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 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
810 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
813 cwnd
= rfc3390_bytes_to_packets(tp
->mss_cache
);
814 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
817 /* Set slow start threshold and cwnd not falling to slow start */
818 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
820 struct tcp_sock
*tp
= tcp_sk(sk
);
821 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
823 tp
->prior_ssthresh
= 0;
825 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
828 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
829 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
830 tcp_packets_in_flight(tp
) + 1U);
831 tp
->snd_cwnd_cnt
= 0;
832 tp
->high_seq
= tp
->snd_nxt
;
833 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
834 TCP_ECN_queue_cwr(tp
);
836 tcp_set_ca_state(sk
, TCP_CA_CWR
);
841 * Packet counting of FACK is based on in-order assumptions, therefore TCP
842 * disables it when reordering is detected
844 static void tcp_disable_fack(struct tcp_sock
*tp
)
846 /* RFC3517 uses different metric in lost marker => reset on change */
848 tp
->lost_skb_hint
= NULL
;
849 tp
->rx_opt
.sack_ok
&= ~2;
852 /* Take a notice that peer is sending D-SACKs */
853 static void tcp_dsack_seen(struct tcp_sock
*tp
)
855 tp
->rx_opt
.sack_ok
|= 4;
858 /* Initialize metrics on socket. */
860 static void tcp_init_metrics(struct sock
*sk
)
862 struct tcp_sock
*tp
= tcp_sk(sk
);
863 struct dst_entry
*dst
= __sk_dst_get(sk
);
870 if (dst_metric_locked(dst
, RTAX_CWND
))
871 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
872 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
873 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
874 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
875 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
877 if (dst_metric(dst
, RTAX_REORDERING
) &&
878 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
879 tcp_disable_fack(tp
);
880 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
883 if (dst_metric(dst
, RTAX_RTT
) == 0)
886 if (!tp
->srtt
&& dst_metric_rtt(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
889 /* Initial rtt is determined from SYN,SYN-ACK.
890 * The segment is small and rtt may appear much
891 * less than real one. Use per-dst memory
892 * to make it more realistic.
894 * A bit of theory. RTT is time passed after "normal" sized packet
895 * is sent until it is ACKed. In normal circumstances sending small
896 * packets force peer to delay ACKs and calculation is correct too.
897 * The algorithm is adaptive and, provided we follow specs, it
898 * NEVER underestimate RTT. BUT! If peer tries to make some clever
899 * tricks sort of "quick acks" for time long enough to decrease RTT
900 * to low value, and then abruptly stops to do it and starts to delay
901 * ACKs, wait for troubles.
903 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
904 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
905 tp
->rtt_seq
= tp
->snd_nxt
;
907 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
908 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
909 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
912 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
916 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
917 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
921 /* Play conservative. If timestamps are not
922 * supported, TCP will fail to recalculate correct
923 * rtt, if initial rto is too small. FORGET ALL AND RESET!
925 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
927 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
928 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
933 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
936 struct tcp_sock
*tp
= tcp_sk(sk
);
937 if (metric
> tp
->reordering
) {
940 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
942 /* This exciting event is worth to be remembered. 8) */
944 mib_idx
= LINUX_MIB_TCPTSREORDER
;
945 else if (tcp_is_reno(tp
))
946 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
947 else if (tcp_is_fack(tp
))
948 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
950 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
952 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
953 #if FASTRETRANS_DEBUG > 1
954 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
955 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
959 tp
->undo_marker
? tp
->undo_retrans
: 0);
961 tcp_disable_fack(tp
);
965 /* This must be called before lost_out is incremented */
966 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
968 if ((tp
->retransmit_skb_hint
== NULL
) ||
969 before(TCP_SKB_CB(skb
)->seq
,
970 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
971 tp
->retransmit_skb_hint
= skb
;
974 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
975 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
978 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
980 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
981 tcp_verify_retransmit_hint(tp
, skb
);
983 tp
->lost_out
+= tcp_skb_pcount(skb
);
984 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
988 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
991 tcp_verify_retransmit_hint(tp
, skb
);
993 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
994 tp
->lost_out
+= tcp_skb_pcount(skb
);
995 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
999 /* This procedure tags the retransmission queue when SACKs arrive.
1001 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1002 * Packets in queue with these bits set are counted in variables
1003 * sacked_out, retrans_out and lost_out, correspondingly.
1005 * Valid combinations are:
1006 * Tag InFlight Description
1007 * 0 1 - orig segment is in flight.
1008 * S 0 - nothing flies, orig reached receiver.
1009 * L 0 - nothing flies, orig lost by net.
1010 * R 2 - both orig and retransmit are in flight.
1011 * L|R 1 - orig is lost, retransmit is in flight.
1012 * S|R 1 - orig reached receiver, retrans is still in flight.
1013 * (L|S|R is logically valid, it could occur when L|R is sacked,
1014 * but it is equivalent to plain S and code short-curcuits it to S.
1015 * L|S is logically invalid, it would mean -1 packet in flight 8))
1017 * These 6 states form finite state machine, controlled by the following events:
1018 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1019 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1020 * 3. Loss detection event of one of three flavors:
1021 * A. Scoreboard estimator decided the packet is lost.
1022 * A'. Reno "three dupacks" marks head of queue lost.
1023 * A''. Its FACK modfication, head until snd.fack is lost.
1024 * B. SACK arrives sacking data transmitted after never retransmitted
1025 * hole was sent out.
1026 * C. SACK arrives sacking SND.NXT at the moment, when the
1027 * segment was retransmitted.
1028 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1030 * It is pleasant to note, that state diagram turns out to be commutative,
1031 * so that we are allowed not to be bothered by order of our actions,
1032 * when multiple events arrive simultaneously. (see the function below).
1034 * Reordering detection.
1035 * --------------------
1036 * Reordering metric is maximal distance, which a packet can be displaced
1037 * in packet stream. With SACKs we can estimate it:
1039 * 1. SACK fills old hole and the corresponding segment was not
1040 * ever retransmitted -> reordering. Alas, we cannot use it
1041 * when segment was retransmitted.
1042 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1043 * for retransmitted and already SACKed segment -> reordering..
1044 * Both of these heuristics are not used in Loss state, when we cannot
1045 * account for retransmits accurately.
1047 * SACK block validation.
1048 * ----------------------
1050 * SACK block range validation checks that the received SACK block fits to
1051 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1052 * Note that SND.UNA is not included to the range though being valid because
1053 * it means that the receiver is rather inconsistent with itself reporting
1054 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1055 * perfectly valid, however, in light of RFC2018 which explicitly states
1056 * that "SACK block MUST reflect the newest segment. Even if the newest
1057 * segment is going to be discarded ...", not that it looks very clever
1058 * in case of head skb. Due to potentional receiver driven attacks, we
1059 * choose to avoid immediate execution of a walk in write queue due to
1060 * reneging and defer head skb's loss recovery to standard loss recovery
1061 * procedure that will eventually trigger (nothing forbids us doing this).
1063 * Implements also blockage to start_seq wrap-around. Problem lies in the
1064 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1065 * there's no guarantee that it will be before snd_nxt (n). The problem
1066 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1069 * <- outs wnd -> <- wrapzone ->
1070 * u e n u_w e_w s n_w
1072 * |<------------+------+----- TCP seqno space --------------+---------->|
1073 * ...-- <2^31 ->| |<--------...
1074 * ...---- >2^31 ------>| |<--------...
1076 * Current code wouldn't be vulnerable but it's better still to discard such
1077 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1078 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1079 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1080 * equal to the ideal case (infinite seqno space without wrap caused issues).
1082 * With D-SACK the lower bound is extended to cover sequence space below
1083 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1084 * again, D-SACK block must not to go across snd_una (for the same reason as
1085 * for the normal SACK blocks, explained above). But there all simplicity
1086 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1087 * fully below undo_marker they do not affect behavior in anyway and can
1088 * therefore be safely ignored. In rare cases (which are more or less
1089 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1090 * fragmentation and packet reordering past skb's retransmission. To consider
1091 * them correctly, the acceptable range must be extended even more though
1092 * the exact amount is rather hard to quantify. However, tp->max_window can
1093 * be used as an exaggerated estimate.
1095 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1096 u32 start_seq
, u32 end_seq
)
1098 /* Too far in future, or reversed (interpretation is ambiguous) */
1099 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1102 /* Nasty start_seq wrap-around check (see comments above) */
1103 if (!before(start_seq
, tp
->snd_nxt
))
1106 /* In outstanding window? ...This is valid exit for D-SACKs too.
1107 * start_seq == snd_una is non-sensical (see comments above)
1109 if (after(start_seq
, tp
->snd_una
))
1112 if (!is_dsack
|| !tp
->undo_marker
)
1115 /* ...Then it's D-SACK, and must reside below snd_una completely */
1116 if (!after(end_seq
, tp
->snd_una
))
1119 if (!before(start_seq
, tp
->undo_marker
))
1123 if (!after(end_seq
, tp
->undo_marker
))
1126 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1127 * start_seq < undo_marker and end_seq >= undo_marker.
1129 return !before(start_seq
, end_seq
- tp
->max_window
);
1132 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1133 * Event "C". Later note: FACK people cheated me again 8), we have to account
1134 * for reordering! Ugly, but should help.
1136 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1137 * less than what is now known to be received by the other end (derived from
1138 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1139 * retransmitted skbs to avoid some costly processing per ACKs.
1141 static void tcp_mark_lost_retrans(struct sock
*sk
)
1143 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1144 struct tcp_sock
*tp
= tcp_sk(sk
);
1145 struct sk_buff
*skb
;
1147 u32 new_low_seq
= tp
->snd_nxt
;
1148 u32 received_upto
= tcp_highest_sack_seq(tp
);
1150 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1151 !after(received_upto
, tp
->lost_retrans_low
) ||
1152 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1155 tcp_for_write_queue(skb
, sk
) {
1156 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1158 if (skb
== tcp_send_head(sk
))
1160 if (cnt
== tp
->retrans_out
)
1162 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1165 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1168 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1169 * constraint here (see above) but figuring out that at
1170 * least tp->reordering SACK blocks reside between ack_seq
1171 * and received_upto is not easy task to do cheaply with
1172 * the available datastructures.
1174 * Whether FACK should check here for tp->reordering segs
1175 * in-between one could argue for either way (it would be
1176 * rather simple to implement as we could count fack_count
1177 * during the walk and do tp->fackets_out - fack_count).
1179 if (after(received_upto
, ack_seq
)) {
1180 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1181 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1183 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1184 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1186 if (before(ack_seq
, new_low_seq
))
1187 new_low_seq
= ack_seq
;
1188 cnt
+= tcp_skb_pcount(skb
);
1192 if (tp
->retrans_out
)
1193 tp
->lost_retrans_low
= new_low_seq
;
1196 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1197 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1200 struct tcp_sock
*tp
= tcp_sk(sk
);
1201 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1202 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1205 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1208 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1209 } else if (num_sacks
> 1) {
1210 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1211 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1213 if (!after(end_seq_0
, end_seq_1
) &&
1214 !before(start_seq_0
, start_seq_1
)) {
1217 NET_INC_STATS_BH(sock_net(sk
),
1218 LINUX_MIB_TCPDSACKOFORECV
);
1222 /* D-SACK for already forgotten data... Do dumb counting. */
1224 !after(end_seq_0
, prior_snd_una
) &&
1225 after(end_seq_0
, tp
->undo_marker
))
1231 struct tcp_sacktag_state
{
1237 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1238 * the incoming SACK may not exactly match but we can find smaller MSS
1239 * aligned portion of it that matches. Therefore we might need to fragment
1240 * which may fail and creates some hassle (caller must handle error case
1243 * FIXME: this could be merged to shift decision code
1245 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1246 u32 start_seq
, u32 end_seq
)
1249 unsigned int pkt_len
;
1252 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1253 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1255 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1256 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1257 mss
= tcp_skb_mss(skb
);
1258 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1261 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1265 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1270 /* Round if necessary so that SACKs cover only full MSSes
1271 * and/or the remaining small portion (if present)
1273 if (pkt_len
> mss
) {
1274 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1275 if (!in_sack
&& new_len
< pkt_len
) {
1277 if (new_len
> skb
->len
)
1282 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1290 static u8
tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1291 struct tcp_sacktag_state
*state
,
1292 int dup_sack
, int pcount
)
1294 struct tcp_sock
*tp
= tcp_sk(sk
);
1295 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1296 int fack_count
= state
->fack_count
;
1298 /* Account D-SACK for retransmitted packet. */
1299 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1300 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1302 if (sacked
& TCPCB_SACKED_ACKED
)
1303 state
->reord
= min(fack_count
, state
->reord
);
1306 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1307 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1310 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1311 if (sacked
& TCPCB_SACKED_RETRANS
) {
1312 /* If the segment is not tagged as lost,
1313 * we do not clear RETRANS, believing
1314 * that retransmission is still in flight.
1316 if (sacked
& TCPCB_LOST
) {
1317 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1318 tp
->lost_out
-= pcount
;
1319 tp
->retrans_out
-= pcount
;
1322 if (!(sacked
& TCPCB_RETRANS
)) {
1323 /* New sack for not retransmitted frame,
1324 * which was in hole. It is reordering.
1326 if (before(TCP_SKB_CB(skb
)->seq
,
1327 tcp_highest_sack_seq(tp
)))
1328 state
->reord
= min(fack_count
,
1331 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1332 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1333 state
->flag
|= FLAG_ONLY_ORIG_SACKED
;
1336 if (sacked
& TCPCB_LOST
) {
1337 sacked
&= ~TCPCB_LOST
;
1338 tp
->lost_out
-= pcount
;
1342 sacked
|= TCPCB_SACKED_ACKED
;
1343 state
->flag
|= FLAG_DATA_SACKED
;
1344 tp
->sacked_out
+= pcount
;
1346 fack_count
+= pcount
;
1348 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1349 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1350 before(TCP_SKB_CB(skb
)->seq
,
1351 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1352 tp
->lost_cnt_hint
+= pcount
;
1354 if (fack_count
> tp
->fackets_out
)
1355 tp
->fackets_out
= fack_count
;
1358 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1359 * frames and clear it. undo_retrans is decreased above, L|R frames
1360 * are accounted above as well.
1362 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1363 sacked
&= ~TCPCB_SACKED_RETRANS
;
1364 tp
->retrans_out
-= pcount
;
1370 static int tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1371 struct tcp_sacktag_state
*state
,
1372 unsigned int pcount
, int shifted
, int mss
,
1375 struct tcp_sock
*tp
= tcp_sk(sk
);
1376 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1380 /* Tweak before seqno plays */
1381 if (!tcp_is_fack(tp
) && tcp_is_sack(tp
) && tp
->lost_skb_hint
&&
1382 !before(TCP_SKB_CB(tp
->lost_skb_hint
)->seq
, TCP_SKB_CB(skb
)->seq
))
1383 tp
->lost_cnt_hint
+= pcount
;
1385 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1386 TCP_SKB_CB(skb
)->seq
+= shifted
;
1388 skb_shinfo(prev
)->gso_segs
+= pcount
;
1389 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1390 skb_shinfo(skb
)->gso_segs
-= pcount
;
1392 /* When we're adding to gso_segs == 1, gso_size will be zero,
1393 * in theory this shouldn't be necessary but as long as DSACK
1394 * code can come after this skb later on it's better to keep
1395 * setting gso_size to something.
