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 IsSackFrto() (sysctl_tcp_frto == 0x2)
116 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
117 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
119 /* Adapt the MSS value used to make delayed ack decision to the
122 static void tcp_measure_rcv_mss(struct sock
*sk
,
123 const struct sk_buff
*skb
)
125 struct inet_connection_sock
*icsk
= inet_csk(sk
);
126 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
129 icsk
->icsk_ack
.last_seg_size
= 0;
131 /* skb->len may jitter because of SACKs, even if peer
132 * sends good full-sized frames.
134 len
= skb_shinfo(skb
)->gso_size
?: skb
->len
;
135 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
136 icsk
->icsk_ack
.rcv_mss
= len
;
138 /* Otherwise, we make more careful check taking into account,
139 * that SACKs block is variable.
141 * "len" is invariant segment length, including TCP header.
143 len
+= skb
->data
- skb_transport_header(skb
);
144 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
145 /* If PSH is not set, packet should be
146 * full sized, provided peer TCP is not badly broken.
147 * This observation (if it is correct 8)) allows
148 * to handle super-low mtu links fairly.
150 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
151 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
152 /* Subtract also invariant (if peer is RFC compliant),
153 * tcp header plus fixed timestamp option length.
154 * Resulting "len" is MSS free of SACK jitter.
156 len
-= tcp_sk(sk
)->tcp_header_len
;
157 icsk
->icsk_ack
.last_seg_size
= len
;
159 icsk
->icsk_ack
.rcv_mss
= len
;
163 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
164 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
165 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
169 static void tcp_incr_quickack(struct sock
*sk
)
171 struct inet_connection_sock
*icsk
= inet_csk(sk
);
172 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
176 if (quickacks
> icsk
->icsk_ack
.quick
)
177 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
180 void tcp_enter_quickack_mode(struct sock
*sk
)
182 struct inet_connection_sock
*icsk
= inet_csk(sk
);
183 tcp_incr_quickack(sk
);
184 icsk
->icsk_ack
.pingpong
= 0;
185 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
188 /* Send ACKs quickly, if "quick" count is not exhausted
189 * and the session is not interactive.
192 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
194 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
195 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
198 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
200 if (tp
->ecn_flags
&TCP_ECN_OK
)
201 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
204 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
206 if (tcp_hdr(skb
)->cwr
)
207 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
210 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
212 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
215 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
217 if (tp
->ecn_flags
&TCP_ECN_OK
) {
218 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
219 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
220 /* Funny extension: if ECT is not set on a segment,
221 * it is surely retransmit. It is not in ECN RFC,
222 * but Linux follows this rule. */
223 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
224 tcp_enter_quickack_mode((struct sock
*)tp
);
228 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
230 if ((tp
->ecn_flags
&TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
231 tp
->ecn_flags
&= ~TCP_ECN_OK
;
234 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
236 if ((tp
->ecn_flags
&TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
237 tp
->ecn_flags
&= ~TCP_ECN_OK
;
240 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
242 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
&TCP_ECN_OK
))
247 /* Buffer size and advertised window tuning.
249 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
252 static void tcp_fixup_sndbuf(struct sock
*sk
)
254 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
255 sizeof(struct sk_buff
);
257 if (sk
->sk_sndbuf
< 3 * sndmem
)
258 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
261 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
263 * All tcp_full_space() is split to two parts: "network" buffer, allocated
264 * forward and advertised in receiver window (tp->rcv_wnd) and
265 * "application buffer", required to isolate scheduling/application
266 * latencies from network.
267 * window_clamp is maximal advertised window. It can be less than
268 * tcp_full_space(), in this case tcp_full_space() - window_clamp
269 * is reserved for "application" buffer. The less window_clamp is
270 * the smoother our behaviour from viewpoint of network, but the lower
271 * throughput and the higher sensitivity of the connection to losses. 8)
273 * rcv_ssthresh is more strict window_clamp used at "slow start"
274 * phase to predict further behaviour of this connection.
275 * It is used for two goals:
276 * - to enforce header prediction at sender, even when application
277 * requires some significant "application buffer". It is check #1.
278 * - to prevent pruning of receive queue because of misprediction
279 * of receiver window. Check #2.
281 * The scheme does not work when sender sends good segments opening
282 * window and then starts to feed us spaghetti. But it should work
283 * in common situations. Otherwise, we have to rely on queue collapsing.
286 /* Slow part of check#2. */
287 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
289 struct tcp_sock
*tp
= tcp_sk(sk
);
291 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
292 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2])/2;
294 while (tp
->rcv_ssthresh
<= window
) {
295 if (truesize
<= skb
->len
)
296 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
304 static void tcp_grow_window(struct sock
*sk
,
307 struct tcp_sock
*tp
= tcp_sk(sk
);
310 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
311 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
312 !tcp_memory_pressure
) {
315 /* Check #2. Increase window, if skb with such overhead
316 * will fit to rcvbuf in future.
318 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
321 incr
= __tcp_grow_window(sk
, skb
);
324 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
325 inet_csk(sk
)->icsk_ack
.quick
|= 1;
330 /* 3. Tuning rcvbuf, when connection enters established state. */
332 static void tcp_fixup_rcvbuf(struct sock
*sk
)
334 struct tcp_sock
*tp
= tcp_sk(sk
);
335 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
337 /* Try to select rcvbuf so that 4 mss-sized segments
338 * will fit to window and corresponding skbs will fit to our rcvbuf.
339 * (was 3; 4 is minimum to allow fast retransmit to work.)
341 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
343 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
344 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
347 /* 4. Try to fixup all. It is made immediately after connection enters
350 static void tcp_init_buffer_space(struct sock
*sk
)
352 struct tcp_sock
*tp
= tcp_sk(sk
);
355 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
356 tcp_fixup_rcvbuf(sk
);
357 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
358 tcp_fixup_sndbuf(sk
);
360 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
362 maxwin
= tcp_full_space(sk
);
364 if (tp
->window_clamp
>= maxwin
) {
365 tp
->window_clamp
= maxwin
;
367 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
368 tp
->window_clamp
= max(maxwin
-
369 (maxwin
>> sysctl_tcp_app_win
),
373 /* Force reservation of one segment. */
374 if (sysctl_tcp_app_win
&&
375 tp
->window_clamp
> 2 * tp
->advmss
&&
376 tp
->window_clamp
+ tp
->advmss
> maxwin
)
377 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
379 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
380 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
383 /* 5. Recalculate window clamp after socket hit its memory bounds. */
384 static void tcp_clamp_window(struct sock
*sk
)
386 struct tcp_sock
*tp
= tcp_sk(sk
);
387 struct inet_connection_sock
*icsk
= inet_csk(sk
);
389 icsk
->icsk_ack
.quick
= 0;
391 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
392 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
393 !tcp_memory_pressure
&&
394 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
395 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
398 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
399 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
403 /* Initialize RCV_MSS value.
404 * RCV_MSS is an our guess about MSS used by the peer.
405 * We haven't any direct information about the MSS.
406 * It's better to underestimate the RCV_MSS rather than overestimate.
407 * Overestimations make us ACKing less frequently than needed.
408 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
410 void tcp_initialize_rcv_mss(struct sock
*sk
)
412 struct tcp_sock
*tp
= tcp_sk(sk
);
413 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
415 hint
= min(hint
, tp
->rcv_wnd
/2);
416 hint
= min(hint
, TCP_MIN_RCVMSS
);
417 hint
= max(hint
, TCP_MIN_MSS
);
419 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
422 /* Receiver "autotuning" code.
424 * The algorithm for RTT estimation w/o timestamps is based on
425 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
426 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
428 * More detail on this code can be found at
429 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
430 * though this reference is out of date. A new paper
433 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
435 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
441 if (new_sample
!= 0) {
442 /* If we sample in larger samples in the non-timestamp
443 * case, we could grossly overestimate the RTT especially
444 * with chatty applications or bulk transfer apps which
445 * are stalled on filesystem I/O.
447 * Also, since we are only going for a minimum in the
448 * non-timestamp case, we do not smooth things out
449 * else with timestamps disabled convergence takes too
453 m
-= (new_sample
>> 3);
455 } else if (m
< new_sample
)
458 /* No previous measure. */
462 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
463 tp
->rcv_rtt_est
.rtt
= new_sample
;
466 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
468 if (tp
->rcv_rtt_est
.time
== 0)
470 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
472 tcp_rcv_rtt_update(tp
,
473 jiffies
- tp
->rcv_rtt_est
.time
,
477 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
478 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
481 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
483 struct tcp_sock
*tp
= tcp_sk(sk
);
484 if (tp
->rx_opt
.rcv_tsecr
&&
485 (TCP_SKB_CB(skb
)->end_seq
-
486 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
487 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
491 * This function should be called every time data is copied to user space.
492 * It calculates the appropriate TCP receive buffer space.
494 void tcp_rcv_space_adjust(struct sock
*sk
)
496 struct tcp_sock
*tp
= tcp_sk(sk
);
500 if (tp
->rcvq_space
.time
== 0)
503 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
504 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
505 tp
->rcv_rtt_est
.rtt
== 0)
508 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
510 space
= max(tp
->rcvq_space
.space
, space
);
512 if (tp
->rcvq_space
.space
!= space
) {
515 tp
->rcvq_space
.space
= space
;
517 if (sysctl_tcp_moderate_rcvbuf
&&
518 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
519 int new_clamp
= space
;
521 /* Receive space grows, normalize in order to
522 * take into account packet headers and sk_buff
523 * structure overhead.
528 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
529 16 + sizeof(struct sk_buff
));
530 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
533 space
= min(space
, sysctl_tcp_rmem
[2]);
534 if (space
> sk
->sk_rcvbuf
) {
535 sk
->sk_rcvbuf
= space
;
537 /* Make the window clamp follow along. */
538 tp
->window_clamp
= new_clamp
;
544 tp
->rcvq_space
.seq
= tp
->copied_seq
;
545 tp
->rcvq_space
.time
= tcp_time_stamp
;
548 /* There is something which you must keep in mind when you analyze the
549 * behavior of the tp->ato delayed ack timeout interval. When a
550 * connection starts up, we want to ack as quickly as possible. The
551 * problem is that "good" TCP's do slow start at the beginning of data
552 * transmission. The means that until we send the first few ACK's the
553 * sender will sit on his end and only queue most of his data, because
554 * he can only send snd_cwnd unacked packets at any given time. For
555 * each ACK we send, he increments snd_cwnd and transmits more of his
558 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
560 struct tcp_sock
*tp
= tcp_sk(sk
);
561 struct inet_connection_sock
*icsk
= inet_csk(sk
);
564 inet_csk_schedule_ack(sk
);
566 tcp_measure_rcv_mss(sk
, skb
);
568 tcp_rcv_rtt_measure(tp
);
570 now
= tcp_time_stamp
;
572 if (!icsk
->icsk_ack
.ato
) {
573 /* The _first_ data packet received, initialize
574 * delayed ACK engine.
576 tcp_incr_quickack(sk
);
577 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
579 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
581 if (m
<= TCP_ATO_MIN
/2) {
582 /* The fastest case is the first. */
583 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
584 } else if (m
< icsk
->icsk_ack
.ato
) {
585 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
586 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
587 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
588 } else if (m
> icsk
->icsk_rto
) {
589 /* Too long gap. Apparently sender failed to
590 * restart window, so that we send ACKs quickly.
592 tcp_incr_quickack(sk
);
593 sk_stream_mem_reclaim(sk
);
596 icsk
->icsk_ack
.lrcvtime
= now
;
598 TCP_ECN_check_ce(tp
, skb
);
601 tcp_grow_window(sk
, skb
);
604 static u32
tcp_rto_min(struct sock
*sk
)
606 struct dst_entry
*dst
= __sk_dst_get(sk
);
607 u32 rto_min
= TCP_RTO_MIN
;
609 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
610 rto_min
= dst
->metrics
[RTAX_RTO_MIN
-1];
614 /* Called to compute a smoothed rtt estimate. The data fed to this
615 * routine either comes from timestamps, or from segments that were
616 * known _not_ to have been retransmitted [see Karn/Partridge
617 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
618 * piece by Van Jacobson.
619 * NOTE: the next three routines used to be one big routine.
620 * To save cycles in the RFC 1323 implementation it was better to break
621 * it up into three procedures. -- erics
623 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
625 struct tcp_sock
*tp
= tcp_sk(sk
);
626 long m
= mrtt
; /* RTT */
628 /* The following amusing code comes from Jacobson's
629 * article in SIGCOMM '88. Note that rtt and mdev
630 * are scaled versions of rtt and mean deviation.
631 * This is designed to be as fast as possible
632 * m stands for "measurement".
634 * On a 1990 paper the rto value is changed to:
635 * RTO = rtt + 4 * mdev
637 * Funny. This algorithm seems to be very broken.
638 * These formulae increase RTO, when it should be decreased, increase
639 * too slowly, when it should be increased quickly, decrease too quickly
640 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
641 * does not matter how to _calculate_ it. Seems, it was trap
642 * that VJ failed to avoid. 8)
647 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
648 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
650 m
= -m
; /* m is now abs(error) */
651 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
652 /* This is similar to one of Eifel findings.
653 * Eifel blocks mdev updates when rtt decreases.
654 * This solution is a bit different: we use finer gain
655 * for mdev in this case (alpha*beta).
656 * Like Eifel it also prevents growth of rto,
657 * but also it limits too fast rto decreases,
658 * happening in pure Eifel.
663 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
665 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
666 if (tp
->mdev
> tp
->mdev_max
) {
667 tp
->mdev_max
= tp
->mdev
;
668 if (tp
->mdev_max
> tp
->rttvar
)
669 tp
->rttvar
= tp
->mdev_max
;
671 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
672 if (tp
->mdev_max
< tp
->rttvar
)
673 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
674 tp
->rtt_seq
= tp
->snd_nxt
;
675 tp
->mdev_max
= tcp_rto_min(sk
);
678 /* no previous measure. */
679 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
680 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
681 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
682 tp
->rtt_seq
= tp
->snd_nxt
;
686 /* Calculate rto without backoff. This is the second half of Van Jacobson's
687 * routine referred to above.
689 static inline void tcp_set_rto(struct sock
*sk
)
691 const struct tcp_sock
*tp
= tcp_sk(sk
);
692 /* Old crap is replaced with new one. 8)
695 * 1. If rtt variance happened to be less 50msec, it is hallucination.
696 * It cannot be less due to utterly erratic ACK generation made
697 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
698 * to do with delayed acks, because at cwnd>2 true delack timeout
699 * is invisible. Actually, Linux-2.4 also generates erratic
700 * ACKs in some circumstances.
702 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
704 /* 2. Fixups made earlier cannot be right.
