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
)/2;
246 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2])/2;
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 static u32
tcp_rto_min(struct sock
*sk
)
560 struct dst_entry
*dst
= __sk_dst_get(sk
);
561 u32 rto_min
= TCP_RTO_MIN
;
563 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
564 rto_min
= dst
->metrics
[RTAX_RTO_MIN
-1];
568 /* Called to compute a smoothed rtt estimate. The data fed to this
569 * routine either comes from timestamps, or from segments that were
570 * known _not_ to have been retransmitted [see Karn/Partridge
571 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
572 * piece by Van Jacobson.
573 * NOTE: the next three routines used to be one big routine.
574 * To save cycles in the RFC 1323 implementation it was better to break
575 * it up into three procedures. -- erics
577 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
579 struct tcp_sock
*tp
= tcp_sk(sk
);
580 long m
= mrtt
; /* RTT */
582 /* The following amusing code comes from Jacobson's
583 * article in SIGCOMM '88. Note that rtt and mdev
584 * are scaled versions of rtt and mean deviation.
585 * This is designed to be as fast as possible
586 * m stands for "measurement".
588 * On a 1990 paper the rto value is changed to:
589 * RTO = rtt + 4 * mdev
591 * Funny. This algorithm seems to be very broken.
592 * These formulae increase RTO, when it should be decreased, increase
593 * too slowly, when it should be increased quickly, decrease too quickly
594 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
595 * does not matter how to _calculate_ it. Seems, it was trap
596 * that VJ failed to avoid. 8)
601 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
602 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
604 m
= -m
; /* m is now abs(error) */
605 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
606 /* This is similar to one of Eifel findings.
607 * Eifel blocks mdev updates when rtt decreases.
608 * This solution is a bit different: we use finer gain
609 * for mdev in this case (alpha*beta).
610 * Like Eifel it also prevents growth of rto,
611 * but also it limits too fast rto decreases,
612 * happening in pure Eifel.
617 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
619 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
620 if (tp
->mdev
> tp
->mdev_max
) {
621 tp
->mdev_max
= tp
->mdev
;
622 if (tp
->mdev_max
> tp
->rttvar
)
623 tp
->rttvar
= tp
->mdev_max
;
625 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
626 if (tp
->mdev_max
< tp
->rttvar
)
627 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
628 tp
->rtt_seq
= tp
->snd_nxt
;
629 tp
->mdev_max
= tcp_rto_min(sk
);
632 /* no previous measure. */
633 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
634 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
635 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
636 tp
->rtt_seq
= tp
->snd_nxt
;
640 /* Calculate rto without backoff. This is the second half of Van Jacobson's
641 * routine referred to above.
643 static inline void tcp_set_rto(struct sock
*sk
)
645 const struct tcp_sock
*tp
= tcp_sk(sk
);
646 /* Old crap is replaced with new one. 8)
649 * 1. If rtt variance happened to be less 50msec, it is hallucination.
650 * It cannot be less due to utterly erratic ACK generation made
651 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
652 * to do with delayed acks, because at cwnd>2 true delack timeout
653 * is invisible. Actually, Linux-2.4 also generates erratic
654 * ACKs in some circumstances.
656 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
658 /* 2. Fixups made earlier cannot be right.
659 * If we do not estimate RTO correctly without them,
660 * all the algo is pure shit and should be replaced
661 * with correct one. It is exactly, which we pretend to do.
665 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
666 * guarantees that rto is higher.
668 static inline void tcp_bound_rto(struct sock
*sk
)
670 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
671 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
674 /* Save metrics learned by this TCP session.
675 This function is called only, when TCP finishes successfully
676 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
678 void tcp_update_metrics(struct sock
*sk
)
680 struct tcp_sock
*tp
= tcp_sk(sk
);
681 struct dst_entry
*dst
= __sk_dst_get(sk
);
683 if (sysctl_tcp_nometrics_save
)
688 if (dst
&& (dst
->flags
&DST_HOST
)) {
689 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
692 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
693 /* This session failed to estimate rtt. Why?
694 * Probably, no packets returned in time.
697 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
698 dst
->metrics
[RTAX_RTT
-1] = 0;
702 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
704 /* If newly calculated rtt larger than stored one,
705 * store new one. Otherwise, use EWMA. Remember,
706 * rtt overestimation is always better than underestimation.
708 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
710 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
712 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
715 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
719 /* Scale deviation to rttvar fixed point */
724 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
725 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
727 dst
->metrics
[RTAX_RTTVAR
-1] -=
728 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
731 if (tp
->snd_ssthresh
>= 0xFFFF) {
732 /* Slow start still did not finish. */
733 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
734 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
735 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
736 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
737 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
738 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
739 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
740 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
741 icsk
->icsk_ca_state
== TCP_CA_Open
) {
742 /* Cong. avoidance phase, cwnd is reliable. */
743 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
744 dst
->metrics
[RTAX_SSTHRESH
-1] =
745 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
746 if (!dst_metric_locked(dst
, RTAX_CWND
))
747 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
749 /* Else slow start did not finish, cwnd is non-sense,
750 ssthresh may be also invalid.
752 if (!dst_metric_locked(dst
, RTAX_CWND
))
753 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
754 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
755 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
756 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
757 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
760 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
761 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
762 tp
->reordering
!= sysctl_tcp_reordering
)
763 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
768 /* Numbers are taken from RFC3390.
770 * John Heffner states:
772 * The RFC specifies a window of no more than 4380 bytes
773 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
774 * is a bit misleading because they use a clamp at 4380 bytes
775 * rather than use a multiplier in the relevant range.
777 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
779 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
782 if (tp
->mss_cache
> 1460)
785 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
787 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
790 /* Set slow start threshold and cwnd not falling to slow start */
791 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
793 struct tcp_sock
*tp
= tcp_sk(sk
);
794 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
796 tp
->prior_ssthresh
= 0;
798 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
801 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
802 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
803 tcp_packets_in_flight(tp
) + 1U);
804 tp
->snd_cwnd_cnt
= 0;
805 tp
->high_seq
= tp
->snd_nxt
;
806 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
807 TCP_ECN_queue_cwr(tp
);
809 tcp_set_ca_state(sk
, TCP_CA_CWR
);
813 /* Initialize metrics on socket. */
815 static void tcp_init_metrics(struct sock
*sk
)
817 struct tcp_sock
*tp
= tcp_sk(sk
);
818 struct dst_entry
*dst
= __sk_dst_get(sk
);
825 if (dst_metric_locked(dst
, RTAX_CWND
))
826 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
827 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
828 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
829 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
830 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
832 if (dst_metric(dst
, RTAX_REORDERING
) &&
833 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
834 tp
->rx_opt
.sack_ok
&= ~2;
835 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
838 if (dst_metric(dst
, RTAX_RTT
) == 0)
841 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
844 /* Initial rtt is determined from SYN,SYN-ACK.
845 * The segment is small and rtt may appear much
846 * less than real one. Use per-dst memory
847 * to make it more realistic.
849 * A bit of theory. RTT is time passed after "normal" sized packet
850 * is sent until it is ACKed. In normal circumstances sending small
851 * packets force peer to delay ACKs and calculation is correct too.
852 * The algorithm is adaptive and, provided we follow specs, it
853 * NEVER underestimate RTT. BUT! If peer tries to make some clever
854 * tricks sort of "quick acks" for time long enough to decrease RTT
855 * to low value, and then abruptly stops to do it and starts to delay
856 * ACKs, wait for troubles.
858 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
859 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
860 tp
->rtt_seq
= tp
->snd_nxt
;
862 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
863 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
864 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
868 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
870 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
871 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
875 /* Play conservative. If timestamps are not
876 * supported, TCP will fail to recalculate correct
877 * rtt, if initial rto is too small. FORGET ALL AND RESET!
879 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
881 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
882 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
886 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
889 struct tcp_sock
*tp
= tcp_sk(sk
);
890 if (metric
> tp
->reordering
) {
891 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
893 /* This exciting event is worth to be remembered. 8) */
895 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
897 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
899 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
901 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
902 #if FASTRETRANS_DEBUG > 1
903 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
904 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
908 tp
->undo_marker
? tp
->undo_retrans
: 0);
910 /* Disable FACK yet. */
911 tp
->rx_opt
.sack_ok
&= ~2;
915 /* This procedure tags the retransmission queue when SACKs arrive.
917 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
918 * Packets in queue with these bits set are counted in variables
919 * sacked_out, retrans_out and lost_out, correspondingly.
921 * Valid combinations are:
922 * Tag InFlight Description
923 * 0 1 - orig segment is in flight.
924 * S 0 - nothing flies, orig reached receiver.
925 * L 0 - nothing flies, orig lost by net.
926 * R 2 - both orig and retransmit are in flight.
927 * L|R 1 - orig is lost, retransmit is in flight.
928 * S|R 1 - orig reached receiver, retrans is still in flight.
929 * (L|S|R is logically valid, it could occur when L|R is sacked,
930 * but it is equivalent to plain S and code short-curcuits it to S.
931 * L|S is logically invalid, it would mean -1 packet in flight 8))
933 * These 6 states form finite state machine, controlled by the following events:
934 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
935 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
936 * 3. Loss detection event of one of three flavors:
937 * A. Scoreboard estimator decided the packet is lost.
938 * A'. Reno "three dupacks" marks head of queue lost.
939 * A''. Its FACK modfication, head until snd.fack is lost.
940 * B. SACK arrives sacking data transmitted after never retransmitted
942 * C. SACK arrives sacking SND.NXT at the moment, when the
943 * segment was retransmitted.
944 * 4. D-SACK added new rule: D-SACK changes any tag to S.
946 * It is pleasant to note, that state diagram turns out to be commutative,
947 * so that we are allowed not to be bothered by order of our actions,
948 * when multiple events arrive simultaneously. (see the function below).
950 * Reordering detection.
951 * --------------------
952 * Reordering metric is maximal distance, which a packet can be displaced
953 * in packet stream. With SACKs we can estimate it:
955 * 1. SACK fills old hole and the corresponding segment was not
956 * ever retransmitted -> reordering. Alas, we cannot use it
957 * when segment was retransmitted.
958 * 2. The last flaw is solved with D-SACK. D-SACK arrives
959 * for retransmitted and already SACKed segment -> reordering..
960 * Both of these heuristics are not used in Loss state, when we cannot
961 * account for retransmits accurately.
963 static int tcp_check_dsack(struct tcp_sock
*tp
, struct sk_buff
*ack_skb
,
964 struct tcp_sack_block_wire
*sp
, int num_sacks
,
967 u32 start_seq_0
= ntohl(get_unaligned(&sp
[0].start_seq
));
968 u32 end_seq_0
= ntohl(get_unaligned(&sp
[0].end_seq
));
971 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
973 tp
->rx_opt
.sack_ok
|= 4;
974 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
975 } else if (num_sacks
> 1) {
976 u32 end_seq_1
= ntohl(get_unaligned(&sp
[1].end_seq
));
977 u32 start_seq_1
= ntohl(get_unaligned(&sp
[1].start_seq
));
979 if (!after(end_seq_0
, end_seq_1
) &&
980 !before(start_seq_0
, start_seq_1
)) {
982 tp
->rx_opt
.sack_ok
|= 4;
983 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
987 /* D-SACK for already forgotten data... Do dumb counting. */
989 !after(end_seq_0
, prior_snd_una
) &&
990 after(end_seq_0
, tp
->undo_marker
))
997 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
999 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1000 struct tcp_sock
*tp
= tcp_sk(sk
);
1001 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1002 TCP_SKB_CB(ack_skb
)->sacked
);
1003 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
1004 struct sk_buff
*cached_skb
;
1005 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
1006 int reord
= tp
->packets_out
;
1008 u32 lost_retrans
= 0;
1010 int found_dup_sack
= 0;
1011 int cached_fack_count
;
1013 int first_sack_index
;
1015 if (!tp
->sacked_out
) {
1016 tp
->fackets_out
= 0;
1017 tp
->highest_sack
= tp
->snd_una
;
1019 prior_fackets
= tp
->fackets_out
;
1021 found_dup_sack
= tcp_check_dsack(tp
, ack_skb
, sp
,
1022 num_sacks
, prior_snd_una
);
1024 flag
|= FLAG_DSACKING_ACK
;
1026 /* Eliminate too old ACKs, but take into
1027 * account more or less fresh ones, they can
1028 * contain valid SACK info.
