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(sk
));
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
))
1272 if (!tp
->packets_out
)
1276 * if the only SACK change is the increase of the end_seq of
1277 * the first block then only apply that SACK block
1278 * and use retrans queue hinting otherwise slowpath */
1280 for (i
= 0; i
< num_sacks
; i
++) {
1281 __be32 start_seq
= sp
[i
].start_seq
;
1282 __be32 end_seq
= sp
[i
].end_seq
;
1285 if (tp
->recv_sack_cache
[i
].start_seq
!= start_seq
)
1288 if ((tp
->recv_sack_cache
[i
].start_seq
!= start_seq
) ||
1289 (tp
->recv_sack_cache
[i
].end_seq
!= end_seq
))
1292 tp
->recv_sack_cache
[i
].start_seq
= start_seq
;
1293 tp
->recv_sack_cache
[i
].end_seq
= end_seq
;
1295 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1296 for (; i
< ARRAY_SIZE(tp
->recv_sack_cache
); i
++) {
1297 tp
->recv_sack_cache
[i
].start_seq
= 0;
1298 tp
->recv_sack_cache
[i
].end_seq
= 0;
1301 first_sack_index
= 0;
1306 tp
->fastpath_skb_hint
= NULL
;
1308 /* order SACK blocks to allow in order walk of the retrans queue */
1309 for (i
= num_sacks
-1; i
> 0; i
--) {
1310 for (j
= 0; j
< i
; j
++){
1311 if (after(ntohl(sp
[j
].start_seq
),
1312 ntohl(sp
[j
+1].start_seq
))){
1313 struct tcp_sack_block_wire tmp
;
1319 /* Track where the first SACK block goes to */
1320 if (j
== first_sack_index
)
1321 first_sack_index
= j
+1;
1328 /* Use SACK fastpath hint if valid */
1329 cached_skb
= tp
->fastpath_skb_hint
;
1330 cached_fack_count
= tp
->fastpath_cnt_hint
;
1332 cached_skb
= tcp_write_queue_head(sk
);
1333 cached_fack_count
= 0;
1336 for (i
= 0; i
< num_sacks
; i
++) {
1337 struct sk_buff
*skb
;
1338 __u32 start_seq
= ntohl(sp
->start_seq
);
1339 __u32 end_seq
= ntohl(sp
->end_seq
);
1341 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1342 int next_dup
= (found_dup_sack
&& (i
+1 == first_sack_index
));
1346 if (!tcp_is_sackblock_valid(tp
, dup_sack
, start_seq
, end_seq
)) {
1348 if (!tp
->undo_marker
)
1349 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO
);
1351 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD
);
1353 /* Don't count olds caused by ACK reordering */
1354 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1355 !after(end_seq
, tp
->snd_una
))
1357 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD
);
1363 fack_count
= cached_fack_count
;
1365 /* Event "B" in the comment above. */
1366 if (after(end_seq
, tp
->high_seq
))
1367 flag
|= FLAG_DATA_LOST
;
1369 tcp_for_write_queue_from(skb
, sk
) {
1373 if (skb
== tcp_send_head(sk
))
1377 cached_fack_count
= fack_count
;
1378 if (i
== first_sack_index
) {
1379 tp
->fastpath_skb_hint
= skb
;
1380 tp
->fastpath_cnt_hint
= fack_count
;
1383 /* The retransmission queue is always in order, so
1384 * we can short-circuit the walk early.
1386 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1389 dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1391 /* Due to sorting DSACK may reside within this SACK block! */
1393 u32 dup_start
= ntohl(sp
->start_seq
);
1394 u32 dup_end
= ntohl(sp
->end_seq
);
1396 if (before(TCP_SKB_CB(skb
)->seq
, dup_end
)) {
1397 in_sack
= tcp_match_skb_to_sack(sk
, skb
, dup_start
, dup_end
);
1403 /* DSACK info lost if out-of-mem, try SACK still */
1405 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
, end_seq
);
1406 if (unlikely(in_sack
< 0))
1410 fack_count
+= tcp_skb_pcount(skb
);
1414 sacked
= TCP_SKB_CB(skb
)->sacked
;
1416 /* Account D-SACK for retransmitted packet. */
1417 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1418 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1420 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
) &&
1421 (sacked
& TCPCB_SACKED_ACKED
))
1422 reord
= min(fack_count
, reord
);
1426 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1427 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
)) {
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
) &&
1515 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1516 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1517 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1521 #if FASTRETRANS_DEBUG > 0
1522 BUG_TRAP((int)tp
->sacked_out
>= 0);
1523 BUG_TRAP((int)tp
->lost_out
>= 0);
1524 BUG_TRAP((int)tp
->retrans_out
>= 0);
1525 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1530 /* If we receive more dupacks than we expected counting segments
1531 * in assumption of absent reordering, interpret this as reordering.
1532 * The only another reason could be bug in receiver TCP.
1534 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1536 struct tcp_sock
*tp
= tcp_sk(sk
);
1539 holes
= max(tp
->lost_out
, 1U);
1540 holes
= min(holes
, tp
->packets_out
);
1542 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1543 tp
->sacked_out
= tp
->packets_out
- holes
;
1544 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1548 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1550 static void tcp_add_reno_sack(struct sock
*sk
)
1552 struct tcp_sock
*tp
= tcp_sk(sk
);
1554 tcp_check_reno_reordering(sk
, 0);
1555 tcp_verify_left_out(tp
);
1558 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1560 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1562 struct tcp_sock
*tp
= tcp_sk(sk
);
1565 /* One ACK acked hole. The rest eat duplicate ACKs. */
1566 if (acked
-1 >= tp
->sacked_out
)
1569 tp
->sacked_out
-= acked
-1;
1571 tcp_check_reno_reordering(sk
, acked
);
1572 tcp_verify_left_out(tp
);
1575 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1580 /* F-RTO can only be used if TCP has never retransmitted anything other than
1581 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1583 int tcp_use_frto(struct sock
*sk
)
1585 const struct tcp_sock
*tp
= tcp_sk(sk
);
1586 struct sk_buff
*skb
;
1588 if (!sysctl_tcp_frto
)
1594 /* Avoid expensive walking of rexmit queue if possible */
1595 if (tp
->retrans_out
> 1)
1598 skb
= tcp_write_queue_head(sk
);
1599 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1600 tcp_for_write_queue_from(skb
, sk
) {
1601 if (skb
== tcp_send_head(sk
))
1603 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1605 /* Short-circuit when first non-SACKed skb has been checked */
1606 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1612 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1613 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1614 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1615 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1616 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1617 * bits are handled if the Loss state is really to be entered (in
1618 * tcp_enter_frto_loss).
1620 * Do like tcp_enter_loss() would; when RTO expires the second time it
1622 * "Reduce ssthresh if it has not yet been made inside this window."
1624 void tcp_enter_frto(struct sock
*sk
)
1626 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1627 struct tcp_sock
*tp
= tcp_sk(sk
);
1628 struct sk_buff
*skb
;
1630 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1631 tp
->snd_una
== tp
->high_seq
||
1632 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1633 !icsk
->icsk_retransmits
)) {
1634 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1635 /* Our state is too optimistic in ssthresh() call because cwnd
1636 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1637 * recovery has not yet completed. Pattern would be this: RTO,
1638 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1640 * RFC4138 should be more specific on what to do, even though
1641 * RTO is quite unlikely to occur after the first Cumulative ACK
1642 * due to back-off and complexity of triggering events ...
1644 if (tp
->frto_counter
) {
1646 stored_cwnd
= tp
->snd_cwnd
;
1648 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1649 tp
->snd_cwnd
= stored_cwnd
;
1651 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1653 /* ... in theory, cong.control module could do "any tricks" in
1654 * ssthresh(), which means that ca_state, lost bits and lost_out
1655 * counter would have to be faked before the call occurs. We
1656 * consider that too expensive, unlikely and hacky, so modules
1657 * using these in ssthresh() must deal these incompatibility
1658 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1660 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1663 tp
->undo_marker
= tp
->snd_una
;
1664 tp
->undo_retrans
= 0;
1666 skb
= tcp_write_queue_head(sk
);
1667 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1668 tp
->undo_marker
= 0;
1669 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1670 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1671 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1673 tcp_verify_left_out(tp
);
1675 /* Too bad if TCP was application limited */
1676 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1678 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1679 * The last condition is necessary at least in tp->frto_counter case.
1681 if (IsSackFrto() && (tp
->frto_counter
||
1682 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1683 after(tp
->high_seq
, tp
->snd_una
)) {
1684 tp
->frto_highmark
= tp
->high_seq
;
1686 tp
->frto_highmark
= tp
->snd_nxt
;
1688 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1689 tp
->high_seq
= tp
->snd_nxt
;
1690 tp
->frto_counter
= 1;
1693 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1694 * which indicates that we should follow the traditional RTO recovery,
1695 * i.e. mark everything lost and do go-back-N retransmission.
1697 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1699 struct tcp_sock
*tp
= tcp_sk(sk
);
1700 struct sk_buff
*skb
;
1703 tp
->retrans_out
= 0;
1704 if (tcp_is_reno(tp
))
1705 tcp_reset_reno_sack(tp
);
1707 tcp_for_write_queue(skb
, sk
) {
1708 if (skb
== tcp_send_head(sk
))
1711 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1713 * Count the retransmission made on RTO correctly (only when
1714 * waiting for the first ACK and did not get it)...
1716 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1717 /* For some reason this R-bit might get cleared? */
1718 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1719 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1720 /* ...enter this if branch just for the first segment */
1721 flag
|= FLAG_DATA_ACKED
;
1723 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1724 tp
->undo_marker
= 0;
1725 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1728 /* Don't lost mark skbs that were fwd transmitted after RTO */
1729 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) &&
1730 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
)) {
1731 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1732 tp
->lost_out
+= tcp_skb_pcount(skb
);
1735 tcp_verify_left_out(tp
);
1737 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1738 tp
->snd_cwnd_cnt
= 0;
1739 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1740 tp
->frto_counter
= 0;
1741 tp
->bytes_acked
= 0;
1743 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1744 sysctl_tcp_reordering
);
1745 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1746 tp
->high_seq
= tp
->frto_highmark
;
1747 TCP_ECN_queue_cwr(tp
);
1749 tcp_clear_retrans_hints_partial(tp
);
1752 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1754 tp
->retrans_out
= 0;
1757 tp
->undo_marker
= 0;
1758 tp
->undo_retrans
= 0;
1761 void tcp_clear_retrans(struct tcp_sock
*tp
)
1763 tcp_clear_retrans_partial(tp
);
1765 tp
->fackets_out
= 0;
1769 /* Enter Loss state. If "how" is not zero, forget all SACK information
1770 * and reset tags completely, otherwise preserve SACKs. If receiver
1771 * dropped its ofo queue, we will know this due to reneging detection.
