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.
964 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
966 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
967 struct tcp_sock
*tp
= tcp_sk(sk
);
968 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
969 TCP_SKB_CB(ack_skb
)->sacked
);
970 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
971 struct sk_buff
*cached_skb
;
972 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
973 int reord
= tp
->packets_out
;
975 u32 lost_retrans
= 0;
977 int found_dup_sack
= 0;
978 int cached_fack_count
;
980 int first_sack_index
;
982 if (!tp
->sacked_out
) {
984 tp
->highest_sack
= tp
->snd_una
;
986 prior_fackets
= tp
->fackets_out
;
988 /* Check for D-SACK. */
989 if (before(ntohl(sp
[0].start_seq
), TCP_SKB_CB(ack_skb
)->ack_seq
)) {
990 flag
|= FLAG_DSACKING_ACK
;
992 tp
->rx_opt
.sack_ok
|= 4;
993 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
994 } else if (num_sacks
> 1 &&
995 !after(ntohl(sp
[0].end_seq
), ntohl(sp
[1].end_seq
)) &&
996 !before(ntohl(sp
[0].start_seq
), ntohl(sp
[1].start_seq
))) {
997 flag
|= FLAG_DSACKING_ACK
;
999 tp
->rx_opt
.sack_ok
|= 4;
1000 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
1003 /* D-SACK for already forgotten data...
1004 * Do dumb counting. */
1005 if (found_dup_sack
&&
1006 !after(ntohl(sp
[0].end_seq
), prior_snd_una
) &&
1007 after(ntohl(sp
[0].end_seq
), tp
->undo_marker
))
1010 /* Eliminate too old ACKs, but take into
1011 * account more or less fresh ones, they can
1012 * contain valid SACK info.
1014 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1018 * if the only SACK change is the increase of the end_seq of
1019 * the first block then only apply that SACK block
1020 * and use retrans queue hinting otherwise slowpath */
1022 for (i
= 0; i
< num_sacks
; i
++) {
1023 __be32 start_seq
= sp
[i
].start_seq
;
1024 __be32 end_seq
= sp
[i
].end_seq
;
1027 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
1030 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
1031 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1034 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1035 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1037 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1038 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1039 tp
->recv_sack_cache
[i
].start_seq
= 0;
1040 tp
->recv_sack_cache
[i
].end_seq
= 0;
1043 first_sack_index
= 0;
1048 tp
->fastpath_skb_hint
= NULL
;
1050 /* order SACK blocks to allow in order walk of the retrans queue */
1051 for (i
= num_sacks
-1; i
> 0; i
--) {
1052 for (j
= 0; j
< i
; j
++){
1053 if (after(ntohl(sp
[j
].start_seq
),
1054 ntohl(sp
[j
+1].start_seq
))){
1055 struct tcp_sack_block_wire tmp
;
1061 /* Track where the first SACK block goes to */
1062 if (j
== first_sack_index
)
1063 first_sack_index
= j
+1;
1070 /* clear flag as used for different purpose in following code */
1073 /* Use SACK fastpath hint if valid */
1074 cached_skb
= tp
->fastpath_skb_hint
;
1075 cached_fack_count
= tp
->fastpath_cnt_hint
;
1077 cached_skb
= tcp_write_queue_head(sk
);
1078 cached_fack_count
= 0;
1081 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1082 struct sk_buff
*skb
;
1083 __u32 start_seq
= ntohl(sp
->start_seq
);
1084 __u32 end_seq
= ntohl(sp
->end_seq
);
1086 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1089 fack_count
= cached_fack_count
;
1091 /* Event "B" in the comment above. */
1092 if (after(end_seq
, tp
->high_seq
))
1093 flag
|= FLAG_DATA_LOST
;
1095 tcp_for_write_queue_from(skb
, sk
) {
1096 int in_sack
, pcount
;
1099 if (skb
== tcp_send_head(sk
))
1103 cached_fack_count
= fack_count
;
1104 if (i
== first_sack_index
) {
1105 tp
->fastpath_skb_hint
= skb
;
1106 tp
->fastpath_cnt_hint
= fack_count
;
1109 /* The retransmission queue is always in order, so
1110 * we can short-circuit the walk early.
1112 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1115 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1116 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1118 pcount
= tcp_skb_pcount(skb
);
1120 if (pcount
> 1 && !in_sack
&&
1121 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1122 unsigned int pkt_len
;
1124 in_sack
= !after(start_seq
,
1125 TCP_SKB_CB(skb
)->seq
);
1128 pkt_len
= (start_seq
-
1129 TCP_SKB_CB(skb
)->seq
);
1131 pkt_len
= (end_seq
-
1132 TCP_SKB_CB(skb
)->seq
);
1133 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1135 pcount
= tcp_skb_pcount(skb
);
1138 fack_count
+= pcount
;
1140 sacked
= TCP_SKB_CB(skb
)->sacked
;
1142 /* Account D-SACK for retransmitted packet. */
1143 if ((dup_sack
&& in_sack
) &&
1144 (sacked
& TCPCB_RETRANS
) &&
1145 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1148 /* The frame is ACKed. */
1149 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1150 if (sacked
&TCPCB_RETRANS
) {
1151 if ((dup_sack
&& in_sack
) &&
1152 (sacked
&TCPCB_SACKED_ACKED
))
1153 reord
= min(fack_count
, reord
);
1155 /* If it was in a hole, we detected reordering. */
1156 if (fack_count
< prior_fackets
&&
1157 !(sacked
&TCPCB_SACKED_ACKED
))
1158 reord
= min(fack_count
, reord
);
1161 /* Nothing to do; acked frame is about to be dropped. */
1165 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1166 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1167 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1168 lost_retrans
= end_seq
;
1173 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1174 if (sacked
& TCPCB_SACKED_RETRANS
) {
1175 /* If the segment is not tagged as lost,
1176 * we do not clear RETRANS, believing
1177 * that retransmission is still in flight.
1179 if (sacked
& TCPCB_LOST
) {
1180 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1181 tp
->lost_out
-= tcp_skb_pcount(skb
);
1182 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1184 /* clear lost hint */
1185 tp
->retransmit_skb_hint
= NULL
;
1188 /* New sack for not retransmitted frame,
1189 * which was in hole. It is reordering.
1191 if (!(sacked
& TCPCB_RETRANS
) &&
1192 fack_count
< prior_fackets
)
1193 reord
= min(fack_count
, reord
);
1195 if (sacked
& TCPCB_LOST
) {
1196 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1197 tp
->lost_out
-= tcp_skb_pcount(skb
);
1199 /* clear lost hint */
1200 tp
->retransmit_skb_hint
= NULL
;
1202 /* SACK enhanced F-RTO detection.
1203 * Set flag if and only if non-rexmitted
1204 * segments below frto_highmark are
1205 * SACKed (RFC4138; Appendix B).
1206 * Clearing correct due to in-order walk
1208 if (after(end_seq
, tp
->frto_highmark
)) {
1209 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1211 if (!(sacked
& TCPCB_RETRANS
))
1212 flag
|= FLAG_ONLY_ORIG_SACKED
;
1216 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1217 flag
|= FLAG_DATA_SACKED
;
1218 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1220 if (fack_count
> tp
->fackets_out
)
1221 tp
->fackets_out
= fack_count
;
1223 if (after(TCP_SKB_CB(skb
)->seq
,
1225 tp
->highest_sack
= TCP_SKB_CB(skb
)->seq
;
1227 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1228 reord
= min(fack_count
, reord
);
1231 /* D-SACK. We can detect redundant retransmission
1232 * in S|R and plain R frames and clear it.
1233 * undo_retrans is decreased above, L|R frames
1234 * are accounted above as well.
1237 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1238 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1239 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1240 tp
->retransmit_skb_hint
= NULL
;
1245 /* Check for lost retransmit. This superb idea is
1246 * borrowed from "ratehalving". Event "C".
1247 * Later note: FACK people cheated me again 8),
1248 * we have to account for reordering! Ugly,
1251 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1252 struct sk_buff
*skb
;
1254 tcp_for_write_queue(skb
, sk
) {
1255 if (skb
== tcp_send_head(sk
))
1257 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1259 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1261 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1262 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1264 !before(lost_retrans
,
1265 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1267 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1268 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1270 /* clear lost hint */
1271 tp
->retransmit_skb_hint
= NULL
;
1273 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1274 tp
->lost_out
+= tcp_skb_pcount(skb
);
1275 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1276 flag
|= FLAG_DATA_SACKED
;
1277 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1283 tp
->left_out
= tp
->sacked_out
+ tp
->lost_out
;
1285 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
&&
1286 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1287 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1289 #if FASTRETRANS_DEBUG > 0
1290 BUG_TRAP((int)tp
->sacked_out
>= 0);
1291 BUG_TRAP((int)tp
->lost_out
>= 0);
1292 BUG_TRAP((int)tp
->retrans_out
>= 0);
1293 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1298 /* F-RTO can only be used if TCP has never retransmitted anything other than
1299 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1301 int tcp_use_frto(struct sock
*sk
)
1303 const struct tcp_sock
*tp
= tcp_sk(sk
);
1304 struct sk_buff
*skb
;
1306 if (!sysctl_tcp_frto
)
1312 /* Avoid expensive walking of rexmit queue if possible */
1313 if (tp
->retrans_out
> 1)
1316 skb
= tcp_write_queue_head(sk
);
1317 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1318 tcp_for_write_queue_from(skb
, sk
) {
1319 if (skb
== tcp_send_head(sk
))
1321 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1323 /* Short-circuit when first non-SACKed skb has been checked */
1324 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1330 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1331 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1332 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1333 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1334 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1335 * bits are handled if the Loss state is really to be entered (in
1336 * tcp_enter_frto_loss).
1338 * Do like tcp_enter_loss() would; when RTO expires the second time it
1340 * "Reduce ssthresh if it has not yet been made inside this window."
1342 void tcp_enter_frto(struct sock
*sk
)
1344 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1345 struct tcp_sock
*tp
= tcp_sk(sk
);
1346 struct sk_buff
*skb
;
1348 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1349 tp
->snd_una
== tp
->high_seq
||
1350 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1351 !icsk
->icsk_retransmits
)) {
1352 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1353 /* Our state is too optimistic in ssthresh() call because cwnd
1354 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1355 * recovery has not yet completed. Pattern would be this: RTO,
1356 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1358 * RFC4138 should be more specific on what to do, even though
1359 * RTO is quite unlikely to occur after the first Cumulative ACK
1360 * due to back-off and complexity of triggering events ...
1362 if (tp
->frto_counter
) {
1364 stored_cwnd
= tp
->snd_cwnd
;
1366 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1367 tp
->snd_cwnd
= stored_cwnd
;
1369 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1371 /* ... in theory, cong.control module could do "any tricks" in
1372 * ssthresh(), which means that ca_state, lost bits and lost_out
1373 * counter would have to be faked before the call occurs. We
1374 * consider that too expensive, unlikely and hacky, so modules
1375 * using these in ssthresh() must deal these incompatibility
1376 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1378 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1381 tp
->undo_marker
= tp
->snd_una
;
1382 tp
->undo_retrans
= 0;
1384 skb
= tcp_write_queue_head(sk
);
1385 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1386 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1387 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1389 tcp_sync_left_out(tp
);
1391 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1392 * The last condition is necessary at least in tp->frto_counter case.
1394 if (IsSackFrto() && (tp
->frto_counter
||
1395 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1396 after(tp
->high_seq
, tp
->snd_una
)) {
1397 tp
->frto_highmark
= tp
->high_seq
;
1399 tp
->frto_highmark
= tp
->snd_nxt
;
1401 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1402 tp
->high_seq
= tp
->snd_nxt
;
1403 tp
->frto_counter
= 1;
1406 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1407 * which indicates that we should follow the traditional RTO recovery,
1408 * i.e. mark everything lost and do go-back-N retransmission.
