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/module.h>
66 #include <linux/sysctl.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71 #include <asm/unaligned.h>
72 #include <net/netdma.h>
74 int sysctl_tcp_timestamps __read_mostly
= 1;
75 int sysctl_tcp_window_scaling __read_mostly
= 1;
76 int sysctl_tcp_sack __read_mostly
= 1;
77 int sysctl_tcp_fack __read_mostly
= 1;
78 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
79 int sysctl_tcp_ecn __read_mostly
;
80 int sysctl_tcp_dsack __read_mostly
= 1;
81 int sysctl_tcp_app_win __read_mostly
= 31;
82 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
84 int sysctl_tcp_stdurg __read_mostly
;
85 int sysctl_tcp_rfc1337 __read_mostly
;
86 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
87 int sysctl_tcp_frto __read_mostly
= 2;
88 int sysctl_tcp_frto_response __read_mostly
;
89 int sysctl_tcp_nometrics_save __read_mostly
;
91 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
92 int sysctl_tcp_abc __read_mostly
;
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
104 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
105 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
106 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
107 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
118 /* Adapt the MSS value used to make delayed ack decision to the
121 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
123 struct inet_connection_sock
*icsk
= inet_csk(sk
);
124 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
127 icsk
->icsk_ack
.last_seg_size
= 0;
129 /* skb->len may jitter because of SACKs, even if peer
130 * sends good full-sized frames.
132 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
133 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
134 icsk
->icsk_ack
.rcv_mss
= len
;
136 /* Otherwise, we make more careful check taking into account,
137 * that SACKs block is variable.
139 * "len" is invariant segment length, including TCP header.
141 len
+= skb
->data
- skb_transport_header(skb
);
142 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
143 /* If PSH is not set, packet should be
144 * full sized, provided peer TCP is not badly broken.
145 * This observation (if it is correct 8)) allows
146 * to handle super-low mtu links fairly.
148 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
149 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
150 /* Subtract also invariant (if peer is RFC compliant),
151 * tcp header plus fixed timestamp option length.
152 * Resulting "len" is MSS free of SACK jitter.
154 len
-= tcp_sk(sk
)->tcp_header_len
;
155 icsk
->icsk_ack
.last_seg_size
= len
;
157 icsk
->icsk_ack
.rcv_mss
= len
;
161 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
162 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
163 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
167 static void tcp_incr_quickack(struct sock
*sk
)
169 struct inet_connection_sock
*icsk
= inet_csk(sk
);
170 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
174 if (quickacks
> icsk
->icsk_ack
.quick
)
175 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
178 void tcp_enter_quickack_mode(struct sock
*sk
)
180 struct inet_connection_sock
*icsk
= inet_csk(sk
);
181 tcp_incr_quickack(sk
);
182 icsk
->icsk_ack
.pingpong
= 0;
183 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
186 /* Send ACKs quickly, if "quick" count is not exhausted
187 * and the session is not interactive.
190 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
192 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
193 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
196 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
198 if (tp
->ecn_flags
& TCP_ECN_OK
)
199 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
202 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
204 if (tcp_hdr(skb
)->cwr
)
205 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
208 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
210 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
213 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
215 if (tp
->ecn_flags
& TCP_ECN_OK
) {
216 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
217 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
218 /* Funny extension: if ECT is not set on a segment,
219 * it is surely retransmit. It is not in ECN RFC,
220 * but Linux follows this rule. */
221 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
222 tcp_enter_quickack_mode((struct sock
*)tp
);
226 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
228 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
229 tp
->ecn_flags
&= ~TCP_ECN_OK
;
232 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
234 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
235 tp
->ecn_flags
&= ~TCP_ECN_OK
;
238 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
240 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
245 /* Buffer size and advertised window tuning.
247 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
250 static void tcp_fixup_sndbuf(struct sock
*sk
)
252 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
253 sizeof(struct sk_buff
);
255 if (sk
->sk_sndbuf
< 3 * sndmem
)
256 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
259 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261 * All tcp_full_space() is split to two parts: "network" buffer, allocated
262 * forward and advertised in receiver window (tp->rcv_wnd) and
263 * "application buffer", required to isolate scheduling/application
264 * latencies from network.
265 * window_clamp is maximal advertised window. It can be less than
266 * tcp_full_space(), in this case tcp_full_space() - window_clamp
267 * is reserved for "application" buffer. The less window_clamp is
268 * the smoother our behaviour from viewpoint of network, but the lower
269 * throughput and the higher sensitivity of the connection to losses. 8)
271 * rcv_ssthresh is more strict window_clamp used at "slow start"
272 * phase to predict further behaviour of this connection.
273 * It is used for two goals:
274 * - to enforce header prediction at sender, even when application
275 * requires some significant "application buffer". It is check #1.
276 * - to prevent pruning of receive queue because of misprediction
277 * of receiver window. Check #2.
279 * The scheme does not work when sender sends good segments opening
280 * window and then starts to feed us spaghetti. But it should work
281 * in common situations. Otherwise, we have to rely on queue collapsing.
284 /* Slow part of check#2. */
285 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
287 struct tcp_sock
*tp
= tcp_sk(sk
);
289 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
290 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
292 while (tp
->rcv_ssthresh
<= window
) {
293 if (truesize
<= skb
->len
)
294 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
302 static void tcp_grow_window(struct sock
*sk
, struct sk_buff
*skb
)
304 struct tcp_sock
*tp
= tcp_sk(sk
);
307 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
308 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
309 !tcp_memory_pressure
) {
312 /* Check #2. Increase window, if skb with such overhead
313 * will fit to rcvbuf in future.
315 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
316 incr
= 2 * tp
->advmss
;
318 incr
= __tcp_grow_window(sk
, skb
);
321 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
323 inet_csk(sk
)->icsk_ack
.quick
|= 1;
328 /* 3. Tuning rcvbuf, when connection enters established state. */
330 static void tcp_fixup_rcvbuf(struct sock
*sk
)
332 struct tcp_sock
*tp
= tcp_sk(sk
);
333 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
335 /* Try to select rcvbuf so that 4 mss-sized segments
336 * will fit to window and corresponding skbs will fit to our rcvbuf.
337 * (was 3; 4 is minimum to allow fast retransmit to work.)
339 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
341 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
342 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
345 /* 4. Try to fixup all. It is made immediately after connection enters
348 static void tcp_init_buffer_space(struct sock
*sk
)
350 struct tcp_sock
*tp
= tcp_sk(sk
);
353 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
354 tcp_fixup_rcvbuf(sk
);
355 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
356 tcp_fixup_sndbuf(sk
);
358 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
360 maxwin
= tcp_full_space(sk
);
362 if (tp
->window_clamp
>= maxwin
) {
363 tp
->window_clamp
= maxwin
;
365 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
366 tp
->window_clamp
= max(maxwin
-
367 (maxwin
>> sysctl_tcp_app_win
),
371 /* Force reservation of one segment. */
372 if (sysctl_tcp_app_win
&&
373 tp
->window_clamp
> 2 * tp
->advmss
&&
374 tp
->window_clamp
+ tp
->advmss
> maxwin
)
375 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
377 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
378 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
381 /* 5. Recalculate window clamp after socket hit its memory bounds. */
382 static void tcp_clamp_window(struct sock
*sk
)
384 struct tcp_sock
*tp
= tcp_sk(sk
);
385 struct inet_connection_sock
*icsk
= inet_csk(sk
);
387 icsk
->icsk_ack
.quick
= 0;
389 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
390 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
391 !tcp_memory_pressure
&&
392 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
393 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
396 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
397 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
400 /* Initialize RCV_MSS value.
401 * RCV_MSS is an our guess about MSS used by the peer.
402 * We haven't any direct information about the MSS.
403 * It's better to underestimate the RCV_MSS rather than overestimate.
404 * Overestimations make us ACKing less frequently than needed.
405 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407 void tcp_initialize_rcv_mss(struct sock
*sk
)
409 struct tcp_sock
*tp
= tcp_sk(sk
);
410 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
412 hint
= min(hint
, tp
->rcv_wnd
/ 2);
413 hint
= min(hint
, TCP_MIN_RCVMSS
);
414 hint
= max(hint
, TCP_MIN_MSS
);
416 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
419 /* Receiver "autotuning" code.
421 * The algorithm for RTT estimation w/o timestamps is based on
422 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
423 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425 * More detail on this code can be found at
426 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
427 * though this reference is out of date. A new paper
430 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
432 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
438 if (new_sample
!= 0) {
439 /* If we sample in larger samples in the non-timestamp
440 * case, we could grossly overestimate the RTT especially
441 * with chatty applications or bulk transfer apps which
442 * are stalled on filesystem I/O.
444 * Also, since we are only going for a minimum in the
445 * non-timestamp case, we do not smooth things out
446 * else with timestamps disabled convergence takes too
450 m
-= (new_sample
>> 3);
452 } else if (m
< new_sample
)
455 /* No previous measure. */
459 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
460 tp
->rcv_rtt_est
.rtt
= new_sample
;
463 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
465 if (tp
->rcv_rtt_est
.time
== 0)
467 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
469 tcp_rcv_rtt_update(tp
, jiffies
- tp
->rcv_rtt_est
.time
, 1);
472 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
473 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
476 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
477 const struct sk_buff
*skb
)
479 struct tcp_sock
*tp
= tcp_sk(sk
);
480 if (tp
->rx_opt
.rcv_tsecr
&&
481 (TCP_SKB_CB(skb
)->end_seq
-
482 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
483 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
487 * This function should be called every time data is copied to user space.
488 * It calculates the appropriate TCP receive buffer space.
490 void tcp_rcv_space_adjust(struct sock
*sk
)
492 struct tcp_sock
*tp
= tcp_sk(sk
);
496 if (tp
->rcvq_space
.time
== 0)
499 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
500 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
503 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
505 space
= max(tp
->rcvq_space
.space
, space
);
507 if (tp
->rcvq_space
.space
!= space
) {
510 tp
->rcvq_space
.space
= space
;
512 if (sysctl_tcp_moderate_rcvbuf
&&
513 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
514 int new_clamp
= space
;
516 /* Receive space grows, normalize in order to
517 * take into account packet headers and sk_buff
518 * structure overhead.
523 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
524 16 + sizeof(struct sk_buff
));
525 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
528 space
= min(space
, sysctl_tcp_rmem
[2]);
529 if (space
> sk
->sk_rcvbuf
) {
530 sk
->sk_rcvbuf
= space
;
532 /* Make the window clamp follow along. */
533 tp
->window_clamp
= new_clamp
;
539 tp
->rcvq_space
.seq
= tp
->copied_seq
;
540 tp
->rcvq_space
.time
= tcp_time_stamp
;
543 /* There is something which you must keep in mind when you analyze the
544 * behavior of the tp->ato delayed ack timeout interval. When a
545 * connection starts up, we want to ack as quickly as possible. The
546 * problem is that "good" TCP's do slow start at the beginning of data
547 * transmission. The means that until we send the first few ACK's the
548 * sender will sit on his end and only queue most of his data, because
549 * he can only send snd_cwnd unacked packets at any given time. For
550 * each ACK we send, he increments snd_cwnd and transmits more of his
553 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
555 struct tcp_sock
*tp
= tcp_sk(sk
);
556 struct inet_connection_sock
*icsk
= inet_csk(sk
);
559 inet_csk_schedule_ack(sk
);
561 tcp_measure_rcv_mss(sk
, skb
);
563 tcp_rcv_rtt_measure(tp
);
565 now
= tcp_time_stamp
;
567 if (!icsk
->icsk_ack
.ato
) {
568 /* The _first_ data packet received, initialize
569 * delayed ACK engine.
571 tcp_incr_quickack(sk
);
572 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
574 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
576 if (m
<= TCP_ATO_MIN
/ 2) {
577 /* The fastest case is the first. */
578 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
579 } else if (m
< icsk
->icsk_ack
.ato
) {
580 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
581 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
582 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
583 } else if (m
> icsk
->icsk_rto
) {
584 /* Too long gap. Apparently sender failed to
585 * restart window, so that we send ACKs quickly.
587 tcp_incr_quickack(sk
);
591 icsk
->icsk_ack
.lrcvtime
= now
;
593 TCP_ECN_check_ce(tp
, skb
);
596 tcp_grow_window(sk
, skb
);
599 static u32
tcp_rto_min(struct sock
*sk
)
601 struct dst_entry
*dst
= __sk_dst_get(sk
);
602 u32 rto_min
= TCP_RTO_MIN
;
604 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
605 rto_min
= dst_metric_rtt(dst
, RTAX_RTO_MIN
);
609 /* Called to compute a smoothed rtt estimate. The data fed to this
610 * routine either comes from timestamps, or from segments that were
611 * known _not_ to have been retransmitted [see Karn/Partridge
612 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
613 * piece by Van Jacobson.
614 * NOTE: the next three routines used to be one big routine.
615 * To save cycles in the RFC 1323 implementation it was better to break
616 * it up into three procedures. -- erics
618 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
620 struct tcp_sock
*tp
= tcp_sk(sk
);
621 long m
= mrtt
; /* RTT */
623 /* The following amusing code comes from Jacobson's
624 * article in SIGCOMM '88. Note that rtt and mdev
625 * are scaled versions of rtt and mean deviation.
626 * This is designed to be as fast as possible
627 * m stands for "measurement".
629 * On a 1990 paper the rto value is changed to:
630 * RTO = rtt + 4 * mdev
632 * Funny. This algorithm seems to be very broken.
633 * These formulae increase RTO, when it should be decreased, increase
634 * too slowly, when it should be increased quickly, decrease too quickly
635 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
636 * does not matter how to _calculate_ it. Seems, it was trap
637 * that VJ failed to avoid. 8)
642 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
643 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
645 m
= -m
; /* m is now abs(error) */
646 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
647 /* This is similar to one of Eifel findings.
648 * Eifel blocks mdev updates when rtt decreases.
649 * This solution is a bit different: we use finer gain
650 * for mdev in this case (alpha*beta).
651 * Like Eifel it also prevents growth of rto,
652 * but also it limits too fast rto decreases,
653 * happening in pure Eifel.
658 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
660 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
661 if (tp
->mdev
> tp
->mdev_max
) {
662 tp
->mdev_max
= tp
->mdev
;
663 if (tp
->mdev_max
> tp
->rttvar
)
664 tp
->rttvar
= tp
->mdev_max
;
666 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
667 if (tp
->mdev_max
< tp
->rttvar
)
668 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
669 tp
->rtt_seq
= tp
->snd_nxt
;
670 tp
->mdev_max
= tcp_rto_min(sk
);
673 /* no previous measure. */
674 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
675 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
676 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
677 tp
->rtt_seq
= tp
->snd_nxt
;
681 /* Calculate rto without backoff. This is the second half of Van Jacobson's
682 * routine referred to above.
684 static inline void tcp_set_rto(struct sock
*sk
)
686 const struct tcp_sock
*tp
= tcp_sk(sk
);
687 /* Old crap is replaced with new one. 8)
690 * 1. If rtt variance happened to be less 50msec, it is hallucination.
691 * It cannot be less due to utterly erratic ACK generation made
692 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
693 * to do with delayed acks, because at cwnd>2 true delack timeout
694 * is invisible. Actually, Linux-2.4 also generates erratic
695 * ACKs in some circumstances.
697 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
699 /* 2. Fixups made earlier cannot be right.
700 * If we do not estimate RTO correctly without them,
701 * all the algo is pure shit and should be replaced
702 * with correct one. It is exactly, which we pretend to do.
706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707 * guarantees that rto is higher.
709 static inline void tcp_bound_rto(struct sock
*sk
)
711 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
712 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
715 /* Save metrics learned by this TCP session.
