2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presence of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
75 int sysctl_tcp_timestamps __read_mostly
= 1;
76 int sysctl_tcp_window_scaling __read_mostly
= 1;
77 int sysctl_tcp_sack __read_mostly
= 1;
78 int sysctl_tcp_fack __read_mostly
= 1;
79 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
80 int sysctl_tcp_ecn __read_mostly
;
81 int sysctl_tcp_dsack __read_mostly
= 1;
82 int sysctl_tcp_app_win __read_mostly
= 31;
83 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
85 int sysctl_tcp_stdurg __read_mostly
;
86 int sysctl_tcp_rfc1337 __read_mostly
;
87 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
88 int sysctl_tcp_frto __read_mostly
= 2;
89 int sysctl_tcp_frto_response __read_mostly
;
90 int sysctl_tcp_nometrics_save __read_mostly
;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
93 int sysctl_tcp_abc __read_mostly
;
95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
101 #define FLAG_ECE 0x40 /* ECE in this ACK */
102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
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 D-SACKs */
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 reporting
1031 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1032 * perfectly valid, however, in light of RFC2018 which explicitly states
1033 * that "SACK block MUST reflect the newest segment. Even if the newest
1034 * segment is going to be discarded ...", not that it looks very clever
1035 * in case of head skb. Due to potentional receiver driven attacks, we
1036 * choose to avoid immediate execution of a walk in write queue due to
1037 * reneging and defer head skb's loss recovery to standard loss recovery
1038 * procedure that will eventually trigger (nothing forbids us doing this).
1040 * Implements also blockage to start_seq wrap-around. Problem lies in the
1041 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1042 * there's no guarantee that it will be before snd_nxt (n). The problem
1043 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1046 * <- outs wnd -> <- wrapzone ->
1047 * u e n u_w e_w s n_w
1049 * |<------------+------+----- TCP seqno space --------------+---------->|
1050 * ...-- <2^31 ->| |<--------...
1051 * ...---- >2^31 ------>| |<--------...
1053 * Current code wouldn't be vulnerable but it's better still to discard such
1054 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1055 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1056 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1057 * equal to the ideal case (infinite seqno space without wrap caused issues).
1059 * With D-SACK the lower bound is extended to cover sequence space below
1060 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1061 * again, D-SACK block must not to go across snd_una (for the same reason as
1062 * for the normal SACK blocks, explained above). But there all simplicity
1063 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1064 * fully below undo_marker they do not affect behavior in anyway and can
1065 * therefore be safely ignored. In rare cases (which are more or less
1066 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1067 * fragmentation and packet reordering past skb's retransmission. To consider
1068 * them correctly, the acceptable range must be extended even more though
1069 * the exact amount is rather hard to quantify. However, tp->max_window can
1070 * be used as an exaggerated estimate.
1072 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1073 u32 start_seq
, u32 end_seq
)
1075 /* Too far in future, or reversed (interpretation is ambiguous) */
1076 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1079 /* Nasty start_seq wrap-around check (see comments above) */
1080 if (!before(start_seq
, tp
->snd_nxt
))
1083 /* In outstanding window? ...This is valid exit for D-SACKs too.
1084 * start_seq == snd_una is non-sensical (see comments above)
1086 if (after(start_seq
, tp
->snd_una
))
1089 if (!is_dsack
|| !tp
->undo_marker
)
1092 /* ...Then it's D-SACK, and must reside below snd_una completely */
1093 if (!after(end_seq
, tp
->snd_una
))
1096 if (!before(start_seq
, tp
->undo_marker
))
1100 if (!after(end_seq
, tp
->undo_marker
))
1103 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1104 * start_seq < undo_marker and end_seq >= undo_marker.
1106 return !before(start_seq
, end_seq
- tp
->max_window
);
1109 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1110 * Event "C". Later note: FACK people cheated me again 8), we have to account
1111 * for reordering! Ugly, but should help.
1113 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1114 * less than what is now known to be received by the other end (derived from
1115 * SACK blocks by the caller). Also calculate the lowest snd_nxt among the
1116 * remaining retransmitted skbs to avoid some costly processing per ACKs.
1118 static int tcp_mark_lost_retrans(struct sock
*sk
, u32 received_upto
)
1120 struct tcp_sock
*tp
= tcp_sk(sk
);
1121 struct sk_buff
*skb
;
1124 u32 new_low_seq
= tp
->snd_nxt
;
1126 tcp_for_write_queue(skb
, sk
) {
1127 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1129 if (skb
== tcp_send_head(sk
))
1131 if (cnt
== tp
->retrans_out
)
1133 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1136 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1139 if (after(received_upto
, ack_seq
) &&
1141 !before(received_upto
,
1142 ack_seq
+ tp
->reordering
* tp
->mss_cache
))) {
1143 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1144 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1146 /* clear lost hint */
1147 tp
->retransmit_skb_hint
= NULL
;
1149 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1150 tp
->lost_out
+= tcp_skb_pcount(skb
);
1151 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1152 flag
|= FLAG_DATA_SACKED
;
1153 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1156 if (before(ack_seq
, new_low_seq
))
1157 new_low_seq
= ack_seq
;
1158 cnt
+= tcp_skb_pcount(skb
);
1162 if (tp
->retrans_out
)
1163 tp
->lost_retrans_low
= new_low_seq
;
1168 static int tcp_check_dsack(struct tcp_sock
*tp
, struct sk_buff
*ack_skb
,
1169 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1172 u32 start_seq_0
= ntohl(get_unaligned(&sp
[0].start_seq
));
1173 u32 end_seq_0
= ntohl(get_unaligned(&sp
[0].end_seq
));
1176 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1179 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
1180 } else if (num_sacks
> 1) {
1181 u32 end_seq_1
= ntohl(get_unaligned(&sp
[1].end_seq
));
1182 u32 start_seq_1
= ntohl(get_unaligned(&sp
[1].start_seq
));
1184 if (!after(end_seq_0
, end_seq_1
) &&
1185 !before(start_seq_0
, start_seq_1
)) {
1188 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
1192 /* D-SACK for already forgotten data... Do dumb counting. */
1194 !after(end_seq_0
, prior_snd_una
) &&
1195 after(end_seq_0
, tp
->undo_marker
))
1201 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1202 * the incoming SACK may not exactly match but we can find smaller MSS
1203 * aligned portion of it that matches. Therefore we might need to fragment
1204 * which may fail and creates some hassle (caller must handle error case
1207 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1208 u32 start_seq
, u32 end_seq
)
1211 unsigned int pkt_len
;
1213 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1214 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1216 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1217 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1219 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1222 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1224 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1225 err
= tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
);
1234 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
1236 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1237 struct tcp_sock
*tp
= tcp_sk(sk
);
1238 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1239 TCP_SKB_CB(ack_skb
)->sacked
);
1240 struct tcp_sack_block_wire
*sp
= (struct tcp_sack_block_wire
*)(ptr
+2);
1241 struct sk_buff
*cached_skb
;
1242 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
1243 int reord
= tp
->packets_out
;
1245 u32 highest_sack_end_seq
= tp
->lost_retrans_low
;
1247 int found_dup_sack
= 0;
1248 int cached_fack_count
;
1250 int first_sack_index
;
1253 if (!tp
->sacked_out
) {
1254 if (WARN_ON(tp
->fackets_out
))
1255 tp
->fackets_out
= 0;
1256 tp
->highest_sack
= tp
->snd_una
;
1258 prior_fackets
= tp
->fackets_out
;
1260 found_dup_sack
= tcp_check_dsack(tp
, ack_skb
, sp
,
1261 num_sacks
, prior_snd_una
);
1263 flag
|= FLAG_DSACKING_ACK
;
1265 /* Eliminate too old ACKs, but take into
1266 * account more or less fresh ones, they can
1267 * contain valid SACK info.
1269 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1273 * if the only SACK change is the increase of the end_seq of
1274 * the first block then only apply that SACK block
1275 * and use retrans queue hinting otherwise slowpath */
1277 for (i
= 0; i
< num_sacks
; i
++) {
1278 __be32 start_seq
= sp
[i
].start_seq
;
1279 __be32 end_seq
= sp
[i
].end_seq
;
1282 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
1285 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
1286 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1289 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1290 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1292 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1293 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1294 tp
->recv_sack_cache
[i
].start_seq
= 0;
1295 tp
->recv_sack_cache
[i
].end_seq
= 0;
1298 first_sack_index
= 0;
1303 tp
->fastpath_skb_hint
= NULL
;
1305 /* order SACK blocks to allow in order walk of the retrans queue */
1306 for (i
= num_sacks
-1; i
> 0; i
--) {
1307 for (j
= 0; j
< i
; j
++){
1308 if (after(ntohl(sp
[j
].start_seq
),
1309 ntohl(sp
[j
+1].start_seq
))){
1310 struct tcp_sack_block_wire tmp
;
1316 /* Track where the first SACK block goes to */
1317 if (j
== first_sack_index
)
1318 first_sack_index
= j
+1;
1325 /* Use SACK fastpath hint if valid */
1326 cached_skb
= tp
->fastpath_skb_hint
;
1327 cached_fack_count
= tp
->fastpath_cnt_hint
;
1329 cached_skb
= tcp_write_queue_head(sk
);
1330 cached_fack_count
= 0;
1333 for (i
= 0; i
< num_sacks
; i
++) {
1334 struct sk_buff
*skb
;
1335 __u32 start_seq
= ntohl(sp
->start_seq
);
1336 __u32 end_seq
= ntohl(sp
->end_seq
);
1338 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1339 int next_dup
= (found_dup_sack
&& (i
+1 == first_sack_index
));
1343 if (!tcp_is_sackblock_valid(tp
, dup_sack
, start_seq
, end_seq
)) {
1345 if (!tp
->undo_marker
)
1346 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO
);
1348 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD
);
1350 /* Don't count olds caused by ACK reordering */
1351 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1352 !after(end_seq
, tp
->snd_una
))
1354 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD
);
1360 fack_count
= cached_fack_count
;
1362 /* Event "B" in the comment above. */
1363 if (after(end_seq
, tp
->high_seq
))
1364 flag
|= FLAG_DATA_LOST
;
1366 tcp_for_write_queue_from(skb
, sk
) {
1370 if (skb
== tcp_send_head(sk
))
1374 cached_fack_count
= fack_count
;
1375 if (i
== first_sack_index
) {
1376 tp
->fastpath_skb_hint
= skb
;
1377 tp
->fastpath_cnt_hint
= fack_count
;
1380 /* The retransmission queue is always in order, so
1381 * we can short-circuit the walk early.
1383 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1386 dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1388 /* Due to sorting DSACK may reside within this SACK block! */
1390 u32 dup_start
= ntohl(sp
->start_seq
);
1391 u32 dup_end
= ntohl(sp
->end_seq
);
1393 if (before(TCP_SKB_CB(skb
)->seq
, dup_end
)) {
1394 in_sack
= tcp_match_skb_to_sack(sk
, skb
, dup_start
, dup_end
);
1400 /* DSACK info lost if out-of-mem, try SACK still */
1402 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
, end_seq
);
1403 if (unlikely(in_sack
< 0))
1406 sacked
= TCP_SKB_CB(skb
)->sacked
;
1408 /* Account D-SACK for retransmitted packet. */
1409 if ((dup_sack
&& in_sack
) &&
1410 (sacked
& TCPCB_RETRANS
) &&
1411 after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1414 /* The frame is ACKed. */
1415 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
1416 if (sacked
&TCPCB_RETRANS
) {
1417 if ((dup_sack
&& in_sack
) &&
1418 (sacked
&TCPCB_SACKED_ACKED
))
1419 reord
= min(fack_count
, reord
);
1422 /* Nothing to do; acked frame is about to be dropped. */
1423 fack_count
+= tcp_skb_pcount(skb
);
1428 fack_count
+= tcp_skb_pcount(skb
);
1432 if (!(sacked
&TCPCB_SACKED_ACKED
)) {
1433 if (sacked
& TCPCB_SACKED_RETRANS
) {
1434 /* If the segment is not tagged as lost,
1435 * we do not clear RETRANS, believing
1436 * that retransmission is still in flight.
1438 if (sacked
& TCPCB_LOST
) {
1439 TCP_SKB_CB(skb
)->sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1440 tp
->lost_out
-= tcp_skb_pcount(skb
);
1441 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1443 /* clear lost hint */
1444 tp
->retransmit_skb_hint
= NULL
;
1447 if (!(sacked
& TCPCB_RETRANS
)) {
1448 /* New sack for not retransmitted frame,
1449 * which was in hole. It is reordering.