1410 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1412 struct tcp_sock
*tp
= tcp_sk(sk
);
1413 struct sk_buff
*skb
;
1418 tp
->fackets_out
= 0;
1419 tp
->retrans_out
= 0;
1421 tcp_for_write_queue(skb
, sk
) {
1422 if (skb
== tcp_send_head(sk
))
1424 cnt
+= tcp_skb_pcount(skb
);
1426 * Count the retransmission made on RTO correctly (only when
1427 * waiting for the first ACK and did not get it)...
1429 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1430 /* For some reason this R-bit might get cleared? */
1431 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1432 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1433 /* ...enter this if branch just for the first segment */
1434 flag
|= FLAG_DATA_ACKED
;
1436 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1438 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
)) {
1440 /* Do not mark those segments lost that were
1441 * forward transmitted after RTO
1443 if (!after(TCP_SKB_CB(skb
)->end_seq
,
1444 tp
->frto_highmark
)) {
1445 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1446 tp
->lost_out
+= tcp_skb_pcount(skb
);
1449 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1450 tp
->fackets_out
= cnt
;
1453 tcp_sync_left_out(tp
);
1455 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1456 tp
->snd_cwnd_cnt
= 0;
1457 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1458 tp
->undo_marker
= 0;
1459 tp
->frto_counter
= 0;
1461 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1462 sysctl_tcp_reordering
);
1463 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1464 tp
->high_seq
= tp
->frto_highmark
;
1465 TCP_ECN_queue_cwr(tp
);
1467 clear_all_retrans_hints(tp
);
1470 void tcp_clear_retrans(struct tcp_sock
*tp
)
1473 tp
->retrans_out
= 0;
1475 tp
->fackets_out
= 0;
1479 tp
->undo_marker
= 0;
1480 tp
->undo_retrans
= 0;
1483 /* Enter Loss state. If "how" is not zero, forget all SACK information
1484 * and reset tags completely, otherwise preserve SACKs. If receiver
1485 * dropped its ofo queue, we will know this due to reneging detection.
1487 void tcp_enter_loss(struct sock
*sk
, int how
)
1489 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1490 struct tcp_sock
*tp
= tcp_sk(sk
);
1491 struct sk_buff
*skb
;
1494 /* Reduce ssthresh if it has not yet been made inside this window. */
1495 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1496 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1497 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1498 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1499 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1502 tp
->snd_cwnd_cnt
= 0;
1503 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1505 tp
->bytes_acked
= 0;
1506 tcp_clear_retrans(tp
);
1508 /* Push undo marker, if it was plain RTO and nothing
1509 * was retransmitted. */
1511 tp
->undo_marker
= tp
->snd_una
;
1513 tcp_for_write_queue(skb
, sk
) {
1514 if (skb
== tcp_send_head(sk
))
1516 cnt
+= tcp_skb_pcount(skb
);
1517 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1518 tp
->undo_marker
= 0;
1519 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1520 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1521 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1522 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1523 tp
->lost_out
+= tcp_skb_pcount(skb
);
1525 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1526 tp
->fackets_out
= cnt
;
1529 tcp_sync_left_out(tp
);
1531 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1532 sysctl_tcp_reordering
);
1533 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1534 tp
->high_seq
= tp
->snd_nxt
;
1535 TCP_ECN_queue_cwr(tp
);
1536 /* Abort FRTO algorithm if one is in progress */
1537 tp
->frto_counter
= 0;
1539 clear_all_retrans_hints(tp
);
1542 static int tcp_check_sack_reneging(struct sock
*sk
)
1544 struct sk_buff
*skb
;
1546 /* If ACK arrived pointing to a remembered SACK,
1547 * it means that our remembered SACKs do not reflect
1548 * real state of receiver i.e.
1549 * receiver _host_ is heavily congested (or buggy).
1550 * Do processing similar to RTO timeout.
1552 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1553 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1554 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1555 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1557 tcp_enter_loss(sk
, 1);
1558 icsk
->icsk_retransmits
++;
1559 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1560 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1561 icsk
->icsk_rto
, TCP_RTO_MAX
);
1567 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1569 return IsReno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1572 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1574 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1577 static inline int tcp_head_timedout(struct sock
*sk
)
1579 struct tcp_sock
*tp
= tcp_sk(sk
);
1581 return tp
->packets_out
&&
1582 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1585 /* Linux NewReno/SACK/FACK/ECN state machine.
1586 * --------------------------------------
1588 * "Open" Normal state, no dubious events, fast path.
1589 * "Disorder" In all the respects it is "Open",
1590 * but requires a bit more attention. It is entered when
1591 * we see some SACKs or dupacks. It is split of "Open"
1592 * mainly to move some processing from fast path to slow one.
1593 * "CWR" CWND was reduced due to some Congestion Notification event.
1594 * It can be ECN, ICMP source quench, local device congestion.
1595 * "Recovery" CWND was reduced, we are fast-retransmitting.
1596 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1598 * tcp_fastretrans_alert() is entered:
1599 * - each incoming ACK, if state is not "Open"
1600 * - when arrived ACK is unusual, namely:
1605 * Counting packets in flight is pretty simple.
1607 * in_flight = packets_out - left_out + retrans_out
1609 * packets_out is SND.NXT-SND.UNA counted in packets.
1611 * retrans_out is number of retransmitted segments.
1613 * left_out is number of segments left network, but not ACKed yet.
1615 * left_out = sacked_out + lost_out
1617 * sacked_out: Packets, which arrived to receiver out of order
1618 * and hence not ACKed. With SACKs this number is simply
1619 * amount of SACKed data. Even without SACKs
1620 * it is easy to give pretty reliable estimate of this number,
1621 * counting duplicate ACKs.
1623 * lost_out: Packets lost by network. TCP has no explicit
1624 * "loss notification" feedback from network (for now).
1625 * It means that this number can be only _guessed_.
1626 * Actually, it is the heuristics to predict lossage that
1627 * distinguishes different algorithms.
1629 * F.e. after RTO, when all the queue is considered as lost,
1630 * lost_out = packets_out and in_flight = retrans_out.
1632 * Essentially, we have now two algorithms counting
1635 * FACK: It is the simplest heuristics. As soon as we decided
1636 * that something is lost, we decide that _all_ not SACKed
1637 * packets until the most forward SACK are lost. I.e.
1638 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1639 * It is absolutely correct estimate, if network does not reorder
1640 * packets. And it loses any connection to reality when reordering
1641 * takes place. We use FACK by default until reordering
1642 * is suspected on the path to this destination.
1644 * NewReno: when Recovery is entered, we assume that one segment
1645 * is lost (classic Reno). While we are in Recovery and
1646 * a partial ACK arrives, we assume that one more packet
1647 * is lost (NewReno). This heuristics are the same in NewReno
1650 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1651 * deflation etc. CWND is real congestion window, never inflated, changes
1652 * only according to classic VJ rules.
1654 * Really tricky (and requiring careful tuning) part of algorithm
1655 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1656 * The first determines the moment _when_ we should reduce CWND and,
1657 * hence, slow down forward transmission. In fact, it determines the moment
1658 * when we decide that hole is caused by loss, rather than by a reorder.
1660 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1661 * holes, caused by lost packets.
1663 * And the most logically complicated part of algorithm is undo
1664 * heuristics. We detect false retransmits due to both too early
1665 * fast retransmit (reordering) and underestimated RTO, analyzing
1666 * timestamps and D-SACKs. When we detect that some segments were
1667 * retransmitted by mistake and CWND reduction was wrong, we undo
1668 * window reduction and abort recovery phase. This logic is hidden
1669 * inside several functions named tcp_try_undo_<something>.
1672 /* This function decides, when we should leave Disordered state
1673 * and enter Recovery phase, reducing congestion window.
1675 * Main question: may we further continue forward transmission
1676 * with the same cwnd?
1678 static int tcp_time_to_recover(struct sock
*sk
)
1680 struct tcp_sock
*tp
= tcp_sk(sk
);
1683 /* Do not perform any recovery during FRTO algorithm */
1684 if (tp
->frto_counter
)
1687 /* Trick#1: The loss is proven. */
1691 /* Not-A-Trick#2 : Classic rule... */
1692 if (tcp_fackets_out(tp
) > tp
->reordering
)
1695 /* Trick#3 : when we use RFC2988 timer restart, fast
1696 * retransmit can be triggered by timeout of queue head.
1698 if (tcp_head_timedout(sk
))
1701 /* Trick#4: It is still not OK... But will it be useful to delay
1704 packets_out
= tp
->packets_out
;
1705 if (packets_out
<= tp
->reordering
&&
1706 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1707 !tcp_may_send_now(sk
)) {
1708 /* We have nothing to send. This connection is limited
1709 * either by receiver window or by application.
1717 /* If we receive more dupacks than we expected counting segments
1718 * in assumption of absent reordering, interpret this as reordering.
1719 * The only another reason could be bug in receiver TCP.
1721 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1723 struct tcp_sock
*tp
= tcp_sk(sk
);
1726 holes
= max(tp
->lost_out
, 1U);
1727 holes
= min(holes
, tp
->packets_out
);
1729 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1730 tp
->sacked_out
= tp
->packets_out
- holes
;
1731 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1735 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1737 static void tcp_add_reno_sack(struct sock
*sk
)
1739 struct tcp_sock
*tp
= tcp_sk(sk
);
1741 tcp_check_reno_reordering(sk
, 0);
1742 tcp_sync_left_out(tp
);
1745 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1747 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1749 struct tcp_sock
*tp
= tcp_sk(sk
);
1752 /* One ACK acked hole. The rest eat duplicate ACKs. */
1753 if (acked
-1 >= tp
->sacked_out
)
1756 tp
->sacked_out
-= acked
-1;
1758 tcp_check_reno_reordering(sk
, acked
);
1759 tcp_sync_left_out(tp
);
1762 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1765 tp
->left_out
= tp
->lost_out
;
1768 /* RFC: This is from the original, I doubt that this is necessary at all:
1769 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1770 * retransmitted past LOST markings in the first place? I'm not fully sure
1771 * about undo and end of connection cases, which can cause R without L?
1773 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
1774 struct sk_buff
*skb
)
1776 if ((tp
->retransmit_skb_hint
!= NULL
) &&
1777 before(TCP_SKB_CB(skb
)->seq
,
1778 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1779 tp
->retransmit_skb_hint
= skb
;
1782 /* Mark head of queue up as lost. */
1783 static void tcp_mark_head_lost(struct sock
*sk
,
1784 int packets
, u32 high_seq
)
1786 struct tcp_sock
*tp
= tcp_sk(sk
);
1787 struct sk_buff
*skb
;
1790 BUG_TRAP(packets
<= tp
->packets_out
);
1791 if (tp
->lost_skb_hint
) {
1792 skb
= tp
->lost_skb_hint
;
1793 cnt
= tp
->lost_cnt_hint
;
1795 skb
= tcp_write_queue_head(sk
);
1799 tcp_for_write_queue_from(skb
, sk
) {
1800 if (skb
== tcp_send_head(sk
))
1802 /* TODO: do this better */
1803 /* this is not the most efficient way to do this... */
1804 tp
->lost_skb_hint
= skb
;
1805 tp
->lost_cnt_hint
= cnt
;
1806 cnt
+= tcp_skb_pcount(skb
);
1807 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1809 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1810 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1811 tp
->lost_out
+= tcp_skb_pcount(skb
);
1812 tcp_verify_retransmit_hint(tp
, skb
);
1815 tcp_sync_left_out(tp
);
1818 /* Account newly detected lost packet(s) */
1820 static void tcp_update_scoreboard(struct sock
*sk
)
1822 struct tcp_sock
*tp
= tcp_sk(sk
);
1825 int lost
= tp
->fackets_out
- tp
->reordering
;
1828 tcp_mark_head_lost(sk
, lost
, tp
->high_seq
);
1830 tcp_mark_head_lost(sk
, 1, tp
->high_seq
);
1833 /* New heuristics: it is possible only after we switched
1834 * to restart timer each time when something is ACKed.
1835 * Hence, we can detect timed out packets during fast
1836 * retransmit without falling to slow start.
