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).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presnce of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
66 #include <linux/config.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
75 int sysctl_tcp_timestamps
= 1;
76 int sysctl_tcp_window_scaling
= 1;
77 int sysctl_tcp_sack
= 1;
78 int sysctl_tcp_fack
= 1;
79 int sysctl_tcp_reordering
= TCP_FASTRETRANS_THRESH
;
81 int sysctl_tcp_dsack
= 1;
82 int sysctl_tcp_app_win
= 31;
83 int sysctl_tcp_adv_win_scale
= 2;
85 int sysctl_tcp_stdurg
;
86 int sysctl_tcp_rfc1337
;
87 int sysctl_tcp_max_orphans
= NR_FILE
;
89 int sysctl_tcp_nometrics_save
;
91 int sysctl_tcp_moderate_rcvbuf
= 1;
93 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
94 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
95 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
96 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
97 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
98 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
99 #define FLAG_ECE 0x40 /* ECE in this ACK */
100 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
101 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
106 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
108 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
109 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
110 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
112 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
114 /* Adapt the MSS value used to make delayed ack decision to the
117 static inline void tcp_measure_rcv_mss(struct sock
*sk
,
118 const struct sk_buff
*skb
)
120 struct inet_connection_sock
*icsk
= inet_csk(sk
);
121 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
124 icsk
->icsk_ack
.last_seg_size
= 0;
126 /* skb->len may jitter because of SACKs, even if peer
127 * sends good full-sized frames.
130 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
131 icsk
->icsk_ack
.rcv_mss
= len
;
133 /* Otherwise, we make more careful check taking into account,
134 * that SACKs block is variable.
136 * "len" is invariant segment length, including TCP header.
138 len
+= skb
->data
- skb
->h
.raw
;
139 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
140 /* If PSH is not set, packet should be
141 * full sized, provided peer TCP is not badly broken.
142 * This observation (if it is correct 8)) allows
143 * to handle super-low mtu links fairly.
145 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
146 !(tcp_flag_word(skb
->h
.th
)&TCP_REMNANT
))) {
147 /* Subtract also invariant (if peer is RFC compliant),
148 * tcp header plus fixed timestamp option length.
149 * Resulting "len" is MSS free of SACK jitter.
151 len
-= tcp_sk(sk
)->tcp_header_len
;
152 icsk
->icsk_ack
.last_seg_size
= len
;
154 icsk
->icsk_ack
.rcv_mss
= len
;
158 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
162 static void tcp_incr_quickack(struct sock
*sk
)
164 struct inet_connection_sock
*icsk
= inet_csk(sk
);
165 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
169 if (quickacks
> icsk
->icsk_ack
.quick
)
170 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
173 void tcp_enter_quickack_mode(struct sock
*sk
)
175 struct inet_connection_sock
*icsk
= inet_csk(sk
);
176 tcp_incr_quickack(sk
);
177 icsk
->icsk_ack
.pingpong
= 0;
178 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
181 /* Send ACKs quickly, if "quick" count is not exhausted
182 * and the session is not interactive.
185 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
187 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
188 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
191 /* Buffer size and advertised window tuning.
193 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
196 static void tcp_fixup_sndbuf(struct sock
*sk
)
198 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
199 sizeof(struct sk_buff
);
201 if (sk
->sk_sndbuf
< 3 * sndmem
)
202 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
205 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
207 * All tcp_full_space() is split to two parts: "network" buffer, allocated
208 * forward and advertised in receiver window (tp->rcv_wnd) and
209 * "application buffer", required to isolate scheduling/application
210 * latencies from network.
211 * window_clamp is maximal advertised window. It can be less than
212 * tcp_full_space(), in this case tcp_full_space() - window_clamp
213 * is reserved for "application" buffer. The less window_clamp is
214 * the smoother our behaviour from viewpoint of network, but the lower
215 * throughput and the higher sensitivity of the connection to losses. 8)
217 * rcv_ssthresh is more strict window_clamp used at "slow start"
218 * phase to predict further behaviour of this connection.
219 * It is used for two goals:
220 * - to enforce header prediction at sender, even when application
221 * requires some significant "application buffer". It is check #1.
222 * - to prevent pruning of receive queue because of misprediction
223 * of receiver window. Check #2.
225 * The scheme does not work when sender sends good segments opening
226 * window and then starts to feed us spagetti. But it should work
227 * in common situations. Otherwise, we have to rely on queue collapsing.
230 /* Slow part of check#2. */
231 static int __tcp_grow_window(const struct sock
*sk
, struct tcp_sock
*tp
,
232 const struct sk_buff
*skb
)
235 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
236 int window
= tcp_full_space(sk
)/2;
238 while (tp
->rcv_ssthresh
<= window
) {
239 if (truesize
<= skb
->len
)
240 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
248 static inline void tcp_grow_window(struct sock
*sk
, struct tcp_sock
*tp
,
252 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
253 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
254 !tcp_memory_pressure
) {
257 /* Check #2. Increase window, if skb with such overhead
258 * will fit to rcvbuf in future.
260 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
263 incr
= __tcp_grow_window(sk
, tp
, skb
);
266 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
267 inet_csk(sk
)->icsk_ack
.quick
|= 1;
272 /* 3. Tuning rcvbuf, when connection enters established state. */
274 static void tcp_fixup_rcvbuf(struct sock
*sk
)
276 struct tcp_sock
*tp
= tcp_sk(sk
);
277 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
279 /* Try to select rcvbuf so that 4 mss-sized segments
280 * will fit to window and correspoding skbs will fit to our rcvbuf.
281 * (was 3; 4 is minimum to allow fast retransmit to work.)
283 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
285 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
286 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
289 /* 4. Try to fixup all. It is made iimediately after connection enters
292 static void tcp_init_buffer_space(struct sock
*sk
)
294 struct tcp_sock
*tp
= tcp_sk(sk
);
297 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
298 tcp_fixup_rcvbuf(sk
);
299 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
300 tcp_fixup_sndbuf(sk
);
302 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
304 maxwin
= tcp_full_space(sk
);
306 if (tp
->window_clamp
>= maxwin
) {
307 tp
->window_clamp
= maxwin
;
309 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
310 tp
->window_clamp
= max(maxwin
-
311 (maxwin
>> sysctl_tcp_app_win
),
315 /* Force reservation of one segment. */
316 if (sysctl_tcp_app_win
&&
317 tp
->window_clamp
> 2 * tp
->advmss
&&
318 tp
->window_clamp
+ tp
->advmss
> maxwin
)
319 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
321 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
322 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
325 /* 5. Recalculate window clamp after socket hit its memory bounds. */
326 static void tcp_clamp_window(struct sock
*sk
, struct tcp_sock
*tp
)
328 struct inet_connection_sock
*icsk
= inet_csk(sk
);
330 unsigned int app_win
= tp
->rcv_nxt
- tp
->copied_seq
;
333 icsk
->icsk_ack
.quick
= 0;
335 skb_queue_walk(&tp
->out_of_order_queue
, skb
) {
339 /* If overcommit is due to out of order segments,
340 * do not clamp window. Try to expand rcvbuf instead.
343 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
344 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
345 !tcp_memory_pressure
&&
346 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0])
347 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
350 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
) {
352 if (atomic_read(&sk
->sk_rmem_alloc
) >= 2 * sk
->sk_rcvbuf
)
354 if (app_win
> icsk
->icsk_ack
.rcv_mss
)
355 app_win
-= icsk
->icsk_ack
.rcv_mss
;
356 app_win
= max(app_win
, 2U*tp
->advmss
);
358 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
362 /* Receiver "autotuning" code.
364 * The algorithm for RTT estimation w/o timestamps is based on
365 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
366 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
368 * More detail on this code can be found at
369 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
370 * though this reference is out of date. A new paper
373 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
375 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
381 if (new_sample
!= 0) {
382 /* If we sample in larger samples in the non-timestamp
383 * case, we could grossly overestimate the RTT especially
384 * with chatty applications or bulk transfer apps which
385 * are stalled on filesystem I/O.
387 * Also, since we are only going for a minimum in the
388 * non-timestamp case, we do not smoothe things out
389 * else with timestamps disabled convergance takes too
393 m
-= (new_sample
>> 3);
395 } else if (m
< new_sample
)
398 /* No previous mesaure. */
402 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
403 tp
->rcv_rtt_est
.rtt
= new_sample
;
406 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
408 if (tp
->rcv_rtt_est
.time
== 0)
410 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
412 tcp_rcv_rtt_update(tp
,
413 jiffies
- tp
->rcv_rtt_est
.time
,
417 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
418 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
421 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
423 struct tcp_sock
*tp
= tcp_sk(sk
);
424 if (tp
->rx_opt
.rcv_tsecr
&&
425 (TCP_SKB_CB(skb
)->end_seq
-
426 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
427 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
431 * This function should be called every time data is copied to user space.
432 * It calculates the appropriate TCP receive buffer space.
434 void tcp_rcv_space_adjust(struct sock
*sk
)
436 struct tcp_sock
*tp
= tcp_sk(sk
);
440 if (tp
->rcvq_space
.time
== 0)
443 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
444 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
445 tp
->rcv_rtt_est
.rtt
== 0)
448 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
450 space
= max(tp
->rcvq_space
.space
, space
);
452 if (tp
->rcvq_space
.space
!= space
) {
455 tp
->rcvq_space
.space
= space
;
457 if (sysctl_tcp_moderate_rcvbuf
) {
458 int new_clamp
= space
;
460 /* Receive space grows, normalize in order to
461 * take into account packet headers and sk_buff
462 * structure overhead.
467 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
468 16 + sizeof(struct sk_buff
));
469 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
472 space
= min(space
, sysctl_tcp_rmem
[2]);
473 if (space
> sk
->sk_rcvbuf
) {
474 sk
->sk_rcvbuf
= space
;
476 /* Make the window clamp follow along. */
477 tp
->window_clamp
= new_clamp
;
483 tp
->rcvq_space
.seq
= tp
->copied_seq
;
484 tp
->rcvq_space
.time
= tcp_time_stamp
;
487 /* There is something which you must keep in mind when you analyze the
488 * behavior of the tp->ato delayed ack timeout interval. When a
489 * connection starts up, we want to ack as quickly as possible. The
490 * problem is that "good" TCP's do slow start at the beginning of data
491 * transmission. The means that until we send the first few ACK's the
492 * sender will sit on his end and only queue most of his data, because
493 * he can only send snd_cwnd unacked packets at any given time. For
494 * each ACK we send, he increments snd_cwnd and transmits more of his
497 static void tcp_event_data_recv(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
499 struct inet_connection_sock
*icsk
= inet_csk(sk
);
502 inet_csk_schedule_ack(sk
);
504 tcp_measure_rcv_mss(sk
, skb
);
506 tcp_rcv_rtt_measure(tp
);
508 now
= tcp_time_stamp
;
510 if (!icsk
->icsk_ack
.ato
) {
511 /* The _first_ data packet received, initialize
512 * delayed ACK engine.
514 tcp_incr_quickack(sk
);
515 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
517 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
519 if (m
<= TCP_ATO_MIN
/2) {
520 /* The fastest case is the first. */
521 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
522 } else if (m
< icsk
->icsk_ack
.ato
) {
523 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
524 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
525 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
526 } else if (m
> icsk
->icsk_rto
) {
527 /* Too long gap. Apparently sender falled to
528 * restart window, so that we send ACKs quickly.
530 tcp_incr_quickack(sk
);
531 sk_stream_mem_reclaim(sk
);
534 icsk
->icsk_ack
.lrcvtime
= now
;
536 TCP_ECN_check_ce(tp
, skb
);
539 tcp_grow_window(sk
, tp
, skb
);
542 /* Called to compute a smoothed rtt estimate. The data fed to this
543 * routine either comes from timestamps, or from segments that were
544 * known _not_ to have been retransmitted [see Karn/Partridge
545 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
546 * piece by Van Jacobson.
547 * NOTE: the next three routines used to be one big routine.
548 * To save cycles in the RFC 1323 implementation it was better to break
549 * it up into three procedures. -- erics
551 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
553 struct tcp_sock
*tp
= tcp_sk(sk
);
554 long m
= mrtt
; /* RTT */
556 /* The following amusing code comes from Jacobson's
557 * article in SIGCOMM '88. Note that rtt and mdev
558 * are scaled versions of rtt and mean deviation.
559 * This is designed to be as fast as possible
560 * m stands for "measurement".
562 * On a 1990 paper the rto value is changed to:
563 * RTO = rtt + 4 * mdev
565 * Funny. This algorithm seems to be very broken.
566 * These formulae increase RTO, when it should be decreased, increase
567 * too slowly, when it should be incresed fastly, decrease too fastly
568 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
569 * does not matter how to _calculate_ it. Seems, it was trap
570 * that VJ failed to avoid. 8)
575 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
576 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
578 m
= -m
; /* m is now abs(error) */
579 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
580 /* This is similar to one of Eifel findings.
581 * Eifel blocks mdev updates when rtt decreases.
582 * This solution is a bit different: we use finer gain
583 * for mdev in this case (alpha*beta).
584 * Like Eifel it also prevents growth of rto,
585 * but also it limits too fast rto decreases,
586 * happening in pure Eifel.