1451 if (fack_count
< prior_fackets
)
1452 reord
= min(fack_count
, reord
);
1454 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1455 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1456 flag
|= FLAG_ONLY_ORIG_SACKED
;
1459 if (sacked
& TCPCB_LOST
) {
1460 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1461 tp
->lost_out
-= tcp_skb_pcount(skb
);
1463 /* clear lost hint */
1464 tp
->retransmit_skb_hint
= NULL
;
1468 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1469 flag
|= FLAG_DATA_SACKED
;
1470 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1472 fack_count
+= tcp_skb_pcount(skb
);
1473 if (fack_count
> tp
->fackets_out
)
1474 tp
->fackets_out
= fack_count
;
1476 if (after(TCP_SKB_CB(skb
)->seq
, tp
->highest_sack
)) {
1477 tp
->highest_sack
= TCP_SKB_CB(skb
)->seq
;
1478 highest_sack_end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1481 if (dup_sack
&& (sacked
&TCPCB_RETRANS
))
1482 reord
= min(fack_count
, reord
);
1484 fack_count
+= tcp_skb_pcount(skb
);
1487 /* D-SACK. We can detect redundant retransmission
1488 * in S|R and plain R frames and clear it.
1489 * undo_retrans is decreased above, L|R frames
1490 * are accounted above as well.
1493 (TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_RETRANS
)) {
1494 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1495 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1496 tp
->retransmit_skb_hint
= NULL
;
1500 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1501 * due to in-order walk
1503 if (after(end_seq
, tp
->frto_highmark
))
1504 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1507 if (tp
->retrans_out
&&
1508 after(highest_sack_end_seq
, tp
->lost_retrans_low
) &&
1509 icsk
->icsk_ca_state
== TCP_CA_Recovery
)
1510 flag
|= tcp_mark_lost_retrans(sk
, highest_sack_end_seq
);
1512 tcp_verify_left_out(tp
);
1514 if ((reord
< tp
->fackets_out
) && icsk
->icsk_ca_state
!= TCP_CA_Loss
&&
1515 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1516 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1518 #if FASTRETRANS_DEBUG > 0
1519 BUG_TRAP((int)tp
->sacked_out
>= 0);
1520 BUG_TRAP((int)tp
->lost_out
>= 0);
1521 BUG_TRAP((int)tp
->retrans_out
>= 0);
1522 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1527 /* If we receive more dupacks than we expected counting segments
1528 * in assumption of absent reordering, interpret this as reordering.
1529 * The only another reason could be bug in receiver TCP.
1531 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1533 struct tcp_sock
*tp
= tcp_sk(sk
);
1536 holes
= max(tp
->lost_out
, 1U);
1537 holes
= min(holes
, tp
->packets_out
);
1539 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1540 tp
->sacked_out
= tp
->packets_out
- holes
;
1541 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1545 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1547 static void tcp_add_reno_sack(struct sock
*sk
)
1549 struct tcp_sock
*tp
= tcp_sk(sk
);
1551 tcp_check_reno_reordering(sk
, 0);
1552 tcp_verify_left_out(tp
);
1555 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1557 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1559 struct tcp_sock
*tp
= tcp_sk(sk
);
1562 /* One ACK acked hole. The rest eat duplicate ACKs. */
1563 if (acked
-1 >= tp
->sacked_out
)
1566 tp
->sacked_out
-= acked
-1;
1568 tcp_check_reno_reordering(sk
, acked
);
1569 tcp_verify_left_out(tp
);
1572 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1577 /* F-RTO can only be used if TCP has never retransmitted anything other than
1578 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1580 int tcp_use_frto(struct sock
*sk
)
1582 const struct tcp_sock
*tp
= tcp_sk(sk
);
1583 struct sk_buff
*skb
;
1585 if (!sysctl_tcp_frto
)
1591 /* Avoid expensive walking of rexmit queue if possible */
1592 if (tp
->retrans_out
> 1)
1595 skb
= tcp_write_queue_head(sk
);
1596 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1597 tcp_for_write_queue_from(skb
, sk
) {
1598 if (skb
== tcp_send_head(sk
))
1600 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1602 /* Short-circuit when first non-SACKed skb has been checked */
1603 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1609 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1610 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1611 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1612 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1613 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1614 * bits are handled if the Loss state is really to be entered (in
1615 * tcp_enter_frto_loss).
1617 * Do like tcp_enter_loss() would; when RTO expires the second time it
1619 * "Reduce ssthresh if it has not yet been made inside this window."
1621 void tcp_enter_frto(struct sock
*sk
)
1623 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1624 struct tcp_sock
*tp
= tcp_sk(sk
);
1625 struct sk_buff
*skb
;
1627 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1628 tp
->snd_una
== tp
->high_seq
||
1629 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1630 !icsk
->icsk_retransmits
)) {
1631 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1632 /* Our state is too optimistic in ssthresh() call because cwnd
1633 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1634 * recovery has not yet completed. Pattern would be this: RTO,
1635 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1637 * RFC4138 should be more specific on what to do, even though
1638 * RTO is quite unlikely to occur after the first Cumulative ACK
1639 * due to back-off and complexity of triggering events ...
1641 if (tp
->frto_counter
) {
1643 stored_cwnd
= tp
->snd_cwnd
;
1645 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1646 tp
->snd_cwnd
= stored_cwnd
;
1648 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1650 /* ... in theory, cong.control module could do "any tricks" in
1651 * ssthresh(), which means that ca_state, lost bits and lost_out
1652 * counter would have to be faked before the call occurs. We
1653 * consider that too expensive, unlikely and hacky, so modules
1654 * using these in ssthresh() must deal these incompatibility
1655 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1657 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1660 tp
->undo_marker
= tp
->snd_una
;
1661 tp
->undo_retrans
= 0;
1663 skb
= tcp_write_queue_head(sk
);
1664 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1665 tp
->undo_marker
= 0;
1666 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1667 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1668 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1670 tcp_verify_left_out(tp
);
1672 /* Too bad if TCP was application limited */
1673 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1675 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1676 * The last condition is necessary at least in tp->frto_counter case.
1678 if (IsSackFrto() && (tp
->frto_counter
||
1679 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1680 after(tp
->high_seq
, tp
->snd_una
)) {
1681 tp
->frto_highmark
= tp
->high_seq
;
1683 tp
->frto_highmark
= tp
->snd_nxt
;
1685 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1686 tp
->high_seq
= tp
->snd_nxt
;
1687 tp
->frto_counter
= 1;
1690 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1691 * which indicates that we should follow the traditional RTO recovery,
1692 * i.e. mark everything lost and do go-back-N retransmission.
1694 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1696 struct tcp_sock
*tp
= tcp_sk(sk
);
1697 struct sk_buff
*skb
;
1700 tp
->retrans_out
= 0;
1701 if (tcp_is_reno(tp
))
1702 tcp_reset_reno_sack(tp
);
1704 tcp_for_write_queue(skb
, sk
) {
1705 if (skb
== tcp_send_head(sk
))
1708 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1710 * Count the retransmission made on RTO correctly (only when
1711 * waiting for the first ACK and did not get it)...
1713 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1714 /* For some reason this R-bit might get cleared? */
1715 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1716 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1717 /* ...enter this if branch just for the first segment */
1718 flag
|= FLAG_DATA_ACKED
;
1720 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1721 tp
->undo_marker
= 0;
1722 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1725 /* Don't lost mark skbs that were fwd transmitted after RTO */
1726 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) &&
1727 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
)) {
1728 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1729 tp
->lost_out
+= tcp_skb_pcount(skb
);
1732 tcp_verify_left_out(tp
);
1734 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1735 tp
->snd_cwnd_cnt
= 0;
1736 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1737 tp
->frto_counter
= 0;
1738 tp
->bytes_acked
= 0;
1740 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1741 sysctl_tcp_reordering
);
1742 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1743 tp
->high_seq
= tp
->frto_highmark
;
1744 TCP_ECN_queue_cwr(tp
);
1746 tcp_clear_retrans_hints_partial(tp
);
1749 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1751 tp
->retrans_out
= 0;
1754 tp
->undo_marker
= 0;
1755 tp
->undo_retrans
= 0;
1758 void tcp_clear_retrans(struct tcp_sock
*tp
)
1760 tcp_clear_retrans_partial(tp
);
1762 tp
->fackets_out
= 0;
1766 /* Enter Loss state. If "how" is not zero, forget all SACK information
1767 * and reset tags completely, otherwise preserve SACKs. If receiver
1768 * dropped its ofo queue, we will know this due to reneging detection.
1770 void tcp_enter_loss(struct sock
*sk
, int how
)
1772 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1773 struct tcp_sock
*tp
= tcp_sk(sk
);
1774 struct sk_buff
*skb
;
1776 /* Reduce ssthresh if it has not yet been made inside this window. */
1777 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1778 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1779 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1780 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1781 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1784 tp
->snd_cwnd_cnt
= 0;
1785 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1787 tp
->bytes_acked
= 0;
1788 tcp_clear_retrans_partial(tp
);
1790 if (tcp_is_reno(tp
))
1791 tcp_reset_reno_sack(tp
);
1794 /* Push undo marker, if it was plain RTO and nothing
1795 * was retransmitted. */
1796 tp
->undo_marker
= tp
->snd_una
;
1797 tcp_clear_retrans_hints_partial(tp
);
1800 tp
->fackets_out
= 0;
1801 tcp_clear_all_retrans_hints(tp
);
1804 tcp_for_write_queue(skb
, sk
) {
1805 if (skb
== tcp_send_head(sk
))
1808 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1809 tp
->undo_marker
= 0;
1810 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1811 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1812 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1813 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1814 tp
->lost_out
+= tcp_skb_pcount(skb
);
1817 tcp_verify_left_out(tp
);
1819 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1820 sysctl_tcp_reordering
);
1821 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1822 tp
->high_seq
= tp
->snd_nxt
;
1823 TCP_ECN_queue_cwr(tp
);
1824 /* Abort F-RTO algorithm if one is in progress */
1825 tp
->frto_counter
= 0;
1828 static int tcp_check_sack_reneging(struct sock
*sk
)
1830 struct sk_buff
*skb
;
1832 /* If ACK arrived pointing to a remembered SACK,
1833 * it means that our remembered SACKs do not reflect
1834 * real state of receiver i.e.
1835 * receiver _host_ is heavily congested (or buggy).
1836 * Do processing similar to RTO timeout.
1838 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1839 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1840 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1841 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1843 tcp_enter_loss(sk
, 1);
1844 icsk
->icsk_retransmits
++;
1845 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1846 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1847 icsk
->icsk_rto
, TCP_RTO_MAX
);
1853 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1855 return tcp_is_reno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1858 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1860 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1863 static inline int tcp_head_timedout(struct sock
*sk
)
1865 struct tcp_sock
*tp
= tcp_sk(sk
);
1867 return tp
->packets_out
&&
1868 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1871 /* Linux NewReno/SACK/FACK/ECN state machine.
1872 * --------------------------------------
1874 * "Open" Normal state, no dubious events, fast path.
1875 * "Disorder" In all the respects it is "Open",
1876 * but requires a bit more attention. It is entered when
1877 * we see some SACKs or dupacks. It is split of "Open"
1878 * mainly to move some processing from fast path to slow one.
1879 * "CWR" CWND was reduced due to some Congestion Notification event.
1880 * It can be ECN, ICMP source quench, local device congestion.
1881 * "Recovery" CWND was reduced, we are fast-retransmitting.
1882 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1884 * tcp_fastretrans_alert() is entered:
1885 * - each incoming ACK, if state is not "Open"
1886 * - when arrived ACK is unusual, namely:
1891 * Counting packets in flight is pretty simple.
1893 * in_flight = packets_out - left_out + retrans_out
1895 * packets_out is SND.NXT-SND.UNA counted in packets.
1897 * retrans_out is number of retransmitted segments.
1899 * left_out is number of segments left network, but not ACKed yet.
1901 * left_out = sacked_out + lost_out
1903 * sacked_out: Packets, which arrived to receiver out of order
1904 * and hence not ACKed. With SACKs this number is simply
1905 * amount of SACKed data. Even without SACKs
1906 * it is easy to give pretty reliable estimate of this number,
1907 * counting duplicate ACKs.
1909 * lost_out: Packets lost by network. TCP has no explicit
1910 * "loss notification" feedback from network (for now).
1911 * It means that this number can be only _guessed_.
1912 * Actually, it is the heuristics to predict lossage that
1913 * distinguishes different algorithms.
1915 * F.e. after RTO, when all the queue is considered as lost,
1916 * lost_out = packets_out and in_flight = retrans_out.
