2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presence of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
75 int sysctl_tcp_timestamps __read_mostly
= 1;
76 int sysctl_tcp_window_scaling __read_mostly
= 1;
77 int sysctl_tcp_sack __read_mostly
= 1;
78 int sysctl_tcp_fack __read_mostly
= 1;
79 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
80 int sysctl_tcp_ecn __read_mostly
;
81 int sysctl_tcp_dsack __read_mostly
= 1;
82 int sysctl_tcp_app_win __read_mostly
= 31;
83 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
85 int sysctl_tcp_stdurg __read_mostly
;
86 int sysctl_tcp_rfc1337 __read_mostly
;
87 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
88 int sysctl_tcp_frto __read_mostly
= 2;
89 int sysctl_tcp_frto_response __read_mostly
;
90 int sysctl_tcp_nometrics_save __read_mostly
;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
93 int sysctl_tcp_abc __read_mostly
;
95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
101 #define FLAG_ECE 0x40 /* ECE in this ACK */
102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
118 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
120 /* Adapt the MSS value used to make delayed ack decision to the
123 static void tcp_measure_rcv_mss(struct sock
*sk
,
124 const struct sk_buff
*skb
)
126 struct inet_connection_sock
*icsk
= inet_csk(sk
);
127 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
130 icsk
->icsk_ack
.last_seg_size
= 0;
132 /* skb->len may jitter because of SACKs, even if peer
133 * sends good full-sized frames.
135 len
= skb_shinfo(skb
)->gso_size
?: skb
->len
;
136 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
137 icsk
->icsk_ack
.rcv_mss
= len
;
139 /* Otherwise, we make more careful check taking into account,
140 * that SACKs block is variable.
142 * "len" is invariant segment length, including TCP header.
144 len
+= skb
->data
- skb_transport_header(skb
);
145 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
146 /* If PSH is not set, packet should be
147 * full sized, provided peer TCP is not badly broken.
148 * This observation (if it is correct 8)) allows
149 * to handle super-low mtu links fairly.
151 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
152 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
153 /* Subtract also invariant (if peer is RFC compliant),
154 * tcp header plus fixed timestamp option length.
155 * Resulting "len" is MSS free of SACK jitter.
157 len
-= tcp_sk(sk
)->tcp_header_len
;
158 icsk
->icsk_ack
.last_seg_size
= len
;
160 icsk
->icsk_ack
.rcv_mss
= len
;
164 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
165 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
166 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
170 static void tcp_incr_quickack(struct sock
*sk
)
172 struct inet_connection_sock
*icsk
= inet_csk(sk
);
173 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
177 if (quickacks
> icsk
->icsk_ack
.quick
)
178 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
181 void tcp_enter_quickack_mode(struct sock
*sk
)
183 struct inet_connection_sock
*icsk
= inet_csk(sk
);
184 tcp_incr_quickack(sk
);
185 icsk
->icsk_ack
.pingpong
= 0;
186 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
189 /* Send ACKs quickly, if "quick" count is not exhausted
190 * and the session is not interactive.
193 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
195 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
196 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
199 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
201 if (tp
->ecn_flags
&TCP_ECN_OK
)
202 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
205 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
207 if (tcp_hdr(skb
)->cwr
)
208 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
213 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
216 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
218 if (tp
->ecn_flags
&TCP_ECN_OK
) {
219 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
220 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
221 /* Funny extension: if ECT is not set on a segment,
222 * it is surely retransmit. It is not in ECN RFC,
223 * but Linux follows this rule. */
224 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
225 tcp_enter_quickack_mode((struct sock
*)tp
);
229 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
231 if ((tp
->ecn_flags
&TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
232 tp
->ecn_flags
&= ~TCP_ECN_OK
;
235 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
237 if ((tp
->ecn_flags
&TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
238 tp
->ecn_flags
&= ~TCP_ECN_OK
;
241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
243 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
&TCP_ECN_OK
))
248 /* Buffer size and advertised window tuning.
250 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
253 static void tcp_fixup_sndbuf(struct sock
*sk
)
255 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
256 sizeof(struct sk_buff
);
258 if (sk
->sk_sndbuf
< 3 * sndmem
)
259 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
264 * All tcp_full_space() is split to two parts: "network" buffer, allocated
265 * forward and advertised in receiver window (tp->rcv_wnd) and
266 * "application buffer", required to isolate scheduling/application
267 * latencies from network.
268 * window_clamp is maximal advertised window. It can be less than
269 * tcp_full_space(), in this case tcp_full_space() - window_clamp
270 * is reserved for "application" buffer. The less window_clamp is
271 * the smoother our behaviour from viewpoint of network, but the lower
272 * throughput and the higher sensitivity of the connection to losses. 8)
274 * rcv_ssthresh is more strict window_clamp used at "slow start"
275 * phase to predict further behaviour of this connection.
276 * It is used for two goals:
277 * - to enforce header prediction at sender, even when application
278 * requires some significant "application buffer". It is check #1.
279 * - to prevent pruning of receive queue because of misprediction
280 * of receiver window. Check #2.
282 * The scheme does not work when sender sends good segments opening
283 * window and then starts to feed us spaghetti. But it should work
284 * in common situations. Otherwise, we have to rely on queue collapsing.
287 /* Slow part of check#2. */
288 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
290 struct tcp_sock
*tp
= tcp_sk(sk
);
292 int truesize
= tcp_win_from_space(skb
->truesize
)/2;
293 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2])/2;
295 while (tp
->rcv_ssthresh
<= window
) {
296 if (truesize
<= skb
->len
)
297 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
305 static void tcp_grow_window(struct sock
*sk
,
308 struct tcp_sock
*tp
= tcp_sk(sk
);
311 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
312 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
313 !tcp_memory_pressure
) {
316 /* Check #2. Increase window, if skb with such overhead
317 * will fit to rcvbuf in future.
319 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
322 incr
= __tcp_grow_window(sk
, skb
);
325 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
, tp
->window_clamp
);
326 inet_csk(sk
)->icsk_ack
.quick
|= 1;
331 /* 3. Tuning rcvbuf, when connection enters established state. */
333 static void tcp_fixup_rcvbuf(struct sock
*sk
)
335 struct tcp_sock
*tp
= tcp_sk(sk
);
336 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
338 /* Try to select rcvbuf so that 4 mss-sized segments
339 * will fit to window and corresponding skbs will fit to our rcvbuf.
340 * (was 3; 4 is minimum to allow fast retransmit to work.)
342 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
344 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
345 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
348 /* 4. Try to fixup all. It is made immediately after connection enters
351 static void tcp_init_buffer_space(struct sock
*sk
)
353 struct tcp_sock
*tp
= tcp_sk(sk
);
356 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
357 tcp_fixup_rcvbuf(sk
);
358 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
359 tcp_fixup_sndbuf(sk
);
361 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
363 maxwin
= tcp_full_space(sk
);
365 if (tp
->window_clamp
>= maxwin
) {
366 tp
->window_clamp
= maxwin
;
368 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
369 tp
->window_clamp
= max(maxwin
-
370 (maxwin
>> sysctl_tcp_app_win
),
374 /* Force reservation of one segment. */
375 if (sysctl_tcp_app_win
&&
376 tp
->window_clamp
> 2 * tp
->advmss
&&
377 tp
->window_clamp
+ tp
->advmss
> maxwin
)
378 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
380 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
381 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
384 /* 5. Recalculate window clamp after socket hit its memory bounds. */
385 static void tcp_clamp_window(struct sock
*sk
)
387 struct tcp_sock
*tp
= tcp_sk(sk
);
388 struct inet_connection_sock
*icsk
= inet_csk(sk
);
390 icsk
->icsk_ack
.quick
= 0;
392 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
393 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
394 !tcp_memory_pressure
&&
395 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
396 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
399 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
400 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U*tp
->advmss
);
404 /* Initialize RCV_MSS value.
