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.*/
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
110 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
111 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
113 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
115 /* Adapt the MSS value used to make delayed ack decision to the
118 static void tcp_measure_rcv_mss(struct sock
*sk
,
119 const struct sk_buff
*skb
)
121 struct inet_connection_sock
*icsk
= inet_csk(sk
);
122 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
125 icsk
->icsk_ack
.last_seg_size
= 0;
127 /* skb->len may jitter because of SACKs, even if peer
128 * sends good full-sized frames.
130 len
= skb_shinfo(skb
)->gso_size
?: skb
->len
;
131 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
132 icsk
->icsk_ack
.rcv_mss
= len
;
134 /* Otherwise, we make more careful check taking into account,
135 * that SACKs block is variable.
137 * "len" is invariant segment length, including TCP header.
139 len
+= skb
->data
- skb
->h
.raw
;
140 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
141 /* If PSH is not set, packet should be
142 * full sized, provided peer TCP is not badly broken.
143 * This observation (if it is correct 8)) allows
144 * to handle super-low mtu links fairly.
146 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
147 !(tcp_flag_word(skb
->h
.th
)&TCP_REMNANT
))) {
148 /* Subtract also invariant (if peer is RFC compliant),
149 * tcp header plus fixed timestamp option length.
150 * Resulting "len" is MSS free of SACK jitter.
152 len
-= tcp_sk(sk
)->tcp_header_len
;
153 icsk
->icsk_ack
.last_seg_size
= len
;
155 icsk
->icsk_ack
.rcv_mss
= len
;
159 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
160 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
161 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
165 static void tcp_incr_quickack(struct sock
*sk
)
167 struct inet_connection_sock
*icsk
= inet_csk(sk
);
168 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
172 if (quickacks
> icsk
->icsk_ack
.quick
)
173 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
176 void tcp_enter_quickack_mode(struct sock
*sk
)
178 struct inet_connection_sock
*icsk
= inet_csk(sk
);
179 tcp_incr_quickack(sk
);
180 icsk
->icsk_ack
.pingpong
= 0;
181 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
184 /* Send ACKs quickly, if "quick" count is not exhausted
185 * and the session is not interactive.
188 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
190 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
191 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
194 /* Buffer size and advertised window tuning.
196 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
199 static void tcp_fixup_sndbuf(struct sock
*sk
)
201 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
202 sizeof(struct sk_buff
);
204 if (sk
->sk_sndbuf
< 3 * sndmem
)
205 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
208 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
210 * All tcp_full_space() is split to two parts: "network" buffer, allocated
211 * forward and advertised in receiver window (tp->rcv_wnd) and
212 * "application buffer", required to isolate scheduling/application
213 * latencies from network.
214 * window_clamp is maximal advertised window. It can be less than
215 * tcp_full_space(), in this case tcp_full_space() - window_clamp
216 * is reserved for "application" buffer. The less window_clamp is
217 * the smoother our behaviour from viewpoint of network, but the lower
218 * throughput and the higher sensitivity of the connection to losses. 8)
220 * rcv_ssthresh is more strict window_clamp used at "slow start"
221 * phase to predict further behaviour of this connection.
222 * It is used for two goals:
223 * - to enforce header prediction at sender, even when application
224 * requires some significant "application buffer". It is check #1.
225 * - to prevent pruning of receive queue because of misprediction
226 * of receiver window. Check #2.
228 * The scheme does not work when sender sends good segments opening
229 * window and then starts to feed us spaghetti. But it should work
230 * in common situations. Otherwise, we have to rely on queue collapsing.
233 /* Slow part of check#2. */
234 static int __tcp_grow_window(const struct sock
*sk
, struct tcp_sock
*tp
,
235 const struct sk_buff
*skb
)
238 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
239 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2])/2;
241 while (tp
->rcv_ssthresh
<= window
) {
242 if (truesize
<= skb
->len
)
243 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
251 static void tcp_grow_window(struct sock
*sk
, struct tcp_sock
*tp
,
255 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
256 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
257 !tcp_memory_pressure
) {
260 /* Check #2. Increase window, if skb with such overhead
261 * will fit to rcvbuf in future.
263 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
266 incr
= __tcp_grow_window(sk
, tp
, skb
);
269 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
270 inet_csk(sk
)->icsk_ack
.quick
|= 1;
275 /* 3. Tuning rcvbuf, when connection enters established state. */
277 static void tcp_fixup_rcvbuf(struct sock
*sk
)
279 struct tcp_sock
*tp
= tcp_sk(sk
);
280 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
282 /* Try to select rcvbuf so that 4 mss-sized segments
283 * will fit to window and corresponding skbs will fit to our rcvbuf.
284 * (was 3; 4 is minimum to allow fast retransmit to work.)
286 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
288 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
289 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
292 /* 4. Try to fixup all. It is made immediately after connection enters
295 static void tcp_init_buffer_space(struct sock
*sk
)
297 struct tcp_sock
*tp
= tcp_sk(sk
);
300 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
301 tcp_fixup_rcvbuf(sk
);
302 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
303 tcp_fixup_sndbuf(sk
);
305 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
307 maxwin
= tcp_full_space(sk
);
309 if (tp
->window_clamp
>= maxwin
) {
310 tp
->window_clamp
= maxwin
;
312 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
313 tp
->window_clamp
= max(maxwin
-
314 (maxwin
>> sysctl_tcp_app_win
),
318 /* Force reservation of one segment. */
319 if (sysctl_tcp_app_win
&&
320 tp
->window_clamp
> 2 * tp
->advmss
&&
321 tp
->window_clamp
+ tp
->advmss
> maxwin
)
322 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
324 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
325 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
328 /* 5. Recalculate window clamp after socket hit its memory bounds. */
329 static void tcp_clamp_window(struct sock
*sk
, struct tcp_sock
*tp
)
331 struct inet_connection_sock
*icsk
= inet_csk(sk
);
333 icsk
->icsk_ack
.quick
= 0;
335 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
336 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
337 !tcp_memory_pressure
&&
338 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
339 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
342 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
343 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
347 /* Initialize RCV_MSS value.
348 * RCV_MSS is an our guess about MSS used by the peer.
349 * We haven't any direct information about the MSS.
350 * It's better to underestimate the RCV_MSS rather than overestimate.
351 * Overestimations make us ACKing less frequently than needed.
352 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
354 void tcp_initialize_rcv_mss(struct sock
*sk
)
356 struct tcp_sock
*tp
= tcp_sk(sk
);
357 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
359 hint
= min(hint
, tp
->rcv_wnd
/2);
360 hint
= min(hint
, TCP_MIN_RCVMSS
);
361 hint
= max(hint
, TCP_MIN_MSS
);
363 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
366 /* Receiver "autotuning" code.
368 * The algorithm for RTT estimation w/o timestamps is based on
369 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
370 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
372 * More detail on this code can be found at
373 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
374 * though this reference is out of date. A new paper
377 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
379 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
385 if (new_sample
!= 0) {
386 /* If we sample in larger samples in the non-timestamp
387 * case, we could grossly overestimate the RTT especially
388 * with chatty applications or bulk transfer apps which
389 * are stalled on filesystem I/O.
391 * Also, since we are only going for a minimum in the
392 * non-timestamp case, we do not smooth things out
393 * else with timestamps disabled convergence takes too
397 m
-= (new_sample
>> 3);
399 } else if (m
< new_sample
)
402 /* No previous measure. */
406 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
407 tp
->rcv_rtt_est
.rtt
= new_sample
;
410 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
412 if (tp
->rcv_rtt_est
.time
== 0)
414 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
416 tcp_rcv_rtt_update(tp
,
417 jiffies
- tp
->rcv_rtt_est
.time
,
421 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
422 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
425 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
427 struct tcp_sock
*tp
= tcp_sk(sk
);
428 if (tp
->rx_opt
.rcv_tsecr
&&
429 (TCP_SKB_CB(skb
)->end_seq
-
430 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
431 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
435 * This function should be called every time data is copied to user space.
436 * It calculates the appropriate TCP receive buffer space.
438 void tcp_rcv_space_adjust(struct sock
*sk
)
440 struct tcp_sock
*tp
= tcp_sk(sk
);
444 if (tp
->rcvq_space
.time
== 0)
447 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
448 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
449 tp
->rcv_rtt_est
.rtt
== 0)
452 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
454 space
= max(tp
->rcvq_space
.space
, space
);
456 if (tp
->rcvq_space
.space
!= space
) {
459 tp
->rcvq_space
.space
= space
;
461 if (sysctl_tcp_moderate_rcvbuf
&&
462 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
463 int new_clamp
= space
;
465 /* Receive space grows, normalize in order to
466 * take into account packet headers and sk_buff
467 * structure overhead.
472 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
473 16 + sizeof(struct sk_buff
));
474 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
477 space
= min(space
, sysctl_tcp_rmem
[2]);
478 if (space
> sk
->sk_rcvbuf
) {
479 sk
->sk_rcvbuf
= space
;
481 /* Make the window clamp follow along. */
482 tp
->window_clamp
= new_clamp
;
488 tp
->rcvq_space
.seq
= tp
->copied_seq
;
489 tp
->rcvq_space
.time
= tcp_time_stamp
;
492 /* There is something which you must keep in mind when you analyze the
493 * behavior of the tp->ato delayed ack timeout interval. When a
494 * connection starts up, we want to ack as quickly as possible. The
495 * problem is that "good" TCP's do slow start at the beginning of data
496 * transmission. The means that until we send the first few ACK's the
497 * sender will sit on his end and only queue most of his data, because
498 * he can only send snd_cwnd unacked packets at any given time. For
499 * each ACK we send, he increments snd_cwnd and transmits more of his
502 static void tcp_event_data_recv(struct sock
*sk
, struct tcp_sock
*tp
, struct sk_buff
*skb
)
504 struct inet_connection_sock
*icsk
= inet_csk(sk
);
507 inet_csk_schedule_ack(sk
);
509 tcp_measure_rcv_mss(sk
, skb
);
511 tcp_rcv_rtt_measure(tp
);
513 now
= tcp_time_stamp
;
515 if (!icsk
->icsk_ack
.ato
) {
516 /* The _first_ data packet received, initialize
517 * delayed ACK engine.
519 tcp_incr_quickack(sk
);
520 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
522 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
524 if (m
<= TCP_ATO_MIN
/2) {
525 /* The fastest case is the first. */
526 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
527 } else if (m
< icsk
->icsk_ack
.ato
) {
528 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
529 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
530 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
531 } else if (m
> icsk
->icsk_rto
) {
532 /* Too long gap. Apparently sender failed to
533 * restart window, so that we send ACKs quickly.
535 tcp_incr_quickack(sk
);
536 sk_stream_mem_reclaim(sk
);
539 icsk
->icsk_ack
.lrcvtime
= now
;
541 TCP_ECN_check_ce(tp
, skb
);
544 tcp_grow_window(sk
, tp
, skb
);
547 /* Called to compute a smoothed rtt estimate. The data fed to this
548 * routine either comes from timestamps, or from segments that were
549 * known _not_ to have been retransmitted [see Karn/Partridge
550 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
551 * piece by Van Jacobson.
552 * NOTE: the next three routines used to be one big routine.
553 * To save cycles in the RFC 1323 implementation it was better to break
554 * it up into three procedures. -- erics
556 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
558 struct tcp_sock
*tp
= tcp_sk(sk
);
559 long m
= mrtt
; /* RTT */
561 /* The following amusing code comes from Jacobson's
562 * article in SIGCOMM '88. Note that rtt and mdev
563 * are scaled versions of rtt and mean deviation.
564 * This is designed to be as fast as possible
565 * m stands for "measurement".
567 * On a 1990 paper the rto value is changed to:
568 * RTO = rtt + 4 * mdev
570 * Funny. This algorithm seems to be very broken.
571 * These formulae increase RTO, when it should be decreased, increase
572 * too slowly, when it should be increased quickly, decrease too quickly
573 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
574 * does not matter how to _calculate_ it. Seems, it was trap
575 * that VJ failed to avoid. 8)
580 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
581 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
583 m
= -m
; /* m is now abs(error) */
584 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
585 /* This is similar to one of Eifel findings.
586 * Eifel blocks mdev updates when rtt decreases.
587 * This solution is a bit different: we use finer gain
588 * for mdev in this case (alpha*beta).
589 * Like Eifel it also prevents growth of rto,
590 * but also it limits too fast rto decreases,
591 * happening in pure Eifel.
596 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
598 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
599 if (tp
->mdev
> tp
->mdev_max
) {
600 tp
->mdev_max
= tp
->mdev
;
601 if (tp
->mdev_max
> tp
->rttvar
)
602 tp
->rttvar
= tp
->mdev_max
;
604 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
605 if (tp
->mdev_max
< tp
->rttvar
)
606 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
607 tp
->rtt_seq
= tp
->snd_nxt
;
608 tp
->mdev_max
= TCP_RTO_MIN
;
611 /* no previous measure. */
612 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
613 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
614 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
615 tp
->rtt_seq
= tp
->snd_nxt
;
619 /* Calculate rto without backoff. This is the second half of Van Jacobson's
620 * routine referred to above.
622 static inline void tcp_set_rto(struct sock
*sk
)
624 const struct tcp_sock
*tp
= tcp_sk(sk
);
625 /* Old crap is replaced with new one. 8)
628 * 1. If rtt variance happened to be less 50msec, it is hallucination.
629 * It cannot be less due to utterly erratic ACK generation made
630 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
631 * to do with delayed acks, because at cwnd>2 true delack timeout
632 * is invisible. Actually, Linux-2.4 also generates erratic
633 * ACKs in some circumstances.
635 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
637 /* 2. Fixups made earlier cannot be right.
638 * If we do not estimate RTO correctly without them,
639 * all the algo is pure shit and should be replaced
640 * with correct one. It is exactly, which we pretend to do.
644 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
645 * guarantees that rto is higher.
647 static inline void tcp_bound_rto(struct sock
*sk
)
649 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
650 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
653 /* Save metrics learned by this TCP session.
