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 presence 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;
92 int sysctl_tcp_abc
= 1;
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
110 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
111 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
113 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
115 /* Adapt the MSS value used to make delayed ack decision to the
118 static inline void tcp_measure_rcv_mss(struct sock
*sk
,
119 const struct sk_buff
*skb
)
121 struct inet_connection_sock
*icsk
= inet_csk(sk
);
122 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
125 icsk
->icsk_ack
.last_seg_size
= 0;
127 /* skb->len may jitter because of SACKs, even if peer
128 * sends good full-sized frames.
131 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
132 icsk
->icsk_ack
.rcv_mss
= len
;
134 /* Otherwise, we make more careful check taking into account,
135 * that SACKs block is variable.
137 * "len" is invariant segment length, including TCP header.
139 len
+= skb
->data
- skb
->h
.raw
;
140 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
141 /* If PSH is not set, packet should be
142 * full sized, provided peer TCP is not badly broken.
143 * This observation (if it is correct 8)) allows
144 * to handle super-low mtu links fairly.
146 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
147 !(tcp_flag_word(skb
->h
.th
)&TCP_REMNANT
))) {
148 /* Subtract also invariant (if peer is RFC compliant),
149 * tcp header plus fixed timestamp option length.
150 * Resulting "len" is MSS free of SACK jitter.
152 len
-= tcp_sk(sk
)->tcp_header_len
;
153 icsk
->icsk_ack
.last_seg_size
= len
;
155 icsk
->icsk_ack
.rcv_mss
= len
;
159 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
163 static void tcp_incr_quickack(struct sock
*sk
)
165 struct inet_connection_sock
*icsk
= inet_csk(sk
);
166 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
170 if (quickacks
> icsk
->icsk_ack
.quick
)
171 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
174 void tcp_enter_quickack_mode(struct sock
*sk
)
176 struct inet_connection_sock
*icsk
= inet_csk(sk
);
177 tcp_incr_quickack(sk
);
178 icsk
->icsk_ack
.pingpong
= 0;
179 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
182 /* Send ACKs quickly, if "quick" count is not exhausted
183 * and the session is not interactive.
186 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
188 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
189 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
192 /* Buffer size and advertised window tuning.
194 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
197 static void tcp_fixup_sndbuf(struct sock
*sk
)
199 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
200 sizeof(struct sk_buff
);
202 if (sk
->sk_sndbuf
< 3 * sndmem
)
203 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
206 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
208 * All tcp_full_space() is split to two parts: "network" buffer, allocated
209 * forward and advertised in receiver window (tp->rcv_wnd) and
210 * "application buffer", required to isolate scheduling/application
211 * latencies from network.
212 * window_clamp is maximal advertised window. It can be less than
213 * tcp_full_space(), in this case tcp_full_space() - window_clamp
214 * is reserved for "application" buffer. The less window_clamp is
215 * the smoother our behaviour from viewpoint of network, but the lower
216 * throughput and the higher sensitivity of the connection to losses. 8)
218 * rcv_ssthresh is more strict window_clamp used at "slow start"
219 * phase to predict further behaviour of this connection.
220 * It is used for two goals:
221 * - to enforce header prediction at sender, even when application
222 * requires some significant "application buffer". It is check #1.
223 * - to prevent pruning of receive queue because of misprediction
224 * of receiver window. Check #2.
226 * The scheme does not work when sender sends good segments opening
227 * window and then starts to feed us spaghetti. But it should work
228 * in common situations. Otherwise, we have to rely on queue collapsing.
231 /* Slow part of check#2. */
232 static int __tcp_grow_window(const struct sock
*sk
, struct tcp_sock
*tp
,
233 const struct sk_buff
*skb
)
236 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
237 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2])/2;
239 while (tp
->rcv_ssthresh
<= window
) {
240 if (truesize
<= skb
->len
)
241 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
249 static inline void tcp_grow_window(struct sock
*sk
, struct tcp_sock
*tp
,
253 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
254 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
255 !tcp_memory_pressure
) {
258 /* Check #2. Increase window, if skb with such overhead
259 * will fit to rcvbuf in future.
261 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
264 incr
= __tcp_grow_window(sk
, tp
, skb
);
267 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
268 inet_csk(sk
)->icsk_ack
.quick
|= 1;
273 /* 3. Tuning rcvbuf, when connection enters established state. */
275 static void tcp_fixup_rcvbuf(struct sock
*sk
)
277 struct tcp_sock
*tp
= tcp_sk(sk
);
278 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
280 /* Try to select rcvbuf so that 4 mss-sized segments
281 * will fit to window and corresponding skbs will fit to our rcvbuf.
282 * (was 3; 4 is minimum to allow fast retransmit to work.)
284 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
286 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
287 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
290 /* 4. Try to fixup all. It is made immediately after connection enters
293 static void tcp_init_buffer_space(struct sock
*sk
)
295 struct tcp_sock
*tp
= tcp_sk(sk
);
298 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
299 tcp_fixup_rcvbuf(sk
);
300 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
301 tcp_fixup_sndbuf(sk
);
303 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
305 maxwin
= tcp_full_space(sk
);
307 if (tp
->window_clamp
>= maxwin
) {
308 tp
->window_clamp
= maxwin
;
310 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
311 tp
->window_clamp
= max(maxwin
-
312 (maxwin
>> sysctl_tcp_app_win
),
316 /* Force reservation of one segment. */
317 if (sysctl_tcp_app_win
&&
318 tp
->window_clamp
> 2 * tp
->advmss
&&
319 tp
->window_clamp
+ tp
->advmss
> maxwin
)
320 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
322 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
323 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
326 /* 5. Recalculate window clamp after socket hit its memory bounds. */
327 static void tcp_clamp_window(struct sock
*sk
, struct tcp_sock
*tp
)
329 struct inet_connection_sock
*icsk
= inet_csk(sk
);
331 icsk
->icsk_ack
.quick
= 0;
333 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
334 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
335 !tcp_memory_pressure
&&
336 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
337 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
340 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
341 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
344 /* Receiver "autotuning" code.
346 * The algorithm for RTT estimation w/o timestamps is based on
347 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
348 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
350 * More detail on this code can be found at
351 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
352 * though this reference is out of date. A new paper
355 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
357 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
363 if (new_sample
!= 0) {
364 /* If we sample in larger samples in the non-timestamp
365 * case, we could grossly overestimate the RTT especially
366 * with chatty applications or bulk transfer apps which
367 * are stalled on filesystem I/O.
369 * Also, since we are only going for a minimum in the
370 * non-timestamp case, we do not smooth things out
371 * else with timestamps disabled convergence takes too
375 m
-= (new_sample
>> 3);
377 } else if (m
< new_sample
)
380 /* No previous measure. */
384 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
385 tp
->rcv_rtt_est
.rtt
= new_sample
;
388 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
390 if (tp
->rcv_rtt_est
.time
== 0)
392 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
394 tcp_rcv_rtt_update(tp
,
395 jiffies
- tp
->rcv_rtt_est
.time
,
399 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
400 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
403 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
405 struct tcp_sock
*tp
= tcp_sk(sk
);
406 if (tp
->rx_opt
.rcv_tsecr
&&
407 (TCP_SKB_CB(skb
)->end_seq
-
408 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
409 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
413 * This function should be called every time data is copied to user space.
414 * It calculates the appropriate TCP receive buffer space.
416 void tcp_rcv_space_adjust(struct sock
*sk
)
418 struct tcp_sock
*tp
= tcp_sk(sk
);
422 if (tp
->rcvq_space
.time
== 0)
425 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
426 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
427 tp
->rcv_rtt_est
.rtt
== 0)
430 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
432 space
= max(tp
->rcvq_space
.space
, space
);
434 if (tp
->rcvq_space
.space
!= space
) {
437 tp
->rcvq_space
.space
= space
;
439 if (sysctl_tcp_moderate_rcvbuf
) {
440 int new_clamp
= space
;
442 /* Receive space grows, normalize in order to
443 * take into account packet headers and sk_buff
444 * structure overhead.
449 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
450 16 + sizeof(struct sk_buff
));
451 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
454 space
= min(space
, sysctl_tcp_rmem
[2]);
455 if (space
> sk
->sk_rcvbuf
) {
456 sk
->sk_rcvbuf
= space
;
458 /* Make the window clamp follow along. */
459 tp
->window_clamp
= new_clamp
;
465 tp
->rcvq_space
.seq
= tp
->copied_seq
;
466 tp
->rcvq_space
.time
= tcp_time_stamp
;
469 /* There is something which you must keep in mind when you analyze the
470 * behavior of the tp->ato delayed ack timeout interval. When a
471 * connection starts up, we want to ack as quickly as possible. The
472 * problem is that "good" TCP's do slow start at the beginning of data
473 * transmission. The means that until we send the first few ACK's the
474 * sender will sit on his end and only queue most of his data, because
475 * he can only send snd_cwnd unacked packets at any given time. For
476 * each ACK we send, he increments snd_cwnd and transmits more of his
479 static void tcp_event_data_recv(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
481 struct inet_connection_sock
*icsk
= inet_csk(sk
);
484 inet_csk_schedule_ack(sk
);
486 tcp_measure_rcv_mss(sk
, skb
);
488 tcp_rcv_rtt_measure(tp
);
490 now
= tcp_time_stamp
;
492 if (!icsk
->icsk_ack
.ato
) {
493 /* The _first_ data packet received, initialize
494 * delayed ACK engine.
496 tcp_incr_quickack(sk
);
497 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
499 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
501 if (m
<= TCP_ATO_MIN
/2) {
502 /* The fastest case is the first. */
503 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
504 } else if (m
< icsk
->icsk_ack
.ato
) {
505 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
506 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
507 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
508 } else if (m
> icsk
->icsk_rto
) {
509 /* Too long gap. Apparently sender failed to
510 * restart window, so that we send ACKs quickly.
512 tcp_incr_quickack(sk
);
513 sk_stream_mem_reclaim(sk
);
516 icsk
->icsk_ack
.lrcvtime
= now
;
518 TCP_ECN_check_ce(tp
, skb
);
521 tcp_grow_window(sk
, tp
, skb
);
524 /* Called to compute a smoothed rtt estimate. The data fed to this
525 * routine either comes from timestamps, or from segments that were
526 * known _not_ to have been retransmitted [see Karn/Partridge
527 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
528 * piece by Van Jacobson.
529 * NOTE: the next three routines used to be one big routine.
530 * To save cycles in the RFC 1323 implementation it was better to break
531 * it up into three procedures. -- erics
533 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
535 struct tcp_sock
*tp
= tcp_sk(sk
);
536 long m
= mrtt
; /* RTT */
538 /* The following amusing code comes from Jacobson's
539 * article in SIGCOMM '88. Note that rtt and mdev
540 * are scaled versions of rtt and mean deviation.
541 * This is designed to be as fast as possible
542 * m stands for "measurement".
544 * On a 1990 paper the rto value is changed to:
545 * RTO = rtt + 4 * mdev
547 * Funny. This algorithm seems to be very broken.
548 * These formulae increase RTO, when it should be decreased, increase
549 * too slowly, when it should be increased quickly, decrease too quickly
550 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
551 * does not matter how to _calculate_ it. Seems, it was trap
552 * that VJ failed to avoid. 8)
557 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
558 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
560 m
= -m
; /* m is now abs(error) */
561 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
562 /* This is similar to one of Eifel findings.
563 * Eifel blocks mdev updates when rtt decreases.
564 * This solution is a bit different: we use finer gain
565 * for mdev in this case (alpha*beta).
566 * Like Eifel it also prevents growth of rto,
567 * but also it limits too fast rto decreases,
568 * happening in pure Eifel.
573 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
575 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
576 if (tp
->mdev
> tp
->mdev_max
) {
577 tp
->mdev_max
= tp
->mdev
;
578 if (tp
->mdev_max
> tp
->rttvar
)
579 tp
->rttvar
= tp
->mdev_max
;
581 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
582 if (tp
->mdev_max
< tp
->rttvar
)
583 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
584 tp
->rtt_seq
= tp
->snd_nxt
;
585 tp
->mdev_max
= TCP_RTO_MIN
;
588 /* no previous measure. */
589 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
590 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
591 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
592 tp
->rtt_seq
= tp
->snd_nxt
;
596 /* Calculate rto without backoff. This is the second half of Van Jacobson's
597 * routine referred to above.
599 static inline void tcp_set_rto(struct sock
*sk
)
601 const struct tcp_sock
*tp
= tcp_sk(sk
);
602 /* Old crap is replaced with new one. 8)
605 * 1. If rtt variance happened to be less 50msec, it is hallucination.
606 * It cannot be less due to utterly erratic ACK generation made
607 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
608 * to do with delayed acks, because at cwnd>2 true delack timeout
609 * is invisible. Actually, Linux-2.4 also generates erratic
610 * ACKs in some circumstances.
612 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
614 /* 2. Fixups made earlier cannot be right.
615 * If we do not estimate RTO correctly without them,
616 * all the algo is pure shit and should be replaced
617 * with correct one. It is exactly, which we pretend to do.
621 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
622 * guarantees that rto is higher.
624 static inline void tcp_bound_rto(struct sock
*sk
)
626 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
627 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
630 /* Save metrics learned by this TCP session.
631 This function is called only, when TCP finishes successfully
632 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
634 void tcp_update_metrics(struct sock
*sk
)
636 struct tcp_sock
*tp
= tcp_sk(sk
);
637 struct dst_entry
*dst
= __sk_dst_get(sk
);
639 if (sysctl_tcp_nometrics_save
)
644 if (dst
&& (dst
->flags
&DST_HOST
)) {
645 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
648 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
649 /* This session failed to estimate rtt. Why?
