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
;
955 * if the only SACK change is the increase of the end_seq of
956 * the first block then only apply that SACK block
957 * and use retrans queue hinting otherwise slowpath */
959 for (i
= 0; i
< num_sacks
; i
++) {
960 __u32 start_seq
= ntohl(sp
[i
].start_seq
);
961 __u32 end_seq
= ntohl(sp
[i
].end_seq
);
964 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
967 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
968 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
971 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
972 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
974 /* Check for D-SACK. */
976 u32 ack
= TCP_SKB_CB(ack_skb
)->ack_seq
;
978 if (before(start_seq
, ack
)) {
980 tp
->rx_opt
.sack_ok
|= 4;
981 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
982 } else if (num_sacks
> 1 &&
983 !after(end_seq
, ntohl(sp
[1].end_seq
)) &&
984 !before(start_seq
, ntohl(sp
[1].start_seq
))) {
986 tp
->rx_opt
.sack_ok
|= 4;
987 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
990 /* D-SACK for already forgotten data...
991 * Do dumb counting. */
993 !after(end_seq
, prior_snd_una
) &&
994 after(end_seq
, tp
->undo_marker
))
997 /* Eliminate too old ACKs, but take into
998 * account more or less fresh ones, they can
999 * contain valid SACK info.
1001 if (before(ack
, prior_snd_una
- tp
->max_window
))
1006 first_sack_index
= 0;
1011 tp
->fastpath_skb_hint
= NULL
;
1013 /* order SACK blocks to allow in order walk of the retrans queue */
1014 for (i
= num_sacks
-1; i
> 0; i
--) {
1015 for (j
= 0; j
< i
; j
++){
1016 if (after(ntohl(sp
[j
].start_seq
),
1017 ntohl(sp
[j
+1].start_seq
))){
1018 struct tcp_sack_block_wire tmp
;
1024 /* Track where the first SACK block goes to */
1025 if (j
== first_sack_index
)
1026 first_sack_index
= j
+1;
1033 /* clear flag as used for different purpose in following code */
1036 /* Use SACK fastpath hint if valid */
1037 cached_skb
= tp
->fastpath_skb_hint
;
1038 cached_fack_count
= tp
->fastpath_cnt_hint
;
1040 cached_skb
= sk
->sk_write_queue
.next
;
1041 cached_fack_count
= 0;
1044 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1045 struct sk_buff
*skb
;
1046 __u32 start_seq
= ntohl(sp
->start_seq
);
1047 __u32 end_seq
= ntohl(sp
->end_seq
);
1051 fack_count
= cached_fack_count
;
1053 /* Event "B" in the comment above. */
1054 if (after(end_seq
, tp
->high_seq
))
1055 flag
|= FLAG_DATA_LOST
;
1057 sk_stream_for_retrans_queue_from(skb
, sk
) {
1058 int in_sack
, pcount
;
1062 cached_fack_count
= fack_count
;
1063 if (i
== first_sack_index
) {
1064 tp
->fastpath_skb_hint
= skb
;
1065 tp
->fastpath_cnt_hint
= fack_count
;
1068 /* The retransmission queue is always in order, so
1069 * we can short-circuit the walk early.
1071 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1074 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1075 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1077 pcount
= tcp_skb_pcount(skb
);
1079 if (pcount
> 1 && !in_sack
&&
1080 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1081 unsigned int pkt_len
;
1083 in_sack
= !after(start_seq
,
1084 TCP_SKB_CB(skb
)->seq
);
1087 pkt_len
= (start_seq
-
1088 TCP_SKB_CB(skb
)->seq
);
1090 pkt_len
= (end_seq
-
1091 TCP_SKB_CB(skb
)->seq
);
1092 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1094 pcount
= tcp_skb_pcount(skb
);
1097 fack_count
+= pcount
;
1099 sacked
= TCP_SKB_CB(skb
)->sacked
;
1101 /* Account D-SACK for retransmitted packet. */
1102 if ((dup_sack
&& in_sack
) &&
1103 (sacked
& TCPCB_RETRANS
) &&
1104 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1107 /* The frame is ACKed. */
1108 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1109 if (sacked
&TCPCB_RETRANS
) {
1110 if ((dup_sack
&& in_sack
) &&
1111 (sacked
&TCPCB_SACKED_ACKED
))
1112 reord
= min(fack_count
, reord
);
1114 /* If it was in a hole, we detected reordering. */
1115 if (fack_count
< prior_fackets
&&
1116 !(sacked
&TCPCB_SACKED_ACKED
))
1117 reord
= min(fack_count
, reord
);
1120 /* Nothing to do; acked frame is about to be dropped. */
1124 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1125 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1126 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1127 lost_retrans
= end_seq
;
1132 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1133 if (sacked
& TCPCB_SACKED_RETRANS
) {
1134 /* If the segment is not tagged as lost,
1135 * we do not clear RETRANS, believing
1136 * that retransmission is still in flight.
1138 if (sacked
& TCPCB_LOST
) {
1139 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1140 tp
->lost_out
-= tcp_skb_pcount(skb
);
1141 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1143 /* clear lost hint */
1144 tp
->retransmit_skb_hint
= NULL
;
1147 /* New sack for not retransmitted frame,
1148 * which was in hole. It is reordering.
1150 if (!(sacked
& TCPCB_RETRANS
) &&
1151 fack_count
< prior_fackets
)
1152 reord
= min(fack_count
, reord
);
1154 if (sacked
& TCPCB_LOST
) {
1155 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1156 tp
->lost_out
-= tcp_skb_pcount(skb
);
1158 /* clear lost hint */
1159 tp
->retransmit_skb_hint
= NULL
;
1163 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1164 flag
|= FLAG_DATA_SACKED
;
1165 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1167 if (fack_count
> tp
->fackets_out
)
1168 tp
->fackets_out
= fack_count
;
1170 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1171 reord
= min(fack_count
, reord
);
1174 /* D-SACK. We can detect redundant retransmission
1175 * in S|R and plain R frames and clear it.
1176 * undo_retrans is decreased above, L|R frames
1177 * are accounted above as well.
1180 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1181 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1182 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1183 tp
->retransmit_skb_hint
= NULL
;
1188 /* Check for lost retransmit. This superb idea is
1189 * borrowed from "ratehalving". Event "C".
1190 * Later note: FACK people cheated me again 8),
1191 * we have to account for reordering! Ugly,
1194 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1195 struct sk_buff
*skb
;
1197 sk_stream_for_retrans_queue(skb
, sk
) {
1198 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1200 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1202 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1203 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1205 !before(lost_retrans
,
1206 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1208 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1209 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1211 /* clear lost hint */
1212 tp
->retransmit_skb_hint
= NULL
;
1214 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1215 tp
->lost_out
+= tcp_skb_pcount(skb
);
1216 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1217 flag
|= FLAG_DATA_SACKED
;
1218 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1224 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1226 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
)
1227 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1229 #if FASTRETRANS_DEBUG > 0
1230 BUG_TRAP((int)tp
->sacked_out
>= 0);
1231 BUG_TRAP((int)tp
->lost_out
>= 0);
1232 BUG_TRAP((int)tp
->retrans_out
>= 0);
1233 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1238 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1239 * segments to see from the next ACKs whether any data was really missing.
1240 * If the RTO was spurious, new ACKs should arrive.
1242 void tcp_enter_frto(struct sock
*sk
)
1244 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1245 struct tcp_sock
*tp
= tcp_sk(sk
);
1246 struct sk_buff
*skb
;
1248 tp
->frto_counter
= 1;
1250 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1251 tp
->snd_una
== tp
->high_seq
||
1252 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1253 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1254 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1255 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1258 /* Have to clear retransmission markers here to keep the bookkeeping
1259 * in shape, even though we are not yet in Loss state.
1260 * If something was really lost, it is eventually caught up
1261 * in tcp_enter_frto_loss.
1263 tp
->retrans_out
= 0;
1264 tp
->undo_marker
= tp
->snd_una
;
1265 tp
->undo_retrans
= 0;
1267 sk_stream_for_retrans_queue(skb
, sk
) {
1268 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_RETRANS
;
1270 tcp_sync_left_out(tp
);
1272 tcp_set_ca_state(sk
, TCP_CA_Open
);
1273 tp
->frto_highmark
= tp
->snd_nxt
;
1276 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1277 * which indicates that we should follow the traditional RTO recovery,
1278 * i.e. mark everything lost and do go-back-N retransmission.
1280 static void tcp_enter_frto_loss(struct sock
*sk
)
1282 struct tcp_sock
*tp
= tcp_sk(sk
);
1283 struct sk_buff
*skb
;
1288 tp
->fackets_out
= 0;
1290 sk_stream_for_retrans_queue(skb
, sk
) {
1291 cnt
+= tcp_skb_pcount(skb
);
1292 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1293 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1295 /* Do not mark those segments lost that were
1296 * forward transmitted after RTO
1298 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1299 tp
->frto_highmark
)) {
1300 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1301 tp
->lost_out
+= tcp_skb_pcount(skb
);
1304 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1305 tp
->fackets_out
= cnt
;
1308 tcp_sync_left_out(tp
);
1310 tp
->snd_cwnd
= tp
->frto_counter
+ tcp_packets_in_flight(tp
)+1;
1311 tp
->snd_cwnd_cnt
= 0;
1312 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1313 tp
->undo_marker
= 0;
1314 tp
->frto_counter
= 0;
1316 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1317 sysctl_tcp_reordering
);
1318 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1319 tp
->high_seq
= tp
->frto_highmark
;
1320 TCP_ECN_queue_cwr(tp
);
1322 clear_all_retrans_hints(tp
);
1325 void tcp_clear_retrans(struct tcp_sock
*tp
)
1328 tp
->retrans_out
= 0;
1330 tp
->fackets_out
= 0;
1334 tp
->undo_marker
= 0;
1335 tp
->undo_retrans
= 0;
1338 /* Enter Loss state. If "how" is not zero, forget all SACK information
1339 * and reset tags completely, otherwise preserve SACKs. If receiver
1340 * dropped its ofo queue, we will know this due to reneging detection.
1342 void tcp_enter_loss(struct sock
*sk
, int how
)
1344 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1345 struct tcp_sock
*tp
= tcp_sk(sk
);
1346 struct sk_buff
*skb
;
1349 /* Reduce ssthresh if it has not yet been made inside this window. */
1350 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1351 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1352 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1353 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1354 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1357 tp
->snd_cwnd_cnt
= 0;
1358 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1360 tp
->bytes_acked
= 0;
1361 tcp_clear_retrans(tp
);
1363 /* Push undo marker, if it was plain RTO and nothing
1364 * was retransmitted. */
1366 tp
->undo_marker
= tp
->snd_una
;
1368 sk_stream_for_retrans_queue(skb
, sk
) {
1369 cnt
+= tcp_skb_pcount(skb
);
1370 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1371 tp
->undo_marker
= 0;
1372 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1373 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1374 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1375 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1376 tp
->lost_out
+= tcp_skb_pcount(skb
);
1378 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1379 tp
->fackets_out
= cnt
;
1382 tcp_sync_left_out(tp
);
1384 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1385 sysctl_tcp_reordering
);
1386 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1387 tp
->high_seq
= tp
->snd_nxt
;
1388 TCP_ECN_queue_cwr(tp
);
1390 clear_all_retrans_hints(tp
);
1393 static int tcp_check_sack_reneging(struct sock
*sk
)
1395 struct sk_buff
*skb
;
1397 /* If ACK arrived pointing to a remembered SACK,
1398 * it means that our remembered SACKs do not reflect
1399 * real state of receiver i.e.
