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 tcp_sock
*tp
,
120 unsigned int len
, lss
;
122 lss
= tp
->ack
.last_seg_size
;
123 tp
->ack
.last_seg_size
= 0;
125 /* skb->len may jitter because of SACKs, even if peer
126 * sends good full-sized frames.
129 if (len
>= tp
->ack
.rcv_mss
) {
130 tp
->ack
.rcv_mss
= len
;
132 /* Otherwise, we make more careful check taking into account,
133 * that SACKs block is variable.
135 * "len" is invariant segment length, including TCP header.
137 len
+= skb
->data
- skb
->h
.raw
;
138 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
139 /* If PSH is not set, packet should be
140 * full sized, provided peer TCP is not badly broken.
141 * This observation (if it is correct 8)) allows
142 * to handle super-low mtu links fairly.
144 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
145 !(tcp_flag_word(skb
->h
.th
)&TCP_REMNANT
))) {
146 /* Subtract also invariant (if peer is RFC compliant),
147 * tcp header plus fixed timestamp option length.
148 * Resulting "len" is MSS free of SACK jitter.
150 len
-= tp
->tcp_header_len
;
151 tp
->ack
.last_seg_size
= len
;
153 tp
->ack
.rcv_mss
= len
;
157 tp
->ack
.pending
|= TCP_ACK_PUSHED
;
161 static void tcp_incr_quickack(struct tcp_sock
*tp
)
163 unsigned quickacks
= tp
->rcv_wnd
/(2*tp
->ack
.rcv_mss
);
167 if (quickacks
> tp
->ack
.quick
)
168 tp
->ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
171 void tcp_enter_quickack_mode(struct tcp_sock
*tp
)
173 tcp_incr_quickack(tp
);
174 tp
->ack
.pingpong
= 0;
175 tp
->ack
.ato
= TCP_ATO_MIN
;
178 /* Send ACKs quickly, if "quick" count is not exhausted
179 * and the session is not interactive.
182 static __inline__
int tcp_in_quickack_mode(struct tcp_sock
*tp
)
184 return (tp
->ack
.quick
&& !tp
->ack
.pingpong
);
187 /* Buffer size and advertised window tuning.
189 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
192 static void tcp_fixup_sndbuf(struct sock
*sk
)
194 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
195 sizeof(struct sk_buff
);
197 if (sk
->sk_sndbuf
< 3 * sndmem
)
198 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
201 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
203 * All tcp_full_space() is split to two parts: "network" buffer, allocated
204 * forward and advertised in receiver window (tp->rcv_wnd) and
205 * "application buffer", required to isolate scheduling/application
206 * latencies from network.
207 * window_clamp is maximal advertised window. It can be less than
208 * tcp_full_space(), in this case tcp_full_space() - window_clamp
209 * is reserved for "application" buffer. The less window_clamp is
210 * the smoother our behaviour from viewpoint of network, but the lower
211 * throughput and the higher sensitivity of the connection to losses. 8)
213 * rcv_ssthresh is more strict window_clamp used at "slow start"
214 * phase to predict further behaviour of this connection.
215 * It is used for two goals:
216 * - to enforce header prediction at sender, even when application
217 * requires some significant "application buffer". It is check #1.
218 * - to prevent pruning of receive queue because of misprediction
219 * of receiver window. Check #2.
221 * The scheme does not work when sender sends good segments opening
222 * window and then starts to feed us spagetti. But it should work
223 * in common situations. Otherwise, we have to rely on queue collapsing.
226 /* Slow part of check#2. */
227 static int __tcp_grow_window(struct sock
*sk
, struct tcp_sock
*tp
,
231 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
232 int window
= tcp_full_space(sk
)/2;
234 while (tp
->rcv_ssthresh
<= window
) {
235 if (truesize
<= skb
->len
)
236 return 2*tp
->ack
.rcv_mss
;
244 static inline void tcp_grow_window(struct sock
*sk
, struct tcp_sock
*tp
,
248 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
249 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
250 !tcp_memory_pressure
) {
253 /* Check #2. Increase window, if skb with such overhead
254 * will fit to rcvbuf in future.
256 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
259 incr
= __tcp_grow_window(sk
, tp
, skb
);
262 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
268 /* 3. Tuning rcvbuf, when connection enters established state. */
270 static void tcp_fixup_rcvbuf(struct sock
*sk
)
272 struct tcp_sock
*tp
= tcp_sk(sk
);
273 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
275 /* Try to select rcvbuf so that 4 mss-sized segments
276 * will fit to window and correspoding skbs will fit to our rcvbuf.
277 * (was 3; 4 is minimum to allow fast retransmit to work.)
279 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
281 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
282 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
285 /* 4. Try to fixup all. It is made iimediately after connection enters
288 static void tcp_init_buffer_space(struct sock
*sk
)
290 struct tcp_sock
*tp
= tcp_sk(sk
);
293 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
294 tcp_fixup_rcvbuf(sk
);
295 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
296 tcp_fixup_sndbuf(sk
);
298 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
300 maxwin
= tcp_full_space(sk
);
302 if (tp
->window_clamp
>= maxwin
) {
303 tp
->window_clamp
= maxwin
;
305 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
306 tp
->window_clamp
= max(maxwin
-
307 (maxwin
>> sysctl_tcp_app_win
),
311 /* Force reservation of one segment. */
312 if (sysctl_tcp_app_win
&&
313 tp
->window_clamp
> 2 * tp
->advmss
&&
314 tp
->window_clamp
+ tp
->advmss
> maxwin
)
315 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
317 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
318 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
321 /* 5. Recalculate window clamp after socket hit its memory bounds. */
322 static void tcp_clamp_window(struct sock
*sk
, struct tcp_sock
*tp
)
325 unsigned int app_win
= tp
->rcv_nxt
- tp
->copied_seq
;
330 skb_queue_walk(&tp
->out_of_order_queue
, skb
) {
334 /* If overcommit is due to out of order segments,
335 * do not clamp window. Try to expand rcvbuf instead.
338 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
339 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
340 !tcp_memory_pressure
&&
341 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0])
342 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
345 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
) {
347 if (atomic_read(&sk
->sk_rmem_alloc
) >= 2 * sk
->sk_rcvbuf
)
349 if (app_win
> tp
->ack
.rcv_mss
)
350 app_win
-= tp
->ack
.rcv_mss
;
351 app_win
= max(app_win
, 2U*tp
->advmss
);
354 tp
->window_clamp
= min(tp
->window_clamp
, app_win
);
355 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
359 /* Receiver "autotuning" code.
361 * The algorithm for RTT estimation w/o timestamps is based on
362 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
363 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
365 * More detail on this code can be found at
366 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
367 * though this reference is out of date. A new paper
370 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
372 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
378 if (new_sample
!= 0) {
379 /* If we sample in larger samples in the non-timestamp
380 * case, we could grossly overestimate the RTT especially
381 * with chatty applications or bulk transfer apps which
382 * are stalled on filesystem I/O.
384 * Also, since we are only going for a minimum in the
385 * non-timestamp case, we do not smoothe things out
386 * else with timestamps disabled convergance takes too
390 m
-= (new_sample
>> 3);
392 } else if (m
< new_sample
)
395 /* No previous mesaure. */
399 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
400 tp
->rcv_rtt_est
.rtt
= new_sample
;
403 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
405 if (tp
->rcv_rtt_est
.time
== 0)
407 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
409 tcp_rcv_rtt_update(tp
,
410 jiffies
- tp
->rcv_rtt_est
.time
,
414 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
415 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
418 static inline void tcp_rcv_rtt_measure_ts(struct tcp_sock
*tp
, struct sk_buff
*skb
)
420 if (tp
->rx_opt
.rcv_tsecr
&&
421 (TCP_SKB_CB(skb
)->end_seq
-
422 TCP_SKB_CB(skb
)->seq
>= tp
->ack
.rcv_mss
))
423 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
427 * This function should be called every time data is copied to user space.
428 * It calculates the appropriate TCP receive buffer space.
430 void tcp_rcv_space_adjust(struct sock
*sk
)
432 struct tcp_sock
*tp
= tcp_sk(sk
);
436 if (tp
->rcvq_space
.time
== 0)
439 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
440 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
441 tp
->rcv_rtt_est
.rtt
== 0)
444 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
446 space
= max(tp
->rcvq_space
.space
, space
);
448 if (tp
->rcvq_space
.space
!= space
) {
451 tp
->rcvq_space
.space
= space
;
453 if (sysctl_tcp_moderate_rcvbuf
) {
454 int new_clamp
= space
;
456 /* Receive space grows, normalize in order to
457 * take into account packet headers and sk_buff
458 * structure overhead.
463 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
464 16 + sizeof(struct sk_buff
));
465 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
468 space
= min(space
, sysctl_tcp_rmem
[2]);
469 if (space
> sk
->sk_rcvbuf
) {
470 sk
->sk_rcvbuf
= space
;
472 /* Make the window clamp follow along. */
473 tp
->window_clamp
= new_clamp
;
479 tp
->rcvq_space
.seq
= tp
->copied_seq
;
480 tp
->rcvq_space
.time
= tcp_time_stamp
;
483 /* There is something which you must keep in mind when you analyze the
484 * behavior of the tp->ato delayed ack timeout interval. When a
485 * connection starts up, we want to ack as quickly as possible. The
486 * problem is that "good" TCP's do slow start at the beginning of data
487 * transmission. The means that until we send the first few ACK's the
488 * sender will sit on his end and only queue most of his data, because
489 * he can only send snd_cwnd unacked packets at any given time. For
490 * each ACK we send, he increments snd_cwnd and transmits more of his
493 static void tcp_event_data_recv(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
497 tcp_schedule_ack(tp
);
499 tcp_measure_rcv_mss(tp
, skb
);
501 tcp_rcv_rtt_measure(tp
);
503 now
= tcp_time_stamp
;
506 /* The _first_ data packet received, initialize
507 * delayed ACK engine.
509 tcp_incr_quickack(tp
);
510 tp
->ack
.ato
= TCP_ATO_MIN
;
512 int m
= now
- tp
->ack
.lrcvtime
;
514 if (m
<= TCP_ATO_MIN
/2) {
515 /* The fastest case is the first. */
516 tp
->ack
.ato
= (tp
->ack
.ato
>>1) + TCP_ATO_MIN
/2;
517 } else if (m
< tp
->ack
.ato
) {
518 tp
->ack
.ato
= (tp
->ack
.ato
>>1) + m
;
519 if (tp
->ack
.ato
> tp
->rto
)
520 tp
->ack
.ato
= tp
->rto
;
521 } else if (m
> tp
->rto
) {
522 /* Too long gap. Apparently sender falled to
523 * restart window, so that we send ACKs quickly.
525 tcp_incr_quickack(tp
);
526 sk_stream_mem_reclaim(sk
);
529 tp
->ack
.lrcvtime
= now
;
531 TCP_ECN_check_ce(tp
, skb
);
534 tcp_grow_window(sk
, tp
, skb
);
537 /* Called to compute a smoothed rtt estimate. The data fed to this
538 * routine either comes from timestamps, or from segments that were
539 * known _not_ to have been retransmitted [see Karn/Partridge
540 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
541 * piece by Van Jacobson.
542 * NOTE: the next three routines used to be one big routine.
543 * To save cycles in the RFC 1323 implementation it was better to break
544 * it up into three procedures. -- erics
546 static void tcp_rtt_estimator(struct tcp_sock
*tp
, __u32 mrtt
, u32
*usrtt
)
548 long m
= mrtt
; /* RTT */
550 /* The following amusing code comes from Jacobson's
551 * article in SIGCOMM '88. Note that rtt and mdev
552 * are scaled versions of rtt and mean deviation.
553 * This is designed to be as fast as possible
554 * m stands for "measurement".
556 * On a 1990 paper the rto value is changed to:
557 * RTO = rtt + 4 * mdev
559 * Funny. This algorithm seems to be very broken.
560 * These formulae increase RTO, when it should be decreased, increase
561 * too slowly, when it should be incresed fastly, decrease too fastly
562 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
563 * does not matter how to _calculate_ it. Seems, it was trap
564 * that VJ failed to avoid. 8)
569 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
570 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
572 m
= -m
; /* m is now abs(error) */
573 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
574 /* This is similar to one of Eifel findings.
575 * Eifel blocks mdev updates when rtt decreases.
576 * This solution is a bit different: we use finer gain
577 * for mdev in this case (alpha*beta).
578 * Like Eifel it also prevents growth of rto,
579 * but also it limits too fast rto decreases,
580 * happening in pure Eifel.
585 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
587 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
588 if (tp
->mdev
> tp
->mdev_max
) {
589 tp
->mdev_max
= tp
->mdev
;
590 if (tp
->mdev_max
> tp
->rttvar
)
591 tp
->rttvar
= tp
->mdev_max
;
593 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
594 if (tp
->mdev_max
< tp
->rttvar
)
595 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
596 tp
->rtt_seq
= tp
->snd_nxt
;
597 tp
->mdev_max
= TCP_RTO_MIN
;
600 /* no previous measure. */
601 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
602 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
603 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
604 tp
->rtt_seq
= tp
->snd_nxt
;
607 if (tp
->ca_ops
->rtt_sample
)
608 tp
->ca_ops
->rtt_sample(tp
, *usrtt
);
611 /* Calculate rto without backoff. This is the second half of Van Jacobson's
612 * routine referred to above.
614 static inline void tcp_set_rto(struct tcp_sock
*tp
)
616 /* Old crap is replaced with new one. 8)
619 * 1. If rtt variance happened to be less 50msec, it is hallucination.
620 * It cannot be less due to utterly erratic ACK generation made
621 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
622 * to do with delayed acks, because at cwnd>2 true delack timeout
623 * is invisible. Actually, Linux-2.4 also generates erratic
624 * ACKs in some curcumstances.