591 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
593 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
594 if (tp
->mdev
> tp
->mdev_max
) {
595 tp
->mdev_max
= tp
->mdev
;
596 if (tp
->mdev_max
> tp
->rttvar
)
597 tp
->rttvar
= tp
->mdev_max
;
599 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
600 if (tp
->mdev_max
< tp
->rttvar
)
601 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
602 tp
->rtt_seq
= tp
->snd_nxt
;
603 tp
->mdev_max
= TCP_RTO_MIN
;
606 /* no previous measure. */
607 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
608 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
609 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
610 tp
->rtt_seq
= tp
->snd_nxt
;
614 /* Calculate rto without backoff. This is the second half of Van Jacobson's
615 * routine referred to above.
617 static inline void tcp_set_rto(struct sock
*sk
)
619 const struct tcp_sock
*tp
= tcp_sk(sk
);
620 /* Old crap is replaced with new one. 8)
623 * 1. If rtt variance happened to be less 50msec, it is hallucination.
624 * It cannot be less due to utterly erratic ACK generation made
625 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
626 * to do with delayed acks, because at cwnd>2 true delack timeout
627 * is invisible. Actually, Linux-2.4 also generates erratic
628 * ACKs in some curcumstances.
630 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
632 /* 2. Fixups made earlier cannot be right.
633 * If we do not estimate RTO correctly without them,
634 * all the algo is pure shit and should be replaced
635 * with correct one. It is exaclty, which we pretend to do.
639 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
640 * guarantees that rto is higher.
642 static inline void tcp_bound_rto(struct sock
*sk
)
644 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
645 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
648 /* Save metrics learned by this TCP session.
649 This function is called only, when TCP finishes successfully
650 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
652 void tcp_update_metrics(struct sock
*sk
)
654 struct tcp_sock
*tp
= tcp_sk(sk
);
655 struct dst_entry
*dst
= __sk_dst_get(sk
);
657 if (sysctl_tcp_nometrics_save
)
662 if (dst
&& (dst
->flags
&DST_HOST
)) {
663 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
666 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
667 /* This session failed to estimate rtt. Why?
668 * Probably, no packets returned in time.
671 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
672 dst
->metrics
[RTAX_RTT
-1] = 0;
676 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
678 /* If newly calculated rtt larger than stored one,
679 * store new one. Otherwise, use EWMA. Remember,
680 * rtt overestimation is always better than underestimation.
682 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
684 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
686 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
689 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
693 /* Scale deviation to rttvar fixed point */
698 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
699 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
701 dst
->metrics
[RTAX_RTTVAR
-1] -=
702 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
705 if (tp
->snd_ssthresh
>= 0xFFFF) {
706 /* Slow start still did not finish. */
707 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
708 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
709 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
710 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
711 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
712 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
713 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
714 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
715 icsk
->icsk_ca_state
== TCP_CA_Open
) {
716 /* Cong. avoidance phase, cwnd is reliable. */
717 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
718 dst
->metrics
[RTAX_SSTHRESH
-1] =
719 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
720 if (!dst_metric_locked(dst
, RTAX_CWND
))
721 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
723 /* Else slow start did not finish, cwnd is non-sense,
724 ssthresh may be also invalid.
726 if (!dst_metric_locked(dst
, RTAX_CWND
))
727 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
728 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
729 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
730 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
731 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
734 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
735 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
736 tp
->reordering
!= sysctl_tcp_reordering
)
737 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
742 /* Numbers are taken from RFC2414. */
743 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
745 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
748 if (tp
->mss_cache
> 1460)
751 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
753 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
756 /* Initialize metrics on socket. */
758 static void tcp_init_metrics(struct sock
*sk
)
760 struct tcp_sock
*tp
= tcp_sk(sk
);
761 struct dst_entry
*dst
= __sk_dst_get(sk
);
768 if (dst_metric_locked(dst
, RTAX_CWND
))
769 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
770 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
771 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
772 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
773 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
775 if (dst_metric(dst
, RTAX_REORDERING
) &&
776 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
777 tp
->rx_opt
.sack_ok
&= ~2;
778 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
781 if (dst_metric(dst
, RTAX_RTT
) == 0)
784 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
787 /* Initial rtt is determined from SYN,SYN-ACK.
788 * The segment is small and rtt may appear much
789 * less than real one. Use per-dst memory
790 * to make it more realistic.
792 * A bit of theory. RTT is time passed after "normal" sized packet
793 * is sent until it is ACKed. In normal curcumstances sending small
794 * packets force peer to delay ACKs and calculation is correct too.
795 * The algorithm is adaptive and, provided we follow specs, it
796 * NEVER underestimate RTT. BUT! If peer tries to make some clever
797 * tricks sort of "quick acks" for time long enough to decrease RTT
798 * to low value, and then abruptly stops to do it and starts to delay
799 * ACKs, wait for troubles.
801 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
802 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
803 tp
->rtt_seq
= tp
->snd_nxt
;
805 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
806 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
807 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
811 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
813 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
814 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
818 /* Play conservative. If timestamps are not
819 * supported, TCP will fail to recalculate correct
820 * rtt, if initial rto is too small. FORGET ALL AND RESET!
822 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
824 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
825 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
829 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
832 struct tcp_sock
*tp
= tcp_sk(sk
);
833 if (metric
> tp
->reordering
) {
834 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
836 /* This exciting event is worth to be remembered. 8) */
838 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
840 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
842 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
844 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
845 #if FASTRETRANS_DEBUG > 1
846 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
847 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
851 tp
->undo_marker
? tp
->undo_retrans
: 0);
853 /* Disable FACK yet. */
854 tp
->rx_opt
.sack_ok
&= ~2;
858 /* This procedure tags the retransmission queue when SACKs arrive.
860 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
861 * Packets in queue with these bits set are counted in variables
862 * sacked_out, retrans_out and lost_out, correspondingly.
864 * Valid combinations are:
865 * Tag InFlight Description
866 * 0 1 - orig segment is in flight.
867 * S 0 - nothing flies, orig reached receiver.
868 * L 0 - nothing flies, orig lost by net.
869 * R 2 - both orig and retransmit are in flight.
870 * L|R 1 - orig is lost, retransmit is in flight.
871 * S|R 1 - orig reached receiver, retrans is still in flight.
872 * (L|S|R is logically valid, it could occur when L|R is sacked,
873 * but it is equivalent to plain S and code short-curcuits it to S.
874 * L|S is logically invalid, it would mean -1 packet in flight 8))
876 * These 6 states form finite state machine, controlled by the following events:
877 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
878 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
879 * 3. Loss detection event of one of three flavors:
880 * A. Scoreboard estimator decided the packet is lost.
881 * A'. Reno "three dupacks" marks head of queue lost.
882 * A''. Its FACK modfication, head until snd.fack is lost.
883 * B. SACK arrives sacking data transmitted after never retransmitted
885 * C. SACK arrives sacking SND.NXT at the moment, when the
886 * segment was retransmitted.
887 * 4. D-SACK added new rule: D-SACK changes any tag to S.
889 * It is pleasant to note, that state diagram turns out to be commutative,
890 * so that we are allowed not to be bothered by order of our actions,
891 * when multiple events arrive simultaneously. (see the function below).
893 * Reordering detection.
894 * --------------------
895 * Reordering metric is maximal distance, which a packet can be displaced
896 * in packet stream. With SACKs we can estimate it:
898 * 1. SACK fills old hole and the corresponding segment was not
899 * ever retransmitted -> reordering. Alas, we cannot use it
900 * when segment was retransmitted.
901 * 2. The last flaw is solved with D-SACK. D-SACK arrives
902 * for retransmitted and already SACKed segment -> reordering..
903 * Both of these heuristics are not used in Loss state, when we cannot
904 * account for retransmits accurately.
907 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
909 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
910 struct tcp_sock
*tp
= tcp_sk(sk
);
911 unsigned char *ptr
= ack_skb
->h
.raw
+ TCP_SKB_CB(ack_skb
)->sacked
;
912 struct tcp_sack_block
*sp
= (struct tcp_sack_block
*)(ptr
+2);
913 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
914 int reord
= tp
->packets_out
;
916 u32 lost_retrans
= 0;
922 prior_fackets
= tp
->fackets_out
;
924 for (i
=0; i
<num_sacks
; i
++, sp
++) {
926 __u32 start_seq
= ntohl(sp
->start_seq
);
927 __u32 end_seq
= ntohl(sp
->end_seq
);
931 /* Check for D-SACK. */
933 u32 ack
= TCP_SKB_CB(ack_skb
)->ack_seq
;
935 if (before(start_seq
, ack
)) {
937 tp
->rx_opt
.sack_ok
|= 4;
938 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
939 } else if (num_sacks
> 1 &&
940 !after(end_seq
, ntohl(sp
[1].end_seq
)) &&
941 !before(start_seq
, ntohl(sp
[1].start_seq
))) {
943 tp
->rx_opt
.sack_ok
|= 4;
944 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
947 /* D-SACK for already forgotten data...
948 * Do dumb counting. */
950 !after(end_seq
, prior_snd_una
) &&
951 after(end_seq
, tp
->undo_marker
))
954 /* Eliminate too old ACKs, but take into
955 * account more or less fresh ones, they can
956 * contain valid SACK info.
958 if (before(ack
, prior_snd_una
- tp
->max_window
))
962 /* Event "B" in the comment above. */
963 if (after(end_seq
, tp
->high_seq
))
964 flag
|= FLAG_DATA_LOST
;
966 sk_stream_for_retrans_queue(skb
, sk
) {
970 /* The retransmission queue is always in order, so
971 * we can short-circuit the walk early.
973 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
976 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
977 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
979 pcount
= tcp_skb_pcount(skb
);
981 if (pcount
> 1 && !in_sack
&&
982 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
983 unsigned int pkt_len
;
985 in_sack
= !after(start_seq
,
986 TCP_SKB_CB(skb
)->seq
);
989 pkt_len
= (start_seq
-
990 TCP_SKB_CB(skb
)->seq
);
993 TCP_SKB_CB(skb
)->seq
);
994 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->tso_size
))
996 pcount
= tcp_skb_pcount(skb
);
999 fack_count
+= pcount
;
1001 sacked
= TCP_SKB_CB(skb
)->sacked
;
1003 /* Account D-SACK for retransmitted packet. */
1004 if ((dup_sack
&& in_sack
) &&
1005 (sacked
& TCPCB_RETRANS
) &&
1006 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1009 /* The frame is ACKed. */
1010 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1011 if (sacked
&TCPCB_RETRANS
) {
1012 if ((dup_sack
&& in_sack
) &&
1013 (sacked
&TCPCB_SACKED_ACKED
))
1014 reord
= min(fack_count
, reord
);
1016 /* If it was in a hole, we detected reordering. */
1017 if (fack_count
< prior_fackets
&&
1018 !(sacked
&TCPCB_SACKED_ACKED
))
1019 reord
= min(fack_count
, reord
);
1022 /* Nothing to do; acked frame is about to be dropped. */
1026 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1027 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1028 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1029 lost_retrans
= end_seq
;
1034 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1035 if (sacked
& TCPCB_SACKED_RETRANS
) {
1036 /* If the segment is not tagged as lost,
1037 * we do not clear RETRANS, believing
1038 * that retransmission is still in flight.
1040 if (sacked
& TCPCB_LOST
) {
1041 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1042 tp
->lost_out
-= tcp_skb_pcount(skb
);
1043 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1046 /* New sack for not retransmitted frame,
1047 * which was in hole. It is reordering.
1049 if (!(sacked
& TCPCB_RETRANS
) &&
1050 fack_count
< prior_fackets
)
1051 reord
= min(fack_count
, reord
);
1053 if (sacked
& TCPCB_LOST
) {
1054 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1055 tp
->lost_out
-= tcp_skb_pcount(skb
);
1059 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1060 flag
|= FLAG_DATA_SACKED
;
1061 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1063 if (fack_count
> tp
->fackets_out
)
1064 tp
->fackets_out
= fack_count
;
1066 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1067 reord
= min(fack_count
, reord
);
1070 /* D-SACK. We can detect redundant retransmission
1071 * in S|R and plain R frames and clear it.
1072 * undo_retrans is decreased above, L|R frames
1073 * are accounted above as well.
1076 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1077 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1078 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1083 /* Check for lost retransmit. This superb idea is
1084 * borrowed from "ratehalving". Event "C".
1085 * Later note: FACK people cheated me again 8),
1086 * we have to account for reordering! Ugly,
1089 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1090 struct sk_buff
*skb
;
1092 sk_stream_for_retrans_queue(skb
, sk
) {
1093 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1095 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1097 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1098 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1100 !before(lost_retrans
,
1101 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1103 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1104 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1106 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1107 tp
->lost_out
+= tcp_skb_pcount(skb
);
1108 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1109 flag
|= FLAG_DATA_SACKED
;
1110 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1116 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1118 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1119 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1121 #if FASTRETRANS_DEBUG > 0
1122 BUG_TRAP((int)tp
->sacked_out
>= 0);
1123 BUG_TRAP((int)tp
->lost_out
>= 0);
1124 BUG_TRAP((int)tp
->retrans_out
>= 0);
1125 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1130 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1131 * segments to see from the next ACKs whether any data was really missing.
1132 * If the RTO was spurious, new ACKs should arrive.