1773 void tcp_enter_loss(struct sock
*sk
, int how
)
1775 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1776 struct tcp_sock
*tp
= tcp_sk(sk
);
1777 struct sk_buff
*skb
;
1779 /* Reduce ssthresh if it has not yet been made inside this window. */
1780 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1781 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1782 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1783 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1784 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1787 tp
->snd_cwnd_cnt
= 0;
1788 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1790 tp
->bytes_acked
= 0;
1791 tcp_clear_retrans_partial(tp
);
1793 if (tcp_is_reno(tp
))
1794 tcp_reset_reno_sack(tp
);
1797 /* Push undo marker, if it was plain RTO and nothing
1798 * was retransmitted. */
1799 tp
->undo_marker
= tp
->snd_una
;
1800 tcp_clear_retrans_hints_partial(tp
);
1803 tp
->fackets_out
= 0;
1804 tcp_clear_all_retrans_hints(tp
);
1807 tcp_for_write_queue(skb
, sk
) {
1808 if (skb
== tcp_send_head(sk
))
1811 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1812 tp
->undo_marker
= 0;
1813 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1814 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1815 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1816 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1817 tp
->lost_out
+= tcp_skb_pcount(skb
);
1820 tcp_verify_left_out(tp
);
1822 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1823 sysctl_tcp_reordering
);
1824 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1825 tp
->high_seq
= tp
->snd_nxt
;
1826 TCP_ECN_queue_cwr(tp
);
1827 /* Abort F-RTO algorithm if one is in progress */
1828 tp
->frto_counter
= 0;
1831 static int tcp_check_sack_reneging(struct sock
*sk
)
1833 struct sk_buff
*skb
;
1835 /* If ACK arrived pointing to a remembered SACK,
1836 * it means that our remembered SACKs do not reflect
1837 * real state of receiver i.e.
1838 * receiver _host_ is heavily congested (or buggy).
1839 * Do processing similar to RTO timeout.
1841 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1842 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1843 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1844 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1846 tcp_enter_loss(sk
, 1);
1847 icsk
->icsk_retransmits
++;
1848 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1849 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1850 icsk
->icsk_rto
, TCP_RTO_MAX
);
1856 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1858 return tcp_is_reno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1861 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1863 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1866 static inline int tcp_head_timedout(struct sock
*sk
)
1868 struct tcp_sock
*tp
= tcp_sk(sk
);
1870 return tp
->packets_out
&&
1871 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1874 /* Linux NewReno/SACK/FACK/ECN state machine.
1875 * --------------------------------------
1877 * "Open" Normal state, no dubious events, fast path.
1878 * "Disorder" In all the respects it is "Open",
1879 * but requires a bit more attention. It is entered when
1880 * we see some SACKs or dupacks. It is split of "Open"
1881 * mainly to move some processing from fast path to slow one.
1882 * "CWR" CWND was reduced due to some Congestion Notification event.
1883 * It can be ECN, ICMP source quench, local device congestion.
1884 * "Recovery" CWND was reduced, we are fast-retransmitting.
1885 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1887 * tcp_fastretrans_alert() is entered:
1888 * - each incoming ACK, if state is not "Open"
1889 * - when arrived ACK is unusual, namely:
1894 * Counting packets in flight is pretty simple.
1896 * in_flight = packets_out - left_out + retrans_out
1898 * packets_out is SND.NXT-SND.UNA counted in packets.
1900 * retrans_out is number of retransmitted segments.
1902 * left_out is number of segments left network, but not ACKed yet.
1904 * left_out = sacked_out + lost_out
1906 * sacked_out: Packets, which arrived to receiver out of order
1907 * and hence not ACKed. With SACKs this number is simply
1908 * amount of SACKed data. Even without SACKs
1909 * it is easy to give pretty reliable estimate of this number,
1910 * counting duplicate ACKs.
1912 * lost_out: Packets lost by network. TCP has no explicit
1913 * "loss notification" feedback from network (for now).
1914 * It means that this number can be only _guessed_.
1915 * Actually, it is the heuristics to predict lossage that
1916 * distinguishes different algorithms.
1918 * F.e. after RTO, when all the queue is considered as lost,
1919 * lost_out = packets_out and in_flight = retrans_out.
1921 * Essentially, we have now two algorithms counting
1924 * FACK: It is the simplest heuristics. As soon as we decided
1925 * that something is lost, we decide that _all_ not SACKed
1926 * packets until the most forward SACK are lost. I.e.
1927 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1928 * It is absolutely correct estimate, if network does not reorder
1929 * packets. And it loses any connection to reality when reordering
1930 * takes place. We use FACK by default until reordering
1931 * is suspected on the path to this destination.
1933 * NewReno: when Recovery is entered, we assume that one segment
1934 * is lost (classic Reno). While we are in Recovery and
1935 * a partial ACK arrives, we assume that one more packet
1936 * is lost (NewReno). This heuristics are the same in NewReno
1939 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1940 * deflation etc. CWND is real congestion window, never inflated, changes
1941 * only according to classic VJ rules.
1943 * Really tricky (and requiring careful tuning) part of algorithm
1944 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1945 * The first determines the moment _when_ we should reduce CWND and,
1946 * hence, slow down forward transmission. In fact, it determines the moment
1947 * when we decide that hole is caused by loss, rather than by a reorder.
1949 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1950 * holes, caused by lost packets.
1952 * And the most logically complicated part of algorithm is undo
1953 * heuristics. We detect false retransmits due to both too early
1954 * fast retransmit (reordering) and underestimated RTO, analyzing
1955 * timestamps and D-SACKs. When we detect that some segments were
1956 * retransmitted by mistake and CWND reduction was wrong, we undo
1957 * window reduction and abort recovery phase. This logic is hidden
1958 * inside several functions named tcp_try_undo_<something>.
1961 /* This function decides, when we should leave Disordered state
1962 * and enter Recovery phase, reducing congestion window.
1964 * Main question: may we further continue forward transmission
1965 * with the same cwnd?
1967 static int tcp_time_to_recover(struct sock
*sk
)
1969 struct tcp_sock
*tp
= tcp_sk(sk
);
1972 /* Do not perform any recovery during F-RTO algorithm */
1973 if (tp
->frto_counter
)
1976 /* Trick#1: The loss is proven. */
1980 /* Not-A-Trick#2 : Classic rule... */
1981 if (tcp_fackets_out(tp
) > tp
->reordering
)
1984 /* Trick#3 : when we use RFC2988 timer restart, fast
1985 * retransmit can be triggered by timeout of queue head.
1987 if (tcp_head_timedout(sk
))
1990 /* Trick#4: It is still not OK... But will it be useful to delay
1993 packets_out
= tp
->packets_out
;
1994 if (packets_out
<= tp
->reordering
&&
1995 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
1996 !tcp_may_send_now(sk
)) {
1997 /* We have nothing to send. This connection is limited
1998 * either by receiver window or by application.
2006 /* RFC: This is from the original, I doubt that this is necessary at all:
2007 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2008 * retransmitted past LOST markings in the first place? I'm not fully sure
2009 * about undo and end of connection cases, which can cause R without L?
2011 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
2012 struct sk_buff
*skb
)
2014 if ((tp
->retransmit_skb_hint
!= NULL
) &&
2015 before(TCP_SKB_CB(skb
)->seq
,
2016 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
2017 tp
->retransmit_skb_hint
= NULL
;
2020 /* Mark head of queue up as lost. */
2021 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2023 struct tcp_sock
*tp
= tcp_sk(sk
);
2024 struct sk_buff
*skb
;
2027 BUG_TRAP(packets
<= tp
->packets_out
);
2028 if (tp
->lost_skb_hint
) {
2029 skb
= tp
->lost_skb_hint
;
2030 cnt
= tp
->lost_cnt_hint
;
2032 skb
= tcp_write_queue_head(sk
);
2036 tcp_for_write_queue_from(skb
, sk
) {
2037 if (skb
== tcp_send_head(sk
))
2039 /* TODO: do this better */
2040 /* this is not the most efficient way to do this... */
2041 tp
->lost_skb_hint
= skb
;
2042 tp
->lost_cnt_hint
= cnt
;
2043 cnt
+= tcp_skb_pcount(skb
);
2044 if (cnt
> packets
|| after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2046 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2047 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2048 tp
->lost_out
+= tcp_skb_pcount(skb
);
2049 tcp_verify_retransmit_hint(tp
, skb
);
2052 tcp_verify_left_out(tp
);
2055 /* Account newly detected lost packet(s) */
2057 static void tcp_update_scoreboard(struct sock
*sk
)
2059 struct tcp_sock
*tp
= tcp_sk(sk
);
2061 if (tcp_is_fack(tp
)) {
2062 int lost
= tp
->fackets_out
- tp
->reordering
;
2065 tcp_mark_head_lost(sk
, lost
);
2067 tcp_mark_head_lost(sk
, 1);
2070 /* New heuristics: it is possible only after we switched
2071 * to restart timer each time when something is ACKed.
2072 * Hence, we can detect timed out packets during fast
2073 * retransmit without falling to slow start.
2075 if (!tcp_is_reno(tp
) && tcp_head_timedout(sk
)) {
2076 struct sk_buff
*skb
;
2078 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2079 : tcp_write_queue_head(sk
);
2081 tcp_for_write_queue_from(skb
, sk
) {
2082 if (skb
== tcp_send_head(sk
))
2084 if (!tcp_skb_timedout(sk
, skb
))
2087 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2088 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2089 tp
->lost_out
+= tcp_skb_pcount(skb
);
2090 tcp_verify_retransmit_hint(tp
, skb
);
2094 tp
->scoreboard_skb_hint
= skb
;
2096 tcp_verify_left_out(tp
);
2100 /* CWND moderation, preventing bursts due to too big ACKs
2101 * in dubious situations.
2103 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2105 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2106 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
2107 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2110 /* Lower bound on congestion window is slow start threshold
2111 * unless congestion avoidance choice decides to overide it.
2113 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2115 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2117 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2120 /* Decrease cwnd each second ack. */
2121 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2123 struct tcp_sock
*tp
= tcp_sk(sk
);
2124 int decr
= tp
->snd_cwnd_cnt
+ 1;
2126 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
2127 (tcp_is_reno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
2128 tp
->snd_cwnd_cnt
= decr
&1;
2131 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2132 tp
->snd_cwnd
-= decr
;
2134 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
2135 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2139 /* Nothing was retransmitted or returned timestamp is less
2140 * than timestamp of the first retransmission.
2142 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2144 return !tp
->retrans_stamp
||
2145 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2146 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
2149 /* Undo procedures. */
2151 #if FASTRETRANS_DEBUG > 1
2152 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2154 struct tcp_sock
*tp
= tcp_sk(sk
);
2155 struct inet_sock
*inet
= inet_sk(sk
);
2157 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2159 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2160 tp
->snd_cwnd
, tcp_left_out(tp
),
2161 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2165 #define DBGUNDO(x...) do { } while (0)
2168 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2170 struct tcp_sock
*tp
= tcp_sk(sk
);
2172 if (tp
->prior_ssthresh
) {
2173 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2175 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2176 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2178 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
2180 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2181 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2182 TCP_ECN_withdraw_cwr(tp
);
2185 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2187 tcp_moderate_cwnd(tp
);
2188 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2190 /* There is something screwy going on with the retrans hints after
2192 tcp_clear_all_retrans_hints(tp
);
2195 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2197 return tp
->undo_marker
&&
2198 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2201 /* People celebrate: "We love our President!" */
2202 static int tcp_try_undo_recovery(struct sock
*sk
)
2204 struct tcp_sock
*tp
= tcp_sk(sk
);
2206 if (tcp_may_undo(tp
)) {
2207 /* Happy end! We did not retransmit anything
2208 * or our original transmission succeeded.
