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 __read_mostly
= 1;
76 int sysctl_tcp_window_scaling __read_mostly
= 1;
77 int sysctl_tcp_sack __read_mostly
= 1;
78 int sysctl_tcp_fack __read_mostly
= 1;
79 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
80 int sysctl_tcp_ecn __read_mostly
;
81 int sysctl_tcp_dsack __read_mostly
= 1;
82 int sysctl_tcp_app_win __read_mostly
= 31;
83 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
85 int sysctl_tcp_stdurg __read_mostly
;
86 int sysctl_tcp_rfc1337 __read_mostly
;
87 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
88 int sysctl_tcp_frto __read_mostly
;
89 int sysctl_tcp_nometrics_save __read_mostly
;
91 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
92 int sysctl_tcp_abc __read_mostly
;
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.*/
103 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
106 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
107 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
108 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
110 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
111 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
112 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
114 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
116 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118 /* Adapt the MSS value used to make delayed ack decision to the
121 static void tcp_measure_rcv_mss(struct sock
*sk
,
122 const struct sk_buff
*skb
)
124 struct inet_connection_sock
*icsk
= inet_csk(sk
);
125 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
128 icsk
->icsk_ack
.last_seg_size
= 0;
130 /* skb->len may jitter because of SACKs, even if peer
131 * sends good full-sized frames.
133 len
= skb_shinfo(skb
)->gso_size
?: skb
->len
;
134 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
135 icsk
->icsk_ack
.rcv_mss
= len
;
137 /* Otherwise, we make more careful check taking into account,
138 * that SACKs block is variable.
140 * "len" is invariant segment length, including TCP header.
142 len
+= skb
->data
- skb
->h
.raw
;
143 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
144 /* If PSH is not set, packet should be
145 * full sized, provided peer TCP is not badly broken.
146 * This observation (if it is correct 8)) allows
147 * to handle super-low mtu links fairly.
149 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
150 !(tcp_flag_word(skb
->h
.th
)&TCP_REMNANT
))) {
151 /* Subtract also invariant (if peer is RFC compliant),
152 * tcp header plus fixed timestamp option length.
153 * Resulting "len" is MSS free of SACK jitter.
155 len
-= tcp_sk(sk
)->tcp_header_len
;
156 icsk
->icsk_ack
.last_seg_size
= len
;
158 icsk
->icsk_ack
.rcv_mss
= len
;
162 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
163 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
164 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
168 static void tcp_incr_quickack(struct sock
*sk
)
170 struct inet_connection_sock
*icsk
= inet_csk(sk
);
171 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
175 if (quickacks
> icsk
->icsk_ack
.quick
)
176 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
179 void tcp_enter_quickack_mode(struct sock
*sk
)
181 struct inet_connection_sock
*icsk
= inet_csk(sk
);
182 tcp_incr_quickack(sk
);
183 icsk
->icsk_ack
.pingpong
= 0;
184 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
187 /* Send ACKs quickly, if "quick" count is not exhausted
188 * and the session is not interactive.
191 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
193 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
194 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
197 /* Buffer size and advertised window tuning.
199 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
202 static void tcp_fixup_sndbuf(struct sock
*sk
)
204 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
205 sizeof(struct sk_buff
);
207 if (sk
->sk_sndbuf
< 3 * sndmem
)
208 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
211 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
213 * All tcp_full_space() is split to two parts: "network" buffer, allocated
214 * forward and advertised in receiver window (tp->rcv_wnd) and
215 * "application buffer", required to isolate scheduling/application
216 * latencies from network.
217 * window_clamp is maximal advertised window. It can be less than
218 * tcp_full_space(), in this case tcp_full_space() - window_clamp
219 * is reserved for "application" buffer. The less window_clamp is
220 * the smoother our behaviour from viewpoint of network, but the lower
221 * throughput and the higher sensitivity of the connection to losses. 8)
223 * rcv_ssthresh is more strict window_clamp used at "slow start"
224 * phase to predict further behaviour of this connection.
225 * It is used for two goals:
226 * - to enforce header prediction at sender, even when application
227 * requires some significant "application buffer". It is check #1.
228 * - to prevent pruning of receive queue because of misprediction
229 * of receiver window. Check #2.
231 * The scheme does not work when sender sends good segments opening
232 * window and then starts to feed us spaghetti. But it should work
233 * in common situations. Otherwise, we have to rely on queue collapsing.
236 /* Slow part of check#2. */
237 static int __tcp_grow_window(const struct sock
*sk
, struct tcp_sock
*tp
,
238 const struct sk_buff
*skb
)
241 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
242 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2])/2;
244 while (tp
->rcv_ssthresh
<= window
) {
245 if (truesize
<= skb
->len
)
246 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
254 static void tcp_grow_window(struct sock
*sk
, struct tcp_sock
*tp
,
258 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
259 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
260 !tcp_memory_pressure
) {
263 /* Check #2. Increase window, if skb with such overhead
264 * will fit to rcvbuf in future.
266 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
269 incr
= __tcp_grow_window(sk
, tp
, skb
);
272 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
273 inet_csk(sk
)->icsk_ack
.quick
|= 1;
278 /* 3. Tuning rcvbuf, when connection enters established state. */
280 static void tcp_fixup_rcvbuf(struct sock
*sk
)
282 struct tcp_sock
*tp
= tcp_sk(sk
);
283 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
285 /* Try to select rcvbuf so that 4 mss-sized segments
286 * will fit to window and corresponding skbs will fit to our rcvbuf.
287 * (was 3; 4 is minimum to allow fast retransmit to work.)
289 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
291 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
292 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
295 /* 4. Try to fixup all. It is made immediately after connection enters
298 static void tcp_init_buffer_space(struct sock
*sk
)
300 struct tcp_sock
*tp
= tcp_sk(sk
);
303 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
304 tcp_fixup_rcvbuf(sk
);
305 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
306 tcp_fixup_sndbuf(sk
);
308 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
310 maxwin
= tcp_full_space(sk
);
312 if (tp
->window_clamp
>= maxwin
) {
313 tp
->window_clamp
= maxwin
;
315 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
316 tp
->window_clamp
= max(maxwin
-
317 (maxwin
>> sysctl_tcp_app_win
),
321 /* Force reservation of one segment. */
322 if (sysctl_tcp_app_win
&&
323 tp
->window_clamp
> 2 * tp
->advmss
&&
324 tp
->window_clamp
+ tp
->advmss
> maxwin
)
325 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
327 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
328 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
331 /* 5. Recalculate window clamp after socket hit its memory bounds. */
332 static void tcp_clamp_window(struct sock
*sk
, struct tcp_sock
*tp
)
334 struct inet_connection_sock
*icsk
= inet_csk(sk
);
336 icsk
->icsk_ack
.quick
= 0;
338 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
339 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
340 !tcp_memory_pressure
&&
341 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
342 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
345 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
346 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
350 /* Initialize RCV_MSS value.
351 * RCV_MSS is an our guess about MSS used by the peer.
352 * We haven't any direct information about the MSS.
353 * It's better to underestimate the RCV_MSS rather than overestimate.
354 * Overestimations make us ACKing less frequently than needed.
355 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
357 void tcp_initialize_rcv_mss(struct sock
*sk
)
359 struct tcp_sock
*tp
= tcp_sk(sk
);
360 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
362 hint
= min(hint
, tp
->rcv_wnd
/2);
363 hint
= min(hint
, TCP_MIN_RCVMSS
);
364 hint
= max(hint
, TCP_MIN_MSS
);
366 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
369 /* Receiver "autotuning" code.
371 * The algorithm for RTT estimation w/o timestamps is based on
372 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
373 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
375 * More detail on this code can be found at
376 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
377 * though this reference is out of date. A new paper
380 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
382 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
388 if (new_sample
!= 0) {
389 /* If we sample in larger samples in the non-timestamp
390 * case, we could grossly overestimate the RTT especially
391 * with chatty applications or bulk transfer apps which
392 * are stalled on filesystem I/O.
394 * Also, since we are only going for a minimum in the
395 * non-timestamp case, we do not smooth things out
396 * else with timestamps disabled convergence takes too
400 m
-= (new_sample
>> 3);
402 } else if (m
< new_sample
)
405 /* No previous measure. */
409 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
410 tp
->rcv_rtt_est
.rtt
= new_sample
;
413 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
415 if (tp
->rcv_rtt_est
.time
== 0)
417 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
419 tcp_rcv_rtt_update(tp
,
420 jiffies
- tp
->rcv_rtt_est
.time
,
424 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
425 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
428 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
430 struct tcp_sock
*tp
= tcp_sk(sk
);
431 if (tp
->rx_opt
.rcv_tsecr
&&
432 (TCP_SKB_CB(skb
)->end_seq
-
433 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
434 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
438 * This function should be called every time data is copied to user space.
439 * It calculates the appropriate TCP receive buffer space.
441 void tcp_rcv_space_adjust(struct sock
*sk
)
443 struct tcp_sock
*tp
= tcp_sk(sk
);
447 if (tp
->rcvq_space
.time
== 0)
450 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
451 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
452 tp
->rcv_rtt_est
.rtt
== 0)
455 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
457 space
= max(tp
->rcvq_space
.space
, space
);
459 if (tp
->rcvq_space
.space
!= space
) {
462 tp
->rcvq_space
.space
= space
;
464 if (sysctl_tcp_moderate_rcvbuf
&&
465 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
466 int new_clamp
= space
;
468 /* Receive space grows, normalize in order to
469 * take into account packet headers and sk_buff
470 * structure overhead.
475 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
476 16 + sizeof(struct sk_buff
));
477 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
480 space
= min(space
, sysctl_tcp_rmem
[2]);
481 if (space
> sk
->sk_rcvbuf
) {
482 sk
->sk_rcvbuf
= space
;
484 /* Make the window clamp follow along. */
485 tp
->window_clamp
= new_clamp
;
491 tp
->rcvq_space
.seq
= tp
->copied_seq
;
492 tp
->rcvq_space
.time
= tcp_time_stamp
;
495 /* There is something which you must keep in mind when you analyze the
496 * behavior of the tp->ato delayed ack timeout interval. When a
497 * connection starts up, we want to ack as quickly as possible. The
498 * problem is that "good" TCP's do slow start at the beginning of data
499 * transmission. The means that until we send the first few ACK's the
500 * sender will sit on his end and only queue most of his data, because
501 * he can only send snd_cwnd unacked packets at any given time. For
502 * each ACK we send, he increments snd_cwnd and transmits more of his
505 static void tcp_event_data_recv(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
507 struct inet_connection_sock
*icsk
= inet_csk(sk
);
510 inet_csk_schedule_ack(sk
);
512 tcp_measure_rcv_mss(sk
, skb
);
514 tcp_rcv_rtt_measure(tp
);
516 now
= tcp_time_stamp
;
518 if (!icsk
->icsk_ack
.ato
) {
519 /* The _first_ data packet received, initialize
520 * delayed ACK engine.
522 tcp_incr_quickack(sk
);
523 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
525 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
527 if (m
<= TCP_ATO_MIN
/2) {
528 /* The fastest case is the first. */
529 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
530 } else if (m
< icsk
->icsk_ack
.ato
) {
531 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
532 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
533 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
534 } else if (m
> icsk
->icsk_rto
) {
535 /* Too long gap. Apparently sender failed to
536 * restart window, so that we send ACKs quickly.
538 tcp_incr_quickack(sk
);
539 sk_stream_mem_reclaim(sk
);
542 icsk
->icsk_ack
.lrcvtime
= now
;
544 TCP_ECN_check_ce(tp
, skb
);
547 tcp_grow_window(sk
, tp
, skb
);
550 /* Called to compute a smoothed rtt estimate. The data fed to this
551 * routine either comes from timestamps, or from segments that were
552 * known _not_ to have been retransmitted [see Karn/Partridge
553 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
554 * piece by Van Jacobson.
555 * NOTE: the next three routines used to be one big routine.
556 * To save cycles in the RFC 1323 implementation it was better to break
557 * it up into three procedures. -- erics
559 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
561 struct tcp_sock
*tp
= tcp_sk(sk
);
562 long m
= mrtt
; /* RTT */
564 /* The following amusing code comes from Jacobson's
565 * article in SIGCOMM '88. Note that rtt and mdev
566 * are scaled versions of rtt and mean deviation.
567 * This is designed to be as fast as possible
568 * m stands for "measurement".
570 * On a 1990 paper the rto value is changed to:
571 * RTO = rtt + 4 * mdev
573 * Funny. This algorithm seems to be very broken.
574 * These formulae increase RTO, when it should be decreased, increase
575 * too slowly, when it should be increased quickly, decrease too quickly
576 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
577 * does not matter how to _calculate_ it. Seems, it was trap
578 * that VJ failed to avoid. 8)
583 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
584 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
586 m
= -m
; /* m is now abs(error) */
587 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
588 /* This is similar to one of Eifel findings.
589 * Eifel blocks mdev updates when rtt decreases.
590 * This solution is a bit different: we use finer gain
591 * for mdev in this case (alpha*beta).
592 * Like Eifel it also prevents growth of rto,
593 * but also it limits too fast rto decreases,
594 * happening in pure Eifel.
599 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
601 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
602 if (tp
->mdev
> tp
->mdev_max
) {
603 tp
->mdev_max
= tp
->mdev
;
604 if (tp
->mdev_max
> tp
->rttvar
)
605 tp
->rttvar
= tp
->mdev_max
;
607 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
608 if (tp
->mdev_max
< tp
->rttvar
)
609 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
610 tp
->rtt_seq
= tp
->snd_nxt
;
611 tp
->mdev_max
= TCP_RTO_MIN
;
614 /* no previous measure. */
615 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
616 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
617 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
618 tp
->rtt_seq
= tp
->snd_nxt
;
622 /* Calculate rto without backoff. This is the second half of Van Jacobson's
623 * routine referred to above.
625 static inline void tcp_set_rto(struct sock
*sk
)
627 const struct tcp_sock
*tp
= tcp_sk(sk
);
628 /* Old crap is replaced with new one. 8)
631 * 1. If rtt variance happened to be less 50msec, it is hallucination.
632 * It cannot be less due to utterly erratic ACK generation made
633 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
634 * to do with delayed acks, because at cwnd>2 true delack timeout
635 * is invisible. Actually, Linux-2.4 also generates erratic
636 * ACKs in some circumstances.
638 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
640 /* 2. Fixups made earlier cannot be right.
641 * If we do not estimate RTO correctly without them,
642 * all the algo is pure shit and should be replaced
643 * with correct one. It is exactly, which we pretend to do.