705 * If we do not estimate RTO correctly without them,
706 * all the algo is pure shit and should be replaced
707 * with correct one. It is exactly, which we pretend to do.
711 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
712 * guarantees that rto is higher.
714 static inline void tcp_bound_rto(struct sock
*sk
)
716 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
717 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
720 /* Save metrics learned by this TCP session.
721 This function is called only, when TCP finishes successfully
722 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
724 void tcp_update_metrics(struct sock
*sk
)
726 struct tcp_sock
*tp
= tcp_sk(sk
);
727 struct dst_entry
*dst
= __sk_dst_get(sk
);
729 if (sysctl_tcp_nometrics_save
)
734 if (dst
&& (dst
->flags
&DST_HOST
)) {
735 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
738 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
739 /* This session failed to estimate rtt. Why?
740 * Probably, no packets returned in time.
743 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
744 dst
->metrics
[RTAX_RTT
-1] = 0;
748 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
750 /* If newly calculated rtt larger than stored one,
751 * store new one. Otherwise, use EWMA. Remember,
752 * rtt overestimation is always better than underestimation.
754 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
756 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
758 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
761 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
765 /* Scale deviation to rttvar fixed point */
770 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
771 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
773 dst
->metrics
[RTAX_RTTVAR
-1] -=
774 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
777 if (tp
->snd_ssthresh
>= 0xFFFF) {
778 /* Slow start still did not finish. */
779 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
780 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
781 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
782 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
783 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
784 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
785 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
786 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
787 icsk
->icsk_ca_state
== TCP_CA_Open
) {
788 /* Cong. avoidance phase, cwnd is reliable. */
789 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
790 dst
->metrics
[RTAX_SSTHRESH
-1] =
791 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
792 if (!dst_metric_locked(dst
, RTAX_CWND
))
793 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
795 /* Else slow start did not finish, cwnd is non-sense,
796 ssthresh may be also invalid.
798 if (!dst_metric_locked(dst
, RTAX_CWND
))
799 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
800 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
801 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
802 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
803 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
806 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
807 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
808 tp
->reordering
!= sysctl_tcp_reordering
)
809 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
814 /* Numbers are taken from RFC3390.
816 * John Heffner states:
818 * The RFC specifies a window of no more than 4380 bytes
819 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
820 * is a bit misleading because they use a clamp at 4380 bytes
821 * rather than use a multiplier in the relevant range.
823 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
825 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
828 if (tp
->mss_cache
> 1460)
831 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
833 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
836 /* Set slow start threshold and cwnd not falling to slow start */
837 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
839 struct tcp_sock
*tp
= tcp_sk(sk
);
840 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
842 tp
->prior_ssthresh
= 0;
844 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
847 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
848 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
849 tcp_packets_in_flight(tp
) + 1U);
850 tp
->snd_cwnd_cnt
= 0;
851 tp
->high_seq
= tp
->snd_nxt
;
852 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
853 TCP_ECN_queue_cwr(tp
);
855 tcp_set_ca_state(sk
, TCP_CA_CWR
);
860 * Packet counting of FACK is based on in-order assumptions, therefore TCP
861 * disables it when reordering is detected
863 static void tcp_disable_fack(struct tcp_sock
*tp
)
865 tp
->rx_opt
.sack_ok
&= ~2;
868 /* Take a notice that peer is sending DSACKs */
869 static void tcp_dsack_seen(struct tcp_sock
*tp
)
871 tp
->rx_opt
.sack_ok
|= 4;
874 /* Initialize metrics on socket. */
876 static void tcp_init_metrics(struct sock
*sk
)
878 struct tcp_sock
*tp
= tcp_sk(sk
);
879 struct dst_entry
*dst
= __sk_dst_get(sk
);
886 if (dst_metric_locked(dst
, RTAX_CWND
))
887 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
888 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
889 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
890 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
891 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
893 if (dst_metric(dst
, RTAX_REORDERING
) &&
894 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
895 tcp_disable_fack(tp
);
896 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
899 if (dst_metric(dst
, RTAX_RTT
) == 0)
902 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
905 /* Initial rtt is determined from SYN,SYN-ACK.
906 * The segment is small and rtt may appear much
907 * less than real one. Use per-dst memory
908 * to make it more realistic.
910 * A bit of theory. RTT is time passed after "normal" sized packet
911 * is sent until it is ACKed. In normal circumstances sending small
912 * packets force peer to delay ACKs and calculation is correct too.
913 * The algorithm is adaptive and, provided we follow specs, it
914 * NEVER underestimate RTT. BUT! If peer tries to make some clever
915 * tricks sort of "quick acks" for time long enough to decrease RTT
916 * to low value, and then abruptly stops to do it and starts to delay
917 * ACKs, wait for troubles.
919 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
920 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
921 tp
->rtt_seq
= tp
->snd_nxt
;
923 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
924 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
925 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
929 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
931 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
932 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
936 /* Play conservative. If timestamps are not
937 * supported, TCP will fail to recalculate correct
938 * rtt, if initial rto is too small. FORGET ALL AND RESET!
940 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
942 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
943 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
947 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
950 struct tcp_sock
*tp
= tcp_sk(sk
);
951 if (metric
> tp
->reordering
) {
952 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
954 /* This exciting event is worth to be remembered. 8) */
956 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
957 else if (tcp_is_reno(tp
))
958 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
959 else if (tcp_is_fack(tp
))
960 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
962 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
963 #if FASTRETRANS_DEBUG > 1
964 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
965 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
969 tp
->undo_marker
? tp
->undo_retrans
: 0);
971 tcp_disable_fack(tp
);
975 /* This procedure tags the retransmission queue when SACKs arrive.
977 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
978 * Packets in queue with these bits set are counted in variables
979 * sacked_out, retrans_out and lost_out, correspondingly.
981 * Valid combinations are:
982 * Tag InFlight Description
983 * 0 1 - orig segment is in flight.
984 * S 0 - nothing flies, orig reached receiver.
985 * L 0 - nothing flies, orig lost by net.
986 * R 2 - both orig and retransmit are in flight.
987 * L|R 1 - orig is lost, retransmit is in flight.
988 * S|R 1 - orig reached receiver, retrans is still in flight.
989 * (L|S|R is logically valid, it could occur when L|R is sacked,
990 * but it is equivalent to plain S and code short-curcuits it to S.
991 * L|S is logically invalid, it would mean -1 packet in flight 8))
993 * These 6 states form finite state machine, controlled by the following events:
994 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
995 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
996 * 3. Loss detection event of one of three flavors:
997 * A. Scoreboard estimator decided the packet is lost.
998 * A'. Reno "three dupacks" marks head of queue lost.
999 * A''. Its FACK modfication, head until snd.fack is lost.
1000 * B. SACK arrives sacking data transmitted after never retransmitted
1001 * hole was sent out.
1002 * C. SACK arrives sacking SND.NXT at the moment, when the
1003 * segment was retransmitted.
1004 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1006 * It is pleasant to note, that state diagram turns out to be commutative,
1007 * so that we are allowed not to be bothered by order of our actions,
1008 * when multiple events arrive simultaneously. (see the function below).
1010 * Reordering detection.
1011 * --------------------
1012 * Reordering metric is maximal distance, which a packet can be displaced
1013 * in packet stream. With SACKs we can estimate it:
1015 * 1. SACK fills old hole and the corresponding segment was not
1016 * ever retransmitted -> reordering. Alas, we cannot use it
1017 * when segment was retransmitted.
1018 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1019 * for retransmitted and already SACKed segment -> reordering..
1020 * Both of these heuristics are not used in Loss state, when we cannot
1021 * account for retransmits accurately.
1023 static int tcp_check_dsack(struct tcp_sock
*tp
, struct sk_buff
*ack_skb
,
1024 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1027 u32 start_seq_0
= ntohl(get_unaligned(&sp
[0].start_seq
));
1028 u32 end_seq_0
= ntohl(get_unaligned(&sp
[0].end_seq
));
1031 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1034 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
1035 } else if (num_sacks
> 1) {
1036 u32 end_seq_1
= ntohl(get_unaligned(&sp
[1].end_seq
));
1037 u32 start_seq_1
= ntohl(get_unaligned(&sp
[1].start_seq
));
1039 if (!after(end_seq_0
, end_seq_1
) &&
1040 !before(start_seq_0
, start_seq_1
)) {
1043 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
1047 /* D-SACK for already forgotten data... Do dumb counting. */
1049 !after(end_seq_0
, prior_snd_una
) &&
1050 after(end_seq_0
, tp
->undo_marker
))
1057 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
1059 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1060 struct tcp_sock
*tp
= tcp_sk(sk
);
1061 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1062 TCP_SKB_CB(ack_skb
)->sacked
);
1063 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
1064 struct sk_buff
*cached_skb
;
1065 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
1066 int reord
= tp
->packets_out
;
1068 u32 lost_retrans
= 0;
1070 int found_dup_sack
= 0;
1071 int cached_fack_count
;
1073 int first_sack_index
;
1075 if (!tp
->sacked_out
) {
1076 tp
->fackets_out
= 0;
1077 tp
->highest_sack
= tp
->snd_una
;
1079 prior_fackets
= tp
->fackets_out
;
1081 found_dup_sack
= tcp_check_dsack(tp
, ack_skb
, sp
,
1082 num_sacks
, prior_snd_una
);
1084 flag
|= FLAG_DSACKING_ACK
;
1086 /* Eliminate too old ACKs, but take into
1087 * account more or less fresh ones, they can
1088 * contain valid SACK info.
1090 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1094 * if the only SACK change is the increase of the end_seq of
1095 * the first block then only apply that SACK block
1096 * and use retrans queue hinting otherwise slowpath */
1098 for (i
= 0; i
< num_sacks
; i
++) {
1099 __be32 start_seq
= sp
[i
].start_seq
;
1100 __be32 end_seq
= sp
[i
].end_seq
;
1103 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
1106 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
1107 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1110 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1111 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1113 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1114 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1115 tp
->recv_sack_cache
[i
].start_seq
= 0;
1116 tp
->recv_sack_cache
[i
].end_seq
= 0;
1119 first_sack_index
= 0;
1124 tp
->fastpath_skb_hint
= NULL
;
1126 /* order SACK blocks to allow in order walk of the retrans queue */
1127 for (i
= num_sacks
-1; i
> 0; i
--) {
1128 for (j
= 0; j
< i
; j
++){
1129 if (after(ntohl(sp
[j
].start_seq
),
1130 ntohl(sp
[j
+1].start_seq
))){
1131 struct tcp_sack_block_wire tmp
;
1137 /* Track where the first SACK block goes to */
1138 if (j
== first_sack_index
)
1139 first_sack_index
= j
+1;
1146 /* clear flag as used for different purpose in following code */
1149 /* Use SACK fastpath hint if valid */
1150 cached_skb
= tp
->fastpath_skb_hint
;
1151 cached_fack_count
= tp
->fastpath_cnt_hint
;
1153 cached_skb
= tcp_write_queue_head(sk
);
1154 cached_fack_count
= 0;
1157 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1158 struct sk_buff
*skb
;
1159 __u32 start_seq
= ntohl(sp
->start_seq
);
1160 __u32 end_seq
= ntohl(sp
->end_seq
);
1162 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1165 fack_count
= cached_fack_count
;
1167 /* Event "B" in the comment above. */
1168 if (after(end_seq
, tp
->high_seq
))
1169 flag
|= FLAG_DATA_LOST
;
1171 tcp_for_write_queue_from(skb
, sk
) {
1172 int in_sack
, pcount
;
1175 if (skb
== tcp_send_head(sk
))
1179 cached_fack_count
= fack_count
;
1180 if (i
== first_sack_index
) {
1181 tp
->fastpath_skb_hint
= skb
;
1182 tp
->fastpath_cnt_hint
= fack_count
;
1185 /* The retransmission queue is always in order, so
1186 * we can short-circuit the walk early.
1188 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1191 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1192 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1194 pcount
= tcp_skb_pcount(skb
);
1196 if (pcount
> 1 && !in_sack
&&
1197 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1198 unsigned int pkt_len
;
1200 in_sack
= !after(start_seq
,
1201 TCP_SKB_CB(skb
)->seq
);
1204 pkt_len
= (start_seq
-
1205 TCP_SKB_CB(skb
)->seq
);
1207 pkt_len
= (end_seq
-
1208 TCP_SKB_CB(skb
)->seq
);
1209 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1211 pcount
= tcp_skb_pcount(skb
);
1214 fack_count
+= pcount
;
1216 sacked
= TCP_SKB_CB(skb
)->sacked
;
1218 /* Account D-SACK for retransmitted packet. */
1219 if ((dup_sack
&& in_sack
) &&
1220 (sacked
& TCPCB_RETRANS
) &&
1221 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1224 /* The frame is ACKed. */
1225 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1226 if (sacked
&TCPCB_RETRANS
) {
1227 if ((dup_sack
&& in_sack
) &&
1228 (sacked
&TCPCB_SACKED_ACKED
))
1229 reord
= min(fack_count
, reord
);
1231 /* If it was in a hole, we detected reordering. */
1232 if (fack_count
< prior_fackets
&&
1233 !(sacked
&TCPCB_SACKED_ACKED
))
1234 reord
= min(fack_count
, reord
);
1237 /* Nothing to do; acked frame is about to be dropped. */
1241 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1242 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1243 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1244 lost_retrans
= end_seq
;
1249 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1250 if (sacked
& TCPCB_SACKED_RETRANS
) {
1251 /* If the segment is not tagged as lost,
1252 * we do not clear RETRANS, believing
1253 * that retransmission is still in flight.
1255 if (sacked
& TCPCB_LOST
) {
1256 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1257 tp
->lost_out
-= tcp_skb_pcount(skb
);
1258 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1260 /* clear lost hint */
1261 tp
->retransmit_skb_hint
= NULL
;
1264 /* New sack for not retransmitted frame,
1265 * which was in hole. It is reordering.
1267 if (!(sacked
& TCPCB_RETRANS
) &&
1268 fack_count
< prior_fackets
)
1269 reord
= min(fack_count
, reord
);
1271 if (sacked
& TCPCB_LOST
) {
1272 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1273 tp
->lost_out
-= tcp_skb_pcount(skb
);
1275 /* clear lost hint */
1276 tp
->retransmit_skb_hint
= NULL
;
1278 /* SACK enhanced F-RTO detection.
1279 * Set flag if and only if non-rexmitted
1280 * segments below frto_highmark are
1281 * SACKed (RFC4138; Appendix B).