2210 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2211 tcp_undo_cwr(sk
, 1);
2212 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2213 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2215 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
2216 tp
->undo_marker
= 0;
2218 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2219 /* Hold old state until something *above* high_seq
2220 * is ACKed. For Reno it is MUST to prevent false
2221 * fast retransmits (RFC2582). SACK TCP is safe. */
2222 tcp_moderate_cwnd(tp
);
2225 tcp_set_ca_state(sk
, TCP_CA_Open
);
2229 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2230 static void tcp_try_undo_dsack(struct sock
*sk
)
2232 struct tcp_sock
*tp
= tcp_sk(sk
);
2234 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2235 DBGUNDO(sk
, "D-SACK");
2236 tcp_undo_cwr(sk
, 1);
2237 tp
->undo_marker
= 0;
2238 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2242 /* Undo during fast recovery after partial ACK. */
2244 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2246 struct tcp_sock
*tp
= tcp_sk(sk
);
2247 /* Partial ACK arrived. Force Hoe's retransmit. */
2248 int failed
= tcp_is_reno(tp
) || tp
->fackets_out
>tp
->reordering
;
2250 if (tcp_may_undo(tp
)) {
2251 /* Plain luck! Hole if filled with delayed
2252 * packet, rather than with a retransmit.
2254 if (tp
->retrans_out
== 0)
2255 tp
->retrans_stamp
= 0;
2257 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2260 tcp_undo_cwr(sk
, 0);
2261 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2263 /* So... Do not make Hoe's retransmit yet.
2264 * If the first packet was delayed, the rest
2265 * ones are most probably delayed as well.
2272 /* Undo during loss recovery after partial ACK. */
2273 static int tcp_try_undo_loss(struct sock
*sk
)
2275 struct tcp_sock
*tp
= tcp_sk(sk
);
2277 if (tcp_may_undo(tp
)) {
2278 struct sk_buff
*skb
;
2279 tcp_for_write_queue(skb
, sk
) {
2280 if (skb
== tcp_send_head(sk
))
2282 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2285 tcp_clear_all_retrans_hints(tp
);
2287 DBGUNDO(sk
, "partial loss");
2289 tcp_undo_cwr(sk
, 1);
2290 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2291 inet_csk(sk
)->icsk_retransmits
= 0;
2292 tp
->undo_marker
= 0;
2293 if (tcp_is_sack(tp
))
2294 tcp_set_ca_state(sk
, TCP_CA_Open
);
2300 static inline void tcp_complete_cwr(struct sock
*sk
)
2302 struct tcp_sock
*tp
= tcp_sk(sk
);
2303 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2304 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2305 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2308 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2310 struct tcp_sock
*tp
= tcp_sk(sk
);
2312 tcp_verify_left_out(tp
);
2314 if (tp
->retrans_out
== 0)
2315 tp
->retrans_stamp
= 0;
2318 tcp_enter_cwr(sk
, 1);
2320 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2321 int state
= TCP_CA_Open
;
2323 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2324 state
= TCP_CA_Disorder
;
2326 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2327 tcp_set_ca_state(sk
, state
);
2328 tp
->high_seq
= tp
->snd_nxt
;
2330 tcp_moderate_cwnd(tp
);
2332 tcp_cwnd_down(sk
, flag
);
2336 static void tcp_mtup_probe_failed(struct sock
*sk
)
2338 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2340 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2341 icsk
->icsk_mtup
.probe_size
= 0;
2344 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2346 struct tcp_sock
*tp
= tcp_sk(sk
);
2347 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2349 /* FIXME: breaks with very large cwnd */
2350 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2351 tp
->snd_cwnd
= tp
->snd_cwnd
*
2352 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2353 icsk
->icsk_mtup
.probe_size
;
2354 tp
->snd_cwnd_cnt
= 0;
2355 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2356 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2358 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2359 icsk
->icsk_mtup
.probe_size
= 0;
2360 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2364 /* Process an event, which can update packets-in-flight not trivially.
2365 * Main goal of this function is to calculate new estimate for left_out,
2366 * taking into account both packets sitting in receiver's buffer and
2367 * packets lost by network.
2369 * Besides that it does CWND reduction, when packet loss is detected
2370 * and changes state of machine.
2372 * It does _not_ decide what to send, it is made in function
2373 * tcp_xmit_retransmit_queue().
2376 tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2378 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2379 struct tcp_sock
*tp
= tcp_sk(sk
);
2380 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2381 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2382 (tp
->fackets_out
> tp
->reordering
));
2384 /* Some technical things:
2385 * 1. Reno does not count dupacks (sacked_out) automatically. */
2386 if (!tp
->packets_out
)
2389 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2390 tp
->fackets_out
= 0;
2392 /* Now state machine starts.
2393 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2395 tp
->prior_ssthresh
= 0;
2397 /* B. In all the states check for reneging SACKs. */
2398 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2401 /* C. Process data loss notification, provided it is valid. */
2402 if ((flag
&FLAG_DATA_LOST
) &&
2403 before(tp
->snd_una
, tp
->high_seq
) &&
2404 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2405 tp
->fackets_out
> tp
->reordering
) {
2406 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2407 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2410 /* D. Check consistency of the current state. */
2411 tcp_verify_left_out(tp
);
2413 /* E. Check state exit conditions. State can be terminated
2414 * when high_seq is ACKed. */
2415 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2416 BUG_TRAP(tp
->retrans_out
== 0);
2417 tp
->retrans_stamp
= 0;
2418 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2419 switch (icsk
->icsk_ca_state
) {
2421 icsk
->icsk_retransmits
= 0;
2422 if (tcp_try_undo_recovery(sk
))
2427 /* CWR is to be held something *above* high_seq
2428 * is ACKed for CWR bit to reach receiver. */
2429 if (tp
->snd_una
!= tp
->high_seq
) {
2430 tcp_complete_cwr(sk
);
2431 tcp_set_ca_state(sk
, TCP_CA_Open
);
2435 case TCP_CA_Disorder
:
2436 tcp_try_undo_dsack(sk
);
2437 if (!tp
->undo_marker
||
2438 /* For SACK case do not Open to allow to undo
2439 * catching for all duplicate ACKs. */
2440 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2441 tp
->undo_marker
= 0;
2442 tcp_set_ca_state(sk
, TCP_CA_Open
);
2446 case TCP_CA_Recovery
:
2447 if (tcp_is_reno(tp
))
2448 tcp_reset_reno_sack(tp
);
2449 if (tcp_try_undo_recovery(sk
))
2451 tcp_complete_cwr(sk
);
2456 /* F. Process state. */
2457 switch (icsk
->icsk_ca_state
) {
2458 case TCP_CA_Recovery
:
2459 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2460 if (tcp_is_reno(tp
) && is_dupack
)
2461 tcp_add_reno_sack(sk
);
2463 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2466 if (flag
&FLAG_DATA_ACKED
)
2467 icsk
->icsk_retransmits
= 0;
2468 if (!tcp_try_undo_loss(sk
)) {
2469 tcp_moderate_cwnd(tp
);
2470 tcp_xmit_retransmit_queue(sk
);
2473 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2475 /* Loss is undone; fall through to processing in Open state. */
2477 if (tcp_is_reno(tp
)) {
2478 if (flag
& FLAG_SND_UNA_ADVANCED
)
2479 tcp_reset_reno_sack(tp
);
2481 tcp_add_reno_sack(sk
);
2484 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2485 tcp_try_undo_dsack(sk
);
2487 if (!tcp_time_to_recover(sk
)) {
2488 tcp_try_to_open(sk
, flag
);
2492 /* MTU probe failure: don't reduce cwnd */
2493 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2494 icsk
->icsk_mtup
.probe_size
&&
2495 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2496 tcp_mtup_probe_failed(sk
);
2497 /* Restores the reduction we did in tcp_mtup_probe() */
2499 tcp_simple_retransmit(sk
);
2503 /* Otherwise enter Recovery state */
2505 if (tcp_is_reno(tp
))
2506 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2508 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2510 tp
->high_seq
= tp
->snd_nxt
;
2511 tp
->prior_ssthresh
= 0;
2512 tp
->undo_marker
= tp
->snd_una
;
2513 tp
->undo_retrans
= tp
->retrans_out
;
2515 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2516 if (!(flag
&FLAG_ECE
))
2517 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2518 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2519 TCP_ECN_queue_cwr(tp
);
2522 tp
->bytes_acked
= 0;
2523 tp
->snd_cwnd_cnt
= 0;
2524 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2527 if (do_lost
|| tcp_head_timedout(sk
))
2528 tcp_update_scoreboard(sk
);
2529 tcp_cwnd_down(sk
, flag
);
2530 tcp_xmit_retransmit_queue(sk
);
2533 /* Read draft-ietf-tcplw-high-performance before mucking
2534 * with this code. (Supersedes RFC1323)
2536 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2538 /* RTTM Rule: A TSecr value received in a segment is used to
2539 * update the averaged RTT measurement only if the segment
2540 * acknowledges some new data, i.e., only if it advances the
2541 * left edge of the send window.
2543 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2544 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2546 * Changed: reset backoff as soon as we see the first valid sample.
2547 * If we do not, we get strongly overestimated rto. With timestamps
2548 * samples are accepted even from very old segments: f.e., when rtt=1
2549 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2550 * answer arrives rto becomes 120 seconds! If at least one of segments
2551 * in window is lost... Voila. --ANK (010210)
2553 struct tcp_sock
*tp
= tcp_sk(sk
);
2554 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2555 tcp_rtt_estimator(sk
, seq_rtt
);
2557 inet_csk(sk
)->icsk_backoff
= 0;
2561 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2563 /* We don't have a timestamp. Can only use
2564 * packets that are not retransmitted to determine
2565 * rtt estimates. Also, we must not reset the
2566 * backoff for rto until we get a non-retransmitted
2567 * packet. This allows us to deal with a situation
2568 * where the network delay has increased suddenly.
2569 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2572 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2575 tcp_rtt_estimator(sk
, seq_rtt
);
2577 inet_csk(sk
)->icsk_backoff
= 0;
2581 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2584 const struct tcp_sock
*tp
= tcp_sk(sk
);
2585 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2586 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2587 tcp_ack_saw_tstamp(sk
, flag
);
2588 else if (seq_rtt
>= 0)
2589 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2592 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
,
2593 u32 in_flight
, int good
)
2595 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2596 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
, good
);
2597 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2600 /* Restart timer after forward progress on connection.
2601 * RFC2988 recommends to restart timer to now+rto.
2603 static void tcp_rearm_rto(struct sock
*sk
)
2605 struct tcp_sock
*tp
= tcp_sk(sk
);
2607 if (!tp
->packets_out
) {
2608 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2610 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2614 /* If we get here, the whole TSO packet has not been acked. */
2615 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2617 struct tcp_sock
*tp
= tcp_sk(sk
);
2620 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2622 packets_acked
= tcp_skb_pcount(skb
);
2623 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2625 packets_acked
-= tcp_skb_pcount(skb
);
2627 if (packets_acked
) {
2628 BUG_ON(tcp_skb_pcount(skb
) == 0);
2629 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2632 return packets_acked
;
2635 /* Remove acknowledged frames from the retransmission queue. If our packet
2636 * is before the ack sequence we can discard it as it's confirmed to have
2637 * arrived at the other end.