650 * Probably, no packets returned in time.
653 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
654 dst
->metrics
[RTAX_RTT
-1] = 0;
658 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
660 /* If newly calculated rtt larger than stored one,
661 * store new one. Otherwise, use EWMA. Remember,
662 * rtt overestimation is always better than underestimation.
664 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
666 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
668 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
671 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
675 /* Scale deviation to rttvar fixed point */
680 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
681 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
683 dst
->metrics
[RTAX_RTTVAR
-1] -=
684 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
687 if (tp
->snd_ssthresh
>= 0xFFFF) {
688 /* Slow start still did not finish. */
689 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
690 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
691 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
692 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
693 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
694 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
695 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
696 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
697 icsk
->icsk_ca_state
== TCP_CA_Open
) {
698 /* Cong. avoidance phase, cwnd is reliable. */
699 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
700 dst
->metrics
[RTAX_SSTHRESH
-1] =
701 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
702 if (!dst_metric_locked(dst
, RTAX_CWND
))
703 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
705 /* Else slow start did not finish, cwnd is non-sense,
706 ssthresh may be also invalid.
708 if (!dst_metric_locked(dst
, RTAX_CWND
))
709 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
710 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
711 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
712 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
713 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
716 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
717 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
718 tp
->reordering
!= sysctl_tcp_reordering
)
719 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
724 /* Numbers are taken from RFC2414. */
725 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
727 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
730 if (tp
->mss_cache
> 1460)
733 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
735 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
738 /* Initialize metrics on socket. */
740 static void tcp_init_metrics(struct sock
*sk
)
742 struct tcp_sock
*tp
= tcp_sk(sk
);
743 struct dst_entry
*dst
= __sk_dst_get(sk
);
750 if (dst_metric_locked(dst
, RTAX_CWND
))
751 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
752 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
753 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
754 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
755 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
757 if (dst_metric(dst
, RTAX_REORDERING
) &&
758 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
759 tp
->rx_opt
.sack_ok
&= ~2;
760 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
763 if (dst_metric(dst
, RTAX_RTT
) == 0)
766 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
769 /* Initial rtt is determined from SYN,SYN-ACK.
770 * The segment is small and rtt may appear much
771 * less than real one. Use per-dst memory
772 * to make it more realistic.
774 * A bit of theory. RTT is time passed after "normal" sized packet
775 * is sent until it is ACKed. In normal circumstances sending small
776 * packets force peer to delay ACKs and calculation is correct too.
777 * The algorithm is adaptive and, provided we follow specs, it
778 * NEVER underestimate RTT. BUT! If peer tries to make some clever
779 * tricks sort of "quick acks" for time long enough to decrease RTT
780 * to low value, and then abruptly stops to do it and starts to delay
781 * ACKs, wait for troubles.
783 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
784 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
785 tp
->rtt_seq
= tp
->snd_nxt
;
787 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
788 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
789 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
793 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
795 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
796 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
800 /* Play conservative. If timestamps are not
801 * supported, TCP will fail to recalculate correct
802 * rtt, if initial rto is too small. FORGET ALL AND RESET!
804 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
806 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
807 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
811 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
814 struct tcp_sock
*tp
= tcp_sk(sk
);
815 if (metric
> tp
->reordering
) {
816 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
818 /* This exciting event is worth to be remembered. 8) */
820 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
822 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
824 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
826 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
827 #if FASTRETRANS_DEBUG > 1
828 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
829 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
833 tp
->undo_marker
? tp
->undo_retrans
: 0);
835 /* Disable FACK yet. */
836 tp
->rx_opt
.sack_ok
&= ~2;
840 /* This procedure tags the retransmission queue when SACKs arrive.
842 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
843 * Packets in queue with these bits set are counted in variables
844 * sacked_out, retrans_out and lost_out, correspondingly.
846 * Valid combinations are:
847 * Tag InFlight Description
848 * 0 1 - orig segment is in flight.
849 * S 0 - nothing flies, orig reached receiver.
850 * L 0 - nothing flies, orig lost by net.
851 * R 2 - both orig and retransmit are in flight.
852 * L|R 1 - orig is lost, retransmit is in flight.
853 * S|R 1 - orig reached receiver, retrans is still in flight.
854 * (L|S|R is logically valid, it could occur when L|R is sacked,
855 * but it is equivalent to plain S and code short-curcuits it to S.
856 * L|S is logically invalid, it would mean -1 packet in flight 8))
858 * These 6 states form finite state machine, controlled by the following events:
859 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
860 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
861 * 3. Loss detection event of one of three flavors:
862 * A. Scoreboard estimator decided the packet is lost.
863 * A'. Reno "three dupacks" marks head of queue lost.
864 * A''. Its FACK modfication, head until snd.fack is lost.
865 * B. SACK arrives sacking data transmitted after never retransmitted
867 * C. SACK arrives sacking SND.NXT at the moment, when the
868 * segment was retransmitted.
869 * 4. D-SACK added new rule: D-SACK changes any tag to S.
871 * It is pleasant to note, that state diagram turns out to be commutative,
872 * so that we are allowed not to be bothered by order of our actions,
873 * when multiple events arrive simultaneously. (see the function below).
875 * Reordering detection.
876 * --------------------
877 * Reordering metric is maximal distance, which a packet can be displaced
878 * in packet stream. With SACKs we can estimate it:
880 * 1. SACK fills old hole and the corresponding segment was not
881 * ever retransmitted -> reordering. Alas, we cannot use it
882 * when segment was retransmitted.
883 * 2. The last flaw is solved with D-SACK. D-SACK arrives
884 * for retransmitted and already SACKed segment -> reordering..
885 * Both of these heuristics are not used in Loss state, when we cannot
886 * account for retransmits accurately.
889 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
891 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
892 struct tcp_sock
*tp
= tcp_sk(sk
);
893 unsigned char *ptr
= ack_skb
->h
.raw
+ TCP_SKB_CB(ack_skb
)->sacked
;
894 struct tcp_sack_block
*sp
= (struct tcp_sack_block
*)(ptr
+2);
895 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
896 int reord
= tp
->packets_out
;
898 u32 lost_retrans
= 0;
905 prior_fackets
= tp
->fackets_out
;
908 * if the only SACK change is the increase of the end_seq of
909 * the first block then only apply that SACK block
910 * and use retrans queue hinting otherwise slowpath */
912 for (i
= 0; i
< num_sacks
; i
++) {
913 __u32 start_seq
= ntohl(sp
[i
].start_seq
);
914 __u32 end_seq
= ntohl(sp
[i
].end_seq
);
917 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
920 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
921 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
924 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
925 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
927 /* Check for D-SACK. */
929 u32 ack
= TCP_SKB_CB(ack_skb
)->ack_seq
;
931 if (before(start_seq
, ack
)) {
933 tp
->rx_opt
.sack_ok
|= 4;
934 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
935 } else if (num_sacks
> 1 &&
936 !after(end_seq
, ntohl(sp
[1].end_seq
)) &&
937 !before(start_seq
, ntohl(sp
[1].start_seq
))) {
939 tp
->rx_opt
.sack_ok
|= 4;
940 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
943 /* D-SACK for already forgotten data...
944 * Do dumb counting. */
946 !after(end_seq
, prior_snd_una
) &&
947 after(end_seq
, tp
->undo_marker
))
950 /* Eliminate too old ACKs, but take into
951 * account more or less fresh ones, they can
952 * contain valid SACK info.
954 if (before(ack
, prior_snd_una
- tp
->max_window
))
963 tp
->fastpath_skb_hint
= NULL
;
965 /* order SACK blocks to allow in order walk of the retrans queue */
966 for (i
= num_sacks
-1; i
> 0; i
--) {
967 for (j
= 0; j
< i
; j
++){
968 if (after(ntohl(sp
[j
].start_seq
),
969 ntohl(sp
[j
+1].start_seq
))){
970 sp
[j
].start_seq
= htonl(tp
->recv_sack_cache
[j
+1].start_seq
);
971 sp
[j
].end_seq
= htonl(tp
->recv_sack_cache
[j
+1].end_seq
);
972 sp
[j
+1].start_seq
= htonl(tp
->recv_sack_cache
[j
].start_seq
);
973 sp
[j
+1].end_seq
= htonl(tp
->recv_sack_cache
[j
].end_seq
);
980 /* clear flag as used for different purpose in following code */
983 for (i
=0; i
<num_sacks
; i
++, sp
++) {
985 __u32 start_seq
= ntohl(sp
->start_seq
);
986 __u32 end_seq
= ntohl(sp
->end_seq
);
989 /* Use SACK fastpath hint if valid */
990 if (tp
->fastpath_skb_hint
) {
991 skb
= tp
->fastpath_skb_hint
;
992 fack_count
= tp
->fastpath_cnt_hint
;
994 skb
= sk
->sk_write_queue
.next
;
998 /* Event "B" in the comment above. */
999 if (after(end_seq
, tp
->high_seq
))
1000 flag
|= FLAG_DATA_LOST
;
1002 sk_stream_for_retrans_queue_from(skb
, sk
) {
1003 int in_sack
, pcount
;
1006 tp
->fastpath_skb_hint
= skb
;
1007 tp
->fastpath_cnt_hint
= fack_count
;
1009 /* The retransmission queue is always in order, so
1010 * we can short-circuit the walk early.
1012 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1015 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1016 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1018 pcount
= tcp_skb_pcount(skb
);
1020 if (pcount
> 1 && !in_sack
&&
1021 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1022 unsigned int pkt_len
;
1024 in_sack
= !after(start_seq
,
1025 TCP_SKB_CB(skb
)->seq
);
1028 pkt_len
= (start_seq
-
1029 TCP_SKB_CB(skb
)->seq
);
1031 pkt_len
= (end_seq
-
1032 TCP_SKB_CB(skb
)->seq
);
1033 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->tso_size
))
1035 pcount
= tcp_skb_pcount(skb
);
1038 fack_count
+= pcount
;
1040 sacked
= TCP_SKB_CB(skb
)->sacked
;
1042 /* Account D-SACK for retransmitted packet. */
1043 if ((dup_sack
&& in_sack
) &&
1044 (sacked
& TCPCB_RETRANS
) &&
1045 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1048 /* The frame is ACKed. */
1049 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1050 if (sacked
&TCPCB_RETRANS
) {
1051 if ((dup_sack
&& in_sack
) &&
1052 (sacked
&TCPCB_SACKED_ACKED
))
1053 reord
= min(fack_count
, reord
);
1055 /* If it was in a hole, we detected reordering. */
1056 if (fack_count
< prior_fackets
&&
1057 !(sacked
&TCPCB_SACKED_ACKED
))
1058 reord
= min(fack_count
, reord
);
1061 /* Nothing to do; acked frame is about to be dropped. */
1065 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1066 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1067 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1068 lost_retrans
= end_seq
;
1073 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1074 if (sacked
& TCPCB_SACKED_RETRANS
) {
1075 /* If the segment is not tagged as lost,
1076 * we do not clear RETRANS, believing
1077 * that retransmission is still in flight.
1079 if (sacked
& TCPCB_LOST
) {
1080 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1081 tp
->lost_out
-= tcp_skb_pcount(skb
);
1082 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1084 /* clear lost hint */
1085 tp
->retransmit_skb_hint
= NULL
;
1088 /* New sack for not retransmitted frame,
1089 * which was in hole. It is reordering.
1091 if (!(sacked
& TCPCB_RETRANS
) &&
1092 fack_count
< prior_fackets
)
1093 reord
= min(fack_count
, reord
);
1095 if (sacked
& TCPCB_LOST
) {
1096 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1097 tp
->lost_out
-= tcp_skb_pcount(skb
);
1099 /* clear lost hint */
1100 tp
->retransmit_skb_hint
= NULL
;
1104 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1105 flag
|= FLAG_DATA_SACKED
;
1106 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1108 if (fack_count
> tp
->fackets_out
)
1109 tp
->fackets_out
= fack_count
;
1111 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1112 reord
= min(fack_count
, reord
);
1115 /* D-SACK. We can detect redundant retransmission
1116 * in S|R and plain R frames and clear it.
1117 * undo_retrans is decreased above, L|R frames
1118 * are accounted above as well.
1121 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1122 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1123 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1124 tp
->retransmit_skb_hint
= NULL
;
1129 /* Check for lost retransmit. This superb idea is
1130 * borrowed from "ratehalving". Event "C".
1131 * Later note: FACK people cheated me again 8),
1132 * we have to account for reordering! Ugly,
1135 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1136 struct sk_buff
*skb
;
1138 sk_stream_for_retrans_queue(skb
, sk
) {
1139 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1141 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1143 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1144 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1146 !before(lost_retrans
,
1147 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1149 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1150 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1152 /* clear lost hint */
1153 tp
->retransmit_skb_hint
= NULL
;
1155 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1156 tp
->lost_out
+= tcp_skb_pcount(skb
);
1157 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1158 flag
|= FLAG_DATA_SACKED
;
1159 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1165 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1167 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1168 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1170 #if FASTRETRANS_DEBUG > 0
1171 BUG_TRAP((int)tp
->sacked_out
>= 0);
1172 BUG_TRAP((int)tp
->lost_out
>= 0);
1173 BUG_TRAP((int)tp
->retrans_out
>= 0);
1174 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1179 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1180 * segments to see from the next ACKs whether any data was really missing.
1181 * If the RTO was spurious, new ACKs should arrive.