654 This function is called only, when TCP finishes successfully
655 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
657 void tcp_update_metrics(struct sock
*sk
)
659 struct tcp_sock
*tp
= tcp_sk(sk
);
660 struct dst_entry
*dst
= __sk_dst_get(sk
);
662 if (sysctl_tcp_nometrics_save
)
667 if (dst
&& (dst
->flags
&DST_HOST
)) {
668 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
671 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
672 /* This session failed to estimate rtt. Why?
673 * Probably, no packets returned in time.
676 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
677 dst
->metrics
[RTAX_RTT
-1] = 0;
681 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
683 /* If newly calculated rtt larger than stored one,
684 * store new one. Otherwise, use EWMA. Remember,
685 * rtt overestimation is always better than underestimation.
687 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
689 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
691 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
694 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
698 /* Scale deviation to rttvar fixed point */
703 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
704 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
706 dst
->metrics
[RTAX_RTTVAR
-1] -=
707 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
710 if (tp
->snd_ssthresh
>= 0xFFFF) {
711 /* Slow start still did not finish. */
712 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
713 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
714 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
715 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
716 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
717 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
718 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
719 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
720 icsk
->icsk_ca_state
== TCP_CA_Open
) {
721 /* Cong. avoidance phase, cwnd is reliable. */
722 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
723 dst
->metrics
[RTAX_SSTHRESH
-1] =
724 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
725 if (!dst_metric_locked(dst
, RTAX_CWND
))
726 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
728 /* Else slow start did not finish, cwnd is non-sense,
729 ssthresh may be also invalid.
731 if (!dst_metric_locked(dst
, RTAX_CWND
))
732 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
733 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
734 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
735 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
736 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
739 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
740 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
741 tp
->reordering
!= sysctl_tcp_reordering
)
742 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
747 /* Numbers are taken from RFC2414. */
748 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
750 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
753 if (tp
->mss_cache
> 1460)
756 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
758 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
761 /* Set slow start threshold and cwnd not falling to slow start */
762 void tcp_enter_cwr(struct sock
*sk
)
764 struct tcp_sock
*tp
= tcp_sk(sk
);
766 tp
->prior_ssthresh
= 0;
768 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
770 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
771 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
772 tcp_packets_in_flight(tp
) + 1U);
773 tp
->snd_cwnd_cnt
= 0;
774 tp
->high_seq
= tp
->snd_nxt
;
775 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
776 TCP_ECN_queue_cwr(tp
);
778 tcp_set_ca_state(sk
, TCP_CA_CWR
);
782 /* Initialize metrics on socket. */
784 static void tcp_init_metrics(struct sock
*sk
)
786 struct tcp_sock
*tp
= tcp_sk(sk
);
787 struct dst_entry
*dst
= __sk_dst_get(sk
);
794 if (dst_metric_locked(dst
, RTAX_CWND
))
795 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
796 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
797 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
798 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
799 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
801 if (dst_metric(dst
, RTAX_REORDERING
) &&
802 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
803 tp
->rx_opt
.sack_ok
&= ~2;
804 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
807 if (dst_metric(dst
, RTAX_RTT
) == 0)
810 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
813 /* Initial rtt is determined from SYN,SYN-ACK.
814 * The segment is small and rtt may appear much
815 * less than real one. Use per-dst memory
816 * to make it more realistic.
818 * A bit of theory. RTT is time passed after "normal" sized packet
819 * is sent until it is ACKed. In normal circumstances sending small
820 * packets force peer to delay ACKs and calculation is correct too.
821 * The algorithm is adaptive and, provided we follow specs, it
822 * NEVER underestimate RTT. BUT! If peer tries to make some clever
823 * tricks sort of "quick acks" for time long enough to decrease RTT
824 * to low value, and then abruptly stops to do it and starts to delay
825 * ACKs, wait for troubles.
827 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
828 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
829 tp
->rtt_seq
= tp
->snd_nxt
;
831 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
832 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
833 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
837 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
839 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
840 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
844 /* Play conservative. If timestamps are not
845 * supported, TCP will fail to recalculate correct
846 * rtt, if initial rto is too small. FORGET ALL AND RESET!
848 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
850 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
851 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
855 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
858 struct tcp_sock
*tp
= tcp_sk(sk
);
859 if (metric
> tp
->reordering
) {
860 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
862 /* This exciting event is worth to be remembered. 8) */
864 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
866 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
868 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
870 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
871 #if FASTRETRANS_DEBUG > 1
872 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
873 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
877 tp
->undo_marker
? tp
->undo_retrans
: 0);
879 /* Disable FACK yet. */
880 tp
->rx_opt
.sack_ok
&= ~2;
884 /* This procedure tags the retransmission queue when SACKs arrive.
886 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
887 * Packets in queue with these bits set are counted in variables
888 * sacked_out, retrans_out and lost_out, correspondingly.
890 * Valid combinations are:
891 * Tag InFlight Description
892 * 0 1 - orig segment is in flight.
893 * S 0 - nothing flies, orig reached receiver.
894 * L 0 - nothing flies, orig lost by net.
895 * R 2 - both orig and retransmit are in flight.
896 * L|R 1 - orig is lost, retransmit is in flight.
897 * S|R 1 - orig reached receiver, retrans is still in flight.
898 * (L|S|R is logically valid, it could occur when L|R is sacked,
899 * but it is equivalent to plain S and code short-curcuits it to S.
900 * L|S is logically invalid, it would mean -1 packet in flight 8))
902 * These 6 states form finite state machine, controlled by the following events:
903 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
904 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
905 * 3. Loss detection event of one of three flavors:
906 * A. Scoreboard estimator decided the packet is lost.
907 * A'. Reno "three dupacks" marks head of queue lost.
908 * A''. Its FACK modfication, head until snd.fack is lost.
909 * B. SACK arrives sacking data transmitted after never retransmitted
911 * C. SACK arrives sacking SND.NXT at the moment, when the
912 * segment was retransmitted.
913 * 4. D-SACK added new rule: D-SACK changes any tag to S.
915 * It is pleasant to note, that state diagram turns out to be commutative,
916 * so that we are allowed not to be bothered by order of our actions,
917 * when multiple events arrive simultaneously. (see the function below).
919 * Reordering detection.
920 * --------------------
921 * Reordering metric is maximal distance, which a packet can be displaced
922 * in packet stream. With SACKs we can estimate it:
924 * 1. SACK fills old hole and the corresponding segment was not
925 * ever retransmitted -> reordering. Alas, we cannot use it
926 * when segment was retransmitted.
927 * 2. The last flaw is solved with D-SACK. D-SACK arrives
928 * for retransmitted and already SACKed segment -> reordering..
929 * Both of these heuristics are not used in Loss state, when we cannot
930 * account for retransmits accurately.
933 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
935 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
936 struct tcp_sock
*tp
= tcp_sk(sk
);
937 unsigned char *ptr
= ack_skb
->h
.raw
+ TCP_SKB_CB(ack_skb
)->sacked
;
938 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
939 struct sk_buff
*cached_skb
;
940 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
941 int reord
= tp
->packets_out
;
943 u32 lost_retrans
= 0;
946 int cached_fack_count
;
948 int first_sack_index
;
952 prior_fackets
= tp
->fackets_out
;
954 /* Check for D-SACK. */
955 if (before(ntohl(sp
[0].start_seq
), TCP_SKB_CB(ack_skb
)->ack_seq
)) {
957 tp
->rx_opt
.sack_ok
|= 4;
958 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
959 } else if (num_sacks
> 1 &&
960 !after(ntohl(sp
[0].end_seq
), ntohl(sp
[1].end_seq
)) &&
961 !before(ntohl(sp
[0].start_seq
), ntohl(sp
[1].start_seq
))) {
963 tp
->rx_opt
.sack_ok
|= 4;
964 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
967 /* D-SACK for already forgotten data...
968 * Do dumb counting. */
970 !after(ntohl(sp
[0].end_seq
), prior_snd_una
) &&
971 after(ntohl(sp
[0].end_seq
), tp
->undo_marker
))
974 /* Eliminate too old ACKs, but take into
975 * account more or less fresh ones, they can
976 * contain valid SACK info.
978 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
982 * if the only SACK change is the increase of the end_seq of
983 * the first block then only apply that SACK block
984 * and use retrans queue hinting otherwise slowpath */
986 for (i
= 0; i
< num_sacks
; i
++) {
987 __be32 start_seq
= sp
[i
].start_seq
;
988 __be32 end_seq
= sp
[i
].end_seq
;
991 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
994 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
995 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
998 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
999 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1001 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1002 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1003 tp
->recv_sack_cache
[i
].start_seq
= 0;
1004 tp
->recv_sack_cache
[i
].end_seq
= 0;
1007 first_sack_index
= 0;
1012 tp
->fastpath_skb_hint
= NULL
;
1014 /* order SACK blocks to allow in order walk of the retrans queue */
1015 for (i
= num_sacks
-1; i
> 0; i
--) {
1016 for (j
= 0; j
< i
; j
++){
1017 if (after(ntohl(sp
[j
].start_seq
),
1018 ntohl(sp
[j
+1].start_seq
))){
1019 struct tcp_sack_block_wire tmp
;
1025 /* Track where the first SACK block goes to */
1026 if (j
== first_sack_index
)
1027 first_sack_index
= j
+1;
1034 /* clear flag as used for different purpose in following code */
1037 /* Use SACK fastpath hint if valid */
1038 cached_skb
= tp
->fastpath_skb_hint
;
1039 cached_fack_count
= tp
->fastpath_cnt_hint
;
1041 cached_skb
= sk
->sk_write_queue
.next
;
1042 cached_fack_count
= 0;
1045 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1046 struct sk_buff
*skb
;
1047 __u32 start_seq
= ntohl(sp
->start_seq
);
1048 __u32 end_seq
= ntohl(sp
->end_seq
);
1052 fack_count
= cached_fack_count
;
1054 /* Event "B" in the comment above. */
1055 if (after(end_seq
, tp
->high_seq
))
1056 flag
|= FLAG_DATA_LOST
;
1058 sk_stream_for_retrans_queue_from(skb
, sk
) {
1059 int in_sack
, pcount
;
1063 cached_fack_count
= fack_count
;
1064 if (i
== first_sack_index
) {
1065 tp
->fastpath_skb_hint
= skb
;
1066 tp
->fastpath_cnt_hint
= fack_count
;
1069 /* The retransmission queue is always in order, so
1070 * we can short-circuit the walk early.
1072 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1075 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1076 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1078 pcount
= tcp_skb_pcount(skb
);
1080 if (pcount
> 1 && !in_sack
&&
1081 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1082 unsigned int pkt_len
;
1084 in_sack
= !after(start_seq
,
1085 TCP_SKB_CB(skb
)->seq
);
1088 pkt_len
= (start_seq
-
1089 TCP_SKB_CB(skb
)->seq
);
1091 pkt_len
= (end_seq
-
1092 TCP_SKB_CB(skb
)->seq
);
1093 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1095 pcount
= tcp_skb_pcount(skb
);
1098 fack_count
+= pcount
;
1100 sacked
= TCP_SKB_CB(skb
)->sacked
;
1102 /* Account D-SACK for retransmitted packet. */
1103 if ((dup_sack
&& in_sack
) &&
1104 (sacked
& TCPCB_RETRANS
) &&
1105 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1108 /* The frame is ACKed. */
1109 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1110 if (sacked
&TCPCB_RETRANS
) {
1111 if ((dup_sack
&& in_sack
) &&
1112 (sacked
&TCPCB_SACKED_ACKED
))
1113 reord
= min(fack_count
, reord
);
1115 /* If it was in a hole, we detected reordering. */
1116 if (fack_count
< prior_fackets
&&
1117 !(sacked
&TCPCB_SACKED_ACKED
))
1118 reord
= min(fack_count
, reord
);
1121 /* Nothing to do; acked frame is about to be dropped. */
1125 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1126 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1127 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1128 lost_retrans
= end_seq
;
1133 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1134 if (sacked
& TCPCB_SACKED_RETRANS
) {
1135 /* If the segment is not tagged as lost,
1136 * we do not clear RETRANS, believing
1137 * that retransmission is still in flight.
1139 if (sacked
& TCPCB_LOST
) {
1140 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1141 tp
->lost_out
-= tcp_skb_pcount(skb
);
1142 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1144 /* clear lost hint */
1145 tp
->retransmit_skb_hint
= NULL
;
1148 /* New sack for not retransmitted frame,
1149 * which was in hole. It is reordering.
1151 if (!(sacked
& TCPCB_RETRANS
) &&
1152 fack_count
< prior_fackets
)
1153 reord
= min(fack_count
, reord
);
1155 if (sacked
& TCPCB_LOST
) {
1156 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1157 tp
->lost_out
-= tcp_skb_pcount(skb
);
1159 /* clear lost hint */
1160 tp
->retransmit_skb_hint
= NULL
;
1164 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1165 flag
|= FLAG_DATA_SACKED
;
1166 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1168 if (fack_count
> tp
->fackets_out
)
1169 tp
->fackets_out
= fack_count
;
1171 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1172 reord
= min(fack_count
, reord
);
1175 /* D-SACK. We can detect redundant retransmission
1176 * in S|R and plain R frames and clear it.
1177 * undo_retrans is decreased above, L|R frames
1178 * are accounted above as well.
1181 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1182 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1183 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1184 tp
->retransmit_skb_hint
= NULL
;
1189 /* Check for lost retransmit. This superb idea is
1190 * borrowed from "ratehalving". Event "C".