1134 void tcp_enter_frto(struct sock
*sk
)
1136 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1137 struct tcp_sock
*tp
= tcp_sk(sk
);
1138 struct sk_buff
*skb
;
1140 tp
->frto_counter
= 1;
1142 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1143 tp
->snd_una
== tp
->high_seq
||
1144 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1145 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1146 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1147 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1150 /* Have to clear retransmission markers here to keep the bookkeeping
1151 * in shape, even though we are not yet in Loss state.
1152 * If something was really lost, it is eventually caught up
1153 * in tcp_enter_frto_loss.
1155 tp
->retrans_out
= 0;
1156 tp
->undo_marker
= tp
->snd_una
;
1157 tp
->undo_retrans
= 0;
1159 sk_stream_for_retrans_queue(skb
, sk
) {
1160 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_RETRANS
;
1162 tcp_sync_left_out(tp
);
1164 tcp_set_ca_state(sk
, TCP_CA_Open
);
1165 tp
->frto_highmark
= tp
->snd_nxt
;
1168 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1169 * which indicates that we should follow the traditional RTO recovery,
1170 * i.e. mark everything lost and do go-back-N retransmission.
1172 static void tcp_enter_frto_loss(struct sock
*sk
)
1174 struct tcp_sock
*tp
= tcp_sk(sk
);
1175 struct sk_buff
*skb
;
1180 tp
->fackets_out
= 0;
1182 sk_stream_for_retrans_queue(skb
, sk
) {
1183 cnt
+= tcp_skb_pcount(skb
);
1184 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1185 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1187 /* Do not mark those segments lost that were
1188 * forward transmitted after RTO
1190 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1191 tp
->frto_highmark
)) {
1192 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1193 tp
->lost_out
+= tcp_skb_pcount(skb
);
1196 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1197 tp
->fackets_out
= cnt
;
1200 tcp_sync_left_out(tp
);
1202 tp
->snd_cwnd
= tp
->frto_counter
+ tcp_packets_in_flight(tp
)+1;
1203 tp
->snd_cwnd_cnt
= 0;
1204 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1205 tp
->undo_marker
= 0;
1206 tp
->frto_counter
= 0;
1208 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1209 sysctl_tcp_reordering
);
1210 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1211 tp
->high_seq
= tp
->frto_highmark
;
1212 TCP_ECN_queue_cwr(tp
);
1215 void tcp_clear_retrans(struct tcp_sock
*tp
)
1218 tp
->retrans_out
= 0;
1220 tp
->fackets_out
= 0;
1224 tp
->undo_marker
= 0;
1225 tp
->undo_retrans
= 0;
1228 /* Enter Loss state. If "how" is not zero, forget all SACK information
1229 * and reset tags completely, otherwise preserve SACKs. If receiver
1230 * dropped its ofo queue, we will know this due to reneging detection.
1232 void tcp_enter_loss(struct sock
*sk
, int how
)
1234 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1235 struct tcp_sock
*tp
= tcp_sk(sk
);
1236 struct sk_buff
*skb
;
1239 /* Reduce ssthresh if it has not yet been made inside this window. */
1240 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1241 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1242 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1243 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1244 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1247 tp
->snd_cwnd_cnt
= 0;
1248 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1250 tcp_clear_retrans(tp
);
1252 /* Push undo marker, if it was plain RTO and nothing
1253 * was retransmitted. */
1255 tp
->undo_marker
= tp
->snd_una
;
1257 sk_stream_for_retrans_queue(skb
, sk
) {
1258 cnt
+= tcp_skb_pcount(skb
);
1259 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1260 tp
->undo_marker
= 0;
1261 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1262 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1263 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1264 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1265 tp
->lost_out
+= tcp_skb_pcount(skb
);
1267 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1268 tp
->fackets_out
= cnt
;
1271 tcp_sync_left_out(tp
);
1273 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1274 sysctl_tcp_reordering
);
1275 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1276 tp
->high_seq
= tp
->snd_nxt
;
1277 TCP_ECN_queue_cwr(tp
);
1280 static int tcp_check_sack_reneging(struct sock
*sk
)
1282 struct sk_buff
*skb
;
1284 /* If ACK arrived pointing to a remembered SACK,
1285 * it means that our remembered SACKs do not reflect
1286 * real state of receiver i.e.
1287 * receiver _host_ is heavily congested (or buggy).
1288 * Do processing similar to RTO timeout.
1290 if ((skb
= skb_peek(&sk
->sk_write_queue
)) != NULL
&&
1291 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1292 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1293 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1295 tcp_enter_loss(sk
, 1);
1296 icsk
->icsk_retransmits
++;
1297 tcp_retransmit_skb(sk
, skb_peek(&sk
->sk_write_queue
));
1298 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1299 icsk
->icsk_rto
, TCP_RTO_MAX
);
1305 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1307 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1310 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1312 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1315 static inline int tcp_head_timedout(struct sock
*sk
, struct tcp_sock
*tp
)
1317 return tp
->packets_out
&&
1318 tcp_skb_timedout(sk
, skb_peek(&sk
->sk_write_queue
));
1321 /* Linux NewReno/SACK/FACK/ECN state machine.
1322 * --------------------------------------
1324 * "Open" Normal state, no dubious events, fast path.
1325 * "Disorder" In all the respects it is "Open",
1326 * but requires a bit more attention. It is entered when
1327 * we see some SACKs or dupacks. It is split of "Open"
1328 * mainly to move some processing from fast path to slow one.
1329 * "CWR" CWND was reduced due to some Congestion Notification event.
1330 * It can be ECN, ICMP source quench, local device congestion.
1331 * "Recovery" CWND was reduced, we are fast-retransmitting.
1332 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1334 * tcp_fastretrans_alert() is entered:
1335 * - each incoming ACK, if state is not "Open"
1336 * - when arrived ACK is unusual, namely:
1341 * Counting packets in flight is pretty simple.
1343 * in_flight = packets_out - left_out + retrans_out
1345 * packets_out is SND.NXT-SND.UNA counted in packets.
1347 * retrans_out is number of retransmitted segments.
1349 * left_out is number of segments left network, but not ACKed yet.
1351 * left_out = sacked_out + lost_out
1353 * sacked_out: Packets, which arrived to receiver out of order
1354 * and hence not ACKed. With SACKs this number is simply
1355 * amount of SACKed data. Even without SACKs
1356 * it is easy to give pretty reliable estimate of this number,
1357 * counting duplicate ACKs.
1359 * lost_out: Packets lost by network. TCP has no explicit
1360 * "loss notification" feedback from network (for now).
1361 * It means that this number can be only _guessed_.
1362 * Actually, it is the heuristics to predict lossage that
1363 * distinguishes different algorithms.
1365 * F.e. after RTO, when all the queue is considered as lost,
1366 * lost_out = packets_out and in_flight = retrans_out.
1368 * Essentially, we have now two algorithms counting
1371 * FACK: It is the simplest heuristics. As soon as we decided
1372 * that something is lost, we decide that _all_ not SACKed
1373 * packets until the most forward SACK are lost. I.e.
1374 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1375 * It is absolutely correct estimate, if network does not reorder
1376 * packets. And it loses any connection to reality when reordering
1377 * takes place. We use FACK by default until reordering
1378 * is suspected on the path to this destination.
1380 * NewReno: when Recovery is entered, we assume that one segment
1381 * is lost (classic Reno). While we are in Recovery and
1382 * a partial ACK arrives, we assume that one more packet
1383 * is lost (NewReno). This heuristics are the same in NewReno
1386 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1387 * deflation etc. CWND is real congestion window, never inflated, changes
1388 * only according to classic VJ rules.
1390 * Really tricky (and requiring careful tuning) part of algorithm
1391 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1392 * The first determines the moment _when_ we should reduce CWND and,
1393 * hence, slow down forward transmission. In fact, it determines the moment
1394 * when we decide that hole is caused by loss, rather than by a reorder.
1396 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1397 * holes, caused by lost packets.
1399 * And the most logically complicated part of algorithm is undo
1400 * heuristics. We detect false retransmits due to both too early
1401 * fast retransmit (reordering) and underestimated RTO, analyzing
1402 * timestamps and D-SACKs. When we detect that some segments were
1403 * retransmitted by mistake and CWND reduction was wrong, we undo
1404 * window reduction and abort recovery phase. This logic is hidden
1405 * inside several functions named tcp_try_undo_<something>.
1408 /* This function decides, when we should leave Disordered state
1409 * and enter Recovery phase, reducing congestion window.
1411 * Main question: may we further continue forward transmission
1412 * with the same cwnd?
1414 static int tcp_time_to_recover(struct sock
*sk
, struct tcp_sock
*tp
)
1418 /* Trick#1: The loss is proven. */
1422 /* Not-A-Trick#2 : Classic rule... */
1423 if (tcp_fackets_out(tp
) > tp
->reordering
)
1426 /* Trick#3 : when we use RFC2988 timer restart, fast
1427 * retransmit can be triggered by timeout of queue head.
1429 if (tcp_head_timedout(sk
, tp
))
1432 /* Trick#4: It is still not OK... But will it be useful to delay
1435 packets_out
= tp
->packets_out
;
1436 if (packets_out
<= tp
->reordering
&&
1437 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1438 !tcp_may_send_now(sk
, tp
)) {
1439 /* We have nothing to send. This connection is limited
1440 * either by receiver window or by application.
1448 /* If we receive more dupacks than we expected counting segments
1449 * in assumption of absent reordering, interpret this as reordering.
1450 * The only another reason could be bug in receiver TCP.
1452 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1454 struct tcp_sock
*tp
= tcp_sk(sk
);
1457 holes
= max(tp
->lost_out
, 1U);
1458 holes
= min(holes
, tp
->packets_out
);
1460 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1461 tp
->sacked_out
= tp
->packets_out
- holes
;
1462 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1466 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1468 static void tcp_add_reno_sack(struct sock
*sk
)
1470 struct tcp_sock
*tp
= tcp_sk(sk
);
1472 tcp_check_reno_reordering(sk
, 0);
1473 tcp_sync_left_out(tp
);
1476 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1478 static void tcp_remove_reno_sacks(struct sock
*sk
, struct tcp_sock
*tp
, int acked
)
1481 /* One ACK acked hole. The rest eat duplicate ACKs. */
1482 if (acked
-1 >= tp
->sacked_out
)
1485 tp
->sacked_out
-= acked
-1;
1487 tcp_check_reno_reordering(sk
, acked
);
1488 tcp_sync_left_out(tp
);
1491 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1494 tp
->left_out
= tp
->lost_out
;
1497 /* Mark head of queue up as lost. */
1498 static void tcp_mark_head_lost(struct sock
*sk
, struct tcp_sock
*tp
,
1499 int packets
, u32 high_seq
)
1501 struct sk_buff
*skb
;
1504 BUG_TRAP(cnt
<= tp
->packets_out
);
1506 sk_stream_for_retrans_queue(skb
, sk
) {
1507 cnt
-= tcp_skb_pcount(skb
);
1508 if (cnt
< 0 || after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1510 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1511 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1512 tp
->lost_out
+= tcp_skb_pcount(skb
);
1515 tcp_sync_left_out(tp
);
1518 /* Account newly detected lost packet(s) */
1520 static void tcp_update_scoreboard(struct sock
*sk
, struct tcp_sock
*tp
)
1523 int lost
= tp
->fackets_out
- tp
->reordering
;
1526 tcp_mark_head_lost(sk
, tp
, lost
, tp
->high_seq
);
1528 tcp_mark_head_lost(sk
, tp
, 1, tp
->high_seq
);
1531 /* New heuristics: it is possible only after we switched
1532 * to restart timer each time when something is ACKed.
1533 * Hence, we can detect timed out packets during fast
1534 * retransmit without falling to slow start.
1536 if (tcp_head_timedout(sk
, tp
)) {
1537 struct sk_buff
*skb
;
1539 sk_stream_for_retrans_queue(skb
, sk
) {
1540 if (tcp_skb_timedout(sk
, skb
) &&
1541 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1542 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1543 tp
->lost_out
+= tcp_skb_pcount(skb
);
1546 tcp_sync_left_out(tp
);
1550 /* CWND moderation, preventing bursts due to too big ACKs
1551 * in dubious situations.
1553 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1555 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1556 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1557 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1560 /* Decrease cwnd each second ack. */
1561 static void tcp_cwnd_down(struct sock
*sk
)
1563 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1564 struct tcp_sock
*tp
= tcp_sk(sk
);
1565 int decr
= tp
->snd_cwnd_cnt
+ 1;
1567 tp
->snd_cwnd_cnt
= decr
&1;
1570 if (decr
&& tp
->snd_cwnd
> icsk
->icsk_ca_ops
->min_cwnd(sk
))
1571 tp
->snd_cwnd
-= decr
;
1573 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1574 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1577 /* Nothing was retransmitted or returned timestamp is less
1578 * than timestamp of the first retransmission.
