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
);
359 tp
->window_clamp
= min(tp
->window_clamp
, app_win
);
360 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
364 /* Receiver "autotuning" code.
366 * The algorithm for RTT estimation w/o timestamps is based on
367 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
368 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
370 * More detail on this code can be found at
371 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
372 * though this reference is out of date. A new paper
375 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
377 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
383 if (new_sample
!= 0) {
384 /* If we sample in larger samples in the non-timestamp
385 * case, we could grossly overestimate the RTT especially
386 * with chatty applications or bulk transfer apps which
387 * are stalled on filesystem I/O.
389 * Also, since we are only going for a minimum in the
390 * non-timestamp case, we do not smoothe things out
391 * else with timestamps disabled convergance takes too
395 m
-= (new_sample
>> 3);
397 } else if (m
< new_sample
)
400 /* No previous mesaure. */
404 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
405 tp
->rcv_rtt_est
.rtt
= new_sample
;
408 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
410 if (tp
->rcv_rtt_est
.time
== 0)
412 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
414 tcp_rcv_rtt_update(tp
,
415 jiffies
- tp
->rcv_rtt_est
.time
,
419 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
420 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
423 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
425 struct tcp_sock
*tp
= tcp_sk(sk
);
426 if (tp
->rx_opt
.rcv_tsecr
&&
427 (TCP_SKB_CB(skb
)->end_seq
-
428 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
429 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
433 * This function should be called every time data is copied to user space.
434 * It calculates the appropriate TCP receive buffer space.
436 void tcp_rcv_space_adjust(struct sock
*sk
)
438 struct tcp_sock
*tp
= tcp_sk(sk
);
442 if (tp
->rcvq_space
.time
== 0)
445 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
446 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
447 tp
->rcv_rtt_est
.rtt
== 0)
450 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
452 space
= max(tp
->rcvq_space
.space
, space
);
454 if (tp
->rcvq_space
.space
!= space
) {
457 tp
->rcvq_space
.space
= space
;
459 if (sysctl_tcp_moderate_rcvbuf
) {
460 int new_clamp
= space
;
462 /* Receive space grows, normalize in order to
463 * take into account packet headers and sk_buff
464 * structure overhead.
469 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
470 16 + sizeof(struct sk_buff
));
471 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
474 space
= min(space
, sysctl_tcp_rmem
[2]);
475 if (space
> sk
->sk_rcvbuf
) {
476 sk
->sk_rcvbuf
= space
;
478 /* Make the window clamp follow along. */
479 tp
->window_clamp
= new_clamp
;
485 tp
->rcvq_space
.seq
= tp
->copied_seq
;
486 tp
->rcvq_space
.time
= tcp_time_stamp
;
489 /* There is something which you must keep in mind when you analyze the
490 * behavior of the tp->ato delayed ack timeout interval. When a
491 * connection starts up, we want to ack as quickly as possible. The
492 * problem is that "good" TCP's do slow start at the beginning of data
493 * transmission. The means that until we send the first few ACK's the
494 * sender will sit on his end and only queue most of his data, because
495 * he can only send snd_cwnd unacked packets at any given time. For
496 * each ACK we send, he increments snd_cwnd and transmits more of his
499 static void tcp_event_data_recv(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
501 struct inet_connection_sock
*icsk
= inet_csk(sk
);
504 inet_csk_schedule_ack(sk
);
506 tcp_measure_rcv_mss(sk
, skb
);
508 tcp_rcv_rtt_measure(tp
);
510 now
= tcp_time_stamp
;
512 if (!icsk
->icsk_ack
.ato
) {
513 /* The _first_ data packet received, initialize
514 * delayed ACK engine.
516 tcp_incr_quickack(sk
);
517 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
519 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
521 if (m
<= TCP_ATO_MIN
/2) {
522 /* The fastest case is the first. */
523 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
524 } else if (m
< icsk
->icsk_ack
.ato
) {
525 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
526 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
527 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
528 } else if (m
> icsk
->icsk_rto
) {
529 /* Too long gap. Apparently sender falled to
530 * restart window, so that we send ACKs quickly.
532 tcp_incr_quickack(sk
);
533 sk_stream_mem_reclaim(sk
);
536 icsk
->icsk_ack
.lrcvtime
= now
;
538 TCP_ECN_check_ce(tp
, skb
);
541 tcp_grow_window(sk
, tp
, skb
);
544 /* Called to compute a smoothed rtt estimate. The data fed to this
545 * routine either comes from timestamps, or from segments that were
546 * known _not_ to have been retransmitted [see Karn/Partridge
547 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
548 * piece by Van Jacobson.
549 * NOTE: the next three routines used to be one big routine.
550 * To save cycles in the RFC 1323 implementation it was better to break
551 * it up into three procedures. -- erics
553 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
, u32
*usrtt
)
555 struct tcp_sock
*tp
= tcp_sk(sk
);
556 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
557 long m
= mrtt
; /* RTT */
559 /* The following amusing code comes from Jacobson's
560 * article in SIGCOMM '88. Note that rtt and mdev
561 * are scaled versions of rtt and mean deviation.
562 * This is designed to be as fast as possible
563 * m stands for "measurement".
565 * On a 1990 paper the rto value is changed to:
566 * RTO = rtt + 4 * mdev
568 * Funny. This algorithm seems to be very broken.
569 * These formulae increase RTO, when it should be decreased, increase
570 * too slowly, when it should be incresed fastly, decrease too fastly
571 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
572 * does not matter how to _calculate_ it. Seems, it was trap
573 * that VJ failed to avoid. 8)
578 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
579 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
581 m
= -m
; /* m is now abs(error) */
582 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
583 /* This is similar to one of Eifel findings.
584 * Eifel blocks mdev updates when rtt decreases.
585 * This solution is a bit different: we use finer gain
586 * for mdev in this case (alpha*beta).
587 * Like Eifel it also prevents growth of rto,
588 * but also it limits too fast rto decreases,
589 * happening in pure Eifel.
594 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
596 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
597 if (tp
->mdev
> tp
->mdev_max
) {
598 tp
->mdev_max
= tp
->mdev
;
599 if (tp
->mdev_max
> tp
->rttvar
)
600 tp
->rttvar
= tp
->mdev_max
;
602 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
603 if (tp
->mdev_max
< tp
->rttvar
)
604 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
605 tp
->rtt_seq
= tp
->snd_nxt
;
606 tp
->mdev_max
= TCP_RTO_MIN
;
609 /* no previous measure. */
610 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
611 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
612 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
613 tp
->rtt_seq
= tp
->snd_nxt
;
616 if (icsk
->icsk_ca_ops
->rtt_sample
)
617 icsk
->icsk_ca_ops
->rtt_sample(sk
, *usrtt
);
620 /* Calculate rto without backoff. This is the second half of Van Jacobson's
621 * routine referred to above.
623 static inline void tcp_set_rto(struct sock
*sk
)
625 const struct tcp_sock
*tp
= tcp_sk(sk
);
626 /* Old crap is replaced with new one. 8)
629 * 1. If rtt variance happened to be less 50msec, it is hallucination.
630 * It cannot be less due to utterly erratic ACK generation made
631 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
632 * to do with delayed acks, because at cwnd>2 true delack timeout
633 * is invisible. Actually, Linux-2.4 also generates erratic
634 * ACKs in some curcumstances.
636 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
638 /* 2. Fixups made earlier cannot be right.
639 * If we do not estimate RTO correctly without them,
640 * all the algo is pure shit and should be replaced
641 * with correct one. It is exaclty, which we pretend to do.
645 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
646 * guarantees that rto is higher.
648 static inline void tcp_bound_rto(struct sock
*sk
)
650 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
651 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
654 /* Save metrics learned by this TCP session.
655 This function is called only, when TCP finishes successfully
656 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
658 void tcp_update_metrics(struct sock
*sk
)
660 struct tcp_sock
*tp
= tcp_sk(sk
);
661 struct dst_entry
*dst
= __sk_dst_get(sk
);
663 if (sysctl_tcp_nometrics_save
)
668 if (dst
&& (dst
->flags
&DST_HOST
)) {
669 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
672 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
673 /* This session failed to estimate rtt. Why?
674 * Probably, no packets returned in time.
677 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
678 dst
->metrics
[RTAX_RTT
-1] = 0;
682 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
684 /* If newly calculated rtt larger than stored one,
685 * store new one. Otherwise, use EWMA. Remember,
686 * rtt overestimation is always better than underestimation.
688 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
690 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
692 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
695 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
699 /* Scale deviation to rttvar fixed point */
704 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
705 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
707 dst
->metrics
[RTAX_RTTVAR
-1] -=
708 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
711 if (tp
->snd_ssthresh
>= 0xFFFF) {
712 /* Slow start still did not finish. */
713 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
714 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
715 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
716 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
717 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
718 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
719 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
720 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
721 icsk
->icsk_ca_state
== TCP_CA_Open
) {
722 /* Cong. avoidance phase, cwnd is reliable. */
723 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
724 dst
->metrics
[RTAX_SSTHRESH
-1] =
725 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
726 if (!dst_metric_locked(dst
, RTAX_CWND
))
727 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
729 /* Else slow start did not finish, cwnd is non-sense,
730 ssthresh may be also invalid.
732 if (!dst_metric_locked(dst
, RTAX_CWND
))
733 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
734 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
735 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
736 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
737 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
740 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
741 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
742 tp
->reordering
!= sysctl_tcp_reordering
)
743 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
748 /* Numbers are taken from RFC2414. */
749 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
751 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
754 if (tp
->mss_cache
> 1460)
757 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
759 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
762 /* Initialize metrics on socket. */
764 static void tcp_init_metrics(struct sock
*sk
)
766 struct tcp_sock
*tp
= tcp_sk(sk
);
767 struct dst_entry
*dst
= __sk_dst_get(sk
);
774 if (dst_metric_locked(dst
, RTAX_CWND
))
775 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
776 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
777 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
778 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
779 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
781 if (dst_metric(dst
, RTAX_REORDERING
) &&
782 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
783 tp
->rx_opt
.sack_ok
&= ~2;
784 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
787 if (dst_metric(dst
, RTAX_RTT
) == 0)
790 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
793 /* Initial rtt is determined from SYN,SYN-ACK.
794 * The segment is small and rtt may appear much
795 * less than real one. Use per-dst memory
796 * to make it more realistic.
798 * A bit of theory. RTT is time passed after "normal" sized packet
799 * is sent until it is ACKed. In normal curcumstances sending small
800 * packets force peer to delay ACKs and calculation is correct too.
801 * The algorithm is adaptive and, provided we follow specs, it
802 * NEVER underestimate RTT. BUT! If peer tries to make some clever
803 * tricks sort of "quick acks" for time long enough to decrease RTT
804 * to low value, and then abruptly stops to do it and starts to delay
805 * ACKs, wait for troubles.
807 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
808 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
809 tp
->rtt_seq
= tp
->snd_nxt
;
811 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
812 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
813 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
817 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
819 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
820 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
824 /* Play conservative. If timestamps are not
825 * supported, TCP will fail to recalculate correct
826 * rtt, if initial rto is too small. FORGET ALL AND RESET!
828 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
830 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
831 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
835 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
838 struct tcp_sock
*tp
= tcp_sk(sk
);
839 if (metric
> tp
->reordering
) {
840 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
842 /* This exciting event is worth to be remembered. 8) */
844 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
846 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
848 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
850 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
851 #if FASTRETRANS_DEBUG > 1
852 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
853 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
857 tp
->undo_marker
? tp
->undo_retrans
: 0);
859 /* Disable FACK yet. */
860 tp
->rx_opt
.sack_ok
&= ~2;
864 /* This procedure tags the retransmission queue when SACKs arrive.
866 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
867 * Packets in queue with these bits set are counted in variables
868 * sacked_out, retrans_out and lost_out, correspondingly.
870 * Valid combinations are:
871 * Tag InFlight Description
872 * 0 1 - orig segment is in flight.
873 * S 0 - nothing flies, orig reached receiver.
874 * L 0 - nothing flies, orig lost by net.
875 * R 2 - both orig and retransmit are in flight.
876 * L|R 1 - orig is lost, retransmit is in flight.
877 * S|R 1 - orig reached receiver, retrans is still in flight.
878 * (L|S|R is logically valid, it could occur when L|R is sacked,
879 * but it is equivalent to plain S and code short-curcuits it to S.