1918 * Essentially, we have now two algorithms counting
1921 * FACK: It is the simplest heuristics. As soon as we decided
1922 * that something is lost, we decide that _all_ not SACKed
1923 * packets until the most forward SACK are lost. I.e.
1924 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1925 * It is absolutely correct estimate, if network does not reorder
1926 * packets. And it loses any connection to reality when reordering
1927 * takes place. We use FACK by default until reordering
1928 * is suspected on the path to this destination.
1930 * NewReno: when Recovery is entered, we assume that one segment
1931 * is lost (classic Reno). While we are in Recovery and
1932 * a partial ACK arrives, we assume that one more packet
1933 * is lost (NewReno). This heuristics are the same in NewReno
1936 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1937 * deflation etc. CWND is real congestion window, never inflated, changes
1938 * only according to classic VJ rules.
1940 * Really tricky (and requiring careful tuning) part of algorithm
1941 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1942 * The first determines the moment _when_ we should reduce CWND and,
1943 * hence, slow down forward transmission. In fact, it determines the moment
1944 * when we decide that hole is caused by loss, rather than by a reorder.
1946 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1947 * holes, caused by lost packets.
1949 * And the most logically complicated part of algorithm is undo
1950 * heuristics. We detect false retransmits due to both too early
1951 * fast retransmit (reordering) and underestimated RTO, analyzing
1952 * timestamps and D-SACKs. When we detect that some segments were
1953 * retransmitted by mistake and CWND reduction was wrong, we undo
1954 * window reduction and abort recovery phase. This logic is hidden
1955 * inside several functions named tcp_try_undo_<something>.
1958 /* This function decides, when we should leave Disordered state
1959 * and enter Recovery phase, reducing congestion window.
1961 * Main question: may we further continue forward transmission
1962 * with the same cwnd?
1964 static int tcp_time_to_recover(struct sock
*sk
)
1966 struct tcp_sock
*tp
= tcp_sk(sk
);
1969 /* Do not perform any recovery during F-RTO algorithm */
1970 if (tp
->frto_counter
)
1973 /* Trick#1: The loss is proven. */
1977 /* Not-A-Trick#2 : Classic rule... */
1978 if (tcp_fackets_out(tp
) > tp
->reordering
)
1981 /* Trick#3 : when we use RFC2988 timer restart, fast
1982 * retransmit can be triggered by timeout of queue head.
1984 if (tcp_head_timedout(sk
))
1987 /* Trick#4: It is still not OK... But will it be useful to delay
1990 packets_out
= tp
->packets_out
;
1991 if (packets_out
<= tp
->reordering
&&
1992 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1993 !tcp_may_send_now(sk
)) {
1994 /* We have nothing to send. This connection is limited
1995 * either by receiver window or by application.
2003 /* RFC: This is from the original, I doubt that this is necessary at all:
2004 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2005 * retransmitted past LOST markings in the first place? I'm not fully sure
2006 * about undo and end of connection cases, which can cause R without L?
2008 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
2009 struct sk_buff
*skb
)
2011 if ((tp
->retransmit_skb_hint
!= NULL
) &&
2012 before(TCP_SKB_CB(skb
)->seq
,
2013 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
2014 tp
->retransmit_skb_hint
= NULL
;
2017 /* Mark head of queue up as lost. */
2018 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2020 struct tcp_sock
*tp
= tcp_sk(sk
);
2021 struct sk_buff
*skb
;
2024 BUG_TRAP(packets
<= tp
->packets_out
);
2025 if (tp
->lost_skb_hint
) {
2026 skb
= tp
->lost_skb_hint
;
2027 cnt
= tp
->lost_cnt_hint
;
2029 skb
= tcp_write_queue_head(sk
);
2033 tcp_for_write_queue_from(skb
, sk
) {
2034 if (skb
== tcp_send_head(sk
))
2036 /* TODO: do this better */
2037 /* this is not the most efficient way to do this... */
2038 tp
->lost_skb_hint
= skb
;
2039 tp
->lost_cnt_hint
= cnt
;
2040 cnt
+= tcp_skb_pcount(skb
);
2041 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2043 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2044 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2045 tp
->lost_out
+= tcp_skb_pcount(skb
);
2046 tcp_verify_retransmit_hint(tp
, skb
);
2049 tcp_verify_left_out(tp
);
2052 /* Account newly detected lost packet(s) */
2054 static void tcp_update_scoreboard(struct sock
*sk
)
2056 struct tcp_sock
*tp
= tcp_sk(sk
);
2058 if (tcp_is_fack(tp
)) {
2059 int lost
= tp
->fackets_out
- tp
->reordering
;
2062 tcp_mark_head_lost(sk
, lost
);
2064 tcp_mark_head_lost(sk
, 1);
2067 /* New heuristics: it is possible only after we switched
2068 * to restart timer each time when something is ACKed.
2069 * Hence, we can detect timed out packets during fast
2070 * retransmit without falling to slow start.
2072 if (!tcp_is_reno(tp
) && tcp_head_timedout(sk
)) {
2073 struct sk_buff
*skb
;
2075 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2076 : tcp_write_queue_head(sk
);
2078 tcp_for_write_queue_from(skb
, sk
) {
2079 if (skb
== tcp_send_head(sk
))
2081 if (!tcp_skb_timedout(sk
, skb
))
2084 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2085 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2086 tp
->lost_out
+= tcp_skb_pcount(skb
);
2087 tcp_verify_retransmit_hint(tp
, skb
);
2091 tp
->scoreboard_skb_hint
= skb
;
2093 tcp_verify_left_out(tp
);
2097 /* CWND moderation, preventing bursts due to too big ACKs
2098 * in dubious situations.
2100 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2102 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2103 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
2104 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2107 /* Lower bound on congestion window is slow start threshold
2108 * unless congestion avoidance choice decides to overide it.
2110 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2112 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2114 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2117 /* Decrease cwnd each second ack. */
2118 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2120 struct tcp_sock
*tp
= tcp_sk(sk
);
2121 int decr
= tp
->snd_cwnd_cnt
+ 1;
2123 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
2124 (tcp_is_reno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
2125 tp
->snd_cwnd_cnt
= decr
&1;
2128 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2129 tp
->snd_cwnd
-= decr
;
2131 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
2132 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2136 /* Nothing was retransmitted or returned timestamp is less
2137 * than timestamp of the first retransmission.
2139 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2141 return !tp
->retrans_stamp
||
2142 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2143 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
2146 /* Undo procedures. */
2148 #if FASTRETRANS_DEBUG > 1
2149 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2151 struct tcp_sock
*tp
= tcp_sk(sk
);
2152 struct inet_sock
*inet
= inet_sk(sk
);
2154 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2156 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2157 tp
->snd_cwnd
, tcp_left_out(tp
),
2158 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2162 #define DBGUNDO(x...) do { } while (0)
2165 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2167 struct tcp_sock
*tp
= tcp_sk(sk
);
2169 if (tp
->prior_ssthresh
) {
2170 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2172 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2173 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2175 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
2177 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2178 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2179 TCP_ECN_withdraw_cwr(tp
);
2182 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2184 tcp_moderate_cwnd(tp
);
2185 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2187 /* There is something screwy going on with the retrans hints after
2189 tcp_clear_all_retrans_hints(tp
);
2192 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2194 return tp
->undo_marker
&&
2195 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2198 /* People celebrate: "We love our President!" */
2199 static int tcp_try_undo_recovery(struct sock
*sk
)
2201 struct tcp_sock
*tp
= tcp_sk(sk
);
2203 if (tcp_may_undo(tp
)) {
2204 /* Happy end! We did not retransmit anything
2205 * or our original transmission succeeded.
2207 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2208 tcp_undo_cwr(sk
, 1);
2209 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2210 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2212 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
2213 tp
->undo_marker
= 0;
2215 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2216 /* Hold old state until something *above* high_seq
2217 * is ACKed. For Reno it is MUST to prevent false
2218 * fast retransmits (RFC2582). SACK TCP is safe. */
2219 tcp_moderate_cwnd(tp
);
2222 tcp_set_ca_state(sk
, TCP_CA_Open
);
2226 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2227 static void tcp_try_undo_dsack(struct sock
*sk
)
2229 struct tcp_sock
*tp
= tcp_sk(sk
);
2231 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2232 DBGUNDO(sk
, "D-SACK");
2233 tcp_undo_cwr(sk
, 1);
2234 tp
->undo_marker
= 0;
2235 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2239 /* Undo during fast recovery after partial ACK. */
2241 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2243 struct tcp_sock
*tp
= tcp_sk(sk
);
2244 /* Partial ACK arrived. Force Hoe's retransmit. */
2245 int failed
= tcp_is_reno(tp
) || tp
->fackets_out
>tp
->reordering
;
2247 if (tcp_may_undo(tp
)) {
2248 /* Plain luck! Hole if filled with delayed
2249 * packet, rather than with a retransmit.
2251 if (tp
->retrans_out
== 0)
2252 tp
->retrans_stamp
= 0;
2254 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2257 tcp_undo_cwr(sk
, 0);
2258 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2260 /* So... Do not make Hoe's retransmit yet.
2261 * If the first packet was delayed, the rest
2262 * ones are most probably delayed as well.
2269 /* Undo during loss recovery after partial ACK. */
2270 static int tcp_try_undo_loss(struct sock
*sk
)
2272 struct tcp_sock
*tp
= tcp_sk(sk
);
2274 if (tcp_may_undo(tp
)) {
2275 struct sk_buff
*skb
;
2276 tcp_for_write_queue(skb
, sk
) {
2277 if (skb
== tcp_send_head(sk
))
2279 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2282 tcp_clear_all_retrans_hints(tp
);
2284 DBGUNDO(sk
, "partial loss");
2286 tcp_undo_cwr(sk
, 1);
2287 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2288 inet_csk(sk
)->icsk_retransmits
= 0;
2289 tp
->undo_marker
= 0;
2290 if (tcp_is_sack(tp
))
2291 tcp_set_ca_state(sk
, TCP_CA_Open
);
2297 static inline void tcp_complete_cwr(struct sock
*sk
)
2299 struct tcp_sock
*tp
= tcp_sk(sk
);
2300 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2301 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2302 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2305 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2307 struct tcp_sock
*tp
= tcp_sk(sk
);
2309 tcp_verify_left_out(tp
);
2311 if (tp
->retrans_out
== 0)
2312 tp
->retrans_stamp
= 0;
2315 tcp_enter_cwr(sk
, 1);
2317 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2318 int state
= TCP_CA_Open
;
2320 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2321 state
= TCP_CA_Disorder
;
2323 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2324 tcp_set_ca_state(sk
, state
);
2325 tp
->high_seq
= tp
->snd_nxt
;
2327 tcp_moderate_cwnd(tp
);
2329 tcp_cwnd_down(sk
, flag
);
2333 static void tcp_mtup_probe_failed(struct sock
*sk
)
2335 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2337 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2338 icsk
->icsk_mtup
.probe_size
= 0;
2341 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2343 struct tcp_sock
*tp
= tcp_sk(sk
);
2344 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2346 /* FIXME: breaks with very large cwnd */
2347 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2348 tp
->snd_cwnd
= tp
->snd_cwnd
*
2349 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2350 icsk
->icsk_mtup
.probe_size
;
2351 tp
->snd_cwnd_cnt
= 0;
2352 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2353 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2355 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2356 icsk
->icsk_mtup
.probe_size
= 0;
2357 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2361 /* Process an event, which can update packets-in-flight not trivially.
2362 * Main goal of this function is to calculate new estimate for left_out,
2363 * taking into account both packets sitting in receiver's buffer and
2364 * packets lost by network.
2366 * Besides that it does CWND reduction, when packet loss is detected
2367 * and changes state of machine.
2369 * It does _not_ decide what to send, it is made in function
2370 * tcp_xmit_retransmit_queue().
2373 tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2375 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2376 struct tcp_sock
*tp
= tcp_sk(sk
);
2377 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2378 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2379 (tp
->fackets_out
> tp
->reordering
));
2381 /* Some technical things:
2382 * 1. Reno does not count dupacks (sacked_out) automatically. */
2383 if (!tp
->packets_out
)
2386 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2387 tp
->fackets_out
= 0;
2389 /* Now state machine starts.