1191 * Later note: FACK people cheated me again 8),
1192 * we have to account for reordering! Ugly,
1195 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1196 struct sk_buff
*skb
;
1198 sk_stream_for_retrans_queue(skb
, sk
) {
1199 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1201 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1203 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1204 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1206 !before(lost_retrans
,
1207 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1209 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1210 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1212 /* clear lost hint */
1213 tp
->retransmit_skb_hint
= NULL
;
1215 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1216 tp
->lost_out
+= tcp_skb_pcount(skb
);
1217 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1218 flag
|= FLAG_DATA_SACKED
;
1219 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1225 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1227 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1228 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1230 #if FASTRETRANS_DEBUG > 0
1231 BUG_TRAP((int)tp
->sacked_out
>= 0);
1232 BUG_TRAP((int)tp
->lost_out
>= 0);
1233 BUG_TRAP((int)tp
->retrans_out
>= 0);
1234 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1239 /* F-RTO can only be used if these conditions are satisfied:
1240 * - there must be some unsent new data
1241 * - the advertised window should allow sending it
1242 * - TCP has never retransmitted anything other than head
1244 int tcp_use_frto(struct sock
*sk
)
1246 const struct tcp_sock
*tp
= tcp_sk(sk
);
1247 struct sk_buff
*skb
;
1249 if (!sysctl_tcp_frto
|| !sk
->sk_send_head
||
1250 after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
1251 tp
->snd_una
+ tp
->snd_wnd
))
1254 /* Avoid expensive walking of rexmit queue if possible */
1255 if (tp
->retrans_out
> 1)
1258 skb
= skb_peek(&sk
->sk_write_queue
)->next
; /* Skips head */
1259 sk_stream_for_retrans_queue_from(skb
, sk
) {
1260 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1262 /* Short-circuit when first non-SACKed skb has been checked */
1263 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1269 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1270 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1271 * the RTO was spurious.
1273 * Do like tcp_enter_loss() would; when RTO expires the second time it
1275 * "Reduce ssthresh if it has not yet been made inside this window."
1277 void tcp_enter_frto(struct sock
*sk
)
1279 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1280 struct tcp_sock
*tp
= tcp_sk(sk
);
1281 struct sk_buff
*skb
;
1283 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1284 tp
->snd_una
== tp
->high_seq
||
1285 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1286 !icsk
->icsk_retransmits
)) {
1287 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1288 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1289 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1292 /* Have to clear retransmission markers here to keep the bookkeeping
1293 * in shape, even though we are not yet in Loss state.
1294 * If something was really lost, it is eventually caught up
1295 * in tcp_enter_frto_loss.
1297 tp
->retrans_out
= 0;
1298 tp
->undo_marker
= tp
->snd_una
;
1299 tp
->undo_retrans
= 0;
1301 sk_stream_for_retrans_queue(skb
, sk
) {
1302 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1304 tcp_sync_left_out(tp
);
1306 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1307 tp
->high_seq
= tp
->snd_nxt
;
1308 tp
->frto_highmark
= tp
->snd_nxt
;
1309 tp
->frto_counter
= 1;
1312 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1313 * which indicates that we should follow the traditional RTO recovery,
1314 * i.e. mark everything lost and do go-back-N retransmission.
1316 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
)
1318 struct tcp_sock
*tp
= tcp_sk(sk
);
1319 struct sk_buff
*skb
;
1324 tp
->fackets_out
= 0;
1326 sk_stream_for_retrans_queue(skb
, sk
) {
1327 cnt
+= tcp_skb_pcount(skb
);
1328 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1329 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1331 /* Do not mark those segments lost that were
1332 * forward transmitted after RTO
1334 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1335 tp
->frto_highmark
)) {
1336 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1337 tp
->lost_out
+= tcp_skb_pcount(skb
);
1340 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1341 tp
->fackets_out
= cnt
;
1344 tcp_sync_left_out(tp
);
1346 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1347 tp
->snd_cwnd_cnt
= 0;
1348 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1349 tp
->undo_marker
= 0;
1350 tp
->frto_counter
= 0;
1352 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1353 sysctl_tcp_reordering
);
1354 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1355 tp
->high_seq
= tp
->frto_highmark
;
1356 TCP_ECN_queue_cwr(tp
);
1358 clear_all_retrans_hints(tp
);
1361 void tcp_clear_retrans(struct tcp_sock
*tp
)
1364 tp
->retrans_out
= 0;
1366 tp
->fackets_out
= 0;
1370 tp
->undo_marker
= 0;
1371 tp
->undo_retrans
= 0;
1374 /* Enter Loss state. If "how" is not zero, forget all SACK information
1375 * and reset tags completely, otherwise preserve SACKs. If receiver
1376 * dropped its ofo queue, we will know this due to reneging detection.
1378 void tcp_enter_loss(struct sock
*sk
, int how
)
1380 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1381 struct tcp_sock
*tp
= tcp_sk(sk
);
1382 struct sk_buff
*skb
;
1385 /* Reduce ssthresh if it has not yet been made inside this window. */
1386 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1387 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1388 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1389 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1390 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1393 tp
->snd_cwnd_cnt
= 0;
1394 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1396 tp
->bytes_acked
= 0;
1397 tcp_clear_retrans(tp
);
1399 /* Push undo marker, if it was plain RTO and nothing
1400 * was retransmitted. */
1402 tp
->undo_marker
= tp
->snd_una
;
1404 sk_stream_for_retrans_queue(skb
, sk
) {
1405 cnt
+= tcp_skb_pcount(skb
);
1406 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1407 tp
->undo_marker
= 0;
1408 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1409 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1410 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1411 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1412 tp
->lost_out
+= tcp_skb_pcount(skb
);
1414 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1415 tp
->fackets_out
= cnt
;
1418 tcp_sync_left_out(tp
);
1420 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1421 sysctl_tcp_reordering
);
1422 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1423 tp
->high_seq
= tp
->snd_nxt
;
1424 TCP_ECN_queue_cwr(tp
);
1426 clear_all_retrans_hints(tp
);
1429 static int tcp_check_sack_reneging(struct sock
*sk
)
1431 struct sk_buff
*skb
;
1433 /* If ACK arrived pointing to a remembered SACK,
1434 * it means that our remembered SACKs do not reflect
1435 * real state of receiver i.e.
1436 * receiver _host_ is heavily congested (or buggy).
1437 * Do processing similar to RTO timeout.
1439 if ((skb
= skb_peek(&sk
->sk_write_queue
)) != NULL
&&
1440 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1441 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1442 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1444 tcp_enter_loss(sk
, 1);
1445 icsk
->icsk_retransmits
++;
1446 tcp_retransmit_skb(sk
, skb_peek(&sk
->sk_write_queue
));
1447 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1448 icsk
->icsk_rto
, TCP_RTO_MAX
);
1454 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1456 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1459 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1461 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1464 static inline int tcp_head_timedout(struct sock
*sk
, struct tcp_sock
*tp
)
1466 return tp
->packets_out
&&
1467 tcp_skb_timedout(sk
, skb_peek(&sk
->sk_write_queue
));
1470 /* Linux NewReno/SACK/FACK/ECN state machine.
1471 * --------------------------------------
1473 * "Open" Normal state, no dubious events, fast path.
1474 * "Disorder" In all the respects it is "Open",
1475 * but requires a bit more attention. It is entered when
1476 * we see some SACKs or dupacks. It is split of "Open"
1477 * mainly to move some processing from fast path to slow one.
1478 * "CWR" CWND was reduced due to some Congestion Notification event.
1479 * It can be ECN, ICMP source quench, local device congestion.
1480 * "Recovery" CWND was reduced, we are fast-retransmitting.
1481 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1483 * tcp_fastretrans_alert() is entered:
1484 * - each incoming ACK, if state is not "Open"
1485 * - when arrived ACK is unusual, namely:
1490 * Counting packets in flight is pretty simple.
1492 * in_flight = packets_out - left_out + retrans_out
1494 * packets_out is SND.NXT-SND.UNA counted in packets.
1496 * retrans_out is number of retransmitted segments.
1498 * left_out is number of segments left network, but not ACKed yet.
1500 * left_out = sacked_out + lost_out
1502 * sacked_out: Packets, which arrived to receiver out of order
1503 * and hence not ACKed. With SACKs this number is simply
1504 * amount of SACKed data. Even without SACKs
1505 * it is easy to give pretty reliable estimate of this number,
1506 * counting duplicate ACKs.
1508 * lost_out: Packets lost by network. TCP has no explicit
1509 * "loss notification" feedback from network (for now).
1510 * It means that this number can be only _guessed_.
1511 * Actually, it is the heuristics to predict lossage that
1512 * distinguishes different algorithms.
1514 * F.e. after RTO, when all the queue is considered as lost,
1515 * lost_out = packets_out and in_flight = retrans_out.
1517 * Essentially, we have now two algorithms counting
1520 * FACK: It is the simplest heuristics. As soon as we decided
1521 * that something is lost, we decide that _all_ not SACKed
1522 * packets until the most forward SACK are lost. I.e.
1523 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1524 * It is absolutely correct estimate, if network does not reorder
1525 * packets. And it loses any connection to reality when reordering
1526 * takes place. We use FACK by default until reordering
1527 * is suspected on the path to this destination.
1529 * NewReno: when Recovery is entered, we assume that one segment
1530 * is lost (classic Reno). While we are in Recovery and
1531 * a partial ACK arrives, we assume that one more packet
1532 * is lost (NewReno). This heuristics are the same in NewReno
1535 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1536 * deflation etc. CWND is real congestion window, never inflated, changes
1537 * only according to classic VJ rules.
1539 * Really tricky (and requiring careful tuning) part of algorithm
1540 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1541 * The first determines the moment _when_ we should reduce CWND and,
1542 * hence, slow down forward transmission. In fact, it determines the moment
1543 * when we decide that hole is caused by loss, rather than by a reorder.
1545 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1546 * holes, caused by lost packets.
1548 * And the most logically complicated part of algorithm is undo
1549 * heuristics. We detect false retransmits due to both too early
1550 * fast retransmit (reordering) and underestimated RTO, analyzing
1551 * timestamps and D-SACKs. When we detect that some segments were
1552 * retransmitted by mistake and CWND reduction was wrong, we undo
1553 * window reduction and abort recovery phase. This logic is hidden
1554 * inside several functions named tcp_try_undo_<something>.
1557 /* This function decides, when we should leave Disordered state
1558 * and enter Recovery phase, reducing congestion window.
1560 * Main question: may we further continue forward transmission
1561 * with the same cwnd?
1563 static int tcp_time_to_recover(struct sock
*sk
, struct tcp_sock
*tp
)
1567 /* Do not perform any recovery during FRTO algorithm */
1568 if (tp
->frto_counter
)
1571 /* Trick#1: The loss is proven. */
1575 /* Not-A-Trick#2 : Classic rule... */
1576 if (tcp_fackets_out(tp
) > tp
->reordering
)
1579 /* Trick#3 : when we use RFC2988 timer restart, fast
1580 * retransmit can be triggered by timeout of queue head.
1582 if (tcp_head_timedout(sk
, tp
))
1585 /* Trick#4: It is still not OK... But will it be useful to delay
1588 packets_out
= tp
->packets_out
;
1589 if (packets_out
<= tp
->reordering
&&
1590 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1591 !tcp_may_send_now(sk
, tp
)) {
1592 /* We have nothing to send. This connection is limited
1593 * either by receiver window or by application.
1601 /* If we receive more dupacks than we expected counting segments
1602 * in assumption of absent reordering, interpret this as reordering.
1603 * The only another reason could be bug in receiver TCP.
1605 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1607 struct tcp_sock
*tp
= tcp_sk(sk
);
1610 holes
= max(tp
->lost_out
, 1U);
1611 holes
= min(holes
, tp
->packets_out
);
1613 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1614 tp
->sacked_out
= tp
->packets_out
- holes
;
1615 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1619 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1621 static void tcp_add_reno_sack(struct sock
*sk
)
1623 struct tcp_sock
*tp
= tcp_sk(sk
);
1625 tcp_check_reno_reordering(sk
, 0);
1626 tcp_sync_left_out(tp
);
1629 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1631 static void tcp_remove_reno_sacks(struct sock
*sk
, struct tcp_sock
*tp
, int acked
)
1634 /* One ACK acked hole. The rest eat duplicate ACKs. */
1635 if (acked
-1 >= tp
->sacked_out
)
1638 tp
->sacked_out
-= acked
-1;
1640 tcp_check_reno_reordering(sk
, acked
);
1641 tcp_sync_left_out(tp
);
1644 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1647 tp
->left_out
= tp
->lost_out
;
1650 /* Mark head of queue up as lost. */
1651 static void tcp_mark_head_lost(struct sock
*sk
, struct tcp_sock
*tp
,
1652 int packets
, u32 high_seq
)
1654 struct sk_buff
*skb
;
1657 BUG_TRAP(packets
<= tp
->packets_out
);
1658 if (tp
->lost_skb_hint
) {
1659 skb
= tp
->lost_skb_hint
;
1660 cnt
= tp
->lost_cnt_hint
;
1662 skb
= sk
->sk_write_queue
.next
;
1666 sk_stream_for_retrans_queue_from(skb
, sk
) {
1667 /* TODO: do this better */
1668 /* this is not the most efficient way to do this... */
1669 tp
->lost_skb_hint
= skb
;
1670 tp
->lost_cnt_hint
= cnt
;
1671 cnt
+= tcp_skb_pcount(skb
);
1672 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1674 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1675 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1676 tp
->lost_out
+= tcp_skb_pcount(skb
);
1678 /* clear xmit_retransmit_queue hints
1679 * if this is beyond hint */
1680 if(tp
->retransmit_skb_hint
!= NULL
&&
1681 before(TCP_SKB_CB(skb
)->seq
,
1682 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
)) {
1684 tp
->retransmit_skb_hint
= NULL
;
1688 tcp_sync_left_out(tp
);
1691 /* Account newly detected lost packet(s) */
1693 static void tcp_update_scoreboard(struct sock
*sk
, struct tcp_sock
*tp
)
1696 int lost
= tp
->fackets_out
- tp
->reordering
;
1699 tcp_mark_head_lost(sk
, tp
, lost
, tp
->high_seq
);
1701 tcp_mark_head_lost(sk
, tp
, 1, tp
->high_seq
);
1704 /* New heuristics: it is possible only after we switched
1705 * to restart timer each time when something is ACKed.
1706 * Hence, we can detect timed out packets during fast
1707 * retransmit without falling to slow start.