880 * L|S is logically invalid, it would mean -1 packet in flight 8))
882 * These 6 states form finite state machine, controlled by the following events:
883 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
884 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
885 * 3. Loss detection event of one of three flavors:
886 * A. Scoreboard estimator decided the packet is lost.
887 * A'. Reno "three dupacks" marks head of queue lost.
888 * A''. Its FACK modfication, head until snd.fack is lost.
889 * B. SACK arrives sacking data transmitted after never retransmitted
891 * C. SACK arrives sacking SND.NXT at the moment, when the
892 * segment was retransmitted.
893 * 4. D-SACK added new rule: D-SACK changes any tag to S.
895 * It is pleasant to note, that state diagram turns out to be commutative,
896 * so that we are allowed not to be bothered by order of our actions,
897 * when multiple events arrive simultaneously. (see the function below).
899 * Reordering detection.
900 * --------------------
901 * Reordering metric is maximal distance, which a packet can be displaced
902 * in packet stream. With SACKs we can estimate it:
904 * 1. SACK fills old hole and the corresponding segment was not
905 * ever retransmitted -> reordering. Alas, we cannot use it
906 * when segment was retransmitted.
907 * 2. The last flaw is solved with D-SACK. D-SACK arrives
908 * for retransmitted and already SACKed segment -> reordering..
909 * Both of these heuristics are not used in Loss state, when we cannot
910 * account for retransmits accurately.
913 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
915 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
916 struct tcp_sock
*tp
= tcp_sk(sk
);
917 unsigned char *ptr
= ack_skb
->h
.raw
+ TCP_SKB_CB(ack_skb
)->sacked
;
918 struct tcp_sack_block
*sp
= (struct tcp_sack_block
*)(ptr
+2);
919 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
920 int reord
= tp
->packets_out
;
922 u32 lost_retrans
= 0;
928 prior_fackets
= tp
->fackets_out
;
930 for (i
=0; i
<num_sacks
; i
++, sp
++) {
932 __u32 start_seq
= ntohl(sp
->start_seq
);
933 __u32 end_seq
= ntohl(sp
->end_seq
);
937 /* Check for D-SACK. */
939 u32 ack
= TCP_SKB_CB(ack_skb
)->ack_seq
;
941 if (before(start_seq
, ack
)) {
943 tp
->rx_opt
.sack_ok
|= 4;
944 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
945 } else if (num_sacks
> 1 &&
946 !after(end_seq
, ntohl(sp
[1].end_seq
)) &&
947 !before(start_seq
, ntohl(sp
[1].start_seq
))) {
949 tp
->rx_opt
.sack_ok
|= 4;
950 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
953 /* D-SACK for already forgotten data...
954 * Do dumb counting. */
956 !after(end_seq
, prior_snd_una
) &&
957 after(end_seq
, tp
->undo_marker
))
960 /* Eliminate too old ACKs, but take into
961 * account more or less fresh ones, they can
962 * contain valid SACK info.
964 if (before(ack
, prior_snd_una
- tp
->max_window
))
968 /* Event "B" in the comment above. */
969 if (after(end_seq
, tp
->high_seq
))
970 flag
|= FLAG_DATA_LOST
;
972 sk_stream_for_retrans_queue(skb
, sk
) {
976 /* The retransmission queue is always in order, so
977 * we can short-circuit the walk early.
979 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
982 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
983 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
985 pcount
= tcp_skb_pcount(skb
);
987 if (pcount
> 1 && !in_sack
&&
988 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
989 unsigned int pkt_len
;
991 in_sack
= !after(start_seq
,
992 TCP_SKB_CB(skb
)->seq
);
995 pkt_len
= (start_seq
-
996 TCP_SKB_CB(skb
)->seq
);
999 TCP_SKB_CB(skb
)->seq
);
1000 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->tso_size
))
1002 pcount
= tcp_skb_pcount(skb
);
1005 fack_count
+= pcount
;
1007 sacked
= TCP_SKB_CB(skb
)->sacked
;
1009 /* Account D-SACK for retransmitted packet. */
1010 if ((dup_sack
&& in_sack
) &&
1011 (sacked
& TCPCB_RETRANS
) &&
1012 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1015 /* The frame is ACKed. */
1016 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1017 if (sacked
&TCPCB_RETRANS
) {
1018 if ((dup_sack
&& in_sack
) &&
1019 (sacked
&TCPCB_SACKED_ACKED
))
1020 reord
= min(fack_count
, reord
);
1022 /* If it was in a hole, we detected reordering. */
1023 if (fack_count
< prior_fackets
&&
1024 !(sacked
&TCPCB_SACKED_ACKED
))
1025 reord
= min(fack_count
, reord
);
1028 /* Nothing to do; acked frame is about to be dropped. */
1032 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1033 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1034 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1035 lost_retrans
= end_seq
;
1040 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1041 if (sacked
& TCPCB_SACKED_RETRANS
) {
1042 /* If the segment is not tagged as lost,
1043 * we do not clear RETRANS, believing
1044 * that retransmission is still in flight.
1046 if (sacked
& TCPCB_LOST
) {
1047 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1048 tp
->lost_out
-= tcp_skb_pcount(skb
);
1049 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1052 /* New sack for not retransmitted frame,
1053 * which was in hole. It is reordering.
1055 if (!(sacked
& TCPCB_RETRANS
) &&
1056 fack_count
< prior_fackets
)
1057 reord
= min(fack_count
, reord
);
1059 if (sacked
& TCPCB_LOST
) {
1060 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1061 tp
->lost_out
-= tcp_skb_pcount(skb
);
1065 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1066 flag
|= FLAG_DATA_SACKED
;
1067 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1069 if (fack_count
> tp
->fackets_out
)
1070 tp
->fackets_out
= fack_count
;
1072 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1073 reord
= min(fack_count
, reord
);
1076 /* D-SACK. We can detect redundant retransmission
1077 * in S|R and plain R frames and clear it.
1078 * undo_retrans is decreased above, L|R frames
1079 * are accounted above as well.
1082 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1083 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1084 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1089 /* Check for lost retransmit. This superb idea is
1090 * borrowed from "ratehalving". Event "C".
1091 * Later note: FACK people cheated me again 8),
1092 * we have to account for reordering! Ugly,
1095 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1096 struct sk_buff
*skb
;
1098 sk_stream_for_retrans_queue(skb
, sk
) {
1099 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1101 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1103 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1104 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1106 !before(lost_retrans
,
1107 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1109 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1110 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1112 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1113 tp
->lost_out
+= tcp_skb_pcount(skb
);
1114 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1115 flag
|= FLAG_DATA_SACKED
;
1116 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1122 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1124 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1125 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1127 #if FASTRETRANS_DEBUG > 0
1128 BUG_TRAP((int)tp
->sacked_out
>= 0);
1129 BUG_TRAP((int)tp
->lost_out
>= 0);
1130 BUG_TRAP((int)tp
->retrans_out
>= 0);
1131 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1136 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1137 * segments to see from the next ACKs whether any data was really missing.
1138 * If the RTO was spurious, new ACKs should arrive.
1140 void tcp_enter_frto(struct sock
*sk
)
1142 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1143 struct tcp_sock
*tp
= tcp_sk(sk
);
1144 struct sk_buff
*skb
;
1146 tp
->frto_counter
= 1;
1148 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1149 tp
->snd_una
== tp
->high_seq
||
1150 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1151 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1152 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1153 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1156 /* Have to clear retransmission markers here to keep the bookkeeping
1157 * in shape, even though we are not yet in Loss state.
1158 * If something was really lost, it is eventually caught up
1159 * in tcp_enter_frto_loss.
1161 tp
->retrans_out
= 0;
1162 tp
->undo_marker
= tp
->snd_una
;
1163 tp
->undo_retrans
= 0;
1165 sk_stream_for_retrans_queue(skb
, sk
) {
1166 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_RETRANS
;
1168 tcp_sync_left_out(tp
);
1170 tcp_set_ca_state(sk
, TCP_CA_Open
);
1171 tp
->frto_highmark
= tp
->snd_nxt
;
1174 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1175 * which indicates that we should follow the traditional RTO recovery,
1176 * i.e. mark everything lost and do go-back-N retransmission.
1178 static void tcp_enter_frto_loss(struct sock
*sk
)
1180 struct tcp_sock
*tp
= tcp_sk(sk
);
1181 struct sk_buff
*skb
;
1186 tp
->fackets_out
= 0;
1188 sk_stream_for_retrans_queue(skb
, sk
) {
1189 cnt
+= tcp_skb_pcount(skb
);
1190 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1191 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1193 /* Do not mark those segments lost that were
1194 * forward transmitted after RTO
1196 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1197 tp
->frto_highmark
)) {
1198 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1199 tp
->lost_out
+= tcp_skb_pcount(skb
);
1202 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1203 tp
->fackets_out
= cnt
;
1206 tcp_sync_left_out(tp
);
1208 tp
->snd_cwnd
= tp
->frto_counter
+ tcp_packets_in_flight(tp
)+1;
1209 tp
->snd_cwnd_cnt
= 0;
1210 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1211 tp
->undo_marker
= 0;
1212 tp
->frto_counter
= 0;
1214 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1215 sysctl_tcp_reordering
);
1216 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1217 tp
->high_seq
= tp
->frto_highmark
;
1218 TCP_ECN_queue_cwr(tp
);
1221 void tcp_clear_retrans(struct tcp_sock
*tp
)
1224 tp
->retrans_out
= 0;
1226 tp
->fackets_out
= 0;
1230 tp
->undo_marker
= 0;
1231 tp
->undo_retrans
= 0;
1234 /* Enter Loss state. If "how" is not zero, forget all SACK information
1235 * and reset tags completely, otherwise preserve SACKs. If receiver
1236 * dropped its ofo queue, we will know this due to reneging detection.
1238 void tcp_enter_loss(struct sock
*sk
, int how
)
1240 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1241 struct tcp_sock
*tp
= tcp_sk(sk
);
1242 struct sk_buff
*skb
;
1245 /* Reduce ssthresh if it has not yet been made inside this window. */
1246 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1247 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1248 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1249 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1250 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1253 tp
->snd_cwnd_cnt
= 0;
1254 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1256 tcp_clear_retrans(tp
);
1258 /* Push undo marker, if it was plain RTO and nothing
1259 * was retransmitted. */
1261 tp
->undo_marker
= tp
->snd_una
;
1263 sk_stream_for_retrans_queue(skb
, sk
) {
1264 cnt
+= tcp_skb_pcount(skb
);
1265 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1266 tp
->undo_marker
= 0;
1267 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1268 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1269 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1270 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1271 tp
->lost_out
+= tcp_skb_pcount(skb
);
1273 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1274 tp
->fackets_out
= cnt
;
1277 tcp_sync_left_out(tp
);
1279 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1280 sysctl_tcp_reordering
);
1281 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1282 tp
->high_seq
= tp
->snd_nxt
;
1283 TCP_ECN_queue_cwr(tp
);
1286 static int tcp_check_sack_reneging(struct sock
*sk
)
1288 struct sk_buff
*skb
;
1290 /* If ACK arrived pointing to a remembered SACK,
1291 * it means that our remembered SACKs do not reflect
1292 * real state of receiver i.e.
1293 * receiver _host_ is heavily congested (or buggy).
1294 * Do processing similar to RTO timeout.
1296 if ((skb
= skb_peek(&sk
->sk_write_queue
)) != NULL
&&
1297 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1298 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1299 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1301 tcp_enter_loss(sk
, 1);
1302 icsk
->icsk_retransmits
++;
1303 tcp_retransmit_skb(sk
, skb_peek(&sk
->sk_write_queue
));
1304 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1305 icsk
->icsk_rto
, TCP_RTO_MAX
);
1311 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1313 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1316 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1318 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1321 static inline int tcp_head_timedout(struct sock
*sk
, struct tcp_sock
*tp
)
1323 return tp
->packets_out
&&
1324 tcp_skb_timedout(sk
, skb_peek(&sk
->sk_write_queue
));
1327 /* Linux NewReno/SACK/FACK/ECN state machine.
1328 * --------------------------------------
1330 * "Open" Normal state, no dubious events, fast path.
1331 * "Disorder" In all the respects it is "Open",
1332 * but requires a bit more attention. It is entered when
1333 * we see some SACKs or dupacks. It is split of "Open"
1334 * mainly to move some processing from fast path to slow one.
1335 * "CWR" CWND was reduced due to some Congestion Notification event.
1336 * It can be ECN, ICMP source quench, local device congestion.
1337 * "Recovery" CWND was reduced, we are fast-retransmitting.