2639 static int tcp_clean_rtx_queue(struct sock
*sk
, s32
*seq_rtt_p
,
2642 struct tcp_sock
*tp
= tcp_sk(sk
);
2643 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2644 struct sk_buff
*skb
;
2645 u32 now
= tcp_time_stamp
;
2646 int fully_acked
= 1;
2648 int prior_packets
= tp
->packets_out
;
2650 u32 reord
= tp
->packets_out
;
2652 s32 ca_seq_rtt
= -1;
2653 ktime_t last_ackt
= net_invalid_timestamp();
2655 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2656 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2659 u8 sacked
= scb
->sacked
;
2661 /* Determine how many packets and what bytes were acked, tso and else */
2662 if (after(scb
->end_seq
, tp
->snd_una
)) {
2663 if (tcp_skb_pcount(skb
) == 1 ||
2664 !after(tp
->snd_una
, scb
->seq
))
2667 packets_acked
= tcp_tso_acked(sk
, skb
);
2672 end_seq
= tp
->snd_una
;
2674 packets_acked
= tcp_skb_pcount(skb
);
2675 end_seq
= scb
->end_seq
;
2678 /* MTU probing checks */
2679 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2680 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2681 tcp_mtup_probe_success(sk
, skb
);
2685 if (sacked
& TCPCB_RETRANS
) {
2686 if (sacked
& TCPCB_SACKED_RETRANS
)
2687 tp
->retrans_out
-= packets_acked
;
2688 flag
|= FLAG_RETRANS_DATA_ACKED
;
2691 if ((flag
& FLAG_DATA_ACKED
) ||
2692 (packets_acked
> 1))
2693 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2695 ca_seq_rtt
= now
- scb
->when
;
2696 last_ackt
= skb
->tstamp
;
2698 seq_rtt
= ca_seq_rtt
;
2700 if (!(sacked
& TCPCB_SACKED_ACKED
))
2701 reord
= min(cnt
, reord
);
2704 if (sacked
& TCPCB_SACKED_ACKED
)
2705 tp
->sacked_out
-= packets_acked
;
2706 if (sacked
& TCPCB_LOST
)
2707 tp
->lost_out
-= packets_acked
;
2709 if ((sacked
& TCPCB_URG
) && tp
->urg_mode
&&
2710 !before(end_seq
, tp
->snd_up
))
2713 ca_seq_rtt
= now
- scb
->when
;
2714 last_ackt
= skb
->tstamp
;
2716 seq_rtt
= ca_seq_rtt
;
2718 reord
= min(cnt
, reord
);
2720 tp
->packets_out
-= packets_acked
;
2721 cnt
+= packets_acked
;
2723 /* Initial outgoing SYN's get put onto the write_queue
2724 * just like anything else we transmit. It is not
2725 * true data, and if we misinform our callers that
2726 * this ACK acks real data, we will erroneously exit
2727 * connection startup slow start one packet too
2728 * quickly. This is severely frowned upon behavior.
2730 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2731 flag
|= FLAG_DATA_ACKED
;
2733 flag
|= FLAG_SYN_ACKED
;
2734 tp
->retrans_stamp
= 0;
2740 tcp_unlink_write_queue(skb
, sk
);
2741 sk_stream_free_skb(sk
, skb
);
2742 tcp_clear_all_retrans_hints(tp
);
2745 if (flag
& FLAG_ACKED
) {
2746 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2747 const struct tcp_congestion_ops
*ca_ops
2748 = inet_csk(sk
)->icsk_ca_ops
;
2750 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2753 if (tcp_is_reno(tp
)) {
2754 tcp_remove_reno_sacks(sk
, pkts_acked
);
2756 /* Non-retransmitted hole got filled? That's reordering */
2757 if (reord
< prior_fackets
)
2758 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2761 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2762 /* hint's skb might be NULL but we don't need to care */
2763 tp
->fastpath_cnt_hint
-= min_t(u32
, pkts_acked
,
2764 tp
->fastpath_cnt_hint
);
2765 if (ca_ops
->pkts_acked
) {
2768 /* Is the ACK triggering packet unambiguous? */
2769 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2770 /* High resolution needed and available? */
2771 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2772 !ktime_equal(last_ackt
,
2773 net_invalid_timestamp()))
2774 rtt_us
= ktime_us_delta(ktime_get_real(),
2776 else if (ca_seq_rtt
> 0)
2777 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
2780 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2784 #if FASTRETRANS_DEBUG > 0
2785 BUG_TRAP((int)tp
->sacked_out
>= 0);
2786 BUG_TRAP((int)tp
->lost_out
>= 0);
2787 BUG_TRAP((int)tp
->retrans_out
>= 0);
2788 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2789 icsk
= inet_csk(sk
);
2791 printk(KERN_DEBUG
"Leak l=%u %d\n",
2792 tp
->lost_out
, icsk
->icsk_ca_state
);
2795 if (tp
->sacked_out
) {
2796 printk(KERN_DEBUG
"Leak s=%u %d\n",
2797 tp
->sacked_out
, icsk
->icsk_ca_state
);
2800 if (tp
->retrans_out
) {
2801 printk(KERN_DEBUG
"Leak r=%u %d\n",
2802 tp
->retrans_out
, icsk
->icsk_ca_state
);
2803 tp
->retrans_out
= 0;
2807 *seq_rtt_p
= seq_rtt
;
2811 static void tcp_ack_probe(struct sock
*sk
)
2813 const struct tcp_sock
*tp
= tcp_sk(sk
);
2814 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2816 /* Was it a usable window open? */
2818 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2819 tp
->snd_una
+ tp
->snd_wnd
)) {
2820 icsk
->icsk_backoff
= 0;
2821 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2822 /* Socket must be waked up by subsequent tcp_data_snd_check().
2823 * This function is not for random using!
2826 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2827 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2832 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2834 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2835 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2838 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2840 const struct tcp_sock
*tp
= tcp_sk(sk
);
2841 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2842 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2845 /* Check that window update is acceptable.
2846 * The function assumes that snd_una<=ack<=snd_next.
2848 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2849 const u32 ack_seq
, const u32 nwin
)
2851 return (after(ack
, tp
->snd_una
) ||
2852 after(ack_seq
, tp
->snd_wl1
) ||
2853 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2856 /* Update our send window.
2858 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2859 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2861 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2864 struct tcp_sock
*tp
= tcp_sk(sk
);
2866 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2868 if (likely(!tcp_hdr(skb
)->syn
))
2869 nwin
<<= tp
->rx_opt
.snd_wscale
;
2871 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2872 flag
|= FLAG_WIN_UPDATE
;
2873 tcp_update_wl(tp
, ack
, ack_seq
);
2875 if (tp
->snd_wnd
!= nwin
) {
2878 /* Note, it is the only place, where
2879 * fast path is recovered for sending TCP.
2882 tcp_fast_path_check(sk
);
2884 if (nwin
> tp
->max_window
) {
2885 tp
->max_window
= nwin
;
2886 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
2896 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2897 * continue in congestion avoidance.
2899 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
2901 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2902 tp
->snd_cwnd_cnt
= 0;
2903 tp
->bytes_acked
= 0;
2904 TCP_ECN_queue_cwr(tp
);
2905 tcp_moderate_cwnd(tp
);
2908 /* A conservative spurious RTO response algorithm: reduce cwnd using
2909 * rate halving and continue in congestion avoidance.
2911 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
2913 tcp_enter_cwr(sk
, 0);
2916 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
2919 tcp_ratehalving_spur_to_response(sk
);
2921 tcp_undo_cwr(sk
, 1);
2924 /* F-RTO spurious RTO detection algorithm (RFC4138)
2926 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2927 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2928 * window (but not to or beyond highest sequence sent before RTO):
2929 * On First ACK, send two new segments out.
2930 * On Second ACK, RTO was likely spurious. Do spurious response (response
2931 * algorithm is not part of the F-RTO detection algorithm
2932 * given in RFC4138 but can be selected separately).
2933 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2934 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2935 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2936 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2938 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2939 * original window even after we transmit two new data segments.
2942 * on first step, wait until first cumulative ACK arrives, then move to
2943 * the second step. In second step, the next ACK decides.
2945 * F-RTO is implemented (mainly) in four functions:
2946 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2947 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2948 * called when tcp_use_frto() showed green light
2949 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2950 * - tcp_enter_frto_loss() is called if there is not enough evidence
2951 * to prove that the RTO is indeed spurious. It transfers the control
2952 * from F-RTO to the conventional RTO recovery
2954 static int tcp_process_frto(struct sock
*sk
, int flag
)
2956 struct tcp_sock
*tp
= tcp_sk(sk
);
2958 tcp_verify_left_out(tp
);
2960 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2961 if (flag
&FLAG_DATA_ACKED
)
2962 inet_csk(sk
)->icsk_retransmits
= 0;
2964 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
2965 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
2966 tp
->undo_marker
= 0;
2968 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
2969 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
2973 if (!IsSackFrto() || tcp_is_reno(tp
)) {
2974 /* RFC4138 shortcoming in step 2; should also have case c):
2975 * ACK isn't duplicate nor advances window, e.g., opposite dir
2978 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
2981 if (!(flag
&FLAG_DATA_ACKED
)) {
2982 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
2987 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
2988 /* Prevent sending of new data. */
2989 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2990 tcp_packets_in_flight(tp
));
2994 if ((tp
->frto_counter
>= 2) &&
2995 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
2996 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
2997 /* RFC4138 shortcoming (see comment above) */
2998 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3001 tcp_enter_frto_loss(sk
, 3, flag
);
3006 if (tp
->frto_counter
== 1) {
3007 /* tcp_may_send_now needs to see updated state */
3008 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3009 tp
->frto_counter
= 2;
3011 if (!tcp_may_send_now(sk
))
3012 tcp_enter_frto_loss(sk
, 2, flag
);
3016 switch (sysctl_tcp_frto_response
) {
3018 tcp_undo_spur_to_response(sk
, flag
);
3021 tcp_conservative_spur_to_response(tp
);
3024 tcp_ratehalving_spur_to_response(sk
);
3027 tp
->frto_counter
= 0;
3028 tp
->undo_marker
= 0;
3029 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS
);
3034 /* This routine deals with incoming acks, but not outgoing ones. */
3035 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3037 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3038 struct tcp_sock
*tp
= tcp_sk(sk
);
3039 u32 prior_snd_una
= tp
->snd_una
;
3040 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3041 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3042 u32 prior_in_flight
;
3048 /* If the ack is newer than sent or older than previous acks
3049 * then we can probably ignore it.
3051 if (after(ack
, tp
->snd_nxt
))
3052 goto uninteresting_ack
;
3054 if (before(ack
, prior_snd_una
))
3057 if (after(ack
, prior_snd_una
))
3058 flag
|= FLAG_SND_UNA_ADVANCED
;
3060 if (sysctl_tcp_abc
) {
3061 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3062 tp
->bytes_acked
+= ack
- prior_snd_una
;
3063 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3064 /* we assume just one segment left network */
3065 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
3068 prior_fackets
= tp
->fackets_out
;
3069 prior_in_flight
= tcp_packets_in_flight(tp
);
3071 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3072 /* Window is constant, pure forward advance.
3073 * No more checks are required.
3074 * Note, we use the fact that SND.UNA>=SND.WL2.
3076 tcp_update_wl(tp
, ack
, ack_seq
);
3078 flag
|= FLAG_WIN_UPDATE
;
3080 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3082 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
3084 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3087 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
3089 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3091 if (TCP_SKB_CB(skb
)->sacked
)
3092 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3094 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3097 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3100 /* We passed data and got it acked, remove any soft error
3101 * log. Something worked...
