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 */
105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained DSACK info */
108 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
109 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
110 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
111 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
112 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
114 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
115 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
116 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
118 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
120 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 /* Adapt the MSS value used to make delayed ack decision to the
125 static void tcp_measure_rcv_mss(struct sock
*sk
,
126 const struct sk_buff
*skb
)
128 struct inet_connection_sock
*icsk
= inet_csk(sk
);
129 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
132 icsk
->icsk_ack
.last_seg_size
= 0;
134 /* skb->len may jitter because of SACKs, even if peer
135 * sends good full-sized frames.
137 len
= skb_shinfo(skb
)->gso_size
?: skb
->len
;
138 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
139 icsk
->icsk_ack
.rcv_mss
= len
;
141 /* Otherwise, we make more careful check taking into account,
142 * that SACKs block is variable.
144 * "len" is invariant segment length, including TCP header.
146 len
+= skb
->data
- skb_transport_header(skb
);
147 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
148 /* If PSH is not set, packet should be
149 * full sized, provided peer TCP is not badly broken.
150 * This observation (if it is correct 8)) allows
151 * to handle super-low mtu links fairly.
153 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
154 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
155 /* Subtract also invariant (if peer is RFC compliant),
156 * tcp header plus fixed timestamp option length.
157 * Resulting "len" is MSS free of SACK jitter.
159 len
-= tcp_sk(sk
)->tcp_header_len
;
160 icsk
->icsk_ack
.last_seg_size
= len
;
162 icsk
->icsk_ack
.rcv_mss
= len
;
166 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
167 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
168 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
172 static void tcp_incr_quickack(struct sock
*sk
)
174 struct inet_connection_sock
*icsk
= inet_csk(sk
);
175 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
179 if (quickacks
> icsk
->icsk_ack
.quick
)
180 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
183 void tcp_enter_quickack_mode(struct sock
*sk
)
185 struct inet_connection_sock
*icsk
= inet_csk(sk
);
186 tcp_incr_quickack(sk
);
187 icsk
->icsk_ack
.pingpong
= 0;
188 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
191 /* Send ACKs quickly, if "quick" count is not exhausted
192 * and the session is not interactive.
195 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
197 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
198 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
201 /* Buffer size and advertised window tuning.
203 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
206 static void tcp_fixup_sndbuf(struct sock
*sk
)
208 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
209 sizeof(struct sk_buff
);
211 if (sk
->sk_sndbuf
< 3 * sndmem
)
212 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
215 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
217 * All tcp_full_space() is split to two parts: "network" buffer, allocated
218 * forward and advertised in receiver window (tp->rcv_wnd) and
219 * "application buffer", required to isolate scheduling/application
220 * latencies from network.
221 * window_clamp is maximal advertised window. It can be less than
222 * tcp_full_space(), in this case tcp_full_space() - window_clamp
223 * is reserved for "application" buffer. The less window_clamp is
224 * the smoother our behaviour from viewpoint of network, but the lower
225 * throughput and the higher sensitivity of the connection to losses. 8)
227 * rcv_ssthresh is more strict window_clamp used at "slow start"
228 * phase to predict further behaviour of this connection.
229 * It is used for two goals:
230 * - to enforce header prediction at sender, even when application
231 * requires some significant "application buffer". It is check #1.
232 * - to prevent pruning of receive queue because of misprediction
233 * of receiver window. Check #2.
235 * The scheme does not work when sender sends good segments opening
236 * window and then starts to feed us spaghetti. But it should work
237 * in common situations. Otherwise, we have to rely on queue collapsing.
240 /* Slow part of check#2. */
241 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
243 struct tcp_sock
*tp
= tcp_sk(sk
);
245 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
246 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
248 while (tp
->rcv_ssthresh
<= window
) {
249 if (truesize
<= skb
->len
)
250 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
258 static void tcp_grow_window(struct sock
*sk
,
261 struct tcp_sock
*tp
= tcp_sk(sk
);
264 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
265 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
266 !tcp_memory_pressure
) {
269 /* Check #2. Increase window, if skb with such overhead
270 * will fit to rcvbuf in future.
272 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
275 incr
= __tcp_grow_window(sk
, skb
);
278 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
279 inet_csk(sk
)->icsk_ack
.quick
|= 1;
284 /* 3. Tuning rcvbuf, when connection enters established state. */
286 static void tcp_fixup_rcvbuf(struct sock
*sk
)
288 struct tcp_sock
*tp
= tcp_sk(sk
);
289 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
291 /* Try to select rcvbuf so that 4 mss-sized segments
292 * will fit to window and corresponding skbs will fit to our rcvbuf.
293 * (was 3; 4 is minimum to allow fast retransmit to work.)
295 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
297 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
298 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
301 /* 4. Try to fixup all. It is made immediately after connection enters
304 static void tcp_init_buffer_space(struct sock
*sk
)
306 struct tcp_sock
*tp
= tcp_sk(sk
);
309 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
310 tcp_fixup_rcvbuf(sk
);
311 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
312 tcp_fixup_sndbuf(sk
);
314 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
316 maxwin
= tcp_full_space(sk
);
318 if (tp
->window_clamp
>= maxwin
) {
319 tp
->window_clamp
= maxwin
;
321 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
322 tp
->window_clamp
= max(maxwin
-
323 (maxwin
>> sysctl_tcp_app_win
),
327 /* Force reservation of one segment. */
328 if (sysctl_tcp_app_win
&&
329 tp
->window_clamp
> 2 * tp
->advmss
&&
330 tp
->window_clamp
+ tp
->advmss
> maxwin
)
331 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
333 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
334 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
337 /* 5. Recalculate window clamp after socket hit its memory bounds. */
338 static void tcp_clamp_window(struct sock
*sk
)
340 struct tcp_sock
*tp
= tcp_sk(sk
);
341 struct inet_connection_sock
*icsk
= inet_csk(sk
);
343 icsk
->icsk_ack
.quick
= 0;
345 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
346 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
347 !tcp_memory_pressure
&&
348 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
349 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
352 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
353 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
357 /* Initialize RCV_MSS value.
358 * RCV_MSS is an our guess about MSS used by the peer.
359 * We haven't any direct information about the MSS.
360 * It's better to underestimate the RCV_MSS rather than overestimate.
361 * Overestimations make us ACKing less frequently than needed.
362 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
364 void tcp_initialize_rcv_mss(struct sock
*sk
)
366 struct tcp_sock
*tp
= tcp_sk(sk
);
367 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
369 hint
= min(hint
, tp
->rcv_wnd
/2);
370 hint
= min(hint
, TCP_MIN_RCVMSS
);
371 hint
= max(hint
, TCP_MIN_MSS
);
373 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
376 /* Receiver "autotuning" code.
378 * The algorithm for RTT estimation w/o timestamps is based on
379 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
380 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
382 * More detail on this code can be found at
383 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
384 * though this reference is out of date. A new paper
387 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
389 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
395 if (new_sample
!= 0) {
396 /* If we sample in larger samples in the non-timestamp
397 * case, we could grossly overestimate the RTT especially
398 * with chatty applications or bulk transfer apps which
399 * are stalled on filesystem I/O.
401 * Also, since we are only going for a minimum in the
402 * non-timestamp case, we do not smooth things out
403 * else with timestamps disabled convergence takes too
407 m
-= (new_sample
>> 3);
409 } else if (m
< new_sample
)
412 /* No previous measure. */
416 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
417 tp
->rcv_rtt_est
.rtt
= new_sample
;
420 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
422 if (tp
->rcv_rtt_est
.time
== 0)
424 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
426 tcp_rcv_rtt_update(tp
,
427 jiffies
- tp
->rcv_rtt_est
.time
,
431 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
432 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
435 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
437 struct tcp_sock
*tp
= tcp_sk(sk
);
438 if (tp
->rx_opt
.rcv_tsecr
&&
439 (TCP_SKB_CB(skb
)->end_seq
-
440 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
441 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
445 * This function should be called every time data is copied to user space.
446 * It calculates the appropriate TCP receive buffer space.
448 void tcp_rcv_space_adjust(struct sock
*sk
)
450 struct tcp_sock
*tp
= tcp_sk(sk
);
454 if (tp
->rcvq_space
.time
== 0)
457 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
458 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
459 tp
->rcv_rtt_est
.rtt
== 0)
462 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
464 space
= max(tp
->rcvq_space
.space
, space
);
466 if (tp
->rcvq_space
.space
!= space
) {
469 tp
->rcvq_space
.space
= space
;
471 if (sysctl_tcp_moderate_rcvbuf
&&
472 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
473 int new_clamp
= space
;
475 /* Receive space grows, normalize in order to
476 * take into account packet headers and sk_buff
477 * structure overhead.
482 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
483 16 + sizeof(struct sk_buff
));
484 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
487 space
= min(space
, sysctl_tcp_rmem
[2]);
488 if (space
> sk
->sk_rcvbuf
) {
489 sk
->sk_rcvbuf
= space
;
491 /* Make the window clamp follow along. */
492 tp
->window_clamp
= new_clamp
;
498 tp
->rcvq_space
.seq
= tp
->copied_seq
;
499 tp
->rcvq_space
.time
= tcp_time_stamp
;
502 /* There is something which you must keep in mind when you analyze the
503 * behavior of the tp->ato delayed ack timeout interval. When a
504 * connection starts up, we want to ack as quickly as possible. The
505 * problem is that "good" TCP's do slow start at the beginning of data
506 * transmission. The means that until we send the first few ACK's the
507 * sender will sit on his end and only queue most of his data, because
508 * he can only send snd_cwnd unacked packets at any given time. For
509 * each ACK we send, he increments snd_cwnd and transmits more of his
512 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
514 struct tcp_sock
*tp
= tcp_sk(sk
);
515 struct inet_connection_sock
*icsk
= inet_csk(sk
);
518 inet_csk_schedule_ack(sk
);
520 tcp_measure_rcv_mss(sk
, skb
);
522 tcp_rcv_rtt_measure(tp
);
524 now
= tcp_time_stamp
;
526 if (!icsk
->icsk_ack
.ato
) {
527 /* The _first_ data packet received, initialize
528 * delayed ACK engine.
530 tcp_incr_quickack(sk
);
531 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
533 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
535 if (m
<= TCP_ATO_MIN
/2) {
536 /* The fastest case is the first. */
537 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
538 } else if (m
< icsk
->icsk_ack
.ato
) {
539 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
540 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
541 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
542 } else if (m
> icsk
->icsk_rto
) {
543 /* Too long gap. Apparently sender failed to
544 * restart window, so that we send ACKs quickly.
546 tcp_incr_quickack(sk
);
547 sk_stream_mem_reclaim(sk
);
550 icsk
->icsk_ack
.lrcvtime
= now
;
552 TCP_ECN_check_ce(tp
, skb
);
555 tcp_grow_window(sk
, skb
);
558 /* Called to compute a smoothed rtt estimate. The data fed to this
559 * routine either comes from timestamps, or from segments that were
560 * known _not_ to have been retransmitted [see Karn/Partridge
561 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
562 * piece by Van Jacobson.
563 * NOTE: the next three routines used to be one big routine.
564 * To save cycles in the RFC 1323 implementation it was better to break
565 * it up into three procedures. -- erics
567 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
569 struct tcp_sock
*tp
= tcp_sk(sk
);
570 long m
= mrtt
; /* RTT */
572 /* The following amusing code comes from Jacobson's
573 * article in SIGCOMM '88. Note that rtt and mdev
574 * are scaled versions of rtt and mean deviation.
575 * This is designed to be as fast as possible
576 * m stands for "measurement".
578 * On a 1990 paper the rto value is changed to:
579 * RTO = rtt + 4 * mdev
581 * Funny. This algorithm seems to be very broken.
582 * These formulae increase RTO, when it should be decreased, increase
583 * too slowly, when it should be increased quickly, decrease too quickly
584 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
585 * does not matter how to _calculate_ it. Seems, it was trap
586 * that VJ failed to avoid. 8)
591 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
592 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
594 m
= -m
; /* m is now abs(error) */
595 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
596 /* This is similar to one of Eifel findings.
597 * Eifel blocks mdev updates when rtt decreases.
598 * This solution is a bit different: we use finer gain
599 * for mdev in this case (alpha*beta).
600 * Like Eifel it also prevents growth of rto,
601 * but also it limits too fast rto decreases,
602 * happening in pure Eifel.
607 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
609 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
610 if (tp
->mdev
> tp
->mdev_max
) {
611 tp
->mdev_max
= tp
->mdev
;
612 if (tp
->mdev_max
> tp
->rttvar
)
613 tp
->rttvar
= tp
->mdev_max
;
615 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
616 if (tp
->mdev_max
< tp
->rttvar
)
617 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
618 tp
->rtt_seq
= tp
->snd_nxt
;
619 tp
->mdev_max
= tcp_rto_min(sk
);
622 /* no previous measure. */
623 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
624 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
625 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
626 tp
->rtt_seq
= tp
->snd_nxt
;
630 /* Calculate rto without backoff. This is the second half of Van Jacobson's
631 * routine referred to above.
633 static inline void tcp_set_rto(struct sock
*sk
)
635 const struct tcp_sock
*tp
= tcp_sk(sk
);
636 /* Old crap is replaced with new one. 8)
639 * 1. If rtt variance happened to be less 50msec, it is hallucination.
640 * It cannot be less due to utterly erratic ACK generation made
641 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
642 * to do with delayed acks, because at cwnd>2 true delack timeout
643 * is invisible. Actually, Linux-2.4 also generates erratic
644 * ACKs in some circumstances.
646 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
648 /* 2. Fixups made earlier cannot be right.
649 * If we do not estimate RTO correctly without them,
650 * all the algo is pure shit and should be replaced
651 * with correct one. It is exactly, which we pretend to do.
655 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
656 * guarantees that rto is higher.
658 static inline void tcp_bound_rto(struct sock
*sk
)
660 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
661 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
664 /* Save metrics learned by this TCP session.
665 This function is called only, when TCP finishes successfully
666 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
668 void tcp_update_metrics(struct sock
*sk
)
670 struct tcp_sock
*tp
= tcp_sk(sk
);
671 struct dst_entry
*dst
= __sk_dst_get(sk
);
673 if (sysctl_tcp_nometrics_save
)
678 if (dst
&& (dst
->flags
&DST_HOST
)) {
679 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
682 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
683 /* This session failed to estimate rtt. Why?
684 * Probably, no packets returned in time.