1338 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1340 * tcp_fastretrans_alert() is entered:
1341 * - each incoming ACK, if state is not "Open"
1342 * - when arrived ACK is unusual, namely:
1347 * Counting packets in flight is pretty simple.
1349 * in_flight = packets_out - left_out + retrans_out
1351 * packets_out is SND.NXT-SND.UNA counted in packets.
1353 * retrans_out is number of retransmitted segments.
1355 * left_out is number of segments left network, but not ACKed yet.
1357 * left_out = sacked_out + lost_out
1359 * sacked_out: Packets, which arrived to receiver out of order
1360 * and hence not ACKed. With SACKs this number is simply
1361 * amount of SACKed data. Even without SACKs
1362 * it is easy to give pretty reliable estimate of this number,
1363 * counting duplicate ACKs.
1365 * lost_out: Packets lost by network. TCP has no explicit
1366 * "loss notification" feedback from network (for now).
1367 * It means that this number can be only _guessed_.
1368 * Actually, it is the heuristics to predict lossage that
1369 * distinguishes different algorithms.
1371 * F.e. after RTO, when all the queue is considered as lost,
1372 * lost_out = packets_out and in_flight = retrans_out.
1374 * Essentially, we have now two algorithms counting
1377 * FACK: It is the simplest heuristics. As soon as we decided
1378 * that something is lost, we decide that _all_ not SACKed
1379 * packets until the most forward SACK are lost. I.e.
1380 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1381 * It is absolutely correct estimate, if network does not reorder
1382 * packets. And it loses any connection to reality when reordering
1383 * takes place. We use FACK by default until reordering
1384 * is suspected on the path to this destination.
1386 * NewReno: when Recovery is entered, we assume that one segment
1387 * is lost (classic Reno). While we are in Recovery and
1388 * a partial ACK arrives, we assume that one more packet
1389 * is lost (NewReno). This heuristics are the same in NewReno
1392 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1393 * deflation etc. CWND is real congestion window, never inflated, changes
1394 * only according to classic VJ rules.
1396 * Really tricky (and requiring careful tuning) part of algorithm
1397 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1398 * The first determines the moment _when_ we should reduce CWND and,
1399 * hence, slow down forward transmission. In fact, it determines the moment
1400 * when we decide that hole is caused by loss, rather than by a reorder.
1402 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1403 * holes, caused by lost packets.
1405 * And the most logically complicated part of algorithm is undo
1406 * heuristics. We detect false retransmits due to both too early
1407 * fast retransmit (reordering) and underestimated RTO, analyzing
1408 * timestamps and D-SACKs. When we detect that some segments were
1409 * retransmitted by mistake and CWND reduction was wrong, we undo
1410 * window reduction and abort recovery phase. This logic is hidden
1411 * inside several functions named tcp_try_undo_<something>.
1414 /* This function decides, when we should leave Disordered state
1415 * and enter Recovery phase, reducing congestion window.
1417 * Main question: may we further continue forward transmission
1418 * with the same cwnd?
1420 static int tcp_time_to_recover(struct sock
*sk
, struct tcp_sock
*tp
)
1424 /* Trick#1: The loss is proven. */
1428 /* Not-A-Trick#2 : Classic rule... */
1429 if (tcp_fackets_out(tp
) > tp
->reordering
)
1432 /* Trick#3 : when we use RFC2988 timer restart, fast
1433 * retransmit can be triggered by timeout of queue head.
1435 if (tcp_head_timedout(sk
, tp
))
1438 /* Trick#4: It is still not OK... But will it be useful to delay
1441 packets_out
= tp
->packets_out
;
1442 if (packets_out
<= tp
->reordering
&&
1443 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1444 !tcp_may_send_now(sk
, tp
)) {
1445 /* We have nothing to send. This connection is limited
1446 * either by receiver window or by application.
1454 /* If we receive more dupacks than we expected counting segments
1455 * in assumption of absent reordering, interpret this as reordering.
1456 * The only another reason could be bug in receiver TCP.
1458 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1460 struct tcp_sock
*tp
= tcp_sk(sk
);
1463 holes
= max(tp
->lost_out
, 1U);
1464 holes
= min(holes
, tp
->packets_out
);
1466 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1467 tp
->sacked_out
= tp
->packets_out
- holes
;
1468 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1472 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1474 static void tcp_add_reno_sack(struct sock
*sk
)
1476 struct tcp_sock
*tp
= tcp_sk(sk
);
1478 tcp_check_reno_reordering(sk
, 0);
1479 tcp_sync_left_out(tp
);
1482 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1484 static void tcp_remove_reno_sacks(struct sock
*sk
, struct tcp_sock
*tp
, int acked
)
1487 /* One ACK acked hole. The rest eat duplicate ACKs. */
1488 if (acked
-1 >= tp
->sacked_out
)
1491 tp
->sacked_out
-= acked
-1;
1493 tcp_check_reno_reordering(sk
, acked
);
1494 tcp_sync_left_out(tp
);
1497 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1500 tp
->left_out
= tp
->lost_out
;
1503 /* Mark head of queue up as lost. */
1504 static void tcp_mark_head_lost(struct sock
*sk
, struct tcp_sock
*tp
,
1505 int packets
, u32 high_seq
)
1507 struct sk_buff
*skb
;
1510 BUG_TRAP(cnt
<= tp
->packets_out
);
1512 sk_stream_for_retrans_queue(skb
, sk
) {
1513 cnt
-= tcp_skb_pcount(skb
);
1514 if (cnt
< 0 || after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1516 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1517 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1518 tp
->lost_out
+= tcp_skb_pcount(skb
);
1521 tcp_sync_left_out(tp
);
1524 /* Account newly detected lost packet(s) */
1526 static void tcp_update_scoreboard(struct sock
*sk
, struct tcp_sock
*tp
)
1529 int lost
= tp
->fackets_out
- tp
->reordering
;
1532 tcp_mark_head_lost(sk
, tp
, lost
, tp
->high_seq
);
1534 tcp_mark_head_lost(sk
, tp
, 1, tp
->high_seq
);
1537 /* New heuristics: it is possible only after we switched
1538 * to restart timer each time when something is ACKed.
1539 * Hence, we can detect timed out packets during fast
1540 * retransmit without falling to slow start.
1542 if (tcp_head_timedout(sk
, tp
)) {
1543 struct sk_buff
*skb
;
1545 sk_stream_for_retrans_queue(skb
, sk
) {
1546 if (tcp_skb_timedout(sk
, skb
) &&
1547 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1548 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1549 tp
->lost_out
+= tcp_skb_pcount(skb
);
1552 tcp_sync_left_out(tp
);
1556 /* CWND moderation, preventing bursts due to too big ACKs
1557 * in dubious situations.
1559 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1561 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1562 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1563 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1566 /* Decrease cwnd each second ack. */
1567 static void tcp_cwnd_down(struct sock
*sk
)
1569 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1570 struct tcp_sock
*tp
= tcp_sk(sk
);
1571 int decr
= tp
->snd_cwnd_cnt
+ 1;
1573 tp
->snd_cwnd_cnt
= decr
&1;
1576 if (decr
&& tp
->snd_cwnd
> icsk
->icsk_ca_ops
->min_cwnd(sk
))
1577 tp
->snd_cwnd
-= decr
;
1579 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1580 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1583 /* Nothing was retransmitted or returned timestamp is less
1584 * than timestamp of the first retransmission.
1586 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1588 return !tp
->retrans_stamp
||
1589 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1590 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1593 /* Undo procedures. */
1595 #if FASTRETRANS_DEBUG > 1
1596 static void DBGUNDO(struct sock
*sk
, struct tcp_sock
*tp
, const char *msg
)
1598 struct inet_sock
*inet
= inet_sk(sk
);
1599 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1601 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1602 tp
->snd_cwnd
, tp
->left_out
,
1603 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1607 #define DBGUNDO(x...) do { } while (0)
1610 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1612 struct tcp_sock
*tp
= tcp_sk(sk
);
1614 if (tp
->prior_ssthresh
) {
1615 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1617 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1618 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1620 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1622 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1623 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1624 TCP_ECN_withdraw_cwr(tp
);
1627 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1629 tcp_moderate_cwnd(tp
);
1630 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1633 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1635 return tp
->undo_marker
&&
1636 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1639 /* People celebrate: "We love our President!" */
1640 static int tcp_try_undo_recovery(struct sock
*sk
, struct tcp_sock
*tp
)
1642 if (tcp_may_undo(tp
)) {
1643 /* Happy end! We did not retransmit anything
1644 * or our original transmission succeeded.
1646 DBGUNDO(sk
, tp
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1647 tcp_undo_cwr(sk
, 1);
1648 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1649 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1651 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1652 tp
->undo_marker
= 0;
1654 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1655 /* Hold old state until something *above* high_seq
1656 * is ACKed. For Reno it is MUST to prevent false
1657 * fast retransmits (RFC2582). SACK TCP is safe. */
1658 tcp_moderate_cwnd(tp
);
1661 tcp_set_ca_state(sk
, TCP_CA_Open
);
1665 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1666 static void tcp_try_undo_dsack(struct sock
*sk
, struct tcp_sock
*tp
)
1668 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1669 DBGUNDO(sk
, tp
, "D-SACK");
1670 tcp_undo_cwr(sk
, 1);
1671 tp
->undo_marker
= 0;
1672 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1676 /* Undo during fast recovery after partial ACK. */
1678 static int tcp_try_undo_partial(struct sock
*sk
, struct tcp_sock
*tp
,
1681 /* Partial ACK arrived. Force Hoe's retransmit. */
1682 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1684 if (tcp_may_undo(tp
)) {
1685 /* Plain luck! Hole if filled with delayed
1686 * packet, rather than with a retransmit.
1688 if (tp
->retrans_out
== 0)
1689 tp
->retrans_stamp
= 0;
1691 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
1693 DBGUNDO(sk
, tp
, "Hoe");
1694 tcp_undo_cwr(sk
, 0);
1695 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1697 /* So... Do not make Hoe's retransmit yet.
1698 * If the first packet was delayed, the rest
1699 * ones are most probably delayed as well.
1706 /* Undo during loss recovery after partial ACK. */
1707 static int tcp_try_undo_loss(struct sock
*sk
, struct tcp_sock
*tp
)
1709 if (tcp_may_undo(tp
)) {
1710 struct sk_buff
*skb
;
1711 sk_stream_for_retrans_queue(skb
, sk
) {
1712 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1714 DBGUNDO(sk
, tp
, "partial loss");
1716 tp
->left_out
= tp
->sacked_out
;
1717 tcp_undo_cwr(sk
, 1);
1718 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1719 inet_csk(sk
)->icsk_retransmits
= 0;
1720 tp
->undo_marker
= 0;
1722 tcp_set_ca_state(sk
, TCP_CA_Open
);
1728 static inline void tcp_complete_cwr(struct sock
*sk
)
1730 struct tcp_sock
*tp
= tcp_sk(sk
);
1731 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1732 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1733 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
1736 static void tcp_try_to_open(struct sock
*sk
, struct tcp_sock
*tp
, int flag
)
1738 tp
->left_out
= tp
->sacked_out
;
1740 if (tp
->retrans_out
== 0)
1741 tp
->retrans_stamp
= 0;
1746 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
1747 int state
= TCP_CA_Open
;
1749 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
1750 state
= TCP_CA_Disorder
;
1752 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
1753 tcp_set_ca_state(sk
, state
);
1754 tp
->high_seq
= tp
->snd_nxt
;
1756 tcp_moderate_cwnd(tp
);
1762 /* Process an event, which can update packets-in-flight not trivially.
1763 * Main goal of this function is to calculate new estimate for left_out,
1764 * taking into account both packets sitting in receiver's buffer and
1765 * packets lost by network.
1767 * Besides that it does CWND reduction, when packet loss is detected
1768 * and changes state of machine.
1770 * It does _not_ decide what to send, it is made in function
1771 * tcp_xmit_retransmit_queue().
1774 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
1775 int prior_packets
, int flag
)
1777 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1778 struct tcp_sock
*tp
= tcp_sk(sk
);
1779 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
1781 /* Some technical things:
1782 * 1. Reno does not count dupacks (sacked_out) automatically. */
1783 if (!tp
->packets_out
)
1785 /* 2. SACK counts snd_fack in packets inaccurately. */
1786 if (tp
->sacked_out
== 0)
1787 tp
->fackets_out
= 0;
1789 /* Now state machine starts.
