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).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #include <linux/slab.h>
66 #include <linux/module.h>
67 #include <linux/sysctl.h>
68 #include <linux/kernel.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
74 #include <net/netdma.h>
76 int sysctl_tcp_timestamps __read_mostly
= 1;
77 int sysctl_tcp_window_scaling __read_mostly
= 1;
78 int sysctl_tcp_sack __read_mostly
= 1;
79 int sysctl_tcp_fack __read_mostly
= 1;
80 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
81 EXPORT_SYMBOL(sysctl_tcp_reordering
);
82 int sysctl_tcp_ecn __read_mostly
= 2;
83 EXPORT_SYMBOL(sysctl_tcp_ecn
);
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 int sysctl_tcp_stdurg __read_mostly
;
90 int sysctl_tcp_rfc1337 __read_mostly
;
91 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
92 int sysctl_tcp_frto __read_mostly
= 2;
93 int sysctl_tcp_frto_response __read_mostly
;
94 int sysctl_tcp_nometrics_save __read_mostly
;
96 int sysctl_tcp_thin_dupack __read_mostly
;
98 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
99 int sysctl_tcp_abc __read_mostly
;
101 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
102 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
103 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
104 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
105 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
106 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
107 #define FLAG_ECE 0x40 /* ECE in this ACK */
108 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
130 struct inet_connection_sock
*icsk
= inet_csk(sk
);
131 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
134 icsk
->icsk_ack
.last_seg_size
= 0;
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
139 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
140 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
141 icsk
->icsk_ack
.rcv_mss
= len
;
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
146 * "len" is invariant segment length, including TCP header.
148 len
+= skb
->data
- skb_transport_header(skb
);
149 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
155 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
156 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
161 len
-= tcp_sk(sk
)->tcp_header_len
;
162 icsk
->icsk_ack
.last_seg_size
= len
;
164 icsk
->icsk_ack
.rcv_mss
= len
;
168 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
169 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
170 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
174 static void tcp_incr_quickack(struct sock
*sk
)
176 struct inet_connection_sock
*icsk
= inet_csk(sk
);
177 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
181 if (quickacks
> icsk
->icsk_ack
.quick
)
182 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
185 void tcp_enter_quickack_mode(struct sock
*sk
)
187 struct inet_connection_sock
*icsk
= inet_csk(sk
);
188 tcp_incr_quickack(sk
);
189 icsk
->icsk_ack
.pingpong
= 0;
190 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
193 /* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
197 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
199 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
200 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
205 if (tp
->ecn_flags
& TCP_ECN_OK
)
206 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
211 if (tcp_hdr(skb
)->cwr
)
212 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
217 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
220 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
222 if (tp
->ecn_flags
& TCP_ECN_OK
) {
223 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
224 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
225 /* Funny extension: if ECT is not set on a segment,
226 * it is surely retransmit. It is not in ECN RFC,
227 * but Linux follows this rule. */
228 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
229 tcp_enter_quickack_mode((struct sock
*)tp
);
233 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
235 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
236 tp
->ecn_flags
&= ~TCP_ECN_OK
;
239 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
241 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
242 tp
->ecn_flags
&= ~TCP_ECN_OK
;
245 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
247 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
252 /* Buffer size and advertised window tuning.
254 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
257 static void tcp_fixup_sndbuf(struct sock
*sk
)
259 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
260 sizeof(struct sk_buff
);
262 if (sk
->sk_sndbuf
< 3 * sndmem
)
263 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
266 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
268 * All tcp_full_space() is split to two parts: "network" buffer, allocated
269 * forward and advertised in receiver window (tp->rcv_wnd) and
270 * "application buffer", required to isolate scheduling/application
271 * latencies from network.
272 * window_clamp is maximal advertised window. It can be less than
273 * tcp_full_space(), in this case tcp_full_space() - window_clamp
274 * is reserved for "application" buffer. The less window_clamp is
275 * the smoother our behaviour from viewpoint of network, but the lower
276 * throughput and the higher sensitivity of the connection to losses. 8)
278 * rcv_ssthresh is more strict window_clamp used at "slow start"
279 * phase to predict further behaviour of this connection.
280 * It is used for two goals:
281 * - to enforce header prediction at sender, even when application
282 * requires some significant "application buffer". It is check #1.
283 * - to prevent pruning of receive queue because of misprediction
284 * of receiver window. Check #2.
286 * The scheme does not work when sender sends good segments opening
287 * window and then starts to feed us spaghetti. But it should work
288 * in common situations. Otherwise, we have to rely on queue collapsing.
291 /* Slow part of check#2. */
292 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
294 struct tcp_sock
*tp
= tcp_sk(sk
);
296 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
297 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
299 while (tp
->rcv_ssthresh
<= window
) {
300 if (truesize
<= skb
->len
)
301 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
309 static void tcp_grow_window(struct sock
*sk
, struct sk_buff
*skb
)
311 struct tcp_sock
*tp
= tcp_sk(sk
);
314 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
315 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
316 !tcp_memory_pressure
) {
319 /* Check #2. Increase window, if skb with such overhead
320 * will fit to rcvbuf in future.
322 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
323 incr
= 2 * tp
->advmss
;
325 incr
= __tcp_grow_window(sk
, skb
);
328 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
330 inet_csk(sk
)->icsk_ack
.quick
|= 1;
335 /* 3. Tuning rcvbuf, when connection enters established state. */
337 static void tcp_fixup_rcvbuf(struct sock
*sk
)
339 struct tcp_sock
*tp
= tcp_sk(sk
);
340 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
342 /* Try to select rcvbuf so that 4 mss-sized segments
343 * will fit to window and corresponding skbs will fit to our rcvbuf.
344 * (was 3; 4 is minimum to allow fast retransmit to work.)
346 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
348 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
349 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
352 /* 4. Try to fixup all. It is made immediately after connection enters
355 static void tcp_init_buffer_space(struct sock
*sk
)
357 struct tcp_sock
*tp
= tcp_sk(sk
);
360 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
361 tcp_fixup_rcvbuf(sk
);
362 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
363 tcp_fixup_sndbuf(sk
);
365 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
367 maxwin
= tcp_full_space(sk
);
369 if (tp
->window_clamp
>= maxwin
) {
370 tp
->window_clamp
= maxwin
;
372 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
373 tp
->window_clamp
= max(maxwin
-
374 (maxwin
>> sysctl_tcp_app_win
),
378 /* Force reservation of one segment. */
379 if (sysctl_tcp_app_win
&&
380 tp
->window_clamp
> 2 * tp
->advmss
&&
381 tp
->window_clamp
+ tp
->advmss
> maxwin
)
382 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
384 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
385 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
388 /* 5. Recalculate window clamp after socket hit its memory bounds. */
389 static void tcp_clamp_window(struct sock
*sk
)
391 struct tcp_sock
*tp
= tcp_sk(sk
);
392 struct inet_connection_sock
*icsk
= inet_csk(sk
);
394 icsk
->icsk_ack
.quick
= 0;
396 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
397 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
398 !tcp_memory_pressure
&&
399 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
400 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
403 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
404 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
407 /* Initialize RCV_MSS value.
408 * RCV_MSS is an our guess about MSS used by the peer.
409 * We haven't any direct information about the MSS.
410 * It's better to underestimate the RCV_MSS rather than overestimate.
411 * Overestimations make us ACKing less frequently than needed.
412 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
414 void tcp_initialize_rcv_mss(struct sock
*sk
)
416 struct tcp_sock
*tp
= tcp_sk(sk
);
417 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
419 hint
= min(hint
, tp
->rcv_wnd
/ 2);
420 hint
= min(hint
, TCP_MSS_DEFAULT
);
421 hint
= max(hint
, TCP_MIN_MSS
);
423 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
425 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
427 /* Receiver "autotuning" code.
429 * The algorithm for RTT estimation w/o timestamps is based on
430 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
431 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
433 * More detail on this code can be found at
434 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
435 * though this reference is out of date. A new paper
438 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
440 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
446 if (new_sample
!= 0) {
447 /* If we sample in larger samples in the non-timestamp
448 * case, we could grossly overestimate the RTT especially
449 * with chatty applications or bulk transfer apps which
450 * are stalled on filesystem I/O.
452 * Also, since we are only going for a minimum in the
453 * non-timestamp case, we do not smooth things out
454 * else with timestamps disabled convergence takes too
458 m
-= (new_sample
>> 3);
460 } else if (m
< new_sample
)
463 /* No previous measure. */
467 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
468 tp
->rcv_rtt_est
.rtt
= new_sample
;
471 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
473 if (tp
->rcv_rtt_est
.time
== 0)
475 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
477 tcp_rcv_rtt_update(tp
, jiffies
- tp
->rcv_rtt_est
.time
, 1);
480 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
481 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
484 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
485 const struct sk_buff
*skb
)
487 struct tcp_sock
*tp
= tcp_sk(sk
);
488 if (tp
->rx_opt
.rcv_tsecr
&&
489 (TCP_SKB_CB(skb
)->end_seq
-
490 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
491 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
495 * This function should be called every time data is copied to user space.
496 * It calculates the appropriate TCP receive buffer space.
498 void tcp_rcv_space_adjust(struct sock
*sk
)
500 struct tcp_sock
*tp
= tcp_sk(sk
);
504 if (tp
->rcvq_space
.time
== 0)
507 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
508 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
511 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
513 space
= max(tp
->rcvq_space
.space
, space
);
515 if (tp
->rcvq_space
.space
!= space
) {
518 tp
->rcvq_space
.space
= space
;
520 if (sysctl_tcp_moderate_rcvbuf
&&
521 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
522 int new_clamp
= space
;
524 /* Receive space grows, normalize in order to
525 * take into account packet headers and sk_buff
526 * structure overhead.
531 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
532 16 + sizeof(struct sk_buff
));
533 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
536 space
= min(space
, sysctl_tcp_rmem
[2]);
537 if (space
> sk
->sk_rcvbuf
) {
538 sk
->sk_rcvbuf
= space
;
540 /* Make the window clamp follow along. */
541 tp
->window_clamp
= new_clamp
;
547 tp
->rcvq_space
.seq
= tp
->copied_seq
;
548 tp
->rcvq_space
.time
= tcp_time_stamp
;
551 /* There is something which you must keep in mind when you analyze the
552 * behavior of the tp->ato delayed ack timeout interval. When a
553 * connection starts up, we want to ack as quickly as possible. The
554 * problem is that "good" TCP's do slow start at the beginning of data
555 * transmission. The means that until we send the first few ACK's the
556 * sender will sit on his end and only queue most of his data, because
557 * he can only send snd_cwnd unacked packets at any given time. For
558 * each ACK we send, he increments snd_cwnd and transmits more of his
561 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
563 struct tcp_sock
*tp
= tcp_sk(sk
);
564 struct inet_connection_sock
*icsk
= inet_csk(sk
);
567 inet_csk_schedule_ack(sk
);
569 tcp_measure_rcv_mss(sk
, skb
);
571 tcp_rcv_rtt_measure(tp
);
573 now
= tcp_time_stamp
;
575 if (!icsk
->icsk_ack
.ato
) {
576 /* The _first_ data packet received, initialize
577 * delayed ACK engine.
579 tcp_incr_quickack(sk
);
580 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
582 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
584 if (m
<= TCP_ATO_MIN
/ 2) {
585 /* The fastest case is the first. */
586 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
587 } else if (m
< icsk
->icsk_ack
.ato
) {
588 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
589 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
590 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
591 } else if (m
> icsk
->icsk_rto
) {
592 /* Too long gap. Apparently sender failed to
593 * restart window, so that we send ACKs quickly.
595 tcp_incr_quickack(sk
);
599 icsk
->icsk_ack
.lrcvtime
= now
;
601 TCP_ECN_check_ce(tp
, skb
);
604 tcp_grow_window(sk
, skb
);
607 /* Called to compute a smoothed rtt estimate. The data fed to this
608 * routine either comes from timestamps, or from segments that were
609 * known _not_ to have been retransmitted [see Karn/Partridge
610 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
611 * piece by Van Jacobson.
612 * NOTE: the next three routines used to be one big routine.
613 * To save cycles in the RFC 1323 implementation it was better to break
614 * it up into three procedures. -- erics
616 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
618 struct tcp_sock
*tp
= tcp_sk(sk
);
619 long m
= mrtt
; /* RTT */
621 /* The following amusing code comes from Jacobson's
622 * article in SIGCOMM '88. Note that rtt and mdev
623 * are scaled versions of rtt and mean deviation.
624 * This is designed to be as fast as possible
625 * m stands for "measurement".
627 * On a 1990 paper the rto value is changed to:
628 * RTO = rtt + 4 * mdev
630 * Funny. This algorithm seems to be very broken.
631 * These formulae increase RTO, when it should be decreased, increase
632 * too slowly, when it should be increased quickly, decrease too quickly
633 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
634 * does not matter how to _calculate_ it. Seems, it was trap
635 * that VJ failed to avoid. 8)
640 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
641 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
643 m
= -m
; /* m is now abs(error) */
644 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
645 /* This is similar to one of Eifel findings.
646 * Eifel blocks mdev updates when rtt decreases.
647 * This solution is a bit different: we use finer gain
648 * for mdev in this case (alpha*beta).
649 * Like Eifel it also prevents growth of rto,
650 * but also it limits too fast rto decreases,
651 * happening in pure Eifel.
656 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
658 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
659 if (tp
->mdev
> tp
->mdev_max
) {
660 tp
->mdev_max
= tp
->mdev
;
661 if (tp
->mdev_max
> tp
->rttvar
)
662 tp
->rttvar
= tp
->mdev_max
;
664 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
665 if (tp
->mdev_max
< tp
->rttvar
)
666 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
667 tp
->rtt_seq
= tp
->snd_nxt
;
668 tp
->mdev_max
= tcp_rto_min(sk
);
671 /* no previous measure. */
672 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
673 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
674 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
675 tp
->rtt_seq
= tp
->snd_nxt
;
679 /* Calculate rto without backoff. This is the second half of Van Jacobson's
680 * routine referred to above.
682 static inline void tcp_set_rto(struct sock
*sk
)
684 const struct tcp_sock
*tp
= tcp_sk(sk
);
685 /* Old crap is replaced with new one. 8)
688 * 1. If rtt variance happened to be less 50msec, it is hallucination.
689 * It cannot be less due to utterly erratic ACK generation made
690 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
691 * to do with delayed acks, because at cwnd>2 true delack timeout
692 * is invisible. Actually, Linux-2.4 also generates erratic
693 * ACKs in some circumstances.
695 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
697 /* 2. Fixups made earlier cannot be right.
698 * If we do not estimate RTO correctly without them,
699 * all the algo is pure shit and should be replaced
700 * with correct one. It is exactly, which we pretend to do.
703 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
704 * guarantees that rto is higher.
709 /* Save metrics learned by this TCP session.
710 This function is called only, when TCP finishes successfully
711 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
713 void tcp_update_metrics(struct sock
*sk
)
715 struct tcp_sock
*tp
= tcp_sk(sk
);
716 struct dst_entry
*dst
= __sk_dst_get(sk
);
718 if (sysctl_tcp_nometrics_save
)
723 if (dst
&& (dst
->flags
& DST_HOST
)) {
724 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
728 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
729 /* This session failed to estimate rtt. Why?
730 * Probably, no packets returned in time.
733 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
734 dst
->metrics
[RTAX_RTT
- 1] = 0;
738 rtt
= dst_metric_rtt(dst
, RTAX_RTT
);
741 /* If newly calculated rtt larger than stored one,
742 * store new one. Otherwise, use EWMA. Remember,
743 * rtt overestimation is always better than underestimation.
745 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
747 set_dst_metric_rtt(dst
, RTAX_RTT
, tp
->srtt
);
749 set_dst_metric_rtt(dst
, RTAX_RTT
, rtt
- (m
>> 3));
752 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
757 /* Scale deviation to rttvar fixed point */
762 var
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
766 var
-= (var
- m
) >> 2;
768 set_dst_metric_rtt(dst
, RTAX_RTTVAR
, var
);
771 if (tcp_in_initial_slowstart(tp
)) {
772 /* Slow start still did not finish. */
773 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
774 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
775 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
776 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
777 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
778 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
779 dst
->metrics
[RTAX_CWND
- 1] = tp
->snd_cwnd
;
780 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
781 icsk
->icsk_ca_state
== TCP_CA_Open
) {
782 /* Cong. avoidance phase, cwnd is reliable. */
783 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
784 dst
->metrics
[RTAX_SSTHRESH
-1] =
785 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
786 if (!dst_metric_locked(dst
, RTAX_CWND
))
787 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_cwnd
) >> 1;
789 /* Else slow start did not finish, cwnd is non-sense,
790 ssthresh may be also invalid.
792 if (!dst_metric_locked(dst
, RTAX_CWND
))
793 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_ssthresh
) >> 1;
794 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
795 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
796 tp
->snd_ssthresh
> dst_metric(dst
, RTAX_SSTHRESH
))
797 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
800 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
801 if (dst_metric(dst
, RTAX_REORDERING
) < tp
->reordering
&&
802 tp
->reordering
!= sysctl_tcp_reordering
)
803 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
808 /* Numbers are taken from RFC3390.