1397 if (!skb_shinfo(prev
)->gso_size
) {
1398 skb_shinfo(prev
)->gso_size
= mss
;
1399 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1402 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1403 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1404 skb_shinfo(skb
)->gso_size
= 0;
1405 skb_shinfo(skb
)->gso_type
= 0;
1408 /* We discard results */
1409 tcp_sacktag_one(skb
, sk
, state
, dup_sack
, pcount
);
1411 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1412 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1415 BUG_ON(!tcp_skb_pcount(skb
));
1416 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1420 /* Whole SKB was eaten :-) */
1422 if (skb
== tp
->retransmit_skb_hint
)
1423 tp
->retransmit_skb_hint
= prev
;
1424 if (skb
== tp
->scoreboard_skb_hint
)
1425 tp
->scoreboard_skb_hint
= prev
;
1426 if (skb
== tp
->lost_skb_hint
) {
1427 tp
->lost_skb_hint
= prev
;
1428 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1431 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(prev
)->flags
;
1432 if (skb
== tcp_highest_sack(sk
))
1433 tcp_advance_highest_sack(sk
, skb
);
1435 tcp_unlink_write_queue(skb
, sk
);
1436 sk_wmem_free_skb(sk
, skb
);
1438 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1443 /* I wish gso_size would have a bit more sane initialization than
1444 * something-or-zero which complicates things
1446 static int tcp_skb_seglen(struct sk_buff
*skb
)
1448 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1451 /* Shifting pages past head area doesn't work */
1452 static int skb_can_shift(struct sk_buff
*skb
)
1454 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1457 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1460 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1461 struct tcp_sacktag_state
*state
,
1462 u32 start_seq
, u32 end_seq
,
1465 struct tcp_sock
*tp
= tcp_sk(sk
);
1466 struct sk_buff
*prev
;
1472 if (!sk_can_gso(sk
))
1475 /* Normally R but no L won't result in plain S */
1477 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1479 if (!skb_can_shift(skb
))
1481 /* This frame is about to be dropped (was ACKed). */
1482 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1485 /* Can only happen with delayed DSACK + discard craziness */
1486 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1488 prev
= tcp_write_queue_prev(sk
, skb
);
1490 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1493 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1494 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1498 pcount
= tcp_skb_pcount(skb
);
1499 mss
= tcp_skb_seglen(skb
);
1501 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1502 * drop this restriction as unnecessary
1504 if (mss
!= tcp_skb_seglen(prev
))
1507 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1509 /* CHECKME: This is non-MSS split case only?, this will
1510 * cause skipped skbs due to advancing loop btw, original
1511 * has that feature too
1513 if (tcp_skb_pcount(skb
) <= 1)
1516 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1518 /* TODO: head merge to next could be attempted here
1519 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1520 * though it might not be worth of the additional hassle
1522 * ...we can probably just fallback to what was done
1523 * previously. We could try merging non-SACKed ones
1524 * as well but it probably isn't going to buy off
1525 * because later SACKs might again split them, and
1526 * it would make skb timestamp tracking considerably
1532 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1534 BUG_ON(len
> skb
->len
);
1536 /* MSS boundaries should be honoured or else pcount will
1537 * severely break even though it makes things bit trickier.
1538 * Optimize common case to avoid most of the divides
1540 mss
= tcp_skb_mss(skb
);
1542 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1543 * drop this restriction as unnecessary
1545 if (mss
!= tcp_skb_seglen(prev
))
1550 } else if (len
< mss
) {
1558 if (!skb_shift(prev
, skb
, len
))
1560 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1563 /* Hole filled allows collapsing with the next as well, this is very
1564 * useful when hole on every nth skb pattern happens
1566 if (prev
== tcp_write_queue_tail(sk
))
1568 skb
= tcp_write_queue_next(sk
, prev
);
1570 if (!skb_can_shift(skb
) ||
1571 (skb
== tcp_send_head(sk
)) ||
1572 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1573 (mss
!= tcp_skb_seglen(skb
)))
1577 if (skb_shift(prev
, skb
, len
)) {
1578 pcount
+= tcp_skb_pcount(skb
);
1579 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1583 state
->fack_count
+= pcount
;
1590 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1594 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1595 struct tcp_sack_block
*next_dup
,
1596 struct tcp_sacktag_state
*state
,
1597 u32 start_seq
, u32 end_seq
,
1600 struct tcp_sock
*tp
= tcp_sk(sk
);
1601 struct sk_buff
*tmp
;
1603 tcp_for_write_queue_from(skb
, sk
) {
1605 int dup_sack
= dup_sack_in
;
1607 if (skb
== tcp_send_head(sk
))
1610 /* queue is in-order => we can short-circuit the walk early */
1611 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1614 if ((next_dup
!= NULL
) &&
1615 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1616 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1617 next_dup
->start_seq
,
1623 /* skb reference here is a bit tricky to get right, since
1624 * shifting can eat and free both this skb and the next,
1625 * so not even _safe variant of the loop is enough.
1628 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1629 start_seq
, end_seq
, dup_sack
);
1638 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1644 if (unlikely(in_sack
< 0))
1648 TCP_SKB_CB(skb
)->sacked
= tcp_sacktag_one(skb
, sk
,
1651 tcp_skb_pcount(skb
));
1653 if (!before(TCP_SKB_CB(skb
)->seq
,
1654 tcp_highest_sack_seq(tp
)))
1655 tcp_advance_highest_sack(sk
, skb
);
1658 state
->fack_count
+= tcp_skb_pcount(skb
);
1663 /* Avoid all extra work that is being done by sacktag while walking in
1666 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1667 struct tcp_sacktag_state
*state
,
1670 tcp_for_write_queue_from(skb
, sk
) {
1671 if (skb
== tcp_send_head(sk
))
1674 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1677 state
->fack_count
+= tcp_skb_pcount(skb
);
1682 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1684 struct tcp_sack_block
*next_dup
,
1685 struct tcp_sacktag_state
*state
,
1688 if (next_dup
== NULL
)
1691 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1692 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1693 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1694 next_dup
->start_seq
, next_dup
->end_seq
,
1701 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1703 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1707 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1710 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1711 struct tcp_sock
*tp
= tcp_sk(sk
);
1712 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1713 TCP_SKB_CB(ack_skb
)->sacked
);
1714 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1715 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1716 struct tcp_sack_block
*cache
;
1717 struct tcp_sacktag_state state
;
1718 struct sk_buff
*skb
;
1719 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1721 int found_dup_sack
= 0;
1723 int first_sack_index
;
1726 state
.reord
= tp
->packets_out
;
1728 if (!tp
->sacked_out
) {
1729 if (WARN_ON(tp
->fackets_out
))
1730 tp
->fackets_out
= 0;
1731 tcp_highest_sack_reset(sk
);
1734 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1735 num_sacks
, prior_snd_una
);
1737 state
.flag
|= FLAG_DSACKING_ACK
;
1739 /* Eliminate too old ACKs, but take into
1740 * account more or less fresh ones, they can
1741 * contain valid SACK info.
1743 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1746 if (!tp
->packets_out
)
1750 first_sack_index
= 0;
1751 for (i
= 0; i
< num_sacks
; i
++) {
1752 int dup_sack
= !i
&& found_dup_sack
;
1754 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1755 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1757 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1758 sp
[used_sacks
].start_seq
,
1759 sp
[used_sacks
].end_seq
)) {
1763 if (!tp
->undo_marker
)
1764 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1766 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1768 /* Don't count olds caused by ACK reordering */
1769 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1770 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1772 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1775 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1777 first_sack_index
= -1;
1781 /* Ignore very old stuff early */
1782 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1788 /* order SACK blocks to allow in order walk of the retrans queue */
1789 for (i
= used_sacks
- 1; i
> 0; i
--) {
1790 for (j
= 0; j
< i
; j
++) {
1791 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1792 swap(sp
[j
], sp
[j
+ 1]);
1794 /* Track where the first SACK block goes to */
1795 if (j
== first_sack_index
)
1796 first_sack_index
= j
+ 1;
1801 skb
= tcp_write_queue_head(sk
);
1802 state
.fack_count
= 0;
1805 if (!tp
->sacked_out
) {
1806 /* It's already past, so skip checking against it */
1807 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1809 cache
= tp
->recv_sack_cache
;
1810 /* Skip empty blocks in at head of the cache */
1811 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1816 while (i
< used_sacks
) {
1817 u32 start_seq
= sp
[i
].start_seq
;
1818 u32 end_seq
= sp
[i
].end_seq
;
1819 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1820 struct tcp_sack_block
*next_dup
= NULL
;
1822 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1823 next_dup
= &sp
[i
+ 1];
1825 /* Event "B" in the comment above. */
1826 if (after(end_seq
, tp
->high_seq
))
1827 state
.flag
|= FLAG_DATA_LOST
;
1829 /* Skip too early cached blocks */
1830 while (tcp_sack_cache_ok(tp
, cache
) &&
1831 !before(start_seq
, cache
->end_seq
))
1834 /* Can skip some work by looking recv_sack_cache? */
1835 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1836 after(end_seq
, cache
->start_seq
)) {
1839 if (before(start_seq
, cache
->start_seq
)) {
1840 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1842 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1849 /* Rest of the block already fully processed? */
1850 if (!after(end_seq
, cache
->end_seq
))
1853 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1857 /* ...tail remains todo... */
1858 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1859 /* ...but better entrypoint exists! */
1860 skb
= tcp_highest_sack(sk
);
1863 state
.fack_count
= tp
->fackets_out
;
1868 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1869 /* Check overlap against next cached too (past this one already) */
1874 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1875 skb
= tcp_highest_sack(sk
);
1878 state
.fack_count
= tp
->fackets_out
;
1880 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1883 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1884 start_seq
, end_seq
, dup_sack
);
1887 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1888 * due to in-order walk
1890 if (after(end_seq
, tp
->frto_highmark
))
1891 state
.flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1896 /* Clear the head of the cache sack blocks so we can skip it next time */
1897 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1898 tp
->recv_sack_cache
[i
].start_seq
= 0;
1899 tp
->recv_sack_cache
[i
].end_seq
= 0;
1901 for (j
= 0; j
< used_sacks
; j
++)
1902 tp
->recv_sack_cache
[i
++] = sp
[j
];
1904 tcp_mark_lost_retrans(sk
);
1906 tcp_verify_left_out(tp
);
1908 if ((state
.reord
< tp
->fackets_out
) &&
1909 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1910 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1911 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1915 #if FASTRETRANS_DEBUG > 0
1916 WARN_ON((int)tp
->sacked_out
< 0);
1917 WARN_ON((int)tp
->lost_out
< 0);
1918 WARN_ON((int)tp
->retrans_out
< 0);
1919 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1924 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1925 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1927 static int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1931 holes
= max(tp
->lost_out
, 1U);
1932 holes
= min(holes
, tp
->packets_out
);
1934 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1935 tp
->sacked_out
= tp
->packets_out
- holes
;
1941 /* If we receive more dupacks than we expected counting segments
1942 * in assumption of absent reordering, interpret this as reordering.
1943 * The only another reason could be bug in receiver TCP.
1945 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1947 struct tcp_sock
*tp
= tcp_sk(sk
);
1948 if (tcp_limit_reno_sacked(tp
))
1949 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1952 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1954 static void tcp_add_reno_sack(struct sock
*sk
)
1956 struct tcp_sock
*tp
= tcp_sk(sk
);
1958 tcp_check_reno_reordering(sk
, 0);
1959 tcp_verify_left_out(tp
);
1962 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1964 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1966 struct tcp_sock
*tp
= tcp_sk(sk
);
1969 /* One ACK acked hole. The rest eat duplicate ACKs. */
1970 if (acked
- 1 >= tp
->sacked_out
)
1973 tp
->sacked_out
-= acked
- 1;
1975 tcp_check_reno_reordering(sk
, acked
);
1976 tcp_verify_left_out(tp
);
1979 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1984 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
1986 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
1989 /* F-RTO can only be used if TCP has never retransmitted anything other than
1990 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1992 int tcp_use_frto(struct sock
*sk
)
1994 const struct tcp_sock
*tp
= tcp_sk(sk
);
1995 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1996 struct sk_buff
*skb
;
1998 if (!sysctl_tcp_frto
)
2001 /* MTU probe and F-RTO won't really play nicely along currently */
2002 if (icsk
->icsk_mtup
.probe_size
)
2005 if (tcp_is_sackfrto(tp
))
2008 /* Avoid expensive walking of rexmit queue if possible */
2009 if (tp
->retrans_out
> 1)
2012 skb
= tcp_write_queue_head(sk
);
2013 if (tcp_skb_is_last(sk
, skb
))
2015 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
2016 tcp_for_write_queue_from(skb
, sk
) {
2017 if (skb
== tcp_send_head(sk
))
2019 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2021 /* Short-circuit when first non-SACKed skb has been checked */
2022 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2028 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2029 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2030 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2031 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2032 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2033 * bits are handled if the Loss state is really to be entered (in
2034 * tcp_enter_frto_loss).
2036 * Do like tcp_enter_loss() would; when RTO expires the second time it
2038 * "Reduce ssthresh if it has not yet been made inside this window."
2040 void tcp_enter_frto(struct sock
*sk
)
2042 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2043 struct tcp_sock
*tp
= tcp_sk(sk
);
2044 struct sk_buff
*skb
;
2046 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
2047 tp
->snd_una
== tp
->high_seq
||
2048 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
2049 !icsk
->icsk_retransmits
)) {
2050 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2051 /* Our state is too optimistic in ssthresh() call because cwnd
2052 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2053 * recovery has not yet completed. Pattern would be this: RTO,
2054 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2056 * RFC4138 should be more specific on what to do, even though
2057 * RTO is quite unlikely to occur after the first Cumulative ACK
2058 * due to back-off and complexity of triggering events ...
2060 if (tp
->frto_counter
) {
2062 stored_cwnd
= tp
->snd_cwnd
;
2064 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2065 tp
->snd_cwnd
= stored_cwnd
;
2067 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2069 /* ... in theory, cong.control module could do "any tricks" in
2070 * ssthresh(), which means that ca_state, lost bits and lost_out
2071 * counter would have to be faked before the call occurs. We
2072 * consider that too expensive, unlikely and hacky, so modules
2073 * using these in ssthresh() must deal these incompatibility
2074 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2076 tcp_ca_event(sk
, CA_EVENT_FRTO
);
2079 tp
->undo_marker
= tp
->snd_una
;
2080 tp
->undo_retrans
= 0;
2082 skb
= tcp_write_queue_head(sk
);
2083 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2084 tp
->undo_marker
= 0;
2085 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2086 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2087 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2089 tcp_verify_left_out(tp
);
2091 /* Too bad if TCP was application limited */
2092 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2094 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2095 * The last condition is necessary at least in tp->frto_counter case.