1183 void tcp_enter_frto(struct sock
*sk
)
1185 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1186 struct tcp_sock
*tp
= tcp_sk(sk
);
1187 struct sk_buff
*skb
;
1189 tp
->frto_counter
= 1;
1191 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1192 tp
->snd_una
== tp
->high_seq
||
1193 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1194 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1195 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1196 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1199 /* Have to clear retransmission markers here to keep the bookkeeping
1200 * in shape, even though we are not yet in Loss state.
1201 * If something was really lost, it is eventually caught up
1202 * in tcp_enter_frto_loss.
1204 tp
->retrans_out
= 0;
1205 tp
->undo_marker
= tp
->snd_una
;
1206 tp
->undo_retrans
= 0;
1208 sk_stream_for_retrans_queue(skb
, sk
) {
1209 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_RETRANS
;
1211 tcp_sync_left_out(tp
);
1213 tcp_set_ca_state(sk
, TCP_CA_Open
);
1214 tp
->frto_highmark
= tp
->snd_nxt
;
1217 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1218 * which indicates that we should follow the traditional RTO recovery,
1219 * i.e. mark everything lost and do go-back-N retransmission.
1221 static void tcp_enter_frto_loss(struct sock
*sk
)
1223 struct tcp_sock
*tp
= tcp_sk(sk
);
1224 struct sk_buff
*skb
;
1229 tp
->fackets_out
= 0;
1231 sk_stream_for_retrans_queue(skb
, sk
) {
1232 cnt
+= tcp_skb_pcount(skb
);
1233 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1234 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1236 /* Do not mark those segments lost that were
1237 * forward transmitted after RTO
1239 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1240 tp
->frto_highmark
)) {
1241 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1242 tp
->lost_out
+= tcp_skb_pcount(skb
);
1245 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1246 tp
->fackets_out
= cnt
;
1249 tcp_sync_left_out(tp
);
1251 tp
->snd_cwnd
= tp
->frto_counter
+ tcp_packets_in_flight(tp
)+1;
1252 tp
->snd_cwnd_cnt
= 0;
1253 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1254 tp
->undo_marker
= 0;
1255 tp
->frto_counter
= 0;
1257 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1258 sysctl_tcp_reordering
);
1259 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1260 tp
->high_seq
= tp
->frto_highmark
;
1261 TCP_ECN_queue_cwr(tp
);
1263 clear_all_retrans_hints(tp
);
1266 void tcp_clear_retrans(struct tcp_sock
*tp
)
1269 tp
->retrans_out
= 0;
1271 tp
->fackets_out
= 0;
1275 tp
->undo_marker
= 0;
1276 tp
->undo_retrans
= 0;
1279 /* Enter Loss state. If "how" is not zero, forget all SACK information
1280 * and reset tags completely, otherwise preserve SACKs. If receiver
1281 * dropped its ofo queue, we will know this due to reneging detection.
1283 void tcp_enter_loss(struct sock
*sk
, int how
)
1285 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1286 struct tcp_sock
*tp
= tcp_sk(sk
);
1287 struct sk_buff
*skb
;
1290 /* Reduce ssthresh if it has not yet been made inside this window. */
1291 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1292 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1293 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1294 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1295 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1298 tp
->snd_cwnd_cnt
= 0;
1299 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1301 tp
->bytes_acked
= 0;
1302 tcp_clear_retrans(tp
);
1304 /* Push undo marker, if it was plain RTO and nothing
1305 * was retransmitted. */
1307 tp
->undo_marker
= tp
->snd_una
;
1309 sk_stream_for_retrans_queue(skb
, sk
) {
1310 cnt
+= tcp_skb_pcount(skb
);
1311 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1312 tp
->undo_marker
= 0;
1313 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1314 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1315 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1316 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1317 tp
->lost_out
+= tcp_skb_pcount(skb
);
1319 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1320 tp
->fackets_out
= cnt
;
1323 tcp_sync_left_out(tp
);
1325 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1326 sysctl_tcp_reordering
);
1327 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1328 tp
->high_seq
= tp
->snd_nxt
;
1329 TCP_ECN_queue_cwr(tp
);
1331 clear_all_retrans_hints(tp
);
1334 static int tcp_check_sack_reneging(struct sock
*sk
)
1336 struct sk_buff
*skb
;
1338 /* If ACK arrived pointing to a remembered SACK,
1339 * it means that our remembered SACKs do not reflect
1340 * real state of receiver i.e.
1341 * receiver _host_ is heavily congested (or buggy).
1342 * Do processing similar to RTO timeout.
1344 if ((skb
= skb_peek(&sk
->sk_write_queue
)) != NULL
&&
1345 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1346 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1347 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1349 tcp_enter_loss(sk
, 1);
1350 icsk
->icsk_retransmits
++;
1351 tcp_retransmit_skb(sk
, skb_peek(&sk
->sk_write_queue
));
1352 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1353 icsk
->icsk_rto
, TCP_RTO_MAX
);
1359 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1361 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1364 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1366 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1369 static inline int tcp_head_timedout(struct sock
*sk
, struct tcp_sock
*tp
)
1371 return tp
->packets_out
&&
1372 tcp_skb_timedout(sk
, skb_peek(&sk
->sk_write_queue
));
1375 /* Linux NewReno/SACK/FACK/ECN state machine.
1376 * --------------------------------------
1378 * "Open" Normal state, no dubious events, fast path.
1379 * "Disorder" In all the respects it is "Open",
1380 * but requires a bit more attention. It is entered when
1381 * we see some SACKs or dupacks. It is split of "Open"
1382 * mainly to move some processing from fast path to slow one.
1383 * "CWR" CWND was reduced due to some Congestion Notification event.
1384 * It can be ECN, ICMP source quench, local device congestion.
1385 * "Recovery" CWND was reduced, we are fast-retransmitting.
1386 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1388 * tcp_fastretrans_alert() is entered:
1389 * - each incoming ACK, if state is not "Open"
1390 * - when arrived ACK is unusual, namely:
1395 * Counting packets in flight is pretty simple.
1397 * in_flight = packets_out - left_out + retrans_out
1399 * packets_out is SND.NXT-SND.UNA counted in packets.
1401 * retrans_out is number of retransmitted segments.
1403 * left_out is number of segments left network, but not ACKed yet.
1405 * left_out = sacked_out + lost_out
1407 * sacked_out: Packets, which arrived to receiver out of order
1408 * and hence not ACKed. With SACKs this number is simply
1409 * amount of SACKed data. Even without SACKs
1410 * it is easy to give pretty reliable estimate of this number,
1411 * counting duplicate ACKs.
1413 * lost_out: Packets lost by network. TCP has no explicit
1414 * "loss notification" feedback from network (for now).
1415 * It means that this number can be only _guessed_.
1416 * Actually, it is the heuristics to predict lossage that
1417 * distinguishes different algorithms.
1419 * F.e. after RTO, when all the queue is considered as lost,
1420 * lost_out = packets_out and in_flight = retrans_out.
1422 * Essentially, we have now two algorithms counting
1425 * FACK: It is the simplest heuristics. As soon as we decided
1426 * that something is lost, we decide that _all_ not SACKed
1427 * packets until the most forward SACK are lost. I.e.
1428 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1429 * It is absolutely correct estimate, if network does not reorder
1430 * packets. And it loses any connection to reality when reordering
1431 * takes place. We use FACK by default until reordering
1432 * is suspected on the path to this destination.
1434 * NewReno: when Recovery is entered, we assume that one segment
1435 * is lost (classic Reno). While we are in Recovery and
1436 * a partial ACK arrives, we assume that one more packet
1437 * is lost (NewReno). This heuristics are the same in NewReno
1440 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1441 * deflation etc. CWND is real congestion window, never inflated, changes
1442 * only according to classic VJ rules.
1444 * Really tricky (and requiring careful tuning) part of algorithm
1445 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1446 * The first determines the moment _when_ we should reduce CWND and,
1447 * hence, slow down forward transmission. In fact, it determines the moment
1448 * when we decide that hole is caused by loss, rather than by a reorder.
1450 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1451 * holes, caused by lost packets.
1453 * And the most logically complicated part of algorithm is undo
1454 * heuristics. We detect false retransmits due to both too early
1455 * fast retransmit (reordering) and underestimated RTO, analyzing
1456 * timestamps and D-SACKs. When we detect that some segments were
1457 * retransmitted by mistake and CWND reduction was wrong, we undo
1458 * window reduction and abort recovery phase. This logic is hidden
1459 * inside several functions named tcp_try_undo_<something>.
1462 /* This function decides, when we should leave Disordered state
1463 * and enter Recovery phase, reducing congestion window.
1465 * Main question: may we further continue forward transmission
1466 * with the same cwnd?
1468 static int tcp_time_to_recover(struct sock
*sk
, struct tcp_sock
*tp
)
1472 /* Trick#1: The loss is proven. */
1476 /* Not-A-Trick#2 : Classic rule... */
1477 if (tcp_fackets_out(tp
) > tp
->reordering
)
1480 /* Trick#3 : when we use RFC2988 timer restart, fast
1481 * retransmit can be triggered by timeout of queue head.
1483 if (tcp_head_timedout(sk
, tp
))
1486 /* Trick#4: It is still not OK... But will it be useful to delay
1489 packets_out
= tp
->packets_out
;
1490 if (packets_out
<= tp
->reordering
&&
1491 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1492 !tcp_may_send_now(sk
, tp
)) {
1493 /* We have nothing to send. This connection is limited
1494 * either by receiver window or by application.
1502 /* If we receive more dupacks than we expected counting segments
1503 * in assumption of absent reordering, interpret this as reordering.
1504 * The only another reason could be bug in receiver TCP.
1506 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1508 struct tcp_sock
*tp
= tcp_sk(sk
);
1511 holes
= max(tp
->lost_out
, 1U);
1512 holes
= min(holes
, tp
->packets_out
);
1514 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1515 tp
->sacked_out
= tp
->packets_out
- holes
;
1516 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1520 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1522 static void tcp_add_reno_sack(struct sock
*sk
)
1524 struct tcp_sock
*tp
= tcp_sk(sk
);
1526 tcp_check_reno_reordering(sk
, 0);
1527 tcp_sync_left_out(tp
);
1530 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1532 static void tcp_remove_reno_sacks(struct sock
*sk
, struct tcp_sock
*tp
, int acked
)
1535 /* One ACK acked hole. The rest eat duplicate ACKs. */
1536 if (acked
-1 >= tp
->sacked_out
)
1539 tp
->sacked_out
-= acked
-1;
1541 tcp_check_reno_reordering(sk
, acked
);
1542 tcp_sync_left_out(tp
);
1545 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1548 tp
->left_out
= tp
->lost_out
;
1551 /* Mark head of queue up as lost. */
1552 static void tcp_mark_head_lost(struct sock
*sk
, struct tcp_sock
*tp
,
1553 int packets
, u32 high_seq
)
1555 struct sk_buff
*skb
;
1558 BUG_TRAP(packets
<= tp
->packets_out
);
1559 if (tp
->lost_skb_hint
) {
1560 skb
= tp
->lost_skb_hint
;
1561 cnt
= tp
->lost_cnt_hint
;
1563 skb
= sk
->sk_write_queue
.next
;
1567 sk_stream_for_retrans_queue_from(skb
, sk
) {
1568 /* TODO: do this better */
1569 /* this is not the most efficient way to do this... */
1570 tp
->lost_skb_hint
= skb
;
1571 tp
->lost_cnt_hint
= cnt
;
1572 cnt
+= tcp_skb_pcount(skb
);
1573 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1575 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1576 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1577 tp
->lost_out
+= tcp_skb_pcount(skb
);
1579 /* clear xmit_retransmit_queue hints
1580 * if this is beyond hint */
1581 if(tp
->retransmit_skb_hint
!= NULL
&&
1582 before(TCP_SKB_CB(skb
)->seq
,
1583 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
)) {
1585 tp
->retransmit_skb_hint
= NULL
;
1589 tcp_sync_left_out(tp
);
1592 /* Account newly detected lost packet(s) */
1594 static void tcp_update_scoreboard(struct sock
*sk
, struct tcp_sock
*tp
)
1597 int lost
= tp
->fackets_out
- tp
->reordering
;
1600 tcp_mark_head_lost(sk
, tp
, lost
, tp
->high_seq
);
1602 tcp_mark_head_lost(sk
, tp
, 1, tp
->high_seq
);
1605 /* New heuristics: it is possible only after we switched
1606 * to restart timer each time when something is ACKed.
1607 * Hence, we can detect timed out packets during fast
1608 * retransmit without falling to slow start.
1610 if (tcp_head_timedout(sk
, tp
)) {
1611 struct sk_buff
*skb
;
1613 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1614 : sk
->sk_write_queue
.next
;
1616 sk_stream_for_retrans_queue_from(skb
, sk
) {
1617 if (!tcp_skb_timedout(sk
, skb
))
1620 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1621 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1622 tp
->lost_out
+= tcp_skb_pcount(skb
);
1624 /* clear xmit_retrans hint */
1625 if (tp
->retransmit_skb_hint
&&
1626 before(TCP_SKB_CB(skb
)->seq
,
1627 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1629 tp
->retransmit_skb_hint
= NULL
;
1633 tp
->scoreboard_skb_hint
= skb
;
1635 tcp_sync_left_out(tp
);
1639 /* CWND moderation, preventing bursts due to too big ACKs
1640 * in dubious situations.