1282 * Clearing correct due to in-order walk
1284 if (after(end_seq
, tp
->frto_highmark
)) {
1285 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1287 if (!(sacked
& TCPCB_RETRANS
))
1288 flag
|= FLAG_ONLY_ORIG_SACKED
;
1292 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1293 flag
|= FLAG_DATA_SACKED
;
1294 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1296 if (fack_count
> tp
->fackets_out
)
1297 tp
->fackets_out
= fack_count
;
1299 if (after(TCP_SKB_CB(skb
)->seq
,
1301 tp
->highest_sack
= TCP_SKB_CB(skb
)->seq
;
1303 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1304 reord
= min(fack_count
, reord
);
1307 /* D-SACK. We can detect redundant retransmission
1308 * in S|R and plain R frames and clear it.
1309 * undo_retrans is decreased above, L|R frames
1310 * are accounted above as well.
1313 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1314 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1315 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1316 tp
->retransmit_skb_hint
= NULL
;
1321 /* Check for lost retransmit. This superb idea is
1322 * borrowed from "ratehalving". Event "C".
1323 * Later note: FACK people cheated me again 8),
1324 * we have to account for reordering! Ugly,
1327 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1328 struct sk_buff
*skb
;
1330 tcp_for_write_queue(skb
, sk
) {
1331 if (skb
== tcp_send_head(sk
))
1333 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1335 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1337 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1338 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1340 !before(lost_retrans
,
1341 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1343 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1344 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1346 /* clear lost hint */
1347 tp
->retransmit_skb_hint
= NULL
;
1349 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1350 tp
->lost_out
+= tcp_skb_pcount(skb
);
1351 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1352 flag
|= FLAG_DATA_SACKED
;
1353 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1359 tcp_verify_left_out(tp
);
1361 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
&&
1362 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1363 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1365 #if FASTRETRANS_DEBUG > 0
1366 BUG_TRAP((int)tp
->sacked_out
>= 0);
1367 BUG_TRAP((int)tp
->lost_out
>= 0);
1368 BUG_TRAP((int)tp
->retrans_out
>= 0);
1369 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1374 /* F-RTO can only be used if TCP has never retransmitted anything other than
1375 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1377 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1379 struct tcp_sock
*tp
= tcp_sk(sk
);
1382 holes
= max(tp
->lost_out
, 1U);
1383 holes
= min(holes
, tp
->packets_out
);
1385 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1386 tp
->sacked_out
= tp
->packets_out
- holes
;
1387 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1391 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1393 static void tcp_add_reno_sack(struct sock
*sk
)
1395 struct tcp_sock
*tp
= tcp_sk(sk
);
1397 tcp_check_reno_reordering(sk
, 0);
1398 tcp_verify_left_out(tp
);
1401 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1403 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1405 struct tcp_sock
*tp
= tcp_sk(sk
);
1408 /* One ACK acked hole. The rest eat duplicate ACKs. */
1409 if (acked
-1 >= tp
->sacked_out
)
1412 tp
->sacked_out
-= acked
-1;
1414 tcp_check_reno_reordering(sk
, acked
);
1415 tcp_verify_left_out(tp
);
1418 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1423 int tcp_use_frto(struct sock
*sk
)
1425 const struct tcp_sock
*tp
= tcp_sk(sk
);
1426 struct sk_buff
*skb
;
1428 if (!sysctl_tcp_frto
)
1434 /* Avoid expensive walking of rexmit queue if possible */
1435 if (tp
->retrans_out
> 1)
1438 skb
= tcp_write_queue_head(sk
);
1439 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1440 tcp_for_write_queue_from(skb
, sk
) {
1441 if (skb
== tcp_send_head(sk
))
1443 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1445 /* Short-circuit when first non-SACKed skb has been checked */
1446 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1452 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1453 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1454 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1455 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1456 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1457 * bits are handled if the Loss state is really to be entered (in
1458 * tcp_enter_frto_loss).
1460 * Do like tcp_enter_loss() would; when RTO expires the second time it
1462 * "Reduce ssthresh if it has not yet been made inside this window."
1464 void tcp_enter_frto(struct sock
*sk
)
1466 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1467 struct tcp_sock
*tp
= tcp_sk(sk
);
1468 struct sk_buff
*skb
;
1470 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1471 tp
->snd_una
== tp
->high_seq
||
1472 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1473 !icsk
->icsk_retransmits
)) {
1474 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1475 /* Our state is too optimistic in ssthresh() call because cwnd
1476 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1477 * recovery has not yet completed. Pattern would be this: RTO,
1478 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1480 * RFC4138 should be more specific on what to do, even though
1481 * RTO is quite unlikely to occur after the first Cumulative ACK
1482 * due to back-off and complexity of triggering events ...
1484 if (tp
->frto_counter
) {
1486 stored_cwnd
= tp
->snd_cwnd
;
1488 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1489 tp
->snd_cwnd
= stored_cwnd
;
1491 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1493 /* ... in theory, cong.control module could do "any tricks" in
1494 * ssthresh(), which means that ca_state, lost bits and lost_out
1495 * counter would have to be faked before the call occurs. We
1496 * consider that too expensive, unlikely and hacky, so modules
1497 * using these in ssthresh() must deal these incompatibility
1498 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1500 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1503 tp
->undo_marker
= tp
->snd_una
;
1504 tp
->undo_retrans
= 0;
1506 skb
= tcp_write_queue_head(sk
);
1507 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1508 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1509 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1511 tcp_verify_left_out(tp
);
1513 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1514 * The last condition is necessary at least in tp->frto_counter case.
1516 if (IsSackFrto() && (tp
->frto_counter
||
1517 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1518 after(tp
->high_seq
, tp
->snd_una
)) {
1519 tp
->frto_highmark
= tp
->high_seq
;
1521 tp
->frto_highmark
= tp
->snd_nxt
;
1523 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1524 tp
->high_seq
= tp
->snd_nxt
;
1525 tp
->frto_counter
= 1;
1528 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1529 * which indicates that we should follow the traditional RTO recovery,
1530 * i.e. mark everything lost and do go-back-N retransmission.
1532 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1534 struct tcp_sock
*tp
= tcp_sk(sk
);
1535 struct sk_buff
*skb
;
1538 tp
->retrans_out
= 0;
1539 if (tcp_is_reno(tp
))
1540 tcp_reset_reno_sack(tp
);
1542 tcp_for_write_queue(skb
, sk
) {
1543 if (skb
== tcp_send_head(sk
))
1546 * Count the retransmission made on RTO correctly (only when
1547 * waiting for the first ACK and did not get it)...
1549 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1550 /* For some reason this R-bit might get cleared? */
1551 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1552 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1553 /* ...enter this if branch just for the first segment */
1554 flag
|= FLAG_DATA_ACKED
;
1556 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1559 /* Don't lost mark skbs that were fwd transmitted after RTO */
1560 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) &&
1561 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
)) {
1562 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1563 tp
->lost_out
+= tcp_skb_pcount(skb
);
1566 tcp_verify_left_out(tp
);
1568 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1569 tp
->snd_cwnd_cnt
= 0;
1570 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1571 tp
->undo_marker
= 0;
1572 tp
->frto_counter
= 0;
1574 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1575 sysctl_tcp_reordering
);
1576 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1577 tp
->high_seq
= tp
->frto_highmark
;
1578 TCP_ECN_queue_cwr(tp
);
1580 clear_all_retrans_hints(tp
);
1583 void tcp_clear_retrans(struct tcp_sock
*tp
)
1585 tp
->retrans_out
= 0;
1587 tp
->fackets_out
= 0;
1591 tp
->undo_marker
= 0;
1592 tp
->undo_retrans
= 0;
1595 /* Enter Loss state. If "how" is not zero, forget all SACK information
1596 * and reset tags completely, otherwise preserve SACKs. If receiver
1597 * dropped its ofo queue, we will know this due to reneging detection.
1599 void tcp_enter_loss(struct sock
*sk
, int how
)
1601 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1602 struct tcp_sock
*tp
= tcp_sk(sk
);
1603 struct sk_buff
*skb
;
1606 /* Reduce ssthresh if it has not yet been made inside this window. */
1607 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1608 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1609 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1610 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1611 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1614 tp
->snd_cwnd_cnt
= 0;
1615 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1617 tp
->bytes_acked
= 0;
1618 tcp_clear_retrans(tp
);
1620 /* Push undo marker, if it was plain RTO and nothing
1621 * was retransmitted. */
1623 tp
->undo_marker
= tp
->snd_una
;
1625 tcp_for_write_queue(skb
, sk
) {
1626 if (skb
== tcp_send_head(sk
))
1628 cnt
+= tcp_skb_pcount(skb
);
1629 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1630 tp
->undo_marker
= 0;
1631 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1632 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1633 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1634 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1635 tp
->lost_out
+= tcp_skb_pcount(skb
);
1637 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1638 tp
->fackets_out
= cnt
;
1641 tcp_verify_left_out(tp
);
1643 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1644 sysctl_tcp_reordering
);
1645 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1646 tp
->high_seq
= tp
->snd_nxt
;
1647 TCP_ECN_queue_cwr(tp
);
1648 /* Abort FRTO algorithm if one is in progress */
1649 tp
->frto_counter
= 0;
1651 clear_all_retrans_hints(tp
);
1654 static int tcp_check_sack_reneging(struct sock
*sk
)
1656 struct sk_buff
*skb
;
1658 /* If ACK arrived pointing to a remembered SACK,
1659 * it means that our remembered SACKs do not reflect
1660 * real state of receiver i.e.
1661 * receiver _host_ is heavily congested (or buggy).
1662 * Do processing similar to RTO timeout.
1664 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1665 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1666 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1667 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1669 tcp_enter_loss(sk
, 1);
1670 icsk
->icsk_retransmits
++;
1671 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1672 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1673 icsk
->icsk_rto
, TCP_RTO_MAX
);
1679 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1681 return tcp_is_reno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1684 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1686 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1689 static inline int tcp_head_timedout(struct sock
*sk
)
1691 struct tcp_sock
*tp
= tcp_sk(sk
);
1693 return tp
->packets_out
&&
1694 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1697 /* Linux NewReno/SACK/FACK/ECN state machine.
1698 * --------------------------------------
1700 * "Open" Normal state, no dubious events, fast path.
1701 * "Disorder" In all the respects it is "Open",
1702 * but requires a bit more attention. It is entered when
1703 * we see some SACKs or dupacks. It is split of "Open"
1704 * mainly to move some processing from fast path to slow one.
1705 * "CWR" CWND was reduced due to some Congestion Notification event.
1706 * It can be ECN, ICMP source quench, local device congestion.
1707 * "Recovery" CWND was reduced, we are fast-retransmitting.
1708 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1710 * tcp_fastretrans_alert() is entered:
1711 * - each incoming ACK, if state is not "Open"
1712 * - when arrived ACK is unusual, namely:
1717 * Counting packets in flight is pretty simple.
1719 * in_flight = packets_out - left_out + retrans_out
1721 * packets_out is SND.NXT-SND.UNA counted in packets.
1723 * retrans_out is number of retransmitted segments.
1725 * left_out is number of segments left network, but not ACKed yet.
1727 * left_out = sacked_out + lost_out
1729 * sacked_out: Packets, which arrived to receiver out of order
1730 * and hence not ACKed. With SACKs this number is simply
1731 * amount of SACKed data. Even without SACKs
1732 * it is easy to give pretty reliable estimate of this number,
1733 * counting duplicate ACKs.
1735 * lost_out: Packets lost by network. TCP has no explicit
1736 * "loss notification" feedback from network (for now).
1737 * It means that this number can be only _guessed_.
1738 * Actually, it is the heuristics to predict lossage that
1739 * distinguishes different algorithms.
1741 * F.e. after RTO, when all the queue is considered as lost,
1742 * lost_out = packets_out and in_flight = retrans_out.
1744 * Essentially, we have now two algorithms counting
1747 * FACK: It is the simplest heuristics. As soon as we decided
1748 * that something is lost, we decide that _all_ not SACKed
1749 * packets until the most forward SACK are lost. I.e.
1750 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1751 * It is absolutely correct estimate, if network does not reorder
1752 * packets. And it loses any connection to reality when reordering
1753 * takes place. We use FACK by default until reordering
1754 * is suspected on the path to this destination.
1756 * NewReno: when Recovery is entered, we assume that one segment
1757 * is lost (classic Reno). While we are in Recovery and
1758 * a partial ACK arrives, we assume that one more packet
1759 * is lost (NewReno). This heuristics are the same in NewReno
1762 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1763 * deflation etc. CWND is real congestion window, never inflated, changes
1764 * only according to classic VJ rules.
1766 * Really tricky (and requiring careful tuning) part of algorithm
1767 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1768 * The first determines the moment _when_ we should reduce CWND and,
1769 * hence, slow down forward transmission. In fact, it determines the moment
1770 * when we decide that hole is caused by loss, rather than by a reorder.
1772 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1773 * holes, caused by lost packets.
1775 * And the most logically complicated part of algorithm is undo
1776 * heuristics. We detect false retransmits due to both too early
1777 * fast retransmit (reordering) and underestimated RTO, analyzing
1778 * timestamps and D-SACKs. When we detect that some segments were
1779 * retransmitted by mistake and CWND reduction was wrong, we undo
1780 * window reduction and abort recovery phase. This logic is hidden
1781 * inside several functions named tcp_try_undo_<something>.
1784 /* This function decides, when we should leave Disordered state
1785 * and enter Recovery phase, reducing congestion window.
1787 * Main question: may we further continue forward transmission
1788 * with the same cwnd?
1790 static int tcp_time_to_recover(struct sock
*sk
)
1792 struct tcp_sock
*tp
= tcp_sk(sk
);
1795 /* Do not perform any recovery during FRTO algorithm */
1796 if (tp
->frto_counter
)
1799 /* Trick#1: The loss is proven. */
1803 /* Not-A-Trick#2 : Classic rule... */
1804 if (tcp_fackets_out(tp
) > tp
->reordering
)
1807 /* Trick#3 : when we use RFC2988 timer restart, fast
1808 * retransmit can be triggered by timeout of queue head.
1810 if (tcp_head_timedout(sk
))
1813 /* Trick#4: It is still not OK... But will it be useful to delay
1816 packets_out
= tp
->packets_out
;
1817 if (packets_out
<= tp
->reordering
&&
1818 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1819 !tcp_may_send_now(sk
)) {
1820 /* We have nothing to send. This connection is limited
1821 * either by receiver window or by application.
1829 /* RFC: This is from the original, I doubt that this is necessary at all:
1830 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1831 * retransmitted past LOST markings in the first place? I'm not fully sure
1832 * about undo and end of connection cases, which can cause R without L?