1030 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1034 * if the only SACK change is the increase of the end_seq of
1035 * the first block then only apply that SACK block
1036 * and use retrans queue hinting otherwise slowpath */
1038 for (i
= 0; i
< num_sacks
; i
++) {
1039 __be32 start_seq
= sp
[i
].start_seq
;
1040 __be32 end_seq
= sp
[i
].end_seq
;
1043 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
1046 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
1047 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1050 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1051 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1053 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1054 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1055 tp
->recv_sack_cache
[i
].start_seq
= 0;
1056 tp
->recv_sack_cache
[i
].end_seq
= 0;
1059 first_sack_index
= 0;
1064 tp
->fastpath_skb_hint
= NULL
;
1066 /* order SACK blocks to allow in order walk of the retrans queue */
1067 for (i
= num_sacks
-1; i
> 0; i
--) {
1068 for (j
= 0; j
< i
; j
++){
1069 if (after(ntohl(sp
[j
].start_seq
),
1070 ntohl(sp
[j
+1].start_seq
))){
1071 struct tcp_sack_block_wire tmp
;
1077 /* Track where the first SACK block goes to */
1078 if (j
== first_sack_index
)
1079 first_sack_index
= j
+1;
1086 /* clear flag as used for different purpose in following code */
1089 /* Use SACK fastpath hint if valid */
1090 cached_skb
= tp
->fastpath_skb_hint
;
1091 cached_fack_count
= tp
->fastpath_cnt_hint
;
1093 cached_skb
= tcp_write_queue_head(sk
);
1094 cached_fack_count
= 0;
1097 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1098 struct sk_buff
*skb
;
1099 __u32 start_seq
= ntohl(sp
->start_seq
);
1100 __u32 end_seq
= ntohl(sp
->end_seq
);
1102 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1105 fack_count
= cached_fack_count
;
1107 /* Event "B" in the comment above. */
1108 if (after(end_seq
, tp
->high_seq
))
1109 flag
|= FLAG_DATA_LOST
;
1111 tcp_for_write_queue_from(skb
, sk
) {
1112 int in_sack
, pcount
;
1115 if (skb
== tcp_send_head(sk
))
1119 cached_fack_count
= fack_count
;
1120 if (i
== first_sack_index
) {
1121 tp
->fastpath_skb_hint
= skb
;
1122 tp
->fastpath_cnt_hint
= fack_count
;
1125 /* The retransmission queue is always in order, so
1126 * we can short-circuit the walk early.
1128 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1131 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1132 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1134 pcount
= tcp_skb_pcount(skb
);
1136 if (pcount
> 1 && !in_sack
&&
1137 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1138 unsigned int pkt_len
;
1140 in_sack
= !after(start_seq
,
1141 TCP_SKB_CB(skb
)->seq
);
1144 pkt_len
= (start_seq
-
1145 TCP_SKB_CB(skb
)->seq
);
1147 pkt_len
= (end_seq
-
1148 TCP_SKB_CB(skb
)->seq
);
1149 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1151 pcount
= tcp_skb_pcount(skb
);
1154 fack_count
+= pcount
;
1156 sacked
= TCP_SKB_CB(skb
)->sacked
;
1158 /* Account D-SACK for retransmitted packet. */
1159 if ((dup_sack
&& in_sack
) &&
1160 (sacked
& TCPCB_RETRANS
) &&
1161 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1164 /* The frame is ACKed. */
1165 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1166 if (sacked
&TCPCB_RETRANS
) {
1167 if ((dup_sack
&& in_sack
) &&
1168 (sacked
&TCPCB_SACKED_ACKED
))
1169 reord
= min(fack_count
, reord
);
1171 /* If it was in a hole, we detected reordering. */
1172 if (fack_count
< prior_fackets
&&
1173 !(sacked
&TCPCB_SACKED_ACKED
))
1174 reord
= min(fack_count
, reord
);
1177 /* Nothing to do; acked frame is about to be dropped. */
1181 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1182 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1183 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1184 lost_retrans
= end_seq
;
1189 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1190 if (sacked
& TCPCB_SACKED_RETRANS
) {
1191 /* If the segment is not tagged as lost,
1192 * we do not clear RETRANS, believing
1193 * that retransmission is still in flight.
1195 if (sacked
& TCPCB_LOST
) {
1196 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1197 tp
->lost_out
-= tcp_skb_pcount(skb
);
1198 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1200 /* clear lost hint */
1201 tp
->retransmit_skb_hint
= NULL
;
1204 /* New sack for not retransmitted frame,
1205 * which was in hole. It is reordering.
1207 if (!(sacked
& TCPCB_RETRANS
) &&
1208 fack_count
< prior_fackets
)
1209 reord
= min(fack_count
, reord
);
1211 if (sacked
& TCPCB_LOST
) {
1212 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1213 tp
->lost_out
-= tcp_skb_pcount(skb
);
1215 /* clear lost hint */
1216 tp
->retransmit_skb_hint
= NULL
;
1218 /* SACK enhanced F-RTO detection.
1219 * Set flag if and only if non-rexmitted
1220 * segments below frto_highmark are
1221 * SACKed (RFC4138; Appendix B).
1222 * Clearing correct due to in-order walk
1224 if (after(end_seq
, tp
->frto_highmark
)) {
1225 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1227 if (!(sacked
& TCPCB_RETRANS
))
1228 flag
|= FLAG_ONLY_ORIG_SACKED
;
1232 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1233 flag
|= FLAG_DATA_SACKED
;
1234 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1236 if (fack_count
> tp
->fackets_out
)
1237 tp
->fackets_out
= fack_count
;
1239 if (after(TCP_SKB_CB(skb
)->seq
,
1241 tp
->highest_sack
= TCP_SKB_CB(skb
)->seq
;
1243 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1244 reord
= min(fack_count
, reord
);
1247 /* D-SACK. We can detect redundant retransmission
1248 * in S|R and plain R frames and clear it.
1249 * undo_retrans is decreased above, L|R frames
1250 * are accounted above as well.
1253 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1254 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1255 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1256 tp
->retransmit_skb_hint
= NULL
;
1261 /* Check for lost retransmit. This superb idea is
1262 * borrowed from "ratehalving". Event "C".
1263 * Later note: FACK people cheated me again 8),
1264 * we have to account for reordering! Ugly,
1267 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1268 struct sk_buff
*skb
;
1270 tcp_for_write_queue(skb
, sk
) {
1271 if (skb
== tcp_send_head(sk
))
1273 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1275 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1277 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1278 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1280 !before(lost_retrans
,
1281 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1283 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1284 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1286 /* clear lost hint */
1287 tp
->retransmit_skb_hint
= NULL
;
1289 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1290 tp
->lost_out
+= tcp_skb_pcount(skb
);
1291 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1292 flag
|= FLAG_DATA_SACKED
;
1293 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1299 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1301 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
&&
1302 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1303 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1305 #if FASTRETRANS_DEBUG > 0
1306 BUG_TRAP((int)tp
->sacked_out
>= 0);
1307 BUG_TRAP((int)tp
->lost_out
>= 0);
1308 BUG_TRAP((int)tp
->retrans_out
>= 0);
1309 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1314 /* F-RTO can only be used if TCP has never retransmitted anything other than
1315 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1317 int tcp_use_frto(struct sock
*sk
)
1319 const struct tcp_sock
*tp
= tcp_sk(sk
);
1320 struct sk_buff
*skb
;
1322 if (!sysctl_tcp_frto
)
1328 /* Avoid expensive walking of rexmit queue if possible */
1329 if (tp
->retrans_out
> 1)
1332 skb
= tcp_write_queue_head(sk
);
1333 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1334 tcp_for_write_queue_from(skb
, sk
) {
1335 if (skb
== tcp_send_head(sk
))
1337 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1339 /* Short-circuit when first non-SACKed skb has been checked */
1340 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1346 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1347 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1348 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1349 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1350 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1351 * bits are handled if the Loss state is really to be entered (in
1352 * tcp_enter_frto_loss).
1354 * Do like tcp_enter_loss() would; when RTO expires the second time it
1356 * "Reduce ssthresh if it has not yet been made inside this window."
1358 void tcp_enter_frto(struct sock
*sk
)
1360 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1361 struct tcp_sock
*tp
= tcp_sk(sk
);
1362 struct sk_buff
*skb
;
1364 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1365 tp
->snd_una
== tp
->high_seq
||
1366 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1367 !icsk
->icsk_retransmits
)) {
1368 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1369 /* Our state is too optimistic in ssthresh() call because cwnd
1370 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1371 * recovery has not yet completed. Pattern would be this: RTO,
1372 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1374 * RFC4138 should be more specific on what to do, even though
1375 * RTO is quite unlikely to occur after the first Cumulative ACK
1376 * due to back-off and complexity of triggering events ...
1378 if (tp
->frto_counter
) {
1380 stored_cwnd
= tp
->snd_cwnd
;
1382 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1383 tp
->snd_cwnd
= stored_cwnd
;
1385 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1387 /* ... in theory, cong.control module could do "any tricks" in
1388 * ssthresh(), which means that ca_state, lost bits and lost_out
1389 * counter would have to be faked before the call occurs. We
1390 * consider that too expensive, unlikely and hacky, so modules
1391 * using these in ssthresh() must deal these incompatibility
1392 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1394 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1397 tp
->undo_marker
= tp
->snd_una
;
1398 tp
->undo_retrans
= 0;
1400 skb
= tcp_write_queue_head(sk
);
1401 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1402 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1403 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1405 tcp_sync_left_out(tp
);
1407 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1408 * The last condition is necessary at least in tp->frto_counter case.
1410 if (IsSackFrto() && (tp
->frto_counter
||
1411 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1412 after(tp
->high_seq
, tp
->snd_una
)) {
1413 tp
->frto_highmark
= tp
->high_seq
;
1415 tp
->frto_highmark
= tp
->snd_nxt
;
1417 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1418 tp
->high_seq
= tp
->snd_nxt
;
1419 tp
->frto_counter
= 1;
1422 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1423 * which indicates that we should follow the traditional RTO recovery,
1424 * i.e. mark everything lost and do go-back-N retransmission.
1426 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1428 struct tcp_sock
*tp
= tcp_sk(sk
);
1429 struct sk_buff
*skb
;
1434 tp
->fackets_out
= 0;
1435 tp
->retrans_out
= 0;
1437 tcp_for_write_queue(skb
, sk
) {
1438 if (skb
== tcp_send_head(sk
))
1440 cnt
+= tcp_skb_pcount(skb
);
1442 * Count the retransmission made on RTO correctly (only when
1443 * waiting for the first ACK and did not get it)...
1445 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1446 /* For some reason this R-bit might get cleared? */
1447 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1448 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1449 /* ...enter this if branch just for the first segment */
1450 flag
|= FLAG_DATA_ACKED
;
1452 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1454 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1456 /* Do not mark those segments lost that were
1457 * forward transmitted after RTO
1459 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1460 tp
->frto_highmark
)) {
1461 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1462 tp
->lost_out
+= tcp_skb_pcount(skb
);
1465 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1466 tp
->fackets_out
= cnt
;
1469 tcp_sync_left_out(tp
);
1471 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1472 tp
->snd_cwnd_cnt
= 0;
1473 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1474 tp
->undo_marker
= 0;
1475 tp
->frto_counter
= 0;
1477 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1478 sysctl_tcp_reordering
);
1479 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1480 tp
->high_seq
= tp
->frto_highmark
;
1481 TCP_ECN_queue_cwr(tp
);
1483 clear_all_retrans_hints(tp
);
1486 void tcp_clear_retrans(struct tcp_sock
*tp
)
1489 tp
->retrans_out
= 0;
1491 tp
->fackets_out
= 0;
1495 tp
->undo_marker
= 0;
1496 tp
->undo_retrans
= 0;
1499 /* Enter Loss state. If "how" is not zero, forget all SACK information
1500 * and reset tags completely, otherwise preserve SACKs. If receiver
1501 * dropped its ofo queue, we will know this due to reneging detection.
1503 void tcp_enter_loss(struct sock
*sk
, int how
)
1505 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1506 struct tcp_sock
*tp
= tcp_sk(sk
);
1507 struct sk_buff
*skb
;
1510 /* Reduce ssthresh if it has not yet been made inside this window. */
1511 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1512 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1513 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1514 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1515 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1518 tp
->snd_cwnd_cnt
= 0;
1519 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1521 tp
->bytes_acked
= 0;
1522 tcp_clear_retrans(tp
);
1524 /* Push undo marker, if it was plain RTO and nothing
1525 * was retransmitted. */
1527 tp
->undo_marker
= tp
->snd_una
;
1529 tcp_for_write_queue(skb
, sk
) {
1530 if (skb
== tcp_send_head(sk
))
1532 cnt
+= tcp_skb_pcount(skb
);
1533 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1534 tp
->undo_marker
= 0;
1535 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1536 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1537 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1538 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1539 tp
->lost_out
+= tcp_skb_pcount(skb
);
1541 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1542 tp
->fackets_out
= cnt
;
1545 tcp_sync_left_out(tp
);
1547 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1548 sysctl_tcp_reordering
);
1549 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1550 tp
->high_seq
= tp
->snd_nxt
;
1551 TCP_ECN_queue_cwr(tp
);
1552 /* Abort FRTO algorithm if one is in progress */
1553 tp
->frto_counter
= 0;
1555 clear_all_retrans_hints(tp
);
1558 static int tcp_check_sack_reneging(struct sock
*sk
)
1560 struct sk_buff
*skb
;
1562 /* If ACK arrived pointing to a remembered SACK,
1563 * it means that our remembered SACKs do not reflect
1564 * real state of receiver i.e.
1565 * receiver _host_ is heavily congested (or buggy).
1566 * Do processing similar to RTO timeout.