1838 if (!IsReno(tp
) && tcp_head_timedout(sk
)) {
1839 struct sk_buff
*skb
;
1841 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1842 : tcp_write_queue_head(sk
);
1844 tcp_for_write_queue_from(skb
, sk
) {
1845 if (skb
== tcp_send_head(sk
))
1847 if (!tcp_skb_timedout(sk
, skb
))
1850 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1851 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1852 tp
->lost_out
+= tcp_skb_pcount(skb
);
1853 tcp_verify_retransmit_hint(tp
, skb
);
1857 tp
->scoreboard_skb_hint
= skb
;
1859 tcp_sync_left_out(tp
);
1863 /* CWND moderation, preventing bursts due to too big ACKs
1864 * in dubious situations.
1866 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1868 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1869 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1870 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1873 /* Lower bound on congestion window is slow start threshold
1874 * unless congestion avoidance choice decides to overide it.
1876 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1878 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1880 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1883 /* Decrease cwnd each second ack. */
1884 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
1886 struct tcp_sock
*tp
= tcp_sk(sk
);
1887 int decr
= tp
->snd_cwnd_cnt
+ 1;
1889 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
1890 (IsReno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
1891 tp
->snd_cwnd_cnt
= decr
&1;
1894 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1895 tp
->snd_cwnd
-= decr
;
1897 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1898 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1902 /* Nothing was retransmitted or returned timestamp is less
1903 * than timestamp of the first retransmission.
1905 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1907 return !tp
->retrans_stamp
||
1908 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1909 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1912 /* Undo procedures. */
1914 #if FASTRETRANS_DEBUG > 1
1915 static void DBGUNDO(struct sock
*sk
, const char *msg
)
1917 struct tcp_sock
*tp
= tcp_sk(sk
);
1918 struct inet_sock
*inet
= inet_sk(sk
);
1920 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1922 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1923 tp
->snd_cwnd
, tp
->left_out
,
1924 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1928 #define DBGUNDO(x...) do { } while (0)
1931 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1933 struct tcp_sock
*tp
= tcp_sk(sk
);
1935 if (tp
->prior_ssthresh
) {
1936 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1938 if (icsk
->icsk_ca_ops
->undo_cwnd
)
1939 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
1941 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
1943 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
1944 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
1945 TCP_ECN_withdraw_cwr(tp
);
1948 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
1950 tcp_moderate_cwnd(tp
);
1951 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1953 /* There is something screwy going on with the retrans hints after
1955 clear_all_retrans_hints(tp
);
1958 static inline int tcp_may_undo(struct tcp_sock
*tp
)
1960 return tp
->undo_marker
&&
1961 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
1964 /* People celebrate: "We love our President!" */
1965 static int tcp_try_undo_recovery(struct sock
*sk
)
1967 struct tcp_sock
*tp
= tcp_sk(sk
);
1969 if (tcp_may_undo(tp
)) {
1970 /* Happy end! We did not retransmit anything
1971 * or our original transmission succeeded.
1973 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
1974 tcp_undo_cwr(sk
, 1);
1975 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
1976 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
1978 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
1979 tp
->undo_marker
= 0;
1981 if (tp
->snd_una
== tp
->high_seq
&& IsReno(tp
)) {
1982 /* Hold old state until something *above* high_seq
1983 * is ACKed. For Reno it is MUST to prevent false
1984 * fast retransmits (RFC2582). SACK TCP is safe. */
1985 tcp_moderate_cwnd(tp
);
1988 tcp_set_ca_state(sk
, TCP_CA_Open
);
1992 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1993 static void tcp_try_undo_dsack(struct sock
*sk
)
1995 struct tcp_sock
*tp
= tcp_sk(sk
);
1997 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
1998 DBGUNDO(sk
, "D-SACK");
1999 tcp_undo_cwr(sk
, 1);
2000 tp
->undo_marker
= 0;
2001 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2005 /* Undo during fast recovery after partial ACK. */
2007 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2009 struct tcp_sock
*tp
= tcp_sk(sk
);
2010 /* Partial ACK arrived. Force Hoe's retransmit. */
2011 int failed
= IsReno(tp
) || tp
->fackets_out
>tp
->reordering
;
2013 if (tcp_may_undo(tp
)) {
2014 /* Plain luck! Hole if filled with delayed
2015 * packet, rather than with a retransmit.
2017 if (tp
->retrans_out
== 0)
2018 tp
->retrans_stamp
= 0;
2020 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2023 tcp_undo_cwr(sk
, 0);
2024 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2026 /* So... Do not make Hoe's retransmit yet.
2027 * If the first packet was delayed, the rest
2028 * ones are most probably delayed as well.
2035 /* Undo during loss recovery after partial ACK. */
2036 static int tcp_try_undo_loss(struct sock
*sk
)
2038 struct tcp_sock
*tp
= tcp_sk(sk
);
2040 if (tcp_may_undo(tp
)) {
2041 struct sk_buff
*skb
;
2042 tcp_for_write_queue(skb
, sk
) {
2043 if (skb
== tcp_send_head(sk
))
2045 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2048 clear_all_retrans_hints(tp
);
2050 DBGUNDO(sk
, "partial loss");
2052 tp
->left_out
= tp
->sacked_out
;
2053 tcp_undo_cwr(sk
, 1);
2054 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2055 inet_csk(sk
)->icsk_retransmits
= 0;
2056 tp
->undo_marker
= 0;
2058 tcp_set_ca_state(sk
, TCP_CA_Open
);
2064 static inline void tcp_complete_cwr(struct sock
*sk
)
2066 struct tcp_sock
*tp
= tcp_sk(sk
);
2067 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2068 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2069 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2072 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2074 struct tcp_sock
*tp
= tcp_sk(sk
);
2076 tcp_sync_left_out(tp
);
2078 if (tp
->retrans_out
== 0)
2079 tp
->retrans_stamp
= 0;
2082 tcp_enter_cwr(sk
, 1);
2084 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2085 int state
= TCP_CA_Open
;
2087 if (tp
->left_out
|| tp
->retrans_out
|| tp
->undo_marker
)
2088 state
= TCP_CA_Disorder
;
2090 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2091 tcp_set_ca_state(sk
, state
);
2092 tp
->high_seq
= tp
->snd_nxt
;
2094 tcp_moderate_cwnd(tp
);
2096 tcp_cwnd_down(sk
, flag
);
2100 static void tcp_mtup_probe_failed(struct sock
*sk
)
2102 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2104 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2105 icsk
->icsk_mtup
.probe_size
= 0;
2108 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2110 struct tcp_sock
*tp
= tcp_sk(sk
);
2111 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2113 /* FIXME: breaks with very large cwnd */
2114 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2115 tp
->snd_cwnd
= tp
->snd_cwnd
*
2116 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2117 icsk
->icsk_mtup
.probe_size
;
2118 tp
->snd_cwnd_cnt
= 0;
2119 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2120 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2122 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2123 icsk
->icsk_mtup
.probe_size
= 0;
2124 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2128 /* Process an event, which can update packets-in-flight not trivially.
2129 * Main goal of this function is to calculate new estimate for left_out,
2130 * taking into account both packets sitting in receiver's buffer and
2131 * packets lost by network.
2133 * Besides that it does CWND reduction, when packet loss is detected
2134 * and changes state of machine.
2136 * It does _not_ decide what to send, it is made in function
2137 * tcp_xmit_retransmit_queue().
2140 tcp_fastretrans_alert(struct sock
*sk
, int prior_packets
, int flag
)
2142 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2143 struct tcp_sock
*tp
= tcp_sk(sk
);
2144 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2145 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2146 (tp
->fackets_out
> tp
->reordering
));
2148 /* Some technical things:
2149 * 1. Reno does not count dupacks (sacked_out) automatically. */
2150 if (!tp
->packets_out
)
2152 /* 2. SACK counts snd_fack in packets inaccurately. */
2153 if (tp
->sacked_out
== 0)
2154 tp
->fackets_out
= 0;
2156 /* Now state machine starts.
2157 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2159 tp
->prior_ssthresh
= 0;
2161 /* B. In all the states check for reneging SACKs. */
2162 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2165 /* C. Process data loss notification, provided it is valid. */
2166 if ((flag
&FLAG_DATA_LOST
) &&
2167 before(tp
->snd_una
, tp
->high_seq
) &&
2168 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2169 tp
->fackets_out
> tp
->reordering
) {
2170 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
2171 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2174 /* D. Synchronize left_out to current state. */
2175 tcp_sync_left_out(tp
);
2177 /* E. Check state exit conditions. State can be terminated
2178 * when high_seq is ACKed. */
2179 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2180 BUG_TRAP(tp
->retrans_out
== 0);
2181 tp
->retrans_stamp
= 0;
2182 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2183 switch (icsk
->icsk_ca_state
) {
2185 icsk
->icsk_retransmits
= 0;
2186 if (tcp_try_undo_recovery(sk
))
2191 /* CWR is to be held something *above* high_seq
2192 * is ACKed for CWR bit to reach receiver. */
2193 if (tp
->snd_una
!= tp
->high_seq
) {
2194 tcp_complete_cwr(sk
);
2195 tcp_set_ca_state(sk
, TCP_CA_Open
);
2199 case TCP_CA_Disorder
:
2200 tcp_try_undo_dsack(sk
);
2201 if (!tp
->undo_marker
||
2202 /* For SACK case do not Open to allow to undo
2203 * catching for all duplicate ACKs. */
2204 IsReno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2205 tp
->undo_marker
= 0;
2206 tcp_set_ca_state(sk
, TCP_CA_Open
);
2210 case TCP_CA_Recovery
:
2212 tcp_reset_reno_sack(tp
);
2213 if (tcp_try_undo_recovery(sk
))
2215 tcp_complete_cwr(sk
);
2220 /* F. Process state. */
2221 switch (icsk
->icsk_ca_state
) {
2222 case TCP_CA_Recovery
:
2223 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2224 if (IsReno(tp
) && is_dupack
)
2225 tcp_add_reno_sack(sk
);
2227 int acked
= prior_packets
- tp
->packets_out
;
2229 tcp_remove_reno_sacks(sk
, acked
);
2230 do_lost
= tcp_try_undo_partial(sk
, acked
);
2234 if (flag
&FLAG_DATA_ACKED
)
2235 icsk
->icsk_retransmits
= 0;
2236 if (!tcp_try_undo_loss(sk
)) {
2237 tcp_moderate_cwnd(tp
);
2238 tcp_xmit_retransmit_queue(sk
);
2241 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2243 /* Loss is undone; fall through to processing in Open state. */
2246 if (flag
& FLAG_SND_UNA_ADVANCED
)
2247 tcp_reset_reno_sack(tp
);
2249 tcp_add_reno_sack(sk
);
2252 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2253 tcp_try_undo_dsack(sk
);
2255 if (!tcp_time_to_recover(sk
)) {
2256 tcp_try_to_open(sk
, flag
);
2260 /* MTU probe failure: don't reduce cwnd */
2261 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2262 icsk
->icsk_mtup
.probe_size
&&
2263 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2264 tcp_mtup_probe_failed(sk
);
2265 /* Restores the reduction we did in tcp_mtup_probe() */
2267 tcp_simple_retransmit(sk
);
2271 /* Otherwise enter Recovery state */
2274 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2276 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2278 tp
->high_seq
= tp
->snd_nxt
;
2279 tp
->prior_ssthresh
= 0;
2280 tp
->undo_marker
= tp
->snd_una
;
2281 tp
->undo_retrans
= tp
->retrans_out
;
2283 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2284 if (!(flag
&FLAG_ECE
))
2285 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2286 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2287 TCP_ECN_queue_cwr(tp
);
2290 tp
->bytes_acked
= 0;
2291 tp
->snd_cwnd_cnt
= 0;
2292 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2295 if (do_lost
|| tcp_head_timedout(sk
))
2296 tcp_update_scoreboard(sk
);
2297 tcp_cwnd_down(sk
, flag
);
2298 tcp_xmit_retransmit_queue(sk
);
2301 /* Read draft-ietf-tcplw-high-performance before mucking
2302 * with this code. (Supersedes RFC1323)
2304 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2306 /* RTTM Rule: A TSecr value received in a segment is used to
2307 * update the averaged RTT measurement only if the segment
2308 * acknowledges some new data, i.e., only if it advances the
2309 * left edge of the send window.