687 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
688 dst
->metrics
[RTAX_RTT
-1] = 0;
692 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
694 /* If newly calculated rtt larger than stored one,
695 * store new one. Otherwise, use EWMA. Remember,
696 * rtt overestimation is always better than underestimation.
698 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
700 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
702 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
705 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
709 /* Scale deviation to rttvar fixed point */
714 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
715 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
717 dst
->metrics
[RTAX_RTTVAR
-1] -=
718 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
721 if (tp
->snd_ssthresh
>= 0xFFFF) {
722 /* Slow start still did not finish. */
723 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
724 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
725 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
726 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
727 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
728 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
729 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
730 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
731 icsk
->icsk_ca_state
== TCP_CA_Open
) {
732 /* Cong. avoidance phase, cwnd is reliable. */
733 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
734 dst
->metrics
[RTAX_SSTHRESH
-1] =
735 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
736 if (!dst_metric_locked(dst
, RTAX_CWND
))
737 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
739 /* Else slow start did not finish, cwnd is non-sense,
740 ssthresh may be also invalid.
742 if (!dst_metric_locked(dst
, RTAX_CWND
))
743 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
744 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
745 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
746 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
747 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
750 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
751 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
752 tp
->reordering
!= sysctl_tcp_reordering
)
753 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
758 /* Numbers are taken from RFC2414. */
759 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
761 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
764 if (tp
->mss_cache
> 1460)
767 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
769 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
772 /* Set slow start threshold and cwnd not falling to slow start */
773 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
775 struct tcp_sock
*tp
= tcp_sk(sk
);
776 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
778 tp
->prior_ssthresh
= 0;
780 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
783 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
784 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
785 tcp_packets_in_flight(tp
) + 1U);
786 tp
->snd_cwnd_cnt
= 0;
787 tp
->high_seq
= tp
->snd_nxt
;
788 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
789 TCP_ECN_queue_cwr(tp
);
791 tcp_set_ca_state(sk
, TCP_CA_CWR
);
795 /* Initialize metrics on socket. */
797 static void tcp_init_metrics(struct sock
*sk
)
799 struct tcp_sock
*tp
= tcp_sk(sk
);
800 struct dst_entry
*dst
= __sk_dst_get(sk
);
807 if (dst_metric_locked(dst
, RTAX_CWND
))
808 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
809 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
810 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
811 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
812 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
814 if (dst_metric(dst
, RTAX_REORDERING
) &&
815 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
816 tp
->rx_opt
.sack_ok
&= ~2;
817 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
820 if (dst_metric(dst
, RTAX_RTT
) == 0)
823 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
826 /* Initial rtt is determined from SYN,SYN-ACK.
827 * The segment is small and rtt may appear much
828 * less than real one. Use per-dst memory
829 * to make it more realistic.
831 * A bit of theory. RTT is time passed after "normal" sized packet
832 * is sent until it is ACKed. In normal circumstances sending small
833 * packets force peer to delay ACKs and calculation is correct too.
834 * The algorithm is adaptive and, provided we follow specs, it
835 * NEVER underestimate RTT. BUT! If peer tries to make some clever
836 * tricks sort of "quick acks" for time long enough to decrease RTT
837 * to low value, and then abruptly stops to do it and starts to delay
838 * ACKs, wait for troubles.
840 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
841 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
842 tp
->rtt_seq
= tp
->snd_nxt
;
844 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
845 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
846 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
850 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
852 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
853 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
857 /* Play conservative. If timestamps are not
858 * supported, TCP will fail to recalculate correct
859 * rtt, if initial rto is too small. FORGET ALL AND RESET!
861 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
863 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
864 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
868 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
871 struct tcp_sock
*tp
= tcp_sk(sk
);
872 if (metric
> tp
->reordering
) {
873 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
875 /* This exciting event is worth to be remembered. 8) */
877 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
879 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
881 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
883 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
884 #if FASTRETRANS_DEBUG > 1
885 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
886 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
890 tp
->undo_marker
? tp
->undo_retrans
: 0);
892 /* Disable FACK yet. */
893 tp
->rx_opt
.sack_ok
&= ~2;
897 /* This procedure tags the retransmission queue when SACKs arrive.
899 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
900 * Packets in queue with these bits set are counted in variables
901 * sacked_out, retrans_out and lost_out, correspondingly.
903 * Valid combinations are:
904 * Tag InFlight Description
905 * 0 1 - orig segment is in flight.
906 * S 0 - nothing flies, orig reached receiver.
907 * L 0 - nothing flies, orig lost by net.
908 * R 2 - both orig and retransmit are in flight.
909 * L|R 1 - orig is lost, retransmit is in flight.
910 * S|R 1 - orig reached receiver, retrans is still in flight.
911 * (L|S|R is logically valid, it could occur when L|R is sacked,
912 * but it is equivalent to plain S and code short-curcuits it to S.
913 * L|S is logically invalid, it would mean -1 packet in flight 8))
915 * These 6 states form finite state machine, controlled by the following events:
916 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
917 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
918 * 3. Loss detection event of one of three flavors:
919 * A. Scoreboard estimator decided the packet is lost.
920 * A'. Reno "three dupacks" marks head of queue lost.
921 * A''. Its FACK modfication, head until snd.fack is lost.
922 * B. SACK arrives sacking data transmitted after never retransmitted
924 * C. SACK arrives sacking SND.NXT at the moment, when the
925 * segment was retransmitted.
926 * 4. D-SACK added new rule: D-SACK changes any tag to S.
928 * It is pleasant to note, that state diagram turns out to be commutative,
929 * so that we are allowed not to be bothered by order of our actions,
930 * when multiple events arrive simultaneously. (see the function below).
932 * Reordering detection.
933 * --------------------
934 * Reordering metric is maximal distance, which a packet can be displaced
935 * in packet stream. With SACKs we can estimate it:
937 * 1. SACK fills old hole and the corresponding segment was not
938 * ever retransmitted -> reordering. Alas, we cannot use it
939 * when segment was retransmitted.
940 * 2. The last flaw is solved with D-SACK. D-SACK arrives
941 * for retransmitted and already SACKed segment -> reordering..
942 * Both of these heuristics are not used in Loss state, when we cannot
943 * account for retransmits accurately.
946 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
948 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
949 struct tcp_sock
*tp
= tcp_sk(sk
);
950 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
951 TCP_SKB_CB(ack_skb
)->sacked
);
952 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
953 struct sk_buff
*cached_skb
;
954 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
955 int reord
= tp
->packets_out
;
957 u32 lost_retrans
= 0;
959 int found_dup_sack
= 0;
960 int cached_fack_count
;
962 int first_sack_index
;
966 prior_fackets
= tp
->fackets_out
;
968 /* Check for D-SACK. */
969 if (before(ntohl(sp
[0].start_seq
), TCP_SKB_CB(ack_skb
)->ack_seq
)) {
970 flag
|= FLAG_DSACKING_ACK
;
972 tp
->rx_opt
.sack_ok
|= 4;
973 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
974 } else if (num_sacks
> 1 &&
975 !after(ntohl(sp
[0].end_seq
), ntohl(sp
[1].end_seq
)) &&
976 !before(ntohl(sp
[0].start_seq
), ntohl(sp
[1].start_seq
))) {
977 flag
|= FLAG_DSACKING_ACK
;
979 tp
->rx_opt
.sack_ok
|= 4;
980 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
983 /* D-SACK for already forgotten data...
984 * Do dumb counting. */
985 if (found_dup_sack
&&
986 !after(ntohl(sp
[0].end_seq
), prior_snd_una
) &&
987 after(ntohl(sp
[0].end_seq
), tp
->undo_marker
))
990 /* Eliminate too old ACKs, but take into
991 * account more or less fresh ones, they can
992 * contain valid SACK info.
994 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
997 if (!tp
->packets_out
)
1001 * if the only SACK change is the increase of the end_seq of
1002 * the first block then only apply that SACK block
1003 * and use retrans queue hinting otherwise slowpath */
1005 for (i
= 0; i
< num_sacks
; i
++) {
1006 __be32 start_seq
= sp
[i
].start_seq
;
1007 __be32 end_seq
= sp
[i
].end_seq
;
1010 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
1013 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
1014 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1017 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1018 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1020 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1021 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1022 tp
->recv_sack_cache
[i
].start_seq
= 0;
1023 tp
->recv_sack_cache
[i
].end_seq
= 0;
1026 first_sack_index
= 0;
1031 tp
->fastpath_skb_hint
= NULL
;
1033 /* order SACK blocks to allow in order walk of the retrans queue */
1034 for (i
= num_sacks
-1; i
> 0; i
--) {
1035 for (j
= 0; j
< i
; j
++){
1036 if (after(ntohl(sp
[j
].start_seq
),
1037 ntohl(sp
[j
+1].start_seq
))){
1038 struct tcp_sack_block_wire tmp
;
1044 /* Track where the first SACK block goes to */
1045 if (j
== first_sack_index
)
1046 first_sack_index
= j
+1;
1053 /* clear flag as used for different purpose in following code */
1056 /* Use SACK fastpath hint if valid */
1057 cached_skb
= tp
->fastpath_skb_hint
;
1058 cached_fack_count
= tp
->fastpath_cnt_hint
;
1060 cached_skb
= tcp_write_queue_head(sk
);
1061 cached_fack_count
= 0;
1064 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1065 struct sk_buff
*skb
;
1066 __u32 start_seq
= ntohl(sp
->start_seq
);
1067 __u32 end_seq
= ntohl(sp
->end_seq
);
1069 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1072 fack_count
= cached_fack_count
;
1074 /* Event "B" in the comment above. */
1075 if (after(end_seq
, tp
->high_seq
))
1076 flag
|= FLAG_DATA_LOST
;
1078 tcp_for_write_queue_from(skb
, sk
) {
1079 int in_sack
, pcount
;
1082 if (skb
== tcp_send_head(sk
))
1086 cached_fack_count
= fack_count
;
1087 if (i
== first_sack_index
) {
1088 tp
->fastpath_skb_hint
= skb
;
1089 tp
->fastpath_cnt_hint
= fack_count
;
1092 /* The retransmission queue is always in order, so
1093 * we can short-circuit the walk early.
1095 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1098 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1099 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1101 pcount
= tcp_skb_pcount(skb
);
1103 if (pcount
> 1 && !in_sack
&&
1104 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1105 unsigned int pkt_len
;
1107 in_sack
= !after(start_seq
,
1108 TCP_SKB_CB(skb
)->seq
);
1111 pkt_len
= (start_seq
-
1112 TCP_SKB_CB(skb
)->seq
);
1114 pkt_len
= (end_seq
-
1115 TCP_SKB_CB(skb
)->seq
);
1116 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1118 pcount
= tcp_skb_pcount(skb
);
1121 fack_count
+= pcount
;
1123 sacked
= TCP_SKB_CB(skb
)->sacked
;
1125 /* Account D-SACK for retransmitted packet. */
1126 if ((dup_sack
&& in_sack
) &&
1127 (sacked
& TCPCB_RETRANS
) &&
1128 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1131 /* The frame is ACKed. */
1132 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1133 if (sacked
&TCPCB_RETRANS
) {
1134 if ((dup_sack
&& in_sack
) &&
1135 (sacked
&TCPCB_SACKED_ACKED
))
1136 reord
= min(fack_count
, reord
);
1138 /* If it was in a hole, we detected reordering. */
1139 if (fack_count
< prior_fackets
&&
1140 !(sacked
&TCPCB_SACKED_ACKED
))
1141 reord
= min(fack_count
, reord
);
1144 /* Nothing to do; acked frame is about to be dropped. */
1148 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1149 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1150 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1151 lost_retrans
= end_seq
;
1156 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1157 if (sacked
& TCPCB_SACKED_RETRANS
) {
1158 /* If the segment is not tagged as lost,
1159 * we do not clear RETRANS, believing
1160 * that retransmission is still in flight.
1162 if (sacked
& TCPCB_LOST
) {
1163 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1164 tp
->lost_out
-= tcp_skb_pcount(skb
);
1165 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1167 /* clear lost hint */
1168 tp
->retransmit_skb_hint
= NULL
;
1171 /* New sack for not retransmitted frame,
1172 * which was in hole. It is reordering.
1174 if (!(sacked
& TCPCB_RETRANS
) &&
1175 fack_count
< prior_fackets
)
1176 reord
= min(fack_count
, reord
);
1178 if (sacked
& TCPCB_LOST
) {
1179 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1180 tp
->lost_out
-= tcp_skb_pcount(skb
);
1182 /* clear lost hint */
1183 tp
->retransmit_skb_hint
= NULL
;
1185 /* SACK enhanced F-RTO detection.
1186 * Set flag if and only if non-rexmitted
1187 * segments below frto_highmark are
1188 * SACKed (RFC4138; Appendix B).
1189 * Clearing correct due to in-order walk
1191 if (after(end_seq
, tp
->frto_highmark
)) {
1192 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1194 if (!(sacked
& TCPCB_RETRANS
))
1195 flag
|= FLAG_ONLY_ORIG_SACKED
;
1199 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1200 flag
|= FLAG_DATA_SACKED
;
1201 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1203 if (fack_count
> tp
->fackets_out
)
1204 tp
->fackets_out
= fack_count
;
1206 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1207 reord
= min(fack_count
, reord
);
1210 /* D-SACK. We can detect redundant retransmission
1211 * in S|R and plain R frames and clear it.
1212 * undo_retrans is decreased above, L|R frames
1213 * are accounted above as well.
1216 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1217 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1218 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1219 tp
->retransmit_skb_hint
= NULL
;
1224 /* Check for lost retransmit. This superb idea is
1225 * borrowed from "ratehalving". Event "C".
