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_frto_response __read_mostly
;
90 int sysctl_tcp_nometrics_save __read_mostly
;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
93 int sysctl_tcp_abc __read_mostly
;
95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
101 #define FLAG_ECE 0x40 /* ECE in this ACK */
102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
106 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
107 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
108 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
109 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
111 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
112 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
113 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
119 /* Adapt the MSS value used to make delayed ack decision to the
122 static void tcp_measure_rcv_mss(struct sock
*sk
,
123 const struct sk_buff
*skb
)
125 struct inet_connection_sock
*icsk
= inet_csk(sk
);
126 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
129 icsk
->icsk_ack
.last_seg_size
= 0;
131 /* skb->len may jitter because of SACKs, even if peer
132 * sends good full-sized frames.
134 len
= skb_shinfo(skb
)->gso_size
?: skb
->len
;
135 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
136 icsk
->icsk_ack
.rcv_mss
= len
;
138 /* Otherwise, we make more careful check taking into account,
139 * that SACKs block is variable.
141 * "len" is invariant segment length, including TCP header.
143 len
+= skb
->data
- skb_transport_header(skb
);
144 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
145 /* If PSH is not set, packet should be
146 * full sized, provided peer TCP is not badly broken.
147 * This observation (if it is correct 8)) allows
148 * to handle super-low mtu links fairly.
150 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
151 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
152 /* Subtract also invariant (if peer is RFC compliant),
153 * tcp header plus fixed timestamp option length.
154 * Resulting "len" is MSS free of SACK jitter.
156 len
-= tcp_sk(sk
)->tcp_header_len
;
157 icsk
->icsk_ack
.last_seg_size
= len
;
159 icsk
->icsk_ack
.rcv_mss
= len
;
163 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
164 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
165 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
169 static void tcp_incr_quickack(struct sock
*sk
)
171 struct inet_connection_sock
*icsk
= inet_csk(sk
);
172 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
176 if (quickacks
> icsk
->icsk_ack
.quick
)
177 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
180 void tcp_enter_quickack_mode(struct sock
*sk
)
182 struct inet_connection_sock
*icsk
= inet_csk(sk
);
183 tcp_incr_quickack(sk
);
184 icsk
->icsk_ack
.pingpong
= 0;
185 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
188 /* Send ACKs quickly, if "quick" count is not exhausted
189 * and the session is not interactive.
192 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
194 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
195 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
198 /* Buffer size and advertised window tuning.
200 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
203 static void tcp_fixup_sndbuf(struct sock
*sk
)
205 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
206 sizeof(struct sk_buff
);
208 if (sk
->sk_sndbuf
< 3 * sndmem
)
209 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
212 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
214 * All tcp_full_space() is split to two parts: "network" buffer, allocated
215 * forward and advertised in receiver window (tp->rcv_wnd) and
216 * "application buffer", required to isolate scheduling/application
217 * latencies from network.
218 * window_clamp is maximal advertised window. It can be less than
219 * tcp_full_space(), in this case tcp_full_space() - window_clamp
220 * is reserved for "application" buffer. The less window_clamp is
221 * the smoother our behaviour from viewpoint of network, but the lower
222 * throughput and the higher sensitivity of the connection to losses. 8)
224 * rcv_ssthresh is more strict window_clamp used at "slow start"
225 * phase to predict further behaviour of this connection.
226 * It is used for two goals:
227 * - to enforce header prediction at sender, even when application
228 * requires some significant "application buffer". It is check #1.
229 * - to prevent pruning of receive queue because of misprediction
230 * of receiver window. Check #2.
232 * The scheme does not work when sender sends good segments opening
233 * window and then starts to feed us spaghetti. But it should work
234 * in common situations. Otherwise, we have to rely on queue collapsing.
237 /* Slow part of check#2. */
238 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
240 struct tcp_sock
*tp
= tcp_sk(sk
);
242 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
243 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2])/2;
245 while (tp
->rcv_ssthresh
<= window
) {
246 if (truesize
<= skb
->len
)
247 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
255 static void tcp_grow_window(struct sock
*sk
,
258 struct tcp_sock
*tp
= tcp_sk(sk
);
261 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
262 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
263 !tcp_memory_pressure
) {
266 /* Check #2. Increase window, if skb with such overhead
267 * will fit to rcvbuf in future.
269 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
272 incr
= __tcp_grow_window(sk
, skb
);
275 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
276 inet_csk(sk
)->icsk_ack
.quick
|= 1;
281 /* 3. Tuning rcvbuf, when connection enters established state. */
283 static void tcp_fixup_rcvbuf(struct sock
*sk
)
285 struct tcp_sock
*tp
= tcp_sk(sk
);
286 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
288 /* Try to select rcvbuf so that 4 mss-sized segments
289 * will fit to window and corresponding skbs will fit to our rcvbuf.
290 * (was 3; 4 is minimum to allow fast retransmit to work.)
292 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
294 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
295 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
298 /* 4. Try to fixup all. It is made immediately after connection enters
301 static void tcp_init_buffer_space(struct sock
*sk
)
303 struct tcp_sock
*tp
= tcp_sk(sk
);
306 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
307 tcp_fixup_rcvbuf(sk
);
308 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
309 tcp_fixup_sndbuf(sk
);
311 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
313 maxwin
= tcp_full_space(sk
);
315 if (tp
->window_clamp
>= maxwin
) {
316 tp
->window_clamp
= maxwin
;
318 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
319 tp
->window_clamp
= max(maxwin
-
320 (maxwin
>> sysctl_tcp_app_win
),
324 /* Force reservation of one segment. */
325 if (sysctl_tcp_app_win
&&
326 tp
->window_clamp
> 2 * tp
->advmss
&&
327 tp
->window_clamp
+ tp
->advmss
> maxwin
)
328 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
330 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
331 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
334 /* 5. Recalculate window clamp after socket hit its memory bounds. */
335 static void tcp_clamp_window(struct sock
*sk
)
337 struct tcp_sock
*tp
= tcp_sk(sk
);
338 struct inet_connection_sock
*icsk
= inet_csk(sk
);
340 icsk
->icsk_ack
.quick
= 0;
342 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
343 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
344 !tcp_memory_pressure
&&
345 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
346 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
349 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
350 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
354 /* Initialize RCV_MSS value.
355 * RCV_MSS is an our guess about MSS used by the peer.
356 * We haven't any direct information about the MSS.
357 * It's better to underestimate the RCV_MSS rather than overestimate.
358 * Overestimations make us ACKing less frequently than needed.
359 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
361 void tcp_initialize_rcv_mss(struct sock
*sk
)
363 struct tcp_sock
*tp
= tcp_sk(sk
);
364 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
366 hint
= min(hint
, tp
->rcv_wnd
/2);
367 hint
= min(hint
, TCP_MIN_RCVMSS
);
368 hint
= max(hint
, TCP_MIN_MSS
);
370 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
373 /* Receiver "autotuning" code.
375 * The algorithm for RTT estimation w/o timestamps is based on
376 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
377 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
379 * More detail on this code can be found at
380 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
381 * though this reference is out of date. A new paper
384 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
386 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
392 if (new_sample
!= 0) {
393 /* If we sample in larger samples in the non-timestamp
394 * case, we could grossly overestimate the RTT especially
395 * with chatty applications or bulk transfer apps which
396 * are stalled on filesystem I/O.
398 * Also, since we are only going for a minimum in the
399 * non-timestamp case, we do not smooth things out
400 * else with timestamps disabled convergence takes too
404 m
-= (new_sample
>> 3);
406 } else if (m
< new_sample
)
409 /* No previous measure. */
413 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
414 tp
->rcv_rtt_est
.rtt
= new_sample
;
417 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
419 if (tp
->rcv_rtt_est
.time
== 0)
421 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
423 tcp_rcv_rtt_update(tp
,
424 jiffies
- tp
->rcv_rtt_est
.time
,
428 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
429 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
432 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
434 struct tcp_sock
*tp
= tcp_sk(sk
);
435 if (tp
->rx_opt
.rcv_tsecr
&&
436 (TCP_SKB_CB(skb
)->end_seq
-
437 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
438 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
442 * This function should be called every time data is copied to user space.
443 * It calculates the appropriate TCP receive buffer space.
445 void tcp_rcv_space_adjust(struct sock
*sk
)
447 struct tcp_sock
*tp
= tcp_sk(sk
);
451 if (tp
->rcvq_space
.time
== 0)
454 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
455 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
456 tp
->rcv_rtt_est
.rtt
== 0)
459 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
461 space
= max(tp
->rcvq_space
.space
, space
);
463 if (tp
->rcvq_space
.space
!= space
) {
466 tp
->rcvq_space
.space
= space
;
468 if (sysctl_tcp_moderate_rcvbuf
&&
469 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
470 int new_clamp
= space
;
472 /* Receive space grows, normalize in order to
473 * take into account packet headers and sk_buff
474 * structure overhead.
479 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
480 16 + sizeof(struct sk_buff
));
481 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
484 space
= min(space
, sysctl_tcp_rmem
[2]);
485 if (space
> sk
->sk_rcvbuf
) {
486 sk
->sk_rcvbuf
= space
;
488 /* Make the window clamp follow along. */
489 tp
->window_clamp
= new_clamp
;
495 tp
->rcvq_space
.seq
= tp
->copied_seq
;
496 tp
->rcvq_space
.time
= tcp_time_stamp
;
499 /* There is something which you must keep in mind when you analyze the
500 * behavior of the tp->ato delayed ack timeout interval. When a
501 * connection starts up, we want to ack as quickly as possible. The
502 * problem is that "good" TCP's do slow start at the beginning of data
503 * transmission. The means that until we send the first few ACK's the
504 * sender will sit on his end and only queue most of his data, because
505 * he can only send snd_cwnd unacked packets at any given time. For
506 * each ACK we send, he increments snd_cwnd and transmits more of his
509 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
511 struct tcp_sock
*tp
= tcp_sk(sk
);
512 struct inet_connection_sock
*icsk
= inet_csk(sk
);
515 inet_csk_schedule_ack(sk
);
517 tcp_measure_rcv_mss(sk
, skb
);
519 tcp_rcv_rtt_measure(tp
);
521 now
= tcp_time_stamp
;
523 if (!icsk
->icsk_ack
.ato
) {
524 /* The _first_ data packet received, initialize
525 * delayed ACK engine.
527 tcp_incr_quickack(sk
);
528 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
530 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
532 if (m
<= TCP_ATO_MIN
/2) {
533 /* The fastest case is the first. */
534 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
535 } else if (m
< icsk
->icsk_ack
.ato
) {
536 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
537 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
538 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
539 } else if (m
> icsk
->icsk_rto
) {
540 /* Too long gap. Apparently sender failed to
541 * restart window, so that we send ACKs quickly.
543 tcp_incr_quickack(sk
);
544 sk_stream_mem_reclaim(sk
);
547 icsk
->icsk_ack
.lrcvtime
= now
;
549 TCP_ECN_check_ce(tp
, skb
);
552 tcp_grow_window(sk
, skb
);
555 /* Called to compute a smoothed rtt estimate. The data fed to this
556 * routine either comes from timestamps, or from segments that were
557 * known _not_ to have been retransmitted [see Karn/Partridge
558 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
559 * piece by Van Jacobson.
560 * NOTE: the next three routines used to be one big routine.
561 * To save cycles in the RFC 1323 implementation it was better to break
562 * it up into three procedures. -- erics
564 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
566 struct tcp_sock
*tp
= tcp_sk(sk
);
567 long m
= mrtt
; /* RTT */
569 /* The following amusing code comes from Jacobson's
570 * article in SIGCOMM '88. Note that rtt and mdev
571 * are scaled versions of rtt and mean deviation.
572 * This is designed to be as fast as possible
573 * m stands for "measurement".
575 * On a 1990 paper the rto value is changed to:
576 * RTO = rtt + 4 * mdev
578 * Funny. This algorithm seems to be very broken.
579 * These formulae increase RTO, when it should be decreased, increase
580 * too slowly, when it should be increased quickly, decrease too quickly
581 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
582 * does not matter how to _calculate_ it. Seems, it was trap
583 * that VJ failed to avoid. 8)
588 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
589 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
591 m
= -m
; /* m is now abs(error) */
592 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
593 /* This is similar to one of Eifel findings.
594 * Eifel blocks mdev updates when rtt decreases.
595 * This solution is a bit different: we use finer gain
596 * for mdev in this case (alpha*beta).
597 * Like Eifel it also prevents growth of rto,
598 * but also it limits too fast rto decreases,
599 * happening in pure Eifel.
604 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
606 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
607 if (tp
->mdev
> tp
->mdev_max
) {
608 tp
->mdev_max
= tp
->mdev
;
609 if (tp
->mdev_max
> tp
->rttvar
)
610 tp
->rttvar
= tp
->mdev_max
;
612 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
613 if (tp
->mdev_max
< tp
->rttvar
)
614 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
615 tp
->rtt_seq
= tp
->snd_nxt
;
616 tp
->mdev_max
= TCP_RTO_MIN
;
619 /* no previous measure. */
620 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
621 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
622 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
623 tp
->rtt_seq
= tp
->snd_nxt
;
627 /* Calculate rto without backoff. This is the second half of Van Jacobson's
628 * routine referred to above.
630 static inline void tcp_set_rto(struct sock
*sk
)
632 const struct tcp_sock
*tp
= tcp_sk(sk
);
633 /* Old crap is replaced with new one. 8)
636 * 1. If rtt variance happened to be less 50msec, it is hallucination.
637 * It cannot be less due to utterly erratic ACK generation made
638 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
639 * to do with delayed acks, because at cwnd>2 true delack timeout
640 * is invisible. Actually, Linux-2.4 also generates erratic
641 * ACKs in some circumstances.
643 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
645 /* 2. Fixups made earlier cannot be right.
646 * If we do not estimate RTO correctly without them,
647 * all the algo is pure shit and should be replaced
648 * with correct one. It is exactly, which we pretend to do.
652 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
653 * guarantees that rto is higher.
655 static inline void tcp_bound_rto(struct sock
*sk
)
657 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
658 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
661 /* Save metrics learned by this TCP session.
