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)
1311 if (pkt_len >= skb->len && !in_sack)
1314 err = tcp_fragment(sk, skb, pkt_len, mss);
1322 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1323 static u8 tcp_sacktag_one(struct sock *sk,
1324 struct tcp_sacktag_state *state, u8 sacked,
1325 u32 start_seq, u32 end_seq,
1326 int dup_sack, int pcount)
1328 struct tcp_sock *tp = tcp_sk(sk);
1329 int fack_count = state->fack_count;
1331 /* Account D-SACK for retransmitted packet. */
1332 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1333 if (tp->undo_marker && tp->undo_retrans &&
1334 after(end_seq, tp->undo_marker))
1336 if (sacked & TCPCB_SACKED_ACKED)
1337 state->reord = min(fack_count, state->reord);
1340 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1341 if (!after(end_seq, tp->snd_una))
1344 if (!(sacked & TCPCB_SACKED_ACKED)) {
1345 if (sacked & TCPCB_SACKED_RETRANS) {
1346 /* If the segment is not tagged as lost,
1347 * we do not clear RETRANS, believing
1348 * that retransmission is still in flight.
1350 if (sacked & TCPCB_LOST) {
1351 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1352 tp->lost_out -= pcount;
1353 tp->retrans_out -= pcount;
1356 if (!(sacked & TCPCB_RETRANS)) {
1357 /* New sack for not retransmitted frame,
1358 * which was in hole. It is reordering.
1360 if (before(start_seq,
1361 tcp_highest_sack_seq(tp)))
1362 state->reord = min(fack_count,
1365 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1366 if (!after(end_seq, tp->frto_highmark))
1367 state->flag |= FLAG_ONLY_ORIG_SACKED;
1370 if (sacked & TCPCB_LOST) {
1371 sacked &= ~TCPCB_LOST;
1372 tp->lost_out -= pcount;
1376 sacked |= TCPCB_SACKED_ACKED;
1377 state->flag |= FLAG_DATA_SACKED;
1378 tp->sacked_out += pcount;
1380 fack_count += pcount;
1382 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1383 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1384 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1385 tp->lost_cnt_hint += pcount;
1387 if (fack_count > tp->fackets_out)
1388 tp->fackets_out = fack_count;
1391 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1392 * frames and clear it. undo_retrans is decreased above, L|R frames
1393 * are accounted above as well.
1395 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1396 sacked &= ~TCPCB_SACKED_RETRANS;
1397 tp->retrans_out -= pcount;
1403 /* Shift newly-SACKed bytes from this skb to the immediately previous
1404 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1406 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1407 struct tcp_sacktag_state *state,
1408 unsigned int pcount, int shifted, int mss,
1411 struct tcp_sock *tp = tcp_sk(sk);
1412 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1413 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1414 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1418 /* Adjust counters and hints for the newly sacked sequence
1419 * range but discard the return value since prev is already
1420 * marked. We must tag the range first because the seq
1421 * advancement below implicitly advances
1422 * tcp_highest_sack_seq() when skb is highest_sack.
1424 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1425 start_seq, end_seq, dup_sack, pcount);
1427 if (skb == tp->lost_skb_hint)
1428 tp->lost_cnt_hint += pcount;
1430 TCP_SKB_CB(prev)->end_seq += shifted;
1431 TCP_SKB_CB(skb)->seq += shifted;
1433 skb_shinfo(prev)->gso_segs += pcount;
1434 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1435 skb_shinfo(skb)->gso_segs -= pcount;
1437 /* When we're adding to gso_segs == 1, gso_size will be zero,
1438 * in theory this shouldn't be necessary but as long as DSACK
1439 * code can come after this skb later on it's better to keep
1440 * setting gso_size to something.
1442 if (!skb_shinfo(prev)->gso_size) {
1443 skb_shinfo(prev)->gso_size = mss;
1444 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1447 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1448 if (skb_shinfo(skb)->gso_segs <= 1) {
1449 skb_shinfo(skb)->gso_size = 0;
1450 skb_shinfo(skb)->gso_type = 0;
1453 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1454 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1457 BUG_ON(!tcp_skb_pcount(skb));
1458 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1462 /* Whole SKB was eaten :-) */
1464 if (skb == tp->retransmit_skb_hint)
1465 tp->retransmit_skb_hint = prev;
1466 if (skb == tp->scoreboard_skb_hint)
1467 tp->scoreboard_skb_hint = prev;
1468 if (skb == tp->lost_skb_hint) {
1469 tp->lost_skb_hint = prev;
1470 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1473 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1474 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1475 TCP_SKB_CB(prev)->end_seq++;
1477 if (skb == tcp_highest_sack(sk))
1478 tcp_advance_highest_sack(sk, skb);
1480 tcp_unlink_write_queue(skb, sk);
1481 sk_wmem_free_skb(sk, skb);
1483 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1488 /* I wish gso_size would have a bit more sane initialization than
1489 * something-or-zero which complicates things
1491 static int tcp_skb_seglen(const struct sk_buff *skb)
1493 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1496 /* Shifting pages past head area doesn't work */
1497 static int skb_can_shift(const struct sk_buff *skb)
1499 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1502 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1505 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1506 struct tcp_sacktag_state *state,
1507 u32 start_seq, u32 end_seq,
1510 struct tcp_sock *tp = tcp_sk(sk);
1511 struct sk_buff *prev;
1517 if (!sk_can_gso(sk))
1520 /* Normally R but no L won't result in plain S */
1522 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1524 if (!skb_can_shift(skb))
1526 /* This frame is about to be dropped (was ACKed). */
1527 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1530 /* Can only happen with delayed DSACK + discard craziness */
1531 if (unlikely(skb == tcp_write_queue_head(sk)))
1533 prev = tcp_write_queue_prev(sk, skb);
1535 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1538 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1539 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1543 pcount = tcp_skb_pcount(skb);
1544 mss = tcp_skb_seglen(skb);
1546 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1547 * drop this restriction as unnecessary
1549 if (mss != tcp_skb_seglen(prev))
1552 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1554 /* CHECKME: This is non-MSS split case only?, this will
1555 * cause skipped skbs due to advancing loop btw, original
1556 * has that feature too
1558 if (tcp_skb_pcount(skb) <= 1)
1561 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1563 /* TODO: head merge to next could be attempted here
1564 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1565 * though it might not be worth of the additional hassle
1567 * ...we can probably just fallback to what was done
1568 * previously. We could try merging non-SACKed ones
1569 * as well but it probably isn't going to buy off
1570 * because later SACKs might again split them, and
1571 * it would make skb timestamp tracking considerably
1577 len = end_seq - TCP_SKB_CB(skb)->seq;
1579 BUG_ON(len > skb->len);
1581 /* MSS boundaries should be honoured or else pcount will
1582 * severely break even though it makes things bit trickier.
1583 * Optimize common case to avoid most of the divides
1585 mss = tcp_skb_mss(skb);
1587 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1588 * drop this restriction as unnecessary
1590 if (mss != tcp_skb_seglen(prev))
1595 } else if (len < mss) {
1603 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1604 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1607 if (!skb_shift(prev, skb, len))
1609 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1612 /* Hole filled allows collapsing with the next as well, this is very
1613 * useful when hole on every nth skb pattern happens
1615 if (prev == tcp_write_queue_tail(sk))
1617 skb = tcp_write_queue_next(sk, prev);
1619 if (!skb_can_shift(skb) ||
1620 (skb == tcp_send_head(sk)) ||
1621 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1622 (mss != tcp_skb_seglen(skb)))
1626 if (skb_shift(prev, skb, len)) {
1627 pcount += tcp_skb_pcount(skb);
1628 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1632 state->fack_count += pcount;
1639 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1643 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1644 struct tcp_sack_block *next_dup,
1645 struct tcp_sacktag_state *state,
1646 u32 start_seq, u32 end_seq,
1649 struct tcp_sock *tp = tcp_sk(sk);
1650 struct sk_buff *tmp;
1652 tcp_for_write_queue_from(skb, sk) {
1654 int dup_sack = dup_sack_in;
1656 if (skb == tcp_send_head(sk))
1659 /* queue is in-order => we can short-circuit the walk early */
1660 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1663 if ((next_dup != NULL) &&
1664 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1665 in_sack = tcp_match_skb_to_sack(sk, skb,
1666 next_dup->start_seq,
1672 /* skb reference here is a bit tricky to get right, since
1673 * shifting can eat and free both this skb and the next,
1674 * so not even _safe variant of the loop is enough.
1677 tmp = tcp_shift_skb_data(sk, skb, state,
1678 start_seq, end_seq, dup_sack);
1687 in_sack = tcp_match_skb_to_sack(sk, skb,
1693 if (unlikely(in_sack < 0))
1697 TCP_SKB_CB(skb)->sacked =
1700 TCP_SKB_CB(skb)->sacked,
1701 TCP_SKB_CB(skb)->seq,
1702 TCP_SKB_CB(skb)->end_seq,
1704 tcp_skb_pcount(skb));
1706 if (!before(TCP_SKB_CB(skb)->seq,
1707 tcp_highest_sack_seq(tp)))
1708 tcp_advance_highest_sack(sk, skb);
1711 state->fack_count += tcp_skb_pcount(skb);
1716 /* Avoid all extra work that is being done by sacktag while walking in
1719 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1720 struct tcp_sacktag_state *state,
1723 tcp_for_write_queue_from(skb, sk) {
1724 if (skb == tcp_send_head(sk))
1727 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1730 state->fack_count += tcp_skb_pcount(skb);
1735 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1737 struct tcp_sack_block *next_dup,
1738 struct tcp_sacktag_state *state,
1741 if (next_dup == NULL)
1744 if (before(next_dup->start_seq, skip_to_seq)) {
1745 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1746 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1747 next_dup->start_seq, next_dup->end_seq,
1754 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1756 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1760 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1763 const struct inet_connection_sock *icsk = inet_csk(sk);
1764 struct tcp_sock *tp = tcp_sk(sk);
1765 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1766 TCP_SKB_CB(ack_skb)->sacked);
1767 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1768 struct tcp_sack_block sp[TCP_NUM_SACKS];
1769 struct tcp_sack_block *cache;
1770 struct tcp_sacktag_state state;
1771 struct sk_buff *skb;
1772 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1774 int found_dup_sack = 0;
1776 int first_sack_index;
1779 state.reord = tp->packets_out;
1781 if (!tp->sacked_out) {
1782 if (WARN_ON(tp->fackets_out))
1783 tp->fackets_out = 0;
1784 tcp_highest_sack_reset(sk);
1787 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1788 num_sacks, prior_snd_una);
1790 state.flag |= FLAG_DSACKING_ACK;
1792 /* Eliminate too old ACKs, but take into
1793 * account more or less fresh ones, they can
1794 * contain valid SACK info.
1796 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1799 if (!tp->packets_out)
1803 first_sack_index = 0;
1804 for (i = 0; i < num_sacks; i++) {
1805 int dup_sack = !i && found_dup_sack;
1807 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1808 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1810 if (!tcp_is_sackblock_valid(tp, dup_sack,
1811 sp[used_sacks].start_seq,
1812 sp[used_sacks].end_seq)) {
1816 if (!tp->undo_marker)
1817 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1819 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1821 /* Don't count olds caused by ACK reordering */
1822 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1823 !after(sp[used_sacks].end_seq, tp->snd_una))
1825 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1828 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1830 first_sack_index = -1;
1834 /* Ignore very old stuff early */
1835 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1841 /* order SACK blocks to allow in order walk of the retrans queue */
1842 for (i = used_sacks - 1; i > 0; i--) {
1843 for (j = 0; j < i; j++) {
1844 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1845 swap(sp[j], sp[j + 1]);
1847 /* Track where the first SACK block goes to */
1848 if (j == first_sack_index)
1849 first_sack_index = j + 1;
1854 skb = tcp_write_queue_head(sk);
1855 state.fack_count = 0;
1858 if (!tp->sacked_out) {
1859 /* It's already past, so skip checking against it */
1860 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1862 cache = tp->recv_sack_cache;
1863 /* Skip empty blocks in at head of the cache */
1864 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1869 while (i < used_sacks) {
1870 u32 start_seq = sp[i].start_seq;
1871 u32 end_seq = sp[i].end_seq;
1872 int dup_sack = (found_dup_sack && (i == first_sack_index));
1873 struct tcp_sack_block *next_dup = NULL;
1875 if (found_dup_sack && ((i + 1) == first_sack_index))
1876 next_dup = &sp[i + 1];
1878 /* Event "B" in the comment above. */
1879 if (after(end_seq, tp->high_seq))
1880 state.flag |= FLAG_DATA_LOST;
1882 /* Skip too early cached blocks */
1883 while (tcp_sack_cache_ok(tp, cache) &&
1884 !before(start_seq, cache->end_seq))
1887 /* Can skip some work by looking recv_sack_cache? */
1888 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1889 after(end_seq, cache->start_seq)) {
1892 if (before(start_seq, cache->start_seq)) {
1893 skb = tcp_sacktag_skip(skb, sk, &state,
1895 skb = tcp_sacktag_walk(skb, sk, next_dup,
1902 /* Rest of the block already fully processed? */
1903 if (!after(end_seq, cache->end_seq))
1906 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1910 /* ...tail remains todo... */
1911 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1912 /* ...but better entrypoint exists! */
1913 skb = tcp_highest_sack(sk);
1916 state.fack_count = tp->fackets_out;
1921 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1922 /* Check overlap against next cached too (past this one already) */
1927 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1928 skb = tcp_highest_sack(sk);
1931 state.fack_count = tp->fackets_out;
1933 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1936 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1937 start_seq, end_seq, dup_sack);
1940 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1941 * due to in-order walk
1943 if (after(end_seq, tp->frto_highmark))
1944 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1949 /* Clear the head of the cache sack blocks so we can skip it next time */
1950 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1951 tp->recv_sack_cache[i].start_seq = 0;
1952 tp->recv_sack_cache[i].end_seq = 0;
1954 for (j = 0; j < used_sacks; j++)
1955 tp->recv_sack_cache[i++] = sp[j];
1957 tcp_mark_lost_retrans(sk);
1959 tcp_verify_left_out(tp);
1961 if ((state.reord < tp->fackets_out) &&
1962 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1963 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1964 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1968 #if FASTRETRANS_DEBUG > 0
1969 WARN_ON((int)tp->sacked_out < 0);
1970 WARN_ON((int)tp->lost_out < 0);
1971 WARN_ON((int)tp->retrans_out < 0);
1972 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1977 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1978 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1980 static int tcp_limit_reno_sacked(struct tcp_sock *tp)
1984 holes = max(tp->lost_out, 1U);
1985 holes = min(holes, tp->packets_out);
1987 if ((tp->sacked_out + holes) > tp->packets_out) {
1988 tp->sacked_out = tp->packets_out - holes;
1994 /* If we receive more dupacks than we expected counting segments
1995 * in assumption of absent reordering, interpret this as reordering.