1400 * receiver _host_ is heavily congested (or buggy).
1401 * Do processing similar to RTO timeout.
1403 if ((skb
= skb_peek(&sk
->sk_write_queue
)) != NULL
&&
1404 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1405 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1406 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1408 tcp_enter_loss(sk
, 1);
1409 icsk
->icsk_retransmits
++;
1410 tcp_retransmit_skb(sk
, skb_peek(&sk
->sk_write_queue
));
1411 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1412 icsk
->icsk_rto
, TCP_RTO_MAX
);
1418 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1420 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1423 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1425 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1428 static inline int tcp_head_timedout(struct sock
*sk
, struct tcp_sock
*tp
)
1430 return tp
->packets_out
&&
1431 tcp_skb_timedout(sk
, skb_peek(&sk
->sk_write_queue
));
1434 /* Linux NewReno/SACK/FACK/ECN state machine.
1435 * --------------------------------------
1437 * "Open" Normal state, no dubious events, fast path.
1438 * "Disorder" In all the respects it is "Open",
1439 * but requires a bit more attention. It is entered when
1440 * we see some SACKs or dupacks. It is split of "Open"
1441 * mainly to move some processing from fast path to slow one.
1442 * "CWR" CWND was reduced due to some Congestion Notification event.
1443 * It can be ECN, ICMP source quench, local device congestion.
1444 * "Recovery" CWND was reduced, we are fast-retransmitting.
1445 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1447 * tcp_fastretrans_alert() is entered:
1448 * - each incoming ACK, if state is not "Open"
1449 * - when arrived ACK is unusual, namely:
1454 * Counting packets in flight is pretty simple.
1456 * in_flight = packets_out - left_out + retrans_out
1458 * packets_out is SND.NXT-SND.UNA counted in packets.
1460 * retrans_out is number of retransmitted segments.
1462 * left_out is number of segments left network, but not ACKed yet.
1464 * left_out = sacked_out + lost_out
1466 * sacked_out: Packets, which arrived to receiver out of order
1467 * and hence not ACKed. With SACKs this number is simply
1468 * amount of SACKed data. Even without SACKs
1469 * it is easy to give pretty reliable estimate of this number,
1470 * counting duplicate ACKs.
1472 * lost_out: Packets lost by network. TCP has no explicit
1473 * "loss notification" feedback from network (for now).
1474 * It means that this number can be only _guessed_.
1475 * Actually, it is the heuristics to predict lossage that
1476 * distinguishes different algorithms.
1478 * F.e. after RTO, when all the queue is considered as lost,
1479 * lost_out = packets_out and in_flight = retrans_out.
1481 * Essentially, we have now two algorithms counting
1484 * FACK: It is the simplest heuristics. As soon as we decided
1485 * that something is lost, we decide that _all_ not SACKed
1486 * packets until the most forward SACK are lost. I.e.
1487 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1488 * It is absolutely correct estimate, if network does not reorder
1489 * packets. And it loses any connection to reality when reordering
1490 * takes place. We use FACK by default until reordering
1491 * is suspected on the path to this destination.
1493 * NewReno: when Recovery is entered, we assume that one segment
1494 * is lost (classic Reno). While we are in Recovery and
1495 * a partial ACK arrives, we assume that one more packet
1496 * is lost (NewReno). This heuristics are the same in NewReno
1499 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1500 * deflation etc. CWND is real congestion window, never inflated, changes
1501 * only according to classic VJ rules.
1503 * Really tricky (and requiring careful tuning) part of algorithm
1504 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1505 * The first determines the moment _when_ we should reduce CWND and,
1506 * hence, slow down forward transmission. In fact, it determines the moment
1507 * when we decide that hole is caused by loss, rather than by a reorder.
1509 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1510 * holes, caused by lost packets.
1512 * And the most logically complicated part of algorithm is undo
1513 * heuristics. We detect false retransmits due to both too early
1514 * fast retransmit (reordering) and underestimated RTO, analyzing
1515 * timestamps and D-SACKs. When we detect that some segments were
1516 * retransmitted by mistake and CWND reduction was wrong, we undo
1517 * window reduction and abort recovery phase. This logic is hidden
1518 * inside several functions named tcp_try_undo_<something>.
1521 /* This function decides, when we should leave Disordered state
1522 * and enter Recovery phase, reducing congestion window.
1524 * Main question: may we further continue forward transmission
1525 * with the same cwnd?
1527 static int tcp_time_to_recover(struct sock
*sk
, struct tcp_sock
*tp
)
1531 /* Trick#1: The loss is proven. */
1535 /* Not-A-Trick#2 : Classic rule... */
1536 if (tcp_fackets_out(tp
) > tp
->reordering
)
1539 /* Trick#3 : when we use RFC2988 timer restart, fast
1540 * retransmit can be triggered by timeout of queue head.
1542 if (tcp_head_timedout(sk
, tp
))
1545 /* Trick#4: It is still not OK... But will it be useful to delay
1548 packets_out
= tp
->packets_out
;
1549 if (packets_out
<= tp
->reordering
&&
1550 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1551 !tcp_may_send_now(sk
, tp
)) {
1552 /* We have nothing to send. This connection is limited
1553 * either by receiver window or by application.
1561 /* If we receive more dupacks than we expected counting segments
1562 * in assumption of absent reordering, interpret this as reordering.
1563 * The only another reason could be bug in receiver TCP.
1565 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1567 struct tcp_sock
*tp
= tcp_sk(sk
);
1570 holes
= max(tp
->lost_out
, 1U);
1571 holes
= min(holes
, tp
->packets_out
);
1573 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1574 tp
->sacked_out
= tp
->packets_out
- holes
;
1575 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1579 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1581 static void tcp_add_reno_sack(struct sock
*sk
)
1583 struct tcp_sock
*tp
= tcp_sk(sk
);
1585 tcp_check_reno_reordering(sk
, 0);
1586 tcp_sync_left_out(tp
);
1589 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1591 static void tcp_remove_reno_sacks(struct sock
*sk
, struct tcp_sock
*tp
, int acked
)
1594 /* One ACK acked hole. The rest eat duplicate ACKs. */
1595 if (acked
-1 >= tp
->sacked_out
)
1598 tp
->sacked_out
-= acked
-1;
1600 tcp_check_reno_reordering(sk
, acked
);
1601 tcp_sync_left_out(tp
);
1604 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1607 tp
->left_out
= tp
->lost_out
;
1610 /* Mark head of queue up as lost. */
1611 static void tcp_mark_head_lost(struct sock
*sk
, struct tcp_sock
*tp
,
1612 int packets
, u32 high_seq
)
1614 struct sk_buff
*skb
;
1617 BUG_TRAP(packets
<= tp
->packets_out
);
1618 if (tp
->lost_skb_hint
) {
1619 skb
= tp
->lost_skb_hint
;
1620 cnt
= tp
->lost_cnt_hint
;
1622 skb
= sk
->sk_write_queue
.next
;
1626 sk_stream_for_retrans_queue_from(skb
, sk
) {
1627 /* TODO: do this better */
1628 /* this is not the most efficient way to do this... */
1629 tp
->lost_skb_hint
= skb
;
1630 tp
->lost_cnt_hint
= cnt
;
1631 cnt
+= tcp_skb_pcount(skb
);
1632 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1634 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1635 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1636 tp
->lost_out
+= tcp_skb_pcount(skb
);
1638 /* clear xmit_retransmit_queue hints
1639 * if this is beyond hint */
1640 if(tp
->retransmit_skb_hint
!= NULL
&&
1641 before(TCP_SKB_CB(skb
)->seq
,
1642 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
)) {
1644 tp
->retransmit_skb_hint
= NULL
;
1648 tcp_sync_left_out(tp
);
1651 /* Account newly detected lost packet(s) */
1653 static void tcp_update_scoreboard(struct sock
*sk
, struct tcp_sock
*tp
)
1656 int lost
= tp
->fackets_out
- tp
->reordering
;
1659 tcp_mark_head_lost(sk
, tp
, lost
, tp
->high_seq
);
1661 tcp_mark_head_lost(sk
, tp
, 1, tp
->high_seq
);
1664 /* New heuristics: it is possible only after we switched
1665 * to restart timer each time when something is ACKed.
1666 * Hence, we can detect timed out packets during fast
1667 * retransmit without falling to slow start.
1669 if (!IsReno(tp
) && tcp_head_timedout(sk
, tp
)) {
1670 struct sk_buff
*skb
;
1672 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1673 : sk
->sk_write_queue
.next
;
1675 sk_stream_for_retrans_queue_from(skb
, sk
) {
1676 if (!tcp_skb_timedout(sk
, skb
))
1679 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1680 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1681 tp
->lost_out
+= tcp_skb_pcount(skb
);
1683 /* clear xmit_retrans hint */
1684 if (tp
->retransmit_skb_hint
&&
1685 before(TCP_SKB_CB(skb
)->seq
,
1686 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1688 tp
->retransmit_skb_hint
= NULL
;
1692 tp
->scoreboard_skb_hint
= skb
;
1694 tcp_sync_left_out(tp
);
1698 /* CWND moderation, preventing bursts due to too big ACKs
1699 * in dubious situations.
1701 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1703 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1704 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1705 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1708 /* Lower bound on congestion window is slow start threshold
1709 * unless congestion avoidance choice decides to overide it.
1711 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1713 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1715 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1718 /* Decrease cwnd each second ack. */
1719 static void tcp_cwnd_down(struct sock
*sk
)
1721 struct tcp_sock
*tp
= tcp_sk(sk
);
1722 int decr
= tp
->snd_cwnd_cnt
+ 1;
1724 tp
->snd_cwnd_cnt
= decr
&1;
1727 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1728 tp
->snd_cwnd
-= decr
;
1730 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1731 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1734 /* Nothing was retransmitted or returned timestamp is less
1735 * than timestamp of the first retransmission.
1737 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1739 return !tp
->retrans_stamp
||
1740 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1741 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1744 /* Undo procedures. */
1746 #if FASTRETRANS_DEBUG > 1
1747 static void DBGUNDO(struct sock
*sk
, struct tcp_sock
*tp
, const char *msg
)
1749 struct inet_sock
*inet
= inet_sk(sk
);
1750 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1752 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1753 tp
->snd_cwnd
, tp
->left_out
,
1754 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1758 #define DBGUNDO(x...) do { } while (0)
1761 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1763 struct tcp_sock
*tp
= tcp_sk(sk
);
1765 if (tp
->prior_ssthresh
) {
1766 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1768 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1769 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1771 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1773 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1774 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1775 TCP_ECN_withdraw_cwr(tp
);
1778 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1780 tcp_moderate_cwnd(tp
);
1781 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1783 /* There is something screwy going on with the retrans hints after
1785 clear_all_retrans_hints(tp
);
1788 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1790 return tp
->undo_marker
&&
1791 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1794 /* People celebrate: "We love our President!" */
1795 static int tcp_try_undo_recovery(struct sock
*sk
, struct tcp_sock
*tp
)
1797 if (tcp_may_undo(tp
)) {
1798 /* Happy end! We did not retransmit anything
1799 * or our original transmission succeeded.