810 * John Heffner states:
812 * The RFC specifies a window of no more than 4380 bytes
813 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
814 * is a bit misleading because they use a clamp at 4380 bytes
815 * rather than use a multiplier in the relevant range.
817 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
819 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
822 if (tp
->mss_cache
> 1460)
825 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
827 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
830 /* Set slow start threshold and cwnd not falling to slow start */
831 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
833 struct tcp_sock
*tp
= tcp_sk(sk
);
834 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
836 tp
->prior_ssthresh
= 0;
838 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
841 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
842 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
843 tcp_packets_in_flight(tp
) + 1U);
844 tp
->snd_cwnd_cnt
= 0;
845 tp
->high_seq
= tp
->snd_nxt
;
846 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
847 TCP_ECN_queue_cwr(tp
);
849 tcp_set_ca_state(sk
, TCP_CA_CWR
);
854 * Packet counting of FACK is based on in-order assumptions, therefore TCP
855 * disables it when reordering is detected
857 static void tcp_disable_fack(struct tcp_sock
*tp
)
859 /* RFC3517 uses different metric in lost marker => reset on change */
861 tp
->lost_skb_hint
= NULL
;
862 tp
->rx_opt
.sack_ok
&= ~2;
865 /* Take a notice that peer is sending D-SACKs */
866 static void tcp_dsack_seen(struct tcp_sock
*tp
)
868 tp
->rx_opt
.sack_ok
|= 4;
871 /* Initialize metrics on socket. */
873 static void tcp_init_metrics(struct sock
*sk
)
875 struct tcp_sock
*tp
= tcp_sk(sk
);
876 struct dst_entry
*dst
= __sk_dst_get(sk
);
883 if (dst_metric_locked(dst
, RTAX_CWND
))
884 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
885 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
886 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
887 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
888 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
890 if (dst_metric(dst
, RTAX_REORDERING
) &&
891 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
892 tcp_disable_fack(tp
);
893 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
896 if (dst_metric(dst
, RTAX_RTT
) == 0)
899 if (!tp
->srtt
&& dst_metric_rtt(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
902 /* Initial rtt is determined from SYN,SYN-ACK.
903 * The segment is small and rtt may appear much
904 * less than real one. Use per-dst memory
905 * to make it more realistic.
907 * A bit of theory. RTT is time passed after "normal" sized packet
908 * is sent until it is ACKed. In normal circumstances sending small
909 * packets force peer to delay ACKs and calculation is correct too.
910 * The algorithm is adaptive and, provided we follow specs, it
911 * NEVER underestimate RTT. BUT! If peer tries to make some clever
912 * tricks sort of "quick acks" for time long enough to decrease RTT
913 * to low value, and then abruptly stops to do it and starts to delay
914 * ACKs, wait for troubles.
916 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
917 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
918 tp
->rtt_seq
= tp
->snd_nxt
;
920 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
921 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
922 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
925 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
929 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
930 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
934 /* Play conservative. If timestamps are not
935 * supported, TCP will fail to recalculate correct
936 * rtt, if initial rto is too small. FORGET ALL AND RESET!
938 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
940 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
941 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
946 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
949 struct tcp_sock
*tp
= tcp_sk(sk
);
950 if (metric
> tp
->reordering
) {
953 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
955 /* This exciting event is worth to be remembered. 8) */
957 mib_idx
= LINUX_MIB_TCPTSREORDER
;
958 else if (tcp_is_reno(tp
))
959 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
960 else if (tcp_is_fack(tp
))
961 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
963 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
965 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
966 #if FASTRETRANS_DEBUG > 1
967 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
968 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
972 tp
->undo_marker
? tp
->undo_retrans
: 0);
974 tcp_disable_fack(tp
);
978 /* This must be called before lost_out is incremented */
979 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
981 if ((tp
->retransmit_skb_hint
== NULL
) ||
982 before(TCP_SKB_CB(skb
)->seq
,
983 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
984 tp
->retransmit_skb_hint
= skb
;
987 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
988 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
991 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
993 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
994 tcp_verify_retransmit_hint(tp
, skb
);
996 tp
->lost_out
+= tcp_skb_pcount(skb
);
997 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1001 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
1002 struct sk_buff
*skb
)
1004 tcp_verify_retransmit_hint(tp
, skb
);
1006 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1007 tp
->lost_out
+= tcp_skb_pcount(skb
);
1008 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1012 /* This procedure tags the retransmission queue when SACKs arrive.
1014 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1015 * Packets in queue with these bits set are counted in variables
1016 * sacked_out, retrans_out and lost_out, correspondingly.
1018 * Valid combinations are:
1019 * Tag InFlight Description
1020 * 0 1 - orig segment is in flight.
1021 * S 0 - nothing flies, orig reached receiver.
1022 * L 0 - nothing flies, orig lost by net.
1023 * R 2 - both orig and retransmit are in flight.
1024 * L|R 1 - orig is lost, retransmit is in flight.
1025 * S|R 1 - orig reached receiver, retrans is still in flight.
1026 * (L|S|R is logically valid, it could occur when L|R is sacked,
1027 * but it is equivalent to plain S and code short-curcuits it to S.
1028 * L|S is logically invalid, it would mean -1 packet in flight 8))
1030 * These 6 states form finite state machine, controlled by the following events:
1031 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1032 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1033 * 3. Loss detection event of one of three flavors:
1034 * A. Scoreboard estimator decided the packet is lost.
1035 * A'. Reno "three dupacks" marks head of queue lost.
1036 * A''. Its FACK modfication, head until snd.fack is lost.
1037 * B. SACK arrives sacking data transmitted after never retransmitted
1038 * hole was sent out.
1039 * C. SACK arrives sacking SND.NXT at the moment, when the
1040 * segment was retransmitted.
1041 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1043 * It is pleasant to note, that state diagram turns out to be commutative,
1044 * so that we are allowed not to be bothered by order of our actions,
1045 * when multiple events arrive simultaneously. (see the function below).
1047 * Reordering detection.
1048 * --------------------
1049 * Reordering metric is maximal distance, which a packet can be displaced
1050 * in packet stream. With SACKs we can estimate it:
1052 * 1. SACK fills old hole and the corresponding segment was not
1053 * ever retransmitted -> reordering. Alas, we cannot use it
1054 * when segment was retransmitted.
1055 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1056 * for retransmitted and already SACKed segment -> reordering..
1057 * Both of these heuristics are not used in Loss state, when we cannot
1058 * account for retransmits accurately.
1060 * SACK block validation.
1061 * ----------------------
1063 * SACK block range validation checks that the received SACK block fits to
1064 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1065 * Note that SND.UNA is not included to the range though being valid because
1066 * it means that the receiver is rather inconsistent with itself reporting
1067 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1068 * perfectly valid, however, in light of RFC2018 which explicitly states
1069 * that "SACK block MUST reflect the newest segment. Even if the newest
1070 * segment is going to be discarded ...", not that it looks very clever
1071 * in case of head skb. Due to potentional receiver driven attacks, we
1072 * choose to avoid immediate execution of a walk in write queue due to
1073 * reneging and defer head skb's loss recovery to standard loss recovery
1074 * procedure that will eventually trigger (nothing forbids us doing this).
1076 * Implements also blockage to start_seq wrap-around. Problem lies in the
1077 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1078 * there's no guarantee that it will be before snd_nxt (n). The problem
1079 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1082 * <- outs wnd -> <- wrapzone ->
1083 * u e n u_w e_w s n_w
1085 * |<------------+------+----- TCP seqno space --------------+---------->|
1086 * ...-- <2^31 ->| |<--------...
1087 * ...---- >2^31 ------>| |<--------...
1089 * Current code wouldn't be vulnerable but it's better still to discard such
1090 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1091 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1092 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1093 * equal to the ideal case (infinite seqno space without wrap caused issues).
1095 * With D-SACK the lower bound is extended to cover sequence space below
1096 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1097 * again, D-SACK block must not to go across snd_una (for the same reason as
1098 * for the normal SACK blocks, explained above). But there all simplicity
1099 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1100 * fully below undo_marker they do not affect behavior in anyway and can
1101 * therefore be safely ignored. In rare cases (which are more or less
1102 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1103 * fragmentation and packet reordering past skb's retransmission. To consider
1104 * them correctly, the acceptable range must be extended even more though
1105 * the exact amount is rather hard to quantify. However, tp->max_window can
1106 * be used as an exaggerated estimate.
1108 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1109 u32 start_seq
, u32 end_seq
)
1111 /* Too far in future, or reversed (interpretation is ambiguous) */
1112 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1115 /* Nasty start_seq wrap-around check (see comments above) */
1116 if (!before(start_seq
, tp
->snd_nxt
))
1119 /* In outstanding window? ...This is valid exit for D-SACKs too.
1120 * start_seq == snd_una is non-sensical (see comments above)
1122 if (after(start_seq
, tp
->snd_una
))
1125 if (!is_dsack
|| !tp
->undo_marker
)
1128 /* ...Then it's D-SACK, and must reside below snd_una completely */
1129 if (!after(end_seq
, tp
->snd_una
))
1132 if (!before(start_seq
, tp
->undo_marker
))
1136 if (!after(end_seq
, tp
->undo_marker
))
1139 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1140 * start_seq < undo_marker and end_seq >= undo_marker.
1142 return !before(start_seq
, end_seq
- tp
->max_window
);
1145 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1146 * Event "C". Later note: FACK people cheated me again 8), we have to account
1147 * for reordering! Ugly, but should help.
1149 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1150 * less than what is now known to be received by the other end (derived from
1151 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1152 * retransmitted skbs to avoid some costly processing per ACKs.
1154 static void tcp_mark_lost_retrans(struct sock
*sk
)
1156 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1157 struct tcp_sock
*tp
= tcp_sk(sk
);
1158 struct sk_buff
*skb
;
1160 u32 new_low_seq
= tp
->snd_nxt
;
1161 u32 received_upto
= tcp_highest_sack_seq(tp
);
1163 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1164 !after(received_upto
, tp
->lost_retrans_low
) ||
1165 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1168 tcp_for_write_queue(skb
, sk
) {
1169 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1171 if (skb
== tcp_send_head(sk
))
1173 if (cnt
== tp
->retrans_out
)
1175 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1178 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1181 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1182 * constraint here (see above) but figuring out that at
1183 * least tp->reordering SACK blocks reside between ack_seq
1184 * and received_upto is not easy task to do cheaply with
1185 * the available datastructures.
1187 * Whether FACK should check here for tp->reordering segs
1188 * in-between one could argue for either way (it would be
1189 * rather simple to implement as we could count fack_count
1190 * during the walk and do tp->fackets_out - fack_count).
1192 if (after(received_upto
, ack_seq
)) {
1193 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1194 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1196 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1197 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1199 if (before(ack_seq
, new_low_seq
))
1200 new_low_seq
= ack_seq
;
1201 cnt
+= tcp_skb_pcount(skb
);
1205 if (tp
->retrans_out
)
1206 tp
->lost_retrans_low
= new_low_seq
;
1209 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1210 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1213 struct tcp_sock
*tp
= tcp_sk(sk
);
1214 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1215 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1218 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1221 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1222 } else if (num_sacks
> 1) {
1223 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1224 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1226 if (!after(end_seq_0
, end_seq_1
) &&
1227 !before(start_seq_0
, start_seq_1
)) {
1230 NET_INC_STATS_BH(sock_net(sk
),
1231 LINUX_MIB_TCPDSACKOFORECV
);
1235 /* D-SACK for already forgotten data... Do dumb counting. */
1237 !after(end_seq_0
, prior_snd_una
) &&
1238 after(end_seq_0
, tp
->undo_marker
))
1244 struct tcp_sacktag_state
{
1250 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1251 * the incoming SACK may not exactly match but we can find smaller MSS
1252 * aligned portion of it that matches. Therefore we might need to fragment
1253 * which may fail and creates some hassle (caller must handle error case
1256 * FIXME: this could be merged to shift decision code
1258 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1259 u32 start_seq
, u32 end_seq
)
1262 unsigned int pkt_len
;
1265 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1266 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1268 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1269 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1270 mss
= tcp_skb_mss(skb
);
1271 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1274 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1278 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1283 /* Round if necessary so that SACKs cover only full MSSes
1284 * and/or the remaining small portion (if present)
1286 if (pkt_len
> mss
) {
1287 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1288 if (!in_sack
&& new_len
< pkt_len
) {
1290 if (new_len
> skb
->len
)
1295 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1303 static u8
tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1304 struct tcp_sacktag_state
*state
,
1305 int dup_sack
, int pcount
)
1307 struct tcp_sock
*tp
= tcp_sk(sk
);
1308 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1309 int fack_count
= state
->fack_count
;
1311 /* Account D-SACK for retransmitted packet. */
1312 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1313 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1315 if (sacked
& TCPCB_SACKED_ACKED
)
1316 state
->reord
= min(fack_count
, state
->reord
);
1319 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1320 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1323 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1324 if (sacked
& TCPCB_SACKED_RETRANS
) {
1325 /* If the segment is not tagged as lost,
1326 * we do not clear RETRANS, believing
1327 * that retransmission is still in flight.
1329 if (sacked
& TCPCB_LOST
) {
1330 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1331 tp
->lost_out
-= pcount
;
1332 tp
->retrans_out
-= pcount
;
1335 if (!(sacked
& TCPCB_RETRANS
)) {
1336 /* New sack for not retransmitted frame,
1337 * which was in hole. It is reordering.
1339 if (before(TCP_SKB_CB(skb
)->seq
,
1340 tcp_highest_sack_seq(tp
)))
1341 state
->reord
= min(fack_count
,
1344 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1345 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1346 state
->flag
|= FLAG_ONLY_ORIG_SACKED
;
1349 if (sacked
& TCPCB_LOST
) {
1350 sacked
&= ~TCPCB_LOST
;
1351 tp
->lost_out
-= pcount
;
1355 sacked
|= TCPCB_SACKED_ACKED
;
1356 state
->flag
|= FLAG_DATA_SACKED
;
1357 tp
->sacked_out
+= pcount
;
1359 fack_count
+= pcount
;
1361 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1362 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1363 before(TCP_SKB_CB(skb
)->seq
,
1364 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1365 tp
->lost_cnt_hint
+= pcount
;
1367 if (fack_count
> tp
->fackets_out
)
1368 tp
->fackets_out
= fack_count
;
1371 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1372 * frames and clear it. undo_retrans is decreased above, L|R frames
1373 * are accounted above as well.
1375 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1376 sacked
&= ~TCPCB_SACKED_RETRANS
;
1377 tp
->retrans_out
-= pcount
;
1383 static int tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1384 struct tcp_sacktag_state
*state
,
1385 unsigned int pcount
, int shifted
, int mss
,
1388 struct tcp_sock
*tp
= tcp_sk(sk
);
1389 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1393 /* Tweak before seqno plays */
1394 if (!tcp_is_fack(tp
) && tcp_is_sack(tp
) && tp
->lost_skb_hint
&&
1395 !before(TCP_SKB_CB(tp
->lost_skb_hint
)->seq
, TCP_SKB_CB(skb
)->seq
))
1396 tp
->lost_cnt_hint
+= pcount
;
1398 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1399 TCP_SKB_CB(skb
)->seq
+= shifted
;
1401 skb_shinfo(prev
)->gso_segs
+= pcount
;
1402 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1403 skb_shinfo(skb
)->gso_segs
-= pcount
;
1405 /* When we're adding to gso_segs == 1, gso_size will be zero,
1406 * in theory this shouldn't be necessary but as long as DSACK
1407 * code can come after this skb later on it's better to keep
1408 * setting gso_size to something.
