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
= 2;
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 D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
108 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
110 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
111 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
112 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
113 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
114 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
116 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
118 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
119 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
121 /* Adapt the MSS value used to make delayed ack decision to the
124 static void tcp_measure_rcv_mss(struct sock
*sk
,
125 const struct sk_buff
*skb
)
127 struct inet_connection_sock
*icsk
= inet_csk(sk
);
128 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
131 icsk
->icsk_ack
.last_seg_size
= 0;
133 /* skb->len may jitter because of SACKs, even if peer
134 * sends good full-sized frames.
136 len
= skb_shinfo(skb
)->gso_size
?: skb
->len
;
137 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
138 icsk
->icsk_ack
.rcv_mss
= len
;
140 /* Otherwise, we make more careful check taking into account,
141 * that SACKs block is variable.
143 * "len" is invariant segment length, including TCP header.
145 len
+= skb
->data
- skb_transport_header(skb
);
146 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
147 /* If PSH is not set, packet should be
148 * full sized, provided peer TCP is not badly broken.
149 * This observation (if it is correct 8)) allows
150 * to handle super-low mtu links fairly.
152 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
153 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
154 /* Subtract also invariant (if peer is RFC compliant),
155 * tcp header plus fixed timestamp option length.
156 * Resulting "len" is MSS free of SACK jitter.
158 len
-= tcp_sk(sk
)->tcp_header_len
;
159 icsk
->icsk_ack
.last_seg_size
= len
;
161 icsk
->icsk_ack
.rcv_mss
= len
;
165 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
166 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
167 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
171 static void tcp_incr_quickack(struct sock
*sk
)
173 struct inet_connection_sock
*icsk
= inet_csk(sk
);
174 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
178 if (quickacks
> icsk
->icsk_ack
.quick
)
179 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
182 void tcp_enter_quickack_mode(struct sock
*sk
)
184 struct inet_connection_sock
*icsk
= inet_csk(sk
);
185 tcp_incr_quickack(sk
);
186 icsk
->icsk_ack
.pingpong
= 0;
187 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
190 /* Send ACKs quickly, if "quick" count is not exhausted
191 * and the session is not interactive.
194 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
196 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
197 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
200 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
202 if (tp
->ecn_flags
&TCP_ECN_OK
)
203 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
206 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
208 if (tcp_hdr(skb
)->cwr
)
209 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
212 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
214 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
217 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
219 if (tp
->ecn_flags
&TCP_ECN_OK
) {
220 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
221 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
222 /* Funny extension: if ECT is not set on a segment,
223 * it is surely retransmit. It is not in ECN RFC,
224 * but Linux follows this rule. */
225 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
226 tcp_enter_quickack_mode((struct sock
*)tp
);
230 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
232 if ((tp
->ecn_flags
&TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
233 tp
->ecn_flags
&= ~TCP_ECN_OK
;
236 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
238 if ((tp
->ecn_flags
&TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
239 tp
->ecn_flags
&= ~TCP_ECN_OK
;
242 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
244 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
&TCP_ECN_OK
))
249 /* Buffer size and advertised window tuning.
251 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
254 static void tcp_fixup_sndbuf(struct sock
*sk
)
256 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
257 sizeof(struct sk_buff
);
259 if (sk
->sk_sndbuf
< 3 * sndmem
)
260 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
263 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
265 * All tcp_full_space() is split to two parts: "network" buffer, allocated
266 * forward and advertised in receiver window (tp->rcv_wnd) and
267 * "application buffer", required to isolate scheduling/application
268 * latencies from network.
269 * window_clamp is maximal advertised window. It can be less than
270 * tcp_full_space(), in this case tcp_full_space() - window_clamp
271 * is reserved for "application" buffer. The less window_clamp is
272 * the smoother our behaviour from viewpoint of network, but the lower
273 * throughput and the higher sensitivity of the connection to losses. 8)
275 * rcv_ssthresh is more strict window_clamp used at "slow start"
276 * phase to predict further behaviour of this connection.
277 * It is used for two goals:
278 * - to enforce header prediction at sender, even when application
279 * requires some significant "application buffer". It is check #1.
280 * - to prevent pruning of receive queue because of misprediction
281 * of receiver window. Check #2.
283 * The scheme does not work when sender sends good segments opening
284 * window and then starts to feed us spaghetti. But it should work
285 * in common situations. Otherwise, we have to rely on queue collapsing.
288 /* Slow part of check#2. */
289 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
291 struct tcp_sock
*tp
= tcp_sk(sk
);
293 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
294 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
296 while (tp
->rcv_ssthresh
<= window
) {
297 if (truesize
<= skb
->len
)
298 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
306 static void tcp_grow_window(struct sock
*sk
,
309 struct tcp_sock
*tp
= tcp_sk(sk
);
312 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
313 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
314 !tcp_memory_pressure
) {
317 /* Check #2. Increase window, if skb with such overhead
318 * will fit to rcvbuf in future.
320 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
323 incr
= __tcp_grow_window(sk
, skb
);
326 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
327 inet_csk(sk
)->icsk_ack
.quick
|= 1;
332 /* 3. Tuning rcvbuf, when connection enters established state. */
334 static void tcp_fixup_rcvbuf(struct sock
*sk
)
336 struct tcp_sock
*tp
= tcp_sk(sk
);
337 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
339 /* Try to select rcvbuf so that 4 mss-sized segments
340 * will fit to window and corresponding skbs will fit to our rcvbuf.
341 * (was 3; 4 is minimum to allow fast retransmit to work.)
343 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
345 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
346 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
349 /* 4. Try to fixup all. It is made immediately after connection enters
352 static void tcp_init_buffer_space(struct sock
*sk
)
354 struct tcp_sock
*tp
= tcp_sk(sk
);
357 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
358 tcp_fixup_rcvbuf(sk
);
359 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
360 tcp_fixup_sndbuf(sk
);
362 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
364 maxwin
= tcp_full_space(sk
);
366 if (tp
->window_clamp
>= maxwin
) {
367 tp
->window_clamp
= maxwin
;
369 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
370 tp
->window_clamp
= max(maxwin
-
371 (maxwin
>> sysctl_tcp_app_win
),
375 /* Force reservation of one segment. */
376 if (sysctl_tcp_app_win
&&
377 tp
->window_clamp
> 2 * tp
->advmss
&&
378 tp
->window_clamp
+ tp
->advmss
> maxwin
)
379 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
381 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
382 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
385 /* 5. Recalculate window clamp after socket hit its memory bounds. */
386 static void tcp_clamp_window(struct sock
*sk
)
388 struct tcp_sock
*tp
= tcp_sk(sk
);
389 struct inet_connection_sock
*icsk
= inet_csk(sk
);
391 icsk
->icsk_ack
.quick
= 0;
393 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
394 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
395 !tcp_memory_pressure
&&
396 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
397 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
400 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
401 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
405 /* Initialize RCV_MSS value.
406 * RCV_MSS is an our guess about MSS used by the peer.
407 * We haven't any direct information about the MSS.
408 * It's better to underestimate the RCV_MSS rather than overestimate.
409 * Overestimations make us ACKing less frequently than needed.
410 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
412 void tcp_initialize_rcv_mss(struct sock
*sk
)
414 struct tcp_sock
*tp
= tcp_sk(sk
);
415 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
417 hint
= min(hint
, tp
->rcv_wnd
/2);
418 hint
= min(hint
, TCP_MIN_RCVMSS
);
419 hint
= max(hint
, TCP_MIN_MSS
);
421 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
424 /* Receiver "autotuning" code.
426 * The algorithm for RTT estimation w/o timestamps is based on
427 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
428 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
430 * More detail on this code can be found at
431 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
432 * though this reference is out of date. A new paper
435 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
437 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
443 if (new_sample
!= 0) {
444 /* If we sample in larger samples in the non-timestamp
445 * case, we could grossly overestimate the RTT especially
446 * with chatty applications or bulk transfer apps which
447 * are stalled on filesystem I/O.
449 * Also, since we are only going for a minimum in the
450 * non-timestamp case, we do not smooth things out
451 * else with timestamps disabled convergence takes too
455 m
-= (new_sample
>> 3);
457 } else if (m
< new_sample
)
460 /* No previous measure. */
464 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
465 tp
->rcv_rtt_est
.rtt
= new_sample
;
468 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
470 if (tp
->rcv_rtt_est
.time
== 0)
472 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
474 tcp_rcv_rtt_update(tp
,
475 jiffies
- tp
->rcv_rtt_est
.time
,
479 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
480 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
483 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
485 struct tcp_sock
*tp
= tcp_sk(sk
);
486 if (tp
->rx_opt
.rcv_tsecr
&&
487 (TCP_SKB_CB(skb
)->end_seq
-
488 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
489 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
493 * This function should be called every time data is copied to user space.
494 * It calculates the appropriate TCP receive buffer space.
496 void tcp_rcv_space_adjust(struct sock
*sk
)
498 struct tcp_sock
*tp
= tcp_sk(sk
);
502 if (tp
->rcvq_space
.time
== 0)
505 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
506 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
507 tp
->rcv_rtt_est
.rtt
== 0)
510 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
512 space
= max(tp
->rcvq_space
.space
, space
);
514 if (tp
->rcvq_space
.space
!= space
) {
517 tp
->rcvq_space
.space
= space
;
519 if (sysctl_tcp_moderate_rcvbuf
&&
520 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
521 int new_clamp
= space
;
523 /* Receive space grows, normalize in order to
524 * take into account packet headers and sk_buff
525 * structure overhead.
530 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
531 16 + sizeof(struct sk_buff
));
532 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
535 space
= min(space
, sysctl_tcp_rmem
[2]);
536 if (space
> sk
->sk_rcvbuf
) {
537 sk
->sk_rcvbuf
= space
;
539 /* Make the window clamp follow along. */
540 tp
->window_clamp
= new_clamp
;
546 tp
->rcvq_space
.seq
= tp
->copied_seq
;
547 tp
->rcvq_space
.time
= tcp_time_stamp
;
550 /* There is something which you must keep in mind when you analyze the
551 * behavior of the tp->ato delayed ack timeout interval. When a
552 * connection starts up, we want to ack as quickly as possible. The
553 * problem is that "good" TCP's do slow start at the beginning of data
554 * transmission. The means that until we send the first few ACK's the
555 * sender will sit on his end and only queue most of his data, because
556 * he can only send snd_cwnd unacked packets at any given time. For
557 * each ACK we send, he increments snd_cwnd and transmits more of his
560 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
562 struct tcp_sock
*tp
= tcp_sk(sk
);
563 struct inet_connection_sock
*icsk
= inet_csk(sk
);
566 inet_csk_schedule_ack(sk
);
568 tcp_measure_rcv_mss(sk
, skb
);
570 tcp_rcv_rtt_measure(tp
);
572 now
= tcp_time_stamp
;
574 if (!icsk
->icsk_ack
.ato
) {
575 /* The _first_ data packet received, initialize
576 * delayed ACK engine.
578 tcp_incr_quickack(sk
);
579 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
581 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
583 if (m
<= TCP_ATO_MIN
/2) {
584 /* The fastest case is the first. */
585 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
586 } else if (m
< icsk
->icsk_ack
.ato
) {
587 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
588 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
589 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
590 } else if (m
> icsk
->icsk_rto
) {
591 /* Too long gap. Apparently sender failed to
592 * restart window, so that we send ACKs quickly.
594 tcp_incr_quickack(sk
);
598 icsk
->icsk_ack
.lrcvtime
= now
;
600 TCP_ECN_check_ce(tp
, skb
);
603 tcp_grow_window(sk
, skb
);
606 static u32
tcp_rto_min(struct sock
*sk
)
608 struct dst_entry
*dst
= __sk_dst_get(sk
);
609 u32 rto_min
= TCP_RTO_MIN
;
611 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
612 rto_min
= dst
->metrics
[RTAX_RTO_MIN
-1];
616 /* Called to compute a smoothed rtt estimate. The data fed to this
617 * routine either comes from timestamps, or from segments that were
618 * known _not_ to have been retransmitted [see Karn/Partridge
619 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
620 * piece by Van Jacobson.
621 * NOTE: the next three routines used to be one big routine.
622 * To save cycles in the RFC 1323 implementation it was better to break
623 * it up into three procedures. -- erics
625 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
627 struct tcp_sock
*tp
= tcp_sk(sk
);
628 long m
= mrtt
; /* RTT */
630 /* The following amusing code comes from Jacobson's
631 * article in SIGCOMM '88. Note that rtt and mdev
632 * are scaled versions of rtt and mean deviation.
633 * This is designed to be as fast as possible
634 * m stands for "measurement".
636 * On a 1990 paper the rto value is changed to:
637 * RTO = rtt + 4 * mdev
639 * Funny. This algorithm seems to be very broken.
640 * These formulae increase RTO, when it should be decreased, increase
641 * too slowly, when it should be increased quickly, decrease too quickly
642 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
643 * does not matter how to _calculate_ it. Seems, it was trap
644 * that VJ failed to avoid. 8)
649 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
650 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
652 m
= -m
; /* m is now abs(error) */
653 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
654 /* This is similar to one of Eifel findings.
655 * Eifel blocks mdev updates when rtt decreases.
656 * This solution is a bit different: we use finer gain
657 * for mdev in this case (alpha*beta).
658 * Like Eifel it also prevents growth of rto,
659 * but also it limits too fast rto decreases,
660 * happening in pure Eifel.
665 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
667 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
668 if (tp
->mdev
> tp
->mdev_max
) {
669 tp
->mdev_max
= tp
->mdev
;
670 if (tp
->mdev_max
> tp
->rttvar
)
671 tp
->rttvar
= tp
->mdev_max
;
673 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
674 if (tp
->mdev_max
< tp
->rttvar
)
675 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
676 tp
->rtt_seq
= tp
->snd_nxt
;
677 tp
->mdev_max
= tcp_rto_min(sk
);
680 /* no previous measure. */
681 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
682 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
683 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
684 tp
->rtt_seq
= tp
->snd_nxt
;
688 /* Calculate rto without backoff. This is the second half of Van Jacobson's
689 * routine referred to above.
691 static inline void tcp_set_rto(struct sock
*sk
)
693 const struct tcp_sock
*tp
= tcp_sk(sk
);
694 /* Old crap is replaced with new one. 8)
697 * 1. If rtt variance happened to be less 50msec, it is hallucination.
698 * It cannot be less due to utterly erratic ACK generation made
699 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
700 * to do with delayed acks, because at cwnd>2 true delack timeout
701 * is invisible. Actually, Linux-2.4 also generates erratic
702 * ACKs in some circumstances.
704 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
706 /* 2. Fixups made earlier cannot be right.
707 * If we do not estimate RTO correctly without them,
708 * all the algo is pure shit and should be replaced
709 * with correct one. It is exactly, which we pretend to do.
713 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
714 * guarantees that rto is higher.
716 static inline void tcp_bound_rto(struct sock
*sk
)
718 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
719 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
722 /* Save metrics learned by this TCP session.
723 This function is called only, when TCP finishes successfully
724 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
726 void tcp_update_metrics(struct sock
*sk
)
728 struct tcp_sock
*tp
= tcp_sk(sk
);
729 struct dst_entry
*dst
= __sk_dst_get(sk
);
731 if (sysctl_tcp_nometrics_save
)
736 if (dst
&& (dst
->flags
&DST_HOST
)) {
737 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
740 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
741 /* This session failed to estimate rtt. Why?