1834 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
1835 struct sk_buff
*skb
)
1837 if ((tp
->retransmit_skb_hint
!= NULL
) &&
1838 before(TCP_SKB_CB(skb
)->seq
,
1839 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1840 tp
->retransmit_skb_hint
= NULL
;
1843 /* Mark head of queue up as lost. */
1844 static void tcp_mark_head_lost(struct sock
*sk
,
1845 int packets
, u32 high_seq
)
1847 struct tcp_sock
*tp
= tcp_sk(sk
);
1848 struct sk_buff
*skb
;
1851 BUG_TRAP(packets
<= tp
->packets_out
);
1852 if (tp
->lost_skb_hint
) {
1853 skb
= tp
->lost_skb_hint
;
1854 cnt
= tp
->lost_cnt_hint
;
1856 skb
= tcp_write_queue_head(sk
);
1860 tcp_for_write_queue_from(skb
, sk
) {
1861 if (skb
== tcp_send_head(sk
))
1863 /* TODO: do this better */
1864 /* this is not the most efficient way to do this... */
1865 tp
->lost_skb_hint
= skb
;
1866 tp
->lost_cnt_hint
= cnt
;
1867 cnt
+= tcp_skb_pcount(skb
);
1868 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1870 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1871 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1872 tp
->lost_out
+= tcp_skb_pcount(skb
);
1873 tcp_verify_retransmit_hint(tp
, skb
);
1876 tcp_verify_left_out(tp
);
1879 /* Account newly detected lost packet(s) */
1881 static void tcp_update_scoreboard(struct sock
*sk
)
1883 struct tcp_sock
*tp
= tcp_sk(sk
);
1885 if (tcp_is_fack(tp
)) {
1886 int lost
= tp
->fackets_out
- tp
->reordering
;
1889 tcp_mark_head_lost(sk
, lost
, tp
->high_seq
);
1891 tcp_mark_head_lost(sk
, 1, tp
->high_seq
);
1894 /* New heuristics: it is possible only after we switched
1895 * to restart timer each time when something is ACKed.
1896 * Hence, we can detect timed out packets during fast
1897 * retransmit without falling to slow start.
1899 if (!tcp_is_reno(tp
) && tcp_head_timedout(sk
)) {
1900 struct sk_buff
*skb
;
1902 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
1903 : tcp_write_queue_head(sk
);
1905 tcp_for_write_queue_from(skb
, sk
) {
1906 if (skb
== tcp_send_head(sk
))
1908 if (!tcp_skb_timedout(sk
, skb
))
1911 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1912 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1913 tp
->lost_out
+= tcp_skb_pcount(skb
);
1914 tcp_verify_retransmit_hint(tp
, skb
);
1918 tp
->scoreboard_skb_hint
= skb
;
1920 tcp_verify_left_out(tp
);
1924 /* CWND moderation, preventing bursts due to too big ACKs
1925 * in dubious situations.
1927 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
1929 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
1930 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
1931 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1934 /* Lower bound on congestion window is slow start threshold
1935 * unless congestion avoidance choice decides to overide it.
1937 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
1939 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1941 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
1944 /* Decrease cwnd each second ack. */
1945 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
1947 struct tcp_sock
*tp
= tcp_sk(sk
);
1948 int decr
= tp
->snd_cwnd_cnt
+ 1;
1950 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
1951 (tcp_is_reno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
1952 tp
->snd_cwnd_cnt
= decr
&1;
1955 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
1956 tp
->snd_cwnd
-= decr
;
1958 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
1959 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1963 /* Nothing was retransmitted or returned timestamp is less
1964 * than timestamp of the first retransmission.
1966 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
1968 return !tp
->retrans_stamp
||
1969 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
1970 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
1973 /* Undo procedures. */
1975 #if FASTRETRANS_DEBUG > 1
1976 static void DBGUNDO(struct sock
*sk
, const char *msg
)
1978 struct tcp_sock
*tp
= tcp_sk(sk
);
1979 struct inet_sock
*inet
= inet_sk(sk
);
1981 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1983 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
1984 tp
->snd_cwnd
, tcp_left_out(tp
),
1985 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
1989 #define DBGUNDO(x...) do { } while (0)
1992 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
1994 struct tcp_sock
*tp
= tcp_sk(sk
);
1996 if (tp
->prior_ssthresh
) {
1997 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1999 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2000 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2002 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
2004 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2005 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2006 TCP_ECN_withdraw_cwr(tp
);
2009 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2011 tcp_moderate_cwnd(tp
);
2012 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2014 /* There is something screwy going on with the retrans hints after
2016 clear_all_retrans_hints(tp
);
2019 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2021 return tp
->undo_marker
&&
2022 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2025 /* People celebrate: "We love our President!" */
2026 static int tcp_try_undo_recovery(struct sock
*sk
)
2028 struct tcp_sock
*tp
= tcp_sk(sk
);
2030 if (tcp_may_undo(tp
)) {
2031 /* Happy end! We did not retransmit anything
2032 * or our original transmission succeeded.
2034 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2035 tcp_undo_cwr(sk
, 1);
2036 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2037 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2039 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
2040 tp
->undo_marker
= 0;
2042 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2043 /* Hold old state until something *above* high_seq
2044 * is ACKed. For Reno it is MUST to prevent false
2045 * fast retransmits (RFC2582). SACK TCP is safe. */
2046 tcp_moderate_cwnd(tp
);
2049 tcp_set_ca_state(sk
, TCP_CA_Open
);
2053 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2054 static void tcp_try_undo_dsack(struct sock
*sk
)
2056 struct tcp_sock
*tp
= tcp_sk(sk
);
2058 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2059 DBGUNDO(sk
, "D-SACK");
2060 tcp_undo_cwr(sk
, 1);
2061 tp
->undo_marker
= 0;
2062 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2066 /* Undo during fast recovery after partial ACK. */
2068 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2070 struct tcp_sock
*tp
= tcp_sk(sk
);
2071 /* Partial ACK arrived. Force Hoe's retransmit. */
2072 int failed
= tcp_is_reno(tp
) || tp
->fackets_out
>tp
->reordering
;
2074 if (tcp_may_undo(tp
)) {
2075 /* Plain luck! Hole if filled with delayed
2076 * packet, rather than with a retransmit.
2078 if (tp
->retrans_out
== 0)
2079 tp
->retrans_stamp
= 0;
2081 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2084 tcp_undo_cwr(sk
, 0);
2085 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2087 /* So... Do not make Hoe's retransmit yet.
2088 * If the first packet was delayed, the rest
2089 * ones are most probably delayed as well.
2096 /* Undo during loss recovery after partial ACK. */
2097 static int tcp_try_undo_loss(struct sock
*sk
)
2099 struct tcp_sock
*tp
= tcp_sk(sk
);
2101 if (tcp_may_undo(tp
)) {
2102 struct sk_buff
*skb
;
2103 tcp_for_write_queue(skb
, sk
) {
2104 if (skb
== tcp_send_head(sk
))
2106 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2109 clear_all_retrans_hints(tp
);
2111 DBGUNDO(sk
, "partial loss");
2113 tcp_undo_cwr(sk
, 1);
2114 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2115 inet_csk(sk
)->icsk_retransmits
= 0;
2116 tp
->undo_marker
= 0;
2117 if (tcp_is_sack(tp
))
2118 tcp_set_ca_state(sk
, TCP_CA_Open
);
2124 static inline void tcp_complete_cwr(struct sock
*sk
)
2126 struct tcp_sock
*tp
= tcp_sk(sk
);
2127 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2128 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2129 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2132 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2134 struct tcp_sock
*tp
= tcp_sk(sk
);
2136 tcp_verify_left_out(tp
);
2138 if (tp
->retrans_out
== 0)
2139 tp
->retrans_stamp
= 0;
2142 tcp_enter_cwr(sk
, 1);
2144 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2145 int state
= TCP_CA_Open
;
2147 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2148 state
= TCP_CA_Disorder
;
2150 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2151 tcp_set_ca_state(sk
, state
);
2152 tp
->high_seq
= tp
->snd_nxt
;
2154 tcp_moderate_cwnd(tp
);
2156 tcp_cwnd_down(sk
, flag
);
2160 static void tcp_mtup_probe_failed(struct sock
*sk
)
2162 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2164 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2165 icsk
->icsk_mtup
.probe_size
= 0;
2168 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2170 struct tcp_sock
*tp
= tcp_sk(sk
);
2171 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2173 /* FIXME: breaks with very large cwnd */
2174 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2175 tp
->snd_cwnd
= tp
->snd_cwnd
*
2176 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2177 icsk
->icsk_mtup
.probe_size
;
2178 tp
->snd_cwnd_cnt
= 0;
2179 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2180 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2182 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2183 icsk
->icsk_mtup
.probe_size
= 0;
2184 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2188 /* Process an event, which can update packets-in-flight not trivially.
2189 * Main goal of this function is to calculate new estimate for left_out,
2190 * taking into account both packets sitting in receiver's buffer and
2191 * packets lost by network.
2193 * Besides that it does CWND reduction, when packet loss is detected
2194 * and changes state of machine.
2196 * It does _not_ decide what to send, it is made in function
2197 * tcp_xmit_retransmit_queue().
2200 tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2202 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2203 struct tcp_sock
*tp
= tcp_sk(sk
);
2204 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2205 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2206 (tp
->fackets_out
> tp
->reordering
));
2208 /* Some technical things:
2209 * 1. Reno does not count dupacks (sacked_out) automatically. */
2210 if (!tp
->packets_out
)
2212 /* 2. SACK counts snd_fack in packets inaccurately. */
2213 if (tp
->sacked_out
== 0)
2214 tp
->fackets_out
= 0;
2216 /* Now state machine starts.
2217 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2219 tp
->prior_ssthresh
= 0;
2221 /* B. In all the states check for reneging SACKs. */
2222 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2225 /* C. Process data loss notification, provided it is valid. */
2226 if ((flag
&FLAG_DATA_LOST
) &&
2227 before(tp
->snd_una
, tp
->high_seq
) &&
2228 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2229 tp
->fackets_out
> tp
->reordering
) {
2230 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
2231 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2234 /* D. Check consistency of the current state. */
2235 tcp_verify_left_out(tp
);
2237 /* E. Check state exit conditions. State can be terminated
2238 * when high_seq is ACKed. */
2239 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2240 BUG_TRAP(tp
->retrans_out
== 0);
2241 tp
->retrans_stamp
= 0;
2242 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2243 switch (icsk
->icsk_ca_state
) {
2245 icsk
->icsk_retransmits
= 0;
2246 if (tcp_try_undo_recovery(sk
))
2251 /* CWR is to be held something *above* high_seq
2252 * is ACKed for CWR bit to reach receiver. */
2253 if (tp
->snd_una
!= tp
->high_seq
) {
2254 tcp_complete_cwr(sk
);
2255 tcp_set_ca_state(sk
, TCP_CA_Open
);
2259 case TCP_CA_Disorder
:
2260 tcp_try_undo_dsack(sk
);
2261 if (!tp
->undo_marker
||
2262 /* For SACK case do not Open to allow to undo
2263 * catching for all duplicate ACKs. */
2264 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2265 tp
->undo_marker
= 0;
2266 tcp_set_ca_state(sk
, TCP_CA_Open
);
2270 case TCP_CA_Recovery
:
2271 if (tcp_is_reno(tp
))
2272 tcp_reset_reno_sack(tp
);
2273 if (tcp_try_undo_recovery(sk
))
2275 tcp_complete_cwr(sk
);
2280 /* F. Process state. */
2281 switch (icsk
->icsk_ca_state
) {
2282 case TCP_CA_Recovery
:
2283 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2284 if (tcp_is_reno(tp
) && is_dupack
)
2285 tcp_add_reno_sack(sk
);
2287 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2290 if (flag
&FLAG_DATA_ACKED
)
2291 icsk
->icsk_retransmits
= 0;
2292 if (!tcp_try_undo_loss(sk
)) {
2293 tcp_moderate_cwnd(tp
);
2294 tcp_xmit_retransmit_queue(sk
);
2297 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2299 /* Loss is undone; fall through to processing in Open state. */
2301 if (tcp_is_reno(tp
)) {
2302 if (flag
& FLAG_SND_UNA_ADVANCED
)
2303 tcp_reset_reno_sack(tp
);
2305 tcp_add_reno_sack(sk
);
2308 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2309 tcp_try_undo_dsack(sk
);
2311 if (!tcp_time_to_recover(sk
)) {
2312 tcp_try_to_open(sk
, flag
);
2316 /* MTU probe failure: don't reduce cwnd */
2317 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2318 icsk
->icsk_mtup
.probe_size
&&
2319 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2320 tcp_mtup_probe_failed(sk
);
2321 /* Restores the reduction we did in tcp_mtup_probe() */
2323 tcp_simple_retransmit(sk
);
2327 /* Otherwise enter Recovery state */
2329 if (tcp_is_reno(tp
))
2330 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2332 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2334 tp
->high_seq
= tp
->snd_nxt
;
2335 tp
->prior_ssthresh
= 0;
2336 tp
->undo_marker
= tp
->snd_una
;
2337 tp
->undo_retrans
= tp
->retrans_out
;
2339 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2340 if (!(flag
&FLAG_ECE
))
2341 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2342 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2343 TCP_ECN_queue_cwr(tp
);
2346 tp
->bytes_acked
= 0;
2347 tp
->snd_cwnd_cnt
= 0;
2348 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2351 if (do_lost
|| tcp_head_timedout(sk
))
2352 tcp_update_scoreboard(sk
);
2353 tcp_cwnd_down(sk
, flag
);
2354 tcp_xmit_retransmit_queue(sk
);
2357 /* Read draft-ietf-tcplw-high-performance before mucking
2358 * with this code. (Supersedes RFC1323)
2360 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2362 /* RTTM Rule: A TSecr value received in a segment is used to
2363 * update the averaged RTT measurement only if the segment
2364 * acknowledges some new data, i.e., only if it advances the
2365 * left edge of the send window.
2367 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2368 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2370 * Changed: reset backoff as soon as we see the first valid sample.
2371 * If we do not, we get strongly overestimated rto. With timestamps
2372 * samples are accepted even from very old segments: f.e., when rtt=1
2373 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2374 * answer arrives rto becomes 120 seconds! If at least one of segments
2375 * in window is lost... Voila. --ANK (010210)
2377 struct tcp_sock
*tp
= tcp_sk(sk
);
2378 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2379 tcp_rtt_estimator(sk
, seq_rtt
);
2381 inet_csk(sk
)->icsk_backoff
= 0;
2385 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2387 /* We don't have a timestamp. Can only use
2388 * packets that are not retransmitted to determine
2389 * rtt estimates. Also, we must not reset the
2390 * backoff for rto until we get a non-retransmitted
2391 * packet. This allows us to deal with a situation
2392 * where the network delay has increased suddenly.