647 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
648 * guarantees that rto is higher.
650 static inline void tcp_bound_rto(struct sock
*sk
)
652 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
653 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
656 /* Save metrics learned by this TCP session.
657 This function is called only, when TCP finishes successfully
658 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
660 void tcp_update_metrics(struct sock
*sk
)
662 struct tcp_sock
*tp
= tcp_sk(sk
);
663 struct dst_entry
*dst
= __sk_dst_get(sk
);
665 if (sysctl_tcp_nometrics_save
)
670 if (dst
&& (dst
->flags
&DST_HOST
)) {
671 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
674 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
675 /* This session failed to estimate rtt. Why?
676 * Probably, no packets returned in time.
679 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
680 dst
->metrics
[RTAX_RTT
-1] = 0;
684 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
686 /* If newly calculated rtt larger than stored one,
687 * store new one. Otherwise, use EWMA. Remember,
688 * rtt overestimation is always better than underestimation.
690 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
692 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
694 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
697 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
701 /* Scale deviation to rttvar fixed point */
706 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
707 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
709 dst
->metrics
[RTAX_RTTVAR
-1] -=
710 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
713 if (tp
->snd_ssthresh
>= 0xFFFF) {
714 /* Slow start still did not finish. */
715 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
716 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
717 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
718 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
719 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
720 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
721 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
722 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
723 icsk
->icsk_ca_state
== TCP_CA_Open
) {
724 /* Cong. avoidance phase, cwnd is reliable. */
725 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
726 dst
->metrics
[RTAX_SSTHRESH
-1] =
727 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
728 if (!dst_metric_locked(dst
, RTAX_CWND
))
729 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
731 /* Else slow start did not finish, cwnd is non-sense,
732 ssthresh may be also invalid.
734 if (!dst_metric_locked(dst
, RTAX_CWND
))
735 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
736 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
737 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
738 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
739 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
742 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
743 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
744 tp
->reordering
!= sysctl_tcp_reordering
)
745 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
750 /* Numbers are taken from RFC2414. */
751 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
753 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
756 if (tp
->mss_cache
> 1460)
759 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
761 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
764 /* Set slow start threshold and cwnd not falling to slow start */
765 void tcp_enter_cwr(struct sock
*sk
)
767 struct tcp_sock
*tp
= tcp_sk(sk
);
769 tp
->prior_ssthresh
= 0;
771 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
773 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
774 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
775 tcp_packets_in_flight(tp
) + 1U);
776 tp
->snd_cwnd_cnt
= 0;
777 tp
->high_seq
= tp
->snd_nxt
;
778 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
779 TCP_ECN_queue_cwr(tp
);
781 tcp_set_ca_state(sk
, TCP_CA_CWR
);
785 /* Initialize metrics on socket. */
787 static void tcp_init_metrics(struct sock
*sk
)
789 struct tcp_sock
*tp
= tcp_sk(sk
);
790 struct dst_entry
*dst
= __sk_dst_get(sk
);
797 if (dst_metric_locked(dst
, RTAX_CWND
))
798 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
799 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
800 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
801 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
802 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
804 if (dst_metric(dst
, RTAX_REORDERING
) &&
805 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
806 tp
->rx_opt
.sack_ok
&= ~2;
807 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
810 if (dst_metric(dst
, RTAX_RTT
) == 0)
813 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
816 /* Initial rtt is determined from SYN,SYN-ACK.
817 * The segment is small and rtt may appear much
818 * less than real one. Use per-dst memory
819 * to make it more realistic.
821 * A bit of theory. RTT is time passed after "normal" sized packet
822 * is sent until it is ACKed. In normal circumstances sending small
823 * packets force peer to delay ACKs and calculation is correct too.
824 * The algorithm is adaptive and, provided we follow specs, it
825 * NEVER underestimate RTT. BUT! If peer tries to make some clever
826 * tricks sort of "quick acks" for time long enough to decrease RTT
827 * to low value, and then abruptly stops to do it and starts to delay
828 * ACKs, wait for troubles.
830 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
831 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
832 tp
->rtt_seq
= tp
->snd_nxt
;
834 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
835 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
836 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
840 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
842 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
843 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
847 /* Play conservative. If timestamps are not
848 * supported, TCP will fail to recalculate correct
849 * rtt, if initial rto is too small. FORGET ALL AND RESET!
851 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
853 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
854 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
858 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
861 struct tcp_sock
*tp
= tcp_sk(sk
);
862 if (metric
> tp
->reordering
) {
863 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
865 /* This exciting event is worth to be remembered. 8) */
867 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
869 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
871 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
873 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
874 #if FASTRETRANS_DEBUG > 1
875 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
876 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
880 tp
->undo_marker
? tp
->undo_retrans
: 0);
882 /* Disable FACK yet. */
883 tp
->rx_opt
.sack_ok
&= ~2;
887 /* This procedure tags the retransmission queue when SACKs arrive.
889 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
890 * Packets in queue with these bits set are counted in variables
891 * sacked_out, retrans_out and lost_out, correspondingly.
893 * Valid combinations are:
894 * Tag InFlight Description
895 * 0 1 - orig segment is in flight.
896 * S 0 - nothing flies, orig reached receiver.
897 * L 0 - nothing flies, orig lost by net.
898 * R 2 - both orig and retransmit are in flight.
899 * L|R 1 - orig is lost, retransmit is in flight.
900 * S|R 1 - orig reached receiver, retrans is still in flight.
901 * (L|S|R is logically valid, it could occur when L|R is sacked,
902 * but it is equivalent to plain S and code short-curcuits it to S.
903 * L|S is logically invalid, it would mean -1 packet in flight 8))
905 * These 6 states form finite state machine, controlled by the following events:
906 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
907 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
908 * 3. Loss detection event of one of three flavors:
909 * A. Scoreboard estimator decided the packet is lost.
910 * A'. Reno "three dupacks" marks head of queue lost.
911 * A''. Its FACK modfication, head until snd.fack is lost.
912 * B. SACK arrives sacking data transmitted after never retransmitted
914 * C. SACK arrives sacking SND.NXT at the moment, when the
915 * segment was retransmitted.
916 * 4. D-SACK added new rule: D-SACK changes any tag to S.
918 * It is pleasant to note, that state diagram turns out to be commutative,
919 * so that we are allowed not to be bothered by order of our actions,
920 * when multiple events arrive simultaneously. (see the function below).
922 * Reordering detection.
923 * --------------------
924 * Reordering metric is maximal distance, which a packet can be displaced
925 * in packet stream. With SACKs we can estimate it:
927 * 1. SACK fills old hole and the corresponding segment was not
928 * ever retransmitted -> reordering. Alas, we cannot use it
929 * when segment was retransmitted.
930 * 2. The last flaw is solved with D-SACK. D-SACK arrives
931 * for retransmitted and already SACKed segment -> reordering..
932 * Both of these heuristics are not used in Loss state, when we cannot
933 * account for retransmits accurately.
936 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
938 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
939 struct tcp_sock
*tp
= tcp_sk(sk
);
940 unsigned char *ptr
= ack_skb
->h
.raw
+ TCP_SKB_CB(ack_skb
)->sacked
;
941 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
942 struct sk_buff
*cached_skb
;
943 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
944 int reord
= tp
->packets_out
;
946 u32 lost_retrans
= 0;
949 int cached_fack_count
;
951 int first_sack_index
;
955 prior_fackets
= tp
->fackets_out
;
957 /* Check for D-SACK. */
958 if (before(ntohl(sp
[0].start_seq
), TCP_SKB_CB(ack_skb
)->ack_seq
)) {
960 tp
->rx_opt
.sack_ok
|= 4;
961 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
962 } else if (num_sacks
> 1 &&
963 !after(ntohl(sp
[0].end_seq
), ntohl(sp
[1].end_seq
)) &&
964 !before(ntohl(sp
[0].start_seq
), ntohl(sp
[1].start_seq
))) {
966 tp
->rx_opt
.sack_ok
|= 4;
967 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
970 /* D-SACK for already forgotten data...
971 * Do dumb counting. */
973 !after(ntohl(sp
[0].end_seq
), prior_snd_una
) &&
974 after(ntohl(sp
[0].end_seq
), tp
->undo_marker
))
977 /* Eliminate too old ACKs, but take into
978 * account more or less fresh ones, they can
979 * contain valid SACK info.
981 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
985 * if the only SACK change is the increase of the end_seq of
986 * the first block then only apply that SACK block
987 * and use retrans queue hinting otherwise slowpath */
989 for (i
= 0; i
< num_sacks
; i
++) {
990 __be32 start_seq
= sp
[i
].start_seq
;
991 __be32 end_seq
= sp
[i
].end_seq
;
994 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
997 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
998 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1001 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1002 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1004 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1005 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1006 tp
->recv_sack_cache
[i
].start_seq
= 0;
1007 tp
->recv_sack_cache
[i
].end_seq
= 0;
1010 first_sack_index
= 0;
1015 tp
->fastpath_skb_hint
= NULL
;
1017 /* order SACK blocks to allow in order walk of the retrans queue */
1018 for (i
= num_sacks
-1; i
> 0; i
--) {
1019 for (j
= 0; j
< i
; j
++){
1020 if (after(ntohl(sp
[j
].start_seq
),
1021 ntohl(sp
[j
+1].start_seq
))){
1022 struct tcp_sack_block_wire tmp
;
1028 /* Track where the first SACK block goes to */
1029 if (j
== first_sack_index
)
1030 first_sack_index
= j
+1;
1037 /* clear flag as used for different purpose in following code */
1040 /* Use SACK fastpath hint if valid */
1041 cached_skb
= tp
->fastpath_skb_hint
;
1042 cached_fack_count
= tp
->fastpath_cnt_hint
;
1044 cached_skb
= sk
->sk_write_queue
.next
;
1045 cached_fack_count
= 0;
1048 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1049 struct sk_buff
*skb
;
1050 __u32 start_seq
= ntohl(sp
->start_seq
);
1051 __u32 end_seq
= ntohl(sp
->end_seq
);
1055 fack_count
= cached_fack_count
;
1057 /* Event "B" in the comment above. */
1058 if (after(end_seq
, tp
->high_seq
))
1059 flag
|= FLAG_DATA_LOST
;
1061 sk_stream_for_retrans_queue_from(skb
, sk
) {
1062 int in_sack
, pcount
;
1066 cached_fack_count
= fack_count
;
1067 if (i
== first_sack_index
) {
1068 tp
->fastpath_skb_hint
= skb
;
1069 tp
->fastpath_cnt_hint
= fack_count
;
1072 /* The retransmission queue is always in order, so
1073 * we can short-circuit the walk early.
1075 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1078 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1079 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1081 pcount
= tcp_skb_pcount(skb
);
1083 if (pcount
> 1 && !in_sack
&&
1084 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1085 unsigned int pkt_len
;
1087 in_sack
= !after(start_seq
,
1088 TCP_SKB_CB(skb
)->seq
);
1091 pkt_len
= (start_seq
-
1092 TCP_SKB_CB(skb
)->seq
);
1094 pkt_len
= (end_seq
-
1095 TCP_SKB_CB(skb
)->seq
);
1096 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1098 pcount
= tcp_skb_pcount(skb
);
1101 fack_count
+= pcount
;
1103 sacked
= TCP_SKB_CB(skb
)->sacked
;
1105 /* Account D-SACK for retransmitted packet. */
1106 if ((dup_sack
&& in_sack
) &&
1107 (sacked
& TCPCB_RETRANS
) &&
1108 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1111 /* The frame is ACKed. */
1112 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1113 if (sacked
&TCPCB_RETRANS
) {
1114 if ((dup_sack
&& in_sack
) &&
1115 (sacked
&TCPCB_SACKED_ACKED
))
1116 reord
= min(fack_count
, reord
);
1118 /* If it was in a hole, we detected reordering. */
1119 if (fack_count
< prior_fackets
&&
1120 !(sacked
&TCPCB_SACKED_ACKED
))
1121 reord
= min(fack_count
, reord
);
1124 /* Nothing to do; acked frame is about to be dropped. */
1128 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1129 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1130 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1131 lost_retrans
= end_seq
;
1136 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1137 if (sacked
& TCPCB_SACKED_RETRANS
) {
1138 /* If the segment is not tagged as lost,
1139 * we do not clear RETRANS, believing
1140 * that retransmission is still in flight.
1142 if (sacked
& TCPCB_LOST
) {
1143 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1144 tp
->lost_out
-= tcp_skb_pcount(skb
);
1145 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1147 /* clear lost hint */
1148 tp
->retransmit_skb_hint
= NULL
;
1151 /* New sack for not retransmitted frame,
1152 * which was in hole. It is reordering.
1154 if (!(sacked
& TCPCB_RETRANS
) &&
1155 fack_count
< prior_fackets
)
1156 reord
= min(fack_count
, reord
);
1158 if (sacked
& TCPCB_LOST
) {
1159 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1160 tp
->lost_out
-= tcp_skb_pcount(skb
);
1162 /* clear lost hint */
1163 tp
->retransmit_skb_hint
= NULL
;
1165 /* SACK enhanced F-RTO detection.
1166 * Set flag if and only if non-rexmitted
1167 * segments below frto_highmark are
1168 * SACKed (RFC4138; Appendix B).
1169 * Clearing correct due to in-order walk
1171 if (after(end_seq
, tp
->frto_highmark
)) {
1172 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1174 if (!(sacked
& TCPCB_RETRANS
))
1175 flag
|= FLAG_ONLY_ORIG_SACKED
;
1179 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1180 flag
|= FLAG_DATA_SACKED
;
1181 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1183 if (fack_count
> tp
->fackets_out
)
1184 tp
->fackets_out
= fack_count
;
1186 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1187 reord
= min(fack_count
, reord
);
1190 /* D-SACK. We can detect redundant retransmission
1191 * in S|R and plain R frames and clear it.
1192 * undo_retrans is decreased above, L|R frames
1193 * are accounted above as well.
1196 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1197 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1198 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1199 tp
->retransmit_skb_hint
= NULL
;
1204 /* Check for lost retransmit. This superb idea is
1205 * borrowed from "ratehalving". Event "C".