742 * Probably, no packets returned in time.
745 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
746 dst
->metrics
[RTAX_RTT
-1] = 0;
750 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
752 /* If newly calculated rtt larger than stored one,
753 * store new one. Otherwise, use EWMA. Remember,
754 * rtt overestimation is always better than underestimation.
756 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
758 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
760 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
763 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
767 /* Scale deviation to rttvar fixed point */
772 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
773 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
775 dst
->metrics
[RTAX_RTTVAR
-1] -=
776 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
779 if (tp
->snd_ssthresh
>= 0xFFFF) {
780 /* Slow start still did not finish. */
781 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
782 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
783 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
784 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
785 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
786 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
787 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
788 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
789 icsk
->icsk_ca_state
== TCP_CA_Open
) {
790 /* Cong. avoidance phase, cwnd is reliable. */
791 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
792 dst
->metrics
[RTAX_SSTHRESH
-1] =
793 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
794 if (!dst_metric_locked(dst
, RTAX_CWND
))
795 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
797 /* Else slow start did not finish, cwnd is non-sense,
798 ssthresh may be also invalid.
800 if (!dst_metric_locked(dst
, RTAX_CWND
))
801 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
802 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
803 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
804 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
805 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
808 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
809 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
810 tp
->reordering
!= sysctl_tcp_reordering
)
811 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
816 /* Numbers are taken from RFC3390.
818 * John Heffner states:
820 * The RFC specifies a window of no more than 4380 bytes
821 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
822 * is a bit misleading because they use a clamp at 4380 bytes
823 * rather than use a multiplier in the relevant range.
825 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
827 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
830 if (tp
->mss_cache
> 1460)
833 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
835 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
838 /* Set slow start threshold and cwnd not falling to slow start */
839 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
841 struct tcp_sock
*tp
= tcp_sk(sk
);
842 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
844 tp
->prior_ssthresh
= 0;
846 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
849 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
850 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
851 tcp_packets_in_flight(tp
) + 1U);
852 tp
->snd_cwnd_cnt
= 0;
853 tp
->high_seq
= tp
->snd_nxt
;
854 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
855 TCP_ECN_queue_cwr(tp
);
857 tcp_set_ca_state(sk
, TCP_CA_CWR
);
862 * Packet counting of FACK is based on in-order assumptions, therefore TCP
863 * disables it when reordering is detected
865 static void tcp_disable_fack(struct tcp_sock
*tp
)
867 /* RFC3517 uses different metric in lost marker => reset on change */
869 tp
->lost_skb_hint
= NULL
;
870 tp
->rx_opt
.sack_ok
&= ~2;
873 /* Take a notice that peer is sending D-SACKs */
874 static void tcp_dsack_seen(struct tcp_sock
*tp
)
876 tp
->rx_opt
.sack_ok
|= 4;
879 /* Initialize metrics on socket. */
881 static void tcp_init_metrics(struct sock
*sk
)
883 struct tcp_sock
*tp
= tcp_sk(sk
);
884 struct dst_entry
*dst
= __sk_dst_get(sk
);
891 if (dst_metric_locked(dst
, RTAX_CWND
))
892 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
893 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
894 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
895 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
896 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
898 if (dst_metric(dst
, RTAX_REORDERING
) &&
899 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
900 tcp_disable_fack(tp
);
901 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
904 if (dst_metric(dst
, RTAX_RTT
) == 0)
907 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
910 /* Initial rtt is determined from SYN,SYN-ACK.
911 * The segment is small and rtt may appear much
912 * less than real one. Use per-dst memory
913 * to make it more realistic.
915 * A bit of theory. RTT is time passed after "normal" sized packet
916 * is sent until it is ACKed. In normal circumstances sending small
917 * packets force peer to delay ACKs and calculation is correct too.
918 * The algorithm is adaptive and, provided we follow specs, it
919 * NEVER underestimate RTT. BUT! If peer tries to make some clever
920 * tricks sort of "quick acks" for time long enough to decrease RTT
921 * to low value, and then abruptly stops to do it and starts to delay
922 * ACKs, wait for troubles.
924 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
925 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
926 tp
->rtt_seq
= tp
->snd_nxt
;
928 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
929 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
930 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
934 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
936 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
937 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
941 /* Play conservative. If timestamps are not
942 * supported, TCP will fail to recalculate correct
943 * rtt, if initial rto is too small. FORGET ALL AND RESET!
945 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
947 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
948 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
952 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
955 struct tcp_sock
*tp
= tcp_sk(sk
);
956 if (metric
> tp
->reordering
) {
957 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
959 /* This exciting event is worth to be remembered. 8) */
961 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
962 else if (tcp_is_reno(tp
))
963 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
964 else if (tcp_is_fack(tp
))
965 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
967 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
968 #if FASTRETRANS_DEBUG > 1
969 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
970 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
974 tp
->undo_marker
? tp
->undo_retrans
: 0);
976 tcp_disable_fack(tp
);
980 /* This procedure tags the retransmission queue when SACKs arrive.
982 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
983 * Packets in queue with these bits set are counted in variables
984 * sacked_out, retrans_out and lost_out, correspondingly.
986 * Valid combinations are:
987 * Tag InFlight Description
988 * 0 1 - orig segment is in flight.
989 * S 0 - nothing flies, orig reached receiver.
990 * L 0 - nothing flies, orig lost by net.
991 * R 2 - both orig and retransmit are in flight.
992 * L|R 1 - orig is lost, retransmit is in flight.
993 * S|R 1 - orig reached receiver, retrans is still in flight.
994 * (L|S|R is logically valid, it could occur when L|R is sacked,
995 * but it is equivalent to plain S and code short-curcuits it to S.
996 * L|S is logically invalid, it would mean -1 packet in flight 8))
998 * These 6 states form finite state machine, controlled by the following events:
999 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1000 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1001 * 3. Loss detection event of one of three flavors:
1002 * A. Scoreboard estimator decided the packet is lost.
1003 * A'. Reno "three dupacks" marks head of queue lost.
1004 * A''. Its FACK modfication, head until snd.fack is lost.
1005 * B. SACK arrives sacking data transmitted after never retransmitted
1006 * hole was sent out.
1007 * C. SACK arrives sacking SND.NXT at the moment, when the
1008 * segment was retransmitted.
1009 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1011 * It is pleasant to note, that state diagram turns out to be commutative,
1012 * so that we are allowed not to be bothered by order of our actions,
1013 * when multiple events arrive simultaneously. (see the function below).
1015 * Reordering detection.
1016 * --------------------
1017 * Reordering metric is maximal distance, which a packet can be displaced
1018 * in packet stream. With SACKs we can estimate it:
1020 * 1. SACK fills old hole and the corresponding segment was not
1021 * ever retransmitted -> reordering. Alas, we cannot use it
1022 * when segment was retransmitted.
1023 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1024 * for retransmitted and already SACKed segment -> reordering..
1025 * Both of these heuristics are not used in Loss state, when we cannot
1026 * account for retransmits accurately.
1028 * SACK block validation.
1029 * ----------------------
1031 * SACK block range validation checks that the received SACK block fits to
1032 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1033 * Note that SND.UNA is not included to the range though being valid because
1034 * it means that the receiver is rather inconsistent with itself reporting
1035 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1036 * perfectly valid, however, in light of RFC2018 which explicitly states
1037 * that "SACK block MUST reflect the newest segment. Even if the newest
1038 * segment is going to be discarded ...", not that it looks very clever
1039 * in case of head skb. Due to potentional receiver driven attacks, we
1040 * choose to avoid immediate execution of a walk in write queue due to
1041 * reneging and defer head skb's loss recovery to standard loss recovery
1042 * procedure that will eventually trigger (nothing forbids us doing this).
1044 * Implements also blockage to start_seq wrap-around. Problem lies in the
1045 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1046 * there's no guarantee that it will be before snd_nxt (n). The problem
1047 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1050 * <- outs wnd -> <- wrapzone ->
1051 * u e n u_w e_w s n_w
1053 * |<------------+------+----- TCP seqno space --------------+---------->|
1054 * ...-- <2^31 ->| |<--------...
1055 * ...---- >2^31 ------>| |<--------...
1057 * Current code wouldn't be vulnerable but it's better still to discard such
1058 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1059 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1060 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1061 * equal to the ideal case (infinite seqno space without wrap caused issues).
1063 * With D-SACK the lower bound is extended to cover sequence space below
1064 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1065 * again, D-SACK block must not to go across snd_una (for the same reason as
1066 * for the normal SACK blocks, explained above). But there all simplicity
1067 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1068 * fully below undo_marker they do not affect behavior in anyway and can
1069 * therefore be safely ignored. In rare cases (which are more or less
1070 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1071 * fragmentation and packet reordering past skb's retransmission. To consider
1072 * them correctly, the acceptable range must be extended even more though
1073 * the exact amount is rather hard to quantify. However, tp->max_window can
1074 * be used as an exaggerated estimate.
1076 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1077 u32 start_seq
, u32 end_seq
)
1079 /* Too far in future, or reversed (interpretation is ambiguous) */
1080 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1083 /* Nasty start_seq wrap-around check (see comments above) */
1084 if (!before(start_seq
, tp
->snd_nxt
))
1087 /* In outstanding window? ...This is valid exit for D-SACKs too.
1088 * start_seq == snd_una is non-sensical (see comments above)
1090 if (after(start_seq
, tp
->snd_una
))
1093 if (!is_dsack
|| !tp
->undo_marker
)
1096 /* ...Then it's D-SACK, and must reside below snd_una completely */
1097 if (!after(end_seq
, tp
->snd_una
))
1100 if (!before(start_seq
, tp
->undo_marker
))
1104 if (!after(end_seq
, tp
->undo_marker
))
1107 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1108 * start_seq < undo_marker and end_seq >= undo_marker.
1110 return !before(start_seq
, end_seq
- tp
->max_window
);
1113 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1114 * Event "C". Later note: FACK people cheated me again 8), we have to account
1115 * for reordering! Ugly, but should help.
1117 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1118 * less than what is now known to be received by the other end (derived from
1119 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1120 * retransmitted skbs to avoid some costly processing per ACKs.
1122 static void tcp_mark_lost_retrans(struct sock
*sk
)
1124 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1125 struct tcp_sock
*tp
= tcp_sk(sk
);
1126 struct sk_buff
*skb
;
1128 u32 new_low_seq
= tp
->snd_nxt
;
1129 u32 received_upto
= tcp_highest_sack_seq(tp
);
1131 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1132 !after(received_upto
, tp
->lost_retrans_low
) ||
1133 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1136 tcp_for_write_queue(skb
, sk
) {
1137 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1139 if (skb
== tcp_send_head(sk
))
1141 if (cnt
== tp
->retrans_out
)
1143 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1146 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1149 if (after(received_upto
, ack_seq
) &&
1151 !before(received_upto
,
1152 ack_seq
+ tp
->reordering
* tp
->mss_cache
))) {
1153 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1154 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1156 /* clear lost hint */
1157 tp
->retransmit_skb_hint
= NULL
;
1159 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1160 tp
->lost_out
+= tcp_skb_pcount(skb
);
1161 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1163 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1165 if (before(ack_seq
, new_low_seq
))
1166 new_low_seq
= ack_seq
;
1167 cnt
+= tcp_skb_pcount(skb
);
1171 if (tp
->retrans_out
)
1172 tp
->lost_retrans_low
= new_low_seq
;
1175 static int tcp_check_dsack(struct tcp_sock
*tp
, struct sk_buff
*ack_skb
,
1176 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1179 u32 start_seq_0
= ntohl(get_unaligned(&sp
[0].start_seq
));
1180 u32 end_seq_0
= ntohl(get_unaligned(&sp
[0].end_seq
));
1183 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1186 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
1187 } else if (num_sacks
> 1) {
1188 u32 end_seq_1
= ntohl(get_unaligned(&sp
[1].end_seq
));
1189 u32 start_seq_1
= ntohl(get_unaligned(&sp
[1].start_seq
));
1191 if (!after(end_seq_0
, end_seq_1
) &&
1192 !before(start_seq_0
, start_seq_1
)) {
1195 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
1199 /* D-SACK for already forgotten data... Do dumb counting. */
1201 !after(end_seq_0
, prior_snd_una
) &&
1202 after(end_seq_0
, tp
->undo_marker
))
1208 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1209 * the incoming SACK may not exactly match but we can find smaller MSS
1210 * aligned portion of it that matches. Therefore we might need to fragment
1211 * which may fail and creates some hassle (caller must handle error case
1214 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1215 u32 start_seq
, u32 end_seq
)
1218 unsigned int pkt_len
;
1220 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1221 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1223 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1224 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1226 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1229 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1231 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1232 err
= tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
);
1240 static int tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1241 int *reord
, int dup_sack
, int fack_count
)
1243 struct tcp_sock
*tp
= tcp_sk(sk
);
1244 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1247 /* Account D-SACK for retransmitted packet. */
1248 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1249 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1251 if (sacked
& TCPCB_SACKED_ACKED
)
1252 *reord
= min(fack_count
, *reord
);
1255 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1256 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1259 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1260 if (sacked
& TCPCB_SACKED_RETRANS
) {
1261 /* If the segment is not tagged as lost,
1262 * we do not clear RETRANS, believing
1263 * that retransmission is still in flight.
1265 if (sacked
& TCPCB_LOST
) {
1266 TCP_SKB_CB(skb
)->sacked
&=
1267 ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1268 tp
->lost_out
-= tcp_skb_pcount(skb
);
1269 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1271 /* clear lost hint */
1272 tp
->retransmit_skb_hint
= NULL
;
1275 if (!(sacked
& TCPCB_RETRANS
)) {
1276 /* New sack for not retransmitted frame,
1277 * which was in hole. It is reordering.
1279 if (before(TCP_SKB_CB(skb
)->seq
,
1280 tcp_highest_sack_seq(tp
)))
1281 *reord
= min(fack_count
, *reord
);
1283 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1284 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1285 flag
|= FLAG_ONLY_ORIG_SACKED
;
1288 if (sacked
& TCPCB_LOST
) {
1289 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1290 tp
->lost_out
-= tcp_skb_pcount(skb
);
1292 /* clear lost hint */
1293 tp
->retransmit_skb_hint
= NULL
;
1297 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1298 flag
|= FLAG_DATA_SACKED
;
1299 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1301 fack_count
+= tcp_skb_pcount(skb
);
1303 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1304 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1305 before(TCP_SKB_CB(skb
)->seq
,
1306 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1307 tp
->lost_cnt_hint
+= tcp_skb_pcount(skb
);
1309 if (fack_count
> tp
->fackets_out
)
1310 tp
->fackets_out
= fack_count
;
1312 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
1313 tcp_advance_highest_sack(sk
, skb
);
1316 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1317 * frames and clear it. undo_retrans is decreased above, L|R frames
1318 * are accounted above as well.