1996 * The only another reason could be bug in receiver TCP.
1998 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2000 struct tcp_sock *tp = tcp_sk(sk);
2001 if (tcp_limit_reno_sacked(tp))
2002 tcp_update_reordering(sk, tp->packets_out + addend, 0);
2005 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2007 static void tcp_add_reno_sack(struct sock *sk)
2009 struct tcp_sock *tp = tcp_sk(sk);
2011 tcp_check_reno_reordering(sk, 0);
2012 tcp_verify_left_out(tp);
2015 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2017 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
2019 struct tcp_sock *tp = tcp_sk(sk);
2022 /* One ACK acked hole. The rest eat duplicate ACKs. */
2023 if (acked - 1 >= tp->sacked_out)
2026 tp->sacked_out -= acked - 1;
2028 tcp_check_reno_reordering(sk, acked);
2029 tcp_verify_left_out(tp);
2032 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2037 static int tcp_is_sackfrto(const struct tcp_sock *tp)
2039 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2042 /* F-RTO can only be used if TCP has never retransmitted anything other than
2043 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2045 int tcp_use_frto(struct sock *sk)
2047 const struct tcp_sock *tp = tcp_sk(sk);
2048 const struct inet_connection_sock *icsk = inet_csk(sk);
2049 struct sk_buff *skb;
2051 if (!sysctl_tcp_frto)
2054 /* MTU probe and F-RTO won't really play nicely along currently */
2055 if (icsk->icsk_mtup.probe_size)
2058 if (tcp_is_sackfrto(tp))
2061 /* Avoid expensive walking of rexmit queue if possible */
2062 if (tp->retrans_out > 1)
2065 skb = tcp_write_queue_head(sk);
2066 if (tcp_skb_is_last(sk, skb))
2068 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2069 tcp_for_write_queue_from(skb, sk) {
2070 if (skb == tcp_send_head(sk))
2072 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2074 /* Short-circuit when first non-SACKed skb has been checked */
2075 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2081 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2082 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2083 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2084 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2085 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2086 * bits are handled if the Loss state is really to be entered (in
2087 * tcp_enter_frto_loss).
2089 * Do like tcp_enter_loss() would; when RTO expires the second time it
2091 * "Reduce ssthresh if it has not yet been made inside this window."
2093 void tcp_enter_frto(struct sock *sk)
2095 const struct inet_connection_sock *icsk = inet_csk(sk);
2096 struct tcp_sock *tp = tcp_sk(sk);
2097 struct sk_buff *skb;
2099 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2100 tp->snd_una == tp->high_seq ||
2101 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2102 !icsk->icsk_retransmits)) {
2103 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2104 /* Our state is too optimistic in ssthresh() call because cwnd
2105 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2106 * recovery has not yet completed. Pattern would be this: RTO,
2107 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2109 * RFC4138 should be more specific on what to do, even though
2110 * RTO is quite unlikely to occur after the first Cumulative ACK
2111 * due to back-off and complexity of triggering events ...
2113 if (tp->frto_counter) {
2115 stored_cwnd = tp->snd_cwnd;
2117 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2118 tp->snd_cwnd = stored_cwnd;
2120 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2122 /* ... in theory, cong.control module could do "any tricks" in
2123 * ssthresh(), which means that ca_state, lost bits and lost_out
2124 * counter would have to be faked before the call occurs. We
2125 * consider that too expensive, unlikely and hacky, so modules
2126 * using these in ssthresh() must deal these incompatibility
2127 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2129 tcp_ca_event(sk, CA_EVENT_FRTO);
2132 tp->undo_marker = tp->snd_una;
2133 tp->undo_retrans = 0;
2135 skb = tcp_write_queue_head(sk);
2136 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2137 tp->undo_marker = 0;
2138 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2139 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2140 tp->retrans_out -= tcp_skb_pcount(skb);
2142 tcp_verify_left_out(tp);
2144 /* Too bad if TCP was application limited */
2145 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2147 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2148 * The last condition is necessary at least in tp->frto_counter case.
2150 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2151 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2152 after(tp->high_seq, tp->snd_una)) {
2153 tp->frto_highmark = tp->high_seq;
2155 tp->frto_highmark = tp->snd_nxt;
2157 tcp_set_ca_state(sk, TCP_CA_Disorder);
2158 tp->high_seq = tp->snd_nxt;
2159 tp->frto_counter = 1;
2162 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2163 * which indicates that we should follow the traditional RTO recovery,
2164 * i.e. mark everything lost and do go-back-N retransmission.
2166 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2168 struct tcp_sock *tp = tcp_sk(sk);
2169 struct sk_buff *skb;
2172 tp->retrans_out = 0;
2173 if (tcp_is_reno(tp))
2174 tcp_reset_reno_sack(tp);
2176 tcp_for_write_queue(skb, sk) {
2177 if (skb == tcp_send_head(sk))
2180 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2182 * Count the retransmission made on RTO correctly (only when
2183 * waiting for the first ACK and did not get it)...
2185 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2186 /* For some reason this R-bit might get cleared? */
2187 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2188 tp->retrans_out += tcp_skb_pcount(skb);
2189 /* ...enter this if branch just for the first segment */
2190 flag |= FLAG_DATA_ACKED;
2192 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2193 tp->undo_marker = 0;
2194 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2197 /* Marking forward transmissions that were made after RTO lost
2198 * can cause unnecessary retransmissions in some scenarios,
2199 * SACK blocks will mitigate that in some but not in all cases.
2200 * We used to not mark them but it was causing break-ups with
2201 * receivers that do only in-order receival.
2203 * TODO: we could detect presence of such receiver and select
2204 * different behavior per flow.
2206 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2207 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2208 tp->lost_out += tcp_skb_pcount(skb);
2209 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2212 tcp_verify_left_out(tp);
2214 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2215 tp->snd_cwnd_cnt = 0;
2216 tp->snd_cwnd_stamp = tcp_time_stamp;
2217 tp->frto_counter = 0;
2218 tp->bytes_acked = 0;
2220 tp->reordering = min_t(unsigned int, tp->reordering,
2221 sysctl_tcp_reordering);
2222 tcp_set_ca_state(sk, TCP_CA_Loss);
2223 tp->high_seq = tp->snd_nxt;
2224 TCP_ECN_queue_cwr(tp);
2226 tcp_clear_all_retrans_hints(tp);
2229 static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2231 tp->retrans_out = 0;
2234 tp->undo_marker = 0;
2235 tp->undo_retrans = 0;
2238 void tcp_clear_retrans(struct tcp_sock *tp)
2240 tcp_clear_retrans_partial(tp);
2242 tp->fackets_out = 0;
2246 /* Enter Loss state. If "how" is not zero, forget all SACK information
2247 * and reset tags completely, otherwise preserve SACKs. If receiver
2248 * dropped its ofo queue, we will know this due to reneging detection.
2250 void tcp_enter_loss(struct sock *sk, int how)
2252 const struct inet_connection_sock *icsk = inet_csk(sk);
2253 struct tcp_sock *tp = tcp_sk(sk);
2254 struct sk_buff *skb;
2256 /* Reduce ssthresh if it has not yet been made inside this window. */
2257 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2258 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2259 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2260 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2261 tcp_ca_event(sk, CA_EVENT_LOSS);
2264 tp->snd_cwnd_cnt = 0;
2265 tp->snd_cwnd_stamp = tcp_time_stamp;
2267 tp->bytes_acked = 0;
2268 tcp_clear_retrans_partial(tp);
2270 if (tcp_is_reno(tp))
2271 tcp_reset_reno_sack(tp);
2273 tp->undo_marker = tp->snd_una;
2276 tp->fackets_out = 0;
2278 tcp_clear_all_retrans_hints(tp);
2280 tcp_for_write_queue(skb, sk) {
2281 if (skb == tcp_send_head(sk))
2284 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2285 tp->undo_marker = 0;
2286 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2287 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2288 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2289 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2290 tp->lost_out += tcp_skb_pcount(skb);
2291 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2294 tcp_verify_left_out(tp);
2296 tp->reordering = min_t(unsigned int, tp->reordering,
2297 sysctl_tcp_reordering);
2298 tcp_set_ca_state(sk, TCP_CA_Loss);
2299 tp->high_seq = tp->snd_nxt;
2300 TCP_ECN_queue_cwr(tp);
2301 /* Abort F-RTO algorithm if one is in progress */
2302 tp->frto_counter = 0;
2305 /* If ACK arrived pointing to a remembered SACK, it means that our
2306 * remembered SACKs do not reflect real state of receiver i.e.
2307 * receiver _host_ is heavily congested (or buggy).
2309 * Do processing similar to RTO timeout.
2311 static int tcp_check_sack_reneging(struct sock *sk, int flag)
2313 if (flag & FLAG_SACK_RENEGING) {
2314 struct inet_connection_sock *icsk = inet_csk(sk);
2315 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2317 tcp_enter_loss(sk, 1);
2318 icsk->icsk_retransmits++;
2319 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2320 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2321 icsk->icsk_rto, TCP_RTO_MAX);
2327 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2329 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2332 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2333 * counter when SACK is enabled (without SACK, sacked_out is used for
2336 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2337 * segments up to the highest received SACK block so far and holes in
2340 * With reordering, holes may still be in flight, so RFC3517 recovery
2341 * uses pure sacked_out (total number of SACKed segments) even though
2342 * it violates the RFC that uses duplicate ACKs, often these are equal
2343 * but when e.g. out-of-window ACKs or packet duplication occurs,
2344 * they differ. Since neither occurs due to loss, TCP should really
2347 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2349 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2352 static inline int tcp_skb_timedout(const struct sock *sk,
2353 const struct sk_buff *skb)
2355 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2358 static inline int tcp_head_timedout(const struct sock *sk)
2360 const struct tcp_sock *tp = tcp_sk(sk);
2362 return tp->packets_out &&
2363 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2366 /* Linux NewReno/SACK/FACK/ECN state machine.
2367 * --------------------------------------
2369 * "Open" Normal state, no dubious events, fast path.
2370 * "Disorder" In all the respects it is "Open",
2371 * but requires a bit more attention. It is entered when
2372 * we see some SACKs or dupacks. It is split of "Open"
2373 * mainly to move some processing from fast path to slow one.
2374 * "CWR" CWND was reduced due to some Congestion Notification event.
2375 * It can be ECN, ICMP source quench, local device congestion.
2376 * "Recovery" CWND was reduced, we are fast-retransmitting.
2377 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2379 * tcp_fastretrans_alert() is entered:
2380 * - each incoming ACK, if state is not "Open"
2381 * - when arrived ACK is unusual, namely:
2386 * Counting packets in flight is pretty simple.
2388 * in_flight = packets_out - left_out + retrans_out
2390 * packets_out is SND.NXT-SND.UNA counted in packets.
2392 * retrans_out is number of retransmitted segments.
2394 * left_out is number of segments left network, but not ACKed yet.
2396 * left_out = sacked_out + lost_out
2398 * sacked_out: Packets, which arrived to receiver out of order
2399 * and hence not ACKed. With SACKs this number is simply
2400 * amount of SACKed data. Even without SACKs
2401 * it is easy to give pretty reliable estimate of this number,
2402 * counting duplicate ACKs.
2404 * lost_out: Packets lost by network. TCP has no explicit
2405 * "loss notification" feedback from network (for now).
2406 * It means that this number can be only _guessed_.
2407 * Actually, it is the heuristics to predict lossage that
2408 * distinguishes different algorithms.
2410 * F.e. after RTO, when all the queue is considered as lost,
2411 * lost_out = packets_out and in_flight = retrans_out.
2413 * Essentially, we have now two algorithms counting
2416 * FACK: It is the simplest heuristics. As soon as we decided
2417 * that something is lost, we decide that _all_ not SACKed
2418 * packets until the most forward SACK are lost. I.e.
2419 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2420 * It is absolutely correct estimate, if network does not reorder
2421 * packets. And it loses any connection to reality when reordering
2422 * takes place. We use FACK by default until reordering
2423 * is suspected on the path to this destination.
2425 * NewReno: when Recovery is entered, we assume that one segment
2426 * is lost (classic Reno). While we are in Recovery and
2427 * a partial ACK arrives, we assume that one more packet
2428 * is lost (NewReno). This heuristics are the same in NewReno
2431 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2432 * deflation etc. CWND is real congestion window, never inflated, changes
2433 * only according to classic VJ rules.
2435 * Really tricky (and requiring careful tuning) part of algorithm
2436 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2437 * The first determines the moment _when_ we should reduce CWND and,
2438 * hence, slow down forward transmission. In fact, it determines the moment
2439 * when we decide that hole is caused by loss, rather than by a reorder.
2441 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2442 * holes, caused by lost packets.
2444 * And the most logically complicated part of algorithm is undo
2445 * heuristics. We detect false retransmits due to both too early
2446 * fast retransmit (reordering) and underestimated RTO, analyzing
2447 * timestamps and D-SACKs. When we detect that some segments were
2448 * retransmitted by mistake and CWND reduction was wrong, we undo
2449 * window reduction and abort recovery phase. This logic is hidden
2450 * inside several functions named tcp_try_undo_<something>.
2453 /* This function decides, when we should leave Disordered state
2454 * and enter Recovery phase, reducing congestion window.
2456 * Main question: may we further continue forward transmission
2457 * with the same cwnd?
2459 static int tcp_time_to_recover(struct sock *sk)
2461 struct tcp_sock *tp = tcp_sk(sk);
2464 /* Do not perform any recovery during F-RTO algorithm */
2465 if (tp->frto_counter)
2468 /* Trick#1: The loss is proven. */
2472 /* Not-A-Trick#2 : Classic rule... */
2473 if (tcp_dupack_heuristics(tp) > tp->reordering)
2476 /* Trick#3 : when we use RFC2988 timer restart, fast
2477 * retransmit can be triggered by timeout of queue head.