716 This function is called only, when TCP finishes successfully
717 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
719 void tcp_update_metrics(struct sock
*sk
)
721 struct tcp_sock
*tp
= tcp_sk(sk
);
722 struct dst_entry
*dst
= __sk_dst_get(sk
);
724 if (sysctl_tcp_nometrics_save
)
729 if (dst
&& (dst
->flags
& DST_HOST
)) {
730 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
734 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
735 /* This session failed to estimate rtt. Why?
736 * Probably, no packets returned in time.
739 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
740 dst
->metrics
[RTAX_RTT
- 1] = 0;
744 rtt
= dst_metric_rtt(dst
, RTAX_RTT
);
747 /* If newly calculated rtt larger than stored one,
748 * store new one. Otherwise, use EWMA. Remember,
749 * rtt overestimation is always better than underestimation.
751 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
753 set_dst_metric_rtt(dst
, RTAX_RTT
, tp
->srtt
);
755 set_dst_metric_rtt(dst
, RTAX_RTT
, rtt
- (m
>> 3));
758 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
763 /* Scale deviation to rttvar fixed point */
768 var
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
772 var
-= (var
- m
) >> 2;
774 set_dst_metric_rtt(dst
, RTAX_RTTVAR
, var
);
777 if (tp
->snd_ssthresh
>= 0xFFFF) {
778 /* Slow start still did not finish. */
779 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
780 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
781 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
782 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
783 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
784 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
785 dst
->metrics
[RTAX_CWND
- 1] = tp
->snd_cwnd
;
786 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
787 icsk
->icsk_ca_state
== TCP_CA_Open
) {
788 /* Cong. avoidance phase, cwnd is reliable. */
789 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
790 dst
->metrics
[RTAX_SSTHRESH
-1] =
791 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
792 if (!dst_metric_locked(dst
, RTAX_CWND
))
793 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_cwnd
) >> 1;
795 /* Else slow start did not finish, cwnd is non-sense,
796 ssthresh may be also invalid.
798 if (!dst_metric_locked(dst
, RTAX_CWND
))
799 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_ssthresh
) >> 1;
800 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
801 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
802 tp
->snd_ssthresh
> dst_metric(dst
, RTAX_SSTHRESH
))
803 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
806 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
807 if (dst_metric(dst
, RTAX_REORDERING
) < tp
->reordering
&&
808 tp
->reordering
!= sysctl_tcp_reordering
)
809 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
814 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
816 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
819 cwnd
= rfc3390_bytes_to_packets(tp
->mss_cache
);
820 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
823 /* Set slow start threshold and cwnd not falling to slow start */
824 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
826 struct tcp_sock
*tp
= tcp_sk(sk
);
827 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
829 tp
->prior_ssthresh
= 0;
831 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
834 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
835 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
836 tcp_packets_in_flight(tp
) + 1U);
837 tp
->snd_cwnd_cnt
= 0;
838 tp
->high_seq
= tp
->snd_nxt
;
839 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
840 TCP_ECN_queue_cwr(tp
);
842 tcp_set_ca_state(sk
, TCP_CA_CWR
);
847 * Packet counting of FACK is based on in-order assumptions, therefore TCP
848 * disables it when reordering is detected
850 static void tcp_disable_fack(struct tcp_sock
*tp
)
852 /* RFC3517 uses different metric in lost marker => reset on change */
854 tp
->lost_skb_hint
= NULL
;
855 tp
->rx_opt
.sack_ok
&= ~2;
858 /* Take a notice that peer is sending D-SACKs */
859 static void tcp_dsack_seen(struct tcp_sock
*tp
)
861 tp
->rx_opt
.sack_ok
|= 4;
864 /* Initialize metrics on socket. */
866 static void tcp_init_metrics(struct sock
*sk
)
868 struct tcp_sock
*tp
= tcp_sk(sk
);
869 struct dst_entry
*dst
= __sk_dst_get(sk
);
876 if (dst_metric_locked(dst
, RTAX_CWND
))
877 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
878 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
879 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
880 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
881 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
883 if (dst_metric(dst
, RTAX_REORDERING
) &&
884 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
885 tcp_disable_fack(tp
);
886 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
889 if (dst_metric(dst
, RTAX_RTT
) == 0)
892 if (!tp
->srtt
&& dst_metric_rtt(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
895 /* Initial rtt is determined from SYN,SYN-ACK.
896 * The segment is small and rtt may appear much
897 * less than real one. Use per-dst memory
898 * to make it more realistic.
900 * A bit of theory. RTT is time passed after "normal" sized packet
901 * is sent until it is ACKed. In normal circumstances sending small
902 * packets force peer to delay ACKs and calculation is correct too.
903 * The algorithm is adaptive and, provided we follow specs, it
904 * NEVER underestimate RTT. BUT! If peer tries to make some clever
905 * tricks sort of "quick acks" for time long enough to decrease RTT
906 * to low value, and then abruptly stops to do it and starts to delay
907 * ACKs, wait for troubles.
909 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
910 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
911 tp
->rtt_seq
= tp
->snd_nxt
;
913 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
914 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
915 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
919 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
921 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
922 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
926 /* Play conservative. If timestamps are not
927 * supported, TCP will fail to recalculate correct
928 * rtt, if initial rto is too small. FORGET ALL AND RESET!
930 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
932 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
933 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
937 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
940 struct tcp_sock
*tp
= tcp_sk(sk
);
941 if (metric
> tp
->reordering
) {
944 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
946 /* This exciting event is worth to be remembered. 8) */
948 mib_idx
= LINUX_MIB_TCPTSREORDER
;
949 else if (tcp_is_reno(tp
))
950 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
951 else if (tcp_is_fack(tp
))
952 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
954 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
956 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
957 #if FASTRETRANS_DEBUG > 1
958 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
959 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
963 tp
->undo_marker
? tp
->undo_retrans
: 0);
965 tcp_disable_fack(tp
);
969 /* This procedure tags the retransmission queue when SACKs arrive.
971 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
972 * Packets in queue with these bits set are counted in variables
973 * sacked_out, retrans_out and lost_out, correspondingly.
975 * Valid combinations are:
976 * Tag InFlight Description
977 * 0 1 - orig segment is in flight.
978 * S 0 - nothing flies, orig reached receiver.
979 * L 0 - nothing flies, orig lost by net.
980 * R 2 - both orig and retransmit are in flight.
981 * L|R 1 - orig is lost, retransmit is in flight.
982 * S|R 1 - orig reached receiver, retrans is still in flight.
983 * (L|S|R is logically valid, it could occur when L|R is sacked,
984 * but it is equivalent to plain S and code short-curcuits it to S.
985 * L|S is logically invalid, it would mean -1 packet in flight 8))
987 * These 6 states form finite state machine, controlled by the following events:
988 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
989 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
990 * 3. Loss detection event of one of three flavors:
991 * A. Scoreboard estimator decided the packet is lost.
992 * A'. Reno "three dupacks" marks head of queue lost.
993 * A''. Its FACK modfication, head until snd.fack is lost.
994 * B. SACK arrives sacking data transmitted after never retransmitted
996 * C. SACK arrives sacking SND.NXT at the moment, when the
997 * segment was retransmitted.
998 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1000 * It is pleasant to note, that state diagram turns out to be commutative,
1001 * so that we are allowed not to be bothered by order of our actions,
1002 * when multiple events arrive simultaneously. (see the function below).
1004 * Reordering detection.
1005 * --------------------
1006 * Reordering metric is maximal distance, which a packet can be displaced
1007 * in packet stream. With SACKs we can estimate it:
1009 * 1. SACK fills old hole and the corresponding segment was not
1010 * ever retransmitted -> reordering. Alas, we cannot use it
1011 * when segment was retransmitted.
1012 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1013 * for retransmitted and already SACKed segment -> reordering..
1014 * Both of these heuristics are not used in Loss state, when we cannot
1015 * account for retransmits accurately.
1017 * SACK block validation.
1018 * ----------------------
1020 * SACK block range validation checks that the received SACK block fits to
1021 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1022 * Note that SND.UNA is not included to the range though being valid because
1023 * it means that the receiver is rather inconsistent with itself reporting
1024 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1025 * perfectly valid, however, in light of RFC2018 which explicitly states
1026 * that "SACK block MUST reflect the newest segment. Even if the newest
1027 * segment is going to be discarded ...", not that it looks very clever
1028 * in case of head skb. Due to potentional receiver driven attacks, we
1029 * choose to avoid immediate execution of a walk in write queue due to
1030 * reneging and defer head skb's loss recovery to standard loss recovery
1031 * procedure that will eventually trigger (nothing forbids us doing this).
1033 * Implements also blockage to start_seq wrap-around. Problem lies in the
1034 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1035 * there's no guarantee that it will be before snd_nxt (n). The problem
1036 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1039 * <- outs wnd -> <- wrapzone ->
1040 * u e n u_w e_w s n_w
1042 * |<------------+------+----- TCP seqno space --------------+---------->|
1043 * ...-- <2^31 ->| |<--------...
1044 * ...---- >2^31 ------>| |<--------...
1046 * Current code wouldn't be vulnerable but it's better still to discard such
1047 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1048 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1049 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1050 * equal to the ideal case (infinite seqno space without wrap caused issues).
1052 * With D-SACK the lower bound is extended to cover sequence space below
1053 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1054 * again, D-SACK block must not to go across snd_una (for the same reason as
1055 * for the normal SACK blocks, explained above). But there all simplicity
1056 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1057 * fully below undo_marker they do not affect behavior in anyway and can
1058 * therefore be safely ignored. In rare cases (which are more or less
1059 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1060 * fragmentation and packet reordering past skb's retransmission. To consider
1061 * them correctly, the acceptable range must be extended even more though
1062 * the exact amount is rather hard to quantify. However, tp->max_window can
1063 * be used as an exaggerated estimate.
1065 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1066 u32 start_seq
, u32 end_seq
)
1068 /* Too far in future, or reversed (interpretation is ambiguous) */
1069 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1072 /* Nasty start_seq wrap-around check (see comments above) */
1073 if (!before(start_seq
, tp
->snd_nxt
))
1076 /* In outstanding window? ...This is valid exit for D-SACKs too.
1077 * start_seq == snd_una is non-sensical (see comments above)
1079 if (after(start_seq
, tp
->snd_una
))
1082 if (!is_dsack
|| !tp
->undo_marker
)
1085 /* ...Then it's D-SACK, and must reside below snd_una completely */
1086 if (!after(end_seq
, tp
->snd_una
))
1089 if (!before(start_seq
, tp
->undo_marker
))
1093 if (!after(end_seq
, tp
->undo_marker
))
1096 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1097 * start_seq < undo_marker and end_seq >= undo_marker.
1099 return !before(start_seq
, end_seq
- tp
->max_window
);
1102 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1103 * Event "C". Later note: FACK people cheated me again 8), we have to account
1104 * for reordering! Ugly, but should help.
1106 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1107 * less than what is now known to be received by the other end (derived from
1108 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1109 * retransmitted skbs to avoid some costly processing per ACKs.
1111 static void tcp_mark_lost_retrans(struct sock
*sk
)
1113 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1114 struct tcp_sock
*tp
= tcp_sk(sk
);
1115 struct sk_buff
*skb
;
1117 u32 new_low_seq
= tp
->snd_nxt
;
1118 u32 received_upto
= tcp_highest_sack_seq(tp
);
1120 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1121 !after(received_upto
, tp
->lost_retrans_low
) ||
1122 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1125 tcp_for_write_queue(skb
, sk
) {
1126 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1128 if (skb
== tcp_send_head(sk
))
1130 if (cnt
== tp
->retrans_out
)
1132 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1135 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1138 if (after(received_upto
, ack_seq
) &&
1140 !before(received_upto
,
1141 ack_seq
+ tp
->reordering
* tp
->mss_cache
))) {
1142 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1143 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1145 /* clear lost hint */
1146 tp
->retransmit_skb_hint
= NULL
;
1148 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1149 tp
->lost_out
+= tcp_skb_pcount(skb
);
1150 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1152 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1154 if (before(ack_seq
, new_low_seq
))
1155 new_low_seq
= ack_seq
;
1156 cnt
+= tcp_skb_pcount(skb
);
1160 if (tp
->retrans_out
)
1161 tp
->lost_retrans_low
= new_low_seq
;
1164 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1165 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1168 struct tcp_sock
*tp
= tcp_sk(sk
);
1169 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1170 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1173 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1176 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1177 } else if (num_sacks
> 1) {
1178 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1179 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1181 if (!after(end_seq_0
, end_seq_1
) &&
1182 !before(start_seq_0
, start_seq_1
)) {
1185 NET_INC_STATS_BH(sock_net(sk
),
1186 LINUX_MIB_TCPDSACKOFORECV
);
1190 /* D-SACK for already forgotten data... Do dumb counting. */
1192 !after(end_seq_0
, prior_snd_una
) &&
1193 after(end_seq_0
, tp
->undo_marker
))
1199 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1200 * the incoming SACK may not exactly match but we can find smaller MSS
1201 * aligned portion of it that matches. Therefore we might need to fragment
1202 * which may fail and creates some hassle (caller must handle error case
1205 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1206 u32 start_seq
, u32 end_seq
)
1209 unsigned int pkt_len
;
1211 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1212 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1214 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1215 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1217 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1220 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1222 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1223 err
= tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
);
1231 static int tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1232 int *reord
, int dup_sack
, int fack_count
)
1234 struct tcp_sock
*tp
= tcp_sk(sk
);
1235 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1238 /* Account D-SACK for retransmitted packet. */
1239 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1240 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1242 if (sacked
& TCPCB_SACKED_ACKED
)
1243 *reord
= min(fack_count
, *reord
);
1246 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1247 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1250 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1251 if (sacked
& TCPCB_SACKED_RETRANS
) {
1252 /* If the segment is not tagged as lost,
1253 * we do not clear RETRANS, believing
1254 * that retransmission is still in flight.
1256 if (sacked
& TCPCB_LOST
) {
1257 TCP_SKB_CB(skb
)->sacked
&=
1258 ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1259 tp
->lost_out
-= tcp_skb_pcount(skb
);
1260 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1262 /* clear lost hint */
1263 tp
->retransmit_skb_hint
= NULL
;
1266 if (!(sacked
& TCPCB_RETRANS
)) {
1267 /* New sack for not retransmitted frame,
1268 * which was in hole. It is reordering.
1270 if (before(TCP_SKB_CB(skb
)->seq
,
1271 tcp_highest_sack_seq(tp
)))
1272 *reord
= min(fack_count
, *reord
);
1274 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1275 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1276 flag
|= FLAG_ONLY_ORIG_SACKED
;
1279 if (sacked
& TCPCB_LOST
) {
1280 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1281 tp
->lost_out
-= tcp_skb_pcount(skb
);
1283 /* clear lost hint */
1284 tp
->retransmit_skb_hint
= NULL
;
1288 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1289 flag
|= FLAG_DATA_SACKED
;
1290 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1292 fack_count
+= tcp_skb_pcount(skb
);
1294 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1295 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1296 before(TCP_SKB_CB(skb
)->seq
,
1297 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1298 tp
->lost_cnt_hint
+= tcp_skb_pcount(skb
);
1300 if (fack_count
> tp
->fackets_out
)
1301 tp
->fackets_out
= fack_count
;
1303 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
1304 tcp_advance_highest_sack(sk
, skb
);
1307 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1308 * frames and clear it. undo_retrans is decreased above, L|R frames
1309 * are accounted above as well.