1320 if (dup_sack
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)) {
1321 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1322 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1323 tp
->retransmit_skb_hint
= NULL
;
1329 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1330 struct tcp_sack_block
*next_dup
,
1331 u32 start_seq
, u32 end_seq
,
1332 int dup_sack_in
, int *fack_count
,
1333 int *reord
, int *flag
)
1335 tcp_for_write_queue_from(skb
, sk
) {
1337 int dup_sack
= dup_sack_in
;
1339 if (skb
== tcp_send_head(sk
))
1342 /* queue is in-order => we can short-circuit the walk early */
1343 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1346 if ((next_dup
!= NULL
) &&
1347 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1348 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1349 next_dup
->start_seq
,
1356 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
, end_seq
);
1357 if (unlikely(in_sack
< 0))
1361 *flag
|= tcp_sacktag_one(skb
, sk
, reord
, dup_sack
, *fack_count
);
1363 *fack_count
+= tcp_skb_pcount(skb
);
1368 /* Avoid all extra work that is being done by sacktag while walking in
1371 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1374 tcp_for_write_queue_from(skb
, sk
) {
1375 if (skb
== tcp_send_head(sk
))
1378 if (!before(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1384 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1386 struct tcp_sack_block
*next_dup
,
1388 int *fack_count
, int *reord
,
1391 if (next_dup
== NULL
)
1394 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1395 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
);
1396 tcp_sacktag_walk(skb
, sk
, NULL
,
1397 next_dup
->start_seq
, next_dup
->end_seq
,
1398 1, fack_count
, reord
, flag
);
1404 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1406 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1410 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
1412 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1413 struct tcp_sock
*tp
= tcp_sk(sk
);
1414 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1415 TCP_SKB_CB(ack_skb
)->sacked
);
1416 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1417 struct tcp_sack_block sp
[4];
1418 struct tcp_sack_block
*cache
;
1419 struct sk_buff
*skb
;
1420 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
1422 int reord
= tp
->packets_out
;
1424 int found_dup_sack
= 0;
1427 int first_sack_index
;
1429 if (!tp
->sacked_out
) {
1430 if (WARN_ON(tp
->fackets_out
))
1431 tp
->fackets_out
= 0;
1432 tcp_highest_sack_reset(sk
);
1435 found_dup_sack
= tcp_check_dsack(tp
, ack_skb
, sp_wire
,
1436 num_sacks
, prior_snd_una
);
1438 flag
|= FLAG_DSACKING_ACK
;
1440 /* Eliminate too old ACKs, but take into
1441 * account more or less fresh ones, they can
1442 * contain valid SACK info.
1444 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1447 if (!tp
->packets_out
)
1451 first_sack_index
= 0;
1452 for (i
= 0; i
< num_sacks
; i
++) {
1453 int dup_sack
= !i
&& found_dup_sack
;
1455 sp
[used_sacks
].start_seq
= ntohl(get_unaligned(&sp_wire
[i
].start_seq
));
1456 sp
[used_sacks
].end_seq
= ntohl(get_unaligned(&sp_wire
[i
].end_seq
));
1458 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1459 sp
[used_sacks
].start_seq
,
1460 sp
[used_sacks
].end_seq
)) {
1462 if (!tp
->undo_marker
)
1463 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO
);
1465 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD
);
1467 /* Don't count olds caused by ACK reordering */
1468 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1469 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1471 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD
);
1474 first_sack_index
= -1;
1478 /* Ignore very old stuff early */
1479 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1485 /* order SACK blocks to allow in order walk of the retrans queue */
1486 for (i
= used_sacks
- 1; i
> 0; i
--) {
1487 for (j
= 0; j
< i
; j
++){
1488 if (after(sp
[j
].start_seq
, sp
[j
+1].start_seq
)) {
1489 struct tcp_sack_block tmp
;
1495 /* Track where the first SACK block goes to */
1496 if (j
== first_sack_index
)
1497 first_sack_index
= j
+1;
1502 skb
= tcp_write_queue_head(sk
);
1506 if (!tp
->sacked_out
) {
1507 /* It's already past, so skip checking against it */
1508 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1510 cache
= tp
->recv_sack_cache
;
1511 /* Skip empty blocks in at head of the cache */
1512 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1517 while (i
< used_sacks
) {
1518 u32 start_seq
= sp
[i
].start_seq
;
1519 u32 end_seq
= sp
[i
].end_seq
;
1520 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1521 struct tcp_sack_block
*next_dup
= NULL
;
1523 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1524 next_dup
= &sp
[i
+ 1];
1526 /* Event "B" in the comment above. */
1527 if (after(end_seq
, tp
->high_seq
))
1528 flag
|= FLAG_DATA_LOST
;
1530 /* Skip too early cached blocks */
1531 while (tcp_sack_cache_ok(tp
, cache
) &&
1532 !before(start_seq
, cache
->end_seq
))
1535 /* Can skip some work by looking recv_sack_cache? */
1536 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1537 after(end_seq
, cache
->start_seq
)) {
1540 if (before(start_seq
, cache
->start_seq
)) {
1541 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1542 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
,
1543 cache
->start_seq
, dup_sack
,
1544 &fack_count
, &reord
, &flag
);
1547 /* Rest of the block already fully processed? */
1548 if (!after(end_seq
, cache
->end_seq
))
1551 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
, cache
->end_seq
,
1552 &fack_count
, &reord
, &flag
);
1554 /* ...tail remains todo... */
1555 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1556 /* ...but better entrypoint exists! */
1557 skb
= tcp_highest_sack(sk
);
1560 fack_count
= tp
->fackets_out
;
1565 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
);
1566 /* Check overlap against next cached too (past this one already) */
1571 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1572 skb
= tcp_highest_sack(sk
);
1575 fack_count
= tp
->fackets_out
;
1577 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1580 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1581 dup_sack
, &fack_count
, &reord
, &flag
);
1584 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1585 * due to in-order walk
1587 if (after(end_seq
, tp
->frto_highmark
))
1588 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1593 /* Clear the head of the cache sack blocks so we can skip it next time */
1594 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1595 tp
->recv_sack_cache
[i
].start_seq
= 0;
1596 tp
->recv_sack_cache
[i
].end_seq
= 0;
1598 for (j
= 0; j
< used_sacks
; j
++)
1599 tp
->recv_sack_cache
[i
++] = sp
[j
];
1601 tcp_mark_lost_retrans(sk
);
1603 tcp_verify_left_out(tp
);
1605 if ((reord
< tp
->fackets_out
) &&
1606 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1607 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1608 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1612 #if FASTRETRANS_DEBUG > 0
1613 BUG_TRAP((int)tp
->sacked_out
>= 0);
1614 BUG_TRAP((int)tp
->lost_out
>= 0);
1615 BUG_TRAP((int)tp
->retrans_out
>= 0);
1616 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1621 /* If we receive more dupacks than we expected counting segments
1622 * in assumption of absent reordering, interpret this as reordering.
1623 * The only another reason could be bug in receiver TCP.
1625 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1627 struct tcp_sock
*tp
= tcp_sk(sk
);
1630 holes
= max(tp
->lost_out
, 1U);
1631 holes
= min(holes
, tp
->packets_out
);
1633 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1634 tp
->sacked_out
= tp
->packets_out
- holes
;
1635 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1639 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1641 static void tcp_add_reno_sack(struct sock
*sk
)
1643 struct tcp_sock
*tp
= tcp_sk(sk
);
1645 tcp_check_reno_reordering(sk
, 0);
1646 tcp_verify_left_out(tp
);
1649 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1651 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1653 struct tcp_sock
*tp
= tcp_sk(sk
);
1656 /* One ACK acked hole. The rest eat duplicate ACKs. */
1657 if (acked
-1 >= tp
->sacked_out
)
1660 tp
->sacked_out
-= acked
-1;
1662 tcp_check_reno_reordering(sk
, acked
);
1663 tcp_verify_left_out(tp
);
1666 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1671 /* F-RTO can only be used if TCP has never retransmitted anything other than
1672 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1674 int tcp_use_frto(struct sock
*sk
)
1676 const struct tcp_sock
*tp
= tcp_sk(sk
);
1677 struct sk_buff
*skb
;
1679 if (!sysctl_tcp_frto
)
1685 /* Avoid expensive walking of rexmit queue if possible */
1686 if (tp
->retrans_out
> 1)
1689 skb
= tcp_write_queue_head(sk
);
1690 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1691 tcp_for_write_queue_from(skb
, sk
) {
1692 if (skb
== tcp_send_head(sk
))
1694 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1696 /* Short-circuit when first non-SACKed skb has been checked */
1697 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1703 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1704 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1705 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1706 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1707 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1708 * bits are handled if the Loss state is really to be entered (in
1709 * tcp_enter_frto_loss).
1711 * Do like tcp_enter_loss() would; when RTO expires the second time it
1713 * "Reduce ssthresh if it has not yet been made inside this window."
1715 void tcp_enter_frto(struct sock
*sk
)
1717 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1718 struct tcp_sock
*tp
= tcp_sk(sk
);
1719 struct sk_buff
*skb
;
1721 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1722 tp
->snd_una
== tp
->high_seq
||
1723 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1724 !icsk
->icsk_retransmits
)) {
1725 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1726 /* Our state is too optimistic in ssthresh() call because cwnd
1727 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1728 * recovery has not yet completed. Pattern would be this: RTO,
1729 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1731 * RFC4138 should be more specific on what to do, even though
1732 * RTO is quite unlikely to occur after the first Cumulative ACK
1733 * due to back-off and complexity of triggering events ...
1735 if (tp
->frto_counter
) {
1737 stored_cwnd
= tp
->snd_cwnd
;
1739 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1740 tp
->snd_cwnd
= stored_cwnd
;
1742 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1744 /* ... in theory, cong.control module could do "any tricks" in
1745 * ssthresh(), which means that ca_state, lost bits and lost_out
1746 * counter would have to be faked before the call occurs. We
1747 * consider that too expensive, unlikely and hacky, so modules
1748 * using these in ssthresh() must deal these incompatibility
1749 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1751 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1754 tp
->undo_marker
= tp
->snd_una
;
1755 tp
->undo_retrans
= 0;
1757 skb
= tcp_write_queue_head(sk
);
1758 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1759 tp
->undo_marker
= 0;
1760 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1761 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1762 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1764 tcp_verify_left_out(tp
);
1766 /* Too bad if TCP was application limited */
1767 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1769 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1770 * The last condition is necessary at least in tp->frto_counter case.
1772 if (IsSackFrto() && (tp
->frto_counter
||
1773 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1774 after(tp
->high_seq
, tp
->snd_una
)) {
1775 tp
->frto_highmark
= tp
->high_seq
;
1777 tp
->frto_highmark
= tp
->snd_nxt
;
1779 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1780 tp
->high_seq
= tp
->snd_nxt
;
1781 tp
->frto_counter
= 1;
1784 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1785 * which indicates that we should follow the traditional RTO recovery,
1786 * i.e. mark everything lost and do go-back-N retransmission.
1788 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1790 struct tcp_sock
*tp
= tcp_sk(sk
);
1791 struct sk_buff
*skb
;
1794 tp
->retrans_out
= 0;
1795 if (tcp_is_reno(tp
))
1796 tcp_reset_reno_sack(tp
);
1798 tcp_for_write_queue(skb
, sk
) {
1799 if (skb
== tcp_send_head(sk
))
1802 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1804 * Count the retransmission made on RTO correctly (only when
1805 * waiting for the first ACK and did not get it)...
1807 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1808 /* For some reason this R-bit might get cleared? */
1809 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1810 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1811 /* ...enter this if branch just for the first segment */
1812 flag
|= FLAG_DATA_ACKED
;
1814 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1815 tp
->undo_marker
= 0;
1816 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1819 /* Don't lost mark skbs that were fwd transmitted after RTO */
1820 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) &&
1821 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
)) {
1822 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1823 tp
->lost_out
+= tcp_skb_pcount(skb
);
1826 tcp_verify_left_out(tp
);
1828 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1829 tp
->snd_cwnd_cnt
= 0;
1830 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1831 tp
->frto_counter
= 0;
1832 tp
->bytes_acked
= 0;
1834 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1835 sysctl_tcp_reordering
);
1836 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1837 tp
->high_seq
= tp
->frto_highmark
;
1838 TCP_ECN_queue_cwr(tp
);
1840 tcp_clear_retrans_hints_partial(tp
);
1843 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1845 tp
->retrans_out
= 0;
1848 tp
->undo_marker
= 0;
1849 tp
->undo_retrans
= 0;
1852 void tcp_clear_retrans(struct tcp_sock
*tp
)
1854 tcp_clear_retrans_partial(tp
);
1856 tp
->fackets_out
= 0;
1860 /* Enter Loss state. If "how" is not zero, forget all SACK information
1861 * and reset tags completely, otherwise preserve SACKs. If receiver
1862 * dropped its ofo queue, we will know this due to reneging detection.
1864 void tcp_enter_loss(struct sock
*sk
, int how
)
1866 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1867 struct tcp_sock
*tp
= tcp_sk(sk
);
1868 struct sk_buff
*skb
;
1870 /* Reduce ssthresh if it has not yet been made inside this window. */
1871 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1872 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1873 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1874 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1875 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1878 tp
->snd_cwnd_cnt
= 0;
1879 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1881 tp
->bytes_acked
= 0;
1882 tcp_clear_retrans_partial(tp
);
1884 if (tcp_is_reno(tp
))
1885 tcp_reset_reno_sack(tp
);
1888 /* Push undo marker, if it was plain RTO and nothing
1889 * was retransmitted. */
1890 tp
->undo_marker
= tp
->snd_una
;
1891 tcp_clear_retrans_hints_partial(tp
);
1894 tp
->fackets_out
= 0;
1895 tcp_clear_all_retrans_hints(tp
);
1898 tcp_for_write_queue(skb
, sk
) {
1899 if (skb
== tcp_send_head(sk
))
1902 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1903 tp
->undo_marker
= 0;
1904 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1905 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1906 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1907 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1908 tp
->lost_out
+= tcp_skb_pcount(skb
);
1911 tcp_verify_left_out(tp
);
1913 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1914 sysctl_tcp_reordering
);
1915 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1916 tp
->high_seq
= tp
->snd_nxt
;
1917 TCP_ECN_queue_cwr(tp
);
1918 /* Abort F-RTO algorithm if one is in progress */
1919 tp
->frto_counter
= 0;
1922 /* If ACK arrived pointing to a remembered SACK, it means that our
1923 * remembered SACKs do not reflect real state of receiver i.e.
1924 * receiver _host_ is heavily congested (or buggy).
1926 * Do processing similar to RTO timeout.
1928 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1930 if (flag
& FLAG_SACK_RENEGING
) {
1931 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1932 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1934 tcp_enter_loss(sk
, 1);
1935 icsk
->icsk_retransmits
++;
1936 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1937 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1938 icsk
->icsk_rto
, TCP_RTO_MAX
);
1944 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1946 return tcp_is_reno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1949 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1950 * counter when SACK is enabled (without SACK, sacked_out is used for
1953 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1954 * segments up to the highest received SACK block so far and holes in
1957 * With reordering, holes may still be in flight, so RFC3517 recovery
1958 * uses pure sacked_out (total number of SACKed segments) even though
1959 * it violates the RFC that uses duplicate ACKs, often these are equal
1960 * but when e.g. out-of-window ACKs or packet duplication occurs,
1961 * they differ. Since neither occurs due to loss, TCP should really
1964 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
1966 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1969 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1971 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1974 static inline int tcp_head_timedout(struct sock
*sk
)
1976 struct tcp_sock
*tp
= tcp_sk(sk
);
1978 return tp
->packets_out
&&
1979 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1982 /* Linux NewReno/SACK/FACK/ECN state machine.