626 tp
->rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
628 /* 2. Fixups made earlier cannot be right.
629 * If we do not estimate RTO correctly without them,
630 * all the algo is pure shit and should be replaced
631 * with correct one. It is exaclty, which we pretend to do.
635 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
636 * guarantees that rto is higher.
638 static inline void tcp_bound_rto(struct tcp_sock
*tp
)
640 if (tp
->rto
> TCP_RTO_MAX
)
641 tp
->rto
= TCP_RTO_MAX
;
644 /* Save metrics learned by this TCP session.
645 This function is called only, when TCP finishes successfully
646 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
648 void tcp_update_metrics(struct sock
*sk
)
650 struct tcp_sock
*tp
= tcp_sk(sk
);
651 struct dst_entry
*dst
= __sk_dst_get(sk
);
653 if (sysctl_tcp_nometrics_save
)
658 if (dst
&& (dst
->flags
&DST_HOST
)) {
661 if (tp
->backoff
|| !tp
->srtt
) {
662 /* This session failed to estimate rtt. Why?
663 * Probably, no packets returned in time.
666 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
667 dst
->metrics
[RTAX_RTT
-1] = 0;
671 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
673 /* If newly calculated rtt larger than stored one,
674 * store new one. Otherwise, use EWMA. Remember,
675 * rtt overestimation is always better than underestimation.
677 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
679 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
681 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
684 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
688 /* Scale deviation to rttvar fixed point */
693 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
694 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
696 dst
->metrics
[RTAX_RTTVAR
-1] -=
697 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
700 if (tp
->snd_ssthresh
>= 0xFFFF) {
701 /* Slow start still did not finish. */
702 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
703 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
704 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
705 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
706 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
707 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
708 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
709 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
710 tp
->ca_state
== TCP_CA_Open
) {
711 /* Cong. avoidance phase, cwnd is reliable. */
712 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
713 dst
->metrics
[RTAX_SSTHRESH
-1] =
714 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
715 if (!dst_metric_locked(dst
, RTAX_CWND
))
716 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
718 /* Else slow start did not finish, cwnd is non-sense,
719 ssthresh may be also invalid.
721 if (!dst_metric_locked(dst
, RTAX_CWND
))
722 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
723 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
724 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
725 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
726 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
729 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
730 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
731 tp
->reordering
!= sysctl_tcp_reordering
)
732 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
737 /* Numbers are taken from RFC2414. */
738 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
740 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
743 if (tp
->mss_cache_std
> 1460)
746 cwnd
= (tp
->mss_cache_std
> 1095) ? 3 : 4;
748 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
751 /* Initialize metrics on socket. */
753 static void tcp_init_metrics(struct sock
*sk
)
755 struct tcp_sock
*tp
= tcp_sk(sk
);
756 struct dst_entry
*dst
= __sk_dst_get(sk
);
763 if (dst_metric_locked(dst
, RTAX_CWND
))
764 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
765 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
766 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
767 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
768 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
770 if (dst_metric(dst
, RTAX_REORDERING
) &&
771 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
772 tp
->rx_opt
.sack_ok
&= ~2;
773 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
776 if (dst_metric(dst
, RTAX_RTT
) == 0)
779 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
782 /* Initial rtt is determined from SYN,SYN-ACK.
783 * The segment is small and rtt may appear much
784 * less than real one. Use per-dst memory
785 * to make it more realistic.
787 * A bit of theory. RTT is time passed after "normal" sized packet
788 * is sent until it is ACKed. In normal curcumstances sending small
789 * packets force peer to delay ACKs and calculation is correct too.
790 * The algorithm is adaptive and, provided we follow specs, it
791 * NEVER underestimate RTT. BUT! If peer tries to make some clever
792 * tricks sort of "quick acks" for time long enough to decrease RTT
793 * to low value, and then abruptly stops to do it and starts to delay
794 * ACKs, wait for troubles.
796 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
797 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
798 tp
->rtt_seq
= tp
->snd_nxt
;
800 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
801 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
802 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
806 if (tp
->rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
808 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
809 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
813 /* Play conservative. If timestamps are not
814 * supported, TCP will fail to recalculate correct
815 * rtt, if initial rto is too small. FORGET ALL AND RESET!
817 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
819 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
820 tp
->rto
= TCP_TIMEOUT_INIT
;
824 static void tcp_update_reordering(struct tcp_sock
*tp
, int metric
, int ts
)
826 if (metric
> tp
->reordering
) {
827 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
829 /* This exciting event is worth to be remembered. 8) */
831 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
833 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
835 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
837 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
838 #if FASTRETRANS_DEBUG > 1
839 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
840 tp
->rx_opt
.sack_ok
, tp
->ca_state
,
844 tp
->undo_marker
? tp
->undo_retrans
: 0);
846 /* Disable FACK yet. */
847 tp
->rx_opt
.sack_ok
&= ~2;
851 /* This procedure tags the retransmission queue when SACKs arrive.
853 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
854 * Packets in queue with these bits set are counted in variables
855 * sacked_out, retrans_out and lost_out, correspondingly.
857 * Valid combinations are:
858 * Tag InFlight Description
859 * 0 1 - orig segment is in flight.
860 * S 0 - nothing flies, orig reached receiver.
861 * L 0 - nothing flies, orig lost by net.
862 * R 2 - both orig and retransmit are in flight.
863 * L|R 1 - orig is lost, retransmit is in flight.
864 * S|R 1 - orig reached receiver, retrans is still in flight.
865 * (L|S|R is logically valid, it could occur when L|R is sacked,
866 * but it is equivalent to plain S and code short-curcuits it to S.
867 * L|S is logically invalid, it would mean -1 packet in flight 8))
869 * These 6 states form finite state machine, controlled by the following events:
870 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
871 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
872 * 3. Loss detection event of one of three flavors:
873 * A. Scoreboard estimator decided the packet is lost.
874 * A'. Reno "three dupacks" marks head of queue lost.
875 * A''. Its FACK modfication, head until snd.fack is lost.
876 * B. SACK arrives sacking data transmitted after never retransmitted
878 * C. SACK arrives sacking SND.NXT at the moment, when the
879 * segment was retransmitted.
880 * 4. D-SACK added new rule: D-SACK changes any tag to S.
882 * It is pleasant to note, that state diagram turns out to be commutative,
883 * so that we are allowed not to be bothered by order of our actions,
884 * when multiple events arrive simultaneously. (see the function below).
886 * Reordering detection.
887 * --------------------
888 * Reordering metric is maximal distance, which a packet can be displaced
889 * in packet stream. With SACKs we can estimate it:
891 * 1. SACK fills old hole and the corresponding segment was not
892 * ever retransmitted -> reordering. Alas, we cannot use it
893 * when segment was retransmitted.
894 * 2. The last flaw is solved with D-SACK. D-SACK arrives
895 * for retransmitted and already SACKed segment -> reordering..
896 * Both of these heuristics are not used in Loss state, when we cannot
897 * account for retransmits accurately.
900 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
902 struct tcp_sock
*tp
= tcp_sk(sk
);
903 unsigned char *ptr
= ack_skb
->h
.raw
+ TCP_SKB_CB(ack_skb
)->sacked
;
904 struct tcp_sack_block
*sp
= (struct tcp_sack_block
*)(ptr
+2);
905 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
906 int reord
= tp
->packets_out
;
908 u32 lost_retrans
= 0;
912 /* So, SACKs for already sent large segments will be lost.
913 * Not good, but alternative is to resegment the queue. */
914 if (sk
->sk_route_caps
& NETIF_F_TSO
) {
915 sk
->sk_route_caps
&= ~NETIF_F_TSO
;
916 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
917 tp
->mss_cache
= tp
->mss_cache_std
;
922 prior_fackets
= tp
->fackets_out
;
924 for (i
=0; i
<num_sacks
; i
++, sp
++) {
926 __u32 start_seq
= ntohl(sp
->start_seq
);
927 __u32 end_seq
= ntohl(sp
->end_seq
);
931 /* Check for D-SACK. */
933 u32 ack
= TCP_SKB_CB(ack_skb
)->ack_seq
;
935 if (before(start_seq
, ack
)) {
937 tp
->rx_opt
.sack_ok
|= 4;
938 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
939 } else if (num_sacks
> 1 &&
940 !after(end_seq
, ntohl(sp
[1].end_seq
)) &&
941 !before(start_seq
, ntohl(sp
[1].start_seq
))) {
943 tp
->rx_opt
.sack_ok
|= 4;
944 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
947 /* D-SACK for already forgotten data...
948 * Do dumb counting. */
950 !after(end_seq
, prior_snd_una
) &&
951 after(end_seq
, tp
->undo_marker
))
954 /* Eliminate too old ACKs, but take into
955 * account more or less fresh ones, they can
956 * contain valid SACK info.
958 if (before(ack
, prior_snd_una
- tp
->max_window
))
962 /* Event "B" in the comment above. */
963 if (after(end_seq
, tp
->high_seq
))
964 flag
|= FLAG_DATA_LOST
;
966 sk_stream_for_retrans_queue(skb
, sk
) {
967 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
970 /* The retransmission queue is always in order, so
971 * we can short-circuit the walk early.
973 if(!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
976 fack_count
+= tcp_skb_pcount(skb
);
978 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
979 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
981 /* Account D-SACK for retransmitted packet. */
982 if ((dup_sack
&& in_sack
) &&
983 (sacked
& TCPCB_RETRANS
) &&
984 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
987 /* The frame is ACKed. */
988 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
989 if (sacked
&TCPCB_RETRANS
) {
990 if ((dup_sack
&& in_sack
) &&
991 (sacked
&TCPCB_SACKED_ACKED
))
992 reord
= min(fack_count
, reord
);
994 /* If it was in a hole, we detected reordering. */
995 if (fack_count
< prior_fackets
&&
996 !(sacked
&TCPCB_SACKED_ACKED
))
997 reord
= min(fack_count
, reord
);
1000 /* Nothing to do; acked frame is about to be dropped. */
1004 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1005 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1006 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1007 lost_retrans
= end_seq
;
1012 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1013 if (sacked
& TCPCB_SACKED_RETRANS
) {
1014 /* If the segment is not tagged as lost,
1015 * we do not clear RETRANS, believing
1016 * that retransmission is still in flight.
1018 if (sacked
& TCPCB_LOST
) {
1019 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1020 tp
->lost_out
-= tcp_skb_pcount(skb
);
1021 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1024 /* New sack for not retransmitted frame,
1025 * which was in hole. It is reordering.
1027 if (!(sacked
& TCPCB_RETRANS
) &&
1028 fack_count
< prior_fackets
)
1029 reord
= min(fack_count
, reord
);
1031 if (sacked
& TCPCB_LOST
) {
1032 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1033 tp
->lost_out
-= tcp_skb_pcount(skb
);
1037 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1038 flag
|= FLAG_DATA_SACKED
;
1039 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1041 if (fack_count
> tp
->fackets_out
)
1042 tp
->fackets_out
= fack_count
;
1044 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1045 reord
= min(fack_count
, reord
);
1048 /* D-SACK. We can detect redundant retransmission
1049 * in S|R and plain R frames and clear it.
1050 * undo_retrans is decreased above, L|R frames
1051 * are accounted above as well.
1054 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1055 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1056 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1061 /* Check for lost retransmit. This superb idea is
1062 * borrowed from "ratehalving". Event "C".
1063 * Later note: FACK people cheated me again 8),
1064 * we have to account for reordering! Ugly,
1067 if (lost_retrans
&& tp
->ca_state
== TCP_CA_Recovery
) {
1068 struct sk_buff
*skb
;
1070 sk_stream_for_retrans_queue(skb
, sk
) {
1071 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1073 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1075 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1076 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1078 !before(lost_retrans
,
1079 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1080 tp
->mss_cache_std
))) {
1081 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1082 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1084 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1085 tp
->lost_out
+= tcp_skb_pcount(skb
);
1086 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1087 flag
|= FLAG_DATA_SACKED
;
1088 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1094 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1096 if ((reord
< tp
->fackets_out
) && tp
->ca_state
!= TCP_CA_Loss
)
1097 tcp_update_reordering(tp
, ((tp
->fackets_out
+ 1) - reord
), 0);
1099 #if FASTRETRANS_DEBUG > 0
1100 BUG_TRAP((int)tp
->sacked_out
>= 0);
1101 BUG_TRAP((int)tp
->lost_out
>= 0);
1102 BUG_TRAP((int)tp
->retrans_out
>= 0);
1103 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1108 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1109 * segments to see from the next ACKs whether any data was really missing.
1110 * If the RTO was spurious, new ACKs should arrive.
1112 void tcp_enter_frto(struct sock
*sk
)
1114 struct tcp_sock
*tp
= tcp_sk(sk
);
1115 struct sk_buff
*skb
;
1117 tp
->frto_counter
= 1;
1119 if (tp
->ca_state
<= TCP_CA_Disorder
||
1120 tp
->snd_una
== tp
->high_seq
||
1121 (tp
->ca_state
== TCP_CA_Loss
&& !tp
->retransmits
)) {
1122 tp
->prior_ssthresh
= tcp_current_ssthresh(tp
);
1123 tp
->snd_ssthresh
= tp
->ca_ops
->ssthresh(tp
);
1124 tcp_ca_event(tp
, CA_EVENT_FRTO
);
1127 /* Have to clear retransmission markers here to keep the bookkeeping
1128 * in shape, even though we are not yet in Loss state.
1129 * If something was really lost, it is eventually caught up
1130 * in tcp_enter_frto_loss.