2097 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
2098 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
2099 after(tp
->high_seq
, tp
->snd_una
)) {
2100 tp
->frto_highmark
= tp
->high_seq
;
2102 tp
->frto_highmark
= tp
->snd_nxt
;
2104 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
2105 tp
->high_seq
= tp
->snd_nxt
;
2106 tp
->frto_counter
= 1;
2109 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2110 * which indicates that we should follow the traditional RTO recovery,
2111 * i.e. mark everything lost and do go-back-N retransmission.
2113 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
2115 struct tcp_sock
*tp
= tcp_sk(sk
);
2116 struct sk_buff
*skb
;
2119 tp
->retrans_out
= 0;
2120 if (tcp_is_reno(tp
))
2121 tcp_reset_reno_sack(tp
);
2123 tcp_for_write_queue(skb
, sk
) {
2124 if (skb
== tcp_send_head(sk
))
2127 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2129 * Count the retransmission made on RTO correctly (only when
2130 * waiting for the first ACK and did not get it)...
2132 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
2133 /* For some reason this R-bit might get cleared? */
2134 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
2135 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2136 /* ...enter this if branch just for the first segment */
2137 flag
|= FLAG_DATA_ACKED
;
2139 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2140 tp
->undo_marker
= 0;
2141 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2144 /* Marking forward transmissions that were made after RTO lost
2145 * can cause unnecessary retransmissions in some scenarios,
2146 * SACK blocks will mitigate that in some but not in all cases.
2147 * We used to not mark them but it was causing break-ups with
2148 * receivers that do only in-order receival.
2150 * TODO: we could detect presence of such receiver and select
2151 * different behavior per flow.
2153 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2154 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2155 tp
->lost_out
+= tcp_skb_pcount(skb
);
2156 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2159 tcp_verify_left_out(tp
);
2161 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2162 tp
->snd_cwnd_cnt
= 0;
2163 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2164 tp
->frto_counter
= 0;
2165 tp
->bytes_acked
= 0;
2167 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2168 sysctl_tcp_reordering
);
2169 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2170 tp
->high_seq
= tp
->snd_nxt
;
2171 TCP_ECN_queue_cwr(tp
);
2173 tcp_clear_all_retrans_hints(tp
);
2176 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2178 tp
->retrans_out
= 0;
2181 tp
->undo_marker
= 0;
2182 tp
->undo_retrans
= 0;
2185 void tcp_clear_retrans(struct tcp_sock
*tp
)
2187 tcp_clear_retrans_partial(tp
);
2189 tp
->fackets_out
= 0;
2193 /* Enter Loss state. If "how" is not zero, forget all SACK information
2194 * and reset tags completely, otherwise preserve SACKs. If receiver
2195 * dropped its ofo queue, we will know this due to reneging detection.
2197 void tcp_enter_loss(struct sock
*sk
, int how
)
2199 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2200 struct tcp_sock
*tp
= tcp_sk(sk
);
2201 struct sk_buff
*skb
;
2203 /* Reduce ssthresh if it has not yet been made inside this window. */
2204 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2205 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2206 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2207 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2208 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2211 tp
->snd_cwnd_cnt
= 0;
2212 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2214 tp
->bytes_acked
= 0;
2215 tcp_clear_retrans_partial(tp
);
2217 if (tcp_is_reno(tp
))
2218 tcp_reset_reno_sack(tp
);
2221 /* Push undo marker, if it was plain RTO and nothing
2222 * was retransmitted. */
2223 tp
->undo_marker
= tp
->snd_una
;
2226 tp
->fackets_out
= 0;
2228 tcp_clear_all_retrans_hints(tp
);
2230 tcp_for_write_queue(skb
, sk
) {
2231 if (skb
== tcp_send_head(sk
))
2234 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2235 tp
->undo_marker
= 0;
2236 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2237 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2238 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2239 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2240 tp
->lost_out
+= tcp_skb_pcount(skb
);
2241 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2244 tcp_verify_left_out(tp
);
2246 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2247 sysctl_tcp_reordering
);
2248 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2249 tp
->high_seq
= tp
->snd_nxt
;
2250 TCP_ECN_queue_cwr(tp
);
2251 /* Abort F-RTO algorithm if one is in progress */
2252 tp
->frto_counter
= 0;
2255 /* If ACK arrived pointing to a remembered SACK, it means that our
2256 * remembered SACKs do not reflect real state of receiver i.e.
2257 * receiver _host_ is heavily congested (or buggy).
2259 * Do processing similar to RTO timeout.
2261 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2263 if (flag
& FLAG_SACK_RENEGING
) {
2264 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2265 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2267 tcp_enter_loss(sk
, 1);
2268 icsk
->icsk_retransmits
++;
2269 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2270 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2271 icsk
->icsk_rto
, TCP_RTO_MAX
);
2277 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2279 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2282 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2283 * counter when SACK is enabled (without SACK, sacked_out is used for
2286 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2287 * segments up to the highest received SACK block so far and holes in
2290 * With reordering, holes may still be in flight, so RFC3517 recovery
2291 * uses pure sacked_out (total number of SACKed segments) even though
2292 * it violates the RFC that uses duplicate ACKs, often these are equal
2293 * but when e.g. out-of-window ACKs or packet duplication occurs,
2294 * they differ. Since neither occurs due to loss, TCP should really
2297 static inline int tcp_dupack_heuristics(struct tcp_sock
*tp
)
2299 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2302 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2304 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2307 static inline int tcp_head_timedout(struct sock
*sk
)
2309 struct tcp_sock
*tp
= tcp_sk(sk
);
2311 return tp
->packets_out
&&
2312 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2315 /* Linux NewReno/SACK/FACK/ECN state machine.
2316 * --------------------------------------
2318 * "Open" Normal state, no dubious events, fast path.
2319 * "Disorder" In all the respects it is "Open",
2320 * but requires a bit more attention. It is entered when
2321 * we see some SACKs or dupacks. It is split of "Open"
2322 * mainly to move some processing from fast path to slow one.
2323 * "CWR" CWND was reduced due to some Congestion Notification event.
2324 * It can be ECN, ICMP source quench, local device congestion.
2325 * "Recovery" CWND was reduced, we are fast-retransmitting.
2326 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2328 * tcp_fastretrans_alert() is entered:
2329 * - each incoming ACK, if state is not "Open"
2330 * - when arrived ACK is unusual, namely:
2335 * Counting packets in flight is pretty simple.
2337 * in_flight = packets_out - left_out + retrans_out
2339 * packets_out is SND.NXT-SND.UNA counted in packets.
2341 * retrans_out is number of retransmitted segments.
2343 * left_out is number of segments left network, but not ACKed yet.
2345 * left_out = sacked_out + lost_out
2347 * sacked_out: Packets, which arrived to receiver out of order
2348 * and hence not ACKed. With SACKs this number is simply
2349 * amount of SACKed data. Even without SACKs
2350 * it is easy to give pretty reliable estimate of this number,
2351 * counting duplicate ACKs.
2353 * lost_out: Packets lost by network. TCP has no explicit
2354 * "loss notification" feedback from network (for now).
2355 * It means that this number can be only _guessed_.
2356 * Actually, it is the heuristics to predict lossage that
2357 * distinguishes different algorithms.
2359 * F.e. after RTO, when all the queue is considered as lost,
2360 * lost_out = packets_out and in_flight = retrans_out.
2362 * Essentially, we have now two algorithms counting
2365 * FACK: It is the simplest heuristics. As soon as we decided
2366 * that something is lost, we decide that _all_ not SACKed
2367 * packets until the most forward SACK are lost. I.e.
2368 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2369 * It is absolutely correct estimate, if network does not reorder
2370 * packets. And it loses any connection to reality when reordering
2371 * takes place. We use FACK by default until reordering
2372 * is suspected on the path to this destination.
2374 * NewReno: when Recovery is entered, we assume that one segment
2375 * is lost (classic Reno). While we are in Recovery and
2376 * a partial ACK arrives, we assume that one more packet
2377 * is lost (NewReno). This heuristics are the same in NewReno
2380 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2381 * deflation etc. CWND is real congestion window, never inflated, changes
2382 * only according to classic VJ rules.
2384 * Really tricky (and requiring careful tuning) part of algorithm
2385 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2386 * The first determines the moment _when_ we should reduce CWND and,
2387 * hence, slow down forward transmission. In fact, it determines the moment
2388 * when we decide that hole is caused by loss, rather than by a reorder.
2390 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2391 * holes, caused by lost packets.
2393 * And the most logically complicated part of algorithm is undo
2394 * heuristics. We detect false retransmits due to both too early
2395 * fast retransmit (reordering) and underestimated RTO, analyzing
2396 * timestamps and D-SACKs. When we detect that some segments were
2397 * retransmitted by mistake and CWND reduction was wrong, we undo
2398 * window reduction and abort recovery phase. This logic is hidden
2399 * inside several functions named tcp_try_undo_<something>.
2402 /* This function decides, when we should leave Disordered state
2403 * and enter Recovery phase, reducing congestion window.
2405 * Main question: may we further continue forward transmission
2406 * with the same cwnd?
2408 static int tcp_time_to_recover(struct sock
*sk
)
2410 struct tcp_sock
*tp
= tcp_sk(sk
);
2413 /* Do not perform any recovery during F-RTO algorithm */
2414 if (tp
->frto_counter
)
2417 /* Trick#1: The loss is proven. */
2421 /* Not-A-Trick#2 : Classic rule... */
2422 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2425 /* Trick#3 : when we use RFC2988 timer restart, fast
2426 * retransmit can be triggered by timeout of queue head.
2428 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2431 /* Trick#4: It is still not OK... But will it be useful to delay
2434 packets_out
= tp
->packets_out
;
2435 if (packets_out
<= tp
->reordering
&&
2436 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2437 !tcp_may_send_now(sk
)) {
2438 /* We have nothing to send. This connection is limited
2439 * either by receiver window or by application.
2444 /* If a thin stream is detected, retransmit after first
2445 * received dupack. Employ only if SACK is supported in order
2446 * to avoid possible corner-case series of spurious retransmissions
2447 * Use only if there are no unsent data.
2449 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2450 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2451 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2457 /* New heuristics: it is possible only after we switched to restart timer
2458 * each time when something is ACKed. Hence, we can detect timed out packets
2459 * during fast retransmit without falling to slow start.
2461 * Usefulness of this as is very questionable, since we should know which of
2462 * the segments is the next to timeout which is relatively expensive to find
2463 * in general case unless we add some data structure just for that. The
2464 * current approach certainly won't find the right one too often and when it
2465 * finally does find _something_ it usually marks large part of the window
2466 * right away (because a retransmission with a larger timestamp blocks the
2467 * loop from advancing). -ij
2469 static void tcp_timeout_skbs(struct sock
*sk
)
2471 struct tcp_sock
*tp
= tcp_sk(sk
);
2472 struct sk_buff
*skb
;
2474 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2477 skb
= tp
->scoreboard_skb_hint
;
2478 if (tp
->scoreboard_skb_hint
== NULL
)
2479 skb
= tcp_write_queue_head(sk
);
2481 tcp_for_write_queue_from(skb
, sk
) {
2482 if (skb
== tcp_send_head(sk
))
2484 if (!tcp_skb_timedout(sk
, skb
))
2487 tcp_skb_mark_lost(tp
, skb
);
2490 tp
->scoreboard_skb_hint
= skb
;
2492 tcp_verify_left_out(tp
);
2495 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2496 * is against sacked "cnt", otherwise it's against facked "cnt"
2498 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2500 struct tcp_sock
*tp
= tcp_sk(sk
);
2501 struct sk_buff
*skb
;
2509 WARN_ON(packets
> tp
->packets_out
);
2510 if (tp
->lost_skb_hint
) {
2511 skb
= tp
->lost_skb_hint
;
2512 cnt
= tp
->lost_cnt_hint
;
2514 skb
= tcp_write_queue_head(sk
);
2518 tcp_for_write_queue_from(skb
, sk
) {
2519 if (skb
== tcp_send_head(sk
))
2521 /* TODO: do this better */
2522 /* this is not the most efficient way to do this... */
2523 tp
->lost_skb_hint
= skb
;
2524 tp
->lost_cnt_hint
= cnt
;
2526 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2530 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2531 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2532 cnt
+= tcp_skb_pcount(skb
);
2534 if (cnt
> packets
) {
2535 if (tcp_is_sack(tp
) || (oldcnt
>= packets
))
2538 mss
= skb_shinfo(skb
)->gso_size
;
2539 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2545 tcp_skb_mark_lost(tp
, skb
);
2547 tcp_verify_left_out(tp
);
2550 /* Account newly detected lost packet(s) */
2552 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2554 struct tcp_sock
*tp
= tcp_sk(sk
);
2556 if (tcp_is_reno(tp
)) {
2557 tcp_mark_head_lost(sk
, 1);
2558 } else if (tcp_is_fack(tp
)) {
2559 int lost
= tp
->fackets_out
- tp
->reordering
;
2562 tcp_mark_head_lost(sk
, lost
);
2564 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2565 if (sacked_upto
< fast_rexmit
)
2566 sacked_upto
= fast_rexmit
;
2567 tcp_mark_head_lost(sk
, sacked_upto
);
2570 tcp_timeout_skbs(sk
);
2573 /* CWND moderation, preventing bursts due to too big ACKs
2574 * in dubious situations.
2576 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2578 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2579 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2580 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2583 /* Lower bound on congestion window is slow start threshold
2584 * unless congestion avoidance choice decides to overide it.
2586 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2588 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2590 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2593 /* Decrease cwnd each second ack. */
2594 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2596 struct tcp_sock
*tp
= tcp_sk(sk
);
2597 int decr
= tp
->snd_cwnd_cnt
+ 1;
2599 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2600 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2601 tp
->snd_cwnd_cnt
= decr
& 1;
2604 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2605 tp
->snd_cwnd
-= decr
;
2607 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2608 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2612 /* Nothing was retransmitted or returned timestamp is less
2613 * than timestamp of the first retransmission.