1580 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1582 return !tp
->retrans_stamp
||
1583 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1584 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1587 /* Undo procedures. */
1589 #if FASTRETRANS_DEBUG > 1
1590 static void DBGUNDO(struct sock
*sk
, struct tcp_sock
*tp
, const char *msg
)
1592 struct inet_sock
*inet
= inet_sk(sk
);
1593 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1595 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1596 tp
->snd_cwnd
, tp
->left_out
,
1597 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1601 #define DBGUNDO(x...) do { } while (0)
1604 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1606 struct tcp_sock
*tp
= tcp_sk(sk
);
1608 if (tp
->prior_ssthresh
) {
1609 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1611 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1612 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1614 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1616 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1617 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1618 TCP_ECN_withdraw_cwr(tp
);
1621 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1623 tcp_moderate_cwnd(tp
);
1624 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1627 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1629 return tp
->undo_marker
&&
1630 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1633 /* People celebrate: "We love our President!" */
1634 static int tcp_try_undo_recovery(struct sock
*sk
, struct tcp_sock
*tp
)
1636 if (tcp_may_undo(tp
)) {
1637 /* Happy end! We did not retransmit anything
1638 * or our original transmission succeeded.
1640 DBGUNDO(sk
, tp
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1641 tcp_undo_cwr(sk
, 1);
1642 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1643 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1645 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1646 tp
->undo_marker
= 0;
1648 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1649 /* Hold old state until something *above* high_seq
1650 * is ACKed. For Reno it is MUST to prevent false
1651 * fast retransmits (RFC2582). SACK TCP is safe. */
1652 tcp_moderate_cwnd(tp
);
1655 tcp_set_ca_state(sk
, TCP_CA_Open
);
1659 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1660 static void tcp_try_undo_dsack(struct sock
*sk
, struct tcp_sock
*tp
)
1662 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1663 DBGUNDO(sk
, tp
, "D-SACK");
1664 tcp_undo_cwr(sk
, 1);
1665 tp
->undo_marker
= 0;
1666 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1670 /* Undo during fast recovery after partial ACK. */
1672 static int tcp_try_undo_partial(struct sock
*sk
, struct tcp_sock
*tp
,
1675 /* Partial ACK arrived. Force Hoe's retransmit. */
1676 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1678 if (tcp_may_undo(tp
)) {
1679 /* Plain luck! Hole if filled with delayed
1680 * packet, rather than with a retransmit.
1682 if (tp
->retrans_out
== 0)
1683 tp
->retrans_stamp
= 0;
1685 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
1687 DBGUNDO(sk
, tp
, "Hoe");
1688 tcp_undo_cwr(sk
, 0);
1689 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1691 /* So... Do not make Hoe's retransmit yet.
1692 * If the first packet was delayed, the rest
1693 * ones are most probably delayed as well.
1700 /* Undo during loss recovery after partial ACK. */
1701 static int tcp_try_undo_loss(struct sock
*sk
, struct tcp_sock
*tp
)
1703 if (tcp_may_undo(tp
)) {
1704 struct sk_buff
*skb
;
1705 sk_stream_for_retrans_queue(skb
, sk
) {
1706 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1708 DBGUNDO(sk
, tp
, "partial loss");
1710 tp
->left_out
= tp
->sacked_out
;
1711 tcp_undo_cwr(sk
, 1);
1712 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1713 inet_csk(sk
)->icsk_retransmits
= 0;
1714 tp
->undo_marker
= 0;
1716 tcp_set_ca_state(sk
, TCP_CA_Open
);
1722 static inline void tcp_complete_cwr(struct sock
*sk
)
1724 struct tcp_sock
*tp
= tcp_sk(sk
);
1725 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1726 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1727 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
1730 static void tcp_try_to_open(struct sock
*sk
, struct tcp_sock
*tp
, int flag
)
1732 tp
->left_out
= tp
->sacked_out
;
1734 if (tp
->retrans_out
== 0)
1735 tp
->retrans_stamp
= 0;
1740 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
1741 int state
= TCP_CA_Open
;
1743 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
1744 state
= TCP_CA_Disorder
;
1746 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
1747 tcp_set_ca_state(sk
, state
);
1748 tp
->high_seq
= tp
->snd_nxt
;
1750 tcp_moderate_cwnd(tp
);
1756 /* Process an event, which can update packets-in-flight not trivially.
1757 * Main goal of this function is to calculate new estimate for left_out,
1758 * taking into account both packets sitting in receiver's buffer and
1759 * packets lost by network.
1761 * Besides that it does CWND reduction, when packet loss is detected
1762 * and changes state of machine.
1764 * It does _not_ decide what to send, it is made in function
1765 * tcp_xmit_retransmit_queue().
1768 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
1769 int prior_packets
, int flag
)
1771 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1772 struct tcp_sock
*tp
= tcp_sk(sk
);
1773 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
1775 /* Some technical things:
1776 * 1. Reno does not count dupacks (sacked_out) automatically. */
1777 if (!tp
->packets_out
)
1779 /* 2. SACK counts snd_fack in packets inaccurately. */
1780 if (tp
->sacked_out
== 0)
1781 tp
->fackets_out
= 0;
1783 /* Now state machine starts.
1784 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1786 tp
->prior_ssthresh
= 0;
1788 /* B. In all the states check for reneging SACKs. */
1789 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
1792 /* C. Process data loss notification, provided it is valid. */
1793 if ((flag
&FLAG_DATA_LOST
) &&
1794 before(tp
->snd_una
, tp
->high_seq
) &&
1795 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
1796 tp
->fackets_out
> tp
->reordering
) {
1797 tcp_mark_head_lost(sk
, tp
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
1798 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
1801 /* D. Synchronize left_out to current state. */
1802 tcp_sync_left_out(tp
);
1804 /* E. Check state exit conditions. State can be terminated
1805 * when high_seq is ACKed. */
1806 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
1807 if (!sysctl_tcp_frto
)
1808 BUG_TRAP(tp
->retrans_out
== 0);
1809 tp
->retrans_stamp
= 0;
1810 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
1811 switch (icsk
->icsk_ca_state
) {
1813 icsk
->icsk_retransmits
= 0;
1814 if (tcp_try_undo_recovery(sk
, tp
))
1819 /* CWR is to be held something *above* high_seq
1820 * is ACKed for CWR bit to reach receiver. */
1821 if (tp
->snd_una
!= tp
->high_seq
) {
1822 tcp_complete_cwr(sk
);
1823 tcp_set_ca_state(sk
, TCP_CA_Open
);
1827 case TCP_CA_Disorder
:
1828 tcp_try_undo_dsack(sk
, tp
);
1829 if (!tp
->undo_marker
||
1830 /* For SACK case do not Open to allow to undo
1831 * catching for all duplicate ACKs. */
1832 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
1833 tp
->undo_marker
= 0;
1834 tcp_set_ca_state(sk
, TCP_CA_Open
);
1838 case TCP_CA_Recovery
:
1840 tcp_reset_reno_sack(tp
);
1841 if (tcp_try_undo_recovery(sk
, tp
))
1843 tcp_complete_cwr(sk
);
1848 /* F. Process state. */
1849 switch (icsk
->icsk_ca_state
) {
1850 case TCP_CA_Recovery
:
1851 if (prior_snd_una
== tp
->snd_una
) {
1852 if (IsReno(tp
) && is_dupack
)
1853 tcp_add_reno_sack(sk
);
1855 int acked
= prior_packets
- tp
->packets_out
;
1857 tcp_remove_reno_sacks(sk
, tp
, acked
);
1858 is_dupack
= tcp_try_undo_partial(sk
, tp
, acked
);
1862 if (flag
&FLAG_DATA_ACKED
)
1863 icsk
->icsk_retransmits
= 0;
1864 if (!tcp_try_undo_loss(sk
, tp
)) {
1865 tcp_moderate_cwnd(tp
);
1866 tcp_xmit_retransmit_queue(sk
);
1869 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1871 /* Loss is undone; fall through to processing in Open state. */
1874 if (tp
->snd_una
!= prior_snd_una
)
1875 tcp_reset_reno_sack(tp
);
1877 tcp_add_reno_sack(sk
);
1880 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
1881 tcp_try_undo_dsack(sk
, tp
);
1883 if (!tcp_time_to_recover(sk
, tp
)) {
1884 tcp_try_to_open(sk
, tp
, flag
);
1888 /* Otherwise enter Recovery state */
1891 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
1893 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
1895 tp
->high_seq
= tp
->snd_nxt
;
1896 tp
->prior_ssthresh
= 0;
1897 tp
->undo_marker
= tp
->snd_una
;
1898 tp
->undo_retrans
= tp
->retrans_out
;
1900 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
1901 if (!(flag
&FLAG_ECE
))
1902 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1903 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1904 TCP_ECN_queue_cwr(tp
);
1907 tp
->snd_cwnd_cnt
= 0;
1908 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
1911 if (is_dupack
|| tcp_head_timedout(sk
, tp
))
1912 tcp_update_scoreboard(sk
, tp
);
1914 tcp_xmit_retransmit_queue(sk
);
1917 /* Read draft-ietf-tcplw-high-performance before mucking
1918 * with this code. (Superceeds RFC1323)
1920 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
1922 /* RTTM Rule: A TSecr value received in a segment is used to
1923 * update the averaged RTT measurement only if the segment
1924 * acknowledges some new data, i.e., only if it advances the
1925 * left edge of the send window.
1927 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1928 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1930 * Changed: reset backoff as soon as we see the first valid sample.
1931 * If we do not, we get strongly overstimated rto. With timestamps
1932 * samples are accepted even from very old segments: f.e., when rtt=1
1933 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1934 * answer arrives rto becomes 120 seconds! If at least one of segments
1935 * in window is lost... Voila. --ANK (010210)
1937 struct tcp_sock
*tp
= tcp_sk(sk
);
1938 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
1939 tcp_rtt_estimator(sk
, seq_rtt
);
1941 inet_csk(sk
)->icsk_backoff
= 0;
1945 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
1947 /* We don't have a timestamp. Can only use
1948 * packets that are not retransmitted to determine
1949 * rtt estimates. Also, we must not reset the
1950 * backoff for rto until we get a non-retransmitted
1951 * packet. This allows us to deal with a situation
1952 * where the network delay has increased suddenly.
1953 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1956 if (flag
& FLAG_RETRANS_DATA_ACKED
)
1959 tcp_rtt_estimator(sk
, seq_rtt
);
1961 inet_csk(sk
)->icsk_backoff
= 0;
1965 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
1968 const struct tcp_sock
*tp
= tcp_sk(sk
);
1969 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1970 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
1971 tcp_ack_saw_tstamp(sk
, flag
);
1972 else if (seq_rtt
>= 0)
1973 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
1976 static inline void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
1977 u32 in_flight
, int good
)
1979 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1980 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
1981 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
1984 /* Restart timer after forward progress on connection.
1985 * RFC2988 recommends to restart timer to now+rto.
1988 static inline void tcp_ack_packets_out(struct sock
*sk
, struct tcp_sock
*tp
)
1990 if (!tp
->packets_out
) {
1991 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
1993 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
1997 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
1998 __u32 now
, __s32
*seq_rtt
)
2000 struct tcp_sock
*tp
= tcp_sk(sk
);
2001 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2002 __u32 seq
= tp
->snd_una
;
2003 __u32 packets_acked
;
2006 /* If we get here, the whole TSO packet has not been
2009 BUG_ON(!after(scb
->end_seq
, seq
));
2011 packets_acked
= tcp_skb_pcount(skb
);
2012 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2014 packets_acked
-= tcp_skb_pcount(skb
);
2016 if (packets_acked
) {
2017 __u8 sacked
= scb
->sacked
;
2019 acked
|= FLAG_DATA_ACKED
;
2021 if (sacked
& TCPCB_RETRANS
) {
2022 if (sacked
& TCPCB_SACKED_RETRANS
)
2023 tp
->retrans_out
-= packets_acked
;
2024 acked
|= FLAG_RETRANS_DATA_ACKED
;
2026 } else if (*seq_rtt
< 0)
2027 *seq_rtt
= now
- scb
->when
;
2028 if (sacked
& TCPCB_SACKED_ACKED
)
2029 tp
->sacked_out
-= packets_acked
;
2030 if (sacked
& TCPCB_LOST
)
2031 tp
->lost_out
-= packets_acked
;
2032 if (sacked
& TCPCB_URG
) {
2034 !before(seq
, tp
->snd_up
))
2037 } else if (*seq_rtt
< 0)
2038 *seq_rtt
= now
- scb
->when
;
2040 if (tp
->fackets_out
) {
2041 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2042 tp
->fackets_out
-= dval
;
2044 tp
->packets_out
-= packets_acked
;
2046 BUG_ON(tcp_skb_pcount(skb
) == 0);
2047 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2053 static inline u32
tcp_usrtt(const struct sk_buff
*skb
)
2055 struct timeval tv
, now
;
2057 do_gettimeofday(&now
);
2058 skb_get_timestamp(skb
, &tv
);
2059 return (now
.tv_sec
- tv
.tv_sec
) * 1000000 + (now
.tv_usec
- tv
.tv_usec
);
2062 /* Remove acknowledged frames from the retransmission queue. */
2063 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2065 struct tcp_sock
*tp
= tcp_sk(sk
);
2066 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2067 struct sk_buff
*skb
;
2068 __u32 now
= tcp_time_stamp
;
2072 void (*rtt_sample
)(struct sock
*sk
, u32 usrtt
)
2073 = icsk
->icsk_ca_ops
->rtt_sample
;
2075 while ((skb
= skb_peek(&sk
->sk_write_queue
)) &&
2076 skb
!= sk
->sk_send_head
) {
2077 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2078 __u8 sacked
= scb
->sacked
;
2080 /* If our packet is before the ack sequence we can
2081 * discard it as it's confirmed to have arrived at
2084 if (after(scb
->end_seq
, tp
->snd_una
)) {
2085 if (tcp_skb_pcount(skb
) > 1 &&
2086 after(tp
->snd_una
, scb
->seq
))
2087 acked
|= tcp_tso_acked(sk
, skb
,
2092 /* Initial outgoing SYN's get put onto the write_queue
2093 * just like anything else we transmit. It is not
2094 * true data, and if we misinform our callers that
2095 * this ACK acks real data, we will erroneously exit
2096 * connection startup slow start one packet too
2097 * quickly. This is severely frowned upon behavior.