2393 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2396 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2399 tcp_rtt_estimator(sk
, seq_rtt
);
2401 inet_csk(sk
)->icsk_backoff
= 0;
2405 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2408 const struct tcp_sock
*tp
= tcp_sk(sk
);
2409 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2410 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2411 tcp_ack_saw_tstamp(sk
, flag
);
2412 else if (seq_rtt
>= 0)
2413 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2416 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
,
2417 u32 in_flight
, int good
)
2419 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2420 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
, good
);
2421 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2424 /* Restart timer after forward progress on connection.
2425 * RFC2988 recommends to restart timer to now+rto.
2428 static void tcp_ack_packets_out(struct sock
*sk
)
2430 struct tcp_sock
*tp
= tcp_sk(sk
);
2432 if (!tp
->packets_out
) {
2433 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2435 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2439 static int tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
,
2440 __u32 now
, __s32
*seq_rtt
)
2442 struct tcp_sock
*tp
= tcp_sk(sk
);
2443 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2444 __u32 seq
= tp
->snd_una
;
2445 __u32 packets_acked
;
2448 /* If we get here, the whole TSO packet has not been
2451 BUG_ON(!after(scb
->end_seq
, seq
));
2453 packets_acked
= tcp_skb_pcount(skb
);
2454 if (tcp_trim_head(sk
, skb
, seq
- scb
->seq
))
2456 packets_acked
-= tcp_skb_pcount(skb
);
2458 if (packets_acked
) {
2459 __u8 sacked
= scb
->sacked
;
2461 acked
|= FLAG_DATA_ACKED
;
2463 if (sacked
& TCPCB_RETRANS
) {
2464 if (sacked
& TCPCB_SACKED_RETRANS
)
2465 tp
->retrans_out
-= packets_acked
;
2466 acked
|= FLAG_RETRANS_DATA_ACKED
;
2468 } else if (*seq_rtt
< 0)
2469 *seq_rtt
= now
- scb
->when
;
2470 if (sacked
& TCPCB_SACKED_ACKED
)
2471 tp
->sacked_out
-= packets_acked
;
2472 if (sacked
& TCPCB_LOST
)
2473 tp
->lost_out
-= packets_acked
;
2474 if (sacked
& TCPCB_URG
) {
2476 !before(seq
, tp
->snd_up
))
2479 } else if (*seq_rtt
< 0)
2480 *seq_rtt
= now
- scb
->when
;
2482 if (tp
->fackets_out
) {
2483 __u32 dval
= min(tp
->fackets_out
, packets_acked
);
2484 tp
->fackets_out
-= dval
;
2486 /* hint's skb might be NULL but we don't need to care */
2487 tp
->fastpath_cnt_hint
-= min_t(u32
, packets_acked
,
2488 tp
->fastpath_cnt_hint
);
2489 tp
->packets_out
-= packets_acked
;
2491 BUG_ON(tcp_skb_pcount(skb
) == 0);
2492 BUG_ON(!before(scb
->seq
, scb
->end_seq
));
2498 /* Remove acknowledged frames from the retransmission queue. */
2499 static int tcp_clean_rtx_queue(struct sock
*sk
, __s32
*seq_rtt_p
)
2501 struct tcp_sock
*tp
= tcp_sk(sk
);
2502 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2503 struct sk_buff
*skb
;
2504 __u32 now
= tcp_time_stamp
;
2506 int prior_packets
= tp
->packets_out
;
2508 ktime_t last_ackt
= net_invalid_timestamp();
2510 while ((skb
= tcp_write_queue_head(sk
)) &&
2511 skb
!= tcp_send_head(sk
)) {
2512 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2513 __u8 sacked
= scb
->sacked
;
2515 /* If our packet is before the ack sequence we can
2516 * discard it as it's confirmed to have arrived at
2519 if (after(scb
->end_seq
, tp
->snd_una
)) {
2520 if (tcp_skb_pcount(skb
) > 1 &&
2521 after(tp
->snd_una
, scb
->seq
))
2522 acked
|= tcp_tso_acked(sk
, skb
,
2527 /* Initial outgoing SYN's get put onto the write_queue
2528 * just like anything else we transmit. It is not
2529 * true data, and if we misinform our callers that
2530 * this ACK acks real data, we will erroneously exit
2531 * connection startup slow start one packet too
2532 * quickly. This is severely frowned upon behavior.
2534 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2535 acked
|= FLAG_DATA_ACKED
;
2537 acked
|= FLAG_SYN_ACKED
;
2538 tp
->retrans_stamp
= 0;
2541 /* MTU probing checks */
2542 if (icsk
->icsk_mtup
.probe_size
) {
2543 if (!after(tp
->mtu_probe
.probe_seq_end
, TCP_SKB_CB(skb
)->end_seq
)) {
2544 tcp_mtup_probe_success(sk
, skb
);
2549 if (sacked
& TCPCB_RETRANS
) {
2550 if (sacked
& TCPCB_SACKED_RETRANS
)
2551 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2552 acked
|= FLAG_RETRANS_DATA_ACKED
;
2554 } else if (seq_rtt
< 0) {
2555 seq_rtt
= now
- scb
->when
;
2556 last_ackt
= skb
->tstamp
;
2558 if (sacked
& TCPCB_SACKED_ACKED
)
2559 tp
->sacked_out
-= tcp_skb_pcount(skb
);
2560 if (sacked
& TCPCB_LOST
)
2561 tp
->lost_out
-= tcp_skb_pcount(skb
);
2562 if (sacked
& TCPCB_URG
) {
2564 !before(scb
->end_seq
, tp
->snd_up
))
2567 } else if (seq_rtt
< 0) {
2568 seq_rtt
= now
- scb
->when
;
2569 last_ackt
= skb
->tstamp
;
2571 tcp_dec_pcount_approx(&tp
->fackets_out
, skb
);
2572 tp
->packets_out
-= tcp_skb_pcount(skb
);
2573 tcp_unlink_write_queue(skb
, sk
);
2574 sk_stream_free_skb(sk
, skb
);
2575 clear_all_retrans_hints(tp
);
2578 if (acked
&FLAG_ACKED
) {
2579 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2580 const struct tcp_congestion_ops
*ca_ops
2581 = inet_csk(sk
)->icsk_ca_ops
;
2583 tcp_ack_update_rtt(sk
, acked
, seq_rtt
);
2584 tcp_ack_packets_out(sk
);
2586 if (tcp_is_reno(tp
))
2587 tcp_remove_reno_sacks(sk
, pkts_acked
);
2589 if (ca_ops
->pkts_acked
) {
2592 /* Is the ACK triggering packet unambiguous? */
2593 if (!(acked
& FLAG_RETRANS_DATA_ACKED
)) {
2594 /* High resolution needed and available? */
2595 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2596 !ktime_equal(last_ackt
,
2597 net_invalid_timestamp()))
2598 rtt_us
= ktime_us_delta(ktime_get_real(),
2600 else if (seq_rtt
> 0)
2601 rtt_us
= jiffies_to_usecs(seq_rtt
);
2604 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2608 #if FASTRETRANS_DEBUG > 0
2609 BUG_TRAP((int)tp
->sacked_out
>= 0);
2610 BUG_TRAP((int)tp
->lost_out
>= 0);
2611 BUG_TRAP((int)tp
->retrans_out
>= 0);
2612 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2613 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2615 printk(KERN_DEBUG
"Leak l=%u %d\n",
2616 tp
->lost_out
, icsk
->icsk_ca_state
);
2619 if (tp
->sacked_out
) {
2620 printk(KERN_DEBUG
"Leak s=%u %d\n",
2621 tp
->sacked_out
, icsk
->icsk_ca_state
);
2624 if (tp
->retrans_out
) {
2625 printk(KERN_DEBUG
"Leak r=%u %d\n",
2626 tp
->retrans_out
, icsk
->icsk_ca_state
);
2627 tp
->retrans_out
= 0;
2631 *seq_rtt_p
= seq_rtt
;
2635 static void tcp_ack_probe(struct sock
*sk
)
2637 const struct tcp_sock
*tp
= tcp_sk(sk
);
2638 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2640 /* Was it a usable window open? */
2642 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2643 tp
->snd_una
+ tp
->snd_wnd
)) {
2644 icsk
->icsk_backoff
= 0;
2645 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2646 /* Socket must be waked up by subsequent tcp_data_snd_check().
2647 * This function is not for random using!
2650 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2651 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2656 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2658 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2659 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2662 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2664 const struct tcp_sock
*tp
= tcp_sk(sk
);
2665 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2666 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2669 /* Check that window update is acceptable.
2670 * The function assumes that snd_una<=ack<=snd_next.
2672 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2673 const u32 ack_seq
, const u32 nwin
)
2675 return (after(ack
, tp
->snd_una
) ||
2676 after(ack_seq
, tp
->snd_wl1
) ||
2677 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2680 /* Update our send window.
2682 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2683 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2685 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2688 struct tcp_sock
*tp
= tcp_sk(sk
);
2690 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2692 if (likely(!tcp_hdr(skb
)->syn
))
2693 nwin
<<= tp
->rx_opt
.snd_wscale
;
2695 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2696 flag
|= FLAG_WIN_UPDATE
;
2697 tcp_update_wl(tp
, ack
, ack_seq
);
2699 if (tp
->snd_wnd
!= nwin
) {
2702 /* Note, it is the only place, where
2703 * fast path is recovered for sending TCP.
2706 tcp_fast_path_check(sk
);
2708 if (nwin
> tp
->max_window
) {
2709 tp
->max_window
= nwin
;
2710 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2720 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2721 * continue in congestion avoidance.
2723 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2725 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2726 tp
->snd_cwnd_cnt
= 0;
2727 TCP_ECN_queue_cwr(tp
);
2728 tcp_moderate_cwnd(tp
);
2731 /* A conservative spurious RTO response algorithm: reduce cwnd using
2732 * rate halving and continue in congestion avoidance.
2734 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
2736 tcp_enter_cwr(sk
, 0);
2739 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
2742 tcp_ratehalving_spur_to_response(sk
);
2744 tcp_undo_cwr(sk
, 1);
2747 /* F-RTO spurious RTO detection algorithm (RFC4138)
2749 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2750 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2751 * window (but not to or beyond highest sequence sent before RTO):
2752 * On First ACK, send two new segments out.
2753 * On Second ACK, RTO was likely spurious. Do spurious response (response
2754 * algorithm is not part of the F-RTO detection algorithm
2755 * given in RFC4138 but can be selected separately).
2756 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2757 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2758 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2759 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2761 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2762 * original window even after we transmit two new data segments.
2765 * on first step, wait until first cumulative ACK arrives, then move to
2766 * the second step. In second step, the next ACK decides.
2768 * F-RTO is implemented (mainly) in four functions:
2769 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2770 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2771 * called when tcp_use_frto() showed green light
2772 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2773 * - tcp_enter_frto_loss() is called if there is not enough evidence
2774 * to prove that the RTO is indeed spurious. It transfers the control
2775 * from F-RTO to the conventional RTO recovery
2777 static int tcp_process_frto(struct sock
*sk
, int flag
)
2779 struct tcp_sock
*tp
= tcp_sk(sk
);
2781 tcp_verify_left_out(tp
);
2783 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2784 if (flag
&FLAG_DATA_ACKED
)
2785 inet_csk(sk
)->icsk_retransmits
= 0;
2787 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2788 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
2792 if (!IsSackFrto() || tcp_is_reno(tp
)) {
2793 /* RFC4138 shortcoming in step 2; should also have case c):
2794 * ACK isn't duplicate nor advances window, e.g., opposite dir
2797 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2800 if (!(flag
&FLAG_DATA_ACKED
)) {
2801 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2806 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2807 /* Prevent sending of new data. */
2808 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2809 tcp_packets_in_flight(tp
));
2813 if ((tp
->frto_counter
>= 2) &&
2814 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2815 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2816 /* RFC4138 shortcoming (see comment above) */
2817 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2820 tcp_enter_frto_loss(sk
, 3, flag
);
2825 if (tp
->frto_counter
== 1) {
2826 /* Sending of the next skb must be allowed or no FRTO */
2827 if (!tcp_send_head(sk
) ||
2828 after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2829 tp
->snd_una
+ tp
->snd_wnd
)) {
2830 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3),
2835 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2836 tp
->frto_counter
= 2;
2839 switch (sysctl_tcp_frto_response
) {
2841 tcp_undo_spur_to_response(sk
, flag
);
2844 tcp_conservative_spur_to_response(tp
);
2847 tcp_ratehalving_spur_to_response(sk
);
2850 tp
->frto_counter
= 0;
2855 /* This routine deals with incoming acks, but not outgoing ones. */
2856 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2858 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2859 struct tcp_sock
*tp
= tcp_sk(sk
);
2860 u32 prior_snd_una
= tp
->snd_una
;
2861 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2862 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2863 u32 prior_in_flight
;
2868 /* If the ack is newer than sent or older than previous acks
2869 * then we can probably ignore it.
2871 if (after(ack
, tp
->snd_nxt
))
2872 goto uninteresting_ack
;
2874 if (before(ack
, prior_snd_una
))
2877 if (after(ack
, prior_snd_una
))
2878 flag
|= FLAG_SND_UNA_ADVANCED
;
2880 if (sysctl_tcp_abc
) {
2881 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2882 tp
->bytes_acked
+= ack
- prior_snd_una
;
2883 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2884 /* we assume just one segment left network */
2885 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2888 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2889 /* Window is constant, pure forward advance.
2890 * No more checks are required.
2891 * Note, we use the fact that SND.UNA>=SND.WL2.
2893 tcp_update_wl(tp
, ack
, ack_seq
);
2895 flag
|= FLAG_WIN_UPDATE
;
2897 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2899 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2901 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2904 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2906 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
2908 if (TCP_SKB_CB(skb
)->sacked
)
2909 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2911 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
2914 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
2917 /* We passed data and got it acked, remove any soft error
2918 * log. Something worked...
2920 sk
->sk_err_soft
= 0;
2921 tp
->rcv_tstamp
= tcp_time_stamp
;
2922 prior_packets
= tp
->packets_out
;
2926 prior_in_flight
= tcp_packets_in_flight(tp
);
2928 /* See if we can take anything off of the retransmit queue. */
2929 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
2931 if (tp
->frto_counter
)
2932 frto_cwnd
= tcp_process_frto(sk
, flag
);
2934 if (tcp_ack_is_dubious(sk
, flag
)) {
2935 /* Advance CWND, if state allows this. */
2936 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
2937 tcp_may_raise_cwnd(sk
, flag
))
2938 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 0);
2939 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
, flag
);
2941 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
2942 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 1);
2945 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
2946 dst_confirm(sk
->sk_dst_cache
);
2951 icsk
->icsk_probes_out
= 0;
2953 /* If this ack opens up a zero window, clear backoff. It was
2954 * being used to time the probes, and is probably far higher than
2955 * it needs to be for normal retransmission.
