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
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.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 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 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
);
345 /* Initialize RCV_MSS value.
346 * RCV_MSS is an our guess about MSS used by the peer.
347 * We haven't any direct information about the MSS.
348 * It's better to underestimate the RCV_MSS rather than overestimate.
349 * Overestimations make us ACKing less frequently than needed.
350 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
352 void tcp_initialize_rcv_mss(struct sock
*sk
)
354 struct tcp_sock
*tp
= tcp_sk(sk
);
355 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
357 hint
= min(hint
, tp
->rcv_wnd
/2);
358 hint
= min(hint
, TCP_MIN_RCVMSS
);
359 hint
= max(hint
, TCP_MIN_MSS
);
361 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
364 /* Receiver "autotuning" code.
366 * The algorithm for RTT estimation w/o timestamps is based on
367 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
368 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
370 * More detail on this code can be found at
371 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
372 * though this reference is out of date. A new paper
375 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
377 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
383 if (new_sample
!= 0) {
384 /* If we sample in larger samples in the non-timestamp
385 * case, we could grossly overestimate the RTT especially
386 * with chatty applications or bulk transfer apps which
387 * are stalled on filesystem I/O.
389 * Also, since we are only going for a minimum in the
390 * non-timestamp case, we do not smooth things out
391 * else with timestamps disabled convergence takes too
395 m
-= (new_sample
>> 3);
397 } else if (m
< new_sample
)
400 /* No previous measure. */
404 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
405 tp
->rcv_rtt_est
.rtt
= new_sample
;
408 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
410 if (tp
->rcv_rtt_est
.time
== 0)
412 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
414 tcp_rcv_rtt_update(tp
,
415 jiffies
- tp
->rcv_rtt_est
.time
,
419 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
420 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
423 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
425 struct tcp_sock
*tp
= tcp_sk(sk
);
426 if (tp
->rx_opt
.rcv_tsecr
&&
427 (TCP_SKB_CB(skb
)->end_seq
-
428 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
429 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
433 * This function should be called every time data is copied to user space.
434 * It calculates the appropriate TCP receive buffer space.
436 void tcp_rcv_space_adjust(struct sock
*sk
)
438 struct tcp_sock
*tp
= tcp_sk(sk
);
442 if (tp
->rcvq_space
.time
== 0)
445 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
446 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
447 tp
->rcv_rtt_est
.rtt
== 0)
450 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
452 space
= max(tp
->rcvq_space
.space
, space
);
454 if (tp
->rcvq_space
.space
!= space
) {
457 tp
->rcvq_space
.space
= space
;
459 if (sysctl_tcp_moderate_rcvbuf
&&
460 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
461 int new_clamp
= space
;
463 /* Receive space grows, normalize in order to
464 * take into account packet headers and sk_buff
465 * structure overhead.
470 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
471 16 + sizeof(struct sk_buff
));
472 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
475 space
= min(space
, sysctl_tcp_rmem
[2]);
476 if (space
> sk
->sk_rcvbuf
) {
477 sk
->sk_rcvbuf
= space
;
479 /* Make the window clamp follow along. */
480 tp
->window_clamp
= new_clamp
;
486 tp
->rcvq_space
.seq
= tp
->copied_seq
;
487 tp
->rcvq_space
.time
= tcp_time_stamp
;
490 /* There is something which you must keep in mind when you analyze the
491 * behavior of the tp->ato delayed ack timeout interval. When a
492 * connection starts up, we want to ack as quickly as possible. The
493 * problem is that "good" TCP's do slow start at the beginning of data
494 * transmission. The means that until we send the first few ACK's the
495 * sender will sit on his end and only queue most of his data, because
496 * he can only send snd_cwnd unacked packets at any given time. For
497 * each ACK we send, he increments snd_cwnd and transmits more of his
500 static void tcp_event_data_recv(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
502 struct inet_connection_sock
*icsk
= inet_csk(sk
);
505 inet_csk_schedule_ack(sk
);
507 tcp_measure_rcv_mss(sk
, skb
);
509 tcp_rcv_rtt_measure(tp
);
511 now
= tcp_time_stamp
;
513 if (!icsk
->icsk_ack
.ato
) {
514 /* The _first_ data packet received, initialize
515 * delayed ACK engine.
517 tcp_incr_quickack(sk
);
518 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
520 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
522 if (m
<= TCP_ATO_MIN
/2) {
523 /* The fastest case is the first. */
524 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
525 } else if (m
< icsk
->icsk_ack
.ato
) {
526 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
527 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
528 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
529 } else if (m
> icsk
->icsk_rto
) {
530 /* Too long gap. Apparently sender failed to
531 * restart window, so that we send ACKs quickly.
533 tcp_incr_quickack(sk
);
534 sk_stream_mem_reclaim(sk
);
537 icsk
->icsk_ack
.lrcvtime
= now
;
539 TCP_ECN_check_ce(tp
, skb
);
542 tcp_grow_window(sk
, tp
, skb
);
545 /* Called to compute a smoothed rtt estimate. The data fed to this
546 * routine either comes from timestamps, or from segments that were
547 * known _not_ to have been retransmitted [see Karn/Partridge
548 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
549 * piece by Van Jacobson.
550 * NOTE: the next three routines used to be one big routine.
551 * To save cycles in the RFC 1323 implementation it was better to break
552 * it up into three procedures. -- erics
554 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
556 struct tcp_sock
*tp
= tcp_sk(sk
);
557 long m
= mrtt
; /* RTT */
559 /* The following amusing code comes from Jacobson's
560 * article in SIGCOMM '88. Note that rtt and mdev
561 * are scaled versions of rtt and mean deviation.
562 * This is designed to be as fast as possible
563 * m stands for "measurement".
565 * On a 1990 paper the rto value is changed to:
566 * RTO = rtt + 4 * mdev
568 * Funny. This algorithm seems to be very broken.
569 * These formulae increase RTO, when it should be decreased, increase
570 * too slowly, when it should be increased quickly, decrease too quickly
571 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
572 * does not matter how to _calculate_ it. Seems, it was trap
573 * that VJ failed to avoid. 8)
578 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
579 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
581 m
= -m
; /* m is now abs(error) */
582 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
583 /* This is similar to one of Eifel findings.
584 * Eifel blocks mdev updates when rtt decreases.
585 * This solution is a bit different: we use finer gain
586 * for mdev in this case (alpha*beta).
587 * Like Eifel it also prevents growth of rto,
588 * but also it limits too fast rto decreases,
589 * happening in pure Eifel.
594 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
596 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
597 if (tp
->mdev
> tp
->mdev_max
) {
598 tp
->mdev_max
= tp
->mdev
;
599 if (tp
->mdev_max
> tp
->rttvar
)
600 tp
->rttvar
= tp
->mdev_max
;
602 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
603 if (tp
->mdev_max
< tp
->rttvar
)
604 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
605 tp
->rtt_seq
= tp
->snd_nxt
;
606 tp
->mdev_max
= TCP_RTO_MIN
;
609 /* no previous measure. */
610 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
611 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
612 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
613 tp
->rtt_seq
= tp
->snd_nxt
;
617 /* Calculate rto without backoff. This is the second half of Van Jacobson's
618 * routine referred to above.
620 static inline void tcp_set_rto(struct sock
*sk
)
622 const struct tcp_sock
*tp
= tcp_sk(sk
);
623 /* Old crap is replaced with new one. 8)
626 * 1. If rtt variance happened to be less 50msec, it is hallucination.
627 * It cannot be less due to utterly erratic ACK generation made
628 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
629 * to do with delayed acks, because at cwnd>2 true delack timeout
630 * is invisible. Actually, Linux-2.4 also generates erratic
631 * ACKs in some circumstances.
633 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
635 /* 2. Fixups made earlier cannot be right.
636 * If we do not estimate RTO correctly without them,
637 * all the algo is pure shit and should be replaced
638 * with correct one. It is exactly, which we pretend to do.
642 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
643 * guarantees that rto is higher.
645 static inline void tcp_bound_rto(struct sock
*sk
)
647 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
648 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
651 /* Save metrics learned by this TCP session.
652 This function is called only, when TCP finishes successfully
653 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
655 void tcp_update_metrics(struct sock
*sk
)
657 struct tcp_sock
*tp
= tcp_sk(sk
);
658 struct dst_entry
*dst
= __sk_dst_get(sk
);
660 if (sysctl_tcp_nometrics_save
)
665 if (dst
&& (dst
->flags
&DST_HOST
)) {
666 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
669 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
670 /* This session failed to estimate rtt. Why?
671 * Probably, no packets returned in time.
674 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
675 dst
->metrics
[RTAX_RTT
-1] = 0;
679 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
681 /* If newly calculated rtt larger than stored one,
682 * store new one. Otherwise, use EWMA. Remember,
683 * rtt overestimation is always better than underestimation.
685 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
687 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
689 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
692 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
696 /* Scale deviation to rttvar fixed point */
701 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
702 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
704 dst
->metrics
[RTAX_RTTVAR
-1] -=
705 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
708 if (tp
->snd_ssthresh
>= 0xFFFF) {
709 /* Slow start still did not finish. */
710 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
711 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
712 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
713 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
714 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
715 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
716 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
717 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
718 icsk
->icsk_ca_state
== TCP_CA_Open
) {
719 /* Cong. avoidance phase, cwnd is reliable. */
720 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
721 dst
->metrics
[RTAX_SSTHRESH
-1] =
722 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
723 if (!dst_metric_locked(dst
, RTAX_CWND
))
724 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
726 /* Else slow start did not finish, cwnd is non-sense,
727 ssthresh may be also invalid.
729 if (!dst_metric_locked(dst
, RTAX_CWND
))
730 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
731 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
732 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
733 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
734 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
737 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
738 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
739 tp
->reordering
!= sysctl_tcp_reordering
)
740 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
745 /* Numbers are taken from RFC2414. */
746 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
748 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
751 if (tp
->mss_cache
> 1460)
754 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
756 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
759 /* Set slow start threshold and cwnd not falling to slow start */
760 void tcp_enter_cwr(struct sock
*sk
)
762 struct tcp_sock
*tp
= tcp_sk(sk
);
764 tp
->prior_ssthresh
= 0;
766 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
768 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
769 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
770 tcp_packets_in_flight(tp
) + 1U);
771 tp
->snd_cwnd_cnt
= 0;
772 tp
->high_seq
= tp
->snd_nxt
;
773 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
774 TCP_ECN_queue_cwr(tp
);
776 tcp_set_ca_state(sk
, TCP_CA_CWR
);
780 /* Initialize metrics on socket. */
782 static void tcp_init_metrics(struct sock
*sk
)
784 struct tcp_sock
*tp
= tcp_sk(sk
);
785 struct dst_entry
*dst
= __sk_dst_get(sk
);
792 if (dst_metric_locked(dst
, RTAX_CWND
))
793 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
794 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
795 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
796 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
797 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
799 if (dst_metric(dst
, RTAX_REORDERING
) &&
800 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
801 tp
->rx_opt
.sack_ok
&= ~2;
802 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
805 if (dst_metric(dst
, RTAX_RTT
) == 0)
808 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
811 /* Initial rtt is determined from SYN,SYN-ACK.
812 * The segment is small and rtt may appear much
813 * less than real one. Use per-dst memory
814 * to make it more realistic.
816 * A bit of theory. RTT is time passed after "normal" sized packet
817 * is sent until it is ACKed. In normal circumstances sending small
818 * packets force peer to delay ACKs and calculation is correct too.
819 * The algorithm is adaptive and, provided we follow specs, it
820 * NEVER underestimate RTT. BUT! If peer tries to make some clever
821 * tricks sort of "quick acks" for time long enough to decrease RTT
822 * to low value, and then abruptly stops to do it and starts to delay
823 * ACKs, wait for troubles.
825 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
826 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
827 tp
->rtt_seq
= tp
->snd_nxt
;
829 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
830 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
831 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
835 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
837 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
838 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
842 /* Play conservative. If timestamps are not
843 * supported, TCP will fail to recalculate correct
844 * rtt, if initial rto is too small. FORGET ALL AND RESET!
846 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
848 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
849 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
853 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
856 struct tcp_sock
*tp
= tcp_sk(sk
);
857 if (metric
> tp
->reordering
) {
858 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
860 /* This exciting event is worth to be remembered. 8) */
862 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
864 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
866 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
868 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
869 #if FASTRETRANS_DEBUG > 1
870 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
871 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
875 tp
->undo_marker
? tp
->undo_retrans
: 0);
877 /* Disable FACK yet. */
878 tp
->rx_opt
.sack_ok
&= ~2;
882 /* This procedure tags the retransmission queue when SACKs arrive.
884 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
885 * Packets in queue with these bits set are counted in variables
886 * sacked_out, retrans_out and lost_out, correspondingly.
888 * Valid combinations are:
889 * Tag InFlight Description
890 * 0 1 - orig segment is in flight.
891 * S 0 - nothing flies, orig reached receiver.
892 * L 0 - nothing flies, orig lost by net.
893 * R 2 - both orig and retransmit are in flight.
894 * L|R 1 - orig is lost, retransmit is in flight.
895 * S|R 1 - orig reached receiver, retrans is still in flight.