1410 if (!skb_shinfo(prev
)->gso_size
) {
1411 skb_shinfo(prev
)->gso_size
= mss
;
1412 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1415 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1416 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1417 skb_shinfo(skb
)->gso_size
= 0;
1418 skb_shinfo(skb
)->gso_type
= 0;
1421 /* We discard results */
1422 tcp_sacktag_one(skb
, sk
, state
, dup_sack
, pcount
);
1424 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1425 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1428 BUG_ON(!tcp_skb_pcount(skb
));
1429 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1433 /* Whole SKB was eaten :-) */
1435 if (skb
== tp
->retransmit_skb_hint
)
1436 tp
->retransmit_skb_hint
= prev
;
1437 if (skb
== tp
->scoreboard_skb_hint
)
1438 tp
->scoreboard_skb_hint
= prev
;
1439 if (skb
== tp
->lost_skb_hint
) {
1440 tp
->lost_skb_hint
= prev
;
1441 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1444 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(prev
)->flags
;
1445 if (skb
== tcp_highest_sack(sk
))
1446 tcp_advance_highest_sack(sk
, skb
);
1448 tcp_unlink_write_queue(skb
, sk
);
1449 sk_wmem_free_skb(sk
, skb
);
1451 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1456 /* I wish gso_size would have a bit more sane initialization than
1457 * something-or-zero which complicates things
1459 static int tcp_skb_seglen(struct sk_buff
*skb
)
1461 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1464 /* Shifting pages past head area doesn't work */
1465 static int skb_can_shift(struct sk_buff
*skb
)
1467 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1470 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1473 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1474 struct tcp_sacktag_state
*state
,
1475 u32 start_seq
, u32 end_seq
,
1478 struct tcp_sock
*tp
= tcp_sk(sk
);
1479 struct sk_buff
*prev
;
1485 if (!sk_can_gso(sk
))
1488 /* Normally R but no L won't result in plain S */
1490 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1492 if (!skb_can_shift(skb
))
1494 /* This frame is about to be dropped (was ACKed). */
1495 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1498 /* Can only happen with delayed DSACK + discard craziness */
1499 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1501 prev
= tcp_write_queue_prev(sk
, skb
);
1503 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1506 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1507 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1511 pcount
= tcp_skb_pcount(skb
);
1512 mss
= tcp_skb_seglen(skb
);
1514 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1515 * drop this restriction as unnecessary
1517 if (mss
!= tcp_skb_seglen(prev
))
1520 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1522 /* CHECKME: This is non-MSS split case only?, this will
1523 * cause skipped skbs due to advancing loop btw, original
1524 * has that feature too
1526 if (tcp_skb_pcount(skb
) <= 1)
1529 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1531 /* TODO: head merge to next could be attempted here
1532 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1533 * though it might not be worth of the additional hassle
1535 * ...we can probably just fallback to what was done
1536 * previously. We could try merging non-SACKed ones
1537 * as well but it probably isn't going to buy off
1538 * because later SACKs might again split them, and
1539 * it would make skb timestamp tracking considerably
1545 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1547 BUG_ON(len
> skb
->len
);
1549 /* MSS boundaries should be honoured or else pcount will
1550 * severely break even though it makes things bit trickier.
1551 * Optimize common case to avoid most of the divides
1553 mss
= tcp_skb_mss(skb
);
1555 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1556 * drop this restriction as unnecessary
1558 if (mss
!= tcp_skb_seglen(prev
))
1563 } else if (len
< mss
) {
1571 if (!skb_shift(prev
, skb
, len
))
1573 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1576 /* Hole filled allows collapsing with the next as well, this is very
1577 * useful when hole on every nth skb pattern happens
1579 if (prev
== tcp_write_queue_tail(sk
))
1581 skb
= tcp_write_queue_next(sk
, prev
);
1583 if (!skb_can_shift(skb
) ||
1584 (skb
== tcp_send_head(sk
)) ||
1585 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1586 (mss
!= tcp_skb_seglen(skb
)))
1590 if (skb_shift(prev
, skb
, len
)) {
1591 pcount
+= tcp_skb_pcount(skb
);
1592 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1596 state
->fack_count
+= pcount
;
1603 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1607 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1608 struct tcp_sack_block
*next_dup
,
1609 struct tcp_sacktag_state
*state
,
1610 u32 start_seq
, u32 end_seq
,
1613 struct tcp_sock
*tp
= tcp_sk(sk
);
1614 struct sk_buff
*tmp
;
1616 tcp_for_write_queue_from(skb
, sk
) {
1618 int dup_sack
= dup_sack_in
;
1620 if (skb
== tcp_send_head(sk
))
1623 /* queue is in-order => we can short-circuit the walk early */
1624 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1627 if ((next_dup
!= NULL
) &&
1628 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1629 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1630 next_dup
->start_seq
,
1636 /* skb reference here is a bit tricky to get right, since
1637 * shifting can eat and free both this skb and the next,
1638 * so not even _safe variant of the loop is enough.
1641 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1642 start_seq
, end_seq
, dup_sack
);
1651 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1657 if (unlikely(in_sack
< 0))
1661 TCP_SKB_CB(skb
)->sacked
= tcp_sacktag_one(skb
, sk
,
1664 tcp_skb_pcount(skb
));
1666 if (!before(TCP_SKB_CB(skb
)->seq
,
1667 tcp_highest_sack_seq(tp
)))
1668 tcp_advance_highest_sack(sk
, skb
);
1671 state
->fack_count
+= tcp_skb_pcount(skb
);
1676 /* Avoid all extra work that is being done by sacktag while walking in
1679 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1680 struct tcp_sacktag_state
*state
,
1683 tcp_for_write_queue_from(skb
, sk
) {
1684 if (skb
== tcp_send_head(sk
))
1687 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1690 state
->fack_count
+= tcp_skb_pcount(skb
);
1695 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1697 struct tcp_sack_block
*next_dup
,
1698 struct tcp_sacktag_state
*state
,
1701 if (next_dup
== NULL
)
1704 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1705 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1706 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1707 next_dup
->start_seq
, next_dup
->end_seq
,
1714 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1716 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1720 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1723 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1724 struct tcp_sock
*tp
= tcp_sk(sk
);
1725 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1726 TCP_SKB_CB(ack_skb
)->sacked
);
1727 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1728 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1729 struct tcp_sack_block
*cache
;
1730 struct tcp_sacktag_state state
;
1731 struct sk_buff
*skb
;
1732 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1734 int found_dup_sack
= 0;
1736 int first_sack_index
;
1739 state
.reord
= tp
->packets_out
;
1741 if (!tp
->sacked_out
) {
1742 if (WARN_ON(tp
->fackets_out
))
1743 tp
->fackets_out
= 0;
1744 tcp_highest_sack_reset(sk
);
1747 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1748 num_sacks
, prior_snd_una
);
1750 state
.flag
|= FLAG_DSACKING_ACK
;
1752 /* Eliminate too old ACKs, but take into
1753 * account more or less fresh ones, they can
1754 * contain valid SACK info.
1756 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1759 if (!tp
->packets_out
)
1763 first_sack_index
= 0;
1764 for (i
= 0; i
< num_sacks
; i
++) {
1765 int dup_sack
= !i
&& found_dup_sack
;
1767 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1768 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1770 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1771 sp
[used_sacks
].start_seq
,
1772 sp
[used_sacks
].end_seq
)) {
1776 if (!tp
->undo_marker
)
1777 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1779 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1781 /* Don't count olds caused by ACK reordering */
1782 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1783 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1785 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1788 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1790 first_sack_index
= -1;
1794 /* Ignore very old stuff early */
1795 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1801 /* order SACK blocks to allow in order walk of the retrans queue */
1802 for (i
= used_sacks
- 1; i
> 0; i
--) {
1803 for (j
= 0; j
< i
; j
++) {
1804 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1805 swap(sp
[j
], sp
[j
+ 1]);
1807 /* Track where the first SACK block goes to */
1808 if (j
== first_sack_index
)
1809 first_sack_index
= j
+ 1;
1814 skb
= tcp_write_queue_head(sk
);
1815 state
.fack_count
= 0;
1818 if (!tp
->sacked_out
) {
1819 /* It's already past, so skip checking against it */
1820 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1822 cache
= tp
->recv_sack_cache
;
1823 /* Skip empty blocks in at head of the cache */
1824 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1829 while (i
< used_sacks
) {
1830 u32 start_seq
= sp
[i
].start_seq
;
1831 u32 end_seq
= sp
[i
].end_seq
;
1832 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1833 struct tcp_sack_block
*next_dup
= NULL
;
1835 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1836 next_dup
= &sp
[i
+ 1];
1838 /* Event "B" in the comment above. */
1839 if (after(end_seq
, tp
->high_seq
))
1840 state
.flag
|= FLAG_DATA_LOST
;
1842 /* Skip too early cached blocks */
1843 while (tcp_sack_cache_ok(tp
, cache
) &&
1844 !before(start_seq
, cache
->end_seq
))
1847 /* Can skip some work by looking recv_sack_cache? */
1848 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1849 after(end_seq
, cache
->start_seq
)) {
1852 if (before(start_seq
, cache
->start_seq
)) {
1853 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1855 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1862 /* Rest of the block already fully processed? */
1863 if (!after(end_seq
, cache
->end_seq
))
1866 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1870 /* ...tail remains todo... */
1871 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1872 /* ...but better entrypoint exists! */
1873 skb
= tcp_highest_sack(sk
);
1876 state
.fack_count
= tp
->fackets_out
;
1881 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1882 /* Check overlap against next cached too (past this one already) */
1887 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1888 skb
= tcp_highest_sack(sk
);
1891 state
.fack_count
= tp
->fackets_out
;
1893 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1896 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1897 start_seq
, end_seq
, dup_sack
);
1900 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1901 * due to in-order walk
1903 if (after(end_seq
, tp
->frto_highmark
))
1904 state
.flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1909 /* Clear the head of the cache sack blocks so we can skip it next time */
1910 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1911 tp
->recv_sack_cache
[i
].start_seq
= 0;
1912 tp
->recv_sack_cache
[i
].end_seq
= 0;
1914 for (j
= 0; j
< used_sacks
; j
++)
1915 tp
->recv_sack_cache
[i
++] = sp
[j
];
1917 tcp_mark_lost_retrans(sk
);
1919 tcp_verify_left_out(tp
);
1921 if ((state
.reord
< tp
->fackets_out
) &&
1922 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1923 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1924 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1928 #if FASTRETRANS_DEBUG > 0
1929 WARN_ON((int)tp
->sacked_out
< 0);
1930 WARN_ON((int)tp
->lost_out
< 0);
1931 WARN_ON((int)tp
->retrans_out
< 0);
1932 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1937 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1938 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1940 static int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1944 holes
= max(tp
->lost_out
, 1U);
1945 holes
= min(holes
, tp
->packets_out
);
1947 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1948 tp
->sacked_out
= tp
->packets_out
- holes
;
1954 /* If we receive more dupacks than we expected counting segments
1955 * in assumption of absent reordering, interpret this as reordering.
1956 * The only another reason could be bug in receiver TCP.
1958 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1960 struct tcp_sock
*tp
= tcp_sk(sk
);
1961 if (tcp_limit_reno_sacked(tp
))
1962 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1965 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1967 static void tcp_add_reno_sack(struct sock
*sk
)
1969 struct tcp_sock
*tp
= tcp_sk(sk
);
1971 tcp_check_reno_reordering(sk
, 0);
1972 tcp_verify_left_out(tp
);
1975 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1977 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1979 struct tcp_sock
*tp
= tcp_sk(sk
);
1982 /* One ACK acked hole. The rest eat duplicate ACKs. */
1983 if (acked
- 1 >= tp
->sacked_out
)
1986 tp
->sacked_out
-= acked
- 1;
1988 tcp_check_reno_reordering(sk
, acked
);
1989 tcp_verify_left_out(tp
);
1992 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1997 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
1999 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
2002 /* F-RTO can only be used if TCP has never retransmitted anything other than
2003 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2005 int tcp_use_frto(struct sock
*sk
)
2007 const struct tcp_sock
*tp
= tcp_sk(sk
);
2008 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2009 struct sk_buff
*skb
;
2011 if (!sysctl_tcp_frto
)
2014 /* MTU probe and F-RTO won't really play nicely along currently */
2015 if (icsk
->icsk_mtup
.probe_size
)
2018 if (tcp_is_sackfrto(tp
))
2021 /* Avoid expensive walking of rexmit queue if possible */
2022 if (tp
->retrans_out
> 1)
2025 skb
= tcp_write_queue_head(sk
);
2026 if (tcp_skb_is_last(sk
, skb
))
2028 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
2029 tcp_for_write_queue_from(skb
, sk
) {
2030 if (skb
== tcp_send_head(sk
))
2032 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2034 /* Short-circuit when first non-SACKed skb has been checked */
2035 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2041 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2042 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2043 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2044 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2045 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2046 * bits are handled if the Loss state is really to be entered (in
2047 * tcp_enter_frto_loss).
2049 * Do like tcp_enter_loss() would; when RTO expires the second time it
2051 * "Reduce ssthresh if it has not yet been made inside this window."
2053 void tcp_enter_frto(struct sock
*sk
)
2055 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2056 struct tcp_sock
*tp
= tcp_sk(sk
);
2057 struct sk_buff
*skb
;
2059 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
2060 tp
->snd_una
== tp
->high_seq
||
2061 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
2062 !icsk
->icsk_retransmits
)) {
2063 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2064 /* Our state is too optimistic in ssthresh() call because cwnd
2065 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2066 * recovery has not yet completed. Pattern would be this: RTO,
2067 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2069 * RFC4138 should be more specific on what to do, even though
2070 * RTO is quite unlikely to occur after the first Cumulative ACK
2071 * due to back-off and complexity of triggering events ...
2073 if (tp
->frto_counter
) {
2075 stored_cwnd
= tp
->snd_cwnd
;
2077 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2078 tp
->snd_cwnd
= stored_cwnd
;
2080 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2082 /* ... in theory, cong.control module could do "any tricks" in
2083 * ssthresh(), which means that ca_state, lost bits and lost_out
2084 * counter would have to be faked before the call occurs. We
2085 * consider that too expensive, unlikely and hacky, so modules
2086 * using these in ssthresh() must deal these incompatibility
2087 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2089 tcp_ca_event(sk
, CA_EVENT_FRTO
);
2092 tp
->undo_marker
= tp
->snd_una
;
2093 tp
->undo_retrans
= 0;
2095 skb
= tcp_write_queue_head(sk
);
2096 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2097 tp
->undo_marker
= 0;
2098 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2099 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2100 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2102 tcp_verify_left_out(tp
);
2104 /* Too bad if TCP was application limited */
2105 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2107 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2108 * The last condition is necessary at least in tp->frto_counter case.
2110 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
2111 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
2112 after(tp
->high_seq
, tp
->snd_una
)) {
2113 tp
->frto_highmark
= tp
->high_seq
;
2115 tp
->frto_highmark
= tp
->snd_nxt
;
2117 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
2118 tp
->high_seq
= tp
->snd_nxt
;
2119 tp
->frto_counter
= 1;
2122 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2123 * which indicates that we should follow the traditional RTO recovery,
2124 * i.e. mark everything lost and do go-back-N retransmission.
2126 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
2128 struct tcp_sock
*tp
= tcp_sk(sk
);
2129 struct sk_buff
*skb
;
2132 tp
->retrans_out
= 0;
2133 if (tcp_is_reno(tp
))
2134 tcp_reset_reno_sack(tp
);
2136 tcp_for_write_queue(skb
, sk
) {
2137 if (skb
== tcp_send_head(sk
))
2140 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2142 * Count the retransmission made on RTO correctly (only when
2143 * waiting for the first ACK and did not get it)...
2145 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
2146 /* For some reason this R-bit might get cleared? */
2147 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
2148 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2149 /* ...enter this if branch just for the first segment */
2150 flag
|= FLAG_DATA_ACKED
;
2152 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2153 tp
->undo_marker
= 0;
2154 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2157 /* Marking forward transmissions that were made after RTO lost
2158 * can cause unnecessary retransmissions in some scenarios,
2159 * SACK blocks will mitigate that in some but not in all cases.
2160 * We used to not mark them but it was causing break-ups with
2161 * receivers that do only in-order receival.
2163 * TODO: we could detect presence of such receiver and select
2164 * different behavior per flow.
2166 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2167 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2168 tp
->lost_out
+= tcp_skb_pcount(skb
);
2169 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2172 tcp_verify_left_out(tp
);
2174 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2175 tp
->snd_cwnd_cnt
= 0;
2176 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2177 tp
->frto_counter
= 0;
2178 tp
->bytes_acked
= 0;
2180 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2181 sysctl_tcp_reordering
);
2182 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2183 tp
->high_seq
= tp
->snd_nxt
;
2184 TCP_ECN_queue_cwr(tp
);
2186 tcp_clear_all_retrans_hints(tp
);
2189 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2191 tp
->retrans_out
= 0;
2194 tp
->undo_marker
= 0;
2195 tp
->undo_retrans
= 0;
2198 void tcp_clear_retrans(struct tcp_sock
*tp
)
2200 tcp_clear_retrans_partial(tp
);
2202 tp
->fackets_out
= 0;
2206 /* Enter Loss state. If "how" is not zero, forget all SACK information
2207 * and reset tags completely, otherwise preserve SACKs. If receiver
2208 * dropped its ofo queue, we will know this due to reneging detection.
2210 void tcp_enter_loss(struct sock
*sk
, int how
)
2212 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2213 struct tcp_sock
*tp
= tcp_sk(sk
);
2214 struct sk_buff
*skb
;
2216 /* Reduce ssthresh if it has not yet been made inside this window. */
2217 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2218 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2219 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2220 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2221 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2224 tp
->snd_cwnd_cnt
= 0;
2225 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2227 tp
->bytes_acked
= 0;
2228 tcp_clear_retrans_partial(tp
);
2230 if (tcp_is_reno(tp
))
2231 tcp_reset_reno_sack(tp
);
2234 /* Push undo marker, if it was plain RTO and nothing
2235 * was retransmitted. */
2236 tp
->undo_marker
= tp
->snd_una
;
2239 tp
->fackets_out
= 0;
2241 tcp_clear_all_retrans_hints(tp
);
2243 tcp_for_write_queue(skb
, sk
) {
2244 if (skb
== tcp_send_head(sk
))
2247 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2248 tp
->undo_marker
= 0;
2249 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2250 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2251 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2252 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2253 tp
->lost_out
+= tcp_skb_pcount(skb
);
2254 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2257 tcp_verify_left_out(tp
);
2259 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2260 sysctl_tcp_reordering
);
2261 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2262 tp
->high_seq
= tp
->snd_nxt
;
2263 TCP_ECN_queue_cwr(tp
);
2264 /* Abort F-RTO algorithm if one is in progress */
2265 tp
->frto_counter
= 0;
2268 /* If ACK arrived pointing to a remembered SACK, it means that our
2269 * remembered SACKs do not reflect real state of receiver i.e.