1132 tp
->retrans_out
= 0;
1133 tp
->undo_marker
= tp
->snd_una
;
1134 tp
->undo_retrans
= 0;
1136 sk_stream_for_retrans_queue(skb
, sk
) {
1137 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_RETRANS
;
1139 tcp_sync_left_out(tp
);
1141 tcp_set_ca_state(tp
, TCP_CA_Open
);
1142 tp
->frto_highmark
= tp
->snd_nxt
;
1145 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1146 * which indicates that we should follow the traditional RTO recovery,
1147 * i.e. mark everything lost and do go-back-N retransmission.
1149 static void tcp_enter_frto_loss(struct sock
*sk
)
1151 struct tcp_sock
*tp
= tcp_sk(sk
);
1152 struct sk_buff
*skb
;
1157 tp
->fackets_out
= 0;
1159 sk_stream_for_retrans_queue(skb
, sk
) {
1160 cnt
+= tcp_skb_pcount(skb
);
1161 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1162 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1164 /* Do not mark those segments lost that were
1165 * forward transmitted after RTO
1167 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1168 tp
->frto_highmark
)) {
1169 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1170 tp
->lost_out
+= tcp_skb_pcount(skb
);
1173 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1174 tp
->fackets_out
= cnt
;
1177 tcp_sync_left_out(tp
);
1179 tp
->snd_cwnd
= tp
->frto_counter
+ tcp_packets_in_flight(tp
)+1;
1180 tp
->snd_cwnd_cnt
= 0;
1181 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1182 tp
->undo_marker
= 0;
1183 tp
->frto_counter
= 0;
1185 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1186 sysctl_tcp_reordering
);
1187 tcp_set_ca_state(tp
, TCP_CA_Loss
);
1188 tp
->high_seq
= tp
->frto_highmark
;
1189 TCP_ECN_queue_cwr(tp
);
1192 void tcp_clear_retrans(struct tcp_sock
*tp
)
1195 tp
->retrans_out
= 0;
1197 tp
->fackets_out
= 0;
1201 tp
->undo_marker
= 0;
1202 tp
->undo_retrans
= 0;
1205 /* Enter Loss state. If "how" is not zero, forget all SACK information
1206 * and reset tags completely, otherwise preserve SACKs. If receiver
1207 * dropped its ofo queue, we will know this due to reneging detection.
1209 void tcp_enter_loss(struct sock
*sk
, int how
)
1211 struct tcp_sock
*tp
= tcp_sk(sk
);
1212 struct sk_buff
*skb
;
1215 /* Reduce ssthresh if it has not yet been made inside this window. */
1216 if (tp
->ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1217 (tp
->ca_state
== TCP_CA_Loss
&& !tp
->retransmits
)) {
1218 tp
->prior_ssthresh
= tcp_current_ssthresh(tp
);
1219 tp
->snd_ssthresh
= tp
->ca_ops
->ssthresh(tp
);
1220 tcp_ca_event(tp
, CA_EVENT_LOSS
);
1223 tp
->snd_cwnd_cnt
= 0;
1224 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1226 tcp_clear_retrans(tp
);
1228 /* Push undo marker, if it was plain RTO and nothing
1229 * was retransmitted. */
1231 tp
->undo_marker
= tp
->snd_una
;
1233 sk_stream_for_retrans_queue(skb
, sk
) {
1234 cnt
+= tcp_skb_pcount(skb
);
1235 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1236 tp
->undo_marker
= 0;
1237 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1238 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1239 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1240 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1241 tp
->lost_out
+= tcp_skb_pcount(skb
);
1243 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1244 tp
->fackets_out
= cnt
;
1247 tcp_sync_left_out(tp
);
1249 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1250 sysctl_tcp_reordering
);
1251 tcp_set_ca_state(tp
, TCP_CA_Loss
);
1252 tp
->high_seq
= tp
->snd_nxt
;
1253 TCP_ECN_queue_cwr(tp
);
1256 static int tcp_check_sack_reneging(struct sock
*sk
, struct tcp_sock
*tp
)
1258 struct sk_buff
*skb
;
1260 /* If ACK arrived pointing to a remembered SACK,
1261 * it means that our remembered SACKs do not reflect
1262 * real state of receiver i.e.
1263 * receiver _host_ is heavily congested (or buggy).
1264 * Do processing similar to RTO timeout.
1266 if ((skb
= skb_peek(&sk
->sk_write_queue
)) != NULL
&&
1267 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1268 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1270 tcp_enter_loss(sk
, 1);
1272 tcp_retransmit_skb(sk
, skb_peek(&sk
->sk_write_queue
));
1273 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1279 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1281 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1284 static inline int tcp_skb_timedout(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1286 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> tp
->rto
);
1289 static inline int tcp_head_timedout(struct sock
*sk
, struct tcp_sock
*tp
)
1291 return tp
->packets_out
&&
1292 tcp_skb_timedout(tp
, skb_peek(&sk
->sk_write_queue
));
1295 /* Linux NewReno/SACK/FACK/ECN state machine.
1296 * --------------------------------------
1298 * "Open" Normal state, no dubious events, fast path.
1299 * "Disorder" In all the respects it is "Open",
1300 * but requires a bit more attention. It is entered when
1301 * we see some SACKs or dupacks. It is split of "Open"
1302 * mainly to move some processing from fast path to slow one.
1303 * "CWR" CWND was reduced due to some Congestion Notification event.
1304 * It can be ECN, ICMP source quench, local device congestion.
1305 * "Recovery" CWND was reduced, we are fast-retransmitting.
1306 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1308 * tcp_fastretrans_alert() is entered:
1309 * - each incoming ACK, if state is not "Open"
1310 * - when arrived ACK is unusual, namely:
1315 * Counting packets in flight is pretty simple.
1317 * in_flight = packets_out - left_out + retrans_out
1319 * packets_out is SND.NXT-SND.UNA counted in packets.
1321 * retrans_out is number of retransmitted segments.
1323 * left_out is number of segments left network, but not ACKed yet.
1325 * left_out = sacked_out + lost_out
1327 * sacked_out: Packets, which arrived to receiver out of order
1328 * and hence not ACKed. With SACKs this number is simply
1329 * amount of SACKed data. Even without SACKs
1330 * it is easy to give pretty reliable estimate of this number,
1331 * counting duplicate ACKs.
1333 * lost_out: Packets lost by network. TCP has no explicit
1334 * "loss notification" feedback from network (for now).
1335 * It means that this number can be only _guessed_.
1336 * Actually, it is the heuristics to predict lossage that
1337 * distinguishes different algorithms.
1339 * F.e. after RTO, when all the queue is considered as lost,
1340 * lost_out = packets_out and in_flight = retrans_out.
1342 * Essentially, we have now two algorithms counting
1345 * FACK: It is the simplest heuristics. As soon as we decided
1346 * that something is lost, we decide that _all_ not SACKed
1347 * packets until the most forward SACK are lost. I.e.
1348 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1349 * It is absolutely correct estimate, if network does not reorder
1350 * packets. And it loses any connection to reality when reordering
1351 * takes place. We use FACK by default until reordering
1352 * is suspected on the path to this destination.
1354 * NewReno: when Recovery is entered, we assume that one segment
1355 * is lost (classic Reno). While we are in Recovery and
1356 * a partial ACK arrives, we assume that one more packet
1357 * is lost (NewReno). This heuristics are the same in NewReno
1360 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1361 * deflation etc. CWND is real congestion window, never inflated, changes
1362 * only according to classic VJ rules.
1364 * Really tricky (and requiring careful tuning) part of algorithm
1365 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1366 * The first determines the moment _when_ we should reduce CWND and,
1367 * hence, slow down forward transmission. In fact, it determines the moment
1368 * when we decide that hole is caused by loss, rather than by a reorder.
1370 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1371 * holes, caused by lost packets.
1373 * And the most logically complicated part of algorithm is undo
1374 * heuristics. We detect false retransmits due to both too early
1375 * fast retransmit (reordering) and underestimated RTO, analyzing
1376 * timestamps and D-SACKs. When we detect that some segments were
1377 * retransmitted by mistake and CWND reduction was wrong, we undo
1378 * window reduction and abort recovery phase. This logic is hidden
1379 * inside several functions named tcp_try_undo_<something>.
1382 /* This function decides, when we should leave Disordered state
1383 * and enter Recovery phase, reducing congestion window.
1385 * Main question: may we further continue forward transmission
1386 * with the same cwnd?
1388 static int tcp_time_to_recover(struct sock
*sk
, struct tcp_sock
*tp
)
1392 /* Trick#1: The loss is proven. */
1396 /* Not-A-Trick#2 : Classic rule... */
1397 if (tcp_fackets_out(tp
) > tp
->reordering
)
1400 /* Trick#3 : when we use RFC2988 timer restart, fast
1401 * retransmit can be triggered by timeout of queue head.
1403 if (tcp_head_timedout(sk
, tp
))
1406 /* Trick#4: It is still not OK... But will it be useful to delay
1409 packets_out
= tp
->packets_out
;
1410 if (packets_out
<= tp
->reordering
&&
1411 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1412 !tcp_may_send_now(sk
, tp
)) {
1413 /* We have nothing to send. This connection is limited
1414 * either by receiver window or by application.
1422 /* If we receive more dupacks than we expected counting segments
1423 * in assumption of absent reordering, interpret this as reordering.
1424 * The only another reason could be bug in receiver TCP.
1426 static void tcp_check_reno_reordering(struct tcp_sock
*tp
, int addend
)
1430 holes
= max(tp
->lost_out
, 1U);
1431 holes
= min(holes
, tp
->packets_out
);
1433 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1434 tp
->sacked_out
= tp
->packets_out
- holes
;
1435 tcp_update_reordering(tp
, tp
->packets_out
+addend
, 0);
1439 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1441 static void tcp_add_reno_sack(struct tcp_sock
*tp
)
1444 tcp_check_reno_reordering(tp
, 0);
1445 tcp_sync_left_out(tp
);
1448 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1450 static void tcp_remove_reno_sacks(struct sock
*sk
, struct tcp_sock
*tp
, int acked
)
1453 /* One ACK acked hole. The rest eat duplicate ACKs. */
1454 if (acked
-1 >= tp
->sacked_out
)
1457 tp
->sacked_out
-= acked
-1;
1459 tcp_check_reno_reordering(tp
, acked
);
1460 tcp_sync_left_out(tp
);
1463 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1466 tp
->left_out
= tp
->lost_out
;
1469 /* Mark head of queue up as lost. */
1470 static void tcp_mark_head_lost(struct sock
*sk
, struct tcp_sock
*tp
,
1471 int packets
, u32 high_seq
)
1473 struct sk_buff
*skb
;
1476 BUG_TRAP(cnt
<= tp
->packets_out
);
1478 sk_stream_for_retrans_queue(skb
, sk
) {
1479 cnt
-= tcp_skb_pcount(skb
);
1480 if (cnt
< 0 || after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1482 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1483 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1484 tp
->lost_out
+= tcp_skb_pcount(skb
);
1487 tcp_sync_left_out(tp
);
1490 /* Account newly detected lost packet(s) */
1492 static void tcp_update_scoreboard(struct sock
*sk
, struct tcp_sock
*tp
)
1495 int lost
= tp
->fackets_out
- tp
->reordering
;
1498 tcp_mark_head_lost(sk
, tp
, lost
, tp
->high_seq
);
1500 tcp_mark_head_lost(sk
, tp
, 1, tp
->high_seq
);
1503 /* New heuristics: it is possible only after we switched
1504 * to restart timer each time when something is ACKed.
1505 * Hence, we can detect timed out packets during fast
1506 * retransmit without falling to slow start.
1508 if (tcp_head_timedout(sk
, tp
)) {
1509 struct sk_buff
*skb
;
1511 sk_stream_for_retrans_queue(skb
, sk
) {
1512 if (tcp_skb_timedout(tp
, skb
) &&
1513 !(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1514 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1515 tp
->lost_out
+= tcp_skb_pcount(skb
);
1518 tcp_sync_left_out(tp
);
1522 /* CWND moderation, preventing bursts due to too big ACKs
1523 * in dubious situations.
1525 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1527 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1528 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1529 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1532 /* Decrease cwnd each second ack. */
1533 static void tcp_cwnd_down(struct tcp_sock
*tp
)
1535 int decr
= tp
->snd_cwnd_cnt
+ 1;
1537 tp
->snd_cwnd_cnt
= decr
&1;
1540 if (decr
&& tp
->snd_cwnd
> tp
->ca_ops
->min_cwnd(tp
))
1541 tp
->snd_cwnd
-= decr
;
1543 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1544 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1547 /* Nothing was retransmitted or returned timestamp is less
1548 * than timestamp of the first retransmission.
1550 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1552 return !tp
->retrans_stamp
||
1553 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1554 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1557 /* Undo procedures. */
1559 #if FASTRETRANS_DEBUG > 1
1560 static void DBGUNDO(struct sock
*sk
, struct tcp_sock
*tp
, const char *msg
)
1562 struct inet_sock
*inet
= inet_sk(sk
);
1563 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1565 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1566 tp
->snd_cwnd
, tp
->left_out
,
1567 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1571 #define DBGUNDO(x...) do { } while (0)
1574 static void tcp_undo_cwr(struct tcp_sock
*tp
, int undo
)
1576 if (tp
->prior_ssthresh
) {
1577 if (tp
->ca_ops
->undo_cwnd
)
1578 tp
->snd_cwnd
= tp
->ca_ops
->undo_cwnd(tp
);
1580 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1582 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1583 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1584 TCP_ECN_withdraw_cwr(tp
);
1587 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1589 tcp_moderate_cwnd(tp
);
1590 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1593 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1595 return tp
->undo_marker
&&
1596 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1599 /* People celebrate: "We love our President!" */
1600 static int tcp_try_undo_recovery(struct sock
*sk
, struct tcp_sock
*tp
)
1602 if (tcp_may_undo(tp
)) {
1603 /* Happy end! We did not retransmit anything
1604 * or our original transmission succeeded.