896 * (L|S|R is logically valid, it could occur when L|R is sacked,
897 * but it is equivalent to plain S and code short-curcuits it to S.
898 * L|S is logically invalid, it would mean -1 packet in flight 8))
900 * These 6 states form finite state machine, controlled by the following events:
901 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
902 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
903 * 3. Loss detection event of one of three flavors:
904 * A. Scoreboard estimator decided the packet is lost.
905 * A'. Reno "three dupacks" marks head of queue lost.
906 * A''. Its FACK modfication, head until snd.fack is lost.
907 * B. SACK arrives sacking data transmitted after never retransmitted
909 * C. SACK arrives sacking SND.NXT at the moment, when the
910 * segment was retransmitted.
911 * 4. D-SACK added new rule: D-SACK changes any tag to S.
913 * It is pleasant to note, that state diagram turns out to be commutative,
914 * so that we are allowed not to be bothered by order of our actions,
915 * when multiple events arrive simultaneously. (see the function below).
917 * Reordering detection.
918 * --------------------
919 * Reordering metric is maximal distance, which a packet can be displaced
920 * in packet stream. With SACKs we can estimate it:
922 * 1. SACK fills old hole and the corresponding segment was not
923 * ever retransmitted -> reordering. Alas, we cannot use it
924 * when segment was retransmitted.
925 * 2. The last flaw is solved with D-SACK. D-SACK arrives
926 * for retransmitted and already SACKed segment -> reordering..
927 * Both of these heuristics are not used in Loss state, when we cannot
928 * account for retransmits accurately.
931 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
933 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
934 struct tcp_sock
*tp
= tcp_sk(sk
);
935 unsigned char *ptr
= ack_skb
->h
.raw
+ TCP_SKB_CB(ack_skb
)->sacked
;
936 struct tcp_sack_block
*sp
= (struct tcp_sack_block
*)(ptr
+2);
937 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
938 int reord
= tp
->packets_out
;
940 u32 lost_retrans
= 0;
947 prior_fackets
= tp
->fackets_out
;
950 * if the only SACK change is the increase of the end_seq of
951 * the first block then only apply that SACK block
952 * and use retrans queue hinting otherwise slowpath */
954 for (i
= 0; i
< num_sacks
; i
++) {
955 __u32 start_seq
= ntohl(sp
[i
].start_seq
);
956 __u32 end_seq
= ntohl(sp
[i
].end_seq
);
959 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
962 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
963 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
966 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
967 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
969 /* Check for D-SACK. */
971 u32 ack
= TCP_SKB_CB(ack_skb
)->ack_seq
;
973 if (before(start_seq
, ack
)) {
975 tp
->rx_opt
.sack_ok
|= 4;
976 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
977 } else if (num_sacks
> 1 &&
978 !after(end_seq
, ntohl(sp
[1].end_seq
)) &&
979 !before(start_seq
, ntohl(sp
[1].start_seq
))) {
981 tp
->rx_opt
.sack_ok
|= 4;
982 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
985 /* D-SACK for already forgotten data...
986 * Do dumb counting. */
988 !after(end_seq
, prior_snd_una
) &&
989 after(end_seq
, tp
->undo_marker
))
992 /* Eliminate too old ACKs, but take into
993 * account more or less fresh ones, they can
994 * contain valid SACK info.
996 if (before(ack
, prior_snd_una
- tp
->max_window
))
1005 tp
->fastpath_skb_hint
= NULL
;
1007 /* order SACK blocks to allow in order walk of the retrans queue */
1008 for (i
= num_sacks
-1; i
> 0; i
--) {
1009 for (j
= 0; j
< i
; j
++){
1010 if (after(ntohl(sp
[j
].start_seq
),
1011 ntohl(sp
[j
+1].start_seq
))){
1012 sp
[j
].start_seq
= htonl(tp
->recv_sack_cache
[j
+1].start_seq
);
1013 sp
[j
].end_seq
= htonl(tp
->recv_sack_cache
[j
+1].end_seq
);
1014 sp
[j
+1].start_seq
= htonl(tp
->recv_sack_cache
[j
].start_seq
);
1015 sp
[j
+1].end_seq
= htonl(tp
->recv_sack_cache
[j
].end_seq
);
1022 /* clear flag as used for different purpose in following code */
1025 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1026 struct sk_buff
*skb
;
1027 __u32 start_seq
= ntohl(sp
->start_seq
);
1028 __u32 end_seq
= ntohl(sp
->end_seq
);
1031 /* Use SACK fastpath hint if valid */
1032 if (tp
->fastpath_skb_hint
) {
1033 skb
= tp
->fastpath_skb_hint
;
1034 fack_count
= tp
->fastpath_cnt_hint
;
1036 skb
= sk
->sk_write_queue
.next
;
1040 /* Event "B" in the comment above. */
1041 if (after(end_seq
, tp
->high_seq
))
1042 flag
|= FLAG_DATA_LOST
;
1044 sk_stream_for_retrans_queue_from(skb
, sk
) {
1045 int in_sack
, pcount
;
1048 tp
->fastpath_skb_hint
= skb
;
1049 tp
->fastpath_cnt_hint
= fack_count
;
1051 /* The retransmission queue is always in order, so
1052 * we can short-circuit the walk early.
1054 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1057 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1058 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1060 pcount
= tcp_skb_pcount(skb
);
1062 if (pcount
> 1 && !in_sack
&&
1063 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1064 unsigned int pkt_len
;
1066 in_sack
= !after(start_seq
,
1067 TCP_SKB_CB(skb
)->seq
);
1070 pkt_len
= (start_seq
-
1071 TCP_SKB_CB(skb
)->seq
);
1073 pkt_len
= (end_seq
-
1074 TCP_SKB_CB(skb
)->seq
);
1075 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1077 pcount
= tcp_skb_pcount(skb
);
1080 fack_count
+= pcount
;
1082 sacked
= TCP_SKB_CB(skb
)->sacked
;
1084 /* Account D-SACK for retransmitted packet. */
1085 if ((dup_sack
&& in_sack
) &&
1086 (sacked
& TCPCB_RETRANS
) &&
1087 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1090 /* The frame is ACKed. */
1091 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1092 if (sacked
&TCPCB_RETRANS
) {
1093 if ((dup_sack
&& in_sack
) &&
1094 (sacked
&TCPCB_SACKED_ACKED
))
1095 reord
= min(fack_count
, reord
);
1097 /* If it was in a hole, we detected reordering. */
1098 if (fack_count
< prior_fackets
&&
1099 !(sacked
&TCPCB_SACKED_ACKED
))
1100 reord
= min(fack_count
, reord
);
1103 /* Nothing to do; acked frame is about to be dropped. */
1107 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1108 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1109 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1110 lost_retrans
= end_seq
;
1115 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1116 if (sacked
& TCPCB_SACKED_RETRANS
) {
1117 /* If the segment is not tagged as lost,
1118 * we do not clear RETRANS, believing
1119 * that retransmission is still in flight.
1121 if (sacked
& TCPCB_LOST
) {
1122 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1123 tp
->lost_out
-= tcp_skb_pcount(skb
);
1124 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1126 /* clear lost hint */
1127 tp
->retransmit_skb_hint
= NULL
;
1130 /* New sack for not retransmitted frame,
1131 * which was in hole. It is reordering.
1133 if (!(sacked
& TCPCB_RETRANS
) &&
1134 fack_count
< prior_fackets
)
1135 reord
= min(fack_count
, reord
);
1137 if (sacked
& TCPCB_LOST
) {
1138 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1139 tp
->lost_out
-= tcp_skb_pcount(skb
);
1141 /* clear lost hint */
1142 tp
->retransmit_skb_hint
= NULL
;
1146 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1147 flag
|= FLAG_DATA_SACKED
;
1148 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1150 if (fack_count
> tp
->fackets_out
)
1151 tp
->fackets_out
= fack_count
;
1153 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1154 reord
= min(fack_count
, reord
);
1157 /* D-SACK. We can detect redundant retransmission
1158 * in S|R and plain R frames and clear it.
1159 * undo_retrans is decreased above, L|R frames
1160 * are accounted above as well.
1163 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1164 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1165 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1166 tp
->retransmit_skb_hint
= NULL
;
1171 /* Check for lost retransmit. This superb idea is
1172 * borrowed from "ratehalving". Event "C".
1173 * Later note: FACK people cheated me again 8),
1174 * we have to account for reordering! Ugly,
1177 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1178 struct sk_buff
*skb
;
1180 sk_stream_for_retrans_queue(skb
, sk
) {
1181 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1183 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1185 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1186 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1188 !before(lost_retrans
,
1189 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1191 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1192 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1194 /* clear lost hint */
1195 tp
->retransmit_skb_hint
= NULL
;
1197 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1198 tp
->lost_out
+= tcp_skb_pcount(skb
);
1199 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1200 flag
|= FLAG_DATA_SACKED
;
1201 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1207 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1209 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1210 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1212 #if FASTRETRANS_DEBUG > 0
1213 BUG_TRAP((int)tp
->sacked_out
>= 0);
1214 BUG_TRAP((int)tp
->lost_out
>= 0);
1215 BUG_TRAP((int)tp
->retrans_out
>= 0);
1216 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1221 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1222 * segments to see from the next ACKs whether any data was really missing.
1223 * If the RTO was spurious, new ACKs should arrive.
1225 void tcp_enter_frto(struct sock
*sk
)
1227 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1228 struct tcp_sock
*tp
= tcp_sk(sk
);
1229 struct sk_buff
*skb
;
1231 tp
->frto_counter
= 1;
1233 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1234 tp
->snd_una
== tp
->high_seq
||
1235 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1236 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1237 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1238 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1241 /* Have to clear retransmission markers here to keep the bookkeeping
1242 * in shape, even though we are not yet in Loss state.
1243 * If something was really lost, it is eventually caught up
1244 * in tcp_enter_frto_loss.
1246 tp
->retrans_out
= 0;
1247 tp
->undo_marker
= tp
->snd_una
;
1248 tp
->undo_retrans
= 0;
1250 sk_stream_for_retrans_queue(skb
, sk
) {
1251 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_RETRANS
;
1253 tcp_sync_left_out(tp
);
1255 tcp_set_ca_state(sk
, TCP_CA_Open
);
1256 tp
->frto_highmark
= tp
->snd_nxt
;
1259 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1260 * which indicates that we should follow the traditional RTO recovery,
1261 * i.e. mark everything lost and do go-back-N retransmission.
1263 static void tcp_enter_frto_loss(struct sock
*sk
)
1265 struct tcp_sock
*tp
= tcp_sk(sk
);
1266 struct sk_buff
*skb
;
1271 tp
->fackets_out
= 0;
1273 sk_stream_for_retrans_queue(skb
, sk
) {
1274 cnt
+= tcp_skb_pcount(skb
);
1275 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1276 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1278 /* Do not mark those segments lost that were
1279 * forward transmitted after RTO
1281 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1282 tp
->frto_highmark
)) {
1283 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1284 tp
->lost_out
+= tcp_skb_pcount(skb
);
1287 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1288 tp
->fackets_out
= cnt
;
1291 tcp_sync_left_out(tp
);
1293 tp
->snd_cwnd
= tp
->frto_counter
+ tcp_packets_in_flight(tp
)+1;
1294 tp
->snd_cwnd_cnt
= 0;
1295 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1296 tp
->undo_marker
= 0;
1297 tp
->frto_counter
= 0;
1299 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1300 sysctl_tcp_reordering
);
1301 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1302 tp
->high_seq
= tp
->frto_highmark
;
1303 TCP_ECN_queue_cwr(tp
);
1305 clear_all_retrans_hints(tp
);
1308 void tcp_clear_retrans(struct tcp_sock
*tp
)
1311 tp
->retrans_out
= 0;
1313 tp
->fackets_out
= 0;
1317 tp
->undo_marker
= 0;
1318 tp
->undo_retrans
= 0;
1321 /* Enter Loss state. If "how" is not zero, forget all SACK information
1322 * and reset tags completely, otherwise preserve SACKs. If receiver
1323 * dropped its ofo queue, we will know this due to reneging detection.
1325 void tcp_enter_loss(struct sock
*sk
, int how
)
1327 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1328 struct tcp_sock
*tp
= tcp_sk(sk
);
1329 struct sk_buff
*skb
;
1332 /* Reduce ssthresh if it has not yet been made inside this window. */
1333 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1334 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1335 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1336 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1337 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1340 tp
->snd_cwnd_cnt
= 0;
1341 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1343 tp
->bytes_acked
= 0;
1344 tcp_clear_retrans(tp
);
1346 /* Push undo marker, if it was plain RTO and nothing
1347 * was retransmitted. */
1349 tp
->undo_marker
= tp
->snd_una
;
1351 sk_stream_for_retrans_queue(skb
, sk
) {
1352 cnt
+= tcp_skb_pcount(skb
);
1353 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1354 tp
->undo_marker
= 0;
1355 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1356 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1357 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1358 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1359 tp
->lost_out
+= tcp_skb_pcount(skb
);
1361 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1362 tp
->fackets_out
= cnt
;
1365 tcp_sync_left_out(tp
);
1367 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1368 sysctl_tcp_reordering
);
1369 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1370 tp
->high_seq
= tp
->snd_nxt
;
1371 TCP_ECN_queue_cwr(tp
);
1373 clear_all_retrans_hints(tp
);
1376 static int tcp_check_sack_reneging(struct sock
*sk
)
1378 struct sk_buff
*skb
;
1380 /* If ACK arrived pointing to a remembered SACK,
1381 * it means that our remembered SACKs do not reflect
1382 * real state of receiver i.e.