3103 sk
->sk_err_soft
= 0;
3104 tp
->rcv_tstamp
= tcp_time_stamp
;
3105 prior_packets
= tp
->packets_out
;
3109 /* See if we can take anything off of the retransmit queue. */
3110 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
, prior_fackets
);
3112 if (tp
->frto_counter
)
3113 frto_cwnd
= tcp_process_frto(sk
, flag
);
3114 /* Guarantee sacktag reordering detection against wrap-arounds */
3115 if (before(tp
->frto_highmark
, tp
->snd_una
))
3116 tp
->frto_highmark
= 0;
3118 if (tcp_ack_is_dubious(sk
, flag
)) {
3119 /* Advance CWND, if state allows this. */
3120 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3121 tcp_may_raise_cwnd(sk
, flag
))
3122 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 0);
3123 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
, flag
);
3125 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3126 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 1);
3129 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
3130 dst_confirm(sk
->sk_dst_cache
);
3135 icsk
->icsk_probes_out
= 0;
3137 /* If this ack opens up a zero window, clear backoff. It was
3138 * being used to time the probes, and is probably far higher than
3139 * it needs to be for normal retransmission.
3141 if (tcp_send_head(sk
))
3146 if (TCP_SKB_CB(skb
)->sacked
)
3147 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3150 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3155 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3156 * But, this can also be called on packets in the established flow when
3157 * the fast version below fails.
3159 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
3162 struct tcphdr
*th
= tcp_hdr(skb
);
3163 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
3165 ptr
= (unsigned char *)(th
+ 1);
3166 opt_rx
->saw_tstamp
= 0;
3168 while (length
> 0) {
3175 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3180 if (opsize
< 2) /* "silly options" */
3182 if (opsize
> length
)
3183 return; /* don't parse partial options */
3186 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3187 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
3189 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
3190 in_mss
= opt_rx
->user_mss
;
3191 opt_rx
->mss_clamp
= in_mss
;
3196 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
3197 if (sysctl_tcp_window_scaling
) {
3198 __u8 snd_wscale
= *(__u8
*) ptr
;
3199 opt_rx
->wscale_ok
= 1;
3200 if (snd_wscale
> 14) {
3201 if (net_ratelimit())
3202 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3203 "scaling value %d >14 received.\n",
3207 opt_rx
->snd_wscale
= snd_wscale
;
3210 case TCPOPT_TIMESTAMP
:
3211 if (opsize
==TCPOLEN_TIMESTAMP
) {
3212 if ((estab
&& opt_rx
->tstamp_ok
) ||
3213 (!estab
&& sysctl_tcp_timestamps
)) {
3214 opt_rx
->saw_tstamp
= 1;
3215 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
3216 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
3220 case TCPOPT_SACK_PERM
:
3221 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
3222 if (sysctl_tcp_sack
) {
3223 opt_rx
->sack_ok
= 1;
3224 tcp_sack_reset(opt_rx
);
3230 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3231 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3233 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3236 #ifdef CONFIG_TCP_MD5SIG
3239 * The MD5 Hash has already been
3240 * checked (see tcp_v{4,6}_do_rcv()).
3252 /* Fast parse options. This hopes to only see timestamps.
3253 * If it is wrong it falls back on tcp_parse_options().
3255 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3256 struct tcp_sock
*tp
)
3258 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3259 tp
->rx_opt
.saw_tstamp
= 0;
3261 } else if (tp
->rx_opt
.tstamp_ok
&&
3262 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3263 __be32
*ptr
= (__be32
*)(th
+ 1);
3264 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3265 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3266 tp
->rx_opt
.saw_tstamp
= 1;
3268 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3270 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3274 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3278 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3280 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3281 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3284 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3286 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3287 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3288 * extra check below makes sure this can only happen
3289 * for pure ACK frames. -DaveM
3291 * Not only, also it occurs for expired timestamps.
3294 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3295 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3296 tcp_store_ts_recent(tp
);
3300 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3302 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3303 * it can pass through stack. So, the following predicate verifies that
3304 * this segment is not used for anything but congestion avoidance or
3305 * fast retransmit. Moreover, we even are able to eliminate most of such
3306 * second order effects, if we apply some small "replay" window (~RTO)
3307 * to timestamp space.
3309 * All these measures still do not guarantee that we reject wrapped ACKs
3310 * on networks with high bandwidth, when sequence space is recycled fastly,
3311 * but it guarantees that such events will be very rare and do not affect
3312 * connection seriously. This doesn't look nice, but alas, PAWS is really
3315 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3316 * states that events when retransmit arrives after original data are rare.
3317 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3318 * the biggest problem on large power networks even with minor reordering.
3319 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3320 * up to bandwidth of 18Gigabit/sec. 8) ]
3323 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3325 struct tcp_sock
*tp
= tcp_sk(sk
);
3326 struct tcphdr
*th
= tcp_hdr(skb
);
3327 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3328 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3330 return (/* 1. Pure ACK with correct sequence number. */
3331 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3333 /* 2. ... and duplicate ACK. */
3334 ack
== tp
->snd_una
&&
3336 /* 3. ... and does not update window. */
3337 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3339 /* 4. ... and sits in replay window. */
3340 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3343 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3345 const struct tcp_sock
*tp
= tcp_sk(sk
);
3346 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3347 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3348 !tcp_disordered_ack(sk
, skb
));
3351 /* Check segment sequence number for validity.
3353 * Segment controls are considered valid, if the segment
3354 * fits to the window after truncation to the window. Acceptability
3355 * of data (and SYN, FIN, of course) is checked separately.
3356 * See tcp_data_queue(), for example.
3358 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3359 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3360 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3361 * (borrowed from freebsd)
3364 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3366 return !before(end_seq
, tp
->rcv_wup
) &&
3367 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3370 /* When we get a reset we do this. */
3371 static void tcp_reset(struct sock
*sk
)
3373 /* We want the right error as BSD sees it (and indeed as we do). */
3374 switch (sk
->sk_state
) {
3376 sk
->sk_err
= ECONNREFUSED
;
3378 case TCP_CLOSE_WAIT
:
3384 sk
->sk_err
= ECONNRESET
;
3387 if (!sock_flag(sk
, SOCK_DEAD
))
3388 sk
->sk_error_report(sk
);
3394 * Process the FIN bit. This now behaves as it is supposed to work
3395 * and the FIN takes effect when it is validly part of sequence
3396 * space. Not before when we get holes.
3398 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3399 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3402 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3403 * close and we go into CLOSING (and later onto TIME-WAIT)
3405 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3407 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3409 struct tcp_sock
*tp
= tcp_sk(sk
);
3411 inet_csk_schedule_ack(sk
);
3413 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3414 sock_set_flag(sk
, SOCK_DONE
);
3416 switch (sk
->sk_state
) {
3418 case TCP_ESTABLISHED
:
3419 /* Move to CLOSE_WAIT */
3420 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3421 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3424 case TCP_CLOSE_WAIT
:
3426 /* Received a retransmission of the FIN, do
3431 /* RFC793: Remain in the LAST-ACK state. */
3435 /* This case occurs when a simultaneous close
3436 * happens, we must ack the received FIN and
3437 * enter the CLOSING state.
3440 tcp_set_state(sk
, TCP_CLOSING
);
3443 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3445 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3448 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3449 * cases we should never reach this piece of code.
3451 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3452 __FUNCTION__
, sk
->sk_state
);
3456 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3457 * Probably, we should reset in this case. For now drop them.
3459 __skb_queue_purge(&tp
->out_of_order_queue
);
3460 if (tcp_is_sack(tp
))
3461 tcp_sack_reset(&tp
->rx_opt
);
3462 sk_stream_mem_reclaim(sk
);
3464 if (!sock_flag(sk
, SOCK_DEAD
)) {
3465 sk
->sk_state_change(sk
);
3467 /* Do not send POLL_HUP for half duplex close. */
3468 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3469 sk
->sk_state
== TCP_CLOSE
)
3470 sk_wake_async(sk
, 1, POLL_HUP
);
3472 sk_wake_async(sk
, 1, POLL_IN
);
3476 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3478 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3479 if (before(seq
, sp
->start_seq
))
3480 sp
->start_seq
= seq
;
3481 if (after(end_seq
, sp
->end_seq
))
3482 sp
->end_seq
= end_seq
;
3488 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3490 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3491 if (before(seq
, tp
->rcv_nxt
))
3492 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3494 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3496 tp
->rx_opt
.dsack
= 1;
3497 tp
->duplicate_sack
[0].start_seq
= seq
;
3498 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3499 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3503 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3505 if (!tp
->rx_opt
.dsack
)
3506 tcp_dsack_set(tp
, seq
, end_seq
);
3508 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3511 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3513 struct tcp_sock
*tp
= tcp_sk(sk
);
3515 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3516 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3517 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3518 tcp_enter_quickack_mode(sk
);
3520 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3521 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3523 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3524 end_seq
= tp
->rcv_nxt
;
3525 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3532 /* These routines update the SACK block as out-of-order packets arrive or
3533 * in-order packets close up the sequence space.
3535 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3538 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3539 struct tcp_sack_block
*swalk
= sp
+1;
3541 /* See if the recent change to the first SACK eats into
3542 * or hits the sequence space of other SACK blocks, if so coalesce.
3544 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3545 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3548 /* Zap SWALK, by moving every further SACK up by one slot.
3549 * Decrease num_sacks.
3551 tp
->rx_opt
.num_sacks
--;
3552 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3553 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3557 this_sack
++, swalk
++;
3561 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3565 tmp
= sack1
->start_seq
;
3566 sack1
->start_seq
= sack2
->start_seq
;
3567 sack2
->start_seq
= tmp
;
3569 tmp
= sack1
->end_seq
;
3570 sack1
->end_seq
= sack2
->end_seq
;
3571 sack2
->end_seq
= tmp
;
3574 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3576 struct tcp_sock
*tp
= tcp_sk(sk
);
3577 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3578 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3584 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3585 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3586 /* Rotate this_sack to the first one. */
3587 for (; this_sack
>0; this_sack
--, sp
--)
3588 tcp_sack_swap(sp
, sp
-1);
3590 tcp_sack_maybe_coalesce(tp
);
3595 /* Could not find an adjacent existing SACK, build a new one,
3596 * put it at the front, and shift everyone else down. We
3597 * always know there is at least one SACK present already here.
3599 * If the sack array is full, forget about the last one.
3601 if (this_sack
>= 4) {
3603 tp
->rx_opt
.num_sacks
--;
3606 for (; this_sack
> 0; this_sack
--, sp
--)
3610 /* Build the new head SACK, and we're done. */
3611 sp
->start_seq
= seq
;
3612 sp
->end_seq
= end_seq
;
3613 tp
->rx_opt
.num_sacks
++;
3614 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3617 /* RCV.NXT advances, some SACKs should be eaten. */
3619 static void tcp_sack_remove(struct tcp_sock
*tp
)
3621 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3622 int num_sacks
= tp
->rx_opt
.num_sacks
;
3625 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3626 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3627 tp
->rx_opt
.num_sacks
= 0;
3628 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3632 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3633 /* Check if the start of the sack is covered by RCV.NXT. */
3634 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3637 /* RCV.NXT must cover all the block! */
3638 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3640 /* Zap this SACK, by moving forward any other SACKS. */
3641 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3642 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3649 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3650 tp
->rx_opt
.num_sacks
= num_sacks
;
3651 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3655 /* This one checks to see if we can put data from the
3656 * out_of_order queue into the receive_queue.