1606 DBGUNDO(sk
, tp
, tp
->ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1607 tcp_undo_cwr(tp
, 1);
1608 if (tp
->ca_state
== TCP_CA_Loss
)
1609 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1611 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1612 tp
->undo_marker
= 0;
1614 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1615 /* Hold old state until something *above* high_seq
1616 * is ACKed. For Reno it is MUST to prevent false
1617 * fast retransmits (RFC2582). SACK TCP is safe. */
1618 tcp_moderate_cwnd(tp
);
1621 tcp_set_ca_state(tp
, TCP_CA_Open
);
1625 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1626 static void tcp_try_undo_dsack(struct sock
*sk
, struct tcp_sock
*tp
)
1628 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1629 DBGUNDO(sk
, tp
, "D-SACK");
1630 tcp_undo_cwr(tp
, 1);
1631 tp
->undo_marker
= 0;
1632 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1636 /* Undo during fast recovery after partial ACK. */
1638 static int tcp_try_undo_partial(struct sock
*sk
, struct tcp_sock
*tp
,
1641 /* Partial ACK arrived. Force Hoe's retransmit. */
1642 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1644 if (tcp_may_undo(tp
)) {
1645 /* Plain luck! Hole if filled with delayed
1646 * packet, rather than with a retransmit.
1648 if (tp
->retrans_out
== 0)
1649 tp
->retrans_stamp
= 0;
1651 tcp_update_reordering(tp
, tcp_fackets_out(tp
)+acked
, 1);
1653 DBGUNDO(sk
, tp
, "Hoe");
1654 tcp_undo_cwr(tp
, 0);
1655 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1657 /* So... Do not make Hoe's retransmit yet.
1658 * If the first packet was delayed, the rest
1659 * ones are most probably delayed as well.
1666 /* Undo during loss recovery after partial ACK. */
1667 static int tcp_try_undo_loss(struct sock
*sk
, struct tcp_sock
*tp
)
1669 if (tcp_may_undo(tp
)) {
1670 struct sk_buff
*skb
;
1671 sk_stream_for_retrans_queue(skb
, sk
) {
1672 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1674 DBGUNDO(sk
, tp
, "partial loss");
1676 tp
->left_out
= tp
->sacked_out
;
1677 tcp_undo_cwr(tp
, 1);
1678 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1679 tp
->retransmits
= 0;
1680 tp
->undo_marker
= 0;
1682 tcp_set_ca_state(tp
, TCP_CA_Open
);
1688 static inline void tcp_complete_cwr(struct tcp_sock
*tp
)
1690 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1691 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1692 tcp_ca_event(tp
, CA_EVENT_COMPLETE_CWR
);
1695 static void tcp_try_to_open(struct sock
*sk
, struct tcp_sock
*tp
, int flag
)
1697 tp
->left_out
= tp
->sacked_out
;
1699 if (tp
->retrans_out
== 0)
1700 tp
->retrans_stamp
= 0;
1705 if (tp
->ca_state
!= TCP_CA_CWR
) {
1706 int state
= TCP_CA_Open
;
1708 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
1709 state
= TCP_CA_Disorder
;
1711 if (tp
->ca_state
!= state
) {
1712 tcp_set_ca_state(tp
, state
);
1713 tp
->high_seq
= tp
->snd_nxt
;
1715 tcp_moderate_cwnd(tp
);
1721 /* Process an event, which can update packets-in-flight not trivially.
1722 * Main goal of this function is to calculate new estimate for left_out,
1723 * taking into account both packets sitting in receiver's buffer and
1724 * packets lost by network.
1726 * Besides that it does CWND reduction, when packet loss is detected
1727 * and changes state of machine.
1729 * It does _not_ decide what to send, it is made in function
1730 * tcp_xmit_retransmit_queue().
1733 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
1734 int prior_packets
, int flag
)
1736 struct tcp_sock
*tp
= tcp_sk(sk
);
1737 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
1739 /* Some technical things:
1740 * 1. Reno does not count dupacks (sacked_out) automatically. */
1741 if (!tp
->packets_out
)
1743 /* 2. SACK counts snd_fack in packets inaccurately. */
1744 if (tp
->sacked_out
== 0)
1745 tp
->fackets_out
= 0;
1747 /* Now state machine starts.
1748 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1750 tp
->prior_ssthresh
= 0;
1752 /* B. In all the states check for reneging SACKs. */
1753 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
, tp
))
1756 /* C. Process data loss notification, provided it is valid. */
1757 if ((flag
&FLAG_DATA_LOST
) &&
1758 before(tp
->snd_una
, tp
->high_seq
) &&
1759 tp
->ca_state
!= TCP_CA_Open
&&
1760 tp
->fackets_out
> tp
->reordering
) {
1761 tcp_mark_head_lost(sk
, tp
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
1762 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
1765 /* D. Synchronize left_out to current state. */
1766 tcp_sync_left_out(tp
);
1768 /* E. Check state exit conditions. State can be terminated
1769 * when high_seq is ACKed. */
1770 if (tp
->ca_state
== TCP_CA_Open
) {
1771 if (!sysctl_tcp_frto
)
1772 BUG_TRAP(tp
->retrans_out
== 0);
1773 tp
->retrans_stamp
= 0;
1774 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
1775 switch (tp
->ca_state
) {
1777 tp
->retransmits
= 0;
1778 if (tcp_try_undo_recovery(sk
, tp
))
1783 /* CWR is to be held something *above* high_seq
1784 * is ACKed for CWR bit to reach receiver. */
1785 if (tp
->snd_una
!= tp
->high_seq
) {
1786 tcp_complete_cwr(tp
);
1787 tcp_set_ca_state(tp
, TCP_CA_Open
);
1791 case TCP_CA_Disorder
:
1792 tcp_try_undo_dsack(sk
, tp
);
1793 if (!tp
->undo_marker
||
1794 /* For SACK case do not Open to allow to undo
1795 * catching for all duplicate ACKs. */
1796 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
1797 tp
->undo_marker
= 0;
1798 tcp_set_ca_state(tp
, TCP_CA_Open
);
1802 case TCP_CA_Recovery
:
1804 tcp_reset_reno_sack(tp
);
1805 if (tcp_try_undo_recovery(sk
, tp
))
1807 tcp_complete_cwr(tp
);
1812 /* F. Process state. */
1813 switch (tp
->ca_state
) {
1814 case TCP_CA_Recovery
:
1815 if (prior_snd_una
== tp
->snd_una
) {
1816 if (IsReno(tp
) && is_dupack
)
1817 tcp_add_reno_sack(tp
);
1819 int acked
= prior_packets
- tp
->packets_out
;
1821 tcp_remove_reno_sacks(sk
, tp
, acked
);
1822 is_dupack
= tcp_try_undo_partial(sk
, tp
, acked
);
1826 if (flag
&FLAG_DATA_ACKED
)
1827 tp
->retransmits
= 0;
1828 if (!tcp_try_undo_loss(sk
, tp
)) {
1829 tcp_moderate_cwnd(tp
);
1830 tcp_xmit_retransmit_queue(sk
);
1833 if (tp
->ca_state
!= TCP_CA_Open
)
1835 /* Loss is undone; fall through to processing in Open state. */
1838 if (tp
->snd_una
!= prior_snd_una
)
1839 tcp_reset_reno_sack(tp
);
1841 tcp_add_reno_sack(tp
);
1844 if (tp
->ca_state
== TCP_CA_Disorder
)
1845 tcp_try_undo_dsack(sk
, tp
);
1847 if (!tcp_time_to_recover(sk
, tp
)) {
1848 tcp_try_to_open(sk
, tp
, flag
);
1852 /* Otherwise enter Recovery state */
1855 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
1857 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
1859 tp
->high_seq
= tp
->snd_nxt
;
1860 tp
->prior_ssthresh
= 0;
1861 tp
->undo_marker
= tp
->snd_una
;
1862 tp
->undo_retrans
= tp
->retrans_out
;
1864 if (tp
->ca_state
< TCP_CA_CWR
) {
1865 if (!(flag
&FLAG_ECE
))
1866 tp
->prior_ssthresh
= tcp_current_ssthresh(tp
);
1867 tp
->snd_ssthresh
= tp
->ca_ops
->ssthresh(tp
);
1868 TCP_ECN_queue_cwr(tp
);
1871 tp
->snd_cwnd_cnt
= 0;
1872 tcp_set_ca_state(tp
, TCP_CA_Recovery
);
1875 if (is_dupack
|| tcp_head_timedout(sk
, tp
))
1876 tcp_update_scoreboard(sk
, tp
);
1878 tcp_xmit_retransmit_queue(sk
);
1881 /* Read draft-ietf-tcplw-high-performance before mucking
1882 * with this code. (Superceeds RFC1323)
1884 static void tcp_ack_saw_tstamp(struct tcp_sock
*tp
, u32
*usrtt
, int flag
)
1888 /* RTTM Rule: A TSecr value received in a segment is used to
1889 * update the averaged RTT measurement only if the segment
1890 * acknowledges some new data, i.e., only if it advances the
1891 * left edge of the send window.
1893 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1894 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1896 * Changed: reset backoff as soon as we see the first valid sample.
1897 * If we do not, we get strongly overstimated rto. With timestamps
1898 * samples are accepted even from very old segments: f.e., when rtt=1
1899 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1900 * answer arrives rto becomes 120 seconds! If at least one of segments
1901 * in window is lost... Voila. --ANK (010210)
1903 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
1904 tcp_rtt_estimator(tp
, seq_rtt
, usrtt
);
1910 static void tcp_ack_no_tstamp(struct tcp_sock
*tp
, u32 seq_rtt
, u32
*usrtt
, int flag
)
1912 /* We don't have a timestamp. Can only use
1913 * packets that are not retransmitted to determine
1914 * rtt estimates. Also, we must not reset the
1915 * backoff for rto until we get a non-retransmitted
1916 * packet. This allows us to deal with a situation
1917 * where the network delay has increased suddenly.
1918 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1921 if (flag
& FLAG_RETRANS_DATA_ACKED
)
1924 tcp_rtt_estimator(tp
, seq_rtt
, usrtt
);
1930 static inline void tcp_ack_update_rtt(struct tcp_sock
*tp
,
1931 int flag
, s32 seq_rtt
, u32
*usrtt
)
1933 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1934 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
1935 tcp_ack_saw_tstamp(tp
, usrtt
, flag
);
1936 else if (seq_rtt
>= 0)
1937 tcp_ack_no_tstamp(tp
, seq_rtt
, usrtt
, flag
);
1940 static inline void tcp_cong_avoid(struct tcp_sock
*tp
, u32 ack
, u32 rtt
,
1941 u32 in_flight
, int good
)
1943 tp
->ca_ops
->cong_avoid(tp
, ack
, rtt
, in_flight
, good
);
1944 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1947 /* Restart timer after forward progress on connection.
1948 * RFC2988 recommends to restart timer to now+rto.
1951 static inline void tcp_ack_packets_out(struct sock
*sk
, struct tcp_sock
*tp
)
1953 if (!tp
->packets_out
) {
1954 tcp_clear_xmit_timer(sk
, TCP_TIME_RETRANS
);
1956 tcp_reset_xmit_timer(sk
, TCP_TIME_RETRANS
, tp
->rto
);
1960 /* There is one downside to this scheme. Although we keep the
1961 * ACK clock ticking, adjusting packet counters and advancing
1962 * congestion window, we do not liberate socket send buffer
1965 * Mucking with skb->truesize and sk->sk_wmem_alloc et al.
1966 * then making a write space wakeup callback is a possible
1967 * future enhancement. WARNING: it is not trivial to make.
1969 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
1970 __u32 now
, __s32
*seq_rtt
)
1972 struct tcp_sock
*tp
= tcp_sk(sk
);
1973 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
1974 __u32 seq
= tp
->snd_una
;
1975 __u32 packets_acked
;
1978 /* If we get here, the whole TSO packet has not been
1981 BUG_ON(!after(scb
->end_seq
, seq
));
1983 packets_acked
= tcp_skb_pcount(skb
);
1984 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
1986 packets_acked
-= tcp_skb_pcount(skb
);
1988 if (packets_acked
) {
1989 __u8 sacked
= scb
->sacked
;
1991 acked
|= FLAG_DATA_ACKED
;
1993 if (sacked
& TCPCB_RETRANS
) {
1994 if (sacked
& TCPCB_SACKED_RETRANS
)
1995 tp
->retrans_out
-= packets_acked
;
1996 acked
|= FLAG_RETRANS_DATA_ACKED
;
1998 } else if (*seq_rtt
< 0)
1999 *seq_rtt
= now
- scb
->when
;
2000 if (sacked
& TCPCB_SACKED_ACKED
)
2001 tp
->sacked_out
-= packets_acked
;
2002 if (sacked
& TCPCB_LOST
)
2003 tp
->lost_out
-= packets_acked
;
2004 if (sacked
& TCPCB_URG
) {
2006 !before(seq
, tp
->snd_up
))
2009 } else if (*seq_rtt
< 0)
2010 *seq_rtt
= now
- scb
->when
;
2012 if (tp
->fackets_out
) {
2013 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2014 tp
->fackets_out
-= dval
;
2016 tp
->packets_out
-= packets_acked
;
2018 BUG_ON(tcp_skb_pcount(skb
) == 0);
2019 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2026 /* Remove acknowledged frames from the retransmission queue. */
2027 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
, s32
*seq_usrtt
)
2029 struct tcp_sock
*tp
= tcp_sk(sk
);
2030 struct sk_buff
*skb
;
2031 __u32 now
= tcp_time_stamp
;
2034 struct timeval usnow
;
2038 do_gettimeofday(&usnow
);
2040 while ((skb
= skb_peek(&sk
->sk_write_queue
)) &&
2041 skb
!= sk
->sk_send_head
) {
2042 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2043 __u8 sacked
= scb
->sacked
;
2045 /* If our packet is before the ack sequence we can
2046 * discard it as it's confirmed to have arrived at
2049 if (after(scb
->end_seq
, tp
->snd_una
)) {
2050 if (tcp_skb_pcount(skb
) > 1)
2051 acked
|= tcp_tso_acked(sk
, skb
,
2056 /* Initial outgoing SYN's get put onto the write_queue
2057 * just like anything else we transmit. It is not
2058 * true data, and if we misinform our callers that
2059 * this ACK acks real data, we will erroneously exit
2060 * connection startup slow start one packet too
2061 * quickly. This is severely frowned upon behavior.