1709 if (!IsReno(tp
) && tcp_head_timedout(sk
, tp
)) {
1710 struct sk_buff
*skb
;
1712 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1713 : sk
->sk_write_queue
.next
;
1715 sk_stream_for_retrans_queue_from(skb
, sk
) {
1716 if (!tcp_skb_timedout(sk
, skb
))
1719 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1720 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1721 tp
->lost_out
+= tcp_skb_pcount(skb
);
1723 /* clear xmit_retrans hint */
1724 if (tp
->retransmit_skb_hint
&&
1725 before(TCP_SKB_CB(skb
)->seq
,
1726 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1728 tp
->retransmit_skb_hint
= NULL
;
1732 tp
->scoreboard_skb_hint
= skb
;
1734 tcp_sync_left_out(tp
);
1738 /* CWND moderation, preventing bursts due to too big ACKs
1739 * in dubious situations.
1741 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1743 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1744 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1745 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1748 /* Lower bound on congestion window is slow start threshold
1749 * unless congestion avoidance choice decides to overide it.
1751 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1753 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1755 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1758 /* Decrease cwnd each second ack. */
1759 static void tcp_cwnd_down(struct sock
*sk
)
1761 struct tcp_sock
*tp
= tcp_sk(sk
);
1762 int decr
= tp
->snd_cwnd_cnt
+ 1;
1764 tp
->snd_cwnd_cnt
= decr
&1;
1767 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1768 tp
->snd_cwnd
-= decr
;
1770 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1771 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1774 /* Nothing was retransmitted or returned timestamp is less
1775 * than timestamp of the first retransmission.
1777 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1779 return !tp
->retrans_stamp
||
1780 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1781 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1784 /* Undo procedures. */
1786 #if FASTRETRANS_DEBUG > 1
1787 static void DBGUNDO(struct sock
*sk
, struct tcp_sock
*tp
, const char *msg
)
1789 struct inet_sock
*inet
= inet_sk(sk
);
1790 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1792 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1793 tp
->snd_cwnd
, tp
->left_out
,
1794 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1798 #define DBGUNDO(x...) do { } while (0)
1801 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1803 struct tcp_sock
*tp
= tcp_sk(sk
);
1805 if (tp
->prior_ssthresh
) {
1806 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1808 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1809 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1811 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1813 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1814 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1815 TCP_ECN_withdraw_cwr(tp
);
1818 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1820 tcp_moderate_cwnd(tp
);
1821 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1823 /* There is something screwy going on with the retrans hints after
1825 clear_all_retrans_hints(tp
);
1828 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1830 return tp
->undo_marker
&&
1831 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1834 /* People celebrate: "We love our President!" */
1835 static int tcp_try_undo_recovery(struct sock
*sk
, struct tcp_sock
*tp
)
1837 if (tcp_may_undo(tp
)) {
1838 /* Happy end! We did not retransmit anything
1839 * or our original transmission succeeded.
1841 DBGUNDO(sk
, tp
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1842 tcp_undo_cwr(sk
, 1);
1843 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1844 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1846 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1847 tp
->undo_marker
= 0;
1849 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1850 /* Hold old state until something *above* high_seq
1851 * is ACKed. For Reno it is MUST to prevent false
1852 * fast retransmits (RFC2582). SACK TCP is safe. */
1853 tcp_moderate_cwnd(tp
);
1856 tcp_set_ca_state(sk
, TCP_CA_Open
);
1860 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1861 static void tcp_try_undo_dsack(struct sock
*sk
, struct tcp_sock
*tp
)
1863 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1864 DBGUNDO(sk
, tp
, "D-SACK");
1865 tcp_undo_cwr(sk
, 1);
1866 tp
->undo_marker
= 0;
1867 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1871 /* Undo during fast recovery after partial ACK. */
1873 static int tcp_try_undo_partial(struct sock
*sk
, struct tcp_sock
*tp
,
1876 /* Partial ACK arrived. Force Hoe's retransmit. */
1877 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1879 if (tcp_may_undo(tp
)) {
1880 /* Plain luck! Hole if filled with delayed
1881 * packet, rather than with a retransmit.
1883 if (tp
->retrans_out
== 0)
1884 tp
->retrans_stamp
= 0;
1886 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
1888 DBGUNDO(sk
, tp
, "Hoe");
1889 tcp_undo_cwr(sk
, 0);
1890 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1892 /* So... Do not make Hoe's retransmit yet.
1893 * If the first packet was delayed, the rest
1894 * ones are most probably delayed as well.
1901 /* Undo during loss recovery after partial ACK. */
1902 static int tcp_try_undo_loss(struct sock
*sk
, struct tcp_sock
*tp
)
1904 if (tcp_may_undo(tp
)) {
1905 struct sk_buff
*skb
;
1906 sk_stream_for_retrans_queue(skb
, sk
) {
1907 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1910 clear_all_retrans_hints(tp
);
1912 DBGUNDO(sk
, tp
, "partial loss");
1914 tp
->left_out
= tp
->sacked_out
;
1915 tcp_undo_cwr(sk
, 1);
1916 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1917 inet_csk(sk
)->icsk_retransmits
= 0;
1918 tp
->undo_marker
= 0;
1920 tcp_set_ca_state(sk
, TCP_CA_Open
);
1926 static inline void tcp_complete_cwr(struct sock
*sk
)
1928 struct tcp_sock
*tp
= tcp_sk(sk
);
1929 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1930 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1931 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
1934 static void tcp_try_to_open(struct sock
*sk
, struct tcp_sock
*tp
, int flag
)
1936 tp
->left_out
= tp
->sacked_out
;
1938 if (tp
->retrans_out
== 0)
1939 tp
->retrans_stamp
= 0;
1944 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
1945 int state
= TCP_CA_Open
;
1947 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
1948 state
= TCP_CA_Disorder
;
1950 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
1951 tcp_set_ca_state(sk
, state
);
1952 tp
->high_seq
= tp
->snd_nxt
;
1954 tcp_moderate_cwnd(tp
);
1960 static void tcp_mtup_probe_failed(struct sock
*sk
)
1962 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1964 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
1965 icsk
->icsk_mtup
.probe_size
= 0;
1968 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
1970 struct tcp_sock
*tp
= tcp_sk(sk
);
1971 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1973 /* FIXME: breaks with very large cwnd */
1974 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1975 tp
->snd_cwnd
= tp
->snd_cwnd
*
1976 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
1977 icsk
->icsk_mtup
.probe_size
;
1978 tp
->snd_cwnd_cnt
= 0;
1979 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1980 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
1982 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
1983 icsk
->icsk_mtup
.probe_size
= 0;
1984 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
1988 /* Process an event, which can update packets-in-flight not trivially.
1989 * Main goal of this function is to calculate new estimate for left_out,
1990 * taking into account both packets sitting in receiver's buffer and
1991 * packets lost by network.
1993 * Besides that it does CWND reduction, when packet loss is detected
1994 * and changes state of machine.
1996 * It does _not_ decide what to send, it is made in function
1997 * tcp_xmit_retransmit_queue().
2000 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
2001 int prior_packets
, int flag
)
2003 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2004 struct tcp_sock
*tp
= tcp_sk(sk
);
2005 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
2007 /* Some technical things:
2008 * 1. Reno does not count dupacks (sacked_out) automatically. */
2009 if (!tp
->packets_out
)
2011 /* 2. SACK counts snd_fack in packets inaccurately. */
2012 if (tp
->sacked_out
== 0)
2013 tp
->fackets_out
= 0;
2015 /* Now state machine starts.
2016 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2018 tp
->prior_ssthresh
= 0;
2020 /* B. In all the states check for reneging SACKs. */
2021 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2024 /* C. Process data loss notification, provided it is valid. */
2025 if ((flag
&FLAG_DATA_LOST
) &&
2026 before(tp
->snd_una
, tp
->high_seq
) &&
2027 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2028 tp
->fackets_out
> tp
->reordering
) {
2029 tcp_mark_head_lost(sk
, tp
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
2030 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2033 /* D. Synchronize left_out to current state. */
2034 tcp_sync_left_out(tp
);
2036 /* E. Check state exit conditions. State can be terminated
2037 * when high_seq is ACKed. */
2038 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2039 BUG_TRAP(tp
->retrans_out
== 0);
2040 tp
->retrans_stamp
= 0;
2041 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2042 switch (icsk
->icsk_ca_state
) {
2044 icsk
->icsk_retransmits
= 0;
2045 if (tcp_try_undo_recovery(sk
, tp
))
2050 /* CWR is to be held something *above* high_seq
2051 * is ACKed for CWR bit to reach receiver. */
2052 if (tp
->snd_una
!= tp
->high_seq
) {
2053 tcp_complete_cwr(sk
);
2054 tcp_set_ca_state(sk
, TCP_CA_Open
);
2058 case TCP_CA_Disorder
:
2059 tcp_try_undo_dsack(sk
, tp
);
2060 if (!tp
->undo_marker
||
2061 /* For SACK case do not Open to allow to undo
2062 * catching for all duplicate ACKs. */
2063 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2064 tp
->undo_marker
= 0;
2065 tcp_set_ca_state(sk
, TCP_CA_Open
);
2069 case TCP_CA_Recovery
:
2071 tcp_reset_reno_sack(tp
);
2072 if (tcp_try_undo_recovery(sk
, tp
))
2074 tcp_complete_cwr(sk
);
2079 /* F. Process state. */
2080 switch (icsk
->icsk_ca_state
) {
2081 case TCP_CA_Recovery
:
2082 if (prior_snd_una
== tp
->snd_una
) {
2083 if (IsReno(tp
) && is_dupack
)
2084 tcp_add_reno_sack(sk
);
2086 int acked
= prior_packets
- tp
->packets_out
;
2088 tcp_remove_reno_sacks(sk
, tp
, acked
);
2089 is_dupack
= tcp_try_undo_partial(sk
, tp
, acked
);
2093 if (flag
&FLAG_DATA_ACKED
)
2094 icsk
->icsk_retransmits
= 0;
2095 if (!tcp_try_undo_loss(sk
, tp
)) {
2096 tcp_moderate_cwnd(tp
);
2097 tcp_xmit_retransmit_queue(sk
);
2100 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2102 /* Loss is undone; fall through to processing in Open state. */
2105 if (tp
->snd_una
!= prior_snd_una
)
2106 tcp_reset_reno_sack(tp
);
2108 tcp_add_reno_sack(sk
);
2111 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2112 tcp_try_undo_dsack(sk
, tp
);
2114 if (!tcp_time_to_recover(sk
, tp
)) {
2115 tcp_try_to_open(sk
, tp
, flag
);
2119 /* MTU probe failure: don't reduce cwnd */
2120 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2121 icsk
->icsk_mtup
.probe_size
&&
2122 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2123 tcp_mtup_probe_failed(sk
);
2124 /* Restores the reduction we did in tcp_mtup_probe() */
2126 tcp_simple_retransmit(sk
);
2130 /* Otherwise enter Recovery state */
2133 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2135 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2137 tp
->high_seq
= tp
->snd_nxt
;
2138 tp
->prior_ssthresh
= 0;
2139 tp
->undo_marker
= tp
->snd_una
;
2140 tp
->undo_retrans
= tp
->retrans_out
;
2142 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2143 if (!(flag
&FLAG_ECE
))
2144 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2145 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2146 TCP_ECN_queue_cwr(tp
);
2149 tp
->bytes_acked
= 0;
2150 tp
->snd_cwnd_cnt
= 0;
2151 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2154 if (is_dupack
|| tcp_head_timedout(sk
, tp
))
2155 tcp_update_scoreboard(sk
, tp
);
2157 tcp_xmit_retransmit_queue(sk
);
2160 /* Read draft-ietf-tcplw-high-performance before mucking
2161 * with this code. (Supersedes RFC1323)
2163 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2165 /* RTTM Rule: A TSecr value received in a segment is used to
2166 * update the averaged RTT measurement only if the segment
2167 * acknowledges some new data, i.e., only if it advances the
2168 * left edge of the send window.
2170 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2171 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2173 * Changed: reset backoff as soon as we see the first valid sample.
2174 * If we do not, we get strongly overestimated rto. With timestamps
2175 * samples are accepted even from very old segments: f.e., when rtt=1
2176 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2177 * answer arrives rto becomes 120 seconds! If at least one of segments
2178 * in window is lost... Voila. --ANK (010210)
2180 struct tcp_sock
*tp
= tcp_sk(sk
);
2181 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2182 tcp_rtt_estimator(sk
, seq_rtt
);
2184 inet_csk(sk
)->icsk_backoff
= 0;
2188 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2190 /* We don't have a timestamp. Can only use
2191 * packets that are not retransmitted to determine
2192 * rtt estimates. Also, we must not reset the
2193 * backoff for rto until we get a non-retransmitted
2194 * packet. This allows us to deal with a situation
2195 * where the network delay has increased suddenly.
2196 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2199 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2202 tcp_rtt_estimator(sk
, seq_rtt
);
2204 inet_csk(sk
)->icsk_backoff
= 0;
2208 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2211 const struct tcp_sock
*tp
= tcp_sk(sk
);
2212 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2213 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2214 tcp_ack_saw_tstamp(sk
, flag
);
2215 else if (seq_rtt
>= 0)
2216 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2219 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
2220 u32 in_flight
, int good
)
2222 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2223 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
2224 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2227 /* Restart timer after forward progress on connection.
2228 * RFC2988 recommends to restart timer to now+rto.