2311 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2312 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2314 * Changed: reset backoff as soon as we see the first valid sample.
2315 * If we do not, we get strongly overestimated rto. With timestamps
2316 * samples are accepted even from very old segments: f.e., when rtt=1
2317 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2318 * answer arrives rto becomes 120 seconds! If at least one of segments
2319 * in window is lost... Voila. --ANK (010210)
2321 struct tcp_sock
*tp
= tcp_sk(sk
);
2322 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2323 tcp_rtt_estimator(sk
, seq_rtt
);
2325 inet_csk(sk
)->icsk_backoff
= 0;
2329 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2331 /* We don't have a timestamp. Can only use
2332 * packets that are not retransmitted to determine
2333 * rtt estimates. Also, we must not reset the
2334 * backoff for rto until we get a non-retransmitted
2335 * packet. This allows us to deal with a situation
2336 * where the network delay has increased suddenly.
2337 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2340 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2343 tcp_rtt_estimator(sk
, seq_rtt
);
2345 inet_csk(sk
)->icsk_backoff
= 0;
2349 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2352 const struct tcp_sock
*tp
= tcp_sk(sk
);
2353 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2354 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2355 tcp_ack_saw_tstamp(sk
, flag
);
2356 else if (seq_rtt
>= 0)
2357 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2360 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
,
2361 u32 in_flight
, int good
)
2363 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2364 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
, good
);
2365 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2368 /* Restart timer after forward progress on connection.
2369 * RFC2988 recommends to restart timer to now+rto.
2372 static void tcp_ack_packets_out(struct sock
*sk
)
2374 struct tcp_sock
*tp
= tcp_sk(sk
);
2376 if (!tp
->packets_out
) {
2377 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2379 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2383 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2384 __u32 now
, __s32
*seq_rtt
)
2386 struct tcp_sock
*tp
= tcp_sk(sk
);
2387 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2388 __u32 seq
= tp
->snd_una
;
2389 __u32 packets_acked
;
2392 /* If we get here, the whole TSO packet has not been
2395 BUG_ON(!after(scb
->end_seq
, seq
));
2397 packets_acked
= tcp_skb_pcount(skb
);
2398 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2400 packets_acked
-= tcp_skb_pcount(skb
);
2402 if (packets_acked
) {
2403 __u8 sacked
= scb
->sacked
;
2405 acked
|= FLAG_DATA_ACKED
;
2407 if (sacked
& TCPCB_RETRANS
) {
2408 if (sacked
& TCPCB_SACKED_RETRANS
)
2409 tp
->retrans_out
-= packets_acked
;
2410 acked
|= FLAG_RETRANS_DATA_ACKED
;
2412 } else if (*seq_rtt
< 0)
2413 *seq_rtt
= now
- scb
->when
;
2414 if (sacked
& TCPCB_SACKED_ACKED
)
2415 tp
->sacked_out
-= packets_acked
;
2416 if (sacked
& TCPCB_LOST
)
2417 tp
->lost_out
-= packets_acked
;
2418 if (sacked
& TCPCB_URG
) {
2420 !before(seq
, tp
->snd_up
))
2423 } else if (*seq_rtt
< 0)
2424 *seq_rtt
= now
- scb
->when
;
2426 if (tp
->fackets_out
) {
2427 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2428 tp
->fackets_out
-= dval
;
2430 /* hint's skb might be NULL but we don't need to care */
2431 tp
->fastpath_cnt_hint
-= min_t(u32
, packets_acked
,
2432 tp
->fastpath_cnt_hint
);
2433 tp
->packets_out
-= packets_acked
;
2435 BUG_ON(tcp_skb_pcount(skb
) == 0);
2436 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2442 /* Remove acknowledged frames from the retransmission queue. */
2443 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2445 struct tcp_sock
*tp
= tcp_sk(sk
);
2446 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2447 struct sk_buff
*skb
;
2448 __u32 now
= tcp_time_stamp
;
2450 int prior_packets
= tp
->packets_out
;
2452 ktime_t last_ackt
= net_invalid_timestamp();
2454 while ((skb
= tcp_write_queue_head(sk
)) &&
2455 skb
!= tcp_send_head(sk
)) {
2456 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2457 __u8 sacked
= scb
->sacked
;
2459 /* If our packet is before the ack sequence we can
2460 * discard it as it's confirmed to have arrived at
2463 if (after(scb
->end_seq
, tp
->snd_una
)) {
2464 if (tcp_skb_pcount(skb
) > 1 &&
2465 after(tp
->snd_una
, scb
->seq
))
2466 acked
|= tcp_tso_acked(sk
, skb
,
2471 /* Initial outgoing SYN's get put onto the write_queue
2472 * just like anything else we transmit. It is not
2473 * true data, and if we misinform our callers that
2474 * this ACK acks real data, we will erroneously exit
2475 * connection startup slow start one packet too
2476 * quickly. This is severely frowned upon behavior.
2478 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2479 acked
|= FLAG_DATA_ACKED
;
2481 acked
|= FLAG_SYN_ACKED
;
2482 tp
->retrans_stamp
= 0;
2485 /* MTU probing checks */
2486 if (icsk
->icsk_mtup
.probe_size
) {
2487 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2488 tcp_mtup_probe_success(sk
, skb
);
2493 if (sacked
& TCPCB_RETRANS
) {
2494 if (sacked
& TCPCB_SACKED_RETRANS
)
2495 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2496 acked
|= FLAG_RETRANS_DATA_ACKED
;
2498 } else if (seq_rtt
< 0) {
2499 seq_rtt
= now
- scb
->when
;
2500 last_ackt
= skb
->tstamp
;
2502 if (sacked
& TCPCB_SACKED_ACKED
)
2503 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2504 if (sacked
& TCPCB_LOST
)
2505 tp
->lost_out
-= tcp_skb_pcount(skb
);
2506 if (sacked
& TCPCB_URG
) {
2508 !before(scb
->end_seq
, tp
->snd_up
))
2511 } else if (seq_rtt
< 0) {
2512 seq_rtt
= now
- scb
->when
;
2513 last_ackt
= skb
->tstamp
;
2515 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2516 tcp_packets_out_dec(tp
, skb
);
2517 tcp_unlink_write_queue(skb
, sk
);
2518 sk_stream_free_skb(sk
, skb
);
2519 clear_all_retrans_hints(tp
);
2522 if (acked
&FLAG_ACKED
) {
2523 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2524 const struct tcp_congestion_ops
*ca_ops
2525 = inet_csk(sk
)->icsk_ca_ops
;
2527 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2528 tcp_ack_packets_out(sk
);
2530 if (ca_ops
->pkts_acked
) {
2533 /* Is the ACK triggering packet unambiguous? */
2534 if (!(acked
& FLAG_RETRANS_DATA_ACKED
)) {
2535 /* High resolution needed and available? */
2536 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2537 !ktime_equal(last_ackt
,
2538 net_invalid_timestamp()))
2539 rtt_us
= ktime_us_delta(ktime_get_real(),
2541 else if (seq_rtt
> 0)
2542 rtt_us
= jiffies_to_usecs(seq_rtt
);
2545 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2549 #if FASTRETRANS_DEBUG > 0
2550 BUG_TRAP((int)tp
->sacked_out
>= 0);
2551 BUG_TRAP((int)tp
->lost_out
>= 0);
2552 BUG_TRAP((int)tp
->retrans_out
>= 0);
2553 if (!tp
->packets_out
&& tp
->rx_opt
.sack_ok
) {
2554 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2556 printk(KERN_DEBUG
"Leak l=%u %d\n",
2557 tp
->lost_out
, icsk
->icsk_ca_state
);
2560 if (tp
->sacked_out
) {
2561 printk(KERN_DEBUG
"Leak s=%u %d\n",
2562 tp
->sacked_out
, icsk
->icsk_ca_state
);
2565 if (tp
->retrans_out
) {
2566 printk(KERN_DEBUG
"Leak r=%u %d\n",
2567 tp
->retrans_out
, icsk
->icsk_ca_state
);
2568 tp
->retrans_out
= 0;
2572 *seq_rtt_p
= seq_rtt
;
2576 static void tcp_ack_probe(struct sock
*sk
)
2578 const struct tcp_sock
*tp
= tcp_sk(sk
);
2579 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2581 /* Was it a usable window open? */
2583 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2584 tp
->snd_una
+ tp
->snd_wnd
)) {
2585 icsk
->icsk_backoff
= 0;
2586 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2587 /* Socket must be waked up by subsequent tcp_data_snd_check().
2588 * This function is not for random using!
2591 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2592 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2597 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2599 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2600 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2603 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2605 const struct tcp_sock
*tp
= tcp_sk(sk
);
2606 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2607 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2610 /* Check that window update is acceptable.
2611 * The function assumes that snd_una<=ack<=snd_next.
2613 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2614 const u32 ack_seq
, const u32 nwin
)
2616 return (after(ack
, tp
->snd_una
) ||
2617 after(ack_seq
, tp
->snd_wl1
) ||
2618 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2621 /* Update our send window.
2623 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2624 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2626 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2629 struct tcp_sock
*tp
= tcp_sk(sk
);
2631 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2633 if (likely(!tcp_hdr(skb
)->syn
))
2634 nwin
<<= tp
->rx_opt
.snd_wscale
;
2636 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2637 flag
|= FLAG_WIN_UPDATE
;
2638 tcp_update_wl(tp
, ack
, ack_seq
);
2640 if (tp
->snd_wnd
!= nwin
) {
2643 /* Note, it is the only place, where
2644 * fast path is recovered for sending TCP.
2647 tcp_fast_path_check(sk
);
2649 if (nwin
> tp
->max_window
) {
2650 tp
->max_window
= nwin
;
2651 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2661 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2662 * continue in congestion avoidance.
2664 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2666 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2667 tp
->snd_cwnd_cnt
= 0;
2668 TCP_ECN_queue_cwr(tp
);
2669 tcp_moderate_cwnd(tp
);
2672 /* A conservative spurious RTO response algorithm: reduce cwnd using
2673 * rate halving and continue in congestion avoidance.
2675 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
2677 tcp_enter_cwr(sk
, 0);
2680 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
2683 tcp_ratehalving_spur_to_response(sk
);
2685 tcp_undo_cwr(sk
, 1);
2688 /* F-RTO spurious RTO detection algorithm (RFC4138)
2690 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2691 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2692 * window (but not to or beyond highest sequence sent before RTO):
2693 * On First ACK, send two new segments out.
2694 * On Second ACK, RTO was likely spurious. Do spurious response (response
2695 * algorithm is not part of the F-RTO detection algorithm
2696 * given in RFC4138 but can be selected separately).
2697 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2698 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2699 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2700 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2702 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2703 * original window even after we transmit two new data segments.
2706 * on first step, wait until first cumulative ACK arrives, then move to
2707 * the second step. In second step, the next ACK decides.