2063 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2064 acked
|= FLAG_DATA_ACKED
;
2067 acked
|= FLAG_SYN_ACKED
;
2068 tp
->retrans_stamp
= 0;
2072 if (sacked
& TCPCB_RETRANS
) {
2073 if(sacked
& TCPCB_SACKED_RETRANS
)
2074 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2075 acked
|= FLAG_RETRANS_DATA_ACKED
;
2077 } else if (seq_rtt
< 0)
2078 seq_rtt
= now
- scb
->when
;
2080 *seq_usrtt
= (usnow
.tv_sec
- skb
->stamp
.tv_sec
) * 1000000
2081 + (usnow
.tv_usec
- skb
->stamp
.tv_usec
);
2083 if (sacked
& TCPCB_SACKED_ACKED
)
2084 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2085 if (sacked
& TCPCB_LOST
)
2086 tp
->lost_out
-= tcp_skb_pcount(skb
);
2087 if (sacked
& TCPCB_URG
) {
2089 !before(scb
->end_seq
, tp
->snd_up
))
2092 } else if (seq_rtt
< 0)
2093 seq_rtt
= now
- scb
->when
;
2094 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2095 tcp_packets_out_dec(tp
, skb
);
2096 __skb_unlink(skb
, skb
->list
);
2097 sk_stream_free_skb(sk
, skb
);
2100 if (acked
&FLAG_ACKED
) {
2101 tcp_ack_update_rtt(tp
, acked
, seq_rtt
, seq_usrtt
);
2102 tcp_ack_packets_out(sk
, tp
);
2104 if (tp
->ca_ops
->pkts_acked
)
2105 tp
->ca_ops
->pkts_acked(tp
, pkts_acked
);
2108 #if FASTRETRANS_DEBUG > 0
2109 BUG_TRAP((int)tp
->sacked_out
>= 0);
2110 BUG_TRAP((int)tp
->lost_out
>= 0);
2111 BUG_TRAP((int)tp
->retrans_out
>= 0);
2112 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2114 printk(KERN_DEBUG
"Leak l=%u %d\n",
2115 tp
->lost_out
, tp
->ca_state
);
2118 if (tp
->sacked_out
) {
2119 printk(KERN_DEBUG
"Leak s=%u %d\n",
2120 tp
->sacked_out
, tp
->ca_state
);
2123 if (tp
->retrans_out
) {
2124 printk(KERN_DEBUG
"Leak r=%u %d\n",
2125 tp
->retrans_out
, tp
->ca_state
);
2126 tp
->retrans_out
= 0;
2130 *seq_rtt_p
= seq_rtt
;
2134 static void tcp_ack_probe(struct sock
*sk
)
2136 struct tcp_sock
*tp
= tcp_sk(sk
);
2138 /* Was it a usable window open? */
2140 if (!after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
2141 tp
->snd_una
+ tp
->snd_wnd
)) {
2143 tcp_clear_xmit_timer(sk
, TCP_TIME_PROBE0
);
2144 /* Socket must be waked up by subsequent tcp_data_snd_check().
2145 * This function is not for random using!
2148 tcp_reset_xmit_timer(sk
, TCP_TIME_PROBE0
,
2149 min(tp
->rto
<< tp
->backoff
, TCP_RTO_MAX
));
2153 static inline int tcp_ack_is_dubious(struct tcp_sock
*tp
, int flag
)
2155 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2156 tp
->ca_state
!= TCP_CA_Open
);
2159 static inline int tcp_may_raise_cwnd(struct tcp_sock
*tp
, int flag
)
2161 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2162 !((1<<tp
->ca_state
)&(TCPF_CA_Recovery
|TCPF_CA_CWR
));
2165 /* Check that window update is acceptable.
2166 * The function assumes that snd_una<=ack<=snd_next.
2168 static inline int tcp_may_update_window(struct tcp_sock
*tp
, u32 ack
,
2169 u32 ack_seq
, u32 nwin
)
2171 return (after(ack
, tp
->snd_una
) ||
2172 after(ack_seq
, tp
->snd_wl1
) ||
2173 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2176 /* Update our send window.
2178 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2179 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2181 static int tcp_ack_update_window(struct sock
*sk
, struct tcp_sock
*tp
,
2182 struct sk_buff
*skb
, u32 ack
, u32 ack_seq
)
2185 u32 nwin
= ntohs(skb
->h
.th
->window
);
2187 if (likely(!skb
->h
.th
->syn
))
2188 nwin
<<= tp
->rx_opt
.snd_wscale
;
2190 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2191 flag
|= FLAG_WIN_UPDATE
;
2192 tcp_update_wl(tp
, ack
, ack_seq
);
2194 if (tp
->snd_wnd
!= nwin
) {
2197 /* Note, it is the only place, where
2198 * fast path is recovered for sending TCP.
2200 tcp_fast_path_check(sk
, tp
);
2202 if (nwin
> tp
->max_window
) {
2203 tp
->max_window
= nwin
;
2204 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
2214 static void tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
)
2216 struct tcp_sock
*tp
= tcp_sk(sk
);
2218 tcp_sync_left_out(tp
);
2220 if (tp
->snd_una
== prior_snd_una
||
2221 !before(tp
->snd_una
, tp
->frto_highmark
)) {
2222 /* RTO was caused by loss, start retransmitting in
2223 * go-back-N slow start
2225 tcp_enter_frto_loss(sk
);
2229 if (tp
->frto_counter
== 1) {
2230 /* First ACK after RTO advances the window: allow two new
2233 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2235 /* Also the second ACK after RTO advances the window.
2236 * The RTO was likely spurious. Reduce cwnd and continue
2237 * in congestion avoidance
2239 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2240 tcp_moderate_cwnd(tp
);
2243 /* F-RTO affects on two new ACKs following RTO.
2244 * At latest on third ACK the TCP behavor is back to normal.
2246 tp
->frto_counter
= (tp
->frto_counter
+ 1) % 3;
2249 /* This routine deals with incoming acks, but not outgoing ones. */
2250 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2252 struct tcp_sock
*tp
= tcp_sk(sk
);
2253 u32 prior_snd_una
= tp
->snd_una
;
2254 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2255 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2256 u32 prior_in_flight
;
2261 /* If the ack is newer than sent or older than previous acks
2262 * then we can probably ignore it.
2264 if (after(ack
, tp
->snd_nxt
))
2265 goto uninteresting_ack
;
2267 if (before(ack
, prior_snd_una
))
2270 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2271 /* Window is constant, pure forward advance.
2272 * No more checks are required.
2273 * Note, we use the fact that SND.UNA>=SND.WL2.
2275 tcp_update_wl(tp
, ack
, ack_seq
);
2277 flag
|= FLAG_WIN_UPDATE
;
2279 tcp_ca_event(tp
, CA_EVENT_FAST_ACK
);
2281 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2283 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2286 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2288 flag
|= tcp_ack_update_window(sk
, tp
, skb
, ack
, ack_seq
);
2290 if (TCP_SKB_CB(skb
)->sacked
)
2291 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2293 if (TCP_ECN_rcv_ecn_echo(tp
, skb
->h
.th
))
2296 tcp_ca_event(tp
, CA_EVENT_SLOW_ACK
);
2299 /* We passed data and got it acked, remove any soft error
2300 * log. Something worked...
2302 sk
->sk_err_soft
= 0;
2303 tp
->rcv_tstamp
= tcp_time_stamp
;
2304 prior_packets
= tp
->packets_out
;
2308 prior_in_flight
= tcp_packets_in_flight(tp
);
2310 /* See if we can take anything off of the retransmit queue. */
2311 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
,
2312 tp
->ca_ops
->rtt_sample
? &seq_usrtt
: NULL
);
2314 if (tp
->frto_counter
)
2315 tcp_process_frto(sk
, prior_snd_una
);
2317 if (tcp_ack_is_dubious(tp
, flag
)) {
2318 /* Advanve CWND, if state allows this. */
2319 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(tp
, flag
))
2320 tcp_cong_avoid(tp
, ack
, seq_rtt
, prior_in_flight
, 0);
2321 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2323 if ((flag
& FLAG_DATA_ACKED
))
2324 tcp_cong_avoid(tp
, ack
, seq_rtt
, prior_in_flight
, 1);
2327 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2328 dst_confirm(sk
->sk_dst_cache
);
2335 /* If this ack opens up a zero window, clear backoff. It was
2336 * being used to time the probes, and is probably far higher than
2337 * it needs to be for normal retransmission.
2339 if (sk
->sk_send_head
)
2344 if (TCP_SKB_CB(skb
)->sacked
)
2345 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2348 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2353 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2354 * But, this can also be called on packets in the established flow when
2355 * the fast version below fails.
2357 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2360 struct tcphdr
*th
= skb
->h
.th
;
2361 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2363 ptr
= (unsigned char *)(th
+ 1);
2364 opt_rx
->saw_tstamp
= 0;
2373 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2378 if (opsize
< 2) /* "silly options" */
2380 if (opsize
> length
)
2381 return; /* don't parse partial options */
2384 if(opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2385 u16 in_mss
= ntohs(get_unaligned((__u16
*)ptr
));
2387 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2388 in_mss
= opt_rx
->user_mss
;
2389 opt_rx
->mss_clamp
= in_mss
;
2394 if(opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2395 if (sysctl_tcp_window_scaling
) {
2396 __u8 snd_wscale
= *(__u8
*) ptr
;
2397 opt_rx
->wscale_ok
= 1;
2398 if (snd_wscale
> 14) {
2400 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2401 "scaling value %d >14 received.\n",
2405 opt_rx
->snd_wscale
= snd_wscale
;
2408 case TCPOPT_TIMESTAMP
:
2409 if(opsize
==TCPOLEN_TIMESTAMP
) {
2410 if ((estab
&& opt_rx
->tstamp_ok
) ||
2411 (!estab
&& sysctl_tcp_timestamps
)) {
2412 opt_rx
->saw_tstamp
= 1;
2413 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__u32
*)ptr
));
2414 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__u32
*)(ptr
+4)));
2418 case TCPOPT_SACK_PERM
:
2419 if(opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2420 if (sysctl_tcp_sack
) {
2421 opt_rx
->sack_ok
= 1;
2422 tcp_sack_reset(opt_rx
);
2428 if((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2429 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2431 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2440 /* Fast parse options. This hopes to only see timestamps.
2441 * If it is wrong it falls back on tcp_parse_options().
2443 static inline int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2444 struct tcp_sock
*tp
)
2446 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2447 tp
->rx_opt
.saw_tstamp
= 0;
2449 } else if (tp
->rx_opt
.tstamp_ok
&&
2450 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2451 __u32
*ptr
= (__u32
*)(th
+ 1);
2452 if (*ptr
== ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2453 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2454 tp
->rx_opt
.saw_tstamp
= 1;
2456 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2458 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2462 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2466 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2468 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2469 tp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
2472 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2474 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2475 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2476 * extra check below makes sure this can only happen
2477 * for pure ACK frames. -DaveM
2479 * Not only, also it occurs for expired timestamps.
2482 if((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2483 xtime
.tv_sec
>= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2484 tcp_store_ts_recent(tp
);
2488 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2490 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2491 * it can pass through stack. So, the following predicate verifies that
2492 * this segment is not used for anything but congestion avoidance or
2493 * fast retransmit. Moreover, we even are able to eliminate most of such
2494 * second order effects, if we apply some small "replay" window (~RTO)
2495 * to timestamp space.
2497 * All these measures still do not guarantee that we reject wrapped ACKs
2498 * on networks with high bandwidth, when sequence space is recycled fastly,
2499 * but it guarantees that such events will be very rare and do not affect
2500 * connection seriously. This doesn't look nice, but alas, PAWS is really
2503 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2504 * states that events when retransmit arrives after original data are rare.
2505 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2506 * the biggest problem on large power networks even with minor reordering.
2507 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2508 * up to bandwidth of 18Gigabit/sec. 8) ]
2511 static int tcp_disordered_ack(struct tcp_sock
*tp
, struct sk_buff
*skb
)
2513 struct tcphdr
*th
= skb
->h
.th
;
2514 u32 seq
= TCP_SKB_CB(skb
)->seq
;
2515 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2517 return (/* 1. Pure ACK with correct sequence number. */
2518 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
2520 /* 2. ... and duplicate ACK. */
2521 ack
== tp
->snd_una
&&
2523 /* 3. ... and does not update window. */
2524 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
2526 /* 4. ... and sits in replay window. */
2527 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (tp
->rto
*1024)/HZ
);
2530 static inline int tcp_paws_discard(struct tcp_sock
*tp
, struct sk_buff
*skb
)
2532 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
2533 xtime
.tv_sec
< tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
2534 !tcp_disordered_ack(tp
, skb
));
2537 /* Check segment sequence number for validity.
2539 * Segment controls are considered valid, if the segment
2540 * fits to the window after truncation to the window. Acceptability
2541 * of data (and SYN, FIN, of course) is checked separately.
2542 * See tcp_data_queue(), for example.