1642 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1644 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1645 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1646 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1649 /* Decrease cwnd each second ack. */
1650 static void tcp_cwnd_down(struct sock
*sk
)
1652 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1653 struct tcp_sock
*tp
= tcp_sk(sk
);
1654 int decr
= tp
->snd_cwnd_cnt
+ 1;
1656 tp
->snd_cwnd_cnt
= decr
&1;
1659 if (decr
&& tp
->snd_cwnd
> icsk
->icsk_ca_ops
->min_cwnd(sk
))
1660 tp
->snd_cwnd
-= decr
;
1662 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1663 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1666 /* Nothing was retransmitted or returned timestamp is less
1667 * than timestamp of the first retransmission.
1669 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1671 return !tp
->retrans_stamp
||
1672 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1673 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1676 /* Undo procedures. */
1678 #if FASTRETRANS_DEBUG > 1
1679 static void DBGUNDO(struct sock
*sk
, struct tcp_sock
*tp
, const char *msg
)
1681 struct inet_sock
*inet
= inet_sk(sk
);
1682 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1684 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1685 tp
->snd_cwnd
, tp
->left_out
,
1686 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1690 #define DBGUNDO(x...) do { } while (0)
1693 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1695 struct tcp_sock
*tp
= tcp_sk(sk
);
1697 if (tp
->prior_ssthresh
) {
1698 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1700 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1701 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1703 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1705 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1706 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1707 TCP_ECN_withdraw_cwr(tp
);
1710 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1712 tcp_moderate_cwnd(tp
);
1713 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1715 /* There is something screwy going on with the retrans hints after
1717 clear_all_retrans_hints(tp
);
1720 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1722 return tp
->undo_marker
&&
1723 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1726 /* People celebrate: "We love our President!" */
1727 static int tcp_try_undo_recovery(struct sock
*sk
, struct tcp_sock
*tp
)
1729 if (tcp_may_undo(tp
)) {
1730 /* Happy end! We did not retransmit anything
1731 * or our original transmission succeeded.
1733 DBGUNDO(sk
, tp
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1734 tcp_undo_cwr(sk
, 1);
1735 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1736 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1738 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1739 tp
->undo_marker
= 0;
1741 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1742 /* Hold old state until something *above* high_seq
1743 * is ACKed. For Reno it is MUST to prevent false
1744 * fast retransmits (RFC2582). SACK TCP is safe. */
1745 tcp_moderate_cwnd(tp
);
1748 tcp_set_ca_state(sk
, TCP_CA_Open
);
1752 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1753 static void tcp_try_undo_dsack(struct sock
*sk
, struct tcp_sock
*tp
)
1755 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1756 DBGUNDO(sk
, tp
, "D-SACK");
1757 tcp_undo_cwr(sk
, 1);
1758 tp
->undo_marker
= 0;
1759 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1763 /* Undo during fast recovery after partial ACK. */
1765 static int tcp_try_undo_partial(struct sock
*sk
, struct tcp_sock
*tp
,
1768 /* Partial ACK arrived. Force Hoe's retransmit. */
1769 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1771 if (tcp_may_undo(tp
)) {
1772 /* Plain luck! Hole if filled with delayed
1773 * packet, rather than with a retransmit.
1775 if (tp
->retrans_out
== 0)
1776 tp
->retrans_stamp
= 0;
1778 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
1780 DBGUNDO(sk
, tp
, "Hoe");
1781 tcp_undo_cwr(sk
, 0);
1782 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1784 /* So... Do not make Hoe's retransmit yet.
1785 * If the first packet was delayed, the rest
1786 * ones are most probably delayed as well.
1793 /* Undo during loss recovery after partial ACK. */
1794 static int tcp_try_undo_loss(struct sock
*sk
, struct tcp_sock
*tp
)
1796 if (tcp_may_undo(tp
)) {
1797 struct sk_buff
*skb
;
1798 sk_stream_for_retrans_queue(skb
, sk
) {
1799 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1802 clear_all_retrans_hints(tp
);
1804 DBGUNDO(sk
, tp
, "partial loss");
1806 tp
->left_out
= tp
->sacked_out
;
1807 tcp_undo_cwr(sk
, 1);
1808 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1809 inet_csk(sk
)->icsk_retransmits
= 0;
1810 tp
->undo_marker
= 0;
1812 tcp_set_ca_state(sk
, TCP_CA_Open
);
1818 static inline void tcp_complete_cwr(struct sock
*sk
)
1820 struct tcp_sock
*tp
= tcp_sk(sk
);
1821 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1822 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1823 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
1826 static void tcp_try_to_open(struct sock
*sk
, struct tcp_sock
*tp
, int flag
)
1828 tp
->left_out
= tp
->sacked_out
;
1830 if (tp
->retrans_out
== 0)
1831 tp
->retrans_stamp
= 0;
1836 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
1837 int state
= TCP_CA_Open
;
1839 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
1840 state
= TCP_CA_Disorder
;
1842 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
1843 tcp_set_ca_state(sk
, state
);
1844 tp
->high_seq
= tp
->snd_nxt
;
1846 tcp_moderate_cwnd(tp
);
1852 /* Process an event, which can update packets-in-flight not trivially.
1853 * Main goal of this function is to calculate new estimate for left_out,
1854 * taking into account both packets sitting in receiver's buffer and
1855 * packets lost by network.
1857 * Besides that it does CWND reduction, when packet loss is detected
1858 * and changes state of machine.
1860 * It does _not_ decide what to send, it is made in function
1861 * tcp_xmit_retransmit_queue().
1864 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
1865 int prior_packets
, int flag
)
1867 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1868 struct tcp_sock
*tp
= tcp_sk(sk
);
1869 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
1871 /* Some technical things:
1872 * 1. Reno does not count dupacks (sacked_out) automatically. */
1873 if (!tp
->packets_out
)
1875 /* 2. SACK counts snd_fack in packets inaccurately. */
1876 if (tp
->sacked_out
== 0)
1877 tp
->fackets_out
= 0;
1879 /* Now state machine starts.
1880 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1882 tp
->prior_ssthresh
= 0;
1884 /* B. In all the states check for reneging SACKs. */
1885 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
1888 /* C. Process data loss notification, provided it is valid. */
1889 if ((flag
&FLAG_DATA_LOST
) &&
1890 before(tp
->snd_una
, tp
->high_seq
) &&
1891 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
1892 tp
->fackets_out
> tp
->reordering
) {
1893 tcp_mark_head_lost(sk
, tp
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
1894 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
1897 /* D. Synchronize left_out to current state. */
1898 tcp_sync_left_out(tp
);
1900 /* E. Check state exit conditions. State can be terminated
1901 * when high_seq is ACKed. */
1902 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
1903 if (!sysctl_tcp_frto
)
1904 BUG_TRAP(tp
->retrans_out
== 0);
1905 tp
->retrans_stamp
= 0;
1906 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
1907 switch (icsk
->icsk_ca_state
) {
1909 icsk
->icsk_retransmits
= 0;
1910 if (tcp_try_undo_recovery(sk
, tp
))
1915 /* CWR is to be held something *above* high_seq
1916 * is ACKed for CWR bit to reach receiver. */
1917 if (tp
->snd_una
!= tp
->high_seq
) {
1918 tcp_complete_cwr(sk
);
1919 tcp_set_ca_state(sk
, TCP_CA_Open
);
1923 case TCP_CA_Disorder
:
1924 tcp_try_undo_dsack(sk
, tp
);
1925 if (!tp
->undo_marker
||
1926 /* For SACK case do not Open to allow to undo
1927 * catching for all duplicate ACKs. */
1928 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
1929 tp
->undo_marker
= 0;
1930 tcp_set_ca_state(sk
, TCP_CA_Open
);
1934 case TCP_CA_Recovery
:
1936 tcp_reset_reno_sack(tp
);
1937 if (tcp_try_undo_recovery(sk
, tp
))
1939 tcp_complete_cwr(sk
);
1944 /* F. Process state. */
1945 switch (icsk
->icsk_ca_state
) {
1946 case TCP_CA_Recovery
:
1947 if (prior_snd_una
== tp
->snd_una
) {
1948 if (IsReno(tp
) && is_dupack
)
1949 tcp_add_reno_sack(sk
);
1951 int acked
= prior_packets
- tp
->packets_out
;
1953 tcp_remove_reno_sacks(sk
, tp
, acked
);
1954 is_dupack
= tcp_try_undo_partial(sk
, tp
, acked
);
1958 if (flag
&FLAG_DATA_ACKED
)
1959 icsk
->icsk_retransmits
= 0;
1960 if (!tcp_try_undo_loss(sk
, tp
)) {
1961 tcp_moderate_cwnd(tp
);
1962 tcp_xmit_retransmit_queue(sk
);
1965 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1967 /* Loss is undone; fall through to processing in Open state. */
1970 if (tp
->snd_una
!= prior_snd_una
)
1971 tcp_reset_reno_sack(tp
);
1973 tcp_add_reno_sack(sk
);
1976 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
1977 tcp_try_undo_dsack(sk
, tp
);
1979 if (!tcp_time_to_recover(sk
, tp
)) {
1980 tcp_try_to_open(sk
, tp
, flag
);
1984 /* Otherwise enter Recovery state */
1987 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
1989 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
1991 tp
->high_seq
= tp
->snd_nxt
;
1992 tp
->prior_ssthresh
= 0;
1993 tp
->undo_marker
= tp
->snd_una
;
1994 tp
->undo_retrans
= tp
->retrans_out
;
1996 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
1997 if (!(flag
&FLAG_ECE
))
1998 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1999 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2000 TCP_ECN_queue_cwr(tp
);
2003 tp
->bytes_acked
= 0;
2004 tp
->snd_cwnd_cnt
= 0;
2005 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2008 if (is_dupack
|| tcp_head_timedout(sk
, tp
))
2009 tcp_update_scoreboard(sk
, tp
);
2011 tcp_xmit_retransmit_queue(sk
);
2014 /* Read draft-ietf-tcplw-high-performance before mucking
2015 * with this code. (Supersedes RFC1323)
2017 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2019 /* RTTM Rule: A TSecr value received in a segment is used to
2020 * update the averaged RTT measurement only if the segment
2021 * acknowledges some new data, i.e., only if it advances the
2022 * left edge of the send window.
2024 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2025 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2027 * Changed: reset backoff as soon as we see the first valid sample.
2028 * If we do not, we get strongly overestimated rto. With timestamps
2029 * samples are accepted even from very old segments: f.e., when rtt=1
2030 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2031 * answer arrives rto becomes 120 seconds! If at least one of segments
2032 * in window is lost... Voila. --ANK (010210)
2034 struct tcp_sock
*tp
= tcp_sk(sk
);
2035 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2036 tcp_rtt_estimator(sk
, seq_rtt
);
2038 inet_csk(sk
)->icsk_backoff
= 0;
2042 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2044 /* We don't have a timestamp. Can only use
2045 * packets that are not retransmitted to determine
2046 * rtt estimates. Also, we must not reset the
2047 * backoff for rto until we get a non-retransmitted
2048 * packet. This allows us to deal with a situation
2049 * where the network delay has increased suddenly.
2050 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2053 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2056 tcp_rtt_estimator(sk
, seq_rtt
);
2058 inet_csk(sk
)->icsk_backoff
= 0;
2062 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2065 const struct tcp_sock
*tp
= tcp_sk(sk
);
2066 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2067 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2068 tcp_ack_saw_tstamp(sk
, flag
);
2069 else if (seq_rtt
>= 0)
2070 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2073 static inline void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
2074 u32 in_flight
, int good
)
2076 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2077 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
2078 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2081 /* Restart timer after forward progress on connection.
2082 * RFC2988 recommends to restart timer to now+rto.
2085 static inline void tcp_ack_packets_out(struct sock
*sk
, struct tcp_sock
*tp
)
2087 if (!tp
->packets_out
) {
2088 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2090 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2094 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2095 __u32 now
, __s32
*seq_rtt
)
2097 struct tcp_sock
*tp
= tcp_sk(sk
);
2098 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2099 __u32 seq
= tp
->snd_una
;
2100 __u32 packets_acked
;
2103 /* If we get here, the whole TSO packet has not been
2106 BUG_ON(!after(scb
->end_seq
, seq
));
2108 packets_acked
= tcp_skb_pcount(skb
);
2109 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2111 packets_acked
-= tcp_skb_pcount(skb
);
2113 if (packets_acked
) {
2114 __u8 sacked
= scb
->sacked
;
2116 acked
|= FLAG_DATA_ACKED
;
2118 if (sacked
& TCPCB_RETRANS
) {
2119 if (sacked
& TCPCB_SACKED_RETRANS
)
2120 tp
->retrans_out
-= packets_acked
;
2121 acked
|= FLAG_RETRANS_DATA_ACKED
;
2123 } else if (*seq_rtt
< 0)
2124 *seq_rtt
= now
- scb
->when
;
2125 if (sacked
& TCPCB_SACKED_ACKED
)
2126 tp
->sacked_out
-= packets_acked
;
2127 if (sacked
& TCPCB_LOST
)
2128 tp
->lost_out
-= packets_acked
;
2129 if (sacked
& TCPCB_URG
) {
2131 !before(seq
, tp
->snd_up
))
2134 } else if (*seq_rtt
< 0)
2135 *seq_rtt
= now
- scb
->when
;
2137 if (tp
->fackets_out
) {
2138 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2139 tp
->fackets_out
-= dval
;
2141 tp
->packets_out
-= packets_acked
;
2143 BUG_ON(tcp_skb_pcount(skb
) == 0);
2144 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2150 static inline u32
tcp_usrtt(const struct sk_buff
*skb
)
2152 struct timeval tv
, now
;
2154 do_gettimeofday(&now
);
2155 skb_get_timestamp(skb
, &tv
);
2156 return (now
.tv_sec
- tv
.tv_sec
) * 1000000 + (now
.tv_usec
- tv
.tv_usec
);
2159 /* Remove acknowledged frames from the retransmission queue. */
2160 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2162 struct tcp_sock
*tp
= tcp_sk(sk
);
2163 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2164 struct sk_buff
*skb
;
2165 __u32 now
= tcp_time_stamp
;
2169 void (*rtt_sample
)(struct sock
*sk
, u32 usrtt
)
2170 = icsk
->icsk_ca_ops
->rtt_sample
;
2172 while ((skb
= skb_peek(&sk
->sk_write_queue
)) &&
2173 skb
!= sk
->sk_send_head
) {
2174 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2175 __u8 sacked
= scb
->sacked
;
2177 /* If our packet is before the ack sequence we can
2178 * discard it as it's confirmed to have arrived at
2181 if (after(scb
->end_seq
, tp
->snd_una
)) {
2182 if (tcp_skb_pcount(skb
) > 1 &&
2183 after(tp
->snd_una
, scb
->seq
))
2184 acked
|= tcp_tso_acked(sk
, skb
,
2189 /* Initial outgoing SYN's get put onto the write_queue
2190 * just like anything else we transmit. It is not
2191 * true data, and if we misinform our callers that
2192 * this ACK acks real data, we will erroneously exit
2193 * connection startup slow start one packet too
2194 * quickly. This is severely frowned upon behavior.