2390 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2392 tp
->prior_ssthresh
= 0;
2394 /* B. In all the states check for reneging SACKs. */
2395 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2398 /* C. Process data loss notification, provided it is valid. */
2399 if ((flag
&FLAG_DATA_LOST
) &&
2400 before(tp
->snd_una
, tp
->high_seq
) &&
2401 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2402 tp
->fackets_out
> tp
->reordering
) {
2403 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2404 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2407 /* D. Check consistency of the current state. */
2408 tcp_verify_left_out(tp
);
2410 /* E. Check state exit conditions. State can be terminated
2411 * when high_seq is ACKed. */
2412 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2413 BUG_TRAP(tp
->retrans_out
== 0);
2414 tp
->retrans_stamp
= 0;
2415 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2416 switch (icsk
->icsk_ca_state
) {
2418 icsk
->icsk_retransmits
= 0;
2419 if (tcp_try_undo_recovery(sk
))
2424 /* CWR is to be held something *above* high_seq
2425 * is ACKed for CWR bit to reach receiver. */
2426 if (tp
->snd_una
!= tp
->high_seq
) {
2427 tcp_complete_cwr(sk
);
2428 tcp_set_ca_state(sk
, TCP_CA_Open
);
2432 case TCP_CA_Disorder
:
2433 tcp_try_undo_dsack(sk
);
2434 if (!tp
->undo_marker
||
2435 /* For SACK case do not Open to allow to undo
2436 * catching for all duplicate ACKs. */
2437 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2438 tp
->undo_marker
= 0;
2439 tcp_set_ca_state(sk
, TCP_CA_Open
);
2443 case TCP_CA_Recovery
:
2444 if (tcp_is_reno(tp
))
2445 tcp_reset_reno_sack(tp
);
2446 if (tcp_try_undo_recovery(sk
))
2448 tcp_complete_cwr(sk
);
2453 /* F. Process state. */
2454 switch (icsk
->icsk_ca_state
) {
2455 case TCP_CA_Recovery
:
2456 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2457 if (tcp_is_reno(tp
) && is_dupack
)
2458 tcp_add_reno_sack(sk
);
2460 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2463 if (flag
&FLAG_DATA_ACKED
)
2464 icsk
->icsk_retransmits
= 0;
2465 if (!tcp_try_undo_loss(sk
)) {
2466 tcp_moderate_cwnd(tp
);
2467 tcp_xmit_retransmit_queue(sk
);
2470 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2472 /* Loss is undone; fall through to processing in Open state. */
2474 if (tcp_is_reno(tp
)) {
2475 if (flag
& FLAG_SND_UNA_ADVANCED
)
2476 tcp_reset_reno_sack(tp
);
2478 tcp_add_reno_sack(sk
);
2481 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2482 tcp_try_undo_dsack(sk
);
2484 if (!tcp_time_to_recover(sk
)) {
2485 tcp_try_to_open(sk
, flag
);
2489 /* MTU probe failure: don't reduce cwnd */
2490 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2491 icsk
->icsk_mtup
.probe_size
&&
2492 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2493 tcp_mtup_probe_failed(sk
);
2494 /* Restores the reduction we did in tcp_mtup_probe() */
2496 tcp_simple_retransmit(sk
);
2500 /* Otherwise enter Recovery state */
2502 if (tcp_is_reno(tp
))
2503 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2505 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2507 tp
->high_seq
= tp
->snd_nxt
;
2508 tp
->prior_ssthresh
= 0;
2509 tp
->undo_marker
= tp
->snd_una
;
2510 tp
->undo_retrans
= tp
->retrans_out
;
2512 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2513 if (!(flag
&FLAG_ECE
))
2514 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2515 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2516 TCP_ECN_queue_cwr(tp
);
2519 tp
->bytes_acked
= 0;
2520 tp
->snd_cwnd_cnt
= 0;
2521 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2524 if (do_lost
|| tcp_head_timedout(sk
))
2525 tcp_update_scoreboard(sk
);
2526 tcp_cwnd_down(sk
, flag
);
2527 tcp_xmit_retransmit_queue(sk
);
2530 /* Read draft-ietf-tcplw-high-performance before mucking
2531 * with this code. (Supersedes RFC1323)
2533 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2535 /* RTTM Rule: A TSecr value received in a segment is used to
2536 * update the averaged RTT measurement only if the segment
2537 * acknowledges some new data, i.e., only if it advances the
2538 * left edge of the send window.
2540 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2541 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2543 * Changed: reset backoff as soon as we see the first valid sample.
2544 * If we do not, we get strongly overestimated rto. With timestamps
2545 * samples are accepted even from very old segments: f.e., when rtt=1
2546 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2547 * answer arrives rto becomes 120 seconds! If at least one of segments
2548 * in window is lost... Voila. --ANK (010210)
2550 struct tcp_sock
*tp
= tcp_sk(sk
);
2551 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2552 tcp_rtt_estimator(sk
, seq_rtt
);
2554 inet_csk(sk
)->icsk_backoff
= 0;
2558 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2560 /* We don't have a timestamp. Can only use
2561 * packets that are not retransmitted to determine
2562 * rtt estimates. Also, we must not reset the
2563 * backoff for rto until we get a non-retransmitted
2564 * packet. This allows us to deal with a situation
2565 * where the network delay has increased suddenly.
2566 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2569 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2572 tcp_rtt_estimator(sk
, seq_rtt
);
2574 inet_csk(sk
)->icsk_backoff
= 0;
2578 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2581 const struct tcp_sock
*tp
= tcp_sk(sk
);
2582 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2583 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2584 tcp_ack_saw_tstamp(sk
, flag
);
2585 else if (seq_rtt
>= 0)
2586 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2589 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
,
2590 u32 in_flight
, int good
)
2592 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2593 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
, good
);
2594 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2597 /* Restart timer after forward progress on connection.
2598 * RFC2988 recommends to restart timer to now+rto.
2600 static void tcp_rearm_rto(struct sock
*sk
)
2602 struct tcp_sock
*tp
= tcp_sk(sk
);
2604 if (!tp
->packets_out
) {
2605 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2607 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2611 /* If we get here, the whole TSO packet has not been acked. */
2612 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2614 struct tcp_sock
*tp
= tcp_sk(sk
);
2617 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2619 packets_acked
= tcp_skb_pcount(skb
);
2620 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2622 packets_acked
-= tcp_skb_pcount(skb
);
2624 if (packets_acked
) {
2625 BUG_ON(tcp_skb_pcount(skb
) == 0);
2626 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2629 return packets_acked
;
2632 /* Remove acknowledged frames from the retransmission queue. If our packet
2633 * is before the ack sequence we can discard it as it's confirmed to have
2634 * arrived at the other end.
2636 static int tcp_clean_rtx_queue(struct sock
*sk
, s32
*seq_rtt_p
,
2639 struct tcp_sock
*tp
= tcp_sk(sk
);
2640 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2641 struct sk_buff
*skb
;
2642 u32 now
= tcp_time_stamp
;
2643 int fully_acked
= 1;
2645 int prior_packets
= tp
->packets_out
;
2647 u32 reord
= tp
->packets_out
;
2649 ktime_t last_ackt
= net_invalid_timestamp();
2651 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2652 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2655 u8 sacked
= scb
->sacked
;
2657 if (after(scb
->end_seq
, tp
->snd_una
)) {
2658 if (tcp_skb_pcount(skb
) == 1 ||
2659 !after(tp
->snd_una
, scb
->seq
))
2662 packets_acked
= tcp_tso_acked(sk
, skb
);
2667 end_seq
= tp
->snd_una
;
2669 packets_acked
= tcp_skb_pcount(skb
);
2670 end_seq
= scb
->end_seq
;
2673 /* MTU probing checks */
2674 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2675 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2676 tcp_mtup_probe_success(sk
, skb
);
2680 if (sacked
& TCPCB_RETRANS
) {
2681 if (sacked
& TCPCB_SACKED_RETRANS
)
2682 tp
->retrans_out
-= packets_acked
;
2683 flag
|= FLAG_RETRANS_DATA_ACKED
;
2685 if ((flag
& FLAG_DATA_ACKED
) ||
2686 (packets_acked
> 1))
2687 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2690 seq_rtt
= now
- scb
->when
;
2692 last_ackt
= skb
->tstamp
;
2694 if (!(sacked
& TCPCB_SACKED_ACKED
))
2695 reord
= min(cnt
, reord
);
2698 if (sacked
& TCPCB_SACKED_ACKED
)
2699 tp
->sacked_out
-= packets_acked
;
2700 if (sacked
& TCPCB_LOST
)
2701 tp
->lost_out
-= packets_acked
;
2703 if ((sacked
& TCPCB_URG
) && tp
->urg_mode
&&
2704 !before(end_seq
, tp
->snd_up
))
2708 seq_rtt
= now
- scb
->when
;
2710 last_ackt
= skb
->tstamp
;
2712 reord
= min(cnt
, reord
);
2714 tp
->packets_out
-= packets_acked
;
2715 cnt
+= packets_acked
;
2717 /* Initial outgoing SYN's get put onto the write_queue
2718 * just like anything else we transmit. It is not
2719 * true data, and if we misinform our callers that
2720 * this ACK acks real data, we will erroneously exit
2721 * connection startup slow start one packet too
2722 * quickly. This is severely frowned upon behavior.
2724 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2725 flag
|= FLAG_DATA_ACKED
;
2727 flag
|= FLAG_SYN_ACKED
;
2728 tp
->retrans_stamp
= 0;
2734 tcp_unlink_write_queue(skb
, sk
);
2735 sk_stream_free_skb(sk
, skb
);
2736 tcp_clear_all_retrans_hints(tp
);
2739 if (flag
& FLAG_ACKED
) {
2740 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2741 const struct tcp_congestion_ops
*ca_ops
2742 = inet_csk(sk
)->icsk_ca_ops
;
2744 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2747 if (tcp_is_reno(tp
)) {
2748 tcp_remove_reno_sacks(sk
, pkts_acked
);
2750 /* Non-retransmitted hole got filled? That's reordering */
2751 if (reord
< prior_fackets
)
2752 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2755 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2756 /* hint's skb might be NULL but we don't need to care */
2757 tp
->fastpath_cnt_hint
-= min_t(u32
, pkts_acked
,
2758 tp
->fastpath_cnt_hint
);
2759 if (ca_ops
->pkts_acked
) {
2762 /* Is the ACK triggering packet unambiguous? */
2763 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2764 /* High resolution needed and available? */
2765 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2766 !ktime_equal(last_ackt
,
2767 net_invalid_timestamp()))
2768 rtt_us
= ktime_us_delta(ktime_get_real(),
2770 else if (seq_rtt
> 0)
2771 rtt_us
= jiffies_to_usecs(seq_rtt
);
2774 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2778 #if FASTRETRANS_DEBUG > 0
2779 BUG_TRAP((int)tp
->sacked_out
>= 0);
2780 BUG_TRAP((int)tp
->lost_out
>= 0);
2781 BUG_TRAP((int)tp
->retrans_out
>= 0);
2782 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2783 icsk
= inet_csk(sk
);
2785 printk(KERN_DEBUG
"Leak l=%u %d\n",
2786 tp
->lost_out
, icsk
->icsk_ca_state
);
2789 if (tp
->sacked_out
) {
2790 printk(KERN_DEBUG
"Leak s=%u %d\n",
2791 tp
->sacked_out
, icsk
->icsk_ca_state
);
2794 if (tp
->retrans_out
) {
2795 printk(KERN_DEBUG
"Leak r=%u %d\n",
2796 tp
->retrans_out
, icsk
->icsk_ca_state
);
2797 tp
->retrans_out
= 0;
2801 *seq_rtt_p
= seq_rtt
;
2805 static void tcp_ack_probe(struct sock
*sk
)
2807 const struct tcp_sock
*tp
= tcp_sk(sk
);
2808 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2810 /* Was it a usable window open? */
2812 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2813 tp
->snd_una
+ tp
->snd_wnd
)) {
2814 icsk
->icsk_backoff
= 0;
2815 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2816 /* Socket must be waked up by subsequent tcp_data_snd_check().
2817 * This function is not for random using!
2820 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2821 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2826 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2828 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2829 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2832 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2834 const struct tcp_sock
*tp
= tcp_sk(sk
);
2835 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2836 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2839 /* Check that window update is acceptable.
2840 * The function assumes that snd_una<=ack<=snd_next.
2842 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2843 const u32 ack_seq
, const u32 nwin
)
2845 return (after(ack
, tp
->snd_una
) ||
2846 after(ack_seq
, tp
->snd_wl1
) ||
2847 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2850 /* Update our send window.
2852 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2853 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2855 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2858 struct tcp_sock
*tp
= tcp_sk(sk
);
2860 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2862 if (likely(!tcp_hdr(skb
)->syn
))
2863 nwin
<<= tp
->rx_opt
.snd_wscale
;
2865 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2866 flag
|= FLAG_WIN_UPDATE
;
2867 tcp_update_wl(tp
, ack
, ack_seq
);
2869 if (tp
->snd_wnd
!= nwin
) {
2872 /* Note, it is the only place, where
2873 * fast path is recovered for sending TCP.