1311 if (dup_sack
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)) {
1312 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1313 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1314 tp
->retransmit_skb_hint
= NULL
;
1320 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1321 struct tcp_sack_block
*next_dup
,
1322 u32 start_seq
, u32 end_seq
,
1323 int dup_sack_in
, int *fack_count
,
1324 int *reord
, int *flag
)
1326 tcp_for_write_queue_from(skb
, sk
) {
1328 int dup_sack
= dup_sack_in
;
1330 if (skb
== tcp_send_head(sk
))
1333 /* queue is in-order => we can short-circuit the walk early */
1334 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1337 if ((next_dup
!= NULL
) &&
1338 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1339 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1340 next_dup
->start_seq
,
1347 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
,
1349 if (unlikely(in_sack
< 0))
1353 *flag
|= tcp_sacktag_one(skb
, sk
, reord
, dup_sack
,
1356 *fack_count
+= tcp_skb_pcount(skb
);
1361 /* Avoid all extra work that is being done by sacktag while walking in
1364 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1365 u32 skip_to_seq
, int *fack_count
)
1367 tcp_for_write_queue_from(skb
, sk
) {
1368 if (skb
== tcp_send_head(sk
))
1371 if (!before(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1374 *fack_count
+= tcp_skb_pcount(skb
);
1379 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1381 struct tcp_sack_block
*next_dup
,
1383 int *fack_count
, int *reord
,
1386 if (next_dup
== NULL
)
1389 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1390 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
, fack_count
);
1391 skb
= tcp_sacktag_walk(skb
, sk
, NULL
,
1392 next_dup
->start_seq
, next_dup
->end_seq
,
1393 1, fack_count
, reord
, flag
);
1399 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1401 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1405 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1408 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1409 struct tcp_sock
*tp
= tcp_sk(sk
);
1410 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1411 TCP_SKB_CB(ack_skb
)->sacked
);
1412 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1413 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1414 struct tcp_sack_block
*cache
;
1415 struct sk_buff
*skb
;
1416 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1418 int reord
= tp
->packets_out
;
1420 int found_dup_sack
= 0;
1423 int first_sack_index
;
1425 if (!tp
->sacked_out
) {
1426 if (WARN_ON(tp
->fackets_out
))
1427 tp
->fackets_out
= 0;
1428 tcp_highest_sack_reset(sk
);
1431 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1432 num_sacks
, prior_snd_una
);
1434 flag
|= FLAG_DSACKING_ACK
;
1436 /* Eliminate too old ACKs, but take into
1437 * account more or less fresh ones, they can
1438 * contain valid SACK info.
1440 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1443 if (!tp
->packets_out
)
1447 first_sack_index
= 0;
1448 for (i
= 0; i
< num_sacks
; i
++) {
1449 int dup_sack
= !i
&& found_dup_sack
;
1451 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1452 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1454 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1455 sp
[used_sacks
].start_seq
,
1456 sp
[used_sacks
].end_seq
)) {
1460 if (!tp
->undo_marker
)
1461 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1463 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1465 /* Don't count olds caused by ACK reordering */
1466 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1467 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1469 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1472 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1474 first_sack_index
= -1;
1478 /* Ignore very old stuff early */
1479 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1485 /* order SACK blocks to allow in order walk of the retrans queue */
1486 for (i
= used_sacks
- 1; i
> 0; i
--) {
1487 for (j
= 0; j
< i
; j
++) {
1488 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1489 struct tcp_sack_block tmp
;
1495 /* Track where the first SACK block goes to */
1496 if (j
== first_sack_index
)
1497 first_sack_index
= j
+ 1;
1502 skb
= tcp_write_queue_head(sk
);
1506 if (!tp
->sacked_out
) {
1507 /* It's already past, so skip checking against it */
1508 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1510 cache
= tp
->recv_sack_cache
;
1511 /* Skip empty blocks in at head of the cache */
1512 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1517 while (i
< used_sacks
) {
1518 u32 start_seq
= sp
[i
].start_seq
;
1519 u32 end_seq
= sp
[i
].end_seq
;
1520 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1521 struct tcp_sack_block
*next_dup
= NULL
;
1523 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1524 next_dup
= &sp
[i
+ 1];
1526 /* Event "B" in the comment above. */
1527 if (after(end_seq
, tp
->high_seq
))
1528 flag
|= FLAG_DATA_LOST
;
1530 /* Skip too early cached blocks */
1531 while (tcp_sack_cache_ok(tp
, cache
) &&
1532 !before(start_seq
, cache
->end_seq
))
1535 /* Can skip some work by looking recv_sack_cache? */
1536 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1537 after(end_seq
, cache
->start_seq
)) {
1540 if (before(start_seq
, cache
->start_seq
)) {
1541 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
,
1543 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1546 dup_sack
, &fack_count
,
1550 /* Rest of the block already fully processed? */
1551 if (!after(end_seq
, cache
->end_seq
))
1554 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1556 &fack_count
, &reord
,
1559 /* ...tail remains todo... */
1560 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1561 /* ...but better entrypoint exists! */
1562 skb
= tcp_highest_sack(sk
);
1565 fack_count
= tp
->fackets_out
;
1570 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
,
1572 /* Check overlap against next cached too (past this one already) */
1577 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1578 skb
= tcp_highest_sack(sk
);
1581 fack_count
= tp
->fackets_out
;
1583 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
, &fack_count
);
1586 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1587 dup_sack
, &fack_count
, &reord
, &flag
);
1590 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1591 * due to in-order walk
1593 if (after(end_seq
, tp
->frto_highmark
))
1594 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1599 /* Clear the head of the cache sack blocks so we can skip it next time */
1600 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1601 tp
->recv_sack_cache
[i
].start_seq
= 0;
1602 tp
->recv_sack_cache
[i
].end_seq
= 0;
1604 for (j
= 0; j
< used_sacks
; j
++)
1605 tp
->recv_sack_cache
[i
++] = sp
[j
];
1607 tcp_mark_lost_retrans(sk
);
1609 tcp_verify_left_out(tp
);
1611 if ((reord
< tp
->fackets_out
) &&
1612 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1613 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1614 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1618 #if FASTRETRANS_DEBUG > 0
1619 WARN_ON((int)tp
->sacked_out
< 0);
1620 WARN_ON((int)tp
->lost_out
< 0);
1621 WARN_ON((int)tp
->retrans_out
< 0);
1622 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1627 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1628 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1630 int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1634 holes
= max(tp
->lost_out
, 1U);
1635 holes
= min(holes
, tp
->packets_out
);
1637 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1638 tp
->sacked_out
= tp
->packets_out
- holes
;
1644 /* If we receive more dupacks than we expected counting segments
1645 * in assumption of absent reordering, interpret this as reordering.
1646 * The only another reason could be bug in receiver TCP.
1648 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1650 struct tcp_sock
*tp
= tcp_sk(sk
);
1651 if (tcp_limit_reno_sacked(tp
))
1652 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1655 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1657 static void tcp_add_reno_sack(struct sock
*sk
)
1659 struct tcp_sock
*tp
= tcp_sk(sk
);
1661 tcp_check_reno_reordering(sk
, 0);
1662 tcp_verify_left_out(tp
);
1665 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1667 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1669 struct tcp_sock
*tp
= tcp_sk(sk
);
1672 /* One ACK acked hole. The rest eat duplicate ACKs. */
1673 if (acked
- 1 >= tp
->sacked_out
)
1676 tp
->sacked_out
-= acked
- 1;
1678 tcp_check_reno_reordering(sk
, acked
);
1679 tcp_verify_left_out(tp
);
1682 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1687 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
1689 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
1692 /* F-RTO can only be used if TCP has never retransmitted anything other than
1693 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1695 int tcp_use_frto(struct sock
*sk
)
1697 const struct tcp_sock
*tp
= tcp_sk(sk
);
1698 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1699 struct sk_buff
*skb
;
1701 if (!sysctl_tcp_frto
)
1704 /* MTU probe and F-RTO won't really play nicely along currently */
1705 if (icsk
->icsk_mtup
.probe_size
)
1708 if (tcp_is_sackfrto(tp
))
1711 /* Avoid expensive walking of rexmit queue if possible */
1712 if (tp
->retrans_out
> 1)
1715 skb
= tcp_write_queue_head(sk
);
1716 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1717 tcp_for_write_queue_from(skb
, sk
) {
1718 if (skb
== tcp_send_head(sk
))
1720 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1722 /* Short-circuit when first non-SACKed skb has been checked */
1723 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1729 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1730 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1731 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1732 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1733 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1734 * bits are handled if the Loss state is really to be entered (in
1735 * tcp_enter_frto_loss).
1737 * Do like tcp_enter_loss() would; when RTO expires the second time it
1739 * "Reduce ssthresh if it has not yet been made inside this window."
1741 void tcp_enter_frto(struct sock
*sk
)
1743 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1744 struct tcp_sock
*tp
= tcp_sk(sk
);
1745 struct sk_buff
*skb
;
1747 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1748 tp
->snd_una
== tp
->high_seq
||
1749 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1750 !icsk
->icsk_retransmits
)) {
1751 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1752 /* Our state is too optimistic in ssthresh() call because cwnd
1753 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1754 * recovery has not yet completed. Pattern would be this: RTO,
1755 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1757 * RFC4138 should be more specific on what to do, even though
1758 * RTO is quite unlikely to occur after the first Cumulative ACK
1759 * due to back-off and complexity of triggering events ...
1761 if (tp
->frto_counter
) {
1763 stored_cwnd
= tp
->snd_cwnd
;
1765 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1766 tp
->snd_cwnd
= stored_cwnd
;
1768 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1770 /* ... in theory, cong.control module could do "any tricks" in
1771 * ssthresh(), which means that ca_state, lost bits and lost_out
1772 * counter would have to be faked before the call occurs. We
1773 * consider that too expensive, unlikely and hacky, so modules
1774 * using these in ssthresh() must deal these incompatibility
1775 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1777 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1780 tp
->undo_marker
= tp
->snd_una
;
1781 tp
->undo_retrans
= 0;
1783 skb
= tcp_write_queue_head(sk
);
1784 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1785 tp
->undo_marker
= 0;
1786 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1787 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1788 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1790 tcp_verify_left_out(tp
);
1792 /* Too bad if TCP was application limited */
1793 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1795 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1796 * The last condition is necessary at least in tp->frto_counter case.
1798 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
1799 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1800 after(tp
->high_seq
, tp
->snd_una
)) {
1801 tp
->frto_highmark
= tp
->high_seq
;
1803 tp
->frto_highmark
= tp
->snd_nxt
;
1805 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1806 tp
->high_seq
= tp
->snd_nxt
;
1807 tp
->frto_counter
= 1;
1810 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1811 * which indicates that we should follow the traditional RTO recovery,
1812 * i.e. mark everything lost and do go-back-N retransmission.
1814 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1816 struct tcp_sock
*tp
= tcp_sk(sk
);
1817 struct sk_buff
*skb
;
1820 tp
->retrans_out
= 0;
1821 if (tcp_is_reno(tp
))
1822 tcp_reset_reno_sack(tp
);
1824 tcp_for_write_queue(skb
, sk
) {
1825 if (skb
== tcp_send_head(sk
))
1828 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1830 * Count the retransmission made on RTO correctly (only when
1831 * waiting for the first ACK and did not get it)...
1833 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
1834 /* For some reason this R-bit might get cleared? */
1835 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1836 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1837 /* ...enter this if branch just for the first segment */
1838 flag
|= FLAG_DATA_ACKED
;
1840 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1841 tp
->undo_marker
= 0;
1842 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1845 /* Marking forward transmissions that were made after RTO lost
1846 * can cause unnecessary retransmissions in some scenarios,
1847 * SACK blocks will mitigate that in some but not in all cases.
1848 * We used to not mark them but it was causing break-ups with
1849 * receivers that do only in-order receival.
1851 * TODO: we could detect presence of such receiver and select
1852 * different behavior per flow.
1854 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1855 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1856 tp
->lost_out
+= tcp_skb_pcount(skb
);
1859 tcp_verify_left_out(tp
);
1861 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1862 tp
->snd_cwnd_cnt
= 0;
1863 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1864 tp
->frto_counter
= 0;
1865 tp
->bytes_acked
= 0;
1867 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1868 sysctl_tcp_reordering
);
1869 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1870 tp
->high_seq
= tp
->snd_nxt
;
1871 TCP_ECN_queue_cwr(tp
);
1873 tcp_clear_retrans_hints_partial(tp
);
1876 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1878 tp
->retrans_out
= 0;
1881 tp
->undo_marker
= 0;
1882 tp
->undo_retrans
= 0;
1885 void tcp_clear_retrans(struct tcp_sock
*tp
)
1887 tcp_clear_retrans_partial(tp
);
1889 tp
->fackets_out
= 0;
1893 /* Enter Loss state. If "how" is not zero, forget all SACK information
1894 * and reset tags completely, otherwise preserve SACKs. If receiver
1895 * dropped its ofo queue, we will know this due to reneging detection.
1897 void tcp_enter_loss(struct sock
*sk
, int how
)
1899 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1900 struct tcp_sock
*tp
= tcp_sk(sk
);
1901 struct sk_buff
*skb
;
1903 /* Reduce ssthresh if it has not yet been made inside this window. */
1904 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1905 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1906 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1907 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1908 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1911 tp
->snd_cwnd_cnt
= 0;
1912 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1914 tp
->bytes_acked
= 0;
1915 tcp_clear_retrans_partial(tp
);
1917 if (tcp_is_reno(tp
))
1918 tcp_reset_reno_sack(tp
);
1921 /* Push undo marker, if it was plain RTO and nothing
1922 * was retransmitted. */
1923 tp
->undo_marker
= tp
->snd_una
;
1924 tcp_clear_retrans_hints_partial(tp
);
1927 tp
->fackets_out
= 0;
1928 tcp_clear_all_retrans_hints(tp
);
1931 tcp_for_write_queue(skb
, sk
) {
1932 if (skb
== tcp_send_head(sk
))
1935 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1936 tp
->undo_marker
= 0;
1937 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1938 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1939 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1940 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1941 tp
->lost_out
+= tcp_skb_pcount(skb
);
1944 tcp_verify_left_out(tp
);
1946 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1947 sysctl_tcp_reordering
);
1948 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1949 tp
->high_seq
= tp
->snd_nxt
;
1950 TCP_ECN_queue_cwr(tp
);
1951 /* Abort F-RTO algorithm if one is in progress */
1952 tp
->frto_counter
= 0;
1955 /* If ACK arrived pointing to a remembered SACK, it means that our
1956 * remembered SACKs do not reflect real state of receiver i.e.
1957 * receiver _host_ is heavily congested (or buggy).
1959 * Do processing similar to RTO timeout.
1961 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1963 if (flag
& FLAG_SACK_RENEGING
) {
1964 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1965 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1967 tcp_enter_loss(sk
, 1);
1968 icsk
->icsk_retransmits
++;
1969 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1970 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1971 icsk
->icsk_rto
, TCP_RTO_MAX
);
1977 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1979 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1982 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1983 * counter when SACK is enabled (without SACK, sacked_out is used for
1986 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1987 * segments up to the highest received SACK block so far and holes in
1990 * With reordering, holes may still be in flight, so RFC3517 recovery
1991 * uses pure sacked_out (total number of SACKed segments) even though
1992 * it violates the RFC that uses duplicate ACKs, often these are equal
1993 * but when e.g. out-of-window ACKs or packet duplication occurs,
1994 * they differ. Since neither occurs due to loss, TCP should really
1997 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
1999 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2002 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2004 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2007 static inline int tcp_head_timedout(struct sock
*sk
)
2009 struct tcp_sock
*tp
= tcp_sk(sk
);
2011 return tp
->packets_out
&&
2012 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2015 /* Linux NewReno/SACK/FACK/ECN state machine.
2016 * --------------------------------------
2018 * "Open" Normal state, no dubious events, fast path.
2019 * "Disorder" In all the respects it is "Open",
2020 * but requires a bit more attention. It is entered when
2021 * we see some SACKs or dupacks. It is split of "Open"
2022 * mainly to move some processing from fast path to slow one.
