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 DSACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
120 /* Adapt the MSS value used to make delayed ack decision to the
123 static void tcp_measure_rcv_mss(struct sock
*sk
,
124 const struct sk_buff
*skb
)
126 struct inet_connection_sock
*icsk
= inet_csk(sk
);
127 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
130 icsk
->icsk_ack
.last_seg_size
= 0;
132 /* skb->len may jitter because of SACKs, even if peer
133 * sends good full-sized frames.
135 len
= skb_shinfo(skb
)->gso_size
?: skb
->len
;
136 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
137 icsk
->icsk_ack
.rcv_mss
= len
;
139 /* Otherwise, we make more careful check taking into account,
140 * that SACKs block is variable.
142 * "len" is invariant segment length, including TCP header.
144 len
+= skb
->data
- skb_transport_header(skb
);
145 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
146 /* If PSH is not set, packet should be
147 * full sized, provided peer TCP is not badly broken.
148 * This observation (if it is correct 8)) allows
149 * to handle super-low mtu links fairly.
151 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
152 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
153 /* Subtract also invariant (if peer is RFC compliant),
154 * tcp header plus fixed timestamp option length.
155 * Resulting "len" is MSS free of SACK jitter.
157 len
-= tcp_sk(sk
)->tcp_header_len
;
158 icsk
->icsk_ack
.last_seg_size
= len
;
160 icsk
->icsk_ack
.rcv_mss
= len
;
164 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
165 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
166 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
170 static void tcp_incr_quickack(struct sock
*sk
)
172 struct inet_connection_sock
*icsk
= inet_csk(sk
);
173 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
177 if (quickacks
> icsk
->icsk_ack
.quick
)
178 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
181 void tcp_enter_quickack_mode(struct sock
*sk
)
183 struct inet_connection_sock
*icsk
= inet_csk(sk
);
184 tcp_incr_quickack(sk
);
185 icsk
->icsk_ack
.pingpong
= 0;
186 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
189 /* Send ACKs quickly, if "quick" count is not exhausted
190 * and the session is not interactive.
193 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
195 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
196 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
201 if (tp
->ecn_flags
&TCP_ECN_OK
)
202 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
207 if (tcp_hdr(skb
)->cwr
)
208 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
213 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
216 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
218 if (tp
->ecn_flags
&TCP_ECN_OK
) {
219 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
220 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
221 /* Funny extension: if ECT is not set on a segment,
222 * it is surely retransmit. It is not in ECN RFC,
223 * but Linux follows this rule. */
224 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
225 tcp_enter_quickack_mode((struct sock
*)tp
);
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
231 if ((tp
->ecn_flags
&TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
232 tp
->ecn_flags
&= ~TCP_ECN_OK
;
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
237 if ((tp
->ecn_flags
&TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
238 tp
->ecn_flags
&= ~TCP_ECN_OK
;
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
243 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
&TCP_ECN_OK
))
248 /* Buffer size and advertised window tuning.
250 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
253 static void tcp_fixup_sndbuf(struct sock
*sk
)
255 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
256 sizeof(struct sk_buff
);
258 if (sk
->sk_sndbuf
< 3 * sndmem
)
259 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
264 * All tcp_full_space() is split to two parts: "network" buffer, allocated
265 * forward and advertised in receiver window (tp->rcv_wnd) and
266 * "application buffer", required to isolate scheduling/application
267 * latencies from network.
268 * window_clamp is maximal advertised window. It can be less than
269 * tcp_full_space(), in this case tcp_full_space() - window_clamp
270 * is reserved for "application" buffer. The less window_clamp is
271 * the smoother our behaviour from viewpoint of network, but the lower
272 * throughput and the higher sensitivity of the connection to losses. 8)
274 * rcv_ssthresh is more strict window_clamp used at "slow start"
275 * phase to predict further behaviour of this connection.
276 * It is used for two goals:
277 * - to enforce header prediction at sender, even when application
278 * requires some significant "application buffer". It is check #1.
279 * - to prevent pruning of receive queue because of misprediction
280 * of receiver window. Check #2.
282 * The scheme does not work when sender sends good segments opening
283 * window and then starts to feed us spaghetti. But it should work
284 * in common situations. Otherwise, we have to rely on queue collapsing.
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
290 struct tcp_sock
*tp
= tcp_sk(sk
);
292 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
293 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2])/2;
295 while (tp
->rcv_ssthresh
<= window
) {
296 if (truesize
<= skb
->len
)
297 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
305 static void tcp_grow_window(struct sock
*sk
,
308 struct tcp_sock
*tp
= tcp_sk(sk
);
311 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
312 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
313 !tcp_memory_pressure
) {
316 /* Check #2. Increase window, if skb with such overhead
317 * will fit to rcvbuf in future.
319 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
322 incr
= __tcp_grow_window(sk
, skb
);
325 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
326 inet_csk(sk
)->icsk_ack
.quick
|= 1;
331 /* 3. Tuning rcvbuf, when connection enters established state. */
333 static void tcp_fixup_rcvbuf(struct sock
*sk
)
335 struct tcp_sock
*tp
= tcp_sk(sk
);
336 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
338 /* Try to select rcvbuf so that 4 mss-sized segments
339 * will fit to window and corresponding skbs will fit to our rcvbuf.
340 * (was 3; 4 is minimum to allow fast retransmit to work.)
342 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
344 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
345 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
348 /* 4. Try to fixup all. It is made immediately after connection enters
351 static void tcp_init_buffer_space(struct sock
*sk
)
353 struct tcp_sock
*tp
= tcp_sk(sk
);
356 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
357 tcp_fixup_rcvbuf(sk
);
358 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
359 tcp_fixup_sndbuf(sk
);
361 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
363 maxwin
= tcp_full_space(sk
);
365 if (tp
->window_clamp
>= maxwin
) {
366 tp
->window_clamp
= maxwin
;
368 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
369 tp
->window_clamp
= max(maxwin
-
370 (maxwin
>> sysctl_tcp_app_win
),
374 /* Force reservation of one segment. */
375 if (sysctl_tcp_app_win
&&
376 tp
->window_clamp
> 2 * tp
->advmss
&&
377 tp
->window_clamp
+ tp
->advmss
> maxwin
)
378 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
380 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
381 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock
*sk
)
387 struct tcp_sock
*tp
= tcp_sk(sk
);
388 struct inet_connection_sock
*icsk
= inet_csk(sk
);
390 icsk
->icsk_ack
.quick
= 0;
392 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
393 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
394 !tcp_memory_pressure
&&
395 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
396 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
399 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
400 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
404 /* Initialize RCV_MSS value.
405 * RCV_MSS is an our guess about MSS used by the peer.
406 * We haven't any direct information about the MSS.
407 * It's better to underestimate the RCV_MSS rather than overestimate.
408 * Overestimations make us ACKing less frequently than needed.
409 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
411 void tcp_initialize_rcv_mss(struct sock
*sk
)
413 struct tcp_sock
*tp
= tcp_sk(sk
);
414 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
416 hint
= min(hint
, tp
->rcv_wnd
/2);
417 hint
= min(hint
, TCP_MIN_RCVMSS
);
418 hint
= max(hint
, TCP_MIN_MSS
);
420 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
423 /* Receiver "autotuning" code.
425 * The algorithm for RTT estimation w/o timestamps is based on
426 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
427 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
429 * More detail on this code can be found at
430 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
431 * though this reference is out of date. A new paper
434 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
436 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
442 if (new_sample
!= 0) {
443 /* If we sample in larger samples in the non-timestamp
444 * case, we could grossly overestimate the RTT especially
445 * with chatty applications or bulk transfer apps which
446 * are stalled on filesystem I/O.
448 * Also, since we are only going for a minimum in the
449 * non-timestamp case, we do not smooth things out
450 * else with timestamps disabled convergence takes too
454 m
-= (new_sample
>> 3);
456 } else if (m
< new_sample
)
459 /* No previous measure. */
463 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
464 tp
->rcv_rtt_est
.rtt
= new_sample
;
467 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
469 if (tp
->rcv_rtt_est
.time
== 0)
471 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
473 tcp_rcv_rtt_update(tp
,
474 jiffies
- tp
->rcv_rtt_est
.time
,
478 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
479 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
482 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
484 struct tcp_sock
*tp
= tcp_sk(sk
);
485 if (tp
->rx_opt
.rcv_tsecr
&&
486 (TCP_SKB_CB(skb
)->end_seq
-
487 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
488 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
492 * This function should be called every time data is copied to user space.
493 * It calculates the appropriate TCP receive buffer space.
495 void tcp_rcv_space_adjust(struct sock
*sk
)
497 struct tcp_sock
*tp
= tcp_sk(sk
);
501 if (tp
->rcvq_space
.time
== 0)
504 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
505 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
506 tp
->rcv_rtt_est
.rtt
== 0)
509 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
511 space
= max(tp
->rcvq_space
.space
, space
);
513 if (tp
->rcvq_space
.space
!= space
) {
516 tp
->rcvq_space
.space
= space
;
518 if (sysctl_tcp_moderate_rcvbuf
&&
519 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
520 int new_clamp
= space
;
522 /* Receive space grows, normalize in order to
523 * take into account packet headers and sk_buff
524 * structure overhead.
529 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
530 16 + sizeof(struct sk_buff
));
531 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
534 space
= min(space
, sysctl_tcp_rmem
[2]);
535 if (space
> sk
->sk_rcvbuf
) {
536 sk
->sk_rcvbuf
= space
;
538 /* Make the window clamp follow along. */
539 tp
->window_clamp
= new_clamp
;
545 tp
->rcvq_space
.seq
= tp
->copied_seq
;
546 tp
->rcvq_space
.time
= tcp_time_stamp
;
549 /* There is something which you must keep in mind when you analyze the
550 * behavior of the tp->ato delayed ack timeout interval. When a
551 * connection starts up, we want to ack as quickly as possible. The
552 * problem is that "good" TCP's do slow start at the beginning of data
553 * transmission. The means that until we send the first few ACK's the
554 * sender will sit on his end and only queue most of his data, because
555 * he can only send snd_cwnd unacked packets at any given time. For
556 * each ACK we send, he increments snd_cwnd and transmits more of his
559 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
561 struct tcp_sock
*tp
= tcp_sk(sk
);
562 struct inet_connection_sock
*icsk
= inet_csk(sk
);
565 inet_csk_schedule_ack(sk
);
567 tcp_measure_rcv_mss(sk
, skb
);
569 tcp_rcv_rtt_measure(tp
);
571 now
= tcp_time_stamp
;
573 if (!icsk
->icsk_ack
.ato
) {
574 /* The _first_ data packet received, initialize
575 * delayed ACK engine.
577 tcp_incr_quickack(sk
);
578 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
580 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
582 if (m
<= TCP_ATO_MIN
/2) {
583 /* The fastest case is the first. */
584 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
585 } else if (m
< icsk
->icsk_ack
.ato
) {
586 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
587 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
588 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
589 } else if (m
> icsk
->icsk_rto
) {
590 /* Too long gap. Apparently sender failed to
591 * restart window, so that we send ACKs quickly.
593 tcp_incr_quickack(sk
);
594 sk_stream_mem_reclaim(sk
);
597 icsk
->icsk_ack
.lrcvtime
= now
;
599 TCP_ECN_check_ce(tp
, skb
);
602 tcp_grow_window(sk
, skb
);
605 static u32
tcp_rto_min(struct sock
*sk
)
607 struct dst_entry
*dst
= __sk_dst_get(sk
);
608 u32 rto_min
= TCP_RTO_MIN
;
610 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
611 rto_min
= dst
->metrics
[RTAX_RTO_MIN
-1];
615 /* Called to compute a smoothed rtt estimate. The data fed to this
616 * routine either comes from timestamps, or from segments that were
617 * known _not_ to have been retransmitted [see Karn/Partridge
618 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
619 * piece by Van Jacobson.
620 * NOTE: the next three routines used to be one big routine.
621 * To save cycles in the RFC 1323 implementation it was better to break
622 * it up into three procedures. -- erics
624 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
626 struct tcp_sock
*tp
= tcp_sk(sk
);
627 long m
= mrtt
; /* RTT */
629 /* The following amusing code comes from Jacobson's
630 * article in SIGCOMM '88. Note that rtt and mdev
631 * are scaled versions of rtt and mean deviation.
632 * This is designed to be as fast as possible
633 * m stands for "measurement".
635 * On a 1990 paper the rto value is changed to:
636 * RTO = rtt + 4 * mdev
638 * Funny. This algorithm seems to be very broken.
639 * These formulae increase RTO, when it should be decreased, increase
640 * too slowly, when it should be increased quickly, decrease too quickly
641 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
642 * does not matter how to _calculate_ it. Seems, it was trap
643 * that VJ failed to avoid. 8)
648 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
649 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
651 m
= -m
; /* m is now abs(error) */
652 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
653 /* This is similar to one of Eifel findings.
654 * Eifel blocks mdev updates when rtt decreases.
655 * This solution is a bit different: we use finer gain
656 * for mdev in this case (alpha*beta).
657 * Like Eifel it also prevents growth of rto,
658 * but also it limits too fast rto decreases,
659 * happening in pure Eifel.
664 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
666 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
667 if (tp
->mdev
> tp
->mdev_max
) {
668 tp
->mdev_max
= tp
->mdev
;
669 if (tp
->mdev_max
> tp
->rttvar
)
670 tp
->rttvar
= tp
->mdev_max
;
672 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
673 if (tp
->mdev_max
< tp
->rttvar
)
674 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
675 tp
->rtt_seq
= tp
->snd_nxt
;
676 tp
->mdev_max
= tcp_rto_min(sk
);
679 /* no previous measure. */
680 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
681 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
682 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
683 tp
->rtt_seq
= tp
->snd_nxt
;
687 /* Calculate rto without backoff. This is the second half of Van Jacobson's
688 * routine referred to above.
690 static inline void tcp_set_rto(struct sock
*sk
)
692 const struct tcp_sock
*tp
= tcp_sk(sk
);
693 /* Old crap is replaced with new one. 8)
696 * 1. If rtt variance happened to be less 50msec, it is hallucination.
697 * It cannot be less due to utterly erratic ACK generation made
698 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
699 * to do with delayed acks, because at cwnd>2 true delack timeout
700 * is invisible. Actually, Linux-2.4 also generates erratic
701 * ACKs in some circumstances.
703 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
705 /* 2. Fixups made earlier cannot be right.
706 * If we do not estimate RTO correctly without them,
707 * all the algo is pure shit and should be replaced
708 * with correct one. It is exactly, which we pretend to do.
712 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
713 * guarantees that rto is higher.
715 static inline void tcp_bound_rto(struct sock
*sk
)
717 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
718 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
721 /* Save metrics learned by this TCP session.
722 This function is called only, when TCP finishes successfully
723 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
725 void tcp_update_metrics(struct sock
*sk
)
727 struct tcp_sock
*tp
= tcp_sk(sk
);
728 struct dst_entry
*dst
= __sk_dst_get(sk
);
730 if (sysctl_tcp_nometrics_save
)
735 if (dst
&& (dst
->flags
&DST_HOST
)) {
736 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
739 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
740 /* This session failed to estimate rtt. Why?