2709 * F-RTO is implemented (mainly) in four functions:
2710 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2711 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2712 * called when tcp_use_frto() showed green light
2713 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2714 * - tcp_enter_frto_loss() is called if there is not enough evidence
2715 * to prove that the RTO is indeed spurious. It transfers the control
2716 * from F-RTO to the conventional RTO recovery
2718 static int tcp_process_frto(struct sock
*sk
, int flag
)
2720 struct tcp_sock
*tp
= tcp_sk(sk
);
2722 tcp_sync_left_out(tp
);
2724 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2725 if (flag
&FLAG_DATA_ACKED
)
2726 inet_csk(sk
)->icsk_retransmits
= 0;
2728 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2729 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
2733 if (!IsSackFrto() || IsReno(tp
)) {
2734 /* RFC4138 shortcoming in step 2; should also have case c):
2735 * ACK isn't duplicate nor advances window, e.g., opposite dir
2738 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2741 if (!(flag
&FLAG_DATA_ACKED
)) {
2742 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2747 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2748 /* Prevent sending of new data. */
2749 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2750 tcp_packets_in_flight(tp
));
2754 if ((tp
->frto_counter
>= 2) &&
2755 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2756 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2757 /* RFC4138 shortcoming (see comment above) */
2758 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2761 tcp_enter_frto_loss(sk
, 3, flag
);
2766 if (tp
->frto_counter
== 1) {
2767 /* Sending of the next skb must be allowed or no FRTO */
2768 if (!tcp_send_head(sk
) ||
2769 after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2770 tp
->snd_una
+ tp
->snd_wnd
)) {
2771 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3),
2776 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2777 tp
->frto_counter
= 2;
2780 switch (sysctl_tcp_frto_response
) {
2782 tcp_undo_spur_to_response(sk
, flag
);
2785 tcp_conservative_spur_to_response(tp
);
2788 tcp_ratehalving_spur_to_response(sk
);
2791 tp
->frto_counter
= 0;
2796 /* This routine deals with incoming acks, but not outgoing ones. */
2797 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2799 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2800 struct tcp_sock
*tp
= tcp_sk(sk
);
2801 u32 prior_snd_una
= tp
->snd_una
;
2802 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2803 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2804 u32 prior_in_flight
;
2809 /* If the ack is newer than sent or older than previous acks
2810 * then we can probably ignore it.
2812 if (after(ack
, tp
->snd_nxt
))
2813 goto uninteresting_ack
;
2815 if (before(ack
, prior_snd_una
))
2818 if (after(ack
, prior_snd_una
))
2819 flag
|= FLAG_SND_UNA_ADVANCED
;
2821 if (sysctl_tcp_abc
) {
2822 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2823 tp
->bytes_acked
+= ack
- prior_snd_una
;
2824 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2825 /* we assume just one segment left network */
2826 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2829 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2830 /* Window is constant, pure forward advance.
2831 * No more checks are required.
2832 * Note, we use the fact that SND.UNA>=SND.WL2.
2834 tcp_update_wl(tp
, ack
, ack_seq
);
2836 flag
|= FLAG_WIN_UPDATE
;
2838 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2840 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2842 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2845 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2847 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
2849 if (TCP_SKB_CB(skb
)->sacked
)
2850 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2852 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
2855 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2858 /* We passed data and got it acked, remove any soft error
2859 * log. Something worked...
2861 sk
->sk_err_soft
= 0;
2862 tp
->rcv_tstamp
= tcp_time_stamp
;
2863 prior_packets
= tp
->packets_out
;
2867 prior_in_flight
= tcp_packets_in_flight(tp
);
2869 /* See if we can take anything off of the retransmit queue. */
2870 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2872 if (tp
->frto_counter
)
2873 frto_cwnd
= tcp_process_frto(sk
, flag
);
2875 if (tcp_ack_is_dubious(sk
, flag
)) {
2876 /* Advance CWND, if state allows this. */
2877 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
2878 tcp_may_raise_cwnd(sk
, flag
))
2879 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 0);
2880 tcp_fastretrans_alert(sk
, prior_packets
, flag
);
2882 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
2883 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 1);
2886 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2887 dst_confirm(sk
->sk_dst_cache
);
2892 icsk
->icsk_probes_out
= 0;
2894 /* If this ack opens up a zero window, clear backoff. It was
2895 * being used to time the probes, and is probably far higher than
2896 * it needs to be for normal retransmission.
2898 if (tcp_send_head(sk
))
2903 if (TCP_SKB_CB(skb
)->sacked
)
2904 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2907 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2912 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2913 * But, this can also be called on packets in the established flow when
2914 * the fast version below fails.
2916 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2919 struct tcphdr
*th
= tcp_hdr(skb
);
2920 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2922 ptr
= (unsigned char *)(th
+ 1);
2923 opt_rx
->saw_tstamp
= 0;
2925 while (length
> 0) {
2932 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2937 if (opsize
< 2) /* "silly options" */
2939 if (opsize
> length
)
2940 return; /* don't parse partial options */
2943 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
2944 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
2946 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
2947 in_mss
= opt_rx
->user_mss
;
2948 opt_rx
->mss_clamp
= in_mss
;
2953 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
2954 if (sysctl_tcp_window_scaling
) {
2955 __u8 snd_wscale
= *(__u8
*) ptr
;
2956 opt_rx
->wscale_ok
= 1;
2957 if (snd_wscale
> 14) {
2958 if (net_ratelimit())
2959 printk(KERN_INFO
"tcp_parse_options: Illegal window "
2960 "scaling value %d >14 received.\n",
2964 opt_rx
->snd_wscale
= snd_wscale
;
2967 case TCPOPT_TIMESTAMP
:
2968 if (opsize
==TCPOLEN_TIMESTAMP
) {
2969 if ((estab
&& opt_rx
->tstamp_ok
) ||
2970 (!estab
&& sysctl_tcp_timestamps
)) {
2971 opt_rx
->saw_tstamp
= 1;
2972 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
2973 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
2977 case TCPOPT_SACK_PERM
:
2978 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
2979 if (sysctl_tcp_sack
) {
2980 opt_rx
->sack_ok
= 1;
2981 tcp_sack_reset(opt_rx
);
2987 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
2988 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
2990 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
2993 #ifdef CONFIG_TCP_MD5SIG
2996 * The MD5 Hash has already been
2997 * checked (see tcp_v{4,6}_do_rcv()).
3009 /* Fast parse options. This hopes to only see timestamps.
3010 * If it is wrong it falls back on tcp_parse_options().
3012 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3013 struct tcp_sock
*tp
)
3015 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3016 tp
->rx_opt
.saw_tstamp
= 0;
3018 } else if (tp
->rx_opt
.tstamp_ok
&&
3019 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3020 __be32
*ptr
= (__be32
*)(th
+ 1);
3021 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3022 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3023 tp
->rx_opt
.saw_tstamp
= 1;
3025 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3027 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3031 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3035 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3037 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3038 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3041 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3043 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3044 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3045 * extra check below makes sure this can only happen
3046 * for pure ACK frames. -DaveM
3048 * Not only, also it occurs for expired timestamps.
3051 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3052 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3053 tcp_store_ts_recent(tp
);
3057 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3059 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3060 * it can pass through stack. So, the following predicate verifies that
3061 * this segment is not used for anything but congestion avoidance or
3062 * fast retransmit. Moreover, we even are able to eliminate most of such
3063 * second order effects, if we apply some small "replay" window (~RTO)
3064 * to timestamp space.
3066 * All these measures still do not guarantee that we reject wrapped ACKs
3067 * on networks with high bandwidth, when sequence space is recycled fastly,
3068 * but it guarantees that such events will be very rare and do not affect
3069 * connection seriously. This doesn't look nice, but alas, PAWS is really
3072 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3073 * states that events when retransmit arrives after original data are rare.
3074 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3075 * the biggest problem on large power networks even with minor reordering.
3076 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3077 * up to bandwidth of 18Gigabit/sec. 8) ]
3080 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3082 struct tcp_sock
*tp
= tcp_sk(sk
);
3083 struct tcphdr
*th
= tcp_hdr(skb
);
3084 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3085 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3087 return (/* 1. Pure ACK with correct sequence number. */
3088 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3090 /* 2. ... and duplicate ACK. */
3091 ack
== tp
->snd_una
&&
3093 /* 3. ... and does not update window. */
3094 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3096 /* 4. ... and sits in replay window. */
3097 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3100 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3102 const struct tcp_sock
*tp
= tcp_sk(sk
);
3103 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3104 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3105 !tcp_disordered_ack(sk
, skb
));
3108 /* Check segment sequence number for validity.
3110 * Segment controls are considered valid, if the segment
3111 * fits to the window after truncation to the window. Acceptability
3112 * of data (and SYN, FIN, of course) is checked separately.
3113 * See tcp_data_queue(), for example.
3115 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3116 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3117 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3118 * (borrowed from freebsd)
3121 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3123 return !before(end_seq
, tp
->rcv_wup
) &&
3124 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3127 /* When we get a reset we do this. */
3128 static void tcp_reset(struct sock
*sk
)
3130 /* We want the right error as BSD sees it (and indeed as we do). */
3131 switch (sk
->sk_state
) {
3133 sk
->sk_err
= ECONNREFUSED
;
3135 case TCP_CLOSE_WAIT
:
3141 sk
->sk_err
= ECONNRESET
;
3144 if (!sock_flag(sk
, SOCK_DEAD
))
3145 sk
->sk_error_report(sk
);
3151 * Process the FIN bit. This now behaves as it is supposed to work
3152 * and the FIN takes effect when it is validly part of sequence
3153 * space. Not before when we get holes.
3155 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3156 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3159 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3160 * close and we go into CLOSING (and later onto TIME-WAIT)
3162 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3164 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3166 struct tcp_sock
*tp
= tcp_sk(sk
);
3168 inet_csk_schedule_ack(sk
);
3170 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3171 sock_set_flag(sk
, SOCK_DONE
);
3173 switch (sk
->sk_state
) {
3175 case TCP_ESTABLISHED
:
3176 /* Move to CLOSE_WAIT */
3177 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3178 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3181 case TCP_CLOSE_WAIT
:
3183 /* Received a retransmission of the FIN, do
3188 /* RFC793: Remain in the LAST-ACK state. */
3192 /* This case occurs when a simultaneous close
3193 * happens, we must ack the received FIN and
3194 * enter the CLOSING state.
3197 tcp_set_state(sk
, TCP_CLOSING
);
3200 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3202 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3205 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3206 * cases we should never reach this piece of code.
3208 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3209 __FUNCTION__
, sk
->sk_state
);
3213 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3214 * Probably, we should reset in this case. For now drop them.
3216 __skb_queue_purge(&tp
->out_of_order_queue
);
3217 if (tp
->rx_opt
.sack_ok
)
3218 tcp_sack_reset(&tp
->rx_opt
);
3219 sk_stream_mem_reclaim(sk
);
3221 if (!sock_flag(sk
, SOCK_DEAD
)) {
3222 sk
->sk_state_change(sk
);
3224 /* Do not send POLL_HUP for half duplex close. */
3225 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3226 sk
->sk_state
== TCP_CLOSE
)
3227 sk_wake_async(sk
, 1, POLL_HUP
);
3229 sk_wake_async(sk
, 1, POLL_IN
);
3233 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3235 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3236 if (before(seq
, sp
->start_seq
))
3237 sp
->start_seq
= seq
;
3238 if (after(end_seq
, sp
->end_seq
))
3239 sp
->end_seq
= end_seq
;
3245 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3247 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3248 if (before(seq
, tp
->rcv_nxt
))
3249 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3251 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3253 tp
->rx_opt
.dsack
= 1;
3254 tp
->duplicate_sack
[0].start_seq
= seq
;
3255 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3256 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3260 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3262 if (!tp
->rx_opt
.dsack
)
3263 tcp_dsack_set(tp
, seq
, end_seq
);
3265 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3268 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3270 struct tcp_sock
*tp
= tcp_sk(sk
);
3272 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3273 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3274 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3275 tcp_enter_quickack_mode(sk
);
3277 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_dsack
) {
3278 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3280 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3281 end_seq
= tp
->rcv_nxt
;
3282 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3289 /* These routines update the SACK block as out-of-order packets arrive or
3290 * in-order packets close up the sequence space.
3292 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3295 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3296 struct tcp_sack_block
*swalk
= sp
+1;
3298 /* See if the recent change to the first SACK eats into
3299 * or hits the sequence space of other SACK blocks, if so coalesce.
3301 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3302 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3305 /* Zap SWALK, by moving every further SACK up by one slot.
3306 * Decrease num_sacks.