2270 * receiver _host_ is heavily congested (or buggy).
2272 * Do processing similar to RTO timeout.
2274 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2276 if (flag
& FLAG_SACK_RENEGING
) {
2277 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2278 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2280 tcp_enter_loss(sk
, 1);
2281 icsk
->icsk_retransmits
++;
2282 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2283 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2284 icsk
->icsk_rto
, TCP_RTO_MAX
);
2290 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2292 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2295 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2296 * counter when SACK is enabled (without SACK, sacked_out is used for
2299 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2300 * segments up to the highest received SACK block so far and holes in
2303 * With reordering, holes may still be in flight, so RFC3517 recovery
2304 * uses pure sacked_out (total number of SACKed segments) even though
2305 * it violates the RFC that uses duplicate ACKs, often these are equal
2306 * but when e.g. out-of-window ACKs or packet duplication occurs,
2307 * they differ. Since neither occurs due to loss, TCP should really
2310 static inline int tcp_dupack_heuristics(struct tcp_sock
*tp
)
2312 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2315 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2317 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2320 static inline int tcp_head_timedout(struct sock
*sk
)
2322 struct tcp_sock
*tp
= tcp_sk(sk
);
2324 return tp
->packets_out
&&
2325 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2328 /* Linux NewReno/SACK/FACK/ECN state machine.
2329 * --------------------------------------
2331 * "Open" Normal state, no dubious events, fast path.
2332 * "Disorder" In all the respects it is "Open",
2333 * but requires a bit more attention. It is entered when
2334 * we see some SACKs or dupacks. It is split of "Open"
2335 * mainly to move some processing from fast path to slow one.
2336 * "CWR" CWND was reduced due to some Congestion Notification event.
2337 * It can be ECN, ICMP source quench, local device congestion.
2338 * "Recovery" CWND was reduced, we are fast-retransmitting.
2339 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2341 * tcp_fastretrans_alert() is entered:
2342 * - each incoming ACK, if state is not "Open"
2343 * - when arrived ACK is unusual, namely:
2348 * Counting packets in flight is pretty simple.
2350 * in_flight = packets_out - left_out + retrans_out
2352 * packets_out is SND.NXT-SND.UNA counted in packets.
2354 * retrans_out is number of retransmitted segments.
2356 * left_out is number of segments left network, but not ACKed yet.
2358 * left_out = sacked_out + lost_out
2360 * sacked_out: Packets, which arrived to receiver out of order
2361 * and hence not ACKed. With SACKs this number is simply
2362 * amount of SACKed data. Even without SACKs
2363 * it is easy to give pretty reliable estimate of this number,
2364 * counting duplicate ACKs.
2366 * lost_out: Packets lost by network. TCP has no explicit
2367 * "loss notification" feedback from network (for now).
2368 * It means that this number can be only _guessed_.
2369 * Actually, it is the heuristics to predict lossage that
2370 * distinguishes different algorithms.
2372 * F.e. after RTO, when all the queue is considered as lost,
2373 * lost_out = packets_out and in_flight = retrans_out.
2375 * Essentially, we have now two algorithms counting
2378 * FACK: It is the simplest heuristics. As soon as we decided
2379 * that something is lost, we decide that _all_ not SACKed
2380 * packets until the most forward SACK are lost. I.e.
2381 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2382 * It is absolutely correct estimate, if network does not reorder
2383 * packets. And it loses any connection to reality when reordering
2384 * takes place. We use FACK by default until reordering
2385 * is suspected on the path to this destination.
2387 * NewReno: when Recovery is entered, we assume that one segment
2388 * is lost (classic Reno). While we are in Recovery and
2389 * a partial ACK arrives, we assume that one more packet
2390 * is lost (NewReno). This heuristics are the same in NewReno
2393 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2394 * deflation etc. CWND is real congestion window, never inflated, changes
2395 * only according to classic VJ rules.
2397 * Really tricky (and requiring careful tuning) part of algorithm
2398 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2399 * The first determines the moment _when_ we should reduce CWND and,
2400 * hence, slow down forward transmission. In fact, it determines the moment
2401 * when we decide that hole is caused by loss, rather than by a reorder.
2403 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2404 * holes, caused by lost packets.
2406 * And the most logically complicated part of algorithm is undo
2407 * heuristics. We detect false retransmits due to both too early
2408 * fast retransmit (reordering) and underestimated RTO, analyzing
2409 * timestamps and D-SACKs. When we detect that some segments were
2410 * retransmitted by mistake and CWND reduction was wrong, we undo
2411 * window reduction and abort recovery phase. This logic is hidden
2412 * inside several functions named tcp_try_undo_<something>.
2415 /* This function decides, when we should leave Disordered state
2416 * and enter Recovery phase, reducing congestion window.
2418 * Main question: may we further continue forward transmission
2419 * with the same cwnd?
2421 static int tcp_time_to_recover(struct sock
*sk
)
2423 struct tcp_sock
*tp
= tcp_sk(sk
);
2426 /* Do not perform any recovery during F-RTO algorithm */
2427 if (tp
->frto_counter
)
2430 /* Trick#1: The loss is proven. */
2434 /* Not-A-Trick#2 : Classic rule... */
2435 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2438 /* Trick#3 : when we use RFC2988 timer restart, fast
2439 * retransmit can be triggered by timeout of queue head.
2441 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2444 /* Trick#4: It is still not OK... But will it be useful to delay
2447 packets_out
= tp
->packets_out
;
2448 if (packets_out
<= tp
->reordering
&&
2449 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2450 !tcp_may_send_now(sk
)) {
2451 /* We have nothing to send. This connection is limited
2452 * either by receiver window or by application.
2457 /* If a thin stream is detected, retransmit after first
2458 * received dupack. Employ only if SACK is supported in order
2459 * to avoid possible corner-case series of spurious retransmissions
2460 * Use only if there are no unsent data.
2462 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2463 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2464 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2470 /* New heuristics: it is possible only after we switched to restart timer
2471 * each time when something is ACKed. Hence, we can detect timed out packets
2472 * during fast retransmit without falling to slow start.
2474 * Usefulness of this as is very questionable, since we should know which of
2475 * the segments is the next to timeout which is relatively expensive to find
2476 * in general case unless we add some data structure just for that. The
2477 * current approach certainly won't find the right one too often and when it
2478 * finally does find _something_ it usually marks large part of the window
2479 * right away (because a retransmission with a larger timestamp blocks the
2480 * loop from advancing). -ij
2482 static void tcp_timeout_skbs(struct sock
*sk
)
2484 struct tcp_sock
*tp
= tcp_sk(sk
);
2485 struct sk_buff
*skb
;
2487 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2490 skb
= tp
->scoreboard_skb_hint
;
2491 if (tp
->scoreboard_skb_hint
== NULL
)
2492 skb
= tcp_write_queue_head(sk
);
2494 tcp_for_write_queue_from(skb
, sk
) {
2495 if (skb
== tcp_send_head(sk
))
2497 if (!tcp_skb_timedout(sk
, skb
))
2500 tcp_skb_mark_lost(tp
, skb
);
2503 tp
->scoreboard_skb_hint
= skb
;
2505 tcp_verify_left_out(tp
);
2508 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2509 * is against sacked "cnt", otherwise it's against facked "cnt"
2511 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2513 struct tcp_sock
*tp
= tcp_sk(sk
);
2514 struct sk_buff
*skb
;
2522 WARN_ON(packets
> tp
->packets_out
);
2523 if (tp
->lost_skb_hint
) {
2524 skb
= tp
->lost_skb_hint
;
2525 cnt
= tp
->lost_cnt_hint
;
2527 skb
= tcp_write_queue_head(sk
);
2531 tcp_for_write_queue_from(skb
, sk
) {
2532 if (skb
== tcp_send_head(sk
))
2534 /* TODO: do this better */
2535 /* this is not the most efficient way to do this... */
2536 tp
->lost_skb_hint
= skb
;
2537 tp
->lost_cnt_hint
= cnt
;
2539 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2543 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2544 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2545 cnt
+= tcp_skb_pcount(skb
);
2547 if (cnt
> packets
) {
2548 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2549 (oldcnt
>= packets
))
2552 mss
= skb_shinfo(skb
)->gso_size
;
2553 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2559 tcp_skb_mark_lost(tp
, skb
);
2561 tcp_verify_left_out(tp
);
2564 /* Account newly detected lost packet(s) */
2566 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2568 struct tcp_sock
*tp
= tcp_sk(sk
);
2570 if (tcp_is_reno(tp
)) {
2571 tcp_mark_head_lost(sk
, 1);
2572 } else if (tcp_is_fack(tp
)) {
2573 int lost
= tp
->fackets_out
- tp
->reordering
;
2576 tcp_mark_head_lost(sk
, lost
);
2578 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2579 if (sacked_upto
< fast_rexmit
)
2580 sacked_upto
= fast_rexmit
;
2581 tcp_mark_head_lost(sk
, sacked_upto
);
2584 tcp_timeout_skbs(sk
);
2587 /* CWND moderation, preventing bursts due to too big ACKs
2588 * in dubious situations.
2590 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2592 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2593 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2594 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2597 /* Lower bound on congestion window is slow start threshold
2598 * unless congestion avoidance choice decides to overide it.
2600 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2602 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2604 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2607 /* Decrease cwnd each second ack. */
2608 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2610 struct tcp_sock
*tp
= tcp_sk(sk
);
2611 int decr
= tp
->snd_cwnd_cnt
+ 1;
2613 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2614 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2615 tp
->snd_cwnd_cnt
= decr
& 1;
2618 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2619 tp
->snd_cwnd
-= decr
;
2621 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2622 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2626 /* Nothing was retransmitted or returned timestamp is less
2627 * than timestamp of the first retransmission.
2629 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2631 return !tp
->retrans_stamp
||
2632 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2633 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2636 /* Undo procedures. */
2638 #if FASTRETRANS_DEBUG > 1
2639 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2641 struct tcp_sock
*tp
= tcp_sk(sk
);
2642 struct inet_sock
*inet
= inet_sk(sk
);
2644 if (sk
->sk_family
== AF_INET
) {
2645 printk(KERN_DEBUG
"Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2647 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2648 tp
->snd_cwnd
, tcp_left_out(tp
),
2649 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2652 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2653 else if (sk
->sk_family
== AF_INET6
) {
2654 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2655 printk(KERN_DEBUG
"Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2657 &np
->daddr
, ntohs(inet
->inet_dport
),
2658 tp
->snd_cwnd
, tcp_left_out(tp
),
2659 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2665 #define DBGUNDO(x...) do { } while (0)
2668 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2670 struct tcp_sock
*tp
= tcp_sk(sk
);
2672 if (tp
->prior_ssthresh
) {
2673 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2675 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2676 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2678 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2680 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2681 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2682 TCP_ECN_withdraw_cwr(tp
);
2685 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2687 tcp_moderate_cwnd(tp
);
2688 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2691 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2693 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2696 /* People celebrate: "We love our President!" */
2697 static int tcp_try_undo_recovery(struct sock
*sk
)
2699 struct tcp_sock
*tp
= tcp_sk(sk
);
2701 if (tcp_may_undo(tp
)) {
2704 /* Happy end! We did not retransmit anything
2705 * or our original transmission succeeded.
2707 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2708 tcp_undo_cwr(sk
, 1);
2709 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2710 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2712 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2714 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2715 tp
->undo_marker
= 0;
2717 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2718 /* Hold old state until something *above* high_seq
2719 * is ACKed. For Reno it is MUST to prevent false
2720 * fast retransmits (RFC2582). SACK TCP is safe. */
2721 tcp_moderate_cwnd(tp
);
2724 tcp_set_ca_state(sk
, TCP_CA_Open
);
2728 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2729 static void tcp_try_undo_dsack(struct sock
*sk
)
2731 struct tcp_sock
*tp
= tcp_sk(sk
);
2733 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2734 DBGUNDO(sk
, "D-SACK");
2735 tcp_undo_cwr(sk
, 1);
2736 tp
->undo_marker
= 0;
2737 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2741 /* We can clear retrans_stamp when there are no retransmissions in the
2742 * window. It would seem that it is trivially available for us in
2743 * tp->retrans_out, however, that kind of assumptions doesn't consider
2744 * what will happen if errors occur when sending retransmission for the
2745 * second time. ...It could the that such segment has only
2746 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2747 * the head skb is enough except for some reneging corner cases that
2748 * are not worth the effort.
2750 * Main reason for all this complexity is the fact that connection dying
2751 * time now depends on the validity of the retrans_stamp, in particular,
2752 * that successive retransmissions of a segment must not advance
2753 * retrans_stamp under any conditions.
2755 static int tcp_any_retrans_done(struct sock
*sk
)
2757 struct tcp_sock
*tp
= tcp_sk(sk
);
2758 struct sk_buff
*skb
;
2760 if (tp
->retrans_out
)
2763 skb
= tcp_write_queue_head(sk
);
2764 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2770 /* Undo during fast recovery after partial ACK. */
2772 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2774 struct tcp_sock
*tp
= tcp_sk(sk
);
2775 /* Partial ACK arrived. Force Hoe's retransmit. */
2776 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2778 if (tcp_may_undo(tp
)) {
2779 /* Plain luck! Hole if filled with delayed
2780 * packet, rather than with a retransmit.
2782 if (!tcp_any_retrans_done(sk
))
2783 tp
->retrans_stamp
= 0;
2785 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2788 tcp_undo_cwr(sk
, 0);
2789 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2791 /* So... Do not make Hoe's retransmit yet.
2792 * If the first packet was delayed, the rest
2793 * ones are most probably delayed as well.
2800 /* Undo during loss recovery after partial ACK. */
2801 static int tcp_try_undo_loss(struct sock
*sk
)
2803 struct tcp_sock
*tp
= tcp_sk(sk
);
2805 if (tcp_may_undo(tp
)) {
2806 struct sk_buff
*skb
;
2807 tcp_for_write_queue(skb
, sk
) {
2808 if (skb
== tcp_send_head(sk
))
2810 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2813 tcp_clear_all_retrans_hints(tp
);
2815 DBGUNDO(sk
, "partial loss");
2817 tcp_undo_cwr(sk
, 1);
2818 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2819 inet_csk(sk
)->icsk_retransmits
= 0;
2820 tp
->undo_marker
= 0;
2821 if (tcp_is_sack(tp
))
2822 tcp_set_ca_state(sk
, TCP_CA_Open
);
2828 static inline void tcp_complete_cwr(struct sock
*sk
)
2830 struct tcp_sock
*tp
= tcp_sk(sk
);
2831 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2832 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2833 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2836 static void tcp_try_keep_open(struct sock
*sk
)
2838 struct tcp_sock
*tp
= tcp_sk(sk
);
2839 int state
= TCP_CA_Open
;
2841 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
) || tp
->undo_marker
)
2842 state
= TCP_CA_Disorder
;
2844 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2845 tcp_set_ca_state(sk
, state
);
2846 tp
->high_seq
= tp
->snd_nxt
;
2850 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2852 struct tcp_sock
*tp
= tcp_sk(sk
);
2854 tcp_verify_left_out(tp
);
2856 if (!tp
->frto_counter
&& !tcp_any_retrans_done(sk
))
2857 tp
->retrans_stamp
= 0;
2859 if (flag
& FLAG_ECE
)
2860 tcp_enter_cwr(sk
, 1);
2862 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2863 tcp_try_keep_open(sk
);
2864 tcp_moderate_cwnd(tp
);
2866 tcp_cwnd_down(sk
, flag
);
2870 static void tcp_mtup_probe_failed(struct sock
*sk
)
2872 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2874 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2875 icsk
->icsk_mtup
.probe_size
= 0;
2878 static void tcp_mtup_probe_success(struct sock
*sk
)
2880 struct tcp_sock
*tp
= tcp_sk(sk
);
2881 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2883 /* FIXME: breaks with very large cwnd */
2884 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2885 tp
->snd_cwnd
= tp
->snd_cwnd
*
2886 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2887 icsk
->icsk_mtup
.probe_size
;
2888 tp
->snd_cwnd_cnt
= 0;
2889 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2890 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2892 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2893 icsk
->icsk_mtup
.probe_size
= 0;
2894 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2897 /* Do a simple retransmit without using the backoff mechanisms in
2898 * tcp_timer. This is used for path mtu discovery.
2899 * The socket is already locked here.