741 * Probably, no packets returned in time.
744 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
745 dst
->metrics
[RTAX_RTT
-1] = 0;
749 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
751 /* If newly calculated rtt larger than stored one,
752 * store new one. Otherwise, use EWMA. Remember,
753 * rtt overestimation is always better than underestimation.
755 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
757 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
759 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
762 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
766 /* Scale deviation to rttvar fixed point */
771 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
772 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
774 dst
->metrics
[RTAX_RTTVAR
-1] -=
775 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
778 if (tp
->snd_ssthresh
>= 0xFFFF) {
779 /* Slow start still did not finish. */
780 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
781 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
782 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
783 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
784 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
785 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
786 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
787 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
788 icsk
->icsk_ca_state
== TCP_CA_Open
) {
789 /* Cong. avoidance phase, cwnd is reliable. */
790 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
791 dst
->metrics
[RTAX_SSTHRESH
-1] =
792 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
793 if (!dst_metric_locked(dst
, RTAX_CWND
))
794 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
796 /* Else slow start did not finish, cwnd is non-sense,
797 ssthresh may be also invalid.
799 if (!dst_metric_locked(dst
, RTAX_CWND
))
800 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
801 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
802 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
803 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
804 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
807 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
808 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
809 tp
->reordering
!= sysctl_tcp_reordering
)
810 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
815 /* Numbers are taken from RFC3390.
817 * John Heffner states:
819 * The RFC specifies a window of no more than 4380 bytes
820 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
821 * is a bit misleading because they use a clamp at 4380 bytes
822 * rather than use a multiplier in the relevant range.
824 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
826 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
829 if (tp
->mss_cache
> 1460)
832 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
834 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
837 /* Set slow start threshold and cwnd not falling to slow start */
838 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
840 struct tcp_sock
*tp
= tcp_sk(sk
);
841 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
843 tp
->prior_ssthresh
= 0;
845 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
848 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
849 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
850 tcp_packets_in_flight(tp
) + 1U);
851 tp
->snd_cwnd_cnt
= 0;
852 tp
->high_seq
= tp
->snd_nxt
;
853 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
854 TCP_ECN_queue_cwr(tp
);
856 tcp_set_ca_state(sk
, TCP_CA_CWR
);
861 * Packet counting of FACK is based on in-order assumptions, therefore TCP
862 * disables it when reordering is detected
864 static void tcp_disable_fack(struct tcp_sock
*tp
)
866 tp
->rx_opt
.sack_ok
&= ~2;
869 /* Take a notice that peer is sending DSACKs */
870 static void tcp_dsack_seen(struct tcp_sock
*tp
)
872 tp
->rx_opt
.sack_ok
|= 4;
875 /* Initialize metrics on socket. */
877 static void tcp_init_metrics(struct sock
*sk
)
879 struct tcp_sock
*tp
= tcp_sk(sk
);
880 struct dst_entry
*dst
= __sk_dst_get(sk
);
887 if (dst_metric_locked(dst
, RTAX_CWND
))
888 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
889 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
890 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
891 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
892 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
894 if (dst_metric(dst
, RTAX_REORDERING
) &&
895 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
896 tcp_disable_fack(tp
);
897 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
900 if (dst_metric(dst
, RTAX_RTT
) == 0)
903 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
906 /* Initial rtt is determined from SYN,SYN-ACK.
907 * The segment is small and rtt may appear much
908 * less than real one. Use per-dst memory
909 * to make it more realistic.
911 * A bit of theory. RTT is time passed after "normal" sized packet
912 * is sent until it is ACKed. In normal circumstances sending small
913 * packets force peer to delay ACKs and calculation is correct too.
914 * The algorithm is adaptive and, provided we follow specs, it
915 * NEVER underestimate RTT. BUT! If peer tries to make some clever
916 * tricks sort of "quick acks" for time long enough to decrease RTT
917 * to low value, and then abruptly stops to do it and starts to delay
918 * ACKs, wait for troubles.
920 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
921 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
922 tp
->rtt_seq
= tp
->snd_nxt
;
924 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
925 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
926 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, TCP_RTO_MIN
);
930 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
932 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
933 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
937 /* Play conservative. If timestamps are not
938 * supported, TCP will fail to recalculate correct
939 * rtt, if initial rto is too small. FORGET ALL AND RESET!
941 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
943 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
944 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
948 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
951 struct tcp_sock
*tp
= tcp_sk(sk
);
952 if (metric
> tp
->reordering
) {
953 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
955 /* This exciting event is worth to be remembered. 8) */
957 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
958 else if (tcp_is_reno(tp
))
959 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
960 else if (tcp_is_fack(tp
))
961 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
963 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
964 #if FASTRETRANS_DEBUG > 1
965 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
966 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
970 tp
->undo_marker
? tp
->undo_retrans
: 0);
972 tcp_disable_fack(tp
);
976 /* This procedure tags the retransmission queue when SACKs arrive.
978 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
979 * Packets in queue with these bits set are counted in variables
980 * sacked_out, retrans_out and lost_out, correspondingly.
982 * Valid combinations are:
983 * Tag InFlight Description
984 * 0 1 - orig segment is in flight.
985 * S 0 - nothing flies, orig reached receiver.
986 * L 0 - nothing flies, orig lost by net.
987 * R 2 - both orig and retransmit are in flight.
988 * L|R 1 - orig is lost, retransmit is in flight.
989 * S|R 1 - orig reached receiver, retrans is still in flight.
990 * (L|S|R is logically valid, it could occur when L|R is sacked,
991 * but it is equivalent to plain S and code short-curcuits it to S.
992 * L|S is logically invalid, it would mean -1 packet in flight 8))
994 * These 6 states form finite state machine, controlled by the following events:
995 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
996 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
997 * 3. Loss detection event of one of three flavors:
998 * A. Scoreboard estimator decided the packet is lost.
999 * A'. Reno "three dupacks" marks head of queue lost.
1000 * A''. Its FACK modfication, head until snd.fack is lost.
1001 * B. SACK arrives sacking data transmitted after never retransmitted
1002 * hole was sent out.
1003 * C. SACK arrives sacking SND.NXT at the moment, when the
1004 * segment was retransmitted.
1005 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1007 * It is pleasant to note, that state diagram turns out to be commutative,
1008 * so that we are allowed not to be bothered by order of our actions,
1009 * when multiple events arrive simultaneously. (see the function below).
1011 * Reordering detection.
1012 * --------------------
1013 * Reordering metric is maximal distance, which a packet can be displaced
1014 * in packet stream. With SACKs we can estimate it:
1016 * 1. SACK fills old hole and the corresponding segment was not
1017 * ever retransmitted -> reordering. Alas, we cannot use it
1018 * when segment was retransmitted.
1019 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1020 * for retransmitted and already SACKed segment -> reordering..
1021 * Both of these heuristics are not used in Loss state, when we cannot
1022 * account for retransmits accurately.
1024 * SACK block validation.
1025 * ----------------------
1027 * SACK block range validation checks that the received SACK block fits to
1028 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1029 * Note that SND.UNA is not included to the range though being valid because
1030 * it means that the receiver is rather inconsistent with itself (reports
1031 * SACK reneging when it should advance SND.UNA).
1033 * Implements also blockage to start_seq wrap-around. Problem lies in the
1034 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1035 * there's no guarantee that it will be before snd_nxt (n). The problem
1036 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1039 * <- outs wnd -> <- wrapzone ->
1040 * u e n u_w e_w s n_w
1042 * |<------------+------+----- TCP seqno space --------------+---------->|
1043 * ...-- <2^31 ->| |<--------...
1044 * ...---- >2^31 ------>| |<--------...
1046 * Current code wouldn't be vulnerable but it's better still to discard such
1047 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1048 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1049 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1050 * equal to the ideal case (infinite seqno space without wrap caused issues).
1052 * With D-SACK the lower bound is extended to cover sequence space below
1053 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1054 * again, DSACK block must not to go across snd_una (for the same reason as
1055 * for the normal SACK blocks, explained above). But there all simplicity
1056 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1057 * fully below undo_marker they do not affect behavior in anyway and can
1058 * therefore be safely ignored. In rare cases (which are more or less
1059 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1060 * fragmentation and packet reordering past skb's retransmission. To consider
1061 * them correctly, the acceptable range must be extended even more though
1062 * the exact amount is rather hard to quantify. However, tp->max_window can
1063 * be used as an exaggerated estimate.
1065 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1066 u32 start_seq
, u32 end_seq
)
1068 /* Too far in future, or reversed (interpretation is ambiguous) */
1069 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1072 /* Nasty start_seq wrap-around check (see comments above) */
1073 if (!before(start_seq
, tp
->snd_nxt
))
1076 /* In outstanding window? ...This is valid exit for DSACKs too.
1077 * start_seq == snd_una is non-sensical (see comments above)
1079 if (after(start_seq
, tp
->snd_una
))
1082 if (!is_dsack
|| !tp
->undo_marker
)
1085 /* ...Then it's D-SACK, and must reside below snd_una completely */
1086 if (!after(end_seq
, tp
->snd_una
))
1089 if (!before(start_seq
, tp
->undo_marker
))
1093 if (!after(end_seq
, tp
->undo_marker
))
1096 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1097 * start_seq < undo_marker and end_seq >= undo_marker.
1099 return !before(start_seq
, end_seq
- tp
->max_window
);
1103 static int tcp_check_dsack(struct tcp_sock
*tp
, struct sk_buff
*ack_skb
,
1104 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1107 u32 start_seq_0
= ntohl(get_unaligned(&sp
[0].start_seq
));
1108 u32 end_seq_0
= ntohl(get_unaligned(&sp
[0].end_seq
));
1111 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1114 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
1115 } else if (num_sacks
> 1) {
1116 u32 end_seq_1
= ntohl(get_unaligned(&sp
[1].end_seq
));
1117 u32 start_seq_1
= ntohl(get_unaligned(&sp
[1].start_seq
));
1119 if (!after(end_seq_0
, end_seq_1
) &&
1120 !before(start_seq_0
, start_seq_1
)) {
1123 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
1127 /* D-SACK for already forgotten data... Do dumb counting. */
1129 !after(end_seq_0
, prior_snd_una
) &&
1130 after(end_seq_0
, tp
->undo_marker
))
1137 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
1139 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1140 struct tcp_sock
*tp
= tcp_sk(sk
);
1141 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1142 TCP_SKB_CB(ack_skb
)->sacked
);
1143 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
1144 struct sk_buff
*cached_skb
;
1145 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
1146 int reord
= tp
->packets_out
;
1148 u32 lost_retrans
= 0;
1150 int found_dup_sack
= 0;
1151 int cached_fack_count
;
1153 int first_sack_index
;
1155 if (!tp
->sacked_out
) {
1156 tp
->fackets_out
= 0;
1157 tp
->highest_sack
= tp
->snd_una
;
1159 prior_fackets
= tp
->fackets_out
;
1161 found_dup_sack
= tcp_check_dsack(tp
, ack_skb
, sp
,
1162 num_sacks
, prior_snd_una
);
1164 flag
|= FLAG_DSACKING_ACK
;
1166 /* Eliminate too old ACKs, but take into
1167 * account more or less fresh ones, they can
1168 * contain valid SACK info.
1170 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1174 * if the only SACK change is the increase of the end_seq of
1175 * the first block then only apply that SACK block
1176 * and use retrans queue hinting otherwise slowpath */
1178 for (i
= 0; i
< num_sacks
; i
++) {
1179 __be32 start_seq
= sp
[i
].start_seq
;
1180 __be32 end_seq
= sp
[i
].end_seq
;
1183 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
1186 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
1187 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1190 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1191 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1193 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1194 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1195 tp
->recv_sack_cache
[i
].start_seq
= 0;
1196 tp
->recv_sack_cache
[i
].end_seq
= 0;
1199 first_sack_index
= 0;
1204 tp
->fastpath_skb_hint
= NULL
;
1206 /* order SACK blocks to allow in order walk of the retrans queue */
1207 for (i
= num_sacks
-1; i
> 0; i
--) {
1208 for (j
= 0; j
< i
; j
++){
1209 if (after(ntohl(sp
[j
].start_seq
),
1210 ntohl(sp
[j
+1].start_seq
))){
1211 struct tcp_sack_block_wire tmp
;
1217 /* Track where the first SACK block goes to */
1218 if (j
== first_sack_index
)
1219 first_sack_index
= j
+1;
1226 /* clear flag as used for different purpose in following code */
1229 /* Use SACK fastpath hint if valid */
1230 cached_skb
= tp
->fastpath_skb_hint
;
1231 cached_fack_count
= tp
->fastpath_cnt_hint
;
1233 cached_skb
= tcp_write_queue_head(sk
);
1234 cached_fack_count
= 0;
1237 for (i
=0; i
<num_sacks
; i
++, sp
++) {
1238 struct sk_buff
*skb
;
1239 __u32 start_seq
= ntohl(sp
->start_seq
);
1240 __u32 end_seq
= ntohl(sp
->end_seq
);
1242 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1244 if (!tcp_is_sackblock_valid(tp
, dup_sack
, start_seq
, end_seq
)) {
1246 if (!tp
->undo_marker
)
1247 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO
);
1249 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD
);
1251 /* Don't count olds caused by ACK reordering */
1252 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1253 !after(end_seq
, tp
->snd_una
))
1255 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD
);
1261 fack_count
= cached_fack_count
;
1263 /* Event "B" in the comment above. */
1264 if (after(end_seq
, tp
->high_seq
))
1265 flag
|= FLAG_DATA_LOST
;
1267 tcp_for_write_queue_from(skb
, sk
) {
1268 int in_sack
, pcount
;
1271 if (skb
== tcp_send_head(sk
))
1275 cached_fack_count
= fack_count
;
1276 if (i
== first_sack_index
) {
1277 tp
->fastpath_skb_hint
= skb
;
1278 tp
->fastpath_cnt_hint
= fack_count
;
1281 /* The retransmission queue is always in order, so
1282 * we can short-circuit the walk early.