2544 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2545 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2546 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2547 * (borrowed from freebsd)
2550 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2552 return !before(end_seq
, tp
->rcv_wup
) &&
2553 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
2556 /* When we get a reset we do this. */
2557 static void tcp_reset(struct sock
*sk
)
2559 /* We want the right error as BSD sees it (and indeed as we do). */
2560 switch (sk
->sk_state
) {
2562 sk
->sk_err
= ECONNREFUSED
;
2564 case TCP_CLOSE_WAIT
:
2570 sk
->sk_err
= ECONNRESET
;
2573 if (!sock_flag(sk
, SOCK_DEAD
))
2574 sk
->sk_error_report(sk
);
2580 * Process the FIN bit. This now behaves as it is supposed to work
2581 * and the FIN takes effect when it is validly part of sequence
2582 * space. Not before when we get holes.
2584 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2585 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2588 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2589 * close and we go into CLOSING (and later onto TIME-WAIT)
2591 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2593 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
2595 struct tcp_sock
*tp
= tcp_sk(sk
);
2597 tcp_schedule_ack(tp
);
2599 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
2600 sock_set_flag(sk
, SOCK_DONE
);
2602 switch (sk
->sk_state
) {
2604 case TCP_ESTABLISHED
:
2605 /* Move to CLOSE_WAIT */
2606 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
2607 tp
->ack
.pingpong
= 1;
2610 case TCP_CLOSE_WAIT
:
2612 /* Received a retransmission of the FIN, do
2617 /* RFC793: Remain in the LAST-ACK state. */
2621 /* This case occurs when a simultaneous close
2622 * happens, we must ack the received FIN and
2623 * enter the CLOSING state.
2626 tcp_set_state(sk
, TCP_CLOSING
);
2629 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2631 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
2634 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2635 * cases we should never reach this piece of code.
2637 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
2638 __FUNCTION__
, sk
->sk_state
);
2642 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2643 * Probably, we should reset in this case. For now drop them.
2645 __skb_queue_purge(&tp
->out_of_order_queue
);
2646 if (tp
->rx_opt
.sack_ok
)
2647 tcp_sack_reset(&tp
->rx_opt
);
2648 sk_stream_mem_reclaim(sk
);
2650 if (!sock_flag(sk
, SOCK_DEAD
)) {
2651 sk
->sk_state_change(sk
);
2653 /* Do not send POLL_HUP for half duplex close. */
2654 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
2655 sk
->sk_state
== TCP_CLOSE
)
2656 sk_wake_async(sk
, 1, POLL_HUP
);
2658 sk_wake_async(sk
, 1, POLL_IN
);
2662 static __inline__
int
2663 tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
2665 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
2666 if (before(seq
, sp
->start_seq
))
2667 sp
->start_seq
= seq
;
2668 if (after(end_seq
, sp
->end_seq
))
2669 sp
->end_seq
= end_seq
;
2675 static inline void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2677 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2678 if (before(seq
, tp
->rcv_nxt
))
2679 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
2681 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
2683 tp
->rx_opt
.dsack
= 1;
2684 tp
->duplicate_sack
[0].start_seq
= seq
;
2685 tp
->duplicate_sack
[0].end_seq
= end_seq
;
2686 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
2690 static inline void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2692 if (!tp
->rx_opt
.dsack
)
2693 tcp_dsack_set(tp
, seq
, end_seq
);
2695 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
2698 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
2700 struct tcp_sock
*tp
= tcp_sk(sk
);
2702 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
2703 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
2704 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
2705 tcp_enter_quickack_mode(tp
);
2707 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2708 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2710 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
2711 end_seq
= tp
->rcv_nxt
;
2712 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
2719 /* These routines update the SACK block as out-of-order packets arrive or
2720 * in-order packets close up the sequence space.
2722 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
2725 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2726 struct tcp_sack_block
*swalk
= sp
+1;
2728 /* See if the recent change to the first SACK eats into
2729 * or hits the sequence space of other SACK blocks, if so coalesce.
2731 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
2732 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
2735 /* Zap SWALK, by moving every further SACK up by one slot.
2736 * Decrease num_sacks.
2738 tp
->rx_opt
.num_sacks
--;
2739 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2740 for(i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
2744 this_sack
++, swalk
++;
2748 static __inline__
void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
2752 tmp
= sack1
->start_seq
;
2753 sack1
->start_seq
= sack2
->start_seq
;
2754 sack2
->start_seq
= tmp
;
2756 tmp
= sack1
->end_seq
;
2757 sack1
->end_seq
= sack2
->end_seq
;
2758 sack2
->end_seq
= tmp
;
2761 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
2763 struct tcp_sock
*tp
= tcp_sk(sk
);
2764 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2765 int cur_sacks
= tp
->rx_opt
.num_sacks
;
2771 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
2772 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
2773 /* Rotate this_sack to the first one. */
2774 for (; this_sack
>0; this_sack
--, sp
--)
2775 tcp_sack_swap(sp
, sp
-1);
2777 tcp_sack_maybe_coalesce(tp
);
2782 /* Could not find an adjacent existing SACK, build a new one,
2783 * put it at the front, and shift everyone else down. We
2784 * always know there is at least one SACK present already here.
2786 * If the sack array is full, forget about the last one.
2788 if (this_sack
>= 4) {
2790 tp
->rx_opt
.num_sacks
--;
2793 for(; this_sack
> 0; this_sack
--, sp
--)
2797 /* Build the new head SACK, and we're done. */
2798 sp
->start_seq
= seq
;
2799 sp
->end_seq
= end_seq
;
2800 tp
->rx_opt
.num_sacks
++;
2801 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2804 /* RCV.NXT advances, some SACKs should be eaten. */
2806 static void tcp_sack_remove(struct tcp_sock
*tp
)
2808 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2809 int num_sacks
= tp
->rx_opt
.num_sacks
;
2812 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2813 if (skb_queue_len(&tp
->out_of_order_queue
) == 0) {
2814 tp
->rx_opt
.num_sacks
= 0;
2815 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
2819 for(this_sack
= 0; this_sack
< num_sacks
; ) {
2820 /* Check if the start of the sack is covered by RCV.NXT. */
2821 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
2824 /* RCV.NXT must cover all the block! */
2825 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
2827 /* Zap this SACK, by moving forward any other SACKS. */
2828 for (i
=this_sack
+1; i
< num_sacks
; i
++)
2829 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
2836 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
2837 tp
->rx_opt
.num_sacks
= num_sacks
;
2838 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2842 /* This one checks to see if we can put data from the
2843 * out_of_order queue into the receive_queue.
2845 static void tcp_ofo_queue(struct sock
*sk
)
2847 struct tcp_sock
*tp
= tcp_sk(sk
);
2848 __u32 dsack_high
= tp
->rcv_nxt
;
2849 struct sk_buff
*skb
;
2851 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
2852 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
2855 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
2856 __u32 dsack
= dsack_high
;
2857 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
2858 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
2859 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
2862 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
2863 SOCK_DEBUG(sk
, "ofo packet was already received \n");
2864 __skb_unlink(skb
, skb
->list
);
2868 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
2869 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
2870 TCP_SKB_CB(skb
)->end_seq
);
2872 __skb_unlink(skb
, skb
->list
);
2873 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
2874 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
2876 tcp_fin(skb
, sk
, skb
->h
.th
);
2880 static int tcp_prune_queue(struct sock
*sk
);
2882 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
2884 struct tcphdr
*th
= skb
->h
.th
;
2885 struct tcp_sock
*tp
= tcp_sk(sk
);
2888 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
2891 __skb_pull(skb
, th
->doff
*4);
2893 TCP_ECN_accept_cwr(tp
, skb
);
2895 if (tp
->rx_opt
.dsack
) {
2896 tp
->rx_opt
.dsack
= 0;
2897 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
2898 4 - tp
->rx_opt
.tstamp_ok
);
2901 /* Queue data for delivery to the user.
2902 * Packets in sequence go to the receive queue.
2903 * Out of sequence packets to the out_of_order_queue.
2905 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
2906 if (tcp_receive_window(tp
) == 0)
2909 /* Ok. In sequence. In window. */
2910 if (tp
->ucopy
.task
== current
&&
2911 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
2912 sock_owned_by_user(sk
) && !tp
->urg_data
) {
2913 int chunk
= min_t(unsigned int, skb
->len
,
2916 __set_current_state(TASK_RUNNING
);
2919 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
2920 tp
->ucopy
.len
-= chunk
;
2921 tp
->copied_seq
+= chunk
;
2922 eaten
= (chunk
== skb
->len
&& !th
->fin
);
2923 tcp_rcv_space_adjust(sk
);
2931 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
2932 !sk_stream_rmem_schedule(sk
, skb
))) {
2933 if (tcp_prune_queue(sk
) < 0 ||
2934 !sk_stream_rmem_schedule(sk
, skb
))
2937 sk_stream_set_owner_r(skb
, sk
);
2938 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
2940 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
2942 tcp_event_data_recv(sk
, tp
, skb
);
2944 tcp_fin(skb
, sk
, th
);
2946 if (skb_queue_len(&tp
->out_of_order_queue
)) {
2949 /* RFC2581. 4.2. SHOULD send immediate ACK, when
2950 * gap in queue is filled.
2952 if (!skb_queue_len(&tp
->out_of_order_queue
))
2953 tp
->ack
.pingpong
= 0;
2956 if (tp
->rx_opt
.num_sacks
)
2957 tcp_sack_remove(tp
);
2959 tcp_fast_path_check(sk
, tp
);
2963 else if (!sock_flag(sk
, SOCK_DEAD
))
2964 sk
->sk_data_ready(sk
, 0);
2968 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
2969 /* A retransmit, 2nd most common case. Force an immediate ack. */
2970 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
2971 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
2974 tcp_enter_quickack_mode(tp
);
2975 tcp_schedule_ack(tp
);
2981 /* Out of window. F.e. zero window probe. */
2982 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
2985 tcp_enter_quickack_mode(tp
);
2987 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
2988 /* Partial packet, seq < rcv_next < end_seq */
2989 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
2990 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
2991 TCP_SKB_CB(skb
)->end_seq
);
2993 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
2995 /* If window is closed, drop tail of packet. But after
2996 * remembering D-SACK for its head made in previous line.
2998 if (!tcp_receive_window(tp
))
3003 TCP_ECN_check_ce(tp
, skb
);
3005 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3006 !sk_stream_rmem_schedule(sk
, skb
)) {
3007 if (tcp_prune_queue(sk
) < 0 ||
3008 !sk_stream_rmem_schedule(sk
, skb
))
3012 /* Disable header prediction. */
3014 tcp_schedule_ack(tp
);
3016 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3017 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3019 sk_stream_set_owner_r(skb
, sk
);
3021 if (!skb_peek(&tp
->out_of_order_queue
)) {
3022 /* Initial out of order segment, build 1 SACK. */
3023 if (tp
->rx_opt
.sack_ok
) {
3024 tp
->rx_opt
.num_sacks
= 1;
3025 tp
->rx_opt
.dsack
= 0;
3026 tp
->rx_opt
.eff_sacks
= 1;
3027 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3028 tp
->selective_acks
[0].end_seq
=
3029 TCP_SKB_CB(skb
)->end_seq
;
3031 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3033 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3034 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3035 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3037 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3038 __skb_append(skb1
, skb
);
3040 if (!tp
->rx_opt
.num_sacks
||
3041 tp
->selective_acks
[0].end_seq
!= seq
)
3044 /* Common case: data arrive in order after hole. */
3045 tp
->selective_acks
[0].end_seq
= end_seq
;
3049 /* Find place to insert this segment. */
3051 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3053 } while ((skb1
= skb1
->prev
) !=
3054 (struct sk_buff
*)&tp
->out_of_order_queue
);
3056 /* Do skb overlap to previous one? */
3057 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3058 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3059 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3060 /* All the bits are present. Drop. */
3062 tcp_dsack_set(tp
, seq
, end_seq
);
3065 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3066 /* Partial overlap. */
3067 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3072 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3074 /* And clean segments covered by new one as whole. */
3075 while ((skb1
= skb
->next
) !=
3076 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3077 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3078 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3079 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3082 __skb_unlink(skb1
, skb1
->list
);
3083 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3088 if (tp
->rx_opt
.sack_ok
)
3089 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3093 /* Collapse contiguous sequence of skbs head..tail with
3094 * sequence numbers start..end.
3095 * Segments with FIN/SYN are not collapsed (only because this
3099 tcp_collapse(struct sock
*sk
, struct sk_buff
*head
,
3100 struct sk_buff
*tail
, u32 start
, u32 end
)
3102 struct sk_buff
*skb
;
3104 /* First, check that queue is collapsable and find
3105 * the point where collapsing can be useful. */
3106 for (skb
= head
; skb
!= tail
; ) {
3107 /* No new bits? It is possible on ofo queue. */
3108 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3109 struct sk_buff
*next
= skb
->next
;
3110 __skb_unlink(skb
, skb
->list
);
3112 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3117 /* The first skb to collapse is:
3119 * - bloated or contains data before "start" or
3120 * overlaps to the next one.
3122 if (!skb
->h
.th
->syn
&& !skb
->h
.th
->fin
&&
3123 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3124 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3125 (skb
->next
!= tail
&&
3126 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3129 /* Decided to skip this, advance start seq. */
3130 start
= TCP_SKB_CB(skb
)->end_seq
;
3133 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3136 while (before(start
, end
)) {
3137 struct sk_buff
*nskb
;
3138 int header
= skb_headroom(skb
);
3139 int copy
= SKB_MAX_ORDER(header
, 0);
3141 /* Too big header? This can happen with IPv6. */
3144 if (end
-start
< copy
)
3146 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3149 skb_reserve(nskb
, header
);
3150 memcpy(nskb
->head
, skb
->head
, header
);
3151 nskb
->nh
.raw
= nskb
->head
+ (skb
->nh
.raw
-skb
->head
);
3152 nskb
->h
.raw
= nskb
->head
+ (skb
->h
.raw
-skb
->head
);
3153 nskb
->mac
.raw
= nskb
->head
+ (skb
->mac
.raw
-skb
->head
);
3154 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3155 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3156 __skb_insert(nskb
, skb
->prev
, skb
, skb
->list
);
3157 sk_stream_set_owner_r(nskb
, sk
);
3159 /* Copy data, releasing collapsed skbs. */
3161 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3162 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3164 if (offset
< 0) BUG();
3166 size
= min(copy
, size
);
3167 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3169 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3173 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3174 struct sk_buff
*next
= skb
->next
;
3175 __skb_unlink(skb
, skb
->list
);
3177 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3179 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3186 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3187 * and tcp_collapse() them until all the queue is collapsed.