2901 void tcp_simple_retransmit(struct sock
*sk
)
2903 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2904 struct tcp_sock
*tp
= tcp_sk(sk
);
2905 struct sk_buff
*skb
;
2906 unsigned int mss
= tcp_current_mss(sk
);
2907 u32 prior_lost
= tp
->lost_out
;
2909 tcp_for_write_queue(skb
, sk
) {
2910 if (skb
== tcp_send_head(sk
))
2912 if (tcp_skb_seglen(skb
) > mss
&&
2913 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2914 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2915 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2916 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2918 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2922 tcp_clear_retrans_hints_partial(tp
);
2924 if (prior_lost
== tp
->lost_out
)
2927 if (tcp_is_reno(tp
))
2928 tcp_limit_reno_sacked(tp
);
2930 tcp_verify_left_out(tp
);
2932 /* Don't muck with the congestion window here.
2933 * Reason is that we do not increase amount of _data_
2934 * in network, but units changed and effective
2935 * cwnd/ssthresh really reduced now.
2937 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2938 tp
->high_seq
= tp
->snd_nxt
;
2939 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2940 tp
->prior_ssthresh
= 0;
2941 tp
->undo_marker
= 0;
2942 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2944 tcp_xmit_retransmit_queue(sk
);
2946 EXPORT_SYMBOL(tcp_simple_retransmit
);
2948 /* Process an event, which can update packets-in-flight not trivially.
2949 * Main goal of this function is to calculate new estimate for left_out,
2950 * taking into account both packets sitting in receiver's buffer and
2951 * packets lost by network.
2953 * Besides that it does CWND reduction, when packet loss is detected
2954 * and changes state of machine.
2956 * It does _not_ decide what to send, it is made in function
2957 * tcp_xmit_retransmit_queue().
2959 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2961 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2962 struct tcp_sock
*tp
= tcp_sk(sk
);
2963 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2964 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2965 (tcp_fackets_out(tp
) > tp
->reordering
));
2966 int fast_rexmit
= 0, mib_idx
;
2968 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2970 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2971 tp
->fackets_out
= 0;
2973 /* Now state machine starts.
2974 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2975 if (flag
& FLAG_ECE
)
2976 tp
->prior_ssthresh
= 0;
2978 /* B. In all the states check for reneging SACKs. */
2979 if (tcp_check_sack_reneging(sk
, flag
))
2982 /* C. Process data loss notification, provided it is valid. */
2983 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2984 before(tp
->snd_una
, tp
->high_seq
) &&
2985 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2986 tp
->fackets_out
> tp
->reordering
) {
2987 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2988 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2991 /* D. Check consistency of the current state. */
2992 tcp_verify_left_out(tp
);
2994 /* E. Check state exit conditions. State can be terminated
2995 * when high_seq is ACKed. */
2996 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2997 WARN_ON(tp
->retrans_out
!= 0);
2998 tp
->retrans_stamp
= 0;
2999 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
3000 switch (icsk
->icsk_ca_state
) {
3002 icsk
->icsk_retransmits
= 0;
3003 if (tcp_try_undo_recovery(sk
))
3008 /* CWR is to be held something *above* high_seq
3009 * is ACKed for CWR bit to reach receiver. */
3010 if (tp
->snd_una
!= tp
->high_seq
) {
3011 tcp_complete_cwr(sk
);
3012 tcp_set_ca_state(sk
, TCP_CA_Open
);
3016 case TCP_CA_Disorder
:
3017 tcp_try_undo_dsack(sk
);
3018 if (!tp
->undo_marker
||
3019 /* For SACK case do not Open to allow to undo
3020 * catching for all duplicate ACKs. */
3021 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
3022 tp
->undo_marker
= 0;
3023 tcp_set_ca_state(sk
, TCP_CA_Open
);
3027 case TCP_CA_Recovery
:
3028 if (tcp_is_reno(tp
))
3029 tcp_reset_reno_sack(tp
);
3030 if (tcp_try_undo_recovery(sk
))
3032 tcp_complete_cwr(sk
);
3037 /* F. Process state. */
3038 switch (icsk
->icsk_ca_state
) {
3039 case TCP_CA_Recovery
:
3040 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
3041 if (tcp_is_reno(tp
) && is_dupack
)
3042 tcp_add_reno_sack(sk
);
3044 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
3047 if (flag
& FLAG_DATA_ACKED
)
3048 icsk
->icsk_retransmits
= 0;
3049 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
3050 tcp_reset_reno_sack(tp
);
3051 if (!tcp_try_undo_loss(sk
)) {
3052 tcp_moderate_cwnd(tp
);
3053 tcp_xmit_retransmit_queue(sk
);
3056 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
3058 /* Loss is undone; fall through to processing in Open state. */
3060 if (tcp_is_reno(tp
)) {
3061 if (flag
& FLAG_SND_UNA_ADVANCED
)
3062 tcp_reset_reno_sack(tp
);
3064 tcp_add_reno_sack(sk
);
3067 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
3068 tcp_try_undo_dsack(sk
);
3070 if (!tcp_time_to_recover(sk
)) {
3071 tcp_try_to_open(sk
, flag
);
3075 /* MTU probe failure: don't reduce cwnd */
3076 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3077 icsk
->icsk_mtup
.probe_size
&&
3078 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3079 tcp_mtup_probe_failed(sk
);
3080 /* Restores the reduction we did in tcp_mtup_probe() */
3082 tcp_simple_retransmit(sk
);
3086 /* Otherwise enter Recovery state */
3088 if (tcp_is_reno(tp
))
3089 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
3091 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
3093 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3095 tp
->high_seq
= tp
->snd_nxt
;
3096 tp
->prior_ssthresh
= 0;
3097 tp
->undo_marker
= tp
->snd_una
;
3098 tp
->undo_retrans
= tp
->retrans_out
;
3100 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
3101 if (!(flag
& FLAG_ECE
))
3102 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
3103 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
3104 TCP_ECN_queue_cwr(tp
);
3107 tp
->bytes_acked
= 0;
3108 tp
->snd_cwnd_cnt
= 0;
3109 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
3113 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3114 tcp_update_scoreboard(sk
, fast_rexmit
);
3115 tcp_cwnd_down(sk
, flag
);
3116 tcp_xmit_retransmit_queue(sk
);
3119 static void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
3121 tcp_rtt_estimator(sk
, seq_rtt
);
3123 inet_csk(sk
)->icsk_backoff
= 0;
3126 /* Read draft-ietf-tcplw-high-performance before mucking
3127 * with this code. (Supersedes RFC1323)
3129 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3131 /* RTTM Rule: A TSecr value received in a segment is used to
3132 * update the averaged RTT measurement only if the segment
3133 * acknowledges some new data, i.e., only if it advances the
3134 * left edge of the send window.
3136 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3137 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3139 * Changed: reset backoff as soon as we see the first valid sample.
3140 * If we do not, we get strongly overestimated rto. With timestamps
3141 * samples are accepted even from very old segments: f.e., when rtt=1
3142 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3143 * answer arrives rto becomes 120 seconds! If at least one of segments
3144 * in window is lost... Voila. --ANK (010210)
3146 struct tcp_sock
*tp
= tcp_sk(sk
);
3148 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
3151 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3153 /* We don't have a timestamp. Can only use
3154 * packets that are not retransmitted to determine
3155 * rtt estimates. Also, we must not reset the
3156 * backoff for rto until we get a non-retransmitted
3157 * packet. This allows us to deal with a situation
3158 * where the network delay has increased suddenly.
3159 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3162 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3165 tcp_valid_rtt_meas(sk
, seq_rtt
);
3168 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3171 const struct tcp_sock
*tp
= tcp_sk(sk
);
3172 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3173 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3174 tcp_ack_saw_tstamp(sk
, flag
);
3175 else if (seq_rtt
>= 0)
3176 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3179 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3181 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3182 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3183 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3186 /* Restart timer after forward progress on connection.
3187 * RFC2988 recommends to restart timer to now+rto.
3189 static void tcp_rearm_rto(struct sock
*sk
)
3191 struct tcp_sock
*tp
= tcp_sk(sk
);
3193 if (!tp
->packets_out
) {
3194 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3196 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3197 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3201 /* If we get here, the whole TSO packet has not been acked. */
3202 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3204 struct tcp_sock
*tp
= tcp_sk(sk
);
3207 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3209 packets_acked
= tcp_skb_pcount(skb
);
3210 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3212 packets_acked
-= tcp_skb_pcount(skb
);
3214 if (packets_acked
) {
3215 BUG_ON(tcp_skb_pcount(skb
) == 0);
3216 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3219 return packets_acked
;
3222 /* Remove acknowledged frames from the retransmission queue. If our packet
3223 * is before the ack sequence we can discard it as it's confirmed to have
3224 * arrived at the other end.
3226 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3229 struct tcp_sock
*tp
= tcp_sk(sk
);
3230 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3231 struct sk_buff
*skb
;
3232 u32 now
= tcp_time_stamp
;
3233 int fully_acked
= 1;
3236 u32 reord
= tp
->packets_out
;
3237 u32 prior_sacked
= tp
->sacked_out
;
3239 s32 ca_seq_rtt
= -1;
3240 ktime_t last_ackt
= net_invalid_timestamp();
3242 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3243 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3245 u8 sacked
= scb
->sacked
;
3247 /* Determine how many packets and what bytes were acked, tso and else */
3248 if (after(scb
->end_seq
, tp
->snd_una
)) {
3249 if (tcp_skb_pcount(skb
) == 1 ||
3250 !after(tp
->snd_una
, scb
->seq
))
3253 acked_pcount
= tcp_tso_acked(sk
, skb
);
3259 acked_pcount
= tcp_skb_pcount(skb
);
3262 if (sacked
& TCPCB_RETRANS
) {
3263 if (sacked
& TCPCB_SACKED_RETRANS
)
3264 tp
->retrans_out
-= acked_pcount
;
3265 flag
|= FLAG_RETRANS_DATA_ACKED
;
3268 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3269 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3271 ca_seq_rtt
= now
- scb
->when
;
3272 last_ackt
= skb
->tstamp
;
3274 seq_rtt
= ca_seq_rtt
;
3276 if (!(sacked
& TCPCB_SACKED_ACKED
))
3277 reord
= min(pkts_acked
, reord
);
3280 if (sacked
& TCPCB_SACKED_ACKED
)
3281 tp
->sacked_out
-= acked_pcount
;
3282 if (sacked
& TCPCB_LOST
)
3283 tp
->lost_out
-= acked_pcount
;
3285 tp
->packets_out
-= acked_pcount
;
3286 pkts_acked
+= acked_pcount
;
3288 /* Initial outgoing SYN's get put onto the write_queue
3289 * just like anything else we transmit. It is not
3290 * true data, and if we misinform our callers that
3291 * this ACK acks real data, we will erroneously exit
3292 * connection startup slow start one packet too
3293 * quickly. This is severely frowned upon behavior.
3295 if (!(scb
->flags
& TCPHDR_SYN
)) {
3296 flag
|= FLAG_DATA_ACKED
;
3298 flag
|= FLAG_SYN_ACKED
;
3299 tp
->retrans_stamp
= 0;
3305 tcp_unlink_write_queue(skb
, sk
);
3306 sk_wmem_free_skb(sk
, skb
);
3307 tp
->scoreboard_skb_hint
= NULL
;
3308 if (skb
== tp
->retransmit_skb_hint
)
3309 tp
->retransmit_skb_hint
= NULL
;
3310 if (skb
== tp
->lost_skb_hint
)
3311 tp
->lost_skb_hint
= NULL
;
3314 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3315 tp
->snd_up
= tp
->snd_una
;
3317 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3318 flag
|= FLAG_SACK_RENEGING
;
3320 if (flag
& FLAG_ACKED
) {
3321 const struct tcp_congestion_ops
*ca_ops
3322 = inet_csk(sk
)->icsk_ca_ops
;
3324 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3325 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3326 tcp_mtup_probe_success(sk
);
3329 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3332 if (tcp_is_reno(tp
)) {
3333 tcp_remove_reno_sacks(sk
, pkts_acked
);
3337 /* Non-retransmitted hole got filled? That's reordering */
3338 if (reord
< prior_fackets
)
3339 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3341 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3342 prior_sacked
- tp
->sacked_out
;
3343 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3346 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3348 if (ca_ops
->pkts_acked
) {
3351 /* Is the ACK triggering packet unambiguous? */
3352 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3353 /* High resolution needed and available? */
3354 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3355 !ktime_equal(last_ackt
,
3356 net_invalid_timestamp()))
3357 rtt_us
= ktime_us_delta(ktime_get_real(),
3359 else if (ca_seq_rtt
> 0)
3360 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3363 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3367 #if FASTRETRANS_DEBUG > 0
3368 WARN_ON((int)tp
->sacked_out
< 0);
3369 WARN_ON((int)tp
->lost_out
< 0);
3370 WARN_ON((int)tp
->retrans_out
< 0);
3371 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3372 icsk
= inet_csk(sk
);
3374 printk(KERN_DEBUG
"Leak l=%u %d\n",
3375 tp
->lost_out
, icsk
->icsk_ca_state
);
3378 if (tp
->sacked_out
) {
3379 printk(KERN_DEBUG
"Leak s=%u %d\n",
3380 tp
->sacked_out
, icsk
->icsk_ca_state
);
3383 if (tp
->retrans_out
) {
3384 printk(KERN_DEBUG
"Leak r=%u %d\n",
3385 tp
->retrans_out
, icsk
->icsk_ca_state
);
3386 tp
->retrans_out
= 0;
3393 static void tcp_ack_probe(struct sock
*sk
)
3395 const struct tcp_sock
*tp
= tcp_sk(sk
);
3396 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3398 /* Was it a usable window open? */
3400 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3401 icsk
->icsk_backoff
= 0;
3402 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3403 /* Socket must be waked up by subsequent tcp_data_snd_check().
3404 * This function is not for random using!
3407 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3408 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3413 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3415 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3416 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3419 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3421 const struct tcp_sock
*tp
= tcp_sk(sk
);
3422 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3423 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3426 /* Check that window update is acceptable.
3427 * The function assumes that snd_una<=ack<=snd_next.
3429 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3430 const u32 ack
, const u32 ack_seq
,
3433 return (after(ack
, tp
->snd_una
) ||
3434 after(ack_seq
, tp
->snd_wl1
) ||
3435 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3438 /* Update our send window.
3440 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3441 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3443 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3446 struct tcp_sock
*tp
= tcp_sk(sk
);
3448 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3450 if (likely(!tcp_hdr(skb
)->syn
))
3451 nwin
<<= tp
->rx_opt
.snd_wscale
;
3453 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3454 flag
|= FLAG_WIN_UPDATE
;
3455 tcp_update_wl(tp
, ack_seq
);
3457 if (tp
->snd_wnd
!= nwin
) {
3460 /* Note, it is the only place, where
3461 * fast path is recovered for sending TCP.
3464 tcp_fast_path_check(sk
);
3466 if (nwin
> tp
->max_window
) {
3467 tp
->max_window
= nwin
;
3468 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3478 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3479 * continue in congestion avoidance.
3481 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3483 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3484 tp
->snd_cwnd_cnt
= 0;
3485 tp
->bytes_acked
= 0;
3486 TCP_ECN_queue_cwr(tp
);
3487 tcp_moderate_cwnd(tp
);
3490 /* A conservative spurious RTO response algorithm: reduce cwnd using
3491 * rate halving and continue in congestion avoidance.
3493 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3495 tcp_enter_cwr(sk
, 0);
3498 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3500 if (flag
& FLAG_ECE
)
3501 tcp_ratehalving_spur_to_response(sk
);
3503 tcp_undo_cwr(sk
, 1);
3506 /* F-RTO spurious RTO detection algorithm (RFC4138)
3508 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3509 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3510 * window (but not to or beyond highest sequence sent before RTO):
3511 * On First ACK, send two new segments out.
3512 * On Second ACK, RTO was likely spurious. Do spurious response (response
3513 * algorithm is not part of the F-RTO detection algorithm
3514 * given in RFC4138 but can be selected separately).
3515 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3516 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3517 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3518 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3520 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3521 * original window even after we transmit two new data segments.
3524 * on first step, wait until first cumulative ACK arrives, then move to
3525 * the second step. In second step, the next ACK decides.