662 This function is called only, when TCP finishes successfully
663 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
665 void tcp_update_metrics(struct sock
*sk
)
667 struct tcp_sock
*tp
= tcp_sk(sk
);
668 struct dst_entry
*dst
= __sk_dst_get(sk
);
670 if (sysctl_tcp_nometrics_save
)
675 if (dst
&& (dst
->flags
&DST_HOST
)) {
676 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
679 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
680 /* This session failed to estimate rtt. Why?
681 * Probably, no packets returned in time.
684 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
685 dst
->metrics
[RTAX_RTT
-1] = 0;
689 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
691 /* If newly calculated rtt larger than stored one,
692 * store new one. Otherwise, use EWMA. Remember,
693 * rtt overestimation is always better than underestimation.
695 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
697 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
699 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
702 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
706 /* Scale deviation to rttvar fixed point */
711 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
712 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
714 dst
->metrics
[RTAX_RTTVAR
-1] -=
715 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
718 if (tp
->snd_ssthresh
>= 0xFFFF) {
719 /* Slow start still did not finish. */
720 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
721 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
722 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
723 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
724 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
725 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
726 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
727 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
728 icsk
->icsk_ca_state
== TCP_CA_Open
) {
729 /* Cong. avoidance phase, cwnd is reliable. */
730 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
731 dst
->metrics
[RTAX_SSTHRESH
-1] =
732 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
733 if (!dst_metric_locked(dst
, RTAX_CWND
))
734 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
736 /* Else slow start did not finish, cwnd is non-sense,
737 ssthresh may be also invalid.
739 if (!dst_metric_locked(dst
, RTAX_CWND
))
740 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
741 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
742 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
743 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
744 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
747 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
748 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
749 tp
->reordering
!= sysctl_tcp_reordering
)
750 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
755 /* Numbers are taken from RFC2414. */
756 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
758 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
761 if (tp
->mss_cache
> 1460)
764 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
766 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
769 /* Set slow start threshold and cwnd not falling to slow start */
770 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
772 struct tcp_sock
*tp
= tcp_sk(sk
);
773 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
775 tp
->prior_ssthresh
= 0;
777 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
780 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
781 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
782 tcp_packets_in_flight(tp
) + 1U);
783 tp
->snd_cwnd_cnt
= 0;
784 tp
->high_seq
= tp
->snd_nxt
;
785 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
786 TCP_ECN_queue_cwr(tp
);
788 tcp_set_ca_state(sk
, TCP_CA_CWR
);
792 /* Initialize metrics on socket. */
794 static void tcp_init_metrics(struct sock
*sk
)
796 struct tcp_sock
*tp
= tcp_sk(sk
);
797 struct dst_entry
*dst
= __sk_dst_get(sk
);
804 if (dst_metric_locked(dst
, RTAX_CWND
))
805 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
806 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
807 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
808 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
809 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
811 if (dst_metric(dst
, RTAX_REORDERING
) &&
812 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
813 tp
->rx_opt
.sack_ok
&= ~2;
814 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
817 if (dst_metric(dst
, RTAX_RTT
) == 0)
820 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
823 /* Initial rtt is determined from SYN,SYN-ACK.
824 * The segment is small and rtt may appear much
825 * less than real one. Use per-dst memory
826 * to make it more realistic.
828 * A bit of theory. RTT is time passed after "normal" sized packet
829 * is sent until it is ACKed. In normal circumstances sending small
830 * packets force peer to delay ACKs and calculation is correct too.
831 * The algorithm is adaptive and, provided we follow specs, it
832 * NEVER underestimate RTT. BUT! If peer tries to make some clever
833 * tricks sort of "quick acks" for time long enough to decrease RTT
834 * to low value, and then abruptly stops to do it and starts to delay
835 * ACKs, wait for troubles.
837 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
838 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
839 tp
->rtt_seq
= tp
->snd_nxt
;
841 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
842 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
843 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
847 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
849 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
850 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
854 /* Play conservative. If timestamps are not
855 * supported, TCP will fail to recalculate correct
856 * rtt, if initial rto is too small. FORGET ALL AND RESET!
858 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
860 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
861 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
865 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
868 struct tcp_sock
*tp
= tcp_sk(sk
);
869 if (metric
> tp
->reordering
) {
870 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
872 /* This exciting event is worth to be remembered. 8) */
874 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
876 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
878 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
880 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
881 #if FASTRETRANS_DEBUG > 1
882 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
883 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
887 tp
->undo_marker
? tp
->undo_retrans
: 0);
889 /* Disable FACK yet. */
890 tp
->rx_opt
.sack_ok
&= ~2;
894 /* This procedure tags the retransmission queue when SACKs arrive.
896 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
897 * Packets in queue with these bits set are counted in variables
898 * sacked_out, retrans_out and lost_out, correspondingly.
900 * Valid combinations are:
901 * Tag InFlight Description
902 * 0 1 - orig segment is in flight.
903 * S 0 - nothing flies, orig reached receiver.
904 * L 0 - nothing flies, orig lost by net.
905 * R 2 - both orig and retransmit are in flight.
906 * L|R 1 - orig is lost, retransmit is in flight.
907 * S|R 1 - orig reached receiver, retrans is still in flight.
908 * (L|S|R is logically valid, it could occur when L|R is sacked,
909 * but it is equivalent to plain S and code short-curcuits it to S.
910 * L|S is logically invalid, it would mean -1 packet in flight 8))
912 * These 6 states form finite state machine, controlled by the following events:
913 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
914 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
915 * 3. Loss detection event of one of three flavors:
916 * A. Scoreboard estimator decided the packet is lost.
917 * A'. Reno "three dupacks" marks head of queue lost.
918 * A''. Its FACK modfication, head until snd.fack is lost.
919 * B. SACK arrives sacking data transmitted after never retransmitted
921 * C. SACK arrives sacking SND.NXT at the moment, when the
922 * segment was retransmitted.
923 * 4. D-SACK added new rule: D-SACK changes any tag to S.
925 * It is pleasant to note, that state diagram turns out to be commutative,
926 * so that we are allowed not to be bothered by order of our actions,
927 * when multiple events arrive simultaneously. (see the function below).
929 * Reordering detection.
930 * --------------------
931 * Reordering metric is maximal distance, which a packet can be displaced
932 * in packet stream. With SACKs we can estimate it:
934 * 1. SACK fills old hole and the corresponding segment was not
935 * ever retransmitted -> reordering. Alas, we cannot use it
936 * when segment was retransmitted.
937 * 2. The last flaw is solved with D-SACK. D-SACK arrives
938 * for retransmitted and already SACKed segment -> reordering..
939 * Both of these heuristics are not used in Loss state, when we cannot
940 * account for retransmits accurately.
943 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
945 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
946 struct tcp_sock
*tp
= tcp_sk(sk
);
947 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
948 TCP_SKB_CB(ack_skb
)->sacked
);
949 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
950 struct sk_buff
*cached_skb
;
951 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
952 int reord
= tp
->packets_out
;
954 u32 lost_retrans
= 0;
957 int cached_fack_count
;
959 int first_sack_index
;
963 prior_fackets
= tp
->fackets_out
;
965 /* Check for D-SACK. */
966 if (before(ntohl(sp
[0].start_seq
), TCP_SKB_CB(ack_skb
)->ack_seq
)) {
968 tp
->rx_opt
.sack_ok
|= 4;
969 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
970 } else if (num_sacks
> 1 &&
971 !after(ntohl(sp
[0].end_seq
), ntohl(sp
[1].end_seq
)) &&
972 !before(ntohl(sp
[0].start_seq
), ntohl(sp
[1].start_seq
))) {
974 tp
->rx_opt
.sack_ok
|= 4;
975 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
978 /* D-SACK for already forgotten data...
979 * Do dumb counting. */
981 !after(ntohl(sp
[0].end_seq
), prior_snd_una
) &&
982 after(ntohl(sp
[0].end_seq
), tp
->undo_marker
))
985 /* Eliminate too old ACKs, but take into
986 * account more or less fresh ones, they can
987 * contain valid SACK info.
989 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
993 * if the only SACK change is the increase of the end_seq of
994 * the first block then only apply that SACK block
995 * and use retrans queue hinting otherwise slowpath */
997 for (i
= 0; i
< num_sacks
; i
++) {
998 __be32 start_seq
= sp
[i
].start_seq
;
999 __be32 end_seq
= sp
[i
].end_seq
;
1002 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
1005 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
1006 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1009 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1010 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1012 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1013 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1014 tp
->recv_sack_cache
[i
].start_seq
= 0;
1015 tp
->recv_sack_cache
[i
].end_seq
= 0;
1018 first_sack_index
= 0;
1023 tp
->fastpath_skb_hint
= NULL
;
1025 /* order SACK blocks to allow in order walk of the retrans queue */
1026 for (i
= num_sacks
-1; i
> 0; i
--) {
1027 for (j
= 0; j
< i
; j
++){
1028 if (after(ntohl(sp
[j
].start_seq
),
1029 ntohl(sp
[j
+1].start_seq
))){
1030 struct tcp_sack_block_wire tmp
;
1036 /* Track where the first SACK block goes to */
1037 if (j
== first_sack_index
)
1038 first_sack_index
= j
+1;
1045 /* clear flag as used for different purpose in following code */
1048 /* Use SACK fastpath hint if valid */
1049 cached_skb
= tp
->fastpath_skb_hint
;
1050 cached_fack_count
= tp
->fastpath_cnt_hint
;
1052 cached_skb
= tcp_write_queue_head(sk
);
1053 cached_fack_count
= 0;
1056 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1057 struct sk_buff
*skb
;
1058 __u32 start_seq
= ntohl(sp
->start_seq
);
1059 __u32 end_seq
= ntohl(sp
->end_seq
);
1063 fack_count
= cached_fack_count
;
1065 /* Event "B" in the comment above. */
1066 if (after(end_seq
, tp
->high_seq
))
1067 flag
|= FLAG_DATA_LOST
;
1069 tcp_for_write_queue_from(skb
, sk
) {
1070 int in_sack
, pcount
;
1073 if (skb
== tcp_send_head(sk
))
1077 cached_fack_count
= fack_count
;
1078 if (i
== first_sack_index
) {
1079 tp
->fastpath_skb_hint
= skb
;
1080 tp
->fastpath_cnt_hint
= fack_count
;
1083 /* The retransmission queue is always in order, so
1084 * we can short-circuit the walk early.
1086 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1089 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1090 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1092 pcount
= tcp_skb_pcount(skb
);
1094 if (pcount
> 1 && !in_sack
&&
1095 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1096 unsigned int pkt_len
;
1098 in_sack
= !after(start_seq
,
1099 TCP_SKB_CB(skb
)->seq
);
1102 pkt_len
= (start_seq
-
1103 TCP_SKB_CB(skb
)->seq
);
1105 pkt_len
= (end_seq
-
1106 TCP_SKB_CB(skb
)->seq
);
1107 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1109 pcount
= tcp_skb_pcount(skb
);
1112 fack_count
+= pcount
;
1114 sacked
= TCP_SKB_CB(skb
)->sacked
;
1116 /* Account D-SACK for retransmitted packet. */
1117 if ((dup_sack
&& in_sack
) &&
1118 (sacked
& TCPCB_RETRANS
) &&
1119 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1122 /* The frame is ACKed. */
1123 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1124 if (sacked
&TCPCB_RETRANS
) {
1125 if ((dup_sack
&& in_sack
) &&
1126 (sacked
&TCPCB_SACKED_ACKED
))
1127 reord
= min(fack_count
, reord
);
1129 /* If it was in a hole, we detected reordering. */
1130 if (fack_count
< prior_fackets
&&
1131 !(sacked
&TCPCB_SACKED_ACKED
))
1132 reord
= min(fack_count
, reord
);
1135 /* Nothing to do; acked frame is about to be dropped. */
1139 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1140 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1141 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1142 lost_retrans
= end_seq
;
1147 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1148 if (sacked
& TCPCB_SACKED_RETRANS
) {
1149 /* If the segment is not tagged as lost,
1150 * we do not clear RETRANS, believing
1151 * that retransmission is still in flight.
1153 if (sacked
& TCPCB_LOST
) {
1154 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1155 tp
->lost_out
-= tcp_skb_pcount(skb
);
1156 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1158 /* clear lost hint */
1159 tp
->retransmit_skb_hint
= NULL
;
1162 /* New sack for not retransmitted frame,
1163 * which was in hole. It is reordering.
1165 if (!(sacked
& TCPCB_RETRANS
) &&
1166 fack_count
< prior_fackets
)
1167 reord
= min(fack_count
, reord
);
1169 if (sacked
& TCPCB_LOST
) {
1170 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1171 tp
->lost_out
-= tcp_skb_pcount(skb
);
1173 /* clear lost hint */
1174 tp
->retransmit_skb_hint
= NULL
;
1176 /* SACK enhanced F-RTO detection.
1177 * Set flag if and only if non-rexmitted
1178 * segments below frto_highmark are
1179 * SACKed (RFC4138; Appendix B).
1180 * Clearing correct due to in-order walk
1182 if (after(end_seq
, tp
->frto_highmark
)) {
1183 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1185 if (!(sacked
& TCPCB_RETRANS
))
1186 flag
|= FLAG_ONLY_ORIG_SACKED
;
1190 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1191 flag
|= FLAG_DATA_SACKED
;
1192 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1194 if (fack_count
> tp
->fackets_out
)
1195 tp
->fackets_out
= fack_count
;
1197 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1198 reord
= min(fack_count
, reord
);
1201 /* D-SACK. We can detect redundant retransmission
1202 * in S|R and plain R frames and clear it.
1203 * undo_retrans is decreased above, L|R frames
1204 * are accounted above as well.
1207 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1208 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1209 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1210 tp
->retransmit_skb_hint
= NULL
;
1215 /* Check for lost retransmit. This superb idea is
1216 * borrowed from "ratehalving". Event "C".