2231 static void tcp_ack_packets_out(struct sock
*sk
, struct tcp_sock
*tp
)
2233 if (!tp
->packets_out
) {
2234 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2236 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2240 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2241 __u32 now
, __s32
*seq_rtt
)
2243 struct tcp_sock
*tp
= tcp_sk(sk
);
2244 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2245 __u32 seq
= tp
->snd_una
;
2246 __u32 packets_acked
;
2249 /* If we get here, the whole TSO packet has not been
2252 BUG_ON(!after(scb
->end_seq
, seq
));
2254 packets_acked
= tcp_skb_pcount(skb
);
2255 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2257 packets_acked
-= tcp_skb_pcount(skb
);
2259 if (packets_acked
) {
2260 __u8 sacked
= scb
->sacked
;
2262 acked
|= FLAG_DATA_ACKED
;
2264 if (sacked
& TCPCB_RETRANS
) {
2265 if (sacked
& TCPCB_SACKED_RETRANS
)
2266 tp
->retrans_out
-= packets_acked
;
2267 acked
|= FLAG_RETRANS_DATA_ACKED
;
2269 } else if (*seq_rtt
< 0)
2270 *seq_rtt
= now
- scb
->when
;
2271 if (sacked
& TCPCB_SACKED_ACKED
)
2272 tp
->sacked_out
-= packets_acked
;
2273 if (sacked
& TCPCB_LOST
)
2274 tp
->lost_out
-= packets_acked
;
2275 if (sacked
& TCPCB_URG
) {
2277 !before(seq
, tp
->snd_up
))
2280 } else if (*seq_rtt
< 0)
2281 *seq_rtt
= now
- scb
->when
;
2283 if (tp
->fackets_out
) {
2284 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2285 tp
->fackets_out
-= dval
;
2287 tp
->packets_out
-= packets_acked
;
2289 BUG_ON(tcp_skb_pcount(skb
) == 0);
2290 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2296 static u32
tcp_usrtt(struct timeval
*tv
)
2300 do_gettimeofday(&now
);
2301 return (now
.tv_sec
- tv
->tv_sec
) * 1000000 + (now
.tv_usec
- tv
->tv_usec
);
2304 /* Remove acknowledged frames from the retransmission queue. */
2305 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2307 struct tcp_sock
*tp
= tcp_sk(sk
);
2308 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2309 struct sk_buff
*skb
;
2310 __u32 now
= tcp_time_stamp
;
2314 void (*rtt_sample
)(struct sock
*sk
, u32 usrtt
)
2315 = icsk
->icsk_ca_ops
->rtt_sample
;
2316 struct timeval tv
= { .tv_sec
= 0, .tv_usec
= 0 };
2318 while ((skb
= skb_peek(&sk
->sk_write_queue
)) &&
2319 skb
!= sk
->sk_send_head
) {
2320 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2321 __u8 sacked
= scb
->sacked
;
2323 /* If our packet is before the ack sequence we can
2324 * discard it as it's confirmed to have arrived at
2327 if (after(scb
->end_seq
, tp
->snd_una
)) {
2328 if (tcp_skb_pcount(skb
) > 1 &&
2329 after(tp
->snd_una
, scb
->seq
))
2330 acked
|= tcp_tso_acked(sk
, skb
,
2335 /* Initial outgoing SYN's get put onto the write_queue
2336 * just like anything else we transmit. It is not
2337 * true data, and if we misinform our callers that
2338 * this ACK acks real data, we will erroneously exit
2339 * connection startup slow start one packet too
2340 * quickly. This is severely frowned upon behavior.
2342 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2343 acked
|= FLAG_DATA_ACKED
;
2346 acked
|= FLAG_SYN_ACKED
;
2347 tp
->retrans_stamp
= 0;
2350 /* MTU probing checks */
2351 if (icsk
->icsk_mtup
.probe_size
) {
2352 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2353 tcp_mtup_probe_success(sk
, skb
);
2358 if (sacked
& TCPCB_RETRANS
) {
2359 if(sacked
& TCPCB_SACKED_RETRANS
)
2360 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2361 acked
|= FLAG_RETRANS_DATA_ACKED
;
2363 } else if (seq_rtt
< 0) {
2364 seq_rtt
= now
- scb
->when
;
2365 skb_get_timestamp(skb
, &tv
);
2367 if (sacked
& TCPCB_SACKED_ACKED
)
2368 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2369 if (sacked
& TCPCB_LOST
)
2370 tp
->lost_out
-= tcp_skb_pcount(skb
);
2371 if (sacked
& TCPCB_URG
) {
2373 !before(scb
->end_seq
, tp
->snd_up
))
2376 } else if (seq_rtt
< 0) {
2377 seq_rtt
= now
- scb
->when
;
2378 skb_get_timestamp(skb
, &tv
);
2380 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2381 tcp_packets_out_dec(tp
, skb
);
2382 __skb_unlink(skb
, &sk
->sk_write_queue
);
2383 sk_stream_free_skb(sk
, skb
);
2384 clear_all_retrans_hints(tp
);
2387 if (acked
&FLAG_ACKED
) {
2388 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2389 tcp_ack_packets_out(sk
, tp
);
2390 if (rtt_sample
&& !(acked
& FLAG_RETRANS_DATA_ACKED
))
2391 (*rtt_sample
)(sk
, tcp_usrtt(&tv
));
2393 if (icsk
->icsk_ca_ops
->pkts_acked
)
2394 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
);
2397 #if FASTRETRANS_DEBUG > 0
2398 BUG_TRAP((int)tp
->sacked_out
>= 0);
2399 BUG_TRAP((int)tp
->lost_out
>= 0);
2400 BUG_TRAP((int)tp
->retrans_out
>= 0);
2401 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2402 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2404 printk(KERN_DEBUG
"Leak l=%u %d\n",
2405 tp
->lost_out
, icsk
->icsk_ca_state
);
2408 if (tp
->sacked_out
) {
2409 printk(KERN_DEBUG
"Leak s=%u %d\n",
2410 tp
->sacked_out
, icsk
->icsk_ca_state
);
2413 if (tp
->retrans_out
) {
2414 printk(KERN_DEBUG
"Leak r=%u %d\n",
2415 tp
->retrans_out
, icsk
->icsk_ca_state
);
2416 tp
->retrans_out
= 0;
2420 *seq_rtt_p
= seq_rtt
;
2424 static void tcp_ack_probe(struct sock
*sk
)
2426 const struct tcp_sock
*tp
= tcp_sk(sk
);
2427 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2429 /* Was it a usable window open? */
2431 if (!after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
2432 tp
->snd_una
+ tp
->snd_wnd
)) {
2433 icsk
->icsk_backoff
= 0;
2434 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2435 /* Socket must be waked up by subsequent tcp_data_snd_check().
2436 * This function is not for random using!
2439 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2440 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2445 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2447 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2448 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2451 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2453 const struct tcp_sock
*tp
= tcp_sk(sk
);
2454 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2455 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2458 /* Check that window update is acceptable.
2459 * The function assumes that snd_una<=ack<=snd_next.
2461 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2462 const u32 ack_seq
, const u32 nwin
)
2464 return (after(ack
, tp
->snd_una
) ||
2465 after(ack_seq
, tp
->snd_wl1
) ||
2466 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2469 /* Update our send window.
2471 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2472 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2474 static int tcp_ack_update_window(struct sock
*sk
, struct tcp_sock
*tp
,
2475 struct sk_buff
*skb
, u32 ack
, u32 ack_seq
)
2478 u32 nwin
= ntohs(skb
->h
.th
->window
);
2480 if (likely(!skb
->h
.th
->syn
))
2481 nwin
<<= tp
->rx_opt
.snd_wscale
;
2483 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2484 flag
|= FLAG_WIN_UPDATE
;
2485 tcp_update_wl(tp
, ack
, ack_seq
);
2487 if (tp
->snd_wnd
!= nwin
) {
2490 /* Note, it is the only place, where
2491 * fast path is recovered for sending TCP.
2494 tcp_fast_path_check(sk
, tp
);
2496 if (nwin
> tp
->max_window
) {
2497 tp
->max_window
= nwin
;
2498 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2508 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2509 * continue in congestion avoidance.
2511 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2513 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2514 tp
->snd_cwnd_cnt
= 0;
2515 tcp_moderate_cwnd(tp
);
2518 /* F-RTO spurious RTO detection algorithm (RFC4138)
2520 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2521 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2522 * window (but not to or beyond highest sequence sent before RTO):
2523 * On First ACK, send two new segments out.
2524 * On Second ACK, RTO was likely spurious. Do spurious response (response
2525 * algorithm is not part of the F-RTO detection algorithm
2526 * given in RFC4138 but can be selected separately).
2527 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2528 * and TCP falls back to conventional RTO recovery.
2530 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2531 * original window even after we transmit two new data segments.
2533 * F-RTO is implemented (mainly) in four functions:
2534 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2535 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2536 * called when tcp_use_frto() showed green light
2537 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2538 * - tcp_enter_frto_loss() is called if there is not enough evidence
2539 * to prove that the RTO is indeed spurious. It transfers the control
2540 * from F-RTO to the conventional RTO recovery
2542 static int tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
, int flag
)
2544 struct tcp_sock
*tp
= tcp_sk(sk
);
2546 tcp_sync_left_out(tp
);
2548 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2549 if (flag
&FLAG_DATA_ACKED
)
2550 inet_csk(sk
)->icsk_retransmits
= 0;
2552 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2553 tcp_enter_frto_loss(sk
, tp
->frto_counter
+ 1);
2557 /* RFC4138 shortcoming in step 2; should also have case c): ACK isn't
2558 * duplicate nor advances window, e.g., opposite dir data, winupdate
2560 if ((tp
->snd_una
== prior_snd_una
) && (flag
&FLAG_NOT_DUP
) &&
2561 !(flag
&FLAG_FORWARD_PROGRESS
))
2564 if (!(flag
&FLAG_DATA_ACKED
)) {
2565 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3));
2569 if (tp
->frto_counter
== 1) {
2570 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2571 tp
->frto_counter
= 2;
2573 } else /* frto_counter == 2 */ {
2574 tcp_conservative_spur_to_response(tp
);
2575 tp
->frto_counter
= 0;
2580 /* This routine deals with incoming acks, but not outgoing ones. */
2581 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2583 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2584 struct tcp_sock
*tp
= tcp_sk(sk
);
2585 u32 prior_snd_una
= tp
->snd_una
;
2586 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2587 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2588 u32 prior_in_flight
;
2593 /* If the ack is newer than sent or older than previous acks
2594 * then we can probably ignore it.
2596 if (after(ack
, tp
->snd_nxt
))
2597 goto uninteresting_ack
;
2599 if (before(ack
, prior_snd_una
))
2602 if (sysctl_tcp_abc
) {
2603 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2604 tp
->bytes_acked
+= ack
- prior_snd_una
;
2605 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2606 /* we assume just one segment left network */
2607 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2610 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2611 /* Window is constant, pure forward advance.
2612 * No more checks are required.
2613 * Note, we use the fact that SND.UNA>=SND.WL2.
2615 tcp_update_wl(tp
, ack
, ack_seq
);
2617 flag
|= FLAG_WIN_UPDATE
;
2619 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2621 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2623 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2626 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2628 flag
|= tcp_ack_update_window(sk
, tp
, skb
, ack
, ack_seq
);
2630 if (TCP_SKB_CB(skb
)->sacked
)
2631 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2633 if (TCP_ECN_rcv_ecn_echo(tp
, skb
->h
.th
))
2636 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2639 /* We passed data and got it acked, remove any soft error
2640 * log. Something worked...
2642 sk
->sk_err_soft
= 0;
2643 tp
->rcv_tstamp
= tcp_time_stamp
;
2644 prior_packets
= tp
->packets_out
;
2648 prior_in_flight
= tcp_packets_in_flight(tp
);
2650 /* See if we can take anything off of the retransmit queue. */
2651 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2653 if (tp
->frto_counter
)
2654 frto_cwnd
= tcp_process_frto(sk
, prior_snd_una
, flag
);
2656 if (tcp_ack_is_dubious(sk
, flag
)) {
2657 /* Advance CWND, if state allows this. */
2658 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
2659 tcp_may_raise_cwnd(sk
, flag
))
2660 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2661 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2663 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
2664 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2667 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2668 dst_confirm(sk
->sk_dst_cache
);
2673 icsk
->icsk_probes_out
= 0;
2675 /* If this ack opens up a zero window, clear backoff. It was
2676 * being used to time the probes, and is probably far higher than
2677 * it needs to be for normal retransmission.
2679 if (sk
->sk_send_head
)
2684 if (TCP_SKB_CB(skb
)->sacked
)
2685 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2688 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2693 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2694 * But, this can also be called on packets in the established flow when
2695 * the fast version below fails.
2697 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2700 struct tcphdr
*th
= skb
->h
.th
;
2701 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2703 ptr
= (unsigned char *)(th
+ 1);
2704 opt_rx
->saw_tstamp
= 0;
2713 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2718 if (opsize
< 2) /* "silly options" */
2720 if (opsize
> length
)
2721 return; /* don't parse partial options */
2724 if(opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2725 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
2727 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2728 in_mss
= opt_rx
->user_mss
;
2729 opt_rx
->mss_clamp
= in_mss
;
2734 if(opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2735 if (sysctl_tcp_window_scaling
) {
2736 __u8 snd_wscale
= *(__u8
*) ptr
;
2737 opt_rx
->wscale_ok
= 1;
2738 if (snd_wscale
> 14) {
2740 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2741 "scaling value %d >14 received.\n",
2745 opt_rx
->snd_wscale
= snd_wscale
;
2748 case TCPOPT_TIMESTAMP
:
2749 if(opsize
==TCPOLEN_TIMESTAMP
) {
2750 if ((estab
&& opt_rx
->tstamp_ok
) ||
2751 (!estab
&& sysctl_tcp_timestamps
)) {
2752 opt_rx
->saw_tstamp
= 1;
2753 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
2754 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
2758 case TCPOPT_SACK_PERM
:
2759 if(opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2760 if (sysctl_tcp_sack
) {
2761 opt_rx
->sack_ok
= 1;
2762 tcp_sack_reset(opt_rx
);
2768 if((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2769 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2771 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2773 #ifdef CONFIG_TCP_MD5SIG
2776 * The MD5 Hash has already been
2777 * checked (see tcp_v{4,6}_do_rcv()).
2788 /* Fast parse options. This hopes to only see timestamps.
2789 * If it is wrong it falls back on tcp_parse_options().
2791 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2792 struct tcp_sock
*tp
)
2794 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2795 tp
->rx_opt
.saw_tstamp
= 0;
2797 } else if (tp
->rx_opt
.tstamp_ok
&&
2798 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2799 __be32
*ptr
= (__be32
*)(th
+ 1);
2800 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2801 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2802 tp
->rx_opt
.saw_tstamp
= 1;
2804 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2806 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2810 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2814 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2816 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2817 tp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
2820 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2822 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2823 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2824 * extra check below makes sure this can only happen
2825 * for pure ACK frames. -DaveM
2827 * Not only, also it occurs for expired timestamps.