405 * RCV_MSS is an our guess about MSS used by the peer.
406 * We haven't any direct information about the MSS.
407 * It's better to underestimate the RCV_MSS rather than overestimate.
408 * Overestimations make us ACKing less frequently than needed.
409 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
411 void tcp_initialize_rcv_mss(struct sock
*sk
)
413 struct tcp_sock
*tp
= tcp_sk(sk
);
414 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
416 hint
= min(hint
, tp
->rcv_wnd
/2);
417 hint
= min(hint
, TCP_MIN_RCVMSS
);
418 hint
= max(hint
, TCP_MIN_MSS
);
420 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
423 /* Receiver "autotuning" code.
425 * The algorithm for RTT estimation w/o timestamps is based on
426 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
427 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
429 * More detail on this code can be found at
430 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
431 * though this reference is out of date. A new paper
434 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
436 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
442 if (new_sample
!= 0) {
443 /* If we sample in larger samples in the non-timestamp
444 * case, we could grossly overestimate the RTT especially
445 * with chatty applications or bulk transfer apps which
446 * are stalled on filesystem I/O.
448 * Also, since we are only going for a minimum in the
449 * non-timestamp case, we do not smooth things out
450 * else with timestamps disabled convergence takes too
454 m
-= (new_sample
>> 3);
456 } else if (m
< new_sample
)
459 /* No previous measure. */
463 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
464 tp
->rcv_rtt_est
.rtt
= new_sample
;
467 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
469 if (tp
->rcv_rtt_est
.time
== 0)
471 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
473 tcp_rcv_rtt_update(tp
,
474 jiffies
- tp
->rcv_rtt_est
.time
,
478 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
479 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
482 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
, const struct sk_buff
*skb
)
484 struct tcp_sock
*tp
= tcp_sk(sk
);
485 if (tp
->rx_opt
.rcv_tsecr
&&
486 (TCP_SKB_CB(skb
)->end_seq
-
487 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
488 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
492 * This function should be called every time data is copied to user space.
493 * It calculates the appropriate TCP receive buffer space.
495 void tcp_rcv_space_adjust(struct sock
*sk
)
497 struct tcp_sock
*tp
= tcp_sk(sk
);
501 if (tp
->rcvq_space
.time
== 0)
504 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
505 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) ||
506 tp
->rcv_rtt_est
.rtt
== 0)
509 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
511 space
= max(tp
->rcvq_space
.space
, space
);
513 if (tp
->rcvq_space
.space
!= space
) {
516 tp
->rcvq_space
.space
= space
;
518 if (sysctl_tcp_moderate_rcvbuf
&&
519 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
520 int new_clamp
= space
;
522 /* Receive space grows, normalize in order to
523 * take into account packet headers and sk_buff
524 * structure overhead.
529 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
530 16 + sizeof(struct sk_buff
));
531 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
534 space
= min(space
, sysctl_tcp_rmem
[2]);
535 if (space
> sk
->sk_rcvbuf
) {
536 sk
->sk_rcvbuf
= space
;
538 /* Make the window clamp follow along. */
539 tp
->window_clamp
= new_clamp
;
545 tp
->rcvq_space
.seq
= tp
->copied_seq
;
546 tp
->rcvq_space
.time
= tcp_time_stamp
;
549 /* There is something which you must keep in mind when you analyze the
550 * behavior of the tp->ato delayed ack timeout interval. When a
551 * connection starts up, we want to ack as quickly as possible. The
552 * problem is that "good" TCP's do slow start at the beginning of data
553 * transmission. The means that until we send the first few ACK's the
554 * sender will sit on his end and only queue most of his data, because
555 * he can only send snd_cwnd unacked packets at any given time. For
556 * each ACK we send, he increments snd_cwnd and transmits more of his
559 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
561 struct tcp_sock
*tp
= tcp_sk(sk
);
562 struct inet_connection_sock
*icsk
= inet_csk(sk
);
565 inet_csk_schedule_ack(sk
);
567 tcp_measure_rcv_mss(sk
, skb
);
569 tcp_rcv_rtt_measure(tp
);
571 now
= tcp_time_stamp
;
573 if (!icsk
->icsk_ack
.ato
) {
574 /* The _first_ data packet received, initialize
575 * delayed ACK engine.
577 tcp_incr_quickack(sk
);
578 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
580 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
582 if (m
<= TCP_ATO_MIN
/2) {
583 /* The fastest case is the first. */
584 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
585 } else if (m
< icsk
->icsk_ack
.ato
) {
586 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
587 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
588 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
589 } else if (m
> icsk
->icsk_rto
) {
590 /* Too long gap. Apparently sender failed to
591 * restart window, so that we send ACKs quickly.
593 tcp_incr_quickack(sk
);
594 sk_stream_mem_reclaim(sk
);
597 icsk
->icsk_ack
.lrcvtime
= now
;
599 TCP_ECN_check_ce(tp
, skb
);
602 tcp_grow_window(sk
, skb
);
605 static u32
tcp_rto_min(struct sock
*sk
)
607 struct dst_entry
*dst
= __sk_dst_get(sk
);
608 u32 rto_min
= TCP_RTO_MIN
;
610 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
611 rto_min
= dst
->metrics
[RTAX_RTO_MIN
-1];
615 /* Called to compute a smoothed rtt estimate. The data fed to this
616 * routine either comes from timestamps, or from segments that were
617 * known _not_ to have been retransmitted [see Karn/Partridge
618 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
619 * piece by Van Jacobson.
620 * NOTE: the next three routines used to be one big routine.
621 * To save cycles in the RFC 1323 implementation it was better to break
622 * it up into three procedures. -- erics
624 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
626 struct tcp_sock
*tp
= tcp_sk(sk
);
627 long m
= mrtt
; /* RTT */
629 /* The following amusing code comes from Jacobson's
630 * article in SIGCOMM '88. Note that rtt and mdev
631 * are scaled versions of rtt and mean deviation.
632 * This is designed to be as fast as possible
633 * m stands for "measurement".
635 * On a 1990 paper the rto value is changed to:
636 * RTO = rtt + 4 * mdev
638 * Funny. This algorithm seems to be very broken.
639 * These formulae increase RTO, when it should be decreased, increase
640 * too slowly, when it should be increased quickly, decrease too quickly
641 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
642 * does not matter how to _calculate_ it. Seems, it was trap
643 * that VJ failed to avoid. 8)
648 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
649 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
651 m
= -m
; /* m is now abs(error) */
652 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
653 /* This is similar to one of Eifel findings.
654 * Eifel blocks mdev updates when rtt decreases.
655 * This solution is a bit different: we use finer gain
656 * for mdev in this case (alpha*beta).
