);
5498 tcp_rcv_rtt_measure_ts(sk, skb);
5500 __skb_pull(skb, tcp_header_len);
5501 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5502 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5505 tcp_cleanup_rbuf(sk, skb->len);
5508 if (tcp_checksum_complete_user(sk, skb))
5511 if ((int)skb->truesize > sk->sk_forward_alloc)
5514 /* Predicted packet is in window by definition.
5515 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5516 * Hence, check seq<=rcv_wup reduces to:
5518 if (tcp_header_len ==
5519 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5520 tp->rcv_nxt == tp->rcv_wup)
5521 tcp_store_ts_recent(tp);
5523 tcp_rcv_rtt_measure_ts(sk, skb);
5525 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5527 /* Bulk data transfer: receiver */
5528 __skb_pull(skb, tcp_header_len);
5529 __skb_queue_tail(&sk->sk_receive_queue, skb);
5530 skb_set_owner_r(skb, sk);
5531 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5534 tcp_event_data_recv(sk, skb);
5536 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5537 /* Well, only one small jumplet in fast path... */
5538 tcp_ack(sk, skb, FLAG_DATA);
5539 tcp_data_snd_check(sk);
5540 if (!inet_csk_ack_scheduled(sk))
5544 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5545 __tcp_ack_snd_check(sk, 0);
5547 #ifdef CONFIG_NET_DMA
5549 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5555 sk->sk_data_ready(sk, 0);
5561 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5565 * Standard slow path.
5568 if (!tcp_validate_incoming(sk, skb, th, 1))
5573 tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5576 tcp_rcv_rtt_measure_ts(sk, skb);
5578 /* Process urgent data. */
5579 tcp_urg(sk, skb, th);
5581 /* step 7: process the segment text */
5582 tcp_data_queue(sk, skb);
5584 tcp_data_snd_check(sk);
5585 tcp_ack_snd_check(sk);
5589 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5595 EXPORT_SYMBOL(tcp_rcv_established);
5597 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5598 const struct tcphdr *th, unsigned int len)
5600 const u8 *hash_location;
5601 struct inet_connection_sock *icsk = inet_csk(sk);
5602 struct tcp_sock *tp = tcp_sk(sk);
5603 struct tcp_cookie_values *cvp = tp->cookie_values;
5604 int saved_clamp = tp->rx_opt.mss_clamp;
5606 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5610 * "If the state is SYN-SENT then
5611 * first check the ACK bit
5612 * If the ACK bit is set
5613 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5614 * a reset (unless the RST bit is set, if so drop
5615 * the segment and return)"
5617 * We do not send data with SYN, so that RFC-correct
5620 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5621 goto reset_and_undo;
5623 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5624 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5626 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5627 goto reset_and_undo;
5630 /* Now ACK is acceptable.
5632 * "If the RST bit is set
5633 * If the ACK was acceptable then signal the user "error:
5634 * connection reset", drop the segment, enter CLOSED state,
5635 * delete TCB, and return."
5644 * "fifth, if neither of the SYN or RST bits is set then
5645 * drop the segment and return."
5651 goto discard_and_undo;
5654 * "If the SYN bit is on ...
5655 * are acceptable then ...
5656 * (our SYN has been ACKed), change the connection
5657 * state to ESTABLISHED..."
5660 TCP_ECN_rcv_synack(tp, th);
5662 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5663 tcp_ack(sk, skb, FLAG_SLOWPATH);
5665 /* Ok.. it's good. Set up sequence numbers and
5666 * move to established.
5668 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5669 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5671 /* RFC1323: The window in SYN & SYN/ACK segments is
5674 tp->snd_wnd = ntohs(th->window);
5675 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5677 if (!tp->rx_opt.wscale_ok) {
5678 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5679 tp->window_clamp = min(tp->window_clamp, 65535U);
5682 if (tp->rx_opt.saw_tstamp) {
5683 tp->rx_opt.tstamp_ok = 1;
5684 tp->tcp_header_len =
5685 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5686 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5687 tcp_store_ts_recent(tp);
5689 tp->tcp_header_len = sizeof(struct tcphdr);
5692 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5693 tcp_enable_fack(tp);
5696 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5697 tcp_initialize_rcv_mss(sk);
5699 /* Remember, tcp_poll() does not lock socket!