1226 * Later note: FACK people cheated me again 8),
1227 * we have to account for reordering! Ugly,
1230 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1231 struct sk_buff
*skb
;
1233 tcp_for_write_queue(skb
, sk
) {
1234 if (skb
== tcp_send_head(sk
))
1236 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1238 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1240 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1241 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1243 !before(lost_retrans
,
1244 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1246 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1247 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1249 /* clear lost hint */
1250 tp
->retransmit_skb_hint
= NULL
;
1252 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1253 tp
->lost_out
+= tcp_skb_pcount(skb
);
1254 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1255 flag
|= FLAG_DATA_SACKED
;
1256 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1262 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1264 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
&&
1265 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1266 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1270 #if FASTRETRANS_DEBUG > 0
1271 BUG_TRAP((int)tp
->sacked_out
>= 0);
1272 BUG_TRAP((int)tp
->lost_out
>= 0);
1273 BUG_TRAP((int)tp
->retrans_out
>= 0);
1274 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1279 /* F-RTO can only be used if TCP has never retransmitted anything other than
1280 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1282 int tcp_use_frto(struct sock
*sk
)
1284 const struct tcp_sock
*tp
= tcp_sk(sk
);
1285 struct sk_buff
*skb
;
1287 if (!sysctl_tcp_frto
)
1293 /* Avoid expensive walking of rexmit queue if possible */
1294 if (tp
->retrans_out
> 1)
1297 skb
= tcp_write_queue_head(sk
);
1298 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1299 tcp_for_write_queue_from(skb
, sk
) {
1300 if (skb
== tcp_send_head(sk
))
1302 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1304 /* Short-circuit when first non-SACKed skb has been checked */
1305 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1311 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1312 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1313 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1314 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1315 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1316 * bits are handled if the Loss state is really to be entered (in
1317 * tcp_enter_frto_loss).
1319 * Do like tcp_enter_loss() would; when RTO expires the second time it
1321 * "Reduce ssthresh if it has not yet been made inside this window."
1323 void tcp_enter_frto(struct sock
*sk
)
1325 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1326 struct tcp_sock
*tp
= tcp_sk(sk
);
1327 struct sk_buff
*skb
;
1329 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1330 tp
->snd_una
== tp
->high_seq
||
1331 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1332 !icsk
->icsk_retransmits
)) {
1333 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1334 /* Our state is too optimistic in ssthresh() call because cwnd
1335 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1336 * recovery has not yet completed. Pattern would be this: RTO,
1337 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1339 * RFC4138 should be more specific on what to do, even though
1340 * RTO is quite unlikely to occur after the first Cumulative ACK
1341 * due to back-off and complexity of triggering events ...
1343 if (tp
->frto_counter
) {
1345 stored_cwnd
= tp
->snd_cwnd
;
1347 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1348 tp
->snd_cwnd
= stored_cwnd
;
1350 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1352 /* ... in theory, cong.control module could do "any tricks" in
1353 * ssthresh(), which means that ca_state, lost bits and lost_out
1354 * counter would have to be faked before the call occurs. We
1355 * consider that too expensive, unlikely and hacky, so modules
1356 * using these in ssthresh() must deal these incompatibility
1357 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1359 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1362 tp
->undo_marker
= tp
->snd_una
;
1363 tp
->undo_retrans
= 0;
1365 skb
= tcp_write_queue_head(sk
);
1366 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1367 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1368 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1370 tcp_sync_left_out(tp
);
1372 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1373 * The last condition is necessary at least in tp->frto_counter case.
1375 if (IsSackFrto() && (tp
->frto_counter
||
1376 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1377 after(tp
->high_seq
, tp
->snd_una
)) {
1378 tp
->frto_highmark
= tp
->high_seq
;
1380 tp
->frto_highmark
= tp
->snd_nxt
;
1382 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1383 tp
->high_seq
= tp
->snd_nxt
;
1384 tp
->frto_counter
= 1;
1387 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1388 * which indicates that we should follow the traditional RTO recovery,
1389 * i.e. mark everything lost and do go-back-N retransmission.
1391 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1393 struct tcp_sock
*tp
= tcp_sk(sk
);
1394 struct sk_buff
*skb
;
1399 tp
->fackets_out
= 0;
1400 tp
->retrans_out
= 0;
1402 tcp_for_write_queue(skb
, sk
) {
1403 if (skb
== tcp_send_head(sk
))
1405 cnt
+= tcp_skb_pcount(skb
);
1406 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1408 * Count the retransmission made on RTO correctly (only when
1409 * waiting for the first ACK and did not get it)...
1411 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1412 /* For some reason this R-bit might get cleared? */
1413 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1414 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1415 /* ...enter this if branch just for the first segment */
1416 flag
|= FLAG_DATA_ACKED
;
1418 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_SACKED_RETRANS
);
1420 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1422 /* Do not mark those segments lost that were
1423 * forward transmitted after RTO
1425 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1426 tp
->frto_highmark
)) {
1427 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1428 tp
->lost_out
+= tcp_skb_pcount(skb
);
1431 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1432 tp
->fackets_out
= cnt
;
1435 tcp_sync_left_out(tp
);
1437 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1438 tp
->snd_cwnd_cnt
= 0;
1439 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1440 tp
->undo_marker
= 0;
1441 tp
->frto_counter
= 0;
1443 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1444 sysctl_tcp_reordering
);
1445 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1446 tp
->high_seq
= tp
->frto_highmark
;
1447 TCP_ECN_queue_cwr(tp
);
1449 clear_all_retrans_hints(tp
);
1452 void tcp_clear_retrans(struct tcp_sock
*tp
)
1455 tp
->retrans_out
= 0;
1457 tp
->fackets_out
= 0;
1461 tp
->undo_marker
= 0;
1462 tp
->undo_retrans
= 0;
1465 /* Enter Loss state. If "how" is not zero, forget all SACK information
1466 * and reset tags completely, otherwise preserve SACKs. If receiver
1467 * dropped its ofo queue, we will know this due to reneging detection.
1469 void tcp_enter_loss(struct sock
*sk
, int how
)
1471 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1472 struct tcp_sock
*tp
= tcp_sk(sk
);
1473 struct sk_buff
*skb
;
1476 /* Reduce ssthresh if it has not yet been made inside this window. */
1477 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1478 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1479 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1480 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1481 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1484 tp
->snd_cwnd_cnt
= 0;
1485 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1487 tp
->bytes_acked
= 0;
1488 tcp_clear_retrans(tp
);
1490 /* Push undo marker, if it was plain RTO and nothing
1491 * was retransmitted. */
1493 tp
->undo_marker
= tp
->snd_una
;
1495 tcp_for_write_queue(skb
, sk
) {
1496 if (skb
== tcp_send_head(sk
))
1498 cnt
+= tcp_skb_pcount(skb
);
1499 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1500 tp
->undo_marker
= 0;
1501 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1502 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1503 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1504 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1505 tp
->lost_out
+= tcp_skb_pcount(skb
);
1507 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1508 tp
->fackets_out
= cnt
;
1511 tcp_sync_left_out(tp
);
1513 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1514 sysctl_tcp_reordering
);
1515 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1516 tp
->high_seq
= tp
->snd_nxt
;
1517 TCP_ECN_queue_cwr(tp
);
1518 /* Abort FRTO algorithm if one is in progress */
1519 tp
->frto_counter
= 0;
1521 clear_all_retrans_hints(tp
);
1524 static int tcp_check_sack_reneging(struct sock
*sk
)
1526 struct sk_buff
*skb
;
1528 /* If ACK arrived pointing to a remembered SACK,
1529 * it means that our remembered SACKs do not reflect
1530 * real state of receiver i.e.
1531 * receiver _host_ is heavily congested (or buggy).
1532 * Do processing similar to RTO timeout.
1534 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1535 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1536 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1537 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1539 tcp_enter_loss(sk
, 1);
1540 icsk
->icsk_retransmits
++;
1541 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1542 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1543 icsk
->icsk_rto
, TCP_RTO_MAX
);
1549 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1551 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1554 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1556 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1559 static inline int tcp_head_timedout(struct sock
*sk
)
1561 struct tcp_sock
*tp
= tcp_sk(sk
);
1563 return tp
->packets_out
&&
1564 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1567 /* Linux NewReno/SACK/FACK/ECN state machine.
1568 * --------------------------------------
1570 * "Open" Normal state, no dubious events, fast path.
1571 * "Disorder" In all the respects it is "Open",
1572 * but requires a bit more attention. It is entered when
1573 * we see some SACKs or dupacks. It is split of "Open"
1574 * mainly to move some processing from fast path to slow one.
1575 * "CWR" CWND was reduced due to some Congestion Notification event.
1576 * It can be ECN, ICMP source quench, local device congestion.
1577 * "Recovery" CWND was reduced, we are fast-retransmitting.
1578 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1580 * tcp_fastretrans_alert() is entered:
1581 * - each incoming ACK, if state is not "Open"
1582 * - when arrived ACK is unusual, namely:
1587 * Counting packets in flight is pretty simple.
1589 * in_flight = packets_out - left_out + retrans_out
1591 * packets_out is SND.NXT-SND.UNA counted in packets.
1593 * retrans_out is number of retransmitted segments.
1595 * left_out is number of segments left network, but not ACKed yet.
1597 * left_out = sacked_out + lost_out
1599 * sacked_out: Packets, which arrived to receiver out of order
1600 * and hence not ACKed. With SACKs this number is simply
1601 * amount of SACKed data. Even without SACKs
1602 * it is easy to give pretty reliable estimate of this number,
1603 * counting duplicate ACKs.
1605 * lost_out: Packets lost by network. TCP has no explicit
1606 * "loss notification" feedback from network (for now).
1607 * It means that this number can be only _guessed_.
1608 * Actually, it is the heuristics to predict lossage that
1609 * distinguishes different algorithms.
1611 * F.e. after RTO, when all the queue is considered as lost,
1612 * lost_out = packets_out and in_flight = retrans_out.
1614 * Essentially, we have now two algorithms counting
1617 * FACK: It is the simplest heuristics. As soon as we decided
1618 * that something is lost, we decide that _all_ not SACKed
1619 * packets until the most forward SACK are lost. I.e.
1620 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1621 * It is absolutely correct estimate, if network does not reorder
1622 * packets. And it loses any connection to reality when reordering
1623 * takes place. We use FACK by default until reordering
1624 * is suspected on the path to this destination.
1626 * NewReno: when Recovery is entered, we assume that one segment
1627 * is lost (classic Reno). While we are in Recovery and
1628 * a partial ACK arrives, we assume that one more packet
1629 * is lost (NewReno). This heuristics are the same in NewReno
1632 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1633 * deflation etc. CWND is real congestion window, never inflated, changes
1634 * only according to classic VJ rules.
1636 * Really tricky (and requiring careful tuning) part of algorithm
1637 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1638 * The first determines the moment _when_ we should reduce CWND and,
1639 * hence, slow down forward transmission. In fact, it determines the moment
1640 * when we decide that hole is caused by loss, rather than by a reorder.
1642 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1643 * holes, caused by lost packets.
1645 * And the most logically complicated part of algorithm is undo
1646 * heuristics. We detect false retransmits due to both too early
1647 * fast retransmit (reordering) and underestimated RTO, analyzing
1648 * timestamps and D-SACKs. When we detect that some segments were
1649 * retransmitted by mistake and CWND reduction was wrong, we undo
1650 * window reduction and abort recovery phase. This logic is hidden
1651 * inside several functions named tcp_try_undo_<something>.
1654 /* This function decides, when we should leave Disordered state
1655 * and enter Recovery phase, reducing congestion window.
1657 * Main question: may we further continue forward transmission
1658 * with the same cwnd?
1660 static int tcp_time_to_recover(struct sock
*sk
)
1662 struct tcp_sock
*tp
= tcp_sk(sk
);
1665 /* Do not perform any recovery during FRTO algorithm */
1666 if (tp
->frto_counter
)
1669 /* Trick#1: The loss is proven. */
1673 /* Not-A-Trick#2 : Classic rule... */
1674 if (tcp_fackets_out(tp
) > tp
->reordering
)
1677 /* Trick#3 : when we use RFC2988 timer restart, fast
1678 * retransmit can be triggered by timeout of queue head.
1680 if (tcp_head_timedout(sk
))
1683 /* Trick#4: It is still not OK... But will it be useful to delay
1686 packets_out
= tp
->packets_out
;
1687 if (packets_out
<= tp
->reordering
&&
1688 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1689 !tcp_may_send_now(sk
)) {
1690 /* We have nothing to send. This connection is limited
1691 * either by receiver window or by application.
1699 /* If we receive more dupacks than we expected counting segments
1700 * in assumption of absent reordering, interpret this as reordering.
1701 * The only another reason could be bug in receiver TCP.
1703 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1705 struct tcp_sock
*tp
= tcp_sk(sk
);
1708 holes
= max(tp
->lost_out
, 1U);
1709 holes
= min(holes
, tp
->packets_out
);
1711 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1712 tp
->sacked_out
= tp
->packets_out
- holes
;
1713 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1717 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1719 static void tcp_add_reno_sack(struct sock
*sk
)
1721 struct tcp_sock
*tp
= tcp_sk(sk
);
1723 tcp_check_reno_reordering(sk
, 0);
1724 tcp_sync_left_out(tp
);
1727 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1729 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1731 struct tcp_sock
*tp
= tcp_sk(sk
);
1734 /* One ACK acked hole. The rest eat duplicate ACKs. */
1735 if (acked
-1 >= tp
->sacked_out
)
1738 tp
->sacked_out
-= acked
-1;
1740 tcp_check_reno_reordering(sk
, acked
);
1741 tcp_sync_left_out(tp
);
1744 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1747 tp
->left_out
= tp
->lost_out
;
1750 /* Mark head of queue up as lost. */
1751 static void tcp_mark_head_lost(struct sock
*sk
,
1752 int packets
, u32 high_seq
)
1754 struct tcp_sock
*tp
= tcp_sk(sk
);
1755 struct sk_buff
*skb
;
1763 BUG_TRAP(packets
<= tp
->packets_out
);
1764 if (tp
->lost_skb_hint
) {
1765 skb
= tp
->lost_skb_hint
;
1766 cnt
= tp
->lost_cnt_hint
;
1768 skb
= tcp_write_queue_head(sk
);
1772 tcp_for_write_queue_from(skb
, sk
) {
1773 if (skb
== tcp_send_head(sk
))
1775 /* TODO: do this better */
1776 /* this is not the most efficient way to do this... */
1777 tp
->lost_skb_hint
= skb
;
1778 tp
->lost_cnt_hint
= cnt
;
1780 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
1784 if (IsFack(tp
) || IsReno(tp
) ||
1785 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1786 cnt
+= tcp_skb_pcount(skb
);
1788 if (cnt
> packets
) {
1789 if ((tp
->rx_opt
.sack_ok
&& !IsFack(tp
)) ||
1790 (oldcnt
>= packets
))
1793 mss
= skb_shinfo(skb
)->gso_size
;
1794 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
1800 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
1801 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1802 tp
->lost_out
+= tcp_skb_pcount(skb
);
1804 /* clear xmit_retransmit_queue hints
1805 * if this is beyond hint */
1806 if (tp
->retransmit_skb_hint
!= NULL
&&
1807 before(TCP_SKB_CB(skb
)->seq
,
1808 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1809 tp
->retransmit_skb_hint
= NULL
;
1813 tcp_sync_left_out(tp
);
1816 /* Account newly detected lost packet(s) */
1818 static void tcp_update_scoreboard(struct sock
*sk
)
1820 struct tcp_sock
*tp
= tcp_sk(sk
);
1823 int lost
= tp
->fackets_out
- tp
->reordering
;
1826 tcp_mark_head_lost(sk
, lost
, tp
->high_seq
);
1828 tcp_mark_head_lost(sk
, 1, tp
->high_seq
);
1831 /* New heuristics: it is possible only after we switched
1832 * to restart timer each time when something is ACKed.