1383 * receiver _host_ is heavily congested (or buggy).
1384 * Do processing similar to RTO timeout.
1386 if ((skb
= skb_peek(&sk
->sk_write_queue
)) != NULL
&&
1387 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1388 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1389 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1391 tcp_enter_loss(sk
, 1);
1392 icsk
->icsk_retransmits
++;
1393 tcp_retransmit_skb(sk
, skb_peek(&sk
->sk_write_queue
));
1394 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1395 icsk
->icsk_rto
, TCP_RTO_MAX
);
1401 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1403 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1406 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1408 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1411 static inline int tcp_head_timedout(struct sock
*sk
, struct tcp_sock
*tp
)
1413 return tp
->packets_out
&&
1414 tcp_skb_timedout(sk
, skb_peek(&sk
->sk_write_queue
));
1417 /* Linux NewReno/SACK/FACK/ECN state machine.
1418 * --------------------------------------
1420 * "Open" Normal state, no dubious events, fast path.
1421 * "Disorder" In all the respects it is "Open",
1422 * but requires a bit more attention. It is entered when
1423 * we see some SACKs or dupacks. It is split of "Open"
1424 * mainly to move some processing from fast path to slow one.
1425 * "CWR" CWND was reduced due to some Congestion Notification event.
1426 * It can be ECN, ICMP source quench, local device congestion.
1427 * "Recovery" CWND was reduced, we are fast-retransmitting.
1428 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1430 * tcp_fastretrans_alert() is entered:
1431 * - each incoming ACK, if state is not "Open"
1432 * - when arrived ACK is unusual, namely:
1437 * Counting packets in flight is pretty simple.
1439 * in_flight = packets_out - left_out + retrans_out
1441 * packets_out is SND.NXT-SND.UNA counted in packets.
1443 * retrans_out is number of retransmitted segments.
1445 * left_out is number of segments left network, but not ACKed yet.
1447 * left_out = sacked_out + lost_out
1449 * sacked_out: Packets, which arrived to receiver out of order
1450 * and hence not ACKed. With SACKs this number is simply
1451 * amount of SACKed data. Even without SACKs
1452 * it is easy to give pretty reliable estimate of this number,
1453 * counting duplicate ACKs.
1455 * lost_out: Packets lost by network. TCP has no explicit
1456 * "loss notification" feedback from network (for now).
1457 * It means that this number can be only _guessed_.
1458 * Actually, it is the heuristics to predict lossage that
1459 * distinguishes different algorithms.
1461 * F.e. after RTO, when all the queue is considered as lost,
1462 * lost_out = packets_out and in_flight = retrans_out.
1464 * Essentially, we have now two algorithms counting
1467 * FACK: It is the simplest heuristics. As soon as we decided
1468 * that something is lost, we decide that _all_ not SACKed
1469 * packets until the most forward SACK are lost. I.e.
1470 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1471 * It is absolutely correct estimate, if network does not reorder
1472 * packets. And it loses any connection to reality when reordering
1473 * takes place. We use FACK by default until reordering
1474 * is suspected on the path to this destination.
1476 * NewReno: when Recovery is entered, we assume that one segment
1477 * is lost (classic Reno). While we are in Recovery and
1478 * a partial ACK arrives, we assume that one more packet
1479 * is lost (NewReno). This heuristics are the same in NewReno
1482 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1483 * deflation etc. CWND is real congestion window, never inflated, changes
1484 * only according to classic VJ rules.
1486 * Really tricky (and requiring careful tuning) part of algorithm
1487 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1488 * The first determines the moment _when_ we should reduce CWND and,
1489 * hence, slow down forward transmission. In fact, it determines the moment
1490 * when we decide that hole is caused by loss, rather than by a reorder.
1492 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1493 * holes, caused by lost packets.
1495 * And the most logically complicated part of algorithm is undo
1496 * heuristics. We detect false retransmits due to both too early
1497 * fast retransmit (reordering) and underestimated RTO, analyzing
1498 * timestamps and D-SACKs. When we detect that some segments were
1499 * retransmitted by mistake and CWND reduction was wrong, we undo
1500 * window reduction and abort recovery phase. This logic is hidden
1501 * inside several functions named tcp_try_undo_<something>.
1504 /* This function decides, when we should leave Disordered state
1505 * and enter Recovery phase, reducing congestion window.
1507 * Main question: may we further continue forward transmission
1508 * with the same cwnd?
1510 static int tcp_time_to_recover(struct sock
*sk
, struct tcp_sock
*tp
)
1514 /* Trick#1: The loss is proven. */
1518 /* Not-A-Trick#2 : Classic rule... */
1519 if (tcp_fackets_out(tp
) > tp
->reordering
)
1522 /* Trick#3 : when we use RFC2988 timer restart, fast
1523 * retransmit can be triggered by timeout of queue head.
1525 if (tcp_head_timedout(sk
, tp
))
1528 /* Trick#4: It is still not OK... But will it be useful to delay
1531 packets_out
= tp
->packets_out
;
1532 if (packets_out
<= tp
->reordering
&&
1533 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1534 !tcp_may_send_now(sk
, tp
)) {
1535 /* We have nothing to send. This connection is limited
1536 * either by receiver window or by application.
1544 /* If we receive more dupacks than we expected counting segments
1545 * in assumption of absent reordering, interpret this as reordering.
1546 * The only another reason could be bug in receiver TCP.
1548 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1550 struct tcp_sock
*tp
= tcp_sk(sk
);
1553 holes
= max(tp
->lost_out
, 1U);
1554 holes
= min(holes
, tp
->packets_out
);
1556 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1557 tp
->sacked_out
= tp
->packets_out
- holes
;
1558 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1562 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1564 static void tcp_add_reno_sack(struct sock
*sk
)
1566 struct tcp_sock
*tp
= tcp_sk(sk
);
1568 tcp_check_reno_reordering(sk
, 0);
1569 tcp_sync_left_out(tp
);
1572 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1574 static void tcp_remove_reno_sacks(struct sock
*sk
, struct tcp_sock
*tp
, int acked
)
1577 /* One ACK acked hole. The rest eat duplicate ACKs. */
1578 if (acked
-1 >= tp
->sacked_out
)
1581 tp
->sacked_out
-= acked
-1;
1583 tcp_check_reno_reordering(sk
, acked
);
1584 tcp_sync_left_out(tp
);
1587 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1590 tp
->left_out
= tp
->lost_out
;
1593 /* Mark head of queue up as lost. */
1594 static void tcp_mark_head_lost(struct sock
*sk
, struct tcp_sock
*tp
,
1595 int packets
, u32 high_seq
)
1597 struct sk_buff
*skb
;
1600 BUG_TRAP(packets
<= tp
->packets_out
);
1601 if (tp
->lost_skb_hint
) {
1602 skb
= tp
->lost_skb_hint
;
1603 cnt
= tp
->lost_cnt_hint
;
1605 skb
= sk
->sk_write_queue
.next
;
1609 sk_stream_for_retrans_queue_from(skb
, sk
) {
1610 /* TODO: do this better */
1611 /* this is not the most efficient way to do this... */
1612 tp
->lost_skb_hint
= skb
;
1613 tp
->lost_cnt_hint
= cnt
;
1614 cnt
+= tcp_skb_pcount(skb
);
1615 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1617 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1618 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1619 tp
->lost_out
+= tcp_skb_pcount(skb
);
1621 /* clear xmit_retransmit_queue hints
1622 * if this is beyond hint */
1623 if(tp
->retransmit_skb_hint
!= NULL
&&
1624 before(TCP_SKB_CB(skb
)->seq
,
1625 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
)) {
1627 tp
->retransmit_skb_hint
= NULL
;
1631 tcp_sync_left_out(tp
);
1634 /* Account newly detected lost packet(s) */
1636 static void tcp_update_scoreboard(struct sock
*sk
, struct tcp_sock
*tp
)
1639 int lost
= tp
->fackets_out
- tp
->reordering
;
1642 tcp_mark_head_lost(sk
, tp
, lost
, tp
->high_seq
);
1644 tcp_mark_head_lost(sk
, tp
, 1, tp
->high_seq
);
1647 /* New heuristics: it is possible only after we switched
1648 * to restart timer each time when something is ACKed.
1649 * Hence, we can detect timed out packets during fast
1650 * retransmit without falling to slow start.
1652 if (!IsReno(tp
) && tcp_head_timedout(sk
, tp
)) {
1653 struct sk_buff
*skb
;
1655 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1656 : sk
->sk_write_queue
.next
;
1658 sk_stream_for_retrans_queue_from(skb
, sk
) {
1659 if (!tcp_skb_timedout(sk
, skb
))
1662 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1663 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1664 tp
->lost_out
+= tcp_skb_pcount(skb
);
1666 /* clear xmit_retrans hint */
1667 if (tp
->retransmit_skb_hint
&&
1668 before(TCP_SKB_CB(skb
)->seq
,
1669 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1671 tp
->retransmit_skb_hint
= NULL
;
1675 tp
->scoreboard_skb_hint
= skb
;
1677 tcp_sync_left_out(tp
);
1681 /* CWND moderation, preventing bursts due to too big ACKs
1682 * in dubious situations.
1684 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1686 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1687 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1688 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1691 /* Lower bound on congestion window is slow start threshold
1692 * unless congestion avoidance choice decides to overide it.
1694 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1696 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1698 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1701 /* Decrease cwnd each second ack. */
1702 static void tcp_cwnd_down(struct sock
*sk
)
1704 struct tcp_sock
*tp
= tcp_sk(sk
);
1705 int decr
= tp
->snd_cwnd_cnt
+ 1;
1707 tp
->snd_cwnd_cnt
= decr
&1;
1710 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1711 tp
->snd_cwnd
-= decr
;
1713 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1714 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1717 /* Nothing was retransmitted or returned timestamp is less
1718 * than timestamp of the first retransmission.
1720 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1722 return !tp
->retrans_stamp
||
1723 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1724 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1727 /* Undo procedures. */
1729 #if FASTRETRANS_DEBUG > 1
1730 static void DBGUNDO(struct sock
*sk
, struct tcp_sock
*tp
, const char *msg
)
1732 struct inet_sock
*inet
= inet_sk(sk
);
1733 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1735 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1736 tp
->snd_cwnd
, tp
->left_out
,
1737 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1741 #define DBGUNDO(x...) do { } while (0)
1744 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1746 struct tcp_sock
*tp
= tcp_sk(sk
);
1748 if (tp
->prior_ssthresh
) {
1749 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1751 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1752 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1754 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1756 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1757 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1758 TCP_ECN_withdraw_cwr(tp
);
1761 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1763 tcp_moderate_cwnd(tp
);
1764 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1766 /* There is something screwy going on with the retrans hints after
1768 clear_all_retrans_hints(tp
);
1771 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1773 return tp
->undo_marker
&&
1774 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1777 /* People celebrate: "We love our President!" */
1778 static int tcp_try_undo_recovery(struct sock
*sk
, struct tcp_sock
*tp
)
1780 if (tcp_may_undo(tp
)) {
1781 /* Happy end! We did not retransmit anything
1782 * or our original transmission succeeded.
1784 DBGUNDO(sk
, tp
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1785 tcp_undo_cwr(sk
, 1);
1786 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1787 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1789 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1790 tp
->undo_marker
= 0;
1792 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1793 /* Hold old state until something *above* high_seq
1794 * is ACKed. For Reno it is MUST to prevent false
1795 * fast retransmits (RFC2582). SACK TCP is safe. */
1796 tcp_moderate_cwnd(tp
);
1799 tcp_set_ca_state(sk
, TCP_CA_Open
);
1803 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1804 static void tcp_try_undo_dsack(struct sock
*sk
, struct tcp_sock
*tp
)
1806 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1807 DBGUNDO(sk
, tp
, "D-SACK");
1808 tcp_undo_cwr(sk
, 1);
1809 tp
->undo_marker
= 0;
1810 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1814 /* Undo during fast recovery after partial ACK. */
1816 static int tcp_try_undo_partial(struct sock
*sk
, struct tcp_sock
*tp
,
1819 /* Partial ACK arrived. Force Hoe's retransmit. */
1820 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1822 if (tcp_may_undo(tp
)) {
1823 /* Plain luck! Hole if filled with delayed
1824 * packet, rather than with a retransmit.
1826 if (tp
->retrans_out
== 0)
1827 tp
->retrans_stamp
= 0;
1829 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
1831 DBGUNDO(sk
, tp
, "Hoe");
1832 tcp_undo_cwr(sk
, 0);
1833 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1835 /* So... Do not make Hoe's retransmit yet.
1836 * If the first packet was delayed, the rest
1837 * ones are most probably delayed as well.