1206 * Later note: FACK people cheated me again 8),
1207 * we have to account for reordering! Ugly,
1210 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1211 struct sk_buff
*skb
;
1213 sk_stream_for_retrans_queue(skb
, sk
) {
1214 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1216 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1218 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1219 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1221 !before(lost_retrans
,
1222 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1224 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1225 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1227 /* clear lost hint */
1228 tp
->retransmit_skb_hint
= NULL
;
1230 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1231 tp
->lost_out
+= tcp_skb_pcount(skb
);
1232 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1233 flag
|= FLAG_DATA_SACKED
;
1234 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1240 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1242 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
&&
1243 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1244 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1246 #if FASTRETRANS_DEBUG > 0
1247 BUG_TRAP((int)tp
->sacked_out
>= 0);
1248 BUG_TRAP((int)tp
->lost_out
>= 0);
1249 BUG_TRAP((int)tp
->retrans_out
>= 0);
1250 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1255 /* F-RTO can only be used if these conditions are satisfied:
1256 * - there must be some unsent new data
1257 * - the advertised window should allow sending it
1258 * - TCP has never retransmitted anything other than head (SACK enhanced
1259 * variant from Appendix B of RFC4138 is more robust here)
1261 int tcp_use_frto(struct sock
*sk
)
1263 const struct tcp_sock
*tp
= tcp_sk(sk
);
1264 struct sk_buff
*skb
;
1266 if (!sysctl_tcp_frto
|| !sk
->sk_send_head
||
1267 after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
1268 tp
->snd_una
+ tp
->snd_wnd
))
1274 /* Avoid expensive walking of rexmit queue if possible */
1275 if (tp
->retrans_out
> 1)
1278 skb
= skb_peek(&sk
->sk_write_queue
)->next
; /* Skips head */
1279 sk_stream_for_retrans_queue_from(skb
, sk
) {
1280 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1282 /* Short-circuit when first non-SACKed skb has been checked */
1283 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1289 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1290 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1291 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1292 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1293 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1294 * bits are handled if the Loss state is really to be entered (in
1295 * tcp_enter_frto_loss).
1297 * Do like tcp_enter_loss() would; when RTO expires the second time it
1299 * "Reduce ssthresh if it has not yet been made inside this window."
1301 void tcp_enter_frto(struct sock
*sk
)
1303 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1304 struct tcp_sock
*tp
= tcp_sk(sk
);
1305 struct sk_buff
*skb
;
1307 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1308 tp
->snd_una
== tp
->high_seq
||
1309 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1310 !icsk
->icsk_retransmits
)) {
1311 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1312 /* Our state is too optimistic in ssthresh() call because cwnd
1313 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1314 * recovery has not yet completed. Pattern would be this: RTO,
1315 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1317 * RFC4138 should be more specific on what to do, even though
1318 * RTO is quite unlikely to occur after the first Cumulative ACK
1319 * due to back-off and complexity of triggering events ...
1321 if (tp
->frto_counter
) {
1323 stored_cwnd
= tp
->snd_cwnd
;
1325 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1326 tp
->snd_cwnd
= stored_cwnd
;
1328 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1330 /* ... in theory, cong.control module could do "any tricks" in
1331 * ssthresh(), which means that ca_state, lost bits and lost_out
1332 * counter would have to be faked before the call occurs. We
1333 * consider that too expensive, unlikely and hacky, so modules
1334 * using these in ssthresh() must deal these incompatibility
1335 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1337 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1340 tp
->undo_marker
= tp
->snd_una
;
1341 tp
->undo_retrans
= 0;
1343 skb
= skb_peek(&sk
->sk_write_queue
);
1344 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1345 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1346 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1348 tcp_sync_left_out(tp
);
1350 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1351 * The last condition is necessary at least in tp->frto_counter case.
1353 if (IsSackFrto() && (tp
->frto_counter
||
1354 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1355 after(tp
->high_seq
, tp
->snd_una
)) {
1356 tp
->frto_highmark
= tp
->high_seq
;
1358 tp
->frto_highmark
= tp
->snd_nxt
;
1360 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1361 tp
->high_seq
= tp
->snd_nxt
;
1362 tp
->frto_counter
= 1;
1365 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1366 * which indicates that we should follow the traditional RTO recovery,
1367 * i.e. mark everything lost and do go-back-N retransmission.
1369 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1371 struct tcp_sock
*tp
= tcp_sk(sk
);
1372 struct sk_buff
*skb
;
1377 tp
->fackets_out
= 0;
1378 tp
->retrans_out
= 0;
1380 sk_stream_for_retrans_queue(skb
, sk
) {
1381 cnt
+= tcp_skb_pcount(skb
);
1383 * Count the retransmission made on RTO correctly (only when
1384 * waiting for the first ACK and did not get it)...
1386 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1387 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1388 /* ...enter this if branch just for the first segment */
1389 flag
|= FLAG_DATA_ACKED
;
1391 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1393 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1395 /* Do not mark those segments lost that were
1396 * forward transmitted after RTO
1398 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1399 tp
->frto_highmark
)) {
1400 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1401 tp
->lost_out
+= tcp_skb_pcount(skb
);
1404 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1405 tp
->fackets_out
= cnt
;
1408 tcp_sync_left_out(tp
);
1410 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1411 tp
->snd_cwnd_cnt
= 0;
1412 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1413 tp
->undo_marker
= 0;
1414 tp
->frto_counter
= 0;
1416 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1417 sysctl_tcp_reordering
);
1418 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1419 tp
->high_seq
= tp
->frto_highmark
;
1420 TCP_ECN_queue_cwr(tp
);
1422 clear_all_retrans_hints(tp
);
1425 void tcp_clear_retrans(struct tcp_sock
*tp
)
1428 tp
->retrans_out
= 0;
1430 tp
->fackets_out
= 0;
1434 tp
->undo_marker
= 0;
1435 tp
->undo_retrans
= 0;
1438 /* Enter Loss state. If "how" is not zero, forget all SACK information
1439 * and reset tags completely, otherwise preserve SACKs. If receiver
1440 * dropped its ofo queue, we will know this due to reneging detection.
1442 void tcp_enter_loss(struct sock
*sk
, int how
)
1444 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1445 struct tcp_sock
*tp
= tcp_sk(sk
);
1446 struct sk_buff
*skb
;
1449 /* Reduce ssthresh if it has not yet been made inside this window. */
1450 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1451 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1452 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1453 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1454 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1457 tp
->snd_cwnd_cnt
= 0;
1458 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1460 tp
->bytes_acked
= 0;
1461 tcp_clear_retrans(tp
);
1463 /* Push undo marker, if it was plain RTO and nothing
1464 * was retransmitted. */
1466 tp
->undo_marker
= tp
->snd_una
;
1468 sk_stream_for_retrans_queue(skb
, sk
) {
1469 cnt
+= tcp_skb_pcount(skb
);
1470 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1471 tp
->undo_marker
= 0;
1472 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1473 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1474 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1475 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1476 tp
->lost_out
+= tcp_skb_pcount(skb
);
1478 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1479 tp
->fackets_out
= cnt
;
1482 tcp_sync_left_out(tp
);
1484 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1485 sysctl_tcp_reordering
);
1486 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1487 tp
->high_seq
= tp
->snd_nxt
;
1488 TCP_ECN_queue_cwr(tp
);
1490 clear_all_retrans_hints(tp
);
1493 static int tcp_check_sack_reneging(struct sock
*sk
)
1495 struct sk_buff
*skb
;
1497 /* If ACK arrived pointing to a remembered SACK,
1498 * it means that our remembered SACKs do not reflect
1499 * real state of receiver i.e.
1500 * receiver _host_ is heavily congested (or buggy).
1501 * Do processing similar to RTO timeout.
1503 if ((skb
= skb_peek(&sk
->sk_write_queue
)) != NULL
&&
1504 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1505 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1506 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1508 tcp_enter_loss(sk
, 1);
1509 icsk
->icsk_retransmits
++;
1510 tcp_retransmit_skb(sk
, skb_peek(&sk
->sk_write_queue
));
1511 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1512 icsk
->icsk_rto
, TCP_RTO_MAX
);
1518 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1520 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1523 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1525 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1528 static inline int tcp_head_timedout(struct sock
*sk
, struct tcp_sock
*tp
)
1530 return tp
->packets_out
&&
1531 tcp_skb_timedout(sk
, skb_peek(&sk
->sk_write_queue
));
1534 /* Linux NewReno/SACK/FACK/ECN state machine.
1535 * --------------------------------------
1537 * "Open" Normal state, no dubious events, fast path.
1538 * "Disorder" In all the respects it is "Open",
1539 * but requires a bit more attention. It is entered when
1540 * we see some SACKs or dupacks. It is split of "Open"
1541 * mainly to move some processing from fast path to slow one.
1542 * "CWR" CWND was reduced due to some Congestion Notification event.
1543 * It can be ECN, ICMP source quench, local device congestion.
1544 * "Recovery" CWND was reduced, we are fast-retransmitting.
1545 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1547 * tcp_fastretrans_alert() is entered:
1548 * - each incoming ACK, if state is not "Open"
1549 * - when arrived ACK is unusual, namely:
1554 * Counting packets in flight is pretty simple.
1556 * in_flight = packets_out - left_out + retrans_out
1558 * packets_out is SND.NXT-SND.UNA counted in packets.
1560 * retrans_out is number of retransmitted segments.
1562 * left_out is number of segments left network, but not ACKed yet.
1564 * left_out = sacked_out + lost_out
1566 * sacked_out: Packets, which arrived to receiver out of order
1567 * and hence not ACKed. With SACKs this number is simply
1568 * amount of SACKed data. Even without SACKs
1569 * it is easy to give pretty reliable estimate of this number,
1570 * counting duplicate ACKs.
1572 * lost_out: Packets lost by network. TCP has no explicit
1573 * "loss notification" feedback from network (for now).
1574 * It means that this number can be only _guessed_.
1575 * Actually, it is the heuristics to predict lossage that
1576 * distinguishes different algorithms.
1578 * F.e. after RTO, when all the queue is considered as lost,
1579 * lost_out = packets_out and in_flight = retrans_out.
1581 * Essentially, we have now two algorithms counting
1584 * FACK: It is the simplest heuristics. As soon as we decided
1585 * that something is lost, we decide that _all_ not SACKed
1586 * packets until the most forward SACK are lost. I.e.
1587 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1588 * It is absolutely correct estimate, if network does not reorder
1589 * packets. And it loses any connection to reality when reordering
1590 * takes place. We use FACK by default until reordering
1591 * is suspected on the path to this destination.
1593 * NewReno: when Recovery is entered, we assume that one segment
1594 * is lost (classic Reno). While we are in Recovery and
1595 * a partial ACK arrives, we assume that one more packet
1596 * is lost (NewReno). This heuristics are the same in NewReno
1599 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1600 * deflation etc. CWND is real congestion window, never inflated, changes
1601 * only according to classic VJ rules.
1603 * Really tricky (and requiring careful tuning) part of algorithm
1604 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1605 * The first determines the moment _when_ we should reduce CWND and,
1606 * hence, slow down forward transmission. In fact, it determines the moment
1607 * when we decide that hole is caused by loss, rather than by a reorder.
1609 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1610 * holes, caused by lost packets.
1612 * And the most logically complicated part of algorithm is undo
1613 * heuristics. We detect false retransmits due to both too early
1614 * fast retransmit (reordering) and underestimated RTO, analyzing
1615 * timestamps and D-SACKs. When we detect that some segments were
1616 * retransmitted by mistake and CWND reduction was wrong, we undo
1617 * window reduction and abort recovery phase. This logic is hidden
1618 * inside several functions named tcp_try_undo_<something>.
1621 /* This function decides, when we should leave Disordered state
1622 * and enter Recovery phase, reducing congestion window.
1624 * Main question: may we further continue forward transmission
1625 * with the same cwnd?
1627 static int tcp_time_to_recover(struct sock
*sk
, struct tcp_sock
*tp
)
1631 /* Do not perform any recovery during FRTO algorithm */
1632 if (tp
->frto_counter
)
1635 /* Trick#1: The loss is proven. */
1639 /* Not-A-Trick#2 : Classic rule... */
1640 if (tcp_fackets_out(tp
) > tp
->reordering
)
1643 /* Trick#3 : when we use RFC2988 timer restart, fast
1644 * retransmit can be triggered by timeout of queue head.
1646 if (tcp_head_timedout(sk
, tp
))
1649 /* Trick#4: It is still not OK... But will it be useful to delay
1652 packets_out
= tp
->packets_out
;
1653 if (packets_out
<= tp
->reordering
&&
1654 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1655 !tcp_may_send_now(sk
, tp
)) {
1656 /* We have nothing to send. This connection is limited
1657 * either by receiver window or by application.
1665 /* If we receive more dupacks than we expected counting segments
1666 * in assumption of absent reordering, interpret this as reordering.
1667 * The only another reason could be bug in receiver TCP.
1669 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1671 struct tcp_sock
*tp
= tcp_sk(sk
);
1674 holes
= max(tp
->lost_out
, 1U);
1675 holes
= min(holes
, tp
->packets_out
);
1677 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1678 tp
->sacked_out
= tp
->packets_out
- holes
;
1679 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1683 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1685 static void tcp_add_reno_sack(struct sock
*sk
)
1687 struct tcp_sock
*tp
= tcp_sk(sk
);
1689 tcp_check_reno_reordering(sk
, 0);
1690 tcp_sync_left_out(tp
);
1693 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1695 static void tcp_remove_reno_sacks(struct sock
*sk
, struct tcp_sock
*tp
, int acked
)
1698 /* One ACK acked hole. The rest eat duplicate ACKs. */
1699 if (acked
-1 >= tp
->sacked_out
)
1702 tp
->sacked_out
-= acked
-1;
1704 tcp_check_reno_reordering(sk
, acked
);
1705 tcp_sync_left_out(tp
);
1708 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1711 tp
->left_out
= tp
->lost_out
;
1714 /* Mark head of queue up as lost. */
1715 static void tcp_mark_head_lost(struct sock
*sk
, struct tcp_sock
*tp
,
1716 int packets
, u32 high_seq
)
1718 struct sk_buff
*skb
;
1721 BUG_TRAP(packets
<= tp
->packets_out
);
1722 if (tp
->lost_skb_hint
) {
1723 skb
= tp
->lost_skb_hint
;
1724 cnt
= tp
->lost_cnt_hint
;
1726 skb
= sk
->sk_write_queue
.next
;
1730 sk_stream_for_retrans_queue_from(skb
, sk
) {
1731 /* TODO: do this better */
1732 /* this is not the most efficient way to do this... */
1733 tp
->lost_skb_hint
= skb
;
1734 tp
->lost_cnt_hint
= cnt
;
1735 cnt
+= tcp_skb_pcount(skb
);
1736 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1738 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1739 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1740 tp
->lost_out
+= tcp_skb_pcount(skb
);
1742 /* clear xmit_retransmit_queue hints
1743 * if this is beyond hint */
1744 if(tp
->retransmit_skb_hint
!= NULL
&&
1745 before(TCP_SKB_CB(skb
)->seq
,
1746 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
)) {
1748 tp
->retransmit_skb_hint
= NULL
;
1752 tcp_sync_left_out(tp
);
1755 /* Account newly detected lost packet(s) */
1757 static void tcp_update_scoreboard(struct sock
*sk
, struct tcp_sock
*tp
)
1760 int lost
= tp
->fackets_out
- tp
->reordering
;
1763 tcp_mark_head_lost(sk
, tp
, lost
, tp
->high_seq
);
1765 tcp_mark_head_lost(sk
, tp
, 1, tp
->high_seq
);
1768 /* New heuristics: it is possible only after we switched
1769 * to restart timer each time when something is ACKed.