1568 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1569 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1570 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1571 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1573 tcp_enter_loss(sk
, 1);
1574 icsk
->icsk_retransmits
++;
1575 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1576 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1577 icsk
->icsk_rto
, TCP_RTO_MAX
);
1583 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1585 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1588 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1590 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1593 static inline int tcp_head_timedout(struct sock
*sk
)
1595 struct tcp_sock
*tp
= tcp_sk(sk
);
1597 return tp
->packets_out
&&
1598 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1601 /* Linux NewReno/SACK/FACK/ECN state machine.
1602 * --------------------------------------
1604 * "Open" Normal state, no dubious events, fast path.
1605 * "Disorder" In all the respects it is "Open",
1606 * but requires a bit more attention. It is entered when
1607 * we see some SACKs or dupacks. It is split of "Open"
1608 * mainly to move some processing from fast path to slow one.
1609 * "CWR" CWND was reduced due to some Congestion Notification event.
1610 * It can be ECN, ICMP source quench, local device congestion.
1611 * "Recovery" CWND was reduced, we are fast-retransmitting.
1612 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1614 * tcp_fastretrans_alert() is entered:
1615 * - each incoming ACK, if state is not "Open"
1616 * - when arrived ACK is unusual, namely:
1621 * Counting packets in flight is pretty simple.
1623 * in_flight = packets_out - left_out + retrans_out
1625 * packets_out is SND.NXT-SND.UNA counted in packets.
1627 * retrans_out is number of retransmitted segments.
1629 * left_out is number of segments left network, but not ACKed yet.
1631 * left_out = sacked_out + lost_out
1633 * sacked_out: Packets, which arrived to receiver out of order
1634 * and hence not ACKed. With SACKs this number is simply
1635 * amount of SACKed data. Even without SACKs
1636 * it is easy to give pretty reliable estimate of this number,
1637 * counting duplicate ACKs.
1639 * lost_out: Packets lost by network. TCP has no explicit
1640 * "loss notification" feedback from network (for now).
1641 * It means that this number can be only _guessed_.
1642 * Actually, it is the heuristics to predict lossage that
1643 * distinguishes different algorithms.
1645 * F.e. after RTO, when all the queue is considered as lost,
1646 * lost_out = packets_out and in_flight = retrans_out.
1648 * Essentially, we have now two algorithms counting
1651 * FACK: It is the simplest heuristics. As soon as we decided
1652 * that something is lost, we decide that _all_ not SACKed
1653 * packets until the most forward SACK are lost. I.e.
1654 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1655 * It is absolutely correct estimate, if network does not reorder
1656 * packets. And it loses any connection to reality when reordering
1657 * takes place. We use FACK by default until reordering
1658 * is suspected on the path to this destination.
1660 * NewReno: when Recovery is entered, we assume that one segment
1661 * is lost (classic Reno). While we are in Recovery and
1662 * a partial ACK arrives, we assume that one more packet
1663 * is lost (NewReno). This heuristics are the same in NewReno
1666 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1667 * deflation etc. CWND is real congestion window, never inflated, changes
1668 * only according to classic VJ rules.
1670 * Really tricky (and requiring careful tuning) part of algorithm
1671 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1672 * The first determines the moment _when_ we should reduce CWND and,
1673 * hence, slow down forward transmission. In fact, it determines the moment
1674 * when we decide that hole is caused by loss, rather than by a reorder.
1676 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1677 * holes, caused by lost packets.
1679 * And the most logically complicated part of algorithm is undo
1680 * heuristics. We detect false retransmits due to both too early
1681 * fast retransmit (reordering) and underestimated RTO, analyzing
1682 * timestamps and D-SACKs. When we detect that some segments were
1683 * retransmitted by mistake and CWND reduction was wrong, we undo
1684 * window reduction and abort recovery phase. This logic is hidden
1685 * inside several functions named tcp_try_undo_<something>.
1688 /* This function decides, when we should leave Disordered state
1689 * and enter Recovery phase, reducing congestion window.
1691 * Main question: may we further continue forward transmission
1692 * with the same cwnd?
1694 static int tcp_time_to_recover(struct sock
*sk
)
1696 struct tcp_sock
*tp
= tcp_sk(sk
);
1699 /* Do not perform any recovery during FRTO algorithm */
1700 if (tp
->frto_counter
)
1703 /* Trick#1: The loss is proven. */
1707 /* Not-A-Trick#2 : Classic rule... */
1708 if (tcp_fackets_out(tp
) > tp
->reordering
)
1711 /* Trick#3 : when we use RFC2988 timer restart, fast
1712 * retransmit can be triggered by timeout of queue head.
1714 if (tcp_head_timedout(sk
))
1717 /* Trick#4: It is still not OK... But will it be useful to delay
1720 packets_out
= tp
->packets_out
;
1721 if (packets_out
<= tp
->reordering
&&
1722 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1723 !tcp_may_send_now(sk
)) {
1724 /* We have nothing to send. This connection is limited
1725 * either by receiver window or by application.
1733 /* If we receive more dupacks than we expected counting segments
1734 * in assumption of absent reordering, interpret this as reordering.
1735 * The only another reason could be bug in receiver TCP.
1737 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1739 struct tcp_sock
*tp
= tcp_sk(sk
);
1742 holes
= max(tp
->lost_out
, 1U);
1743 holes
= min(holes
, tp
->packets_out
);
1745 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1746 tp
->sacked_out
= tp
->packets_out
- holes
;
1747 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1751 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1753 static void tcp_add_reno_sack(struct sock
*sk
)
1755 struct tcp_sock
*tp
= tcp_sk(sk
);
1757 tcp_check_reno_reordering(sk
, 0);
1758 tcp_sync_left_out(tp
);
1761 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1763 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1765 struct tcp_sock
*tp
= tcp_sk(sk
);
1768 /* One ACK acked hole. The rest eat duplicate ACKs. */
1769 if (acked
-1 >= tp
->sacked_out
)
1772 tp
->sacked_out
-= acked
-1;
1774 tcp_check_reno_reordering(sk
, acked
);
1775 tcp_sync_left_out(tp
);
1778 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1781 tp
->left_out
= tp
->lost_out
;
1784 /* RFC: This is from the original, I doubt that this is necessary at all:
1785 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1786 * retransmitted past LOST markings in the first place? I'm not fully sure
1787 * about undo and end of connection cases, which can cause R without L?
1789 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
1790 struct sk_buff
*skb
)
1792 if ((tp
->retransmit_skb_hint
!= NULL
) &&
1793 before(TCP_SKB_CB(skb
)->seq
,
1794 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1795 tp
->retransmit_skb_hint
= NULL
;
1798 /* Mark head of queue up as lost. */
1799 static void tcp_mark_head_lost(struct sock
*sk
,
1800 int packets
, u32 high_seq
)
1802 struct tcp_sock
*tp
= tcp_sk(sk
);
1803 struct sk_buff
*skb
;
1806 BUG_TRAP(packets
<= tp
->packets_out
);
1807 if (tp
->lost_skb_hint
) {
1808 skb
= tp
->lost_skb_hint
;
1809 cnt
= tp
->lost_cnt_hint
;
1811 skb
= tcp_write_queue_head(sk
);
1815 tcp_for_write_queue_from(skb
, sk
) {
1816 if (skb
== tcp_send_head(sk
))
1818 /* TODO: do this better */
1819 /* this is not the most efficient way to do this... */
1820 tp
->lost_skb_hint
= skb
;
1821 tp
->lost_cnt_hint
= cnt
;
1822 cnt
+= tcp_skb_pcount(skb
);
1823 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1825 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1826 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1827 tp
->lost_out
+= tcp_skb_pcount(skb
);
1828 tcp_verify_retransmit_hint(tp
, skb
);
1831 tcp_sync_left_out(tp
);
1834 /* Account newly detected lost packet(s) */
1836 static void tcp_update_scoreboard(struct sock
*sk
)
1838 struct tcp_sock
*tp
= tcp_sk(sk
);
1841 int lost
= tp
->fackets_out
- tp
->reordering
;
1844 tcp_mark_head_lost(sk
, lost
, tp
->high_seq
);
1846 tcp_mark_head_lost(sk
, 1, tp
->high_seq
);
1849 /* New heuristics: it is possible only after we switched
1850 * to restart timer each time when something is ACKed.
1851 * Hence, we can detect timed out packets during fast
1852 * retransmit without falling to slow start.
1854 if (!IsReno(tp
) && tcp_head_timedout(sk
)) {
1855 struct sk_buff
*skb
;
1857 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1858 : tcp_write_queue_head(sk
);
1860 tcp_for_write_queue_from(skb
, sk
) {
1861 if (skb
== tcp_send_head(sk
))
1863 if (!tcp_skb_timedout(sk
, skb
))
1866 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1867 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1868 tp
->lost_out
+= tcp_skb_pcount(skb
);
1869 tcp_verify_retransmit_hint(tp
, skb
);
1873 tp
->scoreboard_skb_hint
= skb
;
1875 tcp_sync_left_out(tp
);
1879 /* CWND moderation, preventing bursts due to too big ACKs
1880 * in dubious situations.
1882 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1884 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1885 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1886 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1889 /* Lower bound on congestion window is slow start threshold
1890 * unless congestion avoidance choice decides to overide it.
1892 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1894 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1896 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1899 /* Decrease cwnd each second ack. */
1900 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
1902 struct tcp_sock
*tp
= tcp_sk(sk
);
1903 int decr
= tp
->snd_cwnd_cnt
+ 1;
1905 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
1906 (IsReno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
1907 tp
->snd_cwnd_cnt
= decr
&1;
1910 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1911 tp
->snd_cwnd
-= decr
;
1913 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1914 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1918 /* Nothing was retransmitted or returned timestamp is less
1919 * than timestamp of the first retransmission.
1921 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1923 return !tp
->retrans_stamp
||
1924 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1925 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1928 /* Undo procedures. */
1930 #if FASTRETRANS_DEBUG > 1
1931 static void DBGUNDO(struct sock
*sk
, const char *msg
)
1933 struct tcp_sock
*tp
= tcp_sk(sk
);
1934 struct inet_sock
*inet
= inet_sk(sk
);
1936 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1938 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1939 tp
->snd_cwnd
, tp
->left_out
,
1940 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1944 #define DBGUNDO(x...) do { } while (0)
1947 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1949 struct tcp_sock
*tp
= tcp_sk(sk
);
1951 if (tp
->prior_ssthresh
) {
1952 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1954 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1955 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1957 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1959 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1960 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1961 TCP_ECN_withdraw_cwr(tp
);
1964 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1966 tcp_moderate_cwnd(tp
);
1967 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1969 /* There is something screwy going on with the retrans hints after
1971 clear_all_retrans_hints(tp
);
1974 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1976 return tp
->undo_marker
&&
1977 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1980 /* People celebrate: "We love our President!" */
1981 static int tcp_try_undo_recovery(struct sock
*sk
)
1983 struct tcp_sock
*tp
= tcp_sk(sk
);
1985 if (tcp_may_undo(tp
)) {
1986 /* Happy end! We did not retransmit anything
1987 * or our original transmission succeeded.
1989 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1990 tcp_undo_cwr(sk
, 1);
1991 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1992 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1994 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1995 tp
->undo_marker
= 0;
1997 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1998 /* Hold old state until something *above* high_seq
1999 * is ACKed. For Reno it is MUST to prevent false
2000 * fast retransmits (RFC2582). SACK TCP is safe. */
2001 tcp_moderate_cwnd(tp
);
2004 tcp_set_ca_state(sk
, TCP_CA_Open
);
2008 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2009 static void tcp_try_undo_dsack(struct sock
*sk
)
2011 struct tcp_sock
*tp
= tcp_sk(sk
);
2013 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2014 DBGUNDO(sk
, "D-SACK");
2015 tcp_undo_cwr(sk
, 1);
2016 tp
->undo_marker
= 0;
2017 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2021 /* Undo during fast recovery after partial ACK. */
2023 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2025 struct tcp_sock
*tp
= tcp_sk(sk
);
2026 /* Partial ACK arrived. Force Hoe's retransmit. */
2027 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
2029 if (tcp_may_undo(tp
)) {
2030 /* Plain luck! Hole if filled with delayed
2031 * packet, rather than with a retransmit.
2033 if (tp
->retrans_out
== 0)
2034 tp
->retrans_stamp
= 0;
2036 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2039 tcp_undo_cwr(sk
, 0);
2040 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2042 /* So... Do not make Hoe's retransmit yet.
2043 * If the first packet was delayed, the rest
2044 * ones are most probably delayed as well.