3527 * F-RTO is implemented (mainly) in four functions:
3528 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3529 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3530 * called when tcp_use_frto() showed green light
3531 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3532 * - tcp_enter_frto_loss() is called if there is not enough evidence
3533 * to prove that the RTO is indeed spurious. It transfers the control
3534 * from F-RTO to the conventional RTO recovery
3536 static int tcp_process_frto(struct sock
*sk
, int flag
)
3538 struct tcp_sock
*tp
= tcp_sk(sk
);
3540 tcp_verify_left_out(tp
);
3542 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3543 if (flag
& FLAG_DATA_ACKED
)
3544 inet_csk(sk
)->icsk_retransmits
= 0;
3546 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3547 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3548 tp
->undo_marker
= 0;
3550 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3551 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3555 if (!tcp_is_sackfrto(tp
)) {
3556 /* RFC4138 shortcoming in step 2; should also have case c):
3557 * ACK isn't duplicate nor advances window, e.g., opposite dir
3560 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3563 if (!(flag
& FLAG_DATA_ACKED
)) {
3564 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3569 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3570 /* Prevent sending of new data. */
3571 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3572 tcp_packets_in_flight(tp
));
3576 if ((tp
->frto_counter
>= 2) &&
3577 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3578 ((flag
& FLAG_DATA_SACKED
) &&
3579 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3580 /* RFC4138 shortcoming (see comment above) */
3581 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3582 (flag
& FLAG_NOT_DUP
))
3585 tcp_enter_frto_loss(sk
, 3, flag
);
3590 if (tp
->frto_counter
== 1) {
3591 /* tcp_may_send_now needs to see updated state */
3592 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3593 tp
->frto_counter
= 2;
3595 if (!tcp_may_send_now(sk
))
3596 tcp_enter_frto_loss(sk
, 2, flag
);
3600 switch (sysctl_tcp_frto_response
) {
3602 tcp_undo_spur_to_response(sk
, flag
);
3605 tcp_conservative_spur_to_response(tp
);
3608 tcp_ratehalving_spur_to_response(sk
);
3611 tp
->frto_counter
= 0;
3612 tp
->undo_marker
= 0;
3613 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3618 /* This routine deals with incoming acks, but not outgoing ones. */
3619 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3621 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3622 struct tcp_sock
*tp
= tcp_sk(sk
);
3623 u32 prior_snd_una
= tp
->snd_una
;
3624 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3625 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3626 u32 prior_in_flight
;
3631 /* If the ack is older than previous acks
3632 * then we can probably ignore it.
3634 if (before(ack
, prior_snd_una
))
3637 /* If the ack includes data we haven't sent yet, discard
3638 * this segment (RFC793 Section 3.9).
3640 if (after(ack
, tp
->snd_nxt
))
3643 if (after(ack
, prior_snd_una
))
3644 flag
|= FLAG_SND_UNA_ADVANCED
;
3646 if (sysctl_tcp_abc
) {
3647 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3648 tp
->bytes_acked
+= ack
- prior_snd_una
;
3649 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3650 /* we assume just one segment left network */
3651 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3655 prior_fackets
= tp
->fackets_out
;
3656 prior_in_flight
= tcp_packets_in_flight(tp
);
3658 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3659 /* Window is constant, pure forward advance.
3660 * No more checks are required.
3661 * Note, we use the fact that SND.UNA>=SND.WL2.
3663 tcp_update_wl(tp
, ack_seq
);
3665 flag
|= FLAG_WIN_UPDATE
;
3667 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3669 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3671 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3674 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3676 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3678 if (TCP_SKB_CB(skb
)->sacked
)
3679 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3681 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3684 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3687 /* We passed data and got it acked, remove any soft error
3688 * log. Something worked...
3690 sk
->sk_err_soft
= 0;
3691 icsk
->icsk_probes_out
= 0;
3692 tp
->rcv_tstamp
= tcp_time_stamp
;
3693 prior_packets
= tp
->packets_out
;
3697 /* See if we can take anything off of the retransmit queue. */
3698 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3700 if (tp
->frto_counter
)
3701 frto_cwnd
= tcp_process_frto(sk
, flag
);
3702 /* Guarantee sacktag reordering detection against wrap-arounds */
3703 if (before(tp
->frto_highmark
, tp
->snd_una
))
3704 tp
->frto_highmark
= 0;
3706 if (tcp_ack_is_dubious(sk
, flag
)) {
3707 /* Advance CWND, if state allows this. */
3708 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3709 tcp_may_raise_cwnd(sk
, flag
))
3710 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3711 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3714 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3715 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3718 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3719 dst_confirm(__sk_dst_get(sk
));
3724 /* If this ack opens up a zero window, clear backoff. It was
3725 * being used to time the probes, and is probably far higher than
3726 * it needs to be for normal retransmission.
3728 if (tcp_send_head(sk
))
3733 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3737 if (TCP_SKB_CB(skb
)->sacked
) {
3738 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3739 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3740 tcp_try_keep_open(sk
);
3743 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3747 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3748 * But, this can also be called on packets in the established flow when
3749 * the fast version below fails.
3751 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3752 u8
**hvpp
, int estab
)
3755 struct tcphdr
*th
= tcp_hdr(skb
);
3756 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3758 ptr
= (unsigned char *)(th
+ 1);
3759 opt_rx
->saw_tstamp
= 0;
3761 while (length
> 0) {
3762 int opcode
= *ptr
++;
3768 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3773 if (opsize
< 2) /* "silly options" */
3775 if (opsize
> length
)
3776 return; /* don't parse partial options */
3779 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3780 u16 in_mss
= get_unaligned_be16(ptr
);
3782 if (opt_rx
->user_mss
&&
3783 opt_rx
->user_mss
< in_mss
)
3784 in_mss
= opt_rx
->user_mss
;
3785 opt_rx
->mss_clamp
= in_mss
;
3790 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3791 !estab
&& sysctl_tcp_window_scaling
) {
3792 __u8 snd_wscale
= *(__u8
*)ptr
;
3793 opt_rx
->wscale_ok
= 1;
3794 if (snd_wscale
> 14) {
3795 if (net_ratelimit())
3796 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3797 "scaling value %d >14 received.\n",
3801 opt_rx
->snd_wscale
= snd_wscale
;
3804 case TCPOPT_TIMESTAMP
:
3805 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3806 ((estab
&& opt_rx
->tstamp_ok
) ||
3807 (!estab
&& sysctl_tcp_timestamps
))) {
3808 opt_rx
->saw_tstamp
= 1;
3809 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3810 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3813 case TCPOPT_SACK_PERM
:
3814 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3815 !estab
&& sysctl_tcp_sack
) {
3816 opt_rx
->sack_ok
= 1;
3817 tcp_sack_reset(opt_rx
);
3822 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3823 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3825 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3828 #ifdef CONFIG_TCP_MD5SIG
3831 * The MD5 Hash has already been
3832 * checked (see tcp_v{4,6}_do_rcv()).
3837 /* This option is variable length.
3840 case TCPOLEN_COOKIE_BASE
:
3841 /* not yet implemented */
3843 case TCPOLEN_COOKIE_PAIR
:
3844 /* not yet implemented */
3846 case TCPOLEN_COOKIE_MIN
+0:
3847 case TCPOLEN_COOKIE_MIN
+2:
3848 case TCPOLEN_COOKIE_MIN
+4:
3849 case TCPOLEN_COOKIE_MIN
+6:
3850 case TCPOLEN_COOKIE_MAX
:
3851 /* 16-bit multiple */
3852 opt_rx
->cookie_plus
= opsize
;
3867 EXPORT_SYMBOL(tcp_parse_options
);
3869 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3871 __be32
*ptr
= (__be32
*)(th
+ 1);
3873 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3874 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3875 tp
->rx_opt
.saw_tstamp
= 1;
3877 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3879 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3885 /* Fast parse options. This hopes to only see timestamps.
3886 * If it is wrong it falls back on tcp_parse_options().
3888 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3889 struct tcp_sock
*tp
, u8
**hvpp
)
3891 /* In the spirit of fast parsing, compare doff directly to constant
3892 * values. Because equality is used, short doff can be ignored here.
3894 if (th
->doff
== (sizeof(*th
) / 4)) {
3895 tp
->rx_opt
.saw_tstamp
= 0;
3897 } else if (tp
->rx_opt
.tstamp_ok
&&
3898 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3899 if (tcp_parse_aligned_timestamp(tp
, th
))
3902 tcp_parse_options(skb
, &tp
->rx_opt
, hvpp
, 1);
3906 #ifdef CONFIG_TCP_MD5SIG
3908 * Parse MD5 Signature option
3910 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3912 int length
= (th
->doff
<< 2) - sizeof (*th
);
3913 u8
*ptr
= (u8
*)(th
+ 1);
3915 /* If the TCP option is too short, we can short cut */
3916 if (length
< TCPOLEN_MD5SIG
)
3919 while (length
> 0) {
3920 int opcode
= *ptr
++;
3931 if (opsize
< 2 || opsize
> length
)
3933 if (opcode
== TCPOPT_MD5SIG
)
3934 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3941 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3944 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3946 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3947 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3950 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3952 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3953 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3954 * extra check below makes sure this can only happen
3955 * for pure ACK frames. -DaveM
3957 * Not only, also it occurs for expired timestamps.
3960 if (tcp_paws_check(&tp
->rx_opt
, 0))
3961 tcp_store_ts_recent(tp
);
3965 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3967 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3968 * it can pass through stack. So, the following predicate verifies that
3969 * this segment is not used for anything but congestion avoidance or
3970 * fast retransmit. Moreover, we even are able to eliminate most of such
3971 * second order effects, if we apply some small "replay" window (~RTO)
3972 * to timestamp space.
3974 * All these measures still do not guarantee that we reject wrapped ACKs
3975 * on networks with high bandwidth, when sequence space is recycled fastly,
3976 * but it guarantees that such events will be very rare and do not affect
3977 * connection seriously. This doesn't look nice, but alas, PAWS is really
3980 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3981 * states that events when retransmit arrives after original data are rare.
3982 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3983 * the biggest problem on large power networks even with minor reordering.
3984 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3985 * up to bandwidth of 18Gigabit/sec. 8) ]
3988 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3990 struct tcp_sock
*tp
= tcp_sk(sk
);
3991 struct tcphdr
*th
= tcp_hdr(skb
);
3992 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3993 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3995 return (/* 1. Pure ACK with correct sequence number. */
3996 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3998 /* 2. ... and duplicate ACK. */
3999 ack
== tp
->snd_una
&&
4001 /* 3. ... and does not update window. */
4002 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
4004 /* 4. ... and sits in replay window. */
4005 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
4008 static inline int tcp_paws_discard(const struct sock
*sk
,
4009 const struct sk_buff
*skb
)
4011 const struct tcp_sock
*tp
= tcp_sk(sk
);
4013 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
4014 !tcp_disordered_ack(sk
, skb
);
4017 /* Check segment sequence number for validity.
4019 * Segment controls are considered valid, if the segment
4020 * fits to the window after truncation to the window. Acceptability
4021 * of data (and SYN, FIN, of course) is checked separately.
4022 * See tcp_data_queue(), for example.
4024 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4025 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4026 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4027 * (borrowed from freebsd)
4030 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4032 return !before(end_seq
, tp
->rcv_wup
) &&
4033 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4036 /* When we get a reset we do this. */
4037 static void tcp_reset(struct sock
*sk
)
4039 /* We want the right error as BSD sees it (and indeed as we do). */
4040 switch (sk
->sk_state
) {
4042 sk
->sk_err
= ECONNREFUSED
;
4044 case TCP_CLOSE_WAIT
:
4050 sk
->sk_err
= ECONNRESET
;
4052 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4055 if (!sock_flag(sk
, SOCK_DEAD
))
4056 sk
->sk_error_report(sk
);
4062 * Process the FIN bit. This now behaves as it is supposed to work
4063 * and the FIN takes effect when it is validly part of sequence
4064 * space. Not before when we get holes.
4066 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4067 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4070 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4071 * close and we go into CLOSING (and later onto TIME-WAIT)
4073 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4075 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
4077 struct tcp_sock
*tp
= tcp_sk(sk
);
4079 inet_csk_schedule_ack(sk
);
4081 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4082 sock_set_flag(sk
, SOCK_DONE
);
4084 switch (sk
->sk_state
) {
4086 case TCP_ESTABLISHED
:
4087 /* Move to CLOSE_WAIT */
4088 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4089 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4092 case TCP_CLOSE_WAIT
:
4094 /* Received a retransmission of the FIN, do
4099 /* RFC793: Remain in the LAST-ACK state. */
4103 /* This case occurs when a simultaneous close
4104 * happens, we must ack the received FIN and
4105 * enter the CLOSING state.
4108 tcp_set_state(sk
, TCP_CLOSING
);
4111 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4113 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4116 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4117 * cases we should never reach this piece of code.
4119 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
4120 __func__
, sk
->sk_state
);
4124 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4125 * Probably, we should reset in this case. For now drop them.
4127 __skb_queue_purge(&tp
->out_of_order_queue
);
4128 if (tcp_is_sack(tp
))
4129 tcp_sack_reset(&tp
->rx_opt
);
4132 if (!sock_flag(sk
, SOCK_DEAD
)) {
4133 sk
->sk_state_change(sk
);
4135 /* Do not send POLL_HUP for half duplex close. */
4136 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4137 sk
->sk_state
== TCP_CLOSE
)
4138 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4140 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4144 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4147 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4148 if (before(seq
, sp
->start_seq
))
4149 sp
->start_seq
= seq
;
4150 if (after(end_seq
, sp
->end_seq
))
4151 sp
->end_seq
= end_seq
;
4157 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4159 struct tcp_sock
*tp
= tcp_sk(sk
);
4161 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4164 if (before(seq
, tp
->rcv_nxt
))
4165 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4167 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4169 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4171 tp
->rx_opt
.dsack
= 1;
4172 tp
->duplicate_sack
[0].start_seq
= seq
;
4173 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4177 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4179 struct tcp_sock
*tp
= tcp_sk(sk
);
4181 if (!tp
->rx_opt
.dsack
)
4182 tcp_dsack_set(sk
, seq
, end_seq
);
4184 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4187 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
4189 struct tcp_sock
*tp
= tcp_sk(sk
);
4191 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4192 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4193 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4194 tcp_enter_quickack_mode(sk
);
4196 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4197 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4199 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4200 end_seq
= tp
->rcv_nxt
;
4201 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4208 /* These routines update the SACK block as out-of-order packets arrive or
4209 * in-order packets close up the sequence space.
4211 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4214 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4215 struct tcp_sack_block
*swalk
= sp
+ 1;
4217 /* See if the recent change to the first SACK eats into
4218 * or hits the sequence space of other SACK blocks, if so coalesce.
4220 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4221 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4224 /* Zap SWALK, by moving every further SACK up by one slot.
4225 * Decrease num_sacks.
4227 tp
->rx_opt
.num_sacks
--;
4228 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4232 this_sack
++, swalk
++;
4236 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4238 struct tcp_sock
*tp
= tcp_sk(sk
);
4239 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4240 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4246 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4247 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4248 /* Rotate this_sack to the first one. */
4249 for (; this_sack
> 0; this_sack
--, sp
--)
4250 swap(*sp
, *(sp
- 1));
4252 tcp_sack_maybe_coalesce(tp
);
4257 /* Could not find an adjacent existing SACK, build a new one,
4258 * put it at the front, and shift everyone else down. We
4259 * always know there is at least one SACK present already here.
4261 * If the sack array is full, forget about the last one.
4263 if (this_sack
>= TCP_NUM_SACKS
) {
4265 tp
->rx_opt
.num_sacks
--;
4268 for (; this_sack
> 0; this_sack
--, sp
--)
4272 /* Build the new head SACK, and we're done. */
4273 sp
->start_seq
= seq
;
4274 sp
->end_seq
= end_seq
;
4275 tp
->rx_opt
.num_sacks
++;
4278 /* RCV.NXT advances, some SACKs should be eaten. */
4280 static void tcp_sack_remove(struct tcp_sock
*tp
)
4282 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4283 int num_sacks
= tp
->rx_opt
.num_sacks
;
4286 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4287 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4288 tp
->rx_opt
.num_sacks
= 0;
4292 for (this_sack
= 0; this_sack
< num_sacks
;) {
4293 /* Check if the start of the sack is covered by RCV.NXT. */
4294 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4297 /* RCV.NXT must cover all the block! */
4298 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4300 /* Zap this SACK, by moving forward any other SACKS. */
4301 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4302 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4309 tp
->rx_opt
.num_sacks
= num_sacks
;
4312 /* This one checks to see if we can put data from the
4313 * out_of_order queue into the receive_queue.
4315 static void tcp_ofo_queue(struct sock
*sk
)
4317 struct tcp_sock
*tp
= tcp_sk(sk
);
4318 __u32 dsack_high
= tp
->rcv_nxt
;
4319 struct sk_buff
*skb
;
4321 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4322 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4325 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4326 __u32 dsack
= dsack_high
;
4327 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4328 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4329 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4332 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4333 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4334 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4338 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4339 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4340 TCP_SKB_CB(skb
)->end_seq
);
4342 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4343 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4344 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4345 if (tcp_hdr(skb
)->fin
)
4346 tcp_fin(skb
, sk
, tcp_hdr(skb
));
4350 static int tcp_prune_ofo_queue(struct sock
*sk
);
4351 static int tcp_prune_queue(struct sock
*sk
);
4353 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
4355 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4356 !sk_rmem_schedule(sk
, size
)) {
4358 if (tcp_prune_queue(sk
) < 0)
4361 if (!sk_rmem_schedule(sk
, size
)) {
4362 if (!tcp_prune_ofo_queue(sk
))
4365 if (!sk_rmem_schedule(sk
, size
))
4372 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4374 struct tcphdr
*th
= tcp_hdr(skb
);
4375 struct tcp_sock
*tp
= tcp_sk(sk
);
4378 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4382 __skb_pull(skb
, th
->doff
* 4);
4384 TCP_ECN_accept_cwr(tp
, skb
);
4386 tp
->rx_opt
.dsack
= 0;
4388 /* Queue data for delivery to the user.