1284 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1287 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1288 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1290 pcount
= tcp_skb_pcount(skb
);
1292 if (pcount
> 1 && !in_sack
&&
1293 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1294 unsigned int pkt_len
;
1296 in_sack
= !after(start_seq
,
1297 TCP_SKB_CB(skb
)->seq
);
1300 pkt_len
= (start_seq
-
1301 TCP_SKB_CB(skb
)->seq
);
1303 pkt_len
= (end_seq
-
1304 TCP_SKB_CB(skb
)->seq
);
1305 if (tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
))
1307 pcount
= tcp_skb_pcount(skb
);
1310 fack_count
+= pcount
;
1312 sacked
= TCP_SKB_CB(skb
)->sacked
;
1314 /* Account D-SACK for retransmitted packet. */
1315 if ((dup_sack
&& in_sack
) &&
1316 (sacked
& TCPCB_RETRANS
) &&
1317 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1320 /* The frame is ACKed. */
1321 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1322 if (sacked
&TCPCB_RETRANS
) {
1323 if ((dup_sack
&& in_sack
) &&
1324 (sacked
&TCPCB_SACKED_ACKED
))
1325 reord
= min(fack_count
, reord
);
1327 /* If it was in a hole, we detected reordering. */
1328 if (fack_count
< prior_fackets
&&
1329 !(sacked
&TCPCB_SACKED_ACKED
))
1330 reord
= min(fack_count
, reord
);
1333 /* Nothing to do; acked frame is about to be dropped. */
1337 if ((sacked
&TCPCB_SACKED_RETRANS
) &&
1338 after(end_seq
, TCP_SKB_CB(skb
)->ack_seq
) &&
1339 (!lost_retrans
|| after(end_seq
, lost_retrans
)))
1340 lost_retrans
= end_seq
;
1345 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1346 if (sacked
& TCPCB_SACKED_RETRANS
) {
1347 /* If the segment is not tagged as lost,
1348 * we do not clear RETRANS, believing
1349 * that retransmission is still in flight.
1351 if (sacked
& TCPCB_LOST
) {
1352 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1353 tp
->lost_out
-= tcp_skb_pcount(skb
);
1354 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1356 /* clear lost hint */
1357 tp
->retransmit_skb_hint
= NULL
;
1360 /* New sack for not retransmitted frame,
1361 * which was in hole. It is reordering.
1363 if (!(sacked
& TCPCB_RETRANS
) &&
1364 fack_count
< prior_fackets
)
1365 reord
= min(fack_count
, reord
);
1367 if (sacked
& TCPCB_LOST
) {
1368 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1369 tp
->lost_out
-= tcp_skb_pcount(skb
);
1371 /* clear lost hint */
1372 tp
->retransmit_skb_hint
= NULL
;
1374 /* SACK enhanced F-RTO detection.
1375 * Set flag if and only if non-rexmitted
1376 * segments below frto_highmark are
1377 * SACKed (RFC4138; Appendix B).
1378 * Clearing correct due to in-order walk
1380 if (after(end_seq
, tp
->frto_highmark
)) {
1381 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1383 if (!(sacked
& TCPCB_RETRANS
))
1384 flag
|= FLAG_ONLY_ORIG_SACKED
;
1388 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1389 flag
|= FLAG_DATA_SACKED
;
1390 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1392 if (fack_count
> tp
->fackets_out
)
1393 tp
->fackets_out
= fack_count
;
1395 if (after(TCP_SKB_CB(skb
)->seq
,
1397 tp
->highest_sack
= TCP_SKB_CB(skb
)->seq
;
1399 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1400 reord
= min(fack_count
, reord
);
1403 /* D-SACK. We can detect redundant retransmission
1404 * in S|R and plain R frames and clear it.
1405 * undo_retrans is decreased above, L|R frames
1406 * are accounted above as well.
1409 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1410 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1411 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1412 tp
->retransmit_skb_hint
= NULL
;
1417 /* Check for lost retransmit. This superb idea is
1418 * borrowed from "ratehalving". Event "C".
1419 * Later note: FACK people cheated me again 8),
1420 * we have to account for reordering! Ugly,
1423 if (lost_retrans
&& icsk
->icsk_ca_state
== TCP_CA_Recovery
) {
1424 struct sk_buff
*skb
;
1426 tcp_for_write_queue(skb
, sk
) {
1427 if (skb
== tcp_send_head(sk
))
1429 if (after(TCP_SKB_CB(skb
)->seq
, lost_retrans
))
1431 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1433 if ((TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
) &&
1434 after(lost_retrans
, TCP_SKB_CB(skb
)->ack_seq
) &&
1436 !before(lost_retrans
,
1437 TCP_SKB_CB(skb
)->ack_seq
+ tp
->reordering
*
1439 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1440 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1442 /* clear lost hint */
1443 tp
->retransmit_skb_hint
= NULL
;
1445 if (!(TCP_SKB_CB(skb
)->sacked
&(TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1446 tp
->lost_out
+= tcp_skb_pcount(skb
);
1447 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1448 flag
|= FLAG_DATA_SACKED
;
1449 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1455 tcp_verify_left_out(tp
);
1457 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
&&
1458 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1459 tcp_update_reordering(sk
, ((tp
->fackets_out
+ 1) - reord
), 0);
1461 #if FASTRETRANS_DEBUG > 0
1462 BUG_TRAP((int)tp
->sacked_out
>= 0);
1463 BUG_TRAP((int)tp
->lost_out
>= 0);
1464 BUG_TRAP((int)tp
->retrans_out
>= 0);
1465 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1470 /* If we receive more dupacks than we expected counting segments
1471 * in assumption of absent reordering, interpret this as reordering.
1472 * The only another reason could be bug in receiver TCP.
1474 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1476 struct tcp_sock
*tp
= tcp_sk(sk
);
1479 holes
= max(tp
->lost_out
, 1U);
1480 holes
= min(holes
, tp
->packets_out
);
1482 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1483 tp
->sacked_out
= tp
->packets_out
- holes
;
1484 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1488 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1490 static void tcp_add_reno_sack(struct sock
*sk
)
1492 struct tcp_sock
*tp
= tcp_sk(sk
);
1494 tcp_check_reno_reordering(sk
, 0);
1495 tcp_verify_left_out(tp
);
1498 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1500 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1502 struct tcp_sock
*tp
= tcp_sk(sk
);
1505 /* One ACK acked hole. The rest eat duplicate ACKs. */
1506 if (acked
-1 >= tp
->sacked_out
)
1509 tp
->sacked_out
-= acked
-1;
1511 tcp_check_reno_reordering(sk
, acked
);
1512 tcp_verify_left_out(tp
);
1515 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1520 /* F-RTO can only be used if TCP has never retransmitted anything other than
1521 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1523 int tcp_use_frto(struct sock
*sk
)
1525 const struct tcp_sock
*tp
= tcp_sk(sk
);
1526 struct sk_buff
*skb
;
1528 if (!sysctl_tcp_frto
)
1534 /* Avoid expensive walking of rexmit queue if possible */
1535 if (tp
->retrans_out
> 1)
1538 skb
= tcp_write_queue_head(sk
);
1539 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1540 tcp_for_write_queue_from(skb
, sk
) {
1541 if (skb
== tcp_send_head(sk
))
1543 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1545 /* Short-circuit when first non-SACKed skb has been checked */
1546 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1552 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1553 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1554 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1555 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1556 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1557 * bits are handled if the Loss state is really to be entered (in
1558 * tcp_enter_frto_loss).
1560 * Do like tcp_enter_loss() would; when RTO expires the second time it
1562 * "Reduce ssthresh if it has not yet been made inside this window."
1564 void tcp_enter_frto(struct sock
*sk
)
1566 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1567 struct tcp_sock
*tp
= tcp_sk(sk
);
1568 struct sk_buff
*skb
;
1570 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1571 tp
->snd_una
== tp
->high_seq
||
1572 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1573 !icsk
->icsk_retransmits
)) {
1574 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1575 /* Our state is too optimistic in ssthresh() call because cwnd
1576 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1577 * recovery has not yet completed. Pattern would be this: RTO,
1578 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1580 * RFC4138 should be more specific on what to do, even though
1581 * RTO is quite unlikely to occur after the first Cumulative ACK
1582 * due to back-off and complexity of triggering events ...
1584 if (tp
->frto_counter
) {
1586 stored_cwnd
= tp
->snd_cwnd
;
1588 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1589 tp
->snd_cwnd
= stored_cwnd
;
1591 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1593 /* ... in theory, cong.control module could do "any tricks" in
1594 * ssthresh(), which means that ca_state, lost bits and lost_out
1595 * counter would have to be faked before the call occurs. We
1596 * consider that too expensive, unlikely and hacky, so modules
1597 * using these in ssthresh() must deal these incompatibility
1598 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1600 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1603 tp
->undo_marker
= tp
->snd_una
;
1604 tp
->undo_retrans
= 0;
1606 skb
= tcp_write_queue_head(sk
);
1607 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1608 tp
->undo_marker
= 0;
1609 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1610 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1611 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1613 tcp_verify_left_out(tp
);
1615 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1616 * The last condition is necessary at least in tp->frto_counter case.
1618 if (IsSackFrto() && (tp
->frto_counter
||
1619 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1620 after(tp
->high_seq
, tp
->snd_una
)) {
1621 tp
->frto_highmark
= tp
->high_seq
;
1623 tp
->frto_highmark
= tp
->snd_nxt
;
1625 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1626 tp
->high_seq
= tp
->snd_nxt
;
1627 tp
->frto_counter
= 1;
1630 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1631 * which indicates that we should follow the traditional RTO recovery,
1632 * i.e. mark everything lost and do go-back-N retransmission.
1634 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1636 struct tcp_sock
*tp
= tcp_sk(sk
);
1637 struct sk_buff
*skb
;
1640 tp
->retrans_out
= 0;
1641 if (tcp_is_reno(tp
))
1642 tcp_reset_reno_sack(tp
);
1644 tcp_for_write_queue(skb
, sk
) {
1645 if (skb
== tcp_send_head(sk
))
1648 * Count the retransmission made on RTO correctly (only when
1649 * waiting for the first ACK and did not get it)...
1651 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1652 /* For some reason this R-bit might get cleared? */
1653 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1654 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1655 /* ...enter this if branch just for the first segment */
1656 flag
|= FLAG_DATA_ACKED
;
1658 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1659 tp
->undo_marker
= 0;
1660 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1663 /* Don't lost mark skbs that were fwd transmitted after RTO */
1664 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) &&
1665 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
)) {
1666 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1667 tp
->lost_out
+= tcp_skb_pcount(skb
);
1670 tcp_verify_left_out(tp
);
1672 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1673 tp
->snd_cwnd_cnt
= 0;
1674 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1675 tp
->frto_counter
= 0;
1677 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1678 sysctl_tcp_reordering
);
1679 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1680 tp
->high_seq
= tp
->frto_highmark
;
1681 TCP_ECN_queue_cwr(tp
);
1683 tcp_clear_retrans_hints_partial(tp
);
1686 void tcp_clear_retrans(struct tcp_sock
*tp
)
1688 tp
->retrans_out
= 0;
1690 tp
->fackets_out
= 0;
1694 tp
->undo_marker
= 0;
1695 tp
->undo_retrans
= 0;
1698 /* Enter Loss state. If "how" is not zero, forget all SACK information
1699 * and reset tags completely, otherwise preserve SACKs. If receiver
1700 * dropped its ofo queue, we will know this due to reneging detection.
1702 void tcp_enter_loss(struct sock
*sk
, int how
)
1704 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1705 struct tcp_sock
*tp
= tcp_sk(sk
);
1706 struct sk_buff
*skb
;
1709 /* Reduce ssthresh if it has not yet been made inside this window. */
1710 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1711 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1712 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1713 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1714 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1717 tp
->snd_cwnd_cnt
= 0;
1718 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1720 tp
->bytes_acked
= 0;
1721 tcp_clear_retrans(tp
);
1724 /* Push undo marker, if it was plain RTO and nothing
1725 * was retransmitted. */
1726 tp
->undo_marker
= tp
->snd_una
;
1727 tcp_clear_retrans_hints_partial(tp
);
1729 tcp_clear_all_retrans_hints(tp
);
1732 tcp_for_write_queue(skb
, sk
) {
1733 if (skb
== tcp_send_head(sk
))
1735 cnt
+= tcp_skb_pcount(skb
);
1736 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1737 tp
->undo_marker
= 0;
1738 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1739 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1740 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1741 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1742 tp
->lost_out
+= tcp_skb_pcount(skb
);
1744 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1745 tp
->fackets_out
= cnt
;
1748 tcp_verify_left_out(tp
);
1750 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1751 sysctl_tcp_reordering
);
1752 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1753 tp
->high_seq
= tp
->snd_nxt
;
1754 TCP_ECN_queue_cwr(tp
);
1755 /* Abort FRTO algorithm if one is in progress */
1756 tp
->frto_counter
= 0;
1759 static int tcp_check_sack_reneging(struct sock
*sk
)
1761 struct sk_buff
*skb
;
1763 /* If ACK arrived pointing to a remembered SACK,
1764 * it means that our remembered SACKs do not reflect
1765 * real state of receiver i.e.
1766 * receiver _host_ is heavily congested (or buggy).
1767 * Do processing similar to RTO timeout.
1769 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1770 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1771 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1772 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1774 tcp_enter_loss(sk
, 1);
1775 icsk
->icsk_retransmits
++;
1776 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1777 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1778 icsk
->icsk_rto
, TCP_RTO_MAX
);
1784 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1786 return tcp_is_reno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1789 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1791 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1794 static inline int tcp_head_timedout(struct sock
*sk
)
1796 struct tcp_sock
*tp
= tcp_sk(sk
);
1798 return tp
->packets_out
&&
1799 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1802 /* Linux NewReno/SACK/FACK/ECN state machine.
1803 * --------------------------------------
1805 * "Open" Normal state, no dubious events, fast path.
1806 * "Disorder" In all the respects it is "Open",
1807 * but requires a bit more attention. It is entered when
1808 * we see some SACKs or dupacks. It is split of "Open"
1809 * mainly to move some processing from fast path to slow one.
1810 * "CWR" CWND was reduced due to some Congestion Notification event.
1811 * It can be ECN, ICMP source quench, local device congestion.
1812 * "Recovery" CWND was reduced, we are fast-retransmitting.
1813 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1815 * tcp_fastretrans_alert() is entered:
1816 * - each incoming ACK, if state is not "Open"
1817 * - when arrived ACK is unusual, namely:
1822 * Counting packets in flight is pretty simple.
1824 * in_flight = packets_out - left_out + retrans_out
1826 * packets_out is SND.NXT-SND.UNA counted in packets.
1828 * retrans_out is number of retransmitted segments.
1830 * left_out is number of segments left network, but not ACKed yet.
1832 * left_out = sacked_out + lost_out
1834 * sacked_out: Packets, which arrived to receiver out of order
1835 * and hence not ACKed. With SACKs this number is simply
1836 * amount of SACKed data. Even without SACKs
1837 * it is easy to give pretty reliable estimate of this number,
1838 * counting duplicate ACKs.
1840 * lost_out: Packets lost by network. TCP has no explicit
1841 * "loss notification" feedback from network (for now).
1842 * It means that this number can be only _guessed_.
1843 * Actually, it is the heuristics to predict lossage that
1844 * distinguishes different algorithms.
1846 * F.e. after RTO, when all the queue is considered as lost,
1847 * lost_out = packets_out and in_flight = retrans_out.
1849 * Essentially, we have now two algorithms counting
1852 * FACK: It is the simplest heuristics. As soon as we decided
1853 * that something is lost, we decide that _all_ not SACKed
1854 * packets until the most forward SACK are lost. I.e.