2615 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2617 return !tp
->retrans_stamp
||
2618 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2619 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2622 /* Undo procedures. */
2624 #if FASTRETRANS_DEBUG > 1
2625 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2627 struct tcp_sock
*tp
= tcp_sk(sk
);
2628 struct inet_sock
*inet
= inet_sk(sk
);
2630 if (sk
->sk_family
== AF_INET
) {
2631 printk(KERN_DEBUG
"Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2633 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2634 tp
->snd_cwnd
, tcp_left_out(tp
),
2635 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2638 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2639 else if (sk
->sk_family
== AF_INET6
) {
2640 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2641 printk(KERN_DEBUG
"Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2643 &np
->daddr
, ntohs(inet
->inet_dport
),
2644 tp
->snd_cwnd
, tcp_left_out(tp
),
2645 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2651 #define DBGUNDO(x...) do { } while (0)
2654 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2656 struct tcp_sock
*tp
= tcp_sk(sk
);
2658 if (tp
->prior_ssthresh
) {
2659 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2661 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2662 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2664 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2666 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2667 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2668 TCP_ECN_withdraw_cwr(tp
);
2671 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2673 tcp_moderate_cwnd(tp
);
2674 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2677 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2679 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2682 /* People celebrate: "We love our President!" */
2683 static int tcp_try_undo_recovery(struct sock
*sk
)
2685 struct tcp_sock
*tp
= tcp_sk(sk
);
2687 if (tcp_may_undo(tp
)) {
2690 /* Happy end! We did not retransmit anything
2691 * or our original transmission succeeded.
2693 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2694 tcp_undo_cwr(sk
, 1);
2695 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2696 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2698 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2700 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2701 tp
->undo_marker
= 0;
2703 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2704 /* Hold old state until something *above* high_seq
2705 * is ACKed. For Reno it is MUST to prevent false
2706 * fast retransmits (RFC2582). SACK TCP is safe. */
2707 tcp_moderate_cwnd(tp
);
2710 tcp_set_ca_state(sk
, TCP_CA_Open
);
2714 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2715 static void tcp_try_undo_dsack(struct sock
*sk
)
2717 struct tcp_sock
*tp
= tcp_sk(sk
);
2719 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2720 DBGUNDO(sk
, "D-SACK");
2721 tcp_undo_cwr(sk
, 1);
2722 tp
->undo_marker
= 0;
2723 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2727 /* We can clear retrans_stamp when there are no retransmissions in the
2728 * window. It would seem that it is trivially available for us in
2729 * tp->retrans_out, however, that kind of assumptions doesn't consider
2730 * what will happen if errors occur when sending retransmission for the
2731 * second time. ...It could the that such segment has only
2732 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2733 * the head skb is enough except for some reneging corner cases that
2734 * are not worth the effort.
2736 * Main reason for all this complexity is the fact that connection dying
2737 * time now depends on the validity of the retrans_stamp, in particular,
2738 * that successive retransmissions of a segment must not advance
2739 * retrans_stamp under any conditions.
2741 static int tcp_any_retrans_done(struct sock
*sk
)
2743 struct tcp_sock
*tp
= tcp_sk(sk
);
2744 struct sk_buff
*skb
;
2746 if (tp
->retrans_out
)
2749 skb
= tcp_write_queue_head(sk
);
2750 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2756 /* Undo during fast recovery after partial ACK. */
2758 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2760 struct tcp_sock
*tp
= tcp_sk(sk
);
2761 /* Partial ACK arrived. Force Hoe's retransmit. */
2762 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2764 if (tcp_may_undo(tp
)) {
2765 /* Plain luck! Hole if filled with delayed
2766 * packet, rather than with a retransmit.
2768 if (!tcp_any_retrans_done(sk
))
2769 tp
->retrans_stamp
= 0;
2771 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2774 tcp_undo_cwr(sk
, 0);
2775 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2777 /* So... Do not make Hoe's retransmit yet.
2778 * If the first packet was delayed, the rest
2779 * ones are most probably delayed as well.
2786 /* Undo during loss recovery after partial ACK. */
2787 static int tcp_try_undo_loss(struct sock
*sk
)
2789 struct tcp_sock
*tp
= tcp_sk(sk
);
2791 if (tcp_may_undo(tp
)) {
2792 struct sk_buff
*skb
;
2793 tcp_for_write_queue(skb
, sk
) {
2794 if (skb
== tcp_send_head(sk
))
2796 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2799 tcp_clear_all_retrans_hints(tp
);
2801 DBGUNDO(sk
, "partial loss");
2803 tcp_undo_cwr(sk
, 1);
2804 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2805 inet_csk(sk
)->icsk_retransmits
= 0;
2806 tp
->undo_marker
= 0;
2807 if (tcp_is_sack(tp
))
2808 tcp_set_ca_state(sk
, TCP_CA_Open
);
2814 static inline void tcp_complete_cwr(struct sock
*sk
)
2816 struct tcp_sock
*tp
= tcp_sk(sk
);
2817 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2818 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2819 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2822 static void tcp_try_keep_open(struct sock
*sk
)
2824 struct tcp_sock
*tp
= tcp_sk(sk
);
2825 int state
= TCP_CA_Open
;
2827 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
) || tp
->undo_marker
)
2828 state
= TCP_CA_Disorder
;
2830 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2831 tcp_set_ca_state(sk
, state
);
2832 tp
->high_seq
= tp
->snd_nxt
;
2836 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2838 struct tcp_sock
*tp
= tcp_sk(sk
);
2840 tcp_verify_left_out(tp
);
2842 if (!tp
->frto_counter
&& !tcp_any_retrans_done(sk
))
2843 tp
->retrans_stamp
= 0;
2845 if (flag
& FLAG_ECE
)
2846 tcp_enter_cwr(sk
, 1);
2848 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2849 tcp_try_keep_open(sk
);
2850 tcp_moderate_cwnd(tp
);
2852 tcp_cwnd_down(sk
, flag
);
2856 static void tcp_mtup_probe_failed(struct sock
*sk
)
2858 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2860 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2861 icsk
->icsk_mtup
.probe_size
= 0;
2864 static void tcp_mtup_probe_success(struct sock
*sk
)
2866 struct tcp_sock
*tp
= tcp_sk(sk
);
2867 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2869 /* FIXME: breaks with very large cwnd */
2870 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2871 tp
->snd_cwnd
= tp
->snd_cwnd
*
2872 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2873 icsk
->icsk_mtup
.probe_size
;
2874 tp
->snd_cwnd_cnt
= 0;
2875 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2876 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2878 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2879 icsk
->icsk_mtup
.probe_size
= 0;
2880 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2883 /* Do a simple retransmit without using the backoff mechanisms in
2884 * tcp_timer. This is used for path mtu discovery.
2885 * The socket is already locked here.
2887 void tcp_simple_retransmit(struct sock
*sk
)
2889 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2890 struct tcp_sock
*tp
= tcp_sk(sk
);
2891 struct sk_buff
*skb
;
2892 unsigned int mss
= tcp_current_mss(sk
);
2893 u32 prior_lost
= tp
->lost_out
;
2895 tcp_for_write_queue(skb
, sk
) {
2896 if (skb
== tcp_send_head(sk
))
2898 if (tcp_skb_seglen(skb
) > mss
&&
2899 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2900 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2901 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2902 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2904 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2908 tcp_clear_retrans_hints_partial(tp
);
2910 if (prior_lost
== tp
->lost_out
)
2913 if (tcp_is_reno(tp
))
2914 tcp_limit_reno_sacked(tp
);
2916 tcp_verify_left_out(tp
);
2918 /* Don't muck with the congestion window here.
2919 * Reason is that we do not increase amount of _data_
2920 * in network, but units changed and effective
2921 * cwnd/ssthresh really reduced now.
2923 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2924 tp
->high_seq
= tp
->snd_nxt
;
2925 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2926 tp
->prior_ssthresh
= 0;
2927 tp
->undo_marker
= 0;
2928 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2930 tcp_xmit_retransmit_queue(sk
);
2932 EXPORT_SYMBOL(tcp_simple_retransmit
);
2934 /* Process an event, which can update packets-in-flight not trivially.
2935 * Main goal of this function is to calculate new estimate for left_out,
2936 * taking into account both packets sitting in receiver's buffer and
2937 * packets lost by network.
2939 * Besides that it does CWND reduction, when packet loss is detected
2940 * and changes state of machine.
2942 * It does _not_ decide what to send, it is made in function
2943 * tcp_xmit_retransmit_queue().
2945 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2947 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2948 struct tcp_sock
*tp
= tcp_sk(sk
);
2949 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2950 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2951 (tcp_fackets_out(tp
) > tp
->reordering
));
2952 int fast_rexmit
= 0, mib_idx
;
2954 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2956 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2957 tp
->fackets_out
= 0;
2959 /* Now state machine starts.
2960 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2961 if (flag
& FLAG_ECE
)
2962 tp
->prior_ssthresh
= 0;
2964 /* B. In all the states check for reneging SACKs. */
2965 if (tcp_check_sack_reneging(sk
, flag
))
2968 /* C. Process data loss notification, provided it is valid. */
2969 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2970 before(tp
->snd_una
, tp
->high_seq
) &&
2971 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2972 tp
->fackets_out
> tp
->reordering
) {
2973 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2974 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2977 /* D. Check consistency of the current state. */
2978 tcp_verify_left_out(tp
);
2980 /* E. Check state exit conditions. State can be terminated
2981 * when high_seq is ACKed. */
2982 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2983 WARN_ON(tp
->retrans_out
!= 0);
2984 tp
->retrans_stamp
= 0;
2985 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2986 switch (icsk
->icsk_ca_state
) {
2988 icsk
->icsk_retransmits
= 0;
2989 if (tcp_try_undo_recovery(sk
))
2994 /* CWR is to be held something *above* high_seq
2995 * is ACKed for CWR bit to reach receiver. */
2996 if (tp
->snd_una
!= tp
->high_seq
) {
2997 tcp_complete_cwr(sk
);
2998 tcp_set_ca_state(sk
, TCP_CA_Open
);
3002 case TCP_CA_Disorder
:
3003 tcp_try_undo_dsack(sk
);
3004 if (!tp
->undo_marker
||
3005 /* For SACK case do not Open to allow to undo
3006 * catching for all duplicate ACKs. */
3007 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
3008 tp
->undo_marker
= 0;
3009 tcp_set_ca_state(sk
, TCP_CA_Open
);
3013 case TCP_CA_Recovery
:
3014 if (tcp_is_reno(tp
))
3015 tcp_reset_reno_sack(tp
);
3016 if (tcp_try_undo_recovery(sk
))
3018 tcp_complete_cwr(sk
);
3023 /* F. Process state. */
3024 switch (icsk
->icsk_ca_state
) {
3025 case TCP_CA_Recovery
:
3026 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
3027 if (tcp_is_reno(tp
) && is_dupack
)
3028 tcp_add_reno_sack(sk
);
3030 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
3033 if (flag
& FLAG_DATA_ACKED
)
3034 icsk
->icsk_retransmits
= 0;
3035 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
3036 tcp_reset_reno_sack(tp
);
3037 if (!tcp_try_undo_loss(sk
)) {
3038 tcp_moderate_cwnd(tp
);
3039 tcp_xmit_retransmit_queue(sk
);
3042 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
3044 /* Loss is undone; fall through to processing in Open state. */
3046 if (tcp_is_reno(tp
)) {
3047 if (flag
& FLAG_SND_UNA_ADVANCED
)
3048 tcp_reset_reno_sack(tp
);
3050 tcp_add_reno_sack(sk
);
3053 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
3054 tcp_try_undo_dsack(sk
);
3056 if (!tcp_time_to_recover(sk
)) {
3057 tcp_try_to_open(sk
, flag
);
3061 /* MTU probe failure: don't reduce cwnd */
3062 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3063 icsk
->icsk_mtup
.probe_size
&&
3064 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3065 tcp_mtup_probe_failed(sk
);
3066 /* Restores the reduction we did in tcp_mtup_probe() */
3068 tcp_simple_retransmit(sk
);
3072 /* Otherwise enter Recovery state */
3074 if (tcp_is_reno(tp
))
3075 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
3077 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
3079 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3081 tp
->high_seq
= tp
->snd_nxt
;
3082 tp
->prior_ssthresh
= 0;
3083 tp
->undo_marker
= tp
->snd_una
;
3084 tp
->undo_retrans
= tp
->retrans_out
;
3086 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
3087 if (!(flag
& FLAG_ECE
))
3088 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
3089 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
3090 TCP_ECN_queue_cwr(tp
);
3093 tp
->bytes_acked
= 0;
3094 tp
->snd_cwnd_cnt
= 0;
3095 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
3099 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3100 tcp_update_scoreboard(sk
, fast_rexmit
);
3101 tcp_cwnd_down(sk
, flag
);
3102 tcp_xmit_retransmit_queue(sk
);
3105 static void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
3107 tcp_rtt_estimator(sk
, seq_rtt
);
3109 inet_csk(sk
)->icsk_backoff
= 0;
3112 /* Read draft-ietf-tcplw-high-performance before mucking
3113 * with this code. (Supersedes RFC1323)
3115 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3117 /* RTTM Rule: A TSecr value received in a segment is used to
3118 * update the averaged RTT measurement only if the segment
3119 * acknowledges some new data, i.e., only if it advances the
3120 * left edge of the send window.
3122 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3123 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3125 * Changed: reset backoff as soon as we see the first valid sample.
3126 * If we do not, we get strongly overestimated rto. With timestamps
3127 * samples are accepted even from very old segments: f.e., when rtt=1
3128 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3129 * answer arrives rto becomes 120 seconds! If at least one of segments
3130 * in window is lost... Voila. --ANK (010210)
3132 struct tcp_sock
*tp
= tcp_sk(sk
);
3134 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
3137 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3139 /* We don't have a timestamp. Can only use
3140 * packets that are not retransmitted to determine
3141 * rtt estimates. Also, we must not reset the
3142 * backoff for rto until we get a non-retransmitted
3143 * packet. This allows us to deal with a situation
3144 * where the network delay has increased suddenly.
3145 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3148 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3151 tcp_valid_rtt_meas(sk
, seq_rtt
);
3154 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3157 const struct tcp_sock
*tp
= tcp_sk(sk
);
3158 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3159 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3160 tcp_ack_saw_tstamp(sk
, flag
);
3161 else if (seq_rtt
>= 0)
3162 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3165 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3167 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3168 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3169 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3172 /* Restart timer after forward progress on connection.
3173 * RFC2988 recommends to restart timer to now+rto.
3175 static void tcp_rearm_rto(struct sock
*sk
)
3177 struct tcp_sock
*tp
= tcp_sk(sk
);
3179 if (!tp
->packets_out
) {
3180 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3182 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3183 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3187 /* If we get here, the whole TSO packet has not been acked. */
3188 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3190 struct tcp_sock
*tp
= tcp_sk(sk
);
3193 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3195 packets_acked
= tcp_skb_pcount(skb
);
3196 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3198 packets_acked
-= tcp_skb_pcount(skb
);
3200 if (packets_acked
) {
3201 BUG_ON(tcp_skb_pcount(skb
) == 0);
3202 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3205 return packets_acked
;
3208 /* Remove acknowledged frames from the retransmission queue. If our packet
3209 * is before the ack sequence we can discard it as it's confirmed to have
3210 * arrived at the other end.