2099 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2100 acked
|= FLAG_DATA_ACKED
;
2103 acked
|= FLAG_SYN_ACKED
;
2104 tp
->retrans_stamp
= 0;
2108 if (sacked
& TCPCB_RETRANS
) {
2109 if(sacked
& TCPCB_SACKED_RETRANS
)
2110 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2111 acked
|= FLAG_RETRANS_DATA_ACKED
;
2113 } else if (seq_rtt
< 0) {
2114 seq_rtt
= now
- scb
->when
;
2116 (*rtt_sample
)(sk
, tcp_usrtt(skb
));
2118 if (sacked
& TCPCB_SACKED_ACKED
)
2119 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2120 if (sacked
& TCPCB_LOST
)
2121 tp
->lost_out
-= tcp_skb_pcount(skb
);
2122 if (sacked
& TCPCB_URG
) {
2124 !before(scb
->end_seq
, tp
->snd_up
))
2127 } else if (seq_rtt
< 0) {
2128 seq_rtt
= now
- scb
->when
;
2130 (*rtt_sample
)(sk
, tcp_usrtt(skb
));
2132 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2133 tcp_packets_out_dec(tp
, skb
);
2134 __skb_unlink(skb
, &sk
->sk_write_queue
);
2135 sk_stream_free_skb(sk
, skb
);
2138 if (acked
&FLAG_ACKED
) {
2139 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2140 tcp_ack_packets_out(sk
, tp
);
2142 if (icsk
->icsk_ca_ops
->pkts_acked
)
2143 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
);
2146 #if FASTRETRANS_DEBUG > 0
2147 BUG_TRAP((int)tp
->sacked_out
>= 0);
2148 BUG_TRAP((int)tp
->lost_out
>= 0);
2149 BUG_TRAP((int)tp
->retrans_out
>= 0);
2150 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2151 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2153 printk(KERN_DEBUG
"Leak l=%u %d\n",
2154 tp
->lost_out
, icsk
->icsk_ca_state
);
2157 if (tp
->sacked_out
) {
2158 printk(KERN_DEBUG
"Leak s=%u %d\n",
2159 tp
->sacked_out
, icsk
->icsk_ca_state
);
2162 if (tp
->retrans_out
) {
2163 printk(KERN_DEBUG
"Leak r=%u %d\n",
2164 tp
->retrans_out
, icsk
->icsk_ca_state
);
2165 tp
->retrans_out
= 0;
2169 *seq_rtt_p
= seq_rtt
;
2173 static void tcp_ack_probe(struct sock
*sk
)
2175 const struct tcp_sock
*tp
= tcp_sk(sk
);
2176 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2178 /* Was it a usable window open? */
2180 if (!after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
2181 tp
->snd_una
+ tp
->snd_wnd
)) {
2182 icsk
->icsk_backoff
= 0;
2183 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2184 /* Socket must be waked up by subsequent tcp_data_snd_check().
2185 * This function is not for random using!
2188 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2189 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2194 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2196 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2197 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2200 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2202 const struct tcp_sock
*tp
= tcp_sk(sk
);
2203 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2204 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2207 /* Check that window update is acceptable.
2208 * The function assumes that snd_una<=ack<=snd_next.
2210 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2211 const u32 ack_seq
, const u32 nwin
)
2213 return (after(ack
, tp
->snd_una
) ||
2214 after(ack_seq
, tp
->snd_wl1
) ||
2215 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2218 /* Update our send window.
2220 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2221 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2223 static int tcp_ack_update_window(struct sock
*sk
, struct tcp_sock
*tp
,
2224 struct sk_buff
*skb
, u32 ack
, u32 ack_seq
)
2227 u32 nwin
= ntohs(skb
->h
.th
->window
);
2229 if (likely(!skb
->h
.th
->syn
))
2230 nwin
<<= tp
->rx_opt
.snd_wscale
;
2232 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2233 flag
|= FLAG_WIN_UPDATE
;
2234 tcp_update_wl(tp
, ack
, ack_seq
);
2236 if (tp
->snd_wnd
!= nwin
) {
2239 /* Note, it is the only place, where
2240 * fast path is recovered for sending TCP.
2243 tcp_fast_path_check(sk
, tp
);
2245 if (nwin
> tp
->max_window
) {
2246 tp
->max_window
= nwin
;
2247 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
2257 static void tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
)
2259 struct tcp_sock
*tp
= tcp_sk(sk
);
2261 tcp_sync_left_out(tp
);
2263 if (tp
->snd_una
== prior_snd_una
||
2264 !before(tp
->snd_una
, tp
->frto_highmark
)) {
2265 /* RTO was caused by loss, start retransmitting in
2266 * go-back-N slow start
2268 tcp_enter_frto_loss(sk
);
2272 if (tp
->frto_counter
== 1) {
2273 /* First ACK after RTO advances the window: allow two new
2276 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2278 /* Also the second ACK after RTO advances the window.
2279 * The RTO was likely spurious. Reduce cwnd and continue
2280 * in congestion avoidance
2282 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2283 tcp_moderate_cwnd(tp
);
2286 /* F-RTO affects on two new ACKs following RTO.
2287 * At latest on third ACK the TCP behavor is back to normal.
2289 tp
->frto_counter
= (tp
->frto_counter
+ 1) % 3;
2292 /* This routine deals with incoming acks, but not outgoing ones. */
2293 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2295 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2296 struct tcp_sock
*tp
= tcp_sk(sk
);
2297 u32 prior_snd_una
= tp
->snd_una
;
2298 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2299 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2300 u32 prior_in_flight
;
2304 /* If the ack is newer than sent or older than previous acks
2305 * then we can probably ignore it.
2307 if (after(ack
, tp
->snd_nxt
))
2308 goto uninteresting_ack
;
2310 if (before(ack
, prior_snd_una
))
2313 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2314 /* Window is constant, pure forward advance.
2315 * No more checks are required.
2316 * Note, we use the fact that SND.UNA>=SND.WL2.
2318 tcp_update_wl(tp
, ack
, ack_seq
);
2320 flag
|= FLAG_WIN_UPDATE
;
2322 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2324 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2326 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2329 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2331 flag
|= tcp_ack_update_window(sk
, tp
, skb
, ack
, ack_seq
);
2333 if (TCP_SKB_CB(skb
)->sacked
)
2334 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2336 if (TCP_ECN_rcv_ecn_echo(tp
, skb
->h
.th
))
2339 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2342 /* We passed data and got it acked, remove any soft error
2343 * log. Something worked...
2345 sk
->sk_err_soft
= 0;
2346 tp
->rcv_tstamp
= tcp_time_stamp
;
2347 prior_packets
= tp
->packets_out
;
2351 prior_in_flight
= tcp_packets_in_flight(tp
);
2353 /* See if we can take anything off of the retransmit queue. */
2354 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2356 if (tp
->frto_counter
)
2357 tcp_process_frto(sk
, prior_snd_una
);
2359 if (tcp_ack_is_dubious(sk
, flag
)) {
2360 /* Advanve CWND, if state allows this. */
2361 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
2362 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2363 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2365 if ((flag
& FLAG_DATA_ACKED
))
2366 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2369 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2370 dst_confirm(sk
->sk_dst_cache
);
2375 icsk
->icsk_probes_out
= 0;
2377 /* If this ack opens up a zero window, clear backoff. It was
2378 * being used to time the probes, and is probably far higher than
2379 * it needs to be for normal retransmission.
2381 if (sk
->sk_send_head
)
2386 if (TCP_SKB_CB(skb
)->sacked
)
2387 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2390 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2395 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2396 * But, this can also be called on packets in the established flow when
2397 * the fast version below fails.
2399 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2402 struct tcphdr
*th
= skb
->h
.th
;
2403 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2405 ptr
= (unsigned char *)(th
+ 1);
2406 opt_rx
->saw_tstamp
= 0;
2415 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2420 if (opsize
< 2) /* "silly options" */
2422 if (opsize
> length
)
2423 return; /* don't parse partial options */
2426 if(opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2427 u16 in_mss
= ntohs(get_unaligned((__u16
*)ptr
));
2429 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2430 in_mss
= opt_rx
->user_mss
;
2431 opt_rx
->mss_clamp
= in_mss
;
2436 if(opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2437 if (sysctl_tcp_window_scaling
) {
2438 __u8 snd_wscale
= *(__u8
*) ptr
;
2439 opt_rx
->wscale_ok
= 1;
2440 if (snd_wscale
> 14) {
2442 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2443 "scaling value %d >14 received.\n",
2447 opt_rx
->snd_wscale
= snd_wscale
;
2450 case TCPOPT_TIMESTAMP
:
2451 if(opsize
==TCPOLEN_TIMESTAMP
) {
2452 if ((estab
&& opt_rx
->tstamp_ok
) ||
2453 (!estab
&& sysctl_tcp_timestamps
)) {
2454 opt_rx
->saw_tstamp
= 1;
2455 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__u32
*)ptr
));
2456 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__u32
*)(ptr
+4)));
2460 case TCPOPT_SACK_PERM
:
2461 if(opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2462 if (sysctl_tcp_sack
) {
2463 opt_rx
->sack_ok
= 1;
2464 tcp_sack_reset(opt_rx
);
2470 if((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2471 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2473 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2482 /* Fast parse options. This hopes to only see timestamps.
2483 * If it is wrong it falls back on tcp_parse_options().
2485 static inline int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2486 struct tcp_sock
*tp
)
2488 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2489 tp
->rx_opt
.saw_tstamp
= 0;
2491 } else if (tp
->rx_opt
.tstamp_ok
&&
2492 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2493 __u32
*ptr
= (__u32
*)(th
+ 1);
2494 if (*ptr
== ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2495 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2496 tp
->rx_opt
.saw_tstamp
= 1;
2498 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2500 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2504 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2508 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2510 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2511 tp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
2514 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2516 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2517 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2518 * extra check below makes sure this can only happen
2519 * for pure ACK frames. -DaveM
2521 * Not only, also it occurs for expired timestamps.
2524 if((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2525 xtime
.tv_sec
>= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2526 tcp_store_ts_recent(tp
);
2530 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2532 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2533 * it can pass through stack. So, the following predicate verifies that
2534 * this segment is not used for anything but congestion avoidance or
2535 * fast retransmit. Moreover, we even are able to eliminate most of such
2536 * second order effects, if we apply some small "replay" window (~RTO)
2537 * to timestamp space.
2539 * All these measures still do not guarantee that we reject wrapped ACKs
2540 * on networks with high bandwidth, when sequence space is recycled fastly,
2541 * but it guarantees that such events will be very rare and do not affect
2542 * connection seriously. This doesn't look nice, but alas, PAWS is really
2545 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2546 * states that events when retransmit arrives after original data are rare.
2547 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2548 * the biggest problem on large power networks even with minor reordering.
2549 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2550 * up to bandwidth of 18Gigabit/sec. 8) ]
2553 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
2555 struct tcp_sock
*tp
= tcp_sk(sk
);
2556 struct tcphdr
*th
= skb
->h
.th
;
2557 u32 seq
= TCP_SKB_CB(skb
)->seq
;
2558 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2560 return (/* 1. Pure ACK with correct sequence number. */
2561 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
2563 /* 2. ... and duplicate ACK. */
2564 ack
== tp
->snd_una
&&
2566 /* 3. ... and does not update window. */
2567 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
2569 /* 4. ... and sits in replay window. */
2570 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
2573 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
2575 const struct tcp_sock
*tp
= tcp_sk(sk
);
2576 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
2577 xtime
.tv_sec
< tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
2578 !tcp_disordered_ack(sk
, skb
));
2581 /* Check segment sequence number for validity.
2583 * Segment controls are considered valid, if the segment
2584 * fits to the window after truncation to the window. Acceptability
2585 * of data (and SYN, FIN, of course) is checked separately.
2586 * See tcp_data_queue(), for example.
2588 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2589 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2590 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2591 * (borrowed from freebsd)
2594 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2596 return !before(end_seq
, tp
->rcv_wup
) &&
2597 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
2600 /* When we get a reset we do this. */
2601 static void tcp_reset(struct sock
*sk
)
2603 /* We want the right error as BSD sees it (and indeed as we do). */
2604 switch (sk
->sk_state
) {
2606 sk
->sk_err
= ECONNREFUSED
;
2608 case TCP_CLOSE_WAIT
:
2614 sk
->sk_err
= ECONNRESET
;
2617 if (!sock_flag(sk
, SOCK_DEAD
))
2618 sk
->sk_error_report(sk
);
2624 * Process the FIN bit. This now behaves as it is supposed to work
2625 * and the FIN takes effect when it is validly part of sequence
2626 * space. Not before when we get holes.