1801 DBGUNDO(sk
, tp
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1802 tcp_undo_cwr(sk
, 1);
1803 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1804 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1806 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1807 tp
->undo_marker
= 0;
1809 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1810 /* Hold old state until something *above* high_seq
1811 * is ACKed. For Reno it is MUST to prevent false
1812 * fast retransmits (RFC2582). SACK TCP is safe. */
1813 tcp_moderate_cwnd(tp
);
1816 tcp_set_ca_state(sk
, TCP_CA_Open
);
1820 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1821 static void tcp_try_undo_dsack(struct sock
*sk
, struct tcp_sock
*tp
)
1823 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1824 DBGUNDO(sk
, tp
, "D-SACK");
1825 tcp_undo_cwr(sk
, 1);
1826 tp
->undo_marker
= 0;
1827 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1831 /* Undo during fast recovery after partial ACK. */
1833 static int tcp_try_undo_partial(struct sock
*sk
, struct tcp_sock
*tp
,
1836 /* Partial ACK arrived. Force Hoe's retransmit. */
1837 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1839 if (tcp_may_undo(tp
)) {
1840 /* Plain luck! Hole if filled with delayed
1841 * packet, rather than with a retransmit.
1843 if (tp
->retrans_out
== 0)
1844 tp
->retrans_stamp
= 0;
1846 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
1848 DBGUNDO(sk
, tp
, "Hoe");
1849 tcp_undo_cwr(sk
, 0);
1850 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1852 /* So... Do not make Hoe's retransmit yet.
1853 * If the first packet was delayed, the rest
1854 * ones are most probably delayed as well.
1861 /* Undo during loss recovery after partial ACK. */
1862 static int tcp_try_undo_loss(struct sock
*sk
, struct tcp_sock
*tp
)
1864 if (tcp_may_undo(tp
)) {
1865 struct sk_buff
*skb
;
1866 sk_stream_for_retrans_queue(skb
, sk
) {
1867 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1870 clear_all_retrans_hints(tp
);
1872 DBGUNDO(sk
, tp
, "partial loss");
1874 tp
->left_out
= tp
->sacked_out
;
1875 tcp_undo_cwr(sk
, 1);
1876 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1877 inet_csk(sk
)->icsk_retransmits
= 0;
1878 tp
->undo_marker
= 0;
1880 tcp_set_ca_state(sk
, TCP_CA_Open
);
1886 static inline void tcp_complete_cwr(struct sock
*sk
)
1888 struct tcp_sock
*tp
= tcp_sk(sk
);
1889 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1890 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1891 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
1894 static void tcp_try_to_open(struct sock
*sk
, struct tcp_sock
*tp
, int flag
)
1896 tp
->left_out
= tp
->sacked_out
;
1898 if (tp
->retrans_out
== 0)
1899 tp
->retrans_stamp
= 0;
1904 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
1905 int state
= TCP_CA_Open
;
1907 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
1908 state
= TCP_CA_Disorder
;
1910 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
1911 tcp_set_ca_state(sk
, state
);
1912 tp
->high_seq
= tp
->snd_nxt
;
1914 tcp_moderate_cwnd(tp
);
1920 static void tcp_mtup_probe_failed(struct sock
*sk
)
1922 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1924 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
1925 icsk
->icsk_mtup
.probe_size
= 0;
1928 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
1930 struct tcp_sock
*tp
= tcp_sk(sk
);
1931 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1933 /* FIXME: breaks with very large cwnd */
1934 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1935 tp
->snd_cwnd
= tp
->snd_cwnd
*
1936 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
1937 icsk
->icsk_mtup
.probe_size
;
1938 tp
->snd_cwnd_cnt
= 0;
1939 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1940 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
1942 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
1943 icsk
->icsk_mtup
.probe_size
= 0;
1944 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
1948 /* Process an event, which can update packets-in-flight not trivially.
1949 * Main goal of this function is to calculate new estimate for left_out,
1950 * taking into account both packets sitting in receiver's buffer and
1951 * packets lost by network.
1953 * Besides that it does CWND reduction, when packet loss is detected
1954 * and changes state of machine.
1956 * It does _not_ decide what to send, it is made in function
1957 * tcp_xmit_retransmit_queue().
1960 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
1961 int prior_packets
, int flag
)
1963 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1964 struct tcp_sock
*tp
= tcp_sk(sk
);
1965 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
1967 /* Some technical things:
1968 * 1. Reno does not count dupacks (sacked_out) automatically. */
1969 if (!tp
->packets_out
)
1971 /* 2. SACK counts snd_fack in packets inaccurately. */
1972 if (tp
->sacked_out
== 0)
1973 tp
->fackets_out
= 0;
1975 /* Now state machine starts.
1976 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1978 tp
->prior_ssthresh
= 0;
1980 /* B. In all the states check for reneging SACKs. */
1981 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
1984 /* C. Process data loss notification, provided it is valid. */
1985 if ((flag
&FLAG_DATA_LOST
) &&
1986 before(tp
->snd_una
, tp
->high_seq
) &&
1987 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
1988 tp
->fackets_out
> tp
->reordering
) {
1989 tcp_mark_head_lost(sk
, tp
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
1990 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
1993 /* D. Synchronize left_out to current state. */
1994 tcp_sync_left_out(tp
);
1996 /* E. Check state exit conditions. State can be terminated
1997 * when high_seq is ACKed. */
1998 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
1999 if (!sysctl_tcp_frto
)
2000 BUG_TRAP(tp
->retrans_out
== 0);
2001 tp
->retrans_stamp
= 0;
2002 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2003 switch (icsk
->icsk_ca_state
) {
2005 icsk
->icsk_retransmits
= 0;
2006 if (tcp_try_undo_recovery(sk
, tp
))
2011 /* CWR is to be held something *above* high_seq
2012 * is ACKed for CWR bit to reach receiver. */
2013 if (tp
->snd_una
!= tp
->high_seq
) {
2014 tcp_complete_cwr(sk
);
2015 tcp_set_ca_state(sk
, TCP_CA_Open
);
2019 case TCP_CA_Disorder
:
2020 tcp_try_undo_dsack(sk
, tp
);
2021 if (!tp
->undo_marker
||
2022 /* For SACK case do not Open to allow to undo
2023 * catching for all duplicate ACKs. */
2024 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2025 tp
->undo_marker
= 0;
2026 tcp_set_ca_state(sk
, TCP_CA_Open
);
2030 case TCP_CA_Recovery
:
2032 tcp_reset_reno_sack(tp
);
2033 if (tcp_try_undo_recovery(sk
, tp
))
2035 tcp_complete_cwr(sk
);
2040 /* F. Process state. */
2041 switch (icsk
->icsk_ca_state
) {
2042 case TCP_CA_Recovery
:
2043 if (prior_snd_una
== tp
->snd_una
) {
2044 if (IsReno(tp
) && is_dupack
)
2045 tcp_add_reno_sack(sk
);
2047 int acked
= prior_packets
- tp
->packets_out
;
2049 tcp_remove_reno_sacks(sk
, tp
, acked
);
2050 is_dupack
= tcp_try_undo_partial(sk
, tp
, acked
);
2054 if (flag
&FLAG_DATA_ACKED
)
2055 icsk
->icsk_retransmits
= 0;
2056 if (!tcp_try_undo_loss(sk
, tp
)) {
2057 tcp_moderate_cwnd(tp
);
2058 tcp_xmit_retransmit_queue(sk
);
2061 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2063 /* Loss is undone; fall through to processing in Open state. */
2066 if (tp
->snd_una
!= prior_snd_una
)
2067 tcp_reset_reno_sack(tp
);
2069 tcp_add_reno_sack(sk
);
2072 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2073 tcp_try_undo_dsack(sk
, tp
);
2075 if (!tcp_time_to_recover(sk
, tp
)) {
2076 tcp_try_to_open(sk
, tp
, flag
);
2080 /* MTU probe failure: don't reduce cwnd */
2081 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2082 icsk
->icsk_mtup
.probe_size
&&
2083 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2084 tcp_mtup_probe_failed(sk
);
2085 /* Restores the reduction we did in tcp_mtup_probe() */
2087 tcp_simple_retransmit(sk
);
2091 /* Otherwise enter Recovery state */
2094 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2096 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2098 tp
->high_seq
= tp
->snd_nxt
;
2099 tp
->prior_ssthresh
= 0;
2100 tp
->undo_marker
= tp
->snd_una
;
2101 tp
->undo_retrans
= tp
->retrans_out
;
2103 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2104 if (!(flag
&FLAG_ECE
))
2105 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2106 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2107 TCP_ECN_queue_cwr(tp
);
2110 tp
->bytes_acked
= 0;
2111 tp
->snd_cwnd_cnt
= 0;
2112 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2115 if (is_dupack
|| tcp_head_timedout(sk
, tp
))
2116 tcp_update_scoreboard(sk
, tp
);
2118 tcp_xmit_retransmit_queue(sk
);
2121 /* Read draft-ietf-tcplw-high-performance before mucking
2122 * with this code. (Supersedes RFC1323)
2124 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2126 /* RTTM Rule: A TSecr value received in a segment is used to
2127 * update the averaged RTT measurement only if the segment
2128 * acknowledges some new data, i.e., only if it advances the
2129 * left edge of the send window.
2131 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2132 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2134 * Changed: reset backoff as soon as we see the first valid sample.
2135 * If we do not, we get strongly overestimated rto. With timestamps
2136 * samples are accepted even from very old segments: f.e., when rtt=1
2137 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2138 * answer arrives rto becomes 120 seconds! If at least one of segments
2139 * in window is lost... Voila. --ANK (010210)
2141 struct tcp_sock
*tp
= tcp_sk(sk
);
2142 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2143 tcp_rtt_estimator(sk
, seq_rtt
);
2145 inet_csk(sk
)->icsk_backoff
= 0;
2149 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2151 /* We don't have a timestamp. Can only use
2152 * packets that are not retransmitted to determine
2153 * rtt estimates. Also, we must not reset the
2154 * backoff for rto until we get a non-retransmitted
2155 * packet. This allows us to deal with a situation
2156 * where the network delay has increased suddenly.
2157 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2160 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2163 tcp_rtt_estimator(sk
, seq_rtt
);
2165 inet_csk(sk
)->icsk_backoff
= 0;
2169 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2172 const struct tcp_sock
*tp
= tcp_sk(sk
);
2173 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2174 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2175 tcp_ack_saw_tstamp(sk
, flag
);
2176 else if (seq_rtt
>= 0)
2177 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2180 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
2181 u32 in_flight
, int good
)
2183 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2184 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
2185 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2188 /* Restart timer after forward progress on connection.
2189 * RFC2988 recommends to restart timer to now+rto.