1217 * Later note: FACK people cheated me again 8),
1218 * we have to account for reordering! Ugly,
1221 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1222 struct sk_buff
*skb
;
1224 tcp_for_write_queue(skb
, sk
) {
1225 if (skb
== tcp_send_head(sk
))
1227 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1229 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1231 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1232 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1234 !before(lost_retrans
,
1235 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1237 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1238 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1240 /* clear lost hint */
1241 tp
->retransmit_skb_hint
= NULL
;
1243 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1244 tp
->lost_out
+= tcp_skb_pcount(skb
);
1245 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1246 flag
|= FLAG_DATA_SACKED
;
1247 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1253 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1255 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
&&
1256 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1257 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1259 #if FASTRETRANS_DEBUG > 0
1260 BUG_TRAP((int)tp
->sacked_out
>= 0);
1261 BUG_TRAP((int)tp
->lost_out
>= 0);
1262 BUG_TRAP((int)tp
->retrans_out
>= 0);
1263 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1268 /* F-RTO can only be used if these conditions are satisfied:
1269 * - there must be some unsent new data
1270 * - the advertised window should allow sending it
1271 * - TCP has never retransmitted anything other than head (SACK enhanced
1272 * variant from Appendix B of RFC4138 is more robust here)
1274 int tcp_use_frto(struct sock
*sk
)
1276 const struct tcp_sock
*tp
= tcp_sk(sk
);
1277 struct sk_buff
*skb
;
1279 if (!sysctl_tcp_frto
|| !tcp_send_head(sk
) ||
1280 after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
1281 tp
->snd_una
+ tp
->snd_wnd
))
1287 /* Avoid expensive walking of rexmit queue if possible */
1288 if (tp
->retrans_out
> 1)
1291 skb
= tcp_write_queue_head(sk
);
1292 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1293 tcp_for_write_queue_from(skb
, sk
) {
1294 if (skb
== tcp_send_head(sk
))
1296 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1298 /* Short-circuit when first non-SACKed skb has been checked */
1299 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1305 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1306 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1307 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1308 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1309 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1310 * bits are handled if the Loss state is really to be entered (in
1311 * tcp_enter_frto_loss).
1313 * Do like tcp_enter_loss() would; when RTO expires the second time it
1315 * "Reduce ssthresh if it has not yet been made inside this window."
1317 void tcp_enter_frto(struct sock
*sk
)
1319 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1320 struct tcp_sock
*tp
= tcp_sk(sk
);
1321 struct sk_buff
*skb
;
1323 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1324 tp
->snd_una
== tp
->high_seq
||
1325 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1326 !icsk
->icsk_retransmits
)) {
1327 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1328 /* Our state is too optimistic in ssthresh() call because cwnd
1329 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1330 * recovery has not yet completed. Pattern would be this: RTO,
1331 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1333 * RFC4138 should be more specific on what to do, even though
1334 * RTO is quite unlikely to occur after the first Cumulative ACK
1335 * due to back-off and complexity of triggering events ...
1337 if (tp
->frto_counter
) {
1339 stored_cwnd
= tp
->snd_cwnd
;
1341 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1342 tp
->snd_cwnd
= stored_cwnd
;
1344 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1346 /* ... in theory, cong.control module could do "any tricks" in
1347 * ssthresh(), which means that ca_state, lost bits and lost_out
1348 * counter would have to be faked before the call occurs. We
1349 * consider that too expensive, unlikely and hacky, so modules
1350 * using these in ssthresh() must deal these incompatibility
1351 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1353 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1356 tp
->undo_marker
= tp
->snd_una
;
1357 tp
->undo_retrans
= 0;
1359 skb
= tcp_write_queue_head(sk
);
1360 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1361 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1362 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1364 tcp_sync_left_out(tp
);
1366 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1367 * The last condition is necessary at least in tp->frto_counter case.
1369 if (IsSackFrto() && (tp
->frto_counter
||
1370 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1371 after(tp
->high_seq
, tp
->snd_una
)) {
1372 tp
->frto_highmark
= tp
->high_seq
;
1374 tp
->frto_highmark
= tp
->snd_nxt
;
1376 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1377 tp
->high_seq
= tp
->snd_nxt
;
1378 tp
->frto_counter
= 1;
1381 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1382 * which indicates that we should follow the traditional RTO recovery,
1383 * i.e. mark everything lost and do go-back-N retransmission.
1385 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1387 struct tcp_sock
*tp
= tcp_sk(sk
);
1388 struct sk_buff
*skb
;
1393 tp
->fackets_out
= 0;
1394 tp
->retrans_out
= 0;
1396 tcp_for_write_queue(skb
, sk
) {
1397 if (skb
== tcp_send_head(sk
))
1399 cnt
+= tcp_skb_pcount(skb
);
1401 * Count the retransmission made on RTO correctly (only when
1402 * waiting for the first ACK and did not get it)...
1404 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1405 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1406 /* ...enter this if branch just for the first segment */
1407 flag
|= FLAG_DATA_ACKED
;
1409 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1411 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1413 /* Do not mark those segments lost that were
1414 * forward transmitted after RTO
1416 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1417 tp
->frto_highmark
)) {
1418 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1419 tp
->lost_out
+= tcp_skb_pcount(skb
);
1422 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1423 tp
->fackets_out
= cnt
;
1426 tcp_sync_left_out(tp
);
1428 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1429 tp
->snd_cwnd_cnt
= 0;
1430 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1431 tp
->undo_marker
= 0;
1432 tp
->frto_counter
= 0;
1434 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1435 sysctl_tcp_reordering
);
1436 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1437 tp
->high_seq
= tp
->frto_highmark
;
1438 TCP_ECN_queue_cwr(tp
);
1440 clear_all_retrans_hints(tp
);
1443 void tcp_clear_retrans(struct tcp_sock
*tp
)
1446 tp
->retrans_out
= 0;
1448 tp
->fackets_out
= 0;
1452 tp
->undo_marker
= 0;
1453 tp
->undo_retrans
= 0;
1456 /* Enter Loss state. If "how" is not zero, forget all SACK information
1457 * and reset tags completely, otherwise preserve SACKs. If receiver
1458 * dropped its ofo queue, we will know this due to reneging detection.
1460 void tcp_enter_loss(struct sock
*sk
, int how
)
1462 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1463 struct tcp_sock
*tp
= tcp_sk(sk
);
1464 struct sk_buff
*skb
;
1467 /* Reduce ssthresh if it has not yet been made inside this window. */
1468 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1469 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1470 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1471 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1472 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1475 tp
->snd_cwnd_cnt
= 0;
1476 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1478 tp
->bytes_acked
= 0;
1479 tcp_clear_retrans(tp
);
1481 /* Push undo marker, if it was plain RTO and nothing
1482 * was retransmitted. */
1484 tp
->undo_marker
= tp
->snd_una
;
1486 tcp_for_write_queue(skb
, sk
) {
1487 if (skb
== tcp_send_head(sk
))
1489 cnt
+= tcp_skb_pcount(skb
);
1490 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1491 tp
->undo_marker
= 0;
1492 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1493 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1494 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1495 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1496 tp
->lost_out
+= tcp_skb_pcount(skb
);
1498 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1499 tp
->fackets_out
= cnt
;
1502 tcp_sync_left_out(tp
);
1504 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1505 sysctl_tcp_reordering
);
1506 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1507 tp
->high_seq
= tp
->snd_nxt
;
1508 TCP_ECN_queue_cwr(tp
);
1510 clear_all_retrans_hints(tp
);
1513 static int tcp_check_sack_reneging(struct sock
*sk
)
1515 struct sk_buff
*skb
;
1517 /* If ACK arrived pointing to a remembered SACK,
1518 * it means that our remembered SACKs do not reflect
1519 * real state of receiver i.e.
1520 * receiver _host_ is heavily congested (or buggy).
1521 * Do processing similar to RTO timeout.
1523 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1524 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1525 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1526 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1528 tcp_enter_loss(sk
, 1);
1529 icsk
->icsk_retransmits
++;
1530 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1531 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1532 icsk
->icsk_rto
, TCP_RTO_MAX
);
1538 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1540 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1543 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1545 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1548 static inline int tcp_head_timedout(struct sock
*sk
)
1550 struct tcp_sock
*tp
= tcp_sk(sk
);
1552 return tp
->packets_out
&&
1553 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1556 /* Linux NewReno/SACK/FACK/ECN state machine.
1557 * --------------------------------------
1559 * "Open" Normal state, no dubious events, fast path.
1560 * "Disorder" In all the respects it is "Open",
1561 * but requires a bit more attention. It is entered when
1562 * we see some SACKs or dupacks. It is split of "Open"
1563 * mainly to move some processing from fast path to slow one.
1564 * "CWR" CWND was reduced due to some Congestion Notification event.
1565 * It can be ECN, ICMP source quench, local device congestion.
1566 * "Recovery" CWND was reduced, we are fast-retransmitting.
1567 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1569 * tcp_fastretrans_alert() is entered:
1570 * - each incoming ACK, if state is not "Open"
1571 * - when arrived ACK is unusual, namely:
1576 * Counting packets in flight is pretty simple.
1578 * in_flight = packets_out - left_out + retrans_out
1580 * packets_out is SND.NXT-SND.UNA counted in packets.
1582 * retrans_out is number of retransmitted segments.
1584 * left_out is number of segments left network, but not ACKed yet.
1586 * left_out = sacked_out + lost_out
1588 * sacked_out: Packets, which arrived to receiver out of order
1589 * and hence not ACKed. With SACKs this number is simply
1590 * amount of SACKed data. Even without SACKs
1591 * it is easy to give pretty reliable estimate of this number,
1592 * counting duplicate ACKs.
1594 * lost_out: Packets lost by network. TCP has no explicit
1595 * "loss notification" feedback from network (for now).
1596 * It means that this number can be only _guessed_.
1597 * Actually, it is the heuristics to predict lossage that
1598 * distinguishes different algorithms.
1600 * F.e. after RTO, when all the queue is considered as lost,
1601 * lost_out = packets_out and in_flight = retrans_out.
1603 * Essentially, we have now two algorithms counting
1606 * FACK: It is the simplest heuristics. As soon as we decided
1607 * that something is lost, we decide that _all_ not SACKed
1608 * packets until the most forward SACK are lost. I.e.
1609 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1610 * It is absolutely correct estimate, if network does not reorder
1611 * packets. And it loses any connection to reality when reordering
1612 * takes place. We use FACK by default until reordering
1613 * is suspected on the path to this destination.
1615 * NewReno: when Recovery is entered, we assume that one segment
1616 * is lost (classic Reno). While we are in Recovery and
1617 * a partial ACK arrives, we assume that one more packet
1618 * is lost (NewReno). This heuristics are the same in NewReno
1621 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1622 * deflation etc. CWND is real congestion window, never inflated, changes
1623 * only according to classic VJ rules.
1625 * Really tricky (and requiring careful tuning) part of algorithm
1626 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1627 * The first determines the moment _when_ we should reduce CWND and,
1628 * hence, slow down forward transmission. In fact, it determines the moment
1629 * when we decide that hole is caused by loss, rather than by a reorder.
1631 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1632 * holes, caused by lost packets.
1634 * And the most logically complicated part of algorithm is undo
1635 * heuristics. We detect false retransmits due to both too early
1636 * fast retransmit (reordering) and underestimated RTO, analyzing
1637 * timestamps and D-SACKs. When we detect that some segments were
1638 * retransmitted by mistake and CWND reduction was wrong, we undo
1639 * window reduction and abort recovery phase. This logic is hidden
1640 * inside several functions named tcp_try_undo_<something>.
1643 /* This function decides, when we should leave Disordered state
1644 * and enter Recovery phase, reducing congestion window.
1646 * Main question: may we further continue forward transmission
1647 * with the same cwnd?
1649 static int tcp_time_to_recover(struct sock
*sk
)
1651 struct tcp_sock
*tp
= tcp_sk(sk
);
1654 /* Do not perform any recovery during FRTO algorithm */
1655 if (tp
->frto_counter
)
1658 /* Trick#1: The loss is proven. */
1662 /* Not-A-Trick#2 : Classic rule... */
1663 if (tcp_fackets_out(tp
) > tp
->reordering
)
1666 /* Trick#3 : when we use RFC2988 timer restart, fast
1667 * retransmit can be triggered by timeout of queue head.
1669 if (tcp_head_timedout(sk
))
1672 /* Trick#4: It is still not OK... But will it be useful to delay
1675 packets_out
= tp
->packets_out
;
1676 if (packets_out
<= tp
->reordering
&&
1677 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1678 !tcp_may_send_now(sk
)) {
1679 /* We have nothing to send. This connection is limited
1680 * either by receiver window or by application.
1688 /* If we receive more dupacks than we expected counting segments
1689 * in assumption of absent reordering, interpret this as reordering.
1690 * The only another reason could be bug in receiver TCP.
1692 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1694 struct tcp_sock
*tp
= tcp_sk(sk
);
1697 holes
= max(tp
->lost_out
, 1U);
1698 holes
= min(holes
, tp
->packets_out
);
1700 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1701 tp
->sacked_out
= tp
->packets_out
- holes
;
1702 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1706 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1708 static void tcp_add_reno_sack(struct sock
*sk
)
1710 struct tcp_sock
*tp
= tcp_sk(sk
);
1712 tcp_check_reno_reordering(sk
, 0);
1713 tcp_sync_left_out(tp
);
1716 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1718 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1720 struct tcp_sock
*tp
= tcp_sk(sk
);
1723 /* One ACK acked hole. The rest eat duplicate ACKs. */
1724 if (acked
-1 >= tp
->sacked_out
)
1727 tp
->sacked_out
-= acked
-1;
1729 tcp_check_reno_reordering(sk
, acked
);
1730 tcp_sync_left_out(tp
);
1733 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1736 tp
->left_out
= tp
->lost_out
;
1739 /* Mark head of queue up as lost. */
1740 static void tcp_mark_head_lost(struct sock
*sk
,
1741 int packets
, u32 high_seq
)
1743 struct tcp_sock
*tp
= tcp_sk(sk
);
1744 struct sk_buff
*skb
;
1747 BUG_TRAP(packets
<= tp
->packets_out
);
1748 if (tp
->lost_skb_hint
) {
1749 skb
= tp
->lost_skb_hint
;
1750 cnt
= tp
->lost_cnt_hint
;
1752 skb
= tcp_write_queue_head(sk
);
1756 tcp_for_write_queue_from(skb
, sk
) {
1757 if (skb
== tcp_send_head(sk
))
1759 /* TODO: do this better */
1760 /* this is not the most efficient way to do this... */
1761 tp
->lost_skb_hint
= skb
;
1762 tp
->lost_cnt_hint
= cnt
;
1763 cnt
+= tcp_skb_pcount(skb
);
1764 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1766 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1767 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1768 tp
->lost_out
+= tcp_skb_pcount(skb
);
1770 /* clear xmit_retransmit_queue hints
1771 * if this is beyond hint */
1772 if (tp
->retransmit_skb_hint
!= NULL
&&
1773 before(TCP_SKB_CB(skb
)->seq
,
1774 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1775 tp
->retransmit_skb_hint
= NULL
;
1779 tcp_sync_left_out(tp
);
1782 /* Account newly detected lost packet(s) */
1784 static void tcp_update_scoreboard(struct sock
*sk
)
1786 struct tcp_sock
*tp
= tcp_sk(sk
);
1789 int lost
= tp
->fackets_out
- tp
->reordering
;
1792 tcp_mark_head_lost(sk
, lost
, tp
->high_seq
);
1794 tcp_mark_head_lost(sk
, 1, tp
->high_seq
);
1797 /* New heuristics: it is possible only after we switched
1798 * to restart timer each time when something is ACKed.