1833 * Hence, we can detect timed out packets during fast
1834 * retransmit without falling to slow start.
1836 if (!IsReno(tp
) && tcp_head_timedout(sk
)) {
1837 struct sk_buff
*skb
;
1839 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1840 : tcp_write_queue_head(sk
);
1842 tcp_for_write_queue_from(skb
, sk
) {
1843 if (skb
== tcp_send_head(sk
))
1845 if (!tcp_skb_timedout(sk
, skb
))
1848 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
1849 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1850 tp
->lost_out
+= tcp_skb_pcount(skb
);
1852 /* clear xmit_retrans hint */
1853 if (tp
->retransmit_skb_hint
&&
1854 before(TCP_SKB_CB(skb
)->seq
,
1855 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1857 tp
->retransmit_skb_hint
= NULL
;
1861 tp
->scoreboard_skb_hint
= skb
;
1863 tcp_sync_left_out(tp
);
1867 /* CWND moderation, preventing bursts due to too big ACKs
1868 * in dubious situations.
1870 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1872 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1873 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1874 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1877 /* Lower bound on congestion window is slow start threshold
1878 * unless congestion avoidance choice decides to overide it.
1880 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1882 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1884 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1887 /* Decrease cwnd each second ack. */
1888 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
1890 struct tcp_sock
*tp
= tcp_sk(sk
);
1891 int decr
= tp
->snd_cwnd_cnt
+ 1;
1893 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
1894 (IsReno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
1895 tp
->snd_cwnd_cnt
= decr
&1;
1898 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1899 tp
->snd_cwnd
-= decr
;
1901 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1902 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1906 /* Nothing was retransmitted or returned timestamp is less
1907 * than timestamp of the first retransmission.
1909 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1911 return !tp
->retrans_stamp
||
1912 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1913 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1916 /* Undo procedures. */
1918 #if FASTRETRANS_DEBUG > 1
1919 static void DBGUNDO(struct sock
*sk
, const char *msg
)
1921 struct tcp_sock
*tp
= tcp_sk(sk
);
1922 struct inet_sock
*inet
= inet_sk(sk
);
1924 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1926 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1927 tp
->snd_cwnd
, tp
->left_out
,
1928 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1932 #define DBGUNDO(x...) do { } while (0)
1935 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1937 struct tcp_sock
*tp
= tcp_sk(sk
);
1939 if (tp
->prior_ssthresh
) {
1940 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1942 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1943 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1945 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1947 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1948 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1949 TCP_ECN_withdraw_cwr(tp
);
1952 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1954 tcp_moderate_cwnd(tp
);
1955 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1957 /* There is something screwy going on with the retrans hints after
1959 clear_all_retrans_hints(tp
);
1962 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1964 return tp
->undo_marker
&&
1965 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1968 /* People celebrate: "We love our President!" */
1969 static int tcp_try_undo_recovery(struct sock
*sk
)
1971 struct tcp_sock
*tp
= tcp_sk(sk
);
1973 if (tcp_may_undo(tp
)) {
1974 /* Happy end! We did not retransmit anything
1975 * or our original transmission succeeded.
1977 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1978 tcp_undo_cwr(sk
, 1);
1979 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1980 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1982 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1983 tp
->undo_marker
= 0;
1985 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1986 /* Hold old state until something *above* high_seq
1987 * is ACKed. For Reno it is MUST to prevent false
1988 * fast retransmits (RFC2582). SACK TCP is safe. */
1989 tcp_moderate_cwnd(tp
);
1992 tcp_set_ca_state(sk
, TCP_CA_Open
);
1996 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1997 static void tcp_try_undo_dsack(struct sock
*sk
)
1999 struct tcp_sock
*tp
= tcp_sk(sk
);
2001 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2002 DBGUNDO(sk
, "D-SACK");
2003 tcp_undo_cwr(sk
, 1);
2004 tp
->undo_marker
= 0;
2005 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2009 /* Undo during fast recovery after partial ACK. */
2011 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2013 struct tcp_sock
*tp
= tcp_sk(sk
);
2014 /* Partial ACK arrived. Force Hoe's retransmit. */
2015 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
2017 if (tcp_may_undo(tp
)) {
2018 /* Plain luck! Hole if filled with delayed
2019 * packet, rather than with a retransmit.
2021 if (tp
->retrans_out
== 0)
2022 tp
->retrans_stamp
= 0;
2024 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2027 tcp_undo_cwr(sk
, 0);
2028 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2030 /* So... Do not make Hoe's retransmit yet.
2031 * If the first packet was delayed, the rest
2032 * ones are most probably delayed as well.
2039 /* Undo during loss recovery after partial ACK. */
2040 static int tcp_try_undo_loss(struct sock
*sk
)
2042 struct tcp_sock
*tp
= tcp_sk(sk
);
2044 if (tcp_may_undo(tp
)) {
2045 struct sk_buff
*skb
;
2046 tcp_for_write_queue(skb
, sk
) {
2047 if (skb
== tcp_send_head(sk
))
2049 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2052 clear_all_retrans_hints(tp
);
2054 DBGUNDO(sk
, "partial loss");
2056 tp
->left_out
= tp
->sacked_out
;
2057 tcp_undo_cwr(sk
, 1);
2058 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2059 inet_csk(sk
)->icsk_retransmits
= 0;
2060 tp
->undo_marker
= 0;
2062 tcp_set_ca_state(sk
, TCP_CA_Open
);
2068 static inline void tcp_complete_cwr(struct sock
*sk
)
2070 struct tcp_sock
*tp
= tcp_sk(sk
);
2071 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2072 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2073 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2076 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2078 struct tcp_sock
*tp
= tcp_sk(sk
);
2080 tcp_sync_left_out(tp
);
2082 if (tp
->retrans_out
== 0)
2083 tp
->retrans_stamp
= 0;
2086 tcp_enter_cwr(sk
, 1);
2088 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2089 int state
= TCP_CA_Open
;
2091 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
2092 state
= TCP_CA_Disorder
;
2094 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2095 tcp_set_ca_state(sk
, state
);
2096 tp
->high_seq
= tp
->snd_nxt
;
2098 tcp_moderate_cwnd(tp
);
2100 tcp_cwnd_down(sk
, flag
);
2104 static void tcp_mtup_probe_failed(struct sock
*sk
)
2106 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2108 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2109 icsk
->icsk_mtup
.probe_size
= 0;
2112 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2114 struct tcp_sock
*tp
= tcp_sk(sk
);
2115 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2117 /* FIXME: breaks with very large cwnd */
2118 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2119 tp
->snd_cwnd
= tp
->snd_cwnd
*
2120 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2121 icsk
->icsk_mtup
.probe_size
;
2122 tp
->snd_cwnd_cnt
= 0;
2123 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2124 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2126 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2127 icsk
->icsk_mtup
.probe_size
= 0;
2128 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2132 /* Process an event, which can update packets-in-flight not trivially.
2133 * Main goal of this function is to calculate new estimate for left_out,
2134 * taking into account both packets sitting in receiver's buffer and
2135 * packets lost by network.
2137 * Besides that it does CWND reduction, when packet loss is detected
2138 * and changes state of machine.
2140 * It does _not_ decide what to send, it is made in function
2141 * tcp_xmit_retransmit_queue().
2144 tcp_fastretrans_alert(struct sock
*sk
, u32 prior_snd_una
,
2145 int pkts_acked
, int flag
)
2147 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2148 struct tcp_sock
*tp
= tcp_sk(sk
);
2149 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2150 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2151 (tp
->fackets_out
> tp
->reordering
));
2153 /* Some technical things:
2154 * 1. Reno does not count dupacks (sacked_out) automatically. */
2155 if (!tp
->packets_out
)
2157 /* 2. SACK counts snd_fack in packets inaccurately. */
2158 if (tp
->sacked_out
== 0)
2159 tp
->fackets_out
= 0;
2161 /* Now state machine starts.
2162 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2164 tp
->prior_ssthresh
= 0;
2166 /* B. In all the states check for reneging SACKs. */
2167 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2170 /* C. Process data loss notification, provided it is valid. */
2171 if ((flag
&FLAG_DATA_LOST
) &&
2172 before(tp
->snd_una
, tp
->high_seq
) &&
2173 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2174 tp
->fackets_out
> tp
->reordering
) {
2175 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
2176 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2179 /* D. Synchronize left_out to current state. */
2180 tcp_sync_left_out(tp
);
2182 /* E. Check state exit conditions. State can be terminated
2183 * when high_seq is ACKed. */
2184 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2185 BUG_TRAP(tp
->retrans_out
== 0);
2186 tp
->retrans_stamp
= 0;
2187 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2188 switch (icsk
->icsk_ca_state
) {
2190 icsk
->icsk_retransmits
= 0;
2191 if (tcp_try_undo_recovery(sk
))
2196 /* CWR is to be held something *above* high_seq
2197 * is ACKed for CWR bit to reach receiver. */
2198 if (tp
->snd_una
!= tp
->high_seq
) {
2199 tcp_complete_cwr(sk
);
2200 tcp_set_ca_state(sk
, TCP_CA_Open
);
2204 case TCP_CA_Disorder
:
2205 tcp_try_undo_dsack(sk
);
2206 if (!tp
->undo_marker
||
2207 /* For SACK case do not Open to allow to undo
2208 * catching for all duplicate ACKs. */
2209 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2210 tp
->undo_marker
= 0;
2211 tcp_set_ca_state(sk
, TCP_CA_Open
);
2215 case TCP_CA_Recovery
:
2217 tcp_reset_reno_sack(tp
);
2218 if (tcp_try_undo_recovery(sk
))
2220 tcp_complete_cwr(sk
);
2225 /* F. Process state. */
2226 switch (icsk
->icsk_ca_state
) {
2227 case TCP_CA_Recovery
:
2228 if (prior_snd_una
== tp
->snd_una
) {
2229 if (IsReno(tp
) && is_dupack
)
2230 tcp_add_reno_sack(sk
);
2232 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2235 if (flag
&FLAG_DATA_ACKED
)
2236 icsk
->icsk_retransmits
= 0;
2237 if (!tcp_try_undo_loss(sk
)) {
2238 tcp_moderate_cwnd(tp
);
2239 tcp_xmit_retransmit_queue(sk
);
2242 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2244 /* Loss is undone; fall through to processing in Open state. */
2247 if (tp
->snd_una
!= prior_snd_una
)
2248 tcp_reset_reno_sack(tp
);
2250 tcp_add_reno_sack(sk
);
2253 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2254 tcp_try_undo_dsack(sk
);
2256 if (!tcp_time_to_recover(sk
)) {
2257 tcp_try_to_open(sk
, flag
);
2261 /* MTU probe failure: don't reduce cwnd */
2262 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2263 icsk
->icsk_mtup
.probe_size
&&
2264 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2265 tcp_mtup_probe_failed(sk
);
2266 /* Restores the reduction we did in tcp_mtup_probe() */
2268 tcp_simple_retransmit(sk
);
2272 /* Otherwise enter Recovery state */
2275 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2277 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2279 tp
->high_seq
= tp
->snd_nxt
;
2280 tp
->prior_ssthresh
= 0;
2281 tp
->undo_marker
= tp
->snd_una
;
2282 tp
->undo_retrans
= tp
->retrans_out
;
2284 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2285 if (!(flag
&FLAG_ECE
))
2286 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2287 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2288 TCP_ECN_queue_cwr(tp
);
2291 tp
->bytes_acked
= 0;
2292 tp
->snd_cwnd_cnt
= 0;
2293 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2296 if (do_lost
|| tcp_head_timedout(sk
))
2297 tcp_update_scoreboard(sk
);
2298 tcp_cwnd_down(sk
, flag
);
2299 tcp_xmit_retransmit_queue(sk
);
2302 /* Read draft-ietf-tcplw-high-performance before mucking
2303 * with this code. (Supersedes RFC1323)
2305 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2307 /* RTTM Rule: A TSecr value received in a segment is used to
2308 * update the averaged RTT measurement only if the segment
2309 * acknowledges some new data, i.e., only if it advances the
2310 * left edge of the send window.
2312 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2313 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2315 * Changed: reset backoff as soon as we see the first valid sample.
2316 * If we do not, we get strongly overestimated rto. With timestamps
2317 * samples are accepted even from very old segments: f.e., when rtt=1
2318 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2319 * answer arrives rto becomes 120 seconds! If at least one of segments
2320 * in window is lost... Voila. --ANK (010210)
2322 struct tcp_sock
*tp
= tcp_sk(sk
);
2323 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2324 tcp_rtt_estimator(sk
, seq_rtt
);
2326 inet_csk(sk
)->icsk_backoff
= 0;
2330 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2332 /* We don't have a timestamp. Can only use
2333 * packets that are not retransmitted to determine
2334 * rtt estimates. Also, we must not reset the
2335 * backoff for rto until we get a non-retransmitted
2336 * packet. This allows us to deal with a situation
2337 * where the network delay has increased suddenly.
2338 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2341 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2344 tcp_rtt_estimator(sk
, seq_rtt
);
2346 inet_csk(sk
)->icsk_backoff
= 0;
2350 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2353 const struct tcp_sock
*tp
= tcp_sk(sk
);
2354 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2355 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2356 tcp_ack_saw_tstamp(sk
, flag
);
2357 else if (seq_rtt
>= 0)
2358 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2361 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 rtt
,
2362 u32 in_flight
, int good
)
2364 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2365 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, rtt
, in_flight
, good
);
2366 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2369 /* Restart timer after forward progress on connection.
2370 * RFC2988 recommends to restart timer to now+rto.