3189 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3191 struct tcp_sock
*tp
= tcp_sk(sk
);
3192 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3193 struct sk_buff
*head
;
3199 start
= TCP_SKB_CB(skb
)->seq
;
3200 end
= TCP_SKB_CB(skb
)->end_seq
;
3206 /* Segment is terminated when we see gap or when
3207 * we are at the end of all the queue. */
3208 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3209 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3210 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3211 tcp_collapse(sk
, head
, skb
, start
, end
);
3213 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3215 /* Start new segment */
3216 start
= TCP_SKB_CB(skb
)->seq
;
3217 end
= TCP_SKB_CB(skb
)->end_seq
;
3219 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3220 start
= TCP_SKB_CB(skb
)->seq
;
3221 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3222 end
= TCP_SKB_CB(skb
)->end_seq
;
3227 /* Reduce allocated memory if we can, trying to get
3228 * the socket within its memory limits again.
3230 * Return less than zero if we should start dropping frames
3231 * until the socket owning process reads some of the data
3232 * to stabilize the situation.
3234 static int tcp_prune_queue(struct sock
*sk
)
3236 struct tcp_sock
*tp
= tcp_sk(sk
);
3238 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3240 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3242 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3243 tcp_clamp_window(sk
, tp
);
3244 else if (tcp_memory_pressure
)
3245 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3247 tcp_collapse_ofo_queue(sk
);
3248 tcp_collapse(sk
, sk
->sk_receive_queue
.next
,
3249 (struct sk_buff
*)&sk
->sk_receive_queue
,
3250 tp
->copied_seq
, tp
->rcv_nxt
);
3251 sk_stream_mem_reclaim(sk
);
3253 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3256 /* Collapsing did not help, destructive actions follow.
3257 * This must not ever occur. */
3259 /* First, purge the out_of_order queue. */
3260 if (skb_queue_len(&tp
->out_of_order_queue
)) {
3261 NET_ADD_STATS_BH(LINUX_MIB_OFOPRUNED
,
3262 skb_queue_len(&tp
->out_of_order_queue
));
3263 __skb_queue_purge(&tp
->out_of_order_queue
);
3265 /* Reset SACK state. A conforming SACK implementation will
3266 * do the same at a timeout based retransmit. When a connection
3267 * is in a sad state like this, we care only about integrity
3268 * of the connection not performance.
3270 if (tp
->rx_opt
.sack_ok
)
3271 tcp_sack_reset(&tp
->rx_opt
);
3272 sk_stream_mem_reclaim(sk
);
3275 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3278 /* If we are really being abused, tell the caller to silently
3279 * drop receive data on the floor. It will get retransmitted
3280 * and hopefully then we'll have sufficient space.
3282 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3284 /* Massive buffer overcommit. */
3290 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3291 * As additional protections, we do not touch cwnd in retransmission phases,
3292 * and if application hit its sndbuf limit recently.
3294 void tcp_cwnd_application_limited(struct sock
*sk
)
3296 struct tcp_sock
*tp
= tcp_sk(sk
);
3298 if (tp
->ca_state
== TCP_CA_Open
&&
3299 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3300 /* Limited by application or receiver window. */
3301 u32 win_used
= max(tp
->snd_cwnd_used
, 2U);
3302 if (win_used
< tp
->snd_cwnd
) {
3303 tp
->snd_ssthresh
= tcp_current_ssthresh(tp
);
3304 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3306 tp
->snd_cwnd_used
= 0;
3308 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3312 /* When incoming ACK allowed to free some skb from write_queue,
3313 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3314 * on the exit from tcp input handler.
3316 * PROBLEM: sndbuf expansion does not work well with largesend.
3318 static void tcp_new_space(struct sock
*sk
)
3320 struct tcp_sock
*tp
= tcp_sk(sk
);
3322 if (tp
->packets_out
< tp
->snd_cwnd
&&
3323 !(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
) &&
3324 !tcp_memory_pressure
&&
3325 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
3326 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache_std
) +
3327 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3328 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3329 tp
->reordering
+ 1);
3330 sndmem
*= 2*demanded
;
3331 if (sndmem
> sk
->sk_sndbuf
)
3332 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3333 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3336 sk
->sk_write_space(sk
);
3339 static inline void tcp_check_space(struct sock
*sk
)
3341 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3342 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3343 if (sk
->sk_socket
&&
3344 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3349 static void __tcp_data_snd_check(struct sock
*sk
, struct sk_buff
*skb
)
3351 struct tcp_sock
*tp
= tcp_sk(sk
);
3353 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
+ tp
->snd_wnd
) ||
3354 tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
||
3355 tcp_write_xmit(sk
, tp
->nonagle
))
3356 tcp_check_probe_timer(sk
, tp
);
3359 static __inline__
void tcp_data_snd_check(struct sock
*sk
)
3361 struct sk_buff
*skb
= sk
->sk_send_head
;
3364 __tcp_data_snd_check(sk
, skb
);
3365 tcp_check_space(sk
);
3369 * Check if sending an ack is needed.
3371 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3373 struct tcp_sock
*tp
= tcp_sk(sk
);
3375 /* More than one full frame received... */
3376 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > tp
->ack
.rcv_mss
3377 /* ... and right edge of window advances far enough.
3378 * (tcp_recvmsg() will send ACK otherwise). Or...
3380 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3381 /* We ACK each frame or... */
3382 tcp_in_quickack_mode(tp
) ||
3383 /* We have out of order data. */
3385 skb_peek(&tp
->out_of_order_queue
))) {
3386 /* Then ack it now */
3389 /* Else, send delayed ack. */
3390 tcp_send_delayed_ack(sk
);
3394 static __inline__
void tcp_ack_snd_check(struct sock
*sk
)
3396 struct tcp_sock
*tp
= tcp_sk(sk
);
3397 if (!tcp_ack_scheduled(tp
)) {
3398 /* We sent a data segment already. */
3401 __tcp_ack_snd_check(sk
, 1);
3405 * This routine is only called when we have urgent data
3406 * signalled. Its the 'slow' part of tcp_urg. It could be
3407 * moved inline now as tcp_urg is only called from one
3408 * place. We handle URGent data wrong. We have to - as
3409 * BSD still doesn't use the correction from RFC961.
3410 * For 1003.1g we should support a new option TCP_STDURG to permit
3411 * either form (or just set the sysctl tcp_stdurg).
3414 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3416 struct tcp_sock
*tp
= tcp_sk(sk
);
3417 u32 ptr
= ntohs(th
->urg_ptr
);
3419 if (ptr
&& !sysctl_tcp_stdurg
)
3421 ptr
+= ntohl(th
->seq
);
3423 /* Ignore urgent data that we've already seen and read. */
3424 if (after(tp
->copied_seq
, ptr
))
3427 /* Do not replay urg ptr.
3429 * NOTE: interesting situation not covered by specs.
3430 * Misbehaving sender may send urg ptr, pointing to segment,
3431 * which we already have in ofo queue. We are not able to fetch
3432 * such data and will stay in TCP_URG_NOTYET until will be eaten
3433 * by recvmsg(). Seems, we are not obliged to handle such wicked
3434 * situations. But it is worth to think about possibility of some
3435 * DoSes using some hypothetical application level deadlock.
3437 if (before(ptr
, tp
->rcv_nxt
))
3440 /* Do we already have a newer (or duplicate) urgent pointer? */
3441 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3444 /* Tell the world about our new urgent pointer. */
3447 /* We may be adding urgent data when the last byte read was
3448 * urgent. To do this requires some care. We cannot just ignore
3449 * tp->copied_seq since we would read the last urgent byte again
3450 * as data, nor can we alter copied_seq until this data arrives
3451 * or we break the sematics of SIOCATMARK (and thus sockatmark())
3453 * NOTE. Double Dutch. Rendering to plain English: author of comment
3454 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3455 * and expect that both A and B disappear from stream. This is _wrong_.
3456 * Though this happens in BSD with high probability, this is occasional.
3457 * Any application relying on this is buggy. Note also, that fix "works"
3458 * only in this artificial test. Insert some normal data between A and B and we will
3459 * decline of BSD again. Verdict: it is better to remove to trap
3462 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3463 !sock_flag(sk
, SOCK_URGINLINE
) &&
3464 tp
->copied_seq
!= tp
->rcv_nxt
) {
3465 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3467 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3468 __skb_unlink(skb
, skb
->list
);
3473 tp
->urg_data
= TCP_URG_NOTYET
;
3476 /* Disable header prediction. */
3480 /* This is the 'fast' part of urgent handling. */
3481 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
3483 struct tcp_sock
*tp
= tcp_sk(sk
);
3485 /* Check if we get a new urgent pointer - normally not. */
3487 tcp_check_urg(sk
,th
);
3489 /* Do we wait for any urgent data? - normally not... */
3490 if (tp
->urg_data
== TCP_URG_NOTYET
) {
3491 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
3494 /* Is the urgent pointer pointing into this packet? */
3495 if (ptr
< skb
->len
) {
3497 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
3499 tp
->urg_data
= TCP_URG_VALID
| tmp
;
3500 if (!sock_flag(sk
, SOCK_DEAD
))
3501 sk
->sk_data_ready(sk
, 0);
3506 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3508 struct tcp_sock
*tp
= tcp_sk(sk
);
3509 int chunk
= skb
->len
- hlen
;
3513 if (skb
->ip_summed
==CHECKSUM_UNNECESSARY
)
3514 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
3516 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
3520 tp
->ucopy
.len
-= chunk
;
3521 tp
->copied_seq
+= chunk
;
3522 tcp_rcv_space_adjust(sk
);
3529 static int __tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3533 if (sock_owned_by_user(sk
)) {
3535 result
= __tcp_checksum_complete(skb
);
3538 result
= __tcp_checksum_complete(skb
);
3543 static __inline__
int
3544 tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3546 return skb
->ip_summed
!= CHECKSUM_UNNECESSARY
&&
3547 __tcp_checksum_complete_user(sk
, skb
);
3551 * TCP receive function for the ESTABLISHED state.
3553 * It is split into a fast path and a slow path. The fast path is
3555 * - A zero window was announced from us - zero window probing
3556 * is only handled properly in the slow path.
3557 * - Out of order segments arrived.
3558 * - Urgent data is expected.
3559 * - There is no buffer space left
3560 * - Unexpected TCP flags/window values/header lengths are received
3561 * (detected by checking the TCP header against pred_flags)
3562 * - Data is sent in both directions. Fast path only supports pure senders
3563 * or pure receivers (this means either the sequence number or the ack
3564 * value must stay constant)
3565 * - Unexpected TCP option.
3567 * When these conditions are not satisfied it drops into a standard
3568 * receive procedure patterned after RFC793 to handle all cases.
3569 * The first three cases are guaranteed by proper pred_flags setting,
3570 * the rest is checked inline. Fast processing is turned on in
3571 * tcp_data_queue when everything is OK.
3573 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
3574 struct tcphdr
*th
, unsigned len
)
3576 struct tcp_sock
*tp
= tcp_sk(sk
);
3579 * Header prediction.
3580 * The code loosely follows the one in the famous
3581 * "30 instruction TCP receive" Van Jacobson mail.
3583 * Van's trick is to deposit buffers into socket queue
3584 * on a device interrupt, to call tcp_recv function
3585 * on the receive process context and checksum and copy
3586 * the buffer to user space. smart...
3588 * Our current scheme is not silly either but we take the
3589 * extra cost of the net_bh soft interrupt processing...
3590 * We do checksum and copy also but from device to kernel.
3593 tp
->rx_opt
.saw_tstamp
= 0;
3595 /* pred_flags is 0xS?10 << 16 + snd_wnd
3596 * if header_predition is to be made
3597 * 'S' will always be tp->tcp_header_len >> 2
3598 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3599 * turn it off (when there are holes in the receive
3600 * space for instance)
3601 * PSH flag is ignored.
3604 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
3605 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3606 int tcp_header_len
= tp
->tcp_header_len
;
3608 /* Timestamp header prediction: tcp_header_len
3609 * is automatically equal to th->doff*4 due to pred_flags
3613 /* Check timestamp */
3614 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
3615 __u32
*ptr
= (__u32
*)(th
+ 1);
3617 /* No? Slow path! */
3618 if (*ptr
!= ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3619 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
3622 tp
->rx_opt
.saw_tstamp
= 1;
3624 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3626 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3628 /* If PAWS failed, check it more carefully in slow path */
3629 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
3632 /* DO NOT update ts_recent here, if checksum fails
3633 * and timestamp was corrupted part, it will result
3634 * in a hung connection since we will drop all
3635 * future packets due to the PAWS test.
3639 if (len
<= tcp_header_len
) {
3640 /* Bulk data transfer: sender */
3641 if (len
== tcp_header_len
) {
3642 /* Predicted packet is in window by definition.