2957 if (tcp_send_head(sk
))
2962 if (TCP_SKB_CB(skb
)->sacked
)
2963 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
2966 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
2971 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2972 * But, this can also be called on packets in the established flow when
2973 * the fast version below fails.
2975 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
2978 struct tcphdr
*th
= tcp_hdr(skb
);
2979 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
2981 ptr
= (unsigned char *)(th
+ 1);
2982 opt_rx
->saw_tstamp
= 0;
2984 while (length
> 0) {
2991 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
2996 if (opsize
< 2) /* "silly options" */
2998 if (opsize
> length
)
2999 return; /* don't parse partial options */
3002 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3003 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
3005 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
3006 in_mss
= opt_rx
->user_mss
;
3007 opt_rx
->mss_clamp
= in_mss
;
3012 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
3013 if (sysctl_tcp_window_scaling
) {
3014 __u8 snd_wscale
= *(__u8
*) ptr
;
3015 opt_rx
->wscale_ok
= 1;
3016 if (snd_wscale
> 14) {
3017 if (net_ratelimit())
3018 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3019 "scaling value %d >14 received.\n",
3023 opt_rx
->snd_wscale
= snd_wscale
;
3026 case TCPOPT_TIMESTAMP
:
3027 if (opsize
==TCPOLEN_TIMESTAMP
) {
3028 if ((estab
&& opt_rx
->tstamp_ok
) ||
3029 (!estab
&& sysctl_tcp_timestamps
)) {
3030 opt_rx
->saw_tstamp
= 1;
3031 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
3032 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
3036 case TCPOPT_SACK_PERM
:
3037 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
3038 if (sysctl_tcp_sack
) {
3039 opt_rx
->sack_ok
= 1;
3040 tcp_sack_reset(opt_rx
);
3046 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3047 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3049 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3052 #ifdef CONFIG_TCP_MD5SIG
3055 * The MD5 Hash has already been
3056 * checked (see tcp_v{4,6}_do_rcv()).
3068 /* Fast parse options. This hopes to only see timestamps.
3069 * If it is wrong it falls back on tcp_parse_options().
3071 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3072 struct tcp_sock
*tp
)
3074 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3075 tp
->rx_opt
.saw_tstamp
= 0;
3077 } else if (tp
->rx_opt
.tstamp_ok
&&
3078 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3079 __be32
*ptr
= (__be32
*)(th
+ 1);
3080 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3081 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3082 tp
->rx_opt
.saw_tstamp
= 1;
3084 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3086 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3090 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3094 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3096 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3097 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3100 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3102 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3103 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3104 * extra check below makes sure this can only happen
3105 * for pure ACK frames. -DaveM
3107 * Not only, also it occurs for expired timestamps.
3110 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3111 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3112 tcp_store_ts_recent(tp
);
3116 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3118 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3119 * it can pass through stack. So, the following predicate verifies that
3120 * this segment is not used for anything but congestion avoidance or
3121 * fast retransmit. Moreover, we even are able to eliminate most of such
3122 * second order effects, if we apply some small "replay" window (~RTO)
3123 * to timestamp space.
3125 * All these measures still do not guarantee that we reject wrapped ACKs
3126 * on networks with high bandwidth, when sequence space is recycled fastly,
3127 * but it guarantees that such events will be very rare and do not affect
3128 * connection seriously. This doesn't look nice, but alas, PAWS is really
3131 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3132 * states that events when retransmit arrives after original data are rare.
3133 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3134 * the biggest problem on large power networks even with minor reordering.
3135 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3136 * up to bandwidth of 18Gigabit/sec. 8) ]
3139 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3141 struct tcp_sock
*tp
= tcp_sk(sk
);
3142 struct tcphdr
*th
= tcp_hdr(skb
);
3143 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3144 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3146 return (/* 1. Pure ACK with correct sequence number. */
3147 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3149 /* 2. ... and duplicate ACK. */
3150 ack
== tp
->snd_una
&&
3152 /* 3. ... and does not update window. */
3153 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3155 /* 4. ... and sits in replay window. */
3156 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3159 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3161 const struct tcp_sock
*tp
= tcp_sk(sk
);
3162 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3163 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3164 !tcp_disordered_ack(sk
, skb
));
3167 /* Check segment sequence number for validity.
3169 * Segment controls are considered valid, if the segment
3170 * fits to the window after truncation to the window. Acceptability
3171 * of data (and SYN, FIN, of course) is checked separately.
3172 * See tcp_data_queue(), for example.
3174 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3175 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3176 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3177 * (borrowed from freebsd)
3180 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3182 return !before(end_seq
, tp
->rcv_wup
) &&
3183 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3186 /* When we get a reset we do this. */
3187 static void tcp_reset(struct sock
*sk
)
3189 /* We want the right error as BSD sees it (and indeed as we do). */
3190 switch (sk
->sk_state
) {
3192 sk
->sk_err
= ECONNREFUSED
;
3194 case TCP_CLOSE_WAIT
:
3200 sk
->sk_err
= ECONNRESET
;
3203 if (!sock_flag(sk
, SOCK_DEAD
))
3204 sk
->sk_error_report(sk
);
3210 * Process the FIN bit. This now behaves as it is supposed to work
3211 * and the FIN takes effect when it is validly part of sequence
3212 * space. Not before when we get holes.
3214 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3215 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3218 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3219 * close and we go into CLOSING (and later onto TIME-WAIT)
3221 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3223 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3225 struct tcp_sock
*tp
= tcp_sk(sk
);
3227 inet_csk_schedule_ack(sk
);
3229 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3230 sock_set_flag(sk
, SOCK_DONE
);
3232 switch (sk
->sk_state
) {
3234 case TCP_ESTABLISHED
:
3235 /* Move to CLOSE_WAIT */
3236 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3237 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3240 case TCP_CLOSE_WAIT
:
3242 /* Received a retransmission of the FIN, do
3247 /* RFC793: Remain in the LAST-ACK state. */
3251 /* This case occurs when a simultaneous close
3252 * happens, we must ack the received FIN and
3253 * enter the CLOSING state.
3256 tcp_set_state(sk
, TCP_CLOSING
);
3259 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3261 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3264 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3265 * cases we should never reach this piece of code.
3267 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3268 __FUNCTION__
, sk
->sk_state
);
3272 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3273 * Probably, we should reset in this case. For now drop them.
3275 __skb_queue_purge(&tp
->out_of_order_queue
);
3276 if (tcp_is_sack(tp
))
3277 tcp_sack_reset(&tp
->rx_opt
);
3278 sk_stream_mem_reclaim(sk
);
3280 if (!sock_flag(sk
, SOCK_DEAD
)) {
3281 sk
->sk_state_change(sk
);
3283 /* Do not send POLL_HUP for half duplex close. */
3284 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3285 sk
->sk_state
== TCP_CLOSE
)
3286 sk_wake_async(sk
, 1, POLL_HUP
);
3288 sk_wake_async(sk
, 1, POLL_IN
);
3292 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3294 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3295 if (before(seq
, sp
->start_seq
))
3296 sp
->start_seq
= seq
;
3297 if (after(end_seq
, sp
->end_seq
))
3298 sp
->end_seq
= end_seq
;
3304 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3306 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3307 if (before(seq
, tp
->rcv_nxt
))
3308 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3310 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3312 tp
->rx_opt
.dsack
= 1;
3313 tp
->duplicate_sack
[0].start_seq
= seq
;
3314 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3315 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3319 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3321 if (!tp
->rx_opt
.dsack
)
3322 tcp_dsack_set(tp
, seq
, end_seq
);
3324 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3327 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3329 struct tcp_sock
*tp
= tcp_sk(sk
);
3331 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3332 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3333 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3334 tcp_enter_quickack_mode(sk
);
3336 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3337 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3339 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3340 end_seq
= tp
->rcv_nxt
;
3341 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3348 /* These routines update the SACK block as out-of-order packets arrive or
3349 * in-order packets close up the sequence space.
3351 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3354 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3355 struct tcp_sack_block
*swalk
= sp
+1;
3357 /* See if the recent change to the first SACK eats into
3358 * or hits the sequence space of other SACK blocks, if so coalesce.
3360 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3361 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3364 /* Zap SWALK, by moving every further SACK up by one slot.
3365 * Decrease num_sacks.
3367 tp
->rx_opt
.num_sacks
--;
3368 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3369 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3373 this_sack
++, swalk
++;
3377 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3381 tmp
= sack1
->start_seq
;
3382 sack1
->start_seq
= sack2
->start_seq
;
3383 sack2
->start_seq
= tmp
;
3385 tmp
= sack1
->end_seq
;
3386 sack1
->end_seq
= sack2
->end_seq
;
3387 sack2
->end_seq
= tmp
;
3390 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3392 struct tcp_sock
*tp
= tcp_sk(sk
);
3393 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3394 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3400 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3401 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3402 /* Rotate this_sack to the first one. */
3403 for (; this_sack
>0; this_sack
--, sp
--)
3404 tcp_sack_swap(sp
, sp
-1);
3406 tcp_sack_maybe_coalesce(tp
);
3411 /* Could not find an adjacent existing SACK, build a new one,
3412 * put it at the front, and shift everyone else down. We
3413 * always know there is at least one SACK present already here.
3415 * If the sack array is full, forget about the last one.
3417 if (this_sack
>= 4) {
3419 tp
->rx_opt
.num_sacks
--;
3422 for (; this_sack
> 0; this_sack
--, sp
--)
3426 /* Build the new head SACK, and we're done. */
3427 sp
->start_seq
= seq
;
3428 sp
->end_seq
= end_seq
;
3429 tp
->rx_opt
.num_sacks
++;
3430 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3433 /* RCV.NXT advances, some SACKs should be eaten. */
3435 static void tcp_sack_remove(struct tcp_sock
*tp
)
3437 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3438 int num_sacks
= tp
->rx_opt
.num_sacks
;
3441 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3442 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3443 tp
->rx_opt
.num_sacks
= 0;
3444 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3448 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3449 /* Check if the start of the sack is covered by RCV.NXT. */
3450 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3453 /* RCV.NXT must cover all the block! */
3454 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3456 /* Zap this SACK, by moving forward any other SACKS. */
3457 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3458 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3465 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3466 tp
->rx_opt
.num_sacks
= num_sacks
;
3467 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3471 /* This one checks to see if we can put data from the
3472 * out_of_order queue into the receive_queue.
3474 static void tcp_ofo_queue(struct sock
*sk
)
3476 struct tcp_sock
*tp
= tcp_sk(sk
);
3477 __u32 dsack_high
= tp
->rcv_nxt
;
3478 struct sk_buff
*skb
;
3480 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3481 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3484 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3485 __u32 dsack
= dsack_high
;
3486 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3487 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3488 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3491 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3492 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3493 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3497 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3498 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3499 TCP_SKB_CB(skb
)->end_seq
);
3501 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3502 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3503 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3504 if (tcp_hdr(skb
)->fin
)
3505 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3509 static int tcp_prune_queue(struct sock
*sk
);
3511 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3513 struct tcphdr
*th
= tcp_hdr(skb
);
3514 struct tcp_sock
*tp
= tcp_sk(sk
);
3517 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3520 __skb_pull(skb
, th
->doff
*4);
3522 TCP_ECN_accept_cwr(tp
, skb
);
3524 if (tp
->rx_opt
.dsack
) {
3525 tp
->rx_opt
.dsack
= 0;
3526 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3527 4 - tp
->rx_opt
.tstamp_ok
);
3530 /* Queue data for delivery to the user.
3531 * Packets in sequence go to the receive queue.
3532 * Out of sequence packets to the out_of_order_queue.
3534 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3535 if (tcp_receive_window(tp
) == 0)
3538 /* Ok. In sequence. In window. */
3539 if (tp
->ucopy
.task
== current
&&
3540 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3541 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3542 int chunk
= min_t(unsigned int, skb
->len
,
3545 __set_current_state(TASK_RUNNING
);
3548 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3549 tp
->ucopy
.len
-= chunk
;
3550 tp
->copied_seq
+= chunk
;
3551 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3552 tcp_rcv_space_adjust(sk
);
3560 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3561 !sk_stream_rmem_schedule(sk
, skb
))) {
3562 if (tcp_prune_queue(sk
) < 0 ||
3563 !sk_stream_rmem_schedule(sk
, skb
))
3566 sk_stream_set_owner_r(skb
, sk
);
3567 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3569 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3571 tcp_event_data_recv(sk
, skb
);
3573 tcp_fin(skb
, sk
, th
);
3575 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3578 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3579 * gap in queue is filled.
3581 if (skb_queue_empty(&tp
->out_of_order_queue
))
3582 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3585 if (tp
->rx_opt
.num_sacks
)
3586 tcp_sack_remove(tp
);
3588 tcp_fast_path_check(sk
);
3592 else if (!sock_flag(sk
, SOCK_DEAD
))
3593 sk
->sk_data_ready(sk
, 0);
3597 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3598 /* A retransmit, 2nd most common case. Force an immediate ack. */
3599 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3600 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3603 tcp_enter_quickack_mode(sk
);
3604 inet_csk_schedule_ack(sk
);
3610 /* Out of window. F.e. zero window probe. */
3611 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3614 tcp_enter_quickack_mode(sk
);
3616 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3617 /* Partial packet, seq < rcv_next < end_seq */
3618 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3619 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3620 TCP_SKB_CB(skb
)->end_seq
);
3622 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3624 /* If window is closed, drop tail of packet. But after
3625 * remembering D-SACK for its head made in previous line.
3627 if (!tcp_receive_window(tp
))
3632 TCP_ECN_check_ce(tp
, skb
);
3634 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3635 !sk_stream_rmem_schedule(sk
, skb
)) {
3636 if (tcp_prune_queue(sk
) < 0 ||
3637 !sk_stream_rmem_schedule(sk
, skb
))
3641 /* Disable header prediction. */
3643 inet_csk_schedule_ack(sk
);
3645 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3646 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3648 sk_stream_set_owner_r(skb
, sk
);
3650 if (!skb_peek(&tp
->out_of_order_queue
)) {
3651 /* Initial out of order segment, build 1 SACK. */
3652 if (tcp_is_sack(tp
)) {
3653 tp
->rx_opt
.num_sacks
= 1;
3654 tp
->rx_opt
.dsack
= 0;
3655 tp
->rx_opt
.eff_sacks
= 1;
3656 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3657 tp
->selective_acks
[0].end_seq
=
3658 TCP_SKB_CB(skb
)->end_seq
;
3660 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3662 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3663 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3664 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3666 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3667 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3669 if (!tp
->rx_opt
.num_sacks
||
3670 tp
->selective_acks
[0].end_seq
!= seq
)
3673 /* Common case: data arrive in order after hole. */
3674 tp
->selective_acks
[0].end_seq
= end_seq
;
3678 /* Find place to insert this segment. */
3680 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3682 } while ((skb1
= skb1
->prev
) !=
3683 (struct sk_buff
*)&tp
->out_of_order_queue
);
3685 /* Do skb overlap to previous one? */
3686 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3687 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3688 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3689 /* All the bits are present. Drop. */
3691 tcp_dsack_set(tp
, seq
, end_seq
);
3694 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3695 /* Partial overlap. */
3696 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3701 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3703 /* And clean segments covered by new one as whole. */
3704 while ((skb1
= skb
->next
) !=
3705 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3706 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3707 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3708 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3711 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3712 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3717 if (tcp_is_sack(tp
))
3718 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3722 /* Collapse contiguous sequence of skbs head..tail with
3723 * sequence numbers start..end.