657 * Like Eifel it also prevents growth of rto,
658 * but also it limits too fast rto decreases,
659 * happening in pure Eifel.
664 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
666 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
667 if (tp
->mdev
> tp
->mdev_max
) {
668 tp
->mdev_max
= tp
->mdev
;
669 if (tp
->mdev_max
> tp
->rttvar
)
670 tp
->rttvar
= tp
->mdev_max
;
672 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
673 if (tp
->mdev_max
< tp
->rttvar
)
674 tp
->rttvar
-= (tp
->rttvar
-tp
->mdev_max
)>>2;
675 tp
->rtt_seq
= tp
->snd_nxt
;
676 tp
->mdev_max
= tcp_rto_min(sk
);
679 /* no previous measure. */
680 tp
->srtt
= m
<<3; /* take the measured time to be rtt */
681 tp
->mdev
= m
<<1; /* make sure rto = 3*rtt */
682 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
683 tp
->rtt_seq
= tp
->snd_nxt
;
687 /* Calculate rto without backoff. This is the second half of Van Jacobson's
688 * routine referred to above.
690 static inline void tcp_set_rto(struct sock
*sk
)
692 const struct tcp_sock
*tp
= tcp_sk(sk
);
693 /* Old crap is replaced with new one. 8)
696 * 1. If rtt variance happened to be less 50msec, it is hallucination.
697 * It cannot be less due to utterly erratic ACK generation made
698 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
699 * to do with delayed acks, because at cwnd>2 true delack timeout
700 * is invisible. Actually, Linux-2.4 also generates erratic
701 * ACKs in some circumstances.
703 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
705 /* 2. Fixups made earlier cannot be right.
706 * If we do not estimate RTO correctly without them,
707 * all the algo is pure shit and should be replaced
708 * with correct one. It is exactly, which we pretend to do.
712 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
713 * guarantees that rto is higher.
715 static inline void tcp_bound_rto(struct sock
*sk
)
717 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
718 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
721 /* Save metrics learned by this TCP session.
722 This function is called only, when TCP finishes successfully
723 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
725 void tcp_update_metrics(struct sock
*sk
)
727 struct tcp_sock
*tp
= tcp_sk(sk
);
728 struct dst_entry
*dst
= __sk_dst_get(sk
);
730 if (sysctl_tcp_nometrics_save
)
735 if (dst
&& (dst
->flags
&DST_HOST
)) {
736 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
739 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
740 /* This session failed to estimate rtt. Why?
741 * Probably, no packets returned in time.
744 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
745 dst
->metrics
[RTAX_RTT
-1] = 0;
749 m
= dst_metric(dst
, RTAX_RTT
) - tp
->srtt
;
751 /* If newly calculated rtt larger than stored one,
752 * store new one. Otherwise, use EWMA. Remember,
753 * rtt overestimation is always better than underestimation.
755 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
757 dst
->metrics
[RTAX_RTT
-1] = tp
->srtt
;
759 dst
->metrics
[RTAX_RTT
-1] -= (m
>>3);
762 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
766 /* Scale deviation to rttvar fixed point */
771 if (m
>= dst_metric(dst
, RTAX_RTTVAR
))
772 dst
->metrics
[RTAX_RTTVAR
-1] = m
;
774 dst
->metrics
[RTAX_RTTVAR
-1] -=
775 (dst
->metrics
[RTAX_RTTVAR
-1] - m
)>>2;
778 if (tp
->snd_ssthresh
>= 0xFFFF) {
779 /* Slow start still did not finish. */
780 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
781 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
782 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
783 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
784 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
785 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
786 dst
->metrics
[RTAX_CWND
-1] = tp
->snd_cwnd
;
787 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
788 icsk
->icsk_ca_state
== TCP_CA_Open
) {
789 /* Cong. avoidance phase, cwnd is reliable. */
790 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
791 dst
->metrics
[RTAX_SSTHRESH
-1] =
792 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
793 if (!dst_metric_locked(dst
, RTAX_CWND
))
794 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_cwnd
) >> 1;
796 /* Else slow start did not finish, cwnd is non-sense,
797 ssthresh may be also invalid.
799 if (!dst_metric_locked(dst
, RTAX_CWND
))
800 dst
->metrics
[RTAX_CWND
-1] = (dst
->metrics
[RTAX_CWND
-1] + tp
->snd_ssthresh
) >> 1;
801 if (dst
->metrics
[RTAX_SSTHRESH
-1] &&
802 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
803 tp
->snd_ssthresh
> dst
->metrics
[RTAX_SSTHRESH
-1])
804 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
807 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
808 if (dst
->metrics
[RTAX_REORDERING
-1] < tp
->reordering
&&
809 tp
->reordering
!= sysctl_tcp_reordering
)
810 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
815 /* Numbers are taken from RFC3390.
817 * John Heffner states:
819 * The RFC specifies a window of no more than 4380 bytes
820 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
821 * is a bit misleading because they use a clamp at 4380 bytes
822 * rather than use a multiplier in the relevant range.
824 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
826 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
829 if (tp
->mss_cache
> 1460)
832 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
834 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
837 /* Set slow start threshold and cwnd not falling to slow start */
838 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
840 struct tcp_sock
*tp
= tcp_sk(sk
);
841 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
843 tp
->prior_ssthresh
= 0;
845 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
848 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
849 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
850 tcp_packets_in_flight(tp
) + 1U);
851 tp
->snd_cwnd_cnt
= 0;
852 tp
->high_seq
= tp
->snd_nxt
;
853 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
854 TCP_ECN_queue_cwr(tp
);
856 tcp_set_ca_state(sk
, TCP_CA_CWR
);
861 * Packet counting of FACK is based on in-order assumptions, therefore TCP
862 * disables it when reordering is detected
864 static void tcp_disable_fack(struct tcp_sock
*tp
)
866 /* RFC3517 uses different metric in lost marker => reset on change */
868 tp
->lost_skb_hint
= NULL
;
869 tp
->rx_opt
.sack_ok
&= ~2;
872 /* Take a notice that peer is sending D-SACKs */
873 static void tcp_dsack_seen(struct tcp_sock
*tp
)
875 tp
->rx_opt
.sack_ok
|= 4;
878 /* Initialize metrics on socket. */
880 static void tcp_init_metrics(struct sock
*sk
)
882 struct tcp_sock
*tp
= tcp_sk(sk
);
883 struct dst_entry
*dst
= __sk_dst_get(sk
);
890 if (dst_metric_locked(dst
, RTAX_CWND
))
891 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
892 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
893 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
894 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
895 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
897 if (dst_metric(dst
, RTAX_REORDERING
) &&
898 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
899 tcp_disable_fack(tp
);
900 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
903 if (dst_metric(dst
, RTAX_RTT
) == 0)
906 if (!tp
->srtt
&& dst_metric(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
909 /* Initial rtt is determined from SYN,SYN-ACK.
910 * The segment is small and rtt may appear much
911 * less than real one. Use per-dst memory
912 * to make it more realistic.
914 * A bit of theory. RTT is time passed after "normal" sized packet
915 * is sent until it is ACKed. In normal circumstances sending small
916 * packets force peer to delay ACKs and calculation is correct too.