1844 /* Undo during loss recovery after partial ACK. */
1845 static int tcp_try_undo_loss(struct sock
*sk
, struct tcp_sock
*tp
)
1847 if (tcp_may_undo(tp
)) {
1848 struct sk_buff
*skb
;
1849 sk_stream_for_retrans_queue(skb
, sk
) {
1850 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1853 clear_all_retrans_hints(tp
);
1855 DBGUNDO(sk
, tp
, "partial loss");
1857 tp
->left_out
= tp
->sacked_out
;
1858 tcp_undo_cwr(sk
, 1);
1859 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1860 inet_csk(sk
)->icsk_retransmits
= 0;
1861 tp
->undo_marker
= 0;
1863 tcp_set_ca_state(sk
, TCP_CA_Open
);
1869 static inline void tcp_complete_cwr(struct sock
*sk
)
1871 struct tcp_sock
*tp
= tcp_sk(sk
);
1872 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1873 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1874 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
1877 static void tcp_try_to_open(struct sock
*sk
, struct tcp_sock
*tp
, int flag
)
1879 tp
->left_out
= tp
->sacked_out
;
1881 if (tp
->retrans_out
== 0)
1882 tp
->retrans_stamp
= 0;
1887 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
1888 int state
= TCP_CA_Open
;
1890 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
1891 state
= TCP_CA_Disorder
;
1893 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
1894 tcp_set_ca_state(sk
, state
);
1895 tp
->high_seq
= tp
->snd_nxt
;
1897 tcp_moderate_cwnd(tp
);
1903 static void tcp_mtup_probe_failed(struct sock
*sk
)
1905 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1907 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
1908 icsk
->icsk_mtup
.probe_size
= 0;
1911 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
1913 struct tcp_sock
*tp
= tcp_sk(sk
);
1914 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1916 /* FIXME: breaks with very large cwnd */
1917 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1918 tp
->snd_cwnd
= tp
->snd_cwnd
*
1919 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
1920 icsk
->icsk_mtup
.probe_size
;
1921 tp
->snd_cwnd_cnt
= 0;
1922 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1923 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
1925 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
1926 icsk
->icsk_mtup
.probe_size
= 0;
1927 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
1931 /* Process an event, which can update packets-in-flight not trivially.
1932 * Main goal of this function is to calculate new estimate for left_out,
1933 * taking into account both packets sitting in receiver's buffer and
1934 * packets lost by network.
1936 * Besides that it does CWND reduction, when packet loss is detected
1937 * and changes state of machine.
1939 * It does _not_ decide what to send, it is made in function
1940 * tcp_xmit_retransmit_queue().
1943 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
1944 int prior_packets
, int flag
)
1946 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1947 struct tcp_sock
*tp
= tcp_sk(sk
);
1948 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
1950 /* Some technical things:
1951 * 1. Reno does not count dupacks (sacked_out) automatically. */
1952 if (!tp
->packets_out
)
1954 /* 2. SACK counts snd_fack in packets inaccurately. */
1955 if (tp
->sacked_out
== 0)
1956 tp
->fackets_out
= 0;
1958 /* Now state machine starts.
1959 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1961 tp
->prior_ssthresh
= 0;
1963 /* B. In all the states check for reneging SACKs. */
1964 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
1967 /* C. Process data loss notification, provided it is valid. */
1968 if ((flag
&FLAG_DATA_LOST
) &&
1969 before(tp
->snd_una
, tp
->high_seq
) &&
1970 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
1971 tp
->fackets_out
> tp
->reordering
) {
1972 tcp_mark_head_lost(sk
, tp
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
1973 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
1976 /* D. Synchronize left_out to current state. */
1977 tcp_sync_left_out(tp
);
1979 /* E. Check state exit conditions. State can be terminated
1980 * when high_seq is ACKed. */
1981 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
1982 if (!sysctl_tcp_frto
)
1983 BUG_TRAP(tp
->retrans_out
== 0);
1984 tp
->retrans_stamp
= 0;
1985 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
1986 switch (icsk
->icsk_ca_state
) {
1988 icsk
->icsk_retransmits
= 0;
1989 if (tcp_try_undo_recovery(sk
, tp
))
1994 /* CWR is to be held something *above* high_seq
1995 * is ACKed for CWR bit to reach receiver. */
1996 if (tp
->snd_una
!= tp
->high_seq
) {
1997 tcp_complete_cwr(sk
);
1998 tcp_set_ca_state(sk
, TCP_CA_Open
);
2002 case TCP_CA_Disorder
:
2003 tcp_try_undo_dsack(sk
, tp
);
2004 if (!tp
->undo_marker
||
2005 /* For SACK case do not Open to allow to undo
2006 * catching for all duplicate ACKs. */
2007 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2008 tp
->undo_marker
= 0;
2009 tcp_set_ca_state(sk
, TCP_CA_Open
);
2013 case TCP_CA_Recovery
:
2015 tcp_reset_reno_sack(tp
);
2016 if (tcp_try_undo_recovery(sk
, tp
))
2018 tcp_complete_cwr(sk
);
2023 /* F. Process state. */
2024 switch (icsk
->icsk_ca_state
) {
2025 case TCP_CA_Recovery
:
2026 if (prior_snd_una
== tp
->snd_una
) {
2027 if (IsReno(tp
) && is_dupack
)
2028 tcp_add_reno_sack(sk
);
2030 int acked
= prior_packets
- tp
->packets_out
;
2032 tcp_remove_reno_sacks(sk
, tp
, acked
);
2033 is_dupack
= tcp_try_undo_partial(sk
, tp
, acked
);
2037 if (flag
&FLAG_DATA_ACKED
)
2038 icsk
->icsk_retransmits
= 0;
2039 if (!tcp_try_undo_loss(sk
, tp
)) {
2040 tcp_moderate_cwnd(tp
);
2041 tcp_xmit_retransmit_queue(sk
);
2044 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2046 /* Loss is undone; fall through to processing in Open state. */
2049 if (tp
->snd_una
!= prior_snd_una
)
2050 tcp_reset_reno_sack(tp
);
2052 tcp_add_reno_sack(sk
);
2055 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2056 tcp_try_undo_dsack(sk
, tp
);
2058 if (!tcp_time_to_recover(sk
, tp
)) {
2059 tcp_try_to_open(sk
, tp
, flag
);
2063 /* MTU probe failure: don't reduce cwnd */
2064 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2065 icsk
->icsk_mtup
.probe_size
&&
2066 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2067 tcp_mtup_probe_failed(sk
);
2068 /* Restores the reduction we did in tcp_mtup_probe() */
2070 tcp_simple_retransmit(sk
);
2074 /* Otherwise enter Recovery state */
2077 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2079 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2081 tp
->high_seq
= tp
->snd_nxt
;
2082 tp
->prior_ssthresh
= 0;
2083 tp
->undo_marker
= tp
->snd_una
;
2084 tp
->undo_retrans
= tp
->retrans_out
;
2086 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2087 if (!(flag
&FLAG_ECE
))
2088 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2089 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2090 TCP_ECN_queue_cwr(tp
);
2093 tp
->bytes_acked
= 0;
2094 tp
->snd_cwnd_cnt
= 0;
2095 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2098 if (is_dupack
|| tcp_head_timedout(sk
, tp
))
2099 tcp_update_scoreboard(sk
, tp
);
2101 tcp_xmit_retransmit_queue(sk
);
2104 /* Read draft-ietf-tcplw-high-performance before mucking
2105 * with this code. (Supersedes RFC1323)
2107 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2109 /* RTTM Rule: A TSecr value received in a segment is used to
2110 * update the averaged RTT measurement only if the segment
2111 * acknowledges some new data, i.e., only if it advances the
2112 * left edge of the send window.
2114 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2115 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2117 * Changed: reset backoff as soon as we see the first valid sample.
2118 * If we do not, we get strongly overestimated rto. With timestamps
2119 * samples are accepted even from very old segments: f.e., when rtt=1
2120 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2121 * answer arrives rto becomes 120 seconds! If at least one of segments
2122 * in window is lost... Voila. --ANK (010210)
2124 struct tcp_sock
*tp
= tcp_sk(sk
);
2125 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2126 tcp_rtt_estimator(sk
, seq_rtt
);
2128 inet_csk(sk
)->icsk_backoff
= 0;
2132 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2134 /* We don't have a timestamp. Can only use
2135 * packets that are not retransmitted to determine
2136 * rtt estimates. Also, we must not reset the
2137 * backoff for rto until we get a non-retransmitted
2138 * packet. This allows us to deal with a situation
2139 * where the network delay has increased suddenly.
2140 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2143 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2146 tcp_rtt_estimator(sk
, seq_rtt
);
2148 inet_csk(sk
)->icsk_backoff
= 0;
2152 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2155 const struct tcp_sock
*tp
= tcp_sk(sk
);
2156 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2157 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2158 tcp_ack_saw_tstamp(sk
, flag
);
2159 else if (seq_rtt
>= 0)
2160 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2163 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
2164 u32 in_flight
, int good
)
2166 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2167 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
2168 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2171 /* Restart timer after forward progress on connection.
2172 * RFC2988 recommends to restart timer to now+rto.
2175 static void tcp_ack_packets_out(struct sock
*sk
, struct tcp_sock
*tp
)
2177 if (!tp
->packets_out
) {
2178 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2180 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2184 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2185 __u32 now
, __s32
*seq_rtt
)
2187 struct tcp_sock
*tp
= tcp_sk(sk
);
2188 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2189 __u32 seq
= tp
->snd_una
;
2190 __u32 packets_acked
;
2193 /* If we get here, the whole TSO packet has not been
2196 BUG_ON(!after(scb
->end_seq
, seq
));
2198 packets_acked
= tcp_skb_pcount(skb
);
2199 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2201 packets_acked
-= tcp_skb_pcount(skb
);
2203 if (packets_acked
) {
2204 __u8 sacked
= scb
->sacked
;
2206 acked
|= FLAG_DATA_ACKED
;
2208 if (sacked
& TCPCB_RETRANS
) {
2209 if (sacked
& TCPCB_SACKED_RETRANS
)
2210 tp
->retrans_out
-= packets_acked
;
2211 acked
|= FLAG_RETRANS_DATA_ACKED
;
2213 } else if (*seq_rtt
< 0)
2214 *seq_rtt
= now
- scb
->when
;
2215 if (sacked
& TCPCB_SACKED_ACKED
)
2216 tp
->sacked_out
-= packets_acked
;
2217 if (sacked
& TCPCB_LOST
)
2218 tp
->lost_out
-= packets_acked
;
2219 if (sacked
& TCPCB_URG
) {
2221 !before(seq
, tp
->snd_up
))
2224 } else if (*seq_rtt
< 0)
2225 *seq_rtt
= now
- scb
->when
;
2227 if (tp
->fackets_out
) {
2228 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2229 tp
->fackets_out
-= dval
;
2231 tp
->packets_out
-= packets_acked
;
2233 BUG_ON(tcp_skb_pcount(skb
) == 0);
2234 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2240 static u32
tcp_usrtt(const struct sk_buff
*skb
)
2242 struct timeval tv
, now
;
2244 do_gettimeofday(&now
);
2245 skb_get_timestamp(skb
, &tv
);
2246 return (now
.tv_sec
- tv
.tv_sec
) * 1000000 + (now
.tv_usec
- tv
.tv_usec
);
2249 /* Remove acknowledged frames from the retransmission queue. */
2250 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2252 struct tcp_sock
*tp
= tcp_sk(sk
);
2253 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2254 struct sk_buff
*skb
;
2255 __u32 now
= tcp_time_stamp
;
2259 void (*rtt_sample
)(struct sock
*sk
, u32 usrtt
)
2260 = icsk
->icsk_ca_ops
->rtt_sample
;
2262 while ((skb
= skb_peek(&sk
->sk_write_queue
)) &&
2263 skb
!= sk
->sk_send_head
) {
2264 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2265 __u8 sacked
= scb
->sacked
;
2267 /* If our packet is before the ack sequence we can
2268 * discard it as it's confirmed to have arrived at
2271 if (after(scb
->end_seq
, tp
->snd_una
)) {
2272 if (tcp_skb_pcount(skb
) > 1 &&
2273 after(tp
->snd_una
, scb
->seq
))
2274 acked
|= tcp_tso_acked(sk
, skb
,
2279 /* Initial outgoing SYN's get put onto the write_queue
2280 * just like anything else we transmit. It is not
2281 * true data, and if we misinform our callers that
2282 * this ACK acks real data, we will erroneously exit
2283 * connection startup slow start one packet too
2284 * quickly. This is severely frowned upon behavior.