1983 * --------------------------------------
1985 * "Open" Normal state, no dubious events, fast path.
1986 * "Disorder" In all the respects it is "Open",
1987 * but requires a bit more attention. It is entered when
1988 * we see some SACKs or dupacks. It is split of "Open"
1989 * mainly to move some processing from fast path to slow one.
1990 * "CWR" CWND was reduced due to some Congestion Notification event.
1991 * It can be ECN, ICMP source quench, local device congestion.
1992 * "Recovery" CWND was reduced, we are fast-retransmitting.
1993 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1995 * tcp_fastretrans_alert() is entered:
1996 * - each incoming ACK, if state is not "Open"
1997 * - when arrived ACK is unusual, namely:
2002 * Counting packets in flight is pretty simple.
2004 * in_flight = packets_out - left_out + retrans_out
2006 * packets_out is SND.NXT-SND.UNA counted in packets.
2008 * retrans_out is number of retransmitted segments.
2010 * left_out is number of segments left network, but not ACKed yet.
2012 * left_out = sacked_out + lost_out
2014 * sacked_out: Packets, which arrived to receiver out of order
2015 * and hence not ACKed. With SACKs this number is simply
2016 * amount of SACKed data. Even without SACKs
2017 * it is easy to give pretty reliable estimate of this number,
2018 * counting duplicate ACKs.
2020 * lost_out: Packets lost by network. TCP has no explicit
2021 * "loss notification" feedback from network (for now).
2022 * It means that this number can be only _guessed_.
2023 * Actually, it is the heuristics to predict lossage that
2024 * distinguishes different algorithms.
2026 * F.e. after RTO, when all the queue is considered as lost,
2027 * lost_out = packets_out and in_flight = retrans_out.
2029 * Essentially, we have now two algorithms counting
2032 * FACK: It is the simplest heuristics. As soon as we decided
2033 * that something is lost, we decide that _all_ not SACKed
2034 * packets until the most forward SACK are lost. I.e.
2035 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2036 * It is absolutely correct estimate, if network does not reorder
2037 * packets. And it loses any connection to reality when reordering
2038 * takes place. We use FACK by default until reordering
2039 * is suspected on the path to this destination.
2041 * NewReno: when Recovery is entered, we assume that one segment
2042 * is lost (classic Reno). While we are in Recovery and
2043 * a partial ACK arrives, we assume that one more packet
2044 * is lost (NewReno). This heuristics are the same in NewReno
2047 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2048 * deflation etc. CWND is real congestion window, never inflated, changes
2049 * only according to classic VJ rules.
2051 * Really tricky (and requiring careful tuning) part of algorithm
2052 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2053 * The first determines the moment _when_ we should reduce CWND and,
2054 * hence, slow down forward transmission. In fact, it determines the moment
2055 * when we decide that hole is caused by loss, rather than by a reorder.
2057 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2058 * holes, caused by lost packets.
2060 * And the most logically complicated part of algorithm is undo
2061 * heuristics. We detect false retransmits due to both too early
2062 * fast retransmit (reordering) and underestimated RTO, analyzing
2063 * timestamps and D-SACKs. When we detect that some segments were
2064 * retransmitted by mistake and CWND reduction was wrong, we undo
2065 * window reduction and abort recovery phase. This logic is hidden
2066 * inside several functions named tcp_try_undo_<something>.
2069 /* This function decides, when we should leave Disordered state
2070 * and enter Recovery phase, reducing congestion window.
2072 * Main question: may we further continue forward transmission
2073 * with the same cwnd?
2075 static int tcp_time_to_recover(struct sock
*sk
)
2077 struct tcp_sock
*tp
= tcp_sk(sk
);
2080 /* Do not perform any recovery during F-RTO algorithm */
2081 if (tp
->frto_counter
)
2084 /* Trick#1: The loss is proven. */
2088 /* Not-A-Trick#2 : Classic rule... */
2089 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2092 /* Trick#3 : when we use RFC2988 timer restart, fast
2093 * retransmit can be triggered by timeout of queue head.
2095 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2098 /* Trick#4: It is still not OK... But will it be useful to delay
2101 packets_out
= tp
->packets_out
;
2102 if (packets_out
<= tp
->reordering
&&
2103 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2104 !tcp_may_send_now(sk
)) {
2105 /* We have nothing to send. This connection is limited
2106 * either by receiver window or by application.
2114 /* RFC: This is from the original, I doubt that this is necessary at all:
2115 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2116 * retransmitted past LOST markings in the first place? I'm not fully sure
2117 * about undo and end of connection cases, which can cause R without L?
2119 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
2120 struct sk_buff
*skb
)
2122 if ((tp
->retransmit_skb_hint
!= NULL
) &&
2123 before(TCP_SKB_CB(skb
)->seq
,
2124 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
2125 tp
->retransmit_skb_hint
= NULL
;
2128 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2129 * is against sacked "cnt", otherwise it's against facked "cnt"
2131 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int fast_rexmit
)
2133 struct tcp_sock
*tp
= tcp_sk(sk
);
2134 struct sk_buff
*skb
;
2137 BUG_TRAP(packets
<= tp
->packets_out
);
2138 if (tp
->lost_skb_hint
) {
2139 skb
= tp
->lost_skb_hint
;
2140 cnt
= tp
->lost_cnt_hint
;
2142 skb
= tcp_write_queue_head(sk
);
2146 tcp_for_write_queue_from(skb
, sk
) {
2147 if (skb
== tcp_send_head(sk
))
2149 /* TODO: do this better */
2150 /* this is not the most efficient way to do this... */
2151 tp
->lost_skb_hint
= skb
;
2152 tp
->lost_cnt_hint
= cnt
;
2154 if (tcp_is_fack(tp
) ||
2155 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2156 cnt
+= tcp_skb_pcount(skb
);
2158 if (((!fast_rexmit
|| (tp
->lost_out
> 0)) && (cnt
> packets
)) ||
2159 after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2161 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2162 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2163 tp
->lost_out
+= tcp_skb_pcount(skb
);
2164 tcp_verify_retransmit_hint(tp
, skb
);
2167 tcp_verify_left_out(tp
);
2170 /* Account newly detected lost packet(s) */
2172 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2174 struct tcp_sock
*tp
= tcp_sk(sk
);
2176 if (tcp_is_reno(tp
)) {
2177 tcp_mark_head_lost(sk
, 1, fast_rexmit
);
2178 } else if (tcp_is_fack(tp
)) {
2179 int lost
= tp
->fackets_out
- tp
->reordering
;
2182 tcp_mark_head_lost(sk
, lost
, fast_rexmit
);
2184 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2185 if (sacked_upto
< 0)
2187 tcp_mark_head_lost(sk
, sacked_upto
, fast_rexmit
);
2190 /* New heuristics: it is possible only after we switched
2191 * to restart timer each time when something is ACKed.
2192 * Hence, we can detect timed out packets during fast
2193 * retransmit without falling to slow start.
2195 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2196 struct sk_buff
*skb
;
2198 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2199 : tcp_write_queue_head(sk
);
2201 tcp_for_write_queue_from(skb
, sk
) {
2202 if (skb
== tcp_send_head(sk
))
2204 if (!tcp_skb_timedout(sk
, skb
))
2207 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2208 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2209 tp
->lost_out
+= tcp_skb_pcount(skb
);
2210 tcp_verify_retransmit_hint(tp
, skb
);
2214 tp
->scoreboard_skb_hint
= skb
;
2216 tcp_verify_left_out(tp
);
2220 /* CWND moderation, preventing bursts due to too big ACKs
2221 * in dubious situations.
2223 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2225 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2226 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
2227 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2230 /* Lower bound on congestion window is slow start threshold
2231 * unless congestion avoidance choice decides to overide it.
2233 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2235 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2237 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2240 /* Decrease cwnd each second ack. */
2241 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2243 struct tcp_sock
*tp
= tcp_sk(sk
);
2244 int decr
= tp
->snd_cwnd_cnt
+ 1;
2246 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
2247 (tcp_is_reno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
2248 tp
->snd_cwnd_cnt
= decr
&1;
2251 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2252 tp
->snd_cwnd
-= decr
;
2254 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
2255 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2259 /* Nothing was retransmitted or returned timestamp is less
2260 * than timestamp of the first retransmission.
2262 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2264 return !tp
->retrans_stamp
||
2265 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2266 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
2269 /* Undo procedures. */
2271 #if FASTRETRANS_DEBUG > 1
2272 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2274 struct tcp_sock
*tp
= tcp_sk(sk
);
2275 struct inet_sock
*inet
= inet_sk(sk
);
2277 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2279 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2280 tp
->snd_cwnd
, tcp_left_out(tp
),
2281 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2285 #define DBGUNDO(x...) do { } while (0)
2288 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2290 struct tcp_sock
*tp
= tcp_sk(sk
);
2292 if (tp
->prior_ssthresh
) {
2293 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2295 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2296 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2298 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
2300 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2301 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2302 TCP_ECN_withdraw_cwr(tp
);
2305 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2307 tcp_moderate_cwnd(tp
);
2308 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2310 /* There is something screwy going on with the retrans hints after
2312 tcp_clear_all_retrans_hints(tp
);
2315 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2317 return tp
->undo_marker
&&
2318 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2321 /* People celebrate: "We love our President!" */
2322 static int tcp_try_undo_recovery(struct sock
*sk
)
2324 struct tcp_sock
*tp
= tcp_sk(sk
);
2326 if (tcp_may_undo(tp
)) {
2327 /* Happy end! We did not retransmit anything
2328 * or our original transmission succeeded.
2330 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2331 tcp_undo_cwr(sk
, 1);
2332 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2333 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2335 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
2336 tp
->undo_marker
= 0;
2338 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2339 /* Hold old state until something *above* high_seq
2340 * is ACKed. For Reno it is MUST to prevent false
2341 * fast retransmits (RFC2582). SACK TCP is safe. */
2342 tcp_moderate_cwnd(tp
);
2345 tcp_set_ca_state(sk
, TCP_CA_Open
);
2349 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2350 static void tcp_try_undo_dsack(struct sock
*sk
)
2352 struct tcp_sock
*tp
= tcp_sk(sk
);
2354 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2355 DBGUNDO(sk
, "D-SACK");
2356 tcp_undo_cwr(sk
, 1);
2357 tp
->undo_marker
= 0;
2358 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2362 /* Undo during fast recovery after partial ACK. */
2364 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2366 struct tcp_sock
*tp
= tcp_sk(sk
);
2367 /* Partial ACK arrived. Force Hoe's retransmit. */
2368 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2370 if (tcp_may_undo(tp
)) {
2371 /* Plain luck! Hole if filled with delayed
2372 * packet, rather than with a retransmit.
2374 if (tp
->retrans_out
== 0)
2375 tp
->retrans_stamp
= 0;
2377 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2380 tcp_undo_cwr(sk
, 0);
2381 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2383 /* So... Do not make Hoe's retransmit yet.
2384 * If the first packet was delayed, the rest
2385 * ones are most probably delayed as well.
2392 /* Undo during loss recovery after partial ACK. */
2393 static int tcp_try_undo_loss(struct sock
*sk
)
2395 struct tcp_sock
*tp
= tcp_sk(sk
);
2397 if (tcp_may_undo(tp
)) {
2398 struct sk_buff
*skb
;
2399 tcp_for_write_queue(skb
, sk
) {
2400 if (skb
== tcp_send_head(sk
))
2402 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2405 tcp_clear_all_retrans_hints(tp
);
2407 DBGUNDO(sk
, "partial loss");
2409 tcp_undo_cwr(sk
, 1);
2410 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2411 inet_csk(sk
)->icsk_retransmits
= 0;
2412 tp
->undo_marker
= 0;
2413 if (tcp_is_sack(tp
))
2414 tcp_set_ca_state(sk
, TCP_CA_Open
);
2420 static inline void tcp_complete_cwr(struct sock
*sk
)
2422 struct tcp_sock
*tp
= tcp_sk(sk
);
2423 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2424 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2425 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2428 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2430 struct tcp_sock
*tp
= tcp_sk(sk
);
2432 tcp_verify_left_out(tp
);
2434 if (tp
->retrans_out
== 0)
2435 tp
->retrans_stamp
= 0;
2438 tcp_enter_cwr(sk
, 1);
2440 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2441 int state
= TCP_CA_Open
;
2443 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2444 state
= TCP_CA_Disorder
;
2446 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2447 tcp_set_ca_state(sk
, state
);
2448 tp
->high_seq
= tp
->snd_nxt
;
2450 tcp_moderate_cwnd(tp
);
2452 tcp_cwnd_down(sk
, flag
);
2456 static void tcp_mtup_probe_failed(struct sock
*sk
)
2458 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2460 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2461 icsk
->icsk_mtup
.probe_size
= 0;
2464 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2466 struct tcp_sock
*tp
= tcp_sk(sk
);
2467 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2469 /* FIXME: breaks with very large cwnd */
2470 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2471 tp
->snd_cwnd
= tp
->snd_cwnd
*
2472 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2473 icsk
->icsk_mtup
.probe_size
;
2474 tp
->snd_cwnd_cnt
= 0;
2475 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2476 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2478 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2479 icsk
->icsk_mtup
.probe_size
= 0;
2480 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2484 /* Process an event, which can update packets-in-flight not trivially.
2485 * Main goal of this function is to calculate new estimate for left_out,
2486 * taking into account both packets sitting in receiver's buffer and
2487 * packets lost by network.
2489 * Besides that it does CWND reduction, when packet loss is detected
2490 * and changes state of machine.
2492 * It does _not_ decide what to send, it is made in function
2493 * tcp_xmit_retransmit_queue().
2496 tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2498 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2499 struct tcp_sock
*tp
= tcp_sk(sk
);
2500 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2501 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2502 (tcp_fackets_out(tp
) > tp
->reordering
));
2503 int fast_rexmit
= 0;
2505 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2507 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2508 tp
->fackets_out
= 0;
2510 /* Now state machine starts.