2628 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2629 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2632 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2633 * close and we go into CLOSING (and later onto TIME-WAIT)
2635 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2637 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
2639 struct tcp_sock
*tp
= tcp_sk(sk
);
2641 inet_csk_schedule_ack(sk
);
2643 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
2644 sock_set_flag(sk
, SOCK_DONE
);
2646 switch (sk
->sk_state
) {
2648 case TCP_ESTABLISHED
:
2649 /* Move to CLOSE_WAIT */
2650 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
2651 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
2654 case TCP_CLOSE_WAIT
:
2656 /* Received a retransmission of the FIN, do
2661 /* RFC793: Remain in the LAST-ACK state. */
2665 /* This case occurs when a simultaneous close
2666 * happens, we must ack the received FIN and
2667 * enter the CLOSING state.
2670 tcp_set_state(sk
, TCP_CLOSING
);
2673 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2675 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
2678 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2679 * cases we should never reach this piece of code.
2681 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
2682 __FUNCTION__
, sk
->sk_state
);
2686 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2687 * Probably, we should reset in this case. For now drop them.
2689 __skb_queue_purge(&tp
->out_of_order_queue
);
2690 if (tp
->rx_opt
.sack_ok
)
2691 tcp_sack_reset(&tp
->rx_opt
);
2692 sk_stream_mem_reclaim(sk
);
2694 if (!sock_flag(sk
, SOCK_DEAD
)) {
2695 sk
->sk_state_change(sk
);
2697 /* Do not send POLL_HUP for half duplex close. */
2698 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
2699 sk
->sk_state
== TCP_CLOSE
)
2700 sk_wake_async(sk
, 1, POLL_HUP
);
2702 sk_wake_async(sk
, 1, POLL_IN
);
2706 static __inline__
int
2707 tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
2709 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
2710 if (before(seq
, sp
->start_seq
))
2711 sp
->start_seq
= seq
;
2712 if (after(end_seq
, sp
->end_seq
))
2713 sp
->end_seq
= end_seq
;
2719 static inline void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2721 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2722 if (before(seq
, tp
->rcv_nxt
))
2723 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
2725 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
2727 tp
->rx_opt
.dsack
= 1;
2728 tp
->duplicate_sack
[0].start_seq
= seq
;
2729 tp
->duplicate_sack
[0].end_seq
= end_seq
;
2730 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
2734 static inline void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2736 if (!tp
->rx_opt
.dsack
)
2737 tcp_dsack_set(tp
, seq
, end_seq
);
2739 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
2742 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
2744 struct tcp_sock
*tp
= tcp_sk(sk
);
2746 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
2747 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
2748 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
2749 tcp_enter_quickack_mode(sk
);
2751 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2752 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2754 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
2755 end_seq
= tp
->rcv_nxt
;
2756 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
2763 /* These routines update the SACK block as out-of-order packets arrive or
2764 * in-order packets close up the sequence space.
2766 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
2769 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2770 struct tcp_sack_block
*swalk
= sp
+1;
2772 /* See if the recent change to the first SACK eats into
2773 * or hits the sequence space of other SACK blocks, if so coalesce.
2775 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
2776 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
2779 /* Zap SWALK, by moving every further SACK up by one slot.
2780 * Decrease num_sacks.
2782 tp
->rx_opt
.num_sacks
--;
2783 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2784 for(i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
2788 this_sack
++, swalk
++;
2792 static __inline__
void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
2796 tmp
= sack1
->start_seq
;
2797 sack1
->start_seq
= sack2
->start_seq
;
2798 sack2
->start_seq
= tmp
;
2800 tmp
= sack1
->end_seq
;
2801 sack1
->end_seq
= sack2
->end_seq
;
2802 sack2
->end_seq
= tmp
;
2805 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
2807 struct tcp_sock
*tp
= tcp_sk(sk
);
2808 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2809 int cur_sacks
= tp
->rx_opt
.num_sacks
;
2815 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
2816 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
2817 /* Rotate this_sack to the first one. */
2818 for (; this_sack
>0; this_sack
--, sp
--)
2819 tcp_sack_swap(sp
, sp
-1);
2821 tcp_sack_maybe_coalesce(tp
);
2826 /* Could not find an adjacent existing SACK, build a new one,
2827 * put it at the front, and shift everyone else down. We
2828 * always know there is at least one SACK present already here.
2830 * If the sack array is full, forget about the last one.
2832 if (this_sack
>= 4) {
2834 tp
->rx_opt
.num_sacks
--;
2837 for(; this_sack
> 0; this_sack
--, sp
--)
2841 /* Build the new head SACK, and we're done. */
2842 sp
->start_seq
= seq
;
2843 sp
->end_seq
= end_seq
;
2844 tp
->rx_opt
.num_sacks
++;
2845 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2848 /* RCV.NXT advances, some SACKs should be eaten. */
2850 static void tcp_sack_remove(struct tcp_sock
*tp
)
2852 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2853 int num_sacks
= tp
->rx_opt
.num_sacks
;
2856 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2857 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
2858 tp
->rx_opt
.num_sacks
= 0;
2859 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
2863 for(this_sack
= 0; this_sack
< num_sacks
; ) {
2864 /* Check if the start of the sack is covered by RCV.NXT. */
2865 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
2868 /* RCV.NXT must cover all the block! */
2869 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
2871 /* Zap this SACK, by moving forward any other SACKS. */
2872 for (i
=this_sack
+1; i
< num_sacks
; i
++)
2873 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
2880 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
2881 tp
->rx_opt
.num_sacks
= num_sacks
;
2882 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2886 /* This one checks to see if we can put data from the
2887 * out_of_order queue into the receive_queue.
2889 static void tcp_ofo_queue(struct sock
*sk
)
2891 struct tcp_sock
*tp
= tcp_sk(sk
);
2892 __u32 dsack_high
= tp
->rcv_nxt
;
2893 struct sk_buff
*skb
;
2895 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
2896 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
2899 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
2900 __u32 dsack
= dsack_high
;
2901 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
2902 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
2903 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
2906 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
2907 SOCK_DEBUG(sk
, "ofo packet was already received \n");
2908 __skb_unlink(skb
, &tp
->out_of_order_queue
);
2912 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
2913 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
2914 TCP_SKB_CB(skb
)->end_seq
);
2916 __skb_unlink(skb
, &tp
->out_of_order_queue
);
2917 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
2918 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
2920 tcp_fin(skb
, sk
, skb
->h
.th
);
2924 static int tcp_prune_queue(struct sock
*sk
);
2926 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
2928 struct tcphdr
*th
= skb
->h
.th
;
2929 struct tcp_sock
*tp
= tcp_sk(sk
);
2932 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
2935 __skb_pull(skb
, th
->doff
*4);
2937 TCP_ECN_accept_cwr(tp
, skb
);
2939 if (tp
->rx_opt
.dsack
) {
2940 tp
->rx_opt
.dsack
= 0;
2941 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
2942 4 - tp
->rx_opt
.tstamp_ok
);
2945 /* Queue data for delivery to the user.
2946 * Packets in sequence go to the receive queue.
2947 * Out of sequence packets to the out_of_order_queue.
2949 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
2950 if (tcp_receive_window(tp
) == 0)
2953 /* Ok. In sequence. In window. */
2954 if (tp
->ucopy
.task
== current
&&
2955 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
2956 sock_owned_by_user(sk
) && !tp
->urg_data
) {
2957 int chunk
= min_t(unsigned int, skb
->len
,
2960 __set_current_state(TASK_RUNNING
);
2963 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
2964 tp
->ucopy
.len
-= chunk
;
2965 tp
->copied_seq
+= chunk
;
2966 eaten
= (chunk
== skb
->len
&& !th
->fin
);
2967 tcp_rcv_space_adjust(sk
);
2975 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
2976 !sk_stream_rmem_schedule(sk
, skb
))) {
2977 if (tcp_prune_queue(sk
) < 0 ||
2978 !sk_stream_rmem_schedule(sk
, skb
))
2981 sk_stream_set_owner_r(skb
, sk
);
2982 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
2984 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
2986 tcp_event_data_recv(sk
, tp
, skb
);
2988 tcp_fin(skb
, sk
, th
);
2990 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
2993 /* RFC2581. 4.2. SHOULD send immediate ACK, when
2994 * gap in queue is filled.
2996 if (skb_queue_empty(&tp
->out_of_order_queue
))
2997 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3000 if (tp
->rx_opt
.num_sacks
)
3001 tcp_sack_remove(tp
);
3003 tcp_fast_path_check(sk
, tp
);
3007 else if (!sock_flag(sk
, SOCK_DEAD
))
3008 sk
->sk_data_ready(sk
, 0);
3012 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3013 /* A retransmit, 2nd most common case. Force an immediate ack. */
3014 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3015 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3018 tcp_enter_quickack_mode(sk
);
3019 inet_csk_schedule_ack(sk
);
3025 /* Out of window. F.e. zero window probe. */
3026 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3029 tcp_enter_quickack_mode(sk
);
3031 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3032 /* Partial packet, seq < rcv_next < end_seq */
3033 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3034 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3035 TCP_SKB_CB(skb
)->end_seq
);
3037 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3039 /* If window is closed, drop tail of packet. But after
3040 * remembering D-SACK for its head made in previous line.
3042 if (!tcp_receive_window(tp
))
3047 TCP_ECN_check_ce(tp
, skb
);
3049 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3050 !sk_stream_rmem_schedule(sk
, skb
)) {
3051 if (tcp_prune_queue(sk
) < 0 ||
3052 !sk_stream_rmem_schedule(sk
, skb
))
3056 /* Disable header prediction. */
3058 inet_csk_schedule_ack(sk
);
3060 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3061 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3063 sk_stream_set_owner_r(skb
, sk
);
3065 if (!skb_peek(&tp
->out_of_order_queue
)) {
3066 /* Initial out of order segment, build 1 SACK. */
3067 if (tp
->rx_opt
.sack_ok
) {
3068 tp
->rx_opt
.num_sacks
= 1;
3069 tp
->rx_opt
.dsack
= 0;
3070 tp
->rx_opt
.eff_sacks
= 1;
3071 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3072 tp
->selective_acks
[0].end_seq
=
3073 TCP_SKB_CB(skb
)->end_seq
;
3075 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3077 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3078 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3079 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3081 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3082 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3084 if (!tp
->rx_opt
.num_sacks
||
3085 tp
->selective_acks
[0].end_seq
!= seq
)
3088 /* Common case: data arrive in order after hole. */
3089 tp
->selective_acks
[0].end_seq
= end_seq
;
3093 /* Find place to insert this segment. */
3095 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3097 } while ((skb1
= skb1
->prev
) !=
3098 (struct sk_buff
*)&tp
->out_of_order_queue
);
3100 /* Do skb overlap to previous one? */
3101 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3102 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3103 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3104 /* All the bits are present. Drop. */
3106 tcp_dsack_set(tp
, seq
, end_seq
);
3109 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3110 /* Partial overlap. */
3111 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3116 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3118 /* And clean segments covered by new one as whole. */
3119 while ((skb1
= skb
->next
) !=
3120 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3121 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3122 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3123 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3126 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3127 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3132 if (tp
->rx_opt
.sack_ok
)
3133 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3137 /* Collapse contiguous sequence of skbs head..tail with
3138 * sequence numbers start..end.
3139 * Segments with FIN/SYN are not collapsed (only because this
3143 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3144 struct sk_buff
*head
, struct sk_buff
*tail
,
3147 struct sk_buff
*skb
;
3149 /* First, check that queue is collapsable and find
3150 * the point where collapsing can be useful. */
3151 for (skb
= head
; skb
!= tail
; ) {
3152 /* No new bits? It is possible on ofo queue. */
3153 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3154 struct sk_buff
*next
= skb
->next
;
3155 __skb_unlink(skb
, list
);
3157 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3162 /* The first skb to collapse is:
3164 * - bloated or contains data before "start" or
3165 * overlaps to the next one.
3167 if (!skb
->h
.th
->syn
&& !skb
->h
.th
->fin
&&
3168 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3169 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3170 (skb
->next
!= tail
&&
3171 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3174 /* Decided to skip this, advance start seq. */
3175 start
= TCP_SKB_CB(skb
)->end_seq
;
3178 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3181 while (before(start
, end
)) {
3182 struct sk_buff
*nskb
;
3183 int header
= skb_headroom(skb
);
3184 int copy
= SKB_MAX_ORDER(header
, 0);
3186 /* Too big header? This can happen with IPv6. */
3189 if (end
-start
< copy
)
3191 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3194 skb_reserve(nskb
, header
);
3195 memcpy(nskb
->head
, skb
->head
, header
);
3196 nskb
->nh
.raw
= nskb
->head
+ (skb
->nh
.raw
-skb
->head
);
3197 nskb
->h
.raw
= nskb
->head
+ (skb
->h
.raw
-skb
->head
);
3198 nskb
->mac
.raw
= nskb
->head
+ (skb
->mac
.raw
-skb
->head
);
3199 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3200 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3201 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3202 sk_stream_set_owner_r(nskb
, sk
);
3204 /* Copy data, releasing collapsed skbs. */
3206 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3207 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3209 if (offset
< 0) BUG();
3211 size
= min(copy
, size
);
3212 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3214 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3218 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3219 struct sk_buff
*next
= skb
->next
;
3220 __skb_unlink(skb
, list
);
3222 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3224 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3231 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3232 * and tcp_collapse() them until all the queue is collapsed.