2286 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2287 acked
|= FLAG_DATA_ACKED
;
2290 acked
|= FLAG_SYN_ACKED
;
2291 tp
->retrans_stamp
= 0;
2294 /* MTU probing checks */
2295 if (icsk
->icsk_mtup
.probe_size
) {
2296 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2297 tcp_mtup_probe_success(sk
, skb
);
2302 if (sacked
& TCPCB_RETRANS
) {
2303 if(sacked
& TCPCB_SACKED_RETRANS
)
2304 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2305 acked
|= FLAG_RETRANS_DATA_ACKED
;
2307 } else if (seq_rtt
< 0) {
2308 seq_rtt
= now
- scb
->when
;
2310 (*rtt_sample
)(sk
, tcp_usrtt(skb
));
2312 if (sacked
& TCPCB_SACKED_ACKED
)
2313 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2314 if (sacked
& TCPCB_LOST
)
2315 tp
->lost_out
-= tcp_skb_pcount(skb
);
2316 if (sacked
& TCPCB_URG
) {
2318 !before(scb
->end_seq
, tp
->snd_up
))
2321 } else if (seq_rtt
< 0) {
2322 seq_rtt
= now
- scb
->when
;
2324 (*rtt_sample
)(sk
, tcp_usrtt(skb
));
2326 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2327 tcp_packets_out_dec(tp
, skb
);
2328 __skb_unlink(skb
, &sk
->sk_write_queue
);
2329 sk_stream_free_skb(sk
, skb
);
2330 clear_all_retrans_hints(tp
);
2333 if (acked
&FLAG_ACKED
) {
2334 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2335 tcp_ack_packets_out(sk
, tp
);
2337 if (icsk
->icsk_ca_ops
->pkts_acked
)
2338 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
);
2341 #if FASTRETRANS_DEBUG > 0
2342 BUG_TRAP((int)tp
->sacked_out
>= 0);
2343 BUG_TRAP((int)tp
->lost_out
>= 0);
2344 BUG_TRAP((int)tp
->retrans_out
>= 0);
2345 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2346 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2348 printk(KERN_DEBUG
"Leak l=%u %d\n",
2349 tp
->lost_out
, icsk
->icsk_ca_state
);
2352 if (tp
->sacked_out
) {
2353 printk(KERN_DEBUG
"Leak s=%u %d\n",
2354 tp
->sacked_out
, icsk
->icsk_ca_state
);
2357 if (tp
->retrans_out
) {
2358 printk(KERN_DEBUG
"Leak r=%u %d\n",
2359 tp
->retrans_out
, icsk
->icsk_ca_state
);
2360 tp
->retrans_out
= 0;
2364 *seq_rtt_p
= seq_rtt
;
2368 static void tcp_ack_probe(struct sock
*sk
)
2370 const struct tcp_sock
*tp
= tcp_sk(sk
);
2371 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2373 /* Was it a usable window open? */
2375 if (!after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
2376 tp
->snd_una
+ tp
->snd_wnd
)) {
2377 icsk
->icsk_backoff
= 0;
2378 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2379 /* Socket must be waked up by subsequent tcp_data_snd_check().
2380 * This function is not for random using!
2383 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2384 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2389 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2391 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2392 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2395 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2397 const struct tcp_sock
*tp
= tcp_sk(sk
);
2398 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2399 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2402 /* Check that window update is acceptable.
2403 * The function assumes that snd_una<=ack<=snd_next.
2405 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2406 const u32 ack_seq
, const u32 nwin
)
2408 return (after(ack
, tp
->snd_una
) ||
2409 after(ack_seq
, tp
->snd_wl1
) ||
2410 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2413 /* Update our send window.
2415 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2416 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2418 static int tcp_ack_update_window(struct sock
*sk
, struct tcp_sock
*tp
,
2419 struct sk_buff
*skb
, u32 ack
, u32 ack_seq
)
2422 u32 nwin
= ntohs(skb
->h
.th
->window
);
2424 if (likely(!skb
->h
.th
->syn
))
2425 nwin
<<= tp
->rx_opt
.snd_wscale
;
2427 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2428 flag
|= FLAG_WIN_UPDATE
;
2429 tcp_update_wl(tp
, ack
, ack_seq
);
2431 if (tp
->snd_wnd
!= nwin
) {
2434 /* Note, it is the only place, where
2435 * fast path is recovered for sending TCP.
2438 tcp_fast_path_check(sk
, tp
);
2440 if (nwin
> tp
->max_window
) {
2441 tp
->max_window
= nwin
;
2442 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2452 static void tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
)
2454 struct tcp_sock
*tp
= tcp_sk(sk
);
2456 tcp_sync_left_out(tp
);
2458 if (tp
->snd_una
== prior_snd_una
||
2459 !before(tp
->snd_una
, tp
->frto_highmark
)) {
2460 /* RTO was caused by loss, start retransmitting in
2461 * go-back-N slow start
2463 tcp_enter_frto_loss(sk
);
2467 if (tp
->frto_counter
== 1) {
2468 /* First ACK after RTO advances the window: allow two new
2471 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2473 /* Also the second ACK after RTO advances the window.
2474 * The RTO was likely spurious. Reduce cwnd and continue
2475 * in congestion avoidance
2477 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2478 tcp_moderate_cwnd(tp
);
2481 /* F-RTO affects on two new ACKs following RTO.
2482 * At latest on third ACK the TCP behavior is back to normal.
2484 tp
->frto_counter
= (tp
->frto_counter
+ 1) % 3;
2487 /* This routine deals with incoming acks, but not outgoing ones. */
2488 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2490 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2491 struct tcp_sock
*tp
= tcp_sk(sk
);
2492 u32 prior_snd_una
= tp
->snd_una
;
2493 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2494 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2495 u32 prior_in_flight
;
2499 /* If the ack is newer than sent or older than previous acks
2500 * then we can probably ignore it.
2502 if (after(ack
, tp
->snd_nxt
))
2503 goto uninteresting_ack
;
2505 if (before(ack
, prior_snd_una
))
2508 if (sysctl_tcp_abc
&& icsk
->icsk_ca_state
< TCP_CA_CWR
)
2509 tp
->bytes_acked
+= ack
- prior_snd_una
;
2511 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2512 /* Window is constant, pure forward advance.
2513 * No more checks are required.
2514 * Note, we use the fact that SND.UNA>=SND.WL2.
2516 tcp_update_wl(tp
, ack
, ack_seq
);
2518 flag
|= FLAG_WIN_UPDATE
;
2520 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2522 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2524 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2527 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2529 flag
|= tcp_ack_update_window(sk
, tp
, skb
, ack
, ack_seq
);
2531 if (TCP_SKB_CB(skb
)->sacked
)
2532 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2534 if (TCP_ECN_rcv_ecn_echo(tp
, skb
->h
.th
))
2537 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2540 /* We passed data and got it acked, remove any soft error
2541 * log. Something worked...
2543 sk
->sk_err_soft
= 0;
2544 tp
->rcv_tstamp
= tcp_time_stamp
;
2545 prior_packets
= tp
->packets_out
;
2549 prior_in_flight
= tcp_packets_in_flight(tp
);
2551 /* See if we can take anything off of the retransmit queue. */
2552 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2554 if (tp
->frto_counter
)
2555 tcp_process_frto(sk
, prior_snd_una
);
2557 if (tcp_ack_is_dubious(sk
, flag
)) {
2558 /* Advance CWND, if state allows this. */
2559 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
2560 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2561 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2563 if ((flag
& FLAG_DATA_ACKED
))
2564 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2567 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2568 dst_confirm(sk
->sk_dst_cache
);
2573 icsk
->icsk_probes_out
= 0;
2575 /* If this ack opens up a zero window, clear backoff. It was
2576 * being used to time the probes, and is probably far higher than
2577 * it needs to be for normal retransmission.
2579 if (sk
->sk_send_head
)
2584 if (TCP_SKB_CB(skb
)->sacked
)
2585 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2588 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2593 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2594 * But, this can also be called on packets in the established flow when
2595 * the fast version below fails.
2597 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2600 struct tcphdr
*th
= skb
->h
.th
;
2601 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2603 ptr
= (unsigned char *)(th
+ 1);
2604 opt_rx
->saw_tstamp
= 0;
2613 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2618 if (opsize
< 2) /* "silly options" */
2620 if (opsize
> length
)
2621 return; /* don't parse partial options */
2624 if(opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2625 u16 in_mss
= ntohs(get_unaligned((__u16
*)ptr
));
2627 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2628 in_mss
= opt_rx
->user_mss
;
2629 opt_rx
->mss_clamp
= in_mss
;
2634 if(opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2635 if (sysctl_tcp_window_scaling
) {
2636 __u8 snd_wscale
= *(__u8
*) ptr
;
2637 opt_rx
->wscale_ok
= 1;
2638 if (snd_wscale
> 14) {
2640 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2641 "scaling value %d >14 received.\n",
2645 opt_rx
->snd_wscale
= snd_wscale
;
2648 case TCPOPT_TIMESTAMP
:
2649 if(opsize
==TCPOLEN_TIMESTAMP
) {
2650 if ((estab
&& opt_rx
->tstamp_ok
) ||
2651 (!estab
&& sysctl_tcp_timestamps
)) {
2652 opt_rx
->saw_tstamp
= 1;
2653 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__u32
*)ptr
));
2654 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__u32
*)(ptr
+4)));
2658 case TCPOPT_SACK_PERM
:
2659 if(opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2660 if (sysctl_tcp_sack
) {
2661 opt_rx
->sack_ok
= 1;
2662 tcp_sack_reset(opt_rx
);
2668 if((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2669 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2671 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2680 /* Fast parse options. This hopes to only see timestamps.
2681 * If it is wrong it falls back on tcp_parse_options().
2683 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2684 struct tcp_sock
*tp
)
2686 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2687 tp
->rx_opt
.saw_tstamp
= 0;
2689 } else if (tp
->rx_opt
.tstamp_ok
&&
2690 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2691 __u32
*ptr
= (__u32
*)(th
+ 1);
2692 if (*ptr
== ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2693 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2694 tp
->rx_opt
.saw_tstamp
= 1;
2696 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2698 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2702 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2706 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2708 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2709 tp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
2712 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2714 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2715 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2716 * extra check below makes sure this can only happen
2717 * for pure ACK frames. -DaveM
2719 * Not only, also it occurs for expired timestamps.
2722 if((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2723 xtime
.tv_sec
>= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2724 tcp_store_ts_recent(tp
);
2728 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2730 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2731 * it can pass through stack. So, the following predicate verifies that
2732 * this segment is not used for anything but congestion avoidance or
2733 * fast retransmit. Moreover, we even are able to eliminate most of such
2734 * second order effects, if we apply some small "replay" window (~RTO)
2735 * to timestamp space.
2737 * All these measures still do not guarantee that we reject wrapped ACKs
2738 * on networks with high bandwidth, when sequence space is recycled fastly,
2739 * but it guarantees that such events will be very rare and do not affect
2740 * connection seriously. This doesn't look nice, but alas, PAWS is really
2743 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2744 * states that events when retransmit arrives after original data are rare.
2745 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2746 * the biggest problem on large power networks even with minor reordering.
2747 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2748 * up to bandwidth of 18Gigabit/sec. 8) ]
2751 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
2753 struct tcp_sock
*tp
= tcp_sk(sk
);
2754 struct tcphdr
*th
= skb
->h
.th
;
2755 u32 seq
= TCP_SKB_CB(skb
)->seq
;
2756 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2758 return (/* 1. Pure ACK with correct sequence number. */
2759 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
2761 /* 2. ... and duplicate ACK. */
2762 ack
== tp
->snd_una
&&
2764 /* 3. ... and does not update window. */
2765 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
2767 /* 4. ... and sits in replay window. */
2768 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
2771 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
2773 const struct tcp_sock
*tp
= tcp_sk(sk
);
2774 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
2775 xtime
.tv_sec
< tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
2776 !tcp_disordered_ack(sk
, skb
));
2779 /* Check segment sequence number for validity.
2781 * Segment controls are considered valid, if the segment
2782 * fits to the window after truncation to the window. Acceptability
2783 * of data (and SYN, FIN, of course) is checked separately.
2784 * See tcp_data_queue(), for example.
2786 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2787 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2788 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2789 * (borrowed from freebsd)
2792 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2794 return !before(end_seq
, tp
->rcv_wup
) &&
2795 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
2798 /* When we get a reset we do this. */
2799 static void tcp_reset(struct sock
*sk
)
2801 /* We want the right error as BSD sees it (and indeed as we do). */
2802 switch (sk
->sk_state
) {
2804 sk
->sk_err
= ECONNREFUSED
;
2806 case TCP_CLOSE_WAIT
:
2812 sk
->sk_err
= ECONNRESET
;
2815 if (!sock_flag(sk
, SOCK_DEAD
))
2816 sk
->sk_error_report(sk
);
2822 * Process the FIN bit. This now behaves as it is supposed to work
2823 * and the FIN takes effect when it is validly part of sequence
2824 * space. Not before when we get holes.
2826 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2827 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2830 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2831 * close and we go into CLOSING (and later onto TIME-WAIT)
2833 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2835 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
2837 struct tcp_sock
*tp
= tcp_sk(sk
);
2839 inet_csk_schedule_ack(sk
);
2841 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
2842 sock_set_flag(sk
, SOCK_DONE
);
2844 switch (sk
->sk_state
) {
2846 case TCP_ESTABLISHED
:
2847 /* Move to CLOSE_WAIT */
2848 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
2849 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
2852 case TCP_CLOSE_WAIT
:
2854 /* Received a retransmission of the FIN, do
2859 /* RFC793: Remain in the LAST-ACK state. */
2863 /* This case occurs when a simultaneous close
2864 * happens, we must ack the received FIN and
2865 * enter the CLOSING state.