2479 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2482 /* Trick#4: It is still not OK... But will it be useful to delay
2485 packets_out = tp->packets_out;
2486 if (packets_out <= tp->reordering &&
2487 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2488 !tcp_may_send_now(sk)) {
2489 /* We have nothing to send. This connection is limited
2490 * either by receiver window or by application.
2495 /* If a thin stream is detected, retransmit after first
2496 * received dupack. Employ only if SACK is supported in order
2497 * to avoid possible corner-case series of spurious retransmissions
2498 * Use only if there are no unsent data.
2500 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2501 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2502 tcp_is_sack(tp) && !tcp_send_head(sk))
2508 /* New heuristics: it is possible only after we switched to restart timer
2509 * each time when something is ACKed. Hence, we can detect timed out packets
2510 * during fast retransmit without falling to slow start.
2512 * Usefulness of this as is very questionable, since we should know which of
2513 * the segments is the next to timeout which is relatively expensive to find
2514 * in general case unless we add some data structure just for that. The
2515 * current approach certainly won't find the right one too often and when it
2516 * finally does find _something_ it usually marks large part of the window
2517 * right away (because a retransmission with a larger timestamp blocks the
2518 * loop from advancing). -ij
2520 static void tcp_timeout_skbs(struct sock *sk)
2522 struct tcp_sock *tp = tcp_sk(sk);
2523 struct sk_buff *skb;
2525 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2528 skb = tp->scoreboard_skb_hint;
2529 if (tp->scoreboard_skb_hint == NULL)
2530 skb = tcp_write_queue_head(sk);
2532 tcp_for_write_queue_from(skb, sk) {
2533 if (skb == tcp_send_head(sk))
2535 if (!tcp_skb_timedout(sk, skb))
2538 tcp_skb_mark_lost(tp, skb);
2541 tp->scoreboard_skb_hint = skb;
2543 tcp_verify_left_out(tp);
2546 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2547 * is against sacked "cnt", otherwise it's against facked "cnt"
2549 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2551 struct tcp_sock *tp = tcp_sk(sk);
2552 struct sk_buff *skb;
2557 WARN_ON(packets > tp->packets_out);
2558 if (tp->lost_skb_hint) {
2559 skb = tp->lost_skb_hint;
2560 cnt = tp->lost_cnt_hint;
2561 /* Head already handled? */
2562 if (mark_head && skb != tcp_write_queue_head(sk))
2565 skb = tcp_write_queue_head(sk);
2569 tcp_for_write_queue_from(skb, sk) {
2570 if (skb == tcp_send_head(sk))
2572 /* TODO: do this better */
2573 /* this is not the most efficient way to do this... */
2574 tp->lost_skb_hint = skb;
2575 tp->lost_cnt_hint = cnt;
2577 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2581 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2582 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2583 cnt += tcp_skb_pcount(skb);
2585 if (cnt > packets) {
2586 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2587 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2588 (oldcnt >= packets))
2591 mss = skb_shinfo(skb)->gso_size;
2592 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2598 tcp_skb_mark_lost(tp, skb);
2603 tcp_verify_left_out(tp);
2606 /* Account newly detected lost packet(s) */
2608 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2610 struct tcp_sock *tp = tcp_sk(sk);
2612 if (tcp_is_reno(tp)) {
2613 tcp_mark_head_lost(sk, 1, 1);
2614 } else if (tcp_is_fack(tp)) {
2615 int lost = tp->fackets_out - tp->reordering;
2618 tcp_mark_head_lost(sk, lost, 0);
2620 int sacked_upto = tp->sacked_out - tp->reordering;
2621 if (sacked_upto >= 0)
2622 tcp_mark_head_lost(sk, sacked_upto, 0);
2623 else if (fast_rexmit)
2624 tcp_mark_head_lost(sk, 1, 1);
2627 tcp_timeout_skbs(sk);
2630 /* CWND moderation, preventing bursts due to too big ACKs
2631 * in dubious situations.
2633 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2635 tp->snd_cwnd = min(tp->snd_cwnd,
2636 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2637 tp->snd_cwnd_stamp = tcp_time_stamp;
2640 /* Lower bound on congestion window is slow start threshold
2641 * unless congestion avoidance choice decides to overide it.
2643 static inline u32 tcp_cwnd_min(const struct sock *sk)
2645 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2647 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2650 /* Decrease cwnd each second ack. */
2651 static void tcp_cwnd_down(struct sock *sk, int flag)
2653 struct tcp_sock *tp = tcp_sk(sk);
2654 int decr = tp->snd_cwnd_cnt + 1;
2656 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2657 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2658 tp->snd_cwnd_cnt = decr & 1;
2661 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2662 tp->snd_cwnd -= decr;
2664 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2665 tp->snd_cwnd_stamp = tcp_time_stamp;
2669 /* Nothing was retransmitted or returned timestamp is less
2670 * than timestamp of the first retransmission.
2672 static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2674 return !tp->retrans_stamp ||
2675 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2676 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2679 /* Undo procedures. */
2681 #if FASTRETRANS_DEBUG > 1
2682 static void DBGUNDO(struct sock *sk, const char *msg)
2684 struct tcp_sock *tp = tcp_sk(sk);
2685 struct inet_sock *inet = inet_sk(sk);
2687 if (sk->sk_family == AF_INET) {
2688 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2690 &inet->inet_daddr, ntohs(inet->inet_dport),
2691 tp->snd_cwnd, tcp_left_out(tp),
2692 tp->snd_ssthresh, tp->prior_ssthresh,
2695 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2696 else if (sk->sk_family == AF_INET6) {
2697 struct ipv6_pinfo *np = inet6_sk(sk);
2698 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2700 &np->daddr, ntohs(inet->inet_dport),
2701 tp->snd_cwnd, tcp_left_out(tp),
2702 tp->snd_ssthresh, tp->prior_ssthresh,
2708 #define DBGUNDO(x...) do { } while (0)
2711 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2713 struct tcp_sock *tp = tcp_sk(sk);
2715 if (tp->prior_ssthresh) {
2716 const struct inet_connection_sock *icsk = inet_csk(sk);
2718 if (icsk->icsk_ca_ops->undo_cwnd)
2719 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2721 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2723 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2724 tp->snd_ssthresh = tp->prior_ssthresh;
2725 TCP_ECN_withdraw_cwr(tp);
2728 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2730 tp->snd_cwnd_stamp = tcp_time_stamp;
2733 static inline int tcp_may_undo(const struct tcp_sock *tp)
2735 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2738 /* People celebrate: "We love our President!" */
2739 static int tcp_try_undo_recovery(struct sock *sk)
2741 struct tcp_sock *tp = tcp_sk(sk);
2743 if (tcp_may_undo(tp)) {
2746 /* Happy end! We did not retransmit anything
2747 * or our original transmission succeeded.
2749 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2750 tcp_undo_cwr(sk, true);
2751 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2752 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2754 mib_idx = LINUX_MIB_TCPFULLUNDO;
2756 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2757 tp->undo_marker = 0;
2759 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2760 /* Hold old state until something *above* high_seq
2761 * is ACKed. For Reno it is MUST to prevent false
2762 * fast retransmits (RFC2582). SACK TCP is safe. */
2763 tcp_moderate_cwnd(tp);
2766 tcp_set_ca_state(sk, TCP_CA_Open);
2770 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2771 static void tcp_try_undo_dsack(struct sock *sk)
2773 struct tcp_sock *tp = tcp_sk(sk);
2775 if (tp->undo_marker && !tp->undo_retrans) {
2776 DBGUNDO(sk, "D-SACK");
2777 tcp_undo_cwr(sk, true);
2778 tp->undo_marker = 0;
2779 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2783 /* We can clear retrans_stamp when there are no retransmissions in the
2784 * window. It would seem that it is trivially available for us in
2785 * tp->retrans_out, however, that kind of assumptions doesn't consider
2786 * what will happen if errors occur when sending retransmission for the
2787 * second time. ...It could the that such segment has only
2788 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2789 * the head skb is enough except for some reneging corner cases that
2790 * are not worth the effort.
2792 * Main reason for all this complexity is the fact that connection dying
2793 * time now depends on the validity of the retrans_stamp, in particular,
2794 * that successive retransmissions of a segment must not advance
2795 * retrans_stamp under any conditions.
2797 static int tcp_any_retrans_done(const struct sock *sk)
2799 const struct tcp_sock *tp = tcp_sk(sk);
2800 struct sk_buff *skb;
2802 if (tp->retrans_out)
2805 skb = tcp_write_queue_head(sk);
2806 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2812 /* Undo during fast recovery after partial ACK. */
2814 static int tcp_try_undo_partial(struct sock *sk, int acked)
2816 struct tcp_sock *tp = tcp_sk(sk);
2817 /* Partial ACK arrived. Force Hoe's retransmit. */
2818 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2820 if (tcp_may_undo(tp)) {
2821 /* Plain luck! Hole if filled with delayed
2822 * packet, rather than with a retransmit.
2824 if (!tcp_any_retrans_done(sk))
2825 tp->retrans_stamp = 0;
2827 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2830 tcp_undo_cwr(sk, false);
2831 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2833 /* So... Do not make Hoe's retransmit yet.
2834 * If the first packet was delayed, the rest
2835 * ones are most probably delayed as well.
2842 /* Undo during loss recovery after partial ACK. */
2843 static int tcp_try_undo_loss(struct sock *sk)
2845 struct tcp_sock *tp = tcp_sk(sk);
2847 if (tcp_may_undo(tp)) {
2848 struct sk_buff *skb;
2849 tcp_for_write_queue(skb, sk) {
2850 if (skb == tcp_send_head(sk))
2852 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2855 tcp_clear_all_retrans_hints(tp);
2857 DBGUNDO(sk, "partial loss");
2859 tcp_undo_cwr(sk, true);
2860 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2861 inet_csk(sk)->icsk_retransmits = 0;
2862 tp->undo_marker = 0;
2863 if (tcp_is_sack(tp))
2864 tcp_set_ca_state(sk, TCP_CA_Open);
2870 static inline void tcp_complete_cwr(struct sock *sk)
2872 struct tcp_sock *tp = tcp_sk(sk);
2874 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2875 if (tp->undo_marker) {
2876 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
2877 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2878 tp->snd_cwnd_stamp = tcp_time_stamp;
2879 } else if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH) {
2880 /* PRR algorithm. */
2881 tp->snd_cwnd = tp->snd_ssthresh;
2882 tp->snd_cwnd_stamp = tcp_time_stamp;
2885 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2888 static void tcp_try_keep_open(struct sock *sk)
2890 struct tcp_sock *tp = tcp_sk(sk);
2891 int state = TCP_CA_Open;
2893 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)
2894 state = TCP_CA_Disorder;
2896 if (inet_csk(sk)->icsk_ca_state != state) {
2897 tcp_set_ca_state(sk, state);
2898 tp->high_seq = tp->snd_nxt;
2902 static void tcp_try_to_open(struct sock *sk, int flag)
2904 struct tcp_sock *tp = tcp_sk(sk);
2906 tcp_verify_left_out(tp);
2908 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2909 tp->retrans_stamp = 0;
2911 if (flag & FLAG_ECE)
2912 tcp_enter_cwr(sk, 1);
2914 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2915 tcp_try_keep_open(sk);
2916 tcp_moderate_cwnd(tp);
2918 tcp_cwnd_down(sk, flag);
2922 static void tcp_mtup_probe_failed(struct sock *sk)
2924 struct inet_connection_sock *icsk = inet_csk(sk);
2926 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2927 icsk->icsk_mtup.probe_size = 0;
2930 static void tcp_mtup_probe_success(struct sock *sk)
2932 struct tcp_sock *tp = tcp_sk(sk);
2933 struct inet_connection_sock *icsk = inet_csk(sk);
2935 /* FIXME: breaks with very large cwnd */
2936 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2937 tp->snd_cwnd = tp->snd_cwnd *
2938 tcp_mss_to_mtu(sk, tp->mss_cache) /
2939 icsk->icsk_mtup.probe_size;
2940 tp->snd_cwnd_cnt = 0;
2941 tp->snd_cwnd_stamp = tcp_time_stamp;
2942 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2944 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2945 icsk->icsk_mtup.probe_size = 0;
2946 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2949 /* Do a simple retransmit without using the backoff mechanisms in
2950 * tcp_timer. This is used for path mtu discovery.
2951 * The socket is already locked here.
2953 void tcp_simple_retransmit(struct sock *sk)
2955 const struct inet_connection_sock *icsk = inet_csk(sk);
2956 struct tcp_sock *tp = tcp_sk(sk);
2957 struct sk_buff *skb;
2958 unsigned int mss = tcp_current_mss(sk);
2959 u32 prior_lost = tp->lost_out;
2961 tcp_for_write_queue(skb, sk) {
2962 if (skb == tcp_send_head(sk))
2964 if (tcp_skb_seglen(skb) > mss &&
2965 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2966 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2967 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2968 tp->retrans_out -= tcp_skb_pcount(skb);
2970 tcp_skb_mark_lost_uncond_verify(tp, skb);
2974 tcp_clear_retrans_hints_partial(tp);
2976 if (prior_lost == tp->lost_out)
2979 if (tcp_is_reno(tp))
2980 tcp_limit_reno_sacked(tp);
2982 tcp_verify_left_out(tp);
2984 /* Don't muck with the congestion window here.
2985 * Reason is that we do not increase amount of _data_
2986 * in network, but units changed and effective
2987 * cwnd/ssthresh really reduced now.
2989 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2990 tp->high_seq = tp->snd_nxt;
2991 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2992 tp->prior_ssthresh = 0;
2993 tp->undo_marker = 0;
2994 tcp_set_ca_state(sk, TCP_CA_Loss);
2996 tcp_xmit_retransmit_queue(sk);
2998 EXPORT_SYMBOL(tcp_simple_retransmit);
3000 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
3001 * (proportional rate reduction with slow start reduction bound) as described in
3002 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
3003 * It computes the number of packets to send (sndcnt) based on packets newly
3005 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
3006 * cwnd reductions across a full RTT.
3007 * 2) If packets in flight is lower than ssthresh (such as due to excess
3008 * losses and/or application stalls), do not perform any further cwnd
3009 * reductions, but instead slow start up to ssthresh.
3011 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
3012 int fast_rexmit, int flag)
3014 struct tcp_sock *tp = tcp_sk(sk);
3016 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
3018 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
3019 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
3021 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
3023 sndcnt = min_t(int, delta,
3024 max_t(int, tp->prr_delivered - tp->prr_out,
3025 newly_acked_sacked) + 1);
3028 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
3029 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3032 /* Process an event, which can update packets-in-flight not trivially.