2196 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2197 acked
|= FLAG_DATA_ACKED
;
2200 acked
|= FLAG_SYN_ACKED
;
2201 tp
->retrans_stamp
= 0;
2205 if (sacked
& TCPCB_RETRANS
) {
2206 if(sacked
& TCPCB_SACKED_RETRANS
)
2207 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2208 acked
|= FLAG_RETRANS_DATA_ACKED
;
2210 } else if (seq_rtt
< 0) {
2211 seq_rtt
= now
- scb
->when
;
2213 (*rtt_sample
)(sk
, tcp_usrtt(skb
));
2215 if (sacked
& TCPCB_SACKED_ACKED
)
2216 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2217 if (sacked
& TCPCB_LOST
)
2218 tp
->lost_out
-= tcp_skb_pcount(skb
);
2219 if (sacked
& TCPCB_URG
) {
2221 !before(scb
->end_seq
, tp
->snd_up
))
2224 } else if (seq_rtt
< 0) {
2225 seq_rtt
= now
- scb
->when
;
2227 (*rtt_sample
)(sk
, tcp_usrtt(skb
));
2229 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2230 tcp_packets_out_dec(tp
, skb
);
2231 __skb_unlink(skb
, &sk
->sk_write_queue
);
2232 sk_stream_free_skb(sk
, skb
);
2233 clear_all_retrans_hints(tp
);
2236 if (acked
&FLAG_ACKED
) {
2237 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2238 tcp_ack_packets_out(sk
, tp
);
2240 if (icsk
->icsk_ca_ops
->pkts_acked
)
2241 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
);
2244 #if FASTRETRANS_DEBUG > 0
2245 BUG_TRAP((int)tp
->sacked_out
>= 0);
2246 BUG_TRAP((int)tp
->lost_out
>= 0);
2247 BUG_TRAP((int)tp
->retrans_out
>= 0);
2248 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2249 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2251 printk(KERN_DEBUG
"Leak l=%u %d\n",
2252 tp
->lost_out
, icsk
->icsk_ca_state
);
2255 if (tp
->sacked_out
) {
2256 printk(KERN_DEBUG
"Leak s=%u %d\n",
2257 tp
->sacked_out
, icsk
->icsk_ca_state
);
2260 if (tp
->retrans_out
) {
2261 printk(KERN_DEBUG
"Leak r=%u %d\n",
2262 tp
->retrans_out
, icsk
->icsk_ca_state
);
2263 tp
->retrans_out
= 0;
2267 *seq_rtt_p
= seq_rtt
;
2271 static void tcp_ack_probe(struct sock
*sk
)
2273 const struct tcp_sock
*tp
= tcp_sk(sk
);
2274 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2276 /* Was it a usable window open? */
2278 if (!after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
2279 tp
->snd_una
+ tp
->snd_wnd
)) {
2280 icsk
->icsk_backoff
= 0;
2281 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2282 /* Socket must be waked up by subsequent tcp_data_snd_check().
2283 * This function is not for random using!
2286 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2287 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2292 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2294 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2295 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2298 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2300 const struct tcp_sock
*tp
= tcp_sk(sk
);
2301 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2302 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2305 /* Check that window update is acceptable.
2306 * The function assumes that snd_una<=ack<=snd_next.
2308 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2309 const u32 ack_seq
, const u32 nwin
)
2311 return (after(ack
, tp
->snd_una
) ||
2312 after(ack_seq
, tp
->snd_wl1
) ||
2313 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2316 /* Update our send window.
2318 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2319 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2321 static int tcp_ack_update_window(struct sock
*sk
, struct tcp_sock
*tp
,
2322 struct sk_buff
*skb
, u32 ack
, u32 ack_seq
)
2325 u32 nwin
= ntohs(skb
->h
.th
->window
);
2327 if (likely(!skb
->h
.th
->syn
))
2328 nwin
<<= tp
->rx_opt
.snd_wscale
;
2330 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2331 flag
|= FLAG_WIN_UPDATE
;
2332 tcp_update_wl(tp
, ack
, ack_seq
);
2334 if (tp
->snd_wnd
!= nwin
) {
2337 /* Note, it is the only place, where
2338 * fast path is recovered for sending TCP.
2341 tcp_fast_path_check(sk
, tp
);
2343 if (nwin
> tp
->max_window
) {
2344 tp
->max_window
= nwin
;
2345 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
2355 static void tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
)
2357 struct tcp_sock
*tp
= tcp_sk(sk
);
2359 tcp_sync_left_out(tp
);
2361 if (tp
->snd_una
== prior_snd_una
||
2362 !before(tp
->snd_una
, tp
->frto_highmark
)) {
2363 /* RTO was caused by loss, start retransmitting in
2364 * go-back-N slow start
2366 tcp_enter_frto_loss(sk
);
2370 if (tp
->frto_counter
== 1) {
2371 /* First ACK after RTO advances the window: allow two new
2374 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2376 /* Also the second ACK after RTO advances the window.
2377 * The RTO was likely spurious. Reduce cwnd and continue
2378 * in congestion avoidance
2380 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2381 tcp_moderate_cwnd(tp
);
2384 /* F-RTO affects on two new ACKs following RTO.
2385 * At latest on third ACK the TCP behavior is back to normal.
2387 tp
->frto_counter
= (tp
->frto_counter
+ 1) % 3;
2390 /* This routine deals with incoming acks, but not outgoing ones. */
2391 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2393 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2394 struct tcp_sock
*tp
= tcp_sk(sk
);
2395 u32 prior_snd_una
= tp
->snd_una
;
2396 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2397 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2398 u32 prior_in_flight
;
2402 /* If the ack is newer than sent or older than previous acks
2403 * then we can probably ignore it.
2405 if (after(ack
, tp
->snd_nxt
))
2406 goto uninteresting_ack
;
2408 if (before(ack
, prior_snd_una
))
2411 if (sysctl_tcp_abc
&& icsk
->icsk_ca_state
< TCP_CA_CWR
)
2412 tp
->bytes_acked
+= ack
- prior_snd_una
;
2414 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2415 /* Window is constant, pure forward advance.
2416 * No more checks are required.
2417 * Note, we use the fact that SND.UNA>=SND.WL2.
2419 tcp_update_wl(tp
, ack
, ack_seq
);
2421 flag
|= FLAG_WIN_UPDATE
;
2423 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2425 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2427 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2430 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2432 flag
|= tcp_ack_update_window(sk
, tp
, skb
, ack
, ack_seq
);
2434 if (TCP_SKB_CB(skb
)->sacked
)
2435 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2437 if (TCP_ECN_rcv_ecn_echo(tp
, skb
->h
.th
))
2440 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2443 /* We passed data and got it acked, remove any soft error
2444 * log. Something worked...
2446 sk
->sk_err_soft
= 0;
2447 tp
->rcv_tstamp
= tcp_time_stamp
;
2448 prior_packets
= tp
->packets_out
;
2452 prior_in_flight
= tcp_packets_in_flight(tp
);
2454 /* See if we can take anything off of the retransmit queue. */
2455 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2457 if (tp
->frto_counter
)
2458 tcp_process_frto(sk
, prior_snd_una
);
2460 if (tcp_ack_is_dubious(sk
, flag
)) {
2461 /* Advance CWND, if state allows this. */
2462 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
2463 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2464 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2466 if ((flag
& FLAG_DATA_ACKED
))
2467 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2470 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2471 dst_confirm(sk
->sk_dst_cache
);
2476 icsk
->icsk_probes_out
= 0;
2478 /* If this ack opens up a zero window, clear backoff. It was
2479 * being used to time the probes, and is probably far higher than
2480 * it needs to be for normal retransmission.
2482 if (sk
->sk_send_head
)
2487 if (TCP_SKB_CB(skb
)->sacked
)
2488 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2491 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2496 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2497 * But, this can also be called on packets in the established flow when
2498 * the fast version below fails.
2500 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2503 struct tcphdr
*th
= skb
->h
.th
;
2504 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2506 ptr
= (unsigned char *)(th
+ 1);
2507 opt_rx
->saw_tstamp
= 0;
2516 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2521 if (opsize
< 2) /* "silly options" */
2523 if (opsize
> length
)
2524 return; /* don't parse partial options */
2527 if(opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2528 u16 in_mss
= ntohs(get_unaligned((__u16
*)ptr
));
2530 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2531 in_mss
= opt_rx
->user_mss
;
2532 opt_rx
->mss_clamp
= in_mss
;
2537 if(opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2538 if (sysctl_tcp_window_scaling
) {
2539 __u8 snd_wscale
= *(__u8
*) ptr
;
2540 opt_rx
->wscale_ok
= 1;
2541 if (snd_wscale
> 14) {
2543 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2544 "scaling value %d >14 received.\n",
2548 opt_rx
->snd_wscale
= snd_wscale
;
2551 case TCPOPT_TIMESTAMP
:
2552 if(opsize
==TCPOLEN_TIMESTAMP
) {
2553 if ((estab
&& opt_rx
->tstamp_ok
) ||
2554 (!estab
&& sysctl_tcp_timestamps
)) {
2555 opt_rx
->saw_tstamp
= 1;
2556 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__u32
*)ptr
));
2557 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__u32
*)(ptr
+4)));
2561 case TCPOPT_SACK_PERM
:
2562 if(opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2563 if (sysctl_tcp_sack
) {
2564 opt_rx
->sack_ok
= 1;
2565 tcp_sack_reset(opt_rx
);
2571 if((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2572 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2574 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2583 /* Fast parse options. This hopes to only see timestamps.
2584 * If it is wrong it falls back on tcp_parse_options().
2586 static inline int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2587 struct tcp_sock
*tp
)
2589 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2590 tp
->rx_opt
.saw_tstamp
= 0;
2592 } else if (tp
->rx_opt
.tstamp_ok
&&
2593 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2594 __u32
*ptr
= (__u32
*)(th
+ 1);
2595 if (*ptr
== ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2596 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2597 tp
->rx_opt
.saw_tstamp
= 1;
2599 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2601 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2605 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2609 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2611 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2612 tp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
2615 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2617 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2618 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2619 * extra check below makes sure this can only happen
2620 * for pure ACK frames. -DaveM
2622 * Not only, also it occurs for expired timestamps.
2625 if((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2626 xtime
.tv_sec
>= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2627 tcp_store_ts_recent(tp
);
2631 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2633 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2634 * it can pass through stack. So, the following predicate verifies that
2635 * this segment is not used for anything but congestion avoidance or
2636 * fast retransmit. Moreover, we even are able to eliminate most of such
2637 * second order effects, if we apply some small "replay" window (~RTO)
2638 * to timestamp space.
2640 * All these measures still do not guarantee that we reject wrapped ACKs
2641 * on networks with high bandwidth, when sequence space is recycled fastly,
2642 * but it guarantees that such events will be very rare and do not affect
2643 * connection seriously. This doesn't look nice, but alas, PAWS is really
2646 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2647 * states that events when retransmit arrives after original data are rare.
2648 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2649 * the biggest problem on large power networks even with minor reordering.
2650 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2651 * up to bandwidth of 18Gigabit/sec. 8) ]
2654 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
2656 struct tcp_sock
*tp
= tcp_sk(sk
);
2657 struct tcphdr
*th
= skb
->h
.th
;
2658 u32 seq
= TCP_SKB_CB(skb
)->seq
;
2659 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2661 return (/* 1. Pure ACK with correct sequence number. */
2662 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
2664 /* 2. ... and duplicate ACK. */
2665 ack
== tp
->snd_una
&&
2667 /* 3. ... and does not update window. */
2668 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
2670 /* 4. ... and sits in replay window. */
2671 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
2674 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
2676 const struct tcp_sock
*tp
= tcp_sk(sk
);
2677 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
2678 xtime
.tv_sec
< tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
2679 !tcp_disordered_ack(sk
, skb
));
2682 /* Check segment sequence number for validity.
2684 * Segment controls are considered valid, if the segment
2685 * fits to the window after truncation to the window. Acceptability
2686 * of data (and SYN, FIN, of course) is checked separately.