917 * The algorithm is adaptive and, provided we follow specs, it
918 * NEVER underestimate RTT. BUT! If peer tries to make some clever
919 * tricks sort of "quick acks" for time long enough to decrease RTT
920 * to low value, and then abruptly stops to do it and starts to delay
921 * ACKs, wait for troubles.
923 if (dst_metric(dst
, RTAX_RTT
) > tp
->srtt
) {
924 tp
->srtt
= dst_metric(dst
, RTAX_RTT
);
925 tp
->rtt_seq
= tp
->snd_nxt
;
927 if (dst_metric(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
928 tp
->mdev
= dst_metric(dst
, RTAX_RTTVAR
);
929 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
933 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
935 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
936 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
940 /* Play conservative. If timestamps are not
941 * supported, TCP will fail to recalculate correct
942 * rtt, if initial rto is too small. FORGET ALL AND RESET!
944 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
946 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
947 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
951 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
954 struct tcp_sock
*tp
= tcp_sk(sk
);
955 if (metric
> tp
->reordering
) {
956 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
958 /* This exciting event is worth to be remembered. 8) */
960 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER
);
961 else if (tcp_is_reno(tp
))
962 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER
);
963 else if (tcp_is_fack(tp
))
964 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER
);
966 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER
);
967 #if FASTRETRANS_DEBUG > 1
968 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
969 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
973 tp
->undo_marker
? tp
->undo_retrans
: 0);
975 tcp_disable_fack(tp
);
979 /* This procedure tags the retransmission queue when SACKs arrive.
981 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
982 * Packets in queue with these bits set are counted in variables
983 * sacked_out, retrans_out and lost_out, correspondingly.
985 * Valid combinations are:
986 * Tag InFlight Description
987 * 0 1 - orig segment is in flight.
988 * S 0 - nothing flies, orig reached receiver.
989 * L 0 - nothing flies, orig lost by net.
990 * R 2 - both orig and retransmit are in flight.
991 * L|R 1 - orig is lost, retransmit is in flight.
992 * S|R 1 - orig reached receiver, retrans is still in flight.
993 * (L|S|R is logically valid, it could occur when L|R is sacked,
994 * but it is equivalent to plain S and code short-curcuits it to S.
995 * L|S is logically invalid, it would mean -1 packet in flight 8))
997 * These 6 states form finite state machine, controlled by the following events:
998 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
999 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1000 * 3. Loss detection event of one of three flavors:
1001 * A. Scoreboard estimator decided the packet is lost.
1002 * A'. Reno "three dupacks" marks head of queue lost.
1003 * A''. Its FACK modfication, head until snd.fack is lost.
1004 * B. SACK arrives sacking data transmitted after never retransmitted
1005 * hole was sent out.
1006 * C. SACK arrives sacking SND.NXT at the moment, when the
1007 * segment was retransmitted.
1008 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1010 * It is pleasant to note, that state diagram turns out to be commutative,
1011 * so that we are allowed not to be bothered by order of our actions,
1012 * when multiple events arrive simultaneously. (see the function below).
1014 * Reordering detection.
1015 * --------------------
1016 * Reordering metric is maximal distance, which a packet can be displaced
1017 * in packet stream. With SACKs we can estimate it:
1019 * 1. SACK fills old hole and the corresponding segment was not
1020 * ever retransmitted -> reordering. Alas, we cannot use it
1021 * when segment was retransmitted.
1022 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1023 * for retransmitted and already SACKed segment -> reordering..
1024 * Both of these heuristics are not used in Loss state, when we cannot
1025 * account for retransmits accurately.
1027 * SACK block validation.
1028 * ----------------------
1030 * SACK block range validation checks that the received SACK block fits to
1031 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1032 * Note that SND.UNA is not included to the range though being valid because
1033 * it means that the receiver is rather inconsistent with itself reporting
1034 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1035 * perfectly valid, however, in light of RFC2018 which explicitly states
1036 * that "SACK block MUST reflect the newest segment. Even if the newest
1037 * segment is going to be discarded ...", not that it looks very clever
1038 * in case of head skb. Due to potentional receiver driven attacks, we
1039 * choose to avoid immediate execution of a walk in write queue due to
1040 * reneging and defer head skb's loss recovery to standard loss recovery
1041 * procedure that will eventually trigger (nothing forbids us doing this).
1043 * Implements also blockage to start_seq wrap-around. Problem lies in the
1044 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1045 * there's no guarantee that it will be before snd_nxt (n). The problem
1046 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1049 * <- outs wnd -> <- wrapzone ->
1050 * u e n u_w e_w s n_w
1052 * |<------------+------+----- TCP seqno space --------------+---------->|
1053 * ...-- <2^31 ->| |<--------...
1054 * ...---- >2^31 ------>| |<--------...
1056 * Current code wouldn't be vulnerable but it's better still to discard such
1057 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1058 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1059 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1060 * equal to the ideal case (infinite seqno space without wrap caused issues).
1062 * With D-SACK the lower bound is extended to cover sequence space below
1063 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1064 * again, D-SACK block must not to go across snd_una (for the same reason as
1065 * for the normal SACK blocks, explained above). But there all simplicity
1066 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1067 * fully below undo_marker they do not affect behavior in anyway and can
1068 * therefore be safely ignored. In rare cases (which are more or less
1069 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1070 * fragmentation and packet reordering past skb's retransmission. To consider
1071 * them correctly, the acceptable range must be extended even more though
1072 * the exact amount is rather hard to quantify. However, tp->max_window can
1073 * be used as an exaggerated estimate.
1075 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1076 u32 start_seq
, u32 end_seq
)
1078 /* Too far in future, or reversed (interpretation is ambiguous) */
1079 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1082 /* Nasty start_seq wrap-around check (see comments above) */
1083 if (!before(start_seq
, tp
->snd_nxt
))
1086 /* In outstanding window? ...This is valid exit for D-SACKs too.
1087 * start_seq == snd_una is non-sensical (see comments above)
1089 if (after(start_seq
, tp
->snd_una
))
1092 if (!is_dsack
|| !tp
->undo_marker
)
1095 /* ...Then it's D-SACK, and must reside below snd_una completely */
1096 if (!after(end_seq
, tp
->snd_una
))
1099 if (!before(start_seq
, tp
->undo_marker
))
1103 if (!after(end_seq
, tp
->undo_marker
))
1106 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1107 * start_seq < undo_marker and end_seq >= undo_marker.
1109 return !before(start_seq
, end_seq
- tp
->max_window
);
1112 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1113 * Event "C". Later note: FACK people cheated me again 8), we have to account
1114 * for reordering! Ugly, but should help.
1116 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1117 * less than what is now known to be received by the other end (derived from
1118 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1119 * retransmitted skbs to avoid some costly processing per ACKs.