3308 tp
->rx_opt
.num_sacks
--;
3309 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3310 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3314 this_sack
++, swalk
++;
3318 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3322 tmp
= sack1
->start_seq
;
3323 sack1
->start_seq
= sack2
->start_seq
;
3324 sack2
->start_seq
= tmp
;
3326 tmp
= sack1
->end_seq
;
3327 sack1
->end_seq
= sack2
->end_seq
;
3328 sack2
->end_seq
= tmp
;
3331 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3333 struct tcp_sock
*tp
= tcp_sk(sk
);
3334 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3335 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3341 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3342 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3343 /* Rotate this_sack to the first one. */
3344 for (; this_sack
>0; this_sack
--, sp
--)
3345 tcp_sack_swap(sp
, sp
-1);
3347 tcp_sack_maybe_coalesce(tp
);
3352 /* Could not find an adjacent existing SACK, build a new one,
3353 * put it at the front, and shift everyone else down. We
3354 * always know there is at least one SACK present already here.
3356 * If the sack array is full, forget about the last one.
3358 if (this_sack
>= 4) {
3360 tp
->rx_opt
.num_sacks
--;
3363 for (; this_sack
> 0; this_sack
--, sp
--)
3367 /* Build the new head SACK, and we're done. */
3368 sp
->start_seq
= seq
;
3369 sp
->end_seq
= end_seq
;
3370 tp
->rx_opt
.num_sacks
++;
3371 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3374 /* RCV.NXT advances, some SACKs should be eaten. */
3376 static void tcp_sack_remove(struct tcp_sock
*tp
)
3378 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3379 int num_sacks
= tp
->rx_opt
.num_sacks
;
3382 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3383 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3384 tp
->rx_opt
.num_sacks
= 0;
3385 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3389 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3390 /* Check if the start of the sack is covered by RCV.NXT. */
3391 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3394 /* RCV.NXT must cover all the block! */
3395 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3397 /* Zap this SACK, by moving forward any other SACKS. */
3398 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3399 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3406 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3407 tp
->rx_opt
.num_sacks
= num_sacks
;
3408 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3412 /* This one checks to see if we can put data from the
3413 * out_of_order queue into the receive_queue.
3415 static void tcp_ofo_queue(struct sock
*sk
)
3417 struct tcp_sock
*tp
= tcp_sk(sk
);
3418 __u32 dsack_high
= tp
->rcv_nxt
;
3419 struct sk_buff
*skb
;
3421 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3422 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3425 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3426 __u32 dsack
= dsack_high
;
3427 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3428 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3429 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3432 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3433 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3434 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3438 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3439 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3440 TCP_SKB_CB(skb
)->end_seq
);
3442 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3443 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3444 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3445 if (tcp_hdr(skb
)->fin
)
3446 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3450 static int tcp_prune_queue(struct sock
*sk
);
3452 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3454 struct tcphdr
*th
= tcp_hdr(skb
);
3455 struct tcp_sock
*tp
= tcp_sk(sk
);
3458 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3461 __skb_pull(skb
, th
->doff
*4);
3463 TCP_ECN_accept_cwr(tp
, skb
);
3465 if (tp
->rx_opt
.dsack
) {
3466 tp
->rx_opt
.dsack
= 0;
3467 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3468 4 - tp
->rx_opt
.tstamp_ok
);
3471 /* Queue data for delivery to the user.
3472 * Packets in sequence go to the receive queue.
3473 * Out of sequence packets to the out_of_order_queue.
3475 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3476 if (tcp_receive_window(tp
) == 0)
3479 /* Ok. In sequence. In window. */
3480 if (tp
->ucopy
.task
== current
&&
3481 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3482 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3483 int chunk
= min_t(unsigned int, skb
->len
,
3486 __set_current_state(TASK_RUNNING
);
3489 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3490 tp
->ucopy
.len
-= chunk
;
3491 tp
->copied_seq
+= chunk
;
3492 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3493 tcp_rcv_space_adjust(sk
);
3501 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3502 !sk_stream_rmem_schedule(sk
, skb
))) {
3503 if (tcp_prune_queue(sk
) < 0 ||
3504 !sk_stream_rmem_schedule(sk
, skb
))
3507 sk_stream_set_owner_r(skb
, sk
);
3508 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3510 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3512 tcp_event_data_recv(sk
, skb
);
3514 tcp_fin(skb
, sk
, th
);
3516 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3519 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3520 * gap in queue is filled.
3522 if (skb_queue_empty(&tp
->out_of_order_queue
))
3523 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3526 if (tp
->rx_opt
.num_sacks
)
3527 tcp_sack_remove(tp
);
3529 tcp_fast_path_check(sk
);
3533 else if (!sock_flag(sk
, SOCK_DEAD
))
3534 sk
->sk_data_ready(sk
, 0);
3538 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3539 /* A retransmit, 2nd most common case. Force an immediate ack. */
3540 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3541 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3544 tcp_enter_quickack_mode(sk
);
3545 inet_csk_schedule_ack(sk
);
3551 /* Out of window. F.e. zero window probe. */
3552 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3555 tcp_enter_quickack_mode(sk
);
3557 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3558 /* Partial packet, seq < rcv_next < end_seq */
3559 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3560 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3561 TCP_SKB_CB(skb
)->end_seq
);
3563 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3565 /* If window is closed, drop tail of packet. But after
3566 * remembering D-SACK for its head made in previous line.
3568 if (!tcp_receive_window(tp
))
3573 TCP_ECN_check_ce(tp
, skb
);
3575 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3576 !sk_stream_rmem_schedule(sk
, skb
)) {
3577 if (tcp_prune_queue(sk
) < 0 ||
3578 !sk_stream_rmem_schedule(sk
, skb
))
3582 /* Disable header prediction. */
3584 inet_csk_schedule_ack(sk
);
3586 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3587 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3589 sk_stream_set_owner_r(skb
, sk
);
3591 if (!skb_peek(&tp
->out_of_order_queue
)) {
3592 /* Initial out of order segment, build 1 SACK. */
3593 if (tp
->rx_opt
.sack_ok
) {
3594 tp
->rx_opt
.num_sacks
= 1;
3595 tp
->rx_opt
.dsack
= 0;
3596 tp
->rx_opt
.eff_sacks
= 1;
3597 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3598 tp
->selective_acks
[0].end_seq
=
3599 TCP_SKB_CB(skb
)->end_seq
;
3601 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3603 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3604 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3605 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3607 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3608 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3610 if (!tp
->rx_opt
.num_sacks
||
3611 tp
->selective_acks
[0].end_seq
!= seq
)
3614 /* Common case: data arrive in order after hole. */
3615 tp
->selective_acks
[0].end_seq
= end_seq
;
3619 /* Find place to insert this segment. */
3621 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3623 } while ((skb1
= skb1
->prev
) !=
3624 (struct sk_buff
*)&tp
->out_of_order_queue
);
3626 /* Do skb overlap to previous one? */
3627 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3628 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3629 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3630 /* All the bits are present. Drop. */
3632 tcp_dsack_set(tp
, seq
, end_seq
);
3635 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3636 /* Partial overlap. */
3637 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3642 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3644 /* And clean segments covered by new one as whole. */
3645 while ((skb1
= skb
->next
) !=
3646 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3647 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3648 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3649 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3652 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3653 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3658 if (tp
->rx_opt
.sack_ok
)
3659 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3663 /* Collapse contiguous sequence of skbs head..tail with
3664 * sequence numbers start..end.
3665 * Segments with FIN/SYN are not collapsed (only because this
3669 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3670 struct sk_buff
*head
, struct sk_buff
*tail
,
3673 struct sk_buff
*skb
;
3675 /* First, check that queue is collapsible and find
3676 * the point where collapsing can be useful. */
3677 for (skb
= head
; skb
!= tail
; ) {
3678 /* No new bits? It is possible on ofo queue. */
3679 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3680 struct sk_buff
*next
= skb
->next
;
3681 __skb_unlink(skb
, list
);
3683 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3688 /* The first skb to collapse is:
3690 * - bloated or contains data before "start" or
3691 * overlaps to the next one.
3693 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
3694 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3695 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3696 (skb
->next
!= tail
&&
3697 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3700 /* Decided to skip this, advance start seq. */
3701 start
= TCP_SKB_CB(skb
)->end_seq
;
3704 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
3707 while (before(start
, end
)) {
3708 struct sk_buff
*nskb
;
3709 int header
= skb_headroom(skb
);
3710 int copy
= SKB_MAX_ORDER(header
, 0);
3712 /* Too big header? This can happen with IPv6. */
3715 if (end
-start
< copy
)
3717 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3721 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
3722 skb_set_network_header(nskb
, (skb_network_header(skb
) -
3724 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
3726 skb_reserve(nskb
, header
);
3727 memcpy(nskb
->head
, skb
->head
, header
);
3728 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3729 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3730 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3731 sk_stream_set_owner_r(nskb
, sk
);
3733 /* Copy data, releasing collapsed skbs. */
3735 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3736 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3740 size
= min(copy
, size
);
3741 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3743 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3747 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3748 struct sk_buff
*next
= skb
->next
;
3749 __skb_unlink(skb
, list
);
3751 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3754 tcp_hdr(skb
)->syn
||
3762 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3763 * and tcp_collapse() them until all the queue is collapsed.
3765 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3767 struct tcp_sock
*tp
= tcp_sk(sk
);
3768 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3769 struct sk_buff
*head
;
3775 start
= TCP_SKB_CB(skb
)->seq
;
3776 end
= TCP_SKB_CB(skb
)->end_seq
;
3782 /* Segment is terminated when we see gap or when
3783 * we are at the end of all the queue. */
3784 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3785 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3786 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3787 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3788 head
, skb
, start
, end
);
3790 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3792 /* Start new segment */
3793 start
= TCP_SKB_CB(skb
)->seq
;
3794 end
= TCP_SKB_CB(skb
)->end_seq
;
3796 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3797 start
= TCP_SKB_CB(skb
)->seq
;
3798 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3799 end
= TCP_SKB_CB(skb
)->end_seq
;
3804 /* Reduce allocated memory if we can, trying to get
3805 * the socket within its memory limits again.
3807 * Return less than zero if we should start dropping frames
3808 * until the socket owning process reads some of the data
3809 * to stabilize the situation.
3811 static int tcp_prune_queue(struct sock
*sk
)
3813 struct tcp_sock
*tp
= tcp_sk(sk
);
3815 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3817 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3819 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3820 tcp_clamp_window(sk
);
3821 else if (tcp_memory_pressure
)
3822 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3824 tcp_collapse_ofo_queue(sk
);
3825 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3826 sk
->sk_receive_queue
.next
,
3827 (struct sk_buff
*)&sk
->sk_receive_queue
,
3828 tp
->copied_seq
, tp
->rcv_nxt
);
3829 sk_stream_mem_reclaim(sk
);
3831 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3834 /* Collapsing did not help, destructive actions follow.
3835 * This must not ever occur. */
3837 /* First, purge the out_of_order queue. */
3838 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3839 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3840 __skb_queue_purge(&tp
->out_of_order_queue
);
3842 /* Reset SACK state. A conforming SACK implementation will
3843 * do the same at a timeout based retransmit. When a connection
3844 * is in a sad state like this, we care only about integrity
3845 * of the connection not performance.
3847 if (tp
->rx_opt
.sack_ok
)
3848 tcp_sack_reset(&tp
->rx_opt
);
3849 sk_stream_mem_reclaim(sk
);
3852 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3855 /* If we are really being abused, tell the caller to silently
3856 * drop receive data on the floor. It will get retransmitted
3857 * and hopefully then we'll have sufficient space.
3859 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3861 /* Massive buffer overcommit. */
3867 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3868 * As additional protections, we do not touch cwnd in retransmission phases,
3869 * and if application hit its sndbuf limit recently.