3234 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3236 struct tcp_sock
*tp
= tcp_sk(sk
);
3237 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3238 struct sk_buff
*head
;
3244 start
= TCP_SKB_CB(skb
)->seq
;
3245 end
= TCP_SKB_CB(skb
)->end_seq
;
3251 /* Segment is terminated when we see gap or when
3252 * we are at the end of all the queue. */
3253 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3254 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3255 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3256 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3257 head
, skb
, start
, end
);
3259 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3261 /* Start new segment */
3262 start
= TCP_SKB_CB(skb
)->seq
;
3263 end
= TCP_SKB_CB(skb
)->end_seq
;
3265 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3266 start
= TCP_SKB_CB(skb
)->seq
;
3267 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3268 end
= TCP_SKB_CB(skb
)->end_seq
;
3273 /* Reduce allocated memory if we can, trying to get
3274 * the socket within its memory limits again.
3276 * Return less than zero if we should start dropping frames
3277 * until the socket owning process reads some of the data
3278 * to stabilize the situation.
3280 static int tcp_prune_queue(struct sock
*sk
)
3282 struct tcp_sock
*tp
= tcp_sk(sk
);
3284 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3286 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3288 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3289 tcp_clamp_window(sk
, tp
);
3290 else if (tcp_memory_pressure
)
3291 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3293 tcp_collapse_ofo_queue(sk
);
3294 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3295 sk
->sk_receive_queue
.next
,
3296 (struct sk_buff
*)&sk
->sk_receive_queue
,
3297 tp
->copied_seq
, tp
->rcv_nxt
);
3298 sk_stream_mem_reclaim(sk
);
3300 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3303 /* Collapsing did not help, destructive actions follow.
3304 * This must not ever occur. */
3306 /* First, purge the out_of_order queue. */
3307 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3308 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3309 __skb_queue_purge(&tp
->out_of_order_queue
);
3311 /* Reset SACK state. A conforming SACK implementation will
3312 * do the same at a timeout based retransmit. When a connection
3313 * is in a sad state like this, we care only about integrity
3314 * of the connection not performance.
3316 if (tp
->rx_opt
.sack_ok
)
3317 tcp_sack_reset(&tp
->rx_opt
);
3318 sk_stream_mem_reclaim(sk
);
3321 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3324 /* If we are really being abused, tell the caller to silently
3325 * drop receive data on the floor. It will get retransmitted
3326 * and hopefully then we'll have sufficient space.
3328 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3330 /* Massive buffer overcommit. */
3336 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3337 * As additional protections, we do not touch cwnd in retransmission phases,
3338 * and if application hit its sndbuf limit recently.
3340 void tcp_cwnd_application_limited(struct sock
*sk
)
3342 struct tcp_sock
*tp
= tcp_sk(sk
);
3344 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3345 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3346 /* Limited by application or receiver window. */
3347 u32 win_used
= max(tp
->snd_cwnd_used
, 2U);
3348 if (win_used
< tp
->snd_cwnd
) {
3349 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3350 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3352 tp
->snd_cwnd_used
= 0;
3354 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3357 static inline int tcp_should_expand_sndbuf(struct sock
*sk
, struct tcp_sock
*tp
)
3359 /* If the user specified a specific send buffer setting, do
3362 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3365 /* If we are under global TCP memory pressure, do not expand. */
3366 if (tcp_memory_pressure
)
3369 /* If we are under soft global TCP memory pressure, do not expand. */
3370 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3373 /* If we filled the congestion window, do not expand. */
3374 if (tp
->packets_out
>= tp
->snd_cwnd
)
3380 /* When incoming ACK allowed to free some skb from write_queue,
3381 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3382 * on the exit from tcp input handler.
3384 * PROBLEM: sndbuf expansion does not work well with largesend.
3386 static void tcp_new_space(struct sock
*sk
)
3388 struct tcp_sock
*tp
= tcp_sk(sk
);
3390 if (tcp_should_expand_sndbuf(sk
, tp
)) {
3391 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3392 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3393 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3394 tp
->reordering
+ 1);
3395 sndmem
*= 2*demanded
;
3396 if (sndmem
> sk
->sk_sndbuf
)
3397 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3398 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3401 sk
->sk_write_space(sk
);
3404 static inline void tcp_check_space(struct sock
*sk
)
3406 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3407 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3408 if (sk
->sk_socket
&&
3409 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3414 static __inline__
void tcp_data_snd_check(struct sock
*sk
, struct tcp_sock
*tp
)
3416 tcp_push_pending_frames(sk
, tp
);
3417 tcp_check_space(sk
);
3421 * Check if sending an ack is needed.
3423 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3425 struct tcp_sock
*tp
= tcp_sk(sk
);
3427 /* More than one full frame received... */
3428 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3429 /* ... and right edge of window advances far enough.
3430 * (tcp_recvmsg() will send ACK otherwise). Or...
3432 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3433 /* We ACK each frame or... */
3434 tcp_in_quickack_mode(sk
) ||
3435 /* We have out of order data. */
3437 skb_peek(&tp
->out_of_order_queue
))) {
3438 /* Then ack it now */
3441 /* Else, send delayed ack. */
3442 tcp_send_delayed_ack(sk
);
3446 static __inline__
void tcp_ack_snd_check(struct sock
*sk
)
3448 if (!inet_csk_ack_scheduled(sk
)) {
3449 /* We sent a data segment already. */
3452 __tcp_ack_snd_check(sk
, 1);
3456 * This routine is only called when we have urgent data
3457 * signalled. Its the 'slow' part of tcp_urg. It could be
3458 * moved inline now as tcp_urg is only called from one
3459 * place. We handle URGent data wrong. We have to - as
3460 * BSD still doesn't use the correction from RFC961.
3461 * For 1003.1g we should support a new option TCP_STDURG to permit
3462 * either form (or just set the sysctl tcp_stdurg).
3465 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3467 struct tcp_sock
*tp
= tcp_sk(sk
);
3468 u32 ptr
= ntohs(th
->urg_ptr
);
3470 if (ptr
&& !sysctl_tcp_stdurg
)
3472 ptr
+= ntohl(th
->seq
);
3474 /* Ignore urgent data that we've already seen and read. */
3475 if (after(tp
->copied_seq
, ptr
))
3478 /* Do not replay urg ptr.
3480 * NOTE: interesting situation not covered by specs.
3481 * Misbehaving sender may send urg ptr, pointing to segment,
3482 * which we already have in ofo queue. We are not able to fetch
3483 * such data and will stay in TCP_URG_NOTYET until will be eaten
3484 * by recvmsg(). Seems, we are not obliged to handle such wicked
3485 * situations. But it is worth to think about possibility of some
3486 * DoSes using some hypothetical application level deadlock.
3488 if (before(ptr
, tp
->rcv_nxt
))
3491 /* Do we already have a newer (or duplicate) urgent pointer? */
3492 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3495 /* Tell the world about our new urgent pointer. */
3498 /* We may be adding urgent data when the last byte read was
3499 * urgent. To do this requires some care. We cannot just ignore
3500 * tp->copied_seq since we would read the last urgent byte again
3501 * as data, nor can we alter copied_seq until this data arrives
3502 * or we break the sematics of SIOCATMARK (and thus sockatmark())
3504 * NOTE. Double Dutch. Rendering to plain English: author of comment
3505 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3506 * and expect that both A and B disappear from stream. This is _wrong_.
3507 * Though this happens in BSD with high probability, this is occasional.
3508 * Any application relying on this is buggy. Note also, that fix "works"
3509 * only in this artificial test. Insert some normal data between A and B and we will
3510 * decline of BSD again. Verdict: it is better to remove to trap
3513 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3514 !sock_flag(sk
, SOCK_URGINLINE
) &&
3515 tp
->copied_seq
!= tp
->rcv_nxt
) {
3516 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3518 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3519 __skb_unlink(skb
, &sk
->sk_receive_queue
);
3524 tp
->urg_data
= TCP_URG_NOTYET
;
3527 /* Disable header prediction. */
3531 /* This is the 'fast' part of urgent handling. */
3532 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
3534 struct tcp_sock
*tp
= tcp_sk(sk
);
3536 /* Check if we get a new urgent pointer - normally not. */
3538 tcp_check_urg(sk
,th
);
3540 /* Do we wait for any urgent data? - normally not... */
3541 if (tp
->urg_data
== TCP_URG_NOTYET
) {
3542 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
3545 /* Is the urgent pointer pointing into this packet? */
3546 if (ptr
< skb
->len
) {
3548 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
3550 tp
->urg_data
= TCP_URG_VALID
| tmp
;
3551 if (!sock_flag(sk
, SOCK_DEAD
))
3552 sk
->sk_data_ready(sk
, 0);
3557 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3559 struct tcp_sock
*tp
= tcp_sk(sk
);
3560 int chunk
= skb
->len
- hlen
;
3564 if (skb
->ip_summed
==CHECKSUM_UNNECESSARY
)
3565 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
3567 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
3571 tp
->ucopy
.len
-= chunk
;
3572 tp
->copied_seq
+= chunk
;
3573 tcp_rcv_space_adjust(sk
);
3580 static int __tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3584 if (sock_owned_by_user(sk
)) {
3586 result
= __tcp_checksum_complete(skb
);
3589 result
= __tcp_checksum_complete(skb
);
3594 static __inline__
int
3595 tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3597 return skb
->ip_summed
!= CHECKSUM_UNNECESSARY
&&
3598 __tcp_checksum_complete_user(sk
, skb
);
3602 * TCP receive function for the ESTABLISHED state.
3604 * It is split into a fast path and a slow path. The fast path is
3606 * - A zero window was announced from us - zero window probing
3607 * is only handled properly in the slow path.
3608 * - Out of order segments arrived.
3609 * - Urgent data is expected.
3610 * - There is no buffer space left
3611 * - Unexpected TCP flags/window values/header lengths are received
3612 * (detected by checking the TCP header against pred_flags)
3613 * - Data is sent in both directions. Fast path only supports pure senders
3614 * or pure receivers (this means either the sequence number or the ack
3615 * value must stay constant)
3616 * - Unexpected TCP option.
3618 * When these conditions are not satisfied it drops into a standard
3619 * receive procedure patterned after RFC793 to handle all cases.
3620 * The first three cases are guaranteed by proper pred_flags setting,
3621 * the rest is checked inline. Fast processing is turned on in
3622 * tcp_data_queue when everything is OK.
3624 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
3625 struct tcphdr
*th
, unsigned len
)
3627 struct tcp_sock
*tp
= tcp_sk(sk
);
3630 * Header prediction.
3631 * The code loosely follows the one in the famous
3632 * "30 instruction TCP receive" Van Jacobson mail.
3634 * Van's trick is to deposit buffers into socket queue
3635 * on a device interrupt, to call tcp_recv function
3636 * on the receive process context and checksum and copy
3637 * the buffer to user space. smart...
3639 * Our current scheme is not silly either but we take the
3640 * extra cost of the net_bh soft interrupt processing...
3641 * We do checksum and copy also but from device to kernel.
3644 tp
->rx_opt
.saw_tstamp
= 0;
3646 /* pred_flags is 0xS?10 << 16 + snd_wnd
3647 * if header_predition is to be made
3648 * 'S' will always be tp->tcp_header_len >> 2
3649 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3650 * turn it off (when there are holes in the receive
3651 * space for instance)
3652 * PSH flag is ignored.
3655 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
3656 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3657 int tcp_header_len
= tp
->tcp_header_len
;
3659 /* Timestamp header prediction: tcp_header_len
3660 * is automatically equal to th->doff*4 due to pred_flags
3664 /* Check timestamp */
3665 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
3666 __u32
*ptr
= (__u32
*)(th
+ 1);
3668 /* No? Slow path! */
3669 if (*ptr
!= ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3670 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
3673 tp
->rx_opt
.saw_tstamp
= 1;
3675 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3677 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3679 /* If PAWS failed, check it more carefully in slow path */
3680 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
3683 /* DO NOT update ts_recent here, if checksum fails
3684 * and timestamp was corrupted part, it will result
3685 * in a hung connection since we will drop all
3686 * future packets due to the PAWS test.
3690 if (len
<= tcp_header_len
) {
3691 /* Bulk data transfer: sender */
3692 if (len
== tcp_header_len
) {
3693 /* Predicted packet is in window by definition.
3694 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3695 * Hence, check seq<=rcv_wup reduces to:
3697 if (tcp_header_len
==
3698 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3699 tp
->rcv_nxt
== tp
->rcv_wup
)
3700 tcp_store_ts_recent(tp
);
3702 tcp_rcv_rtt_measure_ts(sk
, skb
);
3704 /* We know that such packets are checksummed
3707 tcp_ack(sk
, skb
, 0);
3709 tcp_data_snd_check(sk
, tp
);
3711 } else { /* Header too small */
3712 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3718 if (tp
->ucopy
.task
== current
&&
3719 tp
->copied_seq
== tp
->rcv_nxt
&&
3720 len
- tcp_header_len
<= tp
->ucopy
.len
&&
3721 sock_owned_by_user(sk
)) {
3722 __set_current_state(TASK_RUNNING
);
3724 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
3725 /* Predicted packet is in window by definition.