2051 /* Undo during loss recovery after partial ACK. */
2052 static int tcp_try_undo_loss(struct sock
*sk
)
2054 struct tcp_sock
*tp
= tcp_sk(sk
);
2056 if (tcp_may_undo(tp
)) {
2057 struct sk_buff
*skb
;
2058 tcp_for_write_queue(skb
, sk
) {
2059 if (skb
== tcp_send_head(sk
))
2061 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2064 clear_all_retrans_hints(tp
);
2066 DBGUNDO(sk
, "partial loss");
2068 tp
->left_out
= tp
->sacked_out
;
2069 tcp_undo_cwr(sk
, 1);
2070 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2071 inet_csk(sk
)->icsk_retransmits
= 0;
2072 tp
->undo_marker
= 0;
2074 tcp_set_ca_state(sk
, TCP_CA_Open
);
2080 static inline void tcp_complete_cwr(struct sock
*sk
)
2082 struct tcp_sock
*tp
= tcp_sk(sk
);
2083 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2084 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2085 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2088 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2090 struct tcp_sock
*tp
= tcp_sk(sk
);
2092 tcp_sync_left_out(tp
);
2094 if (tp
->retrans_out
== 0)
2095 tp
->retrans_stamp
= 0;
2098 tcp_enter_cwr(sk
, 1);
2100 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2101 int state
= TCP_CA_Open
;
2103 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
2104 state
= TCP_CA_Disorder
;
2106 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2107 tcp_set_ca_state(sk
, state
);
2108 tp
->high_seq
= tp
->snd_nxt
;
2110 tcp_moderate_cwnd(tp
);
2112 tcp_cwnd_down(sk
, flag
);
2116 static void tcp_mtup_probe_failed(struct sock
*sk
)
2118 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2120 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2121 icsk
->icsk_mtup
.probe_size
= 0;
2124 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2126 struct tcp_sock
*tp
= tcp_sk(sk
);
2127 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2129 /* FIXME: breaks with very large cwnd */
2130 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2131 tp
->snd_cwnd
= tp
->snd_cwnd
*
2132 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2133 icsk
->icsk_mtup
.probe_size
;
2134 tp
->snd_cwnd_cnt
= 0;
2135 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2136 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2138 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2139 icsk
->icsk_mtup
.probe_size
= 0;
2140 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2144 /* Process an event, which can update packets-in-flight not trivially.
2145 * Main goal of this function is to calculate new estimate for left_out,
2146 * taking into account both packets sitting in receiver's buffer and
2147 * packets lost by network.
2149 * Besides that it does CWND reduction, when packet loss is detected
2150 * and changes state of machine.
2152 * It does _not_ decide what to send, it is made in function
2153 * tcp_xmit_retransmit_queue().
2156 tcp_fastretrans_alert(struct sock
*sk
, int prior_packets
, int flag
)
2158 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2159 struct tcp_sock
*tp
= tcp_sk(sk
);
2160 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2161 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2162 (tp
->fackets_out
> tp
->reordering
));
2164 /* Some technical things:
2165 * 1. Reno does not count dupacks (sacked_out) automatically. */
2166 if (!tp
->packets_out
)
2168 /* 2. SACK counts snd_fack in packets inaccurately. */
2169 if (tp
->sacked_out
== 0)
2170 tp
->fackets_out
= 0;
2172 /* Now state machine starts.
2173 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2175 tp
->prior_ssthresh
= 0;
2177 /* B. In all the states check for reneging SACKs. */
2178 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2181 /* C. Process data loss notification, provided it is valid. */
2182 if ((flag
&FLAG_DATA_LOST
) &&
2183 before(tp
->snd_una
, tp
->high_seq
) &&
2184 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2185 tp
->fackets_out
> tp
->reordering
) {
2186 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
2187 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2190 /* D. Synchronize left_out to current state. */
2191 tcp_sync_left_out(tp
);
2193 /* E. Check state exit conditions. State can be terminated
2194 * when high_seq is ACKed. */
2195 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2196 BUG_TRAP(tp
->retrans_out
== 0);
2197 tp
->retrans_stamp
= 0;
2198 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2199 switch (icsk
->icsk_ca_state
) {
2201 icsk
->icsk_retransmits
= 0;
2202 if (tcp_try_undo_recovery(sk
))
2207 /* CWR is to be held something *above* high_seq
2208 * is ACKed for CWR bit to reach receiver. */
2209 if (tp
->snd_una
!= tp
->high_seq
) {
2210 tcp_complete_cwr(sk
);
2211 tcp_set_ca_state(sk
, TCP_CA_Open
);
2215 case TCP_CA_Disorder
:
2216 tcp_try_undo_dsack(sk
);
2217 if (!tp
->undo_marker
||
2218 /* For SACK case do not Open to allow to undo
2219 * catching for all duplicate ACKs. */
2220 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2221 tp
->undo_marker
= 0;
2222 tcp_set_ca_state(sk
, TCP_CA_Open
);
2226 case TCP_CA_Recovery
:
2228 tcp_reset_reno_sack(tp
);
2229 if (tcp_try_undo_recovery(sk
))
2231 tcp_complete_cwr(sk
);
2236 /* F. Process state. */
2237 switch (icsk
->icsk_ca_state
) {
2238 case TCP_CA_Recovery
:
2239 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2240 if (IsReno(tp
) && is_dupack
)
2241 tcp_add_reno_sack(sk
);
2243 int acked
= prior_packets
- tp
->packets_out
;
2245 tcp_remove_reno_sacks(sk
, acked
);
2246 do_lost
= tcp_try_undo_partial(sk
, acked
);
2250 if (flag
&FLAG_DATA_ACKED
)
2251 icsk
->icsk_retransmits
= 0;
2252 if (!tcp_try_undo_loss(sk
)) {
2253 tcp_moderate_cwnd(tp
);
2254 tcp_xmit_retransmit_queue(sk
);
2257 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2259 /* Loss is undone; fall through to processing in Open state. */
2262 if (flag
& FLAG_SND_UNA_ADVANCED
)
2263 tcp_reset_reno_sack(tp
);
2265 tcp_add_reno_sack(sk
);
2268 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2269 tcp_try_undo_dsack(sk
);
2271 if (!tcp_time_to_recover(sk
)) {
2272 tcp_try_to_open(sk
, flag
);
2276 /* MTU probe failure: don't reduce cwnd */
2277 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2278 icsk
->icsk_mtup
.probe_size
&&
2279 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2280 tcp_mtup_probe_failed(sk
);
2281 /* Restores the reduction we did in tcp_mtup_probe() */
2283 tcp_simple_retransmit(sk
);
2287 /* Otherwise enter Recovery state */
2290 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2292 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2294 tp
->high_seq
= tp
->snd_nxt
;
2295 tp
->prior_ssthresh
= 0;
2296 tp
->undo_marker
= tp
->snd_una
;
2297 tp
->undo_retrans
= tp
->retrans_out
;
2299 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2300 if (!(flag
&FLAG_ECE
))
2301 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2302 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2303 TCP_ECN_queue_cwr(tp
);
2306 tp
->bytes_acked
= 0;
2307 tp
->snd_cwnd_cnt
= 0;
2308 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2311 if (do_lost
|| tcp_head_timedout(sk
))
2312 tcp_update_scoreboard(sk
);
2313 tcp_cwnd_down(sk
, flag
);
2314 tcp_xmit_retransmit_queue(sk
);
2317 /* Read draft-ietf-tcplw-high-performance before mucking
2318 * with this code. (Supersedes RFC1323)
2320 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2322 /* RTTM Rule: A TSecr value received in a segment is used to
2323 * update the averaged RTT measurement only if the segment
2324 * acknowledges some new data, i.e., only if it advances the
2325 * left edge of the send window.
2327 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2328 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2330 * Changed: reset backoff as soon as we see the first valid sample.
2331 * If we do not, we get strongly overestimated rto. With timestamps
2332 * samples are accepted even from very old segments: f.e., when rtt=1
2333 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2334 * answer arrives rto becomes 120 seconds! If at least one of segments
2335 * in window is lost... Voila. --ANK (010210)
2337 struct tcp_sock
*tp
= tcp_sk(sk
);
2338 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2339 tcp_rtt_estimator(sk
, seq_rtt
);
2341 inet_csk(sk
)->icsk_backoff
= 0;
2345 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2347 /* We don't have a timestamp. Can only use
2348 * packets that are not retransmitted to determine
2349 * rtt estimates. Also, we must not reset the
2350 * backoff for rto until we get a non-retransmitted
2351 * packet. This allows us to deal with a situation
2352 * where the network delay has increased suddenly.
2353 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2356 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2359 tcp_rtt_estimator(sk
, seq_rtt
);
2361 inet_csk(sk
)->icsk_backoff
= 0;
2365 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2368 const struct tcp_sock
*tp
= tcp_sk(sk
);
2369 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2370 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2371 tcp_ack_saw_tstamp(sk
, flag
);
2372 else if (seq_rtt
>= 0)
2373 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2376 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
,
2377 u32 in_flight
, int good
)
2379 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2380 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
, good
);
2381 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2384 /* Restart timer after forward progress on connection.
2385 * RFC2988 recommends to restart timer to now+rto.
2388 static void tcp_ack_packets_out(struct sock
*sk
)
2390 struct tcp_sock
*tp
= tcp_sk(sk
);
2392 if (!tp
->packets_out
) {
2393 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2395 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2399 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2400 __u32 now
, __s32
*seq_rtt
)
2402 struct tcp_sock
*tp
= tcp_sk(sk
);
2403 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2404 __u32 seq
= tp
->snd_una
;
2405 __u32 packets_acked
;
2408 /* If we get here, the whole TSO packet has not been
2411 BUG_ON(!after(scb
->end_seq
, seq
));
2413 packets_acked
= tcp_skb_pcount(skb
);
2414 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2416 packets_acked
-= tcp_skb_pcount(skb
);
2418 if (packets_acked
) {
2419 __u8 sacked
= scb
->sacked
;
2421 acked
|= FLAG_DATA_ACKED
;
2423 if (sacked
& TCPCB_RETRANS
) {
2424 if (sacked
& TCPCB_SACKED_RETRANS
)
2425 tp
->retrans_out
-= packets_acked
;
2426 acked
|= FLAG_RETRANS_DATA_ACKED
;
2428 } else if (*seq_rtt
< 0)
2429 *seq_rtt
= now
- scb
->when
;
2430 if (sacked
& TCPCB_SACKED_ACKED
)
2431 tp
->sacked_out
-= packets_acked
;
2432 if (sacked
& TCPCB_LOST
)
2433 tp
->lost_out
-= packets_acked
;
2434 if (sacked
& TCPCB_URG
) {
2436 !before(seq
, tp
->snd_up
))
2439 } else if (*seq_rtt
< 0)
2440 *seq_rtt
= now
- scb
->when
;
2442 if (tp
->fackets_out
) {
2443 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2444 tp
->fackets_out
-= dval
;
2446 /* hint's skb might be NULL but we don't need to care */
2447 tp
->fastpath_cnt_hint
-= min_t(u32
, packets_acked
,
2448 tp
->fastpath_cnt_hint
);
2449 tp
->packets_out
-= packets_acked
;
2451 BUG_ON(tcp_skb_pcount(skb
) == 0);
2452 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2458 /* Remove acknowledged frames from the retransmission queue. */
2459 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2461 struct tcp_sock
*tp
= tcp_sk(sk
);
2462 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2463 struct sk_buff
*skb
;
2464 __u32 now
= tcp_time_stamp
;
2466 int prior_packets
= tp
->packets_out
;
2468 ktime_t last_ackt
= net_invalid_timestamp();
2470 while ((skb
= tcp_write_queue_head(sk
)) &&
2471 skb
!= tcp_send_head(sk
)) {
2472 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2473 __u8 sacked
= scb
->sacked
;
2475 /* If our packet is before the ack sequence we can
2476 * discard it as it's confirmed to have arrived at
2479 if (after(scb
->end_seq
, tp
->snd_una
)) {
2480 if (tcp_skb_pcount(skb
) > 1 &&
2481 after(tp
->snd_una
, scb
->seq
))
2482 acked
|= tcp_tso_acked(sk
, skb
,
2487 /* Initial outgoing SYN's get put onto the write_queue
2488 * just like anything else we transmit. It is not
2489 * true data, and if we misinform our callers that
2490 * this ACK acks real data, we will erroneously exit
2491 * connection startup slow start one packet too
2492 * quickly. This is severely frowned upon behavior.