1855 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1856 * It is absolutely correct estimate, if network does not reorder
1857 * packets. And it loses any connection to reality when reordering
1858 * takes place. We use FACK by default until reordering
1859 * is suspected on the path to this destination.
1861 * NewReno: when Recovery is entered, we assume that one segment
1862 * is lost (classic Reno). While we are in Recovery and
1863 * a partial ACK arrives, we assume that one more packet
1864 * is lost (NewReno). This heuristics are the same in NewReno
1867 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1868 * deflation etc. CWND is real congestion window, never inflated, changes
1869 * only according to classic VJ rules.
1871 * Really tricky (and requiring careful tuning) part of algorithm
1872 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1873 * The first determines the moment _when_ we should reduce CWND and,
1874 * hence, slow down forward transmission. In fact, it determines the moment
1875 * when we decide that hole is caused by loss, rather than by a reorder.
1877 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1878 * holes, caused by lost packets.
1880 * And the most logically complicated part of algorithm is undo
1881 * heuristics. We detect false retransmits due to both too early
1882 * fast retransmit (reordering) and underestimated RTO, analyzing
1883 * timestamps and D-SACKs. When we detect that some segments were
1884 * retransmitted by mistake and CWND reduction was wrong, we undo
1885 * window reduction and abort recovery phase. This logic is hidden
1886 * inside several functions named tcp_try_undo_<something>.
1889 /* This function decides, when we should leave Disordered state
1890 * and enter Recovery phase, reducing congestion window.
1892 * Main question: may we further continue forward transmission
1893 * with the same cwnd?
1895 static int tcp_time_to_recover(struct sock
*sk
)
1897 struct tcp_sock
*tp
= tcp_sk(sk
);
1900 /* Do not perform any recovery during FRTO algorithm */
1901 if (tp
->frto_counter
)
1904 /* Trick#1: The loss is proven. */
1908 /* Not-A-Trick#2 : Classic rule... */
1909 if (tcp_fackets_out(tp
) > tp
->reordering
)
1912 /* Trick#3 : when we use RFC2988 timer restart, fast
1913 * retransmit can be triggered by timeout of queue head.
1915 if (tcp_head_timedout(sk
))
1918 /* Trick#4: It is still not OK... But will it be useful to delay
1921 packets_out
= tp
->packets_out
;
1922 if (packets_out
<= tp
->reordering
&&
1923 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1924 !tcp_may_send_now(sk
)) {
1925 /* We have nothing to send. This connection is limited
1926 * either by receiver window or by application.
1934 /* RFC: This is from the original, I doubt that this is necessary at all:
1935 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1936 * retransmitted past LOST markings in the first place? I'm not fully sure
1937 * about undo and end of connection cases, which can cause R without L?
1939 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
1940 struct sk_buff
*skb
)
1942 if ((tp
->retransmit_skb_hint
!= NULL
) &&
1943 before(TCP_SKB_CB(skb
)->seq
,
1944 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
1945 tp
->retransmit_skb_hint
= NULL
;
1948 /* Mark head of queue up as lost. */
1949 static void tcp_mark_head_lost(struct sock
*sk
,
1950 int packets
, u32 high_seq
)
1952 struct tcp_sock
*tp
= tcp_sk(sk
);
1953 struct sk_buff
*skb
;
1956 BUG_TRAP(packets
<= tp
->packets_out
);
1957 if (tp
->lost_skb_hint
) {
1958 skb
= tp
->lost_skb_hint
;
1959 cnt
= tp
->lost_cnt_hint
;
1961 skb
= tcp_write_queue_head(sk
);
1965 tcp_for_write_queue_from(skb
, sk
) {
1966 if (skb
== tcp_send_head(sk
))
1968 /* TODO: do this better */
1969 /* this is not the most efficient way to do this... */
1970 tp
->lost_skb_hint
= skb
;
1971 tp
->lost_cnt_hint
= cnt
;
1972 cnt
+= tcp_skb_pcount(skb
);
1973 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, high_seq
))
1975 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
1976 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1977 tp
->lost_out
+= tcp_skb_pcount(skb
);
1978 tcp_verify_retransmit_hint(tp
, skb
);
1981 tcp_verify_left_out(tp
);
1984 /* Account newly detected lost packet(s) */
1986 static void tcp_update_scoreboard(struct sock
*sk
)
1988 struct tcp_sock
*tp
= tcp_sk(sk
);
1990 if (tcp_is_fack(tp
)) {
1991 int lost
= tp
->fackets_out
- tp
->reordering
;
1994 tcp_mark_head_lost(sk
, lost
, tp
->high_seq
);
1996 tcp_mark_head_lost(sk
, 1, tp
->high_seq
);
1999 /* New heuristics: it is possible only after we switched
2000 * to restart timer each time when something is ACKed.
2001 * Hence, we can detect timed out packets during fast
2002 * retransmit without falling to slow start.
2004 if (!tcp_is_reno(tp
) && tcp_head_timedout(sk
)) {
2005 struct sk_buff
*skb
;
2007 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2008 : tcp_write_queue_head(sk
);
2010 tcp_for_write_queue_from(skb
, sk
) {
2011 if (skb
== tcp_send_head(sk
))
2013 if (!tcp_skb_timedout(sk
, skb
))
2016 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_TAGBITS
)) {
2017 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2018 tp
->lost_out
+= tcp_skb_pcount(skb
);
2019 tcp_verify_retransmit_hint(tp
, skb
);
2023 tp
->scoreboard_skb_hint
= skb
;
2025 tcp_verify_left_out(tp
);
2029 /* CWND moderation, preventing bursts due to too big ACKs
2030 * in dubious situations.
2032 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2034 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2035 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
2036 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2039 /* Lower bound on congestion window is slow start threshold
2040 * unless congestion avoidance choice decides to overide it.
2042 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2044 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2046 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2049 /* Decrease cwnd each second ack. */
2050 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2052 struct tcp_sock
*tp
= tcp_sk(sk
);
2053 int decr
= tp
->snd_cwnd_cnt
+ 1;
2055 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
2056 (tcp_is_reno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
2057 tp
->snd_cwnd_cnt
= decr
&1;
2060 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2061 tp
->snd_cwnd
-= decr
;
2063 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
2064 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2068 /* Nothing was retransmitted or returned timestamp is less
2069 * than timestamp of the first retransmission.
2071 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2073 return !tp
->retrans_stamp
||
2074 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2075 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
2078 /* Undo procedures. */
2080 #if FASTRETRANS_DEBUG > 1
2081 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2083 struct tcp_sock
*tp
= tcp_sk(sk
);
2084 struct inet_sock
*inet
= inet_sk(sk
);
2086 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2088 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2089 tp
->snd_cwnd
, tcp_left_out(tp
),
2090 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2094 #define DBGUNDO(x...) do { } while (0)
2097 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2099 struct tcp_sock
*tp
= tcp_sk(sk
);
2101 if (tp
->prior_ssthresh
) {
2102 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2104 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2105 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2107 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
2109 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2110 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2111 TCP_ECN_withdraw_cwr(tp
);
2114 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2116 tcp_moderate_cwnd(tp
);
2117 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2119 /* There is something screwy going on with the retrans hints after
2121 tcp_clear_all_retrans_hints(tp
);
2124 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2126 return tp
->undo_marker
&&
2127 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2130 /* People celebrate: "We love our President!" */
2131 static int tcp_try_undo_recovery(struct sock
*sk
)
2133 struct tcp_sock
*tp
= tcp_sk(sk
);
2135 if (tcp_may_undo(tp
)) {
2136 /* Happy end! We did not retransmit anything
2137 * or our original transmission succeeded.
2139 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2140 tcp_undo_cwr(sk
, 1);
2141 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2142 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2144 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
2145 tp
->undo_marker
= 0;
2147 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2148 /* Hold old state until something *above* high_seq
2149 * is ACKed. For Reno it is MUST to prevent false
2150 * fast retransmits (RFC2582). SACK TCP is safe. */
2151 tcp_moderate_cwnd(tp
);
2154 tcp_set_ca_state(sk
, TCP_CA_Open
);
2158 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2159 static void tcp_try_undo_dsack(struct sock
*sk
)
2161 struct tcp_sock
*tp
= tcp_sk(sk
);
2163 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2164 DBGUNDO(sk
, "D-SACK");
2165 tcp_undo_cwr(sk
, 1);
2166 tp
->undo_marker
= 0;
2167 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2171 /* Undo during fast recovery after partial ACK. */
2173 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2175 struct tcp_sock
*tp
= tcp_sk(sk
);
2176 /* Partial ACK arrived. Force Hoe's retransmit. */
2177 int failed
= tcp_is_reno(tp
) || tp
->fackets_out
>tp
->reordering
;
2179 if (tcp_may_undo(tp
)) {
2180 /* Plain luck! Hole if filled with delayed
2181 * packet, rather than with a retransmit.
2183 if (tp
->retrans_out
== 0)
2184 tp
->retrans_stamp
= 0;
2186 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2189 tcp_undo_cwr(sk
, 0);
2190 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2192 /* So... Do not make Hoe's retransmit yet.
2193 * If the first packet was delayed, the rest
2194 * ones are most probably delayed as well.
2201 /* Undo during loss recovery after partial ACK. */
2202 static int tcp_try_undo_loss(struct sock
*sk
)
2204 struct tcp_sock
*tp
= tcp_sk(sk
);
2206 if (tcp_may_undo(tp
)) {
2207 struct sk_buff
*skb
;
2208 tcp_for_write_queue(skb
, sk
) {
2209 if (skb
== tcp_send_head(sk
))
2211 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2214 tcp_clear_all_retrans_hints(tp
);
2216 DBGUNDO(sk
, "partial loss");
2218 tcp_undo_cwr(sk
, 1);
2219 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2220 inet_csk(sk
)->icsk_retransmits
= 0;
2221 tp
->undo_marker
= 0;
2222 if (tcp_is_sack(tp
))
2223 tcp_set_ca_state(sk
, TCP_CA_Open
);
2229 static inline void tcp_complete_cwr(struct sock
*sk
)
2231 struct tcp_sock
*tp
= tcp_sk(sk
);
2232 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2233 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2234 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2237 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2239 struct tcp_sock
*tp
= tcp_sk(sk
);
2241 tcp_verify_left_out(tp
);
2243 if (tp
->retrans_out
== 0)
2244 tp
->retrans_stamp
= 0;
2247 tcp_enter_cwr(sk
, 1);
2249 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2250 int state
= TCP_CA_Open
;
2252 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2253 state
= TCP_CA_Disorder
;
2255 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2256 tcp_set_ca_state(sk
, state
);
2257 tp
->high_seq
= tp
->snd_nxt
;
2259 tcp_moderate_cwnd(tp
);
2261 tcp_cwnd_down(sk
, flag
);
2265 static void tcp_mtup_probe_failed(struct sock
*sk
)
2267 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2269 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2270 icsk
->icsk_mtup
.probe_size
= 0;
2273 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2275 struct tcp_sock
*tp
= tcp_sk(sk
);
2276 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2278 /* FIXME: breaks with very large cwnd */
2279 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2280 tp
->snd_cwnd
= tp
->snd_cwnd
*
2281 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2282 icsk
->icsk_mtup
.probe_size
;
2283 tp
->snd_cwnd_cnt
= 0;
2284 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2285 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2287 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2288 icsk
->icsk_mtup
.probe_size
= 0;
2289 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2293 /* Process an event, which can update packets-in-flight not trivially.
2294 * Main goal of this function is to calculate new estimate for left_out,
2295 * taking into account both packets sitting in receiver's buffer and
2296 * packets lost by network.
2298 * Besides that it does CWND reduction, when packet loss is detected
2299 * and changes state of machine.
2301 * It does _not_ decide what to send, it is made in function
2302 * tcp_xmit_retransmit_queue().
2305 tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2307 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2308 struct tcp_sock
*tp
= tcp_sk(sk
);
2309 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2310 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2311 (tp
->fackets_out
> tp
->reordering
));
2313 /* Some technical things:
2314 * 1. Reno does not count dupacks (sacked_out) automatically. */
2315 if (!tp
->packets_out
)
2318 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2319 tp
->fackets_out
= 0;
2321 /* Now state machine starts.
2322 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2324 tp
->prior_ssthresh
= 0;
2326 /* B. In all the states check for reneging SACKs. */
2327 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2330 /* C. Process data loss notification, provided it is valid. */
2331 if ((flag
&FLAG_DATA_LOST
) &&
2332 before(tp
->snd_una
, tp
->high_seq
) &&
2333 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2334 tp
->fackets_out
> tp
->reordering
) {
2335 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, tp
->high_seq
);
2336 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2339 /* D. Check consistency of the current state. */
2340 tcp_verify_left_out(tp
);
2342 /* E. Check state exit conditions. State can be terminated
2343 * when high_seq is ACKed. */
2344 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2345 BUG_TRAP(tp
->retrans_out
== 0);
2346 tp
->retrans_stamp
= 0;
2347 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2348 switch (icsk
->icsk_ca_state
) {
2350 icsk
->icsk_retransmits
= 0;
2351 if (tcp_try_undo_recovery(sk
))
2356 /* CWR is to be held something *above* high_seq
2357 * is ACKed for CWR bit to reach receiver. */
2358 if (tp
->snd_una
!= tp
->high_seq
) {
2359 tcp_complete_cwr(sk
);
2360 tcp_set_ca_state(sk
, TCP_CA_Open
);
2364 case TCP_CA_Disorder
:
2365 tcp_try_undo_dsack(sk
);
2366 if (!tp
->undo_marker
||
2367 /* For SACK case do not Open to allow to undo
2368 * catching for all duplicate ACKs. */
2369 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2370 tp
->undo_marker
= 0;
2371 tcp_set_ca_state(sk
, TCP_CA_Open
);
2375 case TCP_CA_Recovery
:
2376 if (tcp_is_reno(tp
))
2377 tcp_reset_reno_sack(tp
);
2378 if (tcp_try_undo_recovery(sk
))
2380 tcp_complete_cwr(sk
);
2385 /* F. Process state. */
2386 switch (icsk
->icsk_ca_state
) {
2387 case TCP_CA_Recovery
:
2388 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2389 if (tcp_is_reno(tp
) && is_dupack
)
2390 tcp_add_reno_sack(sk
);
2392 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2395 if (flag
&FLAG_DATA_ACKED
)
2396 icsk
->icsk_retransmits
= 0;
2397 if (!tcp_try_undo_loss(sk
)) {
2398 tcp_moderate_cwnd(tp
);
2399 tcp_xmit_retransmit_queue(sk
);
2402 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2404 /* Loss is undone; fall through to processing in Open state. */
2406 if (tcp_is_reno(tp
)) {
2407 if (flag
& FLAG_SND_UNA_ADVANCED
)
2408 tcp_reset_reno_sack(tp
);
2410 tcp_add_reno_sack(sk
);
2413 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2414 tcp_try_undo_dsack(sk
);
2416 if (!tcp_time_to_recover(sk
)) {
2417 tcp_try_to_open(sk
, flag
);
2421 /* MTU probe failure: don't reduce cwnd */
2422 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2423 icsk
->icsk_mtup
.probe_size
&&
2424 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2425 tcp_mtup_probe_failed(sk
);
2426 /* Restores the reduction we did in tcp_mtup_probe() */
2428 tcp_simple_retransmit(sk
);
2432 /* Otherwise enter Recovery state */
2434 if (tcp_is_reno(tp
))
2435 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2437 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2439 tp
->high_seq
= tp
->snd_nxt
;
2440 tp
->prior_ssthresh
= 0;
2441 tp
->undo_marker
= tp
->snd_una
;
2442 tp
->undo_retrans
= tp
->retrans_out
;
2444 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2445 if (!(flag
&FLAG_ECE
))
2446 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2447 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2448 TCP_ECN_queue_cwr(tp
);
2451 tp
->bytes_acked
= 0;
2452 tp
->snd_cwnd_cnt
= 0;
2453 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2456 if (do_lost
|| tcp_head_timedout(sk
))
2457 tcp_update_scoreboard(sk
);
2458 tcp_cwnd_down(sk
, flag
);
2459 tcp_xmit_retransmit_queue(sk
);
2462 /* Read draft-ietf-tcplw-high-performance before mucking
2463 * with this code. (Supersedes RFC1323)
2465 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2467 /* RTTM Rule: A TSecr value received in a segment is used to
2468 * update the averaged RTT measurement only if the segment
2469 * acknowledges some new data, i.e., only if it advances the
2470 * left edge of the send window.