1770 * Hence, we can detect timed out packets during fast
1771 * retransmit without falling to slow start.
1773 if (!IsReno(tp
) && tcp_head_timedout(sk
, tp
)) {
1774 struct sk_buff
*skb
;
1776 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1777 : sk
->sk_write_queue
.next
;
1779 sk_stream_for_retrans_queue_from(skb
, sk
) {
1780 if (!tcp_skb_timedout(sk
, skb
))
1783 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1784 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1785 tp
->lost_out
+= tcp_skb_pcount(skb
);
1787 /* clear xmit_retrans hint */
1788 if (tp
->retransmit_skb_hint
&&
1789 before(TCP_SKB_CB(skb
)->seq
,
1790 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1792 tp
->retransmit_skb_hint
= NULL
;
1796 tp
->scoreboard_skb_hint
= skb
;
1798 tcp_sync_left_out(tp
);
1802 /* CWND moderation, preventing bursts due to too big ACKs
1803 * in dubious situations.
1805 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1807 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1808 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1809 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1812 /* Lower bound on congestion window is slow start threshold
1813 * unless congestion avoidance choice decides to overide it.
1815 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1817 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1819 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1822 /* Decrease cwnd each second ack. */
1823 static void tcp_cwnd_down(struct sock
*sk
)
1825 struct tcp_sock
*tp
= tcp_sk(sk
);
1826 int decr
= tp
->snd_cwnd_cnt
+ 1;
1828 tp
->snd_cwnd_cnt
= decr
&1;
1831 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1832 tp
->snd_cwnd
-= decr
;
1834 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1835 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1838 /* Nothing was retransmitted or returned timestamp is less
1839 * than timestamp of the first retransmission.
1841 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1843 return !tp
->retrans_stamp
||
1844 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1845 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1848 /* Undo procedures. */
1850 #if FASTRETRANS_DEBUG > 1
1851 static void DBGUNDO(struct sock
*sk
, struct tcp_sock
*tp
, const char *msg
)
1853 struct inet_sock
*inet
= inet_sk(sk
);
1854 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1856 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1857 tp
->snd_cwnd
, tp
->left_out
,
1858 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1862 #define DBGUNDO(x...) do { } while (0)
1865 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1867 struct tcp_sock
*tp
= tcp_sk(sk
);
1869 if (tp
->prior_ssthresh
) {
1870 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1872 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1873 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1875 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1877 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1878 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1879 TCP_ECN_withdraw_cwr(tp
);
1882 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1884 tcp_moderate_cwnd(tp
);
1885 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1887 /* There is something screwy going on with the retrans hints after
1889 clear_all_retrans_hints(tp
);
1892 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1894 return tp
->undo_marker
&&
1895 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1898 /* People celebrate: "We love our President!" */
1899 static int tcp_try_undo_recovery(struct sock
*sk
, struct tcp_sock
*tp
)
1901 if (tcp_may_undo(tp
)) {
1902 /* Happy end! We did not retransmit anything
1903 * or our original transmission succeeded.
1905 DBGUNDO(sk
, tp
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1906 tcp_undo_cwr(sk
, 1);
1907 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1908 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1910 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1911 tp
->undo_marker
= 0;
1913 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1914 /* Hold old state until something *above* high_seq
1915 * is ACKed. For Reno it is MUST to prevent false
1916 * fast retransmits (RFC2582). SACK TCP is safe. */
1917 tcp_moderate_cwnd(tp
);
1920 tcp_set_ca_state(sk
, TCP_CA_Open
);
1924 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1925 static void tcp_try_undo_dsack(struct sock
*sk
, struct tcp_sock
*tp
)
1927 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1928 DBGUNDO(sk
, tp
, "D-SACK");
1929 tcp_undo_cwr(sk
, 1);
1930 tp
->undo_marker
= 0;
1931 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1935 /* Undo during fast recovery after partial ACK. */
1937 static int tcp_try_undo_partial(struct sock
*sk
, struct tcp_sock
*tp
,
1940 /* Partial ACK arrived. Force Hoe's retransmit. */
1941 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1943 if (tcp_may_undo(tp
)) {
1944 /* Plain luck! Hole if filled with delayed
1945 * packet, rather than with a retransmit.
1947 if (tp
->retrans_out
== 0)
1948 tp
->retrans_stamp
= 0;
1950 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
1952 DBGUNDO(sk
, tp
, "Hoe");
1953 tcp_undo_cwr(sk
, 0);
1954 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1956 /* So... Do not make Hoe's retransmit yet.
1957 * If the first packet was delayed, the rest
1958 * ones are most probably delayed as well.
1965 /* Undo during loss recovery after partial ACK. */
1966 static int tcp_try_undo_loss(struct sock
*sk
, struct tcp_sock
*tp
)
1968 if (tcp_may_undo(tp
)) {
1969 struct sk_buff
*skb
;
1970 sk_stream_for_retrans_queue(skb
, sk
) {
1971 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1974 clear_all_retrans_hints(tp
);
1976 DBGUNDO(sk
, tp
, "partial loss");
1978 tp
->left_out
= tp
->sacked_out
;
1979 tcp_undo_cwr(sk
, 1);
1980 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1981 inet_csk(sk
)->icsk_retransmits
= 0;
1982 tp
->undo_marker
= 0;
1984 tcp_set_ca_state(sk
, TCP_CA_Open
);
1990 static inline void tcp_complete_cwr(struct sock
*sk
)
1992 struct tcp_sock
*tp
= tcp_sk(sk
);
1993 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1994 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1995 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
1998 static void tcp_try_to_open(struct sock
*sk
, struct tcp_sock
*tp
, int flag
)
2000 tp
->left_out
= tp
->sacked_out
;
2002 if (tp
->retrans_out
== 0)
2003 tp
->retrans_stamp
= 0;
2008 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2009 int state
= TCP_CA_Open
;
2011 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
2012 state
= TCP_CA_Disorder
;
2014 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2015 tcp_set_ca_state(sk
, state
);
2016 tp
->high_seq
= tp
->snd_nxt
;
2018 tcp_moderate_cwnd(tp
);
2024 static void tcp_mtup_probe_failed(struct sock
*sk
)
2026 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2028 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2029 icsk
->icsk_mtup
.probe_size
= 0;
2032 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2034 struct tcp_sock
*tp
= tcp_sk(sk
);
2035 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2037 /* FIXME: breaks with very large cwnd */
2038 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2039 tp
->snd_cwnd
= tp
->snd_cwnd
*
2040 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2041 icsk
->icsk_mtup
.probe_size
;
2042 tp
->snd_cwnd_cnt
= 0;
2043 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2044 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2046 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2047 icsk
->icsk_mtup
.probe_size
= 0;
2048 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2052 /* Process an event, which can update packets-in-flight not trivially.
2053 * Main goal of this function is to calculate new estimate for left_out,
2054 * taking into account both packets sitting in receiver's buffer and
2055 * packets lost by network.
2057 * Besides that it does CWND reduction, when packet loss is detected
2058 * and changes state of machine.
2060 * It does _not_ decide what to send, it is made in function
2061 * tcp_xmit_retransmit_queue().
2064 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
2065 int prior_packets
, int flag
)
2067 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2068 struct tcp_sock
*tp
= tcp_sk(sk
);
2069 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
2071 /* Some technical things:
2072 * 1. Reno does not count dupacks (sacked_out) automatically. */
2073 if (!tp
->packets_out
)
2075 /* 2. SACK counts snd_fack in packets inaccurately. */
2076 if (tp
->sacked_out
== 0)
2077 tp
->fackets_out
= 0;
2079 /* Now state machine starts.
2080 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2082 tp
->prior_ssthresh
= 0;
2084 /* B. In all the states check for reneging SACKs. */
2085 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2088 /* C. Process data loss notification, provided it is valid. */
2089 if ((flag
&FLAG_DATA_LOST
) &&
2090 before(tp
->snd_una
, tp
->high_seq
) &&
2091 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2092 tp
->fackets_out
> tp
->reordering
) {
2093 tcp_mark_head_lost(sk
, tp
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
2094 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2097 /* D. Synchronize left_out to current state. */
2098 tcp_sync_left_out(tp
);
2100 /* E. Check state exit conditions. State can be terminated
2101 * when high_seq is ACKed. */
2102 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2103 BUG_TRAP(tp
->retrans_out
== 0);
2104 tp
->retrans_stamp
= 0;
2105 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2106 switch (icsk
->icsk_ca_state
) {
2108 icsk
->icsk_retransmits
= 0;
2109 if (tcp_try_undo_recovery(sk
, tp
))
2114 /* CWR is to be held something *above* high_seq
2115 * is ACKed for CWR bit to reach receiver. */
2116 if (tp
->snd_una
!= tp
->high_seq
) {
2117 tcp_complete_cwr(sk
);
2118 tcp_set_ca_state(sk
, TCP_CA_Open
);
2122 case TCP_CA_Disorder
:
2123 tcp_try_undo_dsack(sk
, tp
);
2124 if (!tp
->undo_marker
||
2125 /* For SACK case do not Open to allow to undo
2126 * catching for all duplicate ACKs. */
2127 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2128 tp
->undo_marker
= 0;
2129 tcp_set_ca_state(sk
, TCP_CA_Open
);
2133 case TCP_CA_Recovery
:
2135 tcp_reset_reno_sack(tp
);
2136 if (tcp_try_undo_recovery(sk
, tp
))
2138 tcp_complete_cwr(sk
);
2143 /* F. Process state. */
2144 switch (icsk
->icsk_ca_state
) {
2145 case TCP_CA_Recovery
:
2146 if (prior_snd_una
== tp
->snd_una
) {
2147 if (IsReno(tp
) && is_dupack
)
2148 tcp_add_reno_sack(sk
);
2150 int acked
= prior_packets
- tp
->packets_out
;
2152 tcp_remove_reno_sacks(sk
, tp
, acked
);
2153 is_dupack
= tcp_try_undo_partial(sk
, tp
, acked
);
2157 if (flag
&FLAG_DATA_ACKED
)
2158 icsk
->icsk_retransmits
= 0;
2159 if (!tcp_try_undo_loss(sk
, tp
)) {
2160 tcp_moderate_cwnd(tp
);
2161 tcp_xmit_retransmit_queue(sk
);
2164 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2166 /* Loss is undone; fall through to processing in Open state. */
2169 if (tp
->snd_una
!= prior_snd_una
)
2170 tcp_reset_reno_sack(tp
);
2172 tcp_add_reno_sack(sk
);
2175 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2176 tcp_try_undo_dsack(sk
, tp
);
2178 if (!tcp_time_to_recover(sk
, tp
)) {
2179 tcp_try_to_open(sk
, tp
, flag
);
2183 /* MTU probe failure: don't reduce cwnd */
2184 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2185 icsk
->icsk_mtup
.probe_size
&&
2186 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2187 tcp_mtup_probe_failed(sk
);
2188 /* Restores the reduction we did in tcp_mtup_probe() */
2190 tcp_simple_retransmit(sk
);
2194 /* Otherwise enter Recovery state */
2197 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2199 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2201 tp
->high_seq
= tp
->snd_nxt
;
2202 tp
->prior_ssthresh
= 0;
2203 tp
->undo_marker
= tp
->snd_una
;
2204 tp
->undo_retrans
= tp
->retrans_out
;
2206 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2207 if (!(flag
&FLAG_ECE
))
2208 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2209 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2210 TCP_ECN_queue_cwr(tp
);
2213 tp
->bytes_acked
= 0;
2214 tp
->snd_cwnd_cnt
= 0;
2215 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2218 if (is_dupack
|| tcp_head_timedout(sk
, tp
))
2219 tcp_update_scoreboard(sk
, tp
);
2221 tcp_xmit_retransmit_queue(sk
);
2224 /* Read draft-ietf-tcplw-high-performance before mucking
2225 * with this code. (Supersedes RFC1323)
2227 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2229 /* RTTM Rule: A TSecr value received in a segment is used to
2230 * update the averaged RTT measurement only if the segment
2231 * acknowledges some new data, i.e., only if it advances the
2232 * left edge of the send window.
2234 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2235 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2237 * Changed: reset backoff as soon as we see the first valid sample.
2238 * If we do not, we get strongly overestimated rto. With timestamps
2239 * samples are accepted even from very old segments: f.e., when rtt=1
2240 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2241 * answer arrives rto becomes 120 seconds! If at least one of segments
2242 * in window is lost... Voila. --ANK (010210)
2244 struct tcp_sock
*tp
= tcp_sk(sk
);
2245 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2246 tcp_rtt_estimator(sk
, seq_rtt
);
2248 inet_csk(sk
)->icsk_backoff
= 0;
2252 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2254 /* We don't have a timestamp. Can only use
2255 * packets that are not retransmitted to determine
2256 * rtt estimates. Also, we must not reset the
2257 * backoff for rto until we get a non-retransmitted
2258 * packet. This allows us to deal with a situation
2259 * where the network delay has increased suddenly.
2260 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2263 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2266 tcp_rtt_estimator(sk
, seq_rtt
);
2268 inet_csk(sk
)->icsk_backoff
= 0;
2272 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2275 const struct tcp_sock
*tp
= tcp_sk(sk
);
2276 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2277 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2278 tcp_ack_saw_tstamp(sk
, flag
);
2279 else if (seq_rtt
>= 0)
2280 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2283 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
2284 u32 in_flight
, int good
)
2286 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2287 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
2288 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2291 /* Restart timer after forward progress on connection.
2292 * RFC2988 recommends to restart timer to now+rto.