2494 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2495 acked
|= FLAG_DATA_ACKED
;
2497 acked
|= FLAG_SYN_ACKED
;
2498 tp
->retrans_stamp
= 0;
2501 /* MTU probing checks */
2502 if (icsk
->icsk_mtup
.probe_size
) {
2503 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2504 tcp_mtup_probe_success(sk
, skb
);
2509 if (sacked
& TCPCB_RETRANS
) {
2510 if (sacked
& TCPCB_SACKED_RETRANS
)
2511 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2512 acked
|= FLAG_RETRANS_DATA_ACKED
;
2514 } else if (seq_rtt
< 0) {
2515 seq_rtt
= now
- scb
->when
;
2516 last_ackt
= skb
->tstamp
;
2518 if (sacked
& TCPCB_SACKED_ACKED
)
2519 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2520 if (sacked
& TCPCB_LOST
)
2521 tp
->lost_out
-= tcp_skb_pcount(skb
);
2522 if (sacked
& TCPCB_URG
) {
2524 !before(scb
->end_seq
, tp
->snd_up
))
2527 } else if (seq_rtt
< 0) {
2528 seq_rtt
= now
- scb
->when
;
2529 last_ackt
= skb
->tstamp
;
2531 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2532 tcp_packets_out_dec(tp
, skb
);
2533 tcp_unlink_write_queue(skb
, sk
);
2534 sk_stream_free_skb(sk
, skb
);
2535 clear_all_retrans_hints(tp
);
2538 if (acked
&FLAG_ACKED
) {
2539 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2540 const struct tcp_congestion_ops
*ca_ops
2541 = inet_csk(sk
)->icsk_ca_ops
;
2543 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2544 tcp_ack_packets_out(sk
);
2546 if (ca_ops
->pkts_acked
) {
2549 /* Is the ACK triggering packet unambiguous? */
2550 if (!(acked
& FLAG_RETRANS_DATA_ACKED
)) {
2551 /* High resolution needed and available? */
2552 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2553 !ktime_equal(last_ackt
,
2554 net_invalid_timestamp()))
2555 rtt_us
= ktime_us_delta(ktime_get_real(),
2557 else if (seq_rtt
> 0)
2558 rtt_us
= jiffies_to_usecs(seq_rtt
);
2561 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2565 #if FASTRETRANS_DEBUG > 0
2566 BUG_TRAP((int)tp
->sacked_out
>= 0);
2567 BUG_TRAP((int)tp
->lost_out
>= 0);
2568 BUG_TRAP((int)tp
->retrans_out
>= 0);
2569 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2570 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2572 printk(KERN_DEBUG
"Leak l=%u %d\n",
2573 tp
->lost_out
, icsk
->icsk_ca_state
);
2576 if (tp
->sacked_out
) {
2577 printk(KERN_DEBUG
"Leak s=%u %d\n",
2578 tp
->sacked_out
, icsk
->icsk_ca_state
);
2581 if (tp
->retrans_out
) {
2582 printk(KERN_DEBUG
"Leak r=%u %d\n",
2583 tp
->retrans_out
, icsk
->icsk_ca_state
);
2584 tp
->retrans_out
= 0;
2588 *seq_rtt_p
= seq_rtt
;
2592 static void tcp_ack_probe(struct sock
*sk
)
2594 const struct tcp_sock
*tp
= tcp_sk(sk
);
2595 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2597 /* Was it a usable window open? */
2599 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2600 tp
->snd_una
+ tp
->snd_wnd
)) {
2601 icsk
->icsk_backoff
= 0;
2602 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2603 /* Socket must be waked up by subsequent tcp_data_snd_check().
2604 * This function is not for random using!
2607 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2608 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2613 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2615 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2616 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2619 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2621 const struct tcp_sock
*tp
= tcp_sk(sk
);
2622 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2623 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2626 /* Check that window update is acceptable.
2627 * The function assumes that snd_una<=ack<=snd_next.
2629 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2630 const u32 ack_seq
, const u32 nwin
)
2632 return (after(ack
, tp
->snd_una
) ||
2633 after(ack_seq
, tp
->snd_wl1
) ||
2634 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2637 /* Update our send window.
2639 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2640 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2642 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2645 struct tcp_sock
*tp
= tcp_sk(sk
);
2647 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2649 if (likely(!tcp_hdr(skb
)->syn
))
2650 nwin
<<= tp
->rx_opt
.snd_wscale
;
2652 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2653 flag
|= FLAG_WIN_UPDATE
;
2654 tcp_update_wl(tp
, ack
, ack_seq
);
2656 if (tp
->snd_wnd
!= nwin
) {
2659 /* Note, it is the only place, where
2660 * fast path is recovered for sending TCP.
2663 tcp_fast_path_check(sk
);
2665 if (nwin
> tp
->max_window
) {
2666 tp
->max_window
= nwin
;
2667 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2677 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2678 * continue in congestion avoidance.
2680 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2682 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2683 tp
->snd_cwnd_cnt
= 0;
2684 TCP_ECN_queue_cwr(tp
);
2685 tcp_moderate_cwnd(tp
);
2688 /* A conservative spurious RTO response algorithm: reduce cwnd using
2689 * rate halving and continue in congestion avoidance.
2691 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
2693 tcp_enter_cwr(sk
, 0);
2696 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
2699 tcp_ratehalving_spur_to_response(sk
);
2701 tcp_undo_cwr(sk
, 1);
2704 /* F-RTO spurious RTO detection algorithm (RFC4138)
2706 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2707 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2708 * window (but not to or beyond highest sequence sent before RTO):
2709 * On First ACK, send two new segments out.
2710 * On Second ACK, RTO was likely spurious. Do spurious response (response
2711 * algorithm is not part of the F-RTO detection algorithm
2712 * given in RFC4138 but can be selected separately).
2713 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2714 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2715 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2716 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2718 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2719 * original window even after we transmit two new data segments.
2722 * on first step, wait until first cumulative ACK arrives, then move to
2723 * the second step. In second step, the next ACK decides.
2725 * F-RTO is implemented (mainly) in four functions:
2726 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2727 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2728 * called when tcp_use_frto() showed green light
2729 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2730 * - tcp_enter_frto_loss() is called if there is not enough evidence
2731 * to prove that the RTO is indeed spurious. It transfers the control
2732 * from F-RTO to the conventional RTO recovery
2734 static int tcp_process_frto(struct sock
*sk
, int flag
)
2736 struct tcp_sock
*tp
= tcp_sk(sk
);
2738 tcp_sync_left_out(tp
);
2740 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2741 if (flag
&FLAG_DATA_ACKED
)
2742 inet_csk(sk
)->icsk_retransmits
= 0;
2744 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2745 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
2749 if (!IsSackFrto() || IsReno(tp
)) {
2750 /* RFC4138 shortcoming in step 2; should also have case c):
2751 * ACK isn't duplicate nor advances window, e.g., opposite dir
2754 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2757 if (!(flag
&FLAG_DATA_ACKED
)) {
2758 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2763 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2764 /* Prevent sending of new data. */
2765 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2766 tcp_packets_in_flight(tp
));
2770 if ((tp
->frto_counter
>= 2) &&
2771 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2772 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2773 /* RFC4138 shortcoming (see comment above) */
2774 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2777 tcp_enter_frto_loss(sk
, 3, flag
);
2782 if (tp
->frto_counter
== 1) {
2783 /* Sending of the next skb must be allowed or no FRTO */
2784 if (!tcp_send_head(sk
) ||
2785 after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2786 tp
->snd_una
+ tp
->snd_wnd
)) {
2787 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3),
2792 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2793 tp
->frto_counter
= 2;
2796 switch (sysctl_tcp_frto_response
) {
2798 tcp_undo_spur_to_response(sk
, flag
);
2801 tcp_conservative_spur_to_response(tp
);
2804 tcp_ratehalving_spur_to_response(sk
);
2807 tp
->frto_counter
= 0;
2812 /* This routine deals with incoming acks, but not outgoing ones. */
2813 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2815 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2816 struct tcp_sock
*tp
= tcp_sk(sk
);
2817 u32 prior_snd_una
= tp
->snd_una
;
2818 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2819 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2820 u32 prior_in_flight
;
2825 /* If the ack is newer than sent or older than previous acks
2826 * then we can probably ignore it.
2828 if (after(ack
, tp
->snd_nxt
))
2829 goto uninteresting_ack
;
2831 if (before(ack
, prior_snd_una
))
2834 if (after(ack
, prior_snd_una
))
2835 flag
|= FLAG_SND_UNA_ADVANCED
;
2837 if (sysctl_tcp_abc
) {
2838 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2839 tp
->bytes_acked
+= ack
- prior_snd_una
;
2840 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2841 /* we assume just one segment left network */
2842 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2845 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2846 /* Window is constant, pure forward advance.
2847 * No more checks are required.
2848 * Note, we use the fact that SND.UNA>=SND.WL2.
2850 tcp_update_wl(tp
, ack
, ack_seq
);
2852 flag
|= FLAG_WIN_UPDATE
;
2854 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2856 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2858 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2861 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2863 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
2865 if (TCP_SKB_CB(skb
)->sacked
)
2866 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2868 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
2871 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2874 /* We passed data and got it acked, remove any soft error
2875 * log. Something worked...
2877 sk
->sk_err_soft
= 0;
2878 tp
->rcv_tstamp
= tcp_time_stamp
;
2879 prior_packets
= tp
->packets_out
;
2883 prior_in_flight
= tcp_packets_in_flight(tp
);
2885 /* See if we can take anything off of the retransmit queue. */
2886 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2888 if (tp
->frto_counter
)
2889 frto_cwnd
= tcp_process_frto(sk
, flag
);
2891 if (tcp_ack_is_dubious(sk
, flag
)) {
2892 /* Advance CWND, if state allows this. */
2893 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
2894 tcp_may_raise_cwnd(sk
, flag
))
2895 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 0);
2896 tcp_fastretrans_alert(sk
, prior_packets
, flag
);
2898 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
2899 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 1);
2902 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2903 dst_confirm(sk
->sk_dst_cache
);
2908 icsk
->icsk_probes_out
= 0;
2910 /* If this ack opens up a zero window, clear backoff. It was
2911 * being used to time the probes, and is probably far higher than
2912 * it needs to be for normal retransmission.
2914 if (tcp_send_head(sk
))
2919 if (TCP_SKB_CB(skb
)->sacked
)
2920 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2923 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2928 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2929 * But, this can also be called on packets in the established flow when
2930 * the fast version below fails.
2932 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2935 struct tcphdr
*th
= tcp_hdr(skb
);
2936 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2938 ptr
= (unsigned char *)(th
+ 1);
2939 opt_rx
->saw_tstamp
= 0;
2941 while (length
> 0) {
2948 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2953 if (opsize
< 2) /* "silly options" */
2955 if (opsize
> length
)
2956 return; /* don't parse partial options */
2959 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2960 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
2962 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2963 in_mss
= opt_rx
->user_mss
;
2964 opt_rx
->mss_clamp
= in_mss
;
2969 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2970 if (sysctl_tcp_window_scaling
) {
2971 __u8 snd_wscale
= *(__u8
*) ptr
;
2972 opt_rx
->wscale_ok
= 1;
2973 if (snd_wscale
> 14) {
2974 if (net_ratelimit())
2975 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2976 "scaling value %d >14 received.\n",
2980 opt_rx
->snd_wscale
= snd_wscale
;
2983 case TCPOPT_TIMESTAMP
:
2984 if (opsize
==TCPOLEN_TIMESTAMP
) {
2985 if ((estab
&& opt_rx
->tstamp_ok
) ||
2986 (!estab
&& sysctl_tcp_timestamps
)) {
2987 opt_rx
->saw_tstamp
= 1;
2988 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
2989 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
2993 case TCPOPT_SACK_PERM
:
2994 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2995 if (sysctl_tcp_sack
) {
2996 opt_rx
->sack_ok
= 1;
2997 tcp_sack_reset(opt_rx
);
3003 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3004 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3006 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3009 #ifdef CONFIG_TCP_MD5SIG
3012 * The MD5 Hash has already been
3013 * checked (see tcp_v{4,6}_do_rcv()).
3025 /* Fast parse options. This hopes to only see timestamps.
3026 * If it is wrong it falls back on tcp_parse_options().
3028 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3029 struct tcp_sock
*tp
)
3031 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3032 tp
->rx_opt
.saw_tstamp
= 0;
3034 } else if (tp
->rx_opt
.tstamp_ok
&&
3035 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3036 __be32
*ptr
= (__be32
*)(th
+ 1);
3037 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3038 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3039 tp
->rx_opt
.saw_tstamp
= 1;
3041 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3043 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3047 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3051 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3053 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3054 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3057 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3059 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3060 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3061 * extra check below makes sure this can only happen
3062 * for pure ACK frames. -DaveM
3064 * Not only, also it occurs for expired timestamps.
3067 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3068 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3069 tcp_store_ts_recent(tp
);
3073 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3075 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3076 * it can pass through stack. So, the following predicate verifies that
3077 * this segment is not used for anything but congestion avoidance or
3078 * fast retransmit. Moreover, we even are able to eliminate most of such
3079 * second order effects, if we apply some small "replay" window (~RTO)
3080 * to timestamp space.
3082 * All these measures still do not guarantee that we reject wrapped ACKs
3083 * on networks with high bandwidth, when sequence space is recycled fastly,
3084 * but it guarantees that such events will be very rare and do not affect
3085 * connection seriously. This doesn't look nice, but alas, PAWS is really
3088 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3089 * states that events when retransmit arrives after original data are rare.
3090 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3091 * the biggest problem on large power networks even with minor reordering.
3092 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3093 * up to bandwidth of 18Gigabit/sec. 8) ]
3096 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3098 struct tcp_sock
*tp
= tcp_sk(sk
);
3099 struct tcphdr
*th
= tcp_hdr(skb
);
3100 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3101 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3103 return (/* 1. Pure ACK with correct sequence number. */
3104 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3106 /* 2. ... and duplicate ACK. */
3107 ack
== tp
->snd_una
&&
3109 /* 3. ... and does not update window. */
3110 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3112 /* 4. ... and sits in replay window. */
3113 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3116 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3118 const struct tcp_sock
*tp
= tcp_sk(sk
);
3119 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3120 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3121 !tcp_disordered_ack(sk
, skb
));
3124 /* Check segment sequence number for validity.
3126 * Segment controls are considered valid, if the segment
3127 * fits to the window after truncation to the window. Acceptability
3128 * of data (and SYN, FIN, of course) is checked separately.
3129 * See tcp_data_queue(), for example.