2868 tcp_set_state(sk
, TCP_CLOSING
);
2871 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2873 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
2876 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2877 * cases we should never reach this piece of code.
2879 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
2880 __FUNCTION__
, sk
->sk_state
);
2884 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2885 * Probably, we should reset in this case. For now drop them.
2887 __skb_queue_purge(&tp
->out_of_order_queue
);
2888 if (tp
->rx_opt
.sack_ok
)
2889 tcp_sack_reset(&tp
->rx_opt
);
2890 sk_stream_mem_reclaim(sk
);
2892 if (!sock_flag(sk
, SOCK_DEAD
)) {
2893 sk
->sk_state_change(sk
);
2895 /* Do not send POLL_HUP for half duplex close. */
2896 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
2897 sk
->sk_state
== TCP_CLOSE
)
2898 sk_wake_async(sk
, 1, POLL_HUP
);
2900 sk_wake_async(sk
, 1, POLL_IN
);
2904 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
2906 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
2907 if (before(seq
, sp
->start_seq
))
2908 sp
->start_seq
= seq
;
2909 if (after(end_seq
, sp
->end_seq
))
2910 sp
->end_seq
= end_seq
;
2916 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2918 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2919 if (before(seq
, tp
->rcv_nxt
))
2920 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
2922 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
2924 tp
->rx_opt
.dsack
= 1;
2925 tp
->duplicate_sack
[0].start_seq
= seq
;
2926 tp
->duplicate_sack
[0].end_seq
= end_seq
;
2927 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
2931 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2933 if (!tp
->rx_opt
.dsack
)
2934 tcp_dsack_set(tp
, seq
, end_seq
);
2936 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
2939 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
2941 struct tcp_sock
*tp
= tcp_sk(sk
);
2943 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
2944 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
2945 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
2946 tcp_enter_quickack_mode(sk
);
2948 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2949 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2951 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
2952 end_seq
= tp
->rcv_nxt
;
2953 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
2960 /* These routines update the SACK block as out-of-order packets arrive or
2961 * in-order packets close up the sequence space.
2963 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
2966 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2967 struct tcp_sack_block
*swalk
= sp
+1;
2969 /* See if the recent change to the first SACK eats into
2970 * or hits the sequence space of other SACK blocks, if so coalesce.
2972 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
2973 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
2976 /* Zap SWALK, by moving every further SACK up by one slot.
2977 * Decrease num_sacks.
2979 tp
->rx_opt
.num_sacks
--;
2980 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
2981 for(i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
2985 this_sack
++, swalk
++;
2989 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
2993 tmp
= sack1
->start_seq
;
2994 sack1
->start_seq
= sack2
->start_seq
;
2995 sack2
->start_seq
= tmp
;
2997 tmp
= sack1
->end_seq
;
2998 sack1
->end_seq
= sack2
->end_seq
;
2999 sack2
->end_seq
= tmp
;
3002 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3004 struct tcp_sock
*tp
= tcp_sk(sk
);
3005 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3006 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3012 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3013 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3014 /* Rotate this_sack to the first one. */
3015 for (; this_sack
>0; this_sack
--, sp
--)
3016 tcp_sack_swap(sp
, sp
-1);
3018 tcp_sack_maybe_coalesce(tp
);
3023 /* Could not find an adjacent existing SACK, build a new one,
3024 * put it at the front, and shift everyone else down. We
3025 * always know there is at least one SACK present already here.
3027 * If the sack array is full, forget about the last one.
3029 if (this_sack
>= 4) {
3031 tp
->rx_opt
.num_sacks
--;
3034 for(; this_sack
> 0; this_sack
--, sp
--)
3038 /* Build the new head SACK, and we're done. */
3039 sp
->start_seq
= seq
;
3040 sp
->end_seq
= end_seq
;
3041 tp
->rx_opt
.num_sacks
++;
3042 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3045 /* RCV.NXT advances, some SACKs should be eaten. */
3047 static void tcp_sack_remove(struct tcp_sock
*tp
)
3049 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3050 int num_sacks
= tp
->rx_opt
.num_sacks
;
3053 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3054 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3055 tp
->rx_opt
.num_sacks
= 0;
3056 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3060 for(this_sack
= 0; this_sack
< num_sacks
; ) {
3061 /* Check if the start of the sack is covered by RCV.NXT. */
3062 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3065 /* RCV.NXT must cover all the block! */
3066 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3068 /* Zap this SACK, by moving forward any other SACKS. */
3069 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3070 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3077 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3078 tp
->rx_opt
.num_sacks
= num_sacks
;
3079 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3083 /* This one checks to see if we can put data from the
3084 * out_of_order queue into the receive_queue.
3086 static void tcp_ofo_queue(struct sock
*sk
)
3088 struct tcp_sock
*tp
= tcp_sk(sk
);
3089 __u32 dsack_high
= tp
->rcv_nxt
;
3090 struct sk_buff
*skb
;
3092 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3093 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3096 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3097 __u32 dsack
= dsack_high
;
3098 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3099 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3100 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3103 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3104 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3105 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3109 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3110 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3111 TCP_SKB_CB(skb
)->end_seq
);
3113 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3114 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3115 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3117 tcp_fin(skb
, sk
, skb
->h
.th
);
3121 static int tcp_prune_queue(struct sock
*sk
);
3123 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3125 struct tcphdr
*th
= skb
->h
.th
;
3126 struct tcp_sock
*tp
= tcp_sk(sk
);
3129 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3132 __skb_pull(skb
, th
->doff
*4);
3134 TCP_ECN_accept_cwr(tp
, skb
);
3136 if (tp
->rx_opt
.dsack
) {
3137 tp
->rx_opt
.dsack
= 0;
3138 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3139 4 - tp
->rx_opt
.tstamp_ok
);
3142 /* Queue data for delivery to the user.
3143 * Packets in sequence go to the receive queue.
3144 * Out of sequence packets to the out_of_order_queue.
3146 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3147 if (tcp_receive_window(tp
) == 0)
3150 /* Ok. In sequence. In window. */
3151 if (tp
->ucopy
.task
== current
&&
3152 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3153 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3154 int chunk
= min_t(unsigned int, skb
->len
,
3157 __set_current_state(TASK_RUNNING
);
3160 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3161 tp
->ucopy
.len
-= chunk
;
3162 tp
->copied_seq
+= chunk
;
3163 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3164 tcp_rcv_space_adjust(sk
);
3172 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3173 !sk_stream_rmem_schedule(sk
, skb
))) {
3174 if (tcp_prune_queue(sk
) < 0 ||
3175 !sk_stream_rmem_schedule(sk
, skb
))
3178 sk_stream_set_owner_r(skb
, sk
);
3179 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3181 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3183 tcp_event_data_recv(sk
, tp
, skb
);
3185 tcp_fin(skb
, sk
, th
);
3187 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3190 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3191 * gap in queue is filled.
3193 if (skb_queue_empty(&tp
->out_of_order_queue
))
3194 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3197 if (tp
->rx_opt
.num_sacks
)
3198 tcp_sack_remove(tp
);
3200 tcp_fast_path_check(sk
, tp
);
3204 else if (!sock_flag(sk
, SOCK_DEAD
))
3205 sk
->sk_data_ready(sk
, 0);
3209 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3210 /* A retransmit, 2nd most common case. Force an immediate ack. */
3211 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3212 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3215 tcp_enter_quickack_mode(sk
);
3216 inet_csk_schedule_ack(sk
);
3222 /* Out of window. F.e. zero window probe. */
3223 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3226 tcp_enter_quickack_mode(sk
);
3228 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3229 /* Partial packet, seq < rcv_next < end_seq */
3230 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3231 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3232 TCP_SKB_CB(skb
)->end_seq
);
3234 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3236 /* If window is closed, drop tail of packet. But after
3237 * remembering D-SACK for its head made in previous line.
3239 if (!tcp_receive_window(tp
))
3244 TCP_ECN_check_ce(tp
, skb
);
3246 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3247 !sk_stream_rmem_schedule(sk
, skb
)) {
3248 if (tcp_prune_queue(sk
) < 0 ||
3249 !sk_stream_rmem_schedule(sk
, skb
))
3253 /* Disable header prediction. */
3255 inet_csk_schedule_ack(sk
);
3257 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3258 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3260 sk_stream_set_owner_r(skb
, sk
);
3262 if (!skb_peek(&tp
->out_of_order_queue
)) {
3263 /* Initial out of order segment, build 1 SACK. */
3264 if (tp
->rx_opt
.sack_ok
) {
3265 tp
->rx_opt
.num_sacks
= 1;
3266 tp
->rx_opt
.dsack
= 0;
3267 tp
->rx_opt
.eff_sacks
= 1;
3268 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3269 tp
->selective_acks
[0].end_seq
=
3270 TCP_SKB_CB(skb
)->end_seq
;
3272 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3274 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3275 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3276 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3278 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3279 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3281 if (!tp
->rx_opt
.num_sacks
||
3282 tp
->selective_acks
[0].end_seq
!= seq
)
3285 /* Common case: data arrive in order after hole. */
3286 tp
->selective_acks
[0].end_seq
= end_seq
;
3290 /* Find place to insert this segment. */
3292 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3294 } while ((skb1
= skb1
->prev
) !=
3295 (struct sk_buff
*)&tp
->out_of_order_queue
);
3297 /* Do skb overlap to previous one? */
3298 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3299 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3300 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3301 /* All the bits are present. Drop. */
3303 tcp_dsack_set(tp
, seq
, end_seq
);
3306 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3307 /* Partial overlap. */
3308 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3313 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3315 /* And clean segments covered by new one as whole. */
3316 while ((skb1
= skb
->next
) !=
3317 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3318 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3319 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3320 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3323 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3324 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3329 if (tp
->rx_opt
.sack_ok
)
3330 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3334 /* Collapse contiguous sequence of skbs head..tail with
3335 * sequence numbers start..end.
3336 * Segments with FIN/SYN are not collapsed (only because this
3340 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3341 struct sk_buff
*head
, struct sk_buff
*tail
,
3344 struct sk_buff
*skb
;
3346 /* First, check that queue is collapsible and find
3347 * the point where collapsing can be useful. */
3348 for (skb
= head
; skb
!= tail
; ) {
3349 /* No new bits? It is possible on ofo queue. */
3350 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3351 struct sk_buff
*next
= skb
->next
;
3352 __skb_unlink(skb
, list
);
3354 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3359 /* The first skb to collapse is:
3361 * - bloated or contains data before "start" or
3362 * overlaps to the next one.
3364 if (!skb
->h
.th
->syn
&& !skb
->h
.th
->fin
&&
3365 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3366 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3367 (skb
->next
!= tail
&&
3368 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3371 /* Decided to skip this, advance start seq. */
3372 start
= TCP_SKB_CB(skb
)->end_seq
;
3375 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3378 while (before(start
, end
)) {
3379 struct sk_buff
*nskb
;
3380 int header
= skb_headroom(skb
);
3381 int copy
= SKB_MAX_ORDER(header
, 0);
3383 /* Too big header? This can happen with IPv6. */
3386 if (end
-start
< copy
)
3388 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3391 skb_reserve(nskb
, header
);
3392 memcpy(nskb
->head
, skb
->head
, header
);
3393 nskb
->nh
.raw
= nskb
->head
+ (skb
->nh
.raw
-skb
->head
);
3394 nskb
->h
.raw
= nskb
->head
+ (skb
->h
.raw
-skb
->head
);
3395 nskb
->mac
.raw
= nskb
->head
+ (skb
->mac
.raw
-skb
->head
);
3396 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3397 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3398 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3399 sk_stream_set_owner_r(nskb
, sk
);
3401 /* Copy data, releasing collapsed skbs. */
3403 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3404 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3408 size
= min(copy
, size
);
3409 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3411 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3415 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3416 struct sk_buff
*next
= skb
->next
;
3417 __skb_unlink(skb
, list
);
3419 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3421 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3428 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3429 * and tcp_collapse() them until all the queue is collapsed.
3431 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3433 struct tcp_sock
*tp
= tcp_sk(sk
);
3434 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3435 struct sk_buff
*head
;
3441 start
= TCP_SKB_CB(skb
)->seq
;
3442 end
= TCP_SKB_CB(skb
)->end_seq
;
3448 /* Segment is terminated when we see gap or when
3449 * we are at the end of all the queue. */
3450 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3451 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3452 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3453 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3454 head
, skb
, start
, end
);
3456 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3458 /* Start new segment */
3459 start
= TCP_SKB_CB(skb
)->seq
;
3460 end
= TCP_SKB_CB(skb
)->end_seq
;
3462 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3463 start
= TCP_SKB_CB(skb
)->seq
;
3464 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3465 end
= TCP_SKB_CB(skb
)->end_seq
;
3470 /* Reduce allocated memory if we can, trying to get
3471 * the socket within its memory limits again.
3473 * Return less than zero if we should start dropping frames
3474 * until the socket owning process reads some of the data
3475 * to stabilize the situation.