1121 static void tcp_mark_lost_retrans(struct sock
*sk
)
1123 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1124 struct tcp_sock
*tp
= tcp_sk(sk
);
1125 struct sk_buff
*skb
;
1127 u32 new_low_seq
= tp
->snd_nxt
;
1128 u32 received_upto
= TCP_SKB_CB(tp
->highest_sack
)->end_seq
;
1130 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1131 !after(received_upto
, tp
->lost_retrans_low
) ||
1132 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1135 tcp_for_write_queue(skb
, sk
) {
1136 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1138 if (skb
== tcp_send_head(sk
))
1140 if (cnt
== tp
->retrans_out
)
1142 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1145 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1148 if (after(received_upto
, ack_seq
) &&
1150 !before(received_upto
,
1151 ack_seq
+ tp
->reordering
* tp
->mss_cache
))) {
1152 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1153 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1155 /* clear lost hint */
1156 tp
->retransmit_skb_hint
= NULL
;
1158 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1159 tp
->lost_out
+= tcp_skb_pcount(skb
);
1160 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1162 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1164 if (before(ack_seq
, new_low_seq
))
1165 new_low_seq
= ack_seq
;
1166 cnt
+= tcp_skb_pcount(skb
);
1170 if (tp
->retrans_out
)
1171 tp
->lost_retrans_low
= new_low_seq
;
1174 static int tcp_check_dsack(struct tcp_sock
*tp
, struct sk_buff
*ack_skb
,
1175 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1178 u32 start_seq_0
= ntohl(get_unaligned(&sp
[0].start_seq
));
1179 u32 end_seq_0
= ntohl(get_unaligned(&sp
[0].end_seq
));
1182 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1185 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
1186 } else if (num_sacks
> 1) {
1187 u32 end_seq_1
= ntohl(get_unaligned(&sp
[1].end_seq
));
1188 u32 start_seq_1
= ntohl(get_unaligned(&sp
[1].start_seq
));
1190 if (!after(end_seq_0
, end_seq_1
) &&
1191 !before(start_seq_0
, start_seq_1
)) {
1194 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
1198 /* D-SACK for already forgotten data... Do dumb counting. */
1200 !after(end_seq_0
, prior_snd_una
) &&
1201 after(end_seq_0
, tp
->undo_marker
))
1207 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1208 * the incoming SACK may not exactly match but we can find smaller MSS
1209 * aligned portion of it that matches. Therefore we might need to fragment
1210 * which may fail and creates some hassle (caller must handle error case
1213 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1214 u32 start_seq
, u32 end_seq
)
1217 unsigned int pkt_len
;
1219 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1220 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1222 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1223 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1225 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1228 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1230 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1231 err
= tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
);
1239 static int tcp_sacktag_one(struct sk_buff
*skb
, struct tcp_sock
*tp
,
1240 int *reord
, int dup_sack
, int fack_count
)
1242 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1245 /* Account D-SACK for retransmitted packet. */
1246 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1247 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1249 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
) &&
1250 (sacked
& TCPCB_SACKED_ACKED
))
1251 *reord
= min(fack_count
, *reord
);
1254 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1255 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1258 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1259 if (sacked
& TCPCB_SACKED_RETRANS
) {
1260 /* If the segment is not tagged as lost,
1261 * we do not clear RETRANS, believing
1262 * that retransmission is still in flight.
1264 if (sacked
& TCPCB_LOST
) {
1265 TCP_SKB_CB(skb
)->sacked
&=
1266 ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1267 tp
->lost_out
-= tcp_skb_pcount(skb
);
1268 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1270 /* clear lost hint */
1271 tp
->retransmit_skb_hint
= NULL
;
1274 if (!(sacked
& TCPCB_RETRANS
)) {
1275 /* New sack for not retransmitted frame,
1276 * which was in hole. It is reordering.
1278 if (before(TCP_SKB_CB(skb
)->seq
,
1279 tcp_highest_sack_seq(tp
)))
1280 *reord
= min(fack_count
, *reord
);
1282 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1283 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1284 flag
|= FLAG_ONLY_ORIG_SACKED
;
1287 if (sacked
& TCPCB_LOST
) {
1288 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1289 tp
->lost_out
-= tcp_skb_pcount(skb
);
1291 /* clear lost hint */
1292 tp
->retransmit_skb_hint
= NULL
;
1296 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1297 flag
|= FLAG_DATA_SACKED
;
1298 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1300 fack_count
+= tcp_skb_pcount(skb
);
1302 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1303 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1304 before(TCP_SKB_CB(skb
)->seq
,
1305 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1306 tp
->lost_cnt_hint
+= tcp_skb_pcount(skb
);
1308 if (fack_count
> tp
->fackets_out
)
1309 tp
->fackets_out
= fack_count
;
1311 if (after(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
1312 tp
->highest_sack
= skb
;
1315 if (dup_sack
&& (sacked
& TCPCB_RETRANS
))
1316 *reord
= min(fack_count
, *reord
);
1319 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1320 * frames and clear it. undo_retrans is decreased above, L|R frames
1321 * are accounted above as well.
1323 if (dup_sack
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)) {
1324 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1325 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1326 tp
->retransmit_skb_hint
= NULL
;
1332 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1333 struct tcp_sack_block
*next_dup
,
1334 u32 start_seq
, u32 end_seq
,
1335 int dup_sack_in
, int *fack_count
,
1336 int *reord
, int *flag
)
1338 struct tcp_sock
*tp
= tcp_sk(sk
);
1340 tcp_for_write_queue_from(skb
, sk
) {
1342 int dup_sack
= dup_sack_in
;
1344 if (skb
== tcp_send_head(sk
))
1347 /* queue is in-order => we can short-circuit the walk early */
1348 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1351 if ((next_dup
!= NULL
) &&
1352 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1353 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1354 next_dup
->start_seq
,
1361 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
, end_seq
);
1362 if (unlikely(in_sack
< 0))
1366 *flag
|= tcp_sacktag_one(skb
, tp
, reord
, dup_sack
, *fack_count
);
1368 *fack_count
+= tcp_skb_pcount(skb
);
1373 /* Avoid all extra work that is being done by sacktag while walking in
1376 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1379 tcp_for_write_queue_from(skb
, sk
) {
1380 if (skb
== tcp_send_head(sk
))
1383 if (!before(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1389 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1391 struct tcp_sack_block
*next_dup
,
1393 int *fack_count
, int *reord
,
1396 if (next_dup
== NULL
)
1399 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1400 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
);
1401 tcp_sacktag_walk(skb
, sk
, NULL
,
1402 next_dup
->start_seq
, next_dup
->end_seq
,
1403 1, fack_count
, reord
, flag
);
1409 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1411 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1415 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
1417 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1418 struct tcp_sock
*tp
= tcp_sk(sk
);
1419 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1420 TCP_SKB_CB(ack_skb
)->sacked
);
1421 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1422 struct tcp_sack_block sp
[4];
1423 struct tcp_sack_block
*cache
;
1424 struct sk_buff
*skb
;
1425 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
1427 int reord
= tp
->packets_out
;
1429 int found_dup_sack
= 0;
1432 int first_sack_index
;
1434 if (!tp
->sacked_out
) {
1435 if (WARN_ON(tp
->fackets_out
))
1436 tp
->fackets_out
= 0;
1437 tp
->highest_sack
= tcp_write_queue_head(sk
);
1440 found_dup_sack
= tcp_check_dsack(tp
, ack_skb
, sp_wire
,
1441 num_sacks
, prior_snd_una
);
1443 flag
|= FLAG_DSACKING_ACK
;
1445 /* Eliminate too old ACKs, but take into
1446 * account more or less fresh ones, they can
1447 * contain valid SACK info.