2023 * "CWR" CWND was reduced due to some Congestion Notification event.
2024 * It can be ECN, ICMP source quench, local device congestion.
2025 * "Recovery" CWND was reduced, we are fast-retransmitting.
2026 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2028 * tcp_fastretrans_alert() is entered:
2029 * - each incoming ACK, if state is not "Open"
2030 * - when arrived ACK is unusual, namely:
2035 * Counting packets in flight is pretty simple.
2037 * in_flight = packets_out - left_out + retrans_out
2039 * packets_out is SND.NXT-SND.UNA counted in packets.
2041 * retrans_out is number of retransmitted segments.
2043 * left_out is number of segments left network, but not ACKed yet.
2045 * left_out = sacked_out + lost_out
2047 * sacked_out: Packets, which arrived to receiver out of order
2048 * and hence not ACKed. With SACKs this number is simply
2049 * amount of SACKed data. Even without SACKs
2050 * it is easy to give pretty reliable estimate of this number,
2051 * counting duplicate ACKs.
2053 * lost_out: Packets lost by network. TCP has no explicit
2054 * "loss notification" feedback from network (for now).
2055 * It means that this number can be only _guessed_.
2056 * Actually, it is the heuristics to predict lossage that
2057 * distinguishes different algorithms.
2059 * F.e. after RTO, when all the queue is considered as lost,
2060 * lost_out = packets_out and in_flight = retrans_out.
2062 * Essentially, we have now two algorithms counting
2065 * FACK: It is the simplest heuristics. As soon as we decided
2066 * that something is lost, we decide that _all_ not SACKed
2067 * packets until the most forward SACK are lost. I.e.
2068 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2069 * It is absolutely correct estimate, if network does not reorder
2070 * packets. And it loses any connection to reality when reordering
2071 * takes place. We use FACK by default until reordering
2072 * is suspected on the path to this destination.
2074 * NewReno: when Recovery is entered, we assume that one segment
2075 * is lost (classic Reno). While we are in Recovery and
2076 * a partial ACK arrives, we assume that one more packet
2077 * is lost (NewReno). This heuristics are the same in NewReno
2080 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2081 * deflation etc. CWND is real congestion window, never inflated, changes
2082 * only according to classic VJ rules.
2084 * Really tricky (and requiring careful tuning) part of algorithm
2085 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2086 * The first determines the moment _when_ we should reduce CWND and,
2087 * hence, slow down forward transmission. In fact, it determines the moment
2088 * when we decide that hole is caused by loss, rather than by a reorder.
2090 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2091 * holes, caused by lost packets.
2093 * And the most logically complicated part of algorithm is undo
2094 * heuristics. We detect false retransmits due to both too early
2095 * fast retransmit (reordering) and underestimated RTO, analyzing
2096 * timestamps and D-SACKs. When we detect that some segments were
2097 * retransmitted by mistake and CWND reduction was wrong, we undo
2098 * window reduction and abort recovery phase. This logic is hidden
2099 * inside several functions named tcp_try_undo_<something>.
2102 /* This function decides, when we should leave Disordered state
2103 * and enter Recovery phase, reducing congestion window.
2105 * Main question: may we further continue forward transmission
2106 * with the same cwnd?
2108 static int tcp_time_to_recover(struct sock
*sk
)
2110 struct tcp_sock
*tp
= tcp_sk(sk
);
2113 /* Do not perform any recovery during F-RTO algorithm */
2114 if (tp
->frto_counter
)
2117 /* Trick#1: The loss is proven. */
2121 /* Not-A-Trick#2 : Classic rule... */
2122 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2125 /* Trick#3 : when we use RFC2988 timer restart, fast
2126 * retransmit can be triggered by timeout of queue head.
2128 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2131 /* Trick#4: It is still not OK... But will it be useful to delay
2134 packets_out
= tp
->packets_out
;
2135 if (packets_out
<= tp
->reordering
&&
2136 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2137 !tcp_may_send_now(sk
)) {
2138 /* We have nothing to send. This connection is limited
2139 * either by receiver window or by application.
2147 /* RFC: This is from the original, I doubt that this is necessary at all:
2148 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2149 * retransmitted past LOST markings in the first place? I'm not fully sure
2150 * about undo and end of connection cases, which can cause R without L?
2152 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
2154 if ((tp
->retransmit_skb_hint
!= NULL
) &&
2155 before(TCP_SKB_CB(skb
)->seq
,
2156 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
2157 tp
->retransmit_skb_hint
= NULL
;
2160 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2161 * is against sacked "cnt", otherwise it's against facked "cnt"
2163 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2165 struct tcp_sock
*tp
= tcp_sk(sk
);
2166 struct sk_buff
*skb
;
2171 WARN_ON(packets
> tp
->packets_out
);
2172 if (tp
->lost_skb_hint
) {
2173 skb
= tp
->lost_skb_hint
;
2174 cnt
= tp
->lost_cnt_hint
;
2176 skb
= tcp_write_queue_head(sk
);
2180 tcp_for_write_queue_from(skb
, sk
) {
2181 if (skb
== tcp_send_head(sk
))
2183 /* TODO: do this better */
2184 /* this is not the most efficient way to do this... */
2185 tp
->lost_skb_hint
= skb
;
2186 tp
->lost_cnt_hint
= cnt
;
2188 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2192 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2193 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2194 cnt
+= tcp_skb_pcount(skb
);
2196 if (cnt
> packets
) {
2197 if (tcp_is_sack(tp
) || (oldcnt
>= packets
))
2200 mss
= skb_shinfo(skb
)->gso_size
;
2201 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2207 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2208 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2209 tp
->lost_out
+= tcp_skb_pcount(skb
);
2210 tcp_verify_retransmit_hint(tp
, skb
);
2213 tcp_verify_left_out(tp
);
2216 /* Account newly detected lost packet(s) */
2218 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2220 struct tcp_sock
*tp
= tcp_sk(sk
);
2222 if (tcp_is_reno(tp
)) {
2223 tcp_mark_head_lost(sk
, 1);
2224 } else if (tcp_is_fack(tp
)) {
2225 int lost
= tp
->fackets_out
- tp
->reordering
;
2228 tcp_mark_head_lost(sk
, lost
);
2230 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2231 if (sacked_upto
< fast_rexmit
)
2232 sacked_upto
= fast_rexmit
;
2233 tcp_mark_head_lost(sk
, sacked_upto
);
2236 /* New heuristics: it is possible only after we switched
2237 * to restart timer each time when something is ACKed.
2238 * Hence, we can detect timed out packets during fast
2239 * retransmit without falling to slow start.
2241 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2242 struct sk_buff
*skb
;
2244 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2245 : tcp_write_queue_head(sk
);
2247 tcp_for_write_queue_from(skb
, sk
) {
2248 if (skb
== tcp_send_head(sk
))
2250 if (!tcp_skb_timedout(sk
, skb
))
2253 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2254 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2255 tp
->lost_out
+= tcp_skb_pcount(skb
);
2256 tcp_verify_retransmit_hint(tp
, skb
);
2260 tp
->scoreboard_skb_hint
= skb
;
2262 tcp_verify_left_out(tp
);
2266 /* CWND moderation, preventing bursts due to too big ACKs
2267 * in dubious situations.
2269 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2271 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2272 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2273 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2276 /* Lower bound on congestion window is slow start threshold
2277 * unless congestion avoidance choice decides to overide it.
2279 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2281 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2283 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2286 /* Decrease cwnd each second ack. */
2287 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2289 struct tcp_sock
*tp
= tcp_sk(sk
);
2290 int decr
= tp
->snd_cwnd_cnt
+ 1;
2292 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2293 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2294 tp
->snd_cwnd_cnt
= decr
& 1;
2297 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2298 tp
->snd_cwnd
-= decr
;
2300 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2301 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2305 /* Nothing was retransmitted or returned timestamp is less
2306 * than timestamp of the first retransmission.
2308 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2310 return !tp
->retrans_stamp
||
2311 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2312 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2315 /* Undo procedures. */
2317 #if FASTRETRANS_DEBUG > 1
2318 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2320 struct tcp_sock
*tp
= tcp_sk(sk
);
2321 struct inet_sock
*inet
= inet_sk(sk
);
2323 if (sk
->sk_family
== AF_INET
) {
2324 printk(KERN_DEBUG
"Undo %s " NIPQUAD_FMT
"/%u c%u l%u ss%u/%u p%u\n",
2326 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2327 tp
->snd_cwnd
, tcp_left_out(tp
),
2328 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2331 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2332 else if (sk
->sk_family
== AF_INET6
) {
2333 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2334 printk(KERN_DEBUG
"Undo %s " NIP6_FMT
"/%u c%u l%u ss%u/%u p%u\n",
2336 NIP6(np
->daddr
), ntohs(inet
->dport
),
2337 tp
->snd_cwnd
, tcp_left_out(tp
),
2338 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2344 #define DBGUNDO(x...) do { } while (0)
2347 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2349 struct tcp_sock
*tp
= tcp_sk(sk
);
2351 if (tp
->prior_ssthresh
) {
2352 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2354 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2355 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2357 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2359 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2360 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2361 TCP_ECN_withdraw_cwr(tp
);
2364 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2366 tcp_moderate_cwnd(tp
);
2367 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2369 /* There is something screwy going on with the retrans hints after
2371 tcp_clear_all_retrans_hints(tp
);
2374 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2376 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2379 /* People celebrate: "We love our President!" */
2380 static int tcp_try_undo_recovery(struct sock
*sk
)
2382 struct tcp_sock
*tp
= tcp_sk(sk
);
2384 if (tcp_may_undo(tp
)) {
2387 /* Happy end! We did not retransmit anything
2388 * or our original transmission succeeded.
2390 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2391 tcp_undo_cwr(sk
, 1);
2392 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2393 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2395 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2397 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2398 tp
->undo_marker
= 0;
2400 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2401 /* Hold old state until something *above* high_seq
2402 * is ACKed. For Reno it is MUST to prevent false
2403 * fast retransmits (RFC2582). SACK TCP is safe. */
2404 tcp_moderate_cwnd(tp
);
2407 tcp_set_ca_state(sk
, TCP_CA_Open
);
2411 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2412 static void tcp_try_undo_dsack(struct sock
*sk
)
2414 struct tcp_sock
*tp
= tcp_sk(sk
);
2416 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2417 DBGUNDO(sk
, "D-SACK");
2418 tcp_undo_cwr(sk
, 1);
2419 tp
->undo_marker
= 0;
2420 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2424 /* Undo during fast recovery after partial ACK. */
2426 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2428 struct tcp_sock
*tp
= tcp_sk(sk
);
2429 /* Partial ACK arrived. Force Hoe's retransmit. */
2430 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2432 if (tcp_may_undo(tp
)) {
2433 /* Plain luck! Hole if filled with delayed
2434 * packet, rather than with a retransmit.
2436 if (tp
->retrans_out
== 0)
2437 tp
->retrans_stamp
= 0;
2439 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2442 tcp_undo_cwr(sk
, 0);
2443 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2445 /* So... Do not make Hoe's retransmit yet.
2446 * If the first packet was delayed, the rest
2447 * ones are most probably delayed as well.
2454 /* Undo during loss recovery after partial ACK. */
2455 static int tcp_try_undo_loss(struct sock
*sk
)
2457 struct tcp_sock
*tp
= tcp_sk(sk
);
2459 if (tcp_may_undo(tp
)) {
2460 struct sk_buff
*skb
;
2461 tcp_for_write_queue(skb
, sk
) {
2462 if (skb
== tcp_send_head(sk
))
2464 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2467 tcp_clear_all_retrans_hints(tp
);
2469 DBGUNDO(sk
, "partial loss");
2471 tcp_undo_cwr(sk
, 1);
2472 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2473 inet_csk(sk
)->icsk_retransmits
= 0;
2474 tp
->undo_marker
= 0;
2475 if (tcp_is_sack(tp
))
2476 tcp_set_ca_state(sk
, TCP_CA_Open
);
2482 static inline void tcp_complete_cwr(struct sock
*sk
)
2484 struct tcp_sock
*tp
= tcp_sk(sk
);
2485 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2486 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2487 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2490 static void tcp_try_keep_open(struct sock
*sk
)
2492 struct tcp_sock
*tp
= tcp_sk(sk
);
2493 int state
= TCP_CA_Open
;
2495 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2496 state
= TCP_CA_Disorder
;
2498 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2499 tcp_set_ca_state(sk
, state
);
2500 tp
->high_seq
= tp
->snd_nxt
;
2504 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2506 struct tcp_sock
*tp
= tcp_sk(sk
);
2508 tcp_verify_left_out(tp
);
2510 if (!tp
->frto_counter
&& tp
->retrans_out
== 0)
2511 tp
->retrans_stamp
= 0;
2513 if (flag
& FLAG_ECE
)
2514 tcp_enter_cwr(sk
, 1);
2516 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2517 tcp_try_keep_open(sk
);
2518 tcp_moderate_cwnd(tp
);
2520 tcp_cwnd_down(sk
, flag
);
2524 static void tcp_mtup_probe_failed(struct sock
*sk
)
2526 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2528 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2529 icsk
->icsk_mtup
.probe_size
= 0;
2532 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2534 struct tcp_sock
*tp
= tcp_sk(sk
);
2535 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2537 /* FIXME: breaks with very large cwnd */
2538 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2539 tp
->snd_cwnd
= tp
->snd_cwnd
*
2540 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2541 icsk
->icsk_mtup
.probe_size
;
2542 tp
->snd_cwnd_cnt
= 0;
2543 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2544 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2546 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2547 icsk
->icsk_mtup
.probe_size
= 0;
2548 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2551 /* Process an event, which can update packets-in-flight not trivially.
2552 * Main goal of this function is to calculate new estimate for left_out,
2553 * taking into account both packets sitting in receiver's buffer and
2554 * packets lost by network.
2556 * Besides that it does CWND reduction, when packet loss is detected
2557 * and changes state of machine.
2559 * It does _not_ decide what to send, it is made in function
2560 * tcp_xmit_retransmit_queue().
2562 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2564 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2565 struct tcp_sock
*tp
= tcp_sk(sk
);
2566 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2567 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2568 (tcp_fackets_out(tp
) > tp
->reordering
));
2569 int fast_rexmit
= 0, mib_idx
;
2571 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2573 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2574 tp
->fackets_out
= 0;
2576 /* Now state machine starts.