3131 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3132 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3133 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3134 * (borrowed from freebsd)
3137 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3139 return !before(end_seq
, tp
->rcv_wup
) &&
3140 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3143 /* When we get a reset we do this. */
3144 static void tcp_reset(struct sock
*sk
)
3146 /* We want the right error as BSD sees it (and indeed as we do). */
3147 switch (sk
->sk_state
) {
3149 sk
->sk_err
= ECONNREFUSED
;
3151 case TCP_CLOSE_WAIT
:
3157 sk
->sk_err
= ECONNRESET
;
3160 if (!sock_flag(sk
, SOCK_DEAD
))
3161 sk
->sk_error_report(sk
);
3167 * Process the FIN bit. This now behaves as it is supposed to work
3168 * and the FIN takes effect when it is validly part of sequence
3169 * space. Not before when we get holes.
3171 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3172 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3175 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3176 * close and we go into CLOSING (and later onto TIME-WAIT)
3178 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3180 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3182 struct tcp_sock
*tp
= tcp_sk(sk
);
3184 inet_csk_schedule_ack(sk
);
3186 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3187 sock_set_flag(sk
, SOCK_DONE
);
3189 switch (sk
->sk_state
) {
3191 case TCP_ESTABLISHED
:
3192 /* Move to CLOSE_WAIT */
3193 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3194 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3197 case TCP_CLOSE_WAIT
:
3199 /* Received a retransmission of the FIN, do
3204 /* RFC793: Remain in the LAST-ACK state. */
3208 /* This case occurs when a simultaneous close
3209 * happens, we must ack the received FIN and
3210 * enter the CLOSING state.
3213 tcp_set_state(sk
, TCP_CLOSING
);
3216 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3218 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3221 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3222 * cases we should never reach this piece of code.
3224 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3225 __FUNCTION__
, sk
->sk_state
);
3229 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3230 * Probably, we should reset in this case. For now drop them.
3232 __skb_queue_purge(&tp
->out_of_order_queue
);
3233 if (tp
->rx_opt
.sack_ok
)
3234 tcp_sack_reset(&tp
->rx_opt
);
3235 sk_stream_mem_reclaim(sk
);
3237 if (!sock_flag(sk
, SOCK_DEAD
)) {
3238 sk
->sk_state_change(sk
);
3240 /* Do not send POLL_HUP for half duplex close. */
3241 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3242 sk
->sk_state
== TCP_CLOSE
)
3243 sk_wake_async(sk
, 1, POLL_HUP
);
3245 sk_wake_async(sk
, 1, POLL_IN
);
3249 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3251 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3252 if (before(seq
, sp
->start_seq
))
3253 sp
->start_seq
= seq
;
3254 if (after(end_seq
, sp
->end_seq
))
3255 sp
->end_seq
= end_seq
;
3261 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3263 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3264 if (before(seq
, tp
->rcv_nxt
))
3265 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3267 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3269 tp
->rx_opt
.dsack
= 1;
3270 tp
->duplicate_sack
[0].start_seq
= seq
;
3271 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3272 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3276 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3278 if (!tp
->rx_opt
.dsack
)
3279 tcp_dsack_set(tp
, seq
, end_seq
);
3281 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3284 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3286 struct tcp_sock
*tp
= tcp_sk(sk
);
3288 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3289 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3290 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3291 tcp_enter_quickack_mode(sk
);
3293 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3294 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3296 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3297 end_seq
= tp
->rcv_nxt
;
3298 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3305 /* These routines update the SACK block as out-of-order packets arrive or
3306 * in-order packets close up the sequence space.
3308 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3311 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3312 struct tcp_sack_block
*swalk
= sp
+1;
3314 /* See if the recent change to the first SACK eats into
3315 * or hits the sequence space of other SACK blocks, if so coalesce.
3317 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3318 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3321 /* Zap SWALK, by moving every further SACK up by one slot.
3322 * Decrease num_sacks.
3324 tp
->rx_opt
.num_sacks
--;
3325 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3326 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3330 this_sack
++, swalk
++;
3334 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3338 tmp
= sack1
->start_seq
;
3339 sack1
->start_seq
= sack2
->start_seq
;
3340 sack2
->start_seq
= tmp
;
3342 tmp
= sack1
->end_seq
;
3343 sack1
->end_seq
= sack2
->end_seq
;
3344 sack2
->end_seq
= tmp
;
3347 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3349 struct tcp_sock
*tp
= tcp_sk(sk
);
3350 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3351 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3357 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3358 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3359 /* Rotate this_sack to the first one. */
3360 for (; this_sack
>0; this_sack
--, sp
--)
3361 tcp_sack_swap(sp
, sp
-1);
3363 tcp_sack_maybe_coalesce(tp
);
3368 /* Could not find an adjacent existing SACK, build a new one,
3369 * put it at the front, and shift everyone else down. We
3370 * always know there is at least one SACK present already here.
3372 * If the sack array is full, forget about the last one.
3374 if (this_sack
>= 4) {
3376 tp
->rx_opt
.num_sacks
--;
3379 for (; this_sack
> 0; this_sack
--, sp
--)
3383 /* Build the new head SACK, and we're done. */
3384 sp
->start_seq
= seq
;
3385 sp
->end_seq
= end_seq
;
3386 tp
->rx_opt
.num_sacks
++;
3387 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3390 /* RCV.NXT advances, some SACKs should be eaten. */
3392 static void tcp_sack_remove(struct tcp_sock
*tp
)
3394 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3395 int num_sacks
= tp
->rx_opt
.num_sacks
;
3398 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3399 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3400 tp
->rx_opt
.num_sacks
= 0;
3401 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3405 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3406 /* Check if the start of the sack is covered by RCV.NXT. */
3407 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3410 /* RCV.NXT must cover all the block! */
3411 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3413 /* Zap this SACK, by moving forward any other SACKS. */
3414 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3415 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3422 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3423 tp
->rx_opt
.num_sacks
= num_sacks
;
3424 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3428 /* This one checks to see if we can put data from the
3429 * out_of_order queue into the receive_queue.
3431 static void tcp_ofo_queue(struct sock
*sk
)
3433 struct tcp_sock
*tp
= tcp_sk(sk
);
3434 __u32 dsack_high
= tp
->rcv_nxt
;
3435 struct sk_buff
*skb
;
3437 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3438 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3441 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3442 __u32 dsack
= dsack_high
;
3443 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3444 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3445 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3448 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3449 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3450 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3454 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3455 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3456 TCP_SKB_CB(skb
)->end_seq
);
3458 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3459 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3460 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3461 if (tcp_hdr(skb
)->fin
)
3462 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3466 static int tcp_prune_queue(struct sock
*sk
);
3468 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3470 struct tcphdr
*th
= tcp_hdr(skb
);
3471 struct tcp_sock
*tp
= tcp_sk(sk
);
3474 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3477 __skb_pull(skb
, th
->doff
*4);
3479 TCP_ECN_accept_cwr(tp
, skb
);
3481 if (tp
->rx_opt
.dsack
) {
3482 tp
->rx_opt
.dsack
= 0;
3483 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3484 4 - tp
->rx_opt
.tstamp_ok
);
3487 /* Queue data for delivery to the user.
3488 * Packets in sequence go to the receive queue.
3489 * Out of sequence packets to the out_of_order_queue.
3491 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3492 if (tcp_receive_window(tp
) == 0)
3495 /* Ok. In sequence. In window. */
3496 if (tp
->ucopy
.task
== current
&&
3497 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3498 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3499 int chunk
= min_t(unsigned int, skb
->len
,
3502 __set_current_state(TASK_RUNNING
);
3505 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3506 tp
->ucopy
.len
-= chunk
;
3507 tp
->copied_seq
+= chunk
;
3508 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3509 tcp_rcv_space_adjust(sk
);
3517 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3518 !sk_stream_rmem_schedule(sk
, skb
))) {
3519 if (tcp_prune_queue(sk
) < 0 ||
3520 !sk_stream_rmem_schedule(sk
, skb
))
3523 sk_stream_set_owner_r(skb
, sk
);
3524 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3526 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3528 tcp_event_data_recv(sk
, skb
);
3530 tcp_fin(skb
, sk
, th
);
3532 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3535 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3536 * gap in queue is filled.
3538 if (skb_queue_empty(&tp
->out_of_order_queue
))
3539 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3542 if (tp
->rx_opt
.num_sacks
)
3543 tcp_sack_remove(tp
);
3545 tcp_fast_path_check(sk
);
3549 else if (!sock_flag(sk
, SOCK_DEAD
))
3550 sk
->sk_data_ready(sk
, 0);
3554 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3555 /* A retransmit, 2nd most common case. Force an immediate ack. */
3556 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3557 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3560 tcp_enter_quickack_mode(sk
);
3561 inet_csk_schedule_ack(sk
);
3567 /* Out of window. F.e. zero window probe. */
3568 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3571 tcp_enter_quickack_mode(sk
);
3573 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3574 /* Partial packet, seq < rcv_next < end_seq */
3575 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3576 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3577 TCP_SKB_CB(skb
)->end_seq
);
3579 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3581 /* If window is closed, drop tail of packet. But after
3582 * remembering D-SACK for its head made in previous line.
3584 if (!tcp_receive_window(tp
))
3589 TCP_ECN_check_ce(tp
, skb
);
3591 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3592 !sk_stream_rmem_schedule(sk
, skb
)) {
3593 if (tcp_prune_queue(sk
) < 0 ||
3594 !sk_stream_rmem_schedule(sk
, skb
))
3598 /* Disable header prediction. */
3600 inet_csk_schedule_ack(sk
);
3602 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3603 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3605 sk_stream_set_owner_r(skb
, sk
);
3607 if (!skb_peek(&tp
->out_of_order_queue
)) {
3608 /* Initial out of order segment, build 1 SACK. */
3609 if (tp
->rx_opt
.sack_ok
) {
3610 tp
->rx_opt
.num_sacks
= 1;
3611 tp
->rx_opt
.dsack
= 0;
3612 tp
->rx_opt
.eff_sacks
= 1;
3613 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3614 tp
->selective_acks
[0].end_seq
=
3615 TCP_SKB_CB(skb
)->end_seq
;
3617 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3619 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3620 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3621 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3623 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3624 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3626 if (!tp
->rx_opt
.num_sacks
||
3627 tp
->selective_acks
[0].end_seq
!= seq
)
3630 /* Common case: data arrive in order after hole. */
3631 tp
->selective_acks
[0].end_seq
= end_seq
;
3635 /* Find place to insert this segment. */
3637 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3639 } while ((skb1
= skb1
->prev
) !=
3640 (struct sk_buff
*)&tp
->out_of_order_queue
);
3642 /* Do skb overlap to previous one? */
3643 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3644 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3645 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3646 /* All the bits are present. Drop. */
3648 tcp_dsack_set(tp
, seq
, end_seq
);
3651 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3652 /* Partial overlap. */
3653 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3658 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3660 /* And clean segments covered by new one as whole. */
3661 while ((skb1
= skb
->next
) !=
3662 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3663 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3664 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3665 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3668 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3669 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3674 if (tp
->rx_opt
.sack_ok
)
3675 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3679 /* Collapse contiguous sequence of skbs head..tail with
3680 * sequence numbers start..end.
3681 * Segments with FIN/SYN are not collapsed (only because this
3685 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3686 struct sk_buff
*head
, struct sk_buff
*tail
,
3689 struct sk_buff
*skb
;
3691 /* First, check that queue is collapsible and find
3692 * the point where collapsing can be useful. */
3693 for (skb
= head
; skb
!= tail
; ) {
3694 /* No new bits? It is possible on ofo queue. */
3695 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3696 struct sk_buff
*next
= skb
->next
;
3697 __skb_unlink(skb
, list
);
3699 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3704 /* The first skb to collapse is:
3706 * - bloated or contains data before "start" or
3707 * overlaps to the next one.
3709 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
3710 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3711 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3712 (skb
->next
!= tail
&&
3713 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3716 /* Decided to skip this, advance start seq. */
3717 start
= TCP_SKB_CB(skb
)->end_seq
;
3720 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
3723 while (before(start
, end
)) {
3724 struct sk_buff
*nskb
;
3725 int header
= skb_headroom(skb
);
3726 int copy
= SKB_MAX_ORDER(header
, 0);
3728 /* Too big header? This can happen with IPv6. */
3731 if (end
-start
< copy
)
3733 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3737 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
3738 skb_set_network_header(nskb
, (skb_network_header(skb
) -
3740 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
3742 skb_reserve(nskb
, header
);
3743 memcpy(nskb
->head
, skb
->head
, header
);
3744 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3745 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3746 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3747 sk_stream_set_owner_r(nskb
, sk
);
3749 /* Copy data, releasing collapsed skbs. */
3751 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3752 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3756 size
= min(copy
, size
);
3757 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3759 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3763 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3764 struct sk_buff
*next
= skb
->next
;
3765 __skb_unlink(skb
, list
);
3767 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3770 tcp_hdr(skb
)->syn
||
3778 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3779 * and tcp_collapse() them until all the queue is collapsed.