1449 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1452 if (!tp
->packets_out
)
1456 first_sack_index
= 0;
1457 for (i
= 0; i
< num_sacks
; i
++) {
1458 int dup_sack
= !i
&& found_dup_sack
;
1460 sp
[used_sacks
].start_seq
= ntohl(get_unaligned(&sp_wire
[i
].start_seq
));
1461 sp
[used_sacks
].end_seq
= ntohl(get_unaligned(&sp_wire
[i
].end_seq
));
1463 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1464 sp
[used_sacks
].start_seq
,
1465 sp
[used_sacks
].end_seq
)) {
1467 if (!tp
->undo_marker
)
1468 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO
);
1470 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD
);
1472 /* Don't count olds caused by ACK reordering */
1473 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1474 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1476 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD
);
1479 first_sack_index
= -1;
1483 /* Ignore very old stuff early */
1484 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1490 /* order SACK blocks to allow in order walk of the retrans queue */
1491 for (i
= used_sacks
- 1; i
> 0; i
--) {
1492 for (j
= 0; j
< i
; j
++){
1493 if (after(sp
[j
].start_seq
, sp
[j
+1].start_seq
)) {
1494 struct tcp_sack_block tmp
;
1500 /* Track where the first SACK block goes to */
1501 if (j
== first_sack_index
)
1502 first_sack_index
= j
+1;
1507 skb
= tcp_write_queue_head(sk
);
1511 if (!tp
->sacked_out
) {
1512 /* It's already past, so skip checking against it */
1513 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1515 cache
= tp
->recv_sack_cache
;
1516 /* Skip empty blocks in at head of the cache */
1517 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1522 while (i
< used_sacks
) {
1523 u32 start_seq
= sp
[i
].start_seq
;
1524 u32 end_seq
= sp
[i
].end_seq
;
1525 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1526 struct tcp_sack_block
*next_dup
= NULL
;
1528 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1529 next_dup
= &sp
[i
+ 1];
1531 /* Event "B" in the comment above. */
1532 if (after(end_seq
, tp
->high_seq
))
1533 flag
|= FLAG_DATA_LOST
;
1535 /* Skip too early cached blocks */
1536 while (tcp_sack_cache_ok(tp
, cache
) &&
1537 !before(start_seq
, cache
->end_seq
))
1540 /* Can skip some work by looking recv_sack_cache? */
1541 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1542 after(end_seq
, cache
->start_seq
)) {
1545 if (before(start_seq
, cache
->start_seq
)) {
1546 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1547 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
,
1548 cache
->start_seq
, dup_sack
,
1549 &fack_count
, &reord
, &flag
);
1552 /* Rest of the block already fully processed? */
1553 if (!after(end_seq
, cache
->end_seq
))
1556 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
, cache
->end_seq
,
1557 &fack_count
, &reord
, &flag
);
1559 /* ...tail remains todo... */
1560 if (TCP_SKB_CB(tp
->highest_sack
)->end_seq
== cache
->end_seq
) {
1561 /* ...but better entrypoint exists! */
1562 skb
= tcp_write_queue_next(sk
, tp
->highest_sack
);
1563 fack_count
= tp
->fackets_out
;
1568 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
);
1569 /* Check overlap against next cached too (past this one already) */
1574 if (tp
->sacked_out
&& !before(start_seq
, tcp_highest_sack_seq(tp
))) {
1575 skb
= tcp_write_queue_next(sk
, tp
->highest_sack
);
1576 fack_count
= tp
->fackets_out
;
1578 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1581 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1582 dup_sack
, &fack_count
, &reord
, &flag
);
1585 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1586 * due to in-order walk
1588 if (after(end_seq
, tp
->frto_highmark
))
1589 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1594 /* Clear the head of the cache sack blocks so we can skip it next time */
1595 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1596 tp
->recv_sack_cache
[i
].start_seq
= 0;
1597 tp
->recv_sack_cache
[i
].end_seq
= 0;
1599 for (j
= 0; j
< used_sacks
; j
++)
1600 tp
->recv_sack_cache
[i
++] = sp
[j
];
1602 tcp_mark_lost_retrans(sk
);
1604 tcp_verify_left_out(tp
);
1606 if ((reord
< tp
->fackets_out
) &&
1607 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1608 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1609 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1613 #if FASTRETRANS_DEBUG > 0
1614 BUG_TRAP((int)tp
->sacked_out
>= 0);
1615 BUG_TRAP((int)tp
->lost_out
>= 0);
1616 BUG_TRAP((int)tp
->retrans_out
>= 0);
1617 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1622 /* If we receive more dupacks than we expected counting segments
1623 * in assumption of absent reordering, interpret this as reordering.
1624 * The only another reason could be bug in receiver TCP.
1626 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1628 struct tcp_sock
*tp
= tcp_sk(sk
);
1631 holes
= max(tp
->lost_out
, 1U);
1632 holes
= min(holes
, tp
->packets_out
);
1634 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1635 tp
->sacked_out
= tp
->packets_out
- holes
;
1636 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1640 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1642 static void tcp_add_reno_sack(struct sock
*sk
)
1644 struct tcp_sock
*tp
= tcp_sk(sk
);
1646 tcp_check_reno_reordering(sk
, 0);
1647 tcp_verify_left_out(tp
);
1650 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1652 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1654 struct tcp_sock
*tp
= tcp_sk(sk
);
1657 /* One ACK acked hole. The rest eat duplicate ACKs. */
1658 if (acked
-1 >= tp
->sacked_out
)
1661 tp
->sacked_out
-= acked
-1;
1663 tcp_check_reno_reordering(sk
, acked
);
1664 tcp_verify_left_out(tp
);
1667 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1672 /* F-RTO can only be used if TCP has never retransmitted anything other than
1673 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1675 int tcp_use_frto(struct sock
*sk
)
1677 const struct tcp_sock
*tp
= tcp_sk(sk
);
1678 struct sk_buff
*skb
;
1680 if (!sysctl_tcp_frto
)
1686 /* Avoid expensive walking of rexmit queue if possible */
1687 if (tp
->retrans_out
> 1)
1690 skb
= tcp_write_queue_head(sk
);
1691 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1692 tcp_for_write_queue_from(skb
, sk
) {
1693 if (skb
== tcp_send_head(sk
))
1695 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1697 /* Short-circuit when first non-SACKed skb has been checked */
1698 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1704 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1705 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1706 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1707 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1708 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1709 * bits are handled if the Loss state is really to be entered (in
1710 * tcp_enter_frto_loss).
1712 * Do like tcp_enter_loss() would; when RTO expires the second time it
1714 * "Reduce ssthresh if it has not yet been made inside this window."
1716 void tcp_enter_frto(struct sock
*sk
)
1718 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1719 struct tcp_sock
*tp
= tcp_sk(sk
);
1720 struct sk_buff
*skb
;
1722 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1723 tp
->snd_una
== tp
->high_seq
||
1724 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1725 !icsk
->icsk_retransmits
)) {
1726 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1727 /* Our state is too optimistic in ssthresh() call because cwnd
1728 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1729 * recovery has not yet completed. Pattern would be this: RTO,
1730 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1732 * RFC4138 should be more specific on what to do, even though
1733 * RTO is quite unlikely to occur after the first Cumulative ACK
1734 * due to back-off and complexity of triggering events ...