2577 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2578 if (flag
& FLAG_ECE
)
2579 tp
->prior_ssthresh
= 0;
2581 /* B. In all the states check for reneging SACKs. */
2582 if (tcp_check_sack_reneging(sk
, flag
))
2585 /* C. Process data loss notification, provided it is valid. */
2586 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2587 before(tp
->snd_una
, tp
->high_seq
) &&
2588 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2589 tp
->fackets_out
> tp
->reordering
) {
2590 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2591 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2594 /* D. Check consistency of the current state. */
2595 tcp_verify_left_out(tp
);
2597 /* E. Check state exit conditions. State can be terminated
2598 * when high_seq is ACKed. */
2599 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2600 WARN_ON(tp
->retrans_out
!= 0);
2601 tp
->retrans_stamp
= 0;
2602 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2603 switch (icsk
->icsk_ca_state
) {
2605 icsk
->icsk_retransmits
= 0;
2606 if (tcp_try_undo_recovery(sk
))
2611 /* CWR is to be held something *above* high_seq
2612 * is ACKed for CWR bit to reach receiver. */
2613 if (tp
->snd_una
!= tp
->high_seq
) {
2614 tcp_complete_cwr(sk
);
2615 tcp_set_ca_state(sk
, TCP_CA_Open
);
2619 case TCP_CA_Disorder
:
2620 tcp_try_undo_dsack(sk
);
2621 if (!tp
->undo_marker
||
2622 /* For SACK case do not Open to allow to undo
2623 * catching for all duplicate ACKs. */
2624 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2625 tp
->undo_marker
= 0;
2626 tcp_set_ca_state(sk
, TCP_CA_Open
);
2630 case TCP_CA_Recovery
:
2631 if (tcp_is_reno(tp
))
2632 tcp_reset_reno_sack(tp
);
2633 if (tcp_try_undo_recovery(sk
))
2635 tcp_complete_cwr(sk
);
2640 /* F. Process state. */
2641 switch (icsk
->icsk_ca_state
) {
2642 case TCP_CA_Recovery
:
2643 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2644 if (tcp_is_reno(tp
) && is_dupack
)
2645 tcp_add_reno_sack(sk
);
2647 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2650 if (flag
& FLAG_DATA_ACKED
)
2651 icsk
->icsk_retransmits
= 0;
2652 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
2653 tcp_reset_reno_sack(tp
);
2654 if (!tcp_try_undo_loss(sk
)) {
2655 tcp_moderate_cwnd(tp
);
2656 tcp_xmit_retransmit_queue(sk
);
2659 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2661 /* Loss is undone; fall through to processing in Open state. */
2663 if (tcp_is_reno(tp
)) {
2664 if (flag
& FLAG_SND_UNA_ADVANCED
)
2665 tcp_reset_reno_sack(tp
);
2667 tcp_add_reno_sack(sk
);
2670 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2671 tcp_try_undo_dsack(sk
);
2673 if (!tcp_time_to_recover(sk
)) {
2674 tcp_try_to_open(sk
, flag
);
2678 /* MTU probe failure: don't reduce cwnd */
2679 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2680 icsk
->icsk_mtup
.probe_size
&&
2681 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2682 tcp_mtup_probe_failed(sk
);
2683 /* Restores the reduction we did in tcp_mtup_probe() */
2685 tcp_simple_retransmit(sk
);
2689 /* Otherwise enter Recovery state */
2691 if (tcp_is_reno(tp
))
2692 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2694 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2696 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2698 tp
->high_seq
= tp
->snd_nxt
;
2699 tp
->prior_ssthresh
= 0;
2700 tp
->undo_marker
= tp
->snd_una
;
2701 tp
->undo_retrans
= tp
->retrans_out
;
2703 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2704 if (!(flag
& FLAG_ECE
))
2705 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2706 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2707 TCP_ECN_queue_cwr(tp
);
2710 tp
->bytes_acked
= 0;
2711 tp
->snd_cwnd_cnt
= 0;
2712 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2716 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2717 tcp_update_scoreboard(sk
, fast_rexmit
);
2718 tcp_cwnd_down(sk
, flag
);
2719 tcp_xmit_retransmit_queue(sk
);
2722 /* Read draft-ietf-tcplw-high-performance before mucking
2723 * with this code. (Supersedes RFC1323)
2725 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2727 /* RTTM Rule: A TSecr value received in a segment is used to
2728 * update the averaged RTT measurement only if the segment
2729 * acknowledges some new data, i.e., only if it advances the
2730 * left edge of the send window.
2732 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2733 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2735 * Changed: reset backoff as soon as we see the first valid sample.
2736 * If we do not, we get strongly overestimated rto. With timestamps
2737 * samples are accepted even from very old segments: f.e., when rtt=1
2738 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2739 * answer arrives rto becomes 120 seconds! If at least one of segments
2740 * in window is lost... Voila. --ANK (010210)
2742 struct tcp_sock
*tp
= tcp_sk(sk
);
2743 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2744 tcp_rtt_estimator(sk
, seq_rtt
);
2746 inet_csk(sk
)->icsk_backoff
= 0;
2750 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2752 /* We don't have a timestamp. Can only use
2753 * packets that are not retransmitted to determine
2754 * rtt estimates. Also, we must not reset the
2755 * backoff for rto until we get a non-retransmitted
2756 * packet. This allows us to deal with a situation
2757 * where the network delay has increased suddenly.
2758 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2761 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2764 tcp_rtt_estimator(sk
, seq_rtt
);
2766 inet_csk(sk
)->icsk_backoff
= 0;
2770 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2773 const struct tcp_sock
*tp
= tcp_sk(sk
);
2774 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2775 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2776 tcp_ack_saw_tstamp(sk
, flag
);
2777 else if (seq_rtt
>= 0)
2778 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2781 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2783 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2784 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2785 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2788 /* Restart timer after forward progress on connection.
2789 * RFC2988 recommends to restart timer to now+rto.
2791 static void tcp_rearm_rto(struct sock
*sk
)
2793 struct tcp_sock
*tp
= tcp_sk(sk
);
2795 if (!tp
->packets_out
) {
2796 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2798 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2799 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2803 /* If we get here, the whole TSO packet has not been acked. */
2804 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2806 struct tcp_sock
*tp
= tcp_sk(sk
);
2809 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2811 packets_acked
= tcp_skb_pcount(skb
);
2812 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2814 packets_acked
-= tcp_skb_pcount(skb
);
2816 if (packets_acked
) {
2817 BUG_ON(tcp_skb_pcount(skb
) == 0);
2818 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2821 return packets_acked
;
2824 /* Remove acknowledged frames from the retransmission queue. If our packet
2825 * is before the ack sequence we can discard it as it's confirmed to have
2826 * arrived at the other end.
2828 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
)
2830 struct tcp_sock
*tp
= tcp_sk(sk
);
2831 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2832 struct sk_buff
*skb
;
2833 u32 now
= tcp_time_stamp
;
2834 int fully_acked
= 1;
2837 u32 reord
= tp
->packets_out
;
2839 s32 ca_seq_rtt
= -1;
2840 ktime_t last_ackt
= net_invalid_timestamp();
2842 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2843 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2846 u8 sacked
= scb
->sacked
;
2848 /* Determine how many packets and what bytes were acked, tso and else */
2849 if (after(scb
->end_seq
, tp
->snd_una
)) {
2850 if (tcp_skb_pcount(skb
) == 1 ||
2851 !after(tp
->snd_una
, scb
->seq
))
2854 acked_pcount
= tcp_tso_acked(sk
, skb
);
2859 end_seq
= tp
->snd_una
;
2861 acked_pcount
= tcp_skb_pcount(skb
);
2862 end_seq
= scb
->end_seq
;
2865 /* MTU probing checks */
2866 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2867 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2868 tcp_mtup_probe_success(sk
, skb
);
2871 if (sacked
& TCPCB_RETRANS
) {
2872 if (sacked
& TCPCB_SACKED_RETRANS
)
2873 tp
->retrans_out
-= acked_pcount
;
2874 flag
|= FLAG_RETRANS_DATA_ACKED
;
2877 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
2878 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2880 ca_seq_rtt
= now
- scb
->when
;
2881 last_ackt
= skb
->tstamp
;
2883 seq_rtt
= ca_seq_rtt
;
2885 if (!(sacked
& TCPCB_SACKED_ACKED
))
2886 reord
= min(pkts_acked
, reord
);
2889 if (sacked
& TCPCB_SACKED_ACKED
)
2890 tp
->sacked_out
-= acked_pcount
;
2891 if (sacked
& TCPCB_LOST
)
2892 tp
->lost_out
-= acked_pcount
;
2894 if (unlikely(tp
->urg_mode
&& !before(end_seq
, tp
->snd_up
)))
2897 tp
->packets_out
-= acked_pcount
;
2898 pkts_acked
+= acked_pcount
;
2900 /* Initial outgoing SYN's get put onto the write_queue
2901 * just like anything else we transmit. It is not
2902 * true data, and if we misinform our callers that
2903 * this ACK acks real data, we will erroneously exit
2904 * connection startup slow start one packet too
2905 * quickly. This is severely frowned upon behavior.
2907 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2908 flag
|= FLAG_DATA_ACKED
;
2910 flag
|= FLAG_SYN_ACKED
;
2911 tp
->retrans_stamp
= 0;
2917 tcp_unlink_write_queue(skb
, sk
);
2918 sk_wmem_free_skb(sk
, skb
);
2919 tcp_clear_all_retrans_hints(tp
);
2922 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2923 flag
|= FLAG_SACK_RENEGING
;
2925 if (flag
& FLAG_ACKED
) {
2926 const struct tcp_congestion_ops
*ca_ops
2927 = inet_csk(sk
)->icsk_ca_ops
;
2929 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2932 if (tcp_is_reno(tp
)) {
2933 tcp_remove_reno_sacks(sk
, pkts_acked
);
2935 /* Non-retransmitted hole got filled? That's reordering */
2936 if (reord
< prior_fackets
)
2937 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2940 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2942 if (ca_ops
->pkts_acked
) {
2945 /* Is the ACK triggering packet unambiguous? */
2946 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2947 /* High resolution needed and available? */
2948 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2949 !ktime_equal(last_ackt
,
2950 net_invalid_timestamp()))
2951 rtt_us
= ktime_us_delta(ktime_get_real(),
2953 else if (ca_seq_rtt
> 0)
2954 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
2957 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2961 #if FASTRETRANS_DEBUG > 0
2962 WARN_ON((int)tp
->sacked_out
< 0);
2963 WARN_ON((int)tp
->lost_out
< 0);
2964 WARN_ON((int)tp
->retrans_out
< 0);
2965 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2966 icsk
= inet_csk(sk
);
2968 printk(KERN_DEBUG
"Leak l=%u %d\n",
2969 tp
->lost_out
, icsk
->icsk_ca_state
);
2972 if (tp
->sacked_out
) {
2973 printk(KERN_DEBUG
"Leak s=%u %d\n",
2974 tp
->sacked_out
, icsk
->icsk_ca_state
);
2977 if (tp
->retrans_out
) {
2978 printk(KERN_DEBUG
"Leak r=%u %d\n",
2979 tp
->retrans_out
, icsk
->icsk_ca_state
);
2980 tp
->retrans_out
= 0;
2987 static void tcp_ack_probe(struct sock
*sk
)
2989 const struct tcp_sock
*tp
= tcp_sk(sk
);
2990 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2992 /* Was it a usable window open? */
2994 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
2995 icsk
->icsk_backoff
= 0;
2996 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2997 /* Socket must be waked up by subsequent tcp_data_snd_check().
2998 * This function is not for random using!
3001 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3002 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3007 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3009 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3010 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3013 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3015 const struct tcp_sock
*tp
= tcp_sk(sk
);
3016 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3017 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3020 /* Check that window update is acceptable.
3021 * The function assumes that snd_una<=ack<=snd_next.
3023 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3024 const u32 ack
, const u32 ack_seq
,
3027 return (after(ack
, tp
->snd_una
) ||
3028 after(ack_seq
, tp
->snd_wl1
) ||
3029 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3032 /* Update our send window.
3034 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3035 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3037 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3040 struct tcp_sock
*tp
= tcp_sk(sk
);
3042 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3044 if (likely(!tcp_hdr(skb
)->syn
))
3045 nwin
<<= tp
->rx_opt
.snd_wscale
;
3047 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3048 flag
|= FLAG_WIN_UPDATE
;
3049 tcp_update_wl(tp
, ack
, ack_seq
);
3051 if (tp
->snd_wnd
!= nwin
) {
3054 /* Note, it is the only place, where
3055 * fast path is recovered for sending TCP.
3058 tcp_fast_path_check(sk
);
3060 if (nwin
> tp
->max_window
) {
3061 tp
->max_window
= nwin
;
3062 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3072 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3073 * continue in congestion avoidance.
3075 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3077 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3078 tp
->snd_cwnd_cnt
= 0;
3079 tp
->bytes_acked
= 0;
3080 TCP_ECN_queue_cwr(tp
);
3081 tcp_moderate_cwnd(tp
);
3084 /* A conservative spurious RTO response algorithm: reduce cwnd using
3085 * rate halving and continue in congestion avoidance.
3087 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3089 tcp_enter_cwr(sk
, 0);
3092 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3094 if (flag
& FLAG_ECE
)
3095 tcp_ratehalving_spur_to_response(sk
);
3097 tcp_undo_cwr(sk
, 1);
3100 /* F-RTO spurious RTO detection algorithm (RFC4138)
3102 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3103 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3104 * window (but not to or beyond highest sequence sent before RTO):
3105 * On First ACK, send two new segments out.
3106 * On Second ACK, RTO was likely spurious. Do spurious response (response
3107 * algorithm is not part of the F-RTO detection algorithm
3108 * given in RFC4138 but can be selected separately).
3109 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3110 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3111 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3112 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3114 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3115 * original window even after we transmit two new data segments.
3118 * on first step, wait until first cumulative ACK arrives, then move to
3119 * the second step. In second step, the next ACK decides.
3121 * F-RTO is implemented (mainly) in four functions:
3122 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3123 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3124 * called when tcp_use_frto() showed green light
3125 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3126 * - tcp_enter_frto_loss() is called if there is not enough evidence
3127 * to prove that the RTO is indeed spurious. It transfers the control
3128 * from F-RTO to the conventional RTO recovery
3130 static int tcp_process_frto(struct sock
*sk
, int flag
)
3132 struct tcp_sock
*tp
= tcp_sk(sk
);
3134 tcp_verify_left_out(tp
);
3136 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3137 if (flag
& FLAG_DATA_ACKED
)
3138 inet_csk(sk
)->icsk_retransmits
= 0;
3140 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3141 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3142 tp
->undo_marker
= 0;
3144 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3145 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3149 if (!tcp_is_sackfrto(tp
)) {
3150 /* RFC4138 shortcoming in step 2; should also have case c):
3151 * ACK isn't duplicate nor advances window, e.g., opposite dir
3154 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3157 if (!(flag
& FLAG_DATA_ACKED
)) {
3158 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3163 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3164 /* Prevent sending of new data. */
3165 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3166 tcp_packets_in_flight(tp
));
3170 if ((tp
->frto_counter
>= 2) &&
3171 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3172 ((flag
& FLAG_DATA_SACKED
) &&
3173 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3174 /* RFC4138 shortcoming (see comment above) */
3175 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3176 (flag
& FLAG_NOT_DUP
))
3179 tcp_enter_frto_loss(sk
, 3, flag
);
3184 if (tp
->frto_counter
== 1) {
3185 /* tcp_may_send_now needs to see updated state */
3186 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3187 tp
->frto_counter
= 2;
3189 if (!tcp_may_send_now(sk
))
3190 tcp_enter_frto_loss(sk
, 2, flag
);
3194 switch (sysctl_tcp_frto_response
) {
3196 tcp_undo_spur_to_response(sk
, flag
);
3199 tcp_conservative_spur_to_response(tp
);
3202 tcp_ratehalving_spur_to_response(sk
);
3205 tp
->frto_counter
= 0;
3206 tp
->undo_marker
= 0;
3207 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3212 /* This routine deals with incoming acks, but not outgoing ones. */
3213 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3215 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3216 struct tcp_sock
*tp
= tcp_sk(sk
);
3217 u32 prior_snd_una
= tp
->snd_una
;
3218 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3219 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3220 u32 prior_in_flight
;
3225 /* If the ack is newer than sent or older than previous acks
3226 * then we can probably ignore it.
3228 if (after(ack
, tp
->snd_nxt
))
3229 goto uninteresting_ack
;
3231 if (before(ack
, prior_snd_una
))
3234 if (after(ack
, prior_snd_una
))
3235 flag
|= FLAG_SND_UNA_ADVANCED
;
3237 if (sysctl_tcp_abc
) {
3238 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3239 tp
->bytes_acked
+= ack
- prior_snd_una
;
3240 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3241 /* we assume just one segment left network */
3242 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3246 prior_fackets
= tp
->fackets_out
;
3247 prior_in_flight
= tcp_packets_in_flight(tp
);
3249 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3250 /* Window is constant, pure forward advance.