1799 * Hence, we can detect timed out packets during fast
1800 * retransmit without falling to slow start.
1802 if (!IsReno(tp
) && tcp_head_timedout(sk
)) {
1803 struct sk_buff
*skb
;
1805 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1806 : tcp_write_queue_head(sk
);
1808 tcp_for_write_queue_from(skb
, sk
) {
1809 if (skb
== tcp_send_head(sk
))
1811 if (!tcp_skb_timedout(sk
, skb
))
1814 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1815 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1816 tp
->lost_out
+= tcp_skb_pcount(skb
);
1818 /* clear xmit_retrans hint */
1819 if (tp
->retransmit_skb_hint
&&
1820 before(TCP_SKB_CB(skb
)->seq
,
1821 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1823 tp
->retransmit_skb_hint
= NULL
;
1827 tp
->scoreboard_skb_hint
= skb
;
1829 tcp_sync_left_out(tp
);
1833 /* CWND moderation, preventing bursts due to too big ACKs
1834 * in dubious situations.
1836 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1838 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1839 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1840 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1843 /* Lower bound on congestion window is slow start threshold
1844 * unless congestion avoidance choice decides to overide it.
1846 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1848 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1850 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1853 /* Decrease cwnd each second ack. */
1854 static void tcp_cwnd_down(struct sock
*sk
)
1856 struct tcp_sock
*tp
= tcp_sk(sk
);
1857 int decr
= tp
->snd_cwnd_cnt
+ 1;
1859 tp
->snd_cwnd_cnt
= decr
&1;
1862 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1863 tp
->snd_cwnd
-= decr
;
1865 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1866 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1869 /* Nothing was retransmitted or returned timestamp is less
1870 * than timestamp of the first retransmission.
1872 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1874 return !tp
->retrans_stamp
||
1875 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1876 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1879 /* Undo procedures. */
1881 #if FASTRETRANS_DEBUG > 1
1882 static void DBGUNDO(struct sock
*sk
, const char *msg
)
1884 struct tcp_sock
*tp
= tcp_sk(sk
);
1885 struct inet_sock
*inet
= inet_sk(sk
);
1887 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1889 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1890 tp
->snd_cwnd
, tp
->left_out
,
1891 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1895 #define DBGUNDO(x...) do { } while (0)
1898 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1900 struct tcp_sock
*tp
= tcp_sk(sk
);
1902 if (tp
->prior_ssthresh
) {
1903 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1905 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1906 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1908 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1910 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1911 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1912 TCP_ECN_withdraw_cwr(tp
);
1915 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1917 tcp_moderate_cwnd(tp
);
1918 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1920 /* There is something screwy going on with the retrans hints after
1922 clear_all_retrans_hints(tp
);
1925 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1927 return tp
->undo_marker
&&
1928 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1931 /* People celebrate: "We love our President!" */
1932 static int tcp_try_undo_recovery(struct sock
*sk
)
1934 struct tcp_sock
*tp
= tcp_sk(sk
);
1936 if (tcp_may_undo(tp
)) {
1937 /* Happy end! We did not retransmit anything
1938 * or our original transmission succeeded.
1940 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1941 tcp_undo_cwr(sk
, 1);
1942 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1943 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1945 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1946 tp
->undo_marker
= 0;
1948 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1949 /* Hold old state until something *above* high_seq
1950 * is ACKed. For Reno it is MUST to prevent false
1951 * fast retransmits (RFC2582). SACK TCP is safe. */
1952 tcp_moderate_cwnd(tp
);
1955 tcp_set_ca_state(sk
, TCP_CA_Open
);
1959 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1960 static void tcp_try_undo_dsack(struct sock
*sk
)
1962 struct tcp_sock
*tp
= tcp_sk(sk
);
1964 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1965 DBGUNDO(sk
, "D-SACK");
1966 tcp_undo_cwr(sk
, 1);
1967 tp
->undo_marker
= 0;
1968 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
1972 /* Undo during fast recovery after partial ACK. */
1974 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
1976 struct tcp_sock
*tp
= tcp_sk(sk
);
1977 /* Partial ACK arrived. Force Hoe's retransmit. */
1978 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
1980 if (tcp_may_undo(tp
)) {
1981 /* Plain luck! Hole if filled with delayed
1982 * packet, rather than with a retransmit.
1984 if (tp
->retrans_out
== 0)
1985 tp
->retrans_stamp
= 0;
1987 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
1990 tcp_undo_cwr(sk
, 0);
1991 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
1993 /* So... Do not make Hoe's retransmit yet.
1994 * If the first packet was delayed, the rest
1995 * ones are most probably delayed as well.
2002 /* Undo during loss recovery after partial ACK. */
2003 static int tcp_try_undo_loss(struct sock
*sk
)
2005 struct tcp_sock
*tp
= tcp_sk(sk
);
2007 if (tcp_may_undo(tp
)) {
2008 struct sk_buff
*skb
;
2009 tcp_for_write_queue(skb
, sk
) {
2010 if (skb
== tcp_send_head(sk
))
2012 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2015 clear_all_retrans_hints(tp
);
2017 DBGUNDO(sk
, "partial loss");
2019 tp
->left_out
= tp
->sacked_out
;
2020 tcp_undo_cwr(sk
, 1);
2021 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2022 inet_csk(sk
)->icsk_retransmits
= 0;
2023 tp
->undo_marker
= 0;
2025 tcp_set_ca_state(sk
, TCP_CA_Open
);
2031 static inline void tcp_complete_cwr(struct sock
*sk
)
2033 struct tcp_sock
*tp
= tcp_sk(sk
);
2034 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2035 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2036 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2039 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2041 struct tcp_sock
*tp
= tcp_sk(sk
);
2043 tp
->left_out
= tp
->sacked_out
;
2045 if (tp
->retrans_out
== 0)
2046 tp
->retrans_stamp
= 0;
2049 tcp_enter_cwr(sk
, 1);
2051 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2052 int state
= TCP_CA_Open
;
2054 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
2055 state
= TCP_CA_Disorder
;
2057 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2058 tcp_set_ca_state(sk
, state
);
2059 tp
->high_seq
= tp
->snd_nxt
;
2061 tcp_moderate_cwnd(tp
);
2067 static void tcp_mtup_probe_failed(struct sock
*sk
)
2069 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2071 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2072 icsk
->icsk_mtup
.probe_size
= 0;
2075 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2077 struct tcp_sock
*tp
= tcp_sk(sk
);
2078 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2080 /* FIXME: breaks with very large cwnd */
2081 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2082 tp
->snd_cwnd
= tp
->snd_cwnd
*
2083 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2084 icsk
->icsk_mtup
.probe_size
;
2085 tp
->snd_cwnd_cnt
= 0;
2086 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2087 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2089 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2090 icsk
->icsk_mtup
.probe_size
= 0;
2091 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2095 /* Process an event, which can update packets-in-flight not trivially.
2096 * Main goal of this function is to calculate new estimate for left_out,
2097 * taking into account both packets sitting in receiver's buffer and
2098 * packets lost by network.
2100 * Besides that it does CWND reduction, when packet loss is detected
2101 * and changes state of machine.
2103 * It does _not_ decide what to send, it is made in function
2104 * tcp_xmit_retransmit_queue().
2107 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
2108 int prior_packets
, int flag
)
2110 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2111 struct tcp_sock
*tp
= tcp_sk(sk
);
2112 int is_dupack
= (tp
->snd_una
== prior_snd_una
&& !(flag
&FLAG_NOT_DUP
));
2114 /* Some technical things:
2115 * 1. Reno does not count dupacks (sacked_out) automatically. */
2116 if (!tp
->packets_out
)
2118 /* 2. SACK counts snd_fack in packets inaccurately. */
2119 if (tp
->sacked_out
== 0)
2120 tp
->fackets_out
= 0;
2122 /* Now state machine starts.
2123 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2125 tp
->prior_ssthresh
= 0;
2127 /* B. In all the states check for reneging SACKs. */
2128 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2131 /* C. Process data loss notification, provided it is valid. */
2132 if ((flag
&FLAG_DATA_LOST
) &&
2133 before(tp
->snd_una
, tp
->high_seq
) &&
2134 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2135 tp
->fackets_out
> tp
->reordering
) {
2136 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
2137 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2140 /* D. Synchronize left_out to current state. */
2141 tcp_sync_left_out(tp
);
2143 /* E. Check state exit conditions. State can be terminated
2144 * when high_seq is ACKed. */
2145 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2146 BUG_TRAP(tp
->retrans_out
== 0);
2147 tp
->retrans_stamp
= 0;
2148 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2149 switch (icsk
->icsk_ca_state
) {
2151 icsk
->icsk_retransmits
= 0;
2152 if (tcp_try_undo_recovery(sk
))
2157 /* CWR is to be held something *above* high_seq
2158 * is ACKed for CWR bit to reach receiver. */
2159 if (tp
->snd_una
!= tp
->high_seq
) {
2160 tcp_complete_cwr(sk
);
2161 tcp_set_ca_state(sk
, TCP_CA_Open
);
2165 case TCP_CA_Disorder
:
2166 tcp_try_undo_dsack(sk
);
2167 if (!tp
->undo_marker
||
2168 /* For SACK case do not Open to allow to undo
2169 * catching for all duplicate ACKs. */
2170 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2171 tp
->undo_marker
= 0;
2172 tcp_set_ca_state(sk
, TCP_CA_Open
);
2176 case TCP_CA_Recovery
:
2178 tcp_reset_reno_sack(tp
);
2179 if (tcp_try_undo_recovery(sk
))
2181 tcp_complete_cwr(sk
);
2186 /* F. Process state. */
2187 switch (icsk
->icsk_ca_state
) {
2188 case TCP_CA_Recovery
:
2189 if (prior_snd_una
== tp
->snd_una
) {
2190 if (IsReno(tp
) && is_dupack
)
2191 tcp_add_reno_sack(sk
);
2193 int acked
= prior_packets
- tp
->packets_out
;
2195 tcp_remove_reno_sacks(sk
, acked
);
2196 is_dupack
= tcp_try_undo_partial(sk
, acked
);
2200 if (flag
&FLAG_DATA_ACKED
)
2201 icsk
->icsk_retransmits
= 0;
2202 if (!tcp_try_undo_loss(sk
)) {
2203 tcp_moderate_cwnd(tp
);
2204 tcp_xmit_retransmit_queue(sk
);
2207 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2209 /* Loss is undone; fall through to processing in Open state. */
2212 if (tp
->snd_una
!= prior_snd_una
)
2213 tcp_reset_reno_sack(tp
);
2215 tcp_add_reno_sack(sk
);
2218 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2219 tcp_try_undo_dsack(sk
);
2221 if (!tcp_time_to_recover(sk
)) {
2222 tcp_try_to_open(sk
, flag
);
2226 /* MTU probe failure: don't reduce cwnd */
2227 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2228 icsk
->icsk_mtup
.probe_size
&&
2229 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2230 tcp_mtup_probe_failed(sk
);
2231 /* Restores the reduction we did in tcp_mtup_probe() */
2233 tcp_simple_retransmit(sk
);
2237 /* Otherwise enter Recovery state */
2240 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2242 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2244 tp
->high_seq
= tp
->snd_nxt
;
2245 tp
->prior_ssthresh
= 0;
2246 tp
->undo_marker
= tp
->snd_una
;
2247 tp
->undo_retrans
= tp
->retrans_out
;
2249 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2250 if (!(flag
&FLAG_ECE
))
2251 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2252 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2253 TCP_ECN_queue_cwr(tp
);
2256 tp
->bytes_acked
= 0;
2257 tp
->snd_cwnd_cnt
= 0;
2258 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2261 if (is_dupack
|| tcp_head_timedout(sk
))
2262 tcp_update_scoreboard(sk
);
2264 tcp_xmit_retransmit_queue(sk
);
2267 /* Read draft-ietf-tcplw-high-performance before mucking
2268 * with this code. (Supersedes RFC1323)
2270 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2272 /* RTTM Rule: A TSecr value received in a segment is used to
2273 * update the averaged RTT measurement only if the segment
2274 * acknowledges some new data, i.e., only if it advances the
2275 * left edge of the send window.
2277 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2278 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2280 * Changed: reset backoff as soon as we see the first valid sample.
2281 * If we do not, we get strongly overestimated rto. With timestamps
2282 * samples are accepted even from very old segments: f.e., when rtt=1
2283 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2284 * answer arrives rto becomes 120 seconds! If at least one of segments
2285 * in window is lost... Voila. --ANK (010210)
2287 struct tcp_sock
*tp
= tcp_sk(sk
);
2288 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2289 tcp_rtt_estimator(sk
, seq_rtt
);
2291 inet_csk(sk
)->icsk_backoff
= 0;
2295 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2297 /* We don't have a timestamp. Can only use
2298 * packets that are not retransmitted to determine
2299 * rtt estimates. Also, we must not reset the
2300 * backoff for rto until we get a non-retransmitted
2301 * packet. This allows us to deal with a situation
2302 * where the network delay has increased suddenly.
2303 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2306 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2309 tcp_rtt_estimator(sk
, seq_rtt
);
2311 inet_csk(sk
)->icsk_backoff
= 0;
2315 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2318 const struct tcp_sock
*tp
= tcp_sk(sk
);
2319 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2320 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2321 tcp_ack_saw_tstamp(sk
, flag
);
2322 else if (seq_rtt
>= 0)
2323 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2326 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
2327 u32 in_flight
, int good
)
2329 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2330 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
2331 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2334 /* Restart timer after forward progress on connection.
2335 * RFC2988 recommends to restart timer to now+rto.