4389 * Packets in sequence go to the receive queue.
4390 * Out of sequence packets to the out_of_order_queue.
4392 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4393 if (tcp_receive_window(tp
) == 0)
4396 /* Ok. In sequence. In window. */
4397 if (tp
->ucopy
.task
== current
&&
4398 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4399 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4400 int chunk
= min_t(unsigned int, skb
->len
,
4403 __set_current_state(TASK_RUNNING
);
4406 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4407 tp
->ucopy
.len
-= chunk
;
4408 tp
->copied_seq
+= chunk
;
4409 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4410 tcp_rcv_space_adjust(sk
);
4418 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4421 skb_set_owner_r(skb
, sk
);
4422 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4424 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4426 tcp_event_data_recv(sk
, skb
);
4428 tcp_fin(skb
, sk
, th
);
4430 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4433 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4434 * gap in queue is filled.
4436 if (skb_queue_empty(&tp
->out_of_order_queue
))
4437 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4440 if (tp
->rx_opt
.num_sacks
)
4441 tcp_sack_remove(tp
);
4443 tcp_fast_path_check(sk
);
4447 else if (!sock_flag(sk
, SOCK_DEAD
))
4448 sk
->sk_data_ready(sk
, 0);
4452 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4453 /* A retransmit, 2nd most common case. Force an immediate ack. */
4454 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4455 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4458 tcp_enter_quickack_mode(sk
);
4459 inet_csk_schedule_ack(sk
);
4465 /* Out of window. F.e. zero window probe. */
4466 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4469 tcp_enter_quickack_mode(sk
);
4471 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4472 /* Partial packet, seq < rcv_next < end_seq */
4473 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4474 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4475 TCP_SKB_CB(skb
)->end_seq
);
4477 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4479 /* If window is closed, drop tail of packet. But after
4480 * remembering D-SACK for its head made in previous line.
4482 if (!tcp_receive_window(tp
))
4487 TCP_ECN_check_ce(tp
, skb
);
4489 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4492 /* Disable header prediction. */
4494 inet_csk_schedule_ack(sk
);
4496 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4497 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4499 skb_set_owner_r(skb
, sk
);
4501 if (!skb_peek(&tp
->out_of_order_queue
)) {
4502 /* Initial out of order segment, build 1 SACK. */
4503 if (tcp_is_sack(tp
)) {
4504 tp
->rx_opt
.num_sacks
= 1;
4505 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4506 tp
->selective_acks
[0].end_seq
=
4507 TCP_SKB_CB(skb
)->end_seq
;
4509 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4511 struct sk_buff
*skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4512 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4513 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4515 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4516 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4518 if (!tp
->rx_opt
.num_sacks
||
4519 tp
->selective_acks
[0].end_seq
!= seq
)
4522 /* Common case: data arrive in order after hole. */
4523 tp
->selective_acks
[0].end_seq
= end_seq
;
4527 /* Find place to insert this segment. */
4529 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4531 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4535 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4538 /* Do skb overlap to previous one? */
4539 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4540 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4541 /* All the bits are present. Drop. */
4543 tcp_dsack_set(sk
, seq
, end_seq
);
4546 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4547 /* Partial overlap. */
4548 tcp_dsack_set(sk
, seq
,
4549 TCP_SKB_CB(skb1
)->end_seq
);
4551 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4555 skb1
= skb_queue_prev(
4556 &tp
->out_of_order_queue
,
4561 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4563 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4565 /* And clean segments covered by new one as whole. */
4566 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4567 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4569 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4571 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4572 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4576 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4577 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4578 TCP_SKB_CB(skb1
)->end_seq
);
4583 if (tcp_is_sack(tp
))
4584 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4588 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4589 struct sk_buff_head
*list
)
4591 struct sk_buff
*next
= NULL
;
4593 if (!skb_queue_is_last(list
, skb
))
4594 next
= skb_queue_next(list
, skb
);
4596 __skb_unlink(skb
, list
);
4598 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4603 /* Collapse contiguous sequence of skbs head..tail with
4604 * sequence numbers start..end.
4606 * If tail is NULL, this means until the end of the list.
4608 * Segments with FIN/SYN are not collapsed (only because this
4612 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4613 struct sk_buff
*head
, struct sk_buff
*tail
,
4616 struct sk_buff
*skb
, *n
;
4619 /* First, check that queue is collapsible and find
4620 * the point where collapsing can be useful. */
4624 skb_queue_walk_from_safe(list
, skb
, n
) {
4627 /* No new bits? It is possible on ofo queue. */
4628 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4629 skb
= tcp_collapse_one(sk
, skb
, list
);
4635 /* The first skb to collapse is:
4637 * - bloated or contains data before "start" or
4638 * overlaps to the next one.
4640 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4641 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4642 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4643 end_of_skbs
= false;
4647 if (!skb_queue_is_last(list
, skb
)) {
4648 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4650 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4651 end_of_skbs
= false;
4656 /* Decided to skip this, advance start seq. */
4657 start
= TCP_SKB_CB(skb
)->end_seq
;
4659 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4662 while (before(start
, end
)) {
4663 struct sk_buff
*nskb
;
4664 unsigned int header
= skb_headroom(skb
);
4665 int copy
= SKB_MAX_ORDER(header
, 0);
4667 /* Too big header? This can happen with IPv6. */
4670 if (end
- start
< copy
)
4672 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4676 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4677 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4679 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4681 skb_reserve(nskb
, header
);
4682 memcpy(nskb
->head
, skb
->head
, header
);
4683 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4684 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4685 __skb_queue_before(list
, skb
, nskb
);
4686 skb_set_owner_r(nskb
, sk
);
4688 /* Copy data, releasing collapsed skbs. */
4690 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4691 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4695 size
= min(copy
, size
);
4696 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4698 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4702 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4703 skb
= tcp_collapse_one(sk
, skb
, list
);
4706 tcp_hdr(skb
)->syn
||
4714 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4715 * and tcp_collapse() them until all the queue is collapsed.
4717 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4719 struct tcp_sock
*tp
= tcp_sk(sk
);
4720 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4721 struct sk_buff
*head
;
4727 start
= TCP_SKB_CB(skb
)->seq
;
4728 end
= TCP_SKB_CB(skb
)->end_seq
;
4732 struct sk_buff
*next
= NULL
;
4734 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4735 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4738 /* Segment is terminated when we see gap or when
4739 * we are at the end of all the queue. */
4741 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4742 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4743 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4744 head
, skb
, start
, end
);
4748 /* Start new segment */
4749 start
= TCP_SKB_CB(skb
)->seq
;
4750 end
= TCP_SKB_CB(skb
)->end_seq
;
4752 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4753 start
= TCP_SKB_CB(skb
)->seq
;
4754 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4755 end
= TCP_SKB_CB(skb
)->end_seq
;
4761 * Purge the out-of-order queue.
4762 * Return true if queue was pruned.
4764 static int tcp_prune_ofo_queue(struct sock
*sk
)
4766 struct tcp_sock
*tp
= tcp_sk(sk
);
4769 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4770 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4771 __skb_queue_purge(&tp
->out_of_order_queue
);
4773 /* Reset SACK state. A conforming SACK implementation will
4774 * do the same at a timeout based retransmit. When a connection
4775 * is in a sad state like this, we care only about integrity
4776 * of the connection not performance.
4778 if (tp
->rx_opt
.sack_ok
)
4779 tcp_sack_reset(&tp
->rx_opt
);
4786 /* Reduce allocated memory if we can, trying to get
4787 * the socket within its memory limits again.
4789 * Return less than zero if we should start dropping frames
4790 * until the socket owning process reads some of the data
4791 * to stabilize the situation.
4793 static int tcp_prune_queue(struct sock
*sk
)
4795 struct tcp_sock
*tp
= tcp_sk(sk
);
4797 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4799 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4801 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4802 tcp_clamp_window(sk
);
4803 else if (tcp_memory_pressure
)
4804 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4806 tcp_collapse_ofo_queue(sk
);
4807 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4808 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4809 skb_peek(&sk
->sk_receive_queue
),
4811 tp
->copied_seq
, tp
->rcv_nxt
);
4814 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4817 /* Collapsing did not help, destructive actions follow.
4818 * This must not ever occur. */
4820 tcp_prune_ofo_queue(sk
);
4822 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4825 /* If we are really being abused, tell the caller to silently
4826 * drop receive data on the floor. It will get retransmitted
4827 * and hopefully then we'll have sufficient space.
4829 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4831 /* Massive buffer overcommit. */
4836 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4837 * As additional protections, we do not touch cwnd in retransmission phases,
4838 * and if application hit its sndbuf limit recently.
4840 void tcp_cwnd_application_limited(struct sock
*sk
)
4842 struct tcp_sock
*tp
= tcp_sk(sk
);
4844 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4845 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4846 /* Limited by application or receiver window. */
4847 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4848 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4849 if (win_used
< tp
->snd_cwnd
) {
4850 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4851 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4853 tp
->snd_cwnd_used
= 0;
4855 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4858 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4860 struct tcp_sock
*tp
= tcp_sk(sk
);
4862 /* If the user specified a specific send buffer setting, do
4865 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4868 /* If we are under global TCP memory pressure, do not expand. */
4869 if (tcp_memory_pressure
)
4872 /* If we are under soft global TCP memory pressure, do not expand. */
4873 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4876 /* If we filled the congestion window, do not expand. */
4877 if (tp
->packets_out
>= tp
->snd_cwnd
)
4883 /* When incoming ACK allowed to free some skb from write_queue,
4884 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4885 * on the exit from tcp input handler.
4887 * PROBLEM: sndbuf expansion does not work well with largesend.
4889 static void tcp_new_space(struct sock
*sk
)
4891 struct tcp_sock
*tp
= tcp_sk(sk
);
4893 if (tcp_should_expand_sndbuf(sk
)) {
4894 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4895 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
4896 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4897 tp
->reordering
+ 1);
4898 sndmem
*= 2 * demanded
;
4899 if (sndmem
> sk
->sk_sndbuf
)
4900 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4901 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4904 sk
->sk_write_space(sk
);
4907 static void tcp_check_space(struct sock
*sk
)
4909 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4910 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4911 if (sk
->sk_socket
&&
4912 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4917 static inline void tcp_data_snd_check(struct sock
*sk
)
4919 tcp_push_pending_frames(sk
);
4920 tcp_check_space(sk
);
4924 * Check if sending an ack is needed.
4926 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4928 struct tcp_sock
*tp
= tcp_sk(sk
);
4930 /* More than one full frame received... */
4931 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4932 /* ... and right edge of window advances far enough.
4933 * (tcp_recvmsg() will send ACK otherwise). Or...
4935 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4936 /* We ACK each frame or... */
4937 tcp_in_quickack_mode(sk
) ||
4938 /* We have out of order data. */
4939 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4940 /* Then ack it now */
4943 /* Else, send delayed ack. */
4944 tcp_send_delayed_ack(sk
);
4948 static inline void tcp_ack_snd_check(struct sock
*sk
)
4950 if (!inet_csk_ack_scheduled(sk
)) {
4951 /* We sent a data segment already. */
4954 __tcp_ack_snd_check(sk
, 1);
4958 * This routine is only called when we have urgent data
4959 * signaled. Its the 'slow' part of tcp_urg. It could be
4960 * moved inline now as tcp_urg is only called from one
4961 * place. We handle URGent data wrong. We have to - as
4962 * BSD still doesn't use the correction from RFC961.
4963 * For 1003.1g we should support a new option TCP_STDURG to permit
4964 * either form (or just set the sysctl tcp_stdurg).
4967 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4969 struct tcp_sock
*tp
= tcp_sk(sk
);
4970 u32 ptr
= ntohs(th
->urg_ptr
);
4972 if (ptr
&& !sysctl_tcp_stdurg
)
4974 ptr
+= ntohl(th
->seq
);
4976 /* Ignore urgent data that we've already seen and read. */
4977 if (after(tp
->copied_seq
, ptr
))
4980 /* Do not replay urg ptr.
4982 * NOTE: interesting situation not covered by specs.
4983 * Misbehaving sender may send urg ptr, pointing to segment,
4984 * which we already have in ofo queue. We are not able to fetch
4985 * such data and will stay in TCP_URG_NOTYET until will be eaten
4986 * by recvmsg(). Seems, we are not obliged to handle such wicked
4987 * situations. But it is worth to think about possibility of some
4988 * DoSes using some hypothetical application level deadlock.
4990 if (before(ptr
, tp
->rcv_nxt
))
4993 /* Do we already have a newer (or duplicate) urgent pointer? */
4994 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4997 /* Tell the world about our new urgent pointer. */
5000 /* We may be adding urgent data when the last byte read was
5001 * urgent. To do this requires some care. We cannot just ignore
5002 * tp->copied_seq since we would read the last urgent byte again
5003 * as data, nor can we alter copied_seq until this data arrives
5004 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5006 * NOTE. Double Dutch. Rendering to plain English: author of comment
5007 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5008 * and expect that both A and B disappear from stream. This is _wrong_.
5009 * Though this happens in BSD with high probability, this is occasional.
5010 * Any application relying on this is buggy. Note also, that fix "works"
5011 * only in this artificial test. Insert some normal data between A and B and we will
5012 * decline of BSD again. Verdict: it is better to remove to trap
5015 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5016 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5017 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5019 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5020 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5025 tp
->urg_data
= TCP_URG_NOTYET
;
5028 /* Disable header prediction. */
5032 /* This is the 'fast' part of urgent handling. */
5033 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
5035 struct tcp_sock
*tp
= tcp_sk(sk
);
5037 /* Check if we get a new urgent pointer - normally not. */
5039 tcp_check_urg(sk
, th
);
5041 /* Do we wait for any urgent data? - normally not... */
5042 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5043 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5046 /* Is the urgent pointer pointing into this packet? */
5047 if (ptr
< skb
->len
) {
5049 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5051 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5052 if (!sock_flag(sk
, SOCK_DEAD
))
5053 sk
->sk_data_ready(sk
, 0);
5058 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5060 struct tcp_sock
*tp
= tcp_sk(sk
);
5061 int chunk
= skb
->len
- hlen
;
5065 if (skb_csum_unnecessary(skb
))
5066 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
5068 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
5072 tp
->ucopy
.len
-= chunk
;
5073 tp
->copied_seq
+= chunk
;
5074 tcp_rcv_space_adjust(sk
);
5081 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
5082 struct sk_buff
*skb
)
5086 if (sock_owned_by_user(sk
)) {
5088 result
= __tcp_checksum_complete(skb
);
5091 result
= __tcp_checksum_complete(skb
);
5096 static inline int tcp_checksum_complete_user(struct sock
*sk
,
5097 struct sk_buff
*skb
)
5099 return !skb_csum_unnecessary(skb
) &&
5100 __tcp_checksum_complete_user(sk
, skb
);
5103 #ifdef CONFIG_NET_DMA
5104 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5107 struct tcp_sock
*tp
= tcp_sk(sk
);
5108 int chunk
= skb
->len
- hlen
;
5110 int copied_early
= 0;
5112 if (tp
->ucopy
.wakeup
)
5115 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5116 tp
->ucopy
.dma_chan
= dma_find_channel(DMA_MEMCPY
);
5118 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5120 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5122 tp
->ucopy
.iov
, chunk
,
5123 tp
->ucopy
.pinned_list
);
5128 tp
->ucopy
.dma_cookie
= dma_cookie
;
5131 tp
->ucopy
.len
-= chunk
;
5132 tp
->copied_seq
+= chunk
;
5133 tcp_rcv_space_adjust(sk
);
5135 if ((tp
->ucopy
.len
== 0) ||
5136 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5137 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5138 tp
->ucopy
.wakeup
= 1;
5139 sk
->sk_data_ready(sk
, 0);
5141 } else if (chunk
> 0) {
5142 tp
->ucopy
.wakeup
= 1;
5143 sk
->sk_data_ready(sk
, 0);
5146 return copied_early
;
5148 #endif /* CONFIG_NET_DMA */
5150 /* Does PAWS and seqno based validation of an incoming segment, flags will
5151 * play significant role here.
5153 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5154 struct tcphdr
*th
, int syn_inerr
)
5157 struct tcp_sock
*tp
= tcp_sk(sk
);
5159 /* RFC1323: H1. Apply PAWS check first. */
5160 if (tcp_fast_parse_options(skb
, th
, tp
, &hash_location
) &&
5161 tp
->rx_opt
.saw_tstamp
&&
5162 tcp_paws_discard(sk
, skb
)) {
5164 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5165 tcp_send_dupack(sk
, skb
);
5168 /* Reset is accepted even if it did not pass PAWS. */
5171 /* Step 1: check sequence number */
5172 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5173 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5174 * (RST) segments are validated by checking their SEQ-fields."