2511 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2513 tp
->prior_ssthresh
= 0;
2515 /* B. In all the states check for reneging SACKs. */
2516 if (tcp_check_sack_reneging(sk
, flag
))
2519 /* C. Process data loss notification, provided it is valid. */
2520 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2521 before(tp
->snd_una
, tp
->high_seq
) &&
2522 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2523 tp
->fackets_out
> tp
->reordering
) {
2524 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, 0);
2525 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2528 /* D. Check consistency of the current state. */
2529 tcp_verify_left_out(tp
);
2531 /* E. Check state exit conditions. State can be terminated
2532 * when high_seq is ACKed. */
2533 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2534 BUG_TRAP(tp
->retrans_out
== 0);
2535 tp
->retrans_stamp
= 0;
2536 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2537 switch (icsk
->icsk_ca_state
) {
2539 icsk
->icsk_retransmits
= 0;
2540 if (tcp_try_undo_recovery(sk
))
2545 /* CWR is to be held something *above* high_seq
2546 * is ACKed for CWR bit to reach receiver. */
2547 if (tp
->snd_una
!= tp
->high_seq
) {
2548 tcp_complete_cwr(sk
);
2549 tcp_set_ca_state(sk
, TCP_CA_Open
);
2553 case TCP_CA_Disorder
:
2554 tcp_try_undo_dsack(sk
);
2555 if (!tp
->undo_marker
||
2556 /* For SACK case do not Open to allow to undo
2557 * catching for all duplicate ACKs. */
2558 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2559 tp
->undo_marker
= 0;
2560 tcp_set_ca_state(sk
, TCP_CA_Open
);
2564 case TCP_CA_Recovery
:
2565 if (tcp_is_reno(tp
))
2566 tcp_reset_reno_sack(tp
);
2567 if (tcp_try_undo_recovery(sk
))
2569 tcp_complete_cwr(sk
);
2574 /* F. Process state. */
2575 switch (icsk
->icsk_ca_state
) {
2576 case TCP_CA_Recovery
:
2577 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2578 if (tcp_is_reno(tp
) && is_dupack
)
2579 tcp_add_reno_sack(sk
);
2581 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2584 if (flag
&FLAG_DATA_ACKED
)
2585 icsk
->icsk_retransmits
= 0;
2586 if (!tcp_try_undo_loss(sk
)) {
2587 tcp_moderate_cwnd(tp
);
2588 tcp_xmit_retransmit_queue(sk
);
2591 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2593 /* Loss is undone; fall through to processing in Open state. */
2595 if (tcp_is_reno(tp
)) {
2596 if (flag
& FLAG_SND_UNA_ADVANCED
)
2597 tcp_reset_reno_sack(tp
);
2599 tcp_add_reno_sack(sk
);
2602 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2603 tcp_try_undo_dsack(sk
);
2605 if (!tcp_time_to_recover(sk
)) {
2606 tcp_try_to_open(sk
, flag
);
2610 /* MTU probe failure: don't reduce cwnd */
2611 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2612 icsk
->icsk_mtup
.probe_size
&&
2613 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2614 tcp_mtup_probe_failed(sk
);
2615 /* Restores the reduction we did in tcp_mtup_probe() */
2617 tcp_simple_retransmit(sk
);
2621 /* Otherwise enter Recovery state */
2623 if (tcp_is_reno(tp
))
2624 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2626 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2628 tp
->high_seq
= tp
->snd_nxt
;
2629 tp
->prior_ssthresh
= 0;
2630 tp
->undo_marker
= tp
->snd_una
;
2631 tp
->undo_retrans
= tp
->retrans_out
;
2633 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2634 if (!(flag
&FLAG_ECE
))
2635 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2636 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2637 TCP_ECN_queue_cwr(tp
);
2640 tp
->bytes_acked
= 0;
2641 tp
->snd_cwnd_cnt
= 0;
2642 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2646 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2647 tcp_update_scoreboard(sk
, fast_rexmit
);
2648 tcp_cwnd_down(sk
, flag
);
2649 tcp_xmit_retransmit_queue(sk
);
2652 /* Read draft-ietf-tcplw-high-performance before mucking
2653 * with this code. (Supersedes RFC1323)
2655 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2657 /* RTTM Rule: A TSecr value received in a segment is used to
2658 * update the averaged RTT measurement only if the segment
2659 * acknowledges some new data, i.e., only if it advances the
2660 * left edge of the send window.
2662 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2663 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2665 * Changed: reset backoff as soon as we see the first valid sample.
2666 * If we do not, we get strongly overestimated rto. With timestamps
2667 * samples are accepted even from very old segments: f.e., when rtt=1
2668 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2669 * answer arrives rto becomes 120 seconds! If at least one of segments
2670 * in window is lost... Voila. --ANK (010210)
2672 struct tcp_sock
*tp
= tcp_sk(sk
);
2673 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2674 tcp_rtt_estimator(sk
, seq_rtt
);
2676 inet_csk(sk
)->icsk_backoff
= 0;
2680 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2682 /* We don't have a timestamp. Can only use
2683 * packets that are not retransmitted to determine
2684 * rtt estimates. Also, we must not reset the
2685 * backoff for rto until we get a non-retransmitted
2686 * packet. This allows us to deal with a situation
2687 * where the network delay has increased suddenly.
2688 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2691 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2694 tcp_rtt_estimator(sk
, seq_rtt
);
2696 inet_csk(sk
)->icsk_backoff
= 0;
2700 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2703 const struct tcp_sock
*tp
= tcp_sk(sk
);
2704 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2705 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2706 tcp_ack_saw_tstamp(sk
, flag
);
2707 else if (seq_rtt
>= 0)
2708 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2711 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2713 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2714 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2715 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2718 /* Restart timer after forward progress on connection.
2719 * RFC2988 recommends to restart timer to now+rto.
2721 static void tcp_rearm_rto(struct sock
*sk
)
2723 struct tcp_sock
*tp
= tcp_sk(sk
);
2725 if (!tp
->packets_out
) {
2726 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2728 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2732 /* If we get here, the whole TSO packet has not been acked. */
2733 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2735 struct tcp_sock
*tp
= tcp_sk(sk
);
2738 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2740 packets_acked
= tcp_skb_pcount(skb
);
2741 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2743 packets_acked
-= tcp_skb_pcount(skb
);
2745 if (packets_acked
) {
2746 BUG_ON(tcp_skb_pcount(skb
) == 0);
2747 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2750 return packets_acked
;
2753 /* Remove acknowledged frames from the retransmission queue. If our packet
2754 * is before the ack sequence we can discard it as it's confirmed to have
2755 * arrived at the other end.
2757 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
)
2759 struct tcp_sock
*tp
= tcp_sk(sk
);
2760 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2761 struct sk_buff
*skb
;
2762 u32 now
= tcp_time_stamp
;
2763 int fully_acked
= 1;
2766 u32 reord
= tp
->packets_out
;
2768 s32 ca_seq_rtt
= -1;
2769 ktime_t last_ackt
= net_invalid_timestamp();
2771 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2772 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2775 u8 sacked
= scb
->sacked
;
2777 /* Determine how many packets and what bytes were acked, tso and else */
2778 if (after(scb
->end_seq
, tp
->snd_una
)) {
2779 if (tcp_skb_pcount(skb
) == 1 ||
2780 !after(tp
->snd_una
, scb
->seq
))
2783 acked_pcount
= tcp_tso_acked(sk
, skb
);
2788 end_seq
= tp
->snd_una
;
2790 acked_pcount
= tcp_skb_pcount(skb
);
2791 end_seq
= scb
->end_seq
;
2794 /* MTU probing checks */
2795 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2796 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2797 tcp_mtup_probe_success(sk
, skb
);
2800 if (sacked
& TCPCB_RETRANS
) {
2801 if (sacked
& TCPCB_SACKED_RETRANS
)
2802 tp
->retrans_out
-= acked_pcount
;
2803 flag
|= FLAG_RETRANS_DATA_ACKED
;
2806 if ((flag
& FLAG_DATA_ACKED
) ||
2808 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2810 ca_seq_rtt
= now
- scb
->when
;
2811 last_ackt
= skb
->tstamp
;
2813 seq_rtt
= ca_seq_rtt
;
2815 if (!(sacked
& TCPCB_SACKED_ACKED
))
2816 reord
= min(pkts_acked
, reord
);
2819 if (sacked
& TCPCB_SACKED_ACKED
)
2820 tp
->sacked_out
-= acked_pcount
;
2821 if (sacked
& TCPCB_LOST
)
2822 tp
->lost_out
-= acked_pcount
;
2824 if (unlikely((sacked
& TCPCB_URG
) && tp
->urg_mode
&&
2825 !before(end_seq
, tp
->snd_up
)))
2828 tp
->packets_out
-= acked_pcount
;
2829 pkts_acked
+= acked_pcount
;
2831 /* Initial outgoing SYN's get put onto the write_queue
2832 * just like anything else we transmit. It is not
2833 * true data, and if we misinform our callers that
2834 * this ACK acks real data, we will erroneously exit
2835 * connection startup slow start one packet too
2836 * quickly. This is severely frowned upon behavior.
2838 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2839 flag
|= FLAG_DATA_ACKED
;
2841 flag
|= FLAG_SYN_ACKED
;
2842 tp
->retrans_stamp
= 0;
2848 tcp_unlink_write_queue(skb
, sk
);
2849 sk_wmem_free_skb(sk
, skb
);
2850 tcp_clear_all_retrans_hints(tp
);
2853 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2854 flag
|= FLAG_SACK_RENEGING
;
2856 if (flag
& FLAG_ACKED
) {
2857 const struct tcp_congestion_ops
*ca_ops
2858 = inet_csk(sk
)->icsk_ca_ops
;
2860 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2863 if (tcp_is_reno(tp
)) {
2864 tcp_remove_reno_sacks(sk
, pkts_acked
);
2866 /* Non-retransmitted hole got filled? That's reordering */
2867 if (reord
< prior_fackets
)
2868 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2871 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2873 if (ca_ops
->pkts_acked
) {
2876 /* Is the ACK triggering packet unambiguous? */
2877 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2878 /* High resolution needed and available? */
2879 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2880 !ktime_equal(last_ackt
,
2881 net_invalid_timestamp()))
2882 rtt_us
= ktime_us_delta(ktime_get_real(),
2884 else if (ca_seq_rtt
> 0)
2885 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
2888 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2892 #if FASTRETRANS_DEBUG > 0
2893 BUG_TRAP((int)tp
->sacked_out
>= 0);
2894 BUG_TRAP((int)tp
->lost_out
>= 0);
2895 BUG_TRAP((int)tp
->retrans_out
>= 0);
2896 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2897 icsk
= inet_csk(sk
);
2899 printk(KERN_DEBUG
"Leak l=%u %d\n",
2900 tp
->lost_out
, icsk
->icsk_ca_state
);
2903 if (tp
->sacked_out
) {
2904 printk(KERN_DEBUG
"Leak s=%u %d\n",
2905 tp
->sacked_out
, icsk
->icsk_ca_state
);
2908 if (tp
->retrans_out
) {
2909 printk(KERN_DEBUG
"Leak r=%u %d\n",
2910 tp
->retrans_out
, icsk
->icsk_ca_state
);
2911 tp
->retrans_out
= 0;
2918 static void tcp_ack_probe(struct sock
*sk
)
2920 const struct tcp_sock
*tp
= tcp_sk(sk
);
2921 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2923 /* Was it a usable window open? */
2925 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
2926 icsk
->icsk_backoff
= 0;
2927 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2928 /* Socket must be waked up by subsequent tcp_data_snd_check().
2929 * This function is not for random using!
2932 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2933 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2938 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2940 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2941 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2944 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2946 const struct tcp_sock
*tp
= tcp_sk(sk
);
2947 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2948 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2951 /* Check that window update is acceptable.
2952 * The function assumes that snd_una<=ack<=snd_next.
2954 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2955 const u32 ack_seq
, const u32 nwin
)
2957 return (after(ack
, tp
->snd_una
) ||
2958 after(ack_seq
, tp
->snd_wl1
) ||
2959 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2962 /* Update our send window.
2964 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2965 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2967 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2970 struct tcp_sock
*tp
= tcp_sk(sk
);
2972 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2974 if (likely(!tcp_hdr(skb
)->syn
))
2975 nwin
<<= tp
->rx_opt
.snd_wscale
;
2977 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2978 flag
|= FLAG_WIN_UPDATE
;
2979 tcp_update_wl(tp
, ack
, ack_seq
);
2981 if (tp
->snd_wnd
!= nwin
) {
2984 /* Note, it is the only place, where
2985 * fast path is recovered for sending TCP.
2988 tcp_fast_path_check(sk
);
2990 if (nwin
> tp
->max_window
) {
2991 tp
->max_window
= nwin
;
2992 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3002 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3003 * continue in congestion avoidance.
3005 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3007 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3008 tp
->snd_cwnd_cnt
= 0;
3009 tp
->bytes_acked
= 0;
3010 TCP_ECN_queue_cwr(tp
);
3011 tcp_moderate_cwnd(tp
);
3014 /* A conservative spurious RTO response algorithm: reduce cwnd using
3015 * rate halving and continue in congestion avoidance.
3017 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3019 tcp_enter_cwr(sk
, 0);
3022 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3025 tcp_ratehalving_spur_to_response(sk
);
3027 tcp_undo_cwr(sk
, 1);
3030 /* F-RTO spurious RTO detection algorithm (RFC4138)
3032 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3033 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3034 * window (but not to or beyond highest sequence sent before RTO):
3035 * On First ACK, send two new segments out.
3036 * On Second ACK, RTO was likely spurious. Do spurious response (response
3037 * algorithm is not part of the F-RTO detection algorithm
3038 * given in RFC4138 but can be selected separately).
3039 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3040 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3041 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3042 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3044 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3045 * original window even after we transmit two new data segments.
3048 * on first step, wait until first cumulative ACK arrives, then move to
3049 * the second step. In second step, the next ACK decides.
3051 * F-RTO is implemented (mainly) in four functions:
3052 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3053 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3054 * called when tcp_use_frto() showed green light
3055 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3056 * - tcp_enter_frto_loss() is called if there is not enough evidence
3057 * to prove that the RTO is indeed spurious. It transfers the control
3058 * from F-RTO to the conventional RTO recovery
3060 static int tcp_process_frto(struct sock
*sk
, int flag
)
3062 struct tcp_sock
*tp
= tcp_sk(sk
);
3064 tcp_verify_left_out(tp
);
3066 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3067 if (flag
&FLAG_DATA_ACKED
)
3068 inet_csk(sk
)->icsk_retransmits
= 0;
3070 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3071 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3072 tp
->undo_marker
= 0;
3074 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3075 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3079 if (!IsSackFrto() || tcp_is_reno(tp
)) {
3080 /* RFC4138 shortcoming in step 2; should also have case c):
3081 * ACK isn't duplicate nor advances window, e.g., opposite dir
3084 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3087 if (!(flag
&FLAG_DATA_ACKED
)) {
3088 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3093 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3094 /* Prevent sending of new data. */
3095 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3096 tcp_packets_in_flight(tp
));
3100 if ((tp
->frto_counter
>= 2) &&
3101 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
3102 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
3103 /* RFC4138 shortcoming (see comment above) */
3104 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3107 tcp_enter_frto_loss(sk
, 3, flag
);
3112 if (tp
->frto_counter
== 1) {
3113 /* tcp_may_send_now needs to see updated state */
3114 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3115 tp
->frto_counter
= 2;
3117 if (!tcp_may_send_now(sk
))
3118 tcp_enter_frto_loss(sk
, 2, flag
);
3122 switch (sysctl_tcp_frto_response
) {
3124 tcp_undo_spur_to_response(sk
, flag
);
3127 tcp_conservative_spur_to_response(tp
);
3130 tcp_ratehalving_spur_to_response(sk
);
3133 tp
->frto_counter
= 0;
3134 tp
->undo_marker
= 0;
3135 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS
);
3140 /* This routine deals with incoming acks, but not outgoing ones. */
3141 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3143 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3144 struct tcp_sock
*tp
= tcp_sk(sk
);
3145 u32 prior_snd_una
= tp
->snd_una
;
3146 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3147 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3148 u32 prior_in_flight
;
3153 /* If the ack is newer than sent or older than previous acks
3154 * then we can probably ignore it.