1790 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1792 tp
->prior_ssthresh
= 0;
1794 /* B. In all the states check for reneging SACKs. */
1795 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
1798 /* C. Process data loss notification, provided it is valid. */
1799 if ((flag
&FLAG_DATA_LOST
) &&
1800 before(tp
->snd_una
, tp
->high_seq
) &&
1801 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
1802 tp
->fackets_out
> tp
->reordering
) {
1803 tcp_mark_head_lost(sk
, tp
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
1804 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
1807 /* D. Synchronize left_out to current state. */
1808 tcp_sync_left_out(tp
);
1810 /* E. Check state exit conditions. State can be terminated
1811 * when high_seq is ACKed. */
1812 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
1813 if (!sysctl_tcp_frto
)
1814 BUG_TRAP(tp
->retrans_out
== 0);
1815 tp
->retrans_stamp
= 0;
1816 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
1817 switch (icsk
->icsk_ca_state
) {
1819 icsk
->icsk_retransmits
= 0;
1820 if (tcp_try_undo_recovery(sk
, tp
))
1825 /* CWR is to be held something *above* high_seq
1826 * is ACKed for CWR bit to reach receiver. */
1827 if (tp
->snd_una
!= tp
->high_seq
) {
1828 tcp_complete_cwr(sk
);
1829 tcp_set_ca_state(sk
, TCP_CA_Open
);
1833 case TCP_CA_Disorder
:
1834 tcp_try_undo_dsack(sk
, tp
);
1835 if (!tp
->undo_marker
||
1836 /* For SACK case do not Open to allow to undo
1837 * catching for all duplicate ACKs. */
1838 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
1839 tp
->undo_marker
= 0;
1840 tcp_set_ca_state(sk
, TCP_CA_Open
);
1844 case TCP_CA_Recovery
:
1846 tcp_reset_reno_sack(tp
);
1847 if (tcp_try_undo_recovery(sk
, tp
))
1849 tcp_complete_cwr(sk
);
1854 /* F. Process state. */
1855 switch (icsk
->icsk_ca_state
) {
1856 case TCP_CA_Recovery
:
1857 if (prior_snd_una
== tp
->snd_una
) {
1858 if (IsReno(tp
) && is_dupack
)
1859 tcp_add_reno_sack(sk
);
1861 int acked
= prior_packets
- tp
->packets_out
;
1863 tcp_remove_reno_sacks(sk
, tp
, acked
);
1864 is_dupack
= tcp_try_undo_partial(sk
, tp
, acked
);
1868 if (flag
&FLAG_DATA_ACKED
)
1869 icsk
->icsk_retransmits
= 0;
1870 if (!tcp_try_undo_loss(sk
, tp
)) {
1871 tcp_moderate_cwnd(tp
);
1872 tcp_xmit_retransmit_queue(sk
);
1875 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1877 /* Loss is undone; fall through to processing in Open state. */
1880 if (tp
->snd_una
!= prior_snd_una
)
1881 tcp_reset_reno_sack(tp
);
1883 tcp_add_reno_sack(sk
);
1886 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
1887 tcp_try_undo_dsack(sk
, tp
);
1889 if (!tcp_time_to_recover(sk
, tp
)) {
1890 tcp_try_to_open(sk
, tp
, flag
);
1894 /* Otherwise enter Recovery state */
1897 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
1899 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
1901 tp
->high_seq
= tp
->snd_nxt
;
1902 tp
->prior_ssthresh
= 0;
1903 tp
->undo_marker
= tp
->snd_una
;
1904 tp
->undo_retrans
= tp
->retrans_out
;
1906 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
1907 if (!(flag
&FLAG_ECE
))
1908 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1909 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1910 TCP_ECN_queue_cwr(tp
);
1913 tp
->snd_cwnd_cnt
= 0;
1914 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
1917 if (is_dupack
|| tcp_head_timedout(sk
, tp
))
1918 tcp_update_scoreboard(sk
, tp
);
1920 tcp_xmit_retransmit_queue(sk
);
1923 /* Read draft-ietf-tcplw-high-performance before mucking
1924 * with this code. (Superceeds RFC1323)
1926 static void tcp_ack_saw_tstamp(struct sock
*sk
, u32
*usrtt
, int flag
)
1928 /* RTTM Rule: A TSecr value received in a segment is used to
1929 * update the averaged RTT measurement only if the segment
1930 * acknowledges some new data, i.e., only if it advances the
1931 * left edge of the send window.
1933 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1934 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1936 * Changed: reset backoff as soon as we see the first valid sample.
1937 * If we do not, we get strongly overstimated rto. With timestamps
1938 * samples are accepted even from very old segments: f.e., when rtt=1
1939 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1940 * answer arrives rto becomes 120 seconds! If at least one of segments
1941 * in window is lost... Voila. --ANK (010210)
1943 struct tcp_sock
*tp
= tcp_sk(sk
);
1944 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
1945 tcp_rtt_estimator(sk
, seq_rtt
, usrtt
);
1947 inet_csk(sk
)->icsk_backoff
= 0;
1951 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, u32
*usrtt
, int flag
)
1953 /* We don't have a timestamp. Can only use
1954 * packets that are not retransmitted to determine
1955 * rtt estimates. Also, we must not reset the
1956 * backoff for rto until we get a non-retransmitted
1957 * packet. This allows us to deal with a situation
1958 * where the network delay has increased suddenly.
1959 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1962 if (flag
& FLAG_RETRANS_DATA_ACKED
)
1965 tcp_rtt_estimator(sk
, seq_rtt
, usrtt
);
1967 inet_csk(sk
)->icsk_backoff
= 0;
1971 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
1972 const s32 seq_rtt
, u32
*usrtt
)
1974 const struct tcp_sock
*tp
= tcp_sk(sk
);
1975 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1976 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
1977 tcp_ack_saw_tstamp(sk
, usrtt
, flag
);
1978 else if (seq_rtt
>= 0)
1979 tcp_ack_no_tstamp(sk
, seq_rtt
, usrtt
, flag
);
1982 static inline void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
1983 u32 in_flight
, int good
)
1985 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1986 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
1987 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
1990 /* Restart timer after forward progress on connection.
1991 * RFC2988 recommends to restart timer to now+rto.
1994 static inline void tcp_ack_packets_out(struct sock
*sk
, struct tcp_sock
*tp
)
1996 if (!tp
->packets_out
) {
1997 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
1999 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2003 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2004 __u32 now
, __s32
*seq_rtt
)
2006 struct tcp_sock
*tp
= tcp_sk(sk
);
2007 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2008 __u32 seq
= tp
->snd_una
;
2009 __u32 packets_acked
;
2012 /* If we get here, the whole TSO packet has not been
2015 BUG_ON(!after(scb
->end_seq
, seq
));
2017 packets_acked
= tcp_skb_pcount(skb
);
2018 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2020 packets_acked
-= tcp_skb_pcount(skb
);
2022 if (packets_acked
) {
2023 __u8 sacked
= scb
->sacked
;
2025 acked
|= FLAG_DATA_ACKED
;
2027 if (sacked
& TCPCB_RETRANS
) {
2028 if (sacked
& TCPCB_SACKED_RETRANS
)
2029 tp
->retrans_out
-= packets_acked
;
2030 acked
|= FLAG_RETRANS_DATA_ACKED
;
2032 } else if (*seq_rtt
< 0)
2033 *seq_rtt
= now
- scb
->when
;
2034 if (sacked
& TCPCB_SACKED_ACKED
)
2035 tp
->sacked_out
-= packets_acked
;
2036 if (sacked
& TCPCB_LOST
)
2037 tp
->lost_out
-= packets_acked
;
2038 if (sacked
& TCPCB_URG
) {
2040 !before(seq
, tp
->snd_up
))
2043 } else if (*seq_rtt
< 0)
2044 *seq_rtt
= now
- scb
->when
;
2046 if (tp
->fackets_out
) {
2047 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2048 tp
->fackets_out
-= dval
;
2050 tp
->packets_out
-= packets_acked
;
2052 BUG_ON(tcp_skb_pcount(skb
) == 0);
2053 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2060 /* Remove acknowledged frames from the retransmission queue. */
2061 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
, s32
*seq_usrtt
)
2063 struct tcp_sock
*tp
= tcp_sk(sk
);
2064 struct sk_buff
*skb
;
2065 __u32 now
= tcp_time_stamp
;
2068 struct timeval usnow
;
2072 do_gettimeofday(&usnow
);
2074 while ((skb
= skb_peek(&sk
->sk_write_queue
)) &&
2075 skb
!= sk
->sk_send_head
) {
2076 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2077 __u8 sacked
= scb
->sacked
;
2079 /* If our packet is before the ack sequence we can
2080 * discard it as it's confirmed to have arrived at
2083 if (after(scb
->end_seq
, tp
->snd_una
)) {
2084 if (tcp_skb_pcount(skb
) > 1 &&
2085 after(tp
->snd_una
, scb
->seq
))
2086 acked
|= tcp_tso_acked(sk
, skb
,
2091 /* Initial outgoing SYN's get put onto the write_queue
2092 * just like anything else we transmit. It is not
2093 * true data, and if we misinform our callers that
2094 * this ACK acks real data, we will erroneously exit
2095 * connection startup slow start one packet too
2096 * quickly. This is severely frowned upon behavior.
2098 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2099 acked
|= FLAG_DATA_ACKED
;
2102 acked
|= FLAG_SYN_ACKED
;
2103 tp
->retrans_stamp
= 0;
2107 if (sacked
& TCPCB_RETRANS
) {
2108 if(sacked
& TCPCB_SACKED_RETRANS
)
2109 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2110 acked
|= FLAG_RETRANS_DATA_ACKED
;
2112 } else if (seq_rtt
< 0)
2113 seq_rtt
= now
- scb
->when
;
2117 skb_get_timestamp(skb
, &tv
);
2118 *seq_usrtt
= (usnow
.tv_sec
- tv
.tv_sec
) * 1000000
2119 + (usnow
.tv_usec
- tv
.tv_usec
);
2122 if (sacked
& TCPCB_SACKED_ACKED
)
2123 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2124 if (sacked
& TCPCB_LOST
)
2125 tp
->lost_out
-= tcp_skb_pcount(skb
);
2126 if (sacked
& TCPCB_URG
) {
2128 !before(scb
->end_seq
, tp
->snd_up
))
2131 } else if (seq_rtt
< 0)
2132 seq_rtt
= now
- scb
->when
;
2133 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2134 tcp_packets_out_dec(tp
, skb
);
2135 __skb_unlink(skb
, &sk
->sk_write_queue
);
2136 sk_stream_free_skb(sk
, skb
);
2139 if (acked
&FLAG_ACKED
) {
2140 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2141 tcp_ack_update_rtt(sk
, acked
, seq_rtt
, seq_usrtt
);
2142 tcp_ack_packets_out(sk
, tp
);
2144 if (icsk
->icsk_ca_ops
->pkts_acked
)
2145 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
);
2148 #if FASTRETRANS_DEBUG > 0
2149 BUG_TRAP((int)tp
->sacked_out
>= 0);
2150 BUG_TRAP((int)tp
->lost_out
>= 0);
2151 BUG_TRAP((int)tp
->retrans_out
>= 0);
2152 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2153 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2155 printk(KERN_DEBUG
"Leak l=%u %d\n",
2156 tp
->lost_out
, icsk
->icsk_ca_state
);
2159 if (tp
->sacked_out
) {
2160 printk(KERN_DEBUG
"Leak s=%u %d\n",
2161 tp
->sacked_out
, icsk
->icsk_ca_state
);
2164 if (tp
->retrans_out
) {
2165 printk(KERN_DEBUG
"Leak r=%u %d\n",
2166 tp
->retrans_out
, icsk
->icsk_ca_state
);
2167 tp
->retrans_out
= 0;
2171 *seq_rtt_p
= seq_rtt
;
2175 static void tcp_ack_probe(struct sock
*sk
)
2177 const struct tcp_sock
*tp
= tcp_sk(sk
);
2178 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2180 /* Was it a usable window open? */
2182 if (!after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
2183 tp
->snd_una
+ tp
->snd_wnd
)) {
2184 icsk
->icsk_backoff
= 0;
2185 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2186 /* Socket must be waked up by subsequent tcp_data_snd_check().
2187 * This function is not for random using!
2190 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2191 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2196 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2198 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2199 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2202 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2204 const struct tcp_sock
*tp
= tcp_sk(sk
);
2205 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2206 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2209 /* Check that window update is acceptable.