3251 * No more checks are required.
3252 * Note, we use the fact that SND.UNA>=SND.WL2.
3254 tcp_update_wl(tp
, ack
, ack_seq
);
3256 flag
|= FLAG_WIN_UPDATE
;
3258 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3260 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3262 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3265 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3267 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3269 if (TCP_SKB_CB(skb
)->sacked
)
3270 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3272 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3275 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3278 /* We passed data and got it acked, remove any soft error
3279 * log. Something worked...
3281 sk
->sk_err_soft
= 0;
3282 icsk
->icsk_probes_out
= 0;
3283 tp
->rcv_tstamp
= tcp_time_stamp
;
3284 prior_packets
= tp
->packets_out
;
3288 /* See if we can take anything off of the retransmit queue. */
3289 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
);
3291 if (tp
->frto_counter
)
3292 frto_cwnd
= tcp_process_frto(sk
, flag
);
3293 /* Guarantee sacktag reordering detection against wrap-arounds */
3294 if (before(tp
->frto_highmark
, tp
->snd_una
))
3295 tp
->frto_highmark
= 0;
3297 if (tcp_ack_is_dubious(sk
, flag
)) {
3298 /* Advance CWND, if state allows this. */
3299 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3300 tcp_may_raise_cwnd(sk
, flag
))
3301 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3302 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3305 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3306 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3309 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3310 dst_confirm(sk
->sk_dst_cache
);
3315 /* If this ack opens up a zero window, clear backoff. It was
3316 * being used to time the probes, and is probably far higher than
3317 * it needs to be for normal retransmission.
3319 if (tcp_send_head(sk
))
3324 if (TCP_SKB_CB(skb
)->sacked
) {
3325 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3326 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3327 tcp_try_keep_open(sk
);
3331 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3335 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3336 * But, this can also be called on packets in the established flow when
3337 * the fast version below fails.
3339 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3343 struct tcphdr
*th
= tcp_hdr(skb
);
3344 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3346 ptr
= (unsigned char *)(th
+ 1);
3347 opt_rx
->saw_tstamp
= 0;
3349 while (length
> 0) {
3350 int opcode
= *ptr
++;
3356 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3361 if (opsize
< 2) /* "silly options" */
3363 if (opsize
> length
)
3364 return; /* don't parse partial options */
3367 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3368 u16 in_mss
= get_unaligned_be16(ptr
);
3370 if (opt_rx
->user_mss
&&
3371 opt_rx
->user_mss
< in_mss
)
3372 in_mss
= opt_rx
->user_mss
;
3373 opt_rx
->mss_clamp
= in_mss
;
3378 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3379 !estab
&& sysctl_tcp_window_scaling
) {
3380 __u8 snd_wscale
= *(__u8
*)ptr
;
3381 opt_rx
->wscale_ok
= 1;
3382 if (snd_wscale
> 14) {
3383 if (net_ratelimit())
3384 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3385 "scaling value %d >14 received.\n",
3389 opt_rx
->snd_wscale
= snd_wscale
;
3392 case TCPOPT_TIMESTAMP
:
3393 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3394 ((estab
&& opt_rx
->tstamp_ok
) ||
3395 (!estab
&& sysctl_tcp_timestamps
))) {
3396 opt_rx
->saw_tstamp
= 1;
3397 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3398 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3401 case TCPOPT_SACK_PERM
:
3402 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3403 !estab
&& sysctl_tcp_sack
) {
3404 opt_rx
->sack_ok
= 1;
3405 tcp_sack_reset(opt_rx
);
3410 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3411 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3413 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3416 #ifdef CONFIG_TCP_MD5SIG
3419 * The MD5 Hash has already been
3420 * checked (see tcp_v{4,6}_do_rcv()).
3432 /* Fast parse options. This hopes to only see timestamps.
3433 * If it is wrong it falls back on tcp_parse_options().
3435 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3436 struct tcp_sock
*tp
)
3438 if (th
->doff
== sizeof(struct tcphdr
) >> 2) {
3439 tp
->rx_opt
.saw_tstamp
= 0;
3441 } else if (tp
->rx_opt
.tstamp_ok
&&
3442 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3443 __be32
*ptr
= (__be32
*)(th
+ 1);
3444 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3445 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3446 tp
->rx_opt
.saw_tstamp
= 1;
3448 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3450 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3454 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3458 #ifdef CONFIG_TCP_MD5SIG
3460 * Parse MD5 Signature option
3462 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3464 int length
= (th
->doff
<< 2) - sizeof (*th
);
3465 u8
*ptr
= (u8
*)(th
+ 1);
3467 /* If the TCP option is too short, we can short cut */
3468 if (length
< TCPOLEN_MD5SIG
)
3471 while (length
> 0) {
3472 int opcode
= *ptr
++;
3483 if (opsize
< 2 || opsize
> length
)
3485 if (opcode
== TCPOPT_MD5SIG
)
3495 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3497 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3498 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3501 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3503 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3504 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3505 * extra check below makes sure this can only happen
3506 * for pure ACK frames. -DaveM
3508 * Not only, also it occurs for expired timestamps.
3511 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3512 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3513 tcp_store_ts_recent(tp
);
3517 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3519 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3520 * it can pass through stack. So, the following predicate verifies that
3521 * this segment is not used for anything but congestion avoidance or
3522 * fast retransmit. Moreover, we even are able to eliminate most of such
3523 * second order effects, if we apply some small "replay" window (~RTO)
3524 * to timestamp space.
3526 * All these measures still do not guarantee that we reject wrapped ACKs
3527 * on networks with high bandwidth, when sequence space is recycled fastly,
3528 * but it guarantees that such events will be very rare and do not affect
3529 * connection seriously. This doesn't look nice, but alas, PAWS is really
3532 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3533 * states that events when retransmit arrives after original data are rare.
3534 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3535 * the biggest problem on large power networks even with minor reordering.
3536 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3537 * up to bandwidth of 18Gigabit/sec. 8) ]
3540 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3542 struct tcp_sock
*tp
= tcp_sk(sk
);
3543 struct tcphdr
*th
= tcp_hdr(skb
);
3544 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3545 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3547 return (/* 1. Pure ACK with correct sequence number. */
3548 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3550 /* 2. ... and duplicate ACK. */
3551 ack
== tp
->snd_una
&&
3553 /* 3. ... and does not update window. */
3554 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3556 /* 4. ... and sits in replay window. */
3557 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3560 static inline int tcp_paws_discard(const struct sock
*sk
,
3561 const struct sk_buff
*skb
)
3563 const struct tcp_sock
*tp
= tcp_sk(sk
);
3564 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3565 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3566 !tcp_disordered_ack(sk
, skb
));
3569 /* Check segment sequence number for validity.
3571 * Segment controls are considered valid, if the segment
3572 * fits to the window after truncation to the window. Acceptability
3573 * of data (and SYN, FIN, of course) is checked separately.
3574 * See tcp_data_queue(), for example.
3576 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3577 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3578 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3579 * (borrowed from freebsd)
3582 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3584 return !before(end_seq
, tp
->rcv_wup
) &&
3585 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3588 /* When we get a reset we do this. */
3589 static void tcp_reset(struct sock
*sk
)
3591 /* We want the right error as BSD sees it (and indeed as we do). */
3592 switch (sk
->sk_state
) {
3594 sk
->sk_err
= ECONNREFUSED
;
3596 case TCP_CLOSE_WAIT
:
3602 sk
->sk_err
= ECONNRESET
;
3605 if (!sock_flag(sk
, SOCK_DEAD
))
3606 sk
->sk_error_report(sk
);
3612 * Process the FIN bit. This now behaves as it is supposed to work
3613 * and the FIN takes effect when it is validly part of sequence
3614 * space. Not before when we get holes.
3616 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3617 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3620 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3621 * close and we go into CLOSING (and later onto TIME-WAIT)
3623 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3625 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3627 struct tcp_sock
*tp
= tcp_sk(sk
);
3629 inet_csk_schedule_ack(sk
);
3631 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3632 sock_set_flag(sk
, SOCK_DONE
);
3634 switch (sk
->sk_state
) {
3636 case TCP_ESTABLISHED
:
3637 /* Move to CLOSE_WAIT */
3638 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3639 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3642 case TCP_CLOSE_WAIT
:
3644 /* Received a retransmission of the FIN, do
3649 /* RFC793: Remain in the LAST-ACK state. */
3653 /* This case occurs when a simultaneous close
3654 * happens, we must ack the received FIN and
3655 * enter the CLOSING state.
3658 tcp_set_state(sk
, TCP_CLOSING
);
3661 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3663 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3666 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3667 * cases we should never reach this piece of code.
3669 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3670 __func__
, sk
->sk_state
);
3674 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3675 * Probably, we should reset in this case. For now drop them.
3677 __skb_queue_purge(&tp
->out_of_order_queue
);
3678 if (tcp_is_sack(tp
))
3679 tcp_sack_reset(&tp
->rx_opt
);
3682 if (!sock_flag(sk
, SOCK_DEAD
)) {
3683 sk
->sk_state_change(sk
);
3685 /* Do not send POLL_HUP for half duplex close. */
3686 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3687 sk
->sk_state
== TCP_CLOSE
)
3688 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3690 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3694 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3697 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3698 if (before(seq
, sp
->start_seq
))
3699 sp
->start_seq
= seq
;
3700 if (after(end_seq
, sp
->end_seq
))
3701 sp
->end_seq
= end_seq
;
3707 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3709 struct tcp_sock
*tp
= tcp_sk(sk
);
3711 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3714 if (before(seq
, tp
->rcv_nxt
))
3715 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3717 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3719 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3721 tp
->rx_opt
.dsack
= 1;
3722 tp
->duplicate_sack
[0].start_seq
= seq
;
3723 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3724 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ 1;
3728 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3730 struct tcp_sock
*tp
= tcp_sk(sk
);
3732 if (!tp
->rx_opt
.dsack
)
3733 tcp_dsack_set(sk
, seq
, end_seq
);
3735 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3738 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3740 struct tcp_sock
*tp
= tcp_sk(sk
);
3742 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3743 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3744 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3745 tcp_enter_quickack_mode(sk
);
3747 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3748 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3750 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3751 end_seq
= tp
->rcv_nxt
;
3752 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3759 /* These routines update the SACK block as out-of-order packets arrive or
3760 * in-order packets close up the sequence space.
3762 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3765 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3766 struct tcp_sack_block
*swalk
= sp
+ 1;
3768 /* See if the recent change to the first SACK eats into
3769 * or hits the sequence space of other SACK blocks, if so coalesce.
3771 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3772 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3775 /* Zap SWALK, by moving every further SACK up by one slot.
3776 * Decrease num_sacks.
3778 tp
->rx_opt
.num_sacks
--;
3779 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
3781 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3785 this_sack
++, swalk
++;
3789 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
,
3790 struct tcp_sack_block
*sack2
)
3794 tmp
= sack1
->start_seq
;
3795 sack1
->start_seq
= sack2
->start_seq
;
3796 sack2
->start_seq
= tmp
;
3798 tmp
= sack1
->end_seq
;
3799 sack1
->end_seq
= sack2
->end_seq
;
3800 sack2
->end_seq
= tmp
;
3803 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3805 struct tcp_sock
*tp
= tcp_sk(sk
);
3806 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3807 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3813 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3814 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3815 /* Rotate this_sack to the first one. */
3816 for (; this_sack
> 0; this_sack
--, sp
--)
3817 tcp_sack_swap(sp
, sp
- 1);
3819 tcp_sack_maybe_coalesce(tp
);
3824 /* Could not find an adjacent existing SACK, build a new one,
3825 * put it at the front, and shift everyone else down. We
3826 * always know there is at least one SACK present already here.
3828 * If the sack array is full, forget about the last one.
3830 if (this_sack
>= TCP_NUM_SACKS
) {
3832 tp
->rx_opt
.num_sacks
--;
3835 for (; this_sack
> 0; this_sack
--, sp
--)
3839 /* Build the new head SACK, and we're done. */
3840 sp
->start_seq
= seq
;
3841 sp
->end_seq
= end_seq
;
3842 tp
->rx_opt
.num_sacks
++;
3843 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
3846 /* RCV.NXT advances, some SACKs should be eaten. */
3848 static void tcp_sack_remove(struct tcp_sock
*tp
)
3850 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3851 int num_sacks
= tp
->rx_opt
.num_sacks
;
3854 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3855 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3856 tp
->rx_opt
.num_sacks
= 0;
3857 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3861 for (this_sack
= 0; this_sack
< num_sacks
;) {
3862 /* Check if the start of the sack is covered by RCV.NXT. */
3863 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3866 /* RCV.NXT must cover all the block! */
3867 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
3869 /* Zap this SACK, by moving forward any other SACKS. */
3870 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3871 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3878 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3879 tp
->rx_opt
.num_sacks
= num_sacks
;
3880 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
3885 /* This one checks to see if we can put data from the
3886 * out_of_order queue into the receive_queue.
3888 static void tcp_ofo_queue(struct sock
*sk
)
3890 struct tcp_sock
*tp
= tcp_sk(sk
);
3891 __u32 dsack_high
= tp
->rcv_nxt
;
3892 struct sk_buff
*skb
;
3894 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3895 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3898 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3899 __u32 dsack
= dsack_high
;
3900 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3901 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3902 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
3905 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3906 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3907 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3911 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3912 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3913 TCP_SKB_CB(skb
)->end_seq
);
3915 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3916 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3917 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3918 if (tcp_hdr(skb
)->fin
)
3919 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3923 static int tcp_prune_ofo_queue(struct sock
*sk
);
3924 static int tcp_prune_queue(struct sock
*sk
);
3926 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
3928 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3929 !sk_rmem_schedule(sk
, size
)) {
3931 if (tcp_prune_queue(sk
) < 0)
3934 if (!sk_rmem_schedule(sk
, size
)) {
3935 if (!tcp_prune_ofo_queue(sk
))
3938 if (!sk_rmem_schedule(sk
, size
))
3945 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3947 struct tcphdr
*th
= tcp_hdr(skb
);
3948 struct tcp_sock
*tp
= tcp_sk(sk
);
3951 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3954 __skb_pull(skb
, th
->doff
* 4);
3956 TCP_ECN_accept_cwr(tp
, skb
);
3958 if (tp
->rx_opt
.dsack
) {
3959 tp
->rx_opt
.dsack
= 0;
3960 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
;
3963 /* Queue data for delivery to the user.
3964 * Packets in sequence go to the receive queue.
3965 * Out of sequence packets to the out_of_order_queue.
3967 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3968 if (tcp_receive_window(tp
) == 0)
3971 /* Ok. In sequence. In window. */
3972 if (tp
->ucopy
.task
== current
&&
3973 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3974 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3975 int chunk
= min_t(unsigned int, skb
->len
,
3978 __set_current_state(TASK_RUNNING
);
3981 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3982 tp
->ucopy
.len
-= chunk
;
3983 tp
->copied_seq
+= chunk
;
3984 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3985 tcp_rcv_space_adjust(sk
);
3993 tcp_try_rmem_schedule(sk
, skb
->truesize
))
3996 skb_set_owner_r(skb
, sk
);
3997 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3999 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4001 tcp_event_data_recv(sk
, skb
);
4003 tcp_fin(skb
, sk
, th
);
4005 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4008 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4009 * gap in queue is filled.