2472 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2473 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2475 * Changed: reset backoff as soon as we see the first valid sample.
2476 * If we do not, we get strongly overestimated rto. With timestamps
2477 * samples are accepted even from very old segments: f.e., when rtt=1
2478 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2479 * answer arrives rto becomes 120 seconds! If at least one of segments
2480 * in window is lost... Voila. --ANK (010210)
2482 struct tcp_sock
*tp
= tcp_sk(sk
);
2483 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2484 tcp_rtt_estimator(sk
, seq_rtt
);
2486 inet_csk(sk
)->icsk_backoff
= 0;
2490 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2492 /* We don't have a timestamp. Can only use
2493 * packets that are not retransmitted to determine
2494 * rtt estimates. Also, we must not reset the
2495 * backoff for rto until we get a non-retransmitted
2496 * packet. This allows us to deal with a situation
2497 * where the network delay has increased suddenly.
2498 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2501 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2504 tcp_rtt_estimator(sk
, seq_rtt
);
2506 inet_csk(sk
)->icsk_backoff
= 0;
2510 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2513 const struct tcp_sock
*tp
= tcp_sk(sk
);
2514 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2515 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2516 tcp_ack_saw_tstamp(sk
, flag
);
2517 else if (seq_rtt
>= 0)
2518 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2521 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
,
2522 u32 in_flight
, int good
)
2524 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2525 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
, good
);
2526 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2529 /* Restart timer after forward progress on connection.
2530 * RFC2988 recommends to restart timer to now+rto.
2532 static void tcp_rearm_rto(struct sock
*sk
)
2534 struct tcp_sock
*tp
= tcp_sk(sk
);
2536 if (!tp
->packets_out
) {
2537 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2539 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2543 /* If we get here, the whole TSO packet has not been acked. */
2544 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2546 struct tcp_sock
*tp
= tcp_sk(sk
);
2549 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2551 packets_acked
= tcp_skb_pcount(skb
);
2552 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2554 packets_acked
-= tcp_skb_pcount(skb
);
2556 if (packets_acked
) {
2557 BUG_ON(tcp_skb_pcount(skb
) == 0);
2558 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2561 return packets_acked
;
2564 /* Remove acknowledged frames from the retransmission queue. If our packet
2565 * is before the ack sequence we can discard it as it's confirmed to have
2566 * arrived at the other end.
2568 static int tcp_clean_rtx_queue(struct sock
*sk
, s32
*seq_rtt_p
)
2570 struct tcp_sock
*tp
= tcp_sk(sk
);
2571 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2572 struct sk_buff
*skb
;
2573 u32 now
= tcp_time_stamp
;
2574 int fully_acked
= 1;
2576 int prior_packets
= tp
->packets_out
;
2578 ktime_t last_ackt
= net_invalid_timestamp();
2580 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2581 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2584 u8 sacked
= scb
->sacked
;
2586 if (after(scb
->end_seq
, tp
->snd_una
)) {
2587 if (tcp_skb_pcount(skb
) == 1 ||
2588 !after(tp
->snd_una
, scb
->seq
))
2591 packets_acked
= tcp_tso_acked(sk
, skb
);
2596 end_seq
= tp
->snd_una
;
2598 packets_acked
= tcp_skb_pcount(skb
);
2599 end_seq
= scb
->end_seq
;
2602 /* MTU probing checks */
2603 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2604 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2605 tcp_mtup_probe_success(sk
, skb
);
2609 if (sacked
& TCPCB_RETRANS
) {
2610 if (sacked
& TCPCB_SACKED_RETRANS
)
2611 tp
->retrans_out
-= packets_acked
;
2612 flag
|= FLAG_RETRANS_DATA_ACKED
;
2614 if ((flag
& FLAG_DATA_ACKED
) ||
2615 (packets_acked
> 1))
2616 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2617 } else if (seq_rtt
< 0) {
2618 seq_rtt
= now
- scb
->when
;
2620 last_ackt
= skb
->tstamp
;
2623 if (sacked
& TCPCB_SACKED_ACKED
)
2624 tp
->sacked_out
-= packets_acked
;
2625 if (sacked
& TCPCB_LOST
)
2626 tp
->lost_out
-= packets_acked
;
2628 if ((sacked
& TCPCB_URG
) && tp
->urg_mode
&&
2629 !before(end_seq
, tp
->snd_up
))
2631 } else if (seq_rtt
< 0) {
2632 seq_rtt
= now
- scb
->when
;
2634 last_ackt
= skb
->tstamp
;
2636 tp
->packets_out
-= packets_acked
;
2638 /* Initial outgoing SYN's get put onto the write_queue
2639 * just like anything else we transmit. It is not
2640 * true data, and if we misinform our callers that
2641 * this ACK acks real data, we will erroneously exit
2642 * connection startup slow start one packet too
2643 * quickly. This is severely frowned upon behavior.
2645 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2646 flag
|= FLAG_DATA_ACKED
;
2648 flag
|= FLAG_SYN_ACKED
;
2649 tp
->retrans_stamp
= 0;
2655 tcp_unlink_write_queue(skb
, sk
);
2656 sk_stream_free_skb(sk
, skb
);
2657 tcp_clear_all_retrans_hints(tp
);
2660 if (flag
& FLAG_ACKED
) {
2661 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2662 const struct tcp_congestion_ops
*ca_ops
2663 = inet_csk(sk
)->icsk_ca_ops
;
2665 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2668 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2669 /* hint's skb might be NULL but we don't need to care */
2670 tp
->fastpath_cnt_hint
-= min_t(u32
, pkts_acked
,
2671 tp
->fastpath_cnt_hint
);
2672 if (tcp_is_reno(tp
))
2673 tcp_remove_reno_sacks(sk
, pkts_acked
);
2675 if (ca_ops
->pkts_acked
) {
2678 /* Is the ACK triggering packet unambiguous? */
2679 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2680 /* High resolution needed and available? */
2681 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2682 !ktime_equal(last_ackt
,
2683 net_invalid_timestamp()))
2684 rtt_us
= ktime_us_delta(ktime_get_real(),
2686 else if (seq_rtt
> 0)
2687 rtt_us
= jiffies_to_usecs(seq_rtt
);
2690 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2694 #if FASTRETRANS_DEBUG > 0
2695 BUG_TRAP((int)tp
->sacked_out
>= 0);
2696 BUG_TRAP((int)tp
->lost_out
>= 0);
2697 BUG_TRAP((int)tp
->retrans_out
>= 0);
2698 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2699 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2701 printk(KERN_DEBUG
"Leak l=%u %d\n",
2702 tp
->lost_out
, icsk
->icsk_ca_state
);
2705 if (tp
->sacked_out
) {
2706 printk(KERN_DEBUG
"Leak s=%u %d\n",
2707 tp
->sacked_out
, icsk
->icsk_ca_state
);
2710 if (tp
->retrans_out
) {
2711 printk(KERN_DEBUG
"Leak r=%u %d\n",
2712 tp
->retrans_out
, icsk
->icsk_ca_state
);
2713 tp
->retrans_out
= 0;
2717 *seq_rtt_p
= seq_rtt
;
2721 static void tcp_ack_probe(struct sock
*sk
)
2723 const struct tcp_sock
*tp
= tcp_sk(sk
);
2724 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2726 /* Was it a usable window open? */
2728 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2729 tp
->snd_una
+ tp
->snd_wnd
)) {
2730 icsk
->icsk_backoff
= 0;
2731 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2732 /* Socket must be waked up by subsequent tcp_data_snd_check().
2733 * This function is not for random using!
2736 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2737 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2742 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2744 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2745 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2748 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2750 const struct tcp_sock
*tp
= tcp_sk(sk
);
2751 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2752 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2755 /* Check that window update is acceptable.
2756 * The function assumes that snd_una<=ack<=snd_next.
2758 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2759 const u32 ack_seq
, const u32 nwin
)
2761 return (after(ack
, tp
->snd_una
) ||
2762 after(ack_seq
, tp
->snd_wl1
) ||
2763 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2766 /* Update our send window.
2768 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2769 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2771 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2774 struct tcp_sock
*tp
= tcp_sk(sk
);
2776 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2778 if (likely(!tcp_hdr(skb
)->syn
))
2779 nwin
<<= tp
->rx_opt
.snd_wscale
;
2781 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2782 flag
|= FLAG_WIN_UPDATE
;
2783 tcp_update_wl(tp
, ack
, ack_seq
);
2785 if (tp
->snd_wnd
!= nwin
) {
2788 /* Note, it is the only place, where
2789 * fast path is recovered for sending TCP.
2792 tcp_fast_path_check(sk
);
2794 if (nwin
> tp
->max_window
) {
2795 tp
->max_window
= nwin
;
2796 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2806 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2807 * continue in congestion avoidance.
2809 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2811 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2812 tp
->snd_cwnd_cnt
= 0;
2813 TCP_ECN_queue_cwr(tp
);
2814 tcp_moderate_cwnd(tp
);
2817 /* A conservative spurious RTO response algorithm: reduce cwnd using
2818 * rate halving and continue in congestion avoidance.
2820 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
2822 tcp_enter_cwr(sk
, 0);
2825 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
2828 tcp_ratehalving_spur_to_response(sk
);
2830 tcp_undo_cwr(sk
, 1);
2833 /* F-RTO spurious RTO detection algorithm (RFC4138)
2835 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2836 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2837 * window (but not to or beyond highest sequence sent before RTO):
2838 * On First ACK, send two new segments out.
2839 * On Second ACK, RTO was likely spurious. Do spurious response (response
2840 * algorithm is not part of the F-RTO detection algorithm
2841 * given in RFC4138 but can be selected separately).
2842 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2843 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2844 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2845 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2847 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2848 * original window even after we transmit two new data segments.
2851 * on first step, wait until first cumulative ACK arrives, then move to
2852 * the second step. In second step, the next ACK decides.
2854 * F-RTO is implemented (mainly) in four functions:
2855 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2856 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2857 * called when tcp_use_frto() showed green light
2858 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2859 * - tcp_enter_frto_loss() is called if there is not enough evidence
2860 * to prove that the RTO is indeed spurious. It transfers the control
2861 * from F-RTO to the conventional RTO recovery
2863 static int tcp_process_frto(struct sock
*sk
, int flag
)
2865 struct tcp_sock
*tp
= tcp_sk(sk
);
2867 tcp_verify_left_out(tp
);
2869 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2870 if (flag
&FLAG_DATA_ACKED
)
2871 inet_csk(sk
)->icsk_retransmits
= 0;
2873 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
2874 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
2875 tp
->undo_marker
= 0;
2877 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2878 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
2882 if (!IsSackFrto() || tcp_is_reno(tp
)) {
2883 /* RFC4138 shortcoming in step 2; should also have case c):
2884 * ACK isn't duplicate nor advances window, e.g., opposite dir
2887 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2890 if (!(flag
&FLAG_DATA_ACKED
)) {
2891 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2896 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2897 /* Prevent sending of new data. */
2898 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2899 tcp_packets_in_flight(tp
));
2903 if ((tp
->frto_counter
>= 2) &&
2904 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2905 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2906 /* RFC4138 shortcoming (see comment above) */
2907 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2910 tcp_enter_frto_loss(sk
, 3, flag
);
2915 if (tp
->frto_counter
== 1) {
2916 /* Sending of the next skb must be allowed or no FRTO */
2917 if (!tcp_send_head(sk
) ||
2918 after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2919 tp
->snd_una
+ tp
->snd_wnd
)) {
2920 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3),
2925 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
2926 tp
->frto_counter
= 2;
2929 switch (sysctl_tcp_frto_response
) {
2931 tcp_undo_spur_to_response(sk
, flag
);
2934 tcp_conservative_spur_to_response(tp
);
2937 tcp_ratehalving_spur_to_response(sk
);
2940 tp
->frto_counter
= 0;
2941 tp
->undo_marker
= 0;
2942 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS
);
2947 /* This routine deals with incoming acks, but not outgoing ones. */
2948 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
2950 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2951 struct tcp_sock
*tp
= tcp_sk(sk
);
2952 u32 prior_snd_una
= tp
->snd_una
;
2953 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
2954 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
2955 u32 prior_in_flight
;
2960 /* If the ack is newer than sent or older than previous acks
2961 * then we can probably ignore it.
2963 if (after(ack
, tp
->snd_nxt
))
2964 goto uninteresting_ack
;
2966 if (before(ack
, prior_snd_una
))
2969 if (after(ack
, prior_snd_una
))
2970 flag
|= FLAG_SND_UNA_ADVANCED
;
2972 if (sysctl_tcp_abc
) {
2973 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
2974 tp
->bytes_acked
+= ack
- prior_snd_una
;
2975 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
2976 /* we assume just one segment left network */
2977 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
2980 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
2981 /* Window is constant, pure forward advance.
2982 * No more checks are required.
2983 * Note, we use the fact that SND.UNA>=SND.WL2.
2985 tcp_update_wl(tp
, ack
, ack_seq
);
2987 flag
|= FLAG_WIN_UPDATE
;
2989 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
2991 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
2993 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
2996 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
2998 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3000 if (TCP_SKB_CB(skb
)->sacked
)
3001 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3003 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3006 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3009 /* We passed data and got it acked, remove any soft error
3010 * log. Something worked...