5700 * Change state from SYN-SENT only after copied_seq
5701 * is initialized. */
5702 tp->copied_seq = tp->rcv_nxt;
5705 cvp->cookie_pair_size > 0 &&
5706 tp->rx_opt.cookie_plus > 0) {
5707 int cookie_size = tp->rx_opt.cookie_plus
5708 - TCPOLEN_COOKIE_BASE;
5709 int cookie_pair_size = cookie_size
5710 + cvp->cookie_desired;
5712 /* A cookie extension option was sent and returned.
5713 * Note that each incoming SYNACK replaces the
5714 * Responder cookie. The initial exchange is most
5715 * fragile, as protection against spoofing relies
5716 * entirely upon the sequence and timestamp (above).
5717 * This replacement strategy allows the correct pair to
5718 * pass through, while any others will be filtered via
5719 * Responder verification later.
5721 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5722 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5723 hash_location, cookie_size);
5724 cvp->cookie_pair_size = cookie_pair_size;
5729 tcp_set_state(sk, TCP_ESTABLISHED);
5731 security_inet_conn_established(sk, skb);
5733 /* Make sure socket is routed, for correct metrics. */
5734 icsk->icsk_af_ops->rebuild_header(sk);
5736 tcp_init_metrics(sk);
5738 tcp_init_congestion_control(sk);
5740 /* Prevent spurious tcp_cwnd_restart() on first data
5743 tp->lsndtime = tcp_time_stamp;
5745 tcp_init_buffer_space(sk);
5747 if (sock_flag(sk, SOCK_KEEPOPEN))
5748 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5750 if (!tp->rx_opt.snd_wscale)
5751 __tcp_fast_path_on(tp, tp->snd_wnd);
5755 if (!sock_flag(sk, SOCK_DEAD)) {
5756 sk->sk_state_change(sk);
5757 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5760 if (sk->sk_write_pending ||
5761 icsk->icsk_accept_queue.rskq_defer_accept ||
5762 icsk->icsk_ack.pingpong) {
5763 /* Save one ACK. Data will be ready after
5764 * several ticks, if write_pending is set.
5766 * It may be deleted, but with this feature tcpdumps
5767 * look so _wonderfully_ clever, that I was not able
5768 * to stand against the temptation 8) --ANK
5770 inet_csk_schedule_ack(sk);
5771 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5772 icsk->icsk_ack.ato = TCP_ATO_MIN;
5773 tcp_incr_quickack(sk);
5774 tcp_enter_quickack_mode(sk);
5775 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5776 TCP_DELACK_MAX, TCP_RTO_MAX);
5787 /* No ACK in the segment */
5791 * "If the RST bit is set
5793 * Otherwise (no ACK) drop the segment and return."
5796 goto discard_and_undo;
5800 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5801 tcp_paws_reject(&tp->rx_opt, 0))
5802 goto discard_and_undo;
5805 /* We see SYN without ACK. It is attempt of
5806 * simultaneous connect with crossed SYNs.
5807 * Particularly, it can be connect to self.
5809 tcp_set_state(sk, TCP_SYN_RECV);
5811 if (tp->rx_opt.saw_tstamp) {
5812 tp->rx_opt.tstamp_ok = 1;
5813 tcp_store_ts_recent(tp);
5814 tp->tcp_header_len =
5815 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5817 tp->tcp_header_len = sizeof(struct tcphdr);
5820 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5821 tp->copied_seq = tp->rcv_nxt;
5822 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5824 /* RFC1323: The window in SYN & SYN/ACK segments is
5827 tp->snd_wnd = ntohs(th->window);
5828 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5829 tp->max_window = tp->snd_wnd;
5831 TCP_ECN_rcv_syn(tp, th);
5834 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5835 tcp_initialize_rcv_mss(sk);
5837 tcp_send_synack(sk);
5839 /* Note, we could accept data and URG from this segment.
5840 * There are no obstacles to make this.
5842 * However, if we ignore data in ACKless segments sometimes,
5843 * we have no reasons to accept it sometimes.
5844 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5845 * is not flawless. So, discard packet for sanity.
5846 * Uncomment this return to process the data.
5853 /* "fifth, if neither of the SYN or RST bits is set then
5854 * drop the segment and return."
5858 tcp_clear_options(&tp->rx_opt);
5859 tp->rx_opt.mss_clamp = saved_clamp;
5863 tcp_clear_options(&tp->rx_opt);
5864 tp->rx_opt.mss_clamp = saved_clamp;
5869 * This function implements the receiving procedure of RFC 793 for
5870 * all states except ESTABLISHED and TIME_WAIT.
5871 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5872 * address independent.