3156 if (after(ack
, tp
->snd_nxt
))
3157 goto uninteresting_ack
;
3159 if (before(ack
, prior_snd_una
))
3162 if (after(ack
, prior_snd_una
))
3163 flag
|= FLAG_SND_UNA_ADVANCED
;
3165 if (sysctl_tcp_abc
) {
3166 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3167 tp
->bytes_acked
+= ack
- prior_snd_una
;
3168 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3169 /* we assume just one segment left network */
3170 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
3173 prior_fackets
= tp
->fackets_out
;
3174 prior_in_flight
= tcp_packets_in_flight(tp
);
3176 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3177 /* Window is constant, pure forward advance.
3178 * No more checks are required.
3179 * Note, we use the fact that SND.UNA>=SND.WL2.
3181 tcp_update_wl(tp
, ack
, ack_seq
);
3183 flag
|= FLAG_WIN_UPDATE
;
3185 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3187 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
3189 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3192 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
3194 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3196 if (TCP_SKB_CB(skb
)->sacked
)
3197 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3199 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3202 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3205 /* We passed data and got it acked, remove any soft error
3206 * log. Something worked...
3208 sk
->sk_err_soft
= 0;
3209 tp
->rcv_tstamp
= tcp_time_stamp
;
3210 prior_packets
= tp
->packets_out
;
3214 /* See if we can take anything off of the retransmit queue. */
3215 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
);
3217 if (tp
->frto_counter
)
3218 frto_cwnd
= tcp_process_frto(sk
, flag
);
3219 /* Guarantee sacktag reordering detection against wrap-arounds */
3220 if (before(tp
->frto_highmark
, tp
->snd_una
))
3221 tp
->frto_highmark
= 0;
3223 if (tcp_ack_is_dubious(sk
, flag
)) {
3224 /* Advance CWND, if state allows this. */
3225 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3226 tcp_may_raise_cwnd(sk
, flag
))
3227 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3228 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
, flag
);
3230 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3231 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3234 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
3235 dst_confirm(sk
->sk_dst_cache
);
3240 icsk
->icsk_probes_out
= 0;
3242 /* If this ack opens up a zero window, clear backoff. It was
3243 * being used to time the probes, and is probably far higher than
3244 * it needs to be for normal retransmission.
3246 if (tcp_send_head(sk
))
3251 if (TCP_SKB_CB(skb
)->sacked
)
3252 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3255 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3260 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3261 * But, this can also be called on packets in the established flow when
3262 * the fast version below fails.
3264 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
3267 struct tcphdr
*th
= tcp_hdr(skb
);
3268 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
3270 ptr
= (unsigned char *)(th
+ 1);
3271 opt_rx
->saw_tstamp
= 0;
3273 while (length
> 0) {
3280 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3285 if (opsize
< 2) /* "silly options" */
3287 if (opsize
> length
)
3288 return; /* don't parse partial options */
3291 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3292 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
3294 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
3295 in_mss
= opt_rx
->user_mss
;
3296 opt_rx
->mss_clamp
= in_mss
;
3301 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
3302 if (sysctl_tcp_window_scaling
) {
3303 __u8 snd_wscale
= *(__u8
*) ptr
;
3304 opt_rx
->wscale_ok
= 1;
3305 if (snd_wscale
> 14) {
3306 if (net_ratelimit())
3307 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3308 "scaling value %d >14 received.\n",
3312 opt_rx
->snd_wscale
= snd_wscale
;
3315 case TCPOPT_TIMESTAMP
:
3316 if (opsize
==TCPOLEN_TIMESTAMP
) {
3317 if ((estab
&& opt_rx
->tstamp_ok
) ||
3318 (!estab
&& sysctl_tcp_timestamps
)) {
3319 opt_rx
->saw_tstamp
= 1;
3320 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
3321 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
3325 case TCPOPT_SACK_PERM
:
3326 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
3327 if (sysctl_tcp_sack
) {
3328 opt_rx
->sack_ok
= 1;
3329 tcp_sack_reset(opt_rx
);
3335 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3336 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3338 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3341 #ifdef CONFIG_TCP_MD5SIG
3344 * The MD5 Hash has already been
3345 * checked (see tcp_v{4,6}_do_rcv()).
3357 /* Fast parse options. This hopes to only see timestamps.
3358 * If it is wrong it falls back on tcp_parse_options().
3360 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3361 struct tcp_sock
*tp
)
3363 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3364 tp
->rx_opt
.saw_tstamp
= 0;
3366 } else if (tp
->rx_opt
.tstamp_ok
&&
3367 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3368 __be32
*ptr
= (__be32
*)(th
+ 1);
3369 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3370 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3371 tp
->rx_opt
.saw_tstamp
= 1;
3373 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3375 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3379 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3383 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3385 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3386 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3389 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3391 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3392 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3393 * extra check below makes sure this can only happen
3394 * for pure ACK frames. -DaveM
3396 * Not only, also it occurs for expired timestamps.
3399 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3400 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3401 tcp_store_ts_recent(tp
);
3405 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3407 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3408 * it can pass through stack. So, the following predicate verifies that
3409 * this segment is not used for anything but congestion avoidance or
3410 * fast retransmit. Moreover, we even are able to eliminate most of such
3411 * second order effects, if we apply some small "replay" window (~RTO)
3412 * to timestamp space.
3414 * All these measures still do not guarantee that we reject wrapped ACKs
3415 * on networks with high bandwidth, when sequence space is recycled fastly,
3416 * but it guarantees that such events will be very rare and do not affect
3417 * connection seriously. This doesn't look nice, but alas, PAWS is really
3420 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3421 * states that events when retransmit arrives after original data are rare.
3422 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3423 * the biggest problem on large power networks even with minor reordering.
3424 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3425 * up to bandwidth of 18Gigabit/sec. 8) ]
3428 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3430 struct tcp_sock
*tp
= tcp_sk(sk
);
3431 struct tcphdr
*th
= tcp_hdr(skb
);
3432 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3433 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3435 return (/* 1. Pure ACK with correct sequence number. */
3436 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3438 /* 2. ... and duplicate ACK. */
3439 ack
== tp
->snd_una
&&
3441 /* 3. ... and does not update window. */
3442 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3444 /* 4. ... and sits in replay window. */
3445 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3448 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3450 const struct tcp_sock
*tp
= tcp_sk(sk
);
3451 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3452 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3453 !tcp_disordered_ack(sk
, skb
));
3456 /* Check segment sequence number for validity.
3458 * Segment controls are considered valid, if the segment
3459 * fits to the window after truncation to the window. Acceptability
3460 * of data (and SYN, FIN, of course) is checked separately.
3461 * See tcp_data_queue(), for example.
3463 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3464 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3465 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3466 * (borrowed from freebsd)
3469 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3471 return !before(end_seq
, tp
->rcv_wup
) &&
3472 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3475 /* When we get a reset we do this. */
3476 static void tcp_reset(struct sock
*sk
)
3478 /* We want the right error as BSD sees it (and indeed as we do). */
3479 switch (sk
->sk_state
) {
3481 sk
->sk_err
= ECONNREFUSED
;
3483 case TCP_CLOSE_WAIT
:
3489 sk
->sk_err
= ECONNRESET
;
3492 if (!sock_flag(sk
, SOCK_DEAD
))
3493 sk
->sk_error_report(sk
);
3499 * Process the FIN bit. This now behaves as it is supposed to work
3500 * and the FIN takes effect when it is validly part of sequence
3501 * space. Not before when we get holes.
3503 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3504 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3507 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3508 * close and we go into CLOSING (and later onto TIME-WAIT)
3510 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3512 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3514 struct tcp_sock
*tp
= tcp_sk(sk
);
3516 inet_csk_schedule_ack(sk
);
3518 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3519 sock_set_flag(sk
, SOCK_DONE
);
3521 switch (sk
->sk_state
) {
3523 case TCP_ESTABLISHED
:
3524 /* Move to CLOSE_WAIT */
3525 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3526 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3529 case TCP_CLOSE_WAIT
:
3531 /* Received a retransmission of the FIN, do
3536 /* RFC793: Remain in the LAST-ACK state. */
3540 /* This case occurs when a simultaneous close
3541 * happens, we must ack the received FIN and
3542 * enter the CLOSING state.
3545 tcp_set_state(sk
, TCP_CLOSING
);
3548 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3550 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3553 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3554 * cases we should never reach this piece of code.
3556 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3557 __FUNCTION__
, sk
->sk_state
);
3561 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3562 * Probably, we should reset in this case. For now drop them.
3564 __skb_queue_purge(&tp
->out_of_order_queue
);
3565 if (tcp_is_sack(tp
))
3566 tcp_sack_reset(&tp
->rx_opt
);
3569 if (!sock_flag(sk
, SOCK_DEAD
)) {
3570 sk
->sk_state_change(sk
);
3572 /* Do not send POLL_HUP for half duplex close. */
3573 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3574 sk
->sk_state
== TCP_CLOSE
)
3575 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3577 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3581 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3583 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3584 if (before(seq
, sp
->start_seq
))
3585 sp
->start_seq
= seq
;
3586 if (after(end_seq
, sp
->end_seq
))
3587 sp
->end_seq
= end_seq
;
3593 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3595 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3596 if (before(seq
, tp
->rcv_nxt
))
3597 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3599 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3601 tp
->rx_opt
.dsack
= 1;
3602 tp
->duplicate_sack
[0].start_seq
= seq
;
3603 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3604 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3608 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3610 if (!tp
->rx_opt
.dsack
)
3611 tcp_dsack_set(tp
, seq
, end_seq
);
3613 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3616 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3618 struct tcp_sock
*tp
= tcp_sk(sk
);
3620 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3621 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3622 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3623 tcp_enter_quickack_mode(sk
);
3625 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3626 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3628 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3629 end_seq
= tp
->rcv_nxt
;
3630 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3637 /* These routines update the SACK block as out-of-order packets arrive or
3638 * in-order packets close up the sequence space.
3640 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3643 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3644 struct tcp_sack_block
*swalk
= sp
+1;
3646 /* See if the recent change to the first SACK eats into
3647 * or hits the sequence space of other SACK blocks, if so coalesce.
3649 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3650 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3653 /* Zap SWALK, by moving every further SACK up by one slot.
3654 * Decrease num_sacks.
3656 tp
->rx_opt
.num_sacks
--;
3657 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3658 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3662 this_sack
++, swalk
++;
3666 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3670 tmp
= sack1
->start_seq
;
3671 sack1
->start_seq
= sack2
->start_seq
;
3672 sack2
->start_seq
= tmp
;
3674 tmp
= sack1
->end_seq
;
3675 sack1
->end_seq
= sack2
->end_seq
;
3676 sack2
->end_seq
= tmp
;
3679 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3681 struct tcp_sock
*tp
= tcp_sk(sk
);
3682 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3683 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3689 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3690 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3691 /* Rotate this_sack to the first one. */
3692 for (; this_sack
>0; this_sack
--, sp
--)
3693 tcp_sack_swap(sp
, sp
-1);
3695 tcp_sack_maybe_coalesce(tp
);
3700 /* Could not find an adjacent existing SACK, build a new one,
3701 * put it at the front, and shift everyone else down. We
3702 * always know there is at least one SACK present already here.
3704 * If the sack array is full, forget about the last one.
3706 if (this_sack
>= 4) {
3708 tp
->rx_opt
.num_sacks
--;
3711 for (; this_sack
> 0; this_sack
--, sp
--)
3715 /* Build the new head SACK, and we're done. */
3716 sp
->start_seq
= seq
;
3717 sp
->end_seq
= end_seq
;
3718 tp
->rx_opt
.num_sacks
++;
3719 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3722 /* RCV.NXT advances, some SACKs should be eaten. */
3724 static void tcp_sack_remove(struct tcp_sock
*tp
)
3726 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3727 int num_sacks
= tp
->rx_opt
.num_sacks
;
3730 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3731 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3732 tp
->rx_opt
.num_sacks
= 0;
3733 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3737 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3738 /* Check if the start of the sack is covered by RCV.NXT. */
3739 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3742 /* RCV.NXT must cover all the block! */
3743 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3745 /* Zap this SACK, by moving forward any other SACKS. */
3746 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3747 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3754 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3755 tp
->rx_opt
.num_sacks
= num_sacks
;
3756 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3760 /* This one checks to see if we can put data from the
3761 * out_of_order queue into the receive_queue.
3763 static void tcp_ofo_queue(struct sock
*sk
)
3765 struct tcp_sock
*tp
= tcp_sk(sk
);
3766 __u32 dsack_high
= tp
->rcv_nxt
;
3767 struct sk_buff
*skb
;
3769 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3770 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3773 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3774 __u32 dsack
= dsack_high
;
3775 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3776 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3777 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3780 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3781 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3782 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3786 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3787 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3788 TCP_SKB_CB(skb
)->end_seq
);
3790 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3791 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3792 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3793 if (tcp_hdr(skb
)->fin
)
3794 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3798 static int tcp_prune_queue(struct sock
*sk
);
3800 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3802 struct tcphdr
*th
= tcp_hdr(skb
);
3803 struct tcp_sock
*tp
= tcp_sk(sk
);
3806 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3809 __skb_pull(skb
, th
->doff
*4);
3811 TCP_ECN_accept_cwr(tp
, skb
);
3813 if (tp
->rx_opt
.dsack
) {
3814 tp
->rx_opt
.dsack
= 0;
3815 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3816 4 - tp
->rx_opt
.tstamp_ok
);
3819 /* Queue data for delivery to the user.
3820 * Packets in sequence go to the receive queue.
3821 * Out of sequence packets to the out_of_order_queue.
3823 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3824 if (tcp_receive_window(tp
) == 0)
3827 /* Ok. In sequence. In window. */
3828 if (tp
->ucopy
.task
== current
&&
3829 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3830 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3831 int chunk
= min_t(unsigned int, skb
->len
,
3834 __set_current_state(TASK_RUNNING
);
3837 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3838 tp
->ucopy
.len
-= chunk
;
3839 tp
->copied_seq
+= chunk
;
3840 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3841 tcp_rcv_space_adjust(sk
);
3849 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3850 !sk_rmem_schedule(sk
, skb
->truesize
))) {
3851 if (tcp_prune_queue(sk
) < 0 ||
3852 !sk_rmem_schedule(sk
, skb
->truesize
))
3855 skb_set_owner_r(skb
, sk
);
3856 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3858 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3860 tcp_event_data_recv(sk
, skb
);
3862 tcp_fin(skb
, sk
, th
);
3864 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3867 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3868 * gap in queue is filled.
3870 if (skb_queue_empty(&tp
->out_of_order_queue
))
3871 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3874 if (tp
->rx_opt
.num_sacks
)
3875 tcp_sack_remove(tp
);
3877 tcp_fast_path_check(sk
);
3881 else if (!sock_flag(sk
, SOCK_DEAD
))
3882 sk
->sk_data_ready(sk
, 0);
3886 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3887 /* A retransmit, 2nd most common case. Force an immediate ack. */
3888 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3889 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3892 tcp_enter_quickack_mode(sk
);
3893 inet_csk_schedule_ack(sk
);
3899 /* Out of window. F.e. zero window probe. */
3900 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3903 tcp_enter_quickack_mode(sk
);
3905 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3906 /* Partial packet, seq < rcv_next < end_seq */
3907 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3908 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3909 TCP_SKB_CB(skb
)->end_seq
);
3911 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3913 /* If window is closed, drop tail of packet. But after
3914 * remembering D-SACK for its head made in previous line.