3012 sk
->sk_err_soft
= 0;
3013 tp
->rcv_tstamp
= tcp_time_stamp
;
3014 prior_packets
= tp
->packets_out
;
3018 prior_in_flight
= tcp_packets_in_flight(tp
);
3020 /* See if we can take anything off of the retransmit queue. */
3021 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
);
3023 if (tp
->frto_counter
)
3024 frto_cwnd
= tcp_process_frto(sk
, flag
);
3026 if (tcp_ack_is_dubious(sk
, flag
)) {
3027 /* Advance CWND, if state allows this. */
3028 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3029 tcp_may_raise_cwnd(sk
, flag
))
3030 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 0);
3031 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
, flag
);
3033 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3034 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 1);
3037 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
3038 dst_confirm(sk
->sk_dst_cache
);
3043 icsk
->icsk_probes_out
= 0;
3045 /* If this ack opens up a zero window, clear backoff. It was
3046 * being used to time the probes, and is probably far higher than
3047 * it needs to be for normal retransmission.
3049 if (tcp_send_head(sk
))
3054 if (TCP_SKB_CB(skb
)->sacked
)
3055 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3058 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3063 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3064 * But, this can also be called on packets in the established flow when
3065 * the fast version below fails.
3067 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
3070 struct tcphdr
*th
= tcp_hdr(skb
);
3071 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
3073 ptr
= (unsigned char *)(th
+ 1);
3074 opt_rx
->saw_tstamp
= 0;
3076 while (length
> 0) {
3083 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3088 if (opsize
< 2) /* "silly options" */
3090 if (opsize
> length
)
3091 return; /* don't parse partial options */
3094 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3095 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
3097 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
3098 in_mss
= opt_rx
->user_mss
;
3099 opt_rx
->mss_clamp
= in_mss
;
3104 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
3105 if (sysctl_tcp_window_scaling
) {
3106 __u8 snd_wscale
= *(__u8
*) ptr
;
3107 opt_rx
->wscale_ok
= 1;
3108 if (snd_wscale
> 14) {
3109 if (net_ratelimit())
3110 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3111 "scaling value %d >14 received.\n",
3115 opt_rx
->snd_wscale
= snd_wscale
;
3118 case TCPOPT_TIMESTAMP
:
3119 if (opsize
==TCPOLEN_TIMESTAMP
) {
3120 if ((estab
&& opt_rx
->tstamp_ok
) ||
3121 (!estab
&& sysctl_tcp_timestamps
)) {
3122 opt_rx
->saw_tstamp
= 1;
3123 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
3124 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
3128 case TCPOPT_SACK_PERM
:
3129 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
3130 if (sysctl_tcp_sack
) {
3131 opt_rx
->sack_ok
= 1;
3132 tcp_sack_reset(opt_rx
);
3138 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3139 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3141 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3144 #ifdef CONFIG_TCP_MD5SIG
3147 * The MD5 Hash has already been
3148 * checked (see tcp_v{4,6}_do_rcv()).
3160 /* Fast parse options. This hopes to only see timestamps.
3161 * If it is wrong it falls back on tcp_parse_options().
3163 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3164 struct tcp_sock
*tp
)
3166 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3167 tp
->rx_opt
.saw_tstamp
= 0;
3169 } else if (tp
->rx_opt
.tstamp_ok
&&
3170 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3171 __be32
*ptr
= (__be32
*)(th
+ 1);
3172 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3173 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3174 tp
->rx_opt
.saw_tstamp
= 1;
3176 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3178 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3182 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3186 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3188 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3189 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3192 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3194 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3195 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3196 * extra check below makes sure this can only happen
3197 * for pure ACK frames. -DaveM
3199 * Not only, also it occurs for expired timestamps.
3202 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3203 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3204 tcp_store_ts_recent(tp
);
3208 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3210 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3211 * it can pass through stack. So, the following predicate verifies that
3212 * this segment is not used for anything but congestion avoidance or
3213 * fast retransmit. Moreover, we even are able to eliminate most of such
3214 * second order effects, if we apply some small "replay" window (~RTO)
3215 * to timestamp space.
3217 * All these measures still do not guarantee that we reject wrapped ACKs
3218 * on networks with high bandwidth, when sequence space is recycled fastly,
3219 * but it guarantees that such events will be very rare and do not affect
3220 * connection seriously. This doesn't look nice, but alas, PAWS is really
3223 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3224 * states that events when retransmit arrives after original data are rare.
3225 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3226 * the biggest problem on large power networks even with minor reordering.
3227 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3228 * up to bandwidth of 18Gigabit/sec. 8) ]
3231 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3233 struct tcp_sock
*tp
= tcp_sk(sk
);
3234 struct tcphdr
*th
= tcp_hdr(skb
);
3235 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3236 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3238 return (/* 1. Pure ACK with correct sequence number. */
3239 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3241 /* 2. ... and duplicate ACK. */
3242 ack
== tp
->snd_una
&&
3244 /* 3. ... and does not update window. */
3245 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3247 /* 4. ... and sits in replay window. */
3248 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3251 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3253 const struct tcp_sock
*tp
= tcp_sk(sk
);
3254 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3255 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3256 !tcp_disordered_ack(sk
, skb
));
3259 /* Check segment sequence number for validity.
3261 * Segment controls are considered valid, if the segment
3262 * fits to the window after truncation to the window. Acceptability
3263 * of data (and SYN, FIN, of course) is checked separately.
3264 * See tcp_data_queue(), for example.
3266 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3267 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3268 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3269 * (borrowed from freebsd)
3272 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3274 return !before(end_seq
, tp
->rcv_wup
) &&
3275 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3278 /* When we get a reset we do this. */
3279 static void tcp_reset(struct sock
*sk
)
3281 /* We want the right error as BSD sees it (and indeed as we do). */
3282 switch (sk
->sk_state
) {
3284 sk
->sk_err
= ECONNREFUSED
;
3286 case TCP_CLOSE_WAIT
:
3292 sk
->sk_err
= ECONNRESET
;
3295 if (!sock_flag(sk
, SOCK_DEAD
))
3296 sk
->sk_error_report(sk
);
3302 * Process the FIN bit. This now behaves as it is supposed to work
3303 * and the FIN takes effect when it is validly part of sequence
3304 * space. Not before when we get holes.
3306 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3307 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3310 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3311 * close and we go into CLOSING (and later onto TIME-WAIT)
3313 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3315 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3317 struct tcp_sock
*tp
= tcp_sk(sk
);
3319 inet_csk_schedule_ack(sk
);
3321 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3322 sock_set_flag(sk
, SOCK_DONE
);
3324 switch (sk
->sk_state
) {
3326 case TCP_ESTABLISHED
:
3327 /* Move to CLOSE_WAIT */
3328 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3329 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3332 case TCP_CLOSE_WAIT
:
3334 /* Received a retransmission of the FIN, do
3339 /* RFC793: Remain in the LAST-ACK state. */
3343 /* This case occurs when a simultaneous close
3344 * happens, we must ack the received FIN and
3345 * enter the CLOSING state.
3348 tcp_set_state(sk
, TCP_CLOSING
);
3351 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3353 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3356 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3357 * cases we should never reach this piece of code.
3359 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3360 __FUNCTION__
, sk
->sk_state
);
3364 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3365 * Probably, we should reset in this case. For now drop them.
3367 __skb_queue_purge(&tp
->out_of_order_queue
);
3368 if (tcp_is_sack(tp
))
3369 tcp_sack_reset(&tp
->rx_opt
);
3370 sk_stream_mem_reclaim(sk
);
3372 if (!sock_flag(sk
, SOCK_DEAD
)) {
3373 sk
->sk_state_change(sk
);
3375 /* Do not send POLL_HUP for half duplex close. */
3376 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3377 sk
->sk_state
== TCP_CLOSE
)
3378 sk_wake_async(sk
, 1, POLL_HUP
);
3380 sk_wake_async(sk
, 1, POLL_IN
);
3384 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3386 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3387 if (before(seq
, sp
->start_seq
))
3388 sp
->start_seq
= seq
;
3389 if (after(end_seq
, sp
->end_seq
))
3390 sp
->end_seq
= end_seq
;
3396 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3398 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3399 if (before(seq
, tp
->rcv_nxt
))
3400 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3402 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3404 tp
->rx_opt
.dsack
= 1;
3405 tp
->duplicate_sack
[0].start_seq
= seq
;
3406 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3407 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3411 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3413 if (!tp
->rx_opt
.dsack
)
3414 tcp_dsack_set(tp
, seq
, end_seq
);
3416 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3419 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3421 struct tcp_sock
*tp
= tcp_sk(sk
);
3423 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3424 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3425 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3426 tcp_enter_quickack_mode(sk
);
3428 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3429 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3431 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3432 end_seq
= tp
->rcv_nxt
;
3433 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3440 /* These routines update the SACK block as out-of-order packets arrive or
3441 * in-order packets close up the sequence space.
3443 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3446 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3447 struct tcp_sack_block
*swalk
= sp
+1;
3449 /* See if the recent change to the first SACK eats into
3450 * or hits the sequence space of other SACK blocks, if so coalesce.
3452 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3453 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3456 /* Zap SWALK, by moving every further SACK up by one slot.
3457 * Decrease num_sacks.
3459 tp
->rx_opt
.num_sacks
--;
3460 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3461 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3465 this_sack
++, swalk
++;
3469 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3473 tmp
= sack1
->start_seq
;
3474 sack1
->start_seq
= sack2
->start_seq
;
3475 sack2
->start_seq
= tmp
;
3477 tmp
= sack1
->end_seq
;
3478 sack1
->end_seq
= sack2
->end_seq
;
3479 sack2
->end_seq
= tmp
;
3482 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3484 struct tcp_sock
*tp
= tcp_sk(sk
);
3485 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3486 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3492 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3493 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3494 /* Rotate this_sack to the first one. */
3495 for (; this_sack
>0; this_sack
--, sp
--)
3496 tcp_sack_swap(sp
, sp
-1);
3498 tcp_sack_maybe_coalesce(tp
);
3503 /* Could not find an adjacent existing SACK, build a new one,
3504 * put it at the front, and shift everyone else down. We
3505 * always know there is at least one SACK present already here.
3507 * If the sack array is full, forget about the last one.
3509 if (this_sack
>= 4) {
3511 tp
->rx_opt
.num_sacks
--;
3514 for (; this_sack
> 0; this_sack
--, sp
--)
3518 /* Build the new head SACK, and we're done. */
3519 sp
->start_seq
= seq
;
3520 sp
->end_seq
= end_seq
;
3521 tp
->rx_opt
.num_sacks
++;
3522 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3525 /* RCV.NXT advances, some SACKs should be eaten. */
3527 static void tcp_sack_remove(struct tcp_sock
*tp
)
3529 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3530 int num_sacks
= tp
->rx_opt
.num_sacks
;
3533 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3534 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3535 tp
->rx_opt
.num_sacks
= 0;
3536 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3540 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3541 /* Check if the start of the sack is covered by RCV.NXT. */
3542 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3545 /* RCV.NXT must cover all the block! */
3546 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3548 /* Zap this SACK, by moving forward any other SACKS. */
3549 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3550 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3557 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3558 tp
->rx_opt
.num_sacks
= num_sacks
;
3559 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3563 /* This one checks to see if we can put data from the
3564 * out_of_order queue into the receive_queue.
3566 static void tcp_ofo_queue(struct sock
*sk
)
3568 struct tcp_sock
*tp
= tcp_sk(sk
);
3569 __u32 dsack_high
= tp
->rcv_nxt
;
3570 struct sk_buff
*skb
;
3572 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3573 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3576 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3577 __u32 dsack
= dsack_high
;
3578 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3579 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3580 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3583 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3584 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3585 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3589 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3590 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3591 TCP_SKB_CB(skb
)->end_seq
);
3593 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3594 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3595 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3596 if (tcp_hdr(skb
)->fin
)
3597 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3601 static int tcp_prune_queue(struct sock
*sk
);
3603 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3605 struct tcphdr
*th
= tcp_hdr(skb
);
3606 struct tcp_sock
*tp
= tcp_sk(sk
);
3609 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3612 __skb_pull(skb
, th
->doff
*4);
3614 TCP_ECN_accept_cwr(tp
, skb
);
3616 if (tp
->rx_opt
.dsack
) {
3617 tp
->rx_opt
.dsack
= 0;
3618 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3619 4 - tp
->rx_opt
.tstamp_ok
);
3622 /* Queue data for delivery to the user.
3623 * Packets in sequence go to the receive queue.
3624 * Out of sequence packets to the out_of_order_queue.
3626 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3627 if (tcp_receive_window(tp
) == 0)
3630 /* Ok. In sequence. In window. */
3631 if (tp
->ucopy
.task
== current
&&
3632 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3633 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3634 int chunk
= min_t(unsigned int, skb
->len
,
3637 __set_current_state(TASK_RUNNING
);
3640 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3641 tp
->ucopy
.len
-= chunk
;
3642 tp
->copied_seq
+= chunk
;
3643 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3644 tcp_rcv_space_adjust(sk
);
3652 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3653 !sk_stream_rmem_schedule(sk
, skb
))) {
3654 if (tcp_prune_queue(sk
) < 0 ||
3655 !sk_stream_rmem_schedule(sk
, skb
))
3658 sk_stream_set_owner_r(skb
, sk
);
3659 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3661 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3663 tcp_event_data_recv(sk
, skb
);
3665 tcp_fin(skb
, sk
, th
);
3667 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3670 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3671 * gap in queue is filled.
3673 if (skb_queue_empty(&tp
->out_of_order_queue
))
3674 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3677 if (tp
->rx_opt
.num_sacks
)
3678 tcp_sack_remove(tp
);
3680 tcp_fast_path_check(sk
);
3684 else if (!sock_flag(sk
, SOCK_DEAD
))
3685 sk
->sk_data_ready(sk
, 0);
3689 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3690 /* A retransmit, 2nd most common case. Force an immediate ack. */
3691 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3692 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3695 tcp_enter_quickack_mode(sk
);
3696 inet_csk_schedule_ack(sk
);
3702 /* Out of window. F.e. zero window probe. */
3703 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3706 tcp_enter_quickack_mode(sk
);
3708 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3709 /* Partial packet, seq < rcv_next < end_seq */
3710 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3711 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3712 TCP_SKB_CB(skb
)->end_seq
);
3714 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3716 /* If window is closed, drop tail of packet. But after
3717 * remembering D-SACK for its head made in previous line.