2295 static void tcp_ack_packets_out(struct sock
*sk
, struct tcp_sock
*tp
)
2297 if (!tp
->packets_out
) {
2298 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2300 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2304 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2305 __u32 now
, __s32
*seq_rtt
)
2307 struct tcp_sock
*tp
= tcp_sk(sk
);
2308 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2309 __u32 seq
= tp
->snd_una
;
2310 __u32 packets_acked
;
2313 /* If we get here, the whole TSO packet has not been
2316 BUG_ON(!after(scb
->end_seq
, seq
));
2318 packets_acked
= tcp_skb_pcount(skb
);
2319 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2321 packets_acked
-= tcp_skb_pcount(skb
);
2323 if (packets_acked
) {
2324 __u8 sacked
= scb
->sacked
;
2326 acked
|= FLAG_DATA_ACKED
;
2328 if (sacked
& TCPCB_RETRANS
) {
2329 if (sacked
& TCPCB_SACKED_RETRANS
)
2330 tp
->retrans_out
-= packets_acked
;
2331 acked
|= FLAG_RETRANS_DATA_ACKED
;
2333 } else if (*seq_rtt
< 0)
2334 *seq_rtt
= now
- scb
->when
;
2335 if (sacked
& TCPCB_SACKED_ACKED
)
2336 tp
->sacked_out
-= packets_acked
;
2337 if (sacked
& TCPCB_LOST
)
2338 tp
->lost_out
-= packets_acked
;
2339 if (sacked
& TCPCB_URG
) {
2341 !before(seq
, tp
->snd_up
))
2344 } else if (*seq_rtt
< 0)
2345 *seq_rtt
= now
- scb
->when
;
2347 if (tp
->fackets_out
) {
2348 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2349 tp
->fackets_out
-= dval
;
2351 tp
->packets_out
-= packets_acked
;
2353 BUG_ON(tcp_skb_pcount(skb
) == 0);
2354 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2360 static u32
tcp_usrtt(struct timeval
*tv
)
2364 do_gettimeofday(&now
);
2365 return (now
.tv_sec
- tv
->tv_sec
) * 1000000 + (now
.tv_usec
- tv
->tv_usec
);
2368 /* Remove acknowledged frames from the retransmission queue. */
2369 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2371 struct tcp_sock
*tp
= tcp_sk(sk
);
2372 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2373 struct sk_buff
*skb
;
2374 __u32 now
= tcp_time_stamp
;
2378 void (*rtt_sample
)(struct sock
*sk
, u32 usrtt
)
2379 = icsk
->icsk_ca_ops
->rtt_sample
;
2380 struct timeval tv
= { .tv_sec
= 0, .tv_usec
= 0 };
2382 while ((skb
= skb_peek(&sk
->sk_write_queue
)) &&
2383 skb
!= sk
->sk_send_head
) {
2384 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2385 __u8 sacked
= scb
->sacked
;
2387 /* If our packet is before the ack sequence we can
2388 * discard it as it's confirmed to have arrived at
2391 if (after(scb
->end_seq
, tp
->snd_una
)) {
2392 if (tcp_skb_pcount(skb
) > 1 &&
2393 after(tp
->snd_una
, scb
->seq
))
2394 acked
|= tcp_tso_acked(sk
, skb
,
2399 /* Initial outgoing SYN's get put onto the write_queue
2400 * just like anything else we transmit. It is not
2401 * true data, and if we misinform our callers that
2402 * this ACK acks real data, we will erroneously exit
2403 * connection startup slow start one packet too
2404 * quickly. This is severely frowned upon behavior.
2406 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2407 acked
|= FLAG_DATA_ACKED
;
2410 acked
|= FLAG_SYN_ACKED
;
2411 tp
->retrans_stamp
= 0;
2414 /* MTU probing checks */
2415 if (icsk
->icsk_mtup
.probe_size
) {
2416 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2417 tcp_mtup_probe_success(sk
, skb
);
2422 if (sacked
& TCPCB_RETRANS
) {
2423 if(sacked
& TCPCB_SACKED_RETRANS
)
2424 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2425 acked
|= FLAG_RETRANS_DATA_ACKED
;
2427 } else if (seq_rtt
< 0) {
2428 seq_rtt
= now
- scb
->when
;
2429 skb_get_timestamp(skb
, &tv
);
2431 if (sacked
& TCPCB_SACKED_ACKED
)
2432 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2433 if (sacked
& TCPCB_LOST
)
2434 tp
->lost_out
-= tcp_skb_pcount(skb
);
2435 if (sacked
& TCPCB_URG
) {
2437 !before(scb
->end_seq
, tp
->snd_up
))
2440 } else if (seq_rtt
< 0) {
2441 seq_rtt
= now
- scb
->when
;
2442 skb_get_timestamp(skb
, &tv
);
2444 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2445 tcp_packets_out_dec(tp
, skb
);
2446 __skb_unlink(skb
, &sk
->sk_write_queue
);
2447 sk_stream_free_skb(sk
, skb
);
2448 clear_all_retrans_hints(tp
);
2451 if (acked
&FLAG_ACKED
) {
2452 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2453 tcp_ack_packets_out(sk
, tp
);
2454 if (rtt_sample
&& !(acked
& FLAG_RETRANS_DATA_ACKED
))
2455 (*rtt_sample
)(sk
, tcp_usrtt(&tv
));
2457 if (icsk
->icsk_ca_ops
->pkts_acked
)
2458 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
);
2461 #if FASTRETRANS_DEBUG > 0
2462 BUG_TRAP((int)tp
->sacked_out
>= 0);
2463 BUG_TRAP((int)tp
->lost_out
>= 0);
2464 BUG_TRAP((int)tp
->retrans_out
>= 0);
2465 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2466 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2468 printk(KERN_DEBUG
"Leak l=%u %d\n",
2469 tp
->lost_out
, icsk
->icsk_ca_state
);
2472 if (tp
->sacked_out
) {
2473 printk(KERN_DEBUG
"Leak s=%u %d\n",
2474 tp
->sacked_out
, icsk
->icsk_ca_state
);
2477 if (tp
->retrans_out
) {
2478 printk(KERN_DEBUG
"Leak r=%u %d\n",
2479 tp
->retrans_out
, icsk
->icsk_ca_state
);
2480 tp
->retrans_out
= 0;
2484 *seq_rtt_p
= seq_rtt
;
2488 static void tcp_ack_probe(struct sock
*sk
)
2490 const struct tcp_sock
*tp
= tcp_sk(sk
);
2491 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2493 /* Was it a usable window open? */
2495 if (!after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
2496 tp
->snd_una
+ tp
->snd_wnd
)) {
2497 icsk
->icsk_backoff
= 0;
2498 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2499 /* Socket must be waked up by subsequent tcp_data_snd_check().
2500 * This function is not for random using!
2503 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2504 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2509 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2511 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2512 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2515 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2517 const struct tcp_sock
*tp
= tcp_sk(sk
);
2518 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2519 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2522 /* Check that window update is acceptable.
2523 * The function assumes that snd_una<=ack<=snd_next.
2525 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2526 const u32 ack_seq
, const u32 nwin
)
2528 return (after(ack
, tp
->snd_una
) ||
2529 after(ack_seq
, tp
->snd_wl1
) ||
2530 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2533 /* Update our send window.
2535 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2536 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2538 static int tcp_ack_update_window(struct sock
*sk
, struct tcp_sock
*tp
,
2539 struct sk_buff
*skb
, u32 ack
, u32 ack_seq
)
2542 u32 nwin
= ntohs(skb
->h
.th
->window
);
2544 if (likely(!skb
->h
.th
->syn
))
2545 nwin
<<= tp
->rx_opt
.snd_wscale
;
2547 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2548 flag
|= FLAG_WIN_UPDATE
;
2549 tcp_update_wl(tp
, ack
, ack_seq
);
2551 if (tp
->snd_wnd
!= nwin
) {
2554 /* Note, it is the only place, where
2555 * fast path is recovered for sending TCP.
2558 tcp_fast_path_check(sk
, tp
);
2560 if (nwin
> tp
->max_window
) {
2561 tp
->max_window
= nwin
;
2562 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2572 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2573 * continue in congestion avoidance.
2575 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2577 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2578 tp
->snd_cwnd_cnt
= 0;
2579 tcp_moderate_cwnd(tp
);
2582 /* F-RTO spurious RTO detection algorithm (RFC4138)
2584 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2585 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2586 * window (but not to or beyond highest sequence sent before RTO):
2587 * On First ACK, send two new segments out.
2588 * On Second ACK, RTO was likely spurious. Do spurious response (response
2589 * algorithm is not part of the F-RTO detection algorithm
2590 * given in RFC4138 but can be selected separately).
2591 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2592 * and TCP falls back to conventional RTO recovery.
2594 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2595 * original window even after we transmit two new data segments.
2598 * on first step, wait until first cumulative ACK arrives, then move to
2599 * the second step. In second step, the next ACK decides.
2601 * F-RTO is implemented (mainly) in four functions:
2602 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2603 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2604 * called when tcp_use_frto() showed green light
2605 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2606 * - tcp_enter_frto_loss() is called if there is not enough evidence
2607 * to prove that the RTO is indeed spurious. It transfers the control
2608 * from F-RTO to the conventional RTO recovery
2610 static int tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
, int flag
)
2612 struct tcp_sock
*tp
= tcp_sk(sk
);
2614 tcp_sync_left_out(tp
);
2616 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2617 if (flag
&FLAG_DATA_ACKED
)
2618 inet_csk(sk
)->icsk_retransmits
= 0;
2620 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2621 tcp_enter_frto_loss(sk
, tp
->frto_counter
+ 1, flag
);
2625 if (!IsSackFrto() || IsReno(tp
)) {
2626 /* RFC4138 shortcoming in step 2; should also have case c):
2627 * ACK isn't duplicate nor advances window, e.g., opposite dir
2630 if ((tp
->snd_una
== prior_snd_una
) && (flag
&FLAG_NOT_DUP
) &&
2631 !(flag
&FLAG_FORWARD_PROGRESS
))
2634 if (!(flag
&FLAG_DATA_ACKED
)) {
2635 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2640 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2641 /* Prevent sending of new data. */
2642 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2643 tcp_packets_in_flight(tp
));
2647 if ((tp
->frto_counter
== 2) &&
2648 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2649 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2650 /* RFC4138 shortcoming (see comment above) */
2651 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2654 tcp_enter_frto_loss(sk
, 3, flag
);
2659 if (tp
->frto_counter
== 1) {
2660 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2661 tp
->frto_counter
= 2;
2663 } else /* frto_counter == 2 */ {
2664 tcp_conservative_spur_to_response(tp
);
2665 tp
->frto_counter
= 0;
2670 /* This routine deals with incoming acks, but not outgoing ones. */
2671 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2673 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2674 struct tcp_sock
*tp
= tcp_sk(sk
);
2675 u32 prior_snd_una
= tp
->snd_una
;
2676 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2677 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2678 u32 prior_in_flight
;
2683 /* If the ack is newer than sent or older than previous acks
2684 * then we can probably ignore it.
2686 if (after(ack
, tp
->snd_nxt
))
2687 goto uninteresting_ack
;
2689 if (before(ack
, prior_snd_una
))
2692 if (sysctl_tcp_abc
) {
2693 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2694 tp
->bytes_acked
+= ack
- prior_snd_una
;
2695 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2696 /* we assume just one segment left network */
2697 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2700 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2701 /* Window is constant, pure forward advance.
2702 * No more checks are required.
2703 * Note, we use the fact that SND.UNA>=SND.WL2.
2705 tcp_update_wl(tp
, ack
, ack_seq
);
2707 flag
|= FLAG_WIN_UPDATE
;
2709 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2711 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2713 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2716 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2718 flag
|= tcp_ack_update_window(sk
, tp
, skb
, ack
, ack_seq
);
2720 if (TCP_SKB_CB(skb
)->sacked
)
2721 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2723 if (TCP_ECN_rcv_ecn_echo(tp
, skb
->h
.th
))
2726 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2729 /* We passed data and got it acked, remove any soft error
2730 * log. Something worked...
2732 sk
->sk_err_soft
= 0;
2733 tp
->rcv_tstamp
= tcp_time_stamp
;
2734 prior_packets
= tp
->packets_out
;
2738 prior_in_flight
= tcp_packets_in_flight(tp
);
2740 /* See if we can take anything off of the retransmit queue. */
2741 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2743 if (tp
->frto_counter
)
2744 frto_cwnd
= tcp_process_frto(sk
, prior_snd_una
, flag
);
2746 if (tcp_ack_is_dubious(sk
, flag
)) {
2747 /* Advance CWND, if state allows this. */
2748 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
2749 tcp_may_raise_cwnd(sk
, flag
))
2750 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2751 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2753 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
2754 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2757 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2758 dst_confirm(sk
->sk_dst_cache
);
2763 icsk
->icsk_probes_out
= 0;
2765 /* If this ack opens up a zero window, clear backoff. It was
2766 * being used to time the probes, and is probably far higher than
2767 * it needs to be for normal retransmission.
2769 if (sk
->sk_send_head
)
2774 if (TCP_SKB_CB(skb
)->sacked
)
2775 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2778 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2783 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2784 * But, this can also be called on packets in the established flow when
2785 * the fast version below fails.
2787 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2790 struct tcphdr
*th
= skb
->h
.th
;
2791 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2793 ptr
= (unsigned char *)(th
+ 1);
2794 opt_rx
->saw_tstamp
= 0;
2803 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2808 if (opsize
< 2) /* "silly options" */
2810 if (opsize
> length
)
2811 return; /* don't parse partial options */
2814 if(opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2815 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
2817 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2818 in_mss
= opt_rx
->user_mss
;
2819 opt_rx
->mss_clamp
= in_mss
;
2824 if(opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2825 if (sysctl_tcp_window_scaling
) {
2826 __u8 snd_wscale
= *(__u8
*) ptr
;
2827 opt_rx
->wscale_ok
= 1;
2828 if (snd_wscale
> 14) {
2830 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2831 "scaling value %d >14 received.\n",
2835 opt_rx
->snd_wscale
= snd_wscale
;
2838 case TCPOPT_TIMESTAMP
:
2839 if(opsize
==TCPOLEN_TIMESTAMP
) {
2840 if ((estab
&& opt_rx
->tstamp_ok
) ||
2841 (!estab
&& sysctl_tcp_timestamps
)) {
2842 opt_rx
->saw_tstamp
= 1;
2843 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
2844 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
2848 case TCPOPT_SACK_PERM
:
2849 if(opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2850 if (sysctl_tcp_sack
) {
2851 opt_rx
->sack_ok
= 1;
2852 tcp_sack_reset(opt_rx
);
2858 if((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2859 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2861 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2863 #ifdef CONFIG_TCP_MD5SIG
2866 * The MD5 Hash has already been
2867 * checked (see tcp_v{4,6}_do_rcv()).