3916 if (!tcp_receive_window(tp
))
3921 TCP_ECN_check_ce(tp
, skb
);
3923 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3924 !sk_rmem_schedule(sk
, skb
->truesize
)) {
3925 if (tcp_prune_queue(sk
) < 0 ||
3926 !sk_rmem_schedule(sk
, skb
->truesize
))
3930 /* Disable header prediction. */
3932 inet_csk_schedule_ack(sk
);
3934 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3935 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3937 skb_set_owner_r(skb
, sk
);
3939 if (!skb_peek(&tp
->out_of_order_queue
)) {
3940 /* Initial out of order segment, build 1 SACK. */
3941 if (tcp_is_sack(tp
)) {
3942 tp
->rx_opt
.num_sacks
= 1;
3943 tp
->rx_opt
.dsack
= 0;
3944 tp
->rx_opt
.eff_sacks
= 1;
3945 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3946 tp
->selective_acks
[0].end_seq
=
3947 TCP_SKB_CB(skb
)->end_seq
;
3949 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3951 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3952 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3953 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3955 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3956 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3958 if (!tp
->rx_opt
.num_sacks
||
3959 tp
->selective_acks
[0].end_seq
!= seq
)
3962 /* Common case: data arrive in order after hole. */
3963 tp
->selective_acks
[0].end_seq
= end_seq
;
3967 /* Find place to insert this segment. */
3969 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3971 } while ((skb1
= skb1
->prev
) !=
3972 (struct sk_buff
*)&tp
->out_of_order_queue
);
3974 /* Do skb overlap to previous one? */
3975 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3976 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3977 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3978 /* All the bits are present. Drop. */
3980 tcp_dsack_set(tp
, seq
, end_seq
);
3983 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3984 /* Partial overlap. */
3985 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3990 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3992 /* And clean segments covered by new one as whole. */
3993 while ((skb1
= skb
->next
) !=
3994 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3995 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3996 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3997 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
4000 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4001 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
4006 if (tcp_is_sack(tp
))
4007 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4011 /* Collapse contiguous sequence of skbs head..tail with
4012 * sequence numbers start..end.
4013 * Segments with FIN/SYN are not collapsed (only because this
4017 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4018 struct sk_buff
*head
, struct sk_buff
*tail
,
4021 struct sk_buff
*skb
;
4023 /* First, check that queue is collapsible and find
4024 * the point where collapsing can be useful. */
4025 for (skb
= head
; skb
!= tail
; ) {
4026 /* No new bits? It is possible on ofo queue. */
4027 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4028 struct sk_buff
*next
= skb
->next
;
4029 __skb_unlink(skb
, list
);
4031 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
4036 /* The first skb to collapse is:
4038 * - bloated or contains data before "start" or
4039 * overlaps to the next one.
4041 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4042 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4043 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4044 (skb
->next
!= tail
&&
4045 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4048 /* Decided to skip this, advance start seq. */
4049 start
= TCP_SKB_CB(skb
)->end_seq
;
4052 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4055 while (before(start
, end
)) {
4056 struct sk_buff
*nskb
;
4057 unsigned int header
= skb_headroom(skb
);
4058 int copy
= SKB_MAX_ORDER(header
, 0);
4060 /* Too big header? This can happen with IPv6. */
4063 if (end
-start
< copy
)
4065 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
4069 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4070 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4072 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4074 skb_reserve(nskb
, header
);
4075 memcpy(nskb
->head
, skb
->head
, header
);
4076 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4077 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4078 __skb_insert(nskb
, skb
->prev
, skb
, list
);
4079 skb_set_owner_r(nskb
, sk
);
4081 /* Copy data, releasing collapsed skbs. */
4083 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4084 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4088 size
= min(copy
, size
);
4089 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4091 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4095 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4096 struct sk_buff
*next
= skb
->next
;
4097 __skb_unlink(skb
, list
);
4099 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
4102 tcp_hdr(skb
)->syn
||
4110 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4111 * and tcp_collapse() them until all the queue is collapsed.
4113 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4115 struct tcp_sock
*tp
= tcp_sk(sk
);
4116 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4117 struct sk_buff
*head
;
4123 start
= TCP_SKB_CB(skb
)->seq
;
4124 end
= TCP_SKB_CB(skb
)->end_seq
;
4130 /* Segment is terminated when we see gap or when
4131 * we are at the end of all the queue. */
4132 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4133 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4134 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4135 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4136 head
, skb
, start
, end
);
4138 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4140 /* Start new segment */
4141 start
= TCP_SKB_CB(skb
)->seq
;
4142 end
= TCP_SKB_CB(skb
)->end_seq
;
4144 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4145 start
= TCP_SKB_CB(skb
)->seq
;
4146 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4147 end
= TCP_SKB_CB(skb
)->end_seq
;
4152 /* Reduce allocated memory if we can, trying to get
4153 * the socket within its memory limits again.
4155 * Return less than zero if we should start dropping frames
4156 * until the socket owning process reads some of the data
4157 * to stabilize the situation.
4159 static int tcp_prune_queue(struct sock
*sk
)
4161 struct tcp_sock
*tp
= tcp_sk(sk
);
4163 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4165 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
4167 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4168 tcp_clamp_window(sk
);
4169 else if (tcp_memory_pressure
)
4170 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4172 tcp_collapse_ofo_queue(sk
);
4173 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4174 sk
->sk_receive_queue
.next
,
4175 (struct sk_buff
*)&sk
->sk_receive_queue
,
4176 tp
->copied_seq
, tp
->rcv_nxt
);
4179 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4182 /* Collapsing did not help, destructive actions follow.
4183 * This must not ever occur. */
4185 /* First, purge the out_of_order queue. */
4186 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4187 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
4188 __skb_queue_purge(&tp
->out_of_order_queue
);
4190 /* Reset SACK state. A conforming SACK implementation will
4191 * do the same at a timeout based retransmit. When a connection
4192 * is in a sad state like this, we care only about integrity
4193 * of the connection not performance.
4195 if (tcp_is_sack(tp
))
4196 tcp_sack_reset(&tp
->rx_opt
);
4200 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4203 /* If we are really being abused, tell the caller to silently
4204 * drop receive data on the floor. It will get retransmitted
4205 * and hopefully then we'll have sufficient space.
4207 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
4209 /* Massive buffer overcommit. */
4215 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4216 * As additional protections, we do not touch cwnd in retransmission phases,
4217 * and if application hit its sndbuf limit recently.
4219 void tcp_cwnd_application_limited(struct sock
*sk
)
4221 struct tcp_sock
*tp
= tcp_sk(sk
);
4223 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4224 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4225 /* Limited by application or receiver window. */
4226 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4227 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4228 if (win_used
< tp
->snd_cwnd
) {
4229 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4230 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4232 tp
->snd_cwnd_used
= 0;
4234 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4237 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4239 struct tcp_sock
*tp
= tcp_sk(sk
);
4241 /* If the user specified a specific send buffer setting, do
4244 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4247 /* If we are under global TCP memory pressure, do not expand. */
4248 if (tcp_memory_pressure
)
4251 /* If we are under soft global TCP memory pressure, do not expand. */
4252 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4255 /* If we filled the congestion window, do not expand. */
4256 if (tp
->packets_out
>= tp
->snd_cwnd
)
4262 /* When incoming ACK allowed to free some skb from write_queue,
4263 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4264 * on the exit from tcp input handler.
4266 * PROBLEM: sndbuf expansion does not work well with largesend.
4268 static void tcp_new_space(struct sock
*sk
)
4270 struct tcp_sock
*tp
= tcp_sk(sk
);
4272 if (tcp_should_expand_sndbuf(sk
)) {
4273 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4274 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4275 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4276 tp
->reordering
+ 1);
4277 sndmem
*= 2*demanded
;
4278 if (sndmem
> sk
->sk_sndbuf
)
4279 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4280 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4283 sk
->sk_write_space(sk
);
4286 static void tcp_check_space(struct sock
*sk
)
4288 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4289 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4290 if (sk
->sk_socket
&&
4291 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4296 static inline void tcp_data_snd_check(struct sock
*sk
)
4298 tcp_push_pending_frames(sk
);
4299 tcp_check_space(sk
);
4303 * Check if sending an ack is needed.
4305 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4307 struct tcp_sock
*tp
= tcp_sk(sk
);
4309 /* More than one full frame received... */
4310 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4311 /* ... and right edge of window advances far enough.
4312 * (tcp_recvmsg() will send ACK otherwise). Or...
4314 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4315 /* We ACK each frame or... */
4316 tcp_in_quickack_mode(sk
) ||
4317 /* We have out of order data. */
4319 skb_peek(&tp
->out_of_order_queue
))) {
4320 /* Then ack it now */
4323 /* Else, send delayed ack. */
4324 tcp_send_delayed_ack(sk
);
4328 static inline void tcp_ack_snd_check(struct sock
*sk
)
4330 if (!inet_csk_ack_scheduled(sk
)) {
4331 /* We sent a data segment already. */
4334 __tcp_ack_snd_check(sk
, 1);
4338 * This routine is only called when we have urgent data
4339 * signaled. Its the 'slow' part of tcp_urg. It could be
4340 * moved inline now as tcp_urg is only called from one
4341 * place. We handle URGent data wrong. We have to - as
4342 * BSD still doesn't use the correction from RFC961.
4343 * For 1003.1g we should support a new option TCP_STDURG to permit
4344 * either form (or just set the sysctl tcp_stdurg).
4347 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4349 struct tcp_sock
*tp
= tcp_sk(sk
);
4350 u32 ptr
= ntohs(th
->urg_ptr
);
4352 if (ptr
&& !sysctl_tcp_stdurg
)
4354 ptr
+= ntohl(th
->seq
);
4356 /* Ignore urgent data that we've already seen and read. */
4357 if (after(tp
->copied_seq
, ptr
))
4360 /* Do not replay urg ptr.
4362 * NOTE: interesting situation not covered by specs.
4363 * Misbehaving sender may send urg ptr, pointing to segment,
4364 * which we already have in ofo queue. We are not able to fetch
4365 * such data and will stay in TCP_URG_NOTYET until will be eaten
4366 * by recvmsg(). Seems, we are not obliged to handle such wicked
4367 * situations. But it is worth to think about possibility of some
4368 * DoSes using some hypothetical application level deadlock.
4370 if (before(ptr
, tp
->rcv_nxt
))
4373 /* Do we already have a newer (or duplicate) urgent pointer? */
4374 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4377 /* Tell the world about our new urgent pointer. */
4380 /* We may be adding urgent data when the last byte read was
4381 * urgent. To do this requires some care. We cannot just ignore
4382 * tp->copied_seq since we would read the last urgent byte again
4383 * as data, nor can we alter copied_seq until this data arrives
4384 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4386 * NOTE. Double Dutch. Rendering to plain English: author of comment
4387 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4388 * and expect that both A and B disappear from stream. This is _wrong_.
4389 * Though this happens in BSD with high probability, this is occasional.
4390 * Any application relying on this is buggy. Note also, that fix "works"
4391 * only in this artificial test. Insert some normal data between A and B and we will
4392 * decline of BSD again. Verdict: it is better to remove to trap
4395 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4396 !sock_flag(sk
, SOCK_URGINLINE
) &&
4397 tp
->copied_seq
!= tp
->rcv_nxt
) {
4398 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4400 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4401 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4406 tp
->urg_data
= TCP_URG_NOTYET
;
4409 /* Disable header prediction. */
4413 /* This is the 'fast' part of urgent handling. */
4414 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4416 struct tcp_sock
*tp
= tcp_sk(sk
);
4418 /* Check if we get a new urgent pointer - normally not. */
4420 tcp_check_urg(sk
,th
);
4422 /* Do we wait for any urgent data? - normally not... */
4423 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4424 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4427 /* Is the urgent pointer pointing into this packet? */
4428 if (ptr
< skb
->len
) {
4430 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4432 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4433 if (!sock_flag(sk
, SOCK_DEAD
))
4434 sk
->sk_data_ready(sk
, 0);
4439 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4441 struct tcp_sock
*tp
= tcp_sk(sk
);
4442 int chunk
= skb
->len
- hlen
;
4446 if (skb_csum_unnecessary(skb
))
4447 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4449 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4453 tp
->ucopy
.len
-= chunk
;
4454 tp
->copied_seq
+= chunk
;
4455 tcp_rcv_space_adjust(sk
);
4462 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4466 if (sock_owned_by_user(sk
)) {
4468 result
= __tcp_checksum_complete(skb
);
4471 result
= __tcp_checksum_complete(skb
);
4476 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4478 return !skb_csum_unnecessary(skb
) &&
4479 __tcp_checksum_complete_user(sk
, skb
);
4482 #ifdef CONFIG_NET_DMA
4483 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4485 struct tcp_sock
*tp
= tcp_sk(sk
);
4486 int chunk
= skb
->len
- hlen
;
4488 int copied_early
= 0;
4490 if (tp
->ucopy
.wakeup
)
4493 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4494 tp
->ucopy
.dma_chan
= get_softnet_dma();
4496 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4498 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4499 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4504 tp
->ucopy
.dma_cookie
= dma_cookie
;
4507 tp
->ucopy
.len
-= chunk
;
4508 tp
->copied_seq
+= chunk
;
4509 tcp_rcv_space_adjust(sk
);
4511 if ((tp
->ucopy
.len
== 0) ||
4512 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4513 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4514 tp
->ucopy
.wakeup
= 1;
4515 sk
->sk_data_ready(sk
, 0);
4517 } else if (chunk
> 0) {
4518 tp
->ucopy
.wakeup
= 1;
4519 sk
->sk_data_ready(sk
, 0);
4522 return copied_early
;
4524 #endif /* CONFIG_NET_DMA */
4527 * TCP receive function for the ESTABLISHED state.
4529 * It is split into a fast path and a slow path. The fast path is
4531 * - A zero window was announced from us - zero window probing
4532 * is only handled properly in the slow path.
4533 * - Out of order segments arrived.
4534 * - Urgent data is expected.
4535 * - There is no buffer space left
4536 * - Unexpected TCP flags/window values/header lengths are received
4537 * (detected by checking the TCP header against pred_flags)
4538 * - Data is sent in both directions. Fast path only supports pure senders
4539 * or pure receivers (this means either the sequence number or the ack
4540 * value must stay constant)
4541 * - Unexpected TCP option.
4543 * When these conditions are not satisfied it drops into a standard
4544 * receive procedure patterned after RFC793 to handle all cases.
4545 * The first three cases are guaranteed by proper pred_flags setting,
4546 * the rest is checked inline. Fast processing is turned on in
4547 * tcp_data_queue when everything is OK.
4549 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4550 struct tcphdr
*th
, unsigned len
)
4552 struct tcp_sock
*tp
= tcp_sk(sk
);
4555 * Header prediction.
4556 * The code loosely follows the one in the famous
4557 * "30 instruction TCP receive" Van Jacobson mail.