3724 * Segments with FIN/SYN are not collapsed (only because this
3728 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3729 struct sk_buff
*head
, struct sk_buff
*tail
,
3732 struct sk_buff
*skb
;
3734 /* First, check that queue is collapsible and find
3735 * the point where collapsing can be useful. */
3736 for (skb
= head
; skb
!= tail
; ) {
3737 /* No new bits? It is possible on ofo queue. */
3738 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3739 struct sk_buff
*next
= skb
->next
;
3740 __skb_unlink(skb
, list
);
3742 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3747 /* The first skb to collapse is:
3749 * - bloated or contains data before "start" or
3750 * overlaps to the next one.
3752 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
3753 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3754 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3755 (skb
->next
!= tail
&&
3756 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3759 /* Decided to skip this, advance start seq. */
3760 start
= TCP_SKB_CB(skb
)->end_seq
;
3763 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
3766 while (before(start
, end
)) {
3767 struct sk_buff
*nskb
;
3768 int header
= skb_headroom(skb
);
3769 int copy
= SKB_MAX_ORDER(header
, 0);
3771 /* Too big header? This can happen with IPv6. */
3774 if (end
-start
< copy
)
3776 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3780 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
3781 skb_set_network_header(nskb
, (skb_network_header(skb
) -
3783 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
3785 skb_reserve(nskb
, header
);
3786 memcpy(nskb
->head
, skb
->head
, header
);
3787 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3788 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3789 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3790 sk_stream_set_owner_r(nskb
, sk
);
3792 /* Copy data, releasing collapsed skbs. */
3794 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3795 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3799 size
= min(copy
, size
);
3800 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3802 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3806 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3807 struct sk_buff
*next
= skb
->next
;
3808 __skb_unlink(skb
, list
);
3810 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3813 tcp_hdr(skb
)->syn
||
3821 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3822 * and tcp_collapse() them until all the queue is collapsed.
3824 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3826 struct tcp_sock
*tp
= tcp_sk(sk
);
3827 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3828 struct sk_buff
*head
;
3834 start
= TCP_SKB_CB(skb
)->seq
;
3835 end
= TCP_SKB_CB(skb
)->end_seq
;
3841 /* Segment is terminated when we see gap or when
3842 * we are at the end of all the queue. */
3843 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3844 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3845 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3846 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3847 head
, skb
, start
, end
);
3849 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3851 /* Start new segment */
3852 start
= TCP_SKB_CB(skb
)->seq
;
3853 end
= TCP_SKB_CB(skb
)->end_seq
;
3855 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3856 start
= TCP_SKB_CB(skb
)->seq
;
3857 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3858 end
= TCP_SKB_CB(skb
)->end_seq
;
3863 /* Reduce allocated memory if we can, trying to get
3864 * the socket within its memory limits again.
3866 * Return less than zero if we should start dropping frames
3867 * until the socket owning process reads some of the data
3868 * to stabilize the situation.
3870 static int tcp_prune_queue(struct sock
*sk
)
3872 struct tcp_sock
*tp
= tcp_sk(sk
);
3874 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3876 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3878 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3879 tcp_clamp_window(sk
);
3880 else if (tcp_memory_pressure
)
3881 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3883 tcp_collapse_ofo_queue(sk
);
3884 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3885 sk
->sk_receive_queue
.next
,
3886 (struct sk_buff
*)&sk
->sk_receive_queue
,
3887 tp
->copied_seq
, tp
->rcv_nxt
);
3888 sk_stream_mem_reclaim(sk
);
3890 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3893 /* Collapsing did not help, destructive actions follow.
3894 * This must not ever occur. */
3896 /* First, purge the out_of_order queue. */
3897 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3898 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3899 __skb_queue_purge(&tp
->out_of_order_queue
);
3901 /* Reset SACK state. A conforming SACK implementation will
3902 * do the same at a timeout based retransmit. When a connection
3903 * is in a sad state like this, we care only about integrity
3904 * of the connection not performance.
3906 if (tcp_is_sack(tp
))
3907 tcp_sack_reset(&tp
->rx_opt
);
3908 sk_stream_mem_reclaim(sk
);
3911 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3914 /* If we are really being abused, tell the caller to silently
3915 * drop receive data on the floor. It will get retransmitted
3916 * and hopefully then we'll have sufficient space.
3918 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
3920 /* Massive buffer overcommit. */
3926 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3927 * As additional protections, we do not touch cwnd in retransmission phases,
3928 * and if application hit its sndbuf limit recently.
3930 void tcp_cwnd_application_limited(struct sock
*sk
)
3932 struct tcp_sock
*tp
= tcp_sk(sk
);
3934 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
3935 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
3936 /* Limited by application or receiver window. */
3937 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
3938 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
3939 if (win_used
< tp
->snd_cwnd
) {
3940 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
3941 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
3943 tp
->snd_cwnd_used
= 0;
3945 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3948 static int tcp_should_expand_sndbuf(struct sock
*sk
)
3950 struct tcp_sock
*tp
= tcp_sk(sk
);
3952 /* If the user specified a specific send buffer setting, do
3955 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
3958 /* If we are under global TCP memory pressure, do not expand. */
3959 if (tcp_memory_pressure
)
3962 /* If we are under soft global TCP memory pressure, do not expand. */
3963 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
3966 /* If we filled the congestion window, do not expand. */
3967 if (tp
->packets_out
>= tp
->snd_cwnd
)
3973 /* When incoming ACK allowed to free some skb from write_queue,
3974 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3975 * on the exit from tcp input handler.
3977 * PROBLEM: sndbuf expansion does not work well with largesend.
3979 static void tcp_new_space(struct sock
*sk
)
3981 struct tcp_sock
*tp
= tcp_sk(sk
);
3983 if (tcp_should_expand_sndbuf(sk
)) {
3984 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
3985 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
3986 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
3987 tp
->reordering
+ 1);
3988 sndmem
*= 2*demanded
;
3989 if (sndmem
> sk
->sk_sndbuf
)
3990 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
3991 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
3994 sk
->sk_write_space(sk
);
3997 static void tcp_check_space(struct sock
*sk
)
3999 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4000 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4001 if (sk
->sk_socket
&&
4002 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4007 static inline void tcp_data_snd_check(struct sock
*sk
)
4009 tcp_push_pending_frames(sk
);
4010 tcp_check_space(sk
);
4014 * Check if sending an ack is needed.
4016 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4018 struct tcp_sock
*tp
= tcp_sk(sk
);
4020 /* More than one full frame received... */
4021 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4022 /* ... and right edge of window advances far enough.
4023 * (tcp_recvmsg() will send ACK otherwise). Or...
4025 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4026 /* We ACK each frame or... */
4027 tcp_in_quickack_mode(sk
) ||
4028 /* We have out of order data. */
4030 skb_peek(&tp
->out_of_order_queue
))) {
4031 /* Then ack it now */
4034 /* Else, send delayed ack. */
4035 tcp_send_delayed_ack(sk
);
4039 static inline void tcp_ack_snd_check(struct sock
*sk
)
4041 if (!inet_csk_ack_scheduled(sk
)) {
4042 /* We sent a data segment already. */
4045 __tcp_ack_snd_check(sk
, 1);
4049 * This routine is only called when we have urgent data
4050 * signaled. Its the 'slow' part of tcp_urg. It could be
4051 * moved inline now as tcp_urg is only called from one
4052 * place. We handle URGent data wrong. We have to - as
4053 * BSD still doesn't use the correction from RFC961.
4054 * For 1003.1g we should support a new option TCP_STDURG to permit
4055 * either form (or just set the sysctl tcp_stdurg).
4058 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4060 struct tcp_sock
*tp
= tcp_sk(sk
);
4061 u32 ptr
= ntohs(th
->urg_ptr
);
4063 if (ptr
&& !sysctl_tcp_stdurg
)
4065 ptr
+= ntohl(th
->seq
);
4067 /* Ignore urgent data that we've already seen and read. */
4068 if (after(tp
->copied_seq
, ptr
))
4071 /* Do not replay urg ptr.
4073 * NOTE: interesting situation not covered by specs.
4074 * Misbehaving sender may send urg ptr, pointing to segment,
4075 * which we already have in ofo queue. We are not able to fetch
4076 * such data and will stay in TCP_URG_NOTYET until will be eaten
4077 * by recvmsg(). Seems, we are not obliged to handle such wicked
4078 * situations. But it is worth to think about possibility of some
4079 * DoSes using some hypothetical application level deadlock.
4081 if (before(ptr
, tp
->rcv_nxt
))
4084 /* Do we already have a newer (or duplicate) urgent pointer? */
4085 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4088 /* Tell the world about our new urgent pointer. */
4091 /* We may be adding urgent data when the last byte read was
4092 * urgent. To do this requires some care. We cannot just ignore
4093 * tp->copied_seq since we would read the last urgent byte again
4094 * as data, nor can we alter copied_seq until this data arrives
4095 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4097 * NOTE. Double Dutch. Rendering to plain English: author of comment
4098 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4099 * and expect that both A and B disappear from stream. This is _wrong_.
4100 * Though this happens in BSD with high probability, this is occasional.
4101 * Any application relying on this is buggy. Note also, that fix "works"
4102 * only in this artificial test. Insert some normal data between A and B and we will
4103 * decline of BSD again. Verdict: it is better to remove to trap
4106 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4107 !sock_flag(sk
, SOCK_URGINLINE
) &&
4108 tp
->copied_seq
!= tp
->rcv_nxt
) {
4109 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4111 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4112 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4117 tp
->urg_data
= TCP_URG_NOTYET
;
4120 /* Disable header prediction. */
4124 /* This is the 'fast' part of urgent handling. */
4125 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4127 struct tcp_sock
*tp
= tcp_sk(sk
);
4129 /* Check if we get a new urgent pointer - normally not. */
4131 tcp_check_urg(sk
,th
);
4133 /* Do we wait for any urgent data? - normally not... */
4134 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4135 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4138 /* Is the urgent pointer pointing into this packet? */
4139 if (ptr
< skb
->len
) {
4141 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4143 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4144 if (!sock_flag(sk
, SOCK_DEAD
))
4145 sk
->sk_data_ready(sk
, 0);
4150 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4152 struct tcp_sock
*tp
= tcp_sk(sk
);
4153 int chunk
= skb
->len
- hlen
;
4157 if (skb_csum_unnecessary(skb
))
4158 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4160 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4164 tp
->ucopy
.len
-= chunk
;
4165 tp
->copied_seq
+= chunk
;
4166 tcp_rcv_space_adjust(sk
);
4173 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4177 if (sock_owned_by_user(sk
)) {
4179 result
= __tcp_checksum_complete(skb
);
4182 result
= __tcp_checksum_complete(skb
);
4187 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4189 return !skb_csum_unnecessary(skb
) &&
4190 __tcp_checksum_complete_user(sk
, skb
);
4193 #ifdef CONFIG_NET_DMA
4194 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4196 struct tcp_sock
*tp
= tcp_sk(sk
);
4197 int chunk
= skb
->len
- hlen
;
4199 int copied_early
= 0;
4201 if (tp
->ucopy
.wakeup
)
4204 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4205 tp
->ucopy
.dma_chan
= get_softnet_dma();
4207 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4209 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4210 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4215 tp
->ucopy
.dma_cookie
= dma_cookie
;
4218 tp
->ucopy
.len
-= chunk
;
4219 tp
->copied_seq
+= chunk
;
4220 tcp_rcv_space_adjust(sk
);
4222 if ((tp
->ucopy
.len
== 0) ||
4223 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4224 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4225 tp
->ucopy
.wakeup
= 1;
4226 sk
->sk_data_ready(sk
, 0);
4228 } else if (chunk
> 0) {
4229 tp
->ucopy
.wakeup
= 1;
4230 sk
->sk_data_ready(sk
, 0);
4233 return copied_early
;
4235 #endif /* CONFIG_NET_DMA */
4238 * TCP receive function for the ESTABLISHED state.
4240 * It is split into a fast path and a slow path. The fast path is
4242 * - A zero window was announced from us - zero window probing
4243 * is only handled properly in the slow path.
4244 * - Out of order segments arrived.
4245 * - Urgent data is expected.
4246 * - There is no buffer space left
4247 * - Unexpected TCP flags/window values/header lengths are received
4248 * (detected by checking the TCP header against pred_flags)
4249 * - Data is sent in both directions. Fast path only supports pure senders
4250 * or pure receivers (this means either the sequence number or the ack
4251 * value must stay constant)
4252 * - Unexpected TCP option.
4254 * When these conditions are not satisfied it drops into a standard
4255 * receive procedure patterned after RFC793 to handle all cases.
4256 * The first three cases are guaranteed by proper pred_flags setting,
4257 * the rest is checked inline. Fast processing is turned on in
4258 * tcp_data_queue when everything is OK.
4260 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4261 struct tcphdr
*th
, unsigned len
)
4263 struct tcp_sock
*tp
= tcp_sk(sk
);
4266 * Header prediction.
4267 * The code loosely follows the one in the famous
4268 * "30 instruction TCP receive" Van Jacobson mail.
4270 * Van's trick is to deposit buffers into socket queue
4271 * on a device interrupt, to call tcp_recv function
4272 * on the receive process context and checksum and copy
4273 * the buffer to user space. smart...
4275 * Our current scheme is not silly either but we take the
4276 * extra cost of the net_bh soft interrupt processing...
4277 * We do checksum and copy also but from device to kernel.
4280 tp
->rx_opt
.saw_tstamp
= 0;
4282 /* pred_flags is 0xS?10 << 16 + snd_wnd
4283 * if header_prediction is to be made
4284 * 'S' will always be tp->tcp_header_len >> 2
4285 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4286 * turn it off (when there are holes in the receive
4287 * space for instance)
4288 * PSH flag is ignored.