5875 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5876 const struct tcphdr *th, unsigned int len)
5878 struct tcp_sock *tp = tcp_sk(sk);
5879 struct inet_connection_sock *icsk = inet_csk(sk);
5882 tp->rx_opt.saw_tstamp = 0;
5884 switch (sk->sk_state) {
5898 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5901 /* Now we have several options: In theory there is
5902 * nothing else in the frame. KA9Q has an option to
5903 * send data with the syn, BSD accepts data with the
5904 * syn up to the [to be] advertised window and
5905 * Solaris 2.1 gives you a protocol error. For now
5906 * we just ignore it, that fits the spec precisely
5907 * and avoids incompatibilities. It would be nice in
5908 * future to drop through and process the data.
5910 * Now that TTCP is starting to be used we ought to
5912 * But, this leaves one open to an easy denial of
5913 * service attack, and SYN cookies can't defend
5914 * against this problem. So, we drop the data
5915 * in the interest of security over speed unless
5916 * it's still in use.
5924 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5928 /* Do step6 onward by hand. */
5929 tcp_urg(sk, skb, th);
5931 tcp_data_snd_check(sk);
5935 if (!tcp_validate_incoming(sk, skb, th, 0))
5938 /* step 5: check the ACK field */
5940 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5941 FLAG_UPDATE_TS_RECENT) > 0;
5943 switch (sk->sk_state) {
5946 tp->copied_seq = tp->rcv_nxt;
5948 tcp_set_state(sk, TCP_ESTABLISHED);
5949 sk->sk_state_change(sk);
5951 /* Note, that this wakeup is only for marginal
5952 * crossed SYN case. Passively open sockets
5953 * are not waked up, because sk->sk_sleep ==
5954 * NULL and sk->sk_socket == NULL.
5958 SOCK_WAKE_IO, POLL_OUT);
5960 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5961 tp->snd_wnd = ntohs(th->window) <<
5962 tp->rx_opt.snd_wscale;
5963 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5965 if (tp->rx_opt.tstamp_ok)
5966 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5968 /* Make sure socket is routed, for
5971 icsk->icsk_af_ops->rebuild_header(sk);
5973 tcp_init_metrics(sk);
5975 tcp_init_congestion_control(sk);
5977 /* Prevent spurious tcp_cwnd_restart() on
5978 * first data packet.
5980 tp->lsndtime = tcp_time_stamp;
5983 tcp_initialize_rcv_mss(sk);
5984 tcp_init_buffer_space(sk);
5985 tcp_fast_path_on(tp);
5992 if (tp->snd_una == tp->write_seq) {
5993 tcp_set_state(sk, TCP_FIN_WAIT2);
5994 sk->sk_shutdown |= SEND_SHUTDOWN;
5995 dst_confirm(__sk_dst_get(sk));
5997 if (!sock_flag(sk, SOCK_DEAD))
5998 /* Wake up lingering close() */
5999 sk->sk_state_change(sk);
6003 if (tp->linger2 < 0 ||
6004 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6005 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6007 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6011 tmo = tcp_fin_time(sk);
6012 if (tmo > TCP_TIMEWAIT_LEN) {
6013 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6014 } else if (th->fin || sock_owned_by_user(sk)) {
6015 /* Bad case. We could lose such FIN otherwise.
6016 * It is not a big problem, but it looks confusing
6017 * and not so rare event. We still can lose it now,
6018 * if it spins in bh_lock_sock(), but it is really
6021 inet_csk_reset_keepalive_timer(sk, tmo);
6023 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6031 if (tp->snd_una == tp->write_seq) {
6032 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6038 if (tp->snd_una == tp->write_seq) {
6039 tcp_update_metrics(sk);
6048 /* step 6: check the URG bit */
6049 tcp_urg(sk, skb, th);
6051 /* step 7: process the segment text */
6052 switch (sk->sk_state) {
6053 case TCP_CLOSE_WAIT:
6056 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6060 /* RFC 793 says to queue data in these states,
6061 * RFC 1122 says we MUST send a reset.
6062 * BSD 4.4 also does reset.
6064 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6065 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6066 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6067 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6073 case TCP_ESTABLISHED:
6074 tcp_data_queue(sk, skb);
6079 /* tcp_data could move socket to TIME-WAIT */
6080 if (sk->sk_state != TCP_CLOSE) {
6081 tcp_data_snd_check(sk);
6082 tcp_ack_snd_check(sk);
6091 EXPORT_SYMBOL(tcp_rcv_state_process);