2192 static void tcp_ack_packets_out(struct sock
*sk
, struct tcp_sock
*tp
)
2194 if (!tp
->packets_out
) {
2195 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2197 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2201 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2202 __u32 now
, __s32
*seq_rtt
)
2204 struct tcp_sock
*tp
= tcp_sk(sk
);
2205 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2206 __u32 seq
= tp
->snd_una
;
2207 __u32 packets_acked
;
2210 /* If we get here, the whole TSO packet has not been
2213 BUG_ON(!after(scb
->end_seq
, seq
));
2215 packets_acked
= tcp_skb_pcount(skb
);
2216 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2218 packets_acked
-= tcp_skb_pcount(skb
);
2220 if (packets_acked
) {
2221 __u8 sacked
= scb
->sacked
;
2223 acked
|= FLAG_DATA_ACKED
;
2225 if (sacked
& TCPCB_RETRANS
) {
2226 if (sacked
& TCPCB_SACKED_RETRANS
)
2227 tp
->retrans_out
-= packets_acked
;
2228 acked
|= FLAG_RETRANS_DATA_ACKED
;
2230 } else if (*seq_rtt
< 0)
2231 *seq_rtt
= now
- scb
->when
;
2232 if (sacked
& TCPCB_SACKED_ACKED
)
2233 tp
->sacked_out
-= packets_acked
;
2234 if (sacked
& TCPCB_LOST
)
2235 tp
->lost_out
-= packets_acked
;
2236 if (sacked
& TCPCB_URG
) {
2238 !before(seq
, tp
->snd_up
))
2241 } else if (*seq_rtt
< 0)
2242 *seq_rtt
= now
- scb
->when
;
2244 if (tp
->fackets_out
) {
2245 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2246 tp
->fackets_out
-= dval
;
2248 tp
->packets_out
-= packets_acked
;
2250 BUG_ON(tcp_skb_pcount(skb
) == 0);
2251 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2257 static u32
tcp_usrtt(struct timeval
*tv
)
2261 do_gettimeofday(&now
);
2262 return (now
.tv_sec
- tv
->tv_sec
) * 1000000 + (now
.tv_usec
- tv
->tv_usec
);
2265 /* Remove acknowledged frames from the retransmission queue. */
2266 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2268 struct tcp_sock
*tp
= tcp_sk(sk
);
2269 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2270 struct sk_buff
*skb
;
2271 __u32 now
= tcp_time_stamp
;
2275 void (*rtt_sample
)(struct sock
*sk
, u32 usrtt
)
2276 = icsk
->icsk_ca_ops
->rtt_sample
;
2277 struct timeval tv
= { .tv_sec
= 0, .tv_usec
= 0 };
2279 while ((skb
= skb_peek(&sk
->sk_write_queue
)) &&
2280 skb
!= sk
->sk_send_head
) {
2281 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2282 __u8 sacked
= scb
->sacked
;
2284 /* If our packet is before the ack sequence we can
2285 * discard it as it's confirmed to have arrived at
2288 if (after(scb
->end_seq
, tp
->snd_una
)) {
2289 if (tcp_skb_pcount(skb
) > 1 &&
2290 after(tp
->snd_una
, scb
->seq
))
2291 acked
|= tcp_tso_acked(sk
, skb
,
2296 /* Initial outgoing SYN's get put onto the write_queue
2297 * just like anything else we transmit. It is not
2298 * true data, and if we misinform our callers that
2299 * this ACK acks real data, we will erroneously exit
2300 * connection startup slow start one packet too
2301 * quickly. This is severely frowned upon behavior.
2303 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2304 acked
|= FLAG_DATA_ACKED
;
2307 acked
|= FLAG_SYN_ACKED
;
2308 tp
->retrans_stamp
= 0;
2311 /* MTU probing checks */
2312 if (icsk
->icsk_mtup
.probe_size
) {
2313 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2314 tcp_mtup_probe_success(sk
, skb
);
2319 if (sacked
& TCPCB_RETRANS
) {
2320 if(sacked
& TCPCB_SACKED_RETRANS
)
2321 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2322 acked
|= FLAG_RETRANS_DATA_ACKED
;
2324 } else if (seq_rtt
< 0) {
2325 seq_rtt
= now
- scb
->when
;
2326 skb_get_timestamp(skb
, &tv
);
2328 if (sacked
& TCPCB_SACKED_ACKED
)
2329 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2330 if (sacked
& TCPCB_LOST
)
2331 tp
->lost_out
-= tcp_skb_pcount(skb
);
2332 if (sacked
& TCPCB_URG
) {
2334 !before(scb
->end_seq
, tp
->snd_up
))
2337 } else if (seq_rtt
< 0) {
2338 seq_rtt
= now
- scb
->when
;
2339 skb_get_timestamp(skb
, &tv
);
2341 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2342 tcp_packets_out_dec(tp
, skb
);
2343 __skb_unlink(skb
, &sk
->sk_write_queue
);
2344 sk_stream_free_skb(sk
, skb
);
2345 clear_all_retrans_hints(tp
);
2348 if (acked
&FLAG_ACKED
) {
2349 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2350 tcp_ack_packets_out(sk
, tp
);
2351 if (rtt_sample
&& !(acked
& FLAG_RETRANS_DATA_ACKED
))
2352 (*rtt_sample
)(sk
, tcp_usrtt(&tv
));
2354 if (icsk
->icsk_ca_ops
->pkts_acked
)
2355 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
);
2358 #if FASTRETRANS_DEBUG > 0
2359 BUG_TRAP((int)tp
->sacked_out
>= 0);
2360 BUG_TRAP((int)tp
->lost_out
>= 0);
2361 BUG_TRAP((int)tp
->retrans_out
>= 0);
2362 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2363 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2365 printk(KERN_DEBUG
"Leak l=%u %d\n",
2366 tp
->lost_out
, icsk
->icsk_ca_state
);
2369 if (tp
->sacked_out
) {
2370 printk(KERN_DEBUG
"Leak s=%u %d\n",
2371 tp
->sacked_out
, icsk
->icsk_ca_state
);
2374 if (tp
->retrans_out
) {
2375 printk(KERN_DEBUG
"Leak r=%u %d\n",
2376 tp
->retrans_out
, icsk
->icsk_ca_state
);
2377 tp
->retrans_out
= 0;
2381 *seq_rtt_p
= seq_rtt
;
2385 static void tcp_ack_probe(struct sock
*sk
)
2387 const struct tcp_sock
*tp
= tcp_sk(sk
);
2388 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2390 /* Was it a usable window open? */
2392 if (!after(TCP_SKB_CB(sk
->sk_send_head
)->end_seq
,
2393 tp
->snd_una
+ tp
->snd_wnd
)) {
2394 icsk
->icsk_backoff
= 0;
2395 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2396 /* Socket must be waked up by subsequent tcp_data_snd_check().
2397 * This function is not for random using!
2400 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2401 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2406 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2408 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2409 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2412 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2414 const struct tcp_sock
*tp
= tcp_sk(sk
);
2415 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2416 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2419 /* Check that window update is acceptable.
2420 * The function assumes that snd_una<=ack<=snd_next.
2422 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2423 const u32 ack_seq
, const u32 nwin
)
2425 return (after(ack
, tp
->snd_una
) ||
2426 after(ack_seq
, tp
->snd_wl1
) ||
2427 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2430 /* Update our send window.
2432 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2433 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2435 static int tcp_ack_update_window(struct sock
*sk
, struct tcp_sock
*tp
,
2436 struct sk_buff
*skb
, u32 ack
, u32 ack_seq
)
2439 u32 nwin
= ntohs(skb
->h
.th
->window
);
2441 if (likely(!skb
->h
.th
->syn
))
2442 nwin
<<= tp
->rx_opt
.snd_wscale
;
2444 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2445 flag
|= FLAG_WIN_UPDATE
;
2446 tcp_update_wl(tp
, ack
, ack_seq
);
2448 if (tp
->snd_wnd
!= nwin
) {
2451 /* Note, it is the only place, where
2452 * fast path is recovered for sending TCP.
2455 tcp_fast_path_check(sk
, tp
);
2457 if (nwin
> tp
->max_window
) {
2458 tp
->max_window
= nwin
;
2459 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2469 static void tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
)
2471 struct tcp_sock
*tp
= tcp_sk(sk
);
2473 tcp_sync_left_out(tp
);
2475 if (tp
->snd_una
== prior_snd_una
||
2476 !before(tp
->snd_una
, tp
->frto_highmark
)) {
2477 /* RTO was caused by loss, start retransmitting in
2478 * go-back-N slow start
2480 tcp_enter_frto_loss(sk
);
2484 if (tp
->frto_counter
== 1) {
2485 /* First ACK after RTO advances the window: allow two new
2488 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2490 /* Also the second ACK after RTO advances the window.
2491 * The RTO was likely spurious. Reduce cwnd and continue
2492 * in congestion avoidance
2494 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2495 tcp_moderate_cwnd(tp
);
2498 /* F-RTO affects on two new ACKs following RTO.
2499 * At latest on third ACK the TCP behavior is back to normal.
2501 tp
->frto_counter
= (tp
->frto_counter
+ 1) % 3;
2504 /* This routine deals with incoming acks, but not outgoing ones. */
2505 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2507 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2508 struct tcp_sock
*tp
= tcp_sk(sk
);
2509 u32 prior_snd_una
= tp
->snd_una
;
2510 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2511 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2512 u32 prior_in_flight
;
2516 /* If the ack is newer than sent or older than previous acks
2517 * then we can probably ignore it.
2519 if (after(ack
, tp
->snd_nxt
))
2520 goto uninteresting_ack
;
2522 if (before(ack
, prior_snd_una
))
2525 if (sysctl_tcp_abc
) {
2526 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2527 tp
->bytes_acked
+= ack
- prior_snd_una
;
2528 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2529 /* we assume just one segment left network */
2530 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2533 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2534 /* Window is constant, pure forward advance.
2535 * No more checks are required.
2536 * Note, we use the fact that SND.UNA>=SND.WL2.
2538 tcp_update_wl(tp
, ack
, ack_seq
);
2540 flag
|= FLAG_WIN_UPDATE
;
2542 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2544 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2546 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2549 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2551 flag
|= tcp_ack_update_window(sk
, tp
, skb
, ack
, ack_seq
);
2553 if (TCP_SKB_CB(skb
)->sacked
)
2554 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2556 if (TCP_ECN_rcv_ecn_echo(tp
, skb
->h
.th
))
2559 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2562 /* We passed data and got it acked, remove any soft error
2563 * log. Something worked...
2565 sk
->sk_err_soft
= 0;
2566 tp
->rcv_tstamp
= tcp_time_stamp
;
2567 prior_packets
= tp
->packets_out
;
2571 prior_in_flight
= tcp_packets_in_flight(tp
);
2573 /* See if we can take anything off of the retransmit queue. */
2574 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2576 if (tp
->frto_counter
)
2577 tcp_process_frto(sk
, prior_snd_una
);
2579 if (tcp_ack_is_dubious(sk
, flag
)) {
2580 /* Advance CWND, if state allows this. */
2581 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
2582 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2583 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2585 if ((flag
& FLAG_DATA_ACKED
))
2586 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2589 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2590 dst_confirm(sk
->sk_dst_cache
);
2595 icsk
->icsk_probes_out
= 0;
2597 /* If this ack opens up a zero window, clear backoff. It was
2598 * being used to time the probes, and is probably far higher than
2599 * it needs to be for normal retransmission.
2601 if (sk
->sk_send_head
)
2606 if (TCP_SKB_CB(skb
)->sacked
)
2607 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2610 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2615 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2616 * But, this can also be called on packets in the established flow when
2617 * the fast version below fails.