3033 * Main goal of this function is to calculate new estimate for left_out,
3034 * taking into account both packets sitting in receiver's buffer and
3035 * packets lost by network.
3037 * Besides that it does CWND reduction, when packet loss is detected
3038 * and changes state of machine.
3040 * It does _not_ decide what to send, it is made in function
3041 * tcp_xmit_retransmit_queue().
3043 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3044 int newly_acked_sacked, int flag)
3046 struct inet_connection_sock *icsk = inet_csk(sk);
3047 struct tcp_sock *tp = tcp_sk(sk);
3048 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3049 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3050 (tcp_fackets_out(tp) > tp->reordering));
3051 int fast_rexmit = 0, mib_idx;
3053 if (WARN_ON(!tp->packets_out && tp->sacked_out))
3055 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3056 tp->fackets_out = 0;
3058 /* Now state machine starts.
3059 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3060 if (flag & FLAG_ECE)
3061 tp->prior_ssthresh = 0;
3063 /* B. In all the states check for reneging SACKs. */
3064 if (tcp_check_sack_reneging(sk, flag))
3067 /* C. Process data loss notification, provided it is valid. */
3068 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
3069 before(tp->snd_una, tp->high_seq) &&
3070 icsk->icsk_ca_state != TCP_CA_Open &&
3071 tp->fackets_out > tp->reordering) {
3072 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
3073 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
3076 /* D. Check consistency of the current state. */
3077 tcp_verify_left_out(tp);
3079 /* E. Check state exit conditions. State can be terminated
3080 * when high_seq is ACKed. */
3081 if (icsk->icsk_ca_state == TCP_CA_Open) {
3082 WARN_ON(tp->retrans_out != 0);
3083 tp->retrans_stamp = 0;
3084 } else if (!before(tp->snd_una, tp->high_seq)) {
3085 switch (icsk->icsk_ca_state) {
3087 icsk->icsk_retransmits = 0;
3088 if (tcp_try_undo_recovery(sk))
3093 /* CWR is to be held something *above* high_seq
3094 * is ACKed for CWR bit to reach receiver. */
3095 if (tp->snd_una != tp->high_seq) {
3096 tcp_complete_cwr(sk);
3097 tcp_set_ca_state(sk, TCP_CA_Open);
3101 case TCP_CA_Disorder:
3102 tcp_try_undo_dsack(sk);
3103 if (!tp->undo_marker ||
3104 /* For SACK case do not Open to allow to undo
3105 * catching for all duplicate ACKs. */
3106 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
3107 tp->undo_marker = 0;
3108 tcp_set_ca_state(sk, TCP_CA_Open);
3112 case TCP_CA_Recovery:
3113 if (tcp_is_reno(tp))
3114 tcp_reset_reno_sack(tp);
3115 if (tcp_try_undo_recovery(sk))
3117 tcp_complete_cwr(sk);
3122 /* F. Process state. */
3123 switch (icsk->icsk_ca_state) {
3124 case TCP_CA_Recovery:
3125 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3126 if (tcp_is_reno(tp) && is_dupack)
3127 tcp_add_reno_sack(sk);
3129 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3132 if (flag & FLAG_DATA_ACKED)
3133 icsk->icsk_retransmits = 0;
3134 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3135 tcp_reset_reno_sack(tp);
3136 if (!tcp_try_undo_loss(sk)) {
3137 tcp_moderate_cwnd(tp);
3138 tcp_xmit_retransmit_queue(sk);
3141 if (icsk->icsk_ca_state != TCP_CA_Open)
3143 /* Loss is undone; fall through to processing in Open state. */
3145 if (tcp_is_reno(tp)) {
3146 if (flag & FLAG_SND_UNA_ADVANCED)
3147 tcp_reset_reno_sack(tp);
3149 tcp_add_reno_sack(sk);
3152 if (icsk->icsk_ca_state == TCP_CA_Disorder)
3153 tcp_try_undo_dsack(sk);
3155 if (!tcp_time_to_recover(sk)) {
3156 tcp_try_to_open(sk, flag);
3160 /* MTU probe failure: don't reduce cwnd */
3161 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3162 icsk->icsk_mtup.probe_size &&
3163 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3164 tcp_mtup_probe_failed(sk);
3165 /* Restores the reduction we did in tcp_mtup_probe() */
3167 tcp_simple_retransmit(sk);
3171 /* Otherwise enter Recovery state */
3173 if (tcp_is_reno(tp))
3174 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3176 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3178 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3180 tp->high_seq = tp->snd_nxt;
3181 tp->prior_ssthresh = 0;
3182 tp->undo_marker = tp->snd_una;
3183 tp->undo_retrans = tp->retrans_out;
3185 if (icsk->icsk_ca_state < TCP_CA_CWR) {
3186 if (!(flag & FLAG_ECE))
3187 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3188 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
3189 TCP_ECN_queue_cwr(tp);
3192 tp->bytes_acked = 0;
3193 tp->snd_cwnd_cnt = 0;
3194 tp->prior_cwnd = tp->snd_cwnd;
3195 tp->prr_delivered = 0;
3197 tcp_set_ca_state(sk, TCP_CA_Recovery);
3201 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3202 tcp_update_scoreboard(sk, fast_rexmit);
3203 tp->prr_delivered += newly_acked_sacked;
3204 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3205 tcp_xmit_retransmit_queue(sk);
3208 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3210 tcp_rtt_estimator(sk, seq_rtt);
3212 inet_csk(sk)->icsk_backoff = 0;
3214 EXPORT_SYMBOL(tcp_valid_rtt_meas);
3216 /* Read draft-ietf-tcplw-high-performance before mucking
3217 * with this code. (Supersedes RFC1323)
3219 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3221 /* RTTM Rule: A TSecr value received in a segment is used to
3222 * update the averaged RTT measurement only if the segment
3223 * acknowledges some new data, i.e., only if it advances the
3224 * left edge of the send window.
3226 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3227 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3229 * Changed: reset backoff as soon as we see the first valid sample.
3230 * If we do not, we get strongly overestimated rto. With timestamps
3231 * samples are accepted even from very old segments: f.e., when rtt=1
3232 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3233 * answer arrives rto becomes 120 seconds! If at least one of segments
3234 * in window is lost... Voila. --ANK (010210)
3236 struct tcp_sock *tp = tcp_sk(sk);
3238 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3241 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3243 /* We don't have a timestamp. Can only use
3244 * packets that are not retransmitted to determine
3245 * rtt estimates. Also, we must not reset the
3246 * backoff for rto until we get a non-retransmitted
3247 * packet. This allows us to deal with a situation
3248 * where the network delay has increased suddenly.
3249 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3252 if (flag & FLAG_RETRANS_DATA_ACKED)
3255 tcp_valid_rtt_meas(sk, seq_rtt);
3258 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3261 const struct tcp_sock *tp = tcp_sk(sk);
3262 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3263 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3264 tcp_ack_saw_tstamp(sk, flag);
3265 else if (seq_rtt >= 0)
3266 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3269 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3271 const struct inet_connection_sock *icsk = inet_csk(sk);
3272 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3273 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3276 /* Restart timer after forward progress on connection.
3277 * RFC2988 recommends to restart timer to now+rto.
3279 static void tcp_rearm_rto(struct sock *sk)
3281 const struct tcp_sock *tp = tcp_sk(sk);
3283 if (!tp->packets_out) {
3284 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3286 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3287 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3291 /* If we get here, the whole TSO packet has not been acked. */
3292 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3294 struct tcp_sock *tp = tcp_sk(sk);
3297 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3299 packets_acked = tcp_skb_pcount(skb);
3300 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3302 packets_acked -= tcp_skb_pcount(skb);
3304 if (packets_acked) {
3305 BUG_ON(tcp_skb_pcount(skb) == 0);
3306 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3309 return packets_acked;
3312 /* Remove acknowledged frames from the retransmission queue. If our packet
3313 * is before the ack sequence we can discard it as it's confirmed to have
3314 * arrived at the other end.
3316 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3319 struct tcp_sock *tp = tcp_sk(sk);
3320 const struct inet_connection_sock *icsk = inet_csk(sk);
3321 struct sk_buff *skb;
3322 u32 now = tcp_time_stamp;
3323 int fully_acked = 1;
3326 u32 reord = tp->packets_out;
3327 u32 prior_sacked = tp->sacked_out;
3329 s32 ca_seq_rtt = -1;
3330 ktime_t last_ackt = net_invalid_timestamp();
3332 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3333 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3335 u8 sacked = scb->sacked;
3337 /* Determine how many packets and what bytes were acked, tso and else */
3338 if (after(scb->end_seq, tp->snd_una)) {
3339 if (tcp_skb_pcount(skb) == 1 ||
3340 !after(tp->snd_una, scb->seq))
3343 acked_pcount = tcp_tso_acked(sk, skb);
3349 acked_pcount = tcp_skb_pcount(skb);
3352 if (sacked & TCPCB_RETRANS) {
3353 if (sacked & TCPCB_SACKED_RETRANS)
3354 tp->retrans_out -= acked_pcount;
3355 flag |= FLAG_RETRANS_DATA_ACKED;
3358 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3359 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3361 ca_seq_rtt = now - scb->when;
3362 last_ackt = skb->tstamp;
3364 seq_rtt = ca_seq_rtt;
3366 if (!(sacked & TCPCB_SACKED_ACKED))
3367 reord = min(pkts_acked, reord);
3370 if (sacked & TCPCB_SACKED_ACKED)
3371 tp->sacked_out -= acked_pcount;
3372 if (sacked & TCPCB_LOST)
3373 tp->lost_out -= acked_pcount;
3375 tp->packets_out -= acked_pcount;
3376 pkts_acked += acked_pcount;
3378 /* Initial outgoing SYN's get put onto the write_queue
3379 * just like anything else we transmit. It is not
3380 * true data, and if we misinform our callers that
3381 * this ACK acks real data, we will erroneously exit
3382 * connection startup slow start one packet too
3383 * quickly. This is severely frowned upon behavior.
3385 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3386 flag |= FLAG_DATA_ACKED;
3388 flag |= FLAG_SYN_ACKED;
3389 tp->retrans_stamp = 0;
3395 tcp_unlink_write_queue(skb, sk);
3396 sk_wmem_free_skb(sk, skb);
3397 tp->scoreboard_skb_hint = NULL;
3398 if (skb == tp->retransmit_skb_hint)
3399 tp->retransmit_skb_hint = NULL;
3400 if (skb == tp->lost_skb_hint)
3401 tp->lost_skb_hint = NULL;
3404 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3405 tp->snd_up = tp->snd_una;
3407 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3408 flag |= FLAG_SACK_RENEGING;
3410 if (flag & FLAG_ACKED) {
3411 const struct tcp_congestion_ops *ca_ops
3412 = inet_csk(sk)->icsk_ca_ops;
3414 if (unlikely(icsk->icsk_mtup.probe_size &&
3415 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3416 tcp_mtup_probe_success(sk);
3419 tcp_ack_update_rtt(sk, flag, seq_rtt);
3422 if (tcp_is_reno(tp)) {
3423 tcp_remove_reno_sacks(sk, pkts_acked);
3427 /* Non-retransmitted hole got filled? That's reordering */
3428 if (reord < prior_fackets && reord <= tp->fackets_out)
3429 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3431 delta = tcp_is_fack(tp) ? pkts_acked :
3432 prior_sacked - tp->sacked_out;
3433 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3436 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3438 if (ca_ops->pkts_acked) {
3441 /* Is the ACK triggering packet unambiguous? */
3442 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3443 /* High resolution needed and available? */
3444 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3445 !ktime_equal(last_ackt,
3446 net_invalid_timestamp()))
3447 rtt_us = ktime_us_delta(ktime_get_real(),
3449 else if (ca_seq_rtt >= 0)
3450 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3453 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3457 #if FASTRETRANS_DEBUG > 0
3458 WARN_ON((int)tp->sacked_out < 0);
3459 WARN_ON((int)tp->lost_out < 0);
3460 WARN_ON((int)tp->retrans_out < 0);
3461 if (!tp->packets_out && tcp_is_sack(tp)) {
3462 icsk = inet_csk(sk);
3464 printk(KERN_DEBUG "Leak l=%u %d\n",
3465 tp->lost_out, icsk->icsk_ca_state);
3468 if (tp->sacked_out) {
3469 printk(KERN_DEBUG "Leak s=%u %d\n",
3470 tp->sacked_out, icsk->icsk_ca_state);
3473 if (tp->retrans_out) {
3474 printk(KERN_DEBUG "Leak r=%u %d\n",
3475 tp->retrans_out, icsk->icsk_ca_state);
3476 tp->retrans_out = 0;
3483 static void tcp_ack_probe(struct sock *sk)
3485 const struct tcp_sock *tp = tcp_sk(sk);
3486 struct inet_connection_sock *icsk = inet_csk(sk);
3488 /* Was it a usable window open? */
3490 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3491 icsk->icsk_backoff = 0;
3492 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3493 /* Socket must be waked up by subsequent tcp_data_snd_check().
3494 * This function is not for random using!
3497 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3498 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3503 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3505 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3506 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3509 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3511 const struct tcp_sock *tp = tcp_sk(sk);
3512 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3513 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3516 /* Check that window update is acceptable.
3517 * The function assumes that snd_una<=ack<=snd_next.
3519 static inline int tcp_may_update_window(const struct tcp_sock *tp,
3520 const u32 ack, const u32 ack_seq,
3523 return after(ack, tp->snd_una) ||
3524 after(ack_seq, tp->snd_wl1) ||
3525 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3528 /* Update our send window.
3530 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3531 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3533 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3536 struct tcp_sock *tp = tcp_sk(sk);
3538 u32 nwin = ntohs(tcp_hdr(skb)->window);
3540 if (likely(!tcp_hdr(skb)->syn))
3541 nwin <<= tp->rx_opt.snd_wscale;
3543 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3544 flag |= FLAG_WIN_UPDATE;
3545 tcp_update_wl(tp, ack_seq);
3547 if (tp->snd_wnd != nwin) {
3550 /* Note, it is the only place, where
3551 * fast path is recovered for sending TCP.
3554 tcp_fast_path_check(sk);
3556 if (nwin > tp->max_window) {
3557 tp->max_window = nwin;
3558 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3568 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3569 * continue in congestion avoidance.