2210 * The function assumes that snd_una<=ack<=snd_next.
2212 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2213 const u32 ack_seq
, const u32 nwin
)
2215 return (after(ack
, tp
->snd_una
) ||
2216 after(ack_seq
, tp
->snd_wl1
) ||
2217 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2220 /* Update our send window.
2222 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2223 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2225 static int tcp_ack_update_window(struct sock
*sk
, struct tcp_sock
*tp
,
2226 struct sk_buff
*skb
, u32 ack
, u32 ack_seq
)
2229 u32 nwin
= ntohs(skb
->h
.th
->window
);
2231 if (likely(!skb
->h
.th
->syn
))
2232 nwin
<<= tp
->rx_opt
.snd_wscale
;
2234 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2235 flag
|= FLAG_WIN_UPDATE
;
2236 tcp_update_wl(tp
, ack
, ack_seq
);
2238 if (tp
->snd_wnd
!= nwin
) {
2241 /* Note, it is the only place, where
2242 * fast path is recovered for sending TCP.
2244 tcp_fast_path_check(sk
, tp
);
2246 if (nwin
> tp
->max_window
) {
2247 tp
->max_window
= nwin
;
2248 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
2258 static void tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
)
2260 struct tcp_sock
*tp
= tcp_sk(sk
);
2262 tcp_sync_left_out(tp
);
2264 if (tp
->snd_una
== prior_snd_una
||
2265 !before(tp
->snd_una
, tp
->frto_highmark
)) {
2266 /* RTO was caused by loss, start retransmitting in
2267 * go-back-N slow start
2269 tcp_enter_frto_loss(sk
);
2273 if (tp
->frto_counter
== 1) {
2274 /* First ACK after RTO advances the window: allow two new
2277 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2279 /* Also the second ACK after RTO advances the window.
2280 * The RTO was likely spurious. Reduce cwnd and continue
2281 * in congestion avoidance
2283 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2284 tcp_moderate_cwnd(tp
);
2287 /* F-RTO affects on two new ACKs following RTO.
2288 * At latest on third ACK the TCP behavor is back to normal.
2290 tp
->frto_counter
= (tp
->frto_counter
+ 1) % 3;
2293 /* This routine deals with incoming acks, but not outgoing ones. */
2294 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2296 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2297 struct tcp_sock
*tp
= tcp_sk(sk
);
2298 u32 prior_snd_una
= tp
->snd_una
;
2299 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2300 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2301 u32 prior_in_flight
;
2306 /* If the ack is newer than sent or older than previous acks
2307 * then we can probably ignore it.
2309 if (after(ack
, tp
->snd_nxt
))
2310 goto uninteresting_ack
;
2312 if (before(ack
, prior_snd_una
))
2315 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2316 /* Window is constant, pure forward advance.
2317 * No more checks are required.
2318 * Note, we use the fact that SND.UNA>=SND.WL2.
2320 tcp_update_wl(tp
, ack
, ack_seq
);
2322 flag
|= FLAG_WIN_UPDATE
;
2324 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2326 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2328 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2331 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2333 flag
|= tcp_ack_update_window(sk
, tp
, skb
, ack
, ack_seq
);
2335 if (TCP_SKB_CB(skb
)->sacked
)
2336 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2338 if (TCP_ECN_rcv_ecn_echo(tp
, skb
->h
.th
))
2341 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2344 /* We passed data and got it acked, remove any soft error
2345 * log. Something worked...
2347 sk
->sk_err_soft
= 0;
2348 tp
->rcv_tstamp
= tcp_time_stamp
;
2349 prior_packets
= tp
->packets_out
;
2353 prior_in_flight
= tcp_packets_in_flight(tp
);
2355 /* See if we can take anything off of the retransmit queue. */
2356 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
,
2357 icsk
->icsk_ca_ops
->rtt_sample
? &seq_usrtt
: NULL
);
2359 if (tp
->frto_counter
)
2360 tcp_process_frto(sk
, prior_snd_una
);
2362 if (tcp_ack_is_dubious(sk
, flag
)) {
2363 /* Advanve CWND, if state allows this. */
2364 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
2365 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2366 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2368 if ((flag
& FLAG_DATA_ACKED
))
2369 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2372 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2373 dst_confirm(sk
->sk_dst_cache
);
2378 icsk
->icsk_probes_out
= 0;
2380 /* If this ack opens up a zero window, clear backoff. It was
2381 * being used to time the probes, and is probably far higher than
2382 * it needs to be for normal retransmission.
2384 if (sk
->sk_send_head
)
2389 if (TCP_SKB_CB(skb
)->sacked
)
2390 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2393 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2398 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2399 * But, this can also be called on packets in the established flow when
2400 * the fast version below fails.
2402 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2405 struct tcphdr
*th
= skb
->h
.th
;
2406 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2408 ptr
= (unsigned char *)(th
+ 1);
2409 opt_rx
->saw_tstamp
= 0;
2418 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2423 if (opsize
< 2) /* "silly options" */
2425 if (opsize
> length
)
2426 return; /* don't parse partial options */
2429 if(opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2430 u16 in_mss
= ntohs(get_unaligned((__u16
*)ptr
));
2432 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2433 in_mss
= opt_rx
->user_mss
;
2434 opt_rx
->mss_clamp
= in_mss
;
2439 if(opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2440 if (sysctl_tcp_window_scaling
) {
2441 __u8 snd_wscale
= *(__u8
*) ptr
;
2442 opt_rx
->wscale_ok
= 1;
2443 if (snd_wscale
> 14) {
2445 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2446 "scaling value %d >14 received.\n",
2450 opt_rx
->snd_wscale
= snd_wscale
;
2453 case TCPOPT_TIMESTAMP
:
2454 if(opsize
==TCPOLEN_TIMESTAMP
) {
2455 if ((estab
&& opt_rx
->tstamp_ok
) ||
2456 (!estab
&& sysctl_tcp_timestamps
)) {
2457 opt_rx
->saw_tstamp
= 1;
2458 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__u32
*)ptr
));
2459 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__u32
*)(ptr
+4)));
2463 case TCPOPT_SACK_PERM
:
2464 if(opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2465 if (sysctl_tcp_sack
) {
2466 opt_rx
->sack_ok
= 1;
2467 tcp_sack_reset(opt_rx
);
2473 if((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2474 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2476 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2485 /* Fast parse options. This hopes to only see timestamps.
2486 * If it is wrong it falls back on tcp_parse_options().
2488 static inline int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2489 struct tcp_sock
*tp
)
2491 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2492 tp
->rx_opt
.saw_tstamp
= 0;
2494 } else if (tp
->rx_opt
.tstamp_ok
&&
2495 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2496 __u32
*ptr
= (__u32
*)(th
+ 1);
2497 if (*ptr
== ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2498 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2499 tp
->rx_opt
.saw_tstamp
= 1;
2501 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2503 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2507 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2511 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2513 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2514 tp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
2517 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2519 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2520 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2521 * extra check below makes sure this can only happen
2522 * for pure ACK frames. -DaveM
2524 * Not only, also it occurs for expired timestamps.
2527 if((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2528 xtime
.tv_sec
>= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2529 tcp_store_ts_recent(tp
);
2533 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2535 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2536 * it can pass through stack. So, the following predicate verifies that
2537 * this segment is not used for anything but congestion avoidance or
2538 * fast retransmit. Moreover, we even are able to eliminate most of such
2539 * second order effects, if we apply some small "replay" window (~RTO)
2540 * to timestamp space.
2542 * All these measures still do not guarantee that we reject wrapped ACKs
2543 * on networks with high bandwidth, when sequence space is recycled fastly,
2544 * but it guarantees that such events will be very rare and do not affect
2545 * connection seriously. This doesn't look nice, but alas, PAWS is really
2548 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2549 * states that events when retransmit arrives after original data are rare.
2550 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2551 * the biggest problem on large power networks even with minor reordering.
2552 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2553 * up to bandwidth of 18Gigabit/sec. 8) ]
2556 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
2558 struct tcp_sock
*tp
= tcp_sk(sk
);
2559 struct tcphdr
*th
= skb
->h
.th
;
2560 u32 seq
= TCP_SKB_CB(skb
)->seq
;
2561 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2563 return (/* 1. Pure ACK with correct sequence number. */
2564 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
2566 /* 2. ... and duplicate ACK. */
2567 ack
== tp
->snd_una
&&
2569 /* 3. ... and does not update window. */
2570 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
2572 /* 4. ... and sits in replay window. */
2573 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
2576 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
2578 const struct tcp_sock
*tp
= tcp_sk(sk
);
2579 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
2580 xtime
.tv_sec
< tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
2581 !tcp_disordered_ack(sk
, skb
));
2584 /* Check segment sequence number for validity.
2586 * Segment controls are considered valid, if the segment
2587 * fits to the window after truncation to the window. Acceptability
2588 * of data (and SYN, FIN, of course) is checked separately.
2589 * See tcp_data_queue(), for example.
2591 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2592 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2593 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2594 * (borrowed from freebsd)
2597 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2599 return !before(end_seq
, tp
->rcv_wup
) &&
2600 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
2603 /* When we get a reset we do this. */
2604 static void tcp_reset(struct sock
*sk
)
2606 /* We want the right error as BSD sees it (and indeed as we do). */
2607 switch (sk
->sk_state
) {
2609 sk
->sk_err
= ECONNREFUSED
;
2611 case TCP_CLOSE_WAIT
:
2617 sk
->sk_err
= ECONNRESET
;
2620 if (!sock_flag(sk
, SOCK_DEAD
))
2621 sk
->sk_error_report(sk
);
2627 * Process the FIN bit. This now behaves as it is supposed to work
2628 * and the FIN takes effect when it is validly part of sequence
2629 * space. Not before when we get holes.
2631 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2632 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2635 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2636 * close and we go into CLOSING (and later onto TIME-WAIT)
2638 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2640 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
2642 struct tcp_sock
*tp
= tcp_sk(sk
);
2644 inet_csk_schedule_ack(sk
);
2646 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
2647 sock_set_flag(sk
, SOCK_DONE
);
2649 switch (sk
->sk_state
) {
2651 case TCP_ESTABLISHED
:
2652 /* Move to CLOSE_WAIT */
2653 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
2654 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
2657 case TCP_CLOSE_WAIT
:
2659 /* Received a retransmission of the FIN, do
2664 /* RFC793: Remain in the LAST-ACK state. */
2668 /* This case occurs when a simultaneous close
2669 * happens, we must ack the received FIN and
2670 * enter the CLOSING state.
2673 tcp_set_state(sk
, TCP_CLOSING
);
2676 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2678 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
2681 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2682 * cases we should never reach this piece of code.
2684 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
2685 __FUNCTION__
, sk
->sk_state
);
2689 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2690 * Probably, we should reset in this case. For now drop them.
2692 __skb_queue_purge(&tp
->out_of_order_queue
);
2693 if (tp
->rx_opt
.sack_ok
)
2694 tcp_sack_reset(&tp
->rx_opt
);
2695 sk_stream_mem_reclaim(sk
);
2697 if (!sock_flag(sk
, SOCK_DEAD
)) {
2698 sk
->sk_state_change(sk
);
2700 /* Do not send POLL_HUP for half duplex close. */
2701 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
2702 sk
->sk_state
== TCP_CLOSE
)
2703 sk_wake_async(sk
, 1, POLL_HUP
);
2705 sk_wake_async(sk
, 1, POLL_IN
);
2709 static __inline__
int
2710 tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
2712 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
2713 if (before(seq
, sp
->start_seq
))
2714 sp
->start_seq
= seq
;
2715 if (after(end_seq
, sp
->end_seq
))
2716 sp
->end_seq
= end_seq
;
2722 static inline void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2724 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2725 if (before(seq
, tp
->rcv_nxt
))
2726 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
2728 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
2730 tp
->rx_opt
.dsack
= 1;
2731 tp
->duplicate_sack
[0].start_seq
= seq
;
2732 tp
->duplicate_sack
[0].end_seq
= end_seq
;
2733 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
2737 static inline void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2739 if (!tp
->rx_opt
.dsack
)
2740 tcp_dsack_set(tp
, seq
, end_seq
);
2742 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
2745 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
2747 struct tcp_sock
*tp
= tcp_sk(sk
);
2749 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
2750 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
2751 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
2752 tcp_enter_quickack_mode(sk
);
2754 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2755 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2757 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
2758 end_seq
= tp
->rcv_nxt
;
2759 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
2766 /* These routines update the SACK block as out-of-order packets arrive or
2767 * in-order packets close up the sequence space.