2830 if((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2831 xtime
.tv_sec
>= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2832 tcp_store_ts_recent(tp
);
2836 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2838 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2839 * it can pass through stack. So, the following predicate verifies that
2840 * this segment is not used for anything but congestion avoidance or
2841 * fast retransmit. Moreover, we even are able to eliminate most of such
2842 * second order effects, if we apply some small "replay" window (~RTO)
2843 * to timestamp space.
2845 * All these measures still do not guarantee that we reject wrapped ACKs
2846 * on networks with high bandwidth, when sequence space is recycled fastly,
2847 * but it guarantees that such events will be very rare and do not affect
2848 * connection seriously. This doesn't look nice, but alas, PAWS is really
2851 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2852 * states that events when retransmit arrives after original data are rare.
2853 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2854 * the biggest problem on large power networks even with minor reordering.
2855 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2856 * up to bandwidth of 18Gigabit/sec. 8) ]
2859 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
2861 struct tcp_sock
*tp
= tcp_sk(sk
);
2862 struct tcphdr
*th
= skb
->h
.th
;
2863 u32 seq
= TCP_SKB_CB(skb
)->seq
;
2864 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2866 return (/* 1. Pure ACK with correct sequence number. */
2867 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
2869 /* 2. ... and duplicate ACK. */
2870 ack
== tp
->snd_una
&&
2872 /* 3. ... and does not update window. */
2873 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
2875 /* 4. ... and sits in replay window. */
2876 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
2879 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
2881 const struct tcp_sock
*tp
= tcp_sk(sk
);
2882 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
2883 xtime
.tv_sec
< tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
2884 !tcp_disordered_ack(sk
, skb
));
2887 /* Check segment sequence number for validity.
2889 * Segment controls are considered valid, if the segment
2890 * fits to the window after truncation to the window. Acceptability
2891 * of data (and SYN, FIN, of course) is checked separately.
2892 * See tcp_data_queue(), for example.
2894 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2895 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2896 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2897 * (borrowed from freebsd)
2900 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2902 return !before(end_seq
, tp
->rcv_wup
) &&
2903 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
2906 /* When we get a reset we do this. */
2907 static void tcp_reset(struct sock
*sk
)
2909 /* We want the right error as BSD sees it (and indeed as we do). */
2910 switch (sk
->sk_state
) {
2912 sk
->sk_err
= ECONNREFUSED
;
2914 case TCP_CLOSE_WAIT
:
2920 sk
->sk_err
= ECONNRESET
;
2923 if (!sock_flag(sk
, SOCK_DEAD
))
2924 sk
->sk_error_report(sk
);
2930 * Process the FIN bit. This now behaves as it is supposed to work
2931 * and the FIN takes effect when it is validly part of sequence
2932 * space. Not before when we get holes.
2934 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2935 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2938 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2939 * close and we go into CLOSING (and later onto TIME-WAIT)
2941 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2943 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
2945 struct tcp_sock
*tp
= tcp_sk(sk
);
2947 inet_csk_schedule_ack(sk
);
2949 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
2950 sock_set_flag(sk
, SOCK_DONE
);
2952 switch (sk
->sk_state
) {
2954 case TCP_ESTABLISHED
:
2955 /* Move to CLOSE_WAIT */
2956 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
2957 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
2960 case TCP_CLOSE_WAIT
:
2962 /* Received a retransmission of the FIN, do
2967 /* RFC793: Remain in the LAST-ACK state. */
2971 /* This case occurs when a simultaneous close
2972 * happens, we must ack the received FIN and
2973 * enter the CLOSING state.
2976 tcp_set_state(sk
, TCP_CLOSING
);
2979 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2981 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
2984 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2985 * cases we should never reach this piece of code.
2987 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
2988 __FUNCTION__
, sk
->sk_state
);
2992 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2993 * Probably, we should reset in this case. For now drop them.
2995 __skb_queue_purge(&tp
->out_of_order_queue
);
2996 if (tp
->rx_opt
.sack_ok
)
2997 tcp_sack_reset(&tp
->rx_opt
);
2998 sk_stream_mem_reclaim(sk
);
3000 if (!sock_flag(sk
, SOCK_DEAD
)) {
3001 sk
->sk_state_change(sk
);
3003 /* Do not send POLL_HUP for half duplex close. */
3004 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3005 sk
->sk_state
== TCP_CLOSE
)
3006 sk_wake_async(sk
, 1, POLL_HUP
);
3008 sk_wake_async(sk
, 1, POLL_IN
);
3012 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3014 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3015 if (before(seq
, sp
->start_seq
))
3016 sp
->start_seq
= seq
;
3017 if (after(end_seq
, sp
->end_seq
))
3018 sp
->end_seq
= end_seq
;
3024 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3026 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3027 if (before(seq
, tp
->rcv_nxt
))
3028 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3030 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3032 tp
->rx_opt
.dsack
= 1;
3033 tp
->duplicate_sack
[0].start_seq
= seq
;
3034 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3035 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3039 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3041 if (!tp
->rx_opt
.dsack
)
3042 tcp_dsack_set(tp
, seq
, end_seq
);
3044 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3047 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3049 struct tcp_sock
*tp
= tcp_sk(sk
);
3051 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3052 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3053 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3054 tcp_enter_quickack_mode(sk
);
3056 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3057 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3059 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3060 end_seq
= tp
->rcv_nxt
;
3061 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3068 /* These routines update the SACK block as out-of-order packets arrive or
3069 * in-order packets close up the sequence space.
3071 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3074 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3075 struct tcp_sack_block
*swalk
= sp
+1;
3077 /* See if the recent change to the first SACK eats into
3078 * or hits the sequence space of other SACK blocks, if so coalesce.
3080 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3081 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3084 /* Zap SWALK, by moving every further SACK up by one slot.
3085 * Decrease num_sacks.
3087 tp
->rx_opt
.num_sacks
--;
3088 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3089 for(i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3093 this_sack
++, swalk
++;
3097 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3101 tmp
= sack1
->start_seq
;
3102 sack1
->start_seq
= sack2
->start_seq
;
3103 sack2
->start_seq
= tmp
;
3105 tmp
= sack1
->end_seq
;
3106 sack1
->end_seq
= sack2
->end_seq
;
3107 sack2
->end_seq
= tmp
;
3110 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3112 struct tcp_sock
*tp
= tcp_sk(sk
);
3113 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3114 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3120 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3121 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3122 /* Rotate this_sack to the first one. */
3123 for (; this_sack
>0; this_sack
--, sp
--)
3124 tcp_sack_swap(sp
, sp
-1);
3126 tcp_sack_maybe_coalesce(tp
);
3131 /* Could not find an adjacent existing SACK, build a new one,
3132 * put it at the front, and shift everyone else down. We
3133 * always know there is at least one SACK present already here.
3135 * If the sack array is full, forget about the last one.
3137 if (this_sack
>= 4) {
3139 tp
->rx_opt
.num_sacks
--;
3142 for(; this_sack
> 0; this_sack
--, sp
--)
3146 /* Build the new head SACK, and we're done. */
3147 sp
->start_seq
= seq
;
3148 sp
->end_seq
= end_seq
;
3149 tp
->rx_opt
.num_sacks
++;
3150 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3153 /* RCV.NXT advances, some SACKs should be eaten. */
3155 static void tcp_sack_remove(struct tcp_sock
*tp
)
3157 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3158 int num_sacks
= tp
->rx_opt
.num_sacks
;
3161 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3162 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3163 tp
->rx_opt
.num_sacks
= 0;
3164 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3168 for(this_sack
= 0; this_sack
< num_sacks
; ) {
3169 /* Check if the start of the sack is covered by RCV.NXT. */
3170 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3173 /* RCV.NXT must cover all the block! */
3174 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3176 /* Zap this SACK, by moving forward any other SACKS. */
3177 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3178 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3185 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3186 tp
->rx_opt
.num_sacks
= num_sacks
;
3187 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3191 /* This one checks to see if we can put data from the
3192 * out_of_order queue into the receive_queue.
3194 static void tcp_ofo_queue(struct sock
*sk
)
3196 struct tcp_sock
*tp
= tcp_sk(sk
);
3197 __u32 dsack_high
= tp
->rcv_nxt
;
3198 struct sk_buff
*skb
;
3200 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3201 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3204 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3205 __u32 dsack
= dsack_high
;
3206 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3207 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3208 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3211 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3212 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3213 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3217 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3218 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3219 TCP_SKB_CB(skb
)->end_seq
);
3221 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3222 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3223 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3225 tcp_fin(skb
, sk
, skb
->h
.th
);
3229 static int tcp_prune_queue(struct sock
*sk
);
3231 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3233 struct tcphdr
*th
= skb
->h
.th
;
3234 struct tcp_sock
*tp
= tcp_sk(sk
);
3237 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3240 __skb_pull(skb
, th
->doff
*4);
3242 TCP_ECN_accept_cwr(tp
, skb
);
3244 if (tp
->rx_opt
.dsack
) {
3245 tp
->rx_opt
.dsack
= 0;
3246 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3247 4 - tp
->rx_opt
.tstamp_ok
);
3250 /* Queue data for delivery to the user.
3251 * Packets in sequence go to the receive queue.
3252 * Out of sequence packets to the out_of_order_queue.
3254 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3255 if (tcp_receive_window(tp
) == 0)
3258 /* Ok. In sequence. In window. */
3259 if (tp
->ucopy
.task
== current
&&
3260 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3261 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3262 int chunk
= min_t(unsigned int, skb
->len
,
3265 __set_current_state(TASK_RUNNING
);
3268 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3269 tp
->ucopy
.len
-= chunk
;
3270 tp
->copied_seq
+= chunk
;
3271 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3272 tcp_rcv_space_adjust(sk
);
3280 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3281 !sk_stream_rmem_schedule(sk
, skb
))) {
3282 if (tcp_prune_queue(sk
) < 0 ||
3283 !sk_stream_rmem_schedule(sk
, skb
))
3286 sk_stream_set_owner_r(skb
, sk
);
3287 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3289 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3291 tcp_event_data_recv(sk
, tp
, skb
);
3293 tcp_fin(skb
, sk
, th
);
3295 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3298 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3299 * gap in queue is filled.
3301 if (skb_queue_empty(&tp
->out_of_order_queue
))
3302 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3305 if (tp
->rx_opt
.num_sacks
)
3306 tcp_sack_remove(tp
);
3308 tcp_fast_path_check(sk
, tp
);
3312 else if (!sock_flag(sk
, SOCK_DEAD
))
3313 sk
->sk_data_ready(sk
, 0);
3317 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3318 /* A retransmit, 2nd most common case. Force an immediate ack. */
3319 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3320 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3323 tcp_enter_quickack_mode(sk
);
3324 inet_csk_schedule_ack(sk
);
3330 /* Out of window. F.e. zero window probe. */
3331 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3334 tcp_enter_quickack_mode(sk
);
3336 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3337 /* Partial packet, seq < rcv_next < end_seq */
3338 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3339 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3340 TCP_SKB_CB(skb
)->end_seq
);
3342 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3344 /* If window is closed, drop tail of packet. But after
3345 * remembering D-SACK for its head made in previous line.
3347 if (!tcp_receive_window(tp
))
3352 TCP_ECN_check_ce(tp
, skb
);
3354 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3355 !sk_stream_rmem_schedule(sk
, skb
)) {
3356 if (tcp_prune_queue(sk
) < 0 ||
3357 !sk_stream_rmem_schedule(sk
, skb
))
3361 /* Disable header prediction. */
3363 inet_csk_schedule_ack(sk
);
3365 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3366 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3368 sk_stream_set_owner_r(skb
, sk
);
3370 if (!skb_peek(&tp
->out_of_order_queue
)) {
3371 /* Initial out of order segment, build 1 SACK. */
3372 if (tp
->rx_opt
.sack_ok
) {
3373 tp
->rx_opt
.num_sacks
= 1;
3374 tp
->rx_opt
.dsack
= 0;
3375 tp
->rx_opt
.eff_sacks
= 1;
3376 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3377 tp
->selective_acks
[0].end_seq
=
3378 TCP_SKB_CB(skb
)->end_seq
;
3380 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3382 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3383 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3384 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3386 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3387 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3389 if (!tp
->rx_opt
.num_sacks
||
3390 tp
->selective_acks
[0].end_seq
!= seq
)
3393 /* Common case: data arrive in order after hole. */
3394 tp
->selective_acks
[0].end_seq
= end_seq
;
3398 /* Find place to insert this segment. */
3400 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3402 } while ((skb1
= skb1
->prev
) !=
3403 (struct sk_buff
*)&tp
->out_of_order_queue
);
3405 /* Do skb overlap to previous one? */
3406 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3407 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3408 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3409 /* All the bits are present. Drop. */
3411 tcp_dsack_set(tp
, seq
, end_seq
);
3414 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3415 /* Partial overlap. */
3416 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3421 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3423 /* And clean segments covered by new one as whole. */
3424 while ((skb1
= skb
->next
) !=
3425 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3426 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3427 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3428 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3431 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3432 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3437 if (tp
->rx_opt
.sack_ok
)
3438 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3442 /* Collapse contiguous sequence of skbs head..tail with
3443 * sequence numbers start..end.
3444 * Segments with FIN/SYN are not collapsed (only because this
3448 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3449 struct sk_buff
*head
, struct sk_buff
*tail
,
3452 struct sk_buff
*skb
;
3454 /* First, check that queue is collapsible and find
3455 * the point where collapsing can be useful. */
3456 for (skb
= head
; skb
!= tail
; ) {
3457 /* No new bits? It is possible on ofo queue. */
3458 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3459 struct sk_buff
*next
= skb
->next
;
3460 __skb_unlink(skb
, list
);
3462 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3467 /* The first skb to collapse is:
3469 * - bloated or contains data before "start" or
3470 * overlaps to the next one.