4559 * Van's trick is to deposit buffers into socket queue
4560 * on a device interrupt, to call tcp_recv function
4561 * on the receive process context and checksum and copy
4562 * the buffer to user space. smart...
4564 * Our current scheme is not silly either but we take the
4565 * extra cost of the net_bh soft interrupt processing...
4566 * We do checksum and copy also but from device to kernel.
4569 tp
->rx_opt
.saw_tstamp
= 0;
4571 /* pred_flags is 0xS?10 << 16 + snd_wnd
4572 * if header_prediction is to be made
4573 * 'S' will always be tp->tcp_header_len >> 2
4574 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4575 * turn it off (when there are holes in the receive
4576 * space for instance)
4577 * PSH flag is ignored.
4580 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4581 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4582 int tcp_header_len
= tp
->tcp_header_len
;
4584 /* Timestamp header prediction: tcp_header_len
4585 * is automatically equal to th->doff*4 due to pred_flags
4589 /* Check timestamp */
4590 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4591 __be32
*ptr
= (__be32
*)(th
+ 1);
4593 /* No? Slow path! */
4594 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4595 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4598 tp
->rx_opt
.saw_tstamp
= 1;
4600 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4602 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4604 /* If PAWS failed, check it more carefully in slow path */
4605 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4608 /* DO NOT update ts_recent here, if checksum fails
4609 * and timestamp was corrupted part, it will result
4610 * in a hung connection since we will drop all
4611 * future packets due to the PAWS test.
4615 if (len
<= tcp_header_len
) {
4616 /* Bulk data transfer: sender */
4617 if (len
== tcp_header_len
) {
4618 /* Predicted packet is in window by definition.
4619 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4620 * Hence, check seq<=rcv_wup reduces to:
4622 if (tcp_header_len
==
4623 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4624 tp
->rcv_nxt
== tp
->rcv_wup
)
4625 tcp_store_ts_recent(tp
);
4627 /* We know that such packets are checksummed
4630 tcp_ack(sk
, skb
, 0);
4632 tcp_data_snd_check(sk
);
4634 } else { /* Header too small */
4635 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4640 int copied_early
= 0;
4642 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4643 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4644 #ifdef CONFIG_NET_DMA
4645 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4650 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4651 __set_current_state(TASK_RUNNING
);
4653 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4657 /* Predicted packet is in window by definition.
4658 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4659 * Hence, check seq<=rcv_wup reduces to:
4661 if (tcp_header_len
==
4662 (sizeof(struct tcphdr
) +
4663 TCPOLEN_TSTAMP_ALIGNED
) &&
4664 tp
->rcv_nxt
== tp
->rcv_wup
)
4665 tcp_store_ts_recent(tp
);
4667 tcp_rcv_rtt_measure_ts(sk
, skb
);
4669 __skb_pull(skb
, tcp_header_len
);
4670 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4671 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4674 tcp_cleanup_rbuf(sk
, skb
->len
);
4677 if (tcp_checksum_complete_user(sk
, skb
))
4680 /* Predicted packet is in window by definition.
4681 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4682 * Hence, check seq<=rcv_wup reduces to:
4684 if (tcp_header_len
==
4685 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4686 tp
->rcv_nxt
== tp
->rcv_wup
)
4687 tcp_store_ts_recent(tp
);
4689 tcp_rcv_rtt_measure_ts(sk
, skb
);
4691 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4694 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4696 /* Bulk data transfer: receiver */
4697 __skb_pull(skb
,tcp_header_len
);
4698 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4699 skb_set_owner_r(skb
, sk
);
4700 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4703 tcp_event_data_recv(sk
, skb
);
4705 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4706 /* Well, only one small jumplet in fast path... */
4707 tcp_ack(sk
, skb
, FLAG_DATA
);
4708 tcp_data_snd_check(sk
);
4709 if (!inet_csk_ack_scheduled(sk
))
4713 __tcp_ack_snd_check(sk
, 0);
4715 #ifdef CONFIG_NET_DMA
4717 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4723 sk
->sk_data_ready(sk
, 0);
4729 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4733 * RFC1323: H1. Apply PAWS check first.
4735 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4736 tcp_paws_discard(sk
, skb
)) {
4738 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4739 tcp_send_dupack(sk
, skb
);
4742 /* Resets are accepted even if PAWS failed.
4744 ts_recent update must be made after we are sure
4745 that the packet is in window.
4750 * Standard slow path.
4753 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4754 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4755 * (RST) segments are validated by checking their SEQ-fields."
4756 * And page 69: "If an incoming segment is not acceptable,
4757 * an acknowledgment should be sent in reply (unless the RST bit
4758 * is set, if so drop the segment and return)".
4761 tcp_send_dupack(sk
, skb
);
4770 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4772 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4773 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4774 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4781 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4783 tcp_rcv_rtt_measure_ts(sk
, skb
);
4785 /* Process urgent data. */
4786 tcp_urg(sk
, skb
, th
);
4788 /* step 7: process the segment text */
4789 tcp_data_queue(sk
, skb
);
4791 tcp_data_snd_check(sk
);
4792 tcp_ack_snd_check(sk
);
4796 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4803 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4804 struct tcphdr
*th
, unsigned len
)
4806 struct tcp_sock
*tp
= tcp_sk(sk
);
4807 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4808 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4810 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4814 * "If the state is SYN-SENT then
4815 * first check the ACK bit
4816 * If the ACK bit is set
4817 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4818 * a reset (unless the RST bit is set, if so drop
4819 * the segment and return)"
4821 * We do not send data with SYN, so that RFC-correct
4824 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4825 goto reset_and_undo
;
4827 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4828 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4830 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4831 goto reset_and_undo
;
4834 /* Now ACK is acceptable.
4836 * "If the RST bit is set
4837 * If the ACK was acceptable then signal the user "error:
4838 * connection reset", drop the segment, enter CLOSED state,
4839 * delete TCB, and return."
4848 * "fifth, if neither of the SYN or RST bits is set then
4849 * drop the segment and return."
4855 goto discard_and_undo
;
4858 * "If the SYN bit is on ...
4859 * are acceptable then ...
4860 * (our SYN has been ACKed), change the connection
4861 * state to ESTABLISHED..."
4864 TCP_ECN_rcv_synack(tp
, th
);
4866 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4867 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4869 /* Ok.. it's good. Set up sequence numbers and
4870 * move to established.
4872 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4873 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4875 /* RFC1323: The window in SYN & SYN/ACK segments is
4878 tp
->snd_wnd
= ntohs(th
->window
);
4879 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4881 if (!tp
->rx_opt
.wscale_ok
) {
4882 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4883 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4886 if (tp
->rx_opt
.saw_tstamp
) {
4887 tp
->rx_opt
.tstamp_ok
= 1;
4888 tp
->tcp_header_len
=
4889 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4890 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4891 tcp_store_ts_recent(tp
);
4893 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4896 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
4897 tcp_enable_fack(tp
);
4900 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4901 tcp_initialize_rcv_mss(sk
);
4903 /* Remember, tcp_poll() does not lock socket!
4904 * Change state from SYN-SENT only after copied_seq
4905 * is initialized. */
4906 tp
->copied_seq
= tp
->rcv_nxt
;
4908 tcp_set_state(sk
, TCP_ESTABLISHED
);
4910 security_inet_conn_established(sk
, skb
);
4912 /* Make sure socket is routed, for correct metrics. */
4913 icsk
->icsk_af_ops
->rebuild_header(sk
);
4915 tcp_init_metrics(sk
);
4917 tcp_init_congestion_control(sk
);
4919 /* Prevent spurious tcp_cwnd_restart() on first data
4922 tp
->lsndtime
= tcp_time_stamp
;
4924 tcp_init_buffer_space(sk
);
4926 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4927 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4929 if (!tp
->rx_opt
.snd_wscale
)
4930 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4934 if (!sock_flag(sk
, SOCK_DEAD
)) {
4935 sk
->sk_state_change(sk
);
4936 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
4939 if (sk
->sk_write_pending
||
4940 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4941 icsk
->icsk_ack
.pingpong
) {
4942 /* Save one ACK. Data will be ready after
4943 * several ticks, if write_pending is set.
4945 * It may be deleted, but with this feature tcpdumps
4946 * look so _wonderfully_ clever, that I was not able
4947 * to stand against the temptation 8) --ANK
4949 inet_csk_schedule_ack(sk
);
4950 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4951 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4952 tcp_incr_quickack(sk
);
4953 tcp_enter_quickack_mode(sk
);
4954 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4955 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4966 /* No ACK in the segment */
4970 * "If the RST bit is set
4972 * Otherwise (no ACK) drop the segment and return."
4975 goto discard_and_undo
;
4979 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4980 goto discard_and_undo
;
4983 /* We see SYN without ACK. It is attempt of
4984 * simultaneous connect with crossed SYNs.
4985 * Particularly, it can be connect to self.
4987 tcp_set_state(sk
, TCP_SYN_RECV
);
4989 if (tp
->rx_opt
.saw_tstamp
) {
4990 tp
->rx_opt
.tstamp_ok
= 1;
4991 tcp_store_ts_recent(tp
);
4992 tp
->tcp_header_len
=
4993 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4995 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4998 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4999 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5001 /* RFC1323: The window in SYN & SYN/ACK segments is
5004 tp
->snd_wnd
= ntohs(th
->window
);
5005 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5006 tp
->max_window
= tp
->snd_wnd
;
5008 TCP_ECN_rcv_syn(tp
, th
);
5011 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5012 tcp_initialize_rcv_mss(sk
);
5015 tcp_send_synack(sk
);
5017 /* Note, we could accept data and URG from this segment.
5018 * There are no obstacles to make this.
5020 * However, if we ignore data in ACKless segments sometimes,
5021 * we have no reasons to accept it sometimes.
5022 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5023 * is not flawless. So, discard packet for sanity.
5024 * Uncomment this return to process the data.
5031 /* "fifth, if neither of the SYN or RST bits is set then
5032 * drop the segment and return."
5036 tcp_clear_options(&tp
->rx_opt
);
5037 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5041 tcp_clear_options(&tp
->rx_opt
);
5042 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5048 * This function implements the receiving procedure of RFC 793 for
5049 * all states except ESTABLISHED and TIME_WAIT.
5050 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5051 * address independent.
5054 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5055 struct tcphdr
*th
, unsigned len
)
5057 struct tcp_sock
*tp
= tcp_sk(sk
);
5058 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5061 tp
->rx_opt
.saw_tstamp
= 0;
5063 switch (sk
->sk_state
) {
5075 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5078 /* Now we have several options: In theory there is
5079 * nothing else in the frame. KA9Q has an option to
5080 * send data with the syn, BSD accepts data with the
5081 * syn up to the [to be] advertised window and
5082 * Solaris 2.1 gives you a protocol error. For now
5083 * we just ignore it, that fits the spec precisely
5084 * and avoids incompatibilities. It would be nice in
5085 * future to drop through and process the data.
5087 * Now that TTCP is starting to be used we ought to
5089 * But, this leaves one open to an easy denial of
5090 * service attack, and SYN cookies can't defend
5091 * against this problem. So, we drop the data
5092 * in the interest of security over speed unless
5093 * it's still in use.
5101 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5105 /* Do step6 onward by hand. */
5106 tcp_urg(sk
, skb
, th
);
5108 tcp_data_snd_check(sk
);
5112 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5113 tcp_paws_discard(sk
, skb
)) {
5115 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
5116 tcp_send_dupack(sk
, skb
);
5119 /* Reset is accepted even if it did not pass PAWS. */
5122 /* step 1: check sequence number */
5123 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5125 tcp_send_dupack(sk
, skb
);
5129 /* step 2: check RST bit */
5135 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5137 /* step 3: check security and precedence [ignored] */
5141 * Check for a SYN in window.
5143 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5144 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
5149 /* step 5: check the ACK field */
5151 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5153 switch (sk
->sk_state
) {
5156 tp
->copied_seq
= tp
->rcv_nxt
;
5158 tcp_set_state(sk
, TCP_ESTABLISHED
);
5159 sk
->sk_state_change(sk
);
5161 /* Note, that this wakeup is only for marginal
5162 * crossed SYN case. Passively open sockets
5163 * are not waked up, because sk->sk_sleep ==
5164 * NULL and sk->sk_socket == NULL.
5168 SOCK_WAKE_IO
, POLL_OUT
);
5170 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5171 tp
->snd_wnd
= ntohs(th
->window
) <<
5172 tp
->rx_opt
.snd_wscale
;
5173 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5174 TCP_SKB_CB(skb
)->seq
);
5176 /* tcp_ack considers this ACK as duplicate
5177 * and does not calculate rtt.
5178 * Fix it at least with timestamps.
5180 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5182 tcp_ack_saw_tstamp(sk
, 0);
5184 if (tp
->rx_opt
.tstamp_ok
)
5185 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5187 /* Make sure socket is routed, for
5190 icsk
->icsk_af_ops
->rebuild_header(sk
);
5192 tcp_init_metrics(sk
);
5194 tcp_init_congestion_control(sk
);
5196 /* Prevent spurious tcp_cwnd_restart() on
5197 * first data packet.
5199 tp
->lsndtime
= tcp_time_stamp
;
5202 tcp_initialize_rcv_mss(sk
);
5203 tcp_init_buffer_space(sk
);
5204 tcp_fast_path_on(tp
);
5211 if (tp
->snd_una
== tp
->write_seq
) {
5212 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5213 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5214 dst_confirm(sk
->sk_dst_cache
);
5216 if (!sock_flag(sk
, SOCK_DEAD
))
5217 /* Wake up lingering close() */
5218 sk
->sk_state_change(sk
);
5222 if (tp
->linger2
< 0 ||
5223 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5224 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5226 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5230 tmo
= tcp_fin_time(sk
);
5231 if (tmo
> TCP_TIMEWAIT_LEN
) {
5232 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5233 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5234 /* Bad case. We could lose such FIN otherwise.
5235 * It is not a big problem, but it looks confusing
5236 * and not so rare event. We still can lose it now,
5237 * if it spins in bh_lock_sock(), but it is really
5240 inet_csk_reset_keepalive_timer(sk
, tmo
);
5242 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5250 if (tp
->snd_una
== tp
->write_seq
) {
5251 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5257 if (tp
->snd_una
== tp
->write_seq
) {
5258 tcp_update_metrics(sk
);
5267 /* step 6: check the URG bit */
5268 tcp_urg(sk
, skb
, th
);
5270 /* step 7: process the segment text */
5271 switch (sk
->sk_state
) {
5272 case TCP_CLOSE_WAIT
:
5275 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5279 /* RFC 793 says to queue data in these states,
5280 * RFC 1122 says we MUST send a reset.
5281 * BSD 4.4 also does reset.
5283 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5284 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5285 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5286 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5292 case TCP_ESTABLISHED
:
5293 tcp_data_queue(sk
, skb
);
5298 /* tcp_data could move socket to TIME-WAIT */
5299 if (sk
->sk_state
!= TCP_CLOSE
) {
5300 tcp_data_snd_check(sk
);
5301 tcp_ack_snd_check(sk
);
5311 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5312 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5313 EXPORT_SYMBOL(tcp_parse_options
);
5314 EXPORT_SYMBOL(tcp_rcv_established
);
5315 EXPORT_SYMBOL(tcp_rcv_state_process
);
5316 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);