3871 void tcp_cwnd_application_limited(struct sock
*sk
)
3873 struct tcp_sock
*tp
= tcp_sk(sk
);
3875 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3876 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3877 /* Limited by application or receiver window. */
3878 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
3879 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
3880 if (win_used
< tp
->snd_cwnd
) {
3881 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3882 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3884 tp
->snd_cwnd_used
= 0;
3886 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3889 static int tcp_should_expand_sndbuf(struct sock
*sk
)
3891 struct tcp_sock
*tp
= tcp_sk(sk
);
3893 /* If the user specified a specific send buffer setting, do
3896 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3899 /* If we are under global TCP memory pressure, do not expand. */
3900 if (tcp_memory_pressure
)
3903 /* If we are under soft global TCP memory pressure, do not expand. */
3904 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3907 /* If we filled the congestion window, do not expand. */
3908 if (tp
->packets_out
>= tp
->snd_cwnd
)
3914 /* When incoming ACK allowed to free some skb from write_queue,
3915 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3916 * on the exit from tcp input handler.
3918 * PROBLEM: sndbuf expansion does not work well with largesend.
3920 static void tcp_new_space(struct sock
*sk
)
3922 struct tcp_sock
*tp
= tcp_sk(sk
);
3924 if (tcp_should_expand_sndbuf(sk
)) {
3925 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3926 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3927 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3928 tp
->reordering
+ 1);
3929 sndmem
*= 2*demanded
;
3930 if (sndmem
> sk
->sk_sndbuf
)
3931 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3932 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3935 sk
->sk_write_space(sk
);
3938 static void tcp_check_space(struct sock
*sk
)
3940 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
3941 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
3942 if (sk
->sk_socket
&&
3943 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
3948 static inline void tcp_data_snd_check(struct sock
*sk
)
3950 tcp_push_pending_frames(sk
);
3951 tcp_check_space(sk
);
3955 * Check if sending an ack is needed.
3957 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
3959 struct tcp_sock
*tp
= tcp_sk(sk
);
3961 /* More than one full frame received... */
3962 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
3963 /* ... and right edge of window advances far enough.
3964 * (tcp_recvmsg() will send ACK otherwise). Or...
3966 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
3967 /* We ACK each frame or... */
3968 tcp_in_quickack_mode(sk
) ||
3969 /* We have out of order data. */
3971 skb_peek(&tp
->out_of_order_queue
))) {
3972 /* Then ack it now */
3975 /* Else, send delayed ack. */
3976 tcp_send_delayed_ack(sk
);
3980 static inline void tcp_ack_snd_check(struct sock
*sk
)
3982 if (!inet_csk_ack_scheduled(sk
)) {
3983 /* We sent a data segment already. */
3986 __tcp_ack_snd_check(sk
, 1);
3990 * This routine is only called when we have urgent data
3991 * signaled. Its the 'slow' part of tcp_urg. It could be
3992 * moved inline now as tcp_urg is only called from one
3993 * place. We handle URGent data wrong. We have to - as
3994 * BSD still doesn't use the correction from RFC961.
3995 * For 1003.1g we should support a new option TCP_STDURG to permit
3996 * either form (or just set the sysctl tcp_stdurg).
3999 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4001 struct tcp_sock
*tp
= tcp_sk(sk
);
4002 u32 ptr
= ntohs(th
->urg_ptr
);
4004 if (ptr
&& !sysctl_tcp_stdurg
)
4006 ptr
+= ntohl(th
->seq
);
4008 /* Ignore urgent data that we've already seen and read. */
4009 if (after(tp
->copied_seq
, ptr
))
4012 /* Do not replay urg ptr.
4014 * NOTE: interesting situation not covered by specs.
4015 * Misbehaving sender may send urg ptr, pointing to segment,
4016 * which we already have in ofo queue. We are not able to fetch
4017 * such data and will stay in TCP_URG_NOTYET until will be eaten
4018 * by recvmsg(). Seems, we are not obliged to handle such wicked
4019 * situations. But it is worth to think about possibility of some
4020 * DoSes using some hypothetical application level deadlock.
4022 if (before(ptr
, tp
->rcv_nxt
))
4025 /* Do we already have a newer (or duplicate) urgent pointer? */
4026 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4029 /* Tell the world about our new urgent pointer. */
4032 /* We may be adding urgent data when the last byte read was
4033 * urgent. To do this requires some care. We cannot just ignore
4034 * tp->copied_seq since we would read the last urgent byte again
4035 * as data, nor can we alter copied_seq until this data arrives
4036 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4038 * NOTE. Double Dutch. Rendering to plain English: author of comment
4039 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4040 * and expect that both A and B disappear from stream. This is _wrong_.
4041 * Though this happens in BSD with high probability, this is occasional.
4042 * Any application relying on this is buggy. Note also, that fix "works"
4043 * only in this artificial test. Insert some normal data between A and B and we will
4044 * decline of BSD again. Verdict: it is better to remove to trap
4047 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4048 !sock_flag(sk
, SOCK_URGINLINE
) &&
4049 tp
->copied_seq
!= tp
->rcv_nxt
) {
4050 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4052 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4053 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4058 tp
->urg_data
= TCP_URG_NOTYET
;
4061 /* Disable header prediction. */
4065 /* This is the 'fast' part of urgent handling. */
4066 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4068 struct tcp_sock
*tp
= tcp_sk(sk
);
4070 /* Check if we get a new urgent pointer - normally not. */
4072 tcp_check_urg(sk
,th
);
4074 /* Do we wait for any urgent data? - normally not... */
4075 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4076 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4079 /* Is the urgent pointer pointing into this packet? */
4080 if (ptr
< skb
->len
) {
4082 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4084 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4085 if (!sock_flag(sk
, SOCK_DEAD
))
4086 sk
->sk_data_ready(sk
, 0);
4091 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4093 struct tcp_sock
*tp
= tcp_sk(sk
);
4094 int chunk
= skb
->len
- hlen
;
4098 if (skb_csum_unnecessary(skb
))
4099 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4101 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4105 tp
->ucopy
.len
-= chunk
;
4106 tp
->copied_seq
+= chunk
;
4107 tcp_rcv_space_adjust(sk
);
4114 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4118 if (sock_owned_by_user(sk
)) {
4120 result
= __tcp_checksum_complete(skb
);
4123 result
= __tcp_checksum_complete(skb
);
4128 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4130 return !skb_csum_unnecessary(skb
) &&
4131 __tcp_checksum_complete_user(sk
, skb
);
4134 #ifdef CONFIG_NET_DMA
4135 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4137 struct tcp_sock
*tp
= tcp_sk(sk
);
4138 int chunk
= skb
->len
- hlen
;
4140 int copied_early
= 0;
4142 if (tp
->ucopy
.wakeup
)
4145 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4146 tp
->ucopy
.dma_chan
= get_softnet_dma();
4148 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4150 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4151 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4156 tp
->ucopy
.dma_cookie
= dma_cookie
;
4159 tp
->ucopy
.len
-= chunk
;
4160 tp
->copied_seq
+= chunk
;
4161 tcp_rcv_space_adjust(sk
);
4163 if ((tp
->ucopy
.len
== 0) ||
4164 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4165 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4166 tp
->ucopy
.wakeup
= 1;
4167 sk
->sk_data_ready(sk
, 0);
4169 } else if (chunk
> 0) {
4170 tp
->ucopy
.wakeup
= 1;
4171 sk
->sk_data_ready(sk
, 0);
4174 return copied_early
;
4176 #endif /* CONFIG_NET_DMA */
4179 * TCP receive function for the ESTABLISHED state.
4181 * It is split into a fast path and a slow path. The fast path is
4183 * - A zero window was announced from us - zero window probing
4184 * is only handled properly in the slow path.
4185 * - Out of order segments arrived.
4186 * - Urgent data is expected.
4187 * - There is no buffer space left
4188 * - Unexpected TCP flags/window values/header lengths are received
4189 * (detected by checking the TCP header against pred_flags)
4190 * - Data is sent in both directions. Fast path only supports pure senders
4191 * or pure receivers (this means either the sequence number or the ack
4192 * value must stay constant)
4193 * - Unexpected TCP option.
4195 * When these conditions are not satisfied it drops into a standard
4196 * receive procedure patterned after RFC793 to handle all cases.
4197 * The first three cases are guaranteed by proper pred_flags setting,
4198 * the rest is checked inline. Fast processing is turned on in
4199 * tcp_data_queue when everything is OK.
4201 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4202 struct tcphdr
*th
, unsigned len
)
4204 struct tcp_sock
*tp
= tcp_sk(sk
);
4207 * Header prediction.
4208 * The code loosely follows the one in the famous
4209 * "30 instruction TCP receive" Van Jacobson mail.
4211 * Van's trick is to deposit buffers into socket queue
4212 * on a device interrupt, to call tcp_recv function
4213 * on the receive process context and checksum and copy
4214 * the buffer to user space. smart...
4216 * Our current scheme is not silly either but we take the
4217 * extra cost of the net_bh soft interrupt processing...
4218 * We do checksum and copy also but from device to kernel.
4221 tp
->rx_opt
.saw_tstamp
= 0;
4223 /* pred_flags is 0xS?10 << 16 + snd_wnd
4224 * if header_prediction is to be made
4225 * 'S' will always be tp->tcp_header_len >> 2
4226 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4227 * turn it off (when there are holes in the receive
4228 * space for instance)
4229 * PSH flag is ignored.
4232 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4233 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4234 int tcp_header_len
= tp
->tcp_header_len
;
4236 /* Timestamp header prediction: tcp_header_len
4237 * is automatically equal to th->doff*4 due to pred_flags
4241 /* Check timestamp */
4242 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4243 __be32
*ptr
= (__be32
*)(th
+ 1);
4245 /* No? Slow path! */
4246 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4247 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4250 tp
->rx_opt
.saw_tstamp
= 1;
4252 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4254 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4256 /* If PAWS failed, check it more carefully in slow path */
4257 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4260 /* DO NOT update ts_recent here, if checksum fails
4261 * and timestamp was corrupted part, it will result
4262 * in a hung connection since we will drop all
4263 * future packets due to the PAWS test.
4267 if (len
<= tcp_header_len
) {
4268 /* Bulk data transfer: sender */
4269 if (len
== tcp_header_len
) {
4270 /* Predicted packet is in window by definition.
4271 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4272 * Hence, check seq<=rcv_wup reduces to:
4274 if (tcp_header_len
==
4275 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4276 tp
->rcv_nxt
== tp
->rcv_wup
)
4277 tcp_store_ts_recent(tp
);
4279 /* We know that such packets are checksummed
4282 tcp_ack(sk
, skb
, 0);
4284 tcp_data_snd_check(sk
);
4286 } else { /* Header too small */
4287 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4292 int copied_early
= 0;
4294 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4295 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4296 #ifdef CONFIG_NET_DMA
4297 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4302 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4303 __set_current_state(TASK_RUNNING
);
4305 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4309 /* Predicted packet is in window by definition.
4310 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4311 * Hence, check seq<=rcv_wup reduces to:
4313 if (tcp_header_len
==
4314 (sizeof(struct tcphdr
) +
4315 TCPOLEN_TSTAMP_ALIGNED
) &&
4316 tp
->rcv_nxt
== tp
->rcv_wup
)
4317 tcp_store_ts_recent(tp
);
4319 tcp_rcv_rtt_measure_ts(sk
, skb
);
4321 __skb_pull(skb
, tcp_header_len
);
4322 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4323 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4326 tcp_cleanup_rbuf(sk
, skb
->len
);
4329 if (tcp_checksum_complete_user(sk
, skb
))
4332 /* Predicted packet is in window by definition.
4333 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4334 * Hence, check seq<=rcv_wup reduces to:
4336 if (tcp_header_len
==
4337 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4338 tp
->rcv_nxt
== tp
->rcv_wup
)
4339 tcp_store_ts_recent(tp
);
4341 tcp_rcv_rtt_measure_ts(sk
, skb
);
4343 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4346 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4348 /* Bulk data transfer: receiver */
4349 __skb_pull(skb
,tcp_header_len
);
4350 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4351 sk_stream_set_owner_r(skb
, sk
);
4352 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4355 tcp_event_data_recv(sk
, skb
);
4357 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4358 /* Well, only one small jumplet in fast path... */
4359 tcp_ack(sk
, skb
, FLAG_DATA
);
4360 tcp_data_snd_check(sk
);
4361 if (!inet_csk_ack_scheduled(sk
))
4365 __tcp_ack_snd_check(sk
, 0);
4367 #ifdef CONFIG_NET_DMA
4369 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4375 sk
->sk_data_ready(sk
, 0);
4381 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4385 * RFC1323: H1. Apply PAWS check first.