2373 static void tcp_ack_packets_out(struct sock
*sk
)
2375 struct tcp_sock
*tp
= tcp_sk(sk
);
2377 if (!tp
->packets_out
) {
2378 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2380 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2384 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2385 __u32 now
, __s32
*seq_rtt
)
2387 struct tcp_sock
*tp
= tcp_sk(sk
);
2388 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2389 __u32 seq
= tp
->snd_una
;
2390 __u32 packets_acked
;
2393 /* If we get here, the whole TSO packet has not been
2396 BUG_ON(!after(scb
->end_seq
, seq
));
2398 packets_acked
= tcp_skb_pcount(skb
);
2399 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2401 packets_acked
-= tcp_skb_pcount(skb
);
2403 if (packets_acked
) {
2404 __u8 sacked
= scb
->sacked
;
2406 acked
|= FLAG_DATA_ACKED
;
2408 if (sacked
& TCPCB_RETRANS
) {
2409 if (sacked
& TCPCB_SACKED_RETRANS
)
2410 tp
->retrans_out
-= packets_acked
;
2411 acked
|= FLAG_RETRANS_DATA_ACKED
;
2413 } else if (*seq_rtt
< 0)
2414 *seq_rtt
= now
- scb
->when
;
2415 if (sacked
& TCPCB_SACKED_ACKED
)
2416 tp
->sacked_out
-= packets_acked
;
2417 if (sacked
& TCPCB_LOST
)
2418 tp
->lost_out
-= packets_acked
;
2419 if (sacked
& TCPCB_URG
) {
2421 !before(seq
, tp
->snd_up
))
2424 } else if (*seq_rtt
< 0)
2425 *seq_rtt
= now
- scb
->when
;
2427 if (tp
->fackets_out
) {
2428 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2429 tp
->fackets_out
-= dval
;
2431 /* hint's skb might be NULL but we don't need to care */
2432 tp
->fastpath_cnt_hint
-= min_t(u32
, packets_acked
,
2433 tp
->fastpath_cnt_hint
);
2434 tp
->packets_out
-= packets_acked
;
2436 BUG_ON(tcp_skb_pcount(skb
) == 0);
2437 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2443 /* Remove acknowledged frames from the retransmission queue. */
2444 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2446 struct tcp_sock
*tp
= tcp_sk(sk
);
2447 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2448 struct sk_buff
*skb
;
2449 __u32 now
= tcp_time_stamp
;
2451 int prior_packets
= tp
->packets_out
;
2453 ktime_t last_ackt
= net_invalid_timestamp();
2455 while ((skb
= tcp_write_queue_head(sk
)) &&
2456 skb
!= tcp_send_head(sk
)) {
2457 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2458 __u8 sacked
= scb
->sacked
;
2460 /* If our packet is before the ack sequence we can
2461 * discard it as it's confirmed to have arrived at
2464 if (after(scb
->end_seq
, tp
->snd_una
)) {
2465 if (tcp_skb_pcount(skb
) > 1 &&
2466 after(tp
->snd_una
, scb
->seq
))
2467 acked
|= tcp_tso_acked(sk
, skb
,
2472 /* Initial outgoing SYN's get put onto the write_queue
2473 * just like anything else we transmit. It is not
2474 * true data, and if we misinform our callers that
2475 * this ACK acks real data, we will erroneously exit
2476 * connection startup slow start one packet too
2477 * quickly. This is severely frowned upon behavior.
2479 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2480 acked
|= FLAG_DATA_ACKED
;
2482 acked
|= FLAG_SYN_ACKED
;
2483 tp
->retrans_stamp
= 0;
2486 /* MTU probing checks */
2487 if (icsk
->icsk_mtup
.probe_size
) {
2488 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2489 tcp_mtup_probe_success(sk
, skb
);
2494 if (sacked
& TCPCB_RETRANS
) {
2495 if (sacked
& TCPCB_SACKED_RETRANS
)
2496 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2497 acked
|= FLAG_RETRANS_DATA_ACKED
;
2499 } else if (seq_rtt
< 0) {
2500 seq_rtt
= now
- scb
->when
;
2501 last_ackt
= skb
->tstamp
;
2503 if (sacked
& TCPCB_SACKED_ACKED
)
2504 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2505 if (sacked
& TCPCB_LOST
)
2506 tp
->lost_out
-= tcp_skb_pcount(skb
);
2507 if (sacked
& TCPCB_URG
) {
2509 !before(scb
->end_seq
, tp
->snd_up
))
2512 } else if (seq_rtt
< 0) {
2513 seq_rtt
= now
- scb
->when
;
2514 last_ackt
= skb
->tstamp
;
2516 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2517 tcp_packets_out_dec(tp
, skb
);
2518 tcp_unlink_write_queue(skb
, sk
);
2519 sk_stream_free_skb(sk
, skb
);
2520 clear_all_retrans_hints(tp
);
2523 if (acked
&FLAG_ACKED
) {
2524 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2525 const struct tcp_congestion_ops
*ca_ops
2526 = inet_csk(sk
)->icsk_ca_ops
;
2528 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2529 tcp_ack_packets_out(sk
);
2532 tcp_remove_reno_sacks(sk
, pkts_acked
);
2534 /* Is the ACK triggering packet unambiguous? */
2535 if (acked
& FLAG_RETRANS_DATA_ACKED
)
2536 last_ackt
= net_invalid_timestamp();
2538 if (ca_ops
->pkts_acked
)
2539 ca_ops
->pkts_acked(sk
, pkts_acked
, last_ackt
);
2542 #if FASTRETRANS_DEBUG > 0
2543 BUG_TRAP((int)tp
->sacked_out
>= 0);
2544 BUG_TRAP((int)tp
->lost_out
>= 0);
2545 BUG_TRAP((int)tp
->retrans_out
>= 0);
2546 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2547 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2549 printk(KERN_DEBUG
"Leak l=%u %d\n",
2550 tp
->lost_out
, icsk
->icsk_ca_state
);
2553 if (tp
->sacked_out
) {
2554 printk(KERN_DEBUG
"Leak s=%u %d\n",
2555 tp
->sacked_out
, icsk
->icsk_ca_state
);
2558 if (tp
->retrans_out
) {
2559 printk(KERN_DEBUG
"Leak r=%u %d\n",
2560 tp
->retrans_out
, icsk
->icsk_ca_state
);
2561 tp
->retrans_out
= 0;
2565 *seq_rtt_p
= seq_rtt
;
2569 static void tcp_ack_probe(struct sock
*sk
)
2571 const struct tcp_sock
*tp
= tcp_sk(sk
);
2572 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2574 /* Was it a usable window open? */
2576 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2577 tp
->snd_una
+ tp
->snd_wnd
)) {
2578 icsk
->icsk_backoff
= 0;
2579 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2580 /* Socket must be waked up by subsequent tcp_data_snd_check().
2581 * This function is not for random using!
2584 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2585 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2590 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2592 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2593 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2596 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2598 const struct tcp_sock
*tp
= tcp_sk(sk
);
2599 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2600 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2603 /* Check that window update is acceptable.
2604 * The function assumes that snd_una<=ack<=snd_next.
2606 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2607 const u32 ack_seq
, const u32 nwin
)
2609 return (after(ack
, tp
->snd_una
) ||
2610 after(ack_seq
, tp
->snd_wl1
) ||
2611 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2614 /* Update our send window.
2616 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2617 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2619 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2622 struct tcp_sock
*tp
= tcp_sk(sk
);
2624 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2626 if (likely(!tcp_hdr(skb
)->syn
))
2627 nwin
<<= tp
->rx_opt
.snd_wscale
;
2629 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2630 flag
|= FLAG_WIN_UPDATE
;
2631 tcp_update_wl(tp
, ack
, ack_seq
);
2633 if (tp
->snd_wnd
!= nwin
) {
2636 /* Note, it is the only place, where
2637 * fast path is recovered for sending TCP.
2640 tcp_fast_path_check(sk
);
2642 if (nwin
> tp
->max_window
) {
2643 tp
->max_window
= nwin
;
2644 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2654 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2655 * continue in congestion avoidance.
2657 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2659 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2660 tp
->snd_cwnd_cnt
= 0;
2661 TCP_ECN_queue_cwr(tp
);
2662 tcp_moderate_cwnd(tp
);
2665 /* A conservative spurious RTO response algorithm: reduce cwnd using
2666 * rate halving and continue in congestion avoidance.
2668 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
2670 tcp_enter_cwr(sk
, 0);
2673 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
2676 tcp_ratehalving_spur_to_response(sk
);
2678 tcp_undo_cwr(sk
, 1);
2681 /* F-RTO spurious RTO detection algorithm (RFC4138)
2683 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2684 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2685 * window (but not to or beyond highest sequence sent before RTO):
2686 * On First ACK, send two new segments out.
2687 * On Second ACK, RTO was likely spurious. Do spurious response (response
2688 * algorithm is not part of the F-RTO detection algorithm
2689 * given in RFC4138 but can be selected separately).
2690 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2691 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2692 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2693 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2695 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2696 * original window even after we transmit two new data segments.
2699 * on first step, wait until first cumulative ACK arrives, then move to
2700 * the second step. In second step, the next ACK decides.
2702 * F-RTO is implemented (mainly) in four functions:
2703 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2704 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2705 * called when tcp_use_frto() showed green light
2706 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2707 * - tcp_enter_frto_loss() is called if there is not enough evidence
2708 * to prove that the RTO is indeed spurious. It transfers the control
2709 * from F-RTO to the conventional RTO recovery
2711 static int tcp_process_frto(struct sock
*sk
, u32 prior_snd_una
, int flag
)
2713 struct tcp_sock
*tp
= tcp_sk(sk
);
2715 tcp_sync_left_out(tp
);
2717 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2718 if (flag
&FLAG_DATA_ACKED
)
2719 inet_csk(sk
)->icsk_retransmits
= 0;
2721 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2722 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
2726 if (!IsSackFrto() || IsReno(tp
)) {
2727 /* RFC4138 shortcoming in step 2; should also have case c):
2728 * ACK isn't duplicate nor advances window, e.g., opposite dir
2731 if ((tp
->snd_una
== prior_snd_una
) && (flag
&FLAG_NOT_DUP
) &&
2732 !(flag
&FLAG_FORWARD_PROGRESS
))
2735 if (!(flag
&FLAG_DATA_ACKED
)) {
2736 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2741 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2742 /* Prevent sending of new data. */
2743 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2744 tcp_packets_in_flight(tp
));
2748 if ((tp
->frto_counter
>= 2) &&
2749 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2750 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2751 /* RFC4138 shortcoming (see comment above) */
2752 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2755 tcp_enter_frto_loss(sk
, 3, flag
);
2760 if (tp
->frto_counter
== 1) {
2761 /* Sending of the next skb must be allowed or no FRTO */
2762 if (!tcp_send_head(sk
) ||
2763 after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2764 tp
->snd_una
+ tp
->snd_wnd
)) {
2765 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3),
2770 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2771 tp
->frto_counter
= 2;
2774 switch (sysctl_tcp_frto_response
) {
2776 tcp_undo_spur_to_response(sk
, flag
);
2779 tcp_conservative_spur_to_response(tp
);
2782 tcp_ratehalving_spur_to_response(sk
);
2785 tp
->frto_counter
= 0;
2790 /* This routine deals with incoming acks, but not outgoing ones. */
2791 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2793 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2794 struct tcp_sock
*tp
= tcp_sk(sk
);
2795 u32 prior_snd_una
= tp
->snd_una
;
2796 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2797 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2798 u32 prior_in_flight
;
2803 /* If the ack is newer than sent or older than previous acks
2804 * then we can probably ignore it.
2806 if (after(ack
, tp
->snd_nxt
))
2807 goto uninteresting_ack
;
2809 if (before(ack
, prior_snd_una
))
2812 if (after(ack
, prior_snd_una
))
2813 flag
|= FLAG_SND_UNA_ADVANCED
;
2815 if (sysctl_tcp_abc
) {
2816 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2817 tp
->bytes_acked
+= ack
- prior_snd_una
;
2818 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2819 /* we assume just one segment left network */
2820 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2823 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2824 /* Window is constant, pure forward advance.
2825 * No more checks are required.
2826 * Note, we use the fact that SND.UNA>=SND.WL2.
2828 tcp_update_wl(tp
, ack
, ack_seq
);
2830 flag
|= FLAG_WIN_UPDATE
;
2832 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2834 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2836 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2839 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2841 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
2843 if (TCP_SKB_CB(skb
)->sacked
)
2844 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2846 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
2849 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2852 /* We passed data and got it acked, remove any soft error
2853 * log. Something worked...
2855 sk
->sk_err_soft
= 0;
2856 tp
->rcv_tstamp
= tcp_time_stamp
;
2857 prior_packets
= tp
->packets_out
;
2861 prior_in_flight
= tcp_packets_in_flight(tp
);
2863 /* See if we can take anything off of the retransmit queue. */
2864 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2866 if (tp
->frto_counter
)
2867 frto_cwnd
= tcp_process_frto(sk
, prior_snd_una
, flag
);
2869 if (tcp_ack_is_dubious(sk
, flag
)) {
2870 /* Advance CWND, if state allows this. */
2871 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
2872 tcp_may_raise_cwnd(sk
, flag
))
2873 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 0);
2874 tcp_fastretrans_alert(sk
, prior_snd_una
, prior_packets
- tp
->packets_out
, flag
);
2876 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
2877 tcp_cong_avoid(sk
, ack
, seq_rtt
, prior_in_flight
, 1);
2880 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2881 dst_confirm(sk
->sk_dst_cache
);
2886 icsk
->icsk_probes_out
= 0;
2888 /* If this ack opens up a zero window, clear backoff. It was
2889 * being used to time the probes, and is probably far higher than
2890 * it needs to be for normal retransmission.
2892 if (tcp_send_head(sk
))
2897 if (TCP_SKB_CB(skb
)->sacked
)
2898 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2901 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2906 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2907 * But, this can also be called on packets in the established flow when
2908 * the fast version below fails.