4011 if (skb_queue_empty(&tp
->out_of_order_queue
))
4012 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4015 if (tp
->rx_opt
.num_sacks
)
4016 tcp_sack_remove(tp
);
4018 tcp_fast_path_check(sk
);
4022 else if (!sock_flag(sk
, SOCK_DEAD
))
4023 sk
->sk_data_ready(sk
, 0);
4027 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4028 /* A retransmit, 2nd most common case. Force an immediate ack. */
4029 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4030 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4033 tcp_enter_quickack_mode(sk
);
4034 inet_csk_schedule_ack(sk
);
4040 /* Out of window. F.e. zero window probe. */
4041 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4044 tcp_enter_quickack_mode(sk
);
4046 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4047 /* Partial packet, seq < rcv_next < end_seq */
4048 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4049 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4050 TCP_SKB_CB(skb
)->end_seq
);
4052 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4054 /* If window is closed, drop tail of packet. But after
4055 * remembering D-SACK for its head made in previous line.
4057 if (!tcp_receive_window(tp
))
4062 TCP_ECN_check_ce(tp
, skb
);
4064 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4067 /* Disable header prediction. */
4069 inet_csk_schedule_ack(sk
);
4071 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4072 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4074 skb_set_owner_r(skb
, sk
);
4076 if (!skb_peek(&tp
->out_of_order_queue
)) {
4077 /* Initial out of order segment, build 1 SACK. */
4078 if (tcp_is_sack(tp
)) {
4079 tp
->rx_opt
.num_sacks
= 1;
4080 tp
->rx_opt
.dsack
= 0;
4081 tp
->rx_opt
.eff_sacks
= 1;
4082 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4083 tp
->selective_acks
[0].end_seq
=
4084 TCP_SKB_CB(skb
)->end_seq
;
4086 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4088 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
4089 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4090 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4092 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4093 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4095 if (!tp
->rx_opt
.num_sacks
||
4096 tp
->selective_acks
[0].end_seq
!= seq
)
4099 /* Common case: data arrive in order after hole. */
4100 tp
->selective_acks
[0].end_seq
= end_seq
;
4104 /* Find place to insert this segment. */
4106 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4108 } while ((skb1
= skb1
->prev
) !=
4109 (struct sk_buff
*)&tp
->out_of_order_queue
);
4111 /* Do skb overlap to previous one? */
4112 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
4113 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4114 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4115 /* All the bits are present. Drop. */
4117 tcp_dsack_set(sk
, seq
, end_seq
);
4120 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4121 /* Partial overlap. */
4122 tcp_dsack_set(sk
, seq
,
4123 TCP_SKB_CB(skb1
)->end_seq
);
4128 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
4130 /* And clean segments covered by new one as whole. */
4131 while ((skb1
= skb
->next
) !=
4132 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4133 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4134 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4135 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4139 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4140 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4141 TCP_SKB_CB(skb1
)->end_seq
);
4146 if (tcp_is_sack(tp
))
4147 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4151 /* Collapse contiguous sequence of skbs head..tail with
4152 * sequence numbers start..end.
4153 * Segments with FIN/SYN are not collapsed (only because this
4157 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4158 struct sk_buff
*head
, struct sk_buff
*tail
,
4161 struct sk_buff
*skb
;
4163 /* First, check that queue is collapsible and find
4164 * the point where collapsing can be useful. */
4165 for (skb
= head
; skb
!= tail
;) {
4166 /* No new bits? It is possible on ofo queue. */
4167 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4168 struct sk_buff
*next
= skb
->next
;
4169 __skb_unlink(skb
, list
);
4171 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4176 /* The first skb to collapse is:
4178 * - bloated or contains data before "start" or
4179 * overlaps to the next one.
4181 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4182 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4183 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4184 (skb
->next
!= tail
&&
4185 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4188 /* Decided to skip this, advance start seq. */
4189 start
= TCP_SKB_CB(skb
)->end_seq
;
4192 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4195 while (before(start
, end
)) {
4196 struct sk_buff
*nskb
;
4197 unsigned int header
= skb_headroom(skb
);
4198 int copy
= SKB_MAX_ORDER(header
, 0);
4200 /* Too big header? This can happen with IPv6. */
4203 if (end
- start
< copy
)
4205 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4209 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4210 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4212 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4214 skb_reserve(nskb
, header
);
4215 memcpy(nskb
->head
, skb
->head
, header
);
4216 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4217 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4218 __skb_insert(nskb
, skb
->prev
, skb
, list
);
4219 skb_set_owner_r(nskb
, sk
);
4221 /* Copy data, releasing collapsed skbs. */
4223 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4224 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4228 size
= min(copy
, size
);
4229 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4231 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4235 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4236 struct sk_buff
*next
= skb
->next
;
4237 __skb_unlink(skb
, list
);
4239 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4242 tcp_hdr(skb
)->syn
||
4250 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4251 * and tcp_collapse() them until all the queue is collapsed.
4253 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4255 struct tcp_sock
*tp
= tcp_sk(sk
);
4256 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4257 struct sk_buff
*head
;
4263 start
= TCP_SKB_CB(skb
)->seq
;
4264 end
= TCP_SKB_CB(skb
)->end_seq
;
4270 /* Segment is terminated when we see gap or when
4271 * we are at the end of all the queue. */
4272 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4273 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4274 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4275 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4276 head
, skb
, start
, end
);
4278 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4280 /* Start new segment */
4281 start
= TCP_SKB_CB(skb
)->seq
;
4282 end
= TCP_SKB_CB(skb
)->end_seq
;
4284 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4285 start
= TCP_SKB_CB(skb
)->seq
;
4286 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4287 end
= TCP_SKB_CB(skb
)->end_seq
;
4293 * Purge the out-of-order queue.
4294 * Return true if queue was pruned.
4296 static int tcp_prune_ofo_queue(struct sock
*sk
)
4298 struct tcp_sock
*tp
= tcp_sk(sk
);
4301 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4302 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4303 __skb_queue_purge(&tp
->out_of_order_queue
);
4305 /* Reset SACK state. A conforming SACK implementation will
4306 * do the same at a timeout based retransmit. When a connection
4307 * is in a sad state like this, we care only about integrity
4308 * of the connection not performance.
4310 if (tp
->rx_opt
.sack_ok
)
4311 tcp_sack_reset(&tp
->rx_opt
);
4318 /* Reduce allocated memory if we can, trying to get
4319 * the socket within its memory limits again.
4321 * Return less than zero if we should start dropping frames
4322 * until the socket owning process reads some of the data
4323 * to stabilize the situation.
4325 static int tcp_prune_queue(struct sock
*sk
)
4327 struct tcp_sock
*tp
= tcp_sk(sk
);
4329 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4331 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4333 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4334 tcp_clamp_window(sk
);
4335 else if (tcp_memory_pressure
)
4336 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4338 tcp_collapse_ofo_queue(sk
);
4339 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4340 sk
->sk_receive_queue
.next
,
4341 (struct sk_buff
*)&sk
->sk_receive_queue
,
4342 tp
->copied_seq
, tp
->rcv_nxt
);
4345 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4348 /* Collapsing did not help, destructive actions follow.
4349 * This must not ever occur. */
4351 tcp_prune_ofo_queue(sk
);
4353 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4356 /* If we are really being abused, tell the caller to silently
4357 * drop receive data on the floor. It will get retransmitted
4358 * and hopefully then we'll have sufficient space.
4360 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4362 /* Massive buffer overcommit. */
4367 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4368 * As additional protections, we do not touch cwnd in retransmission phases,
4369 * and if application hit its sndbuf limit recently.
4371 void tcp_cwnd_application_limited(struct sock
*sk
)
4373 struct tcp_sock
*tp
= tcp_sk(sk
);
4375 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4376 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4377 /* Limited by application or receiver window. */
4378 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4379 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4380 if (win_used
< tp
->snd_cwnd
) {
4381 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4382 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4384 tp
->snd_cwnd_used
= 0;
4386 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4389 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4391 struct tcp_sock
*tp
= tcp_sk(sk
);
4393 /* If the user specified a specific send buffer setting, do
4396 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4399 /* If we are under global TCP memory pressure, do not expand. */
4400 if (tcp_memory_pressure
)
4403 /* If we are under soft global TCP memory pressure, do not expand. */
4404 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4407 /* If we filled the congestion window, do not expand. */
4408 if (tp
->packets_out
>= tp
->snd_cwnd
)
4414 /* When incoming ACK allowed to free some skb from write_queue,
4415 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4416 * on the exit from tcp input handler.
4418 * PROBLEM: sndbuf expansion does not work well with largesend.
4420 static void tcp_new_space(struct sock
*sk
)
4422 struct tcp_sock
*tp
= tcp_sk(sk
);
4424 if (tcp_should_expand_sndbuf(sk
)) {
4425 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4426 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4427 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4428 tp
->reordering
+ 1);
4429 sndmem
*= 2 * demanded
;
4430 if (sndmem
> sk
->sk_sndbuf
)
4431 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4432 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4435 sk
->sk_write_space(sk
);
4438 static void tcp_check_space(struct sock
*sk
)
4440 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4441 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4442 if (sk
->sk_socket
&&
4443 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4448 static inline void tcp_data_snd_check(struct sock
*sk
)
4450 tcp_push_pending_frames(sk
);
4451 tcp_check_space(sk
);
4455 * Check if sending an ack is needed.
4457 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4459 struct tcp_sock
*tp
= tcp_sk(sk
);
4461 /* More than one full frame received... */
4462 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4463 /* ... and right edge of window advances far enough.
4464 * (tcp_recvmsg() will send ACK otherwise). Or...
4466 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4467 /* We ACK each frame or... */
4468 tcp_in_quickack_mode(sk
) ||
4469 /* We have out of order data. */
4470 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4471 /* Then ack it now */
4474 /* Else, send delayed ack. */
4475 tcp_send_delayed_ack(sk
);
4479 static inline void tcp_ack_snd_check(struct sock
*sk
)
4481 if (!inet_csk_ack_scheduled(sk
)) {
4482 /* We sent a data segment already. */
4485 __tcp_ack_snd_check(sk
, 1);
4489 * This routine is only called when we have urgent data
4490 * signaled. Its the 'slow' part of tcp_urg. It could be
4491 * moved inline now as tcp_urg is only called from one
4492 * place. We handle URGent data wrong. We have to - as
4493 * BSD still doesn't use the correction from RFC961.
4494 * For 1003.1g we should support a new option TCP_STDURG to permit
4495 * either form (or just set the sysctl tcp_stdurg).
4498 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4500 struct tcp_sock
*tp
= tcp_sk(sk
);
4501 u32 ptr
= ntohs(th
->urg_ptr
);
4503 if (ptr
&& !sysctl_tcp_stdurg
)
4505 ptr
+= ntohl(th
->seq
);
4507 /* Ignore urgent data that we've already seen and read. */
4508 if (after(tp
->copied_seq
, ptr
))
4511 /* Do not replay urg ptr.
4513 * NOTE: interesting situation not covered by specs.
4514 * Misbehaving sender may send urg ptr, pointing to segment,
4515 * which we already have in ofo queue. We are not able to fetch
4516 * such data and will stay in TCP_URG_NOTYET until will be eaten
4517 * by recvmsg(). Seems, we are not obliged to handle such wicked
4518 * situations. But it is worth to think about possibility of some
4519 * DoSes using some hypothetical application level deadlock.
4521 if (before(ptr
, tp
->rcv_nxt
))
4524 /* Do we already have a newer (or duplicate) urgent pointer? */
4525 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4528 /* Tell the world about our new urgent pointer. */
4531 /* We may be adding urgent data when the last byte read was
4532 * urgent. To do this requires some care. We cannot just ignore
4533 * tp->copied_seq since we would read the last urgent byte again
4534 * as data, nor can we alter copied_seq until this data arrives
4535 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4537 * NOTE. Double Dutch. Rendering to plain English: author of comment
4538 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4539 * and expect that both A and B disappear from stream. This is _wrong_.
4540 * Though this happens in BSD with high probability, this is occasional.
4541 * Any application relying on this is buggy. Note also, that fix "works"
4542 * only in this artificial test. Insert some normal data between A and B and we will
4543 * decline of BSD again. Verdict: it is better to remove to trap
4546 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4547 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4548 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4550 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4551 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4556 tp
->urg_data
= TCP_URG_NOTYET
;
4559 /* Disable header prediction. */
4563 /* This is the 'fast' part of urgent handling. */
4564 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4566 struct tcp_sock
*tp
= tcp_sk(sk
);
4568 /* Check if we get a new urgent pointer - normally not. */
4570 tcp_check_urg(sk
, th
);
4572 /* Do we wait for any urgent data? - normally not... */
4573 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4574 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4577 /* Is the urgent pointer pointing into this packet? */
4578 if (ptr
< skb
->len
) {
4580 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4582 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4583 if (!sock_flag(sk
, SOCK_DEAD
))
4584 sk
->sk_data_ready(sk
, 0);
4589 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4591 struct tcp_sock
*tp
= tcp_sk(sk
);
4592 int chunk
= skb
->len
- hlen
;
4596 if (skb_csum_unnecessary(skb
))
4597 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4599 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4603 tp
->ucopy
.len
-= chunk
;
4604 tp
->copied_seq
+= chunk
;
4605 tcp_rcv_space_adjust(sk
);
4612 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4613 struct sk_buff
*skb
)
4617 if (sock_owned_by_user(sk
)) {
4619 result
= __tcp_checksum_complete(skb
);
4622 result
= __tcp_checksum_complete(skb
);
4627 static inline int tcp_checksum_complete_user(struct sock
*sk
,
4628 struct sk_buff
*skb
)
4630 return !skb_csum_unnecessary(skb
) &&
4631 __tcp_checksum_complete_user(sk
, skb
);
4634 #ifdef CONFIG_NET_DMA
4635 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4638 struct tcp_sock
*tp
= tcp_sk(sk
);
4639 int chunk
= skb
->len
- hlen
;
4641 int copied_early
= 0;
4643 if (tp
->ucopy
.wakeup
)
4646 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4647 tp
->ucopy
.dma_chan
= get_softnet_dma();
4649 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4651 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4653 tp
->ucopy
.iov
, chunk
,
4654 tp
->ucopy
.pinned_list
);
4659 tp
->ucopy
.dma_cookie
= dma_cookie
;
4662 tp
->ucopy
.len
-= chunk
;
4663 tp
->copied_seq
+= chunk
;
4664 tcp_rcv_space_adjust(sk
);
4666 if ((tp
->ucopy
.len
== 0) ||
4667 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4668 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4669 tp
->ucopy
.wakeup
= 1;
4670 sk
->sk_data_ready(sk
, 0);
4672 } else if (chunk
> 0) {
4673 tp
->ucopy
.wakeup
= 1;
4674 sk
->sk_data_ready(sk
, 0);
4677 return copied_early
;
4679 #endif /* CONFIG_NET_DMA */
4682 * TCP receive function for the ESTABLISHED state.
4684 * It is split into a fast path and a slow path. The fast path is
4686 * - A zero window was announced from us - zero window probing
4687 * is only handled properly in the slow path.
4688 * - Out of order segments arrived.
4689 * - Urgent data is expected.
4690 * - There is no buffer space left
4691 * - Unexpected TCP flags/window values/header lengths are received
4692 * (detected by checking the TCP header against pred_flags)
4693 * - Data is sent in both directions. Fast path only supports pure senders
4694 * or pure receivers (this means either the sequence number or the ack
4695 * value must stay constant)
4696 * - Unexpected TCP option.
4698 * When these conditions are not satisfied it drops into a standard
4699 * receive procedure patterned after RFC793 to handle all cases.
4700 * The first three cases are guaranteed by proper pred_flags setting,
4701 * the rest is checked inline. Fast processing is turned on in
4702 * tcp_data_queue when everything is OK.
4704 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4705 struct tcphdr
*th
, unsigned len
)
4707 struct tcp_sock
*tp
= tcp_sk(sk
);
4710 * Header prediction.
4711 * The code loosely follows the one in the famous
4712 * "30 instruction TCP receive" Van Jacobson mail.