3212 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3215 struct tcp_sock
*tp
= tcp_sk(sk
);
3216 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3217 struct sk_buff
*skb
;
3218 u32 now
= tcp_time_stamp
;
3219 int fully_acked
= 1;
3222 u32 reord
= tp
->packets_out
;
3223 u32 prior_sacked
= tp
->sacked_out
;
3225 s32 ca_seq_rtt
= -1;
3226 ktime_t last_ackt
= net_invalid_timestamp();
3228 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3229 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3231 u8 sacked
= scb
->sacked
;
3233 /* Determine how many packets and what bytes were acked, tso and else */
3234 if (after(scb
->end_seq
, tp
->snd_una
)) {
3235 if (tcp_skb_pcount(skb
) == 1 ||
3236 !after(tp
->snd_una
, scb
->seq
))
3239 acked_pcount
= tcp_tso_acked(sk
, skb
);
3245 acked_pcount
= tcp_skb_pcount(skb
);
3248 if (sacked
& TCPCB_RETRANS
) {
3249 if (sacked
& TCPCB_SACKED_RETRANS
)
3250 tp
->retrans_out
-= acked_pcount
;
3251 flag
|= FLAG_RETRANS_DATA_ACKED
;
3254 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3255 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3257 ca_seq_rtt
= now
- scb
->when
;
3258 last_ackt
= skb
->tstamp
;
3260 seq_rtt
= ca_seq_rtt
;
3262 if (!(sacked
& TCPCB_SACKED_ACKED
))
3263 reord
= min(pkts_acked
, reord
);
3266 if (sacked
& TCPCB_SACKED_ACKED
)
3267 tp
->sacked_out
-= acked_pcount
;
3268 if (sacked
& TCPCB_LOST
)
3269 tp
->lost_out
-= acked_pcount
;
3271 tp
->packets_out
-= acked_pcount
;
3272 pkts_acked
+= acked_pcount
;
3274 /* Initial outgoing SYN's get put onto the write_queue
3275 * just like anything else we transmit. It is not
3276 * true data, and if we misinform our callers that
3277 * this ACK acks real data, we will erroneously exit
3278 * connection startup slow start one packet too
3279 * quickly. This is severely frowned upon behavior.
3281 if (!(scb
->flags
& TCPHDR_SYN
)) {
3282 flag
|= FLAG_DATA_ACKED
;
3284 flag
|= FLAG_SYN_ACKED
;
3285 tp
->retrans_stamp
= 0;
3291 tcp_unlink_write_queue(skb
, sk
);
3292 sk_wmem_free_skb(sk
, skb
);
3293 tp
->scoreboard_skb_hint
= NULL
;
3294 if (skb
== tp
->retransmit_skb_hint
)
3295 tp
->retransmit_skb_hint
= NULL
;
3296 if (skb
== tp
->lost_skb_hint
)
3297 tp
->lost_skb_hint
= NULL
;
3300 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3301 tp
->snd_up
= tp
->snd_una
;
3303 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3304 flag
|= FLAG_SACK_RENEGING
;
3306 if (flag
& FLAG_ACKED
) {
3307 const struct tcp_congestion_ops
*ca_ops
3308 = inet_csk(sk
)->icsk_ca_ops
;
3310 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3311 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3312 tcp_mtup_probe_success(sk
);
3315 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3318 if (tcp_is_reno(tp
)) {
3319 tcp_remove_reno_sacks(sk
, pkts_acked
);
3323 /* Non-retransmitted hole got filled? That's reordering */
3324 if (reord
< prior_fackets
)
3325 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3327 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3328 prior_sacked
- tp
->sacked_out
;
3329 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3332 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3334 if (ca_ops
->pkts_acked
) {
3337 /* Is the ACK triggering packet unambiguous? */
3338 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3339 /* High resolution needed and available? */
3340 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3341 !ktime_equal(last_ackt
,
3342 net_invalid_timestamp()))
3343 rtt_us
= ktime_us_delta(ktime_get_real(),
3345 else if (ca_seq_rtt
> 0)
3346 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3349 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3353 #if FASTRETRANS_DEBUG > 0
3354 WARN_ON((int)tp
->sacked_out
< 0);
3355 WARN_ON((int)tp
->lost_out
< 0);
3356 WARN_ON((int)tp
->retrans_out
< 0);
3357 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3358 icsk
= inet_csk(sk
);
3360 printk(KERN_DEBUG
"Leak l=%u %d\n",
3361 tp
->lost_out
, icsk
->icsk_ca_state
);
3364 if (tp
->sacked_out
) {
3365 printk(KERN_DEBUG
"Leak s=%u %d\n",
3366 tp
->sacked_out
, icsk
->icsk_ca_state
);
3369 if (tp
->retrans_out
) {
3370 printk(KERN_DEBUG
"Leak r=%u %d\n",
3371 tp
->retrans_out
, icsk
->icsk_ca_state
);
3372 tp
->retrans_out
= 0;
3379 static void tcp_ack_probe(struct sock
*sk
)
3381 const struct tcp_sock
*tp
= tcp_sk(sk
);
3382 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3384 /* Was it a usable window open? */
3386 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3387 icsk
->icsk_backoff
= 0;
3388 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3389 /* Socket must be waked up by subsequent tcp_data_snd_check().
3390 * This function is not for random using!
3393 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3394 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3399 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3401 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3402 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3405 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3407 const struct tcp_sock
*tp
= tcp_sk(sk
);
3408 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3409 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3412 /* Check that window update is acceptable.
3413 * The function assumes that snd_una<=ack<=snd_next.
3415 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3416 const u32 ack
, const u32 ack_seq
,
3419 return (after(ack
, tp
->snd_una
) ||
3420 after(ack_seq
, tp
->snd_wl1
) ||
3421 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3424 /* Update our send window.
3426 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3427 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3429 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3432 struct tcp_sock
*tp
= tcp_sk(sk
);
3434 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3436 if (likely(!tcp_hdr(skb
)->syn
))
3437 nwin
<<= tp
->rx_opt
.snd_wscale
;
3439 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3440 flag
|= FLAG_WIN_UPDATE
;
3441 tcp_update_wl(tp
, ack_seq
);
3443 if (tp
->snd_wnd
!= nwin
) {
3446 /* Note, it is the only place, where
3447 * fast path is recovered for sending TCP.
3450 tcp_fast_path_check(sk
);
3452 if (nwin
> tp
->max_window
) {
3453 tp
->max_window
= nwin
;
3454 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3464 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3465 * continue in congestion avoidance.
3467 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3469 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3470 tp
->snd_cwnd_cnt
= 0;
3471 tp
->bytes_acked
= 0;
3472 TCP_ECN_queue_cwr(tp
);
3473 tcp_moderate_cwnd(tp
);
3476 /* A conservative spurious RTO response algorithm: reduce cwnd using
3477 * rate halving and continue in congestion avoidance.
3479 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3481 tcp_enter_cwr(sk
, 0);
3484 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3486 if (flag
& FLAG_ECE
)
3487 tcp_ratehalving_spur_to_response(sk
);
3489 tcp_undo_cwr(sk
, 1);
3492 /* F-RTO spurious RTO detection algorithm (RFC4138)
3494 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3495 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3496 * window (but not to or beyond highest sequence sent before RTO):
3497 * On First ACK, send two new segments out.
3498 * On Second ACK, RTO was likely spurious. Do spurious response (response
3499 * algorithm is not part of the F-RTO detection algorithm
3500 * given in RFC4138 but can be selected separately).
3501 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3502 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3503 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3504 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3506 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3507 * original window even after we transmit two new data segments.
3510 * on first step, wait until first cumulative ACK arrives, then move to
3511 * the second step. In second step, the next ACK decides.
3513 * F-RTO is implemented (mainly) in four functions:
3514 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3515 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3516 * called when tcp_use_frto() showed green light
3517 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3518 * - tcp_enter_frto_loss() is called if there is not enough evidence
3519 * to prove that the RTO is indeed spurious. It transfers the control
3520 * from F-RTO to the conventional RTO recovery
3522 static int tcp_process_frto(struct sock
*sk
, int flag
)
3524 struct tcp_sock
*tp
= tcp_sk(sk
);
3526 tcp_verify_left_out(tp
);
3528 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3529 if (flag
& FLAG_DATA_ACKED
)
3530 inet_csk(sk
)->icsk_retransmits
= 0;
3532 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3533 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3534 tp
->undo_marker
= 0;
3536 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3537 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3541 if (!tcp_is_sackfrto(tp
)) {
3542 /* RFC4138 shortcoming in step 2; should also have case c):
3543 * ACK isn't duplicate nor advances window, e.g., opposite dir
3546 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3549 if (!(flag
& FLAG_DATA_ACKED
)) {
3550 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3555 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3556 /* Prevent sending of new data. */
3557 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3558 tcp_packets_in_flight(tp
));
3562 if ((tp
->frto_counter
>= 2) &&
3563 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3564 ((flag
& FLAG_DATA_SACKED
) &&
3565 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3566 /* RFC4138 shortcoming (see comment above) */
3567 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3568 (flag
& FLAG_NOT_DUP
))
3571 tcp_enter_frto_loss(sk
, 3, flag
);
3576 if (tp
->frto_counter
== 1) {
3577 /* tcp_may_send_now needs to see updated state */
3578 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3579 tp
->frto_counter
= 2;
3581 if (!tcp_may_send_now(sk
))
3582 tcp_enter_frto_loss(sk
, 2, flag
);
3586 switch (sysctl_tcp_frto_response
) {
3588 tcp_undo_spur_to_response(sk
, flag
);
3591 tcp_conservative_spur_to_response(tp
);
3594 tcp_ratehalving_spur_to_response(sk
);
3597 tp
->frto_counter
= 0;
3598 tp
->undo_marker
= 0;
3599 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3604 /* This routine deals with incoming acks, but not outgoing ones. */
3605 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3607 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3608 struct tcp_sock
*tp
= tcp_sk(sk
);
3609 u32 prior_snd_una
= tp
->snd_una
;
3610 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3611 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3612 u32 prior_in_flight
;
3617 /* If the ack is older than previous acks
3618 * then we can probably ignore it.
3620 if (before(ack
, prior_snd_una
))
3623 /* If the ack includes data we haven't sent yet, discard
3624 * this segment (RFC793 Section 3.9).
3626 if (after(ack
, tp
->snd_nxt
))
3629 if (after(ack
, prior_snd_una
))
3630 flag
|= FLAG_SND_UNA_ADVANCED
;
3632 if (sysctl_tcp_abc
) {
3633 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3634 tp
->bytes_acked
+= ack
- prior_snd_una
;
3635 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3636 /* we assume just one segment left network */
3637 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3641 prior_fackets
= tp
->fackets_out
;
3642 prior_in_flight
= tcp_packets_in_flight(tp
);
3644 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3645 /* Window is constant, pure forward advance.
3646 * No more checks are required.
3647 * Note, we use the fact that SND.UNA>=SND.WL2.
3649 tcp_update_wl(tp
, ack_seq
);
3651 flag
|= FLAG_WIN_UPDATE
;
3653 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3655 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3657 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3660 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3662 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3664 if (TCP_SKB_CB(skb
)->sacked
)
3665 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3667 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3670 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3673 /* We passed data and got it acked, remove any soft error
3674 * log. Something worked...
3676 sk
->sk_err_soft
= 0;
3677 icsk
->icsk_probes_out
= 0;
3678 tp
->rcv_tstamp
= tcp_time_stamp
;
3679 prior_packets
= tp
->packets_out
;
3683 /* See if we can take anything off of the retransmit queue. */
3684 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3686 if (tp
->frto_counter
)
3687 frto_cwnd
= tcp_process_frto(sk
, flag
);
3688 /* Guarantee sacktag reordering detection against wrap-arounds */
3689 if (before(tp
->frto_highmark
, tp
->snd_una
))
3690 tp
->frto_highmark
= 0;
3692 if (tcp_ack_is_dubious(sk
, flag
)) {
3693 /* Advance CWND, if state allows this. */
3694 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3695 tcp_may_raise_cwnd(sk
, flag
))
3696 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3697 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3700 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3701 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3704 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3705 dst_confirm(__sk_dst_get(sk
));
3710 /* If this ack opens up a zero window, clear backoff. It was
3711 * being used to time the probes, and is probably far higher than
3712 * it needs to be for normal retransmission.
3714 if (tcp_send_head(sk
))
3719 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3723 if (TCP_SKB_CB(skb
)->sacked
) {
3724 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3725 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3726 tcp_try_keep_open(sk
);
3729 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3733 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3734 * But, this can also be called on packets in the established flow when
3735 * the fast version below fails.
3737 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3738 u8
**hvpp
, int estab
)
3741 struct tcphdr
*th
= tcp_hdr(skb
);
3742 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3744 ptr
= (unsigned char *)(th
+ 1);
3745 opt_rx
->saw_tstamp
= 0;
3747 while (length
> 0) {
3748 int opcode
= *ptr
++;
3754 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3759 if (opsize
< 2) /* "silly options" */
3761 if (opsize
> length
)
3762 return; /* don't parse partial options */
3765 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3766 u16 in_mss
= get_unaligned_be16(ptr
);
3768 if (opt_rx
->user_mss
&&
3769 opt_rx
->user_mss
< in_mss
)
3770 in_mss
= opt_rx
->user_mss
;
3771 opt_rx
->mss_clamp
= in_mss
;
3776 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3777 !estab
&& sysctl_tcp_window_scaling
) {
3778 __u8 snd_wscale
= *(__u8
*)ptr
;
3779 opt_rx
->wscale_ok
= 1;
3780 if (snd_wscale
> 14) {
3781 if (net_ratelimit())
3782 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3783 "scaling value %d >14 received.\n",
3787 opt_rx
->snd_wscale
= snd_wscale
;
3790 case TCPOPT_TIMESTAMP
:
3791 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3792 ((estab
&& opt_rx
->tstamp_ok
) ||
3793 (!estab
&& sysctl_tcp_timestamps
))) {
3794 opt_rx
->saw_tstamp
= 1;
3795 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3796 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3799 case TCPOPT_SACK_PERM
:
3800 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3801 !estab
&& sysctl_tcp_sack
) {
3802 opt_rx
->sack_ok
= 1;
3803 tcp_sack_reset(opt_rx
);
3808 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3809 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3811 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3814 #ifdef CONFIG_TCP_MD5SIG
3817 * The MD5 Hash has already been
3818 * checked (see tcp_v{4,6}_do_rcv()).
3823 /* This option is variable length.