2769 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
2772 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2773 struct tcp_sack_block
*swalk
= sp
+1;
2775 /* See if the recent change to the first SACK eats into
2776 * or hits the sequence space of other SACK blocks, if so coalesce.
2778 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
2779 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
2782 /* Zap SWALK, by moving every further SACK up by one slot.
2783 * Decrease num_sacks.
2785 tp
->rx_opt
.num_sacks
--;
2786 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2787 for(i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
2791 this_sack
++, swalk
++;
2795 static __inline__
void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
2799 tmp
= sack1
->start_seq
;
2800 sack1
->start_seq
= sack2
->start_seq
;
2801 sack2
->start_seq
= tmp
;
2803 tmp
= sack1
->end_seq
;
2804 sack1
->end_seq
= sack2
->end_seq
;
2805 sack2
->end_seq
= tmp
;
2808 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
2810 struct tcp_sock
*tp
= tcp_sk(sk
);
2811 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2812 int cur_sacks
= tp
->rx_opt
.num_sacks
;
2818 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
2819 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
2820 /* Rotate this_sack to the first one. */
2821 for (; this_sack
>0; this_sack
--, sp
--)
2822 tcp_sack_swap(sp
, sp
-1);
2824 tcp_sack_maybe_coalesce(tp
);
2829 /* Could not find an adjacent existing SACK, build a new one,
2830 * put it at the front, and shift everyone else down. We
2831 * always know there is at least one SACK present already here.
2833 * If the sack array is full, forget about the last one.
2835 if (this_sack
>= 4) {
2837 tp
->rx_opt
.num_sacks
--;
2840 for(; this_sack
> 0; this_sack
--, sp
--)
2844 /* Build the new head SACK, and we're done. */
2845 sp
->start_seq
= seq
;
2846 sp
->end_seq
= end_seq
;
2847 tp
->rx_opt
.num_sacks
++;
2848 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2851 /* RCV.NXT advances, some SACKs should be eaten. */
2853 static void tcp_sack_remove(struct tcp_sock
*tp
)
2855 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2856 int num_sacks
= tp
->rx_opt
.num_sacks
;
2859 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2860 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
2861 tp
->rx_opt
.num_sacks
= 0;
2862 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
2866 for(this_sack
= 0; this_sack
< num_sacks
; ) {
2867 /* Check if the start of the sack is covered by RCV.NXT. */
2868 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
2871 /* RCV.NXT must cover all the block! */
2872 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
2874 /* Zap this SACK, by moving forward any other SACKS. */
2875 for (i
=this_sack
+1; i
< num_sacks
; i
++)
2876 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
2883 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
2884 tp
->rx_opt
.num_sacks
= num_sacks
;
2885 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2889 /* This one checks to see if we can put data from the
2890 * out_of_order queue into the receive_queue.
2892 static void tcp_ofo_queue(struct sock
*sk
)
2894 struct tcp_sock
*tp
= tcp_sk(sk
);
2895 __u32 dsack_high
= tp
->rcv_nxt
;
2896 struct sk_buff
*skb
;
2898 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
2899 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
2902 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
2903 __u32 dsack
= dsack_high
;
2904 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
2905 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
2906 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
2909 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
2910 SOCK_DEBUG(sk
, "ofo packet was already received \n");
2911 __skb_unlink(skb
, &tp
->out_of_order_queue
);
2915 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
2916 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
2917 TCP_SKB_CB(skb
)->end_seq
);
2919 __skb_unlink(skb
, &tp
->out_of_order_queue
);
2920 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
2921 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
2923 tcp_fin(skb
, sk
, skb
->h
.th
);
2927 static int tcp_prune_queue(struct sock
*sk
);
2929 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
2931 struct tcphdr
*th
= skb
->h
.th
;
2932 struct tcp_sock
*tp
= tcp_sk(sk
);
2935 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
2938 __skb_pull(skb
, th
->doff
*4);
2940 TCP_ECN_accept_cwr(tp
, skb
);
2942 if (tp
->rx_opt
.dsack
) {
2943 tp
->rx_opt
.dsack
= 0;
2944 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
2945 4 - tp
->rx_opt
.tstamp_ok
);
2948 /* Queue data for delivery to the user.
2949 * Packets in sequence go to the receive queue.
2950 * Out of sequence packets to the out_of_order_queue.
2952 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
2953 if (tcp_receive_window(tp
) == 0)
2956 /* Ok. In sequence. In window. */
2957 if (tp
->ucopy
.task
== current
&&
2958 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
2959 sock_owned_by_user(sk
) && !tp
->urg_data
) {
2960 int chunk
= min_t(unsigned int, skb
->len
,
2963 __set_current_state(TASK_RUNNING
);
2966 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
2967 tp
->ucopy
.len
-= chunk
;
2968 tp
->copied_seq
+= chunk
;
2969 eaten
= (chunk
== skb
->len
&& !th
->fin
);
2970 tcp_rcv_space_adjust(sk
);
2978 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
2979 !sk_stream_rmem_schedule(sk
, skb
))) {
2980 if (tcp_prune_queue(sk
) < 0 ||
2981 !sk_stream_rmem_schedule(sk
, skb
))
2984 sk_stream_set_owner_r(skb
, sk
);
2985 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
2987 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
2989 tcp_event_data_recv(sk
, tp
, skb
);
2991 tcp_fin(skb
, sk
, th
);
2993 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
2996 /* RFC2581. 4.2. SHOULD send immediate ACK, when
2997 * gap in queue is filled.
2999 if (skb_queue_empty(&tp
->out_of_order_queue
))
3000 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3003 if (tp
->rx_opt
.num_sacks
)
3004 tcp_sack_remove(tp
);
3006 tcp_fast_path_check(sk
, tp
);
3010 else if (!sock_flag(sk
, SOCK_DEAD
))
3011 sk
->sk_data_ready(sk
, 0);
3015 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3016 /* A retransmit, 2nd most common case. Force an immediate ack. */
3017 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3018 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3021 tcp_enter_quickack_mode(sk
);
3022 inet_csk_schedule_ack(sk
);
3028 /* Out of window. F.e. zero window probe. */
3029 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3032 tcp_enter_quickack_mode(sk
);
3034 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3035 /* Partial packet, seq < rcv_next < end_seq */
3036 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3037 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3038 TCP_SKB_CB(skb
)->end_seq
);
3040 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3042 /* If window is closed, drop tail of packet. But after
3043 * remembering D-SACK for its head made in previous line.
3045 if (!tcp_receive_window(tp
))
3050 TCP_ECN_check_ce(tp
, skb
);
3052 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3053 !sk_stream_rmem_schedule(sk
, skb
)) {
3054 if (tcp_prune_queue(sk
) < 0 ||
3055 !sk_stream_rmem_schedule(sk
, skb
))
3059 /* Disable header prediction. */
3061 inet_csk_schedule_ack(sk
);
3063 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3064 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3066 sk_stream_set_owner_r(skb
, sk
);
3068 if (!skb_peek(&tp
->out_of_order_queue
)) {
3069 /* Initial out of order segment, build 1 SACK. */
3070 if (tp
->rx_opt
.sack_ok
) {
3071 tp
->rx_opt
.num_sacks
= 1;
3072 tp
->rx_opt
.dsack
= 0;
3073 tp
->rx_opt
.eff_sacks
= 1;
3074 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3075 tp
->selective_acks
[0].end_seq
=
3076 TCP_SKB_CB(skb
)->end_seq
;
3078 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3080 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3081 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3082 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3084 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3085 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3087 if (!tp
->rx_opt
.num_sacks
||
3088 tp
->selective_acks
[0].end_seq
!= seq
)
3091 /* Common case: data arrive in order after hole. */
3092 tp
->selective_acks
[0].end_seq
= end_seq
;
3096 /* Find place to insert this segment. */
3098 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3100 } while ((skb1
= skb1
->prev
) !=
3101 (struct sk_buff
*)&tp
->out_of_order_queue
);
3103 /* Do skb overlap to previous one? */
3104 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3105 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3106 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3107 /* All the bits are present. Drop. */
3109 tcp_dsack_set(tp
, seq
, end_seq
);
3112 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3113 /* Partial overlap. */
3114 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3119 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3121 /* And clean segments covered by new one as whole. */
3122 while ((skb1
= skb
->next
) !=
3123 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3124 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3125 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3126 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3129 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3130 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3135 if (tp
->rx_opt
.sack_ok
)
3136 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3140 /* Collapse contiguous sequence of skbs head..tail with
3141 * sequence numbers start..end.
3142 * Segments with FIN/SYN are not collapsed (only because this
3146 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3147 struct sk_buff
*head
, struct sk_buff
*tail
,
3150 struct sk_buff
*skb
;
3152 /* First, check that queue is collapsable and find
3153 * the point where collapsing can be useful. */
3154 for (skb
= head
; skb
!= tail
; ) {
3155 /* No new bits? It is possible on ofo queue. */
3156 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3157 struct sk_buff
*next
= skb
->next
;
3158 __skb_unlink(skb
, list
);
3160 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3165 /* The first skb to collapse is:
3167 * - bloated or contains data before "start" or
3168 * overlaps to the next one.
3170 if (!skb
->h
.th
->syn
&& !skb
->h
.th
->fin
&&
3171 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3172 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3173 (skb
->next
!= tail
&&
3174 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3177 /* Decided to skip this, advance start seq. */
3178 start
= TCP_SKB_CB(skb
)->end_seq
;
3181 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3184 while (before(start
, end
)) {
3185 struct sk_buff
*nskb
;
3186 int header
= skb_headroom(skb
);
3187 int copy
= SKB_MAX_ORDER(header
, 0);
3189 /* Too big header? This can happen with IPv6. */
3192 if (end
-start
< copy
)
3194 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3197 skb_reserve(nskb
, header
);
3198 memcpy(nskb
->head
, skb
->head
, header
);
3199 nskb
->nh
.raw
= nskb
->head
+ (skb
->nh
.raw
-skb
->head
);
3200 nskb
->h
.raw
= nskb
->head
+ (skb
->h
.raw
-skb
->head
);
3201 nskb
->mac
.raw
= nskb
->head
+ (skb
->mac
.raw
-skb
->head
);
3202 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3203 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3204 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3205 sk_stream_set_owner_r(nskb
, sk
);
3207 /* Copy data, releasing collapsed skbs. */
3209 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3210 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3212 if (offset
< 0) BUG();
3214 size
= min(copy
, size
);
3215 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3217 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3221 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3222 struct sk_buff
*next
= skb
->next
;
3223 __skb_unlink(skb
, list
);
3225 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3227 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3234 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3235 * and tcp_collapse() them until all the queue is collapsed.
3237 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3239 struct tcp_sock
*tp
= tcp_sk(sk
);
3240 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3241 struct sk_buff
*head
;
3247 start
= TCP_SKB_CB(skb
)->seq
;
3248 end
= TCP_SKB_CB(skb
)->end_seq
;
3254 /* Segment is terminated when we see gap or when
3255 * we are at the end of all the queue. */
3256 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3257 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3258 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3259 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3260 head
, skb
, start
, end
);
3262 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3264 /* Start new segment */
3265 start
= TCP_SKB_CB(skb
)->seq
;
3266 end
= TCP_SKB_CB(skb
)->end_seq
;
3268 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3269 start
= TCP_SKB_CB(skb
)->seq
;
3270 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3271 end
= TCP_SKB_CB(skb
)->end_seq
;
3276 /* Reduce allocated memory if we can, trying to get
3277 * the socket within its memory limits again.
3279 * Return less than zero if we should start dropping frames
3280 * until the socket owning process reads some of the data
3281 * to stabilize the situation.
3283 static int tcp_prune_queue(struct sock
*sk
)
3285 struct tcp_sock
*tp
= tcp_sk(sk
);
3287 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3289 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3291 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3292 tcp_clamp_window(sk
, tp
);
3293 else if (tcp_memory_pressure
)
3294 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3296 tcp_collapse_ofo_queue(sk
);
3297 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3298 sk
->sk_receive_queue
.next
,
3299 (struct sk_buff
*)&sk
->sk_receive_queue
,
3300 tp
->copied_seq
, tp
->rcv_nxt
);
3301 sk_stream_mem_reclaim(sk
);
3303 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3306 /* Collapsing did not help, destructive actions follow.