2910 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2913 struct tcphdr
*th
= tcp_hdr(skb
);
2914 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2916 ptr
= (unsigned char *)(th
+ 1);
2917 opt_rx
->saw_tstamp
= 0;
2919 while (length
> 0) {
2926 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2931 if (opsize
< 2) /* "silly options" */
2933 if (opsize
> length
)
2934 return; /* don't parse partial options */
2937 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2938 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
2940 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2941 in_mss
= opt_rx
->user_mss
;
2942 opt_rx
->mss_clamp
= in_mss
;
2947 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2948 if (sysctl_tcp_window_scaling
) {
2949 __u8 snd_wscale
= *(__u8
*) ptr
;
2950 opt_rx
->wscale_ok
= 1;
2951 if (snd_wscale
> 14) {
2952 if (net_ratelimit())
2953 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2954 "scaling value %d >14 received.\n",
2958 opt_rx
->snd_wscale
= snd_wscale
;
2961 case TCPOPT_TIMESTAMP
:
2962 if (opsize
==TCPOLEN_TIMESTAMP
) {
2963 if ((estab
&& opt_rx
->tstamp_ok
) ||
2964 (!estab
&& sysctl_tcp_timestamps
)) {
2965 opt_rx
->saw_tstamp
= 1;
2966 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
2967 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
2971 case TCPOPT_SACK_PERM
:
2972 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2973 if (sysctl_tcp_sack
) {
2974 opt_rx
->sack_ok
= 1;
2975 tcp_sack_reset(opt_rx
);
2981 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2982 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2984 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2987 #ifdef CONFIG_TCP_MD5SIG
2990 * The MD5 Hash has already been
2991 * checked (see tcp_v{4,6}_do_rcv()).
3003 /* Fast parse options. This hopes to only see timestamps.
3004 * If it is wrong it falls back on tcp_parse_options().
3006 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3007 struct tcp_sock
*tp
)
3009 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3010 tp
->rx_opt
.saw_tstamp
= 0;
3012 } else if (tp
->rx_opt
.tstamp_ok
&&
3013 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3014 __be32
*ptr
= (__be32
*)(th
+ 1);
3015 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3016 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3017 tp
->rx_opt
.saw_tstamp
= 1;
3019 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3021 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3025 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3029 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3031 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3032 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3035 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3037 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3038 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3039 * extra check below makes sure this can only happen
3040 * for pure ACK frames. -DaveM
3042 * Not only, also it occurs for expired timestamps.
3045 if (tcp_paws_check(&tp
->rx_opt
, 0))
3046 tcp_store_ts_recent(tp
);
3050 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3052 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3053 * it can pass through stack. So, the following predicate verifies that
3054 * this segment is not used for anything but congestion avoidance or
3055 * fast retransmit. Moreover, we even are able to eliminate most of such
3056 * second order effects, if we apply some small "replay" window (~RTO)
3057 * to timestamp space.
3059 * All these measures still do not guarantee that we reject wrapped ACKs
3060 * on networks with high bandwidth, when sequence space is recycled fastly,
3061 * but it guarantees that such events will be very rare and do not affect
3062 * connection seriously. This doesn't look nice, but alas, PAWS is really
3065 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3066 * states that events when retransmit arrives after original data are rare.
3067 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3068 * the biggest problem on large power networks even with minor reordering.
3069 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3070 * up to bandwidth of 18Gigabit/sec. 8) ]
3073 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3075 struct tcp_sock
*tp
= tcp_sk(sk
);
3076 struct tcphdr
*th
= tcp_hdr(skb
);
3077 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3078 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3080 return (/* 1. Pure ACK with correct sequence number. */
3081 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3083 /* 2. ... and duplicate ACK. */
3084 ack
== tp
->snd_una
&&
3086 /* 3. ... and does not update window. */
3087 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3089 /* 4. ... and sits in replay window. */
3090 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3093 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3095 const struct tcp_sock
*tp
= tcp_sk(sk
);
3097 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3098 !tcp_disordered_ack(sk
, skb
);
3101 /* Check segment sequence number for validity.
3103 * Segment controls are considered valid, if the segment
3104 * fits to the window after truncation to the window. Acceptability
3105 * of data (and SYN, FIN, of course) is checked separately.
3106 * See tcp_data_queue(), for example.
3108 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3109 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3110 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3111 * (borrowed from freebsd)
3114 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3116 return !before(end_seq
, tp
->rcv_wup
) &&
3117 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3120 /* When we get a reset we do this. */
3121 static void tcp_reset(struct sock
*sk
)
3123 /* We want the right error as BSD sees it (and indeed as we do). */
3124 switch (sk
->sk_state
) {
3126 sk
->sk_err
= ECONNREFUSED
;
3128 case TCP_CLOSE_WAIT
:
3134 sk
->sk_err
= ECONNRESET
;
3137 if (!sock_flag(sk
, SOCK_DEAD
))
3138 sk
->sk_error_report(sk
);
3144 * Process the FIN bit. This now behaves as it is supposed to work
3145 * and the FIN takes effect when it is validly part of sequence
3146 * space. Not before when we get holes.
3148 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3149 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3152 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3153 * close and we go into CLOSING (and later onto TIME-WAIT)
3155 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3157 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3159 struct tcp_sock
*tp
= tcp_sk(sk
);
3161 inet_csk_schedule_ack(sk
);
3163 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3164 sock_set_flag(sk
, SOCK_DONE
);
3166 switch (sk
->sk_state
) {
3168 case TCP_ESTABLISHED
:
3169 /* Move to CLOSE_WAIT */
3170 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3171 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3174 case TCP_CLOSE_WAIT
:
3176 /* Received a retransmission of the FIN, do
3181 /* RFC793: Remain in the LAST-ACK state. */
3185 /* This case occurs when a simultaneous close
3186 * happens, we must ack the received FIN and
3187 * enter the CLOSING state.
3190 tcp_set_state(sk
, TCP_CLOSING
);
3193 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3195 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3198 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3199 * cases we should never reach this piece of code.
3201 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3202 __FUNCTION__
, sk
->sk_state
);
3206 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3207 * Probably, we should reset in this case. For now drop them.
3209 __skb_queue_purge(&tp
->out_of_order_queue
);
3210 if (tp
->rx_opt
.sack_ok
)
3211 tcp_sack_reset(&tp
->rx_opt
);
3212 sk_stream_mem_reclaim(sk
);
3214 if (!sock_flag(sk
, SOCK_DEAD
)) {
3215 sk
->sk_state_change(sk
);
3217 /* Do not send POLL_HUP for half duplex close. */
3218 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3219 sk
->sk_state
== TCP_CLOSE
)
3220 sk_wake_async(sk
, 1, POLL_HUP
);
3222 sk_wake_async(sk
, 1, POLL_IN
);
3226 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3228 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3229 if (before(seq
, sp
->start_seq
))
3230 sp
->start_seq
= seq
;
3231 if (after(end_seq
, sp
->end_seq
))
3232 sp
->end_seq
= end_seq
;
3238 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3240 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3241 if (before(seq
, tp
->rcv_nxt
))
3242 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3244 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3246 tp
->rx_opt
.dsack
= 1;
3247 tp
->duplicate_sack
[0].start_seq
= seq
;
3248 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3249 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3253 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3255 if (!tp
->rx_opt
.dsack
)
3256 tcp_dsack_set(tp
, seq
, end_seq
);
3258 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3261 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3263 struct tcp_sock
*tp
= tcp_sk(sk
);
3265 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3266 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3267 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3268 tcp_enter_quickack_mode(sk
);
3270 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3271 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3273 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3274 end_seq
= tp
->rcv_nxt
;
3275 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3282 /* These routines update the SACK block as out-of-order packets arrive or
3283 * in-order packets close up the sequence space.
3285 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3288 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3289 struct tcp_sack_block
*swalk
= sp
+1;
3291 /* See if the recent change to the first SACK eats into
3292 * or hits the sequence space of other SACK blocks, if so coalesce.
3294 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3295 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3298 /* Zap SWALK, by moving every further SACK up by one slot.
3299 * Decrease num_sacks.
3301 tp
->rx_opt
.num_sacks
--;
3302 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3303 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3307 this_sack
++, swalk
++;
3311 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3315 tmp
= sack1
->start_seq
;
3316 sack1
->start_seq
= sack2
->start_seq
;
3317 sack2
->start_seq
= tmp
;
3319 tmp
= sack1
->end_seq
;
3320 sack1
->end_seq
= sack2
->end_seq
;
3321 sack2
->end_seq
= tmp
;
3324 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3326 struct tcp_sock
*tp
= tcp_sk(sk
);
3327 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3328 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3334 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3335 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3336 /* Rotate this_sack to the first one. */
3337 for (; this_sack
>0; this_sack
--, sp
--)
3338 tcp_sack_swap(sp
, sp
-1);
3340 tcp_sack_maybe_coalesce(tp
);
3345 /* Could not find an adjacent existing SACK, build a new one,
3346 * put it at the front, and shift everyone else down. We
3347 * always know there is at least one SACK present already here.
3349 * If the sack array is full, forget about the last one.
3351 if (this_sack
>= 4) {
3353 tp
->rx_opt
.num_sacks
--;
3356 for (; this_sack
> 0; this_sack
--, sp
--)
3360 /* Build the new head SACK, and we're done. */
3361 sp
->start_seq
= seq
;
3362 sp
->end_seq
= end_seq
;
3363 tp
->rx_opt
.num_sacks
++;
3364 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3367 /* RCV.NXT advances, some SACKs should be eaten. */
3369 static void tcp_sack_remove(struct tcp_sock
*tp
)
3371 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3372 int num_sacks
= tp
->rx_opt
.num_sacks
;
3375 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3376 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3377 tp
->rx_opt
.num_sacks
= 0;
3378 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3382 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3383 /* Check if the start of the sack is covered by RCV.NXT. */
3384 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3387 /* RCV.NXT must cover all the block! */
3388 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3390 /* Zap this SACK, by moving forward any other SACKS. */
3391 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3392 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3399 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3400 tp
->rx_opt
.num_sacks
= num_sacks
;
3401 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3405 /* This one checks to see if we can put data from the
3406 * out_of_order queue into the receive_queue.
3408 static void tcp_ofo_queue(struct sock
*sk
)
3410 struct tcp_sock
*tp
= tcp_sk(sk
);
3411 __u32 dsack_high
= tp
->rcv_nxt
;
3412 struct sk_buff
*skb
;
3414 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3415 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3418 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3419 __u32 dsack
= dsack_high
;
3420 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3421 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3422 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3425 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3426 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3427 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3431 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3432 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3433 TCP_SKB_CB(skb
)->end_seq
);
3435 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3436 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3437 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3438 if (tcp_hdr(skb
)->fin
)
3439 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3443 static int tcp_prune_queue(struct sock
*sk
);
3445 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3447 struct tcphdr
*th
= tcp_hdr(skb
);
3448 struct tcp_sock
*tp
= tcp_sk(sk
);
3451 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3454 __skb_pull(skb
, th
->doff
*4);
3456 TCP_ECN_accept_cwr(tp
, skb
);
3458 if (tp
->rx_opt
.dsack
) {
3459 tp
->rx_opt
.dsack
= 0;
3460 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3461 4 - tp
->rx_opt
.tstamp_ok
);
3464 /* Queue data for delivery to the user.
3465 * Packets in sequence go to the receive queue.
3466 * Out of sequence packets to the out_of_order_queue.
3468 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3469 if (tcp_receive_window(tp
) == 0)
3472 /* Ok. In sequence. In window. */
3473 if (tp
->ucopy
.task
== current
&&
3474 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3475 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3476 int chunk
= min_t(unsigned int, skb
->len
,
3479 __set_current_state(TASK_RUNNING
);
3482 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3483 tp
->ucopy
.len
-= chunk
;
3484 tp
->copied_seq
+= chunk
;
3485 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3486 tcp_rcv_space_adjust(sk
);
3494 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3495 !sk_stream_rmem_schedule(sk
, skb
))) {
3496 if (tcp_prune_queue(sk
) < 0 ||
3497 !sk_stream_rmem_schedule(sk
, skb
))
3500 sk_stream_set_owner_r(skb
, sk
);
3501 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3503 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3505 tcp_event_data_recv(sk
, skb
);
3507 tcp_fin(skb
, sk
, th
);
3509 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3512 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3513 * gap in queue is filled.
3515 if (skb_queue_empty(&tp
->out_of_order_queue
))
3516 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3519 if (tp
->rx_opt
.num_sacks
)
3520 tcp_sack_remove(tp
);
3522 tcp_fast_path_check(sk
);
3526 else if (!sock_flag(sk
, SOCK_DEAD
))
3527 sk
->sk_data_ready(sk
, 0);
3531 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3532 /* A retransmit, 2nd most common case. Force an immediate ack. */
3533 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3534 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3537 tcp_enter_quickack_mode(sk
);
3538 inet_csk_schedule_ack(sk
);
3544 /* Out of window. F.e. zero window probe. */
3545 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3548 tcp_enter_quickack_mode(sk
);
3550 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3551 /* Partial packet, seq < rcv_next < end_seq */
3552 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3553 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3554 TCP_SKB_CB(skb
)->end_seq
);
3556 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3558 /* If window is closed, drop tail of packet. But after
3559 * remembering D-SACK for its head made in previous line.
3561 if (!tcp_receive_window(tp
))
3566 TCP_ECN_check_ce(tp
, skb
);
3568 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3569 !sk_stream_rmem_schedule(sk
, skb
)) {
3570 if (tcp_prune_queue(sk
) < 0 ||
3571 !sk_stream_rmem_schedule(sk
, skb
))
3575 /* Disable header prediction. */
3577 inet_csk_schedule_ack(sk
);
3579 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3580 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3582 sk_stream_set_owner_r(skb
, sk
);
3584 if (!skb_peek(&tp
->out_of_order_queue
)) {
3585 /* Initial out of order segment, build 1 SACK. */
3586 if (tp
->rx_opt
.sack_ok
) {
3587 tp
->rx_opt
.num_sacks
= 1;
3588 tp
->rx_opt
.dsack
= 0;
3589 tp
->rx_opt
.eff_sacks
= 1;
3590 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3591 tp
->selective_acks
[0].end_seq
=
3592 TCP_SKB_CB(skb
)->end_seq
;
3594 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3596 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3597 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3598 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3600 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3601 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3603 if (!tp
->rx_opt
.num_sacks
||
3604 tp
->selective_acks
[0].end_seq
!= seq
)
3607 /* Common case: data arrive in order after hole. */
3608 tp
->selective_acks
[0].end_seq
= end_seq
;
3612 /* Find place to insert this segment. */
3614 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3616 } while ((skb1
= skb1
->prev
) !=
3617 (struct sk_buff
*)&tp
->out_of_order_queue
);
3619 /* Do skb overlap to previous one? */
3620 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3621 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3622 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3623 /* All the bits are present. Drop. */
3625 tcp_dsack_set(tp
, seq
, end_seq
);
3628 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3629 /* Partial overlap. */
3630 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3635 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3637 /* And clean segments covered by new one as whole. */
3638 while ((skb1
= skb
->next
) !=
3639 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3640 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3641 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3642 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3645 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3646 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3651 if (tp
->rx_opt
.sack_ok
)
3652 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3656 /* Collapse contiguous sequence of skbs head..tail with
3657 * sequence numbers start..end.