2338 static void tcp_ack_packets_out(struct sock
*sk
)
2340 struct tcp_sock
*tp
= tcp_sk(sk
);
2342 if (!tp
->packets_out
) {
2343 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2345 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2349 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2350 __u32 now
, __s32
*seq_rtt
)
2352 struct tcp_sock
*tp
= tcp_sk(sk
);
2353 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2354 __u32 seq
= tp
->snd_una
;
2355 __u32 packets_acked
;
2358 /* If we get here, the whole TSO packet has not been
2361 BUG_ON(!after(scb
->end_seq
, seq
));
2363 packets_acked
= tcp_skb_pcount(skb
);
2364 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2366 packets_acked
-= tcp_skb_pcount(skb
);
2368 if (packets_acked
) {
2369 __u8 sacked
= scb
->sacked
;
2371 acked
|= FLAG_DATA_ACKED
;
2373 if (sacked
& TCPCB_RETRANS
) {
2374 if (sacked
& TCPCB_SACKED_RETRANS
)
2375 tp
->retrans_out
-= packets_acked
;
2376 acked
|= FLAG_RETRANS_DATA_ACKED
;
2378 } else if (*seq_rtt
< 0)
2379 *seq_rtt
= now
- scb
->when
;
2380 if (sacked
& TCPCB_SACKED_ACKED
)
2381 tp
->sacked_out
-= packets_acked
;
2382 if (sacked
& TCPCB_LOST
)
2383 tp
->lost_out
-= packets_acked
;
2384 if (sacked
& TCPCB_URG
) {
2386 !before(seq
, tp
->snd_up
))
2389 } else if (*seq_rtt
< 0)
2390 *seq_rtt
= now
- scb
->when
;
2392 if (tp
->fackets_out
) {
2393 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2394 tp
->fackets_out
-= dval
;
2396 tp
->packets_out
-= packets_acked
;
2398 BUG_ON(tcp_skb_pcount(skb
) == 0);
2399 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2405 static u32
tcp_usrtt(struct timeval
*tv
)
2409 do_gettimeofday(&now
);
2410 return (now
.tv_sec
- tv
->tv_sec
) * 1000000 + (now
.tv_usec
- tv
->tv_usec
);
2413 /* Remove acknowledged frames from the retransmission queue. */
2414 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2416 struct tcp_sock
*tp
= tcp_sk(sk
);
2417 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2418 struct sk_buff
*skb
;
2419 __u32 now
= tcp_time_stamp
;
2423 void (*rtt_sample
)(struct sock
*sk
, u32 usrtt
)
2424 = icsk
->icsk_ca_ops
->rtt_sample
;
2425 struct timeval tv
= { .tv_sec
= 0, .tv_usec
= 0 };
2427 while ((skb
= tcp_write_queue_head(sk
)) &&
2428 skb
!= tcp_send_head(sk
)) {
2429 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2430 __u8 sacked
= scb
->sacked
;
2432 /* If our packet is before the ack sequence we can
2433 * discard it as it's confirmed to have arrived at
2436 if (after(scb
->end_seq
, tp
->snd_una
)) {
2437 if (tcp_skb_pcount(skb
) > 1 &&
2438 after(tp
->snd_una
, scb
->seq
))
2439 acked
|= tcp_tso_acked(sk
, skb
,
2444 /* Initial outgoing SYN's get put onto the write_queue
2445 * just like anything else we transmit. It is not
2446 * true data, and if we misinform our callers that
2447 * this ACK acks real data, we will erroneously exit
2448 * connection startup slow start one packet too
2449 * quickly. This is severely frowned upon behavior.
2451 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2452 acked
|= FLAG_DATA_ACKED
;
2455 acked
|= FLAG_SYN_ACKED
;
2456 tp
->retrans_stamp
= 0;
2459 /* MTU probing checks */
2460 if (icsk
->icsk_mtup
.probe_size
) {
2461 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2462 tcp_mtup_probe_success(sk
, skb
);
2467 if (sacked
& TCPCB_RETRANS
) {
2468 if (sacked
& TCPCB_SACKED_RETRANS
)
2469 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2470 acked
|= FLAG_RETRANS_DATA_ACKED
;
2472 } else if (seq_rtt
< 0) {
2473 seq_rtt
= now
- scb
->when
;
2474 skb_get_timestamp(skb
, &tv
);
2476 if (sacked
& TCPCB_SACKED_ACKED
)
2477 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2478 if (sacked
& TCPCB_LOST
)
2479 tp
->lost_out
-= tcp_skb_pcount(skb
);
2480 if (sacked
& TCPCB_URG
) {
2482 !before(scb
->end_seq
, tp
->snd_up
))
2485 } else if (seq_rtt
< 0) {
2486 seq_rtt
= now
- scb
->when
;
2487 skb_get_timestamp(skb
, &tv
);
2489 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2490 tcp_packets_out_dec(tp
, skb
);
2491 tcp_unlink_write_queue(skb
, sk
);
2492 sk_stream_free_skb(sk
, skb
);
2493 clear_all_retrans_hints(tp
);
2496 if (acked
&FLAG_ACKED
) {
2497 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2498 tcp_ack_packets_out(sk
);
2499 if (rtt_sample
&& !(acked
& FLAG_RETRANS_DATA_ACKED
))
2500 (*rtt_sample
)(sk
, tcp_usrtt(&tv
));
2502 if (icsk
->icsk_ca_ops
->pkts_acked
)
2503 icsk
->icsk_ca_ops
->pkts_acked(sk
, pkts_acked
);
2506 #if FASTRETRANS_DEBUG > 0
2507 BUG_TRAP((int)tp
->sacked_out
>= 0);
2508 BUG_TRAP((int)tp
->lost_out
>= 0);
2509 BUG_TRAP((int)tp
->retrans_out
>= 0);
2510 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2511 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2513 printk(KERN_DEBUG
"Leak l=%u %d\n",
2514 tp
->lost_out
, icsk
->icsk_ca_state
);
2517 if (tp
->sacked_out
) {
2518 printk(KERN_DEBUG
"Leak s=%u %d\n",
2519 tp
->sacked_out
, icsk
->icsk_ca_state
);
2522 if (tp
->retrans_out
) {
2523 printk(KERN_DEBUG
"Leak r=%u %d\n",
2524 tp
->retrans_out
, icsk
->icsk_ca_state
);
2525 tp
->retrans_out
= 0;
2529 *seq_rtt_p
= seq_rtt
;
2533 static void tcp_ack_probe(struct sock
*sk
)
2535 const struct tcp_sock
*tp
= tcp_sk(sk
);
2536 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2538 /* Was it a usable window open? */
2540 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2541 tp
->snd_una
+ tp
->snd_wnd
)) {
2542 icsk
->icsk_backoff
= 0;
2543 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2544 /* Socket must be waked up by subsequent tcp_data_snd_check().
2545 * This function is not for random using!
2548 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2549 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2554 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2556 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2557 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2560 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2562 const struct tcp_sock
*tp
= tcp_sk(sk
);
2563 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2564 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2567 /* Check that window update is acceptable.
2568 * The function assumes that snd_una<=ack<=snd_next.
2570 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2571 const u32 ack_seq
, const u32 nwin
)
2573 return (after(ack
, tp
->snd_una
) ||
2574 after(ack_seq
, tp
->snd_wl1
) ||
2575 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2578 /* Update our send window.
2580 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2581 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2583 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2586 struct tcp_sock
*tp
= tcp_sk(sk
);
2588 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2590 if (likely(!tcp_hdr(skb
)->syn
))
2591 nwin
<<= tp
->rx_opt
.snd_wscale
;
2593 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2594 flag
|= FLAG_WIN_UPDATE
;
2595 tcp_update_wl(tp
, ack
, ack_seq
);
2597 if (tp
->snd_wnd
!= nwin
) {
2600 /* Note, it is the only place, where
2601 * fast path is recovered for sending TCP.
2604 tcp_fast_path_check(sk
);
2606 if (nwin
> tp
->max_window
) {
2607 tp
->max_window
= nwin
;
2608 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2618 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2619 * continue in congestion avoidance.
2621 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2623 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2624 tp
->snd_cwnd_cnt
= 0;
2625 tcp_moderate_cwnd(tp
);
2628 /* A conservative spurious RTO response algorithm: reduce cwnd using
2629 * rate halving and continue in congestion avoidance.
2631 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
2633 tcp_enter_cwr(sk
, 0);
2636 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
2639 tcp_ratehalving_spur_to_response(sk
);
2641 tcp_undo_cwr(sk
, 1);
2644 /* F-RTO spurious RTO detection algorithm (RFC4138)
2646 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2647 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2648 * window (but not to or beyond highest sequence sent before RTO):
2649 * On First ACK, send two new segments out.
2650 * On Second ACK, RTO was likely spurious. Do spurious response (response
2651 * algorithm is not part of the F-RTO detection algorithm
2652 * given in RFC4138 but can be selected separately).
2653 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2654 * and TCP falls back to conventional RTO recovery.
2656 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2657 * original window even after we transmit two new data segments.
2660 * on first step, wait until first cumulative ACK arrives, then move to
2661 * the second step. In second step, the next ACK decides.
2663 * F-RTO is implemented (mainly) in four functions:
2664 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2665 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2666 * called when tcp_use_frto() showed green light
2667 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2668 * - tcp_enter_frto_loss() is called if there is not enough evidence
2669 * to prove that the RTO is indeed spurious. It transfers the control
2670 * from F-RTO to the conventional RTO recovery
2672 static int tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
, int flag
)
2674 struct tcp_sock
*tp
= tcp_sk(sk
);
2676 tcp_sync_left_out(tp
);
2678 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2679 if (flag
&FLAG_DATA_ACKED
)
2680 inet_csk(sk
)->icsk_retransmits
= 0;
2682 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2683 tcp_enter_frto_loss(sk
, tp
->frto_counter
+ 1, flag
);
2687 if (!IsSackFrto() || IsReno(tp
)) {
2688 /* RFC4138 shortcoming in step 2; should also have case c):
2689 * ACK isn't duplicate nor advances window, e.g., opposite dir
2692 if ((tp
->snd_una
== prior_snd_una
) && (flag
&FLAG_NOT_DUP
) &&
2693 !(flag
&FLAG_FORWARD_PROGRESS
))
2696 if (!(flag
&FLAG_DATA_ACKED
)) {
2697 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2702 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2703 /* Prevent sending of new data. */
2704 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2705 tcp_packets_in_flight(tp
));
2709 if ((tp
->frto_counter
== 2) &&
2710 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2711 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2712 /* RFC4138 shortcoming (see comment above) */
2713 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2716 tcp_enter_frto_loss(sk
, 3, flag
);
2721 if (tp
->frto_counter
== 1) {
2722 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2723 tp
->frto_counter
= 2;
2725 } else /* frto_counter == 2 */ {
2726 switch (sysctl_tcp_frto_response
) {
2728 tcp_undo_spur_to_response(sk
, flag
);
2731 tcp_conservative_spur_to_response(tp
);
2734 tcp_ratehalving_spur_to_response(sk
);
2737 tp
->frto_counter
= 0;
2742 /* This routine deals with incoming acks, but not outgoing ones. */
2743 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2745 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2746 struct tcp_sock
*tp
= tcp_sk(sk
);
2747 u32 prior_snd_una
= tp
->snd_una
;
2748 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2749 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2750 u32 prior_in_flight
;
2755 /* If the ack is newer than sent or older than previous acks
2756 * then we can probably ignore it.
2758 if (after(ack
, tp
->snd_nxt
))
2759 goto uninteresting_ack
;
2761 if (before(ack
, prior_snd_una
))
2764 if (sysctl_tcp_abc
) {
2765 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2766 tp
->bytes_acked
+= ack
- prior_snd_una
;
2767 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2768 /* we assume just one segment left network */
2769 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2772 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2773 /* Window is constant, pure forward advance.
2774 * No more checks are required.
2775 * Note, we use the fact that SND.UNA>=SND.WL2.
2777 tcp_update_wl(tp
, ack
, ack_seq
);
2779 flag
|= FLAG_WIN_UPDATE
;
2781 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2783 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2785 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2788 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2790 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
2792 if (TCP_SKB_CB(skb
)->sacked
)
2793 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2795 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
2798 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2801 /* We passed data and got it acked, remove any soft error
2802 * log. Something worked...
2804 sk
->sk_err_soft
= 0;
2805 tp
->rcv_tstamp
= tcp_time_stamp
;
2806 prior_packets
= tp
->packets_out
;
2810 prior_in_flight
= tcp_packets_in_flight(tp
);
2812 /* See if we can take anything off of the retransmit queue. */
2813 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2815 if (tp
->frto_counter
)
2816 frto_cwnd
= tcp_process_frto(sk
, prior_snd_una
, flag
);
2818 if (tcp_ack_is_dubious(sk
, flag
)) {
2819 /* Advance CWND, if state allows this. */
2820 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
2821 tcp_may_raise_cwnd(sk
, flag
))
2822 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2823 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
, flag
);
2825 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
2826 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2829 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2830 dst_confirm(sk
->sk_dst_cache
);
2835 icsk
->icsk_probes_out
= 0;
2837 /* If this ack opens up a zero window, clear backoff. It was
2838 * being used to time the probes, and is probably far higher than
2839 * it needs to be for normal retransmission.
2841 if (tcp_send_head(sk
))
2846 if (TCP_SKB_CB(skb
)->sacked
)
2847 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2850 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2855 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2856 * But, this can also be called on packets in the established flow when
2857 * the fast version below fails.
2859 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2862 struct tcphdr
*th
= tcp_hdr(skb
);
2863 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2865 ptr
= (unsigned char *)(th
+ 1);
2866 opt_rx
->saw_tstamp
= 0;
2868 while (length
> 0) {
2875 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2880 if (opsize
< 2) /* "silly options" */
2882 if (opsize
> length
)
2883 return; /* don't parse partial options */
2886 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2887 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
2889 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2890 in_mss
= opt_rx
->user_mss
;
2891 opt_rx
->mss_clamp
= in_mss
;
2896 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2897 if (sysctl_tcp_window_scaling
) {
2898 __u8 snd_wscale
= *(__u8
*) ptr
;
2899 opt_rx
->wscale_ok
= 1;
2900 if (snd_wscale
> 14) {
2901 if (net_ratelimit())
2902 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2903 "scaling value %d >14 received.\n",
2907 opt_rx
->snd_wscale
= snd_wscale
;
2910 case TCPOPT_TIMESTAMP
:
2911 if (opsize
==TCPOLEN_TIMESTAMP
) {
2912 if ((estab
&& opt_rx
->tstamp_ok
) ||
2913 (!estab
&& sysctl_tcp_timestamps
)) {
2914 opt_rx
->saw_tstamp
= 1;
2915 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
2916 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
2920 case TCPOPT_SACK_PERM
:
2921 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2922 if (sysctl_tcp_sack
) {
2923 opt_rx
->sack_ok
= 1;
2924 tcp_sack_reset(opt_rx
);
2930 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2931 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2933 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2935 #ifdef CONFIG_TCP_MD5SIG
2938 * The MD5 Hash has already been
2939 * checked (see tcp_v{4,6}_do_rcv()).