3658 static void tcp_ofo_queue(struct sock
*sk
)
3660 struct tcp_sock
*tp
= tcp_sk(sk
);
3661 __u32 dsack_high
= tp
->rcv_nxt
;
3662 struct sk_buff
*skb
;
3664 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3665 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3668 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3669 __u32 dsack
= dsack_high
;
3670 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3671 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3672 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3675 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3676 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3677 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3681 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3682 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3683 TCP_SKB_CB(skb
)->end_seq
);
3685 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3686 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3687 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3688 if (tcp_hdr(skb
)->fin
)
3689 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3693 static int tcp_prune_queue(struct sock
*sk
);
3695 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3697 struct tcphdr
*th
= tcp_hdr(skb
);
3698 struct tcp_sock
*tp
= tcp_sk(sk
);
3701 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3704 __skb_pull(skb
, th
->doff
*4);
3706 TCP_ECN_accept_cwr(tp
, skb
);
3708 if (tp
->rx_opt
.dsack
) {
3709 tp
->rx_opt
.dsack
= 0;
3710 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3711 4 - tp
->rx_opt
.tstamp_ok
);
3714 /* Queue data for delivery to the user.
3715 * Packets in sequence go to the receive queue.
3716 * Out of sequence packets to the out_of_order_queue.
3718 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3719 if (tcp_receive_window(tp
) == 0)
3722 /* Ok. In sequence. In window. */
3723 if (tp
->ucopy
.task
== current
&&
3724 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3725 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3726 int chunk
= min_t(unsigned int, skb
->len
,
3729 __set_current_state(TASK_RUNNING
);
3732 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3733 tp
->ucopy
.len
-= chunk
;
3734 tp
->copied_seq
+= chunk
;
3735 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3736 tcp_rcv_space_adjust(sk
);
3744 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3745 !sk_stream_rmem_schedule(sk
, skb
))) {
3746 if (tcp_prune_queue(sk
) < 0 ||
3747 !sk_stream_rmem_schedule(sk
, skb
))
3750 sk_stream_set_owner_r(skb
, sk
);
3751 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3753 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3755 tcp_event_data_recv(sk
, skb
);
3757 tcp_fin(skb
, sk
, th
);
3759 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3762 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3763 * gap in queue is filled.
3765 if (skb_queue_empty(&tp
->out_of_order_queue
))
3766 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3769 if (tp
->rx_opt
.num_sacks
)
3770 tcp_sack_remove(tp
);
3772 tcp_fast_path_check(sk
);
3776 else if (!sock_flag(sk
, SOCK_DEAD
))
3777 sk
->sk_data_ready(sk
, 0);
3781 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3782 /* A retransmit, 2nd most common case. Force an immediate ack. */
3783 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3784 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3787 tcp_enter_quickack_mode(sk
);
3788 inet_csk_schedule_ack(sk
);
3794 /* Out of window. F.e. zero window probe. */
3795 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3798 tcp_enter_quickack_mode(sk
);
3800 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3801 /* Partial packet, seq < rcv_next < end_seq */
3802 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3803 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3804 TCP_SKB_CB(skb
)->end_seq
);
3806 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3808 /* If window is closed, drop tail of packet. But after
3809 * remembering D-SACK for its head made in previous line.
3811 if (!tcp_receive_window(tp
))
3816 TCP_ECN_check_ce(tp
, skb
);
3818 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3819 !sk_stream_rmem_schedule(sk
, skb
)) {
3820 if (tcp_prune_queue(sk
) < 0 ||
3821 !sk_stream_rmem_schedule(sk
, skb
))
3825 /* Disable header prediction. */
3827 inet_csk_schedule_ack(sk
);
3829 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3830 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3832 sk_stream_set_owner_r(skb
, sk
);
3834 if (!skb_peek(&tp
->out_of_order_queue
)) {
3835 /* Initial out of order segment, build 1 SACK. */
3836 if (tcp_is_sack(tp
)) {
3837 tp
->rx_opt
.num_sacks
= 1;
3838 tp
->rx_opt
.dsack
= 0;
3839 tp
->rx_opt
.eff_sacks
= 1;
3840 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3841 tp
->selective_acks
[0].end_seq
=
3842 TCP_SKB_CB(skb
)->end_seq
;
3844 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3846 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3847 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3848 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3850 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3851 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3853 if (!tp
->rx_opt
.num_sacks
||
3854 tp
->selective_acks
[0].end_seq
!= seq
)
3857 /* Common case: data arrive in order after hole. */
3858 tp
->selective_acks
[0].end_seq
= end_seq
;
3862 /* Find place to insert this segment. */
3864 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3866 } while ((skb1
= skb1
->prev
) !=
3867 (struct sk_buff
*)&tp
->out_of_order_queue
);
3869 /* Do skb overlap to previous one? */
3870 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3871 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3872 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3873 /* All the bits are present. Drop. */
3875 tcp_dsack_set(tp
, seq
, end_seq
);
3878 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3879 /* Partial overlap. */
3880 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
3885 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
3887 /* And clean segments covered by new one as whole. */
3888 while ((skb1
= skb
->next
) !=
3889 (struct sk_buff
*)&tp
->out_of_order_queue
&&
3890 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
3891 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3892 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
3895 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
3896 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
3901 if (tcp_is_sack(tp
))
3902 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
3906 /* Collapse contiguous sequence of skbs head..tail with
3907 * sequence numbers start..end.
3908 * Segments with FIN/SYN are not collapsed (only because this
3912 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
3913 struct sk_buff
*head
, struct sk_buff
*tail
,
3916 struct sk_buff
*skb
;
3918 /* First, check that queue is collapsible and find
3919 * the point where collapsing can be useful. */
3920 for (skb
= head
; skb
!= tail
; ) {
3921 /* No new bits? It is possible on ofo queue. */
3922 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3923 struct sk_buff
*next
= skb
->next
;
3924 __skb_unlink(skb
, list
);
3926 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3931 /* The first skb to collapse is:
3933 * - bloated or contains data before "start" or
3934 * overlaps to the next one.
3936 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
3937 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
3938 before(TCP_SKB_CB(skb
)->seq
, start
) ||
3939 (skb
->next
!= tail
&&
3940 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
3943 /* Decided to skip this, advance start seq. */
3944 start
= TCP_SKB_CB(skb
)->end_seq
;
3947 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
3950 while (before(start
, end
)) {
3951 struct sk_buff
*nskb
;
3952 unsigned int header
= skb_headroom(skb
);
3953 int copy
= SKB_MAX_ORDER(header
, 0);
3955 /* Too big header? This can happen with IPv6. */
3958 if (end
-start
< copy
)
3960 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
3964 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
3965 skb_set_network_header(nskb
, (skb_network_header(skb
) -
3967 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
3969 skb_reserve(nskb
, header
);
3970 memcpy(nskb
->head
, skb
->head
, header
);
3971 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
3972 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
3973 __skb_insert(nskb
, skb
->prev
, skb
, list
);
3974 sk_stream_set_owner_r(nskb
, sk
);
3976 /* Copy data, releasing collapsed skbs. */
3978 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
3979 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
3983 size
= min(copy
, size
);
3984 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
3986 TCP_SKB_CB(nskb
)->end_seq
+= size
;
3990 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
3991 struct sk_buff
*next
= skb
->next
;
3992 __skb_unlink(skb
, list
);
3994 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
3997 tcp_hdr(skb
)->syn
||
4005 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4006 * and tcp_collapse() them until all the queue is collapsed.
4008 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4010 struct tcp_sock
*tp
= tcp_sk(sk
);
4011 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4012 struct sk_buff
*head
;
4018 start
= TCP_SKB_CB(skb
)->seq
;
4019 end
= TCP_SKB_CB(skb
)->end_seq
;
4025 /* Segment is terminated when we see gap or when
4026 * we are at the end of all the queue. */
4027 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4028 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4029 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4030 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4031 head
, skb
, start
, end
);
4033 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4035 /* Start new segment */
4036 start
= TCP_SKB_CB(skb
)->seq
;
4037 end
= TCP_SKB_CB(skb
)->end_seq
;
4039 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4040 start
= TCP_SKB_CB(skb
)->seq
;
4041 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4042 end
= TCP_SKB_CB(skb
)->end_seq
;
4047 /* Reduce allocated memory if we can, trying to get
4048 * the socket within its memory limits again.
4050 * Return less than zero if we should start dropping frames
4051 * until the socket owning process reads some of the data
4052 * to stabilize the situation.
4054 static int tcp_prune_queue(struct sock
*sk
)
4056 struct tcp_sock
*tp
= tcp_sk(sk
);
4058 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4060 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
4062 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4063 tcp_clamp_window(sk
);
4064 else if (tcp_memory_pressure
)
4065 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4067 tcp_collapse_ofo_queue(sk
);
4068 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4069 sk
->sk_receive_queue
.next
,
4070 (struct sk_buff
*)&sk
->sk_receive_queue
,
4071 tp
->copied_seq
, tp
->rcv_nxt
);
4072 sk_stream_mem_reclaim(sk
);
4074 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4077 /* Collapsing did not help, destructive actions follow.
4078 * This must not ever occur. */
4080 /* First, purge the out_of_order queue. */
4081 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4082 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
4083 __skb_queue_purge(&tp
->out_of_order_queue
);
4085 /* Reset SACK state. A conforming SACK implementation will
4086 * do the same at a timeout based retransmit. When a connection
4087 * is in a sad state like this, we care only about integrity
4088 * of the connection not performance.
4090 if (tcp_is_sack(tp
))
4091 tcp_sack_reset(&tp
->rx_opt
);
4092 sk_stream_mem_reclaim(sk
);
4095 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4098 /* If we are really being abused, tell the caller to silently
4099 * drop receive data on the floor. It will get retransmitted
4100 * and hopefully then we'll have sufficient space.
4102 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
4104 /* Massive buffer overcommit. */
4110 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4111 * As additional protections, we do not touch cwnd in retransmission phases,
4112 * and if application hit its sndbuf limit recently.
4114 void tcp_cwnd_application_limited(struct sock
*sk
)
4116 struct tcp_sock
*tp
= tcp_sk(sk
);
4118 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4119 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4120 /* Limited by application or receiver window. */
4121 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4122 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4123 if (win_used
< tp
->snd_cwnd
) {
4124 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4125 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4127 tp
->snd_cwnd_used
= 0;
4129 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4132 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4134 struct tcp_sock
*tp
= tcp_sk(sk
);
4136 /* If the user specified a specific send buffer setting, do
4139 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4142 /* If we are under global TCP memory pressure, do not expand. */
4143 if (tcp_memory_pressure
)
4146 /* If we are under soft global TCP memory pressure, do not expand. */
4147 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4150 /* If we filled the congestion window, do not expand. */
4151 if (tp
->packets_out
>= tp
->snd_cwnd
)
4157 /* When incoming ACK allowed to free some skb from write_queue,
4158 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4159 * on the exit from tcp input handler.
4161 * PROBLEM: sndbuf expansion does not work well with largesend.