3477 static int tcp_prune_queue(struct sock
*sk
)
3479 struct tcp_sock
*tp
= tcp_sk(sk
);
3481 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3483 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3485 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3486 tcp_clamp_window(sk
, tp
);
3487 else if (tcp_memory_pressure
)
3488 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3490 tcp_collapse_ofo_queue(sk
);
3491 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3492 sk
->sk_receive_queue
.next
,
3493 (struct sk_buff
*)&sk
->sk_receive_queue
,
3494 tp
->copied_seq
, tp
->rcv_nxt
);
3495 sk_stream_mem_reclaim(sk
);
3497 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3500 /* Collapsing did not help, destructive actions follow.
3501 * This must not ever occur. */
3503 /* First, purge the out_of_order queue. */
3504 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3505 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3506 __skb_queue_purge(&tp
->out_of_order_queue
);
3508 /* Reset SACK state. A conforming SACK implementation will
3509 * do the same at a timeout based retransmit. When a connection
3510 * is in a sad state like this, we care only about integrity
3511 * of the connection not performance.
3513 if (tp
->rx_opt
.sack_ok
)
3514 tcp_sack_reset(&tp
->rx_opt
);
3515 sk_stream_mem_reclaim(sk
);
3518 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3521 /* If we are really being abused, tell the caller to silently
3522 * drop receive data on the floor. It will get retransmitted
3523 * and hopefully then we'll have sufficient space.
3525 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3527 /* Massive buffer overcommit. */
3533 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3534 * As additional protections, we do not touch cwnd in retransmission phases,
3535 * and if application hit its sndbuf limit recently.
3537 void tcp_cwnd_application_limited(struct sock
*sk
)
3539 struct tcp_sock
*tp
= tcp_sk(sk
);
3541 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3542 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3543 /* Limited by application or receiver window. */
3544 u32 win_used
= max(tp
->snd_cwnd_used
, 2U);
3545 if (win_used
< tp
->snd_cwnd
) {
3546 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3547 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3549 tp
->snd_cwnd_used
= 0;
3551 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3554 static int tcp_should_expand_sndbuf(struct sock
*sk
, struct tcp_sock
*tp
)
3556 /* If the user specified a specific send buffer setting, do
3559 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3562 /* If we are under global TCP memory pressure, do not expand. */
3563 if (tcp_memory_pressure
)
3566 /* If we are under soft global TCP memory pressure, do not expand. */
3567 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3570 /* If we filled the congestion window, do not expand. */
3571 if (tp
->packets_out
>= tp
->snd_cwnd
)
3577 /* When incoming ACK allowed to free some skb from write_queue,
3578 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3579 * on the exit from tcp input handler.
3581 * PROBLEM: sndbuf expansion does not work well with largesend.
3583 static void tcp_new_space(struct sock
*sk
)
3585 struct tcp_sock
*tp
= tcp_sk(sk
);
3587 if (tcp_should_expand_sndbuf(sk
, tp
)) {
3588 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3589 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3590 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3591 tp
->reordering
+ 1);
3592 sndmem
*= 2*demanded
;
3593 if (sndmem
> sk
->sk_sndbuf
)
3594 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3595 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3598 sk
->sk_write_space(sk
);
3601 static void tcp_check_space(struct sock
*sk
)
3603 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3604 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3605 if (sk
->sk_socket
&&
3606 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3611 static inline void tcp_data_snd_check(struct sock
*sk
, struct tcp_sock
*tp
)
3613 tcp_push_pending_frames(sk
, tp
);
3614 tcp_check_space(sk
);
3618 * Check if sending an ack is needed.
3620 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3622 struct tcp_sock
*tp
= tcp_sk(sk
);
3624 /* More than one full frame received... */
3625 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3626 /* ... and right edge of window advances far enough.
3627 * (tcp_recvmsg() will send ACK otherwise). Or...
3629 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3630 /* We ACK each frame or... */
3631 tcp_in_quickack_mode(sk
) ||
3632 /* We have out of order data. */
3634 skb_peek(&tp
->out_of_order_queue
))) {
3635 /* Then ack it now */
3638 /* Else, send delayed ack. */
3639 tcp_send_delayed_ack(sk
);
3643 static inline void tcp_ack_snd_check(struct sock
*sk
)
3645 if (!inet_csk_ack_scheduled(sk
)) {
3646 /* We sent a data segment already. */
3649 __tcp_ack_snd_check(sk
, 1);
3653 * This routine is only called when we have urgent data
3654 * signaled. Its the 'slow' part of tcp_urg. It could be
3655 * moved inline now as tcp_urg is only called from one
3656 * place. We handle URGent data wrong. We have to - as
3657 * BSD still doesn't use the correction from RFC961.
3658 * For 1003.1g we should support a new option TCP_STDURG to permit
3659 * either form (or just set the sysctl tcp_stdurg).
3662 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3664 struct tcp_sock
*tp
= tcp_sk(sk
);
3665 u32 ptr
= ntohs(th
->urg_ptr
);
3667 if (ptr
&& !sysctl_tcp_stdurg
)
3669 ptr
+= ntohl(th
->seq
);
3671 /* Ignore urgent data that we've already seen and read. */
3672 if (after(tp
->copied_seq
, ptr
))
3675 /* Do not replay urg ptr.
3677 * NOTE: interesting situation not covered by specs.
3678 * Misbehaving sender may send urg ptr, pointing to segment,
3679 * which we already have in ofo queue. We are not able to fetch
3680 * such data and will stay in TCP_URG_NOTYET until will be eaten
3681 * by recvmsg(). Seems, we are not obliged to handle such wicked
3682 * situations. But it is worth to think about possibility of some
3683 * DoSes using some hypothetical application level deadlock.
3685 if (before(ptr
, tp
->rcv_nxt
))
3688 /* Do we already have a newer (or duplicate) urgent pointer? */
3689 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3692 /* Tell the world about our new urgent pointer. */
3695 /* We may be adding urgent data when the last byte read was
3696 * urgent. To do this requires some care. We cannot just ignore
3697 * tp->copied_seq since we would read the last urgent byte again
3698 * as data, nor can we alter copied_seq until this data arrives
3699 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3701 * NOTE. Double Dutch. Rendering to plain English: author of comment
3702 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3703 * and expect that both A and B disappear from stream. This is _wrong_.
3704 * Though this happens in BSD with high probability, this is occasional.
3705 * Any application relying on this is buggy. Note also, that fix "works"
3706 * only in this artificial test. Insert some normal data between A and B and we will
3707 * decline of BSD again. Verdict: it is better to remove to trap
3710 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3711 !sock_flag(sk
, SOCK_URGINLINE
) &&
3712 tp
->copied_seq
!= tp
->rcv_nxt
) {
3713 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3715 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3716 __skb_unlink(skb
, &sk
->sk_receive_queue
);
3721 tp
->urg_data
= TCP_URG_NOTYET
;
3724 /* Disable header prediction. */
3728 /* This is the 'fast' part of urgent handling. */
3729 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
3731 struct tcp_sock
*tp
= tcp_sk(sk
);
3733 /* Check if we get a new urgent pointer - normally not. */
3735 tcp_check_urg(sk
,th
);
3737 /* Do we wait for any urgent data? - normally not... */
3738 if (tp
->urg_data
== TCP_URG_NOTYET
) {
3739 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
3742 /* Is the urgent pointer pointing into this packet? */
3743 if (ptr
< skb
->len
) {
3745 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
3747 tp
->urg_data
= TCP_URG_VALID
| tmp
;
3748 if (!sock_flag(sk
, SOCK_DEAD
))
3749 sk
->sk_data_ready(sk
, 0);
3754 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3756 struct tcp_sock
*tp
= tcp_sk(sk
);
3757 int chunk
= skb
->len
- hlen
;
3761 if (skb
->ip_summed
==CHECKSUM_UNNECESSARY
)
3762 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
3764 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
3768 tp
->ucopy
.len
-= chunk
;
3769 tp
->copied_seq
+= chunk
;
3770 tcp_rcv_space_adjust(sk
);
3777 static int __tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3781 if (sock_owned_by_user(sk
)) {
3783 result
= __tcp_checksum_complete(skb
);
3786 result
= __tcp_checksum_complete(skb
);
3791 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3793 return skb
->ip_summed
!= CHECKSUM_UNNECESSARY
&&
3794 __tcp_checksum_complete_user(sk
, skb
);
3797 #ifdef CONFIG_NET_DMA
3798 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3800 struct tcp_sock
*tp
= tcp_sk(sk
);
3801 int chunk
= skb
->len
- hlen
;
3803 int copied_early
= 0;
3805 if (tp
->ucopy
.wakeup
)
3808 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
3809 tp
->ucopy
.dma_chan
= get_softnet_dma();
3811 if (tp
->ucopy
.dma_chan
&& skb
->ip_summed
== CHECKSUM_UNNECESSARY
) {
3813 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
3814 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
3819 tp
->ucopy
.dma_cookie
= dma_cookie
;
3822 tp
->ucopy
.len
-= chunk
;
3823 tp
->copied_seq
+= chunk
;
3824 tcp_rcv_space_adjust(sk
);
3826 if ((tp
->ucopy
.len
== 0) ||
3827 (tcp_flag_word(skb
->h
.th
) & TCP_FLAG_PSH
) ||
3828 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
3829 tp
->ucopy
.wakeup
= 1;
3830 sk
->sk_data_ready(sk
, 0);
3832 } else if (chunk
> 0) {
3833 tp
->ucopy
.wakeup
= 1;
3834 sk
->sk_data_ready(sk
, 0);
3837 return copied_early
;
3839 #endif /* CONFIG_NET_DMA */
3842 * TCP receive function for the ESTABLISHED state.
3844 * It is split into a fast path and a slow path. The fast path is
3846 * - A zero window was announced from us - zero window probing
3847 * is only handled properly in the slow path.
3848 * - Out of order segments arrived.
3849 * - Urgent data is expected.
3850 * - There is no buffer space left
3851 * - Unexpected TCP flags/window values/header lengths are received
3852 * (detected by checking the TCP header against pred_flags)
3853 * - Data is sent in both directions. Fast path only supports pure senders
3854 * or pure receivers (this means either the sequence number or the ack
3855 * value must stay constant)
3856 * - Unexpected TCP option.
3858 * When these conditions are not satisfied it drops into a standard
3859 * receive procedure patterned after RFC793 to handle all cases.
3860 * The first three cases are guaranteed by proper pred_flags setting,
3861 * the rest is checked inline. Fast processing is turned on in
3862 * tcp_data_queue when everything is OK.
3864 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
3865 struct tcphdr
*th
, unsigned len
)
3867 struct tcp_sock
*tp
= tcp_sk(sk
);
3870 * Header prediction.
3871 * The code loosely follows the one in the famous
3872 * "30 instruction TCP receive" Van Jacobson mail.
3874 * Van's trick is to deposit buffers into socket queue
3875 * on a device interrupt, to call tcp_recv function
3876 * on the receive process context and checksum and copy
3877 * the buffer to user space. smart...
3879 * Our current scheme is not silly either but we take the
3880 * extra cost of the net_bh soft interrupt processing...
3881 * We do checksum and copy also but from device to kernel.
3884 tp
->rx_opt
.saw_tstamp
= 0;
3886 /* pred_flags is 0xS?10 << 16 + snd_wnd
3887 * if header_prediction is to be made
3888 * 'S' will always be tp->tcp_header_len >> 2
3889 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3890 * turn it off (when there are holes in the receive
3891 * space for instance)
3892 * PSH flag is ignored.
3895 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
3896 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3897 int tcp_header_len
= tp
->tcp_header_len
;
3899 /* Timestamp header prediction: tcp_header_len
3900 * is automatically equal to th->doff*4 due to pred_flags
3904 /* Check timestamp */
3905 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
3906 __u32
*ptr
= (__u32
*)(th
+ 1);
3908 /* No? Slow path! */
3909 if (*ptr
!= ntohl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3910 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
3913 tp
->rx_opt
.saw_tstamp
= 1;
3915 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3917 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3919 /* If PAWS failed, check it more carefully in slow path */
3920 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
3923 /* DO NOT update ts_recent here, if checksum fails
3924 * and timestamp was corrupted part, it will result
3925 * in a hung connection since we will drop all
3926 * future packets due to the PAWS test.
3930 if (len
<= tcp_header_len
) {
3931 /* Bulk data transfer: sender */
3932 if (len
== tcp_header_len
) {
3933 /* Predicted packet is in window by definition.
3934 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3935 * Hence, check seq<=rcv_wup reduces to:
3937 if (tcp_header_len
==
3938 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3939 tp
->rcv_nxt
== tp
->rcv_wup
)
3940 tcp_store_ts_recent(tp
);
3942 /* We know that such packets are checksummed
3945 tcp_ack(sk
, skb
, 0);
3947 tcp_data_snd_check(sk
, tp
);
3949 } else { /* Header too small */
3950 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3955 int copied_early
= 0;
3957 if (tp
->copied_seq
== tp
->rcv_nxt
&&
3958 len
- tcp_header_len
<= tp
->ucopy
.len
) {
3959 #ifdef CONFIG_NET_DMA
3960 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
3965 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
3966 __set_current_state(TASK_RUNNING
);
3968 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
3972 /* Predicted packet is in window by definition.