4714 * Van's trick is to deposit buffers into socket queue
4715 * on a device interrupt, to call tcp_recv function
4716 * on the receive process context and checksum and copy
4717 * the buffer to user space. smart...
4719 * Our current scheme is not silly either but we take the
4720 * extra cost of the net_bh soft interrupt processing...
4721 * We do checksum and copy also but from device to kernel.
4724 tp
->rx_opt
.saw_tstamp
= 0;
4726 /* pred_flags is 0xS?10 << 16 + snd_wnd
4727 * if header_prediction is to be made
4728 * 'S' will always be tp->tcp_header_len >> 2
4729 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4730 * turn it off (when there are holes in the receive
4731 * space for instance)
4732 * PSH flag is ignored.
4735 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4736 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4737 int tcp_header_len
= tp
->tcp_header_len
;
4739 /* Timestamp header prediction: tcp_header_len
4740 * is automatically equal to th->doff*4 due to pred_flags
4744 /* Check timestamp */
4745 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4746 __be32
*ptr
= (__be32
*)(th
+ 1);
4748 /* No? Slow path! */
4749 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4750 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4753 tp
->rx_opt
.saw_tstamp
= 1;
4755 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4757 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4759 /* If PAWS failed, check it more carefully in slow path */
4760 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4763 /* DO NOT update ts_recent here, if checksum fails
4764 * and timestamp was corrupted part, it will result
4765 * in a hung connection since we will drop all
4766 * future packets due to the PAWS test.
4770 if (len
<= tcp_header_len
) {
4771 /* Bulk data transfer: sender */
4772 if (len
== tcp_header_len
) {
4773 /* Predicted packet is in window by definition.
4774 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4775 * Hence, check seq<=rcv_wup reduces to:
4777 if (tcp_header_len
==
4778 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4779 tp
->rcv_nxt
== tp
->rcv_wup
)
4780 tcp_store_ts_recent(tp
);
4782 /* We know that such packets are checksummed
4785 tcp_ack(sk
, skb
, 0);
4787 tcp_data_snd_check(sk
);
4789 } else { /* Header too small */
4790 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4795 int copied_early
= 0;
4797 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4798 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4799 #ifdef CONFIG_NET_DMA
4800 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4805 if (tp
->ucopy
.task
== current
&&
4806 sock_owned_by_user(sk
) && !copied_early
) {
4807 __set_current_state(TASK_RUNNING
);
4809 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4813 /* Predicted packet is in window by definition.
4814 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4815 * Hence, check seq<=rcv_wup reduces to:
4817 if (tcp_header_len
==
4818 (sizeof(struct tcphdr
) +
4819 TCPOLEN_TSTAMP_ALIGNED
) &&
4820 tp
->rcv_nxt
== tp
->rcv_wup
)
4821 tcp_store_ts_recent(tp
);
4823 tcp_rcv_rtt_measure_ts(sk
, skb
);
4825 __skb_pull(skb
, tcp_header_len
);
4826 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4827 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
4830 tcp_cleanup_rbuf(sk
, skb
->len
);
4833 if (tcp_checksum_complete_user(sk
, skb
))
4836 /* Predicted packet is in window by definition.
4837 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4838 * Hence, check seq<=rcv_wup reduces to:
4840 if (tcp_header_len
==
4841 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4842 tp
->rcv_nxt
== tp
->rcv_wup
)
4843 tcp_store_ts_recent(tp
);
4845 tcp_rcv_rtt_measure_ts(sk
, skb
);
4847 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4850 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
4852 /* Bulk data transfer: receiver */
4853 __skb_pull(skb
, tcp_header_len
);
4854 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4855 skb_set_owner_r(skb
, sk
);
4856 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4859 tcp_event_data_recv(sk
, skb
);
4861 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4862 /* Well, only one small jumplet in fast path... */
4863 tcp_ack(sk
, skb
, FLAG_DATA
);
4864 tcp_data_snd_check(sk
);
4865 if (!inet_csk_ack_scheduled(sk
))
4869 __tcp_ack_snd_check(sk
, 0);
4871 #ifdef CONFIG_NET_DMA
4873 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4879 sk
->sk_data_ready(sk
, 0);
4885 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
4889 * RFC1323: H1. Apply PAWS check first.
4891 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4892 tcp_paws_discard(sk
, skb
)) {
4894 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
4895 tcp_send_dupack(sk
, skb
);
4898 /* Resets are accepted even if PAWS failed.
4900 ts_recent update must be made after we are sure
4901 that the packet is in window.
4906 * Standard slow path.
4909 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4910 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4911 * (RST) segments are validated by checking their SEQ-fields."
4912 * And page 69: "If an incoming segment is not acceptable,
4913 * an acknowledgment should be sent in reply (unless the RST bit
4914 * is set, if so drop the segment and return)".
4917 tcp_send_dupack(sk
, skb
);
4926 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4928 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4929 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4930 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
4937 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4939 tcp_rcv_rtt_measure_ts(sk
, skb
);
4941 /* Process urgent data. */
4942 tcp_urg(sk
, skb
, th
);
4944 /* step 7: process the segment text */
4945 tcp_data_queue(sk
, skb
);
4947 tcp_data_snd_check(sk
);
4948 tcp_ack_snd_check(sk
);
4952 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4959 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4960 struct tcphdr
*th
, unsigned len
)
4962 struct tcp_sock
*tp
= tcp_sk(sk
);
4963 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4964 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4966 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4970 * "If the state is SYN-SENT then
4971 * first check the ACK bit
4972 * If the ACK bit is set
4973 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4974 * a reset (unless the RST bit is set, if so drop
4975 * the segment and return)"
4977 * We do not send data with SYN, so that RFC-correct
4980 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4981 goto reset_and_undo
;
4983 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4984 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4986 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
4987 goto reset_and_undo
;
4990 /* Now ACK is acceptable.
4992 * "If the RST bit is set
4993 * If the ACK was acceptable then signal the user "error:
4994 * connection reset", drop the segment, enter CLOSED state,
4995 * delete TCB, and return."
5004 * "fifth, if neither of the SYN or RST bits is set then
5005 * drop the segment and return."
5011 goto discard_and_undo
;
5014 * "If the SYN bit is on ...
5015 * are acceptable then ...
5016 * (our SYN has been ACKed), change the connection
5017 * state to ESTABLISHED..."
5020 TCP_ECN_rcv_synack(tp
, th
);
5022 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5023 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5025 /* Ok.. it's good. Set up sequence numbers and
5026 * move to established.
5028 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5029 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5031 /* RFC1323: The window in SYN & SYN/ACK segments is
5034 tp
->snd_wnd
= ntohs(th
->window
);
5035 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
5037 if (!tp
->rx_opt
.wscale_ok
) {
5038 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5039 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5042 if (tp
->rx_opt
.saw_tstamp
) {
5043 tp
->rx_opt
.tstamp_ok
= 1;
5044 tp
->tcp_header_len
=
5045 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5046 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5047 tcp_store_ts_recent(tp
);
5049 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5052 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5053 tcp_enable_fack(tp
);
5056 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5057 tcp_initialize_rcv_mss(sk
);
5059 /* Remember, tcp_poll() does not lock socket!
5060 * Change state from SYN-SENT only after copied_seq
5061 * is initialized. */
5062 tp
->copied_seq
= tp
->rcv_nxt
;
5064 tcp_set_state(sk
, TCP_ESTABLISHED
);
5066 security_inet_conn_established(sk
, skb
);
5068 /* Make sure socket is routed, for correct metrics. */
5069 icsk
->icsk_af_ops
->rebuild_header(sk
);
5071 tcp_init_metrics(sk
);
5073 tcp_init_congestion_control(sk
);
5075 /* Prevent spurious tcp_cwnd_restart() on first data
5078 tp
->lsndtime
= tcp_time_stamp
;
5080 tcp_init_buffer_space(sk
);
5082 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5083 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5085 if (!tp
->rx_opt
.snd_wscale
)
5086 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5090 if (!sock_flag(sk
, SOCK_DEAD
)) {
5091 sk
->sk_state_change(sk
);
5092 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5095 if (sk
->sk_write_pending
||
5096 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5097 icsk
->icsk_ack
.pingpong
) {
5098 /* Save one ACK. Data will be ready after
5099 * several ticks, if write_pending is set.
5101 * It may be deleted, but with this feature tcpdumps
5102 * look so _wonderfully_ clever, that I was not able
5103 * to stand against the temptation 8) --ANK
5105 inet_csk_schedule_ack(sk
);
5106 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5107 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5108 tcp_incr_quickack(sk
);
5109 tcp_enter_quickack_mode(sk
);
5110 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5111 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5122 /* No ACK in the segment */
5126 * "If the RST bit is set
5128 * Otherwise (no ACK) drop the segment and return."
5131 goto discard_and_undo
;
5135 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5136 tcp_paws_check(&tp
->rx_opt
, 0))
5137 goto discard_and_undo
;
5140 /* We see SYN without ACK. It is attempt of
5141 * simultaneous connect with crossed SYNs.
5142 * Particularly, it can be connect to self.
5144 tcp_set_state(sk
, TCP_SYN_RECV
);
5146 if (tp
->rx_opt
.saw_tstamp
) {
5147 tp
->rx_opt
.tstamp_ok
= 1;
5148 tcp_store_ts_recent(tp
);
5149 tp
->tcp_header_len
=
5150 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5152 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5155 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5156 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5158 /* RFC1323: The window in SYN & SYN/ACK segments is
5161 tp
->snd_wnd
= ntohs(th
->window
);
5162 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5163 tp
->max_window
= tp
->snd_wnd
;
5165 TCP_ECN_rcv_syn(tp
, th
);
5168 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5169 tcp_initialize_rcv_mss(sk
);
5171 tcp_send_synack(sk
);
5173 /* Note, we could accept data and URG from this segment.
5174 * There are no obstacles to make this.
5176 * However, if we ignore data in ACKless segments sometimes,
5177 * we have no reasons to accept it sometimes.
5178 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5179 * is not flawless. So, discard packet for sanity.
5180 * Uncomment this return to process the data.
5187 /* "fifth, if neither of the SYN or RST bits is set then
5188 * drop the segment and return."
5192 tcp_clear_options(&tp
->rx_opt
);
5193 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5197 tcp_clear_options(&tp
->rx_opt
);
5198 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5203 * This function implements the receiving procedure of RFC 793 for
5204 * all states except ESTABLISHED and TIME_WAIT.
5205 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5206 * address independent.
5209 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5210 struct tcphdr
*th
, unsigned len
)
5212 struct tcp_sock
*tp
= tcp_sk(sk
);
5213 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5216 tp
->rx_opt
.saw_tstamp
= 0;
5218 switch (sk
->sk_state
) {
5230 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5233 /* Now we have several options: In theory there is
5234 * nothing else in the frame. KA9Q has an option to
5235 * send data with the syn, BSD accepts data with the
5236 * syn up to the [to be] advertised window and
5237 * Solaris 2.1 gives you a protocol error. For now
5238 * we just ignore it, that fits the spec precisely
5239 * and avoids incompatibilities. It would be nice in
5240 * future to drop through and process the data.
5242 * Now that TTCP is starting to be used we ought to
5244 * But, this leaves one open to an easy denial of
5245 * service attack, and SYN cookies can't defend
5246 * against this problem. So, we drop the data
5247 * in the interest of security over speed unless
5248 * it's still in use.
5256 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5260 /* Do step6 onward by hand. */
5261 tcp_urg(sk
, skb
, th
);
5263 tcp_data_snd_check(sk
);
5267 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5268 tcp_paws_discard(sk
, skb
)) {
5270 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5271 tcp_send_dupack(sk
, skb
);
5274 /* Reset is accepted even if it did not pass PAWS. */
5277 /* step 1: check sequence number */
5278 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5280 tcp_send_dupack(sk
, skb
);
5284 /* step 2: check RST bit */
5290 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5292 /* step 3: check security and precedence [ignored] */
5296 * Check for a SYN in window.
5298 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5299 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5304 /* step 5: check the ACK field */
5306 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5308 switch (sk
->sk_state
) {
5311 tp
->copied_seq
= tp
->rcv_nxt
;
5313 tcp_set_state(sk
, TCP_ESTABLISHED
);
5314 sk
->sk_state_change(sk
);
5316 /* Note, that this wakeup is only for marginal
5317 * crossed SYN case. Passively open sockets
5318 * are not waked up, because sk->sk_sleep ==
5319 * NULL and sk->sk_socket == NULL.
5323 SOCK_WAKE_IO
, POLL_OUT
);
5325 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5326 tp
->snd_wnd
= ntohs(th
->window
) <<
5327 tp
->rx_opt
.snd_wscale
;
5328 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5329 TCP_SKB_CB(skb
)->seq
);
5331 /* tcp_ack considers this ACK as duplicate
5332 * and does not calculate rtt.
5333 * Fix it at least with timestamps.
5335 if (tp
->rx_opt
.saw_tstamp
&&
5336 tp
->rx_opt
.rcv_tsecr
&& !tp
->srtt
)
5337 tcp_ack_saw_tstamp(sk
, 0);
5339 if (tp
->rx_opt
.tstamp_ok
)
5340 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5342 /* Make sure socket is routed, for
5345 icsk
->icsk_af_ops
->rebuild_header(sk
);
5347 tcp_init_metrics(sk
);
5349 tcp_init_congestion_control(sk
);
5351 /* Prevent spurious tcp_cwnd_restart() on
5352 * first data packet.
5354 tp
->lsndtime
= tcp_time_stamp
;
5357 tcp_initialize_rcv_mss(sk
);
5358 tcp_init_buffer_space(sk
);
5359 tcp_fast_path_on(tp
);
5366 if (tp
->snd_una
== tp
->write_seq
) {
5367 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5368 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5369 dst_confirm(sk
->sk_dst_cache
);
5371 if (!sock_flag(sk
, SOCK_DEAD
))
5372 /* Wake up lingering close() */
5373 sk
->sk_state_change(sk
);
5377 if (tp
->linger2
< 0 ||
5378 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5379 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5381 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5385 tmo
= tcp_fin_time(sk
);
5386 if (tmo
> TCP_TIMEWAIT_LEN
) {
5387 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5388 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5389 /* Bad case. We could lose such FIN otherwise.
5390 * It is not a big problem, but it looks confusing
5391 * and not so rare event. We still can lose it now,
5392 * if it spins in bh_lock_sock(), but it is really
5395 inet_csk_reset_keepalive_timer(sk
, tmo
);
5397 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5405 if (tp
->snd_una
== tp
->write_seq
) {
5406 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5412 if (tp
->snd_una
== tp
->write_seq
) {
5413 tcp_update_metrics(sk
);
5422 /* step 6: check the URG bit */
5423 tcp_urg(sk
, skb
, th
);
5425 /* step 7: process the segment text */
5426 switch (sk
->sk_state
) {
5427 case TCP_CLOSE_WAIT
:
5430 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5434 /* RFC 793 says to queue data in these states,
5435 * RFC 1122 says we MUST send a reset.
5436 * BSD 4.4 also does reset.
5438 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5439 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5440 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5441 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5447 case TCP_ESTABLISHED
:
5448 tcp_data_queue(sk
, skb
);
5453 /* tcp_data could move socket to TIME-WAIT */
5454 if (sk
->sk_state
!= TCP_CLOSE
) {
5455 tcp_data_snd_check(sk
);
5456 tcp_ack_snd_check(sk
);
5466 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5467 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5468 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
5469 EXPORT_SYMBOL(tcp_parse_options
);
5470 #ifdef CONFIG_TCP_MD5SIG
5471 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
5473 EXPORT_SYMBOL(tcp_rcv_established
);
5474 EXPORT_SYMBOL(tcp_rcv_state_process
);
5475 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);