3571 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3573 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3574 tp->snd_cwnd_cnt = 0;
3575 tp->bytes_acked = 0;
3576 TCP_ECN_queue_cwr(tp);
3577 tcp_moderate_cwnd(tp);
3580 /* A conservative spurious RTO response algorithm: reduce cwnd using
3581 * rate halving and continue in congestion avoidance.
3583 static void tcp_ratehalving_spur_to_response(struct sock *sk)
3585 tcp_enter_cwr(sk, 0);
3588 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3590 if (flag & FLAG_ECE)
3591 tcp_ratehalving_spur_to_response(sk);
3593 tcp_undo_cwr(sk, true);
3596 /* F-RTO spurious RTO detection algorithm (RFC4138)
3598 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3599 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3600 * window (but not to or beyond highest sequence sent before RTO):
3601 * On First ACK, send two new segments out.
3602 * On Second ACK, RTO was likely spurious. Do spurious response (response
3603 * algorithm is not part of the F-RTO detection algorithm
3604 * given in RFC4138 but can be selected separately).
3605 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3606 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3607 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3608 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3610 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3611 * original window even after we transmit two new data segments.
3614 * on first step, wait until first cumulative ACK arrives, then move to
3615 * the second step. In second step, the next ACK decides.
3617 * F-RTO is implemented (mainly) in four functions:
3618 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3619 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3620 * called when tcp_use_frto() showed green light
3621 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3622 * - tcp_enter_frto_loss() is called if there is not enough evidence
3623 * to prove that the RTO is indeed spurious. It transfers the control
3624 * from F-RTO to the conventional RTO recovery
3626 static int tcp_process_frto(struct sock *sk, int flag)
3628 struct tcp_sock *tp = tcp_sk(sk);
3630 tcp_verify_left_out(tp);
3632 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3633 if (flag & FLAG_DATA_ACKED)
3634 inet_csk(sk)->icsk_retransmits = 0;
3636 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3637 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3638 tp->undo_marker = 0;
3640 if (!before(tp->snd_una, tp->frto_highmark)) {
3641 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3645 if (!tcp_is_sackfrto(tp)) {
3646 /* RFC4138 shortcoming in step 2; should also have case c):
3647 * ACK isn't duplicate nor advances window, e.g., opposite dir
3650 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3653 if (!(flag & FLAG_DATA_ACKED)) {
3654 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3659 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3660 if (!tcp_packets_in_flight(tp)) {
3661 tcp_enter_frto_loss(sk, 2, flag);
3665 /* Prevent sending of new data. */
3666 tp->snd_cwnd = min(tp->snd_cwnd,
3667 tcp_packets_in_flight(tp));
3671 if ((tp->frto_counter >= 2) &&
3672 (!(flag & FLAG_FORWARD_PROGRESS) ||
3673 ((flag & FLAG_DATA_SACKED) &&
3674 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3675 /* RFC4138 shortcoming (see comment above) */
3676 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3677 (flag & FLAG_NOT_DUP))
3680 tcp_enter_frto_loss(sk, 3, flag);
3685 if (tp->frto_counter == 1) {
3686 /* tcp_may_send_now needs to see updated state */
3687 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3688 tp->frto_counter = 2;
3690 if (!tcp_may_send_now(sk))
3691 tcp_enter_frto_loss(sk, 2, flag);
3695 switch (sysctl_tcp_frto_response) {
3697 tcp_undo_spur_to_response(sk, flag);
3700 tcp_conservative_spur_to_response(tp);
3703 tcp_ratehalving_spur_to_response(sk);
3706 tp->frto_counter = 0;
3707 tp->undo_marker = 0;
3708 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3713 /* RFC 5961 7 [ACK Throttling] */
3714 static void tcp_send_challenge_ack(struct sock *sk)
3716 /* unprotected vars, we dont care of overwrites */
3717 static u32 challenge_timestamp;
3718 static unsigned int challenge_count;
3719 u32 count, now = jiffies / HZ;
3721 if (now != challenge_timestamp) {
3722 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3724 challenge_timestamp = now;
3725 ACCESS_ONCE(challenge_count) =
3727 ((u64) random32() * sysctl_tcp_challenge_ack_limit)
3730 count = ACCESS_ONCE(challenge_count);
3732 ACCESS_ONCE(challenge_count) = count - 1;
3733 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3738 static void tcp_store_ts_recent(struct tcp_sock *tp)
3740 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3741 tp->rx_opt.ts_recent_stamp = get_seconds();
3744 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3746 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3747 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3748 * extra check below makes sure this can only happen
3749 * for pure ACK frames. -DaveM
3751 * Not only, also it occurs for expired timestamps.
3754 if (tcp_paws_check(&tp->rx_opt, 0))
3755 tcp_store_ts_recent(tp);
3759 /* This routine deals with incoming acks, but not outgoing ones. */
3760 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3762 struct inet_connection_sock *icsk = inet_csk(sk);
3763 struct tcp_sock *tp = tcp_sk(sk);
3764 u32 prior_snd_una = tp->snd_una;
3765 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3766 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3767 u32 prior_in_flight;
3770 int prior_sacked = tp->sacked_out;
3771 int newly_acked_sacked = 0;
3774 /* If the ack is older than previous acks
3775 * then we can probably ignore it.
3777 if (before(ack, prior_snd_una)) {
3778 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3779 if (before(ack, prior_snd_una - tp->max_window)) {
3780 tcp_send_challenge_ack(sk);
3786 /* If the ack includes data we haven't sent yet, discard
3787 * this segment (RFC793 Section 3.9).
3789 if (after(ack, tp->snd_nxt))
3792 if (after(ack, prior_snd_una))
3793 flag |= FLAG_SND_UNA_ADVANCED;
3795 if (sysctl_tcp_abc) {
3796 if (icsk->icsk_ca_state < TCP_CA_CWR)
3797 tp->bytes_acked += ack - prior_snd_una;
3798 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3799 /* we assume just one segment left network */
3800 tp->bytes_acked += min(ack - prior_snd_una,
3804 prior_fackets = tp->fackets_out;
3805 prior_in_flight = tcp_packets_in_flight(tp);
3807 /* ts_recent update must be made after we are sure that the packet
3810 if (flag & FLAG_UPDATE_TS_RECENT)
3811 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3813 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3814 /* Window is constant, pure forward advance.
3815 * No more checks are required.
3816 * Note, we use the fact that SND.UNA>=SND.WL2.
3818 tcp_update_wl(tp, ack_seq);
3820 flag |= FLAG_WIN_UPDATE;
3822 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3824 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3826 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3829 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3831 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3833 if (TCP_SKB_CB(skb)->sacked)
3834 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3836 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3839 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3842 /* We passed data and got it acked, remove any soft error
3843 * log. Something worked...
3845 sk->sk_err_soft = 0;
3846 icsk->icsk_probes_out = 0;
3847 tp->rcv_tstamp = tcp_time_stamp;
3848 prior_packets = tp->packets_out;
3852 /* See if we can take anything off of the retransmit queue. */
3853 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3855 newly_acked_sacked = (prior_packets - prior_sacked) -
3856 (tp->packets_out - tp->sacked_out);
3858 if (tp->frto_counter)
3859 frto_cwnd = tcp_process_frto(sk, flag);
3860 /* Guarantee sacktag reordering detection against wrap-arounds */
3861 if (before(tp->frto_highmark, tp->snd_una))
3862 tp->frto_highmark = 0;
3864 if (tcp_ack_is_dubious(sk, flag)) {
3865 /* Advance CWND, if state allows this. */
3866 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3867 tcp_may_raise_cwnd(sk, flag))
3868 tcp_cong_avoid(sk, ack, prior_in_flight);
3869 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3870 newly_acked_sacked, flag);
3872 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3873 tcp_cong_avoid(sk, ack, prior_in_flight);
3876 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3877 dst_confirm(__sk_dst_get(sk));
3882 /* If this ack opens up a zero window, clear backoff. It was
3883 * being used to time the probes, and is probably far higher than
3884 * it needs to be for normal retransmission.
3886 if (tcp_send_head(sk))
3891 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3895 if (TCP_SKB_CB(skb)->sacked) {
3896 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3897 if (icsk->icsk_ca_state == TCP_CA_Open)
3898 tcp_try_keep_open(sk);
3901 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3905 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3906 * But, this can also be called on packets in the established flow when
3907 * the fast version below fails.
3909 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3910 const u8 **hvpp, int estab)
3912 const unsigned char *ptr;
3913 const struct tcphdr *th = tcp_hdr(skb);
3914 int length = (th->doff * 4) - sizeof(struct tcphdr);
3916 ptr = (const unsigned char *)(th + 1);
3917 opt_rx->saw_tstamp = 0;
3919 while (length > 0) {
3920 int opcode = *ptr++;
3926 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3931 if (opsize < 2) /* "silly options" */
3933 if (opsize > length)
3934 return; /* don't parse partial options */
3937 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3938 u16 in_mss = get_unaligned_be16(ptr);
3940 if (opt_rx->user_mss &&
3941 opt_rx->user_mss < in_mss)
3942 in_mss = opt_rx->user_mss;
3943 opt_rx->mss_clamp = in_mss;
3948 if (opsize == TCPOLEN_WINDOW && th->syn &&
3949 !estab && sysctl_tcp_window_scaling) {
3950 __u8 snd_wscale = *(__u8 *)ptr;
3951 opt_rx->wscale_ok = 1;
3952 if (snd_wscale > 14) {
3953 if (net_ratelimit())
3954 printk(KERN_INFO "tcp_parse_options: Illegal window "
3955 "scaling value %d >14 received.\n",
3959 opt_rx->snd_wscale = snd_wscale;
3962 case TCPOPT_TIMESTAMP:
3963 if ((opsize == TCPOLEN_TIMESTAMP) &&
3964 ((estab && opt_rx->tstamp_ok) ||
3965 (!estab && sysctl_tcp_timestamps))) {
3966 opt_rx->saw_tstamp = 1;
3967 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3968 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3971 case TCPOPT_SACK_PERM:
3972 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3973 !estab && sysctl_tcp_sack) {
3974 opt_rx->sack_ok = 1;
3975 tcp_sack_reset(opt_rx);
3980 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3981 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3983 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3986 #ifdef CONFIG_TCP_MD5SIG
3989 * The MD5 Hash has already been
3990 * checked (see tcp_v{4,6}_do_rcv()).
3995 /* This option is variable length.
3998 case TCPOLEN_COOKIE_BASE:
3999 /* not yet implemented */
4001 case TCPOLEN_COOKIE_PAIR:
4002 /* not yet implemented */
4004 case TCPOLEN_COOKIE_MIN+0:
4005 case TCPOLEN_COOKIE_MIN+2:
4006 case TCPOLEN_COOKIE_MIN+4:
4007 case TCPOLEN_COOKIE_MIN+6:
4008 case TCPOLEN_COOKIE_MAX:
4009 /* 16-bit multiple */
4010 opt_rx->cookie_plus = opsize;
4025 EXPORT_SYMBOL(tcp_parse_options);
4027 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4029 const __be32 *ptr = (const __be32 *)(th + 1);
4031 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4032 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4033 tp->rx_opt.saw_tstamp = 1;
4035 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4037 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4043 /* Fast parse options. This hopes to only see timestamps.
4044 * If it is wrong it falls back on tcp_parse_options().
4046 static int tcp_fast_parse_options(const struct sk_buff *skb,
4047 const struct tcphdr *th,
4048 struct tcp_sock *tp, const u8 **hvpp)
4050 /* In the spirit of fast parsing, compare doff directly to constant
4051 * values. Because equality is used, short doff can be ignored here.
4053 if (th->doff == (sizeof(*th) / 4)) {
4054 tp->rx_opt.saw_tstamp = 0;
4056 } else if (tp->rx_opt.tstamp_ok &&
4057 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4058 if (tcp_parse_aligned_timestamp(tp, th))
4061 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
4065 #ifdef CONFIG_TCP_MD5SIG
4067 * Parse MD5 Signature option
4069 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4071 int length = (th->doff << 2) - sizeof(*th);
4072 const u8 *ptr = (const u8 *)(th + 1);
4074 /* If the TCP option is too short, we can short cut */
4075 if (length < TCPOLEN_MD5SIG)
4078 while (length > 0) {
4079 int opcode = *ptr++;
4090 if (opsize < 2 || opsize > length)
4092 if (opcode == TCPOPT_MD5SIG)
4093 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4100 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4103 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4105 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4106 * it can pass through stack. So, the following predicate verifies that
4107 * this segment is not used for anything but congestion avoidance or
4108 * fast retransmit. Moreover, we even are able to eliminate most of such
4109 * second order effects, if we apply some small "replay" window (~RTO)
4110 * to timestamp space.
4112 * All these measures still do not guarantee that we reject wrapped ACKs
4113 * on networks with high bandwidth, when sequence space is recycled fastly,
4114 * but it guarantees that such events will be very rare and do not affect
4115 * connection seriously. This doesn't look nice, but alas, PAWS is really
4118 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4119 * states that events when retransmit arrives after original data are rare.
4120 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4121 * the biggest problem on large power networks even with minor reordering.
4122 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4123 * up to bandwidth of 18Gigabit/sec. 8) ]
4126 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4128 const struct tcp_sock *tp = tcp_sk(sk);
4129 const struct tcphdr *th = tcp_hdr(skb);
4130 u32 seq = TCP_SKB_CB(skb)->seq;
4131 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4133 return (/* 1. Pure ACK with correct sequence number. */
4134 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4136 /* 2. ... and duplicate ACK. */
4137 ack == tp->snd_una &&
4139 /* 3. ... and does not update window. */
4140 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4142 /* 4. ... and sits in replay window. */
4143 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4146 static inline int tcp_paws_discard(const struct sock *sk,
4147 const struct sk_buff *skb)
4149 const struct tcp_sock *tp = tcp_sk(sk);
4151 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4152 !tcp_disordered_ack(sk, skb);
4155 /* Check segment sequence number for validity.
4157 * Segment controls are considered valid, if the segment
4158 * fits to the window after truncation to the window. Acceptability
4159 * of data (and SYN, FIN, of course) is checked separately.
4160 * See tcp_data_queue(), for example.