3781 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3783 struct tcp_sock
*tp
= tcp_sk(sk
);
3784 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3785 struct sk_buff
*head
;
3791 start
= TCP_SKB_CB(skb
)->seq
;
3792 end
= TCP_SKB_CB(skb
)->end_seq
;
3798 /* Segment is terminated when we see gap or when
3799 * we are at the end of all the queue. */
3800 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3801 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3802 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3803 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3804 head
, skb
, start
, end
);
3806 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3808 /* Start new segment */
3809 start
= TCP_SKB_CB(skb
)->seq
;
3810 end
= TCP_SKB_CB(skb
)->end_seq
;
3812 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3813 start
= TCP_SKB_CB(skb
)->seq
;
3814 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3815 end
= TCP_SKB_CB(skb
)->end_seq
;
3820 /* Reduce allocated memory if we can, trying to get
3821 * the socket within its memory limits again.
3823 * Return less than zero if we should start dropping frames
3824 * until the socket owning process reads some of the data
3825 * to stabilize the situation.
3827 static int tcp_prune_queue(struct sock
*sk
)
3829 struct tcp_sock
*tp
= tcp_sk(sk
);
3831 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3833 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3835 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3836 tcp_clamp_window(sk
);
3837 else if (tcp_memory_pressure
)
3838 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3840 tcp_collapse_ofo_queue(sk
);
3841 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3842 sk
->sk_receive_queue
.next
,
3843 (struct sk_buff
*)&sk
->sk_receive_queue
,
3844 tp
->copied_seq
, tp
->rcv_nxt
);
3845 sk_stream_mem_reclaim(sk
);
3847 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3850 /* Collapsing did not help, destructive actions follow.
3851 * This must not ever occur. */
3853 /* First, purge the out_of_order queue. */
3854 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3855 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3856 __skb_queue_purge(&tp
->out_of_order_queue
);
3858 /* Reset SACK state. A conforming SACK implementation will
3859 * do the same at a timeout based retransmit. When a connection
3860 * is in a sad state like this, we care only about integrity
3861 * of the connection not performance.
3863 if (tp
->rx_opt
.sack_ok
)
3864 tcp_sack_reset(&tp
->rx_opt
);
3865 sk_stream_mem_reclaim(sk
);
3868 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3871 /* If we are really being abused, tell the caller to silently
3872 * drop receive data on the floor. It will get retransmitted
3873 * and hopefully then we'll have sufficient space.
3875 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3877 /* Massive buffer overcommit. */
3883 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3884 * As additional protections, we do not touch cwnd in retransmission phases,
3885 * and if application hit its sndbuf limit recently.
3887 void tcp_cwnd_application_limited(struct sock
*sk
)
3889 struct tcp_sock
*tp
= tcp_sk(sk
);
3891 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3892 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3893 /* Limited by application or receiver window. */
3894 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
3895 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
3896 if (win_used
< tp
->snd_cwnd
) {
3897 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3898 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3900 tp
->snd_cwnd_used
= 0;
3902 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3905 static int tcp_should_expand_sndbuf(struct sock
*sk
)
3907 struct tcp_sock
*tp
= tcp_sk(sk
);
3909 /* If the user specified a specific send buffer setting, do
3912 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3915 /* If we are under global TCP memory pressure, do not expand. */
3916 if (tcp_memory_pressure
)
3919 /* If we are under soft global TCP memory pressure, do not expand. */
3920 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3923 /* If we filled the congestion window, do not expand. */
3924 if (tp
->packets_out
>= tp
->snd_cwnd
)
3930 /* When incoming ACK allowed to free some skb from write_queue,
3931 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3932 * on the exit from tcp input handler.
3934 * PROBLEM: sndbuf expansion does not work well with largesend.
3936 static void tcp_new_space(struct sock
*sk
)
3938 struct tcp_sock
*tp
= tcp_sk(sk
);
3940 if (tcp_should_expand_sndbuf(sk
)) {
3941 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3942 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3943 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3944 tp
->reordering
+ 1);
3945 sndmem
*= 2*demanded
;
3946 if (sndmem
> sk
->sk_sndbuf
)
3947 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3948 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3951 sk
->sk_write_space(sk
);
3954 static void tcp_check_space(struct sock
*sk
)
3956 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3957 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3958 if (sk
->sk_socket
&&
3959 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3964 static inline void tcp_data_snd_check(struct sock
*sk
)
3966 tcp_push_pending_frames(sk
);
3967 tcp_check_space(sk
);
3971 * Check if sending an ack is needed.
3973 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3975 struct tcp_sock
*tp
= tcp_sk(sk
);
3977 /* More than one full frame received... */
3978 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3979 /* ... and right edge of window advances far enough.
3980 * (tcp_recvmsg() will send ACK otherwise). Or...
3982 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3983 /* We ACK each frame or... */
3984 tcp_in_quickack_mode(sk
) ||
3985 /* We have out of order data. */
3987 skb_peek(&tp
->out_of_order_queue
))) {
3988 /* Then ack it now */
3991 /* Else, send delayed ack. */
3992 tcp_send_delayed_ack(sk
);
3996 static inline void tcp_ack_snd_check(struct sock
*sk
)
3998 if (!inet_csk_ack_scheduled(sk
)) {
3999 /* We sent a data segment already. */
4002 __tcp_ack_snd_check(sk
, 1);
4006 * This routine is only called when we have urgent data
4007 * signaled. Its the 'slow' part of tcp_urg. It could be
4008 * moved inline now as tcp_urg is only called from one
4009 * place. We handle URGent data wrong. We have to - as
4010 * BSD still doesn't use the correction from RFC961.
4011 * For 1003.1g we should support a new option TCP_STDURG to permit
4012 * either form (or just set the sysctl tcp_stdurg).
4015 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4017 struct tcp_sock
*tp
= tcp_sk(sk
);
4018 u32 ptr
= ntohs(th
->urg_ptr
);
4020 if (ptr
&& !sysctl_tcp_stdurg
)
4022 ptr
+= ntohl(th
->seq
);
4024 /* Ignore urgent data that we've already seen and read. */
4025 if (after(tp
->copied_seq
, ptr
))
4028 /* Do not replay urg ptr.
4030 * NOTE: interesting situation not covered by specs.
4031 * Misbehaving sender may send urg ptr, pointing to segment,
4032 * which we already have in ofo queue. We are not able to fetch
4033 * such data and will stay in TCP_URG_NOTYET until will be eaten
4034 * by recvmsg(). Seems, we are not obliged to handle such wicked
4035 * situations. But it is worth to think about possibility of some
4036 * DoSes using some hypothetical application level deadlock.
4038 if (before(ptr
, tp
->rcv_nxt
))
4041 /* Do we already have a newer (or duplicate) urgent pointer? */
4042 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4045 /* Tell the world about our new urgent pointer. */
4048 /* We may be adding urgent data when the last byte read was
4049 * urgent. To do this requires some care. We cannot just ignore
4050 * tp->copied_seq since we would read the last urgent byte again
4051 * as data, nor can we alter copied_seq until this data arrives
4052 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4054 * NOTE. Double Dutch. Rendering to plain English: author of comment
4055 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4056 * and expect that both A and B disappear from stream. This is _wrong_.
4057 * Though this happens in BSD with high probability, this is occasional.
4058 * Any application relying on this is buggy. Note also, that fix "works"
4059 * only in this artificial test. Insert some normal data between A and B and we will
4060 * decline of BSD again. Verdict: it is better to remove to trap
4063 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4064 !sock_flag(sk
, SOCK_URGINLINE
) &&
4065 tp
->copied_seq
!= tp
->rcv_nxt
) {
4066 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4068 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4069 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4074 tp
->urg_data
= TCP_URG_NOTYET
;
4077 /* Disable header prediction. */
4081 /* This is the 'fast' part of urgent handling. */
4082 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4084 struct tcp_sock
*tp
= tcp_sk(sk
);
4086 /* Check if we get a new urgent pointer - normally not. */
4088 tcp_check_urg(sk
,th
);
4090 /* Do we wait for any urgent data? - normally not... */
4091 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4092 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4095 /* Is the urgent pointer pointing into this packet? */
4096 if (ptr
< skb
->len
) {
4098 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4100 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4101 if (!sock_flag(sk
, SOCK_DEAD
))
4102 sk
->sk_data_ready(sk
, 0);
4107 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4109 struct tcp_sock
*tp
= tcp_sk(sk
);
4110 int chunk
= skb
->len
- hlen
;
4114 if (skb_csum_unnecessary(skb
))
4115 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4117 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4121 tp
->ucopy
.len
-= chunk
;
4122 tp
->copied_seq
+= chunk
;
4123 tcp_rcv_space_adjust(sk
);
4130 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4134 if (sock_owned_by_user(sk
)) {
4136 result
= __tcp_checksum_complete(skb
);
4139 result
= __tcp_checksum_complete(skb
);
4144 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4146 return !skb_csum_unnecessary(skb
) &&
4147 __tcp_checksum_complete_user(sk
, skb
);
4150 #ifdef CONFIG_NET_DMA
4151 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4153 struct tcp_sock
*tp
= tcp_sk(sk
);
4154 int chunk
= skb
->len
- hlen
;
4156 int copied_early
= 0;
4158 if (tp
->ucopy
.wakeup
)
4161 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4162 tp
->ucopy
.dma_chan
= get_softnet_dma();
4164 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4166 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4167 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4172 tp
->ucopy
.dma_cookie
= dma_cookie
;
4175 tp
->ucopy
.len
-= chunk
;
4176 tp
->copied_seq
+= chunk
;
4177 tcp_rcv_space_adjust(sk
);
4179 if ((tp
->ucopy
.len
== 0) ||
4180 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4181 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4182 tp
->ucopy
.wakeup
= 1;
4183 sk
->sk_data_ready(sk
, 0);
4185 } else if (chunk
> 0) {
4186 tp
->ucopy
.wakeup
= 1;
4187 sk
->sk_data_ready(sk
, 0);
4190 return copied_early
;
4192 #endif /* CONFIG_NET_DMA */
4195 * TCP receive function for the ESTABLISHED state.
4197 * It is split into a fast path and a slow path. The fast path is
4199 * - A zero window was announced from us - zero window probing
4200 * is only handled properly in the slow path.
4201 * - Out of order segments arrived.
4202 * - Urgent data is expected.
4203 * - There is no buffer space left
4204 * - Unexpected TCP flags/window values/header lengths are received
4205 * (detected by checking the TCP header against pred_flags)
4206 * - Data is sent in both directions. Fast path only supports pure senders
4207 * or pure receivers (this means either the sequence number or the ack
4208 * value must stay constant)
4209 * - Unexpected TCP option.
4211 * When these conditions are not satisfied it drops into a standard
4212 * receive procedure patterned after RFC793 to handle all cases.
4213 * The first three cases are guaranteed by proper pred_flags setting,
4214 * the rest is checked inline. Fast processing is turned on in
4215 * tcp_data_queue when everything is OK.
4217 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4218 struct tcphdr
*th
, unsigned len
)
4220 struct tcp_sock
*tp
= tcp_sk(sk
);
4223 * Header prediction.
4224 * The code loosely follows the one in the famous
4225 * "30 instruction TCP receive" Van Jacobson mail.
4227 * Van's trick is to deposit buffers into socket queue
4228 * on a device interrupt, to call tcp_recv function
4229 * on the receive process context and checksum and copy
4230 * the buffer to user space. smart...
4232 * Our current scheme is not silly either but we take the
4233 * extra cost of the net_bh soft interrupt processing...
4234 * We do checksum and copy also but from device to kernel.
4237 tp
->rx_opt
.saw_tstamp
= 0;
4239 /* pred_flags is 0xS?10 << 16 + snd_wnd
4240 * if header_prediction is to be made
4241 * 'S' will always be tp->tcp_header_len >> 2
4242 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4243 * turn it off (when there are holes in the receive
4244 * space for instance)
4245 * PSH flag is ignored.
4248 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4249 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4250 int tcp_header_len
= tp
->tcp_header_len
;
4252 /* Timestamp header prediction: tcp_header_len
4253 * is automatically equal to th->doff*4 due to pred_flags
4257 /* Check timestamp */
4258 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4259 __be32
*ptr
= (__be32
*)(th
+ 1);
4261 /* No? Slow path! */
4262 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4263 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4266 tp
->rx_opt
.saw_tstamp
= 1;
4268 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4270 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4272 /* If PAWS failed, check it more carefully in slow path */
4273 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4276 /* DO NOT update ts_recent here, if checksum fails
4277 * and timestamp was corrupted part, it will result
4278 * in a hung connection since we will drop all
4279 * future packets due to the PAWS test.
4283 if (len
<= tcp_header_len
) {
4284 /* Bulk data transfer: sender */
4285 if (len
== tcp_header_len
) {
4286 /* Predicted packet is in window by definition.