2876 tcp_fast_path_check(sk
);
2878 if (nwin
> tp
->max_window
) {
2879 tp
->max_window
= nwin
;
2880 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2890 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2891 * continue in congestion avoidance.
2893 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2895 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2896 tp
->snd_cwnd_cnt
= 0;
2897 tp
->bytes_acked
= 0;
2898 TCP_ECN_queue_cwr(tp
);
2899 tcp_moderate_cwnd(tp
);
2902 /* A conservative spurious RTO response algorithm: reduce cwnd using
2903 * rate halving and continue in congestion avoidance.
2905 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
2907 tcp_enter_cwr(sk
, 0);
2910 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
2913 tcp_ratehalving_spur_to_response(sk
);
2915 tcp_undo_cwr(sk
, 1);
2918 /* F-RTO spurious RTO detection algorithm (RFC4138)
2920 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2921 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2922 * window (but not to or beyond highest sequence sent before RTO):
2923 * On First ACK, send two new segments out.
2924 * On Second ACK, RTO was likely spurious. Do spurious response (response
2925 * algorithm is not part of the F-RTO detection algorithm
2926 * given in RFC4138 but can be selected separately).
2927 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2928 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2929 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2930 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2932 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2933 * original window even after we transmit two new data segments.
2936 * on first step, wait until first cumulative ACK arrives, then move to
2937 * the second step. In second step, the next ACK decides.
2939 * F-RTO is implemented (mainly) in four functions:
2940 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2941 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2942 * called when tcp_use_frto() showed green light
2943 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2944 * - tcp_enter_frto_loss() is called if there is not enough evidence
2945 * to prove that the RTO is indeed spurious. It transfers the control
2946 * from F-RTO to the conventional RTO recovery
2948 static int tcp_process_frto(struct sock
*sk
, int flag
)
2950 struct tcp_sock
*tp
= tcp_sk(sk
);
2952 tcp_verify_left_out(tp
);
2954 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2955 if (flag
&FLAG_DATA_ACKED
)
2956 inet_csk(sk
)->icsk_retransmits
= 0;
2958 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
2959 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
2960 tp
->undo_marker
= 0;
2962 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2963 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
2967 if (!IsSackFrto() || tcp_is_reno(tp
)) {
2968 /* RFC4138 shortcoming in step 2; should also have case c):
2969 * ACK isn't duplicate nor advances window, e.g., opposite dir
2972 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2975 if (!(flag
&FLAG_DATA_ACKED
)) {
2976 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2981 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2982 /* Prevent sending of new data. */
2983 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2984 tcp_packets_in_flight(tp
));
2988 if ((tp
->frto_counter
>= 2) &&
2989 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2990 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2991 /* RFC4138 shortcoming (see comment above) */
2992 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2995 tcp_enter_frto_loss(sk
, 3, flag
);
3000 if (tp
->frto_counter
== 1) {
3001 /* Sending of the next skb must be allowed or no F-RTO */
3002 if (!tcp_send_head(sk
) ||
3003 after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
3004 tp
->snd_una
+ tp
->snd_wnd
)) {
3005 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3),
3010 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3011 tp
->frto_counter
= 2;
3014 switch (sysctl_tcp_frto_response
) {
3016 tcp_undo_spur_to_response(sk
, flag
);
3019 tcp_conservative_spur_to_response(tp
);
3022 tcp_ratehalving_spur_to_response(sk
);
3025 tp
->frto_counter
= 0;
3026 tp
->undo_marker
= 0;
3027 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS
);
3032 /* This routine deals with incoming acks, but not outgoing ones. */
3033 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3035 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3036 struct tcp_sock
*tp
= tcp_sk(sk
);
3037 u32 prior_snd_una
= tp
->snd_una
;
3038 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3039 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3040 u32 prior_in_flight
;
3046 /* If the ack is newer than sent or older than previous acks
3047 * then we can probably ignore it.
3049 if (after(ack
, tp
->snd_nxt
))
3050 goto uninteresting_ack
;
3052 if (before(ack
, prior_snd_una
))
3055 if (after(ack
, prior_snd_una
))
3056 flag
|= FLAG_SND_UNA_ADVANCED
;
3058 if (sysctl_tcp_abc
) {
3059 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3060 tp
->bytes_acked
+= ack
- prior_snd_una
;
3061 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3062 /* we assume just one segment left network */
3063 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
3066 prior_fackets
= tp
->fackets_out
;
3068 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3069 /* Window is constant, pure forward advance.
3070 * No more checks are required.
3071 * Note, we use the fact that SND.UNA>=SND.WL2.
3073 tcp_update_wl(tp
, ack
, ack_seq
);
3075 flag
|= FLAG_WIN_UPDATE
;
3077 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3079 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
3081 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3084 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
3086 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3088 if (TCP_SKB_CB(skb
)->sacked
)
3089 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3091 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3094 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3097 /* We passed data and got it acked, remove any soft error
3098 * log. Something worked...
3100 sk
->sk_err_soft
= 0;
3101 tp
->rcv_tstamp
= tcp_time_stamp
;
3102 prior_packets
= tp
->packets_out
;
3106 prior_in_flight
= tcp_packets_in_flight(tp
);
3108 /* See if we can take anything off of the retransmit queue. */
3109 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
, prior_fackets
);
3111 /* Guarantee sacktag reordering detection against wrap-arounds */
3112 if (before(tp
->frto_highmark
, tp
->snd_una
))
3113 tp
->frto_highmark
= 0;
3114 if (tp
->frto_counter
)
3115 frto_cwnd
= tcp_process_frto(sk
, flag
);
3117 if (tcp_ack_is_dubious(sk
, flag
)) {
3118 /* Advance CWND, if state allows this. */
3119 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3120 tcp_may_raise_cwnd(sk
, flag
))
3121 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 0);
3122 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
, flag
);
3124 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3125 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 1);
3128 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
3129 dst_confirm(sk
->sk_dst_cache
);
3134 icsk
->icsk_probes_out
= 0;
3136 /* If this ack opens up a zero window, clear backoff. It was
3137 * being used to time the probes, and is probably far higher than
3138 * it needs to be for normal retransmission.
3140 if (tcp_send_head(sk
))
3145 if (TCP_SKB_CB(skb
)->sacked
)
3146 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3149 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3154 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3155 * But, this can also be called on packets in the established flow when
3156 * the fast version below fails.
3158 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
3161 struct tcphdr
*th
= tcp_hdr(skb
);
3162 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
3164 ptr
= (unsigned char *)(th
+ 1);
3165 opt_rx
->saw_tstamp
= 0;
3167 while (length
> 0) {
3174 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3179 if (opsize
< 2) /* "silly options" */
3181 if (opsize
> length
)
3182 return; /* don't parse partial options */
3185 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3186 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
3188 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
3189 in_mss
= opt_rx
->user_mss
;
3190 opt_rx
->mss_clamp
= in_mss
;
3195 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
3196 if (sysctl_tcp_window_scaling
) {
3197 __u8 snd_wscale
= *(__u8
*) ptr
;
3198 opt_rx
->wscale_ok
= 1;
3199 if (snd_wscale
> 14) {
3200 if (net_ratelimit())
3201 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3202 "scaling value %d >14 received.\n",
3206 opt_rx
->snd_wscale
= snd_wscale
;
3209 case TCPOPT_TIMESTAMP
:
3210 if (opsize
==TCPOLEN_TIMESTAMP
) {
3211 if ((estab
&& opt_rx
->tstamp_ok
) ||
3212 (!estab
&& sysctl_tcp_timestamps
)) {
3213 opt_rx
->saw_tstamp
= 1;
3214 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
3215 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
3219 case TCPOPT_SACK_PERM
:
3220 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
3221 if (sysctl_tcp_sack
) {
3222 opt_rx
->sack_ok
= 1;
3223 tcp_sack_reset(opt_rx
);
3229 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3230 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3232 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3235 #ifdef CONFIG_TCP_MD5SIG
3238 * The MD5 Hash has already been
3239 * checked (see tcp_v{4,6}_do_rcv()).
3251 /* Fast parse options. This hopes to only see timestamps.
3252 * If it is wrong it falls back on tcp_parse_options().
3254 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3255 struct tcp_sock
*tp
)
3257 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3258 tp
->rx_opt
.saw_tstamp
= 0;
3260 } else if (tp
->rx_opt
.tstamp_ok
&&
3261 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3262 __be32
*ptr
= (__be32
*)(th
+ 1);
3263 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3264 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3265 tp
->rx_opt
.saw_tstamp
= 1;
3267 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3269 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3273 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3277 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3279 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3280 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3283 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3285 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3286 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3287 * extra check below makes sure this can only happen
3288 * for pure ACK frames. -DaveM
3290 * Not only, also it occurs for expired timestamps.
3293 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3294 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3295 tcp_store_ts_recent(tp
);
3299 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3301 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3302 * it can pass through stack. So, the following predicate verifies that
3303 * this segment is not used for anything but congestion avoidance or
3304 * fast retransmit. Moreover, we even are able to eliminate most of such
3305 * second order effects, if we apply some small "replay" window (~RTO)
3306 * to timestamp space.
3308 * All these measures still do not guarantee that we reject wrapped ACKs
3309 * on networks with high bandwidth, when sequence space is recycled fastly,
3310 * but it guarantees that such events will be very rare and do not affect
3311 * connection seriously. This doesn't look nice, but alas, PAWS is really
3314 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3315 * states that events when retransmit arrives after original data are rare.
3316 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3317 * the biggest problem on large power networks even with minor reordering.
3318 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3319 * up to bandwidth of 18Gigabit/sec. 8) ]
3322 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3324 struct tcp_sock
*tp
= tcp_sk(sk
);
3325 struct tcphdr
*th
= tcp_hdr(skb
);
3326 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3327 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3329 return (/* 1. Pure ACK with correct sequence number. */
3330 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3332 /* 2. ... and duplicate ACK. */
3333 ack
== tp
->snd_una
&&
3335 /* 3. ... and does not update window. */
3336 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3338 /* 4. ... and sits in replay window. */
3339 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3342 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3344 const struct tcp_sock
*tp
= tcp_sk(sk
);
3345 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3346 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3347 !tcp_disordered_ack(sk
, skb
));
3350 /* Check segment sequence number for validity.
3352 * Segment controls are considered valid, if the segment
3353 * fits to the window after truncation to the window. Acceptability
3354 * of data (and SYN, FIN, of course) is checked separately.
3355 * See tcp_data_queue(), for example.
3357 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3358 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3359 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3360 * (borrowed from freebsd)
3363 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3365 return !before(end_seq
, tp
->rcv_wup
) &&
3366 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3369 /* When we get a reset we do this. */
3370 static void tcp_reset(struct sock
*sk
)
3372 /* We want the right error as BSD sees it (and indeed as we do). */
3373 switch (sk
->sk_state
) {
3375 sk
->sk_err
= ECONNREFUSED
;
3377 case TCP_CLOSE_WAIT
:
3383 sk
->sk_err
= ECONNRESET
;
3386 if (!sock_flag(sk
, SOCK_DEAD
))
3387 sk
->sk_error_report(sk
);
3393 * Process the FIN bit. This now behaves as it is supposed to work
3394 * and the FIN takes effect when it is validly part of sequence
3395 * space. Not before when we get holes.
3397 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3398 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3401 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3402 * close and we go into CLOSING (and later onto TIME-WAIT)
3404 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3406 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3408 struct tcp_sock
*tp
= tcp_sk(sk
);
3410 inet_csk_schedule_ack(sk
);
3412 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3413 sock_set_flag(sk
, SOCK_DONE
);
3415 switch (sk
->sk_state
) {
3417 case TCP_ESTABLISHED
:
3418 /* Move to CLOSE_WAIT */
3419 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3420 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3423 case TCP_CLOSE_WAIT
:
3425 /* Received a retransmission of the FIN, do
3430 /* RFC793: Remain in the LAST-ACK state. */
3434 /* This case occurs when a simultaneous close
3435 * happens, we must ack the received FIN and
3436 * enter the CLOSING state.
3439 tcp_set_state(sk
, TCP_CLOSING
);
3442 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3444 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3447 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3448 * cases we should never reach this piece of code.
3450 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3451 __FUNCTION__
, sk
->sk_state
);
3455 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3456 * Probably, we should reset in this case. For now drop them.