1736 if (tp
->frto_counter
) {
1738 stored_cwnd
= tp
->snd_cwnd
;
1740 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1741 tp
->snd_cwnd
= stored_cwnd
;
1743 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1745 /* ... in theory, cong.control module could do "any tricks" in
1746 * ssthresh(), which means that ca_state, lost bits and lost_out
1747 * counter would have to be faked before the call occurs. We
1748 * consider that too expensive, unlikely and hacky, so modules
1749 * using these in ssthresh() must deal these incompatibility
1750 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1752 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1755 tp
->undo_marker
= tp
->snd_una
;
1756 tp
->undo_retrans
= 0;
1758 skb
= tcp_write_queue_head(sk
);
1759 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1760 tp
->undo_marker
= 0;
1761 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1762 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1763 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1765 tcp_verify_left_out(tp
);
1767 /* Too bad if TCP was application limited */
1768 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1770 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1771 * The last condition is necessary at least in tp->frto_counter case.
1773 if (IsSackFrto() && (tp
->frto_counter
||
1774 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1775 after(tp
->high_seq
, tp
->snd_una
)) {
1776 tp
->frto_highmark
= tp
->high_seq
;
1778 tp
->frto_highmark
= tp
->snd_nxt
;
1780 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1781 tp
->high_seq
= tp
->snd_nxt
;
1782 tp
->frto_counter
= 1;
1785 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1786 * which indicates that we should follow the traditional RTO recovery,
1787 * i.e. mark everything lost and do go-back-N retransmission.
1789 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1791 struct tcp_sock
*tp
= tcp_sk(sk
);
1792 struct sk_buff
*skb
;
1795 tp
->retrans_out
= 0;
1796 if (tcp_is_reno(tp
))
1797 tcp_reset_reno_sack(tp
);
1799 tcp_for_write_queue(skb
, sk
) {
1800 if (skb
== tcp_send_head(sk
))
1803 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1805 * Count the retransmission made on RTO correctly (only when
1806 * waiting for the first ACK and did not get it)...
1808 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1809 /* For some reason this R-bit might get cleared? */
1810 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1811 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1812 /* ...enter this if branch just for the first segment */
1813 flag
|= FLAG_DATA_ACKED
;
1815 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1816 tp
->undo_marker
= 0;
1817 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1820 /* Don't lost mark skbs that were fwd transmitted after RTO */
1821 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) &&
1822 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
)) {
1823 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1824 tp
->lost_out
+= tcp_skb_pcount(skb
);
1827 tcp_verify_left_out(tp
);
1829 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1830 tp
->snd_cwnd_cnt
= 0;
1831 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1832 tp
->frto_counter
= 0;
1833 tp
->bytes_acked
= 0;
1835 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1836 sysctl_tcp_reordering
);
1837 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1838 tp
->high_seq
= tp
->frto_highmark
;
1839 TCP_ECN_queue_cwr(tp
);
1841 tcp_clear_retrans_hints_partial(tp
);
1844 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1846 tp
->retrans_out
= 0;
1849 tp
->undo_marker
= 0;
1850 tp
->undo_retrans
= 0;
1853 void tcp_clear_retrans(struct tcp_sock
*tp
)
1855 tcp_clear_retrans_partial(tp
);
1857 tp
->fackets_out
= 0;
1861 /* Enter Loss state. If "how" is not zero, forget all SACK information
1862 * and reset tags completely, otherwise preserve SACKs. If receiver
1863 * dropped its ofo queue, we will know this due to reneging detection.
1865 void tcp_enter_loss(struct sock
*sk
, int how
)
1867 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1868 struct tcp_sock
*tp
= tcp_sk(sk
);
1869 struct sk_buff
*skb
;
1871 /* Reduce ssthresh if it has not yet been made inside this window. */
1872 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1873 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1874 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1875 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1876 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1879 tp
->snd_cwnd_cnt
= 0;
1880 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1882 tp
->bytes_acked
= 0;
1883 tcp_clear_retrans_partial(tp
);
1885 if (tcp_is_reno(tp
))
1886 tcp_reset_reno_sack(tp
);
1889 /* Push undo marker, if it was plain RTO and nothing
1890 * was retransmitted. */
1891 tp
->undo_marker
= tp
->snd_una
;
1892 tcp_clear_retrans_hints_partial(tp
);
1895 tp
->fackets_out
= 0;
1896 tcp_clear_all_retrans_hints(tp
);
1899 tcp_for_write_queue(skb
, sk
) {
1900 if (skb
== tcp_send_head(sk
))
1903 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1904 tp
->undo_marker
= 0;
1905 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1906 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1907 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1908 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1909 tp
->lost_out
+= tcp_skb_pcount(skb
);
1912 tcp_verify_left_out(tp
);
1914 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1915 sysctl_tcp_reordering
);
1916 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1917 tp
->high_seq
= tp
->snd_nxt
;
1918 TCP_ECN_queue_cwr(tp
);
1919 /* Abort F-RTO algorithm if one is in progress */
1920 tp
->frto_counter
= 0;
1923 static int tcp_check_sack_reneging(struct sock
*sk
)
1925 struct sk_buff
*skb
;
1927 /* If ACK arrived pointing to a remembered SACK,
1928 * it means that our remembered SACKs do not reflect
1929 * real state of receiver i.e.
1930 * receiver _host_ is heavily congested (or buggy).
1931 * Do processing similar to RTO timeout.
1933 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1934 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1935 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1936 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1938 tcp_enter_loss(sk
, 1);
1939 icsk
->icsk_retransmits
++;
1940 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1941 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1942 icsk
->icsk_rto
, TCP_RTO_MAX
);
1948 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1950 return tcp_is_reno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1953 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1954 * counter when SACK is enabled (without SACK, sacked_out is used for
1957 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1958 * segments up to the highest received SACK block so far and holes in
1961 * With reordering, holes may still be in flight, so RFC3517 recovery
1962 * uses pure sacked_out (total number of SACKed segments) even though
1963 * it violates the RFC that uses duplicate ACKs, often these are equal
1964 * but when e.g. out-of-window ACKs or packet duplication occurs,
1965 * they differ. Since neither occurs due to loss, TCP should really
1968 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
1970 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1973 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1975 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1978 static inline int tcp_head_timedout(struct sock
*sk
)
1980 struct tcp_sock
*tp
= tcp_sk(sk
);
1982 return tp
->packets_out
&&
1983 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1986 /* Linux NewReno/SACK/FACK/ECN state machine.
1987 * --------------------------------------
1989 * "Open" Normal state, no dubious events, fast path.
1990 * "Disorder" In all the respects it is "Open",
1991 * but requires a bit more attention. It is entered when
1992 * we see some SACKs or dupacks. It is split of "Open"
1993 * mainly to move some processing from fast path to slow one.
1994 * "CWR" CWND was reduced due to some Congestion Notification event.
1995 * It can be ECN, ICMP source quench, local device congestion.
1996 * "Recovery" CWND was reduced, we are fast-retransmitting.
1997 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1999 * tcp_fastretrans_alert() is entered:
2000 * - each incoming ACK, if state is not "Open"
2001 * - when arrived ACK is unusual, namely:
2006 * Counting packets in flight is pretty simple.