2619 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2622 struct tcphdr
*th
= skb
->h
.th
;
2623 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2625 ptr
= (unsigned char *)(th
+ 1);
2626 opt_rx
->saw_tstamp
= 0;
2635 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2640 if (opsize
< 2) /* "silly options" */
2642 if (opsize
> length
)
2643 return; /* don't parse partial options */
2646 if(opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2647 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
2649 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2650 in_mss
= opt_rx
->user_mss
;
2651 opt_rx
->mss_clamp
= in_mss
;
2656 if(opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2657 if (sysctl_tcp_window_scaling
) {
2658 __u8 snd_wscale
= *(__u8
*) ptr
;
2659 opt_rx
->wscale_ok
= 1;
2660 if (snd_wscale
> 14) {
2662 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2663 "scaling value %d >14 received.\n",
2667 opt_rx
->snd_wscale
= snd_wscale
;
2670 case TCPOPT_TIMESTAMP
:
2671 if(opsize
==TCPOLEN_TIMESTAMP
) {
2672 if ((estab
&& opt_rx
->tstamp_ok
) ||
2673 (!estab
&& sysctl_tcp_timestamps
)) {
2674 opt_rx
->saw_tstamp
= 1;
2675 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
2676 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
2680 case TCPOPT_SACK_PERM
:
2681 if(opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2682 if (sysctl_tcp_sack
) {
2683 opt_rx
->sack_ok
= 1;
2684 tcp_sack_reset(opt_rx
);
2690 if((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2691 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2693 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2695 #ifdef CONFIG_TCP_MD5SIG
2698 * The MD5 Hash has already been
2699 * checked (see tcp_v{4,6}_do_rcv()).
2710 /* Fast parse options. This hopes to only see timestamps.
2711 * If it is wrong it falls back on tcp_parse_options().
2713 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2714 struct tcp_sock
*tp
)
2716 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2717 tp
->rx_opt
.saw_tstamp
= 0;
2719 } else if (tp
->rx_opt
.tstamp_ok
&&
2720 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2721 __be32
*ptr
= (__be32
*)(th
+ 1);
2722 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2723 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2724 tp
->rx_opt
.saw_tstamp
= 1;
2726 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2728 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2732 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2736 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2738 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2739 tp
->rx_opt
.ts_recent_stamp
= xtime
.tv_sec
;
2742 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2744 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2745 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2746 * extra check below makes sure this can only happen
2747 * for pure ACK frames. -DaveM
2749 * Not only, also it occurs for expired timestamps.
2752 if((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2753 xtime
.tv_sec
>= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2754 tcp_store_ts_recent(tp
);
2758 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2760 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2761 * it can pass through stack. So, the following predicate verifies that
2762 * this segment is not used for anything but congestion avoidance or
2763 * fast retransmit. Moreover, we even are able to eliminate most of such
2764 * second order effects, if we apply some small "replay" window (~RTO)
2765 * to timestamp space.
2767 * All these measures still do not guarantee that we reject wrapped ACKs
2768 * on networks with high bandwidth, when sequence space is recycled fastly,
2769 * but it guarantees that such events will be very rare and do not affect
2770 * connection seriously. This doesn't look nice, but alas, PAWS is really
2773 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2774 * states that events when retransmit arrives after original data are rare.
2775 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2776 * the biggest problem on large power networks even with minor reordering.
2777 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2778 * up to bandwidth of 18Gigabit/sec. 8) ]
2781 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
2783 struct tcp_sock
*tp
= tcp_sk(sk
);
2784 struct tcphdr
*th
= skb
->h
.th
;
2785 u32 seq
= TCP_SKB_CB(skb
)->seq
;
2786 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2788 return (/* 1. Pure ACK with correct sequence number. */
2789 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
2791 /* 2. ... and duplicate ACK. */
2792 ack
== tp
->snd_una
&&
2794 /* 3. ... and does not update window. */
2795 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
2797 /* 4. ... and sits in replay window. */
2798 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
2801 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
2803 const struct tcp_sock
*tp
= tcp_sk(sk
);
2804 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
2805 xtime
.tv_sec
< tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
2806 !tcp_disordered_ack(sk
, skb
));
2809 /* Check segment sequence number for validity.
2811 * Segment controls are considered valid, if the segment
2812 * fits to the window after truncation to the window. Acceptability
2813 * of data (and SYN, FIN, of course) is checked separately.
2814 * See tcp_data_queue(), for example.
2816 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2817 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2818 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2819 * (borrowed from freebsd)
2822 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2824 return !before(end_seq
, tp
->rcv_wup
) &&
2825 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
2828 /* When we get a reset we do this. */
2829 static void tcp_reset(struct sock
*sk
)
2831 /* We want the right error as BSD sees it (and indeed as we do). */
2832 switch (sk
->sk_state
) {
2834 sk
->sk_err
= ECONNREFUSED
;
2836 case TCP_CLOSE_WAIT
:
2842 sk
->sk_err
= ECONNRESET
;
2845 if (!sock_flag(sk
, SOCK_DEAD
))
2846 sk
->sk_error_report(sk
);
2852 * Process the FIN bit. This now behaves as it is supposed to work
2853 * and the FIN takes effect when it is validly part of sequence
2854 * space. Not before when we get holes.
2856 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2857 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2860 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2861 * close and we go into CLOSING (and later onto TIME-WAIT)
2863 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2865 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
2867 struct tcp_sock
*tp
= tcp_sk(sk
);
2869 inet_csk_schedule_ack(sk
);
2871 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
2872 sock_set_flag(sk
, SOCK_DONE
);
2874 switch (sk
->sk_state
) {
2876 case TCP_ESTABLISHED
:
2877 /* Move to CLOSE_WAIT */
2878 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
2879 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
2882 case TCP_CLOSE_WAIT
:
2884 /* Received a retransmission of the FIN, do
2889 /* RFC793: Remain in the LAST-ACK state. */
2893 /* This case occurs when a simultaneous close
2894 * happens, we must ack the received FIN and
2895 * enter the CLOSING state.
2898 tcp_set_state(sk
, TCP_CLOSING
);
2901 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2903 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
2906 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2907 * cases we should never reach this piece of code.
2909 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
2910 __FUNCTION__
, sk
->sk_state
);
2914 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2915 * Probably, we should reset in this case. For now drop them.
2917 __skb_queue_purge(&tp
->out_of_order_queue
);
2918 if (tp
->rx_opt
.sack_ok
)
2919 tcp_sack_reset(&tp
->rx_opt
);
2920 sk_stream_mem_reclaim(sk
);
2922 if (!sock_flag(sk
, SOCK_DEAD
)) {
2923 sk
->sk_state_change(sk
);
2925 /* Do not send POLL_HUP for half duplex close. */
2926 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
2927 sk
->sk_state
== TCP_CLOSE
)
2928 sk_wake_async(sk
, 1, POLL_HUP
);
2930 sk_wake_async(sk
, 1, POLL_IN
);
2934 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
2936 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
2937 if (before(seq
, sp
->start_seq
))
2938 sp
->start_seq
= seq
;
2939 if (after(end_seq
, sp
->end_seq
))
2940 sp
->end_seq
= end_seq
;
2946 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2948 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2949 if (before(seq
, tp
->rcv_nxt
))
2950 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
2952 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
2954 tp
->rx_opt
.dsack
= 1;
2955 tp
->duplicate_sack
[0].start_seq
= seq
;
2956 tp
->duplicate_sack
[0].end_seq
= end_seq
;
2957 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
2961 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
2963 if (!tp
->rx_opt
.dsack
)
2964 tcp_dsack_set(tp
, seq
, end_seq
);
2966 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
2969 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
2971 struct tcp_sock
*tp
= tcp_sk(sk
);
2973 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
2974 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
2975 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
2976 tcp_enter_quickack_mode(sk
);
2978 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
2979 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2981 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
2982 end_seq
= tp
->rcv_nxt
;
2983 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
2990 /* These routines update the SACK block as out-of-order packets arrive or
2991 * in-order packets close up the sequence space.
2993 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
2996 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
2997 struct tcp_sack_block
*swalk
= sp
+1;
2999 /* See if the recent change to the first SACK eats into
3000 * or hits the sequence space of other SACK blocks, if so coalesce.
3002 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3003 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3006 /* Zap SWALK, by moving every further SACK up by one slot.
3007 * Decrease num_sacks.
3009 tp
->rx_opt
.num_sacks
--;
3010 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3011 for(i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3015 this_sack
++, swalk
++;
3019 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3023 tmp
= sack1
->start_seq
;
3024 sack1
->start_seq
= sack2
->start_seq
;
3025 sack2
->start_seq
= tmp
;
3027 tmp
= sack1
->end_seq
;
3028 sack1
->end_seq
= sack2
->end_seq
;
3029 sack2
->end_seq
= tmp
;
3032 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3034 struct tcp_sock
*tp
= tcp_sk(sk
);
3035 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3036 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3042 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3043 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3044 /* Rotate this_sack to the first one. */
3045 for (; this_sack
>0; this_sack
--, sp
--)
3046 tcp_sack_swap(sp
, sp
-1);
3048 tcp_sack_maybe_coalesce(tp
);
3053 /* Could not find an adjacent existing SACK, build a new one,
3054 * put it at the front, and shift everyone else down. We
3055 * always know there is at least one SACK present already here.
3057 * If the sack array is full, forget about the last one.
3059 if (this_sack
>= 4) {
3061 tp
->rx_opt
.num_sacks
--;
3064 for(; this_sack
> 0; this_sack
--, sp
--)
3068 /* Build the new head SACK, and we're done. */
3069 sp
->start_seq
= seq
;
3070 sp
->end_seq
= end_seq
;
3071 tp
->rx_opt
.num_sacks
++;
3072 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3075 /* RCV.NXT advances, some SACKs should be eaten. */
3077 static void tcp_sack_remove(struct tcp_sock
*tp
)
3079 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3080 int num_sacks
= tp
->rx_opt
.num_sacks
;
3083 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3084 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3085 tp
->rx_opt
.num_sacks
= 0;
3086 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3090 for(this_sack
= 0; this_sack
< num_sacks
; ) {
3091 /* Check if the start of the sack is covered by RCV.NXT. */
3092 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3095 /* RCV.NXT must cover all the block! */
3096 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3098 /* Zap this SACK, by moving forward any other SACKS. */
3099 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3100 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3107 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3108 tp
->rx_opt
.num_sacks
= num_sacks
;
3109 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3113 /* This one checks to see if we can put data from the
3114 * out_of_order queue into the receive_queue.
3116 static void tcp_ofo_queue(struct sock
*sk
)
3118 struct tcp_sock
*tp
= tcp_sk(sk
);
3119 __u32 dsack_high
= tp
->rcv_nxt
;
3120 struct sk_buff
*skb
;
3122 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3123 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3126 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3127 __u32 dsack
= dsack_high
;
3128 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3129 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3130 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3133 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3134 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3135 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3139 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3140 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3141 TCP_SKB_CB(skb
)->end_seq
);
3143 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3144 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3145 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3147 tcp_fin(skb
, sk
, skb
->h
.th
);
3151 static int tcp_prune_queue(struct sock
*sk
);
3153 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3155 struct tcphdr
*th
= skb
->h
.th
;
3156 struct tcp_sock
*tp
= tcp_sk(sk
);
3159 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3162 __skb_pull(skb
, th
->doff
*4);
3164 TCP_ECN_accept_cwr(tp
, skb
);
3166 if (tp
->rx_opt
.dsack
) {
3167 tp
->rx_opt
.dsack
= 0;
3168 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3169 4 - tp
->rx_opt
.tstamp_ok
);
3172 /* Queue data for delivery to the user.