4163 static void tcp_new_space(struct sock
*sk
)
4165 struct tcp_sock
*tp
= tcp_sk(sk
);
4167 if (tcp_should_expand_sndbuf(sk
)) {
4168 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4169 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4170 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4171 tp
->reordering
+ 1);
4172 sndmem
*= 2*demanded
;
4173 if (sndmem
> sk
->sk_sndbuf
)
4174 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4175 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4178 sk
->sk_write_space(sk
);
4181 static void tcp_check_space(struct sock
*sk
)
4183 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4184 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4185 if (sk
->sk_socket
&&
4186 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4191 static inline void tcp_data_snd_check(struct sock
*sk
)
4193 tcp_push_pending_frames(sk
);
4194 tcp_check_space(sk
);
4198 * Check if sending an ack is needed.
4200 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4202 struct tcp_sock
*tp
= tcp_sk(sk
);
4204 /* More than one full frame received... */
4205 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4206 /* ... and right edge of window advances far enough.
4207 * (tcp_recvmsg() will send ACK otherwise). Or...
4209 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4210 /* We ACK each frame or... */
4211 tcp_in_quickack_mode(sk
) ||
4212 /* We have out of order data. */
4214 skb_peek(&tp
->out_of_order_queue
))) {
4215 /* Then ack it now */
4218 /* Else, send delayed ack. */
4219 tcp_send_delayed_ack(sk
);
4223 static inline void tcp_ack_snd_check(struct sock
*sk
)
4225 if (!inet_csk_ack_scheduled(sk
)) {
4226 /* We sent a data segment already. */
4229 __tcp_ack_snd_check(sk
, 1);
4233 * This routine is only called when we have urgent data
4234 * signaled. Its the 'slow' part of tcp_urg. It could be
4235 * moved inline now as tcp_urg is only called from one
4236 * place. We handle URGent data wrong. We have to - as
4237 * BSD still doesn't use the correction from RFC961.
4238 * For 1003.1g we should support a new option TCP_STDURG to permit
4239 * either form (or just set the sysctl tcp_stdurg).
4242 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4244 struct tcp_sock
*tp
= tcp_sk(sk
);
4245 u32 ptr
= ntohs(th
->urg_ptr
);
4247 if (ptr
&& !sysctl_tcp_stdurg
)
4249 ptr
+= ntohl(th
->seq
);
4251 /* Ignore urgent data that we've already seen and read. */
4252 if (after(tp
->copied_seq
, ptr
))
4255 /* Do not replay urg ptr.
4257 * NOTE: interesting situation not covered by specs.
4258 * Misbehaving sender may send urg ptr, pointing to segment,
4259 * which we already have in ofo queue. We are not able to fetch
4260 * such data and will stay in TCP_URG_NOTYET until will be eaten
4261 * by recvmsg(). Seems, we are not obliged to handle such wicked
4262 * situations. But it is worth to think about possibility of some
4263 * DoSes using some hypothetical application level deadlock.
4265 if (before(ptr
, tp
->rcv_nxt
))
4268 /* Do we already have a newer (or duplicate) urgent pointer? */
4269 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4272 /* Tell the world about our new urgent pointer. */
4275 /* We may be adding urgent data when the last byte read was
4276 * urgent. To do this requires some care. We cannot just ignore
4277 * tp->copied_seq since we would read the last urgent byte again
4278 * as data, nor can we alter copied_seq until this data arrives
4279 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4281 * NOTE. Double Dutch. Rendering to plain English: author of comment
4282 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4283 * and expect that both A and B disappear from stream. This is _wrong_.
4284 * Though this happens in BSD with high probability, this is occasional.
4285 * Any application relying on this is buggy. Note also, that fix "works"
4286 * only in this artificial test. Insert some normal data between A and B and we will
4287 * decline of BSD again. Verdict: it is better to remove to trap
4290 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4291 !sock_flag(sk
, SOCK_URGINLINE
) &&
4292 tp
->copied_seq
!= tp
->rcv_nxt
) {
4293 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4295 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4296 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4301 tp
->urg_data
= TCP_URG_NOTYET
;
4304 /* Disable header prediction. */
4308 /* This is the 'fast' part of urgent handling. */
4309 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4311 struct tcp_sock
*tp
= tcp_sk(sk
);
4313 /* Check if we get a new urgent pointer - normally not. */
4315 tcp_check_urg(sk
,th
);
4317 /* Do we wait for any urgent data? - normally not... */
4318 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4319 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4322 /* Is the urgent pointer pointing into this packet? */
4323 if (ptr
< skb
->len
) {
4325 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4327 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4328 if (!sock_flag(sk
, SOCK_DEAD
))
4329 sk
->sk_data_ready(sk
, 0);
4334 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4336 struct tcp_sock
*tp
= tcp_sk(sk
);
4337 int chunk
= skb
->len
- hlen
;
4341 if (skb_csum_unnecessary(skb
))
4342 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4344 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4348 tp
->ucopy
.len
-= chunk
;
4349 tp
->copied_seq
+= chunk
;
4350 tcp_rcv_space_adjust(sk
);
4357 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4361 if (sock_owned_by_user(sk
)) {
4363 result
= __tcp_checksum_complete(skb
);
4366 result
= __tcp_checksum_complete(skb
);
4371 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4373 return !skb_csum_unnecessary(skb
) &&
4374 __tcp_checksum_complete_user(sk
, skb
);
4377 #ifdef CONFIG_NET_DMA
4378 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4380 struct tcp_sock
*tp
= tcp_sk(sk
);
4381 int chunk
= skb
->len
- hlen
;
4383 int copied_early
= 0;
4385 if (tp
->ucopy
.wakeup
)
4388 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4389 tp
->ucopy
.dma_chan
= get_softnet_dma();
4391 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4393 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4394 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4399 tp
->ucopy
.dma_cookie
= dma_cookie
;
4402 tp
->ucopy
.len
-= chunk
;
4403 tp
->copied_seq
+= chunk
;
4404 tcp_rcv_space_adjust(sk
);
4406 if ((tp
->ucopy
.len
== 0) ||
4407 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4408 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4409 tp
->ucopy
.wakeup
= 1;
4410 sk
->sk_data_ready(sk
, 0);
4412 } else if (chunk
> 0) {
4413 tp
->ucopy
.wakeup
= 1;
4414 sk
->sk_data_ready(sk
, 0);
4417 return copied_early
;
4419 #endif /* CONFIG_NET_DMA */
4422 * TCP receive function for the ESTABLISHED state.
4424 * It is split into a fast path and a slow path. The fast path is
4426 * - A zero window was announced from us - zero window probing
4427 * is only handled properly in the slow path.
4428 * - Out of order segments arrived.
4429 * - Urgent data is expected.
4430 * - There is no buffer space left
4431 * - Unexpected TCP flags/window values/header lengths are received
4432 * (detected by checking the TCP header against pred_flags)
4433 * - Data is sent in both directions. Fast path only supports pure senders
4434 * or pure receivers (this means either the sequence number or the ack
4435 * value must stay constant)
4436 * - Unexpected TCP option.
4438 * When these conditions are not satisfied it drops into a standard
4439 * receive procedure patterned after RFC793 to handle all cases.
4440 * The first three cases are guaranteed by proper pred_flags setting,
4441 * the rest is checked inline. Fast processing is turned on in
4442 * tcp_data_queue when everything is OK.
4444 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4445 struct tcphdr
*th
, unsigned len
)
4447 struct tcp_sock
*tp
= tcp_sk(sk
);
4450 * Header prediction.
4451 * The code loosely follows the one in the famous
4452 * "30 instruction TCP receive" Van Jacobson mail.
4454 * Van's trick is to deposit buffers into socket queue
4455 * on a device interrupt, to call tcp_recv function
4456 * on the receive process context and checksum and copy
4457 * the buffer to user space. smart...
4459 * Our current scheme is not silly either but we take the
4460 * extra cost of the net_bh soft interrupt processing...
4461 * We do checksum and copy also but from device to kernel.
4464 tp
->rx_opt
.saw_tstamp
= 0;
4466 /* pred_flags is 0xS?10 << 16 + snd_wnd
4467 * if header_prediction is to be made
4468 * 'S' will always be tp->tcp_header_len >> 2
4469 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4470 * turn it off (when there are holes in the receive
4471 * space for instance)
4472 * PSH flag is ignored.
4475 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4476 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4477 int tcp_header_len
= tp
->tcp_header_len
;
4479 /* Timestamp header prediction: tcp_header_len
4480 * is automatically equal to th->doff*4 due to pred_flags
4484 /* Check timestamp */
4485 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4486 __be32
*ptr
= (__be32
*)(th
+ 1);
4488 /* No? Slow path! */
4489 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4490 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4493 tp
->rx_opt
.saw_tstamp
= 1;
4495 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4497 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4499 /* If PAWS failed, check it more carefully in slow path */
4500 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4503 /* DO NOT update ts_recent here, if checksum fails
4504 * and timestamp was corrupted part, it will result
4505 * in a hung connection since we will drop all
4506 * future packets due to the PAWS test.
4510 if (len
<= tcp_header_len
) {
4511 /* Bulk data transfer: sender */
4512 if (len
== tcp_header_len
) {
4513 /* Predicted packet is in window by definition.
4514 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4515 * Hence, check seq<=rcv_wup reduces to:
4517 if (tcp_header_len
==
4518 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4519 tp
->rcv_nxt
== tp
->rcv_wup
)
4520 tcp_store_ts_recent(tp
);
4522 /* We know that such packets are checksummed
4525 tcp_ack(sk
, skb
, 0);
4527 tcp_data_snd_check(sk
);
4529 } else { /* Header too small */
4530 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4535 int copied_early
= 0;
4537 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4538 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4539 #ifdef CONFIG_NET_DMA
4540 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4545 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4546 __set_current_state(TASK_RUNNING
);
4548 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4552 /* Predicted packet is in window by definition.
4553 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4554 * Hence, check seq<=rcv_wup reduces to:
4556 if (tcp_header_len
==
4557 (sizeof(struct tcphdr
) +
4558 TCPOLEN_TSTAMP_ALIGNED
) &&
4559 tp
->rcv_nxt
== tp
->rcv_wup
)
4560 tcp_store_ts_recent(tp
);
4562 tcp_rcv_rtt_measure_ts(sk
, skb
);
4564 __skb_pull(skb
, tcp_header_len
);
4565 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4566 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4569 tcp_cleanup_rbuf(sk
, skb
->len
);
4572 if (tcp_checksum_complete_user(sk
, skb
))
4575 /* Predicted packet is in window by definition.
4576 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4577 * Hence, check seq<=rcv_wup reduces to:
4579 if (tcp_header_len
==
4580 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4581 tp
->rcv_nxt
== tp
->rcv_wup
)
4582 tcp_store_ts_recent(tp
);
4584 tcp_rcv_rtt_measure_ts(sk
, skb
);
4586 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4589 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4591 /* Bulk data transfer: receiver */
4592 __skb_pull(skb
,tcp_header_len
);
4593 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4594 sk_stream_set_owner_r(skb
, sk
);
4595 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4598 tcp_event_data_recv(sk
, skb
);
4600 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4601 /* Well, only one small jumplet in fast path... */
4602 tcp_ack(sk
, skb
, FLAG_DATA
);
4603 tcp_data_snd_check(sk
);
4604 if (!inet_csk_ack_scheduled(sk
))
4608 __tcp_ack_snd_check(sk
, 0);
4610 #ifdef CONFIG_NET_DMA
4612 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4618 sk
->sk_data_ready(sk
, 0);
4624 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4628 * RFC1323: H1. Apply PAWS check first.
4630 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4631 tcp_paws_discard(sk
, skb
)) {
4633 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4634 tcp_send_dupack(sk
, skb
);
4637 /* Resets are accepted even if PAWS failed.