2687 * See tcp_data_queue(), for example.
2689 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2690 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2691 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2692 * (borrowed from freebsd)
2695 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2697 return !before(end_seq
, tp
->rcv_wup
) &&
2698 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
2701 /* When we get a reset we do this. */
2702 static void tcp_reset(struct sock
*sk
)
2704 /* We want the right error as BSD sees it (and indeed as we do). */
2705 switch (sk
->sk_state
) {
2707 sk
->sk_err
= ECONNREFUSED
;
2709 case TCP_CLOSE_WAIT
:
2715 sk
->sk_err
= ECONNRESET
;
2718 if (!sock_flag(sk
, SOCK_DEAD
))
2719 sk
->sk_error_report(sk
);
2725 * Process the FIN bit. This now behaves as it is supposed to work
2726 * and the FIN takes effect when it is validly part of sequence
2727 * space. Not before when we get holes.
2729 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2730 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2733 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2734 * close and we go into CLOSING (and later onto TIME-WAIT)
2736 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2738 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
2740 struct tcp_sock
*tp
= tcp_sk(sk
);
2742 inet_csk_schedule_ack(sk
);
2744 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
2745 sock_set_flag(sk
, SOCK_DONE
);
2747 switch (sk
->sk_state
) {
2749 case TCP_ESTABLISHED
:
2750 /* Move to CLOSE_WAIT */
2751 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
2752 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
2755 case TCP_CLOSE_WAIT
:
2757 /* Received a retransmission of the FIN, do
2762 /* RFC793: Remain in the LAST-ACK state. */
2766 /* This case occurs when a simultaneous close
2767 * happens, we must ack the received FIN and
2768 * enter the CLOSING state.
2771 tcp_set_state(sk
, TCP_CLOSING
);
2774 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2776 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
2779 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2780 * cases we should never reach this piece of code.
2782 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
2783 __FUNCTION__
, sk
->sk_state
);
2787 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2788 * Probably, we should reset in this case. For now drop them.
2790 __skb_queue_purge(&tp
->out_of_order_queue
);
2791 if (tp
->rx_opt
.sack_ok
)
2792 tcp_sack_reset(&tp
->rx_opt
);
2793 sk_stream_mem_reclaim(sk
);
2795 if (!sock_flag(sk
, SOCK_DEAD
)) {
2796 sk
->sk_state_change(sk
);
2798 /* Do not send POLL_HUP for half duplex close. */
2799 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
2800 sk
->sk_state
== TCP_CLOSE
)
2801 sk_wake_async(sk
, 1, POLL_HUP
);
2803 sk_wake_async(sk
, 1, POLL_IN
);
2807 static __inline__
int
2808 tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
2810 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
2811 if (before(seq
, sp
->start_seq
))
2812 sp
->start_seq
= seq
;
2813 if (after(end_seq
, sp
->end_seq
))
2814 sp
->end_seq
= end_seq
;
2820 static inline void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2822 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2823 if (before(seq
, tp
->rcv_nxt
))
2824 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
2826 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
2828 tp
->rx_opt
.dsack
= 1;
2829 tp
->duplicate_sack
[0].start_seq
= seq
;
2830 tp
->duplicate_sack
[0].end_seq
= end_seq
;
2831 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
2835 static inline void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2837 if (!tp
->rx_opt
.dsack
)
2838 tcp_dsack_set(tp
, seq
, end_seq
);
2840 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
2843 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
2845 struct tcp_sock
*tp
= tcp_sk(sk
);
2847 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
2848 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
2849 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
2850 tcp_enter_quickack_mode(sk
);
2852 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2853 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2855 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
2856 end_seq
= tp
->rcv_nxt
;
2857 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
2864 /* These routines update the SACK block as out-of-order packets arrive or
2865 * in-order packets close up the sequence space.
2867 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
2870 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2871 struct tcp_sack_block
*swalk
= sp
+1;
2873 /* See if the recent change to the first SACK eats into
2874 * or hits the sequence space of other SACK blocks, if so coalesce.
2876 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
2877 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
2880 /* Zap SWALK, by moving every further SACK up by one slot.
2881 * Decrease num_sacks.
2883 tp
->rx_opt
.num_sacks
--;
2884 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2885 for(i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
2889 this_sack
++, swalk
++;
2893 static __inline__
void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
2897 tmp
= sack1
->start_seq
;
2898 sack1
->start_seq
= sack2
->start_seq
;
2899 sack2
->start_seq
= tmp
;
2901 tmp
= sack1
->end_seq
;
2902 sack1
->end_seq
= sack2
->end_seq
;
2903 sack2
->end_seq
= tmp
;
2906 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
2908 struct tcp_sock
*tp
= tcp_sk(sk
);
2909 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2910 int cur_sacks
= tp
->rx_opt
.num_sacks
;
2916 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
2917 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
2918 /* Rotate this_sack to the first one. */
2919 for (; this_sack
>0; this_sack
--, sp
--)
2920 tcp_sack_swap(sp
, sp
-1);
2922 tcp_sack_maybe_coalesce(tp
);
2927 /* Could not find an adjacent existing SACK, build a new one,
2928 * put it at the front, and shift everyone else down. We
2929 * always know there is at least one SACK present already here.
2931 * If the sack array is full, forget about the last one.
2933 if (this_sack
>= 4) {
2935 tp
->rx_opt
.num_sacks
--;
2938 for(; this_sack
> 0; this_sack
--, sp
--)
2942 /* Build the new head SACK, and we're done. */
2943 sp
->start_seq
= seq
;
2944 sp
->end_seq
= end_seq
;
2945 tp
->rx_opt
.num_sacks
++;
2946 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2949 /* RCV.NXT advances, some SACKs should be eaten. */
2951 static void tcp_sack_remove(struct tcp_sock
*tp
)
2953 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2954 int num_sacks
= tp
->rx_opt
.num_sacks
;
2957 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
2958 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
2959 tp
->rx_opt
.num_sacks
= 0;
2960 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
2964 for(this_sack
= 0; this_sack
< num_sacks
; ) {
2965 /* Check if the start of the sack is covered by RCV.NXT. */
2966 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
2969 /* RCV.NXT must cover all the block! */
2970 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
2972 /* Zap this SACK, by moving forward any other SACKS. */
2973 for (i
=this_sack
+1; i
< num_sacks
; i
++)
2974 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
2981 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
2982 tp
->rx_opt
.num_sacks
= num_sacks
;
2983 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2987 /* This one checks to see if we can put data from the
2988 * out_of_order queue into the receive_queue.
2990 static void tcp_ofo_queue(struct sock
*sk
)
2992 struct tcp_sock
*tp
= tcp_sk(sk
);
2993 __u32 dsack_high
= tp
->rcv_nxt
;
2994 struct sk_buff
*skb
;
2996 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
2997 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3000 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3001 __u32 dsack
= dsack_high
;
3002 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3003 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3004 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3007 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3008 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3009 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3013 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3014 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3015 TCP_SKB_CB(skb
)->end_seq
);
3017 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3018 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3019 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3021 tcp_fin(skb
, sk
, skb
->h
.th
);
3025 static int tcp_prune_queue(struct sock
*sk
);
3027 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3029 struct tcphdr
*th
= skb
->h
.th
;
3030 struct tcp_sock
*tp
= tcp_sk(sk
);
3033 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3036 __skb_pull(skb
, th
->doff
*4);
3038 TCP_ECN_accept_cwr(tp
, skb
);
3040 if (tp
->rx_opt
.dsack
) {
3041 tp
->rx_opt
.dsack
= 0;
3042 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3043 4 - tp
->rx_opt
.tstamp_ok
);
3046 /* Queue data for delivery to the user.
3047 * Packets in sequence go to the receive queue.
3048 * Out of sequence packets to the out_of_order_queue.
3050 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3051 if (tcp_receive_window(tp
) == 0)
3054 /* Ok. In sequence. In window. */
3055 if (tp
->ucopy
.task
== current
&&
3056 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3057 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3058 int chunk
= min_t(unsigned int, skb
->len
,
3061 __set_current_state(TASK_RUNNING
);
3064 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3065 tp
->ucopy
.len
-= chunk
;
3066 tp
->copied_seq
+= chunk
;
3067 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3068 tcp_rcv_space_adjust(sk
);
3076 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3077 !sk_stream_rmem_schedule(sk
, skb
))) {
3078 if (tcp_prune_queue(sk
) < 0 ||
3079 !sk_stream_rmem_schedule(sk
, skb
))
3082 sk_stream_set_owner_r(skb
, sk
);
3083 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3085 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3087 tcp_event_data_recv(sk
, tp
, skb
);
3089 tcp_fin(skb
, sk
, th
);
3091 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3094 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3095 * gap in queue is filled.
3097 if (skb_queue_empty(&tp
->out_of_order_queue
))
3098 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3101 if (tp
->rx_opt
.num_sacks
)
3102 tcp_sack_remove(tp
);
3104 tcp_fast_path_check(sk
, tp
);
3108 else if (!sock_flag(sk
, SOCK_DEAD
))
3109 sk
->sk_data_ready(sk
, 0);
3113 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3114 /* A retransmit, 2nd most common case. Force an immediate ack. */
3115 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3116 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3119 tcp_enter_quickack_mode(sk
);
3120 inet_csk_schedule_ack(sk
);
3126 /* Out of window. F.e. zero window probe. */
3127 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3130 tcp_enter_quickack_mode(sk
);
3132 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3133 /* Partial packet, seq < rcv_next < end_seq */
3134 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3135 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3136 TCP_SKB_CB(skb
)->end_seq
);
3138 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3140 /* If window is closed, drop tail of packet. But after
3141 * remembering D-SACK for its head made in previous line.
3143 if (!tcp_receive_window(tp
))
3148 TCP_ECN_check_ce(tp
, skb
);
3150 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3151 !sk_stream_rmem_schedule(sk
, skb
)) {
3152 if (tcp_prune_queue(sk
) < 0 ||
3153 !sk_stream_rmem_schedule(sk
, skb
))
3157 /* Disable header prediction. */
3159 inet_csk_schedule_ack(sk
);
3161 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3162 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3164 sk_stream_set_owner_r(skb
, sk
);
3166 if (!skb_peek(&tp
->out_of_order_queue
)) {
3167 /* Initial out of order segment, build 1 SACK. */
3168 if (tp
->rx_opt
.sack_ok
) {
3169 tp
->rx_opt
.num_sacks
= 1;
3170 tp
->rx_opt
.dsack
= 0;
3171 tp
->rx_opt
.eff_sacks
= 1;
3172 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3173 tp
->selective_acks
[0].end_seq
=
3174 TCP_SKB_CB(skb
)->end_seq
;
3176 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3178 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3179 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3180 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3182 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3183 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3185 if (!tp
->rx_opt
.num_sacks
||
3186 tp
->selective_acks
[0].end_seq
!= seq
)
3189 /* Common case: data arrive in order after hole. */
3190 tp
->selective_acks
[0].end_seq
= end_seq
;
3194 /* Find place to insert this segment. */
3196 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3198 } while ((skb1
= skb1
->prev
) !=
3199 (struct sk_buff
*)&tp
->out_of_order_queue
);
3201 /* Do skb overlap to previous one? */
3202 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3203 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3204 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3205 /* All the bits are present. Drop. */
3207 tcp_dsack_set(tp
, seq
, end_seq
);
3210 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3211 /* Partial overlap. */
3212 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3217 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3219 /* And clean segments covered by new one as whole. */
3220 while ((skb1
= skb
->next
) !=
3221 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3222 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3223 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3224 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3227 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3228 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3233 if (tp
->rx_opt
.sack_ok
)
3234 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3238 /* Collapse contiguous sequence of skbs head..tail with
3239 * sequence numbers start..end.
3240 * Segments with FIN/SYN are not collapsed (only because this
3244 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3245 struct sk_buff
*head
, struct sk_buff
*tail
,
3248 struct sk_buff
*skb
;
3250 /* First, check that queue is collapsible and find
3251 * the point where collapsing can be useful. */
3252 for (skb
= head
; skb
!= tail
; ) {
3253 /* No new bits? It is possible on ofo queue. */
3254 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3255 struct sk_buff
*next
= skb
->next
;
3256 __skb_unlink(skb
, list
);
3258 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3263 /* The first skb to collapse is:
3265 * - bloated or contains data before "start" or
3266 * overlaps to the next one.
3268 if (!skb
->h
.th
->syn
&& !skb
->h
.th
->fin
&&
3269 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3270 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3271 (skb
->next
!= tail
&&
3272 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3275 /* Decided to skip this, advance start seq. */
3276 start
= TCP_SKB_CB(skb
)->end_seq
;
3279 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3282 while (before(start
, end
)) {
3283 struct sk_buff
*nskb
;
3284 int header
= skb_headroom(skb
);
3285 int copy
= SKB_MAX_ORDER(header
, 0);
3287 /* Too big header? This can happen with IPv6. */
3290 if (end
-start
< copy
)
3292 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3295 skb_reserve(nskb
, header
);
3296 memcpy(nskb
->head
, skb
->head
, header
);
3297 nskb
->nh
.raw
= nskb
->head
+ (skb
->nh
.raw
-skb
->head
);
3298 nskb
->h
.raw
= nskb
->head
+ (skb
->h
.raw
-skb
->head
);
3299 nskb
->mac
.raw
= nskb
->head
+ (skb
->mac
.raw
-skb
->head
);
3300 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3301 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3302 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3303 sk_stream_set_owner_r(nskb
, sk
);
3305 /* Copy data, releasing collapsed skbs. */
3307 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3308 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3310 if (offset
< 0) BUG();
3312 size
= min(copy
, size
);
3313 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3315 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3319 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3320 struct sk_buff
*next
= skb
->next
;
3321 __skb_unlink(skb
, list
);
3323 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3325 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3332 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3333 * and tcp_collapse() them until all the queue is collapsed.