3173 * Packets in sequence go to the receive queue.
3174 * Out of sequence packets to the out_of_order_queue.
3176 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3177 if (tcp_receive_window(tp
) == 0)
3180 /* Ok. In sequence. In window. */
3181 if (tp
->ucopy
.task
== current
&&
3182 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3183 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3184 int chunk
= min_t(unsigned int, skb
->len
,
3187 __set_current_state(TASK_RUNNING
);
3190 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3191 tp
->ucopy
.len
-= chunk
;
3192 tp
->copied_seq
+= chunk
;
3193 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3194 tcp_rcv_space_adjust(sk
);
3202 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3203 !sk_stream_rmem_schedule(sk
, skb
))) {
3204 if (tcp_prune_queue(sk
) < 0 ||
3205 !sk_stream_rmem_schedule(sk
, skb
))
3208 sk_stream_set_owner_r(skb
, sk
);
3209 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3211 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3213 tcp_event_data_recv(sk
, tp
, skb
);
3215 tcp_fin(skb
, sk
, th
);
3217 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3220 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3221 * gap in queue is filled.
3223 if (skb_queue_empty(&tp
->out_of_order_queue
))
3224 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3227 if (tp
->rx_opt
.num_sacks
)
3228 tcp_sack_remove(tp
);
3230 tcp_fast_path_check(sk
, tp
);
3234 else if (!sock_flag(sk
, SOCK_DEAD
))
3235 sk
->sk_data_ready(sk
, 0);
3239 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3240 /* A retransmit, 2nd most common case. Force an immediate ack. */
3241 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3242 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3245 tcp_enter_quickack_mode(sk
);
3246 inet_csk_schedule_ack(sk
);
3252 /* Out of window. F.e. zero window probe. */
3253 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3256 tcp_enter_quickack_mode(sk
);
3258 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3259 /* Partial packet, seq < rcv_next < end_seq */
3260 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3261 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3262 TCP_SKB_CB(skb
)->end_seq
);
3264 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3266 /* If window is closed, drop tail of packet. But after
3267 * remembering D-SACK for its head made in previous line.
3269 if (!tcp_receive_window(tp
))
3274 TCP_ECN_check_ce(tp
, skb
);
3276 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3277 !sk_stream_rmem_schedule(sk
, skb
)) {
3278 if (tcp_prune_queue(sk
) < 0 ||
3279 !sk_stream_rmem_schedule(sk
, skb
))
3283 /* Disable header prediction. */
3285 inet_csk_schedule_ack(sk
);
3287 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3288 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3290 sk_stream_set_owner_r(skb
, sk
);
3292 if (!skb_peek(&tp
->out_of_order_queue
)) {
3293 /* Initial out of order segment, build 1 SACK. */
3294 if (tp
->rx_opt
.sack_ok
) {
3295 tp
->rx_opt
.num_sacks
= 1;
3296 tp
->rx_opt
.dsack
= 0;
3297 tp
->rx_opt
.eff_sacks
= 1;
3298 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3299 tp
->selective_acks
[0].end_seq
=
3300 TCP_SKB_CB(skb
)->end_seq
;
3302 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3304 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3305 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3306 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3308 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3309 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3311 if (!tp
->rx_opt
.num_sacks
||
3312 tp
->selective_acks
[0].end_seq
!= seq
)
3315 /* Common case: data arrive in order after hole. */
3316 tp
->selective_acks
[0].end_seq
= end_seq
;
3320 /* Find place to insert this segment. */
3322 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3324 } while ((skb1
= skb1
->prev
) !=
3325 (struct sk_buff
*)&tp
->out_of_order_queue
);
3327 /* Do skb overlap to previous one? */
3328 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3329 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3330 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3331 /* All the bits are present. Drop. */
3333 tcp_dsack_set(tp
, seq
, end_seq
);
3336 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3337 /* Partial overlap. */
3338 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3343 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3345 /* And clean segments covered by new one as whole. */
3346 while ((skb1
= skb
->next
) !=
3347 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3348 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3349 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3350 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3353 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3354 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3359 if (tp
->rx_opt
.sack_ok
)
3360 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3364 /* Collapse contiguous sequence of skbs head..tail with
3365 * sequence numbers start..end.
3366 * Segments with FIN/SYN are not collapsed (only because this
3370 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3371 struct sk_buff
*head
, struct sk_buff
*tail
,
3374 struct sk_buff
*skb
;
3376 /* First, check that queue is collapsible and find
3377 * the point where collapsing can be useful. */
3378 for (skb
= head
; skb
!= tail
; ) {
3379 /* No new bits? It is possible on ofo queue. */
3380 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3381 struct sk_buff
*next
= skb
->next
;
3382 __skb_unlink(skb
, list
);
3384 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3389 /* The first skb to collapse is:
3391 * - bloated or contains data before "start" or
3392 * overlaps to the next one.
3394 if (!skb
->h
.th
->syn
&& !skb
->h
.th
->fin
&&
3395 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3396 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3397 (skb
->next
!= tail
&&
3398 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3401 /* Decided to skip this, advance start seq. */
3402 start
= TCP_SKB_CB(skb
)->end_seq
;
3405 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3408 while (before(start
, end
)) {
3409 struct sk_buff
*nskb
;
3410 int header
= skb_headroom(skb
);
3411 int copy
= SKB_MAX_ORDER(header
, 0);
3413 /* Too big header? This can happen with IPv6. */
3416 if (end
-start
< copy
)
3418 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3421 skb_reserve(nskb
, header
);
3422 memcpy(nskb
->head
, skb
->head
, header
);
3423 nskb
->nh
.raw
= nskb
->head
+ (skb
->nh
.raw
-skb
->head
);
3424 nskb
->h
.raw
= nskb
->head
+ (skb
->h
.raw
-skb
->head
);
3425 nskb
->mac
.raw
= nskb
->head
+ (skb
->mac
.raw
-skb
->head
);
3426 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3427 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3428 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3429 sk_stream_set_owner_r(nskb
, sk
);
3431 /* Copy data, releasing collapsed skbs. */
3433 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3434 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3438 size
= min(copy
, size
);
3439 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3441 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3445 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3446 struct sk_buff
*next
= skb
->next
;
3447 __skb_unlink(skb
, list
);
3449 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3451 if (skb
== tail
|| skb
->h
.th
->syn
|| skb
->h
.th
->fin
)
3458 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3459 * and tcp_collapse() them until all the queue is collapsed.
3461 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3463 struct tcp_sock
*tp
= tcp_sk(sk
);
3464 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3465 struct sk_buff
*head
;
3471 start
= TCP_SKB_CB(skb
)->seq
;
3472 end
= TCP_SKB_CB(skb
)->end_seq
;
3478 /* Segment is terminated when we see gap or when
3479 * we are at the end of all the queue. */
3480 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3481 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3482 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3483 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3484 head
, skb
, start
, end
);
3486 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3488 /* Start new segment */
3489 start
= TCP_SKB_CB(skb
)->seq
;
3490 end
= TCP_SKB_CB(skb
)->end_seq
;
3492 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3493 start
= TCP_SKB_CB(skb
)->seq
;
3494 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3495 end
= TCP_SKB_CB(skb
)->end_seq
;
3500 /* Reduce allocated memory if we can, trying to get
3501 * the socket within its memory limits again.
3503 * Return less than zero if we should start dropping frames
3504 * until the socket owning process reads some of the data
3505 * to stabilize the situation.
3507 static int tcp_prune_queue(struct sock
*sk
)
3509 struct tcp_sock
*tp
= tcp_sk(sk
);
3511 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3513 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3515 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3516 tcp_clamp_window(sk
, tp
);
3517 else if (tcp_memory_pressure
)
3518 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3520 tcp_collapse_ofo_queue(sk
);
3521 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3522 sk
->sk_receive_queue
.next
,
3523 (struct sk_buff
*)&sk
->sk_receive_queue
,
3524 tp
->copied_seq
, tp
->rcv_nxt
);
3525 sk_stream_mem_reclaim(sk
);
3527 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3530 /* Collapsing did not help, destructive actions follow.
3531 * This must not ever occur. */
3533 /* First, purge the out_of_order queue. */
3534 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3535 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3536 __skb_queue_purge(&tp
->out_of_order_queue
);
3538 /* Reset SACK state. A conforming SACK implementation will
3539 * do the same at a timeout based retransmit. When a connection
3540 * is in a sad state like this, we care only about integrity
3541 * of the connection not performance.
3543 if (tp
->rx_opt
.sack_ok
)
3544 tcp_sack_reset(&tp
->rx_opt
);
3545 sk_stream_mem_reclaim(sk
);
3548 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3551 /* If we are really being abused, tell the caller to silently
3552 * drop receive data on the floor. It will get retransmitted
3553 * and hopefully then we'll have sufficient space.
3555 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3557 /* Massive buffer overcommit. */
3563 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3564 * As additional protections, we do not touch cwnd in retransmission phases,
3565 * and if application hit its sndbuf limit recently.
3567 void tcp_cwnd_application_limited(struct sock
*sk
)
3569 struct tcp_sock
*tp
= tcp_sk(sk
);
3571 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3572 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3573 /* Limited by application or receiver window. */
3574 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
3575 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
3576 if (win_used
< tp
->snd_cwnd
) {
3577 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3578 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3580 tp
->snd_cwnd_used
= 0;
3582 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3585 static int tcp_should_expand_sndbuf(struct sock
*sk
, struct tcp_sock
*tp
)
3587 /* If the user specified a specific send buffer setting, do
3590 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3593 /* If we are under global TCP memory pressure, do not expand. */
3594 if (tcp_memory_pressure
)
3597 /* If we are under soft global TCP memory pressure, do not expand. */
3598 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3601 /* If we filled the congestion window, do not expand. */
3602 if (tp
->packets_out
>= tp
->snd_cwnd
)
3608 /* When incoming ACK allowed to free some skb from write_queue,
3609 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3610 * on the exit from tcp input handler.
3612 * PROBLEM: sndbuf expansion does not work well with largesend.
3614 static void tcp_new_space(struct sock
*sk
)
3616 struct tcp_sock
*tp
= tcp_sk(sk
);
3618 if (tcp_should_expand_sndbuf(sk
, tp
)) {
3619 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3620 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3621 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3622 tp
->reordering
+ 1);
3623 sndmem
*= 2*demanded
;
3624 if (sndmem
> sk
->sk_sndbuf
)
3625 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3626 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3629 sk
->sk_write_space(sk
);
3632 static void tcp_check_space(struct sock
*sk
)
3634 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3635 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3636 if (sk
->sk_socket
&&
3637 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3642 static inline void tcp_data_snd_check(struct sock
*sk
, struct tcp_sock
*tp
)
3644 tcp_push_pending_frames(sk
, tp
);
3645 tcp_check_space(sk
);
3649 * Check if sending an ack is needed.
3651 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3653 struct tcp_sock
*tp
= tcp_sk(sk
);
3655 /* More than one full frame received... */
3656 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3657 /* ... and right edge of window advances far enough.
3658 * (tcp_recvmsg() will send ACK otherwise). Or...