3643 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3644 * Hence, check seq<=rcv_wup reduces to:
3646 if (tcp_header_len
==
3647 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3648 tp
->rcv_nxt
== tp
->rcv_wup
)
3649 tcp_store_ts_recent(tp
);
3651 tcp_rcv_rtt_measure_ts(tp
, skb
);
3653 /* We know that such packets are checksummed
3656 tcp_ack(sk
, skb
, 0);
3658 tcp_data_snd_check(sk
);
3660 } else { /* Header too small */
3661 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3667 if (tp
->ucopy
.task
== current
&&
3668 tp
->copied_seq
== tp
->rcv_nxt
&&
3669 len
- tcp_header_len
<= tp
->ucopy
.len
&&
3670 sock_owned_by_user(sk
)) {
3671 __set_current_state(TASK_RUNNING
);
3673 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
3674 /* Predicted packet is in window by definition.
3675 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3676 * Hence, check seq<=rcv_wup reduces to:
3678 if (tcp_header_len
==
3679 (sizeof(struct tcphdr
) +
3680 TCPOLEN_TSTAMP_ALIGNED
) &&
3681 tp
->rcv_nxt
== tp
->rcv_wup
)
3682 tcp_store_ts_recent(tp
);
3684 tcp_rcv_rtt_measure_ts(tp
, skb
);
3686 __skb_pull(skb
, tcp_header_len
);
3687 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3688 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
3693 if (tcp_checksum_complete_user(sk
, skb
))
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(tp
, skb
);
3707 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
3710 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
3712 /* Bulk data transfer: receiver */
3713 __skb_pull(skb
,tcp_header_len
);
3714 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3715 sk_stream_set_owner_r(skb
, sk
);
3716 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3719 tcp_event_data_recv(sk
, tp
, skb
);
3721 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
3722 /* Well, only one small jumplet in fast path... */
3723 tcp_ack(sk
, skb
, FLAG_DATA
);
3724 tcp_data_snd_check(sk
);
3725 if (!tcp_ack_scheduled(tp
))
3729 __tcp_ack_snd_check(sk
, 0);
3734 sk
->sk_data_ready(sk
, 0);
3740 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
3744 * RFC1323: H1. Apply PAWS check first.
3746 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
3747 tcp_paws_discard(tp
, skb
)) {
3749 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
3750 tcp_send_dupack(sk
, skb
);
3753 /* Resets are accepted even if PAWS failed.
3755 ts_recent update must be made after we are sure
3756 that the packet is in window.
3761 * Standard slow path.
3764 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3765 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3766 * (RST) segments are validated by checking their SEQ-fields."
3767 * And page 69: "If an incoming segment is not acceptable,
3768 * an acknowledgment should be sent in reply (unless the RST bit
3769 * is set, if so drop the segment and return)".
3772 tcp_send_dupack(sk
, skb
);
3781 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3783 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3784 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3785 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
3792 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
3794 tcp_rcv_rtt_measure_ts(tp
, skb
);
3796 /* Process urgent data. */
3797 tcp_urg(sk
, skb
, th
);
3799 /* step 7: process the segment text */
3800 tcp_data_queue(sk
, skb
);
3802 tcp_data_snd_check(sk
);
3803 tcp_ack_snd_check(sk
);
3807 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3814 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
3815 struct tcphdr
*th
, unsigned len
)
3817 struct tcp_sock
*tp
= tcp_sk(sk
);
3818 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
3820 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
3824 * "If the state is SYN-SENT then
3825 * first check the ACK bit
3826 * If the ACK bit is set
3827 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3828 * a reset (unless the RST bit is set, if so drop
3829 * the segment and return)"
3831 * We do not send data with SYN, so that RFC-correct
3834 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
3835 goto reset_and_undo
;
3837 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
3838 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
3840 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
3841 goto reset_and_undo
;
3844 /* Now ACK is acceptable.
3846 * "If the RST bit is set
3847 * If the ACK was acceptable then signal the user "error:
3848 * connection reset", drop the segment, enter CLOSED state,
3849 * delete TCB, and return."
3858 * "fifth, if neither of the SYN or RST bits is set then
3859 * drop the segment and return."
3865 goto discard_and_undo
;
3868 * "If the SYN bit is on ...
3869 * are acceptable then ...
3870 * (our SYN has been ACKed), change the connection
3871 * state to ESTABLISHED..."
3874 TCP_ECN_rcv_synack(tp
, th
);
3875 if (tp
->ecn_flags
&TCP_ECN_OK
)
3876 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
3878 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
3879 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
3881 /* Ok.. it's good. Set up sequence numbers and
3882 * move to established.
3884 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
3885 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
3887 /* RFC1323: The window in SYN & SYN/ACK segments is
3890 tp
->snd_wnd
= ntohs(th
->window
);
3891 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
3893 if (!tp
->rx_opt
.wscale_ok
) {
3894 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
3895 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
3898 if (tp
->rx_opt
.saw_tstamp
) {
3899 tp
->rx_opt
.tstamp_ok
= 1;
3900 tp
->tcp_header_len
=
3901 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
3902 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
3903 tcp_store_ts_recent(tp
);
3905 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3908 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
3909 tp
->rx_opt
.sack_ok
|= 2;
3911 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
3912 tcp_initialize_rcv_mss(sk
);
3914 /* Remember, tcp_poll() does not lock socket!
3915 * Change state from SYN-SENT only after copied_seq
3916 * is initialized. */
3917 tp
->copied_seq
= tp
->rcv_nxt
;
3919 tcp_set_state(sk
, TCP_ESTABLISHED
);
3921 /* Make sure socket is routed, for correct metrics. */
3922 tp
->af_specific
->rebuild_header(sk
);
3924 tcp_init_metrics(sk
);
3926 tcp_init_congestion_control(tp
);
3928 /* Prevent spurious tcp_cwnd_restart() on first data
3931 tp
->lsndtime
= tcp_time_stamp
;
3933 tcp_init_buffer_space(sk
);
3935 if (sock_flag(sk
, SOCK_KEEPOPEN
))
3936 tcp_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
3938 if (!tp
->rx_opt
.snd_wscale
)
3939 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
3943 if (!sock_flag(sk
, SOCK_DEAD
)) {
3944 sk
->sk_state_change(sk
);
3945 sk_wake_async(sk
, 0, POLL_OUT
);
3948 if (sk
->sk_write_pending
|| tp
->defer_accept
|| tp
->ack
.pingpong
) {
3949 /* Save one ACK. Data will be ready after
3950 * several ticks, if write_pending is set.
3952 * It may be deleted, but with this feature tcpdumps
3953 * look so _wonderfully_ clever, that I was not able
3954 * to stand against the temptation 8) --ANK
3956 tcp_schedule_ack(tp
);
3957 tp
->ack
.lrcvtime
= tcp_time_stamp
;
3958 tp
->ack
.ato
= TCP_ATO_MIN
;
3959 tcp_incr_quickack(tp
);
3960 tcp_enter_quickack_mode(tp
);
3961 tcp_reset_xmit_timer(sk
, TCP_TIME_DACK
, TCP_DELACK_MAX
);
3972 /* No ACK in the segment */
3976 * "If the RST bit is set
3978 * Otherwise (no ACK) drop the segment and return."
3981 goto discard_and_undo
;
3985 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
3986 goto discard_and_undo
;
3989 /* We see SYN without ACK. It is attempt of
3990 * simultaneous connect with crossed SYNs.
3991 * Particularly, it can be connect to self.
3993 tcp_set_state(sk
, TCP_SYN_RECV
);
3995 if (tp
->rx_opt
.saw_tstamp
) {
3996 tp
->rx_opt
.tstamp_ok
= 1;
3997 tcp_store_ts_recent(tp
);
3998 tp
->tcp_header_len
=
3999 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4001 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4004 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4005 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4007 /* RFC1323: The window in SYN & SYN/ACK segments is
4010 tp
->snd_wnd
= ntohs(th
->window
);
4011 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4012 tp
->max_window
= tp
->snd_wnd
;
4014 TCP_ECN_rcv_syn(tp
, th
);
4015 if (tp
->ecn_flags
&TCP_ECN_OK
)
4016 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
4018 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
4019 tcp_initialize_rcv_mss(sk
);
4022 tcp_send_synack(sk
);
4024 /* Note, we could accept data and URG from this segment.
4025 * There are no obstacles to make this.
4027 * However, if we ignore data in ACKless segments sometimes,
4028 * we have no reasons to accept it sometimes.
4029 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4030 * is not flawless. So, discard packet for sanity.
4031 * Uncomment this return to process the data.
4038 /* "fifth, if neither of the SYN or RST bits is set then
4039 * drop the segment and return."
4043 tcp_clear_options(&tp
->rx_opt
);
4044 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4048 tcp_clear_options(&tp
->rx_opt
);
4049 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4055 * This function implements the receiving procedure of RFC 793 for
4056 * all states except ESTABLISHED and TIME_WAIT.
4057 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4058 * address independent.
4061 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4062 struct tcphdr
*th
, unsigned len
)
4064 struct tcp_sock
*tp
= tcp_sk(sk
);
4067 tp
->rx_opt
.saw_tstamp
= 0;
4069 switch (sk
->sk_state
) {
4081 if(tp
->af_specific
->conn_request(sk
, skb
) < 0)
4084 /* Now we have several options: In theory there is
4085 * nothing else in the frame. KA9Q has an option to
4086 * send data with the syn, BSD accepts data with the
4087 * syn up to the [to be] advertised window and
4088 * Solaris 2.1 gives you a protocol error. For now
4089 * we just ignore it, that fits the spec precisely
4090 * and avoids incompatibilities. It would be nice in
4091 * future to drop through and process the data.
4093 * Now that TTCP is starting to be used we ought to
4095 * But, this leaves one open to an easy denial of
4096 * service attack, and SYN cookies can't defend
4097 * against this problem. So, we drop the data
4098 * in the interest of security over speed.
4105 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4109 /* Do step6 onward by hand. */
4110 tcp_urg(sk
, skb
, th
);
4112 tcp_data_snd_check(sk
);
4116 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4117 tcp_paws_discard(tp
, skb
)) {
4119 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4120 tcp_send_dupack(sk
, skb
);
4123 /* Reset is accepted even if it did not pass PAWS. */
4126 /* step 1: check sequence number */
4127 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4129 tcp_send_dupack(sk
, skb
);
4133 /* step 2: check RST bit */
4139 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4141 /* step 3: check security and precedence [ignored] */
4145 * Check for a SYN in window.
4147 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4148 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4153 /* step 5: check the ACK field */
4155 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4157 switch(sk
->sk_state
) {
4160 tp
->copied_seq
= tp
->rcv_nxt
;
4162 tcp_set_state(sk
, TCP_ESTABLISHED
);
4163 sk
->sk_state_change(sk
);
4165 /* Note, that this wakeup is only for marginal
4166 * crossed SYN case. Passively open sockets
4167 * are not waked up, because sk->sk_sleep ==
4168 * NULL and sk->sk_socket == NULL.
4170 if (sk
->sk_socket
) {
4171 sk_wake_async(sk
,0,POLL_OUT
);
4174 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4175 tp
->snd_wnd
= ntohs(th
->window
) <<
4176 tp
->rx_opt
.snd_wscale
;
4177 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4178 TCP_SKB_CB(skb
)->seq
);
4180 /* tcp_ack considers this ACK as duplicate
4181 * and does not calculate rtt.
4182 * Fix it at least with timestamps.
4184 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4186 tcp_ack_saw_tstamp(tp
, 0, 0);
4188 if (tp
->rx_opt
.tstamp_ok
)
4189 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4191 /* Make sure socket is routed, for
4194 tp
->af_specific
->rebuild_header(sk
);
4196 tcp_init_metrics(sk
);
4198 tcp_init_congestion_control(tp
);
4200 /* Prevent spurious tcp_cwnd_restart() on
4201 * first data packet.
4203 tp
->lsndtime
= tcp_time_stamp
;
4205 tcp_initialize_rcv_mss(sk
);
4206 tcp_init_buffer_space(sk
);
4207 tcp_fast_path_on(tp
);
4214 if (tp
->snd_una
== tp
->write_seq
) {
4215 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4216 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4217 dst_confirm(sk
->sk_dst_cache
);
4219 if (!sock_flag(sk
, SOCK_DEAD
))
4220 /* Wake up lingering close() */
4221 sk
->sk_state_change(sk
);
4225 if (tp
->linger2
< 0 ||
4226 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4227 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4229 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4233 tmo
= tcp_fin_time(tp
);
4234 if (tmo
> TCP_TIMEWAIT_LEN
) {
4235 tcp_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4236 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4237 /* Bad case. We could lose such FIN otherwise.
4238 * It is not a big problem, but it looks confusing
4239 * and not so rare event. We still can lose it now,
4240 * if it spins in bh_lock_sock(), but it is really
4243 tcp_reset_keepalive_timer(sk
, tmo
);
4245 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4253 if (tp
->snd_una
== tp
->write_seq
) {
4254 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4260 if (tp
->snd_una
== tp
->write_seq
) {
4261 tcp_update_metrics(sk
);
4270 /* step 6: check the URG bit */
4271 tcp_urg(sk
, skb
, th
);
4273 /* step 7: process the segment text */
4274 switch (sk
->sk_state
) {
4275 case TCP_CLOSE_WAIT
:
4278 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4282 /* RFC 793 says to queue data in these states,
4283 * RFC 1122 says we MUST send a reset.
4284 * BSD 4.4 also does reset.
4286 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4287 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4288 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4289 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4295 case TCP_ESTABLISHED
:
4296 tcp_data_queue(sk
, skb
);
4301 /* tcp_data could move socket to TIME-WAIT */
4302 if (sk
->sk_state
!= TCP_CLOSE
) {
4303 tcp_data_snd_check(sk
);
4304 tcp_ack_snd_check(sk
);
4314 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4315 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4316 EXPORT_SYMBOL(tcp_parse_options
);
4317 EXPORT_SYMBOL(tcp_rcv_established
);
4318 EXPORT_SYMBOL(tcp_rcv_state_process
);