2951 /* Fast parse options. This hopes to only see timestamps.
2952 * If it is wrong it falls back on tcp_parse_options().
2954 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
2955 struct tcp_sock
*tp
)
2957 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
2958 tp
->rx_opt
.saw_tstamp
= 0;
2960 } else if (tp
->rx_opt
.tstamp_ok
&&
2961 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
2962 __be32
*ptr
= (__be32
*)(th
+ 1);
2963 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
2964 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
2965 tp
->rx_opt
.saw_tstamp
= 1;
2967 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
2969 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
2973 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
2977 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
2979 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
2980 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
2983 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
2985 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
2986 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2987 * extra check below makes sure this can only happen
2988 * for pure ACK frames. -DaveM
2990 * Not only, also it occurs for expired timestamps.
2993 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
2994 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
2995 tcp_store_ts_recent(tp
);
2999 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3001 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3002 * it can pass through stack. So, the following predicate verifies that
3003 * this segment is not used for anything but congestion avoidance or
3004 * fast retransmit. Moreover, we even are able to eliminate most of such
3005 * second order effects, if we apply some small "replay" window (~RTO)
3006 * to timestamp space.
3008 * All these measures still do not guarantee that we reject wrapped ACKs
3009 * on networks with high bandwidth, when sequence space is recycled fastly,
3010 * but it guarantees that such events will be very rare and do not affect
3011 * connection seriously. This doesn't look nice, but alas, PAWS is really
3014 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3015 * states that events when retransmit arrives after original data are rare.
3016 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3017 * the biggest problem on large power networks even with minor reordering.
3018 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3019 * up to bandwidth of 18Gigabit/sec. 8) ]
3022 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3024 struct tcp_sock
*tp
= tcp_sk(sk
);
3025 struct tcphdr
*th
= tcp_hdr(skb
);
3026 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3027 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3029 return (/* 1. Pure ACK with correct sequence number. */
3030 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3032 /* 2. ... and duplicate ACK. */
3033 ack
== tp
->snd_una
&&
3035 /* 3. ... and does not update window. */
3036 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3038 /* 4. ... and sits in replay window. */
3039 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3042 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3044 const struct tcp_sock
*tp
= tcp_sk(sk
);
3045 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3046 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3047 !tcp_disordered_ack(sk
, skb
));
3050 /* Check segment sequence number for validity.
3052 * Segment controls are considered valid, if the segment
3053 * fits to the window after truncation to the window. Acceptability
3054 * of data (and SYN, FIN, of course) is checked separately.
3055 * See tcp_data_queue(), for example.
3057 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3058 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3059 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3060 * (borrowed from freebsd)
3063 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3065 return !before(end_seq
, tp
->rcv_wup
) &&
3066 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3069 /* When we get a reset we do this. */
3070 static void tcp_reset(struct sock
*sk
)
3072 /* We want the right error as BSD sees it (and indeed as we do). */
3073 switch (sk
->sk_state
) {
3075 sk
->sk_err
= ECONNREFUSED
;
3077 case TCP_CLOSE_WAIT
:
3083 sk
->sk_err
= ECONNRESET
;
3086 if (!sock_flag(sk
, SOCK_DEAD
))
3087 sk
->sk_error_report(sk
);
3093 * Process the FIN bit. This now behaves as it is supposed to work
3094 * and the FIN takes effect when it is validly part of sequence
3095 * space. Not before when we get holes.
3097 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3098 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3101 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3102 * close and we go into CLOSING (and later onto TIME-WAIT)
3104 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3106 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3108 struct tcp_sock
*tp
= tcp_sk(sk
);
3110 inet_csk_schedule_ack(sk
);
3112 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3113 sock_set_flag(sk
, SOCK_DONE
);
3115 switch (sk
->sk_state
) {
3117 case TCP_ESTABLISHED
:
3118 /* Move to CLOSE_WAIT */
3119 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3120 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3123 case TCP_CLOSE_WAIT
:
3125 /* Received a retransmission of the FIN, do
3130 /* RFC793: Remain in the LAST-ACK state. */
3134 /* This case occurs when a simultaneous close
3135 * happens, we must ack the received FIN and
3136 * enter the CLOSING state.
3139 tcp_set_state(sk
, TCP_CLOSING
);
3142 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3144 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3147 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3148 * cases we should never reach this piece of code.
3150 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3151 __FUNCTION__
, sk
->sk_state
);
3155 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3156 * Probably, we should reset in this case. For now drop them.
3158 __skb_queue_purge(&tp
->out_of_order_queue
);
3159 if (tp
->rx_opt
.sack_ok
)
3160 tcp_sack_reset(&tp
->rx_opt
);
3161 sk_stream_mem_reclaim(sk
);
3163 if (!sock_flag(sk
, SOCK_DEAD
)) {
3164 sk
->sk_state_change(sk
);
3166 /* Do not send POLL_HUP for half duplex close. */
3167 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3168 sk
->sk_state
== TCP_CLOSE
)
3169 sk_wake_async(sk
, 1, POLL_HUP
);
3171 sk_wake_async(sk
, 1, POLL_IN
);
3175 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3177 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3178 if (before(seq
, sp
->start_seq
))
3179 sp
->start_seq
= seq
;
3180 if (after(end_seq
, sp
->end_seq
))
3181 sp
->end_seq
= end_seq
;
3187 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3189 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3190 if (before(seq
, tp
->rcv_nxt
))
3191 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3193 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3195 tp
->rx_opt
.dsack
= 1;
3196 tp
->duplicate_sack
[0].start_seq
= seq
;
3197 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3198 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3202 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3204 if (!tp
->rx_opt
.dsack
)
3205 tcp_dsack_set(tp
, seq
, end_seq
);
3207 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3210 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3212 struct tcp_sock
*tp
= tcp_sk(sk
);
3214 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3215 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3216 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3217 tcp_enter_quickack_mode(sk
);
3219 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3220 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3222 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3223 end_seq
= tp
->rcv_nxt
;
3224 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3231 /* These routines update the SACK block as out-of-order packets arrive or
3232 * in-order packets close up the sequence space.
3234 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3237 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3238 struct tcp_sack_block
*swalk
= sp
+1;
3240 /* See if the recent change to the first SACK eats into
3241 * or hits the sequence space of other SACK blocks, if so coalesce.
3243 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3244 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3247 /* Zap SWALK, by moving every further SACK up by one slot.
3248 * Decrease num_sacks.
3250 tp
->rx_opt
.num_sacks
--;
3251 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3252 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3256 this_sack
++, swalk
++;
3260 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3264 tmp
= sack1
->start_seq
;
3265 sack1
->start_seq
= sack2
->start_seq
;
3266 sack2
->start_seq
= tmp
;
3268 tmp
= sack1
->end_seq
;
3269 sack1
->end_seq
= sack2
->end_seq
;
3270 sack2
->end_seq
= tmp
;
3273 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3275 struct tcp_sock
*tp
= tcp_sk(sk
);
3276 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3277 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3283 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3284 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3285 /* Rotate this_sack to the first one. */
3286 for (; this_sack
>0; this_sack
--, sp
--)
3287 tcp_sack_swap(sp
, sp
-1);
3289 tcp_sack_maybe_coalesce(tp
);
3294 /* Could not find an adjacent existing SACK, build a new one,
3295 * put it at the front, and shift everyone else down. We
3296 * always know there is at least one SACK present already here.
3298 * If the sack array is full, forget about the last one.
3300 if (this_sack
>= 4) {
3302 tp
->rx_opt
.num_sacks
--;
3305 for (; this_sack
> 0; this_sack
--, sp
--)
3309 /* Build the new head SACK, and we're done. */
3310 sp
->start_seq
= seq
;
3311 sp
->end_seq
= end_seq
;
3312 tp
->rx_opt
.num_sacks
++;
3313 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3316 /* RCV.NXT advances, some SACKs should be eaten. */
3318 static void tcp_sack_remove(struct tcp_sock
*tp
)
3320 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3321 int num_sacks
= tp
->rx_opt
.num_sacks
;
3324 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3325 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3326 tp
->rx_opt
.num_sacks
= 0;
3327 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3331 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3332 /* Check if the start of the sack is covered by RCV.NXT. */
3333 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3336 /* RCV.NXT must cover all the block! */
3337 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3339 /* Zap this SACK, by moving forward any other SACKS. */
3340 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3341 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3348 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3349 tp
->rx_opt
.num_sacks
= num_sacks
;
3350 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3354 /* This one checks to see if we can put data from the
3355 * out_of_order queue into the receive_queue.
3357 static void tcp_ofo_queue(struct sock
*sk
)
3359 struct tcp_sock
*tp
= tcp_sk(sk
);
3360 __u32 dsack_high
= tp
->rcv_nxt
;
3361 struct sk_buff
*skb
;
3363 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3364 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3367 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3368 __u32 dsack
= dsack_high
;
3369 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3370 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3371 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3374 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3375 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3376 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3380 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3381 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3382 TCP_SKB_CB(skb
)->end_seq
);
3384 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3385 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3386 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3387 if (tcp_hdr(skb
)->fin
)
3388 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3392 static int tcp_prune_queue(struct sock
*sk
);
3394 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3396 struct tcphdr
*th
= tcp_hdr(skb
);
3397 struct tcp_sock
*tp
= tcp_sk(sk
);
3400 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3403 __skb_pull(skb
, th
->doff
*4);
3405 TCP_ECN_accept_cwr(tp
, skb
);
3407 if (tp
->rx_opt
.dsack
) {
3408 tp
->rx_opt
.dsack
= 0;
3409 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3410 4 - tp
->rx_opt
.tstamp_ok
);
3413 /* Queue data for delivery to the user.
3414 * Packets in sequence go to the receive queue.
3415 * Out of sequence packets to the out_of_order_queue.
3417 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3418 if (tcp_receive_window(tp
) == 0)
3421 /* Ok. In sequence. In window. */
3422 if (tp
->ucopy
.task
== current
&&
3423 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3424 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3425 int chunk
= min_t(unsigned int, skb
->len
,
3428 __set_current_state(TASK_RUNNING
);
3431 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3432 tp
->ucopy
.len
-= chunk
;
3433 tp
->copied_seq
+= chunk
;
3434 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3435 tcp_rcv_space_adjust(sk
);
3443 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3444 !sk_stream_rmem_schedule(sk
, skb
))) {
3445 if (tcp_prune_queue(sk
) < 0 ||
3446 !sk_stream_rmem_schedule(sk
, skb
))
3449 sk_stream_set_owner_r(skb
, sk
);
3450 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3452 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3454 tcp_event_data_recv(sk
, skb
);
3456 tcp_fin(skb
, sk
, th
);
3458 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3461 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3462 * gap in queue is filled.
3464 if (skb_queue_empty(&tp
->out_of_order_queue
))
3465 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3468 if (tp
->rx_opt
.num_sacks
)
3469 tcp_sack_remove(tp
);
3471 tcp_fast_path_check(sk
);
3475 else if (!sock_flag(sk
, SOCK_DEAD
))
3476 sk
->sk_data_ready(sk
, 0);
3480 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3481 /* A retransmit, 2nd most common case. Force an immediate ack. */
3482 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3483 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3486 tcp_enter_quickack_mode(sk
);
3487 inet_csk_schedule_ack(sk
);
3493 /* Out of window. F.e. zero window probe. */
3494 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3497 tcp_enter_quickack_mode(sk
);
3499 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3500 /* Partial packet, seq < rcv_next < end_seq */
3501 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3502 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3503 TCP_SKB_CB(skb
)->end_seq
);
3505 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3507 /* If window is closed, drop tail of packet. But after
3508 * remembering D-SACK for its head made in previous line.
3510 if (!tcp_receive_window(tp
))
3515 TCP_ECN_check_ce(tp
, skb
);
3517 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3518 !sk_stream_rmem_schedule(sk
, skb
)) {
3519 if (tcp_prune_queue(sk
) < 0 ||
3520 !sk_stream_rmem_schedule(sk
, skb
))
3524 /* Disable header prediction. */
3526 inet_csk_schedule_ack(sk
);
3528 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3529 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3531 sk_stream_set_owner_r(skb
, sk
);
3533 if (!skb_peek(&tp
->out_of_order_queue
)) {
3534 /* Initial out of order segment, build 1 SACK. */
3535 if (tp
->rx_opt
.sack_ok
) {
3536 tp
->rx_opt
.num_sacks
= 1;
3537 tp
->rx_opt
.dsack
= 0;
3538 tp
->rx_opt
.eff_sacks
= 1;
3539 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3540 tp
->selective_acks
[0].end_seq
=
3541 TCP_SKB_CB(skb
)->end_seq
;
3543 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3545 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3546 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3547 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3549 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3550 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3552 if (!tp
->rx_opt
.num_sacks
||
3553 tp
->selective_acks
[0].end_seq
!= seq
)
3556 /* Common case: data arrive in order after hole. */
3557 tp
->selective_acks
[0].end_seq
= end_seq
;
3561 /* Find place to insert this segment. */
3563 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3565 } while ((skb1
= skb1
->prev
) !=
3566 (struct sk_buff
*)&tp
->out_of_order_queue
);
3568 /* Do skb overlap to previous one? */
3569 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3570 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3571 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3572 /* All the bits are present. Drop. */
3574 tcp_dsack_set(tp
, seq
, end_seq
);
3577 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3578 /* Partial overlap. */
3579 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3584 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3586 /* And clean segments covered by new one as whole. */
3587 while ((skb1
= skb
->next
) !=
3588 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3589 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3590 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3591 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3594 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3595 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3600 if (tp
->rx_opt
.sack_ok
)
3601 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3605 /* Collapse contiguous sequence of skbs head..tail with
3606 * sequence numbers start..end.