4287 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4288 * Hence, check seq<=rcv_wup reduces to:
4290 if (tcp_header_len
==
4291 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4292 tp
->rcv_nxt
== tp
->rcv_wup
)
4293 tcp_store_ts_recent(tp
);
4295 /* We know that such packets are checksummed
4298 tcp_ack(sk
, skb
, 0);
4300 tcp_data_snd_check(sk
);
4302 } else { /* Header too small */
4303 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4308 int copied_early
= 0;
4310 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4311 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4312 #ifdef CONFIG_NET_DMA
4313 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4318 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4319 __set_current_state(TASK_RUNNING
);
4321 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
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
) +
4331 TCPOLEN_TSTAMP_ALIGNED
) &&
4332 tp
->rcv_nxt
== tp
->rcv_wup
)
4333 tcp_store_ts_recent(tp
);
4335 tcp_rcv_rtt_measure_ts(sk
, skb
);
4337 __skb_pull(skb
, tcp_header_len
);
4338 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4339 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4342 tcp_cleanup_rbuf(sk
, skb
->len
);
4345 if (tcp_checksum_complete_user(sk
, skb
))
4348 /* Predicted packet is in window by definition.
4349 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4350 * Hence, check seq<=rcv_wup reduces to:
4352 if (tcp_header_len
==
4353 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4354 tp
->rcv_nxt
== tp
->rcv_wup
)
4355 tcp_store_ts_recent(tp
);
4357 tcp_rcv_rtt_measure_ts(sk
, skb
);
4359 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4362 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4364 /* Bulk data transfer: receiver */
4365 __skb_pull(skb
,tcp_header_len
);
4366 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4367 sk_stream_set_owner_r(skb
, sk
);
4368 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4371 tcp_event_data_recv(sk
, skb
);
4373 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4374 /* Well, only one small jumplet in fast path... */
4375 tcp_ack(sk
, skb
, FLAG_DATA
);
4376 tcp_data_snd_check(sk
);
4377 if (!inet_csk_ack_scheduled(sk
))
4381 __tcp_ack_snd_check(sk
, 0);
4383 #ifdef CONFIG_NET_DMA
4385 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4391 sk
->sk_data_ready(sk
, 0);
4397 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4401 * RFC1323: H1. Apply PAWS check first.
4403 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4404 tcp_paws_discard(sk
, skb
)) {
4406 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4407 tcp_send_dupack(sk
, skb
);
4410 /* Resets are accepted even if PAWS failed.
4412 ts_recent update must be made after we are sure
4413 that the packet is in window.
4418 * Standard slow path.
4421 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4422 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4423 * (RST) segments are validated by checking their SEQ-fields."
4424 * And page 69: "If an incoming segment is not acceptable,
4425 * an acknowledgment should be sent in reply (unless the RST bit
4426 * is set, if so drop the segment and return)".
4429 tcp_send_dupack(sk
, skb
);
4438 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4440 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4441 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4442 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4449 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4451 tcp_rcv_rtt_measure_ts(sk
, skb
);
4453 /* Process urgent data. */
4454 tcp_urg(sk
, skb
, th
);
4456 /* step 7: process the segment text */
4457 tcp_data_queue(sk
, skb
);
4459 tcp_data_snd_check(sk
);
4460 tcp_ack_snd_check(sk
);
4464 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4471 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4472 struct tcphdr
*th
, unsigned len
)
4474 struct tcp_sock
*tp
= tcp_sk(sk
);
4475 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4476 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4478 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4482 * "If the state is SYN-SENT then
4483 * first check the ACK bit
4484 * If the ACK bit is set
4485 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4486 * a reset (unless the RST bit is set, if so drop
4487 * the segment and return)"
4489 * We do not send data with SYN, so that RFC-correct
4492 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4493 goto reset_and_undo
;
4495 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4496 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4498 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4499 goto reset_and_undo
;
4502 /* Now ACK is acceptable.
4504 * "If the RST bit is set
4505 * If the ACK was acceptable then signal the user "error:
4506 * connection reset", drop the segment, enter CLOSED state,
4507 * delete TCB, and return."
4516 * "fifth, if neither of the SYN or RST bits is set then
4517 * drop the segment and return."
4523 goto discard_and_undo
;
4526 * "If the SYN bit is on ...
4527 * are acceptable then ...
4528 * (our SYN has been ACKed), change the connection
4529 * state to ESTABLISHED..."
4532 TCP_ECN_rcv_synack(tp
, th
);
4534 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4535 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4537 /* Ok.. it's good. Set up sequence numbers and
4538 * move to established.
4540 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4541 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4543 /* RFC1323: The window in SYN & SYN/ACK segments is
4546 tp
->snd_wnd
= ntohs(th
->window
);
4547 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4549 if (!tp
->rx_opt
.wscale_ok
) {
4550 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4551 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4554 if (tp
->rx_opt
.saw_tstamp
) {
4555 tp
->rx_opt
.tstamp_ok
= 1;
4556 tp
->tcp_header_len
=
4557 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4558 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4559 tcp_store_ts_recent(tp
);
4561 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4564 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4565 tp
->rx_opt
.sack_ok
|= 2;
4568 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4569 tcp_initialize_rcv_mss(sk
);
4571 /* Remember, tcp_poll() does not lock socket!
4572 * Change state from SYN-SENT only after copied_seq
4573 * is initialized. */
4574 tp
->copied_seq
= tp
->rcv_nxt
;
4576 tcp_set_state(sk
, TCP_ESTABLISHED
);
4578 security_inet_conn_established(sk
, skb
);
4580 /* Make sure socket is routed, for correct metrics. */
4581 icsk
->icsk_af_ops
->rebuild_header(sk
);
4583 tcp_init_metrics(sk
);
4585 tcp_init_congestion_control(sk
);
4587 /* Prevent spurious tcp_cwnd_restart() on first data
4590 tp
->lsndtime
= tcp_time_stamp
;
4592 tcp_init_buffer_space(sk
);
4594 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4595 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4597 if (!tp
->rx_opt
.snd_wscale
)
4598 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4602 if (!sock_flag(sk
, SOCK_DEAD
)) {
4603 sk
->sk_state_change(sk
);
4604 sk_wake_async(sk
, 0, POLL_OUT
);
4607 if (sk
->sk_write_pending
||
4608 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4609 icsk
->icsk_ack
.pingpong
) {
4610 /* Save one ACK. Data will be ready after
4611 * several ticks, if write_pending is set.
4613 * It may be deleted, but with this feature tcpdumps
4614 * look so _wonderfully_ clever, that I was not able
4615 * to stand against the temptation 8) --ANK
4617 inet_csk_schedule_ack(sk
);
4618 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4619 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4620 tcp_incr_quickack(sk
);
4621 tcp_enter_quickack_mode(sk
);
4622 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4623 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4634 /* No ACK in the segment */
4638 * "If the RST bit is set
4640 * Otherwise (no ACK) drop the segment and return."
4643 goto discard_and_undo
;
4647 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4648 goto discard_and_undo
;
4651 /* We see SYN without ACK. It is attempt of
4652 * simultaneous connect with crossed SYNs.
4653 * Particularly, it can be connect to self.
4655 tcp_set_state(sk
, TCP_SYN_RECV
);
4657 if (tp
->rx_opt
.saw_tstamp
) {
4658 tp
->rx_opt
.tstamp_ok
= 1;
4659 tcp_store_ts_recent(tp
);
4660 tp
->tcp_header_len
=
4661 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4663 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4666 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4667 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4669 /* RFC1323: The window in SYN & SYN/ACK segments is
4672 tp
->snd_wnd
= ntohs(th
->window
);
4673 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4674 tp
->max_window
= tp
->snd_wnd
;
4676 TCP_ECN_rcv_syn(tp
, th
);
4679 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4680 tcp_initialize_rcv_mss(sk
);
4683 tcp_send_synack(sk
);
4685 /* Note, we could accept data and URG from this segment.
4686 * There are no obstacles to make this.
4688 * However, if we ignore data in ACKless segments sometimes,
4689 * we have no reasons to accept it sometimes.
4690 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4691 * is not flawless. So, discard packet for sanity.
4692 * Uncomment this return to process the data.
4699 /* "fifth, if neither of the SYN or RST bits is set then
4700 * drop the segment and return."
4704 tcp_clear_options(&tp
->rx_opt
);
4705 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4709 tcp_clear_options(&tp
->rx_opt
);
4710 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4716 * This function implements the receiving procedure of RFC 793 for
4717 * all states except ESTABLISHED and TIME_WAIT.
4718 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4719 * address independent.
4722 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4723 struct tcphdr
*th
, unsigned len
)
4725 struct tcp_sock
*tp
= tcp_sk(sk
);
4726 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4729 tp
->rx_opt
.saw_tstamp
= 0;
4731 switch (sk
->sk_state
) {
4743 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4746 /* Now we have several options: In theory there is
4747 * nothing else in the frame. KA9Q has an option to
4748 * send data with the syn, BSD accepts data with the
4749 * syn up to the [to be] advertised window and
4750 * Solaris 2.1 gives you a protocol error. For now
4751 * we just ignore it, that fits the spec precisely
4752 * and avoids incompatibilities. It would be nice in
4753 * future to drop through and process the data.
4755 * Now that TTCP is starting to be used we ought to
4757 * But, this leaves one open to an easy denial of
4758 * service attack, and SYN cookies can't defend
4759 * against this problem. So, we drop the data
4760 * in the interest of security over speed unless
4761 * it's still in use.
4769 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4773 /* Do step6 onward by hand. */
4774 tcp_urg(sk
, skb
, th
);
4776 tcp_data_snd_check(sk
);
4780 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4781 tcp_paws_discard(sk
, skb
)) {
4783 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4784 tcp_send_dupack(sk
, skb
);
4787 /* Reset is accepted even if it did not pass PAWS. */
4790 /* step 1: check sequence number */
4791 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4793 tcp_send_dupack(sk
, skb
);
4797 /* step 2: check RST bit */
4803 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4805 /* step 3: check security and precedence [ignored] */
4809 * Check for a SYN in window.
4811 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4812 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4817 /* step 5: check the ACK field */
4819 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4821 switch (sk
->sk_state
) {
4824 tp
->copied_seq
= tp
->rcv_nxt
;
4826 tcp_set_state(sk
, TCP_ESTABLISHED
);
4827 sk
->sk_state_change(sk
);
4829 /* Note, that this wakeup is only for marginal
4830 * crossed SYN case. Passively open sockets
4831 * are not waked up, because sk->sk_sleep ==
4832 * NULL and sk->sk_socket == NULL.
4834 if (sk
->sk_socket
) {
4835 sk_wake_async(sk
,0,POLL_OUT
);
4838 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4839 tp
->snd_wnd
= ntohs(th
->window
) <<
4840 tp
->rx_opt
.snd_wscale
;
4841 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4842 TCP_SKB_CB(skb
)->seq
);
4844 /* tcp_ack considers this ACK as duplicate
4845 * and does not calculate rtt.
4846 * Fix it at least with timestamps.
4848 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4850 tcp_ack_saw_tstamp(sk
, 0);
4852 if (tp
->rx_opt
.tstamp_ok
)
4853 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4855 /* Make sure socket is routed, for
4858 icsk
->icsk_af_ops
->rebuild_header(sk
);
4860 tcp_init_metrics(sk
);
4862 tcp_init_congestion_control(sk
);
4864 /* Prevent spurious tcp_cwnd_restart() on
4865 * first data packet.
4867 tp
->lsndtime
= tcp_time_stamp
;
4870 tcp_initialize_rcv_mss(sk
);
4871 tcp_init_buffer_space(sk
);
4872 tcp_fast_path_on(tp
);
4879 if (tp
->snd_una
== tp
->write_seq
) {
4880 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4881 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4882 dst_confirm(sk
->sk_dst_cache
);
4884 if (!sock_flag(sk
, SOCK_DEAD
))
4885 /* Wake up lingering close() */
4886 sk
->sk_state_change(sk
);
4890 if (tp
->linger2
< 0 ||
4891 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4892 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4894 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4898 tmo
= tcp_fin_time(sk
);
4899 if (tmo
> TCP_TIMEWAIT_LEN
) {
4900 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4901 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4902 /* Bad case. We could lose such FIN otherwise.
4903 * It is not a big problem, but it looks confusing
4904 * and not so rare event. We still can lose it now,
4905 * if it spins in bh_lock_sock(), but it is really
4908 inet_csk_reset_keepalive_timer(sk
, tmo
);
4910 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4918 if (tp
->snd_una
== tp
->write_seq
) {
4919 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4925 if (tp
->snd_una
== tp
->write_seq
) {
4926 tcp_update_metrics(sk
);
4935 /* step 6: check the URG bit */
4936 tcp_urg(sk
, skb
, th
);
4938 /* step 7: process the segment text */
4939 switch (sk
->sk_state
) {
4940 case TCP_CLOSE_WAIT
:
4943 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4947 /* RFC 793 says to queue data in these states,
4948 * RFC 1122 says we MUST send a reset.
4949 * BSD 4.4 also does reset.
4951 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4952 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4953 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4954 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4960 case TCP_ESTABLISHED
:
4961 tcp_data_queue(sk
, skb
);
4966 /* tcp_data could move socket to TIME-WAIT */
4967 if (sk
->sk_state
!= TCP_CLOSE
) {
4968 tcp_data_snd_check(sk
);
4969 tcp_ack_snd_check(sk
);
4979 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4980 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4981 EXPORT_SYMBOL(tcp_parse_options
);
4982 EXPORT_SYMBOL(tcp_rcv_established
);
4983 EXPORT_SYMBOL(tcp_rcv_state_process
);
4984 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);