2878 /* Fast parse options. This hopes to only see timestamps.
2879 * If it is wrong it falls back on tcp_parse_options().
2881 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2882 struct tcp_sock
*tp
)
2884 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2885 tp
->rx_opt
.saw_tstamp
= 0;
2887 } else if (tp
->rx_opt
.tstamp_ok
&&
2888 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2889 __be32
*ptr
= (__be32
*)(th
+ 1);
2890 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2891 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2892 tp
->rx_opt
.saw_tstamp
= 1;
2894 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2896 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2900 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2904 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2906 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2907 tp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
2910 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2912 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2913 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2914 * extra check below makes sure this can only happen
2915 * for pure ACK frames. -DaveM
2917 * Not only, also it occurs for expired timestamps.
2920 if((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2921 xtime
.tv_sec
>= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2922 tcp_store_ts_recent(tp
);
2926 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2928 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2929 * it can pass through stack. So, the following predicate verifies that
2930 * this segment is not used for anything but congestion avoidance or
2931 * fast retransmit. Moreover, we even are able to eliminate most of such
2932 * second order effects, if we apply some small "replay" window (~RTO)
2933 * to timestamp space.
2935 * All these measures still do not guarantee that we reject wrapped ACKs
2936 * on networks with high bandwidth, when sequence space is recycled fastly,
2937 * but it guarantees that such events will be very rare and do not affect
2938 * connection seriously. This doesn't look nice, but alas, PAWS is really
2941 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2942 * states that events when retransmit arrives after original data are rare.
2943 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2944 * the biggest problem on large power networks even with minor reordering.
2945 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2946 * up to bandwidth of 18Gigabit/sec. 8) ]
2949 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
2951 struct tcp_sock
*tp
= tcp_sk(sk
);
2952 struct tcphdr
*th
= skb
->h
.th
;
2953 u32 seq
= TCP_SKB_CB(skb
)->seq
;
2954 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2956 return (/* 1. Pure ACK with correct sequence number. */
2957 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
2959 /* 2. ... and duplicate ACK. */
2960 ack
== tp
->snd_una
&&
2962 /* 3. ... and does not update window. */
2963 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
2965 /* 4. ... and sits in replay window. */
2966 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
2969 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
2971 const struct tcp_sock
*tp
= tcp_sk(sk
);
2972 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
2973 xtime
.tv_sec
< tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
2974 !tcp_disordered_ack(sk
, skb
));
2977 /* Check segment sequence number for validity.
2979 * Segment controls are considered valid, if the segment
2980 * fits to the window after truncation to the window. Acceptability
2981 * of data (and SYN, FIN, of course) is checked separately.
2982 * See tcp_data_queue(), for example.
2984 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2985 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2986 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2987 * (borrowed from freebsd)
2990 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2992 return !before(end_seq
, tp
->rcv_wup
) &&
2993 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
2996 /* When we get a reset we do this. */
2997 static void tcp_reset(struct sock
*sk
)
2999 /* We want the right error as BSD sees it (and indeed as we do). */
3000 switch (sk
->sk_state
) {
3002 sk
->sk_err
= ECONNREFUSED
;
3004 case TCP_CLOSE_WAIT
:
3010 sk
->sk_err
= ECONNRESET
;
3013 if (!sock_flag(sk
, SOCK_DEAD
))
3014 sk
->sk_error_report(sk
);
3020 * Process the FIN bit. This now behaves as it is supposed to work
3021 * and the FIN takes effect when it is validly part of sequence
3022 * space. Not before when we get holes.
3024 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3025 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3028 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3029 * close and we go into CLOSING (and later onto TIME-WAIT)
3031 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3033 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3035 struct tcp_sock
*tp
= tcp_sk(sk
);
3037 inet_csk_schedule_ack(sk
);
3039 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3040 sock_set_flag(sk
, SOCK_DONE
);
3042 switch (sk
->sk_state
) {
3044 case TCP_ESTABLISHED
:
3045 /* Move to CLOSE_WAIT */
3046 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3047 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3050 case TCP_CLOSE_WAIT
:
3052 /* Received a retransmission of the FIN, do
3057 /* RFC793: Remain in the LAST-ACK state. */
3061 /* This case occurs when a simultaneous close
3062 * happens, we must ack the received FIN and
3063 * enter the CLOSING state.
3066 tcp_set_state(sk
, TCP_CLOSING
);
3069 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3071 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3074 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3075 * cases we should never reach this piece of code.
3077 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3078 __FUNCTION__
, sk
->sk_state
);
3082 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3083 * Probably, we should reset in this case. For now drop them.
3085 __skb_queue_purge(&tp
->out_of_order_queue
);
3086 if (tp
->rx_opt
.sack_ok
)
3087 tcp_sack_reset(&tp
->rx_opt
);
3088 sk_stream_mem_reclaim(sk
);
3090 if (!sock_flag(sk
, SOCK_DEAD
)) {
3091 sk
->sk_state_change(sk
);
3093 /* Do not send POLL_HUP for half duplex close. */
3094 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3095 sk
->sk_state
== TCP_CLOSE
)
3096 sk_wake_async(sk
, 1, POLL_HUP
);
3098 sk_wake_async(sk
, 1, POLL_IN
);
3102 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3104 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3105 if (before(seq
, sp
->start_seq
))
3106 sp
->start_seq
= seq
;
3107 if (after(end_seq
, sp
->end_seq
))
3108 sp
->end_seq
= end_seq
;
3114 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3116 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3117 if (before(seq
, tp
->rcv_nxt
))
3118 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3120 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3122 tp
->rx_opt
.dsack
= 1;
3123 tp
->duplicate_sack
[0].start_seq
= seq
;
3124 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3125 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3129 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3131 if (!tp
->rx_opt
.dsack
)
3132 tcp_dsack_set(tp
, seq
, end_seq
);
3134 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3137 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3139 struct tcp_sock
*tp
= tcp_sk(sk
);
3141 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3142 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3143 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3144 tcp_enter_quickack_mode(sk
);
3146 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3147 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3149 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3150 end_seq
= tp
->rcv_nxt
;
3151 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3158 /* These routines update the SACK block as out-of-order packets arrive or
3159 * in-order packets close up the sequence space.
3161 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3164 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3165 struct tcp_sack_block
*swalk
= sp
+1;
3167 /* See if the recent change to the first SACK eats into
3168 * or hits the sequence space of other SACK blocks, if so coalesce.
3170 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3171 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3174 /* Zap SWALK, by moving every further SACK up by one slot.
3175 * Decrease num_sacks.
3177 tp
->rx_opt
.num_sacks
--;
3178 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3179 for(i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3183 this_sack
++, swalk
++;
3187 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3191 tmp
= sack1
->start_seq
;
3192 sack1
->start_seq
= sack2
->start_seq
;
3193 sack2
->start_seq
= tmp
;
3195 tmp
= sack1
->end_seq
;
3196 sack1
->end_seq
= sack2
->end_seq
;
3197 sack2
->end_seq
= tmp
;
3200 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3202 struct tcp_sock
*tp
= tcp_sk(sk
);
3203 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3204 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3210 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3211 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3212 /* Rotate this_sack to the first one. */
3213 for (; this_sack
>0; this_sack
--, sp
--)
3214 tcp_sack_swap(sp
, sp
-1);
3216 tcp_sack_maybe_coalesce(tp
);
3221 /* Could not find an adjacent existing SACK, build a new one,
3222 * put it at the front, and shift everyone else down. We
3223 * always know there is at least one SACK present already here.
3225 * If the sack array is full, forget about the last one.
3227 if (this_sack
>= 4) {
3229 tp
->rx_opt
.num_sacks
--;
3232 for(; this_sack
> 0; this_sack
--, sp
--)
3236 /* Build the new head SACK, and we're done. */
3237 sp
->start_seq
= seq
;
3238 sp
->end_seq
= end_seq
;
3239 tp
->rx_opt
.num_sacks
++;
3240 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3243 /* RCV.NXT advances, some SACKs should be eaten. */
3245 static void tcp_sack_remove(struct tcp_sock
*tp
)
3247 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3248 int num_sacks
= tp
->rx_opt
.num_sacks
;
3251 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3252 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3253 tp
->rx_opt
.num_sacks
= 0;
3254 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3258 for(this_sack
= 0; this_sack
< num_sacks
; ) {
3259 /* Check if the start of the sack is covered by RCV.NXT. */
3260 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3263 /* RCV.NXT must cover all the block! */
3264 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3266 /* Zap this SACK, by moving forward any other SACKS. */
3267 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3268 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3275 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3276 tp
->rx_opt
.num_sacks
= num_sacks
;
3277 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3281 /* This one checks to see if we can put data from the
3282 * out_of_order queue into the receive_queue.
3284 static void tcp_ofo_queue(struct sock
*sk
)
3286 struct tcp_sock
*tp
= tcp_sk(sk
);
3287 __u32 dsack_high
= tp
->rcv_nxt
;
3288 struct sk_buff
*skb
;
3290 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3291 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3294 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3295 __u32 dsack
= dsack_high
;
3296 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3297 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3298 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3301 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3302 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3303 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3307 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3308 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3309 TCP_SKB_CB(skb
)->end_seq
);
3311 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3312 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3313 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3315 tcp_fin(skb
, sk
, skb
->h
.th
);
3319 static int tcp_prune_queue(struct sock
*sk
);
3321 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3323 struct tcphdr
*th
= skb
->h
.th
;
3324 struct tcp_sock
*tp
= tcp_sk(sk
);
3327 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3330 __skb_pull(skb
, th
->doff
*4);
3332 TCP_ECN_accept_cwr(tp
, skb
);
3334 if (tp
->rx_opt
.dsack
) {
3335 tp
->rx_opt
.dsack
= 0;
3336 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3337 4 - tp
->rx_opt
.tstamp_ok
);
3340 /* Queue data for delivery to the user.
3341 * Packets in sequence go to the receive queue.
3342 * Out of sequence packets to the out_of_order_queue.
3344 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3345 if (tcp_receive_window(tp
) == 0)
3348 /* Ok. In sequence. In window. */
3349 if (tp
->ucopy
.task
== current
&&
3350 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3351 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3352 int chunk
= min_t(unsigned int, skb
->len
,
3355 __set_current_state(TASK_RUNNING
);
3358 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3359 tp
->ucopy
.len
-= chunk
;
3360 tp
->copied_seq
+= chunk
;
3361 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3362 tcp_rcv_space_adjust(sk
);
3370 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3371 !sk_stream_rmem_schedule(sk
, skb
))) {
3372 if (tcp_prune_queue(sk
) < 0 ||
3373 !sk_stream_rmem_schedule(sk
, skb
))
3376 sk_stream_set_owner_r(skb
, sk
);
3377 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3379 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3381 tcp_event_data_recv(sk
, tp
, skb
);
3383 tcp_fin(skb
, sk
, th
);
3385 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3388 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3389 * gap in queue is filled.
3391 if (skb_queue_empty(&tp
->out_of_order_queue
))
3392 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3395 if (tp
->rx_opt
.num_sacks
)
3396 tcp_sack_remove(tp
);
3398 tcp_fast_path_check(sk
, tp
);
3402 else if (!sock_flag(sk
, SOCK_DEAD
))
3403 sk
->sk_data_ready(sk
, 0);
3407 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3408 /* A retransmit, 2nd most common case. Force an immediate ack. */
3409 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3410 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3413 tcp_enter_quickack_mode(sk
);
3414 inet_csk_schedule_ack(sk
);
3420 /* Out of window. F.e. zero window probe. */
3421 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3424 tcp_enter_quickack_mode(sk
);
3426 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3427 /* Partial packet, seq < rcv_next < end_seq */
3428 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3429 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3430 TCP_SKB_CB(skb
)->end_seq
);
3432 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3434 /* If window is closed, drop tail of packet. But after
3435 * remembering D-SACK for its head made in previous line.
3437 if (!tcp_receive_window(tp
))
3442 TCP_ECN_check_ce(tp
, skb
);
3444 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3445 !sk_stream_rmem_schedule(sk
, skb
)) {
3446 if (tcp_prune_queue(sk
) < 0 ||
3447 !sk_stream_rmem_schedule(sk
, skb
))
3451 /* Disable header prediction. */
3453 inet_csk_schedule_ack(sk
);
3455 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3456 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3458 sk_stream_set_owner_r(skb
, sk
);
3460 if (!skb_peek(&tp
->out_of_order_queue
)) {
3461 /* Initial out of order segment, build 1 SACK. */
3462 if (tp
->rx_opt
.sack_ok
) {
3463 tp
->rx_opt
.num_sacks
= 1;
3464 tp
->rx_opt
.dsack
= 0;
3465 tp
->rx_opt
.eff_sacks
= 1;
3466 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3467 tp
->selective_acks
[0].end_seq
=
3468 TCP_SKB_CB(skb
)->end_seq
;
3470 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3472 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3473 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3474 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3476 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3477 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3479 if (!tp
->rx_opt
.num_sacks
||
3480 tp
->selective_acks
[0].end_seq
!= seq
)
3483 /* Common case: data arrive in order after hole. */
3484 tp
->selective_acks
[0].end_seq
= end_seq
;
3488 /* Find place to insert this segment. */
3490 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3492 } while ((skb1
= skb1
->prev
) !=
3493 (struct sk_buff
*)&tp
->out_of_order_queue
);
3495 /* Do skb overlap to previous one? */
3496 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3497 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3498 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3499 /* All the bits are present. Drop. */
3501 tcp_dsack_set(tp
, seq
, end_seq
);
3504 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3505 /* Partial overlap. */
3506 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3511 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3513 /* And clean segments covered by new one as whole. */
3514 while ((skb1
= skb
->next
) !=
3515 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3516 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3517 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3518 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3521 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3522 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3527 if (tp
->rx_opt
.sack_ok
)
3528 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3532 /* Collapse contiguous sequence of skbs head..tail with
3533 * sequence numbers start..end.