3826 case TCPOLEN_COOKIE_BASE
:
3827 /* not yet implemented */
3829 case TCPOLEN_COOKIE_PAIR
:
3830 /* not yet implemented */
3832 case TCPOLEN_COOKIE_MIN
+0:
3833 case TCPOLEN_COOKIE_MIN
+2:
3834 case TCPOLEN_COOKIE_MIN
+4:
3835 case TCPOLEN_COOKIE_MIN
+6:
3836 case TCPOLEN_COOKIE_MAX
:
3837 /* 16-bit multiple */
3838 opt_rx
->cookie_plus
= opsize
;
3853 EXPORT_SYMBOL(tcp_parse_options
);
3855 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3857 __be32
*ptr
= (__be32
*)(th
+ 1);
3859 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3860 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3861 tp
->rx_opt
.saw_tstamp
= 1;
3863 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3865 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3871 /* Fast parse options. This hopes to only see timestamps.
3872 * If it is wrong it falls back on tcp_parse_options().
3874 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3875 struct tcp_sock
*tp
, u8
**hvpp
)
3877 /* In the spirit of fast parsing, compare doff directly to constant
3878 * values. Because equality is used, short doff can be ignored here.
3880 if (th
->doff
== (sizeof(*th
) / 4)) {
3881 tp
->rx_opt
.saw_tstamp
= 0;
3883 } else if (tp
->rx_opt
.tstamp_ok
&&
3884 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3885 if (tcp_parse_aligned_timestamp(tp
, th
))
3888 tcp_parse_options(skb
, &tp
->rx_opt
, hvpp
, 1);
3892 #ifdef CONFIG_TCP_MD5SIG
3894 * Parse MD5 Signature option
3896 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3898 int length
= (th
->doff
<< 2) - sizeof (*th
);
3899 u8
*ptr
= (u8
*)(th
+ 1);
3901 /* If the TCP option is too short, we can short cut */
3902 if (length
< TCPOLEN_MD5SIG
)
3905 while (length
> 0) {
3906 int opcode
= *ptr
++;
3917 if (opsize
< 2 || opsize
> length
)
3919 if (opcode
== TCPOPT_MD5SIG
)
3920 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3927 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3930 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3932 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3933 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3936 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3938 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3939 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3940 * extra check below makes sure this can only happen
3941 * for pure ACK frames. -DaveM
3943 * Not only, also it occurs for expired timestamps.
3946 if (tcp_paws_check(&tp
->rx_opt
, 0))
3947 tcp_store_ts_recent(tp
);
3951 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3953 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3954 * it can pass through stack. So, the following predicate verifies that
3955 * this segment is not used for anything but congestion avoidance or
3956 * fast retransmit. Moreover, we even are able to eliminate most of such
3957 * second order effects, if we apply some small "replay" window (~RTO)
3958 * to timestamp space.
3960 * All these measures still do not guarantee that we reject wrapped ACKs
3961 * on networks with high bandwidth, when sequence space is recycled fastly,
3962 * but it guarantees that such events will be very rare and do not affect
3963 * connection seriously. This doesn't look nice, but alas, PAWS is really
3966 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3967 * states that events when retransmit arrives after original data are rare.
3968 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3969 * the biggest problem on large power networks even with minor reordering.
3970 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3971 * up to bandwidth of 18Gigabit/sec. 8) ]
3974 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3976 struct tcp_sock
*tp
= tcp_sk(sk
);
3977 struct tcphdr
*th
= tcp_hdr(skb
);
3978 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3979 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3981 return (/* 1. Pure ACK with correct sequence number. */
3982 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3984 /* 2. ... and duplicate ACK. */
3985 ack
== tp
->snd_una
&&
3987 /* 3. ... and does not update window. */
3988 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3990 /* 4. ... and sits in replay window. */
3991 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3994 static inline int tcp_paws_discard(const struct sock
*sk
,
3995 const struct sk_buff
*skb
)
3997 const struct tcp_sock
*tp
= tcp_sk(sk
);
3999 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
4000 !tcp_disordered_ack(sk
, skb
);
4003 /* Check segment sequence number for validity.
4005 * Segment controls are considered valid, if the segment
4006 * fits to the window after truncation to the window. Acceptability
4007 * of data (and SYN, FIN, of course) is checked separately.
4008 * See tcp_data_queue(), for example.
4010 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4011 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4012 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4013 * (borrowed from freebsd)
4016 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4018 return !before(end_seq
, tp
->rcv_wup
) &&
4019 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4022 /* When we get a reset we do this. */
4023 static void tcp_reset(struct sock
*sk
)
4025 /* We want the right error as BSD sees it (and indeed as we do). */
4026 switch (sk
->sk_state
) {
4028 sk
->sk_err
= ECONNREFUSED
;
4030 case TCP_CLOSE_WAIT
:
4036 sk
->sk_err
= ECONNRESET
;
4039 if (!sock_flag(sk
, SOCK_DEAD
))
4040 sk
->sk_error_report(sk
);
4046 * Process the FIN bit. This now behaves as it is supposed to work
4047 * and the FIN takes effect when it is validly part of sequence
4048 * space. Not before when we get holes.
4050 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4051 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4054 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4055 * close and we go into CLOSING (and later onto TIME-WAIT)
4057 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4059 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
4061 struct tcp_sock
*tp
= tcp_sk(sk
);
4063 inet_csk_schedule_ack(sk
);
4065 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4066 sock_set_flag(sk
, SOCK_DONE
);
4068 switch (sk
->sk_state
) {
4070 case TCP_ESTABLISHED
:
4071 /* Move to CLOSE_WAIT */
4072 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4073 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4076 case TCP_CLOSE_WAIT
:
4078 /* Received a retransmission of the FIN, do
4083 /* RFC793: Remain in the LAST-ACK state. */
4087 /* This case occurs when a simultaneous close
4088 * happens, we must ack the received FIN and
4089 * enter the CLOSING state.
4092 tcp_set_state(sk
, TCP_CLOSING
);
4095 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4097 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4100 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4101 * cases we should never reach this piece of code.
4103 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
4104 __func__
, sk
->sk_state
);
4108 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4109 * Probably, we should reset in this case. For now drop them.
4111 __skb_queue_purge(&tp
->out_of_order_queue
);
4112 if (tcp_is_sack(tp
))
4113 tcp_sack_reset(&tp
->rx_opt
);
4116 if (!sock_flag(sk
, SOCK_DEAD
)) {
4117 sk
->sk_state_change(sk
);
4119 /* Do not send POLL_HUP for half duplex close. */
4120 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4121 sk
->sk_state
== TCP_CLOSE
)
4122 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4124 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4128 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4131 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4132 if (before(seq
, sp
->start_seq
))
4133 sp
->start_seq
= seq
;
4134 if (after(end_seq
, sp
->end_seq
))
4135 sp
->end_seq
= end_seq
;
4141 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4143 struct tcp_sock
*tp
= tcp_sk(sk
);
4145 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4148 if (before(seq
, tp
->rcv_nxt
))
4149 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4151 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4153 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4155 tp
->rx_opt
.dsack
= 1;
4156 tp
->duplicate_sack
[0].start_seq
= seq
;
4157 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4161 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4163 struct tcp_sock
*tp
= tcp_sk(sk
);
4165 if (!tp
->rx_opt
.dsack
)
4166 tcp_dsack_set(sk
, seq
, end_seq
);
4168 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4171 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
4173 struct tcp_sock
*tp
= tcp_sk(sk
);
4175 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4176 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4177 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4178 tcp_enter_quickack_mode(sk
);
4180 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4181 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4183 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4184 end_seq
= tp
->rcv_nxt
;
4185 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4192 /* These routines update the SACK block as out-of-order packets arrive or
4193 * in-order packets close up the sequence space.
4195 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4198 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4199 struct tcp_sack_block
*swalk
= sp
+ 1;
4201 /* See if the recent change to the first SACK eats into
4202 * or hits the sequence space of other SACK blocks, if so coalesce.
4204 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4205 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4208 /* Zap SWALK, by moving every further SACK up by one slot.
4209 * Decrease num_sacks.
4211 tp
->rx_opt
.num_sacks
--;
4212 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4216 this_sack
++, swalk
++;
4220 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4222 struct tcp_sock
*tp
= tcp_sk(sk
);
4223 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4224 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4230 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4231 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4232 /* Rotate this_sack to the first one. */
4233 for (; this_sack
> 0; this_sack
--, sp
--)
4234 swap(*sp
, *(sp
- 1));
4236 tcp_sack_maybe_coalesce(tp
);
4241 /* Could not find an adjacent existing SACK, build a new one,
4242 * put it at the front, and shift everyone else down. We
4243 * always know there is at least one SACK present already here.
4245 * If the sack array is full, forget about the last one.
4247 if (this_sack
>= TCP_NUM_SACKS
) {
4249 tp
->rx_opt
.num_sacks
--;
4252 for (; this_sack
> 0; this_sack
--, sp
--)
4256 /* Build the new head SACK, and we're done. */
4257 sp
->start_seq
= seq
;
4258 sp
->end_seq
= end_seq
;
4259 tp
->rx_opt
.num_sacks
++;
4262 /* RCV.NXT advances, some SACKs should be eaten. */
4264 static void tcp_sack_remove(struct tcp_sock
*tp
)
4266 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4267 int num_sacks
= tp
->rx_opt
.num_sacks
;
4270 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4271 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4272 tp
->rx_opt
.num_sacks
= 0;
4276 for (this_sack
= 0; this_sack
< num_sacks
;) {
4277 /* Check if the start of the sack is covered by RCV.NXT. */
4278 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4281 /* RCV.NXT must cover all the block! */
4282 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4284 /* Zap this SACK, by moving forward any other SACKS. */
4285 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4286 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4293 tp
->rx_opt
.num_sacks
= num_sacks
;
4296 /* This one checks to see if we can put data from the
4297 * out_of_order queue into the receive_queue.
4299 static void tcp_ofo_queue(struct sock
*sk
)
4301 struct tcp_sock
*tp
= tcp_sk(sk
);
4302 __u32 dsack_high
= tp
->rcv_nxt
;
4303 struct sk_buff
*skb
;
4305 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4306 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4309 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4310 __u32 dsack
= dsack_high
;
4311 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4312 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4313 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4316 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4317 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4318 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4322 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4323 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4324 TCP_SKB_CB(skb
)->end_seq
);
4326 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4327 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4328 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4329 if (tcp_hdr(skb
)->fin
)
4330 tcp_fin(skb
, sk
, tcp_hdr(skb
));
4334 static int tcp_prune_ofo_queue(struct sock
*sk
);
4335 static int tcp_prune_queue(struct sock
*sk
);
4337 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
4339 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4340 !sk_rmem_schedule(sk
, size
)) {
4342 if (tcp_prune_queue(sk
) < 0)
4345 if (!sk_rmem_schedule(sk
, size
)) {
4346 if (!tcp_prune_ofo_queue(sk
))
4349 if (!sk_rmem_schedule(sk
, size
))
4356 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4358 struct tcphdr
*th
= tcp_hdr(skb
);
4359 struct tcp_sock
*tp
= tcp_sk(sk
);
4362 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4366 __skb_pull(skb
, th
->doff
* 4);
4368 TCP_ECN_accept_cwr(tp
, skb
);
4370 tp
->rx_opt
.dsack
= 0;
4372 /* Queue data for delivery to the user.
4373 * Packets in sequence go to the receive queue.
4374 * Out of sequence packets to the out_of_order_queue.
4376 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4377 if (tcp_receive_window(tp
) == 0)
4380 /* Ok. In sequence. In window. */
4381 if (tp
->ucopy
.task
== current
&&
4382 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4383 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4384 int chunk
= min_t(unsigned int, skb
->len
,
4387 __set_current_state(TASK_RUNNING
);
4390 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4391 tp
->ucopy
.len
-= chunk
;
4392 tp
->copied_seq
+= chunk
;
4393 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4394 tcp_rcv_space_adjust(sk
);
4402 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4405 skb_set_owner_r(skb
, sk
);
4406 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4408 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4410 tcp_event_data_recv(sk
, skb
);
4412 tcp_fin(skb
, sk
, th
);
4414 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4417 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4418 * gap in queue is filled.
4420 if (skb_queue_empty(&tp
->out_of_order_queue
))
4421 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4424 if (tp
->rx_opt
.num_sacks
)
4425 tcp_sack_remove(tp
);
4427 tcp_fast_path_check(sk
);
4431 else if (!sock_flag(sk
, SOCK_DEAD
))
4432 sk
->sk_data_ready(sk
, 0);
4436 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4437 /* A retransmit, 2nd most common case. Force an immediate ack. */
4438 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4439 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4442 tcp_enter_quickack_mode(sk
);
4443 inet_csk_schedule_ack(sk
);
4449 /* Out of window. F.e. zero window probe. */
4450 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4453 tcp_enter_quickack_mode(sk
);
4455 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4456 /* Partial packet, seq < rcv_next < end_seq */
4457 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4458 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4459 TCP_SKB_CB(skb
)->end_seq
);
4461 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4463 /* If window is closed, drop tail of packet. But after
4464 * remembering D-SACK for its head made in previous line.
4466 if (!tcp_receive_window(tp
))
4471 TCP_ECN_check_ce(tp
, skb
);
4473 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4476 /* Disable header prediction. */
4478 inet_csk_schedule_ack(sk
);
4480 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4481 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4483 skb_set_owner_r(skb
, sk
);
4485 if (!skb_peek(&tp
->out_of_order_queue
)) {
4486 /* Initial out of order segment, build 1 SACK. */
4487 if (tcp_is_sack(tp
)) {
4488 tp
->rx_opt
.num_sacks
= 1;
4489 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4490 tp
->selective_acks
[0].end_seq
=
4491 TCP_SKB_CB(skb
)->end_seq
;
4493 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4495 struct sk_buff
*skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4496 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4497 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4499 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4500 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4502 if (!tp
->rx_opt
.num_sacks
||
4503 tp
->selective_acks
[0].end_seq
!= seq
)
4506 /* Common case: data arrive in order after hole. */
4507 tp
->selective_acks
[0].end_seq
= end_seq
;
4511 /* Find place to insert this segment. */
4513 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4515 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4519 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4522 /* Do skb overlap to previous one? */
4523 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4524 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4525 /* All the bits are present. Drop. */
4527 tcp_dsack_set(sk
, seq
, end_seq
);
4530 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4531 /* Partial overlap. */
4532 tcp_dsack_set(sk
, seq
,
4533 TCP_SKB_CB(skb1
)->end_seq
);
4535 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4539 skb1
= skb_queue_prev(
4540 &tp
->out_of_order_queue
,
4545 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4547 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4549 /* And clean segments covered by new one as whole. */
4550 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4551 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4553 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4555 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4556 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4560 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4561 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4562 TCP_SKB_CB(skb1
)->end_seq
);
4567 if (tcp_is_sack(tp
))
4568 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4572 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4573 struct sk_buff_head
*list
)
4575 struct sk_buff
*next
= NULL
;
4577 if (!skb_queue_is_last(list
, skb
))
4578 next
= skb_queue_next(list
, skb
);
4580 __skb_unlink(skb
, list
);
4582 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4587 /* Collapse contiguous sequence of skbs head..tail with
4588 * sequence numbers start..end.
4590 * If tail is NULL, this means until the end of the list.