3973 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3974 * Hence, check seq<=rcv_wup reduces to:
3976 if (tcp_header_len
==
3977 (sizeof(struct tcphdr
) +
3978 TCPOLEN_TSTAMP_ALIGNED
) &&
3979 tp
->rcv_nxt
== tp
->rcv_wup
)
3980 tcp_store_ts_recent(tp
);
3982 tcp_rcv_rtt_measure_ts(sk
, skb
);
3984 __skb_pull(skb
, tcp_header_len
);
3985 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3986 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
3989 tcp_cleanup_rbuf(sk
, skb
->len
);
3992 if (tcp_checksum_complete_user(sk
, skb
))
3995 /* Predicted packet is in window by definition.
3996 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3997 * Hence, check seq<=rcv_wup reduces to:
3999 if (tcp_header_len
==
4000 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4001 tp
->rcv_nxt
== tp
->rcv_wup
)
4002 tcp_store_ts_recent(tp
);
4004 tcp_rcv_rtt_measure_ts(sk
, skb
);
4006 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4009 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4011 /* Bulk data transfer: receiver */
4012 __skb_pull(skb
,tcp_header_len
);
4013 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4014 sk_stream_set_owner_r(skb
, sk
);
4015 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4018 tcp_event_data_recv(sk
, tp
, skb
);
4020 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4021 /* Well, only one small jumplet in fast path... */
4022 tcp_ack(sk
, skb
, FLAG_DATA
);
4023 tcp_data_snd_check(sk
, tp
);
4024 if (!inet_csk_ack_scheduled(sk
))
4028 __tcp_ack_snd_check(sk
, 0);
4030 #ifdef CONFIG_NET_DMA
4032 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4038 sk
->sk_data_ready(sk
, 0);
4044 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4048 * RFC1323: H1. Apply PAWS check first.
4050 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4051 tcp_paws_discard(sk
, skb
)) {
4053 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4054 tcp_send_dupack(sk
, skb
);
4057 /* Resets are accepted even if PAWS failed.
4059 ts_recent update must be made after we are sure
4060 that the packet is in window.
4065 * Standard slow path.
4068 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4069 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4070 * (RST) segments are validated by checking their SEQ-fields."
4071 * And page 69: "If an incoming segment is not acceptable,
4072 * an acknowledgment should be sent in reply (unless the RST bit
4073 * is set, if so drop the segment and return)".
4076 tcp_send_dupack(sk
, skb
);
4085 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4087 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4088 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4089 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4096 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4098 tcp_rcv_rtt_measure_ts(sk
, skb
);
4100 /* Process urgent data. */
4101 tcp_urg(sk
, skb
, th
);
4103 /* step 7: process the segment text */
4104 tcp_data_queue(sk
, skb
);
4106 tcp_data_snd_check(sk
, tp
);
4107 tcp_ack_snd_check(sk
);
4111 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4118 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4119 struct tcphdr
*th
, unsigned len
)
4121 struct tcp_sock
*tp
= tcp_sk(sk
);
4122 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4123 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4125 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4129 * "If the state is SYN-SENT then
4130 * first check the ACK bit
4131 * If the ACK bit is set
4132 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4133 * a reset (unless the RST bit is set, if so drop
4134 * the segment and return)"
4136 * We do not send data with SYN, so that RFC-correct
4139 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4140 goto reset_and_undo
;
4142 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4143 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4145 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4146 goto reset_and_undo
;
4149 /* Now ACK is acceptable.
4151 * "If the RST bit is set
4152 * If the ACK was acceptable then signal the user "error:
4153 * connection reset", drop the segment, enter CLOSED state,
4154 * delete TCB, and return."
4163 * "fifth, if neither of the SYN or RST bits is set then
4164 * drop the segment and return."
4170 goto discard_and_undo
;
4173 * "If the SYN bit is on ...
4174 * are acceptable then ...
4175 * (our SYN has been ACKed), change the connection
4176 * state to ESTABLISHED..."
4179 TCP_ECN_rcv_synack(tp
, th
);
4181 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4182 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4184 /* Ok.. it's good. Set up sequence numbers and
4185 * move to established.
4187 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4188 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4190 /* RFC1323: The window in SYN & SYN/ACK segments is
4193 tp
->snd_wnd
= ntohs(th
->window
);
4194 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4196 if (!tp
->rx_opt
.wscale_ok
) {
4197 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4198 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4201 if (tp
->rx_opt
.saw_tstamp
) {
4202 tp
->rx_opt
.tstamp_ok
= 1;
4203 tp
->tcp_header_len
=
4204 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4205 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4206 tcp_store_ts_recent(tp
);
4208 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4211 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4212 tp
->rx_opt
.sack_ok
|= 2;
4215 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4216 tcp_initialize_rcv_mss(sk
);
4218 /* Remember, tcp_poll() does not lock socket!
4219 * Change state from SYN-SENT only after copied_seq
4220 * is initialized. */
4221 tp
->copied_seq
= tp
->rcv_nxt
;
4223 tcp_set_state(sk
, TCP_ESTABLISHED
);
4225 /* Make sure socket is routed, for correct metrics. */
4226 icsk
->icsk_af_ops
->rebuild_header(sk
);
4228 tcp_init_metrics(sk
);
4230 tcp_init_congestion_control(sk
);
4232 /* Prevent spurious tcp_cwnd_restart() on first data
4235 tp
->lsndtime
= tcp_time_stamp
;
4237 tcp_init_buffer_space(sk
);
4239 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4240 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4242 if (!tp
->rx_opt
.snd_wscale
)
4243 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4247 if (!sock_flag(sk
, SOCK_DEAD
)) {
4248 sk
->sk_state_change(sk
);
4249 sk_wake_async(sk
, 0, POLL_OUT
);
4252 if (sk
->sk_write_pending
||
4253 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4254 icsk
->icsk_ack
.pingpong
) {
4255 /* Save one ACK. Data will be ready after
4256 * several ticks, if write_pending is set.
4258 * It may be deleted, but with this feature tcpdumps
4259 * look so _wonderfully_ clever, that I was not able
4260 * to stand against the temptation 8) --ANK
4262 inet_csk_schedule_ack(sk
);
4263 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4264 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4265 tcp_incr_quickack(sk
);
4266 tcp_enter_quickack_mode(sk
);
4267 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4268 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4279 /* No ACK in the segment */
4283 * "If the RST bit is set
4285 * Otherwise (no ACK) drop the segment and return."
4288 goto discard_and_undo
;
4292 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4293 goto discard_and_undo
;
4296 /* We see SYN without ACK. It is attempt of
4297 * simultaneous connect with crossed SYNs.
4298 * Particularly, it can be connect to self.
4300 tcp_set_state(sk
, TCP_SYN_RECV
);
4302 if (tp
->rx_opt
.saw_tstamp
) {
4303 tp
->rx_opt
.tstamp_ok
= 1;
4304 tcp_store_ts_recent(tp
);
4305 tp
->tcp_header_len
=
4306 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4308 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4311 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4312 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4314 /* RFC1323: The window in SYN & SYN/ACK segments is
4317 tp
->snd_wnd
= ntohs(th
->window
);
4318 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4319 tp
->max_window
= tp
->snd_wnd
;
4321 TCP_ECN_rcv_syn(tp
, th
);
4324 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4325 tcp_initialize_rcv_mss(sk
);
4328 tcp_send_synack(sk
);
4330 /* Note, we could accept data and URG from this segment.
4331 * There are no obstacles to make this.
4333 * However, if we ignore data in ACKless segments sometimes,
4334 * we have no reasons to accept it sometimes.
4335 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4336 * is not flawless. So, discard packet for sanity.
4337 * Uncomment this return to process the data.
4344 /* "fifth, if neither of the SYN or RST bits is set then
4345 * drop the segment and return."
4349 tcp_clear_options(&tp
->rx_opt
);
4350 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4354 tcp_clear_options(&tp
->rx_opt
);
4355 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4361 * This function implements the receiving procedure of RFC 793 for
4362 * all states except ESTABLISHED and TIME_WAIT.
4363 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4364 * address independent.
4367 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4368 struct tcphdr
*th
, unsigned len
)
4370 struct tcp_sock
*tp
= tcp_sk(sk
);
4371 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4374 tp
->rx_opt
.saw_tstamp
= 0;
4376 switch (sk
->sk_state
) {
4388 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4391 /* Now we have several options: In theory there is
4392 * nothing else in the frame. KA9Q has an option to
4393 * send data with the syn, BSD accepts data with the
4394 * syn up to the [to be] advertised window and
4395 * Solaris 2.1 gives you a protocol error. For now
4396 * we just ignore it, that fits the spec precisely
4397 * and avoids incompatibilities. It would be nice in
4398 * future to drop through and process the data.
4400 * Now that TTCP is starting to be used we ought to
4402 * But, this leaves one open to an easy denial of
4403 * service attack, and SYN cookies can't defend
4404 * against this problem. So, we drop the data
4405 * in the interest of security over speed.
4412 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4416 /* Do step6 onward by hand. */
4417 tcp_urg(sk
, skb
, th
);
4419 tcp_data_snd_check(sk
, tp
);
4423 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4424 tcp_paws_discard(sk
, skb
)) {
4426 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4427 tcp_send_dupack(sk
, skb
);
4430 /* Reset is accepted even if it did not pass PAWS. */
4433 /* step 1: check sequence number */
4434 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4436 tcp_send_dupack(sk
, skb
);
4440 /* step 2: check RST bit */
4446 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4448 /* step 3: check security and precedence [ignored] */
4452 * Check for a SYN in window.
4454 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4455 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4460 /* step 5: check the ACK field */
4462 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4464 switch(sk
->sk_state
) {
4467 tp
->copied_seq
= tp
->rcv_nxt
;
4469 tcp_set_state(sk
, TCP_ESTABLISHED
);
4470 sk
->sk_state_change(sk
);
4472 /* Note, that this wakeup is only for marginal
4473 * crossed SYN case. Passively open sockets
4474 * are not waked up, because sk->sk_sleep ==
4475 * NULL and sk->sk_socket == NULL.
4477 if (sk
->sk_socket
) {
4478 sk_wake_async(sk
,0,POLL_OUT
);
4481 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4482 tp
->snd_wnd
= ntohs(th
->window
) <<
4483 tp
->rx_opt
.snd_wscale
;
4484 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4485 TCP_SKB_CB(skb
)->seq
);
4487 /* tcp_ack considers this ACK as duplicate
4488 * and does not calculate rtt.
4489 * Fix it at least with timestamps.
4491 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4493 tcp_ack_saw_tstamp(sk
, 0);
4495 if (tp
->rx_opt
.tstamp_ok
)
4496 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4498 /* Make sure socket is routed, for
4501 icsk
->icsk_af_ops
->rebuild_header(sk
);
4503 tcp_init_metrics(sk
);
4505 tcp_init_congestion_control(sk
);
4507 /* Prevent spurious tcp_cwnd_restart() on
4508 * first data packet.
4510 tp
->lsndtime
= tcp_time_stamp
;
4513 tcp_initialize_rcv_mss(sk
);
4514 tcp_init_buffer_space(sk
);
4515 tcp_fast_path_on(tp
);
4522 if (tp
->snd_una
== tp
->write_seq
) {
4523 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4524 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4525 dst_confirm(sk
->sk_dst_cache
);
4527 if (!sock_flag(sk
, SOCK_DEAD
))
4528 /* Wake up lingering close() */
4529 sk
->sk_state_change(sk
);
4533 if (tp
->linger2
< 0 ||
4534 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4535 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4537 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4541 tmo
= tcp_fin_time(sk
);
4542 if (tmo
> TCP_TIMEWAIT_LEN
) {
4543 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4544 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4545 /* Bad case. We could lose such FIN otherwise.
4546 * It is not a big problem, but it looks confusing
4547 * and not so rare event. We still can lose it now,
4548 * if it spins in bh_lock_sock(), but it is really
4551 inet_csk_reset_keepalive_timer(sk
, tmo
);
4553 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4561 if (tp
->snd_una
== tp
->write_seq
) {
4562 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4568 if (tp
->snd_una
== tp
->write_seq
) {
4569 tcp_update_metrics(sk
);
4578 /* step 6: check the URG bit */
4579 tcp_urg(sk
, skb
, th
);
4581 /* step 7: process the segment text */
4582 switch (sk
->sk_state
) {
4583 case TCP_CLOSE_WAIT
:
4586 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4590 /* RFC 793 says to queue data in these states,
4591 * RFC 1122 says we MUST send a reset.
4592 * BSD 4.4 also does reset.
4594 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4595 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4596 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4597 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4603 case TCP_ESTABLISHED
:
4604 tcp_data_queue(sk
, skb
);
4609 /* tcp_data could move socket to TIME-WAIT */
4610 if (sk
->sk_state
!= TCP_CLOSE
) {
4611 tcp_data_snd_check(sk
, tp
);
4612 tcp_ack_snd_check(sk
);
4622 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4623 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4624 EXPORT_SYMBOL(tcp_parse_options
);
4625 EXPORT_SYMBOL(tcp_rcv_established
);
4626 EXPORT_SYMBOL(tcp_rcv_state_process
);
4627 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);