3726 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3727 * Hence, check seq<=rcv_wup reduces to:
3729 if (tcp_header_len
==
3730 (sizeof(struct tcphdr
) +
3731 TCPOLEN_TSTAMP_ALIGNED
) &&
3732 tp
->rcv_nxt
== tp
->rcv_wup
)
3733 tcp_store_ts_recent(tp
);
3735 tcp_rcv_rtt_measure_ts(sk
, skb
);
3737 __skb_pull(skb
, tcp_header_len
);
3738 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3739 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
3744 if (tcp_checksum_complete_user(sk
, skb
))
3747 /* Predicted packet is in window by definition.
3748 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3749 * Hence, check seq<=rcv_wup reduces to:
3751 if (tcp_header_len
==
3752 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3753 tp
->rcv_nxt
== tp
->rcv_wup
)
3754 tcp_store_ts_recent(tp
);
3756 tcp_rcv_rtt_measure_ts(sk
, skb
);
3758 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
3761 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
3763 /* Bulk data transfer: receiver */
3764 __skb_pull(skb
,tcp_header_len
);
3765 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3766 sk_stream_set_owner_r(skb
, sk
);
3767 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3770 tcp_event_data_recv(sk
, tp
, skb
);
3772 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
3773 /* Well, only one small jumplet in fast path... */
3774 tcp_ack(sk
, skb
, FLAG_DATA
);
3775 tcp_data_snd_check(sk
, tp
);
3776 if (!inet_csk_ack_scheduled(sk
))
3780 __tcp_ack_snd_check(sk
, 0);
3785 sk
->sk_data_ready(sk
, 0);
3791 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
3795 * RFC1323: H1. Apply PAWS check first.
3797 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
3798 tcp_paws_discard(sk
, skb
)) {
3800 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
3801 tcp_send_dupack(sk
, skb
);
3804 /* Resets are accepted even if PAWS failed.
3806 ts_recent update must be made after we are sure
3807 that the packet is in window.
3812 * Standard slow path.
3815 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3816 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3817 * (RST) segments are validated by checking their SEQ-fields."
3818 * And page 69: "If an incoming segment is not acceptable,
3819 * an acknowledgment should be sent in reply (unless the RST bit
3820 * is set, if so drop the segment and return)".
3823 tcp_send_dupack(sk
, skb
);
3832 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3834 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3835 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3836 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
3843 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
3845 tcp_rcv_rtt_measure_ts(sk
, skb
);
3847 /* Process urgent data. */
3848 tcp_urg(sk
, skb
, th
);
3850 /* step 7: process the segment text */
3851 tcp_data_queue(sk
, skb
);
3853 tcp_data_snd_check(sk
, tp
);
3854 tcp_ack_snd_check(sk
);
3858 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3865 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
3866 struct tcphdr
*th
, unsigned len
)
3868 struct tcp_sock
*tp
= tcp_sk(sk
);
3869 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
3871 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
3874 struct inet_connection_sock
*icsk
;
3876 * "If the state is SYN-SENT then
3877 * first check the ACK bit
3878 * If the ACK bit is set
3879 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3880 * a reset (unless the RST bit is set, if so drop
3881 * the segment and return)"
3883 * We do not send data with SYN, so that RFC-correct
3886 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
3887 goto reset_and_undo
;
3889 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
3890 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
3892 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
3893 goto reset_and_undo
;
3896 /* Now ACK is acceptable.
3898 * "If the RST bit is set
3899 * If the ACK was acceptable then signal the user "error:
3900 * connection reset", drop the segment, enter CLOSED state,
3901 * delete TCB, and return."
3910 * "fifth, if neither of the SYN or RST bits is set then
3911 * drop the segment and return."
3917 goto discard_and_undo
;
3920 * "If the SYN bit is on ...
3921 * are acceptable then ...
3922 * (our SYN has been ACKed), change the connection
3923 * state to ESTABLISHED..."
3926 TCP_ECN_rcv_synack(tp
, th
);
3927 if (tp
->ecn_flags
&TCP_ECN_OK
)
3928 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
3930 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
3931 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
3933 /* Ok.. it's good. Set up sequence numbers and
3934 * move to established.
3936 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
3937 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
3939 /* RFC1323: The window in SYN & SYN/ACK segments is
3942 tp
->snd_wnd
= ntohs(th
->window
);
3943 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
3945 if (!tp
->rx_opt
.wscale_ok
) {
3946 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
3947 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
3950 if (tp
->rx_opt
.saw_tstamp
) {
3951 tp
->rx_opt
.tstamp_ok
= 1;
3952 tp
->tcp_header_len
=
3953 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
3954 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
3955 tcp_store_ts_recent(tp
);
3957 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3960 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
3961 tp
->rx_opt
.sack_ok
|= 2;
3963 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
3964 tcp_initialize_rcv_mss(sk
);
3966 /* Remember, tcp_poll() does not lock socket!
3967 * Change state from SYN-SENT only after copied_seq
3968 * is initialized. */
3969 tp
->copied_seq
= tp
->rcv_nxt
;
3971 tcp_set_state(sk
, TCP_ESTABLISHED
);
3973 /* Make sure socket is routed, for correct metrics. */
3974 tp
->af_specific
->rebuild_header(sk
);
3976 tcp_init_metrics(sk
);
3978 tcp_init_congestion_control(sk
);
3980 /* Prevent spurious tcp_cwnd_restart() on first data
3983 tp
->lsndtime
= tcp_time_stamp
;
3985 tcp_init_buffer_space(sk
);
3987 if (sock_flag(sk
, SOCK_KEEPOPEN
))
3988 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
3990 if (!tp
->rx_opt
.snd_wscale
)
3991 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
3995 if (!sock_flag(sk
, SOCK_DEAD
)) {
3996 sk
->sk_state_change(sk
);
3997 sk_wake_async(sk
, 0, POLL_OUT
);
4000 icsk
= inet_csk(sk
);
4002 if (sk
->sk_write_pending
||
4003 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4004 icsk
->icsk_ack
.pingpong
) {
4005 /* Save one ACK. Data will be ready after
4006 * several ticks, if write_pending is set.
4008 * It may be deleted, but with this feature tcpdumps
4009 * look so _wonderfully_ clever, that I was not able
4010 * to stand against the temptation 8) --ANK
4012 inet_csk_schedule_ack(sk
);
4013 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4014 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4015 tcp_incr_quickack(sk
);
4016 tcp_enter_quickack_mode(sk
);
4017 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4018 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4029 /* No ACK in the segment */
4033 * "If the RST bit is set
4035 * Otherwise (no ACK) drop the segment and return."
4038 goto discard_and_undo
;
4042 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4043 goto discard_and_undo
;
4046 /* We see SYN without ACK. It is attempt of
4047 * simultaneous connect with crossed SYNs.
4048 * Particularly, it can be connect to self.
4050 tcp_set_state(sk
, TCP_SYN_RECV
);
4052 if (tp
->rx_opt
.saw_tstamp
) {
4053 tp
->rx_opt
.tstamp_ok
= 1;
4054 tcp_store_ts_recent(tp
);
4055 tp
->tcp_header_len
=
4056 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4058 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4061 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4062 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4064 /* RFC1323: The window in SYN & SYN/ACK segments is
4067 tp
->snd_wnd
= ntohs(th
->window
);
4068 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4069 tp
->max_window
= tp
->snd_wnd
;
4071 TCP_ECN_rcv_syn(tp
, th
);
4072 if (tp
->ecn_flags
&TCP_ECN_OK
)
4073 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
4075 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
4076 tcp_initialize_rcv_mss(sk
);
4079 tcp_send_synack(sk
);
4081 /* Note, we could accept data and URG from this segment.
4082 * There are no obstacles to make this.
4084 * However, if we ignore data in ACKless segments sometimes,
4085 * we have no reasons to accept it sometimes.
4086 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4087 * is not flawless. So, discard packet for sanity.
4088 * Uncomment this return to process the data.
4095 /* "fifth, if neither of the SYN or RST bits is set then
4096 * drop the segment and return."
4100 tcp_clear_options(&tp
->rx_opt
);
4101 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4105 tcp_clear_options(&tp
->rx_opt
);
4106 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4112 * This function implements the receiving procedure of RFC 793 for
4113 * all states except ESTABLISHED and TIME_WAIT.
4114 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4115 * address independent.
4118 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4119 struct tcphdr
*th
, unsigned len
)
4121 struct tcp_sock
*tp
= tcp_sk(sk
);
4124 tp
->rx_opt
.saw_tstamp
= 0;
4126 switch (sk
->sk_state
) {
4138 if(tp
->af_specific
->conn_request(sk
, skb
) < 0)
4141 /* Now we have several options: In theory there is
4142 * nothing else in the frame. KA9Q has an option to
4143 * send data with the syn, BSD accepts data with the
4144 * syn up to the [to be] advertised window and
4145 * Solaris 2.1 gives you a protocol error. For now
4146 * we just ignore it, that fits the spec precisely
4147 * and avoids incompatibilities. It would be nice in
4148 * future to drop through and process the data.
4150 * Now that TTCP is starting to be used we ought to
4152 * But, this leaves one open to an easy denial of
4153 * service attack, and SYN cookies can't defend
4154 * against this problem. So, we drop the data
4155 * in the interest of security over speed.
4162 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4166 /* Do step6 onward by hand. */
4167 tcp_urg(sk
, skb
, th
);
4169 tcp_data_snd_check(sk
, tp
);
4173 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4174 tcp_paws_discard(sk
, skb
)) {
4176 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4177 tcp_send_dupack(sk
, skb
);
4180 /* Reset is accepted even if it did not pass PAWS. */
4183 /* step 1: check sequence number */
4184 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4186 tcp_send_dupack(sk
, skb
);
4190 /* step 2: check RST bit */
4196 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4198 /* step 3: check security and precedence [ignored] */
4202 * Check for a SYN in window.
4204 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4205 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4210 /* step 5: check the ACK field */
4212 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4214 switch(sk
->sk_state
) {
4217 tp
->copied_seq
= tp
->rcv_nxt
;
4219 tcp_set_state(sk
, TCP_ESTABLISHED
);
4220 sk
->sk_state_change(sk
);
4222 /* Note, that this wakeup is only for marginal
4223 * crossed SYN case. Passively open sockets
4224 * are not waked up, because sk->sk_sleep ==
4225 * NULL and sk->sk_socket == NULL.
4227 if (sk
->sk_socket
) {
4228 sk_wake_async(sk
,0,POLL_OUT
);
4231 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4232 tp
->snd_wnd
= ntohs(th
->window
) <<
4233 tp
->rx_opt
.snd_wscale
;
4234 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4235 TCP_SKB_CB(skb
)->seq
);
4237 /* tcp_ack considers this ACK as duplicate
4238 * and does not calculate rtt.
4239 * Fix it at least with timestamps.
4241 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4243 tcp_ack_saw_tstamp(sk
, 0);
4245 if (tp
->rx_opt
.tstamp_ok
)
4246 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4248 /* Make sure socket is routed, for
4251 tp
->af_specific
->rebuild_header(sk
);
4253 tcp_init_metrics(sk
);
4255 tcp_init_congestion_control(sk
);
4257 /* Prevent spurious tcp_cwnd_restart() on
4258 * first data packet.
4260 tp
->lsndtime
= tcp_time_stamp
;
4262 tcp_initialize_rcv_mss(sk
);
4263 tcp_init_buffer_space(sk
);
4264 tcp_fast_path_on(tp
);
4271 if (tp
->snd_una
== tp
->write_seq
) {
4272 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4273 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4274 dst_confirm(sk
->sk_dst_cache
);
4276 if (!sock_flag(sk
, SOCK_DEAD
))
4277 /* Wake up lingering close() */
4278 sk
->sk_state_change(sk
);
4282 if (tp
->linger2
< 0 ||
4283 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4284 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4286 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4290 tmo
= tcp_fin_time(sk
);
4291 if (tmo
> TCP_TIMEWAIT_LEN
) {
4292 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4293 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4294 /* Bad case. We could lose such FIN otherwise.
4295 * It is not a big problem, but it looks confusing
4296 * and not so rare event. We still can lose it now,
4297 * if it spins in bh_lock_sock(), but it is really
4300 inet_csk_reset_keepalive_timer(sk
, tmo
);
4302 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4310 if (tp
->snd_una
== tp
->write_seq
) {
4311 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4317 if (tp
->snd_una
== tp
->write_seq
) {
4318 tcp_update_metrics(sk
);
4327 /* step 6: check the URG bit */
4328 tcp_urg(sk
, skb
, th
);
4330 /* step 7: process the segment text */
4331 switch (sk
->sk_state
) {
4332 case TCP_CLOSE_WAIT
:
4335 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4339 /* RFC 793 says to queue data in these states,
4340 * RFC 1122 says we MUST send a reset.
4341 * BSD 4.4 also does reset.
4343 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4344 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4345 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4346 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4352 case TCP_ESTABLISHED
:
4353 tcp_data_queue(sk
, skb
);
4358 /* tcp_data could move socket to TIME-WAIT */
4359 if (sk
->sk_state
!= TCP_CLOSE
) {
4360 tcp_data_snd_check(sk
, tp
);
4361 tcp_ack_snd_check(sk
);
4371 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4372 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4373 EXPORT_SYMBOL(tcp_parse_options
);
4374 EXPORT_SYMBOL(tcp_rcv_established
);
4375 EXPORT_SYMBOL(tcp_rcv_state_process
);