5175 * And page 69: "If an incoming segment is not acceptable,
5176 * an acknowledgment should be sent in reply (unless the RST
5177 * bit is set, if so drop the segment and return)".
5180 tcp_send_dupack(sk
, skb
);
5184 /* Step 2: check RST bit */
5190 /* ts_recent update must be made after we are sure that the packet
5193 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5195 /* step 3: check security and precedence [ignored] */
5197 /* step 4: Check for a SYN in window. */
5198 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5200 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5201 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5214 * TCP receive function for the ESTABLISHED state.
5216 * It is split into a fast path and a slow path. The fast path is
5218 * - A zero window was announced from us - zero window probing
5219 * is only handled properly in the slow path.
5220 * - Out of order segments arrived.
5221 * - Urgent data is expected.
5222 * - There is no buffer space left
5223 * - Unexpected TCP flags/window values/header lengths are received
5224 * (detected by checking the TCP header against pred_flags)
5225 * - Data is sent in both directions. Fast path only supports pure senders
5226 * or pure receivers (this means either the sequence number or the ack
5227 * value must stay constant)
5228 * - Unexpected TCP option.
5230 * When these conditions are not satisfied it drops into a standard
5231 * receive procedure patterned after RFC793 to handle all cases.
5232 * The first three cases are guaranteed by proper pred_flags setting,
5233 * the rest is checked inline. Fast processing is turned on in
5234 * tcp_data_queue when everything is OK.
5236 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5237 struct tcphdr
*th
, unsigned len
)
5239 struct tcp_sock
*tp
= tcp_sk(sk
);
5243 * Header prediction.
5244 * The code loosely follows the one in the famous
5245 * "30 instruction TCP receive" Van Jacobson mail.
5247 * Van's trick is to deposit buffers into socket queue
5248 * on a device interrupt, to call tcp_recv function
5249 * on the receive process context and checksum and copy
5250 * the buffer to user space. smart...
5252 * Our current scheme is not silly either but we take the
5253 * extra cost of the net_bh soft interrupt processing...
5254 * We do checksum and copy also but from device to kernel.
5257 tp
->rx_opt
.saw_tstamp
= 0;
5259 /* pred_flags is 0xS?10 << 16 + snd_wnd
5260 * if header_prediction is to be made
5261 * 'S' will always be tp->tcp_header_len >> 2
5262 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5263 * turn it off (when there are holes in the receive
5264 * space for instance)
5265 * PSH flag is ignored.
5268 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5269 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5270 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5271 int tcp_header_len
= tp
->tcp_header_len
;
5273 /* Timestamp header prediction: tcp_header_len
5274 * is automatically equal to th->doff*4 due to pred_flags
5278 /* Check timestamp */
5279 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5280 /* No? Slow path! */
5281 if (!tcp_parse_aligned_timestamp(tp
, th
))
5284 /* If PAWS failed, check it more carefully in slow path */
5285 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5288 /* DO NOT update ts_recent here, if checksum fails
5289 * and timestamp was corrupted part, it will result
5290 * in a hung connection since we will drop all
5291 * future packets due to the PAWS test.
5295 if (len
<= tcp_header_len
) {
5296 /* Bulk data transfer: sender */
5297 if (len
== tcp_header_len
) {
5298 /* Predicted packet is in window by definition.
5299 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5300 * Hence, check seq<=rcv_wup reduces to:
5302 if (tcp_header_len
==
5303 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5304 tp
->rcv_nxt
== tp
->rcv_wup
)
5305 tcp_store_ts_recent(tp
);
5307 /* We know that such packets are checksummed
5310 tcp_ack(sk
, skb
, 0);
5312 tcp_data_snd_check(sk
);
5314 } else { /* Header too small */
5315 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5320 int copied_early
= 0;
5322 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5323 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5324 #ifdef CONFIG_NET_DMA
5325 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5330 if (tp
->ucopy
.task
== current
&&
5331 sock_owned_by_user(sk
) && !copied_early
) {
5332 __set_current_state(TASK_RUNNING
);
5334 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5338 /* Predicted packet is in window by definition.
5339 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5340 * Hence, check seq<=rcv_wup reduces to:
5342 if (tcp_header_len
==
5343 (sizeof(struct tcphdr
) +
5344 TCPOLEN_TSTAMP_ALIGNED
) &&
5345 tp
->rcv_nxt
== tp
->rcv_wup
)
5346 tcp_store_ts_recent(tp
);
5348 tcp_rcv_rtt_measure_ts(sk
, skb
);
5350 __skb_pull(skb
, tcp_header_len
);
5351 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5352 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5355 tcp_cleanup_rbuf(sk
, skb
->len
);
5358 if (tcp_checksum_complete_user(sk
, skb
))
5361 /* Predicted packet is in window by definition.
5362 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5363 * Hence, check seq<=rcv_wup reduces to:
5365 if (tcp_header_len
==
5366 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5367 tp
->rcv_nxt
== tp
->rcv_wup
)
5368 tcp_store_ts_recent(tp
);
5370 tcp_rcv_rtt_measure_ts(sk
, skb
);
5372 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5375 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5377 /* Bulk data transfer: receiver */
5378 __skb_pull(skb
, tcp_header_len
);
5379 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
5380 skb_set_owner_r(skb
, sk
);
5381 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5384 tcp_event_data_recv(sk
, skb
);
5386 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5387 /* Well, only one small jumplet in fast path... */
5388 tcp_ack(sk
, skb
, FLAG_DATA
);
5389 tcp_data_snd_check(sk
);
5390 if (!inet_csk_ack_scheduled(sk
))
5394 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5395 __tcp_ack_snd_check(sk
, 0);
5397 #ifdef CONFIG_NET_DMA
5399 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5405 sk
->sk_data_ready(sk
, 0);
5411 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5415 * Standard slow path.
5418 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
5423 if (th
->ack
&& tcp_ack(sk
, skb
, FLAG_SLOWPATH
) < 0)
5426 tcp_rcv_rtt_measure_ts(sk
, skb
);
5428 /* Process urgent data. */
5429 tcp_urg(sk
, skb
, th
);
5431 /* step 7: process the segment text */
5432 tcp_data_queue(sk
, skb
);
5434 tcp_data_snd_check(sk
);
5435 tcp_ack_snd_check(sk
);
5439 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5445 EXPORT_SYMBOL(tcp_rcv_established
);
5447 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5448 struct tcphdr
*th
, unsigned len
)
5451 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5452 struct tcp_sock
*tp
= tcp_sk(sk
);
5453 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
5454 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5456 tcp_parse_options(skb
, &tp
->rx_opt
, &hash_location
, 0);
5460 * "If the state is SYN-SENT then
5461 * first check the ACK bit
5462 * If the ACK bit is set
5463 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5464 * a reset (unless the RST bit is set, if so drop
5465 * the segment and return)"
5467 * We do not send data with SYN, so that RFC-correct
5470 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5471 goto reset_and_undo
;
5473 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5474 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5476 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5477 goto reset_and_undo
;
5480 /* Now ACK is acceptable.
5482 * "If the RST bit is set
5483 * If the ACK was acceptable then signal the user "error:
5484 * connection reset", drop the segment, enter CLOSED state,
5485 * delete TCB, and return."
5494 * "fifth, if neither of the SYN or RST bits is set then
5495 * drop the segment and return."
5501 goto discard_and_undo
;
5504 * "If the SYN bit is on ...
5505 * are acceptable then ...
5506 * (our SYN has been ACKed), change the connection
5507 * state to ESTABLISHED..."
5510 TCP_ECN_rcv_synack(tp
, th
);
5512 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5513 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5515 /* Ok.. it's good. Set up sequence numbers and
5516 * move to established.
5518 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5519 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5521 /* RFC1323: The window in SYN & SYN/ACK segments is
5524 tp
->snd_wnd
= ntohs(th
->window
);
5525 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5527 if (!tp
->rx_opt
.wscale_ok
) {
5528 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5529 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5532 if (tp
->rx_opt
.saw_tstamp
) {
5533 tp
->rx_opt
.tstamp_ok
= 1;
5534 tp
->tcp_header_len
=
5535 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5536 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5537 tcp_store_ts_recent(tp
);
5539 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5542 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5543 tcp_enable_fack(tp
);
5546 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5547 tcp_initialize_rcv_mss(sk
);
5549 /* Remember, tcp_poll() does not lock socket!
5550 * Change state from SYN-SENT only after copied_seq
5551 * is initialized. */
5552 tp
->copied_seq
= tp
->rcv_nxt
;
5555 cvp
->cookie_pair_size
> 0 &&
5556 tp
->rx_opt
.cookie_plus
> 0) {
5557 int cookie_size
= tp
->rx_opt
.cookie_plus
5558 - TCPOLEN_COOKIE_BASE
;
5559 int cookie_pair_size
= cookie_size
5560 + cvp
->cookie_desired
;
5562 /* A cookie extension option was sent and returned.
5563 * Note that each incoming SYNACK replaces the
5564 * Responder cookie. The initial exchange is most
5565 * fragile, as protection against spoofing relies
5566 * entirely upon the sequence and timestamp (above).
5567 * This replacement strategy allows the correct pair to
5568 * pass through, while any others will be filtered via
5569 * Responder verification later.
5571 if (sizeof(cvp
->cookie_pair
) >= cookie_pair_size
) {
5572 memcpy(&cvp
->cookie_pair
[cvp
->cookie_desired
],
5573 hash_location
, cookie_size
);
5574 cvp
->cookie_pair_size
= cookie_pair_size
;
5579 tcp_set_state(sk
, TCP_ESTABLISHED
);
5581 security_inet_conn_established(sk
, skb
);
5583 /* Make sure socket is routed, for correct metrics. */
5584 icsk
->icsk_af_ops
->rebuild_header(sk
);
5586 tcp_init_metrics(sk
);
5588 tcp_init_congestion_control(sk
);
5590 /* Prevent spurious tcp_cwnd_restart() on first data
5593 tp
->lsndtime
= tcp_time_stamp
;
5595 tcp_init_buffer_space(sk
);
5597 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5598 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5600 if (!tp
->rx_opt
.snd_wscale
)
5601 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5605 if (!sock_flag(sk
, SOCK_DEAD
)) {
5606 sk
->sk_state_change(sk
);
5607 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5610 if (sk
->sk_write_pending
||
5611 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5612 icsk
->icsk_ack
.pingpong
) {
5613 /* Save one ACK. Data will be ready after
5614 * several ticks, if write_pending is set.
5616 * It may be deleted, but with this feature tcpdumps
5617 * look so _wonderfully_ clever, that I was not able
5618 * to stand against the temptation 8) --ANK
5620 inet_csk_schedule_ack(sk
);
5621 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5622 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5623 tcp_incr_quickack(sk
);
5624 tcp_enter_quickack_mode(sk
);
5625 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5626 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5637 /* No ACK in the segment */
5641 * "If the RST bit is set
5643 * Otherwise (no ACK) drop the segment and return."
5646 goto discard_and_undo
;
5650 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5651 tcp_paws_reject(&tp
->rx_opt
, 0))
5652 goto discard_and_undo
;
5655 /* We see SYN without ACK. It is attempt of
5656 * simultaneous connect with crossed SYNs.
5657 * Particularly, it can be connect to self.
5659 tcp_set_state(sk
, TCP_SYN_RECV
);
5661 if (tp
->rx_opt
.saw_tstamp
) {
5662 tp
->rx_opt
.tstamp_ok
= 1;
5663 tcp_store_ts_recent(tp
);
5664 tp
->tcp_header_len
=
5665 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5667 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5670 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5671 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5673 /* RFC1323: The window in SYN & SYN/ACK segments is
5676 tp
->snd_wnd
= ntohs(th
->window
);
5677 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5678 tp
->max_window
= tp
->snd_wnd
;
5680 TCP_ECN_rcv_syn(tp
, th
);
5683 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5684 tcp_initialize_rcv_mss(sk
);
5686 tcp_send_synack(sk
);
5688 /* Note, we could accept data and URG from this segment.
5689 * There are no obstacles to make this.
5691 * However, if we ignore data in ACKless segments sometimes,
5692 * we have no reasons to accept it sometimes.
5693 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5694 * is not flawless. So, discard packet for sanity.
5695 * Uncomment this return to process the data.
5702 /* "fifth, if neither of the SYN or RST bits is set then
5703 * drop the segment and return."
5707 tcp_clear_options(&tp
->rx_opt
);
5708 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5712 tcp_clear_options(&tp
->rx_opt
);
5713 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5718 * This function implements the receiving procedure of RFC 793 for
5719 * all states except ESTABLISHED and TIME_WAIT.
5720 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5721 * address independent.
5724 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5725 struct tcphdr
*th
, unsigned len
)
5727 struct tcp_sock
*tp
= tcp_sk(sk
);
5728 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5732 tp
->rx_opt
.saw_tstamp
= 0;
5734 switch (sk
->sk_state
) {
5746 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5749 /* Now we have several options: In theory there is
5750 * nothing else in the frame. KA9Q has an option to
5751 * send data with the syn, BSD accepts data with the
5752 * syn up to the [to be] advertised window and
5753 * Solaris 2.1 gives you a protocol error. For now
5754 * we just ignore it, that fits the spec precisely
5755 * and avoids incompatibilities. It would be nice in
5756 * future to drop through and process the data.
5758 * Now that TTCP is starting to be used we ought to
5760 * But, this leaves one open to an easy denial of
5761 * service attack, and SYN cookies can't defend
5762 * against this problem. So, we drop the data
5763 * in the interest of security over speed unless
5764 * it's still in use.
5772 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5776 /* Do step6 onward by hand. */
5777 tcp_urg(sk
, skb
, th
);
5779 tcp_data_snd_check(sk
);
5783 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5787 /* step 5: check the ACK field */
5789 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
) > 0;
5791 switch (sk
->sk_state
) {
5794 tp
->copied_seq
= tp
->rcv_nxt
;
5796 tcp_set_state(sk
, TCP_ESTABLISHED
);
5797 sk
->sk_state_change(sk
);
5799 /* Note, that this wakeup is only for marginal
5800 * crossed SYN case. Passively open sockets
5801 * are not waked up, because sk->sk_sleep ==
5802 * NULL and sk->sk_socket == NULL.
5806 SOCK_WAKE_IO
, POLL_OUT
);
5808 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5809 tp
->snd_wnd
= ntohs(th
->window
) <<
5810 tp
->rx_opt
.snd_wscale
;
5811 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5813 /* tcp_ack considers this ACK as duplicate
5814 * and does not calculate rtt.
5817 tcp_ack_update_rtt(sk
, 0, 0);
5819 if (tp
->rx_opt
.tstamp_ok
)
5820 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5822 /* Make sure socket is routed, for
5825 icsk
->icsk_af_ops
->rebuild_header(sk
);
5827 tcp_init_metrics(sk
);
5829 tcp_init_congestion_control(sk
);
5831 /* Prevent spurious tcp_cwnd_restart() on
5832 * first data packet.
5834 tp
->lsndtime
= tcp_time_stamp
;
5837 tcp_initialize_rcv_mss(sk
);
5838 tcp_init_buffer_space(sk
);
5839 tcp_fast_path_on(tp
);
5846 if (tp
->snd_una
== tp
->write_seq
) {
5847 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5848 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5849 dst_confirm(__sk_dst_get(sk
));
5851 if (!sock_flag(sk
, SOCK_DEAD
))
5852 /* Wake up lingering close() */
5853 sk
->sk_state_change(sk
);
5857 if (tp
->linger2
< 0 ||
5858 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5859 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5861 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5865 tmo
= tcp_fin_time(sk
);
5866 if (tmo
> TCP_TIMEWAIT_LEN
) {
5867 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5868 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5869 /* Bad case. We could lose such FIN otherwise.
5870 * It is not a big problem, but it looks confusing
5871 * and not so rare event. We still can lose it now,
5872 * if it spins in bh_lock_sock(), but it is really
5875 inet_csk_reset_keepalive_timer(sk
, tmo
);
5877 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5885 if (tp
->snd_una
== tp
->write_seq
) {
5886 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5892 if (tp
->snd_una
== tp
->write_seq
) {
5893 tcp_update_metrics(sk
);
5902 /* step 6: check the URG bit */
5903 tcp_urg(sk
, skb
, th
);
5905 /* step 7: process the segment text */
5906 switch (sk
->sk_state
) {
5907 case TCP_CLOSE_WAIT
:
5910 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5914 /* RFC 793 says to queue data in these states,
5915 * RFC 1122 says we MUST send a reset.
5916 * BSD 4.4 also does reset.
5918 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5919 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5920 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5921 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5927 case TCP_ESTABLISHED
:
5928 tcp_data_queue(sk
, skb
);
5933 /* tcp_data could move socket to TIME-WAIT */
5934 if (sk
->sk_state
!= TCP_CLOSE
) {
5935 tcp_data_snd_check(sk
);
5936 tcp_ack_snd_check(sk
);
5945 EXPORT_SYMBOL(tcp_rcv_state_process
);