4162 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4163 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4164 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4165 * (borrowed from freebsd)
4168 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4170 return !before(end_seq, tp->rcv_wup) &&
4171 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4174 /* When we get a reset we do this. */
4175 static void tcp_reset(struct sock *sk)
4177 /* We want the right error as BSD sees it (and indeed as we do). */
4178 switch (sk->sk_state) {
4180 sk->sk_err = ECONNREFUSED;
4182 case TCP_CLOSE_WAIT:
4188 sk->sk_err = ECONNRESET;
4190 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4193 if (!sock_flag(sk, SOCK_DEAD))
4194 sk->sk_error_report(sk);
4200 * Process the FIN bit. This now behaves as it is supposed to work
4201 * and the FIN takes effect when it is validly part of sequence
4202 * space. Not before when we get holes.
4204 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4205 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4208 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4209 * close and we go into CLOSING (and later onto TIME-WAIT)
4211 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4213 static void tcp_fin(struct sock *sk)
4215 struct tcp_sock *tp = tcp_sk(sk);
4217 inet_csk_schedule_ack(sk);
4219 sk->sk_shutdown |= RCV_SHUTDOWN;
4220 sock_set_flag(sk, SOCK_DONE);
4222 switch (sk->sk_state) {
4224 case TCP_ESTABLISHED:
4225 /* Move to CLOSE_WAIT */
4226 tcp_set_state(sk, TCP_CLOSE_WAIT);
4227 inet_csk(sk)->icsk_ack.pingpong = 1;
4230 case TCP_CLOSE_WAIT:
4232 /* Received a retransmission of the FIN, do
4237 /* RFC793: Remain in the LAST-ACK state. */
4241 /* This case occurs when a simultaneous close
4242 * happens, we must ack the received FIN and
4243 * enter the CLOSING state.
4246 tcp_set_state(sk, TCP_CLOSING);
4249 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4251 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4254 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4255 * cases we should never reach this piece of code.
4257 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
4258 __func__, sk->sk_state);
4262 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4263 * Probably, we should reset in this case. For now drop them.
4265 __skb_queue_purge(&tp->out_of_order_queue);
4266 if (tcp_is_sack(tp))
4267 tcp_sack_reset(&tp->rx_opt);
4270 if (!sock_flag(sk, SOCK_DEAD)) {
4271 sk->sk_state_change(sk);
4273 /* Do not send POLL_HUP for half duplex close. */
4274 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4275 sk->sk_state == TCP_CLOSE)
4276 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4278 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4282 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4285 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4286 if (before(seq, sp->start_seq))
4287 sp->start_seq = seq;
4288 if (after(end_seq, sp->end_seq))
4289 sp->end_seq = end_seq;
4295 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4297 struct tcp_sock *tp = tcp_sk(sk);
4299 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4302 if (before(seq, tp->rcv_nxt))
4303 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4305 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4307 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4309 tp->rx_opt.dsack = 1;
4310 tp->duplicate_sack[0].start_seq = seq;
4311 tp->duplicate_sack[0].end_seq = end_seq;
4315 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4317 struct tcp_sock *tp = tcp_sk(sk);
4319 if (!tp->rx_opt.dsack)
4320 tcp_dsack_set(sk, seq, end_seq);
4322 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4325 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4327 struct tcp_sock *tp = tcp_sk(sk);
4329 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4330 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4331 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4332 tcp_enter_quickack_mode(sk);
4334 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4335 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4337 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4338 end_seq = tp->rcv_nxt;
4339 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4346 /* These routines update the SACK block as out-of-order packets arrive or
4347 * in-order packets close up the sequence space.
4349 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4352 struct tcp_sack_block *sp = &tp->selective_acks[0];
4353 struct tcp_sack_block *swalk = sp + 1;
4355 /* See if the recent change to the first SACK eats into
4356 * or hits the sequence space of other SACK blocks, if so coalesce.
4358 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4359 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4362 /* Zap SWALK, by moving every further SACK up by one slot.
4363 * Decrease num_sacks.
4365 tp->rx_opt.num_sacks--;
4366 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4370 this_sack++, swalk++;
4374 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4376 struct tcp_sock *tp = tcp_sk(sk);
4377 struct tcp_sack_block *sp = &tp->selective_acks[0];
4378 int cur_sacks = tp->rx_opt.num_sacks;
4384 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4385 if (tcp_sack_extend(sp, seq, end_seq)) {
4386 /* Rotate this_sack to the first one. */
4387 for (; this_sack > 0; this_sack--, sp--)
4388 swap(*sp, *(sp - 1));
4390 tcp_sack_maybe_coalesce(tp);
4395 /* Could not find an adjacent existing SACK, build a new one,
4396 * put it at the front, and shift everyone else down. We
4397 * always know there is at least one SACK present already here.
4399 * If the sack array is full, forget about the last one.
4401 if (this_sack >= TCP_NUM_SACKS) {
4403 tp->rx_opt.num_sacks--;
4406 for (; this_sack > 0; this_sack--, sp--)
4410 /* Build the new head SACK, and we're done. */
4411 sp->start_seq = seq;
4412 sp->end_seq = end_seq;
4413 tp->rx_opt.num_sacks++;
4416 /* RCV.NXT advances, some SACKs should be eaten. */
4418 static void tcp_sack_remove(struct tcp_sock *tp)
4420 struct tcp_sack_block *sp = &tp->selective_acks[0];
4421 int num_sacks = tp->rx_opt.num_sacks;
4424 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4425 if (skb_queue_empty(&tp->out_of_order_queue)) {
4426 tp->rx_opt.num_sacks = 0;
4430 for (this_sack = 0; this_sack < num_sacks;) {
4431 /* Check if the start of the sack is covered by RCV.NXT. */
4432 if (!before(tp->rcv_nxt, sp->start_seq)) {
4435 /* RCV.NXT must cover all the block! */
4436 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4438 /* Zap this SACK, by moving forward any other SACKS. */
4439 for (i=this_sack+1; i < num_sacks; i++)
4440 tp->selective_acks[i-1] = tp->selective_acks[i];
4447 tp->rx_opt.num_sacks = num_sacks;
4450 /* This one checks to see if we can put data from the
4451 * out_of_order queue into the receive_queue.
4453 static void tcp_ofo_queue(struct sock *sk)
4455 struct tcp_sock *tp = tcp_sk(sk);
4456 __u32 dsack_high = tp->rcv_nxt;
4457 struct sk_buff *skb;
4459 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4460 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4463 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4464 __u32 dsack = dsack_high;
4465 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4466 dsack_high = TCP_SKB_CB(skb)->end_seq;
4467 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4470 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4471 SOCK_DEBUG(sk, "ofo packet was already received\n");
4472 __skb_unlink(skb, &tp->out_of_order_queue);
4476 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4477 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4478 TCP_SKB_CB(skb)->end_seq);
4480 __skb_unlink(skb, &tp->out_of_order_queue);
4481 __skb_queue_tail(&sk->sk_receive_queue, skb);
4482 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4483 if (tcp_hdr(skb)->fin)
4488 static int tcp_prune_ofo_queue(struct sock *sk);
4489 static int tcp_prune_queue(struct sock *sk);
4491 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4493 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4494 !sk_rmem_schedule(sk, size)) {
4496 if (tcp_prune_queue(sk) < 0)
4499 if (!sk_rmem_schedule(sk, size)) {
4500 if (!tcp_prune_ofo_queue(sk))
4503 if (!sk_rmem_schedule(sk, size))
4510 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4512 const struct tcphdr *th = tcp_hdr(skb);
4513 struct tcp_sock *tp = tcp_sk(sk);
4516 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4520 __skb_pull(skb, th->doff * 4);
4522 TCP_ECN_accept_cwr(tp, skb);
4524 tp->rx_opt.dsack = 0;
4526 /* Queue data for delivery to the user.
4527 * Packets in sequence go to the receive queue.
4528 * Out of sequence packets to the out_of_order_queue.
4530 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4531 if (tcp_receive_window(tp) == 0)
4534 /* Ok. In sequence. In window. */
4535 if (tp->ucopy.task == current &&
4536 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4537 sock_owned_by_user(sk) && !tp->urg_data) {
4538 int chunk = min_t(unsigned int, skb->len,
4541 __set_current_state(TASK_RUNNING);
4544 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4545 tp->ucopy.len -= chunk;
4546 tp->copied_seq += chunk;
4547 eaten = (chunk == skb->len);
4548 tcp_rcv_space_adjust(sk);
4556 tcp_try_rmem_schedule(sk, skb->truesize))
4559 skb_set_owner_r(skb, sk);
4560 __skb_queue_tail(&sk->sk_receive_queue, skb);
4562 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4564 tcp_event_data_recv(sk, skb);
4568 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4571 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4572 * gap in queue is filled.
4574 if (skb_queue_empty(&tp->out_of_order_queue))
4575 inet_csk(sk)->icsk_ack.pingpong = 0;
4578 if (tp->rx_opt.num_sacks)
4579 tcp_sack_remove(tp);
4581 tcp_fast_path_check(sk);
4585 else if (!sock_flag(sk, SOCK_DEAD))
4586 sk->sk_data_ready(sk, 0);
4590 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4591 /* A retransmit, 2nd most common case. Force an immediate ack. */
4592 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4593 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4596 tcp_enter_quickack_mode(sk);
4597 inet_csk_schedule_ack(sk);
4603 /* Out of window. F.e. zero window probe. */
4604 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4607 tcp_enter_quickack_mode(sk);
4609 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4610 /* Partial packet, seq < rcv_next < end_seq */
4611 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4612 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4613 TCP_SKB_CB(skb)->end_seq);
4615 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4617 /* If window is closed, drop tail of packet. But after
4618 * remembering D-SACK for its head made in previous line.
4620 if (!tcp_receive_window(tp))
4625 TCP_ECN_check_ce(tp, skb);
4627 if (tcp_try_rmem_schedule(sk, skb->truesize))
4630 /* Disable header prediction. */
4632 inet_csk_schedule_ack(sk);
4634 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4635 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4637 skb_set_owner_r(skb, sk);
4639 if (!skb_peek(&tp->out_of_order_queue)) {
4640 /* Initial out of order segment, build 1 SACK. */
4641 if (tcp_is_sack(tp)) {
4642 tp->rx_opt.num_sacks = 1;
4643 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4644 tp->selective_acks[0].end_seq =
4645 TCP_SKB_CB(skb)->end_seq;
4647 __skb_queue_head(&tp->out_of_order_queue, skb);
4649 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
4650 u32 seq = TCP_SKB_CB(skb)->seq;
4651 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4653 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4654 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4656 if (!tp->rx_opt.num_sacks ||
4657 tp->selective_acks[0].end_seq != seq)
4660 /* Common case: data arrive in order after hole. */
4661 tp->selective_acks[0].end_seq = end_seq;
4665 /* Find place to insert this segment. */
4667 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4669 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4673 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4676 /* Do skb overlap to previous one? */
4677 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4678 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4679 /* All the bits are present. Drop. */
4681 tcp_dsack_set(sk, seq, end_seq);
4684 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4685 /* Partial overlap. */
4686 tcp_dsack_set(sk, seq,
4687 TCP_SKB_CB(skb1)->end_seq);
4689 if (skb_queue_is_first(&tp->out_of_order_queue,
4693 skb1 = skb_queue_prev(
4694 &tp->out_of_order_queue,
4699 __skb_queue_head(&tp->out_of_order_queue, skb);
4701 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4703 /* And clean segments covered by new one as whole. */
4704 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4705 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4707 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4709 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4710 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4714 __skb_unlink(skb1, &tp->out_of_order_queue);
4715 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4716 TCP_SKB_CB(skb1)->end_seq);
4721 if (tcp_is_sack(tp))
4722 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4726 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4727 struct sk_buff_head *list)
4729 struct sk_buff *next = NULL;
4731 if (!skb_queue_is_last(list, skb))
4732 next = skb_queue_next(list, skb);
4734 __skb_unlink(skb, list);
4736 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4741 /* Collapse contiguous sequence of skbs head..tail with
4742 * sequence numbers start..end.
4744 * If tail is NULL, this means until the end of the list.
4746 * Segments with FIN/SYN are not collapsed (only because this
4750 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4751 struct sk_buff *head, struct sk_buff *tail,
4754 struct sk_buff *skb, *n;
4757 /* First, check that queue is collapsible and find
4758 * the point where collapsing can be useful. */
4762 skb_queue_walk_from_safe(list, skb, n) {
4765 /* No new bits? It is possible on ofo queue. */
4766 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4767 skb = tcp_collapse_one(sk, skb, list);
4773 /* The first skb to collapse is:
4775 * - bloated or contains data before "start" or
4776 * overlaps to the next one.
4778 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4779 (tcp_win_from_space(skb->truesize) > skb->len ||
4780 before(TCP_SKB_CB(skb)->seq, start))) {
4781 end_of_skbs = false;
4785 if (!skb_queue_is_last(list, skb)) {
4786 struct sk_buff *next = skb_queue_next(list, skb);
4788 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4789 end_of_skbs = false;
4794 /* Decided to skip this, advance start seq. */
4795 start = TCP_SKB_CB(skb)->end_seq;
4797 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4800 while (before(start, end)) {
4801 struct sk_buff *nskb;
4802 unsigned int header = skb_headroom(skb);
4803 int copy = SKB_MAX_ORDER(header, 0);
4805 /* Too big header? This can happen with IPv6. */
4808 if (end - start < copy)
4810 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4814 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4815 skb_set_network_header(nskb, (skb_network_header(skb) -
4817 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4819 skb_reserve(nskb, header);
4820 memcpy(nskb->head, skb->head, header);
4821 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4822 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4823 __skb_queue_before(list, skb, nskb);
4824 skb_set_owner_r(nskb, sk);
4826 /* Copy data, releasing collapsed skbs. */
4828 int offset = start - TCP_SKB_CB(skb)->seq;
4829 int size = TCP_SKB_CB(skb)->end_seq - start;
4833 size = min(copy, size);
4834 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4836 TCP_SKB_CB(nskb)->end_seq += size;
4840 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4841 skb = tcp_collapse_one(sk, skb, list);
4844 tcp_hdr(skb)->syn ||
4852 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4853 * and tcp_collapse() them until all the queue is collapsed.