3607 * Segments with FIN/SYN are not collapsed (only because this
3611 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3612 struct sk_buff
*head
, struct sk_buff
*tail
,
3615 struct sk_buff
*skb
;
3617 /* First, check that queue is collapsible and find
3618 * the point where collapsing can be useful. */
3619 for (skb
= head
; skb
!= tail
; ) {
3620 /* No new bits? It is possible on ofo queue. */
3621 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3622 struct sk_buff
*next
= skb
->next
;
3623 __skb_unlink(skb
, list
);
3625 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3630 /* The first skb to collapse is:
3632 * - bloated or contains data before "start" or
3633 * overlaps to the next one.
3635 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
3636 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3637 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3638 (skb
->next
!= tail
&&
3639 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3642 /* Decided to skip this, advance start seq. */
3643 start
= TCP_SKB_CB(skb
)->end_seq
;
3646 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
3649 while (before(start
, end
)) {
3650 struct sk_buff
*nskb
;
3651 int header
= skb_headroom(skb
);
3652 int copy
= SKB_MAX_ORDER(header
, 0);
3654 /* Too big header? This can happen with IPv6. */
3657 if (end
-start
< copy
)
3659 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3663 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
3664 skb_set_network_header(nskb
, (skb_network_header(skb
) -
3666 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
3668 skb_reserve(nskb
, header
);
3669 memcpy(nskb
->head
, skb
->head
, header
);
3670 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3671 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3672 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3673 sk_stream_set_owner_r(nskb
, sk
);
3675 /* Copy data, releasing collapsed skbs. */
3677 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3678 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3682 size
= min(copy
, size
);
3683 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3685 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3689 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3690 struct sk_buff
*next
= skb
->next
;
3691 __skb_unlink(skb
, list
);
3693 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3696 tcp_hdr(skb
)->syn
||
3704 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3705 * and tcp_collapse() them until all the queue is collapsed.
3707 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3709 struct tcp_sock
*tp
= tcp_sk(sk
);
3710 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3711 struct sk_buff
*head
;
3717 start
= TCP_SKB_CB(skb
)->seq
;
3718 end
= TCP_SKB_CB(skb
)->end_seq
;
3724 /* Segment is terminated when we see gap or when
3725 * we are at the end of all the queue. */
3726 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3727 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3728 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3729 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3730 head
, skb
, start
, end
);
3732 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3734 /* Start new segment */
3735 start
= TCP_SKB_CB(skb
)->seq
;
3736 end
= TCP_SKB_CB(skb
)->end_seq
;
3738 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3739 start
= TCP_SKB_CB(skb
)->seq
;
3740 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3741 end
= TCP_SKB_CB(skb
)->end_seq
;
3746 /* Reduce allocated memory if we can, trying to get
3747 * the socket within its memory limits again.
3749 * Return less than zero if we should start dropping frames
3750 * until the socket owning process reads some of the data
3751 * to stabilize the situation.
3753 static int tcp_prune_queue(struct sock
*sk
)
3755 struct tcp_sock
*tp
= tcp_sk(sk
);
3757 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3759 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3761 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3762 tcp_clamp_window(sk
);
3763 else if (tcp_memory_pressure
)
3764 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3766 tcp_collapse_ofo_queue(sk
);
3767 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3768 sk
->sk_receive_queue
.next
,
3769 (struct sk_buff
*)&sk
->sk_receive_queue
,
3770 tp
->copied_seq
, tp
->rcv_nxt
);
3771 sk_stream_mem_reclaim(sk
);
3773 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3776 /* Collapsing did not help, destructive actions follow.
3777 * This must not ever occur. */
3779 /* First, purge the out_of_order queue. */
3780 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3781 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3782 __skb_queue_purge(&tp
->out_of_order_queue
);
3784 /* Reset SACK state. A conforming SACK implementation will
3785 * do the same at a timeout based retransmit. When a connection
3786 * is in a sad state like this, we care only about integrity
3787 * of the connection not performance.
3789 if (tp
->rx_opt
.sack_ok
)
3790 tcp_sack_reset(&tp
->rx_opt
);
3791 sk_stream_mem_reclaim(sk
);
3794 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3797 /* If we are really being abused, tell the caller to silently
3798 * drop receive data on the floor. It will get retransmitted
3799 * and hopefully then we'll have sufficient space.
3801 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3803 /* Massive buffer overcommit. */
3809 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3810 * As additional protections, we do not touch cwnd in retransmission phases,
3811 * and if application hit its sndbuf limit recently.
3813 void tcp_cwnd_application_limited(struct sock
*sk
)
3815 struct tcp_sock
*tp
= tcp_sk(sk
);
3817 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3818 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3819 /* Limited by application or receiver window. */
3820 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
3821 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
3822 if (win_used
< tp
->snd_cwnd
) {
3823 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3824 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3826 tp
->snd_cwnd_used
= 0;
3828 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3831 static int tcp_should_expand_sndbuf(struct sock
*sk
)
3833 struct tcp_sock
*tp
= tcp_sk(sk
);
3835 /* If the user specified a specific send buffer setting, do
3838 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3841 /* If we are under global TCP memory pressure, do not expand. */
3842 if (tcp_memory_pressure
)
3845 /* If we are under soft global TCP memory pressure, do not expand. */
3846 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3849 /* If we filled the congestion window, do not expand. */
3850 if (tp
->packets_out
>= tp
->snd_cwnd
)
3856 /* When incoming ACK allowed to free some skb from write_queue,
3857 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3858 * on the exit from tcp input handler.
3860 * PROBLEM: sndbuf expansion does not work well with largesend.
3862 static void tcp_new_space(struct sock
*sk
)
3864 struct tcp_sock
*tp
= tcp_sk(sk
);
3866 if (tcp_should_expand_sndbuf(sk
)) {
3867 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3868 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3869 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3870 tp
->reordering
+ 1);
3871 sndmem
*= 2*demanded
;
3872 if (sndmem
> sk
->sk_sndbuf
)
3873 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3874 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3877 sk
->sk_write_space(sk
);
3880 static void tcp_check_space(struct sock
*sk
)
3882 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3883 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3884 if (sk
->sk_socket
&&
3885 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3890 static inline void tcp_data_snd_check(struct sock
*sk
)
3892 tcp_push_pending_frames(sk
);
3893 tcp_check_space(sk
);
3897 * Check if sending an ack is needed.
3899 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3901 struct tcp_sock
*tp
= tcp_sk(sk
);
3903 /* More than one full frame received... */
3904 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3905 /* ... and right edge of window advances far enough.
3906 * (tcp_recvmsg() will send ACK otherwise). Or...
3908 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3909 /* We ACK each frame or... */
3910 tcp_in_quickack_mode(sk
) ||
3911 /* We have out of order data. */
3913 skb_peek(&tp
->out_of_order_queue
))) {
3914 /* Then ack it now */
3917 /* Else, send delayed ack. */
3918 tcp_send_delayed_ack(sk
);
3922 static inline void tcp_ack_snd_check(struct sock
*sk
)
3924 if (!inet_csk_ack_scheduled(sk
)) {
3925 /* We sent a data segment already. */
3928 __tcp_ack_snd_check(sk
, 1);
3932 * This routine is only called when we have urgent data
3933 * signaled. Its the 'slow' part of tcp_urg. It could be
3934 * moved inline now as tcp_urg is only called from one
3935 * place. We handle URGent data wrong. We have to - as
3936 * BSD still doesn't use the correction from RFC961.
3937 * For 1003.1g we should support a new option TCP_STDURG to permit
3938 * either form (or just set the sysctl tcp_stdurg).
3941 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3943 struct tcp_sock
*tp
= tcp_sk(sk
);
3944 u32 ptr
= ntohs(th
->urg_ptr
);
3946 if (ptr
&& !sysctl_tcp_stdurg
)
3948 ptr
+= ntohl(th
->seq
);
3950 /* Ignore urgent data that we've already seen and read. */
3951 if (after(tp
->copied_seq
, ptr
))
3954 /* Do not replay urg ptr.
3956 * NOTE: interesting situation not covered by specs.
3957 * Misbehaving sender may send urg ptr, pointing to segment,
3958 * which we already have in ofo queue. We are not able to fetch
3959 * such data and will stay in TCP_URG_NOTYET until will be eaten
3960 * by recvmsg(). Seems, we are not obliged to handle such wicked
3961 * situations. But it is worth to think about possibility of some
3962 * DoSes using some hypothetical application level deadlock.
3964 if (before(ptr
, tp
->rcv_nxt
))
3967 /* Do we already have a newer (or duplicate) urgent pointer? */
3968 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
3971 /* Tell the world about our new urgent pointer. */
3974 /* We may be adding urgent data when the last byte read was
3975 * urgent. To do this requires some care. We cannot just ignore
3976 * tp->copied_seq since we would read the last urgent byte again
3977 * as data, nor can we alter copied_seq until this data arrives
3978 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3980 * NOTE. Double Dutch. Rendering to plain English: author of comment
3981 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3982 * and expect that both A and B disappear from stream. This is _wrong_.
3983 * Though this happens in BSD with high probability, this is occasional.
3984 * Any application relying on this is buggy. Note also, that fix "works"
3985 * only in this artificial test. Insert some normal data between A and B and we will
3986 * decline of BSD again. Verdict: it is better to remove to trap
3989 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
3990 !sock_flag(sk
, SOCK_URGINLINE
) &&
3991 tp
->copied_seq
!= tp
->rcv_nxt
) {
3992 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
3994 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
3995 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4000 tp
->urg_data
= TCP_URG_NOTYET
;
4003 /* Disable header prediction. */
4007 /* This is the 'fast' part of urgent handling. */
4008 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4010 struct tcp_sock
*tp
= tcp_sk(sk
);
4012 /* Check if we get a new urgent pointer - normally not. */
4014 tcp_check_urg(sk
,th
);
4016 /* Do we wait for any urgent data? - normally not... */
4017 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4018 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4021 /* Is the urgent pointer pointing into this packet? */
4022 if (ptr
< skb
->len
) {
4024 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4026 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4027 if (!sock_flag(sk
, SOCK_DEAD
))
4028 sk
->sk_data_ready(sk
, 0);
4033 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4035 struct tcp_sock
*tp
= tcp_sk(sk
);
4036 int chunk
= skb
->len
- hlen
;
4040 if (skb_csum_unnecessary(skb
))
4041 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4043 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4047 tp
->ucopy
.len
-= chunk
;
4048 tp
->copied_seq
+= chunk
;
4049 tcp_rcv_space_adjust(sk
);
4056 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4060 if (sock_owned_by_user(sk
)) {
4062 result
= __tcp_checksum_complete(skb
);
4065 result
= __tcp_checksum_complete(skb
);
4070 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4072 return !skb_csum_unnecessary(skb
) &&
4073 __tcp_checksum_complete_user(sk
, skb
);
4076 #ifdef CONFIG_NET_DMA
4077 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4079 struct tcp_sock
*tp
= tcp_sk(sk
);
4080 int chunk
= skb
->len
- hlen
;
4082 int copied_early
= 0;
4084 if (tp
->ucopy
.wakeup
)
4087 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4088 tp
->ucopy
.dma_chan
= get_softnet_dma();
4090 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4092 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4093 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4098 tp
->ucopy
.dma_cookie
= dma_cookie
;
4101 tp
->ucopy
.len
-= chunk
;
4102 tp
->copied_seq
+= chunk
;
4103 tcp_rcv_space_adjust(sk
);
4105 if ((tp
->ucopy
.len
== 0) ||
4106 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4107 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4108 tp
->ucopy
.wakeup
= 1;
4109 sk
->sk_data_ready(sk
, 0);
4111 } else if (chunk
> 0) {
4112 tp
->ucopy
.wakeup
= 1;
4113 sk
->sk_data_ready(sk
, 0);
4116 return copied_early
;
4118 #endif /* CONFIG_NET_DMA */
4121 * TCP receive function for the ESTABLISHED state.
4123 * It is split into a fast path and a slow path. The fast path is
4125 * - A zero window was announced from us - zero window probing
4126 * is only handled properly in the slow path.
4127 * - Out of order segments arrived.
4128 * - Urgent data is expected.
4129 * - There is no buffer space left
4130 * - Unexpected TCP flags/window values/header lengths are received
4131 * (detected by checking the TCP header against pred_flags)
4132 * - Data is sent in both directions. Fast path only supports pure senders
4133 * or pure receivers (this means either the sequence number or the ack
4134 * value must stay constant)
4135 * - Unexpected TCP option.
4137 * When these conditions are not satisfied it drops into a standard
4138 * receive procedure patterned after RFC793 to handle all cases.
4139 * The first three cases are guaranteed by proper pred_flags setting,
4140 * the rest is checked inline. Fast processing is turned on in
4141 * tcp_data_queue when everything is OK.
4143 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4144 struct tcphdr
*th
, unsigned len
)
4146 struct tcp_sock
*tp
= tcp_sk(sk
);
4149 * Header prediction.
4150 * The code loosely follows the one in the famous
4151 * "30 instruction TCP receive" Van Jacobson mail.
4153 * Van's trick is to deposit buffers into socket queue
4154 * on a device interrupt, to call tcp_recv function
4155 * on the receive process context and checksum and copy
4156 * the buffer to user space. smart...
4158 * Our current scheme is not silly either but we take the
4159 * extra cost of the net_bh soft interrupt processing...
4160 * We do checksum and copy also but from device to kernel.
4163 tp
->rx_opt
.saw_tstamp
= 0;
4165 /* pred_flags is 0xS?10 << 16 + snd_wnd
4166 * if header_prediction is to be made
4167 * 'S' will always be tp->tcp_header_len >> 2
4168 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4169 * turn it off (when there are holes in the receive
4170 * space for instance)
4171 * PSH flag is ignored.
4174 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4175 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4176 int tcp_header_len
= tp
->tcp_header_len
;
4178 /* Timestamp header prediction: tcp_header_len
4179 * is automatically equal to th->doff*4 due to pred_flags
4183 /* Check timestamp */
4184 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4185 __be32
*ptr
= (__be32
*)(th
+ 1);
4187 /* No? Slow path! */
4188 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4189 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4192 tp
->rx_opt
.saw_tstamp
= 1;
4194 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4196 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4198 /* If PAWS failed, check it more carefully in slow path */
4199 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4202 /* DO NOT update ts_recent here, if checksum fails
4203 * and timestamp was corrupted part, it will result
4204 * in a hung connection since we will drop all
4205 * future packets due to the PAWS test.