4387 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4388 tcp_paws_discard(sk
, skb
)) {
4390 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4391 tcp_send_dupack(sk
, skb
);
4394 /* Resets are accepted even if PAWS failed.
4396 ts_recent update must be made after we are sure
4397 that the packet is in window.
4402 * Standard slow path.
4405 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4406 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4407 * (RST) segments are validated by checking their SEQ-fields."
4408 * And page 69: "If an incoming segment is not acceptable,
4409 * an acknowledgment should be sent in reply (unless the RST bit
4410 * is set, if so drop the segment and return)".
4413 tcp_send_dupack(sk
, skb
);
4422 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4424 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4425 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4426 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4433 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4435 tcp_rcv_rtt_measure_ts(sk
, skb
);
4437 /* Process urgent data. */
4438 tcp_urg(sk
, skb
, th
);
4440 /* step 7: process the segment text */
4441 tcp_data_queue(sk
, skb
);
4443 tcp_data_snd_check(sk
);
4444 tcp_ack_snd_check(sk
);
4448 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4455 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4456 struct tcphdr
*th
, unsigned len
)
4458 struct tcp_sock
*tp
= tcp_sk(sk
);
4459 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4460 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4462 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4466 * "If the state is SYN-SENT then
4467 * first check the ACK bit
4468 * If the ACK bit is set
4469 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4470 * a reset (unless the RST bit is set, if so drop
4471 * the segment and return)"
4473 * We do not send data with SYN, so that RFC-correct
4476 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4477 goto reset_and_undo
;
4479 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4480 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4482 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4483 goto reset_and_undo
;
4486 /* Now ACK is acceptable.
4488 * "If the RST bit is set
4489 * If the ACK was acceptable then signal the user "error:
4490 * connection reset", drop the segment, enter CLOSED state,
4491 * delete TCB, and return."
4500 * "fifth, if neither of the SYN or RST bits is set then
4501 * drop the segment and return."
4507 goto discard_and_undo
;
4510 * "If the SYN bit is on ...
4511 * are acceptable then ...
4512 * (our SYN has been ACKed), change the connection
4513 * state to ESTABLISHED..."
4516 TCP_ECN_rcv_synack(tp
, th
);
4518 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4519 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4521 /* Ok.. it's good. Set up sequence numbers and
4522 * move to established.
4524 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4525 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4527 /* RFC1323: The window in SYN & SYN/ACK segments is
4530 tp
->snd_wnd
= ntohs(th
->window
);
4531 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4533 if (!tp
->rx_opt
.wscale_ok
) {
4534 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4535 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4538 if (tp
->rx_opt
.saw_tstamp
) {
4539 tp
->rx_opt
.tstamp_ok
= 1;
4540 tp
->tcp_header_len
=
4541 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4542 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4543 tcp_store_ts_recent(tp
);
4545 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4548 if (tp
->rx_opt
.sack_ok
&& sysctl_tcp_fack
)
4549 tp
->rx_opt
.sack_ok
|= 2;
4552 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4553 tcp_initialize_rcv_mss(sk
);
4555 /* Remember, tcp_poll() does not lock socket!
4556 * Change state from SYN-SENT only after copied_seq
4557 * is initialized. */
4558 tp
->copied_seq
= tp
->rcv_nxt
;
4560 tcp_set_state(sk
, TCP_ESTABLISHED
);
4562 security_inet_conn_established(sk
, skb
);
4564 /* Make sure socket is routed, for correct metrics. */
4565 icsk
->icsk_af_ops
->rebuild_header(sk
);
4567 tcp_init_metrics(sk
);
4569 tcp_init_congestion_control(sk
);
4571 /* Prevent spurious tcp_cwnd_restart() on first data
4574 tp
->lsndtime
= tcp_time_stamp
;
4576 tcp_init_buffer_space(sk
);
4578 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4579 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4581 if (!tp
->rx_opt
.snd_wscale
)
4582 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4586 if (!sock_flag(sk
, SOCK_DEAD
)) {
4587 sk
->sk_state_change(sk
);
4588 sk_wake_async(sk
, 0, POLL_OUT
);
4591 if (sk
->sk_write_pending
||
4592 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4593 icsk
->icsk_ack
.pingpong
) {
4594 /* Save one ACK. Data will be ready after
4595 * several ticks, if write_pending is set.
4597 * It may be deleted, but with this feature tcpdumps
4598 * look so _wonderfully_ clever, that I was not able
4599 * to stand against the temptation 8) --ANK
4601 inet_csk_schedule_ack(sk
);
4602 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4603 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4604 tcp_incr_quickack(sk
);
4605 tcp_enter_quickack_mode(sk
);
4606 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4607 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4618 /* No ACK in the segment */
4622 * "If the RST bit is set
4624 * Otherwise (no ACK) drop the segment and return."
4627 goto discard_and_undo
;
4631 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4632 goto discard_and_undo
;
4635 /* We see SYN without ACK. It is attempt of
4636 * simultaneous connect with crossed SYNs.
4637 * Particularly, it can be connect to self.
4639 tcp_set_state(sk
, TCP_SYN_RECV
);
4641 if (tp
->rx_opt
.saw_tstamp
) {
4642 tp
->rx_opt
.tstamp_ok
= 1;
4643 tcp_store_ts_recent(tp
);
4644 tp
->tcp_header_len
=
4645 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4647 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4650 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4651 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4653 /* RFC1323: The window in SYN & SYN/ACK segments is
4656 tp
->snd_wnd
= ntohs(th
->window
);
4657 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4658 tp
->max_window
= tp
->snd_wnd
;
4660 TCP_ECN_rcv_syn(tp
, th
);
4663 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4664 tcp_initialize_rcv_mss(sk
);
4667 tcp_send_synack(sk
);
4669 /* Note, we could accept data and URG from this segment.
4670 * There are no obstacles to make this.
4672 * However, if we ignore data in ACKless segments sometimes,
4673 * we have no reasons to accept it sometimes.
4674 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4675 * is not flawless. So, discard packet for sanity.
4676 * Uncomment this return to process the data.
4683 /* "fifth, if neither of the SYN or RST bits is set then
4684 * drop the segment and return."
4688 tcp_clear_options(&tp
->rx_opt
);
4689 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4693 tcp_clear_options(&tp
->rx_opt
);
4694 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4700 * This function implements the receiving procedure of RFC 793 for
4701 * all states except ESTABLISHED and TIME_WAIT.
4702 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4703 * address independent.
4706 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4707 struct tcphdr
*th
, unsigned len
)
4709 struct tcp_sock
*tp
= tcp_sk(sk
);
4710 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4713 tp
->rx_opt
.saw_tstamp
= 0;
4715 switch (sk
->sk_state
) {
4727 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4730 /* Now we have several options: In theory there is
4731 * nothing else in the frame. KA9Q has an option to
4732 * send data with the syn, BSD accepts data with the
4733 * syn up to the [to be] advertised window and
4734 * Solaris 2.1 gives you a protocol error. For now
4735 * we just ignore it, that fits the spec precisely
4736 * and avoids incompatibilities. It would be nice in
4737 * future to drop through and process the data.
4739 * Now that TTCP is starting to be used we ought to
4741 * But, this leaves one open to an easy denial of
4742 * service attack, and SYN cookies can't defend
4743 * against this problem. So, we drop the data
4744 * in the interest of security over speed unless
4745 * it's still in use.
4753 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4757 /* Do step6 onward by hand. */
4758 tcp_urg(sk
, skb
, th
);
4760 tcp_data_snd_check(sk
);
4764 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4765 tcp_paws_discard(sk
, skb
)) {
4767 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4768 tcp_send_dupack(sk
, skb
);
4771 /* Reset is accepted even if it did not pass PAWS. */
4774 /* step 1: check sequence number */
4775 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4777 tcp_send_dupack(sk
, skb
);
4781 /* step 2: check RST bit */
4787 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4789 /* step 3: check security and precedence [ignored] */
4793 * Check for a SYN in window.
4795 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4796 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4801 /* step 5: check the ACK field */
4803 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4805 switch (sk
->sk_state
) {
4808 tp
->copied_seq
= tp
->rcv_nxt
;
4810 tcp_set_state(sk
, TCP_ESTABLISHED
);
4811 sk
->sk_state_change(sk
);
4813 /* Note, that this wakeup is only for marginal
4814 * crossed SYN case. Passively open sockets
4815 * are not waked up, because sk->sk_sleep ==
4816 * NULL and sk->sk_socket == NULL.
4818 if (sk
->sk_socket
) {
4819 sk_wake_async(sk
,0,POLL_OUT
);
4822 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4823 tp
->snd_wnd
= ntohs(th
->window
) <<
4824 tp
->rx_opt
.snd_wscale
;
4825 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4826 TCP_SKB_CB(skb
)->seq
);
4828 /* tcp_ack considers this ACK as duplicate
4829 * and does not calculate rtt.
4830 * Fix it at least with timestamps.
4832 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4834 tcp_ack_saw_tstamp(sk
, 0);
4836 if (tp
->rx_opt
.tstamp_ok
)
4837 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4839 /* Make sure socket is routed, for
4842 icsk
->icsk_af_ops
->rebuild_header(sk
);
4844 tcp_init_metrics(sk
);
4846 tcp_init_congestion_control(sk
);
4848 /* Prevent spurious tcp_cwnd_restart() on
4849 * first data packet.
4851 tp
->lsndtime
= tcp_time_stamp
;
4854 tcp_initialize_rcv_mss(sk
);
4855 tcp_init_buffer_space(sk
);
4856 tcp_fast_path_on(tp
);
4863 if (tp
->snd_una
== tp
->write_seq
) {
4864 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4865 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4866 dst_confirm(sk
->sk_dst_cache
);
4868 if (!sock_flag(sk
, SOCK_DEAD
))
4869 /* Wake up lingering close() */
4870 sk
->sk_state_change(sk
);
4874 if (tp
->linger2
< 0 ||
4875 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4876 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4878 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4882 tmo
= tcp_fin_time(sk
);
4883 if (tmo
> TCP_TIMEWAIT_LEN
) {
4884 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4885 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4886 /* Bad case. We could lose such FIN otherwise.
4887 * It is not a big problem, but it looks confusing
4888 * and not so rare event. We still can lose it now,
4889 * if it spins in bh_lock_sock(), but it is really
4892 inet_csk_reset_keepalive_timer(sk
, tmo
);
4894 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4902 if (tp
->snd_una
== tp
->write_seq
) {
4903 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4909 if (tp
->snd_una
== tp
->write_seq
) {
4910 tcp_update_metrics(sk
);
4919 /* step 6: check the URG bit */
4920 tcp_urg(sk
, skb
, th
);
4922 /* step 7: process the segment text */
4923 switch (sk
->sk_state
) {
4924 case TCP_CLOSE_WAIT
:
4927 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4931 /* RFC 793 says to queue data in these states,
4932 * RFC 1122 says we MUST send a reset.
4933 * BSD 4.4 also does reset.
4935 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4936 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4937 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4938 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4944 case TCP_ESTABLISHED
:
4945 tcp_data_queue(sk
, skb
);
4950 /* tcp_data could move socket to TIME-WAIT */
4951 if (sk
->sk_state
!= TCP_CLOSE
) {
4952 tcp_data_snd_check(sk
);
4953 tcp_ack_snd_check(sk
);
4963 EXPORT_SYMBOL(sysctl_tcp_ecn
);
4964 EXPORT_SYMBOL(sysctl_tcp_reordering
);
4965 EXPORT_SYMBOL(tcp_parse_options
);
4966 EXPORT_SYMBOL(tcp_rcv_established
);
4967 EXPORT_SYMBOL(tcp_rcv_state_process
);
4968 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);