3307 * This must not ever occur. */
3309 /* First, purge the out_of_order queue. */
3310 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3311 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3312 __skb_queue_purge(&tp
->out_of_order_queue
);
3314 /* Reset SACK state. A conforming SACK implementation will
3315 * do the same at a timeout based retransmit. When a connection
3316 * is in a sad state like this, we care only about integrity
3317 * of the connection not performance.
3319 if (tp
->rx_opt
.sack_ok
)
3320 tcp_sack_reset(&tp
->rx_opt
);
3321 sk_stream_mem_reclaim(sk
);
3324 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3327 /* If we are really being abused, tell the caller to silently
3328 * drop receive data on the floor. It will get retransmitted
3329 * and hopefully then we'll have sufficient space.
3331 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3333 /* Massive buffer overcommit. */
3339 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3340 * As additional protections, we do not touch cwnd in retransmission phases,
3341 * and if application hit its sndbuf limit recently.
3343 void tcp_cwnd_application_limited(struct sock
*sk
)
3345 struct tcp_sock
*tp
= tcp_sk(sk
);
3347 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3348 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3349 /* Limited by application or receiver window. */
3350 u32 win_used
= max(tp
->snd_cwnd_used
, 2U);
3351 if (win_used
< tp
->snd_cwnd
) {
3352 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3353 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3355 tp
->snd_cwnd_used
= 0;
3357 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3360 static inline int tcp_should_expand_sndbuf(struct sock
*sk
, struct tcp_sock
*tp
)
3362 /* If the user specified a specific send buffer setting, do
3365 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3368 /* If we are under global TCP memory pressure, do not expand. */
3369 if (tcp_memory_pressure
)
3372 /* If we are under soft global TCP memory pressure, do not expand. */
3373 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3376 /* If we filled the congestion window, do not expand. */
3377 if (tp
->packets_out
>= tp
->snd_cwnd
)
3383 /* When incoming ACK allowed to free some skb from write_queue,
3384 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3385 * on the exit from tcp input handler.
3387 * PROBLEM: sndbuf expansion does not work well with largesend.
3389 static void tcp_new_space(struct sock
*sk
)
3391 struct tcp_sock
*tp
= tcp_sk(sk
);
3393 if (tcp_should_expand_sndbuf(sk
, tp
)) {
3394 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3395 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3396 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3397 tp
->reordering
+ 1);
3398 sndmem
*= 2*demanded
;
3399 if (sndmem
> sk
->sk_sndbuf
)
3400 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3401 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3404 sk
->sk_write_space(sk
);
3407 static inline void tcp_check_space(struct sock
*sk
)
3409 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3410 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3411 if (sk
->sk_socket
&&
3412 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3417 static __inline__
void tcp_data_snd_check(struct sock
*sk
, struct tcp_sock
*tp
)
3419 tcp_push_pending_frames(sk
, tp
);
3420 tcp_check_space(sk
);
3424 * Check if sending an ack is needed.
3426 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3428 struct tcp_sock
*tp
= tcp_sk(sk
);
3430 /* More than one full frame received... */
3431 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3432 /* ... and right edge of window advances far enough.
3433 * (tcp_recvmsg() will send ACK otherwise). Or...
3435 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3436 /* We ACK each frame or... */
3437 tcp_in_quickack_mode(sk
) ||
3438 /* We have out of order data. */
3440 skb_peek(&tp
->out_of_order_queue
))) {
3441 /* Then ack it now */
3444 /* Else, send delayed ack. */
3445 tcp_send_delayed_ack(sk
);
3449 static __inline__
void tcp_ack_snd_check(struct sock
*sk
)
3451 if (!inet_csk_ack_scheduled(sk
)) {
3452 /* We sent a data segment already. */
3455 __tcp_ack_snd_check(sk
, 1);
3459 * This routine is only called when we have urgent data
3460 * signalled. Its the 'slow' part of tcp_urg. It could be
3461 * moved inline now as tcp_urg is only called from one
3462 * place. We handle URGent data wrong. We have to - as
3463 * BSD still doesn't use the correction from RFC961.
3464 * For 1003.1g we should support a new option TCP_STDURG to permit
3465 * either form (or just set the sysctl tcp_stdurg).
3468 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3470 struct tcp_sock
*tp
= tcp_sk(sk
);
3471 u32 ptr
= ntohs(th
->urg_ptr
);
3473 if (ptr
&& !sysctl_tcp_stdurg
)
3475 ptr
+= ntohl(th
->seq
);
3477 /* Ignore urgent data that we've already seen and read. */
3478 if (after(tp
->copied_seq
, ptr
))
3481 /* Do not replay urg ptr.
3483 * NOTE: interesting situation not covered by specs.
3484 * Misbehaving sender may send urg ptr, pointing to segment,
3485 * which we already have in ofo queue. We are not able to fetch
3486 * such data and will stay in TCP_URG_NOTYET until will be eaten
3487 * by recvmsg(). Seems, we are not obliged to handle such wicked
3488 * situations. But it is worth to think about possibility of some
3489 * DoSes using some hypothetical application level deadlock.
3491 if (before(ptr
, tp
->rcv_nxt
))
3494 /* Do we already have a newer (or duplicate) urgent pointer? */
3495 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3498 /* Tell the world about our new urgent pointer. */
3501 /* We may be adding urgent data when the last byte read was
3502 * urgent. To do this requires some care. We cannot just ignore
3503 * tp->copied_seq since we would read the last urgent byte again
3504 * as data, nor can we alter copied_seq until this data arrives
3505 * or we break the sematics of SIOCATMARK (and thus sockatmark())
3507 * NOTE. Double Dutch. Rendering to plain English: author of comment
3508 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3509 * and expect that both A and B disappear from stream. This is _wrong_.
3510 * Though this happens in BSD with high probability, this is occasional.
3511 * Any application relying on this is buggy. Note also, that fix "works"
3512 * only in this artificial test. Insert some normal data between A and B and we will
3513 * decline of BSD again. Verdict: it is better to remove to trap
3516 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3517 !sock_flag(sk
, SOCK_URGINLINE
) &&
3518 tp
->copied_seq
!= tp
->rcv_nxt
) {
3519 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3521 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3522 __skb_unlink(skb
, &sk
->sk_receive_queue
);
3527 tp
->urg_data
= TCP_URG_NOTYET
;
3530 /* Disable header prediction. */
3534 /* This is the 'fast' part of urgent handling. */
3535 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
3537 struct tcp_sock
*tp
= tcp_sk(sk
);
3539 /* Check if we get a new urgent pointer - normally not. */
3541 tcp_check_urg(sk
,th
);
3543 /* Do we wait for any urgent data? - normally not... */
3544 if (tp
->urg_data
== TCP_URG_NOTYET
) {
3545 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
3548 /* Is the urgent pointer pointing into this packet? */
3549 if (ptr
< skb
->len
) {
3551 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
3553 tp
->urg_data
= TCP_URG_VALID
| tmp
;
3554 if (!sock_flag(sk
, SOCK_DEAD
))
3555 sk
->sk_data_ready(sk
, 0);
3560 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3562 struct tcp_sock
*tp
= tcp_sk(sk
);
3563 int chunk
= skb
->len
- hlen
;
3567 if (skb
->ip_summed
==CHECKSUM_UNNECESSARY
)
3568 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
3570 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
3574 tp
->ucopy
.len
-= chunk
;
3575 tp
->copied_seq
+= chunk
;
3576 tcp_rcv_space_adjust(sk
);
3583 static int __tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3587 if (sock_owned_by_user(sk
)) {
3589 result
= __tcp_checksum_complete(skb
);
3592 result
= __tcp_checksum_complete(skb
);
3597 static __inline__
int
3598 tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3600 return skb
->ip_summed
!= CHECKSUM_UNNECESSARY
&&
3601 __tcp_checksum_complete_user(sk
, skb
);
3605 * TCP receive function for the ESTABLISHED state.
3607 * It is split into a fast path and a slow path. The fast path is
3609 * - A zero window was announced from us - zero window probing
3610 * is only handled properly in the slow path.
3611 * - Out of order segments arrived.
3612 * - Urgent data is expected.
3613 * - There is no buffer space left
3614 * - Unexpected TCP flags/window values/header lengths are received
3615 * (detected by checking the TCP header against pred_flags)
3616 * - Data is sent in both directions. Fast path only supports pure senders
3617 * or pure receivers (this means either the sequence number or the ack
3618 * value must stay constant)
3619 * - Unexpected TCP option.
3621 * When these conditions are not satisfied it drops into a standard
3622 * receive procedure patterned after RFC793 to handle all cases.
3623 * The first three cases are guaranteed by proper pred_flags setting,
3624 * the rest is checked inline. Fast processing is turned on in
3625 * tcp_data_queue when everything is OK.
3627 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
3628 struct tcphdr
*th
, unsigned len
)
3630 struct tcp_sock
*tp
= tcp_sk(sk
);
3633 * Header prediction.
3634 * The code loosely follows the one in the famous
3635 * "30 instruction TCP receive" Van Jacobson mail.
3637 * Van's trick is to deposit buffers into socket queue
3638 * on a device interrupt, to call tcp_recv function
3639 * on the receive process context and checksum and copy
3640 * the buffer to user space. smart...
3642 * Our current scheme is not silly either but we take the
3643 * extra cost of the net_bh soft interrupt processing...
3644 * We do checksum and copy also but from device to kernel.
3647 tp
->rx_opt
.saw_tstamp
= 0;
3649 /* pred_flags is 0xS?10 << 16 + snd_wnd
3650 * if header_predition is to be made
3651 * 'S' will always be tp->tcp_header_len >> 2
3652 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3653 * turn it off (when there are holes in the receive
3654 * space for instance)
3655 * PSH flag is ignored.
3658 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
3659 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3660 int tcp_header_len
= tp
->tcp_header_len
;
3662 /* Timestamp header prediction: tcp_header_len
3663 * is automatically equal to th->doff*4 due to pred_flags
3667 /* Check timestamp */
3668 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
3669 __u32
*ptr
= (__u32
*)(th
+ 1);
3671 /* No? Slow path! */
3672 if (*ptr
!= ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3673 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
3676 tp
->rx_opt
.saw_tstamp
= 1;
3678 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3680 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3682 /* If PAWS failed, check it more carefully in slow path */
3683 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
3686 /* DO NOT update ts_recent here, if checksum fails
3687 * and timestamp was corrupted part, it will result
3688 * in a hung connection since we will drop all
3689 * future packets due to the PAWS test.
3693 if (len
<= tcp_header_len
) {
3694 /* Bulk data transfer: sender */
3695 if (len
== tcp_header_len
) {
3696 /* Predicted packet is in window by definition.
3697 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3698 * Hence, check seq<=rcv_wup reduces to:
3700 if (tcp_header_len
==
3701 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3702 tp
->rcv_nxt
== tp
->rcv_wup
)
3703 tcp_store_ts_recent(tp
);
3705 tcp_rcv_rtt_measure_ts(sk
, skb
);
3707 /* We know that such packets are checksummed
3710 tcp_ack(sk
, skb
, 0);
3712 tcp_data_snd_check(sk
, tp
);
3714 } else { /* Header too small */
3715 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3721 if (tp
->ucopy
.task
== current
&&
3722 tp
->copied_seq
== tp
->rcv_nxt
&&
3723 len
- tcp_header_len
<= tp
->ucopy
.len
&&
3724 sock_owned_by_user(sk
)) {
3725 __set_current_state(TASK_RUNNING
);
3727 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
3728 /* Predicted packet is in window by definition.
3729 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3730 * Hence, check seq<=rcv_wup reduces to:
3732 if (tcp_header_len
==
3733 (sizeof(struct tcphdr
) +
3734 TCPOLEN_TSTAMP_ALIGNED
) &&
3735 tp
->rcv_nxt
== tp
->rcv_wup
)
3736 tcp_store_ts_recent(tp
);
3738 tcp_rcv_rtt_measure_ts(sk
, skb
);
3740 __skb_pull(skb
, tcp_header_len
);
3741 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3742 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
3747 if (tcp_checksum_complete_user(sk
, skb
))
3750 /* Predicted packet is in window by definition.