3660 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3661 /* We ACK each frame or... */
3662 tcp_in_quickack_mode(sk
) ||
3663 /* We have out of order data. */
3665 skb_peek(&tp
->out_of_order_queue
))) {
3666 /* Then ack it now */
3669 /* Else, send delayed ack. */
3670 tcp_send_delayed_ack(sk
);
3674 static inline void tcp_ack_snd_check(struct sock
*sk
)
3676 if (!inet_csk_ack_scheduled(sk
)) {
3677 /* We sent a data segment already. */
3680 __tcp_ack_snd_check(sk
, 1);
3684 * This routine is only called when we have urgent data
3685 * signaled. Its the 'slow' part of tcp_urg. It could be
3686 * moved inline now as tcp_urg is only called from one
3687 * place. We handle URGent data wrong. We have to - as
3688 * BSD still doesn't use the correction from RFC961.
3689 * For 1003.1g we should support a new option TCP_STDURG to permit
3690 * either form (or just set the sysctl tcp_stdurg).
3693 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3695 struct tcp_sock
*tp
= tcp_sk(sk
);
3696 u32 ptr
= ntohs(th
->urg_ptr
);
3698 if (ptr
&& !sysctl_tcp_stdurg
)
3700 ptr
+= ntohl(th
->seq
);
3702 /* Ignore urgent data that we've already seen and read. */
3703 if (after(tp
->copied_seq
, ptr
))
3706 /* Do not replay urg ptr.
3708 * NOTE: interesting situation not covered by specs.
3709 * Misbehaving sender may send urg ptr, pointing to segment,
3710 * which we already have in ofo queue. We are not able to fetch
3711 * such data and will stay in TCP_URG_NOTYET until will be eaten
3712 * by recvmsg(). Seems, we are not obliged to handle such wicked
3713 * situations. But it is worth to think about possibility of some
3714 * DoSes using some hypothetical application level deadlock.
3716 if (before(ptr
, tp
->rcv_nxt
))
3719 /* Do we already have a newer (or duplicate) urgent pointer? */
3720 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3723 /* Tell the world about our new urgent pointer. */
3726 /* We may be adding urgent data when the last byte read was
3727 * urgent. To do this requires some care. We cannot just ignore
3728 * tp->copied_seq since we would read the last urgent byte again
3729 * as data, nor can we alter copied_seq until this data arrives
3730 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3732 * NOTE. Double Dutch. Rendering to plain English: author of comment
3733 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3734 * and expect that both A and B disappear from stream. This is _wrong_.
3735 * Though this happens in BSD with high probability, this is occasional.
3736 * Any application relying on this is buggy. Note also, that fix "works"
3737 * only in this artificial test. Insert some normal data between A and B and we will
3738 * decline of BSD again. Verdict: it is better to remove to trap
3741 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3742 !sock_flag(sk
, SOCK_URGINLINE
) &&
3743 tp
->copied_seq
!= tp
->rcv_nxt
) {
3744 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3746 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3747 __skb_unlink(skb
, &sk
->sk_receive_queue
);
3752 tp
->urg_data
= TCP_URG_NOTYET
;
3755 /* Disable header prediction. */
3759 /* This is the 'fast' part of urgent handling. */
3760 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
3762 struct tcp_sock
*tp
= tcp_sk(sk
);
3764 /* Check if we get a new urgent pointer - normally not. */
3766 tcp_check_urg(sk
,th
);
3768 /* Do we wait for any urgent data? - normally not... */
3769 if (tp
->urg_data
== TCP_URG_NOTYET
) {
3770 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
3773 /* Is the urgent pointer pointing into this packet? */
3774 if (ptr
< skb
->len
) {
3776 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
3778 tp
->urg_data
= TCP_URG_VALID
| tmp
;
3779 if (!sock_flag(sk
, SOCK_DEAD
))
3780 sk
->sk_data_ready(sk
, 0);
3785 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3787 struct tcp_sock
*tp
= tcp_sk(sk
);
3788 int chunk
= skb
->len
- hlen
;
3792 if (skb
->ip_summed
==CHECKSUM_UNNECESSARY
)
3793 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
3795 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
3799 tp
->ucopy
.len
-= chunk
;
3800 tp
->copied_seq
+= chunk
;
3801 tcp_rcv_space_adjust(sk
);
3808 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3812 if (sock_owned_by_user(sk
)) {
3814 result
= __tcp_checksum_complete(skb
);
3817 result
= __tcp_checksum_complete(skb
);
3822 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
3824 return skb
->ip_summed
!= CHECKSUM_UNNECESSARY
&&
3825 __tcp_checksum_complete_user(sk
, skb
);
3828 #ifdef CONFIG_NET_DMA
3829 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
3831 struct tcp_sock
*tp
= tcp_sk(sk
);
3832 int chunk
= skb
->len
- hlen
;
3834 int copied_early
= 0;
3836 if (tp
->ucopy
.wakeup
)
3839 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
3840 tp
->ucopy
.dma_chan
= get_softnet_dma();
3842 if (tp
->ucopy
.dma_chan
&& skb
->ip_summed
== CHECKSUM_UNNECESSARY
) {
3844 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
3845 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
3850 tp
->ucopy
.dma_cookie
= dma_cookie
;
3853 tp
->ucopy
.len
-= chunk
;
3854 tp
->copied_seq
+= chunk
;
3855 tcp_rcv_space_adjust(sk
);
3857 if ((tp
->ucopy
.len
== 0) ||
3858 (tcp_flag_word(skb
->h
.th
) & TCP_FLAG_PSH
) ||
3859 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
3860 tp
->ucopy
.wakeup
= 1;
3861 sk
->sk_data_ready(sk
, 0);
3863 } else if (chunk
> 0) {
3864 tp
->ucopy
.wakeup
= 1;
3865 sk
->sk_data_ready(sk
, 0);
3868 return copied_early
;
3870 #endif /* CONFIG_NET_DMA */
3873 * TCP receive function for the ESTABLISHED state.
3875 * It is split into a fast path and a slow path. The fast path is
3877 * - A zero window was announced from us - zero window probing
3878 * is only handled properly in the slow path.
3879 * - Out of order segments arrived.
3880 * - Urgent data is expected.
3881 * - There is no buffer space left
3882 * - Unexpected TCP flags/window values/header lengths are received
3883 * (detected by checking the TCP header against pred_flags)
3884 * - Data is sent in both directions. Fast path only supports pure senders
3885 * or pure receivers (this means either the sequence number or the ack
3886 * value must stay constant)
3887 * - Unexpected TCP option.
3889 * When these conditions are not satisfied it drops into a standard
3890 * receive procedure patterned after RFC793 to handle all cases.
3891 * The first three cases are guaranteed by proper pred_flags setting,
3892 * the rest is checked inline. Fast processing is turned on in
3893 * tcp_data_queue when everything is OK.
3895 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
3896 struct tcphdr
*th
, unsigned len
)
3898 struct tcp_sock
*tp
= tcp_sk(sk
);
3901 * Header prediction.
3902 * The code loosely follows the one in the famous
3903 * "30 instruction TCP receive" Van Jacobson mail.
3905 * Van's trick is to deposit buffers into socket queue
3906 * on a device interrupt, to call tcp_recv function
3907 * on the receive process context and checksum and copy
3908 * the buffer to user space. smart...
3910 * Our current scheme is not silly either but we take the
3911 * extra cost of the net_bh soft interrupt processing...
3912 * We do checksum and copy also but from device to kernel.
3915 tp
->rx_opt
.saw_tstamp
= 0;
3917 /* pred_flags is 0xS?10 << 16 + snd_wnd
3918 * if header_prediction is to be made
3919 * 'S' will always be tp->tcp_header_len >> 2
3920 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3921 * turn it off (when there are holes in the receive
3922 * space for instance)
3923 * PSH flag is ignored.
3926 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
3927 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3928 int tcp_header_len
= tp
->tcp_header_len
;
3930 /* Timestamp header prediction: tcp_header_len
3931 * is automatically equal to th->doff*4 due to pred_flags
3935 /* Check timestamp */
3936 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
3937 __be32
*ptr
= (__be32
*)(th
+ 1);
3939 /* No? Slow path! */
3940 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3941 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
3944 tp
->rx_opt
.saw_tstamp
= 1;
3946 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3948 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3950 /* If PAWS failed, check it more carefully in slow path */
3951 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
3954 /* DO NOT update ts_recent here, if checksum fails
3955 * and timestamp was corrupted part, it will result
3956 * in a hung connection since we will drop all
3957 * future packets due to the PAWS test.
3961 if (len
<= tcp_header_len
) {
3962 /* Bulk data transfer: sender */
3963 if (len
== tcp_header_len
) {
3964 /* Predicted packet is in window by definition.
3965 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
3966 * Hence, check seq<=rcv_wup reduces to:
3968 if (tcp_header_len
==
3969 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
3970 tp
->rcv_nxt
== tp
->rcv_wup
)
3971 tcp_store_ts_recent(tp
);
3973 /* We know that such packets are checksummed
3976 tcp_ack(sk
, skb
, 0);
3978 tcp_data_snd_check(sk
, tp
);
3980 } else { /* Header too small */
3981 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
3986 int copied_early
= 0;
3988 if (tp
->copied_seq
== tp
->rcv_nxt
&&
3989 len
- tcp_header_len
<= tp
->ucopy
.len
) {
3990 #ifdef CONFIG_NET_DMA
3991 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
3996 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
3997 __set_current_state(TASK_RUNNING
);
3999 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4003 /* Predicted packet is in window by definition.
4004 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4005 * Hence, check seq<=rcv_wup reduces to:
4007 if (tcp_header_len
==
4008 (sizeof(struct tcphdr
) +
4009 TCPOLEN_TSTAMP_ALIGNED
) &&
4010 tp
->rcv_nxt
== tp
->rcv_wup
)
4011 tcp_store_ts_recent(tp
);
4013 tcp_rcv_rtt_measure_ts(sk
, skb
);
4015 __skb_pull(skb
, tcp_header_len
);
4016 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4017 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4020 tcp_cleanup_rbuf(sk
, skb
->len
);
4023 if (tcp_checksum_complete_user(sk
, skb
))
4026 /* Predicted packet is in window by definition.
4027 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4028 * Hence, check seq<=rcv_wup reduces to:
4030 if (tcp_header_len
==
4031 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4032 tp
->rcv_nxt
== tp
->rcv_wup
)
4033 tcp_store_ts_recent(tp
);
4035 tcp_rcv_rtt_measure_ts(sk
, skb
);
4037 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4040 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4042 /* Bulk data transfer: receiver */
4043 __skb_pull(skb
,tcp_header_len
);
4044 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4045 sk_stream_set_owner_r(skb
, sk
);
4046 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4049 tcp_event_data_recv(sk
, tp
, skb
);
4051 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4052 /* Well, only one small jumplet in fast path... */
4053 tcp_ack(sk
, skb
, FLAG_DATA
);
4054 tcp_data_snd_check(sk
, tp
);
4055 if (!inet_csk_ack_scheduled(sk
))
4059 __tcp_ack_snd_check(sk
, 0);
4061 #ifdef CONFIG_NET_DMA
4063 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4069 sk
->sk_data_ready(sk
, 0);
4075 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4079 * RFC1323: H1. Apply PAWS check first.
4081 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4082 tcp_paws_discard(sk
, skb
)) {
4084 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4085 tcp_send_dupack(sk
, skb
);
4088 /* Resets are accepted even if PAWS failed.