4592 * Segments with FIN/SYN are not collapsed (only because this
4596 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4597 struct sk_buff
*head
, struct sk_buff
*tail
,
4600 struct sk_buff
*skb
, *n
;
4603 /* First, check that queue is collapsible and find
4604 * the point where collapsing can be useful. */
4608 skb_queue_walk_from_safe(list
, skb
, n
) {
4611 /* No new bits? It is possible on ofo queue. */
4612 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4613 skb
= tcp_collapse_one(sk
, skb
, list
);
4619 /* The first skb to collapse is:
4621 * - bloated or contains data before "start" or
4622 * overlaps to the next one.
4624 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4625 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4626 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4627 end_of_skbs
= false;
4631 if (!skb_queue_is_last(list
, skb
)) {
4632 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4634 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4635 end_of_skbs
= false;
4640 /* Decided to skip this, advance start seq. */
4641 start
= TCP_SKB_CB(skb
)->end_seq
;
4643 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4646 while (before(start
, end
)) {
4647 struct sk_buff
*nskb
;
4648 unsigned int header
= skb_headroom(skb
);
4649 int copy
= SKB_MAX_ORDER(header
, 0);
4651 /* Too big header? This can happen with IPv6. */
4654 if (end
- start
< copy
)
4656 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4660 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4661 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4663 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4665 skb_reserve(nskb
, header
);
4666 memcpy(nskb
->head
, skb
->head
, header
);
4667 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4668 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4669 __skb_queue_before(list
, skb
, nskb
);
4670 skb_set_owner_r(nskb
, sk
);
4672 /* Copy data, releasing collapsed skbs. */
4674 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4675 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4679 size
= min(copy
, size
);
4680 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4682 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4686 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4687 skb
= tcp_collapse_one(sk
, skb
, list
);
4690 tcp_hdr(skb
)->syn
||
4698 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4699 * and tcp_collapse() them until all the queue is collapsed.
4701 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4703 struct tcp_sock
*tp
= tcp_sk(sk
);
4704 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4705 struct sk_buff
*head
;
4711 start
= TCP_SKB_CB(skb
)->seq
;
4712 end
= TCP_SKB_CB(skb
)->end_seq
;
4716 struct sk_buff
*next
= NULL
;
4718 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4719 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4722 /* Segment is terminated when we see gap or when
4723 * we are at the end of all the queue. */
4725 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4726 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4727 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4728 head
, skb
, start
, end
);
4732 /* Start new segment */
4733 start
= TCP_SKB_CB(skb
)->seq
;
4734 end
= TCP_SKB_CB(skb
)->end_seq
;
4736 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4737 start
= TCP_SKB_CB(skb
)->seq
;
4738 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4739 end
= TCP_SKB_CB(skb
)->end_seq
;
4745 * Purge the out-of-order queue.
4746 * Return true if queue was pruned.
4748 static int tcp_prune_ofo_queue(struct sock
*sk
)
4750 struct tcp_sock
*tp
= tcp_sk(sk
);
4753 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4754 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4755 __skb_queue_purge(&tp
->out_of_order_queue
);
4757 /* Reset SACK state. A conforming SACK implementation will
4758 * do the same at a timeout based retransmit. When a connection
4759 * is in a sad state like this, we care only about integrity
4760 * of the connection not performance.
4762 if (tp
->rx_opt
.sack_ok
)
4763 tcp_sack_reset(&tp
->rx_opt
);
4770 /* Reduce allocated memory if we can, trying to get
4771 * the socket within its memory limits again.
4773 * Return less than zero if we should start dropping frames
4774 * until the socket owning process reads some of the data
4775 * to stabilize the situation.
4777 static int tcp_prune_queue(struct sock
*sk
)
4779 struct tcp_sock
*tp
= tcp_sk(sk
);
4781 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4783 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4785 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4786 tcp_clamp_window(sk
);
4787 else if (tcp_memory_pressure
)
4788 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4790 tcp_collapse_ofo_queue(sk
);
4791 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4792 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4793 skb_peek(&sk
->sk_receive_queue
),
4795 tp
->copied_seq
, tp
->rcv_nxt
);
4798 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4801 /* Collapsing did not help, destructive actions follow.
4802 * This must not ever occur. */
4804 tcp_prune_ofo_queue(sk
);
4806 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4809 /* If we are really being abused, tell the caller to silently
4810 * drop receive data on the floor. It will get retransmitted
4811 * and hopefully then we'll have sufficient space.
4813 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4815 /* Massive buffer overcommit. */
4820 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4821 * As additional protections, we do not touch cwnd in retransmission phases,
4822 * and if application hit its sndbuf limit recently.
4824 void tcp_cwnd_application_limited(struct sock
*sk
)
4826 struct tcp_sock
*tp
= tcp_sk(sk
);
4828 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4829 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4830 /* Limited by application or receiver window. */
4831 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4832 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4833 if (win_used
< tp
->snd_cwnd
) {
4834 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4835 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4837 tp
->snd_cwnd_used
= 0;
4839 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4842 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4844 struct tcp_sock
*tp
= tcp_sk(sk
);
4846 /* If the user specified a specific send buffer setting, do
4849 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4852 /* If we are under global TCP memory pressure, do not expand. */
4853 if (tcp_memory_pressure
)
4856 /* If we are under soft global TCP memory pressure, do not expand. */
4857 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4860 /* If we filled the congestion window, do not expand. */
4861 if (tp
->packets_out
>= tp
->snd_cwnd
)
4867 /* When incoming ACK allowed to free some skb from write_queue,
4868 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4869 * on the exit from tcp input handler.
4871 * PROBLEM: sndbuf expansion does not work well with largesend.
4873 static void tcp_new_space(struct sock
*sk
)
4875 struct tcp_sock
*tp
= tcp_sk(sk
);
4877 if (tcp_should_expand_sndbuf(sk
)) {
4878 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4879 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
4880 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4881 tp
->reordering
+ 1);
4882 sndmem
*= 2 * demanded
;
4883 if (sndmem
> sk
->sk_sndbuf
)
4884 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4885 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4888 sk
->sk_write_space(sk
);
4891 static void tcp_check_space(struct sock
*sk
)
4893 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4894 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4895 if (sk
->sk_socket
&&
4896 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4901 static inline void tcp_data_snd_check(struct sock
*sk
)
4903 tcp_push_pending_frames(sk
);
4904 tcp_check_space(sk
);
4908 * Check if sending an ack is needed.
4910 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4912 struct tcp_sock
*tp
= tcp_sk(sk
);
4914 /* More than one full frame received... */
4915 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4916 /* ... and right edge of window advances far enough.
4917 * (tcp_recvmsg() will send ACK otherwise). Or...
4919 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4920 /* We ACK each frame or... */
4921 tcp_in_quickack_mode(sk
) ||
4922 /* We have out of order data. */
4923 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4924 /* Then ack it now */
4927 /* Else, send delayed ack. */
4928 tcp_send_delayed_ack(sk
);
4932 static inline void tcp_ack_snd_check(struct sock
*sk
)
4934 if (!inet_csk_ack_scheduled(sk
)) {
4935 /* We sent a data segment already. */
4938 __tcp_ack_snd_check(sk
, 1);
4942 * This routine is only called when we have urgent data
4943 * signaled. Its the 'slow' part of tcp_urg. It could be
4944 * moved inline now as tcp_urg is only called from one
4945 * place. We handle URGent data wrong. We have to - as
4946 * BSD still doesn't use the correction from RFC961.
4947 * For 1003.1g we should support a new option TCP_STDURG to permit
4948 * either form (or just set the sysctl tcp_stdurg).
4951 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4953 struct tcp_sock
*tp
= tcp_sk(sk
);
4954 u32 ptr
= ntohs(th
->urg_ptr
);
4956 if (ptr
&& !sysctl_tcp_stdurg
)
4958 ptr
+= ntohl(th
->seq
);
4960 /* Ignore urgent data that we've already seen and read. */
4961 if (after(tp
->copied_seq
, ptr
))
4964 /* Do not replay urg ptr.
4966 * NOTE: interesting situation not covered by specs.
4967 * Misbehaving sender may send urg ptr, pointing to segment,
4968 * which we already have in ofo queue. We are not able to fetch
4969 * such data and will stay in TCP_URG_NOTYET until will be eaten
4970 * by recvmsg(). Seems, we are not obliged to handle such wicked
4971 * situations. But it is worth to think about possibility of some
4972 * DoSes using some hypothetical application level deadlock.
4974 if (before(ptr
, tp
->rcv_nxt
))
4977 /* Do we already have a newer (or duplicate) urgent pointer? */
4978 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4981 /* Tell the world about our new urgent pointer. */
4984 /* We may be adding urgent data when the last byte read was
4985 * urgent. To do this requires some care. We cannot just ignore
4986 * tp->copied_seq since we would read the last urgent byte again
4987 * as data, nor can we alter copied_seq until this data arrives
4988 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4990 * NOTE. Double Dutch. Rendering to plain English: author of comment
4991 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4992 * and expect that both A and B disappear from stream. This is _wrong_.
4993 * Though this happens in BSD with high probability, this is occasional.
4994 * Any application relying on this is buggy. Note also, that fix "works"
4995 * only in this artificial test. Insert some normal data between A and B and we will
4996 * decline of BSD again. Verdict: it is better to remove to trap
4999 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5000 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5001 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5003 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5004 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5009 tp
->urg_data
= TCP_URG_NOTYET
;
5012 /* Disable header prediction. */
5016 /* This is the 'fast' part of urgent handling. */
5017 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
5019 struct tcp_sock
*tp
= tcp_sk(sk
);
5021 /* Check if we get a new urgent pointer - normally not. */
5023 tcp_check_urg(sk
, th
);
5025 /* Do we wait for any urgent data? - normally not... */
5026 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5027 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5030 /* Is the urgent pointer pointing into this packet? */
5031 if (ptr
< skb
->len
) {
5033 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5035 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5036 if (!sock_flag(sk
, SOCK_DEAD
))
5037 sk
->sk_data_ready(sk
, 0);
5042 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5044 struct tcp_sock
*tp
= tcp_sk(sk
);
5045 int chunk
= skb
->len
- hlen
;
5049 if (skb_csum_unnecessary(skb
))
5050 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
5052 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
5056 tp
->ucopy
.len
-= chunk
;
5057 tp
->copied_seq
+= chunk
;
5058 tcp_rcv_space_adjust(sk
);
5065 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
5066 struct sk_buff
*skb
)
5070 if (sock_owned_by_user(sk
)) {
5072 result
= __tcp_checksum_complete(skb
);
5075 result
= __tcp_checksum_complete(skb
);
5080 static inline int tcp_checksum_complete_user(struct sock
*sk
,
5081 struct sk_buff
*skb
)
5083 return !skb_csum_unnecessary(skb
) &&
5084 __tcp_checksum_complete_user(sk
, skb
);
5087 #ifdef CONFIG_NET_DMA
5088 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5091 struct tcp_sock
*tp
= tcp_sk(sk
);
5092 int chunk
= skb
->len
- hlen
;
5094 int copied_early
= 0;
5096 if (tp
->ucopy
.wakeup
)
5099 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5100 tp
->ucopy
.dma_chan
= dma_find_channel(DMA_MEMCPY
);
5102 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5104 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5106 tp
->ucopy
.iov
, chunk
,
5107 tp
->ucopy
.pinned_list
);
5112 tp
->ucopy
.dma_cookie
= dma_cookie
;
5115 tp
->ucopy
.len
-= chunk
;
5116 tp
->copied_seq
+= chunk
;
5117 tcp_rcv_space_adjust(sk
);
5119 if ((tp
->ucopy
.len
== 0) ||
5120 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5121 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5122 tp
->ucopy
.wakeup
= 1;
5123 sk
->sk_data_ready(sk
, 0);
5125 } else if (chunk
> 0) {
5126 tp
->ucopy
.wakeup
= 1;
5127 sk
->sk_data_ready(sk
, 0);
5130 return copied_early
;
5132 #endif /* CONFIG_NET_DMA */
5134 /* Does PAWS and seqno based validation of an incoming segment, flags will
5135 * play significant role here.
5137 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5138 struct tcphdr
*th
, int syn_inerr
)
5141 struct tcp_sock
*tp
= tcp_sk(sk
);
5143 /* RFC1323: H1. Apply PAWS check first. */
5144 if (tcp_fast_parse_options(skb
, th
, tp
, &hash_location
) &&
5145 tp
->rx_opt
.saw_tstamp
&&
5146 tcp_paws_discard(sk
, skb
)) {
5148 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5149 tcp_send_dupack(sk
, skb
);
5152 /* Reset is accepted even if it did not pass PAWS. */
5155 /* Step 1: check sequence number */
5156 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5157 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5158 * (RST) segments are validated by checking their SEQ-fields."
5159 * And page 69: "If an incoming segment is not acceptable,
5160 * an acknowledgment should be sent in reply (unless the RST
5161 * bit is set, if so drop the segment and return)".
5164 tcp_send_dupack(sk
, skb
);
5168 /* Step 2: check RST bit */
5174 /* ts_recent update must be made after we are sure that the packet
5177 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5179 /* step 3: check security and precedence [ignored] */
5181 /* step 4: Check for a SYN in window. */
5182 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5184 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5185 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5198 * TCP receive function for the ESTABLISHED state.
5200 * It is split into a fast path and a slow path. The fast path is
5202 * - A zero window was announced from us - zero window probing
5203 * is only handled properly in the slow path.
5204 * - Out of order segments arrived.
5205 * - Urgent data is expected.
5206 * - There is no buffer space left
5207 * - Unexpected TCP flags/window values/header lengths are received
5208 * (detected by checking the TCP header against pred_flags)
5209 * - Data is sent in both directions. Fast path only supports pure senders
5210 * or pure receivers (this means either the sequence number or the ack
5211 * value must stay constant)
5212 * - Unexpected TCP option.
5214 * When these conditions are not satisfied it drops into a standard
5215 * receive procedure patterned after RFC793 to handle all cases.
5216 * The first three cases are guaranteed by proper pred_flags setting,
5217 * the rest is checked inline. Fast processing is turned on in
5218 * tcp_data_queue when everything is OK.
5220 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5221 struct tcphdr
*th
, unsigned len
)
5223 struct tcp_sock
*tp
= tcp_sk(sk
);
5227 * Header prediction.
5228 * The code loosely follows the one in the famous
5229 * "30 instruction TCP receive" Van Jacobson mail.
5231 * Van's trick is to deposit buffers into socket queue
5232 * on a device interrupt, to call tcp_recv function
5233 * on the receive process context and checksum and copy
5234 * the buffer to user space. smart...
5236 * Our current scheme is not silly either but we take the
5237 * extra cost of the net_bh soft interrupt processing...
5238 * We do checksum and copy also but from device to kernel.