4291 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4292 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4293 int tcp_header_len
= tp
->tcp_header_len
;
4295 /* Timestamp header prediction: tcp_header_len
4296 * is automatically equal to th->doff*4 due to pred_flags
4300 /* Check timestamp */
4301 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4302 __be32
*ptr
= (__be32
*)(th
+ 1);
4304 /* No? Slow path! */
4305 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4306 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4309 tp
->rx_opt
.saw_tstamp
= 1;
4311 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4313 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4315 /* If PAWS failed, check it more carefully in slow path */
4316 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4319 /* DO NOT update ts_recent here, if checksum fails
4320 * and timestamp was corrupted part, it will result
4321 * in a hung connection since we will drop all
4322 * future packets due to the PAWS test.
4326 if (len
<= tcp_header_len
) {
4327 /* Bulk data transfer: sender */
4328 if (len
== tcp_header_len
) {
4329 /* Predicted packet is in window by definition.
4330 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4331 * Hence, check seq<=rcv_wup reduces to:
4333 if (tcp_header_len
==
4334 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4335 tp
->rcv_nxt
== tp
->rcv_wup
)
4336 tcp_store_ts_recent(tp
);
4338 /* We know that such packets are checksummed
4341 tcp_ack(sk
, skb
, 0);
4343 tcp_data_snd_check(sk
);
4345 } else { /* Header too small */
4346 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4351 int copied_early
= 0;
4353 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4354 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4355 #ifdef CONFIG_NET_DMA
4356 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4361 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4362 __set_current_state(TASK_RUNNING
);
4364 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4368 /* Predicted packet is in window by definition.
4369 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4370 * Hence, check seq<=rcv_wup reduces to:
4372 if (tcp_header_len
==
4373 (sizeof(struct tcphdr
) +
4374 TCPOLEN_TSTAMP_ALIGNED
) &&
4375 tp
->rcv_nxt
== tp
->rcv_wup
)
4376 tcp_store_ts_recent(tp
);
4378 tcp_rcv_rtt_measure_ts(sk
, skb
);
4380 __skb_pull(skb
, tcp_header_len
);
4381 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4382 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4385 tcp_cleanup_rbuf(sk
, skb
->len
);
4388 if (tcp_checksum_complete_user(sk
, skb
))
4391 /* Predicted packet is in window by definition.
4392 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4393 * Hence, check seq<=rcv_wup reduces to:
4395 if (tcp_header_len
==
4396 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4397 tp
->rcv_nxt
== tp
->rcv_wup
)
4398 tcp_store_ts_recent(tp
);
4400 tcp_rcv_rtt_measure_ts(sk
, skb
);
4402 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4405 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4407 /* Bulk data transfer: receiver */
4408 __skb_pull(skb
,tcp_header_len
);
4409 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4410 sk_stream_set_owner_r(skb
, sk
);
4411 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4414 tcp_event_data_recv(sk
, skb
);
4416 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4417 /* Well, only one small jumplet in fast path... */
4418 tcp_ack(sk
, skb
, FLAG_DATA
);
4419 tcp_data_snd_check(sk
);
4420 if (!inet_csk_ack_scheduled(sk
))
4424 __tcp_ack_snd_check(sk
, 0);
4426 #ifdef CONFIG_NET_DMA
4428 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4434 sk
->sk_data_ready(sk
, 0);
4440 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4444 * RFC1323: H1. Apply PAWS check first.
4446 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4447 tcp_paws_discard(sk
, skb
)) {
4449 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4450 tcp_send_dupack(sk
, skb
);
4453 /* Resets are accepted even if PAWS failed.
4455 ts_recent update must be made after we are sure
4456 that the packet is in window.
4461 * Standard slow path.
4464 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4465 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4466 * (RST) segments are validated by checking their SEQ-fields."
4467 * And page 69: "If an incoming segment is not acceptable,
4468 * an acknowledgment should be sent in reply (unless the RST bit
4469 * is set, if so drop the segment and return)".
4472 tcp_send_dupack(sk
, skb
);
4481 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4483 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4484 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4485 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4492 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4494 tcp_rcv_rtt_measure_ts(sk
, skb
);
4496 /* Process urgent data. */
4497 tcp_urg(sk
, skb
, th
);
4499 /* step 7: process the segment text */
4500 tcp_data_queue(sk
, skb
);
4502 tcp_data_snd_check(sk
);
4503 tcp_ack_snd_check(sk
);
4507 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4514 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4515 struct tcphdr
*th
, unsigned len
)
4517 struct tcp_sock
*tp
= tcp_sk(sk
);
4518 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4519 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4521 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4525 * "If the state is SYN-SENT then
4526 * first check the ACK bit
4527 * If the ACK bit is set
4528 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4529 * a reset (unless the RST bit is set, if so drop
4530 * the segment and return)"
4532 * We do not send data with SYN, so that RFC-correct
4535 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4536 goto reset_and_undo
;
4538 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4539 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4541 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4542 goto reset_and_undo
;
4545 /* Now ACK is acceptable.
4547 * "If the RST bit is set
4548 * If the ACK was acceptable then signal the user "error:
4549 * connection reset", drop the segment, enter CLOSED state,
4550 * delete TCB, and return."
4559 * "fifth, if neither of the SYN or RST bits is set then
4560 * drop the segment and return."
4566 goto discard_and_undo
;
4569 * "If the SYN bit is on ...
4570 * are acceptable then ...
4571 * (our SYN has been ACKed), change the connection
4572 * state to ESTABLISHED..."
4575 TCP_ECN_rcv_synack(tp
, th
);
4577 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4578 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4580 /* Ok.. it's good. Set up sequence numbers and
4581 * move to established.
4583 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4584 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4586 /* RFC1323: The window in SYN & SYN/ACK segments is
4589 tp
->snd_wnd
= ntohs(th
->window
);
4590 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4592 if (!tp
->rx_opt
.wscale_ok
) {
4593 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4594 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4597 if (tp
->rx_opt
.saw_tstamp
) {
4598 tp
->rx_opt
.tstamp_ok
= 1;
4599 tp
->tcp_header_len
=
4600 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4601 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4602 tcp_store_ts_recent(tp
);
4604 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4607 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
4608 tcp_enable_fack(tp
);
4611 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4612 tcp_initialize_rcv_mss(sk
);
4614 /* Remember, tcp_poll() does not lock socket!
4615 * Change state from SYN-SENT only after copied_seq
4616 * is initialized. */
4617 tp
->copied_seq
= tp
->rcv_nxt
;
4619 tcp_set_state(sk
, TCP_ESTABLISHED
);
4621 security_inet_conn_established(sk
, skb
);
4623 /* Make sure socket is routed, for correct metrics. */
4624 icsk
->icsk_af_ops
->rebuild_header(sk
);
4626 tcp_init_metrics(sk
);
4628 tcp_init_congestion_control(sk
);
4630 /* Prevent spurious tcp_cwnd_restart() on first data
4633 tp
->lsndtime
= tcp_time_stamp
;
4635 tcp_init_buffer_space(sk
);
4637 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4638 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4640 if (!tp
->rx_opt
.snd_wscale
)
4641 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4645 if (!sock_flag(sk
, SOCK_DEAD
)) {
4646 sk
->sk_state_change(sk
);
4647 sk_wake_async(sk
, 0, POLL_OUT
);
4650 if (sk
->sk_write_pending
||
4651 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4652 icsk
->icsk_ack
.pingpong
) {
4653 /* Save one ACK. Data will be ready after
4654 * several ticks, if write_pending is set.
4656 * It may be deleted, but with this feature tcpdumps
4657 * look so _wonderfully_ clever, that I was not able
4658 * to stand against the temptation 8) --ANK
4660 inet_csk_schedule_ack(sk
);
4661 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4662 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4663 tcp_incr_quickack(sk
);
4664 tcp_enter_quickack_mode(sk
);
4665 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4666 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4677 /* No ACK in the segment */
4681 * "If the RST bit is set
4683 * Otherwise (no ACK) drop the segment and return."
4686 goto discard_and_undo
;
4690 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4691 goto discard_and_undo
;
4694 /* We see SYN without ACK. It is attempt of
4695 * simultaneous connect with crossed SYNs.
4696 * Particularly, it can be connect to self.
4698 tcp_set_state(sk
, TCP_SYN_RECV
);
4700 if (tp
->rx_opt
.saw_tstamp
) {
4701 tp
->rx_opt
.tstamp_ok
= 1;
4702 tcp_store_ts_recent(tp
);
4703 tp
->tcp_header_len
=
4704 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4706 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4709 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4710 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4712 /* RFC1323: The window in SYN & SYN/ACK segments is
4715 tp
->snd_wnd
= ntohs(th
->window
);
4716 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4717 tp
->max_window
= tp
->snd_wnd
;
4719 TCP_ECN_rcv_syn(tp
, th
);
4722 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4723 tcp_initialize_rcv_mss(sk
);
4726 tcp_send_synack(sk
);
4728 /* Note, we could accept data and URG from this segment.
4729 * There are no obstacles to make this.
4731 * However, if we ignore data in ACKless segments sometimes,
4732 * we have no reasons to accept it sometimes.
4733 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4734 * is not flawless. So, discard packet for sanity.
4735 * Uncomment this return to process the data.
4742 /* "fifth, if neither of the SYN or RST bits is set then
4743 * drop the segment and return."
4747 tcp_clear_options(&tp
->rx_opt
);
4748 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4752 tcp_clear_options(&tp
->rx_opt
);
4753 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4759 * This function implements the receiving procedure of RFC 793 for
4760 * all states except ESTABLISHED and TIME_WAIT.
4761 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4762 * address independent.
4765 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4766 struct tcphdr
*th
, unsigned len
)
4768 struct tcp_sock
*tp
= tcp_sk(sk
);
4769 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4772 tp
->rx_opt
.saw_tstamp
= 0;
4774 switch (sk
->sk_state
) {
4786 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4789 /* Now we have several options: In theory there is
4790 * nothing else in the frame. KA9Q has an option to
4791 * send data with the syn, BSD accepts data with the
4792 * syn up to the [to be] advertised window and
4793 * Solaris 2.1 gives you a protocol error. For now
4794 * we just ignore it, that fits the spec precisely
4795 * and avoids incompatibilities. It would be nice in
4796 * future to drop through and process the data.
4798 * Now that TTCP is starting to be used we ought to
4800 * But, this leaves one open to an easy denial of
4801 * service attack, and SYN cookies can't defend
4802 * against this problem. So, we drop the data
4803 * in the interest of security over speed unless
4804 * it's still in use.
4812 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4816 /* Do step6 onward by hand. */
4817 tcp_urg(sk
, skb
, th
);
4819 tcp_data_snd_check(sk
);
4823 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4824 tcp_paws_discard(sk
, skb
)) {
4826 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4827 tcp_send_dupack(sk
, skb
);
4830 /* Reset is accepted even if it did not pass PAWS. */
4833 /* step 1: check sequence number */
4834 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4836 tcp_send_dupack(sk
, skb
);
4840 /* step 2: check RST bit */
4846 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4848 /* step 3: check security and precedence [ignored] */
4852 * Check for a SYN in window.
4854 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4855 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4860 /* step 5: check the ACK field */
4862 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4864 switch (sk
->sk_state
) {
4867 tp
->copied_seq
= tp
->rcv_nxt
;
4869 tcp_set_state(sk
, TCP_ESTABLISHED
);
4870 sk
->sk_state_change(sk
);
4872 /* Note, that this wakeup is only for marginal
4873 * crossed SYN case. Passively open sockets
4874 * are not waked up, because sk->sk_sleep ==
4875 * NULL and sk->sk_socket == NULL.
4877 if (sk
->sk_socket
) {
4878 sk_wake_async(sk
,0,POLL_OUT
);
4881 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4882 tp
->snd_wnd
= ntohs(th
->window
) <<
4883 tp
->rx_opt
.snd_wscale
;
4884 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4885 TCP_SKB_CB(skb
)->seq
);
4887 /* tcp_ack considers this ACK as duplicate
4888 * and does not calculate rtt.
4889 * Fix it at least with timestamps.
4891 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4893 tcp_ack_saw_tstamp(sk
, 0);
4895 if (tp
->rx_opt
.tstamp_ok
)
4896 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4898 /* Make sure socket is routed, for
4901 icsk
->icsk_af_ops
->rebuild_header(sk
);
4903 tcp_init_metrics(sk
);
4905 tcp_init_congestion_control(sk
);
4907 /* Prevent spurious tcp_cwnd_restart() on
4908 * first data packet.
4910 tp
->lsndtime
= tcp_time_stamp
;
4913 tcp_initialize_rcv_mss(sk
);
4914 tcp_init_buffer_space(sk
);
4915 tcp_fast_path_on(tp
);
4922 if (tp
->snd_una
== tp
->write_seq
) {
4923 tcp_set_state(sk
, TCP_FIN_WAIT2
);
4924 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
4925 dst_confirm(sk
->sk_dst_cache
);
4927 if (!sock_flag(sk
, SOCK_DEAD
))
4928 /* Wake up lingering close() */
4929 sk
->sk_state_change(sk
);
4933 if (tp
->linger2
< 0 ||
4934 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4935 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
4937 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
4941 tmo
= tcp_fin_time(sk
);
4942 if (tmo
> TCP_TIMEWAIT_LEN
) {
4943 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
4944 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
4945 /* Bad case. We could lose such FIN otherwise.
4946 * It is not a big problem, but it looks confusing
4947 * and not so rare event. We still can lose it now,
4948 * if it spins in bh_lock_sock(), but it is really
4951 inet_csk_reset_keepalive_timer(sk
, tmo
);
4953 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
4961 if (tp
->snd_una
== tp
->write_seq
) {
4962 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4968 if (tp
->snd_una
== tp
->write_seq
) {
4969 tcp_update_metrics(sk
);
4978 /* step 6: check the URG bit */
4979 tcp_urg(sk
, skb
, th
);
4981 /* step 7: process the segment text */
4982 switch (sk
->sk_state
) {
4983 case TCP_CLOSE_WAIT
:
4986 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4990 /* RFC 793 says to queue data in these states,
4991 * RFC 1122 says we MUST send a reset.
4992 * BSD 4.4 also does reset.
4994 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
4995 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4996 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
4997 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5003 case TCP_ESTABLISHED
:
5004 tcp_data_queue(sk
, skb
);
5009 /* tcp_data could move socket to TIME-WAIT */
5010 if (sk
->sk_state
!= TCP_CLOSE
) {
5011 tcp_data_snd_check(sk
);
5012 tcp_ack_snd_check(sk
);
5022 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5023 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5024 EXPORT_SYMBOL(tcp_parse_options
);
5025 EXPORT_SYMBOL(tcp_rcv_established
);
5026 EXPORT_SYMBOL(tcp_rcv_state_process
);
5027 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);