4209 if (len
<= tcp_header_len
) {
4210 /* Bulk data transfer: sender */
4211 if (len
== tcp_header_len
) {
4212 /* Predicted packet is in window by definition.
4213 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4214 * Hence, check seq<=rcv_wup reduces to:
4216 if (tcp_header_len
==
4217 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4218 tp
->rcv_nxt
== tp
->rcv_wup
)
4219 tcp_store_ts_recent(tp
);
4221 /* We know that such packets are checksummed
4224 tcp_ack(sk
, skb
, 0);
4226 tcp_data_snd_check(sk
);
4228 } else { /* Header too small */
4229 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4234 int copied_early
= 0;
4236 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4237 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4238 #ifdef CONFIG_NET_DMA
4239 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4244 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4245 __set_current_state(TASK_RUNNING
);
4247 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4251 /* Predicted packet is in window by definition.
4252 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4253 * Hence, check seq<=rcv_wup reduces to:
4255 if (tcp_header_len
==
4256 (sizeof(struct tcphdr
) +
4257 TCPOLEN_TSTAMP_ALIGNED
) &&
4258 tp
->rcv_nxt
== tp
->rcv_wup
)
4259 tcp_store_ts_recent(tp
);
4261 tcp_rcv_rtt_measure_ts(sk
, skb
);
4263 __skb_pull(skb
, tcp_header_len
);
4264 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4265 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4268 tcp_cleanup_rbuf(sk
, skb
->len
);
4271 if (tcp_checksum_complete_user(sk
, skb
))
4274 /* Predicted packet is in window by definition.
4275 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4276 * Hence, check seq<=rcv_wup reduces to:
4278 if (tcp_header_len
==
4279 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4280 tp
->rcv_nxt
== tp
->rcv_wup
)
4281 tcp_store_ts_recent(tp
);
4283 tcp_rcv_rtt_measure_ts(sk
, skb
);
4285 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4288 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4290 /* Bulk data transfer: receiver */
4291 __skb_pull(skb
,tcp_header_len
);
4292 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4293 sk_stream_set_owner_r(skb
, sk
);
4294 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4297 tcp_event_data_recv(sk
, skb
);
4299 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4300 /* Well, only one small jumplet in fast path... */
4301 tcp_ack(sk
, skb
, FLAG_DATA
);
4302 tcp_data_snd_check(sk
);
4303 if (!inet_csk_ack_scheduled(sk
))
4307 __tcp_ack_snd_check(sk
, 0);
4309 #ifdef CONFIG_NET_DMA
4311 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4317 sk
->sk_data_ready(sk
, 0);
4323 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4327 * RFC1323: H1. Apply PAWS check first.
4329 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4330 tcp_paws_discard(sk
, skb
)) {
4332 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4333 tcp_send_dupack(sk
, skb
);
4336 /* Resets are accepted even if PAWS failed.
4338 ts_recent update must be made after we are sure
4339 that the packet is in window.
4344 * Standard slow path.
4347 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4348 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4349 * (RST) segments are validated by checking their SEQ-fields."
4350 * And page 69: "If an incoming segment is not acceptable,
4351 * an acknowledgment should be sent in reply (unless the RST bit
4352 * is set, if so drop the segment and return)".
4355 tcp_send_dupack(sk
, skb
);
4364 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4366 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4367 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4368 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4375 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4377 tcp_rcv_rtt_measure_ts(sk
, skb
);
4379 /* Process urgent data. */
4380 tcp_urg(sk
, skb
, th
);
4382 /* step 7: process the segment text */
4383 tcp_data_queue(sk
, skb
);
4385 tcp_data_snd_check(sk
);
4386 tcp_ack_snd_check(sk
);
4390 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4397 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4398 struct tcphdr
*th
, unsigned len
)
4400 struct tcp_sock
*tp
= tcp_sk(sk
);
4401 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4402 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4404 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4408 * "If the state is SYN-SENT then
4409 * first check the ACK bit
4410 * If the ACK bit is set
4411 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4412 * a reset (unless the RST bit is set, if so drop
4413 * the segment and return)"
4415 * We do not send data with SYN, so that RFC-correct
4418 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4419 goto reset_and_undo
;
4421 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4422 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4424 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4425 goto reset_and_undo
;
4428 /* Now ACK is acceptable.
4430 * "If the RST bit is set
4431 * If the ACK was acceptable then signal the user "error:
4432 * connection reset", drop the segment, enter CLOSED state,
4433 * delete TCB, and return."
4442 * "fifth, if neither of the SYN or RST bits is set then
4443 * drop the segment and return."
4449 goto discard_and_undo
;
4452 * "If the SYN bit is on ...
4453 * are acceptable then ...
4454 * (our SYN has been ACKed), change the connection
4455 * state to ESTABLISHED..."
4458 TCP_ECN_rcv_synack(tp
, th
);
4460 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4461 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4463 /* Ok.. it's good. Set up sequence numbers and
4464 * move to established.
4466 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4467 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4469 /* RFC1323: The window in SYN & SYN/ACK segments is
4472 tp
->snd_wnd
= ntohs(th
->window
);
4473 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4475 if (!tp
->rx_opt
.wscale_ok
) {
4476 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4477 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4480 if (tp
->rx_opt
.saw_tstamp
) {
4481 tp
->rx_opt
.tstamp_ok
= 1;
4482 tp
->tcp_header_len
=
4483 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4484 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4485 tcp_store_ts_recent(tp
);
4487 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4490 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4491 tp
->rx_opt
.sack_ok
|= 2;
4494 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4495 tcp_initialize_rcv_mss(sk
);
4497 /* Remember, tcp_poll() does not lock socket!
4498 * Change state from SYN-SENT only after copied_seq
4499 * is initialized. */
4500 tp
->copied_seq
= tp
->rcv_nxt
;
4502 tcp_set_state(sk
, TCP_ESTABLISHED
);
4504 security_inet_conn_established(sk
, skb
);
4506 /* Make sure socket is routed, for correct metrics. */
4507 icsk
->icsk_af_ops
->rebuild_header(sk
);
4509 tcp_init_metrics(sk
);
4511 tcp_init_congestion_control(sk
);
4513 /* Prevent spurious tcp_cwnd_restart() on first data
4516 tp
->lsndtime
= tcp_time_stamp
;
4518 tcp_init_buffer_space(sk
);
4520 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4521 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4523 if (!tp
->rx_opt
.snd_wscale
)
4524 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4528 if (!sock_flag(sk
, SOCK_DEAD
)) {
4529 sk
->sk_state_change(sk
);
4530 sk_wake_async(sk
, 0, POLL_OUT
);
4533 if (sk
->sk_write_pending
||
4534 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4535 icsk
->icsk_ack
.pingpong
) {
4536 /* Save one ACK. Data will be ready after
4537 * several ticks, if write_pending is set.
4539 * It may be deleted, but with this feature tcpdumps
4540 * look so _wonderfully_ clever, that I was not able
4541 * to stand against the temptation 8) --ANK
4543 inet_csk_schedule_ack(sk
);
4544 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4545 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4546 tcp_incr_quickack(sk
);
4547 tcp_enter_quickack_mode(sk
);
4548 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4549 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4560 /* No ACK in the segment */
4564 * "If the RST bit is set
4566 * Otherwise (no ACK) drop the segment and return."
4569 goto discard_and_undo
;
4573 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4574 goto discard_and_undo
;
4577 /* We see SYN without ACK. It is attempt of
4578 * simultaneous connect with crossed SYNs.
4579 * Particularly, it can be connect to self.
4581 tcp_set_state(sk
, TCP_SYN_RECV
);
4583 if (tp
->rx_opt
.saw_tstamp
) {
4584 tp
->rx_opt
.tstamp_ok
= 1;
4585 tcp_store_ts_recent(tp
);
4586 tp
->tcp_header_len
=
4587 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4589 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4592 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4593 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4595 /* RFC1323: The window in SYN & SYN/ACK segments is
4598 tp
->snd_wnd
= ntohs(th
->window
);
4599 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4600 tp
->max_window
= tp
->snd_wnd
;
4602 TCP_ECN_rcv_syn(tp
, th
);
4605 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4606 tcp_initialize_rcv_mss(sk
);
4609 tcp_send_synack(sk
);
4611 /* Note, we could accept data and URG from this segment.
4612 * There are no obstacles to make this.
4614 * However, if we ignore data in ACKless segments sometimes,
4615 * we have no reasons to accept it sometimes.
4616 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4617 * is not flawless. So, discard packet for sanity.
4618 * Uncomment this return to process the data.
4625 /* "fifth, if neither of the SYN or RST bits is set then
4626 * drop the segment and return."
4630 tcp_clear_options(&tp
->rx_opt
);
4631 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4635 tcp_clear_options(&tp
->rx_opt
);
4636 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4642 * This function implements the receiving procedure of RFC 793 for
4643 * all states except ESTABLISHED and TIME_WAIT.
4644 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4645 * address independent.
4648 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4649 struct tcphdr
*th
, unsigned len
)
4651 struct tcp_sock
*tp
= tcp_sk(sk
);
4652 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4655 tp
->rx_opt
.saw_tstamp
= 0;
4657 switch (sk
->sk_state
) {
4669 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4672 /* Now we have several options: In theory there is
4673 * nothing else in the frame. KA9Q has an option to
4674 * send data with the syn, BSD accepts data with the
4675 * syn up to the [to be] advertised window and
4676 * Solaris 2.1 gives you a protocol error. For now
4677 * we just ignore it, that fits the spec precisely
4678 * and avoids incompatibilities. It would be nice in
4679 * future to drop through and process the data.
4681 * Now that TTCP is starting to be used we ought to
4683 * But, this leaves one open to an easy denial of
4684 * service attack, and SYN cookies can't defend
4685 * against this problem. So, we drop the data
4686 * in the interest of security over speed unless
4687 * it's still in use.
4695 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4699 /* Do step6 onward by hand. */
4700 tcp_urg(sk
, skb
, th
);
4702 tcp_data_snd_check(sk
);
4706 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4707 tcp_paws_discard(sk
, skb
)) {
4709 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4710 tcp_send_dupack(sk
, skb
);
4713 /* Reset is accepted even if it did not pass PAWS. */
4716 /* step 1: check sequence number */
4717 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4719 tcp_send_dupack(sk
, skb
);
4723 /* step 2: check RST bit */
4729 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4731 /* step 3: check security and precedence [ignored] */
4735 * Check for a SYN in window.
4737 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4738 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4743 /* step 5: check the ACK field */
4745 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4747 switch (sk
->sk_state
) {
4750 tp
->copied_seq
= tp
->rcv_nxt
;
4752 tcp_set_state(sk
, TCP_ESTABLISHED
);
4753 sk
->sk_state_change(sk
);
4755 /* Note, that this wakeup is only for marginal
4756 * crossed SYN case. Passively open sockets
4757 * are not waked up, because sk->sk_sleep ==
4758 * NULL and sk->sk_socket == NULL.
4760 if (sk
->sk_socket
) {
4761 sk_wake_async(sk
,0,POLL_OUT
);
4764 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4765 tp
->snd_wnd
= ntohs(th
->window
) <<
4766 tp
->rx_opt
.snd_wscale
;
4767 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4768 TCP_SKB_CB(skb
)->seq
);
4770 /* tcp_ack considers this ACK as duplicate
4771 * and does not calculate rtt.
4772 * Fix it at least with timestamps.
4774 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4776 tcp_ack_saw_tstamp(sk
, 0);
4778 if (tp
->rx_opt
.tstamp_ok
)
4779 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4781 /* Make sure socket is routed, for
4784 icsk
->icsk_af_ops
->rebuild_header(sk
);
4786 tcp_init_metrics(sk
);
4788 tcp_init_congestion_control(sk
);
4790 /* Prevent spurious tcp_cwnd_restart() on
4791 * first data packet.
4793 tp
->lsndtime
= tcp_time_stamp
;
4796 tcp_initialize_rcv_mss(sk
);
4797 tcp_init_buffer_space(sk
);
4798 tcp_fast_path_on(tp
);
4805 if (tp
->snd_una
== tp
->write_seq
) {
4806 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4807 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4808 dst_confirm(sk
->sk_dst_cache
);
4810 if (!sock_flag(sk
, SOCK_DEAD
))
4811 /* Wake up lingering close() */
4812 sk
->sk_state_change(sk
);
4816 if (tp
->linger2
< 0 ||
4817 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4818 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4820 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4824 tmo
= tcp_fin_time(sk
);
4825 if (tmo
> TCP_TIMEWAIT_LEN
) {
4826 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4827 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4828 /* Bad case. We could lose such FIN otherwise.
4829 * It is not a big problem, but it looks confusing
4830 * and not so rare event. We still can lose it now,
4831 * if it spins in bh_lock_sock(), but it is really
4834 inet_csk_reset_keepalive_timer(sk
, tmo
);
4836 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4844 if (tp
->snd_una
== tp
->write_seq
) {
4845 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4851 if (tp
->snd_una
== tp
->write_seq
) {
4852 tcp_update_metrics(sk
);
4861 /* step 6: check the URG bit */
4862 tcp_urg(sk
, skb
, th
);
4864 /* step 7: process the segment text */
4865 switch (sk
->sk_state
) {
4866 case TCP_CLOSE_WAIT
:
4869 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4873 /* RFC 793 says to queue data in these states,
4874 * RFC 1122 says we MUST send a reset.
4875 * BSD 4.4 also does reset.
4877 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4878 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4879 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4880 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4886 case TCP_ESTABLISHED
:
4887 tcp_data_queue(sk
, skb
);
4892 /* tcp_data could move socket to TIME-WAIT */
4893 if (sk
->sk_state
!= TCP_CLOSE
) {
4894 tcp_data_snd_check(sk
);
4895 tcp_ack_snd_check(sk
);
4905 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4906 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4907 EXPORT_SYMBOL(tcp_parse_options
);
4908 EXPORT_SYMBOL(tcp_rcv_established
);
4909 EXPORT_SYMBOL(tcp_rcv_state_process
);
4910 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);