4090 ts_recent update must be made after we are sure
4091 that the packet is in window.
4096 * Standard slow path.
4099 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4100 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4101 * (RST) segments are validated by checking their SEQ-fields."
4102 * And page 69: "If an incoming segment is not acceptable,
4103 * an acknowledgment should be sent in reply (unless the RST bit
4104 * is set, if so drop the segment and return)".
4107 tcp_send_dupack(sk
, skb
);
4116 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4118 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4119 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4120 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4127 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4129 tcp_rcv_rtt_measure_ts(sk
, skb
);
4131 /* Process urgent data. */
4132 tcp_urg(sk
, skb
, th
);
4134 /* step 7: process the segment text */
4135 tcp_data_queue(sk
, skb
);
4137 tcp_data_snd_check(sk
, tp
);
4138 tcp_ack_snd_check(sk
);
4142 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4149 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4150 struct tcphdr
*th
, unsigned len
)
4152 struct tcp_sock
*tp
= tcp_sk(sk
);
4153 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4154 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4156 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4160 * "If the state is SYN-SENT then
4161 * first check the ACK bit
4162 * If the ACK bit is set
4163 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4164 * a reset (unless the RST bit is set, if so drop
4165 * the segment and return)"
4167 * We do not send data with SYN, so that RFC-correct
4170 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4171 goto reset_and_undo
;
4173 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4174 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4176 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4177 goto reset_and_undo
;
4180 /* Now ACK is acceptable.
4182 * "If the RST bit is set
4183 * If the ACK was acceptable then signal the user "error:
4184 * connection reset", drop the segment, enter CLOSED state,
4185 * delete TCB, and return."
4194 * "fifth, if neither of the SYN or RST bits is set then
4195 * drop the segment and return."
4201 goto discard_and_undo
;
4204 * "If the SYN bit is on ...
4205 * are acceptable then ...
4206 * (our SYN has been ACKed), change the connection
4207 * state to ESTABLISHED..."
4210 TCP_ECN_rcv_synack(tp
, th
);
4212 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4213 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4215 /* Ok.. it's good. Set up sequence numbers and
4216 * move to established.
4218 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4219 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4221 /* RFC1323: The window in SYN & SYN/ACK segments is
4224 tp
->snd_wnd
= ntohs(th
->window
);
4225 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4227 if (!tp
->rx_opt
.wscale_ok
) {
4228 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4229 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4232 if (tp
->rx_opt
.saw_tstamp
) {
4233 tp
->rx_opt
.tstamp_ok
= 1;
4234 tp
->tcp_header_len
=
4235 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4236 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4237 tcp_store_ts_recent(tp
);
4239 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4242 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4243 tp
->rx_opt
.sack_ok
|= 2;
4246 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4247 tcp_initialize_rcv_mss(sk
);
4249 /* Remember, tcp_poll() does not lock socket!
4250 * Change state from SYN-SENT only after copied_seq
4251 * is initialized. */
4252 tp
->copied_seq
= tp
->rcv_nxt
;
4254 tcp_set_state(sk
, TCP_ESTABLISHED
);
4256 security_inet_conn_established(sk
, skb
);
4258 /* Make sure socket is routed, for correct metrics. */
4259 icsk
->icsk_af_ops
->rebuild_header(sk
);
4261 tcp_init_metrics(sk
);
4263 tcp_init_congestion_control(sk
);
4265 /* Prevent spurious tcp_cwnd_restart() on first data
4268 tp
->lsndtime
= tcp_time_stamp
;
4270 tcp_init_buffer_space(sk
);
4272 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4273 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4275 if (!tp
->rx_opt
.snd_wscale
)
4276 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4280 if (!sock_flag(sk
, SOCK_DEAD
)) {
4281 sk
->sk_state_change(sk
);
4282 sk_wake_async(sk
, 0, POLL_OUT
);
4285 if (sk
->sk_write_pending
||
4286 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4287 icsk
->icsk_ack
.pingpong
) {
4288 /* Save one ACK. Data will be ready after
4289 * several ticks, if write_pending is set.
4291 * It may be deleted, but with this feature tcpdumps
4292 * look so _wonderfully_ clever, that I was not able
4293 * to stand against the temptation 8) --ANK
4295 inet_csk_schedule_ack(sk
);
4296 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4297 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4298 tcp_incr_quickack(sk
);
4299 tcp_enter_quickack_mode(sk
);
4300 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4301 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4312 /* No ACK in the segment */
4316 * "If the RST bit is set
4318 * Otherwise (no ACK) drop the segment and return."
4321 goto discard_and_undo
;
4325 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4326 goto discard_and_undo
;
4329 /* We see SYN without ACK. It is attempt of
4330 * simultaneous connect with crossed SYNs.
4331 * Particularly, it can be connect to self.
4333 tcp_set_state(sk
, TCP_SYN_RECV
);
4335 if (tp
->rx_opt
.saw_tstamp
) {
4336 tp
->rx_opt
.tstamp_ok
= 1;
4337 tcp_store_ts_recent(tp
);
4338 tp
->tcp_header_len
=
4339 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4341 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4344 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4345 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4347 /* RFC1323: The window in SYN & SYN/ACK segments is
4350 tp
->snd_wnd
= ntohs(th
->window
);
4351 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4352 tp
->max_window
= tp
->snd_wnd
;
4354 TCP_ECN_rcv_syn(tp
, th
);
4357 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4358 tcp_initialize_rcv_mss(sk
);
4361 tcp_send_synack(sk
);
4363 /* Note, we could accept data and URG from this segment.
4364 * There are no obstacles to make this.
4366 * However, if we ignore data in ACKless segments sometimes,
4367 * we have no reasons to accept it sometimes.
4368 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4369 * is not flawless. So, discard packet for sanity.
4370 * Uncomment this return to process the data.
4377 /* "fifth, if neither of the SYN or RST bits is set then
4378 * drop the segment and return."
4382 tcp_clear_options(&tp
->rx_opt
);
4383 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4387 tcp_clear_options(&tp
->rx_opt
);
4388 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4394 * This function implements the receiving procedure of RFC 793 for
4395 * all states except ESTABLISHED and TIME_WAIT.
4396 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4397 * address independent.
4400 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4401 struct tcphdr
*th
, unsigned len
)
4403 struct tcp_sock
*tp
= tcp_sk(sk
);
4404 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4407 tp
->rx_opt
.saw_tstamp
= 0;
4409 switch (sk
->sk_state
) {
4421 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4424 /* Now we have several options: In theory there is
4425 * nothing else in the frame. KA9Q has an option to
4426 * send data with the syn, BSD accepts data with the
4427 * syn up to the [to be] advertised window and
4428 * Solaris 2.1 gives you a protocol error. For now
4429 * we just ignore it, that fits the spec precisely
4430 * and avoids incompatibilities. It would be nice in
4431 * future to drop through and process the data.
4433 * Now that TTCP is starting to be used we ought to
4435 * But, this leaves one open to an easy denial of
4436 * service attack, and SYN cookies can't defend
4437 * against this problem. So, we drop the data
4438 * in the interest of security over speed unless
4439 * it's still in use.
4447 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4451 /* Do step6 onward by hand. */
4452 tcp_urg(sk
, skb
, th
);
4454 tcp_data_snd_check(sk
, tp
);
4458 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4459 tcp_paws_discard(sk
, skb
)) {
4461 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4462 tcp_send_dupack(sk
, skb
);
4465 /* Reset is accepted even if it did not pass PAWS. */
4468 /* step 1: check sequence number */
4469 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4471 tcp_send_dupack(sk
, skb
);
4475 /* step 2: check RST bit */
4481 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4483 /* step 3: check security and precedence [ignored] */
4487 * Check for a SYN in window.
4489 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4490 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4495 /* step 5: check the ACK field */
4497 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4499 switch(sk
->sk_state
) {
4502 tp
->copied_seq
= tp
->rcv_nxt
;
4504 tcp_set_state(sk
, TCP_ESTABLISHED
);
4505 sk
->sk_state_change(sk
);
4507 /* Note, that this wakeup is only for marginal
4508 * crossed SYN case. Passively open sockets
4509 * are not waked up, because sk->sk_sleep ==
4510 * NULL and sk->sk_socket == NULL.
4512 if (sk
->sk_socket
) {
4513 sk_wake_async(sk
,0,POLL_OUT
);
4516 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4517 tp
->snd_wnd
= ntohs(th
->window
) <<
4518 tp
->rx_opt
.snd_wscale
;
4519 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4520 TCP_SKB_CB(skb
)->seq
);
4522 /* tcp_ack considers this ACK as duplicate
4523 * and does not calculate rtt.
4524 * Fix it at least with timestamps.
4526 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4528 tcp_ack_saw_tstamp(sk
, 0);
4530 if (tp
->rx_opt
.tstamp_ok
)
4531 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4533 /* Make sure socket is routed, for
4536 icsk
->icsk_af_ops
->rebuild_header(sk
);
4538 tcp_init_metrics(sk
);
4540 tcp_init_congestion_control(sk
);
4542 /* Prevent spurious tcp_cwnd_restart() on
4543 * first data packet.
4545 tp
->lsndtime
= tcp_time_stamp
;
4548 tcp_initialize_rcv_mss(sk
);
4549 tcp_init_buffer_space(sk
);
4550 tcp_fast_path_on(tp
);
4557 if (tp
->snd_una
== tp
->write_seq
) {
4558 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4559 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4560 dst_confirm(sk
->sk_dst_cache
);
4562 if (!sock_flag(sk
, SOCK_DEAD
))
4563 /* Wake up lingering close() */
4564 sk
->sk_state_change(sk
);
4568 if (tp
->linger2
< 0 ||
4569 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4570 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4572 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4576 tmo
= tcp_fin_time(sk
);
4577 if (tmo
> TCP_TIMEWAIT_LEN
) {
4578 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4579 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4580 /* Bad case. We could lose such FIN otherwise.
4581 * It is not a big problem, but it looks confusing
4582 * and not so rare event. We still can lose it now,
4583 * if it spins in bh_lock_sock(), but it is really
4586 inet_csk_reset_keepalive_timer(sk
, tmo
);
4588 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4596 if (tp
->snd_una
== tp
->write_seq
) {
4597 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4603 if (tp
->snd_una
== tp
->write_seq
) {
4604 tcp_update_metrics(sk
);
4613 /* step 6: check the URG bit */
4614 tcp_urg(sk
, skb
, th
);
4616 /* step 7: process the segment text */
4617 switch (sk
->sk_state
) {
4618 case TCP_CLOSE_WAIT
:
4621 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4625 /* RFC 793 says to queue data in these states,
4626 * RFC 1122 says we MUST send a reset.
4627 * BSD 4.4 also does reset.
4629 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4630 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4631 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4632 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4638 case TCP_ESTABLISHED
:
4639 tcp_data_queue(sk
, skb
);
4644 /* tcp_data could move socket to TIME-WAIT */
4645 if (sk
->sk_state
!= TCP_CLOSE
) {
4646 tcp_data_snd_check(sk
, tp
);
4647 tcp_ack_snd_check(sk
);
4657 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4658 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4659 EXPORT_SYMBOL(tcp_parse_options
);
4660 EXPORT_SYMBOL(tcp_rcv_established
);
4661 EXPORT_SYMBOL(tcp_rcv_state_process
);
4662 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);