3719 if (!tcp_receive_window(tp
))
3724 TCP_ECN_check_ce(tp
, skb
);
3726 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3727 !sk_stream_rmem_schedule(sk
, skb
)) {
3728 if (tcp_prune_queue(sk
) < 0 ||
3729 !sk_stream_rmem_schedule(sk
, skb
))
3733 /* Disable header prediction. */
3735 inet_csk_schedule_ack(sk
);
3737 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3738 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3740 sk_stream_set_owner_r(skb
, sk
);
3742 if (!skb_peek(&tp
->out_of_order_queue
)) {
3743 /* Initial out of order segment, build 1 SACK. */
3744 if (tcp_is_sack(tp
)) {
3745 tp
->rx_opt
.num_sacks
= 1;
3746 tp
->rx_opt
.dsack
= 0;
3747 tp
->rx_opt
.eff_sacks
= 1;
3748 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3749 tp
->selective_acks
[0].end_seq
=
3750 TCP_SKB_CB(skb
)->end_seq
;
3752 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3754 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3755 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3756 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3758 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3759 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3761 if (!tp
->rx_opt
.num_sacks
||
3762 tp
->selective_acks
[0].end_seq
!= seq
)
3765 /* Common case: data arrive in order after hole. */
3766 tp
->selective_acks
[0].end_seq
= end_seq
;
3770 /* Find place to insert this segment. */
3772 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3774 } while ((skb1
= skb1
->prev
) !=
3775 (struct sk_buff
*)&tp
->out_of_order_queue
);
3777 /* Do skb overlap to previous one? */
3778 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3779 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3780 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3781 /* All the bits are present. Drop. */
3783 tcp_dsack_set(tp
, seq
, end_seq
);
3786 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3787 /* Partial overlap. */
3788 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3793 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3795 /* And clean segments covered by new one as whole. */
3796 while ((skb1
= skb
->next
) !=
3797 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3798 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3799 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3800 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3803 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3804 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3809 if (tcp_is_sack(tp
))
3810 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3814 /* Collapse contiguous sequence of skbs head..tail with
3815 * sequence numbers start..end.
3816 * Segments with FIN/SYN are not collapsed (only because this
3820 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3821 struct sk_buff
*head
, struct sk_buff
*tail
,
3824 struct sk_buff
*skb
;
3826 /* First, check that queue is collapsible and find
3827 * the point where collapsing can be useful. */
3828 for (skb
= head
; skb
!= tail
; ) {
3829 /* No new bits? It is possible on ofo queue. */
3830 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3831 struct sk_buff
*next
= skb
->next
;
3832 __skb_unlink(skb
, list
);
3834 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3839 /* The first skb to collapse is:
3841 * - bloated or contains data before "start" or
3842 * overlaps to the next one.
3844 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
3845 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3846 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3847 (skb
->next
!= tail
&&
3848 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3851 /* Decided to skip this, advance start seq. */
3852 start
= TCP_SKB_CB(skb
)->end_seq
;
3855 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
3858 while (before(start
, end
)) {
3859 struct sk_buff
*nskb
;
3860 int header
= skb_headroom(skb
);
3861 int copy
= SKB_MAX_ORDER(header
, 0);
3863 /* Too big header? This can happen with IPv6. */
3866 if (end
-start
< copy
)
3868 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3872 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
3873 skb_set_network_header(nskb
, (skb_network_header(skb
) -
3875 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
3877 skb_reserve(nskb
, header
);
3878 memcpy(nskb
->head
, skb
->head
, header
);
3879 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3880 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3881 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3882 sk_stream_set_owner_r(nskb
, sk
);
3884 /* Copy data, releasing collapsed skbs. */
3886 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3887 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3891 size
= min(copy
, size
);
3892 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3894 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3898 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3899 struct sk_buff
*next
= skb
->next
;
3900 __skb_unlink(skb
, list
);
3902 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3905 tcp_hdr(skb
)->syn
||
3913 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3914 * and tcp_collapse() them until all the queue is collapsed.
3916 static void tcp_collapse_ofo_queue(struct sock
*sk
)
3918 struct tcp_sock
*tp
= tcp_sk(sk
);
3919 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
3920 struct sk_buff
*head
;
3926 start
= TCP_SKB_CB(skb
)->seq
;
3927 end
= TCP_SKB_CB(skb
)->end_seq
;
3933 /* Segment is terminated when we see gap or when
3934 * we are at the end of all the queue. */
3935 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
3936 after(TCP_SKB_CB(skb
)->seq
, end
) ||
3937 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
3938 tcp_collapse(sk
, &tp
->out_of_order_queue
,
3939 head
, skb
, start
, end
);
3941 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
3943 /* Start new segment */
3944 start
= TCP_SKB_CB(skb
)->seq
;
3945 end
= TCP_SKB_CB(skb
)->end_seq
;
3947 if (before(TCP_SKB_CB(skb
)->seq
, start
))
3948 start
= TCP_SKB_CB(skb
)->seq
;
3949 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
3950 end
= TCP_SKB_CB(skb
)->end_seq
;
3955 /* Reduce allocated memory if we can, trying to get
3956 * the socket within its memory limits again.
3958 * Return less than zero if we should start dropping frames
3959 * until the socket owning process reads some of the data
3960 * to stabilize the situation.
3962 static int tcp_prune_queue(struct sock
*sk
)
3964 struct tcp_sock
*tp
= tcp_sk(sk
);
3966 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
3968 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
3970 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
3971 tcp_clamp_window(sk
);
3972 else if (tcp_memory_pressure
)
3973 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
3975 tcp_collapse_ofo_queue(sk
);
3976 tcp_collapse(sk
, &sk
->sk_receive_queue
,
3977 sk
->sk_receive_queue
.next
,
3978 (struct sk_buff
*)&sk
->sk_receive_queue
,
3979 tp
->copied_seq
, tp
->rcv_nxt
);
3980 sk_stream_mem_reclaim(sk
);
3982 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
3985 /* Collapsing did not help, destructive actions follow.
3986 * This must not ever occur. */
3988 /* First, purge the out_of_order queue. */
3989 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3990 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
3991 __skb_queue_purge(&tp
->out_of_order_queue
);
3993 /* Reset SACK state. A conforming SACK implementation will
3994 * do the same at a timeout based retransmit. When a connection
3995 * is in a sad state like this, we care only about integrity
3996 * of the connection not performance.
3998 if (tcp_is_sack(tp
))
3999 tcp_sack_reset(&tp
->rx_opt
);
4000 sk_stream_mem_reclaim(sk
);
4003 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4006 /* If we are really being abused, tell the caller to silently
4007 * drop receive data on the floor. It will get retransmitted
4008 * and hopefully then we'll have sufficient space.
4010 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
4012 /* Massive buffer overcommit. */
4018 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4019 * As additional protections, we do not touch cwnd in retransmission phases,
4020 * and if application hit its sndbuf limit recently.
4022 void tcp_cwnd_application_limited(struct sock
*sk
)
4024 struct tcp_sock
*tp
= tcp_sk(sk
);
4026 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4027 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4028 /* Limited by application or receiver window. */
4029 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4030 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4031 if (win_used
< tp
->snd_cwnd
) {
4032 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4033 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4035 tp
->snd_cwnd_used
= 0;
4037 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4040 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4042 struct tcp_sock
*tp
= tcp_sk(sk
);
4044 /* If the user specified a specific send buffer setting, do
4047 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4050 /* If we are under global TCP memory pressure, do not expand. */
4051 if (tcp_memory_pressure
)
4054 /* If we are under soft global TCP memory pressure, do not expand. */
4055 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4058 /* If we filled the congestion window, do not expand. */
4059 if (tp
->packets_out
>= tp
->snd_cwnd
)
4065 /* When incoming ACK allowed to free some skb from write_queue,
4066 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4067 * on the exit from tcp input handler.
4069 * PROBLEM: sndbuf expansion does not work well with largesend.
4071 static void tcp_new_space(struct sock
*sk
)
4073 struct tcp_sock
*tp
= tcp_sk(sk
);
4075 if (tcp_should_expand_sndbuf(sk
)) {
4076 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4077 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4078 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4079 tp
->reordering
+ 1);
4080 sndmem
*= 2*demanded
;
4081 if (sndmem
> sk
->sk_sndbuf
)
4082 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4083 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4086 sk
->sk_write_space(sk
);
4089 static void tcp_check_space(struct sock
*sk
)
4091 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4092 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4093 if (sk
->sk_socket
&&
4094 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4099 static inline void tcp_data_snd_check(struct sock
*sk
)
4101 tcp_push_pending_frames(sk
);
4102 tcp_check_space(sk
);
4106 * Check if sending an ack is needed.
4108 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4110 struct tcp_sock
*tp
= tcp_sk(sk
);
4112 /* More than one full frame received... */
4113 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4114 /* ... and right edge of window advances far enough.
4115 * (tcp_recvmsg() will send ACK otherwise). Or...
4117 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4118 /* We ACK each frame or... */
4119 tcp_in_quickack_mode(sk
) ||
4120 /* We have out of order data. */
4122 skb_peek(&tp
->out_of_order_queue
))) {
4123 /* Then ack it now */
4126 /* Else, send delayed ack. */
4127 tcp_send_delayed_ack(sk
);
4131 static inline void tcp_ack_snd_check(struct sock
*sk
)
4133 if (!inet_csk_ack_scheduled(sk
)) {
4134 /* We sent a data segment already. */
4137 __tcp_ack_snd_check(sk
, 1);
4141 * This routine is only called when we have urgent data
4142 * signaled. Its the 'slow' part of tcp_urg. It could be
4143 * moved inline now as tcp_urg is only called from one
4144 * place. We handle URGent data wrong. We have to - as
4145 * BSD still doesn't use the correction from RFC961.
4146 * For 1003.1g we should support a new option TCP_STDURG to permit
4147 * either form (or just set the sysctl tcp_stdurg).
4150 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4152 struct tcp_sock
*tp
= tcp_sk(sk
);
4153 u32 ptr
= ntohs(th
->urg_ptr
);
4155 if (ptr
&& !sysctl_tcp_stdurg
)
4157 ptr
+= ntohl(th
->seq
);
4159 /* Ignore urgent data that we've already seen and read. */
4160 if (after(tp
->copied_seq
, ptr
))
4163 /* Do not replay urg ptr.
4165 * NOTE: interesting situation not covered by specs.
4166 * Misbehaving sender may send urg ptr, pointing to segment,
4167 * which we already have in ofo queue. We are not able to fetch
4168 * such data and will stay in TCP_URG_NOTYET until will be eaten
4169 * by recvmsg(). Seems, we are not obliged to handle such wicked
4170 * situations. But it is worth to think about possibility of some
4171 * DoSes using some hypothetical application level deadlock.
4173 if (before(ptr
, tp
->rcv_nxt
))
4176 /* Do we already have a newer (or duplicate) urgent pointer? */
4177 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4180 /* Tell the world about our new urgent pointer. */
4183 /* We may be adding urgent data when the last byte read was
4184 * urgent. To do this requires some care. We cannot just ignore
4185 * tp->copied_seq since we would read the last urgent byte again
4186 * as data, nor can we alter copied_seq until this data arrives
4187 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4189 * NOTE. Double Dutch. Rendering to plain English: author of comment
4190 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4191 * and expect that both A and B disappear from stream. This is _wrong_.
4192 * Though this happens in BSD with high probability, this is occasional.
4193 * Any application relying on this is buggy. Note also, that fix "works"
4194 * only in this artificial test. Insert some normal data between A and B and we will
4195 * decline of BSD again. Verdict: it is better to remove to trap
4198 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4199 !sock_flag(sk
, SOCK_URGINLINE
) &&
4200 tp
->copied_seq
!= tp
->rcv_nxt
) {
4201 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4203 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4204 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4209 tp
->urg_data
= TCP_URG_NOTYET
;
4212 /* Disable header prediction. */
4216 /* This is the 'fast' part of urgent handling. */
4217 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4219 struct tcp_sock
*tp
= tcp_sk(sk
);
4221 /* Check if we get a new urgent pointer - normally not. */
4223 tcp_check_urg(sk
,th
);
4225 /* Do we wait for any urgent data? - normally not... */
4226 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4227 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4230 /* Is the urgent pointer pointing into this packet? */
4231 if (ptr
< skb
->len
) {
4233 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4235 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4236 if (!sock_flag(sk
, SOCK_DEAD
))
4237 sk
->sk_data_ready(sk
, 0);
4242 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4244 struct tcp_sock
*tp
= tcp_sk(sk
);
4245 int chunk
= skb
->len
- hlen
;
4249 if (skb_csum_unnecessary(skb
))
4250 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4252 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4256 tp
->ucopy
.len
-= chunk
;
4257 tp
->copied_seq
+= chunk
;
4258 tcp_rcv_space_adjust(sk
);
4265 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4269 if (sock_owned_by_user(sk
)) {
4271 result
= __tcp_checksum_complete(skb
);
4274 result
= __tcp_checksum_complete(skb
);
4279 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4281 return !skb_csum_unnecessary(skb
) &&
4282 __tcp_checksum_complete_user(sk
, skb
);
4285 #ifdef CONFIG_NET_DMA
4286 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4288 struct tcp_sock
*tp
= tcp_sk(sk
);
4289 int chunk
= skb
->len
- hlen
;
4291 int copied_early
= 0;
4293 if (tp
->ucopy
.wakeup
)
4296 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4297 tp
->ucopy
.dma_chan
= get_softnet_dma();
4299 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4301 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4302 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4307 tp
->ucopy
.dma_cookie
= dma_cookie
;
4310 tp
->ucopy
.len
-= chunk
;
4311 tp
->copied_seq
+= chunk
;
4312 tcp_rcv_space_adjust(sk
);
4314 if ((tp
->ucopy
.len
== 0) ||
4315 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4316 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4317 tp
->ucopy
.wakeup
= 1;
4318 sk
->sk_data_ready(sk
, 0);
4320 } else if (chunk
> 0) {
4321 tp
->ucopy
.wakeup
= 1;
4322 sk
->sk_data_ready(sk
, 0);
4325 return copied_early
;
4327 #endif /* CONFIG_NET_DMA */
4330 * TCP receive function for the ESTABLISHED state.
4332 * It is split into a fast path and a slow path. The fast path is
4334 * - A zero window was announced from us - zero window probing
4335 * is only handled properly in the slow path.
4336 * - Out of order segments arrived.
4337 * - Urgent data is expected.
4338 * - There is no buffer space left
4339 * - Unexpected TCP flags/window values/header lengths are received
4340 * (detected by checking the TCP header against pred_flags)
4341 * - Data is sent in both directions. Fast path only supports pure senders
4342 * or pure receivers (this means either the sequence number or the ack
4343 * value must stay constant)
4344 * - Unexpected TCP option.
4346 * When these conditions are not satisfied it drops into a standard
4347 * receive procedure patterned after RFC793 to handle all cases.
4348 * The first three cases are guaranteed by proper pred_flags setting,
4349 * the rest is checked inline. Fast processing is turned on in
4350 * tcp_data_queue when everything is OK.
4352 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4353 struct tcphdr
*th
, unsigned len
)
4355 struct tcp_sock
*tp
= tcp_sk(sk
);
4358 * Header prediction.
4359 * The code loosely follows the one in the famous
4360 * "30 instruction TCP receive" Van Jacobson mail.
4362 * Van's trick is to deposit buffers into socket queue
4363 * on a device interrupt, to call tcp_recv function
4364 * on the receive process context and checksum and copy
4365 * the buffer to user space. smart...
4367 * Our current scheme is not silly either but we take the
4368 * extra cost of the net_bh soft interrupt processing...
4369 * We do checksum and copy also but from device to kernel.