3458 __skb_queue_purge(&tp
->out_of_order_queue
);
3459 if (tcp_is_sack(tp
))
3460 tcp_sack_reset(&tp
->rx_opt
);
3461 sk_stream_mem_reclaim(sk
);
3463 if (!sock_flag(sk
, SOCK_DEAD
)) {
3464 sk
->sk_state_change(sk
);
3466 /* Do not send POLL_HUP for half duplex close. */
3467 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3468 sk
->sk_state
== TCP_CLOSE
)
3469 sk_wake_async(sk
, 1, POLL_HUP
);
3471 sk_wake_async(sk
, 1, POLL_IN
);
3475 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3477 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3478 if (before(seq
, sp
->start_seq
))
3479 sp
->start_seq
= seq
;
3480 if (after(end_seq
, sp
->end_seq
))
3481 sp
->end_seq
= end_seq
;
3487 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3489 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3490 if (before(seq
, tp
->rcv_nxt
))
3491 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3493 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3495 tp
->rx_opt
.dsack
= 1;
3496 tp
->duplicate_sack
[0].start_seq
= seq
;
3497 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3498 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3502 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3504 if (!tp
->rx_opt
.dsack
)
3505 tcp_dsack_set(tp
, seq
, end_seq
);
3507 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3510 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3512 struct tcp_sock
*tp
= tcp_sk(sk
);
3514 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3515 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3516 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3517 tcp_enter_quickack_mode(sk
);
3519 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3520 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3522 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3523 end_seq
= tp
->rcv_nxt
;
3524 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3531 /* These routines update the SACK block as out-of-order packets arrive or
3532 * in-order packets close up the sequence space.
3534 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3537 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3538 struct tcp_sack_block
*swalk
= sp
+1;
3540 /* See if the recent change to the first SACK eats into
3541 * or hits the sequence space of other SACK blocks, if so coalesce.
3543 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3544 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3547 /* Zap SWALK, by moving every further SACK up by one slot.
3548 * Decrease num_sacks.
3550 tp
->rx_opt
.num_sacks
--;
3551 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3552 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3556 this_sack
++, swalk
++;
3560 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3564 tmp
= sack1
->start_seq
;
3565 sack1
->start_seq
= sack2
->start_seq
;
3566 sack2
->start_seq
= tmp
;
3568 tmp
= sack1
->end_seq
;
3569 sack1
->end_seq
= sack2
->end_seq
;
3570 sack2
->end_seq
= tmp
;
3573 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3575 struct tcp_sock
*tp
= tcp_sk(sk
);
3576 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3577 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3583 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3584 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3585 /* Rotate this_sack to the first one. */
3586 for (; this_sack
>0; this_sack
--, sp
--)
3587 tcp_sack_swap(sp
, sp
-1);
3589 tcp_sack_maybe_coalesce(tp
);
3594 /* Could not find an adjacent existing SACK, build a new one,
3595 * put it at the front, and shift everyone else down. We
3596 * always know there is at least one SACK present already here.
3598 * If the sack array is full, forget about the last one.
3600 if (this_sack
>= 4) {
3602 tp
->rx_opt
.num_sacks
--;
3605 for (; this_sack
> 0; this_sack
--, sp
--)
3609 /* Build the new head SACK, and we're done. */
3610 sp
->start_seq
= seq
;
3611 sp
->end_seq
= end_seq
;
3612 tp
->rx_opt
.num_sacks
++;
3613 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3616 /* RCV.NXT advances, some SACKs should be eaten. */
3618 static void tcp_sack_remove(struct tcp_sock
*tp
)
3620 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3621 int num_sacks
= tp
->rx_opt
.num_sacks
;
3624 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3625 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3626 tp
->rx_opt
.num_sacks
= 0;
3627 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3631 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3632 /* Check if the start of the sack is covered by RCV.NXT. */
3633 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3636 /* RCV.NXT must cover all the block! */
3637 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3639 /* Zap this SACK, by moving forward any other SACKS. */
3640 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3641 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3648 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3649 tp
->rx_opt
.num_sacks
= num_sacks
;
3650 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3654 /* This one checks to see if we can put data from the
3655 * out_of_order queue into the receive_queue.
3657 static void tcp_ofo_queue(struct sock
*sk
)
3659 struct tcp_sock
*tp
= tcp_sk(sk
);
3660 __u32 dsack_high
= tp
->rcv_nxt
;
3661 struct sk_buff
*skb
;
3663 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3664 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3667 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3668 __u32 dsack
= dsack_high
;
3669 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3670 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3671 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3674 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3675 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3676 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3680 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3681 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3682 TCP_SKB_CB(skb
)->end_seq
);
3684 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3685 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3686 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3687 if (tcp_hdr(skb
)->fin
)
3688 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3692 static int tcp_prune_queue(struct sock
*sk
);
3694 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3696 struct tcphdr
*th
= tcp_hdr(skb
);
3697 struct tcp_sock
*tp
= tcp_sk(sk
);
3700 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3703 __skb_pull(skb
, th
->doff
*4);
3705 TCP_ECN_accept_cwr(tp
, skb
);
3707 if (tp
->rx_opt
.dsack
) {
3708 tp
->rx_opt
.dsack
= 0;
3709 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3710 4 - tp
->rx_opt
.tstamp_ok
);
3713 /* Queue data for delivery to the user.
3714 * Packets in sequence go to the receive queue.
3715 * Out of sequence packets to the out_of_order_queue.
3717 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3718 if (tcp_receive_window(tp
) == 0)
3721 /* Ok. In sequence. In window. */
3722 if (tp
->ucopy
.task
== current
&&
3723 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3724 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3725 int chunk
= min_t(unsigned int, skb
->len
,
3728 __set_current_state(TASK_RUNNING
);
3731 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3732 tp
->ucopy
.len
-= chunk
;
3733 tp
->copied_seq
+= chunk
;
3734 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3735 tcp_rcv_space_adjust(sk
);
3743 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3744 !sk_stream_rmem_schedule(sk
, skb
))) {
3745 if (tcp_prune_queue(sk
) < 0 ||
3746 !sk_stream_rmem_schedule(sk
, skb
))
3749 sk_stream_set_owner_r(skb
, sk
);
3750 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3752 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3754 tcp_event_data_recv(sk
, skb
);
3756 tcp_fin(skb
, sk
, th
);
3758 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3761 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3762 * gap in queue is filled.
3764 if (skb_queue_empty(&tp
->out_of_order_queue
))
3765 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3768 if (tp
->rx_opt
.num_sacks
)
3769 tcp_sack_remove(tp
);
3771 tcp_fast_path_check(sk
);
3775 else if (!sock_flag(sk
, SOCK_DEAD
))
3776 sk
->sk_data_ready(sk
, 0);
3780 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3781 /* A retransmit, 2nd most common case. Force an immediate ack. */
3782 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3783 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3786 tcp_enter_quickack_mode(sk
);
3787 inet_csk_schedule_ack(sk
);
3793 /* Out of window. F.e. zero window probe. */
3794 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3797 tcp_enter_quickack_mode(sk
);
3799 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3800 /* Partial packet, seq < rcv_next < end_seq */
3801 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3802 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3803 TCP_SKB_CB(skb
)->end_seq
);
3805 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3807 /* If window is closed, drop tail of packet. But after
3808 * remembering D-SACK for its head made in previous line.
3810 if (!tcp_receive_window(tp
))
3815 TCP_ECN_check_ce(tp
, skb
);
3817 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3818 !sk_stream_rmem_schedule(sk
, skb
)) {
3819 if (tcp_prune_queue(sk
) < 0 ||
3820 !sk_stream_rmem_schedule(sk
, skb
))
3824 /* Disable header prediction. */
3826 inet_csk_schedule_ack(sk
);
3828 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3829 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3831 sk_stream_set_owner_r(skb
, sk
);
3833 if (!skb_peek(&tp
->out_of_order_queue
)) {
3834 /* Initial out of order segment, build 1 SACK. */
3835 if (tcp_is_sack(tp
)) {
3836 tp
->rx_opt
.num_sacks
= 1;
3837 tp
->rx_opt
.dsack
= 0;
3838 tp
->rx_opt
.eff_sacks
= 1;
3839 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3840 tp
->selective_acks
[0].end_seq
=
3841 TCP_SKB_CB(skb
)->end_seq
;
3843 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3845 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3846 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3847 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3849 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3850 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3852 if (!tp
->rx_opt
.num_sacks
||
3853 tp
->selective_acks
[0].end_seq
!= seq
)
3856 /* Common case: data arrive in order after hole. */
3857 tp
->selective_acks
[0].end_seq
= end_seq
;
3861 /* Find place to insert this segment. */
3863 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3865 } while ((skb1
= skb1
->prev
) !=
3866 (struct sk_buff
*)&tp
->out_of_order_queue
);
3868 /* Do skb overlap to previous one? */
3869 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3870 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3871 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3872 /* All the bits are present. Drop. */
3874 tcp_dsack_set(tp
, seq
, end_seq
);
3877 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3878 /* Partial overlap. */
3879 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3884 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3886 /* And clean segments covered by new one as whole. */
3887 while ((skb1
= skb
->next
) !=
3888 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3889 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3890 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3891 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3894 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3895 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3900 if (tcp_is_sack(tp
))
3901 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3905 /* Collapse contiguous sequence of skbs head..tail with
3906 * sequence numbers start..end.
3907 * Segments with FIN/SYN are not collapsed (only because this
3911 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3912 struct sk_buff
*head
, struct sk_buff
*tail
,
3915 struct sk_buff
*skb
;
3917 /* First, check that queue is collapsible and find
3918 * the point where collapsing can be useful. */
3919 for (skb
= head
; skb
!= tail
; ) {
3920 /* No new bits? It is possible on ofo queue. */
3921 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3922 struct sk_buff
*next
= skb
->next
;
3923 __skb_unlink(skb
, list
);
3925 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3930 /* The first skb to collapse is:
3932 * - bloated or contains data before "start" or
3933 * overlaps to the next one.
3935 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
3936 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3937 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3938 (skb
->next
!= tail
&&
3939 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3942 /* Decided to skip this, advance start seq. */
3943 start
= TCP_SKB_CB(skb
)->end_seq
;
3946 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
3949 while (before(start
, end
)) {
3950 struct sk_buff
*nskb
;
3951 unsigned int header
= skb_headroom(skb
);
3952 int copy
= SKB_MAX_ORDER(header
, 0);
3954 /* Too big header? This can happen with IPv6. */
3957 if (end
-start
< copy
)
3959 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3963 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
3964 skb_set_network_header(nskb
, (skb_network_header(skb
) -
3966 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
3968 skb_reserve(nskb
, header
);
3969 memcpy(nskb
->head
, skb
->head
, header
);
3970 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3971 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3972 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3973 sk_stream_set_owner_r(nskb
, sk
);
3975 /* Copy data, releasing collapsed skbs. */
3977 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3978 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3982 size
= min(copy
, size
);
3983 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3985 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3989 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3990 struct sk_buff
*next
= skb
->next
;
3991 __skb_unlink(skb
, list
);
3993 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3996 tcp_hdr(skb
)->syn
||
4004 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4005 * and tcp_collapse() them until all the queue is collapsed.
4007 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4009 struct tcp_sock
*tp
= tcp_sk(sk
);
4010 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4011 struct sk_buff
*head
;
4017 start
= TCP_SKB_CB(skb
)->seq
;
4018 end
= TCP_SKB_CB(skb
)->end_seq
;
4024 /* Segment is terminated when we see gap or when
4025 * we are at the end of all the queue. */
4026 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4027 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4028 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4029 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4030 head
, skb
, start
, end
);
4032 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4034 /* Start new segment */
4035 start
= TCP_SKB_CB(skb
)->seq
;
4036 end
= TCP_SKB_CB(skb
)->end_seq
;
4038 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4039 start
= TCP_SKB_CB(skb
)->seq
;
4040 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4041 end
= TCP_SKB_CB(skb
)->end_seq
;
4046 /* Reduce allocated memory if we can, trying to get
4047 * the socket within its memory limits again.
4049 * Return less than zero if we should start dropping frames
4050 * until the socket owning process reads some of the data
4051 * to stabilize the situation.
4053 static int tcp_prune_queue(struct sock
*sk
)
4055 struct tcp_sock
*tp
= tcp_sk(sk
);
4057 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4059 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
4061 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4062 tcp_clamp_window(sk
);
4063 else if (tcp_memory_pressure
)
4064 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4066 tcp_collapse_ofo_queue(sk
);
4067 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4068 sk
->sk_receive_queue
.next
,
4069 (struct sk_buff
*)&sk
->sk_receive_queue
,
4070 tp
->copied_seq
, tp
->rcv_nxt
);
4071 sk_stream_mem_reclaim(sk
);
4073 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4076 /* Collapsing did not help, destructive actions follow.