3472 if (!skb
->h
.th
->syn
&& !skb
->h
.th
->fin
&&
3473 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3474 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3475 (skb
->next
!= tail
&&
3476 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3479 /* Decided to skip this, advance start seq. */
3480 start
= TCP_SKB_CB(skb
)->end_seq
;
3483 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3486 while (before(start
, end
)) {
3487 struct sk_buff
*nskb
;
3488 int header
= skb_headroom(skb
);
3489 int copy
= SKB_MAX_ORDER(header
, 0);
3491 /* Too big header? This can happen with IPv6. */
3494 if (end
-start
< copy
)
3496 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3499 skb_reserve(nskb
, header
);
3500 memcpy(nskb
->head
, skb
->head
, header
);
3501 nskb
->nh
.raw
= nskb
->head
+ (skb
->nh
.raw
-skb
->head
);
3502 nskb
->h
.raw
= nskb
->head
+ (skb
->h
.raw
-skb
->head
);
3503 nskb
->mac
.raw
= nskb
->head
+ (skb
->mac
.raw
-skb
->head
);
3504 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3505 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3506 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3507 sk_stream_set_owner_r(nskb
, sk
);
3509 /* Copy data, releasing collapsed skbs. */
3511 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3512 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3516 size
= min(copy
, size
);
3517 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3519 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3523 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3524 struct sk_buff
*next
= skb
->next
;
3525 __skb_unlink(skb
, list
);
3527 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3529 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3536 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3537 * and tcp_collapse() them until all the queue is collapsed.
3539 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3541 struct tcp_sock
*tp
= tcp_sk(sk
);
3542 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3543 struct sk_buff
*head
;
3549 start
= TCP_SKB_CB(skb
)->seq
;
3550 end
= TCP_SKB_CB(skb
)->end_seq
;
3556 /* Segment is terminated when we see gap or when
3557 * we are at the end of all the queue. */
3558 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3559 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3560 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3561 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3562 head
, skb
, start
, end
);
3564 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3566 /* Start new segment */
3567 start
= TCP_SKB_CB(skb
)->seq
;
3568 end
= TCP_SKB_CB(skb
)->end_seq
;
3570 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3571 start
= TCP_SKB_CB(skb
)->seq
;
3572 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3573 end
= TCP_SKB_CB(skb
)->end_seq
;
3578 /* Reduce allocated memory if we can, trying to get
3579 * the socket within its memory limits again.
3581 * Return less than zero if we should start dropping frames
3582 * until the socket owning process reads some of the data
3583 * to stabilize the situation.
3585 static int tcp_prune_queue(struct sock
*sk
)
3587 struct tcp_sock
*tp
= tcp_sk(sk
);
3589 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3591 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3593 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3594 tcp_clamp_window(sk
, tp
);
3595 else if (tcp_memory_pressure
)
3596 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3598 tcp_collapse_ofo_queue(sk
);
3599 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3600 sk
->sk_receive_queue
.next
,
3601 (struct sk_buff
*)&sk
->sk_receive_queue
,
3602 tp
->copied_seq
, tp
->rcv_nxt
);
3603 sk_stream_mem_reclaim(sk
);
3605 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3608 /* Collapsing did not help, destructive actions follow.
3609 * This must not ever occur. */
3611 /* First, purge the out_of_order queue. */
3612 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3613 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3614 __skb_queue_purge(&tp
->out_of_order_queue
);
3616 /* Reset SACK state. A conforming SACK implementation will
3617 * do the same at a timeout based retransmit. When a connection
3618 * is in a sad state like this, we care only about integrity
3619 * of the connection not performance.
3621 if (tp
->rx_opt
.sack_ok
)
3622 tcp_sack_reset(&tp
->rx_opt
);
3623 sk_stream_mem_reclaim(sk
);
3626 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3629 /* If we are really being abused, tell the caller to silently
3630 * drop receive data on the floor. It will get retransmitted
3631 * and hopefully then we'll have sufficient space.
3633 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3635 /* Massive buffer overcommit. */
3641 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3642 * As additional protections, we do not touch cwnd in retransmission phases,
3643 * and if application hit its sndbuf limit recently.
3645 void tcp_cwnd_application_limited(struct sock
*sk
)
3647 struct tcp_sock
*tp
= tcp_sk(sk
);
3649 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3650 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3651 /* Limited by application or receiver window. */
3652 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
3653 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
3654 if (win_used
< tp
->snd_cwnd
) {
3655 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3656 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3658 tp
->snd_cwnd_used
= 0;
3660 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3663 static int tcp_should_expand_sndbuf(struct sock
*sk
, struct tcp_sock
*tp
)
3665 /* If the user specified a specific send buffer setting, do
3668 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3671 /* If we are under global TCP memory pressure, do not expand. */
3672 if (tcp_memory_pressure
)
3675 /* If we are under soft global TCP memory pressure, do not expand. */
3676 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3679 /* If we filled the congestion window, do not expand. */
3680 if (tp
->packets_out
>= tp
->snd_cwnd
)
3686 /* When incoming ACK allowed to free some skb from write_queue,
3687 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3688 * on the exit from tcp input handler.
3690 * PROBLEM: sndbuf expansion does not work well with largesend.
3692 static void tcp_new_space(struct sock
*sk
)
3694 struct tcp_sock
*tp
= tcp_sk(sk
);
3696 if (tcp_should_expand_sndbuf(sk
, tp
)) {
3697 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3698 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3699 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3700 tp
->reordering
+ 1);
3701 sndmem
*= 2*demanded
;
3702 if (sndmem
> sk
->sk_sndbuf
)
3703 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3704 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3707 sk
->sk_write_space(sk
);
3710 static void tcp_check_space(struct sock
*sk
)
3712 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3713 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3714 if (sk
->sk_socket
&&
3715 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3720 static inline void tcp_data_snd_check(struct sock
*sk
, struct tcp_sock
*tp
)
3722 tcp_push_pending_frames(sk
, tp
);
3723 tcp_check_space(sk
);
3727 * Check if sending an ack is needed.
3729 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3731 struct tcp_sock
*tp
= tcp_sk(sk
);
3733 /* More than one full frame received... */
3734 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3735 /* ... and right edge of window advances far enough.
3736 * (tcp_recvmsg() will send ACK otherwise). Or...
3738 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3739 /* We ACK each frame or... */
3740 tcp_in_quickack_mode(sk
) ||
3741 /* We have out of order data. */
3743 skb_peek(&tp
->out_of_order_queue
))) {
3744 /* Then ack it now */
3747 /* Else, send delayed ack. */
3748 tcp_send_delayed_ack(sk
);
3752 static inline void tcp_ack_snd_check(struct sock
*sk
)
3754 if (!inet_csk_ack_scheduled(sk
)) {
3755 /* We sent a data segment already. */
3758 __tcp_ack_snd_check(sk
, 1);
3762 * This routine is only called when we have urgent data
3763 * signaled. Its the 'slow' part of tcp_urg. It could be
3764 * moved inline now as tcp_urg is only called from one
3765 * place. We handle URGent data wrong. We have to - as
3766 * BSD still doesn't use the correction from RFC961.
3767 * For 1003.1g we should support a new option TCP_STDURG to permit
3768 * either form (or just set the sysctl tcp_stdurg).
3771 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3773 struct tcp_sock
*tp
= tcp_sk(sk
);
3774 u32 ptr
= ntohs(th
->urg_ptr
);
3776 if (ptr
&& !sysctl_tcp_stdurg
)
3778 ptr
+= ntohl(th
->seq
);
3780 /* Ignore urgent data that we've already seen and read. */
3781 if (after(tp
->copied_seq
, ptr
))
3784 /* Do not replay urg ptr.
3786 * NOTE: interesting situation not covered by specs.
3787 * Misbehaving sender may send urg ptr, pointing to segment,
3788 * which we already have in ofo queue. We are not able to fetch
3789 * such data and will stay in TCP_URG_NOTYET until will be eaten
3790 * by recvmsg(). Seems, we are not obliged to handle such wicked
3791 * situations. But it is worth to think about possibility of some
3792 * DoSes using some hypothetical application level deadlock.
3794 if (before(ptr
, tp
->rcv_nxt
))
3797 /* Do we already have a newer (or duplicate) urgent pointer? */
3798 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3801 /* Tell the world about our new urgent pointer. */
3804 /* We may be adding urgent data when the last byte read was
3805 * urgent. To do this requires some care. We cannot just ignore
3806 * tp->copied_seq since we would read the last urgent byte again
3807 * as data, nor can we alter copied_seq until this data arrives
3808 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3810 * NOTE. Double Dutch. Rendering to plain English: author of comment
3811 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3812 * and expect that both A and B disappear from stream. This is _wrong_.
3813 * Though this happens in BSD with high probability, this is occasional.
3814 * Any application relying on this is buggy. Note also, that fix "works"
3815 * only in this artificial test. Insert some normal data between A and B and we will
3816 * decline of BSD again. Verdict: it is better to remove to trap
3819 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3820 !sock_flag(sk
, SOCK_URGINLINE
) &&
3821 tp
->copied_seq
!= tp
->rcv_nxt
) {
3822 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3824 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3825 __skb_unlink(skb
, &sk
->sk_receive_queue
);
3830 tp
->urg_data
= TCP_URG_NOTYET
;
3833 /* Disable header prediction. */
3837 /* This is the 'fast' part of urgent handling. */
3838 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
3840 struct tcp_sock
*tp
= tcp_sk(sk
);
3842 /* Check if we get a new urgent pointer - normally not. */
3844 tcp_check_urg(sk
,th
);
3846 /* Do we wait for any urgent data? - normally not... */
3847 if (tp
->urg_data
== TCP_URG_NOTYET
) {
3848 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
3851 /* Is the urgent pointer pointing into this packet? */
3852 if (ptr
< skb
->len
) {
3854 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
3856 tp
->urg_data
= TCP_URG_VALID
| tmp
;
3857 if (!sock_flag(sk
, SOCK_DEAD
))
3858 sk
->sk_data_ready(sk
, 0);
3863 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3865 struct tcp_sock
*tp
= tcp_sk(sk
);
3866 int chunk
= skb
->len
- hlen
;
3870 if (skb
->ip_summed
==CHECKSUM_UNNECESSARY
)
3871 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
3873 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
3877 tp
->ucopy
.len
-= chunk
;
3878 tp
->copied_seq
+= chunk
;
3879 tcp_rcv_space_adjust(sk
);
3886 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3890 if (sock_owned_by_user(sk
)) {
3892 result
= __tcp_checksum_complete(skb
);
3895 result
= __tcp_checksum_complete(skb
);
3900 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3902 return skb
->ip_summed
!= CHECKSUM_UNNECESSARY
&&
3903 __tcp_checksum_complete_user(sk
, skb
);
3906 #ifdef CONFIG_NET_DMA
3907 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3909 struct tcp_sock
*tp
= tcp_sk(sk
);
3910 int chunk
= skb
->len
- hlen
;
3912 int copied_early
= 0;
3914 if (tp
->ucopy
.wakeup
)
3917 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
3918 tp
->ucopy
.dma_chan
= get_softnet_dma();
3920 if (tp
->ucopy
.dma_chan
&& skb
->ip_summed
== CHECKSUM_UNNECESSARY
) {
3922 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
3923 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
3928 tp
->ucopy
.dma_cookie
= dma_cookie
;
3931 tp
->ucopy
.len
-= chunk
;
3932 tp
->copied_seq
+= chunk
;
3933 tcp_rcv_space_adjust(sk
);
3935 if ((tp
->ucopy
.len
== 0) ||
3936 (tcp_flag_word(skb
->h
.th
) & TCP_FLAG_PSH
) ||
3937 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
3938 tp
->ucopy
.wakeup
= 1;
3939 sk
->sk_data_ready(sk
, 0);
3941 } else if (chunk
> 0) {
3942 tp
->ucopy
.wakeup
= 1;
3943 sk
->sk_data_ready(sk
, 0);
3946 return copied_early
;
3948 #endif /* CONFIG_NET_DMA */
3951 * TCP receive function for the ESTABLISHED state.
3953 * It is split into a fast path and a slow path. The fast path is
3955 * - A zero window was announced from us - zero window probing
3956 * is only handled properly in the slow path.
3957 * - Out of order segments arrived.
3958 * - Urgent data is expected.
3959 * - There is no buffer space left
3960 * - Unexpected TCP flags/window values/header lengths are received
3961 * (detected by checking the TCP header against pred_flags)
3962 * - Data is sent in both directions. Fast path only supports pure senders
3963 * or pure receivers (this means either the sequence number or the ack
3964 * value must stay constant)
3965 * - Unexpected TCP option.
3967 * When these conditions are not satisfied it drops into a standard
3968 * receive procedure patterned after RFC793 to handle all cases.
3969 * The first three cases are guaranteed by proper pred_flags setting,
3970 * the rest is checked inline. Fast processing is turned on in
3971 * tcp_data_queue when everything is OK.
3973 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
3974 struct tcphdr
*th
, unsigned len
)
3976 struct tcp_sock
*tp
= tcp_sk(sk
);
3979 * Header prediction.
3980 * The code loosely follows the one in the famous
3981 * "30 instruction TCP receive" Van Jacobson mail.
3983 * Van's trick is to deposit buffers into socket queue
3984 * on a device interrupt, to call tcp_recv function
3985 * on the receive process context and checksum and copy
3986 * the buffer to user space. smart...
3988 * Our current scheme is not silly either but we take the
3989 * extra cost of the net_bh soft interrupt processing...
3990 * We do checksum and copy also but from device to kernel.
3993 tp
->rx_opt
.saw_tstamp
= 0;
3995 /* pred_flags is 0xS?10 << 16 + snd_wnd
3996 * if header_prediction is to be made
3997 * 'S' will always be tp->tcp_header_len >> 2
3998 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3999 * turn it off (when there are holes in the receive
4000 * space for instance)
4001 * PSH flag is ignored.
4004 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4005 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4006 int tcp_header_len
= tp
->tcp_header_len
;
4008 /* Timestamp header prediction: tcp_header_len
4009 * is automatically equal to th->doff*4 due to pred_flags
4013 /* Check timestamp */
4014 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4015 __be32
*ptr
= (__be32
*)(th
+ 1);
4017 /* No? Slow path! */
4018 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4019 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4022 tp
->rx_opt
.saw_tstamp
= 1;
4024 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4026 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4028 /* If PAWS failed, check it more carefully in slow path */
4029 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4032 /* DO NOT update ts_recent here, if checksum fails
4033 * and timestamp was corrupted part, it will result
4034 * in a hung connection since we will drop all
4035 * future packets due to the PAWS test.