4372 tp
->rx_opt
.saw_tstamp
= 0;
4374 /* pred_flags is 0xS?10 << 16 + snd_wnd
4375 * if header_prediction is to be made
4376 * 'S' will always be tp->tcp_header_len >> 2
4377 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4378 * turn it off (when there are holes in the receive
4379 * space for instance)
4380 * PSH flag is ignored.
4383 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4384 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4385 int tcp_header_len
= tp
->tcp_header_len
;
4387 /* Timestamp header prediction: tcp_header_len
4388 * is automatically equal to th->doff*4 due to pred_flags
4392 /* Check timestamp */
4393 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4394 __be32
*ptr
= (__be32
*)(th
+ 1);
4396 /* No? Slow path! */
4397 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4398 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4401 tp
->rx_opt
.saw_tstamp
= 1;
4403 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4405 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4407 /* If PAWS failed, check it more carefully in slow path */
4408 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4411 /* DO NOT update ts_recent here, if checksum fails
4412 * and timestamp was corrupted part, it will result
4413 * in a hung connection since we will drop all
4414 * future packets due to the PAWS test.
4418 if (len
<= tcp_header_len
) {
4419 /* Bulk data transfer: sender */
4420 if (len
== tcp_header_len
) {
4421 /* Predicted packet is in window by definition.
4422 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4423 * Hence, check seq<=rcv_wup reduces to:
4425 if (tcp_header_len
==
4426 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4427 tp
->rcv_nxt
== tp
->rcv_wup
)
4428 tcp_store_ts_recent(tp
);
4430 /* We know that such packets are checksummed
4433 tcp_ack(sk
, skb
, 0);
4435 tcp_data_snd_check(sk
);
4437 } else { /* Header too small */
4438 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4443 int copied_early
= 0;
4445 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4446 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4447 #ifdef CONFIG_NET_DMA
4448 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4453 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4454 __set_current_state(TASK_RUNNING
);
4456 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4460 /* Predicted packet is in window by definition.
4461 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4462 * Hence, check seq<=rcv_wup reduces to:
4464 if (tcp_header_len
==
4465 (sizeof(struct tcphdr
) +
4466 TCPOLEN_TSTAMP_ALIGNED
) &&
4467 tp
->rcv_nxt
== tp
->rcv_wup
)
4468 tcp_store_ts_recent(tp
);
4470 tcp_rcv_rtt_measure_ts(sk
, skb
);
4472 __skb_pull(skb
, tcp_header_len
);
4473 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4474 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4477 tcp_cleanup_rbuf(sk
, skb
->len
);
4480 if (tcp_checksum_complete_user(sk
, skb
))
4483 /* Predicted packet is in window by definition.
4484 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4485 * Hence, check seq<=rcv_wup reduces to:
4487 if (tcp_header_len
==
4488 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4489 tp
->rcv_nxt
== tp
->rcv_wup
)
4490 tcp_store_ts_recent(tp
);
4492 tcp_rcv_rtt_measure_ts(sk
, skb
);
4494 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4497 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4499 /* Bulk data transfer: receiver */
4500 __skb_pull(skb
,tcp_header_len
);
4501 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4502 sk_stream_set_owner_r(skb
, sk
);
4503 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4506 tcp_event_data_recv(sk
, skb
);
4508 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4509 /* Well, only one small jumplet in fast path... */
4510 tcp_ack(sk
, skb
, FLAG_DATA
);
4511 tcp_data_snd_check(sk
);
4512 if (!inet_csk_ack_scheduled(sk
))
4516 __tcp_ack_snd_check(sk
, 0);
4518 #ifdef CONFIG_NET_DMA
4520 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4526 sk
->sk_data_ready(sk
, 0);
4532 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4536 * RFC1323: H1. Apply PAWS check first.
4538 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4539 tcp_paws_discard(sk
, skb
)) {
4541 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4542 tcp_send_dupack(sk
, skb
);
4545 /* Resets are accepted even if PAWS failed.
4547 ts_recent update must be made after we are sure
4548 that the packet is in window.
4553 * Standard slow path.
4556 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4557 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4558 * (RST) segments are validated by checking their SEQ-fields."
4559 * And page 69: "If an incoming segment is not acceptable,
4560 * an acknowledgment should be sent in reply (unless the RST bit
4561 * is set, if so drop the segment and return)".
4564 tcp_send_dupack(sk
, skb
);
4573 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4575 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4576 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4577 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4584 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4586 tcp_rcv_rtt_measure_ts(sk
, skb
);
4588 /* Process urgent data. */
4589 tcp_urg(sk
, skb
, th
);
4591 /* step 7: process the segment text */
4592 tcp_data_queue(sk
, skb
);
4594 tcp_data_snd_check(sk
);
4595 tcp_ack_snd_check(sk
);
4599 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4606 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4607 struct tcphdr
*th
, unsigned len
)
4609 struct tcp_sock
*tp
= tcp_sk(sk
);
4610 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4611 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4613 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4617 * "If the state is SYN-SENT then
4618 * first check the ACK bit
4619 * If the ACK bit is set
4620 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4621 * a reset (unless the RST bit is set, if so drop
4622 * the segment and return)"
4624 * We do not send data with SYN, so that RFC-correct
4627 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4628 goto reset_and_undo
;
4630 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4631 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4633 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4634 goto reset_and_undo
;
4637 /* Now ACK is acceptable.
4639 * "If the RST bit is set
4640 * If the ACK was acceptable then signal the user "error:
4641 * connection reset", drop the segment, enter CLOSED state,
4642 * delete TCB, and return."
4651 * "fifth, if neither of the SYN or RST bits is set then
4652 * drop the segment and return."
4658 goto discard_and_undo
;
4661 * "If the SYN bit is on ...
4662 * are acceptable then ...
4663 * (our SYN has been ACKed), change the connection
4664 * state to ESTABLISHED..."
4667 TCP_ECN_rcv_synack(tp
, th
);
4669 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4670 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4672 /* Ok.. it's good. Set up sequence numbers and
4673 * move to established.
4675 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4676 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4678 /* RFC1323: The window in SYN & SYN/ACK segments is
4681 tp
->snd_wnd
= ntohs(th
->window
);
4682 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4684 if (!tp
->rx_opt
.wscale_ok
) {
4685 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4686 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4689 if (tp
->rx_opt
.saw_tstamp
) {
4690 tp
->rx_opt
.tstamp_ok
= 1;
4691 tp
->tcp_header_len
=
4692 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4693 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4694 tcp_store_ts_recent(tp
);
4696 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4699 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
4700 tcp_enable_fack(tp
);
4703 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4704 tcp_initialize_rcv_mss(sk
);
4706 /* Remember, tcp_poll() does not lock socket!
4707 * Change state from SYN-SENT only after copied_seq
4708 * is initialized. */
4709 tp
->copied_seq
= tp
->rcv_nxt
;
4711 tcp_set_state(sk
, TCP_ESTABLISHED
);
4713 security_inet_conn_established(sk
, skb
);
4715 /* Make sure socket is routed, for correct metrics. */
4716 icsk
->icsk_af_ops
->rebuild_header(sk
);
4718 tcp_init_metrics(sk
);
4720 tcp_init_congestion_control(sk
);
4722 /* Prevent spurious tcp_cwnd_restart() on first data
4725 tp
->lsndtime
= tcp_time_stamp
;
4727 tcp_init_buffer_space(sk
);
4729 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4730 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4732 if (!tp
->rx_opt
.snd_wscale
)
4733 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4737 if (!sock_flag(sk
, SOCK_DEAD
)) {
4738 sk
->sk_state_change(sk
);
4739 sk_wake_async(sk
, 0, POLL_OUT
);
4742 if (sk
->sk_write_pending
||
4743 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4744 icsk
->icsk_ack
.pingpong
) {
4745 /* Save one ACK. Data will be ready after
4746 * several ticks, if write_pending is set.
4748 * It may be deleted, but with this feature tcpdumps
4749 * look so _wonderfully_ clever, that I was not able
4750 * to stand against the temptation 8) --ANK
4752 inet_csk_schedule_ack(sk
);
4753 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4754 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4755 tcp_incr_quickack(sk
);
4756 tcp_enter_quickack_mode(sk
);
4757 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4758 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4769 /* No ACK in the segment */
4773 * "If the RST bit is set
4775 * Otherwise (no ACK) drop the segment and return."
4778 goto discard_and_undo
;
4782 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4783 goto discard_and_undo
;
4786 /* We see SYN without ACK. It is attempt of
4787 * simultaneous connect with crossed SYNs.
4788 * Particularly, it can be connect to self.
4790 tcp_set_state(sk
, TCP_SYN_RECV
);
4792 if (tp
->rx_opt
.saw_tstamp
) {
4793 tp
->rx_opt
.tstamp_ok
= 1;
4794 tcp_store_ts_recent(tp
);
4795 tp
->tcp_header_len
=
4796 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4798 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4801 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4802 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4804 /* RFC1323: The window in SYN & SYN/ACK segments is
4807 tp
->snd_wnd
= ntohs(th
->window
);
4808 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4809 tp
->max_window
= tp
->snd_wnd
;
4811 TCP_ECN_rcv_syn(tp
, th
);
4814 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4815 tcp_initialize_rcv_mss(sk
);
4818 tcp_send_synack(sk
);
4820 /* Note, we could accept data and URG from this segment.
4821 * There are no obstacles to make this.
4823 * However, if we ignore data in ACKless segments sometimes,
4824 * we have no reasons to accept it sometimes.
4825 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4826 * is not flawless. So, discard packet for sanity.
4827 * Uncomment this return to process the data.
4834 /* "fifth, if neither of the SYN or RST bits is set then
4835 * drop the segment and return."
4839 tcp_clear_options(&tp
->rx_opt
);
4840 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4844 tcp_clear_options(&tp
->rx_opt
);
4845 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4851 * This function implements the receiving procedure of RFC 793 for
4852 * all states except ESTABLISHED and TIME_WAIT.
4853 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4854 * address independent.
4857 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4858 struct tcphdr
*th
, unsigned len
)
4860 struct tcp_sock
*tp
= tcp_sk(sk
);
4861 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4864 tp
->rx_opt
.saw_tstamp
= 0;
4866 switch (sk
->sk_state
) {
4878 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4881 /* Now we have several options: In theory there is
4882 * nothing else in the frame. KA9Q has an option to
4883 * send data with the syn, BSD accepts data with the
4884 * syn up to the [to be] advertised window and
4885 * Solaris 2.1 gives you a protocol error. For now
4886 * we just ignore it, that fits the spec precisely
4887 * and avoids incompatibilities. It would be nice in
4888 * future to drop through and process the data.
4890 * Now that TTCP is starting to be used we ought to
4892 * But, this leaves one open to an easy denial of
4893 * service attack, and SYN cookies can't defend
4894 * against this problem. So, we drop the data
4895 * in the interest of security over speed unless
4896 * it's still in use.
4904 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4908 /* Do step6 onward by hand. */
4909 tcp_urg(sk
, skb
, th
);
4911 tcp_data_snd_check(sk
);
4915 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4916 tcp_paws_discard(sk
, skb
)) {
4918 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4919 tcp_send_dupack(sk
, skb
);
4922 /* Reset is accepted even if it did not pass PAWS. */
4925 /* step 1: check sequence number */
4926 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4928 tcp_send_dupack(sk
, skb
);
4932 /* step 2: check RST bit */
4938 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4940 /* step 3: check security and precedence [ignored] */
4944 * Check for a SYN in window.
4946 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4947 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4952 /* step 5: check the ACK field */
4954 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4956 switch (sk
->sk_state
) {
4959 tp
->copied_seq
= tp
->rcv_nxt
;
4961 tcp_set_state(sk
, TCP_ESTABLISHED
);
4962 sk
->sk_state_change(sk
);
4964 /* Note, that this wakeup is only for marginal
4965 * crossed SYN case. Passively open sockets
4966 * are not waked up, because sk->sk_sleep ==
4967 * NULL and sk->sk_socket == NULL.
4969 if (sk
->sk_socket
) {
4970 sk_wake_async(sk
,0,POLL_OUT
);
4973 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
4974 tp
->snd_wnd
= ntohs(th
->window
) <<
4975 tp
->rx_opt
.snd_wscale
;
4976 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
4977 TCP_SKB_CB(skb
)->seq
);
4979 /* tcp_ack considers this ACK as duplicate
4980 * and does not calculate rtt.
4981 * Fix it at least with timestamps.
4983 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4985 tcp_ack_saw_tstamp(sk
, 0);
4987 if (tp
->rx_opt
.tstamp_ok
)
4988 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4990 /* Make sure socket is routed, for
4993 icsk
->icsk_af_ops
->rebuild_header(sk
);
4995 tcp_init_metrics(sk
);
4997 tcp_init_congestion_control(sk
);
4999 /* Prevent spurious tcp_cwnd_restart() on
5000 * first data packet.
5002 tp
->lsndtime
= tcp_time_stamp
;
5005 tcp_initialize_rcv_mss(sk
);
5006 tcp_init_buffer_space(sk
);
5007 tcp_fast_path_on(tp
);
5014 if (tp
->snd_una
== tp
->write_seq
) {
5015 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5016 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5017 dst_confirm(sk
->sk_dst_cache
);
5019 if (!sock_flag(sk
, SOCK_DEAD
))
5020 /* Wake up lingering close() */
5021 sk
->sk_state_change(sk
);
5025 if (tp
->linger2
< 0 ||
5026 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5027 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5029 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5033 tmo
= tcp_fin_time(sk
);
5034 if (tmo
> TCP_TIMEWAIT_LEN
) {
5035 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5036 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5037 /* Bad case. We could lose such FIN otherwise.
5038 * It is not a big problem, but it looks confusing
5039 * and not so rare event. We still can lose it now,
5040 * if it spins in bh_lock_sock(), but it is really
5043 inet_csk_reset_keepalive_timer(sk
, tmo
);
5045 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5053 if (tp
->snd_una
== tp
->write_seq
) {
5054 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5060 if (tp
->snd_una
== tp
->write_seq
) {
5061 tcp_update_metrics(sk
);
5070 /* step 6: check the URG bit */
5071 tcp_urg(sk
, skb
, th
);
5073 /* step 7: process the segment text */
5074 switch (sk
->sk_state
) {
5075 case TCP_CLOSE_WAIT
:
5078 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5082 /* RFC 793 says to queue data in these states,
5083 * RFC 1122 says we MUST send a reset.
5084 * BSD 4.4 also does reset.
5086 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5087 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5088 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5089 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5095 case TCP_ESTABLISHED
:
5096 tcp_data_queue(sk
, skb
);
5101 /* tcp_data could move socket to TIME-WAIT */
5102 if (sk
->sk_state
!= TCP_CLOSE
) {
5103 tcp_data_snd_check(sk
);
5104 tcp_ack_snd_check(sk
);
5114 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5115 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5116 EXPORT_SYMBOL(tcp_parse_options
);
5117 EXPORT_SYMBOL(tcp_rcv_established
);
5118 EXPORT_SYMBOL(tcp_rcv_state_process
);
5119 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);