4077 * This must not ever occur. */
4079 /* First, purge the out_of_order queue. */
4080 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4081 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
4082 __skb_queue_purge(&tp
->out_of_order_queue
);
4084 /* Reset SACK state. A conforming SACK implementation will
4085 * do the same at a timeout based retransmit. When a connection
4086 * is in a sad state like this, we care only about integrity
4087 * of the connection not performance.
4089 if (tcp_is_sack(tp
))
4090 tcp_sack_reset(&tp
->rx_opt
);
4091 sk_stream_mem_reclaim(sk
);
4094 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4097 /* If we are really being abused, tell the caller to silently
4098 * drop receive data on the floor. It will get retransmitted
4099 * and hopefully then we'll have sufficient space.
4101 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
4103 /* Massive buffer overcommit. */
4109 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4110 * As additional protections, we do not touch cwnd in retransmission phases,
4111 * and if application hit its sndbuf limit recently.
4113 void tcp_cwnd_application_limited(struct sock
*sk
)
4115 struct tcp_sock
*tp
= tcp_sk(sk
);
4117 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4118 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4119 /* Limited by application or receiver window. */
4120 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4121 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4122 if (win_used
< tp
->snd_cwnd
) {
4123 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4124 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4126 tp
->snd_cwnd_used
= 0;
4128 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4131 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4133 struct tcp_sock
*tp
= tcp_sk(sk
);
4135 /* If the user specified a specific send buffer setting, do
4138 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4141 /* If we are under global TCP memory pressure, do not expand. */
4142 if (tcp_memory_pressure
)
4145 /* If we are under soft global TCP memory pressure, do not expand. */
4146 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4149 /* If we filled the congestion window, do not expand. */
4150 if (tp
->packets_out
>= tp
->snd_cwnd
)
4156 /* When incoming ACK allowed to free some skb from write_queue,
4157 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4158 * on the exit from tcp input handler.
4160 * PROBLEM: sndbuf expansion does not work well with largesend.
4162 static void tcp_new_space(struct sock
*sk
)
4164 struct tcp_sock
*tp
= tcp_sk(sk
);
4166 if (tcp_should_expand_sndbuf(sk
)) {
4167 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4168 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4169 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4170 tp
->reordering
+ 1);
4171 sndmem
*= 2*demanded
;
4172 if (sndmem
> sk
->sk_sndbuf
)
4173 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4174 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4177 sk
->sk_write_space(sk
);
4180 static void tcp_check_space(struct sock
*sk
)
4182 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4183 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4184 if (sk
->sk_socket
&&
4185 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4190 static inline void tcp_data_snd_check(struct sock
*sk
)
4192 tcp_push_pending_frames(sk
);
4193 tcp_check_space(sk
);
4197 * Check if sending an ack is needed.
4199 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4201 struct tcp_sock
*tp
= tcp_sk(sk
);
4203 /* More than one full frame received... */
4204 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4205 /* ... and right edge of window advances far enough.
4206 * (tcp_recvmsg() will send ACK otherwise). Or...
4208 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4209 /* We ACK each frame or... */
4210 tcp_in_quickack_mode(sk
) ||
4211 /* We have out of order data. */
4213 skb_peek(&tp
->out_of_order_queue
))) {
4214 /* Then ack it now */
4217 /* Else, send delayed ack. */
4218 tcp_send_delayed_ack(sk
);
4222 static inline void tcp_ack_snd_check(struct sock
*sk
)
4224 if (!inet_csk_ack_scheduled(sk
)) {
4225 /* We sent a data segment already. */
4228 __tcp_ack_snd_check(sk
, 1);
4232 * This routine is only called when we have urgent data
4233 * signaled. Its the 'slow' part of tcp_urg. It could be
4234 * moved inline now as tcp_urg is only called from one
4235 * place. We handle URGent data wrong. We have to - as
4236 * BSD still doesn't use the correction from RFC961.
4237 * For 1003.1g we should support a new option TCP_STDURG to permit
4238 * either form (or just set the sysctl tcp_stdurg).
4241 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4243 struct tcp_sock
*tp
= tcp_sk(sk
);
4244 u32 ptr
= ntohs(th
->urg_ptr
);
4246 if (ptr
&& !sysctl_tcp_stdurg
)
4248 ptr
+= ntohl(th
->seq
);
4250 /* Ignore urgent data that we've already seen and read. */
4251 if (after(tp
->copied_seq
, ptr
))
4254 /* Do not replay urg ptr.
4256 * NOTE: interesting situation not covered by specs.
4257 * Misbehaving sender may send urg ptr, pointing to segment,
4258 * which we already have in ofo queue. We are not able to fetch
4259 * such data and will stay in TCP_URG_NOTYET until will be eaten
4260 * by recvmsg(). Seems, we are not obliged to handle such wicked
4261 * situations. But it is worth to think about possibility of some
4262 * DoSes using some hypothetical application level deadlock.
4264 if (before(ptr
, tp
->rcv_nxt
))
4267 /* Do we already have a newer (or duplicate) urgent pointer? */
4268 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4271 /* Tell the world about our new urgent pointer. */
4274 /* We may be adding urgent data when the last byte read was
4275 * urgent. To do this requires some care. We cannot just ignore
4276 * tp->copied_seq since we would read the last urgent byte again
4277 * as data, nor can we alter copied_seq until this data arrives
4278 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4280 * NOTE. Double Dutch. Rendering to plain English: author of comment
4281 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4282 * and expect that both A and B disappear from stream. This is _wrong_.
4283 * Though this happens in BSD with high probability, this is occasional.
4284 * Any application relying on this is buggy. Note also, that fix "works"
4285 * only in this artificial test. Insert some normal data between A and B and we will
4286 * decline of BSD again. Verdict: it is better to remove to trap
4289 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4290 !sock_flag(sk
, SOCK_URGINLINE
) &&
4291 tp
->copied_seq
!= tp
->rcv_nxt
) {
4292 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4294 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4295 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4300 tp
->urg_data
= TCP_URG_NOTYET
;
4303 /* Disable header prediction. */
4307 /* This is the 'fast' part of urgent handling. */
4308 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4310 struct tcp_sock
*tp
= tcp_sk(sk
);
4312 /* Check if we get a new urgent pointer - normally not. */
4314 tcp_check_urg(sk
,th
);
4316 /* Do we wait for any urgent data? - normally not... */
4317 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4318 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4321 /* Is the urgent pointer pointing into this packet? */
4322 if (ptr
< skb
->len
) {
4324 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4326 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4327 if (!sock_flag(sk
, SOCK_DEAD
))
4328 sk
->sk_data_ready(sk
, 0);
4333 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4335 struct tcp_sock
*tp
= tcp_sk(sk
);
4336 int chunk
= skb
->len
- hlen
;
4340 if (skb_csum_unnecessary(skb
))
4341 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4343 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4347 tp
->ucopy
.len
-= chunk
;
4348 tp
->copied_seq
+= chunk
;
4349 tcp_rcv_space_adjust(sk
);
4356 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4360 if (sock_owned_by_user(sk
)) {
4362 result
= __tcp_checksum_complete(skb
);
4365 result
= __tcp_checksum_complete(skb
);
4370 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4372 return !skb_csum_unnecessary(skb
) &&
4373 __tcp_checksum_complete_user(sk
, skb
);
4376 #ifdef CONFIG_NET_DMA
4377 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4379 struct tcp_sock
*tp
= tcp_sk(sk
);
4380 int chunk
= skb
->len
- hlen
;
4382 int copied_early
= 0;
4384 if (tp
->ucopy
.wakeup
)
4387 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4388 tp
->ucopy
.dma_chan
= get_softnet_dma();
4390 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4392 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4393 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4398 tp
->ucopy
.dma_cookie
= dma_cookie
;
4401 tp
->ucopy
.len
-= chunk
;
4402 tp
->copied_seq
+= chunk
;
4403 tcp_rcv_space_adjust(sk
);
4405 if ((tp
->ucopy
.len
== 0) ||
4406 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4407 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4408 tp
->ucopy
.wakeup
= 1;
4409 sk
->sk_data_ready(sk
, 0);
4411 } else if (chunk
> 0) {
4412 tp
->ucopy
.wakeup
= 1;
4413 sk
->sk_data_ready(sk
, 0);
4416 return copied_early
;
4418 #endif /* CONFIG_NET_DMA */
4421 * TCP receive function for the ESTABLISHED state.
4423 * It is split into a fast path and a slow path. The fast path is
4425 * - A zero window was announced from us - zero window probing
4426 * is only handled properly in the slow path.
4427 * - Out of order segments arrived.
4428 * - Urgent data is expected.
4429 * - There is no buffer space left
4430 * - Unexpected TCP flags/window values/header lengths are received
4431 * (detected by checking the TCP header against pred_flags)
4432 * - Data is sent in both directions. Fast path only supports pure senders
4433 * or pure receivers (this means either the sequence number or the ack
4434 * value must stay constant)
4435 * - Unexpected TCP option.
4437 * When these conditions are not satisfied it drops into a standard
4438 * receive procedure patterned after RFC793 to handle all cases.
4439 * The first three cases are guaranteed by proper pred_flags setting,
4440 * the rest is checked inline. Fast processing is turned on in
4441 * tcp_data_queue when everything is OK.
4443 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4444 struct tcphdr
*th
, unsigned len
)
4446 struct tcp_sock
*tp
= tcp_sk(sk
);
4449 * Header prediction.
4450 * The code loosely follows the one in the famous
4451 * "30 instruction TCP receive" Van Jacobson mail.
4453 * Van's trick is to deposit buffers into socket queue
4454 * on a device interrupt, to call tcp_recv function
4455 * on the receive process context and checksum and copy
4456 * the buffer to user space. smart...
4458 * Our current scheme is not silly either but we take the
4459 * extra cost of the net_bh soft interrupt processing...
4460 * We do checksum and copy also but from device to kernel.
4463 tp
->rx_opt
.saw_tstamp
= 0;
4465 /* pred_flags is 0xS?10 << 16 + snd_wnd
4466 * if header_prediction is to be made
4467 * 'S' will always be tp->tcp_header_len >> 2
4468 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4469 * turn it off (when there are holes in the receive
4470 * space for instance)
4471 * PSH flag is ignored.
4474 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4475 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4476 int tcp_header_len
= tp
->tcp_header_len
;
4478 /* Timestamp header prediction: tcp_header_len
4479 * is automatically equal to th->doff*4 due to pred_flags
4483 /* Check timestamp */
4484 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4485 __be32
*ptr
= (__be32
*)(th
+ 1);
4487 /* No? Slow path! */
4488 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4489 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4492 tp
->rx_opt
.saw_tstamp
= 1;
4494 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4496 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4498 /* If PAWS failed, check it more carefully in slow path */
4499 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4502 /* DO NOT update ts_recent here, if checksum fails
4503 * and timestamp was corrupted part, it will result
4504 * in a hung connection since we will drop all
4505 * future packets due to the PAWS test.
4509 if (len
<= tcp_header_len
) {
4510 /* Bulk data transfer: sender */
4511 if (len
== tcp_header_len
) {
4512 /* Predicted packet is in window by definition.
4513 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4514 * Hence, check seq<=rcv_wup reduces to:
4516 if (tcp_header_len
==
4517 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4518 tp
->rcv_nxt
== tp
->rcv_wup
)
4519 tcp_store_ts_recent(tp
);
4521 /* We know that such packets are checksummed
4524 tcp_ack(sk
, skb
, 0);
4526 tcp_data_snd_check(sk
);
4528 } else { /* Header too small */
4529 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4534 int copied_early
= 0;
4536 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4537 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4538 #ifdef CONFIG_NET_DMA
4539 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4544 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4545 __set_current_state(TASK_RUNNING
);
4547 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4551 /* Predicted packet is in window by definition.
4552 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4553 * Hence, check seq<=rcv_wup reduces to:
4555 if (tcp_header_len
==
4556 (sizeof(struct tcphdr
) +
4557 TCPOLEN_TSTAMP_ALIGNED
) &&
4558 tp
->rcv_nxt
== tp
->rcv_wup
)
4559 tcp_store_ts_recent(tp
);
4561 tcp_rcv_rtt_measure_ts(sk
, skb
);
4563 __skb_pull(skb
, tcp_header_len
);
4564 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4565 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4568 tcp_cleanup_rbuf(sk
, skb
->len
);
4571 if (tcp_checksum_complete_user(sk
, skb
))
4574 /* Predicted packet is in window by definition.