4639 ts_recent update must be made after we are sure
4640 that the packet is in window.
4645 * Standard slow path.
4648 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4649 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4650 * (RST) segments are validated by checking their SEQ-fields."
4651 * And page 69: "If an incoming segment is not acceptable,
4652 * an acknowledgment should be sent in reply (unless the RST bit
4653 * is set, if so drop the segment and return)".
4656 tcp_send_dupack(sk
, skb
);
4665 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4667 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4668 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4669 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4676 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4678 tcp_rcv_rtt_measure_ts(sk
, skb
);
4680 /* Process urgent data. */
4681 tcp_urg(sk
, skb
, th
);
4683 /* step 7: process the segment text */
4684 tcp_data_queue(sk
, skb
);
4686 tcp_data_snd_check(sk
);
4687 tcp_ack_snd_check(sk
);
4691 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4698 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4699 struct tcphdr
*th
, unsigned len
)
4701 struct tcp_sock
*tp
= tcp_sk(sk
);
4702 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4703 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4705 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4709 * "If the state is SYN-SENT then
4710 * first check the ACK bit
4711 * If the ACK bit is set
4712 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4713 * a reset (unless the RST bit is set, if so drop
4714 * the segment and return)"
4716 * We do not send data with SYN, so that RFC-correct
4719 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4720 goto reset_and_undo
;
4722 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4723 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4725 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4726 goto reset_and_undo
;
4729 /* Now ACK is acceptable.
4731 * "If the RST bit is set
4732 * If the ACK was acceptable then signal the user "error:
4733 * connection reset", drop the segment, enter CLOSED state,
4734 * delete TCB, and return."
4743 * "fifth, if neither of the SYN or RST bits is set then
4744 * drop the segment and return."
4750 goto discard_and_undo
;
4753 * "If the SYN bit is on ...
4754 * are acceptable then ...
4755 * (our SYN has been ACKed), change the connection
4756 * state to ESTABLISHED..."
4759 TCP_ECN_rcv_synack(tp
, th
);
4761 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4762 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4764 /* Ok.. it's good. Set up sequence numbers and
4765 * move to established.
4767 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4768 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4770 /* RFC1323: The window in SYN & SYN/ACK segments is
4773 tp
->snd_wnd
= ntohs(th
->window
);
4774 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4776 if (!tp
->rx_opt
.wscale_ok
) {
4777 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4778 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4781 if (tp
->rx_opt
.saw_tstamp
) {
4782 tp
->rx_opt
.tstamp_ok
= 1;
4783 tp
->tcp_header_len
=
4784 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4785 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4786 tcp_store_ts_recent(tp
);
4788 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4791 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
4792 tcp_enable_fack(tp
);
4795 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4796 tcp_initialize_rcv_mss(sk
);
4798 /* Remember, tcp_poll() does not lock socket!
4799 * Change state from SYN-SENT only after copied_seq
4800 * is initialized. */
4801 tp
->copied_seq
= tp
->rcv_nxt
;
4803 tcp_set_state(sk
, TCP_ESTABLISHED
);
4805 security_inet_conn_established(sk
, skb
);
4807 /* Make sure socket is routed, for correct metrics. */
4808 icsk
->icsk_af_ops
->rebuild_header(sk
);
4810 tcp_init_metrics(sk
);
4812 tcp_init_congestion_control(sk
);
4814 /* Prevent spurious tcp_cwnd_restart() on first data
4817 tp
->lsndtime
= tcp_time_stamp
;
4819 tcp_init_buffer_space(sk
);
4821 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4822 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4824 if (!tp
->rx_opt
.snd_wscale
)
4825 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4829 if (!sock_flag(sk
, SOCK_DEAD
)) {
4830 sk
->sk_state_change(sk
);
4831 sk_wake_async(sk
, 0, POLL_OUT
);
4834 if (sk
->sk_write_pending
||
4835 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4836 icsk
->icsk_ack
.pingpong
) {
4837 /* Save one ACK. Data will be ready after
4838 * several ticks, if write_pending is set.
4840 * It may be deleted, but with this feature tcpdumps
4841 * look so _wonderfully_ clever, that I was not able
4842 * to stand against the temptation 8) --ANK
4844 inet_csk_schedule_ack(sk
);
4845 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4846 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4847 tcp_incr_quickack(sk
);
4848 tcp_enter_quickack_mode(sk
);
4849 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4850 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4861 /* No ACK in the segment */
4865 * "If the RST bit is set
4867 * Otherwise (no ACK) drop the segment and return."
4870 goto discard_and_undo
;
4874 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4875 goto discard_and_undo
;
4878 /* We see SYN without ACK. It is attempt of
4879 * simultaneous connect with crossed SYNs.
4880 * Particularly, it can be connect to self.
4882 tcp_set_state(sk
, TCP_SYN_RECV
);
4884 if (tp
->rx_opt
.saw_tstamp
) {
4885 tp
->rx_opt
.tstamp_ok
= 1;
4886 tcp_store_ts_recent(tp
);
4887 tp
->tcp_header_len
=
4888 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4890 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4893 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4894 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4896 /* RFC1323: The window in SYN & SYN/ACK segments is
4899 tp
->snd_wnd
= ntohs(th
->window
);
4900 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4901 tp
->max_window
= tp
->snd_wnd
;
4903 TCP_ECN_rcv_syn(tp
, th
);
4906 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4907 tcp_initialize_rcv_mss(sk
);
4910 tcp_send_synack(sk
);
4912 /* Note, we could accept data and URG from this segment.
4913 * There are no obstacles to make this.
4915 * However, if we ignore data in ACKless segments sometimes,
4916 * we have no reasons to accept it sometimes.
4917 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4918 * is not flawless. So, discard packet for sanity.
4919 * Uncomment this return to process the data.
4926 /* "fifth, if neither of the SYN or RST bits is set then
4927 * drop the segment and return."
4931 tcp_clear_options(&tp
->rx_opt
);
4932 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4936 tcp_clear_options(&tp
->rx_opt
);
4937 tp
->rx_opt
.mss_clamp
= saved_clamp
;
4943 * This function implements the receiving procedure of RFC 793 for
4944 * all states except ESTABLISHED and TIME_WAIT.
4945 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4946 * address independent.
4949 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4950 struct tcphdr
*th
, unsigned len
)
4952 struct tcp_sock
*tp
= tcp_sk(sk
);
4953 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4956 tp
->rx_opt
.saw_tstamp
= 0;
4958 switch (sk
->sk_state
) {
4970 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
4973 /* Now we have several options: In theory there is
4974 * nothing else in the frame. KA9Q has an option to
4975 * send data with the syn, BSD accepts data with the
4976 * syn up to the [to be] advertised window and
4977 * Solaris 2.1 gives you a protocol error. For now
4978 * we just ignore it, that fits the spec precisely
4979 * and avoids incompatibilities. It would be nice in
4980 * future to drop through and process the data.
4982 * Now that TTCP is starting to be used we ought to
4984 * But, this leaves one open to an easy denial of
4985 * service attack, and SYN cookies can't defend
4986 * against this problem. So, we drop the data
4987 * in the interest of security over speed unless
4988 * it's still in use.
4996 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5000 /* Do step6 onward by hand. */
5001 tcp_urg(sk
, skb
, th
);
5003 tcp_data_snd_check(sk
);
5007 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5008 tcp_paws_discard(sk
, skb
)) {
5010 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
5011 tcp_send_dupack(sk
, skb
);
5014 /* Reset is accepted even if it did not pass PAWS. */
5017 /* step 1: check sequence number */
5018 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5020 tcp_send_dupack(sk
, skb
);
5024 /* step 2: check RST bit */
5030 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5032 /* step 3: check security and precedence [ignored] */
5036 * Check for a SYN in window.
5038 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5039 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
5044 /* step 5: check the ACK field */
5046 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5048 switch (sk
->sk_state
) {
5051 tp
->copied_seq
= tp
->rcv_nxt
;
5053 tcp_set_state(sk
, TCP_ESTABLISHED
);
5054 sk
->sk_state_change(sk
);
5056 /* Note, that this wakeup is only for marginal
5057 * crossed SYN case. Passively open sockets
5058 * are not waked up, because sk->sk_sleep ==
5059 * NULL and sk->sk_socket == NULL.
5061 if (sk
->sk_socket
) {
5062 sk_wake_async(sk
,0,POLL_OUT
);
5065 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5066 tp
->snd_wnd
= ntohs(th
->window
) <<
5067 tp
->rx_opt
.snd_wscale
;
5068 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5069 TCP_SKB_CB(skb
)->seq
);
5071 /* tcp_ack considers this ACK as duplicate
5072 * and does not calculate rtt.
5073 * Fix it at least with timestamps.
5075 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5077 tcp_ack_saw_tstamp(sk
, 0);
5079 if (tp
->rx_opt
.tstamp_ok
)
5080 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5082 /* Make sure socket is routed, for
5085 icsk
->icsk_af_ops
->rebuild_header(sk
);
5087 tcp_init_metrics(sk
);
5089 tcp_init_congestion_control(sk
);
5091 /* Prevent spurious tcp_cwnd_restart() on
5092 * first data packet.
5094 tp
->lsndtime
= tcp_time_stamp
;
5097 tcp_initialize_rcv_mss(sk
);
5098 tcp_init_buffer_space(sk
);
5099 tcp_fast_path_on(tp
);
5106 if (tp
->snd_una
== tp
->write_seq
) {
5107 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5108 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5109 dst_confirm(sk
->sk_dst_cache
);
5111 if (!sock_flag(sk
, SOCK_DEAD
))
5112 /* Wake up lingering close() */
5113 sk
->sk_state_change(sk
);
5117 if (tp
->linger2
< 0 ||
5118 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5119 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5121 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5125 tmo
= tcp_fin_time(sk
);
5126 if (tmo
> TCP_TIMEWAIT_LEN
) {
5127 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5128 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5129 /* Bad case. We could lose such FIN otherwise.
5130 * It is not a big problem, but it looks confusing
5131 * and not so rare event. We still can lose it now,
5132 * if it spins in bh_lock_sock(), but it is really
5135 inet_csk_reset_keepalive_timer(sk
, tmo
);
5137 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5145 if (tp
->snd_una
== tp
->write_seq
) {
5146 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5152 if (tp
->snd_una
== tp
->write_seq
) {
5153 tcp_update_metrics(sk
);
5162 /* step 6: check the URG bit */
5163 tcp_urg(sk
, skb
, th
);
5165 /* step 7: process the segment text */
5166 switch (sk
->sk_state
) {
5167 case TCP_CLOSE_WAIT
:
5170 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5174 /* RFC 793 says to queue data in these states,
5175 * RFC 1122 says we MUST send a reset.
5176 * BSD 4.4 also does reset.
5178 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5179 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5180 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5181 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5187 case TCP_ESTABLISHED
:
5188 tcp_data_queue(sk
, skb
);
5193 /* tcp_data could move socket to TIME-WAIT */
5194 if (sk
->sk_state
!= TCP_CLOSE
) {
5195 tcp_data_snd_check(sk
);
5196 tcp_ack_snd_check(sk
);
5206 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5207 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5208 EXPORT_SYMBOL(tcp_parse_options
);
5209 EXPORT_SYMBOL(tcp_rcv_established
);
5210 EXPORT_SYMBOL(tcp_rcv_state_process
);
5211 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);