4039 if (len
<= tcp_header_len
) {
4040 /* Bulk data transfer: sender */
4041 if (len
== tcp_header_len
) {
4042 /* Predicted packet is in window by definition.
4043 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4044 * Hence, check seq<=rcv_wup reduces to:
4046 if (tcp_header_len
==
4047 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4048 tp
->rcv_nxt
== tp
->rcv_wup
)
4049 tcp_store_ts_recent(tp
);
4051 /* We know that such packets are checksummed
4054 tcp_ack(sk
, skb
, 0);
4056 tcp_data_snd_check(sk
, tp
);
4058 } else { /* Header too small */
4059 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4064 int copied_early
= 0;
4066 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4067 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4068 #ifdef CONFIG_NET_DMA
4069 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4074 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4075 __set_current_state(TASK_RUNNING
);
4077 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4081 /* Predicted packet is in window by definition.
4082 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4083 * Hence, check seq<=rcv_wup reduces to:
4085 if (tcp_header_len
==
4086 (sizeof(struct tcphdr
) +
4087 TCPOLEN_TSTAMP_ALIGNED
) &&
4088 tp
->rcv_nxt
== tp
->rcv_wup
)
4089 tcp_store_ts_recent(tp
);
4091 tcp_rcv_rtt_measure_ts(sk
, skb
);
4093 __skb_pull(skb
, tcp_header_len
);
4094 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4095 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4098 tcp_cleanup_rbuf(sk
, skb
->len
);
4101 if (tcp_checksum_complete_user(sk
, skb
))
4104 /* Predicted packet is in window by definition.
4105 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4106 * Hence, check seq<=rcv_wup reduces to:
4108 if (tcp_header_len
==
4109 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4110 tp
->rcv_nxt
== tp
->rcv_wup
)
4111 tcp_store_ts_recent(tp
);
4113 tcp_rcv_rtt_measure_ts(sk
, skb
);
4115 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4118 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4120 /* Bulk data transfer: receiver */
4121 __skb_pull(skb
,tcp_header_len
);
4122 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4123 sk_stream_set_owner_r(skb
, sk
);
4124 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4127 tcp_event_data_recv(sk
, tp
, skb
);
4129 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4130 /* Well, only one small jumplet in fast path... */
4131 tcp_ack(sk
, skb
, FLAG_DATA
);
4132 tcp_data_snd_check(sk
, tp
);
4133 if (!inet_csk_ack_scheduled(sk
))
4137 __tcp_ack_snd_check(sk
, 0);
4139 #ifdef CONFIG_NET_DMA
4141 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4147 sk
->sk_data_ready(sk
, 0);
4153 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4157 * RFC1323: H1. Apply PAWS check first.
4159 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4160 tcp_paws_discard(sk
, skb
)) {
4162 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4163 tcp_send_dupack(sk
, skb
);
4166 /* Resets are accepted even if PAWS failed.
4168 ts_recent update must be made after we are sure
4169 that the packet is in window.
4174 * Standard slow path.
4177 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4178 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4179 * (RST) segments are validated by checking their SEQ-fields."
4180 * And page 69: "If an incoming segment is not acceptable,
4181 * an acknowledgment should be sent in reply (unless the RST bit
4182 * is set, if so drop the segment and return)".
4185 tcp_send_dupack(sk
, skb
);
4194 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4196 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4197 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4198 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4205 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4207 tcp_rcv_rtt_measure_ts(sk
, skb
);
4209 /* Process urgent data. */
4210 tcp_urg(sk
, skb
, th
);
4212 /* step 7: process the segment text */
4213 tcp_data_queue(sk
, skb
);
4215 tcp_data_snd_check(sk
, tp
);
4216 tcp_ack_snd_check(sk
);
4220 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4227 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4228 struct tcphdr
*th
, unsigned len
)
4230 struct tcp_sock
*tp
= tcp_sk(sk
);
4231 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4232 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4234 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4238 * "If the state is SYN-SENT then
4239 * first check the ACK bit
4240 * If the ACK bit is set
4241 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4242 * a reset (unless the RST bit is set, if so drop
4243 * the segment and return)"
4245 * We do not send data with SYN, so that RFC-correct
4248 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4249 goto reset_and_undo
;
4251 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4252 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4254 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4255 goto reset_and_undo
;
4258 /* Now ACK is acceptable.
4260 * "If the RST bit is set
4261 * If the ACK was acceptable then signal the user "error:
4262 * connection reset", drop the segment, enter CLOSED state,
4263 * delete TCB, and return."
4272 * "fifth, if neither of the SYN or RST bits is set then
4273 * drop the segment and return."
4279 goto discard_and_undo
;
4282 * "If the SYN bit is on ...
4283 * are acceptable then ...
4284 * (our SYN has been ACKed), change the connection
4285 * state to ESTABLISHED..."
4288 TCP_ECN_rcv_synack(tp
, th
);
4290 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4291 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4293 /* Ok.. it's good. Set up sequence numbers and
4294 * move to established.
4296 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4297 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4299 /* RFC1323: The window in SYN & SYN/ACK segments is
4302 tp
->snd_wnd
= ntohs(th
->window
);
4303 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4305 if (!tp
->rx_opt
.wscale_ok
) {
4306 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4307 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4310 if (tp
->rx_opt
.saw_tstamp
) {
4311 tp
->rx_opt
.tstamp_ok
= 1;
4312 tp
->tcp_header_len
=
4313 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4314 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4315 tcp_store_ts_recent(tp
);
4317 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4320 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4321 tp
->rx_opt
.sack_ok
|= 2;
4324 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4325 tcp_initialize_rcv_mss(sk
);
4327 /* Remember, tcp_poll() does not lock socket!
4328 * Change state from SYN-SENT only after copied_seq
4329 * is initialized. */
4330 tp
->copied_seq
= tp
->rcv_nxt
;
4332 tcp_set_state(sk
, TCP_ESTABLISHED
);
4334 security_inet_conn_established(sk
, skb
);
4336 /* Make sure socket is routed, for correct metrics. */
4337 icsk
->icsk_af_ops
->rebuild_header(sk
);
4339 tcp_init_metrics(sk
);
4341 tcp_init_congestion_control(sk
);
4343 /* Prevent spurious tcp_cwnd_restart() on first data
4346 tp
->lsndtime
= tcp_time_stamp
;
4348 tcp_init_buffer_space(sk
);
4350 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4351 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4353 if (!tp
->rx_opt
.snd_wscale
)
4354 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4358 if (!sock_flag(sk
, SOCK_DEAD
)) {
4359 sk
->sk_state_change(sk
);
4360 sk_wake_async(sk
, 0, POLL_OUT
);
4363 if (sk
->sk_write_pending
||
4364 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4365 icsk
->icsk_ack
.pingpong
) {
4366 /* Save one ACK. Data will be ready after
4367 * several ticks, if write_pending is set.
4369 * It may be deleted, but with this feature tcpdumps
4370 * look so _wonderfully_ clever, that I was not able
4371 * to stand against the temptation 8) --ANK
4373 inet_csk_schedule_ack(sk
);
4374 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4375 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4376 tcp_incr_quickack(sk
);
4377 tcp_enter_quickack_mode(sk
);
4378 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4379 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4390 /* No ACK in the segment */
4394 * "If the RST bit is set
4396 * Otherwise (no ACK) drop the segment and return."
4399 goto discard_and_undo
;
4403 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4404 goto discard_and_undo
;
4407 /* We see SYN without ACK. It is attempt of
4408 * simultaneous connect with crossed SYNs.
4409 * Particularly, it can be connect to self.
4411 tcp_set_state(sk
, TCP_SYN_RECV
);
4413 if (tp
->rx_opt
.saw_tstamp
) {
4414 tp
->rx_opt
.tstamp_ok
= 1;
4415 tcp_store_ts_recent(tp
);
4416 tp
->tcp_header_len
=
4417 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4419 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4422 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4423 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4425 /* RFC1323: The window in SYN & SYN/ACK segments is
4428 tp
->snd_wnd
= ntohs(th
->window
);
4429 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4430 tp
->max_window
= tp
->snd_wnd
;
4432 TCP_ECN_rcv_syn(tp
, th
);
4435 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4436 tcp_initialize_rcv_mss(sk
);
4439 tcp_send_synack(sk
);
4441 /* Note, we could accept data and URG from this segment.
4442 * There are no obstacles to make this.
4444 * However, if we ignore data in ACKless segments sometimes,
4445 * we have no reasons to accept it sometimes.
4446 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4447 * is not flawless. So, discard packet for sanity.
4448 * Uncomment this return to process the data.
4455 /* "fifth, if neither of the SYN or RST bits is set then
4456 * drop the segment and return."
4460 tcp_clear_options(&tp
->rx_opt
);
4461 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4465 tcp_clear_options(&tp
->rx_opt
);
4466 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4472 * This function implements the receiving procedure of RFC 793 for
4473 * all states except ESTABLISHED and TIME_WAIT.
4474 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4475 * address independent.
4478 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4479 struct tcphdr
*th
, unsigned len
)
4481 struct tcp_sock
*tp
= tcp_sk(sk
);
4482 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4485 tp
->rx_opt
.saw_tstamp
= 0;
4487 switch (sk
->sk_state
) {
4499 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4502 /* Now we have several options: In theory there is
4503 * nothing else in the frame. KA9Q has an option to
4504 * send data with the syn, BSD accepts data with the
4505 * syn up to the [to be] advertised window and
4506 * Solaris 2.1 gives you a protocol error. For now
4507 * we just ignore it, that fits the spec precisely
4508 * and avoids incompatibilities. It would be nice in
4509 * future to drop through and process the data.
4511 * Now that TTCP is starting to be used we ought to
4513 * But, this leaves one open to an easy denial of
4514 * service attack, and SYN cookies can't defend
4515 * against this problem. So, we drop the data
4516 * in the interest of security over speed unless
4517 * it's still in use.
4525 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4529 /* Do step6 onward by hand. */
4530 tcp_urg(sk
, skb
, th
);
4532 tcp_data_snd_check(sk
, tp
);
4536 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4537 tcp_paws_discard(sk
, skb
)) {
4539 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4540 tcp_send_dupack(sk
, skb
);
4543 /* Reset is accepted even if it did not pass PAWS. */
4546 /* step 1: check sequence number */
4547 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4549 tcp_send_dupack(sk
, skb
);
4553 /* step 2: check RST bit */
4559 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4561 /* step 3: check security and precedence [ignored] */
4565 * Check for a SYN in window.
4567 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4568 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4573 /* step 5: check the ACK field */
4575 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4577 switch(sk
->sk_state
) {
4580 tp
->copied_seq
= tp
->rcv_nxt
;
4582 tcp_set_state(sk
, TCP_ESTABLISHED
);
4583 sk
->sk_state_change(sk
);
4585 /* Note, that this wakeup is only for marginal
4586 * crossed SYN case. Passively open sockets
4587 * are not waked up, because sk->sk_sleep ==
4588 * NULL and sk->sk_socket == NULL.
4590 if (sk
->sk_socket
) {
4591 sk_wake_async(sk
,0,POLL_OUT
);
4594 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4595 tp
->snd_wnd
= ntohs(th
->window
) <<
4596 tp
->rx_opt
.snd_wscale
;
4597 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4598 TCP_SKB_CB(skb
)->seq
);
4600 /* tcp_ack considers this ACK as duplicate
4601 * and does not calculate rtt.
4602 * Fix it at least with timestamps.
4604 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4606 tcp_ack_saw_tstamp(sk
, 0);
4608 if (tp
->rx_opt
.tstamp_ok
)
4609 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4611 /* Make sure socket is routed, for
4614 icsk
->icsk_af_ops
->rebuild_header(sk
);
4616 tcp_init_metrics(sk
);
4618 tcp_init_congestion_control(sk
);
4620 /* Prevent spurious tcp_cwnd_restart() on
4621 * first data packet.
4623 tp
->lsndtime
= tcp_time_stamp
;
4626 tcp_initialize_rcv_mss(sk
);
4627 tcp_init_buffer_space(sk
);
4628 tcp_fast_path_on(tp
);
4635 if (tp
->snd_una
== tp
->write_seq
) {
4636 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4637 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4638 dst_confirm(sk
->sk_dst_cache
);
4640 if (!sock_flag(sk
, SOCK_DEAD
))
4641 /* Wake up lingering close() */
4642 sk
->sk_state_change(sk
);
4646 if (tp
->linger2
< 0 ||
4647 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4648 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4650 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4654 tmo
= tcp_fin_time(sk
);
4655 if (tmo
> TCP_TIMEWAIT_LEN
) {
4656 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4657 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4658 /* Bad case. We could lose such FIN otherwise.
4659 * It is not a big problem, but it looks confusing
4660 * and not so rare event. We still can lose it now,
4661 * if it spins in bh_lock_sock(), but it is really
4664 inet_csk_reset_keepalive_timer(sk
, tmo
);
4666 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4674 if (tp
->snd_una
== tp
->write_seq
) {
4675 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4681 if (tp
->snd_una
== tp
->write_seq
) {
4682 tcp_update_metrics(sk
);
4691 /* step 6: check the URG bit */
4692 tcp_urg(sk
, skb
, th
);
4694 /* step 7: process the segment text */
4695 switch (sk
->sk_state
) {
4696 case TCP_CLOSE_WAIT
:
4699 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4703 /* RFC 793 says to queue data in these states,
4704 * RFC 1122 says we MUST send a reset.
4705 * BSD 4.4 also does reset.
4707 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4708 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4709 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4710 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4716 case TCP_ESTABLISHED
:
4717 tcp_data_queue(sk
, skb
);
4722 /* tcp_data could move socket to TIME-WAIT */
4723 if (sk
->sk_state
!= TCP_CLOSE
) {
4724 tcp_data_snd_check(sk
, tp
);
4725 tcp_ack_snd_check(sk
);
4735 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4736 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4737 EXPORT_SYMBOL(tcp_parse_options
);
4738 EXPORT_SYMBOL(tcp_rcv_established
);
4739 EXPORT_SYMBOL(tcp_rcv_state_process
);
4740 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);