4575 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4576 * Hence, check seq<=rcv_wup reduces to:
4578 if (tcp_header_len
==
4579 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4580 tp
->rcv_nxt
== tp
->rcv_wup
)
4581 tcp_store_ts_recent(tp
);
4583 tcp_rcv_rtt_measure_ts(sk
, skb
);
4585 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4588 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4590 /* Bulk data transfer: receiver */
4591 __skb_pull(skb
,tcp_header_len
);
4592 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4593 sk_stream_set_owner_r(skb
, sk
);
4594 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4597 tcp_event_data_recv(sk
, skb
);
4599 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4600 /* Well, only one small jumplet in fast path... */
4601 tcp_ack(sk
, skb
, FLAG_DATA
);
4602 tcp_data_snd_check(sk
);
4603 if (!inet_csk_ack_scheduled(sk
))
4607 __tcp_ack_snd_check(sk
, 0);
4609 #ifdef CONFIG_NET_DMA
4611 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4617 sk
->sk_data_ready(sk
, 0);
4623 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4627 * RFC1323: H1. Apply PAWS check first.
4629 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4630 tcp_paws_discard(sk
, skb
)) {
4632 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4633 tcp_send_dupack(sk
, skb
);
4636 /* Resets are accepted even if PAWS failed.
4638 ts_recent update must be made after we are sure
4639 that the packet is in window.
4644 * Standard slow path.
4647 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4648 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4649 * (RST) segments are validated by checking their SEQ-fields."
4650 * And page 69: "If an incoming segment is not acceptable,
4651 * an acknowledgment should be sent in reply (unless the RST bit
4652 * is set, if so drop the segment and return)".
4655 tcp_send_dupack(sk
, skb
);
4664 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4666 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4667 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4668 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4675 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4677 tcp_rcv_rtt_measure_ts(sk
, skb
);
4679 /* Process urgent data. */
4680 tcp_urg(sk
, skb
, th
);
4682 /* step 7: process the segment text */
4683 tcp_data_queue(sk
, skb
);
4685 tcp_data_snd_check(sk
);
4686 tcp_ack_snd_check(sk
);
4690 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4697 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4698 struct tcphdr
*th
, unsigned len
)
4700 struct tcp_sock
*tp
= tcp_sk(sk
);
4701 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4702 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4704 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4708 * "If the state is SYN-SENT then
4709 * first check the ACK bit
4710 * If the ACK bit is set
4711 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4712 * a reset (unless the RST bit is set, if so drop
4713 * the segment and return)"
4715 * We do not send data with SYN, so that RFC-correct
4718 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4719 goto reset_and_undo
;
4721 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4722 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4724 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4725 goto reset_and_undo
;
4728 /* Now ACK is acceptable.
4730 * "If the RST bit is set
4731 * If the ACK was acceptable then signal the user "error:
4732 * connection reset", drop the segment, enter CLOSED state,
4733 * delete TCB, and return."
4742 * "fifth, if neither of the SYN or RST bits is set then
4743 * drop the segment and return."
4749 goto discard_and_undo
;
4752 * "If the SYN bit is on ...
4753 * are acceptable then ...
4754 * (our SYN has been ACKed), change the connection
4755 * state to ESTABLISHED..."
4758 TCP_ECN_rcv_synack(tp
, th
);
4760 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4761 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4763 /* Ok.. it's good. Set up sequence numbers and
4764 * move to established.
4766 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4767 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4769 /* RFC1323: The window in SYN & SYN/ACK segments is
4772 tp
->snd_wnd
= ntohs(th
->window
);
4773 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4775 if (!tp
->rx_opt
.wscale_ok
) {
4776 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4777 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4780 if (tp
->rx_opt
.saw_tstamp
) {
4781 tp
->rx_opt
.tstamp_ok
= 1;
4782 tp
->tcp_header_len
=
4783 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4784 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4785 tcp_store_ts_recent(tp
);
4787 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4790 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
4791 tcp_enable_fack(tp
);
4794 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4795 tcp_initialize_rcv_mss(sk
);
4797 /* Remember, tcp_poll() does not lock socket!
4798 * Change state from SYN-SENT only after copied_seq
4799 * is initialized. */
4800 tp
->copied_seq
= tp
->rcv_nxt
;
4802 tcp_set_state(sk
, TCP_ESTABLISHED
);
4804 security_inet_conn_established(sk
, skb
);
4806 /* Make sure socket is routed, for correct metrics. */
4807 icsk
->icsk_af_ops
->rebuild_header(sk
);
4809 tcp_init_metrics(sk
);
4811 tcp_init_congestion_control(sk
);
4813 /* Prevent spurious tcp_cwnd_restart() on first data
4816 tp
->lsndtime
= tcp_time_stamp
;
4818 tcp_init_buffer_space(sk
);
4820 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4821 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4823 if (!tp
->rx_opt
.snd_wscale
)
4824 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4828 if (!sock_flag(sk
, SOCK_DEAD
)) {
4829 sk
->sk_state_change(sk
);
4830 sk_wake_async(sk
, 0, POLL_OUT
);
4833 if (sk
->sk_write_pending
||
4834 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4835 icsk
->icsk_ack
.pingpong
) {
4836 /* Save one ACK. Data will be ready after
4837 * several ticks, if write_pending is set.
4839 * It may be deleted, but with this feature tcpdumps
4840 * look so _wonderfully_ clever, that I was not able
4841 * to stand against the temptation 8) --ANK
4843 inet_csk_schedule_ack(sk
);
4844 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4845 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4846 tcp_incr_quickack(sk
);
4847 tcp_enter_quickack_mode(sk
);
4848 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4849 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4860 /* No ACK in the segment */
4864 * "If the RST bit is set
4866 * Otherwise (no ACK) drop the segment and return."
4869 goto discard_and_undo
;
4873 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4874 goto discard_and_undo
;
4877 /* We see SYN without ACK. It is attempt of
4878 * simultaneous connect with crossed SYNs.
4879 * Particularly, it can be connect to self.
4881 tcp_set_state(sk
, TCP_SYN_RECV
);
4883 if (tp
->rx_opt
.saw_tstamp
) {
4884 tp
->rx_opt
.tstamp_ok
= 1;
4885 tcp_store_ts_recent(tp
);
4886 tp
->tcp_header_len
=
4887 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4889 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4892 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4893 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4895 /* RFC1323: The window in SYN & SYN/ACK segments is
4898 tp
->snd_wnd
= ntohs(th
->window
);
4899 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4900 tp
->max_window
= tp
->snd_wnd
;
4902 TCP_ECN_rcv_syn(tp
, th
);
4905 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4906 tcp_initialize_rcv_mss(sk
);
4909 tcp_send_synack(sk
);
4911 /* Note, we could accept data and URG from this segment.
4912 * There are no obstacles to make this.
4914 * However, if we ignore data in ACKless segments sometimes,
4915 * we have no reasons to accept it sometimes.
4916 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4917 * is not flawless. So, discard packet for sanity.
4918 * Uncomment this return to process the data.
4925 /* "fifth, if neither of the SYN or RST bits is set then
4926 * drop the segment and return."
4930 tcp_clear_options(&tp
->rx_opt
);
4931 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4935 tcp_clear_options(&tp
->rx_opt
);
4936 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4942 * This function implements the receiving procedure of RFC 793 for
4943 * all states except ESTABLISHED and TIME_WAIT.
4944 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4945 * address independent.
4948 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4949 struct tcphdr
*th
, unsigned len
)
4951 struct tcp_sock
*tp
= tcp_sk(sk
);
4952 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4955 tp
->rx_opt
.saw_tstamp
= 0;
4957 switch (sk
->sk_state
) {
4969 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4972 /* Now we have several options: In theory there is
4973 * nothing else in the frame. KA9Q has an option to
4974 * send data with the syn, BSD accepts data with the
4975 * syn up to the [to be] advertised window and
4976 * Solaris 2.1 gives you a protocol error. For now
4977 * we just ignore it, that fits the spec precisely
4978 * and avoids incompatibilities. It would be nice in
4979 * future to drop through and process the data.
4981 * Now that TTCP is starting to be used we ought to
4983 * But, this leaves one open to an easy denial of
4984 * service attack, and SYN cookies can't defend
4985 * against this problem. So, we drop the data
4986 * in the interest of security over speed unless
4987 * it's still in use.
4995 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
4999 /* Do step6 onward by hand. */
5000 tcp_urg(sk
, skb
, th
);
5002 tcp_data_snd_check(sk
);
5006 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5007 tcp_paws_discard(sk
, skb
)) {
5009 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
5010 tcp_send_dupack(sk
, skb
);
5013 /* Reset is accepted even if it did not pass PAWS. */
5016 /* step 1: check sequence number */
5017 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5019 tcp_send_dupack(sk
, skb
);
5023 /* step 2: check RST bit */
5029 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5031 /* step 3: check security and precedence [ignored] */
5035 * Check for a SYN in window.
5037 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5038 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
5043 /* step 5: check the ACK field */
5045 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5047 switch (sk
->sk_state
) {
5050 tp
->copied_seq
= tp
->rcv_nxt
;
5052 tcp_set_state(sk
, TCP_ESTABLISHED
);
5053 sk
->sk_state_change(sk
);
5055 /* Note, that this wakeup is only for marginal
5056 * crossed SYN case. Passively open sockets
5057 * are not waked up, because sk->sk_sleep ==
5058 * NULL and sk->sk_socket == NULL.
5060 if (sk
->sk_socket
) {
5061 sk_wake_async(sk
,0,POLL_OUT
);
5064 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5065 tp
->snd_wnd
= ntohs(th
->window
) <<
5066 tp
->rx_opt
.snd_wscale
;
5067 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5068 TCP_SKB_CB(skb
)->seq
);
5070 /* tcp_ack considers this ACK as duplicate
5071 * and does not calculate rtt.
5072 * Fix it at least with timestamps.
5074 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5076 tcp_ack_saw_tstamp(sk
, 0);
5078 if (tp
->rx_opt
.tstamp_ok
)
5079 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5081 /* Make sure socket is routed, for
5084 icsk
->icsk_af_ops
->rebuild_header(sk
);
5086 tcp_init_metrics(sk
);
5088 tcp_init_congestion_control(sk
);
5090 /* Prevent spurious tcp_cwnd_restart() on
5091 * first data packet.
5093 tp
->lsndtime
= tcp_time_stamp
;
5096 tcp_initialize_rcv_mss(sk
);
5097 tcp_init_buffer_space(sk
);
5098 tcp_fast_path_on(tp
);
5105 if (tp
->snd_una
== tp
->write_seq
) {
5106 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5107 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5108 dst_confirm(sk
->sk_dst_cache
);
5110 if (!sock_flag(sk
, SOCK_DEAD
))
5111 /* Wake up lingering close() */
5112 sk
->sk_state_change(sk
);
5116 if (tp
->linger2
< 0 ||
5117 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5118 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5120 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5124 tmo
= tcp_fin_time(sk
);
5125 if (tmo
> TCP_TIMEWAIT_LEN
) {
5126 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5127 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5128 /* Bad case. We could lose such FIN otherwise.
5129 * It is not a big problem, but it looks confusing
5130 * and not so rare event. We still can lose it now,
5131 * if it spins in bh_lock_sock(), but it is really
5134 inet_csk_reset_keepalive_timer(sk
, tmo
);
5136 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5144 if (tp
->snd_una
== tp
->write_seq
) {
5145 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5151 if (tp
->snd_una
== tp
->write_seq
) {
5152 tcp_update_metrics(sk
);
5161 /* step 6: check the URG bit */
5162 tcp_urg(sk
, skb
, th
);
5164 /* step 7: process the segment text */
5165 switch (sk
->sk_state
) {
5166 case TCP_CLOSE_WAIT
:
5169 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5173 /* RFC 793 says to queue data in these states,
5174 * RFC 1122 says we MUST send a reset.
5175 * BSD 4.4 also does reset.
5177 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5178 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5179 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5180 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5186 case TCP_ESTABLISHED
:
5187 tcp_data_queue(sk
, skb
);
5192 /* tcp_data could move socket to TIME-WAIT */
5193 if (sk
->sk_state
!= TCP_CLOSE
) {
5194 tcp_data_snd_check(sk
);
5195 tcp_ack_snd_check(sk
);
5205 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5206 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5207 EXPORT_SYMBOL(tcp_parse_options
);
5208 EXPORT_SYMBOL(tcp_rcv_established
);
5209 EXPORT_SYMBOL(tcp_rcv_state_process
);
5210 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);