4855 static void tcp_collapse_ofo_queue(struct sock *sk)
4857 struct tcp_sock *tp = tcp_sk(sk);
4858 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4859 struct sk_buff *head;
4865 start = TCP_SKB_CB(skb)->seq;
4866 end = TCP_SKB_CB(skb)->end_seq;
4870 struct sk_buff *next = NULL;
4872 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4873 next = skb_queue_next(&tp->out_of_order_queue, skb);
4876 /* Segment is terminated when we see gap or when
4877 * we are at the end of all the queue. */
4879 after(TCP_SKB_CB(skb)->seq, end) ||
4880 before(TCP_SKB_CB(skb)->end_seq, start)) {
4881 tcp_collapse(sk, &tp->out_of_order_queue,
4882 head, skb, start, end);
4886 /* Start new segment */
4887 start = TCP_SKB_CB(skb)->seq;
4888 end = TCP_SKB_CB(skb)->end_seq;
4890 if (before(TCP_SKB_CB(skb)->seq, start))
4891 start = TCP_SKB_CB(skb)->seq;
4892 if (after(TCP_SKB_CB(skb)->end_seq, end))
4893 end = TCP_SKB_CB(skb)->end_seq;
4899 * Purge the out-of-order queue.
4900 * Return true if queue was pruned.
4902 static int tcp_prune_ofo_queue(struct sock *sk)
4904 struct tcp_sock *tp = tcp_sk(sk);
4907 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4908 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4909 __skb_queue_purge(&tp->out_of_order_queue);
4911 /* Reset SACK state. A conforming SACK implementation will
4912 * do the same at a timeout based retransmit. When a connection
4913 * is in a sad state like this, we care only about integrity
4914 * of the connection not performance.
4916 if (tp->rx_opt.sack_ok)
4917 tcp_sack_reset(&tp->rx_opt);
4924 /* Reduce allocated memory if we can, trying to get
4925 * the socket within its memory limits again.
4927 * Return less than zero if we should start dropping frames
4928 * until the socket owning process reads some of the data
4929 * to stabilize the situation.
4931 static int tcp_prune_queue(struct sock *sk)
4933 struct tcp_sock *tp = tcp_sk(sk);
4935 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4937 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4939 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4940 tcp_clamp_window(sk);
4941 else if (tcp_memory_pressure)
4942 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4944 tcp_collapse_ofo_queue(sk);
4945 if (!skb_queue_empty(&sk->sk_receive_queue))
4946 tcp_collapse(sk, &sk->sk_receive_queue,
4947 skb_peek(&sk->sk_receive_queue),
4949 tp->copied_seq, tp->rcv_nxt);
4952 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4955 /* Collapsing did not help, destructive actions follow.
4956 * This must not ever occur. */
4958 tcp_prune_ofo_queue(sk);
4960 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4963 /* If we are really being abused, tell the caller to silently
4964 * drop receive data on the floor. It will get retransmitted
4965 * and hopefully then we'll have sufficient space.
4967 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4969 /* Massive buffer overcommit. */
4974 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4975 * As additional protections, we do not touch cwnd in retransmission phases,
4976 * and if application hit its sndbuf limit recently.
4978 void tcp_cwnd_application_limited(struct sock *sk)
4980 struct tcp_sock *tp = tcp_sk(sk);
4982 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4983 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4984 /* Limited by application or receiver window. */
4985 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4986 u32 win_used = max(tp->snd_cwnd_used, init_win);
4987 if (win_used < tp->snd_cwnd) {
4988 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4989 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4991 tp->snd_cwnd_used = 0;
4993 tp->snd_cwnd_stamp = tcp_time_stamp;
4996 static int tcp_should_expand_sndbuf(const struct sock *sk)
4998 const struct tcp_sock *tp = tcp_sk(sk);
5000 /* If the user specified a specific send buffer setting, do
5003 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5006 /* If we are under global TCP memory pressure, do not expand. */
5007 if (tcp_memory_pressure)
5010 /* If we are under soft global TCP memory pressure, do not expand. */
5011 if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
5014 /* If we filled the congestion window, do not expand. */
5015 if (tp->packets_out >= tp->snd_cwnd)
5021 /* When incoming ACK allowed to free some skb from write_queue,
5022 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5023 * on the exit from tcp input handler.
5025 * PROBLEM: sndbuf expansion does not work well with largesend.
5027 static void tcp_new_space(struct sock *sk)
5029 struct tcp_sock *tp = tcp_sk(sk);
5031 if (tcp_should_expand_sndbuf(sk)) {
5032 int sndmem = SKB_TRUESIZE(max_t(u32,
5033 tp->rx_opt.mss_clamp,
5036 int demanded = max_t(unsigned int, tp->snd_cwnd,
5037 tp->reordering + 1);
5038 sndmem *= 2 * demanded;
5039 if (sndmem > sk->sk_sndbuf)
5040 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5041 tp->snd_cwnd_stamp = tcp_time_stamp;
5044 sk->sk_write_space(sk);
5047 static void tcp_check_space(struct sock *sk)
5049 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5050 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5051 if (sk->sk_socket &&
5052 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5057 static inline void tcp_data_snd_check(struct sock *sk)
5059 tcp_push_pending_frames(sk);
5060 tcp_check_space(sk);
5064 * Check if sending an ack is needed.
5066 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5068 struct tcp_sock *tp = tcp_sk(sk);
5070 /* More than one full frame received... */
5071 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5072 /* ... and right edge of window advances far enough.
5073 * (tcp_recvmsg() will send ACK otherwise). Or...
5075 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5076 /* We ACK each frame or... */
5077 tcp_in_quickack_mode(sk) ||
5078 /* We have out of order data. */
5079 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5080 /* Then ack it now */
5083 /* Else, send delayed ack. */
5084 tcp_send_delayed_ack(sk);
5088 static inline void tcp_ack_snd_check(struct sock *sk)
5090 if (!inet_csk_ack_scheduled(sk)) {
5091 /* We sent a data segment already. */
5094 __tcp_ack_snd_check(sk, 1);
5098 * This routine is only called when we have urgent data
5099 * signaled. Its the 'slow' part of tcp_urg. It could be
5100 * moved inline now as tcp_urg is only called from one
5101 * place. We handle URGent data wrong. We have to - as
5102 * BSD still doesn't use the correction from RFC961.
5103 * For 1003.1g we should support a new option TCP_STDURG to permit
5104 * either form (or just set the sysctl tcp_stdurg).
5107 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5109 struct tcp_sock *tp = tcp_sk(sk);
5110 u32 ptr = ntohs(th->urg_ptr);
5112 if (ptr && !sysctl_tcp_stdurg)
5114 ptr += ntohl(th->seq);
5116 /* Ignore urgent data that we've already seen and read. */
5117 if (after(tp->copied_seq, ptr))
5120 /* Do not replay urg ptr.
5122 * NOTE: interesting situation not covered by specs.
5123 * Misbehaving sender may send urg ptr, pointing to segment,
5124 * which we already have in ofo queue. We are not able to fetch
5125 * such data and will stay in TCP_URG_NOTYET until will be eaten
5126 * by recvmsg(). Seems, we are not obliged to handle such wicked
5127 * situations. But it is worth to think about possibility of some
5128 * DoSes using some hypothetical application level deadlock.
5130 if (before(ptr, tp->rcv_nxt))
5133 /* Do we already have a newer (or duplicate) urgent pointer? */
5134 if (tp->urg_data && !after(ptr, tp->urg_seq))
5137 /* Tell the world about our new urgent pointer. */
5140 /* We may be adding urgent data when the last byte read was
5141 * urgent. To do this requires some care. We cannot just ignore
5142 * tp->copied_seq since we would read the last urgent byte again
5143 * as data, nor can we alter copied_seq until this data arrives
5144 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5146 * NOTE. Double Dutch. Rendering to plain English: author of comment
5147 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5148 * and expect that both A and B disappear from stream. This is _wrong_.
5149 * Though this happens in BSD with high probability, this is occasional.
5150 * Any application relying on this is buggy. Note also, that fix "works"
5151 * only in this artificial test. Insert some normal data between A and B and we will
5152 * decline of BSD again. Verdict: it is better to remove to trap
5155 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5156 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5157 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5159 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5160 __skb_unlink(skb, &sk->sk_receive_queue);
5165 tp->urg_data = TCP_URG_NOTYET;
5168 /* Disable header prediction. */
5172 /* This is the 'fast' part of urgent handling. */
5173 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5175 struct tcp_sock *tp = tcp_sk(sk);
5177 /* Check if we get a new urgent pointer - normally not. */
5179 tcp_check_urg(sk, th);
5181 /* Do we wait for any urgent data? - normally not... */
5182 if (tp->urg_data == TCP_URG_NOTYET) {
5183 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5186 /* Is the urgent pointer pointing into this packet? */
5187 if (ptr < skb->len) {
5189 if (skb_copy_bits(skb, ptr, &tmp, 1))
5191 tp->urg_data = TCP_URG_VALID | tmp;
5192 if (!sock_flag(sk, SOCK_DEAD))
5193 sk->sk_data_ready(sk, 0);
5198 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5200 struct tcp_sock *tp = tcp_sk(sk);
5201 int chunk = skb->len - hlen;
5205 if (skb_csum_unnecessary(skb))
5206 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5208 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5212 tp->ucopy.len -= chunk;
5213 tp->copied_seq += chunk;
5214 tcp_rcv_space_adjust(sk);
5221 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5222 struct sk_buff *skb)
5226 if (sock_owned_by_user(sk)) {
5228 result = __tcp_checksum_complete(skb);
5231 result = __tcp_checksum_complete(skb);
5236 static inline int tcp_checksum_complete_user(struct sock *sk,
5237 struct sk_buff *skb)
5239 return !skb_csum_unnecessary(skb) &&
5240 __tcp_checksum_complete_user(sk, skb);
5243 #ifdef CONFIG_NET_DMA
5244 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5247 struct tcp_sock *tp = tcp_sk(sk);
5248 int chunk = skb->len - hlen;
5250 int copied_early = 0;
5252 if (tp->ucopy.wakeup)
5255 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5256 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
5258 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5260 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5262 tp->ucopy.iov, chunk,
5263 tp->ucopy.pinned_list);
5268 tp->ucopy.dma_cookie = dma_cookie;
5271 tp->ucopy.len -= chunk;
5272 tp->copied_seq += chunk;
5273 tcp_rcv_space_adjust(sk);
5275 if ((tp->ucopy.len == 0) ||
5276 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5277 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5278 tp->ucopy.wakeup = 1;
5279 sk->sk_data_ready(sk, 0);
5281 } else if (chunk > 0) {
5282 tp->ucopy.wakeup = 1;
5283 sk->sk_data_ready(sk, 0);
5286 return copied_early;
5288 #endif /* CONFIG_NET_DMA */
5290 /* Does PAWS and seqno based validation of an incoming segment, flags will
5291 * play significant role here.
5293 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5294 const struct tcphdr *th, int syn_inerr)
5296 const u8 *hash_location;
5297 struct tcp_sock *tp = tcp_sk(sk);
5299 /* RFC1323: H1. Apply PAWS check first. */
5300 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5301 tp->rx_opt.saw_tstamp &&
5302 tcp_paws_discard(sk, skb)) {
5304 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5305 tcp_send_dupack(sk, skb);
5308 /* Reset is accepted even if it did not pass PAWS. */
5311 /* Step 1: check sequence number */
5312 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5313 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5314 * (RST) segments are validated by checking their SEQ-fields."
5315 * And page 69: "If an incoming segment is not acceptable,
5316 * an acknowledgment should be sent in reply (unless the RST
5317 * bit is set, if so drop the segment and return)".
5322 tcp_send_dupack(sk, skb);
5327 /* Step 2: check RST bit */
5330 * If sequence number exactly matches RCV.NXT, then
5331 * RESET the connection
5333 * Send a challenge ACK
5335 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5338 tcp_send_challenge_ack(sk);
5342 /* step 3: check security and precedence [ignored] */
5344 /* step 4: Check for a SYN
5345 * RFC 5691 4.2 : Send a challenge ack
5350 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5351 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5352 tcp_send_challenge_ack(sk);
5364 * TCP receive function for the ESTABLISHED state.
5366 * It is split into a fast path and a slow path. The fast path is
5368 * - A zero window was announced from us - zero window probing
5369 * is only handled properly in the slow path.
5370 * - Out of order segments arrived.
5371 * - Urgent data is expected.
5372 * - There is no buffer space left
5373 * - Unexpected TCP flags/window values/header lengths are received
5374 * (detected by checking the TCP header against pred_flags)
5375 * - Data is sent in both directions. Fast path only supports pure senders
5376 * or pure receivers (this means either the sequence number or the ack
5377 * value must stay constant)
5378 * - Unexpected TCP option.
5380 * When these conditions are not satisfied it drops into a standard
5381 * receive procedure patterned after RFC793 to handle all cases.
5382 * The first three cases are guaranteed by proper pred_flags setting,
5383 * the rest is checked inline. Fast processing is turned on in
5384 * tcp_data_queue when everything is OK.
5386 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5387 const struct tcphdr *th, unsigned int len)
5389 struct tcp_sock *tp = tcp_sk(sk);
5392 * Header prediction.
5393 * The code loosely follows the one in the famous
5394 * "30 instruction TCP receive" Van Jacobson mail.
5396 * Van's trick is to deposit buffers into socket queue
5397 * on a device interrupt, to call tcp_recv function
5398 * on the receive process context and checksum and copy
5399 * the buffer to user space. smart...
5401 * Our current scheme is not silly either but we take the
5402 * extra cost of the net_bh soft interrupt processing...
5403 * We do checksum and copy also but from device to kernel.
5406 tp->rx_opt.saw_tstamp = 0;
5408 /* pred_flags is 0xS?10 << 16 + snd_wnd
5409 * if header_prediction is to be made
5410 * 'S' will always be tp->tcp_header_len >> 2
5411 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5412 * turn it off (when there are holes in the receive
5413 * space for instance)
5414 * PSH flag is ignored.
5417 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5418 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5419 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5420 int tcp_header_len = tp->tcp_header_len;
5422 /* Timestamp header prediction: tcp_header_len
5423 * is automatically equal to th->doff*4 due to pred_flags
5427 /* Check timestamp */
5428 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5429 /* No? Slow path! */
5430 if (!tcp_parse_aligned_timestamp(tp, th))
5433 /* If PAWS failed, check it more carefully in slow path */
5434 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5437 /* DO NOT update ts_recent here, if checksum fails
5438 * and timestamp was corrupted part, it will result
5439 * in a hung connection since we will drop all
5440 * future packets due to the PAWS test.
5444 if (len <= tcp_header_len) {
5445 /* Bulk data transfer: sender */
5446 if (len == tcp_header_len) {
5447 /* Predicted packet is in window by definition.