3751 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3752 * Hence, check seq<=rcv_wup reduces to:
3754 if (tcp_header_len
==
3755 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3756 tp
->rcv_nxt
== tp
->rcv_wup
)
3757 tcp_store_ts_recent(tp
);
3759 tcp_rcv_rtt_measure_ts(sk
, skb
);
3761 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
3764 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
3766 /* Bulk data transfer: receiver */
3767 __skb_pull(skb
,tcp_header_len
);
3768 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3769 sk_stream_set_owner_r(skb
, sk
);
3770 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3773 tcp_event_data_recv(sk
, tp
, skb
);
3775 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
3776 /* Well, only one small jumplet in fast path... */
3777 tcp_ack(sk
, skb
, FLAG_DATA
);
3778 tcp_data_snd_check(sk
, tp
);
3779 if (!inet_csk_ack_scheduled(sk
))
3783 __tcp_ack_snd_check(sk
, 0);
3788 sk
->sk_data_ready(sk
, 0);
3794 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
3798 * RFC1323: H1. Apply PAWS check first.
3800 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
3801 tcp_paws_discard(sk
, skb
)) {
3803 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
3804 tcp_send_dupack(sk
, skb
);
3807 /* Resets are accepted even if PAWS failed.
3809 ts_recent update must be made after we are sure
3810 that the packet is in window.
3815 * Standard slow path.
3818 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3819 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3820 * (RST) segments are validated by checking their SEQ-fields."
3821 * And page 69: "If an incoming segment is not acceptable,
3822 * an acknowledgment should be sent in reply (unless the RST bit
3823 * is set, if so drop the segment and return)".
3826 tcp_send_dupack(sk
, skb
);
3835 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3837 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3838 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3839 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
3846 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
3848 tcp_rcv_rtt_measure_ts(sk
, skb
);
3850 /* Process urgent data. */
3851 tcp_urg(sk
, skb
, th
);
3853 /* step 7: process the segment text */
3854 tcp_data_queue(sk
, skb
);
3856 tcp_data_snd_check(sk
, tp
);
3857 tcp_ack_snd_check(sk
);
3861 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3868 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
3869 struct tcphdr
*th
, unsigned len
)
3871 struct tcp_sock
*tp
= tcp_sk(sk
);
3872 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
3874 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
3877 struct inet_connection_sock
*icsk
;
3879 * "If the state is SYN-SENT then
3880 * first check the ACK bit
3881 * If the ACK bit is set
3882 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3883 * a reset (unless the RST bit is set, if so drop
3884 * the segment and return)"
3886 * We do not send data with SYN, so that RFC-correct
3889 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
3890 goto reset_and_undo
;
3892 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
3893 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
3895 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
3896 goto reset_and_undo
;
3899 /* Now ACK is acceptable.
3901 * "If the RST bit is set
3902 * If the ACK was acceptable then signal the user "error:
3903 * connection reset", drop the segment, enter CLOSED state,
3904 * delete TCB, and return."
3913 * "fifth, if neither of the SYN or RST bits is set then
3914 * drop the segment and return."
3920 goto discard_and_undo
;
3923 * "If the SYN bit is on ...
3924 * are acceptable then ...
3925 * (our SYN has been ACKed), change the connection
3926 * state to ESTABLISHED..."
3929 TCP_ECN_rcv_synack(tp
, th
);
3930 if (tp
->ecn_flags
&TCP_ECN_OK
)
3931 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
3933 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
3934 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
3936 /* Ok.. it's good. Set up sequence numbers and
3937 * move to established.
3939 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
3940 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
3942 /* RFC1323: The window in SYN & SYN/ACK segments is
3945 tp
->snd_wnd
= ntohs(th
->window
);
3946 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
3948 if (!tp
->rx_opt
.wscale_ok
) {
3949 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
3950 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
3953 if (tp
->rx_opt
.saw_tstamp
) {
3954 tp
->rx_opt
.tstamp_ok
= 1;
3955 tp
->tcp_header_len
=
3956 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
3957 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
3958 tcp_store_ts_recent(tp
);
3960 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3963 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
3964 tp
->rx_opt
.sack_ok
|= 2;
3966 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
3967 tcp_initialize_rcv_mss(sk
);
3969 /* Remember, tcp_poll() does not lock socket!
3970 * Change state from SYN-SENT only after copied_seq
3971 * is initialized. */
3972 tp
->copied_seq
= tp
->rcv_nxt
;
3974 tcp_set_state(sk
, TCP_ESTABLISHED
);
3976 /* Make sure socket is routed, for correct metrics. */
3977 tp
->af_specific
->rebuild_header(sk
);
3979 tcp_init_metrics(sk
);
3981 tcp_init_congestion_control(sk
);
3983 /* Prevent spurious tcp_cwnd_restart() on first data
3986 tp
->lsndtime
= tcp_time_stamp
;
3988 tcp_init_buffer_space(sk
);
3990 if (sock_flag(sk
, SOCK_KEEPOPEN
))
3991 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
3993 if (!tp
->rx_opt
.snd_wscale
)
3994 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
3998 if (!sock_flag(sk
, SOCK_DEAD
)) {
3999 sk
->sk_state_change(sk
);
4000 sk_wake_async(sk
, 0, POLL_OUT
);
4003 icsk
= inet_csk(sk
);
4005 if (sk
->sk_write_pending
||
4006 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4007 icsk
->icsk_ack
.pingpong
) {
4008 /* Save one ACK. Data will be ready after
4009 * several ticks, if write_pending is set.
4011 * It may be deleted, but with this feature tcpdumps
4012 * look so _wonderfully_ clever, that I was not able
4013 * to stand against the temptation 8) --ANK
4015 inet_csk_schedule_ack(sk
);
4016 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4017 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4018 tcp_incr_quickack(sk
);
4019 tcp_enter_quickack_mode(sk
);
4020 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4021 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4032 /* No ACK in the segment */
4036 * "If the RST bit is set
4038 * Otherwise (no ACK) drop the segment and return."
4041 goto discard_and_undo
;
4045 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4046 goto discard_and_undo
;
4049 /* We see SYN without ACK. It is attempt of
4050 * simultaneous connect with crossed SYNs.
4051 * Particularly, it can be connect to self.
4053 tcp_set_state(sk
, TCP_SYN_RECV
);
4055 if (tp
->rx_opt
.saw_tstamp
) {
4056 tp
->rx_opt
.tstamp_ok
= 1;
4057 tcp_store_ts_recent(tp
);
4058 tp
->tcp_header_len
=
4059 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4061 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4064 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4065 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4067 /* RFC1323: The window in SYN & SYN/ACK segments is
4070 tp
->snd_wnd
= ntohs(th
->window
);
4071 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4072 tp
->max_window
= tp
->snd_wnd
;
4074 TCP_ECN_rcv_syn(tp
, th
);
4075 if (tp
->ecn_flags
&TCP_ECN_OK
)
4076 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
4078 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
4079 tcp_initialize_rcv_mss(sk
);
4082 tcp_send_synack(sk
);
4084 /* Note, we could accept data and URG from this segment.
4085 * There are no obstacles to make this.
4087 * However, if we ignore data in ACKless segments sometimes,
4088 * we have no reasons to accept it sometimes.
4089 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4090 * is not flawless. So, discard packet for sanity.
4091 * Uncomment this return to process the data.
4098 /* "fifth, if neither of the SYN or RST bits is set then
4099 * drop the segment and return."
4103 tcp_clear_options(&tp
->rx_opt
);
4104 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4108 tcp_clear_options(&tp
->rx_opt
);
4109 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4115 * This function implements the receiving procedure of RFC 793 for
4116 * all states except ESTABLISHED and TIME_WAIT.
4117 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4118 * address independent.
4121 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4122 struct tcphdr
*th
, unsigned len
)
4124 struct tcp_sock
*tp
= tcp_sk(sk
);
4127 tp
->rx_opt
.saw_tstamp
= 0;
4129 switch (sk
->sk_state
) {
4141 if(tp
->af_specific
->conn_request(sk
, skb
) < 0)
4144 /* Now we have several options: In theory there is
4145 * nothing else in the frame. KA9Q has an option to
4146 * send data with the syn, BSD accepts data with the
4147 * syn up to the [to be] advertised window and
4148 * Solaris 2.1 gives you a protocol error. For now
4149 * we just ignore it, that fits the spec precisely
4150 * and avoids incompatibilities. It would be nice in
4151 * future to drop through and process the data.
4153 * Now that TTCP is starting to be used we ought to
4155 * But, this leaves one open to an easy denial of
4156 * service attack, and SYN cookies can't defend
4157 * against this problem. So, we drop the data
4158 * in the interest of security over speed.
4165 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4169 /* Do step6 onward by hand. */
4170 tcp_urg(sk
, skb
, th
);
4172 tcp_data_snd_check(sk
, tp
);
4176 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4177 tcp_paws_discard(sk
, skb
)) {
4179 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4180 tcp_send_dupack(sk
, skb
);
4183 /* Reset is accepted even if it did not pass PAWS. */
4186 /* step 1: check sequence number */
4187 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4189 tcp_send_dupack(sk
, skb
);
4193 /* step 2: check RST bit */
4199 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4201 /* step 3: check security and precedence [ignored] */
4205 * Check for a SYN in window.
4207 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4208 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4213 /* step 5: check the ACK field */
4215 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4217 switch(sk
->sk_state
) {
4220 tp
->copied_seq
= tp
->rcv_nxt
;
4222 tcp_set_state(sk
, TCP_ESTABLISHED
);
4223 sk
->sk_state_change(sk
);
4225 /* Note, that this wakeup is only for marginal
4226 * crossed SYN case. Passively open sockets
4227 * are not waked up, because sk->sk_sleep ==
4228 * NULL and sk->sk_socket == NULL.
4230 if (sk
->sk_socket
) {
4231 sk_wake_async(sk
,0,POLL_OUT
);
4234 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4235 tp
->snd_wnd
= ntohs(th
->window
) <<
4236 tp
->rx_opt
.snd_wscale
;
4237 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4238 TCP_SKB_CB(skb
)->seq
);
4240 /* tcp_ack considers this ACK as duplicate
4241 * and does not calculate rtt.
4242 * Fix it at least with timestamps.
4244 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4246 tcp_ack_saw_tstamp(sk
, NULL
, 0);
4248 if (tp
->rx_opt
.tstamp_ok
)
4249 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4251 /* Make sure socket is routed, for
4254 tp
->af_specific
->rebuild_header(sk
);
4256 tcp_init_metrics(sk
);
4258 tcp_init_congestion_control(sk
);
4260 /* Prevent spurious tcp_cwnd_restart() on
4261 * first data packet.
4263 tp
->lsndtime
= tcp_time_stamp
;
4265 tcp_initialize_rcv_mss(sk
);
4266 tcp_init_buffer_space(sk
);
4267 tcp_fast_path_on(tp
);
4274 if (tp
->snd_una
== tp
->write_seq
) {
4275 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4276 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4277 dst_confirm(sk
->sk_dst_cache
);
4279 if (!sock_flag(sk
, SOCK_DEAD
))
4280 /* Wake up lingering close() */
4281 sk
->sk_state_change(sk
);
4285 if (tp
->linger2
< 0 ||
4286 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4287 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4289 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4293 tmo
= tcp_fin_time(sk
);
4294 if (tmo
> TCP_TIMEWAIT_LEN
) {
4295 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4296 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4297 /* Bad case. We could lose such FIN otherwise.
4298 * It is not a big problem, but it looks confusing
4299 * and not so rare event. We still can lose it now,
4300 * if it spins in bh_lock_sock(), but it is really
4303 inet_csk_reset_keepalive_timer(sk
, tmo
);
4305 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4313 if (tp
->snd_una
== tp
->write_seq
) {
4314 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4320 if (tp
->snd_una
== tp
->write_seq
) {
4321 tcp_update_metrics(sk
);
4330 /* step 6: check the URG bit */
4331 tcp_urg(sk
, skb
, th
);
4333 /* step 7: process the segment text */
4334 switch (sk
->sk_state
) {
4335 case TCP_CLOSE_WAIT
:
4338 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4342 /* RFC 793 says to queue data in these states,
4343 * RFC 1122 says we MUST send a reset.
4344 * BSD 4.4 also does reset.
4346 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4347 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4348 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4349 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4355 case TCP_ESTABLISHED
:
4356 tcp_data_queue(sk
, skb
);
4361 /* tcp_data could move socket to TIME-WAIT */
4362 if (sk
->sk_state
!= TCP_CLOSE
) {
4363 tcp_data_snd_check(sk
, tp
);
4364 tcp_ack_snd_check(sk
);
4374 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4375 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4376 EXPORT_SYMBOL(tcp_parse_options
);
4377 EXPORT_SYMBOL(tcp_rcv_established
);
4378 EXPORT_SYMBOL(tcp_rcv_state_process
);