3534 * Segments with FIN/SYN are not collapsed (only because this
3538 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3539 struct sk_buff
*head
, struct sk_buff
*tail
,
3542 struct sk_buff
*skb
;
3544 /* First, check that queue is collapsible and find
3545 * the point where collapsing can be useful. */
3546 for (skb
= head
; skb
!= tail
; ) {
3547 /* No new bits? It is possible on ofo queue. */
3548 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3549 struct sk_buff
*next
= skb
->next
;
3550 __skb_unlink(skb
, list
);
3552 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3557 /* The first skb to collapse is:
3559 * - bloated or contains data before "start" or
3560 * overlaps to the next one.
3562 if (!skb
->h
.th
->syn
&& !skb
->h
.th
->fin
&&
3563 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3564 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3565 (skb
->next
!= tail
&&
3566 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3569 /* Decided to skip this, advance start seq. */
3570 start
= TCP_SKB_CB(skb
)->end_seq
;
3573 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3576 while (before(start
, end
)) {
3577 struct sk_buff
*nskb
;
3578 int header
= skb_headroom(skb
);
3579 int copy
= SKB_MAX_ORDER(header
, 0);
3581 /* Too big header? This can happen with IPv6. */
3584 if (end
-start
< copy
)
3586 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3589 skb_reserve(nskb
, header
);
3590 memcpy(nskb
->head
, skb
->head
, header
);
3591 nskb
->nh
.raw
= nskb
->head
+ (skb
->nh
.raw
-skb
->head
);
3592 nskb
->h
.raw
= nskb
->head
+ (skb
->h
.raw
-skb
->head
);
3593 nskb
->mac
.raw
= nskb
->head
+ (skb
->mac
.raw
-skb
->head
);
3594 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3595 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3596 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3597 sk_stream_set_owner_r(nskb
, sk
);
3599 /* Copy data, releasing collapsed skbs. */
3601 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3602 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3606 size
= min(copy
, size
);
3607 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3609 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3613 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3614 struct sk_buff
*next
= skb
->next
;
3615 __skb_unlink(skb
, list
);
3617 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3619 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3626 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3627 * and tcp_collapse() them until all the queue is collapsed.
3629 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3631 struct tcp_sock
*tp
= tcp_sk(sk
);
3632 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3633 struct sk_buff
*head
;
3639 start
= TCP_SKB_CB(skb
)->seq
;
3640 end
= TCP_SKB_CB(skb
)->end_seq
;
3646 /* Segment is terminated when we see gap or when
3647 * we are at the end of all the queue. */
3648 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3649 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3650 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3651 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3652 head
, skb
, start
, end
);
3654 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3656 /* Start new segment */
3657 start
= TCP_SKB_CB(skb
)->seq
;
3658 end
= TCP_SKB_CB(skb
)->end_seq
;
3660 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3661 start
= TCP_SKB_CB(skb
)->seq
;
3662 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3663 end
= TCP_SKB_CB(skb
)->end_seq
;
3668 /* Reduce allocated memory if we can, trying to get
3669 * the socket within its memory limits again.
3671 * Return less than zero if we should start dropping frames
3672 * until the socket owning process reads some of the data
3673 * to stabilize the situation.
3675 static int tcp_prune_queue(struct sock
*sk
)
3677 struct tcp_sock
*tp
= tcp_sk(sk
);
3679 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3681 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3683 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3684 tcp_clamp_window(sk
, tp
);
3685 else if (tcp_memory_pressure
)
3686 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3688 tcp_collapse_ofo_queue(sk
);
3689 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3690 sk
->sk_receive_queue
.next
,
3691 (struct sk_buff
*)&sk
->sk_receive_queue
,
3692 tp
->copied_seq
, tp
->rcv_nxt
);
3693 sk_stream_mem_reclaim(sk
);
3695 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3698 /* Collapsing did not help, destructive actions follow.
3699 * This must not ever occur. */
3701 /* First, purge the out_of_order queue. */
3702 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3703 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3704 __skb_queue_purge(&tp
->out_of_order_queue
);
3706 /* Reset SACK state. A conforming SACK implementation will
3707 * do the same at a timeout based retransmit. When a connection
3708 * is in a sad state like this, we care only about integrity
3709 * of the connection not performance.
3711 if (tp
->rx_opt
.sack_ok
)
3712 tcp_sack_reset(&tp
->rx_opt
);
3713 sk_stream_mem_reclaim(sk
);
3716 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3719 /* If we are really being abused, tell the caller to silently
3720 * drop receive data on the floor. It will get retransmitted
3721 * and hopefully then we'll have sufficient space.
3723 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3725 /* Massive buffer overcommit. */
3731 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3732 * As additional protections, we do not touch cwnd in retransmission phases,
3733 * and if application hit its sndbuf limit recently.
3735 void tcp_cwnd_application_limited(struct sock
*sk
)
3737 struct tcp_sock
*tp
= tcp_sk(sk
);
3739 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3740 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3741 /* Limited by application or receiver window. */
3742 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
3743 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
3744 if (win_used
< tp
->snd_cwnd
) {
3745 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3746 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3748 tp
->snd_cwnd_used
= 0;
3750 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3753 static int tcp_should_expand_sndbuf(struct sock
*sk
, struct tcp_sock
*tp
)
3755 /* If the user specified a specific send buffer setting, do
3758 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3761 /* If we are under global TCP memory pressure, do not expand. */
3762 if (tcp_memory_pressure
)
3765 /* If we are under soft global TCP memory pressure, do not expand. */
3766 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3769 /* If we filled the congestion window, do not expand. */
3770 if (tp
->packets_out
>= tp
->snd_cwnd
)
3776 /* When incoming ACK allowed to free some skb from write_queue,
3777 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3778 * on the exit from tcp input handler.
3780 * PROBLEM: sndbuf expansion does not work well with largesend.
3782 static void tcp_new_space(struct sock
*sk
)
3784 struct tcp_sock
*tp
= tcp_sk(sk
);
3786 if (tcp_should_expand_sndbuf(sk
, tp
)) {
3787 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3788 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3789 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3790 tp
->reordering
+ 1);
3791 sndmem
*= 2*demanded
;
3792 if (sndmem
> sk
->sk_sndbuf
)
3793 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3794 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3797 sk
->sk_write_space(sk
);
3800 static void tcp_check_space(struct sock
*sk
)
3802 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3803 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3804 if (sk
->sk_socket
&&
3805 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3810 static inline void tcp_data_snd_check(struct sock
*sk
, struct tcp_sock
*tp
)
3812 tcp_push_pending_frames(sk
, tp
);
3813 tcp_check_space(sk
);
3817 * Check if sending an ack is needed.
3819 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3821 struct tcp_sock
*tp
= tcp_sk(sk
);
3823 /* More than one full frame received... */
3824 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3825 /* ... and right edge of window advances far enough.
3826 * (tcp_recvmsg() will send ACK otherwise). Or...
3828 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3829 /* We ACK each frame or... */
3830 tcp_in_quickack_mode(sk
) ||
3831 /* We have out of order data. */
3833 skb_peek(&tp
->out_of_order_queue
))) {
3834 /* Then ack it now */
3837 /* Else, send delayed ack. */
3838 tcp_send_delayed_ack(sk
);
3842 static inline void tcp_ack_snd_check(struct sock
*sk
)
3844 if (!inet_csk_ack_scheduled(sk
)) {
3845 /* We sent a data segment already. */
3848 __tcp_ack_snd_check(sk
, 1);
3852 * This routine is only called when we have urgent data
3853 * signaled. Its the 'slow' part of tcp_urg. It could be
3854 * moved inline now as tcp_urg is only called from one
3855 * place. We handle URGent data wrong. We have to - as
3856 * BSD still doesn't use the correction from RFC961.
3857 * For 1003.1g we should support a new option TCP_STDURG to permit
3858 * either form (or just set the sysctl tcp_stdurg).
3861 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3863 struct tcp_sock
*tp
= tcp_sk(sk
);
3864 u32 ptr
= ntohs(th
->urg_ptr
);
3866 if (ptr
&& !sysctl_tcp_stdurg
)
3868 ptr
+= ntohl(th
->seq
);
3870 /* Ignore urgent data that we've already seen and read. */
3871 if (after(tp
->copied_seq
, ptr
))
3874 /* Do not replay urg ptr.
3876 * NOTE: interesting situation not covered by specs.
3877 * Misbehaving sender may send urg ptr, pointing to segment,
3878 * which we already have in ofo queue. We are not able to fetch
3879 * such data and will stay in TCP_URG_NOTYET until will be eaten
3880 * by recvmsg(). Seems, we are not obliged to handle such wicked
3881 * situations. But it is worth to think about possibility of some
3882 * DoSes using some hypothetical application level deadlock.
3884 if (before(ptr
, tp
->rcv_nxt
))
3887 /* Do we already have a newer (or duplicate) urgent pointer? */
3888 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3891 /* Tell the world about our new urgent pointer. */
3894 /* We may be adding urgent data when the last byte read was
3895 * urgent. To do this requires some care. We cannot just ignore
3896 * tp->copied_seq since we would read the last urgent byte again
3897 * as data, nor can we alter copied_seq until this data arrives
3898 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3900 * NOTE. Double Dutch. Rendering to plain English: author of comment
3901 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3902 * and expect that both A and B disappear from stream. This is _wrong_.
3903 * Though this happens in BSD with high probability, this is occasional.
3904 * Any application relying on this is buggy. Note also, that fix "works"
3905 * only in this artificial test. Insert some normal data between A and B and we will
3906 * decline of BSD again. Verdict: it is better to remove to trap
3909 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3910 !sock_flag(sk
, SOCK_URGINLINE
) &&
3911 tp
->copied_seq
!= tp
->rcv_nxt
) {
3912 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3914 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3915 __skb_unlink(skb
, &sk
->sk_receive_queue
);
3920 tp
->urg_data
= TCP_URG_NOTYET
;
3923 /* Disable header prediction. */
3927 /* This is the 'fast' part of urgent handling. */
3928 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
3930 struct tcp_sock
*tp
= tcp_sk(sk
);
3932 /* Check if we get a new urgent pointer - normally not. */
3934 tcp_check_urg(sk
,th
);
3936 /* Do we wait for any urgent data? - normally not... */
3937 if (tp
->urg_data
== TCP_URG_NOTYET
) {
3938 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
3941 /* Is the urgent pointer pointing into this packet? */
3942 if (ptr
< skb
->len
) {
3944 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
3946 tp
->urg_data
= TCP_URG_VALID
| tmp
;
3947 if (!sock_flag(sk
, SOCK_DEAD
))
3948 sk
->sk_data_ready(sk
, 0);
3953 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3955 struct tcp_sock
*tp
= tcp_sk(sk
);
3956 int chunk
= skb
->len
- hlen
;
3960 if (skb
->ip_summed
==CHECKSUM_UNNECESSARY
)
3961 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
3963 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
3967 tp
->ucopy
.len
-= chunk
;
3968 tp
->copied_seq
+= chunk
;
3969 tcp_rcv_space_adjust(sk
);
3976 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3980 if (sock_owned_by_user(sk
)) {
3982 result
= __tcp_checksum_complete(skb
);
3985 result
= __tcp_checksum_complete(skb
);
3990 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3992 return skb
->ip_summed
!= CHECKSUM_UNNECESSARY
&&
3993 __tcp_checksum_complete_user(sk
, skb
);
3996 #ifdef CONFIG_NET_DMA
3997 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3999 struct tcp_sock
*tp
= tcp_sk(sk
);
4000 int chunk
= skb
->len
- hlen
;
4002 int copied_early
= 0;
4004 if (tp
->ucopy
.wakeup
)
4007 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4008 tp
->ucopy
.dma_chan
= get_softnet_dma();
4010 if (tp
->ucopy
.dma_chan
&& skb
->ip_summed
== CHECKSUM_UNNECESSARY
) {
4012 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4013 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4018 tp
->ucopy
.dma_cookie
= dma_cookie
;
4021 tp
->ucopy
.len
-= chunk
;
4022 tp
->copied_seq
+= chunk
;
4023 tcp_rcv_space_adjust(sk
);
4025 if ((tp
->ucopy
.len
== 0) ||
4026 (tcp_flag_word(skb
->h
.th
) & TCP_FLAG_PSH
) ||
4027 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4028 tp
->ucopy
.wakeup
= 1;
4029 sk
->sk_data_ready(sk
, 0);
4031 } else if (chunk
> 0) {
4032 tp
->ucopy
.wakeup
= 1;
4033 sk
->sk_data_ready(sk
, 0);
4036 return copied_early
;
4038 #endif /* CONFIG_NET_DMA */
4041 * TCP receive function for the ESTABLISHED state.
4043 * It is split into a fast path and a slow path. The fast path is
4045 * - A zero window was announced from us - zero window probing
4046 * is only handled properly in the slow path.
4047 * - Out of order segments arrived.
4048 * - Urgent data is expected.
4049 * - There is no buffer space left
4050 * - Unexpected TCP flags/window values/header lengths are received
4051 * (detected by checking the TCP header against pred_flags)
4052 * - Data is sent in both directions. Fast path only supports pure senders
4053 * or pure receivers (this means either the sequence number or the ack
4054 * value must stay constant)
4055 * - Unexpected TCP option.
4057 * When these conditions are not satisfied it drops into a standard
4058 * receive procedure patterned after RFC793 to handle all cases.
4059 * The first three cases are guaranteed by proper pred_flags setting,
4060 * the rest is checked inline. Fast processing is turned on in
4061 * tcp_data_queue when everything is OK.
4063 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4064 struct tcphdr
*th
, unsigned len
)
4066 struct tcp_sock
*tp
= tcp_sk(sk
);
4069 * Header prediction.
4070 * The code loosely follows the one in the famous
4071 * "30 instruction TCP receive" Van Jacobson mail.
4073 * Van's trick is to deposit buffers into socket queue
4074 * on a device interrupt, to call tcp_recv function
4075 * on the receive process context and checksum and copy
4076 * the buffer to user space. smart...
4078 * Our current scheme is not silly either but we take the
4079 * extra cost of the net_bh soft interrupt processing...
4080 * We do checksum and copy also but from device to kernel.
4083 tp
->rx_opt
.saw_tstamp
= 0;
4085 /* pred_flags is 0xS?10 << 16 + snd_wnd
4086 * if header_prediction is to be made
4087 * 'S' will always be tp->tcp_header_len >> 2
4088 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4089 * turn it off (when there are holes in the receive
4090 * space for instance)
4091 * PSH flag is ignored.