3658 * Segments with FIN/SYN are not collapsed (only because this
3662 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3663 struct sk_buff
*head
, struct sk_buff
*tail
,
3666 struct sk_buff
*skb
;
3668 /* First, check that queue is collapsible and find
3669 * the point where collapsing can be useful. */
3670 for (skb
= head
; skb
!= tail
; ) {
3671 /* No new bits? It is possible on ofo queue. */
3672 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3673 struct sk_buff
*next
= skb
->next
;
3674 __skb_unlink(skb
, list
);
3676 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3681 /* The first skb to collapse is:
3683 * - bloated or contains data before "start" or
3684 * overlaps to the next one.
3686 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
3687 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3688 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3689 (skb
->next
!= tail
&&
3690 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3693 /* Decided to skip this, advance start seq. */
3694 start
= TCP_SKB_CB(skb
)->end_seq
;
3697 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
3700 while (before(start
, end
)) {
3701 struct sk_buff
*nskb
;
3702 int header
= skb_headroom(skb
);
3703 int copy
= SKB_MAX_ORDER(header
, 0);
3705 /* Too big header? This can happen with IPv6. */
3708 if (end
-start
< copy
)
3710 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3714 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
3715 skb_set_network_header(nskb
, (skb_network_header(skb
) -
3717 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
3719 skb_reserve(nskb
, header
);
3720 memcpy(nskb
->head
, skb
->head
, header
);
3721 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3722 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3723 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3724 sk_stream_set_owner_r(nskb
, sk
);
3726 /* Copy data, releasing collapsed skbs. */
3728 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3729 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3733 size
= min(copy
, size
);
3734 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3736 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3740 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3741 struct sk_buff
*next
= skb
->next
;
3742 __skb_unlink(skb
, list
);
3744 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3747 tcp_hdr(skb
)->syn
||
3755 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3756 * and tcp_collapse() them until all the queue is collapsed.
3758 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3760 struct tcp_sock
*tp
= tcp_sk(sk
);
3761 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3762 struct sk_buff
*head
;
3768 start
= TCP_SKB_CB(skb
)->seq
;
3769 end
= TCP_SKB_CB(skb
)->end_seq
;
3775 /* Segment is terminated when we see gap or when
3776 * we are at the end of all the queue. */
3777 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3778 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3779 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3780 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3781 head
, skb
, start
, end
);
3783 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3785 /* Start new segment */
3786 start
= TCP_SKB_CB(skb
)->seq
;
3787 end
= TCP_SKB_CB(skb
)->end_seq
;
3789 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3790 start
= TCP_SKB_CB(skb
)->seq
;
3791 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3792 end
= TCP_SKB_CB(skb
)->end_seq
;
3797 /* Reduce allocated memory if we can, trying to get
3798 * the socket within its memory limits again.
3800 * Return less than zero if we should start dropping frames
3801 * until the socket owning process reads some of the data
3802 * to stabilize the situation.
3804 static int tcp_prune_queue(struct sock
*sk
)
3806 struct tcp_sock
*tp
= tcp_sk(sk
);
3808 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3810 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3812 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3813 tcp_clamp_window(sk
);
3814 else if (tcp_memory_pressure
)
3815 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3817 tcp_collapse_ofo_queue(sk
);
3818 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3819 sk
->sk_receive_queue
.next
,
3820 (struct sk_buff
*)&sk
->sk_receive_queue
,
3821 tp
->copied_seq
, tp
->rcv_nxt
);
3822 sk_stream_mem_reclaim(sk
);
3824 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3827 /* Collapsing did not help, destructive actions follow.
3828 * This must not ever occur. */
3830 /* First, purge the out_of_order queue. */
3831 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3832 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3833 __skb_queue_purge(&tp
->out_of_order_queue
);
3835 /* Reset SACK state. A conforming SACK implementation will
3836 * do the same at a timeout based retransmit. When a connection
3837 * is in a sad state like this, we care only about integrity
3838 * of the connection not performance.
3840 if (tp
->rx_opt
.sack_ok
)
3841 tcp_sack_reset(&tp
->rx_opt
);
3842 sk_stream_mem_reclaim(sk
);
3845 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3848 /* If we are really being abused, tell the caller to silently
3849 * drop receive data on the floor. It will get retransmitted
3850 * and hopefully then we'll have sufficient space.
3852 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3854 /* Massive buffer overcommit. */
3860 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3861 * As additional protections, we do not touch cwnd in retransmission phases,
3862 * and if application hit its sndbuf limit recently.
3864 void tcp_cwnd_application_limited(struct sock
*sk
)
3866 struct tcp_sock
*tp
= tcp_sk(sk
);
3868 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3869 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3870 /* Limited by application or receiver window. */
3871 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
3872 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
3873 if (win_used
< tp
->snd_cwnd
) {
3874 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3875 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3877 tp
->snd_cwnd_used
= 0;
3879 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3882 static int tcp_should_expand_sndbuf(struct sock
*sk
)
3884 struct tcp_sock
*tp
= tcp_sk(sk
);
3886 /* If the user specified a specific send buffer setting, do
3889 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3892 /* If we are under global TCP memory pressure, do not expand. */
3893 if (tcp_memory_pressure
)
3896 /* If we are under soft global TCP memory pressure, do not expand. */
3897 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3900 /* If we filled the congestion window, do not expand. */
3901 if (tp
->packets_out
>= tp
->snd_cwnd
)
3907 /* When incoming ACK allowed to free some skb from write_queue,
3908 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3909 * on the exit from tcp input handler.
3911 * PROBLEM: sndbuf expansion does not work well with largesend.
3913 static void tcp_new_space(struct sock
*sk
)
3915 struct tcp_sock
*tp
= tcp_sk(sk
);
3917 if (tcp_should_expand_sndbuf(sk
)) {
3918 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3919 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3920 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3921 tp
->reordering
+ 1);
3922 sndmem
*= 2*demanded
;
3923 if (sndmem
> sk
->sk_sndbuf
)
3924 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3925 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3928 sk
->sk_write_space(sk
);
3931 static void tcp_check_space(struct sock
*sk
)
3933 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3934 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3935 if (sk
->sk_socket
&&
3936 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3941 static inline void tcp_data_snd_check(struct sock
*sk
)
3943 tcp_push_pending_frames(sk
);
3944 tcp_check_space(sk
);
3948 * Check if sending an ack is needed.
3950 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3952 struct tcp_sock
*tp
= tcp_sk(sk
);
3954 /* More than one full frame received... */
3955 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3956 /* ... and right edge of window advances far enough.
3957 * (tcp_recvmsg() will send ACK otherwise). Or...
3959 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3960 /* We ACK each frame or... */
3961 tcp_in_quickack_mode(sk
) ||
3962 /* We have out of order data. */
3964 skb_peek(&tp
->out_of_order_queue
))) {
3965 /* Then ack it now */
3968 /* Else, send delayed ack. */
3969 tcp_send_delayed_ack(sk
);
3973 static inline void tcp_ack_snd_check(struct sock
*sk
)
3975 if (!inet_csk_ack_scheduled(sk
)) {
3976 /* We sent a data segment already. */
3979 __tcp_ack_snd_check(sk
, 1);
3983 * This routine is only called when we have urgent data
3984 * signaled. Its the 'slow' part of tcp_urg. It could be
3985 * moved inline now as tcp_urg is only called from one
3986 * place. We handle URGent data wrong. We have to - as
3987 * BSD still doesn't use the correction from RFC961.
3988 * For 1003.1g we should support a new option TCP_STDURG to permit
3989 * either form (or just set the sysctl tcp_stdurg).
3992 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
3994 struct tcp_sock
*tp
= tcp_sk(sk
);
3995 u32 ptr
= ntohs(th
->urg_ptr
);
3997 if (ptr
&& !sysctl_tcp_stdurg
)
3999 ptr
+= ntohl(th
->seq
);
4001 /* Ignore urgent data that we've already seen and read. */
4002 if (after(tp
->copied_seq
, ptr
))
4005 /* Do not replay urg ptr.
4007 * NOTE: interesting situation not covered by specs.
4008 * Misbehaving sender may send urg ptr, pointing to segment,
4009 * which we already have in ofo queue. We are not able to fetch
4010 * such data and will stay in TCP_URG_NOTYET until will be eaten
4011 * by recvmsg(). Seems, we are not obliged to handle such wicked
4012 * situations. But it is worth to think about possibility of some
4013 * DoSes using some hypothetical application level deadlock.
4015 if (before(ptr
, tp
->rcv_nxt
))
4018 /* Do we already have a newer (or duplicate) urgent pointer? */
4019 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4022 /* Tell the world about our new urgent pointer. */
4025 /* We may be adding urgent data when the last byte read was
4026 * urgent. To do this requires some care. We cannot just ignore
4027 * tp->copied_seq since we would read the last urgent byte again
4028 * as data, nor can we alter copied_seq until this data arrives
4029 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4031 * NOTE. Double Dutch. Rendering to plain English: author of comment
4032 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4033 * and expect that both A and B disappear from stream. This is _wrong_.
4034 * Though this happens in BSD with high probability, this is occasional.
4035 * Any application relying on this is buggy. Note also, that fix "works"
4036 * only in this artificial test. Insert some normal data between A and B and we will
4037 * decline of BSD again. Verdict: it is better to remove to trap
4040 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4041 !sock_flag(sk
, SOCK_URGINLINE
) &&
4042 tp
->copied_seq
!= tp
->rcv_nxt
) {
4043 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4045 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4046 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4051 tp
->urg_data
= TCP_URG_NOTYET
;
4054 /* Disable header prediction. */
4058 /* This is the 'fast' part of urgent handling. */
4059 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4061 struct tcp_sock
*tp
= tcp_sk(sk
);
4063 /* Check if we get a new urgent pointer - normally not. */
4065 tcp_check_urg(sk
,th
);
4067 /* Do we wait for any urgent data? - normally not... */
4068 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4069 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4072 /* Is the urgent pointer pointing into this packet? */
4073 if (ptr
< skb
->len
) {
4075 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4077 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4078 if (!sock_flag(sk
, SOCK_DEAD
))
4079 sk
->sk_data_ready(sk
, 0);
4084 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4086 struct tcp_sock
*tp
= tcp_sk(sk
);
4087 int chunk
= skb
->len
- hlen
;
4091 if (skb_csum_unnecessary(skb
))
4092 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4094 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4098 tp
->ucopy
.len
-= chunk
;
4099 tp
->copied_seq
+= chunk
;
4100 tcp_rcv_space_adjust(sk
);
4107 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4111 if (sock_owned_by_user(sk
)) {
4113 result
= __tcp_checksum_complete(skb
);
4116 result
= __tcp_checksum_complete(skb
);
4121 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4123 return !skb_csum_unnecessary(skb
) &&
4124 __tcp_checksum_complete_user(sk
, skb
);
4127 #ifdef CONFIG_NET_DMA
4128 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4130 struct tcp_sock
*tp
= tcp_sk(sk
);
4131 int chunk
= skb
->len
- hlen
;
4133 int copied_early
= 0;
4135 if (tp
->ucopy
.wakeup
)
4138 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4139 tp
->ucopy
.dma_chan
= get_softnet_dma();
4141 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4143 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4144 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4149 tp
->ucopy
.dma_cookie
= dma_cookie
;
4152 tp
->ucopy
.len
-= chunk
;
4153 tp
->copied_seq
+= chunk
;
4154 tcp_rcv_space_adjust(sk
);
4156 if ((tp
->ucopy
.len
== 0) ||
4157 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4158 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4159 tp
->ucopy
.wakeup
= 1;
4160 sk
->sk_data_ready(sk
, 0);
4162 } else if (chunk
> 0) {
4163 tp
->ucopy
.wakeup
= 1;
4164 sk
->sk_data_ready(sk
, 0);
4167 return copied_early
;
4169 #endif /* CONFIG_NET_DMA */
4172 * TCP receive function for the ESTABLISHED state.
4174 * It is split into a fast path and a slow path. The fast path is
4176 * - A zero window was announced from us - zero window probing
4177 * is only handled properly in the slow path.
4178 * - Out of order segments arrived.
4179 * - Urgent data is expected.
4180 * - There is no buffer space left
4181 * - Unexpected TCP flags/window values/header lengths are received
4182 * (detected by checking the TCP header against pred_flags)
4183 * - Data is sent in both directions. Fast path only supports pure senders
4184 * or pure receivers (this means either the sequence number or the ack
4185 * value must stay constant)
4186 * - Unexpected TCP option.
4188 * When these conditions are not satisfied it drops into a standard
4189 * receive procedure patterned after RFC793 to handle all cases.
4190 * The first three cases are guaranteed by proper pred_flags setting,
4191 * the rest is checked inline. Fast processing is turned on in
4192 * tcp_data_queue when everything is OK.
4194 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4195 struct tcphdr
*th
, unsigned len
)
4197 struct tcp_sock
*tp
= tcp_sk(sk
);
4200 * Header prediction.
4201 * The code loosely follows the one in the famous
4202 * "30 instruction TCP receive" Van Jacobson mail.
4204 * Van's trick is to deposit buffers into socket queue
4205 * on a device interrupt, to call tcp_recv function
4206 * on the receive process context and checksum and copy
4207 * the buffer to user space. smart...
4209 * Our current scheme is not silly either but we take the
4210 * extra cost of the net_bh soft interrupt processing...
4211 * We do checksum and copy also but from device to kernel.
4214 tp
->rx_opt
.saw_tstamp
= 0;
4216 /* pred_flags is 0xS?10 << 16 + snd_wnd
4217 * if header_prediction is to be made
4218 * 'S' will always be tp->tcp_header_len >> 2
4219 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4220 * turn it off (when there are holes in the receive
4221 * space for instance)
4222 * PSH flag is ignored.
4225 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4226 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4227 int tcp_header_len
= tp
->tcp_header_len
;
4229 /* Timestamp header prediction: tcp_header_len
4230 * is automatically equal to th->doff*4 due to pred_flags
4234 /* Check timestamp */
4235 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4236 __be32
*ptr
= (__be32
*)(th
+ 1);
4238 /* No? Slow path! */
4239 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4240 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4243 tp
->rx_opt
.saw_tstamp
= 1;
4245 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4247 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4249 /* If PAWS failed, check it more carefully in slow path */
4250 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4253 /* DO NOT update ts_recent here, if checksum fails
4254 * and timestamp was corrupted part, it will result
4255 * in a hung connection since we will drop all
4256 * future packets due to the PAWS test.