3335 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3337 struct tcp_sock
*tp
= tcp_sk(sk
);
3338 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3339 struct sk_buff
*head
;
3345 start
= TCP_SKB_CB(skb
)->seq
;
3346 end
= TCP_SKB_CB(skb
)->end_seq
;
3352 /* Segment is terminated when we see gap or when
3353 * we are at the end of all the queue. */
3354 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3355 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3356 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3357 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3358 head
, skb
, start
, end
);
3360 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3362 /* Start new segment */
3363 start
= TCP_SKB_CB(skb
)->seq
;
3364 end
= TCP_SKB_CB(skb
)->end_seq
;
3366 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3367 start
= TCP_SKB_CB(skb
)->seq
;
3368 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3369 end
= TCP_SKB_CB(skb
)->end_seq
;
3374 /* Reduce allocated memory if we can, trying to get
3375 * the socket within its memory limits again.
3377 * Return less than zero if we should start dropping frames
3378 * until the socket owning process reads some of the data
3379 * to stabilize the situation.
3381 static int tcp_prune_queue(struct sock
*sk
)
3383 struct tcp_sock
*tp
= tcp_sk(sk
);
3385 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3387 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3389 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3390 tcp_clamp_window(sk
, tp
);
3391 else if (tcp_memory_pressure
)
3392 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3394 tcp_collapse_ofo_queue(sk
);
3395 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3396 sk
->sk_receive_queue
.next
,
3397 (struct sk_buff
*)&sk
->sk_receive_queue
,
3398 tp
->copied_seq
, tp
->rcv_nxt
);
3399 sk_stream_mem_reclaim(sk
);
3401 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3404 /* Collapsing did not help, destructive actions follow.
3405 * This must not ever occur. */
3407 /* First, purge the out_of_order queue. */
3408 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3409 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3410 __skb_queue_purge(&tp
->out_of_order_queue
);
3412 /* Reset SACK state. A conforming SACK implementation will
3413 * do the same at a timeout based retransmit. When a connection
3414 * is in a sad state like this, we care only about integrity
3415 * of the connection not performance.
3417 if (tp
->rx_opt
.sack_ok
)
3418 tcp_sack_reset(&tp
->rx_opt
);
3419 sk_stream_mem_reclaim(sk
);
3422 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3425 /* If we are really being abused, tell the caller to silently
3426 * drop receive data on the floor. It will get retransmitted
3427 * and hopefully then we'll have sufficient space.
3429 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3431 /* Massive buffer overcommit. */
3437 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3438 * As additional protections, we do not touch cwnd in retransmission phases,
3439 * and if application hit its sndbuf limit recently.
3441 void tcp_cwnd_application_limited(struct sock
*sk
)
3443 struct tcp_sock
*tp
= tcp_sk(sk
);
3445 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3446 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3447 /* Limited by application or receiver window. */
3448 u32 win_used
= max(tp
->snd_cwnd_used
, 2U);
3449 if (win_used
< tp
->snd_cwnd
) {
3450 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3451 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3453 tp
->snd_cwnd_used
= 0;
3455 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3458 static inline int tcp_should_expand_sndbuf(struct sock
*sk
, struct tcp_sock
*tp
)
3460 /* If the user specified a specific send buffer setting, do
3463 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3466 /* If we are under global TCP memory pressure, do not expand. */
3467 if (tcp_memory_pressure
)
3470 /* If we are under soft global TCP memory pressure, do not expand. */
3471 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3474 /* If we filled the congestion window, do not expand. */
3475 if (tp
->packets_out
>= tp
->snd_cwnd
)
3481 /* When incoming ACK allowed to free some skb from write_queue,
3482 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3483 * on the exit from tcp input handler.
3485 * PROBLEM: sndbuf expansion does not work well with largesend.
3487 static void tcp_new_space(struct sock
*sk
)
3489 struct tcp_sock
*tp
= tcp_sk(sk
);
3491 if (tcp_should_expand_sndbuf(sk
, tp
)) {
3492 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3493 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3494 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3495 tp
->reordering
+ 1);
3496 sndmem
*= 2*demanded
;
3497 if (sndmem
> sk
->sk_sndbuf
)
3498 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3499 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3502 sk
->sk_write_space(sk
);
3505 static inline void tcp_check_space(struct sock
*sk
)
3507 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3508 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3509 if (sk
->sk_socket
&&
3510 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3515 static __inline__
void tcp_data_snd_check(struct sock
*sk
, struct tcp_sock
*tp
)
3517 tcp_push_pending_frames(sk
, tp
);
3518 tcp_check_space(sk
);
3522 * Check if sending an ack is needed.
3524 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3526 struct tcp_sock
*tp
= tcp_sk(sk
);
3528 /* More than one full frame received... */
3529 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3530 /* ... and right edge of window advances far enough.
3531 * (tcp_recvmsg() will send ACK otherwise). Or...
3533 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3534 /* We ACK each frame or... */
3535 tcp_in_quickack_mode(sk
) ||
3536 /* We have out of order data. */
3538 skb_peek(&tp
->out_of_order_queue
))) {
3539 /* Then ack it now */
3542 /* Else, send delayed ack. */
3543 tcp_send_delayed_ack(sk
);
3547 static __inline__
void tcp_ack_snd_check(struct sock
*sk
)
3549 if (!inet_csk_ack_scheduled(sk
)) {
3550 /* We sent a data segment already. */
3553 __tcp_ack_snd_check(sk
, 1);
3557 * This routine is only called when we have urgent data
3558 * signaled. Its the 'slow' part of tcp_urg. It could be
3559 * moved inline now as tcp_urg is only called from one
3560 * place. We handle URGent data wrong. We have to - as
3561 * BSD still doesn't use the correction from RFC961.
3562 * For 1003.1g we should support a new option TCP_STDURG to permit
3563 * either form (or just set the sysctl tcp_stdurg).
3566 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3568 struct tcp_sock
*tp
= tcp_sk(sk
);
3569 u32 ptr
= ntohs(th
->urg_ptr
);
3571 if (ptr
&& !sysctl_tcp_stdurg
)
3573 ptr
+= ntohl(th
->seq
);
3575 /* Ignore urgent data that we've already seen and read. */
3576 if (after(tp
->copied_seq
, ptr
))
3579 /* Do not replay urg ptr.
3581 * NOTE: interesting situation not covered by specs.
3582 * Misbehaving sender may send urg ptr, pointing to segment,
3583 * which we already have in ofo queue. We are not able to fetch
3584 * such data and will stay in TCP_URG_NOTYET until will be eaten
3585 * by recvmsg(). Seems, we are not obliged to handle such wicked
3586 * situations. But it is worth to think about possibility of some
3587 * DoSes using some hypothetical application level deadlock.
3589 if (before(ptr
, tp
->rcv_nxt
))
3592 /* Do we already have a newer (or duplicate) urgent pointer? */
3593 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3596 /* Tell the world about our new urgent pointer. */
3599 /* We may be adding urgent data when the last byte read was
3600 * urgent. To do this requires some care. We cannot just ignore
3601 * tp->copied_seq since we would read the last urgent byte again
3602 * as data, nor can we alter copied_seq until this data arrives
3603 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3605 * NOTE. Double Dutch. Rendering to plain English: author of comment
3606 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3607 * and expect that both A and B disappear from stream. This is _wrong_.
3608 * Though this happens in BSD with high probability, this is occasional.
3609 * Any application relying on this is buggy. Note also, that fix "works"
3610 * only in this artificial test. Insert some normal data between A and B and we will
3611 * decline of BSD again. Verdict: it is better to remove to trap
3614 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3615 !sock_flag(sk
, SOCK_URGINLINE
) &&
3616 tp
->copied_seq
!= tp
->rcv_nxt
) {
3617 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3619 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3620 __skb_unlink(skb
, &sk
->sk_receive_queue
);
3625 tp
->urg_data
= TCP_URG_NOTYET
;
3628 /* Disable header prediction. */
3632 /* This is the 'fast' part of urgent handling. */
3633 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
3635 struct tcp_sock
*tp
= tcp_sk(sk
);
3637 /* Check if we get a new urgent pointer - normally not. */
3639 tcp_check_urg(sk
,th
);
3641 /* Do we wait for any urgent data? - normally not... */
3642 if (tp
->urg_data
== TCP_URG_NOTYET
) {
3643 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
3646 /* Is the urgent pointer pointing into this packet? */
3647 if (ptr
< skb
->len
) {
3649 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
3651 tp
->urg_data
= TCP_URG_VALID
| tmp
;
3652 if (!sock_flag(sk
, SOCK_DEAD
))
3653 sk
->sk_data_ready(sk
, 0);
3658 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3660 struct tcp_sock
*tp
= tcp_sk(sk
);
3661 int chunk
= skb
->len
- hlen
;
3665 if (skb
->ip_summed
==CHECKSUM_UNNECESSARY
)
3666 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
3668 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
3672 tp
->ucopy
.len
-= chunk
;
3673 tp
->copied_seq
+= chunk
;
3674 tcp_rcv_space_adjust(sk
);
3681 static int __tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3685 if (sock_owned_by_user(sk
)) {
3687 result
= __tcp_checksum_complete(skb
);
3690 result
= __tcp_checksum_complete(skb
);
3695 static __inline__
int
3696 tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3698 return skb
->ip_summed
!= CHECKSUM_UNNECESSARY
&&
3699 __tcp_checksum_complete_user(sk
, skb
);
3703 * TCP receive function for the ESTABLISHED state.
3705 * It is split into a fast path and a slow path. The fast path is
3707 * - A zero window was announced from us - zero window probing
3708 * is only handled properly in the slow path.
3709 * - Out of order segments arrived.
3710 * - Urgent data is expected.
3711 * - There is no buffer space left
3712 * - Unexpected TCP flags/window values/header lengths are received
3713 * (detected by checking the TCP header against pred_flags)
3714 * - Data is sent in both directions. Fast path only supports pure senders
3715 * or pure receivers (this means either the sequence number or the ack
3716 * value must stay constant)
3717 * - Unexpected TCP option.
3719 * When these conditions are not satisfied it drops into a standard
3720 * receive procedure patterned after RFC793 to handle all cases.
3721 * The first three cases are guaranteed by proper pred_flags setting,
3722 * the rest is checked inline. Fast processing is turned on in
3723 * tcp_data_queue when everything is OK.
3725 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
3726 struct tcphdr
*th
, unsigned len
)
3728 struct tcp_sock
*tp
= tcp_sk(sk
);
3731 * Header prediction.
3732 * The code loosely follows the one in the famous
3733 * "30 instruction TCP receive" Van Jacobson mail.
3735 * Van's trick is to deposit buffers into socket queue
3736 * on a device interrupt, to call tcp_recv function
3737 * on the receive process context and checksum and copy
3738 * the buffer to user space. smart...
3740 * Our current scheme is not silly either but we take the
3741 * extra cost of the net_bh soft interrupt processing...
3742 * We do checksum and copy also but from device to kernel.
3745 tp
->rx_opt
.saw_tstamp
= 0;
3747 /* pred_flags is 0xS?10 << 16 + snd_wnd
3748 * if header_prediction is to be made
3749 * 'S' will always be tp->tcp_header_len >> 2
3750 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3751 * turn it off (when there are holes in the receive
3752 * space for instance)
3753 * PSH flag is ignored.
3756 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
3757 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3758 int tcp_header_len
= tp
->tcp_header_len
;
3760 /* Timestamp header prediction: tcp_header_len
3761 * is automatically equal to th->doff*4 due to pred_flags
3765 /* Check timestamp */
3766 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
3767 __u32
*ptr
= (__u32
*)(th
+ 1);
3769 /* No? Slow path! */
3770 if (*ptr
!= ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3771 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
3774 tp
->rx_opt
.saw_tstamp
= 1;
3776 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3778 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3780 /* If PAWS failed, check it more carefully in slow path */
3781 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
3784 /* DO NOT update ts_recent here, if checksum fails
3785 * and timestamp was corrupted part, it will result
3786 * in a hung connection since we will drop all
3787 * future packets due to the PAWS test.
3791 if (len
<= tcp_header_len
) {
3792 /* Bulk data transfer: sender */
3793 if (len
== tcp_header_len
) {
3794 /* Predicted packet is in window by definition.
3795 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3796 * Hence, check seq<=rcv_wup reduces to:
3798 if (tcp_header_len
==
3799 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3800 tp
->rcv_nxt
== tp
->rcv_wup
)
3801 tcp_store_ts_recent(tp
);
3803 tcp_rcv_rtt_measure_ts(sk
, skb
);
3805 /* We know that such packets are checksummed
3808 tcp_ack(sk
, skb
, 0);
3810 tcp_data_snd_check(sk
, tp
);
3812 } else { /* Header too small */
3813 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3819 if (tp
->ucopy
.task
== current
&&
3820 tp
->copied_seq
== tp
->rcv_nxt
&&
3821 len
- tcp_header_len
<= tp
->ucopy
.len
&&
3822 sock_owned_by_user(sk
)) {
3823 __set_current_state(TASK_RUNNING
);
3825 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
)) {
3826 /* Predicted packet is in window by definition.