4094 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4095 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4096 int tcp_header_len
= tp
->tcp_header_len
;
4098 /* Timestamp header prediction: tcp_header_len
4099 * is automatically equal to th->doff*4 due to pred_flags
4103 /* Check timestamp */
4104 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4105 __be32
*ptr
= (__be32
*)(th
+ 1);
4107 /* No? Slow path! */
4108 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4109 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4112 tp
->rx_opt
.saw_tstamp
= 1;
4114 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4116 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4118 /* If PAWS failed, check it more carefully in slow path */
4119 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4122 /* DO NOT update ts_recent here, if checksum fails
4123 * and timestamp was corrupted part, it will result
4124 * in a hung connection since we will drop all
4125 * future packets due to the PAWS test.
4129 if (len
<= tcp_header_len
) {
4130 /* Bulk data transfer: sender */
4131 if (len
== tcp_header_len
) {
4132 /* Predicted packet is in window by definition.
4133 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4134 * Hence, check seq<=rcv_wup reduces to:
4136 if (tcp_header_len
==
4137 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4138 tp
->rcv_nxt
== tp
->rcv_wup
)
4139 tcp_store_ts_recent(tp
);
4141 /* We know that such packets are checksummed
4144 tcp_ack(sk
, skb
, 0);
4146 tcp_data_snd_check(sk
, tp
);
4148 } else { /* Header too small */
4149 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4154 int copied_early
= 0;
4156 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4157 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4158 #ifdef CONFIG_NET_DMA
4159 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4164 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4165 __set_current_state(TASK_RUNNING
);
4167 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4171 /* Predicted packet is in window by definition.
4172 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4173 * Hence, check seq<=rcv_wup reduces to:
4175 if (tcp_header_len
==
4176 (sizeof(struct tcphdr
) +
4177 TCPOLEN_TSTAMP_ALIGNED
) &&
4178 tp
->rcv_nxt
== tp
->rcv_wup
)
4179 tcp_store_ts_recent(tp
);
4181 tcp_rcv_rtt_measure_ts(sk
, skb
);
4183 __skb_pull(skb
, tcp_header_len
);
4184 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4185 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4188 tcp_cleanup_rbuf(sk
, skb
->len
);
4191 if (tcp_checksum_complete_user(sk
, skb
))
4194 /* Predicted packet is in window by definition.
4195 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4196 * Hence, check seq<=rcv_wup reduces to:
4198 if (tcp_header_len
==
4199 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4200 tp
->rcv_nxt
== tp
->rcv_wup
)
4201 tcp_store_ts_recent(tp
);
4203 tcp_rcv_rtt_measure_ts(sk
, skb
);
4205 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4208 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4210 /* Bulk data transfer: receiver */
4211 __skb_pull(skb
,tcp_header_len
);
4212 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4213 sk_stream_set_owner_r(skb
, sk
);
4214 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4217 tcp_event_data_recv(sk
, tp
, skb
);
4219 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4220 /* Well, only one small jumplet in fast path... */
4221 tcp_ack(sk
, skb
, FLAG_DATA
);
4222 tcp_data_snd_check(sk
, tp
);
4223 if (!inet_csk_ack_scheduled(sk
))
4227 __tcp_ack_snd_check(sk
, 0);
4229 #ifdef CONFIG_NET_DMA
4231 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4237 sk
->sk_data_ready(sk
, 0);
4243 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4247 * RFC1323: H1. Apply PAWS check first.
4249 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4250 tcp_paws_discard(sk
, skb
)) {
4252 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4253 tcp_send_dupack(sk
, skb
);
4256 /* Resets are accepted even if PAWS failed.
4258 ts_recent update must be made after we are sure
4259 that the packet is in window.
4264 * Standard slow path.
4267 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4268 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4269 * (RST) segments are validated by checking their SEQ-fields."
4270 * And page 69: "If an incoming segment is not acceptable,
4271 * an acknowledgment should be sent in reply (unless the RST bit
4272 * is set, if so drop the segment and return)".
4275 tcp_send_dupack(sk
, skb
);
4284 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4286 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4287 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4288 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4295 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4297 tcp_rcv_rtt_measure_ts(sk
, skb
);
4299 /* Process urgent data. */
4300 tcp_urg(sk
, skb
, th
);
4302 /* step 7: process the segment text */
4303 tcp_data_queue(sk
, skb
);
4305 tcp_data_snd_check(sk
, tp
);
4306 tcp_ack_snd_check(sk
);
4310 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4317 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4318 struct tcphdr
*th
, unsigned len
)
4320 struct tcp_sock
*tp
= tcp_sk(sk
);
4321 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4322 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4324 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4328 * "If the state is SYN-SENT then
4329 * first check the ACK bit
4330 * If the ACK bit is set
4331 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4332 * a reset (unless the RST bit is set, if so drop
4333 * the segment and return)"
4335 * We do not send data with SYN, so that RFC-correct
4338 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4339 goto reset_and_undo
;
4341 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4342 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4344 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4345 goto reset_and_undo
;
4348 /* Now ACK is acceptable.
4350 * "If the RST bit is set
4351 * If the ACK was acceptable then signal the user "error:
4352 * connection reset", drop the segment, enter CLOSED state,
4353 * delete TCB, and return."
4362 * "fifth, if neither of the SYN or RST bits is set then
4363 * drop the segment and return."
4369 goto discard_and_undo
;
4372 * "If the SYN bit is on ...
4373 * are acceptable then ...
4374 * (our SYN has been ACKed), change the connection
4375 * state to ESTABLISHED..."
4378 TCP_ECN_rcv_synack(tp
, th
);
4380 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4381 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4383 /* Ok.. it's good. Set up sequence numbers and
4384 * move to established.
4386 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4387 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4389 /* RFC1323: The window in SYN & SYN/ACK segments is
4392 tp
->snd_wnd
= ntohs(th
->window
);
4393 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4395 if (!tp
->rx_opt
.wscale_ok
) {
4396 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4397 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4400 if (tp
->rx_opt
.saw_tstamp
) {
4401 tp
->rx_opt
.tstamp_ok
= 1;
4402 tp
->tcp_header_len
=
4403 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4404 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4405 tcp_store_ts_recent(tp
);
4407 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4410 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4411 tp
->rx_opt
.sack_ok
|= 2;
4414 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4415 tcp_initialize_rcv_mss(sk
);
4417 /* Remember, tcp_poll() does not lock socket!
4418 * Change state from SYN-SENT only after copied_seq
4419 * is initialized. */
4420 tp
->copied_seq
= tp
->rcv_nxt
;
4422 tcp_set_state(sk
, TCP_ESTABLISHED
);
4424 security_inet_conn_established(sk
, skb
);
4426 /* Make sure socket is routed, for correct metrics. */
4427 icsk
->icsk_af_ops
->rebuild_header(sk
);
4429 tcp_init_metrics(sk
);
4431 tcp_init_congestion_control(sk
);
4433 /* Prevent spurious tcp_cwnd_restart() on first data
4436 tp
->lsndtime
= tcp_time_stamp
;
4438 tcp_init_buffer_space(sk
);
4440 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4441 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4443 if (!tp
->rx_opt
.snd_wscale
)
4444 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4448 if (!sock_flag(sk
, SOCK_DEAD
)) {
4449 sk
->sk_state_change(sk
);
4450 sk_wake_async(sk
, 0, POLL_OUT
);
4453 if (sk
->sk_write_pending
||
4454 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4455 icsk
->icsk_ack
.pingpong
) {
4456 /* Save one ACK. Data will be ready after
4457 * several ticks, if write_pending is set.
4459 * It may be deleted, but with this feature tcpdumps
4460 * look so _wonderfully_ clever, that I was not able
4461 * to stand against the temptation 8) --ANK
4463 inet_csk_schedule_ack(sk
);
4464 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4465 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4466 tcp_incr_quickack(sk
);
4467 tcp_enter_quickack_mode(sk
);
4468 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4469 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4480 /* No ACK in the segment */
4484 * "If the RST bit is set
4486 * Otherwise (no ACK) drop the segment and return."
4489 goto discard_and_undo
;
4493 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4494 goto discard_and_undo
;
4497 /* We see SYN without ACK. It is attempt of
4498 * simultaneous connect with crossed SYNs.
4499 * Particularly, it can be connect to self.
4501 tcp_set_state(sk
, TCP_SYN_RECV
);
4503 if (tp
->rx_opt
.saw_tstamp
) {
4504 tp
->rx_opt
.tstamp_ok
= 1;
4505 tcp_store_ts_recent(tp
);
4506 tp
->tcp_header_len
=
4507 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4509 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4512 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4513 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4515 /* RFC1323: The window in SYN & SYN/ACK segments is
4518 tp
->snd_wnd
= ntohs(th
->window
);
4519 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4520 tp
->max_window
= tp
->snd_wnd
;
4522 TCP_ECN_rcv_syn(tp
, th
);
4525 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4526 tcp_initialize_rcv_mss(sk
);
4529 tcp_send_synack(sk
);
4531 /* Note, we could accept data and URG from this segment.
4532 * There are no obstacles to make this.
4534 * However, if we ignore data in ACKless segments sometimes,
4535 * we have no reasons to accept it sometimes.
4536 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4537 * is not flawless. So, discard packet for sanity.
4538 * Uncomment this return to process the data.
4545 /* "fifth, if neither of the SYN or RST bits is set then
4546 * drop the segment and return."
4550 tcp_clear_options(&tp
->rx_opt
);
4551 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4555 tcp_clear_options(&tp
->rx_opt
);
4556 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4562 * This function implements the receiving procedure of RFC 793 for
4563 * all states except ESTABLISHED and TIME_WAIT.
4564 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4565 * address independent.
4568 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4569 struct tcphdr
*th
, unsigned len
)
4571 struct tcp_sock
*tp
= tcp_sk(sk
);
4572 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4575 tp
->rx_opt
.saw_tstamp
= 0;
4577 switch (sk
->sk_state
) {
4589 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4592 /* Now we have several options: In theory there is
4593 * nothing else in the frame. KA9Q has an option to
4594 * send data with the syn, BSD accepts data with the
4595 * syn up to the [to be] advertised window and
4596 * Solaris 2.1 gives you a protocol error. For now
4597 * we just ignore it, that fits the spec precisely
4598 * and avoids incompatibilities. It would be nice in
4599 * future to drop through and process the data.
4601 * Now that TTCP is starting to be used we ought to
4603 * But, this leaves one open to an easy denial of
4604 * service attack, and SYN cookies can't defend
4605 * against this problem. So, we drop the data
4606 * in the interest of security over speed unless
4607 * it's still in use.
4615 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4619 /* Do step6 onward by hand. */
4620 tcp_urg(sk
, skb
, th
);
4622 tcp_data_snd_check(sk
, tp
);
4626 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4627 tcp_paws_discard(sk
, skb
)) {
4629 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4630 tcp_send_dupack(sk
, skb
);
4633 /* Reset is accepted even if it did not pass PAWS. */
4636 /* step 1: check sequence number */
4637 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4639 tcp_send_dupack(sk
, skb
);
4643 /* step 2: check RST bit */
4649 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4651 /* step 3: check security and precedence [ignored] */
4655 * Check for a SYN in window.
4657 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4658 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4663 /* step 5: check the ACK field */
4665 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4667 switch(sk
->sk_state
) {
4670 tp
->copied_seq
= tp
->rcv_nxt
;
4672 tcp_set_state(sk
, TCP_ESTABLISHED
);
4673 sk
->sk_state_change(sk
);
4675 /* Note, that this wakeup is only for marginal
4676 * crossed SYN case. Passively open sockets
4677 * are not waked up, because sk->sk_sleep ==
4678 * NULL and sk->sk_socket == NULL.
4680 if (sk
->sk_socket
) {
4681 sk_wake_async(sk
,0,POLL_OUT
);
4684 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4685 tp
->snd_wnd
= ntohs(th
->window
) <<
4686 tp
->rx_opt
.snd_wscale
;
4687 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4688 TCP_SKB_CB(skb
)->seq
);
4690 /* tcp_ack considers this ACK as duplicate
4691 * and does not calculate rtt.
4692 * Fix it at least with timestamps.
4694 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4696 tcp_ack_saw_tstamp(sk
, 0);
4698 if (tp
->rx_opt
.tstamp_ok
)
4699 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4701 /* Make sure socket is routed, for
4704 icsk
->icsk_af_ops
->rebuild_header(sk
);
4706 tcp_init_metrics(sk
);
4708 tcp_init_congestion_control(sk
);
4710 /* Prevent spurious tcp_cwnd_restart() on
4711 * first data packet.
4713 tp
->lsndtime
= tcp_time_stamp
;
4716 tcp_initialize_rcv_mss(sk
);
4717 tcp_init_buffer_space(sk
);
4718 tcp_fast_path_on(tp
);
4725 if (tp
->snd_una
== tp
->write_seq
) {
4726 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4727 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4728 dst_confirm(sk
->sk_dst_cache
);
4730 if (!sock_flag(sk
, SOCK_DEAD
))
4731 /* Wake up lingering close() */
4732 sk
->sk_state_change(sk
);
4736 if (tp
->linger2
< 0 ||
4737 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4738 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4740 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4744 tmo
= tcp_fin_time(sk
);
4745 if (tmo
> TCP_TIMEWAIT_LEN
) {
4746 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4747 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4748 /* Bad case. We could lose such FIN otherwise.
4749 * It is not a big problem, but it looks confusing
4750 * and not so rare event. We still can lose it now,
4751 * if it spins in bh_lock_sock(), but it is really
4754 inet_csk_reset_keepalive_timer(sk
, tmo
);
4756 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4764 if (tp
->snd_una
== tp
->write_seq
) {
4765 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4771 if (tp
->snd_una
== tp
->write_seq
) {
4772 tcp_update_metrics(sk
);
4781 /* step 6: check the URG bit */
4782 tcp_urg(sk
, skb
, th
);
4784 /* step 7: process the segment text */
4785 switch (sk
->sk_state
) {
4786 case TCP_CLOSE_WAIT
:
4789 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4793 /* RFC 793 says to queue data in these states,
4794 * RFC 1122 says we MUST send a reset.
4795 * BSD 4.4 also does reset.
4797 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4798 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4799 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4800 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4806 case TCP_ESTABLISHED
:
4807 tcp_data_queue(sk
, skb
);
4812 /* tcp_data could move socket to TIME-WAIT */
4813 if (sk
->sk_state
!= TCP_CLOSE
) {
4814 tcp_data_snd_check(sk
, tp
);
4815 tcp_ack_snd_check(sk
);
4825 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4826 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4827 EXPORT_SYMBOL(tcp_parse_options
);
4828 EXPORT_SYMBOL(tcp_rcv_established
);
4829 EXPORT_SYMBOL(tcp_rcv_state_process
);
4830 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);