5448 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5449 * Hence, check seq<=rcv_wup reduces to:
5451 if (tcp_header_len ==
5452 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5453 tp->rcv_nxt == tp->rcv_wup)
5454 tcp_store_ts_recent(tp);
5456 /* We know that such packets are checksummed
5459 tcp_ack(sk, skb, 0);
5461 tcp_data_snd_check(sk);
5463 } else { /* Header too small */
5464 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5469 int copied_early = 0;
5471 if (tp->copied_seq == tp->rcv_nxt &&
5472 len - tcp_header_len <= tp->ucopy.len) {
5473 #ifdef CONFIG_NET_DMA
5474 if (tp->ucopy.task == current &&
5475 sock_owned_by_user(sk) &&
5476 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5481 if (tp->ucopy.task == current &&
5482 sock_owned_by_user(sk) && !copied_early) {
5483 __set_current_state(TASK_RUNNING);
5485 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5489 /* Predicted packet is in window by definition.
5490 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5491 * Hence, check seq<=rcv_wup reduces to:
5493 if (tcp_header_len ==
5494 (sizeof(struct tcphdr) +
5495 TCPOLEN_TSTAMP_ALIGNED) &&
5496 tp->rcv_nxt == tp->rcv_wup)
5497 tcp_store_ts_recent(tp);
5499 tcp_rcv_rtt_measure_ts(sk, skb);
5501 __skb_pull(skb, tcp_header_len);
5502 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5503 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5506 tcp_cleanup_rbuf(sk, skb->len);
5509 if (tcp_checksum_complete_user(sk, skb))
5512 if ((int)skb->truesize > sk->sk_forward_alloc)
5515 /* Predicted packet is in window by definition.
5516 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5517 * Hence, check seq<=rcv_wup reduces to:
5519 if (tcp_header_len ==
5520 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5521 tp->rcv_nxt == tp->rcv_wup)
5522 tcp_store_ts_recent(tp);
5524 tcp_rcv_rtt_measure_ts(sk, skb);
5526 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5528 /* Bulk data transfer: receiver */
5529 __skb_pull(skb, tcp_header_len);
5530 __skb_queue_tail(&sk->sk_receive_queue, skb);
5531 skb_set_owner_r(skb, sk);
5532 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5535 tcp_event_data_recv(sk, skb);
5537 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5538 /* Well, only one small jumplet in fast path... */
5539 tcp_ack(sk, skb, FLAG_DATA);
5540 tcp_data_snd_check(sk);
5541 if (!inet_csk_ack_scheduled(sk))
5545 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5546 __tcp_ack_snd_check(sk, 0);
5548 #ifdef CONFIG_NET_DMA
5550 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5556 sk->sk_data_ready(sk, 0);
5562 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5566 * Standard slow path.
5569 if (!tcp_validate_incoming(sk, skb, th, 1))
5574 tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5577 tcp_rcv_rtt_measure_ts(sk, skb);
5579 /* Process urgent data. */
5580 tcp_urg(sk, skb, th);
5582 /* step 7: process the segment text */
5583 tcp_data_queue(sk, skb);
5585 tcp_data_snd_check(sk);
5586 tcp_ack_snd_check(sk);
5590 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5596 EXPORT_SYMBOL(tcp_rcv_established);
5598 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5599 const struct tcphdr *th, unsigned int len)
5601 const u8 *hash_location;
5602 struct inet_connection_sock *icsk = inet_csk(sk);
5603 struct tcp_sock *tp = tcp_sk(sk);
5604 struct tcp_cookie_values *cvp = tp->cookie_values;
5605 int saved_clamp = tp->rx_opt.mss_clamp;
5607 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5611 * "If the state is SYN-SENT then
5612 * first check the ACK bit
5613 * If the ACK bit is set
5614 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5615 * a reset (unless the RST bit is set, if so drop
5616 * the segment and return)"
5618 * We do not send data with SYN, so that RFC-correct
5621 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5622 goto reset_and_undo;
5624 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5625 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5627 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5628 goto reset_and_undo;
5631 /* Now ACK is acceptable.
5633 * "If the RST bit is set
5634 * If the ACK was acceptable then signal the user "error:
5635 * connection reset", drop the segment, enter CLOSED state,
5636 * delete TCB, and return."
5645 * "fifth, if neither of the SYN or RST bits is set then
5646 * drop the segment and return."
5652 goto discard_and_undo;
5655 * "If the SYN bit is on ...
5656 * are acceptable then ...
5657 * (our SYN has been ACKed), change the connection
5658 * state to ESTABLISHED..."
5661 TCP_ECN_rcv_synack(tp, th);
5663 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5664 tcp_ack(sk, skb, FLAG_SLOWPATH);
5666 /* Ok.. it's good. Set up sequence numbers and
5667 * move to established.
5669 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5670 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5672 /* RFC1323: The window in SYN & SYN/ACK segments is
5675 tp->snd_wnd = ntohs(th->window);
5676 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5678 if (!tp->rx_opt.wscale_ok) {
5679 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5680 tp->window_clamp = min(tp->window_clamp, 65535U);
5683 if (tp->rx_opt.saw_tstamp) {
5684 tp->rx_opt.tstamp_ok = 1;
5685 tp->tcp_header_len =
5686 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5687 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5688 tcp_store_ts_recent(tp);
5690 tp->tcp_header_len = sizeof(struct tcphdr);
5693 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5694 tcp_enable_fack(tp);
5697 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5698 tcp_initialize_rcv_mss(sk);
5700 /* Remember, tcp_poll() does not lock socket!
5701 * Change state from SYN-SENT only after copied_seq
5702 * is initialized. */
5703 tp->copied_seq = tp->rcv_nxt;
5706 cvp->cookie_pair_size > 0 &&
5707 tp->rx_opt.cookie_plus > 0) {
5708 int cookie_size = tp->rx_opt.cookie_plus
5709 - TCPOLEN_COOKIE_BASE;
5710 int cookie_pair_size = cookie_size
5711 + cvp->cookie_desired;
5713 /* A cookie extension option was sent and returned.
5714 * Note that each incoming SYNACK replaces the
5715 * Responder cookie. The initial exchange is most
5716 * fragile, as protection against spoofing relies
5717 * entirely upon the sequence and timestamp (above).
5718 * This replacement strategy allows the correct pair to
5719 * pass through, while any others will be filtered via
5720 * Responder verification later.
5722 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5723 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5724 hash_location, cookie_size);
5725 cvp->cookie_pair_size = cookie_pair_size;
5730 tcp_set_state(sk, TCP_ESTABLISHED);
5732 security_inet_conn_established(sk, skb);
5734 /* Make sure socket is routed, for correct metrics. */
5735 icsk->icsk_af_ops->rebuild_header(sk);
5737 tcp_init_metrics(sk);
5739 tcp_init_congestion_control(sk);
5741 /* Prevent spurious tcp_cwnd_restart() on first data
5744 tp->lsndtime = tcp_time_stamp;
5746 tcp_init_buffer_space(sk);
5748 if (sock_flag(sk, SOCK_KEEPOPEN))
5749 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5751 if (!tp->rx_opt.snd_wscale)
5752 __tcp_fast_path_on(tp, tp->snd_wnd);
5756 if (!sock_flag(sk, SOCK_DEAD)) {
5757 sk->sk_state_change(sk);
5758 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5761 if (sk->sk_write_pending ||
5762 icsk->icsk_accept_queue.rskq_defer_accept ||
5763 icsk->icsk_ack.pingpong) {
5764 /* Save one ACK. Data will be ready after
5765 * several ticks, if write_pending is set.
5767 * It may be deleted, but with this feature tcpdumps
5768 * look so _wonderfully_ clever, that I was not able
5769 * to stand against the temptation 8) --ANK
5771 inet_csk_schedule_ack(sk);
5772 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5773 icsk->icsk_ack.ato = TCP_ATO_MIN;
5774 tcp_incr_quickack(sk);
5775 tcp_enter_quickack_mode(sk);
5776 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5777 TCP_DELACK_MAX, TCP_RTO_MAX);
5788 /* No ACK in the segment */
5792 * "If the RST bit is set
5794 * Otherwise (no ACK) drop the segment and return."
5797 goto discard_and_undo;
5801 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5802 tcp_paws_reject(&tp->rx_opt, 0))
5803 goto discard_and_undo;
5806 /* We see SYN without ACK. It is attempt of
5807 * simultaneous connect with crossed SYNs.
5808 * Particularly, it can be connect to self.
5810 tcp_set_state(sk, TCP_SYN_RECV);
5812 if (tp->rx_opt.saw_tstamp) {
5813 tp->rx_opt.tstamp_ok = 1;
5814 tcp_store_ts_recent(tp);
5815 tp->tcp_header_len =
5816 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5818 tp->tcp_header_len = sizeof(struct tcphdr);
5821 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5822 tp->copied_seq = tp->rcv_nxt;
5823 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5825 /* RFC1323: The window in SYN & SYN/ACK segments is
5828 tp->snd_wnd = ntohs(th->window);
5829 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5830 tp->max_window = tp->snd_wnd;
5832 TCP_ECN_rcv_syn(tp, th);
5835 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5836 tcp_initialize_rcv_mss(sk);
5838 tcp_send_synack(sk);
5840 /* Note, we could accept data and URG from this segment.
5841 * There are no obstacles to make this.
5843 * However, if we ignore data in ACKless segments sometimes,
5844 * we have no reasons to accept it sometimes.
5845 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5846 * is not flawless. So, discard packet for sanity.
5847 * Uncomment this return to process the data.
5854 /* "fifth, if neither of the SYN or RST bits is set then
5855 * drop the segment and return."
5859 tcp_clear_options(&tp->rx_opt);
5860 tp->rx_opt.mss_clamp = saved_clamp;
5864 tcp_clear_options(&tp->rx_opt);
5865 tp->rx_opt.mss_clamp = saved_clamp;
5870 * This function implements the receiving procedure of RFC 793 for
5871 * all states except ESTABLISHED and TIME_WAIT.
5872 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5873 * address independent.
5876 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5877 const struct tcphdr *th, unsigned int len)
5879 struct tcp_sock *tp = tcp_sk(sk);
5880 struct inet_connection_sock *icsk = inet_csk(sk);
5883 tp->rx_opt.saw_tstamp = 0;
5885 switch (sk->sk_state) {
5899 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5902 /* Now we have several options: In theory there is
5903 * nothing else in the frame. KA9Q has an option to
5904 * send data with the syn, BSD accepts data with the
5905 * syn up to the [to be] advertised window and
5906 * Solaris 2.1 gives you a protocol error. For now
5907 * we just ignore it, that fits the spec precisely
5908 * and avoids incompatibilities. It would be nice in
5909 * future to drop through and process the data.
5911 * Now that TTCP is starting to be used we ought to
5913 * But, this leaves one open to an easy denial of
5914 * service attack, and SYN cookies can't defend
5915 * against this problem. So, we drop the data
5916 * in the interest of security over speed unless
5917 * it's still in use.
5925 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5929 /* Do step6 onward by hand. */
5930 tcp_urg(sk, skb, th);
5932 tcp_data_snd_check(sk);
5936 if (!tcp_validate_incoming(sk, skb, th, 0))
5939 /* step 5: check the ACK field */
5941 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5942 FLAG_UPDATE_TS_RECENT) > 0;
5944 switch (sk->sk_state) {
5947 tp->copied_seq = tp->rcv_nxt;
5949 tcp_set_state(sk, TCP_ESTABLISHED);
5950 sk->sk_state_change(sk);
5952 /* Note, that this wakeup is only for marginal
5953 * crossed SYN case. Passively open sockets
5954 * are not waked up, because sk->sk_sleep ==
5955 * NULL and sk->sk_socket == NULL.
5959 SOCK_WAKE_IO, POLL_OUT);
5961 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5962 tp->snd_wnd = ntohs(th->window) <<
5963 tp->rx_opt.snd_wscale;
5964 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5966 if (tp->rx_opt.tstamp_ok)
5967 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5969 /* Make sure socket is routed, for
5972 icsk->icsk_af_ops->rebuild_header(sk);
5974 tcp_init_metrics(sk);
5976 tcp_init_congestion_control(sk);
5978 /* Prevent spurious tcp_cwnd_restart() on
5979 * first data packet.
5981 tp->lsndtime = tcp_time_stamp;
5984 tcp_initialize_rcv_mss(sk);
5985 tcp_init_buffer_space(sk);
5986 tcp_fast_path_on(tp);
5993 if (tp->snd_una == tp->write_seq) {
5994 tcp_set_state(sk, TCP_FIN_WAIT2);
5995 sk->sk_shutdown |= SEND_SHUTDOWN;
5996 dst_confirm(__sk_dst_get(sk));
5998 if (!sock_flag(sk, SOCK_DEAD))
5999 /* Wake up lingering close() */
6000 sk->sk_state_change(sk);
6004 if (tp->linger2 < 0 ||
6005 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6006 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6008 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6012 tmo = tcp_fin_time(sk);
6013 if (tmo > TCP_TIMEWAIT_LEN) {
6014 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6015 } else if (th->fin || sock_owned_by_user(sk)) {
6016 /* Bad case. We could lose such FIN otherwise.
6017 * It is not a big problem, but it looks confusing
6018 * and not so rare event. We still can lose it now,
6019 * if it spins in bh_lock_sock(), but it is really
6022 inet_csk_reset_keepalive_timer(sk, tmo);
6024 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6032 if (tp->snd_una == tp->write_seq) {
6033 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6039 if (tp->snd_una == tp->write_seq) {
6040 tcp_update_metrics(sk);
6049 /* step 6: check the URG bit */
6050 tcp_urg(sk, skb, th);
6052 /* step 7: process the segment text */
6053 switch (sk->sk_state) {
6054 case TCP_CLOSE_WAIT:
6057 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6061 /* RFC 793 says to queue data in these states,
6062 * RFC 1122 says we MUST send a reset.
6063 * BSD 4.4 also does reset.
6065 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6066 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6067 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6068 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6074 case TCP_ESTABLISHED:
6075 tcp_data_queue(sk, skb);
6080 /* tcp_data could move socket to TIME-WAIT */
6081 if (sk->sk_state != TCP_CLOSE) {
6082 tcp_data_snd_check(sk);
6083 tcp_ack_snd_check(sk);
6092 EXPORT_SYMBOL(tcp_rcv_state_process);