5241 tp
->rx_opt
.saw_tstamp
= 0;
5243 /* pred_flags is 0xS?10 << 16 + snd_wnd
5244 * if header_prediction is to be made
5245 * 'S' will always be tp->tcp_header_len >> 2
5246 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5247 * turn it off (when there are holes in the receive
5248 * space for instance)
5249 * PSH flag is ignored.
5252 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5253 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5254 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5255 int tcp_header_len
= tp
->tcp_header_len
;
5257 /* Timestamp header prediction: tcp_header_len
5258 * is automatically equal to th->doff*4 due to pred_flags
5262 /* Check timestamp */
5263 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5264 /* No? Slow path! */
5265 if (!tcp_parse_aligned_timestamp(tp
, th
))
5268 /* If PAWS failed, check it more carefully in slow path */
5269 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5272 /* DO NOT update ts_recent here, if checksum fails
5273 * and timestamp was corrupted part, it will result
5274 * in a hung connection since we will drop all
5275 * future packets due to the PAWS test.
5279 if (len
<= tcp_header_len
) {
5280 /* Bulk data transfer: sender */
5281 if (len
== tcp_header_len
) {
5282 /* Predicted packet is in window by definition.
5283 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5284 * Hence, check seq<=rcv_wup reduces to:
5286 if (tcp_header_len
==
5287 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5288 tp
->rcv_nxt
== tp
->rcv_wup
)
5289 tcp_store_ts_recent(tp
);
5291 /* We know that such packets are checksummed
5294 tcp_ack(sk
, skb
, 0);
5296 tcp_data_snd_check(sk
);
5298 } else { /* Header too small */
5299 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5304 int copied_early
= 0;
5306 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5307 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5308 #ifdef CONFIG_NET_DMA
5309 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5314 if (tp
->ucopy
.task
== current
&&
5315 sock_owned_by_user(sk
) && !copied_early
) {
5316 __set_current_state(TASK_RUNNING
);
5318 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5322 /* Predicted packet is in window by definition.
5323 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5324 * Hence, check seq<=rcv_wup reduces to:
5326 if (tcp_header_len
==
5327 (sizeof(struct tcphdr
) +
5328 TCPOLEN_TSTAMP_ALIGNED
) &&
5329 tp
->rcv_nxt
== tp
->rcv_wup
)
5330 tcp_store_ts_recent(tp
);
5332 tcp_rcv_rtt_measure_ts(sk
, skb
);
5334 __skb_pull(skb
, tcp_header_len
);
5335 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5336 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5339 tcp_cleanup_rbuf(sk
, skb
->len
);
5342 if (tcp_checksum_complete_user(sk
, skb
))
5345 /* Predicted packet is in window by definition.
5346 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5347 * Hence, check seq<=rcv_wup reduces to:
5349 if (tcp_header_len
==
5350 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5351 tp
->rcv_nxt
== tp
->rcv_wup
)
5352 tcp_store_ts_recent(tp
);
5354 tcp_rcv_rtt_measure_ts(sk
, skb
);
5356 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5359 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5361 /* Bulk data transfer: receiver */
5362 __skb_pull(skb
, tcp_header_len
);
5363 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
5364 skb_set_owner_r(skb
, sk
);
5365 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5368 tcp_event_data_recv(sk
, skb
);
5370 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5371 /* Well, only one small jumplet in fast path... */
5372 tcp_ack(sk
, skb
, FLAG_DATA
);
5373 tcp_data_snd_check(sk
);
5374 if (!inet_csk_ack_scheduled(sk
))
5378 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5379 __tcp_ack_snd_check(sk
, 0);
5381 #ifdef CONFIG_NET_DMA
5383 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5389 sk
->sk_data_ready(sk
, 0);
5395 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5399 * Standard slow path.
5402 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
5407 if (th
->ack
&& tcp_ack(sk
, skb
, FLAG_SLOWPATH
) < 0)
5410 tcp_rcv_rtt_measure_ts(sk
, skb
);
5412 /* Process urgent data. */
5413 tcp_urg(sk
, skb
, th
);
5415 /* step 7: process the segment text */
5416 tcp_data_queue(sk
, skb
);
5418 tcp_data_snd_check(sk
);
5419 tcp_ack_snd_check(sk
);
5423 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5429 EXPORT_SYMBOL(tcp_rcv_established
);
5431 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5432 struct tcphdr
*th
, unsigned len
)
5435 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5436 struct tcp_sock
*tp
= tcp_sk(sk
);
5437 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
5438 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5440 tcp_parse_options(skb
, &tp
->rx_opt
, &hash_location
, 0);
5444 * "If the state is SYN-SENT then
5445 * first check the ACK bit
5446 * If the ACK bit is set
5447 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5448 * a reset (unless the RST bit is set, if so drop
5449 * the segment and return)"
5451 * We do not send data with SYN, so that RFC-correct
5454 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5455 goto reset_and_undo
;
5457 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5458 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5460 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5461 goto reset_and_undo
;
5464 /* Now ACK is acceptable.
5466 * "If the RST bit is set
5467 * If the ACK was acceptable then signal the user "error:
5468 * connection reset", drop the segment, enter CLOSED state,
5469 * delete TCB, and return."
5478 * "fifth, if neither of the SYN or RST bits is set then
5479 * drop the segment and return."
5485 goto discard_and_undo
;
5488 * "If the SYN bit is on ...
5489 * are acceptable then ...
5490 * (our SYN has been ACKed), change the connection
5491 * state to ESTABLISHED..."
5494 TCP_ECN_rcv_synack(tp
, th
);
5496 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5497 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5499 /* Ok.. it's good. Set up sequence numbers and
5500 * move to established.
5502 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5503 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5505 /* RFC1323: The window in SYN & SYN/ACK segments is
5508 tp
->snd_wnd
= ntohs(th
->window
);
5509 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5511 if (!tp
->rx_opt
.wscale_ok
) {
5512 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5513 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5516 if (tp
->rx_opt
.saw_tstamp
) {
5517 tp
->rx_opt
.tstamp_ok
= 1;
5518 tp
->tcp_header_len
=
5519 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5520 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5521 tcp_store_ts_recent(tp
);
5523 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5526 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5527 tcp_enable_fack(tp
);
5530 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5531 tcp_initialize_rcv_mss(sk
);
5533 /* Remember, tcp_poll() does not lock socket!
5534 * Change state from SYN-SENT only after copied_seq
5535 * is initialized. */
5536 tp
->copied_seq
= tp
->rcv_nxt
;
5539 cvp
->cookie_pair_size
> 0 &&
5540 tp
->rx_opt
.cookie_plus
> 0) {
5541 int cookie_size
= tp
->rx_opt
.cookie_plus
5542 - TCPOLEN_COOKIE_BASE
;
5543 int cookie_pair_size
= cookie_size
5544 + cvp
->cookie_desired
;
5546 /* A cookie extension option was sent and returned.
5547 * Note that each incoming SYNACK replaces the
5548 * Responder cookie. The initial exchange is most
5549 * fragile, as protection against spoofing relies
5550 * entirely upon the sequence and timestamp (above).
5551 * This replacement strategy allows the correct pair to
5552 * pass through, while any others will be filtered via
5553 * Responder verification later.
5555 if (sizeof(cvp
->cookie_pair
) >= cookie_pair_size
) {
5556 memcpy(&cvp
->cookie_pair
[cvp
->cookie_desired
],
5557 hash_location
, cookie_size
);
5558 cvp
->cookie_pair_size
= cookie_pair_size
;
5563 tcp_set_state(sk
, TCP_ESTABLISHED
);
5565 security_inet_conn_established(sk
, skb
);
5567 /* Make sure socket is routed, for correct metrics. */
5568 icsk
->icsk_af_ops
->rebuild_header(sk
);
5570 tcp_init_metrics(sk
);
5572 tcp_init_congestion_control(sk
);
5574 /* Prevent spurious tcp_cwnd_restart() on first data
5577 tp
->lsndtime
= tcp_time_stamp
;
5579 tcp_init_buffer_space(sk
);
5581 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5582 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5584 if (!tp
->rx_opt
.snd_wscale
)
5585 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5589 if (!sock_flag(sk
, SOCK_DEAD
)) {
5590 sk
->sk_state_change(sk
);
5591 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5594 if (sk
->sk_write_pending
||
5595 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5596 icsk
->icsk_ack
.pingpong
) {
5597 /* Save one ACK. Data will be ready after
5598 * several ticks, if write_pending is set.
5600 * It may be deleted, but with this feature tcpdumps
5601 * look so _wonderfully_ clever, that I was not able
5602 * to stand against the temptation 8) --ANK
5604 inet_csk_schedule_ack(sk
);
5605 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5606 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5607 tcp_incr_quickack(sk
);
5608 tcp_enter_quickack_mode(sk
);
5609 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5610 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5621 /* No ACK in the segment */
5625 * "If the RST bit is set
5627 * Otherwise (no ACK) drop the segment and return."
5630 goto discard_and_undo
;
5634 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5635 tcp_paws_reject(&tp
->rx_opt
, 0))
5636 goto discard_and_undo
;
5639 /* We see SYN without ACK. It is attempt of
5640 * simultaneous connect with crossed SYNs.
5641 * Particularly, it can be connect to self.
5643 tcp_set_state(sk
, TCP_SYN_RECV
);
5645 if (tp
->rx_opt
.saw_tstamp
) {
5646 tp
->rx_opt
.tstamp_ok
= 1;
5647 tcp_store_ts_recent(tp
);
5648 tp
->tcp_header_len
=
5649 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5651 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5654 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5655 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5657 /* RFC1323: The window in SYN & SYN/ACK segments is
5660 tp
->snd_wnd
= ntohs(th
->window
);
5661 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5662 tp
->max_window
= tp
->snd_wnd
;
5664 TCP_ECN_rcv_syn(tp
, th
);
5667 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5668 tcp_initialize_rcv_mss(sk
);
5670 tcp_send_synack(sk
);
5672 /* Note, we could accept data and URG from this segment.
5673 * There are no obstacles to make this.
5675 * However, if we ignore data in ACKless segments sometimes,
5676 * we have no reasons to accept it sometimes.
5677 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5678 * is not flawless. So, discard packet for sanity.
5679 * Uncomment this return to process the data.
5686 /* "fifth, if neither of the SYN or RST bits is set then
5687 * drop the segment and return."
5691 tcp_clear_options(&tp
->rx_opt
);
5692 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5696 tcp_clear_options(&tp
->rx_opt
);
5697 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5702 * This function implements the receiving procedure of RFC 793 for
5703 * all states except ESTABLISHED and TIME_WAIT.
5704 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5705 * address independent.
5708 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5709 struct tcphdr
*th
, unsigned len
)
5711 struct tcp_sock
*tp
= tcp_sk(sk
);
5712 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5716 tp
->rx_opt
.saw_tstamp
= 0;
5718 switch (sk
->sk_state
) {
5730 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5733 /* Now we have several options: In theory there is
5734 * nothing else in the frame. KA9Q has an option to
5735 * send data with the syn, BSD accepts data with the
5736 * syn up to the [to be] advertised window and
5737 * Solaris 2.1 gives you a protocol error. For now
5738 * we just ignore it, that fits the spec precisely
5739 * and avoids incompatibilities. It would be nice in
5740 * future to drop through and process the data.
5742 * Now that TTCP is starting to be used we ought to
5744 * But, this leaves one open to an easy denial of
5745 * service attack, and SYN cookies can't defend
5746 * against this problem. So, we drop the data
5747 * in the interest of security over speed unless
5748 * it's still in use.
5756 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5760 /* Do step6 onward by hand. */
5761 tcp_urg(sk
, skb
, th
);
5763 tcp_data_snd_check(sk
);
5767 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5771 /* step 5: check the ACK field */
5773 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
) > 0;
5775 switch (sk
->sk_state
) {
5778 tp
->copied_seq
= tp
->rcv_nxt
;
5780 tcp_set_state(sk
, TCP_ESTABLISHED
);
5781 sk
->sk_state_change(sk
);
5783 /* Note, that this wakeup is only for marginal
5784 * crossed SYN case. Passively open sockets
5785 * are not waked up, because sk->sk_sleep ==
5786 * NULL and sk->sk_socket == NULL.
5790 SOCK_WAKE_IO
, POLL_OUT
);
5792 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5793 tp
->snd_wnd
= ntohs(th
->window
) <<
5794 tp
->rx_opt
.snd_wscale
;
5795 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5797 /* tcp_ack considers this ACK as duplicate
5798 * and does not calculate rtt.
5801 tcp_ack_update_rtt(sk
, 0, 0);
5803 if (tp
->rx_opt
.tstamp_ok
)
5804 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5806 /* Make sure socket is routed, for
5809 icsk
->icsk_af_ops
->rebuild_header(sk
);
5811 tcp_init_metrics(sk
);
5813 tcp_init_congestion_control(sk
);
5815 /* Prevent spurious tcp_cwnd_restart() on
5816 * first data packet.
5818 tp
->lsndtime
= tcp_time_stamp
;
5821 tcp_initialize_rcv_mss(sk
);
5822 tcp_init_buffer_space(sk
);
5823 tcp_fast_path_on(tp
);
5830 if (tp
->snd_una
== tp
->write_seq
) {
5831 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5832 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5833 dst_confirm(__sk_dst_get(sk
));
5835 if (!sock_flag(sk
, SOCK_DEAD
))
5836 /* Wake up lingering close() */
5837 sk
->sk_state_change(sk
);
5841 if (tp
->linger2
< 0 ||
5842 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5843 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5845 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5849 tmo
= tcp_fin_time(sk
);
5850 if (tmo
> TCP_TIMEWAIT_LEN
) {
5851 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5852 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5853 /* Bad case. We could lose such FIN otherwise.
5854 * It is not a big problem, but it looks confusing
5855 * and not so rare event. We still can lose it now,
5856 * if it spins in bh_lock_sock(), but it is really
5859 inet_csk_reset_keepalive_timer(sk
, tmo
);
5861 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5869 if (tp
->snd_una
== tp
->write_seq
) {
5870 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5876 if (tp
->snd_una
== tp
->write_seq
) {
5877 tcp_update_metrics(sk
);
5886 /* step 6: check the URG bit */
5887 tcp_urg(sk
, skb
, th
);
5889 /* step 7: process the segment text */
5890 switch (sk
->sk_state
) {
5891 case TCP_CLOSE_WAIT
:
5894 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5898 /* RFC 793 says to queue data in these states,
5899 * RFC 1122 says we MUST send a reset.
5900 * BSD 4.4 also does reset.
5902 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5903 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5904 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5905 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5911 case TCP_ESTABLISHED
:
5912 tcp_data_queue(sk
, skb
);
5917 /* tcp_data could move socket to TIME-WAIT */
5918 if (sk
->sk_state
!= TCP_CLOSE
) {
5919 tcp_data_snd_check(sk
);
5920 tcp_ack_snd_check(sk
);
5929 EXPORT_SYMBOL(tcp_rcv_state_process
);