4260 if (len
<= tcp_header_len
) {
4261 /* Bulk data transfer: sender */
4262 if (len
== tcp_header_len
) {
4263 /* Predicted packet is in window by definition.
4264 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4265 * Hence, check seq<=rcv_wup reduces to:
4267 if (tcp_header_len
==
4268 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4269 tp
->rcv_nxt
== tp
->rcv_wup
)
4270 tcp_store_ts_recent(tp
);
4272 /* We know that such packets are checksummed
4275 tcp_ack(sk
, skb
, 0);
4277 tcp_data_snd_check(sk
);
4279 } else { /* Header too small */
4280 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4285 int copied_early
= 0;
4287 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4288 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4289 #ifdef CONFIG_NET_DMA
4290 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4295 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4296 __set_current_state(TASK_RUNNING
);
4298 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4302 /* Predicted packet is in window by definition.
4303 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4304 * Hence, check seq<=rcv_wup reduces to:
4306 if (tcp_header_len
==
4307 (sizeof(struct tcphdr
) +
4308 TCPOLEN_TSTAMP_ALIGNED
) &&
4309 tp
->rcv_nxt
== tp
->rcv_wup
)
4310 tcp_store_ts_recent(tp
);
4312 tcp_rcv_rtt_measure_ts(sk
, skb
);
4314 __skb_pull(skb
, tcp_header_len
);
4315 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4316 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4319 tcp_cleanup_rbuf(sk
, skb
->len
);
4322 if (tcp_checksum_complete_user(sk
, skb
))
4325 /* Predicted packet is in window by definition.
4326 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4327 * Hence, check seq<=rcv_wup reduces to:
4329 if (tcp_header_len
==
4330 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4331 tp
->rcv_nxt
== tp
->rcv_wup
)
4332 tcp_store_ts_recent(tp
);
4334 tcp_rcv_rtt_measure_ts(sk
, skb
);
4336 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4339 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4341 /* Bulk data transfer: receiver */
4342 __skb_pull(skb
,tcp_header_len
);
4343 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4344 sk_stream_set_owner_r(skb
, sk
);
4345 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4348 tcp_event_data_recv(sk
, skb
);
4350 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4351 /* Well, only one small jumplet in fast path... */
4352 tcp_ack(sk
, skb
, FLAG_DATA
);
4353 tcp_data_snd_check(sk
);
4354 if (!inet_csk_ack_scheduled(sk
))
4358 __tcp_ack_snd_check(sk
, 0);
4360 #ifdef CONFIG_NET_DMA
4362 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4368 sk
->sk_data_ready(sk
, 0);
4374 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4378 * RFC1323: H1. Apply PAWS check first.
4380 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4381 tcp_paws_discard(sk
, skb
)) {
4383 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4384 tcp_send_dupack(sk
, skb
);
4387 /* Resets are accepted even if PAWS failed.
4389 ts_recent update must be made after we are sure
4390 that the packet is in window.
4395 * Standard slow path.
4398 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4399 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4400 * (RST) segments are validated by checking their SEQ-fields."
4401 * And page 69: "If an incoming segment is not acceptable,
4402 * an acknowledgment should be sent in reply (unless the RST bit
4403 * is set, if so drop the segment and return)".
4406 tcp_send_dupack(sk
, skb
);
4415 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4417 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4418 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4419 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4426 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4428 tcp_rcv_rtt_measure_ts(sk
, skb
);
4430 /* Process urgent data. */
4431 tcp_urg(sk
, skb
, th
);
4433 /* step 7: process the segment text */
4434 tcp_data_queue(sk
, skb
);
4436 tcp_data_snd_check(sk
);
4437 tcp_ack_snd_check(sk
);
4441 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4448 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4449 struct tcphdr
*th
, unsigned len
)
4451 struct tcp_sock
*tp
= tcp_sk(sk
);
4452 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4453 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4455 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4459 * "If the state is SYN-SENT then
4460 * first check the ACK bit
4461 * If the ACK bit is set
4462 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4463 * a reset (unless the RST bit is set, if so drop
4464 * the segment and return)"
4466 * We do not send data with SYN, so that RFC-correct
4469 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4470 goto reset_and_undo
;
4472 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4473 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4475 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4476 goto reset_and_undo
;
4479 /* Now ACK is acceptable.
4481 * "If the RST bit is set
4482 * If the ACK was acceptable then signal the user "error:
4483 * connection reset", drop the segment, enter CLOSED state,
4484 * delete TCB, and return."
4493 * "fifth, if neither of the SYN or RST bits is set then
4494 * drop the segment and return."
4500 goto discard_and_undo
;
4503 * "If the SYN bit is on ...
4504 * are acceptable then ...
4505 * (our SYN has been ACKed), change the connection
4506 * state to ESTABLISHED..."
4509 TCP_ECN_rcv_synack(tp
, th
);
4511 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4512 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4514 /* Ok.. it's good. Set up sequence numbers and
4515 * move to established.
4517 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4518 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4520 /* RFC1323: The window in SYN & SYN/ACK segments is
4523 tp
->snd_wnd
= ntohs(th
->window
);
4524 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4526 if (!tp
->rx_opt
.wscale_ok
) {
4527 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4528 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4531 if (tp
->rx_opt
.saw_tstamp
) {
4532 tp
->rx_opt
.tstamp_ok
= 1;
4533 tp
->tcp_header_len
=
4534 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4535 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4536 tcp_store_ts_recent(tp
);
4538 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4541 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4542 tp
->rx_opt
.sack_ok
|= 2;
4545 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4546 tcp_initialize_rcv_mss(sk
);
4548 /* Remember, tcp_poll() does not lock socket!
4549 * Change state from SYN-SENT only after copied_seq
4550 * is initialized. */
4551 tp
->copied_seq
= tp
->rcv_nxt
;
4553 tcp_set_state(sk
, TCP_ESTABLISHED
);
4555 security_inet_conn_established(sk
, skb
);
4557 /* Make sure socket is routed, for correct metrics. */
4558 icsk
->icsk_af_ops
->rebuild_header(sk
);
4560 tcp_init_metrics(sk
);
4562 tcp_init_congestion_control(sk
);
4564 /* Prevent spurious tcp_cwnd_restart() on first data
4567 tp
->lsndtime
= tcp_time_stamp
;
4569 tcp_init_buffer_space(sk
);
4571 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4572 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4574 if (!tp
->rx_opt
.snd_wscale
)
4575 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4579 if (!sock_flag(sk
, SOCK_DEAD
)) {
4580 sk
->sk_state_change(sk
);
4581 sk_wake_async(sk
, 0, POLL_OUT
);
4584 if (sk
->sk_write_pending
||
4585 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4586 icsk
->icsk_ack
.pingpong
) {
4587 /* Save one ACK. Data will be ready after
4588 * several ticks, if write_pending is set.
4590 * It may be deleted, but with this feature tcpdumps
4591 * look so _wonderfully_ clever, that I was not able
4592 * to stand against the temptation 8) --ANK
4594 inet_csk_schedule_ack(sk
);
4595 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4596 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4597 tcp_incr_quickack(sk
);
4598 tcp_enter_quickack_mode(sk
);
4599 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4600 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4611 /* No ACK in the segment */
4615 * "If the RST bit is set
4617 * Otherwise (no ACK) drop the segment and return."
4620 goto discard_and_undo
;
4624 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
4625 tcp_paws_reject(&tp
->rx_opt
, 0))
4626 goto discard_and_undo
;
4629 /* We see SYN without ACK. It is attempt of
4630 * simultaneous connect with crossed SYNs.
4631 * Particularly, it can be connect to self.
4633 tcp_set_state(sk
, TCP_SYN_RECV
);
4635 if (tp
->rx_opt
.saw_tstamp
) {
4636 tp
->rx_opt
.tstamp_ok
= 1;
4637 tcp_store_ts_recent(tp
);
4638 tp
->tcp_header_len
=
4639 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4641 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4644 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4645 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4647 /* RFC1323: The window in SYN & SYN/ACK segments is
4650 tp
->snd_wnd
= ntohs(th
->window
);
4651 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4652 tp
->max_window
= tp
->snd_wnd
;
4654 TCP_ECN_rcv_syn(tp
, th
);
4657 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4658 tcp_initialize_rcv_mss(sk
);
4661 tcp_send_synack(sk
);
4663 /* Note, we could accept data and URG from this segment.
4664 * There are no obstacles to make this.
4666 * However, if we ignore data in ACKless segments sometimes,
4667 * we have no reasons to accept it sometimes.
4668 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4669 * is not flawless. So, discard packet for sanity.
4670 * Uncomment this return to process the data.
4677 /* "fifth, if neither of the SYN or RST bits is set then
4678 * drop the segment and return."
4682 tcp_clear_options(&tp
->rx_opt
);
4683 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4687 tcp_clear_options(&tp
->rx_opt
);
4688 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4694 * This function implements the receiving procedure of RFC 793 for
4695 * all states except ESTABLISHED and TIME_WAIT.
4696 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4697 * address independent.
4700 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4701 struct tcphdr
*th
, unsigned len
)
4703 struct tcp_sock
*tp
= tcp_sk(sk
);
4704 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4707 tp
->rx_opt
.saw_tstamp
= 0;
4709 switch (sk
->sk_state
) {
4721 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4724 /* Now we have several options: In theory there is
4725 * nothing else in the frame. KA9Q has an option to
4726 * send data with the syn, BSD accepts data with the
4727 * syn up to the [to be] advertised window and
4728 * Solaris 2.1 gives you a protocol error. For now
4729 * we just ignore it, that fits the spec precisely
4730 * and avoids incompatibilities. It would be nice in
4731 * future to drop through and process the data.
4733 * Now that TTCP is starting to be used we ought to
4735 * But, this leaves one open to an easy denial of
4736 * service attack, and SYN cookies can't defend
4737 * against this problem. So, we drop the data
4738 * in the interest of security over speed unless
4739 * it's still in use.
4747 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4751 /* Do step6 onward by hand. */
4752 tcp_urg(sk
, skb
, th
);
4754 tcp_data_snd_check(sk
);
4758 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4759 tcp_paws_discard(sk
, skb
)) {
4761 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4762 tcp_send_dupack(sk
, skb
);
4765 /* Reset is accepted even if it did not pass PAWS. */
4768 /* step 1: check sequence number */
4769 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4771 tcp_send_dupack(sk
, skb
);
4775 /* step 2: check RST bit */
4781 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4783 /* step 3: check security and precedence [ignored] */
4787 * Check for a SYN in window.
4789 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4790 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4795 /* step 5: check the ACK field */
4797 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4799 switch (sk
->sk_state
) {
4802 tp
->copied_seq
= tp
->rcv_nxt
;
4804 tcp_set_state(sk
, TCP_ESTABLISHED
);
4805 sk
->sk_state_change(sk
);
4807 /* Note, that this wakeup is only for marginal
4808 * crossed SYN case. Passively open sockets
4809 * are not waked up, because sk->sk_sleep ==
4810 * NULL and sk->sk_socket == NULL.
4812 if (sk
->sk_socket
) {
4813 sk_wake_async(sk
,0,POLL_OUT
);
4816 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4817 tp
->snd_wnd
= ntohs(th
->window
) <<
4818 tp
->rx_opt
.snd_wscale
;
4819 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4820 TCP_SKB_CB(skb
)->seq
);
4822 /* tcp_ack considers this ACK as duplicate
4823 * and does not calculate rtt.
4824 * Fix it at least with timestamps.
4826 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4828 tcp_ack_saw_tstamp(sk
, 0);
4830 if (tp
->rx_opt
.tstamp_ok
)
4831 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4833 /* Make sure socket is routed, for
4836 icsk
->icsk_af_ops
->rebuild_header(sk
);
4838 tcp_init_metrics(sk
);
4840 tcp_init_congestion_control(sk
);
4842 /* Prevent spurious tcp_cwnd_restart() on
4843 * first data packet.
4845 tp
->lsndtime
= tcp_time_stamp
;
4848 tcp_initialize_rcv_mss(sk
);
4849 tcp_init_buffer_space(sk
);
4850 tcp_fast_path_on(tp
);
4857 if (tp
->snd_una
== tp
->write_seq
) {
4858 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4859 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4860 dst_confirm(sk
->sk_dst_cache
);
4862 if (!sock_flag(sk
, SOCK_DEAD
))
4863 /* Wake up lingering close() */
4864 sk
->sk_state_change(sk
);
4868 if (tp
->linger2
< 0 ||
4869 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4870 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4872 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4876 tmo
= tcp_fin_time(sk
);
4877 if (tmo
> TCP_TIMEWAIT_LEN
) {
4878 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4879 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4880 /* Bad case. We could lose such FIN otherwise.
4881 * It is not a big problem, but it looks confusing
4882 * and not so rare event. We still can lose it now,
4883 * if it spins in bh_lock_sock(), but it is really
4886 inet_csk_reset_keepalive_timer(sk
, tmo
);
4888 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4896 if (tp
->snd_una
== tp
->write_seq
) {
4897 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4903 if (tp
->snd_una
== tp
->write_seq
) {
4904 tcp_update_metrics(sk
);
4913 /* step 6: check the URG bit */
4914 tcp_urg(sk
, skb
, th
);
4916 /* step 7: process the segment text */
4917 switch (sk
->sk_state
) {
4918 case TCP_CLOSE_WAIT
:
4921 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4925 /* RFC 793 says to queue data in these states,
4926 * RFC 1122 says we MUST send a reset.
4927 * BSD 4.4 also does reset.
4929 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4930 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4931 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4932 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4938 case TCP_ESTABLISHED
:
4939 tcp_data_queue(sk
, skb
);
4944 /* tcp_data could move socket to TIME-WAIT */
4945 if (sk
->sk_state
!= TCP_CLOSE
) {
4946 tcp_data_snd_check(sk
);
4947 tcp_ack_snd_check(sk
);
4957 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4958 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4959 EXPORT_SYMBOL(tcp_parse_options
);
4960 EXPORT_SYMBOL(tcp_rcv_established
);
4961 EXPORT_SYMBOL(tcp_rcv_state_process
);
4962 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);