3827 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3828 * Hence, check seq<=rcv_wup reduces to:
3830 if (tcp_header_len
==
3831 (sizeof(struct tcphdr
) +
3832 TCPOLEN_TSTAMP_ALIGNED
) &&
3833 tp
->rcv_nxt
== tp
->rcv_wup
)
3834 tcp_store_ts_recent(tp
);
3836 tcp_rcv_rtt_measure_ts(sk
, skb
);
3838 __skb_pull(skb
, tcp_header_len
);
3839 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3840 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
3845 if (tcp_checksum_complete_user(sk
, skb
))
3848 /* Predicted packet is in window by definition.
3849 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3850 * Hence, check seq<=rcv_wup reduces to:
3852 if (tcp_header_len
==
3853 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3854 tp
->rcv_nxt
== tp
->rcv_wup
)
3855 tcp_store_ts_recent(tp
);
3857 tcp_rcv_rtt_measure_ts(sk
, skb
);
3859 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
3862 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
3864 /* Bulk data transfer: receiver */
3865 __skb_pull(skb
,tcp_header_len
);
3866 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3867 sk_stream_set_owner_r(skb
, sk
);
3868 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3871 tcp_event_data_recv(sk
, tp
, skb
);
3873 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
3874 /* Well, only one small jumplet in fast path... */
3875 tcp_ack(sk
, skb
, FLAG_DATA
);
3876 tcp_data_snd_check(sk
, tp
);
3877 if (!inet_csk_ack_scheduled(sk
))
3881 __tcp_ack_snd_check(sk
, 0);
3886 sk
->sk_data_ready(sk
, 0);
3892 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
3896 * RFC1323: H1. Apply PAWS check first.
3898 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
3899 tcp_paws_discard(sk
, skb
)) {
3901 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
3902 tcp_send_dupack(sk
, skb
);
3905 /* Resets are accepted even if PAWS failed.
3907 ts_recent update must be made after we are sure
3908 that the packet is in window.
3913 * Standard slow path.
3916 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3917 /* RFC793, page 37: "In all states except SYN-SENT, all reset
3918 * (RST) segments are validated by checking their SEQ-fields."
3919 * And page 69: "If an incoming segment is not acceptable,
3920 * an acknowledgment should be sent in reply (unless the RST bit
3921 * is set, if so drop the segment and return)".
3924 tcp_send_dupack(sk
, skb
);
3933 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3935 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3936 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3937 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
3944 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
3946 tcp_rcv_rtt_measure_ts(sk
, skb
);
3948 /* Process urgent data. */
3949 tcp_urg(sk
, skb
, th
);
3951 /* step 7: process the segment text */
3952 tcp_data_queue(sk
, skb
);
3954 tcp_data_snd_check(sk
, tp
);
3955 tcp_ack_snd_check(sk
);
3959 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3966 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
3967 struct tcphdr
*th
, unsigned len
)
3969 struct tcp_sock
*tp
= tcp_sk(sk
);
3970 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
3972 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
3975 struct inet_connection_sock
*icsk
;
3977 * "If the state is SYN-SENT then
3978 * first check the ACK bit
3979 * If the ACK bit is set
3980 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
3981 * a reset (unless the RST bit is set, if so drop
3982 * the segment and return)"
3984 * We do not send data with SYN, so that RFC-correct
3987 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
3988 goto reset_and_undo
;
3990 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
3991 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
3993 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
3994 goto reset_and_undo
;
3997 /* Now ACK is acceptable.
3999 * "If the RST bit is set
4000 * If the ACK was acceptable then signal the user "error:
4001 * connection reset", drop the segment, enter CLOSED state,
4002 * delete TCB, and return."
4011 * "fifth, if neither of the SYN or RST bits is set then
4012 * drop the segment and return."
4018 goto discard_and_undo
;
4021 * "If the SYN bit is on ...
4022 * are acceptable then ...
4023 * (our SYN has been ACKed), change the connection
4024 * state to ESTABLISHED..."
4027 TCP_ECN_rcv_synack(tp
, th
);
4028 if (tp
->ecn_flags
&TCP_ECN_OK
)
4029 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
4031 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4032 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4034 /* Ok.. it's good. Set up sequence numbers and
4035 * move to established.
4037 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4038 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4040 /* RFC1323: The window in SYN & SYN/ACK segments is
4043 tp
->snd_wnd
= ntohs(th
->window
);
4044 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4046 if (!tp
->rx_opt
.wscale_ok
) {
4047 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4048 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4051 if (tp
->rx_opt
.saw_tstamp
) {
4052 tp
->rx_opt
.tstamp_ok
= 1;
4053 tp
->tcp_header_len
=
4054 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4055 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4056 tcp_store_ts_recent(tp
);
4058 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4061 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4062 tp
->rx_opt
.sack_ok
|= 2;
4064 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
4065 tcp_initialize_rcv_mss(sk
);
4067 /* Remember, tcp_poll() does not lock socket!
4068 * Change state from SYN-SENT only after copied_seq
4069 * is initialized. */
4070 tp
->copied_seq
= tp
->rcv_nxt
;
4072 tcp_set_state(sk
, TCP_ESTABLISHED
);
4074 /* Make sure socket is routed, for correct metrics. */
4075 tp
->af_specific
->rebuild_header(sk
);
4077 tcp_init_metrics(sk
);
4079 tcp_init_congestion_control(sk
);
4081 /* Prevent spurious tcp_cwnd_restart() on first data
4084 tp
->lsndtime
= tcp_time_stamp
;
4086 tcp_init_buffer_space(sk
);
4088 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4089 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4091 if (!tp
->rx_opt
.snd_wscale
)
4092 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4096 if (!sock_flag(sk
, SOCK_DEAD
)) {
4097 sk
->sk_state_change(sk
);
4098 sk_wake_async(sk
, 0, POLL_OUT
);
4101 icsk
= inet_csk(sk
);
4103 if (sk
->sk_write_pending
||
4104 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4105 icsk
->icsk_ack
.pingpong
) {
4106 /* Save one ACK. Data will be ready after
4107 * several ticks, if write_pending is set.
4109 * It may be deleted, but with this feature tcpdumps
4110 * look so _wonderfully_ clever, that I was not able
4111 * to stand against the temptation 8) --ANK
4113 inet_csk_schedule_ack(sk
);
4114 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4115 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4116 tcp_incr_quickack(sk
);
4117 tcp_enter_quickack_mode(sk
);
4118 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4119 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4130 /* No ACK in the segment */
4134 * "If the RST bit is set
4136 * Otherwise (no ACK) drop the segment and return."
4139 goto discard_and_undo
;
4143 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4144 goto discard_and_undo
;
4147 /* We see SYN without ACK. It is attempt of
4148 * simultaneous connect with crossed SYNs.
4149 * Particularly, it can be connect to self.
4151 tcp_set_state(sk
, TCP_SYN_RECV
);
4153 if (tp
->rx_opt
.saw_tstamp
) {
4154 tp
->rx_opt
.tstamp_ok
= 1;
4155 tcp_store_ts_recent(tp
);
4156 tp
->tcp_header_len
=
4157 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4159 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4162 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4163 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4165 /* RFC1323: The window in SYN & SYN/ACK segments is
4168 tp
->snd_wnd
= ntohs(th
->window
);
4169 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4170 tp
->max_window
= tp
->snd_wnd
;
4172 TCP_ECN_rcv_syn(tp
, th
);
4173 if (tp
->ecn_flags
&TCP_ECN_OK
)
4174 sock_set_flag(sk
, SOCK_NO_LARGESEND
);
4176 tcp_sync_mss(sk
, tp
->pmtu_cookie
);
4177 tcp_initialize_rcv_mss(sk
);
4180 tcp_send_synack(sk
);
4182 /* Note, we could accept data and URG from this segment.
4183 * There are no obstacles to make this.
4185 * However, if we ignore data in ACKless segments sometimes,
4186 * we have no reasons to accept it sometimes.
4187 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4188 * is not flawless. So, discard packet for sanity.
4189 * Uncomment this return to process the data.
4196 /* "fifth, if neither of the SYN or RST bits is set then
4197 * drop the segment and return."
4201 tcp_clear_options(&tp
->rx_opt
);
4202 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4206 tcp_clear_options(&tp
->rx_opt
);
4207 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4213 * This function implements the receiving procedure of RFC 793 for
4214 * all states except ESTABLISHED and TIME_WAIT.
4215 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4216 * address independent.
4219 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4220 struct tcphdr
*th
, unsigned len
)
4222 struct tcp_sock
*tp
= tcp_sk(sk
);
4225 tp
->rx_opt
.saw_tstamp
= 0;
4227 switch (sk
->sk_state
) {
4239 if(tp
->af_specific
->conn_request(sk
, skb
) < 0)
4242 /* Now we have several options: In theory there is
4243 * nothing else in the frame. KA9Q has an option to
4244 * send data with the syn, BSD accepts data with the
4245 * syn up to the [to be] advertised window and
4246 * Solaris 2.1 gives you a protocol error. For now
4247 * we just ignore it, that fits the spec precisely
4248 * and avoids incompatibilities. It would be nice in
4249 * future to drop through and process the data.
4251 * Now that TTCP is starting to be used we ought to
4253 * But, this leaves one open to an easy denial of
4254 * service attack, and SYN cookies can't defend
4255 * against this problem. So, we drop the data
4256 * in the interest of security over speed.
4263 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4267 /* Do step6 onward by hand. */
4268 tcp_urg(sk
, skb
, th
);
4270 tcp_data_snd_check(sk
, tp
);
4274 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4275 tcp_paws_discard(sk
, skb
)) {
4277 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4278 tcp_send_dupack(sk
, skb
);
4281 /* Reset is accepted even if it did not pass PAWS. */
4284 /* step 1: check sequence number */
4285 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4287 tcp_send_dupack(sk
, skb
);
4291 /* step 2: check RST bit */
4297 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4299 /* step 3: check security and precedence [ignored] */
4303 * Check for a SYN in window.
4305 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4306 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4311 /* step 5: check the ACK field */
4313 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4315 switch(sk
->sk_state
) {
4318 tp
->copied_seq
= tp
->rcv_nxt
;
4320 tcp_set_state(sk
, TCP_ESTABLISHED
);
4321 sk
->sk_state_change(sk
);
4323 /* Note, that this wakeup is only for marginal
4324 * crossed SYN case. Passively open sockets
4325 * are not waked up, because sk->sk_sleep ==
4326 * NULL and sk->sk_socket == NULL.
4328 if (sk
->sk_socket
) {
4329 sk_wake_async(sk
,0,POLL_OUT
);
4332 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4333 tp
->snd_wnd
= ntohs(th
->window
) <<
4334 tp
->rx_opt
.snd_wscale
;
4335 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4336 TCP_SKB_CB(skb
)->seq
);
4338 /* tcp_ack considers this ACK as duplicate
4339 * and does not calculate rtt.
4340 * Fix it at least with timestamps.
4342 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4344 tcp_ack_saw_tstamp(sk
, 0);
4346 if (tp
->rx_opt
.tstamp_ok
)
4347 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4349 /* Make sure socket is routed, for
4352 tp
->af_specific
->rebuild_header(sk
);
4354 tcp_init_metrics(sk
);
4356 tcp_init_congestion_control(sk
);
4358 /* Prevent spurious tcp_cwnd_restart() on
4359 * first data packet.
4361 tp
->lsndtime
= tcp_time_stamp
;
4363 tcp_initialize_rcv_mss(sk
);
4364 tcp_init_buffer_space(sk
);
4365 tcp_fast_path_on(tp
);
4372 if (tp
->snd_una
== tp
->write_seq
) {
4373 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4374 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4375 dst_confirm(sk
->sk_dst_cache
);
4377 if (!sock_flag(sk
, SOCK_DEAD
))
4378 /* Wake up lingering close() */
4379 sk
->sk_state_change(sk
);
4383 if (tp
->linger2
< 0 ||
4384 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4385 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4387 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4391 tmo
= tcp_fin_time(sk
);
4392 if (tmo
> TCP_TIMEWAIT_LEN
) {
4393 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4394 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4395 /* Bad case. We could lose such FIN otherwise.
4396 * It is not a big problem, but it looks confusing
4397 * and not so rare event. We still can lose it now,
4398 * if it spins in bh_lock_sock(), but it is really
4401 inet_csk_reset_keepalive_timer(sk
, tmo
);
4403 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4411 if (tp
->snd_una
== tp
->write_seq
) {
4412 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4418 if (tp
->snd_una
== tp
->write_seq
) {
4419 tcp_update_metrics(sk
);
4428 /* step 6: check the URG bit */
4429 tcp_urg(sk
, skb
, th
);
4431 /* step 7: process the segment text */
4432 switch (sk
->sk_state
) {
4433 case TCP_CLOSE_WAIT
:
4436 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4440 /* RFC 793 says to queue data in these states,
4441 * RFC 1122 says we MUST send a reset.
4442 * BSD 4.4 also does reset.
4444 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4445 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4446 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4447 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4453 case TCP_ESTABLISHED
:
4454 tcp_data_queue(sk
, skb
);
4459 /* tcp_data could move socket to TIME-WAIT */
4460 if (sk
->sk_state
!= TCP_CLOSE
) {
4461 tcp_data_snd_check(sk
, tp
);
4462 tcp_ack_snd_check(sk
);
4472 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4473 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4474 EXPORT_SYMBOL(sysctl_tcp_abc
);
4475 EXPORT_SYMBOL(tcp_parse_options
);
4476 EXPORT_SYMBOL(tcp_rcv_established
);
4477 EXPORT_SYMBOL(tcp_rcv_state_process
);