2008 * in_flight = packets_out - left_out + retrans_out
2010 * packets_out is SND.NXT-SND.UNA counted in packets.
2012 * retrans_out is number of retransmitted segments.
2014 * left_out is number of segments left network, but not ACKed yet.
2016 * left_out = sacked_out + lost_out
2018 * sacked_out: Packets, which arrived to receiver out of order
2019 * and hence not ACKed. With SACKs this number is simply
2020 * amount of SACKed data. Even without SACKs
2021 * it is easy to give pretty reliable estimate of this number,
2022 * counting duplicate ACKs.
2024 * lost_out: Packets lost by network. TCP has no explicit
2025 * "loss notification" feedback from network (for now).
2026 * It means that this number can be only _guessed_.
2027 * Actually, it is the heuristics to predict lossage that
2028 * distinguishes different algorithms.
2030 * F.e. after RTO, when all the queue is considered as lost,
2031 * lost_out = packets_out and in_flight = retrans_out.
2033 * Essentially, we have now two algorithms counting
2036 * FACK: It is the simplest heuristics. As soon as we decided
2037 * that something is lost, we decide that _all_ not SACKed
2038 * packets until the most forward SACK are lost. I.e.
2039 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2040 * It is absolutely correct estimate, if network does not reorder
2041 * packets. And it loses any connection to reality when reordering
2042 * takes place. We use FACK by default until reordering
2043 * is suspected on the path to this destination.
2045 * NewReno: when Recovery is entered, we assume that one segment
2046 * is lost (classic Reno). While we are in Recovery and
2047 * a partial ACK arrives, we assume that one more packet
2048 * is lost (NewReno). This heuristics are the same in NewReno
2051 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2052 * deflation etc. CWND is real congestion window, never inflated, changes
2053 * only according to classic VJ rules.
2055 * Really tricky (and requiring careful tuning) part of algorithm
2056 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2057 * The first determines the moment _when_ we should reduce CWND and,
2058 * hence, slow down forward transmission. In fact, it determines the moment
2059 * when we decide that hole is caused by loss, rather than by a reorder.
2061 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2062 * holes, caused by lost packets.
2064 * And the most logically complicated part of algorithm is undo
2065 * heuristics. We detect false retransmits due to both too early
2066 * fast retransmit (reordering) and underestimated RTO, analyzing
2067 * timestamps and D-SACKs. When we detect that some segments were
2068 * retransmitted by mistake and CWND reduction was wrong, we undo
2069 * window reduction and abort recovery phase. This logic is hidden
2070 * inside several functions named tcp_try_undo_<something>.
2073 /* This function decides, when we should leave Disordered state
2074 * and enter Recovery phase, reducing congestion window.
2076 * Main question: may we further continue forward transmission
2077 * with the same cwnd?
2079 static int tcp_time_to_recover(struct sock
*sk
)
2081 struct tcp_sock
*tp
= tcp_sk(sk
);
2084 /* Do not perform any recovery during F-RTO algorithm */
2085 if (tp
->frto_counter
)
2088 /* Trick#1: The loss is proven. */
2092 /* Not-A-Trick#2 : Classic rule... */
2093 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2096 /* Trick#3 : when we use RFC2988 timer restart, fast
2097 * retransmit can be triggered by timeout of queue head.
2099 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2102 /* Trick#4: It is still not OK... But will it be useful to delay
2105 packets_out
= tp
->packets_out
;
2106 if (packets_out
<= tp
->reordering
&&
2107 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2108 !tcp_may_send_now(sk
)) {
2109 /* We have nothing to send. This connection is limited
2110 * either by receiver window or by application.
2118 /* RFC: This is from the original, I doubt that this is necessary at all:
2119 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2120 * retransmitted past LOST markings in the first place? I'm not fully sure
2121 * about undo and end of connection cases, which can cause R without L?
2123 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
2124 struct sk_buff
*skb
)
2126 if ((tp
->retransmit_skb_hint
!= NULL
) &&
2127 before(TCP_SKB_CB(skb
)->seq
,
2128 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
2129 tp
->retransmit_skb_hint
= NULL
;
2132 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2133 * is against sacked "cnt", otherwise it's against facked "cnt"
2135 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int fast_rexmit
)
2137 struct tcp_sock
*tp
= tcp_sk(sk
);
2138 struct sk_buff
*skb
;
2141 BUG_TRAP(packets
<= tp
->packets_out
);
2142 if (tp
->lost_skb_hint
) {
2143 skb
= tp
->lost_skb_hint
;
2144 cnt
= tp
->lost_cnt_hint
;
2146 skb
= tcp_write_queue_head(sk
);
2150 tcp_for_write_queue_from(skb
, sk
) {
2151 if (skb
== tcp_send_head(sk
))
2153 /* TODO: do this better */
2154 /* this is not the most efficient way to do this... */
2155 tp
->lost_skb_hint
= skb
;
2156 tp
->lost_cnt_hint
= cnt
;
2158 if (tcp_is_fack(tp
) ||
2159 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2160 cnt
+= tcp_skb_pcount(skb
);
2162 if (((!fast_rexmit
|| (tp
->lost_out
> 0)) && (cnt
> packets
)) ||
2163 after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2165 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2166 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2167 tp
->lost_out
+= tcp_skb_pcount(skb
);
2168 tcp_verify_retransmit_hint(tp
, skb
);
2171 tcp_verify_left_out(tp
);
2174 /* Account newly detected lost packet(s) */
2176 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2178 struct tcp_sock
*tp
= tcp_sk(sk
);
2180 if (tcp_is_reno(tp
)) {
2181 tcp_mark_head_lost(sk
, 1, fast_rexmit
);
2182 } else if (tcp_is_fack(tp
)) {
2183 int lost
= tp
->fackets_out
- tp
->reordering
;
2186 tcp_mark_head_lost(sk
, lost
, fast_rexmit
);
2188 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2189 if (sacked_upto
< 0)
2191 tcp_mark_head_lost(sk
, sacked_upto
, fast_rexmit
);
2194 /* New heuristics: it is possible only after we switched
2195 * to restart timer each time when something is ACKed.
2196 * Hence, we can detect timed out packets during fast
2197 * retransmit without falling to slow start.
2199 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2200 struct sk_buff
*skb
;
2202 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2203 : tcp_write_queue_head(sk
);
2205 tcp_for_write_queue_from(skb
, sk
) {
2206 if (skb
== tcp_send_head(sk
))
2208 if (!tcp_skb_timedout(sk
, skb
))
2211 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2212 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2213 tp
->lost_out
+= tcp_skb_pcount(skb
);
2214 tcp_verify_retransmit_hint(tp
, skb
);
2218 tp
->scoreboard_skb_hint
= skb
;
2220 tcp_verify_left_out(tp
);
2224 /* CWND moderation, preventing bursts due to too big ACKs
2225 * in dubious situations.
2227 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2229 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2230 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
2231 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2234 /* Lower bound on congestion window is slow start threshold
2235 * unless congestion avoidance choice decides to overide it.
2237 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2239 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2241 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2244 /* Decrease cwnd each second ack. */
2245 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2247 struct tcp_sock
*tp
= tcp_sk(sk
);
2248 int decr
= tp
->snd_cwnd_cnt
+ 1;
2250 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
2251 (tcp_is_reno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
2252 tp
->snd_cwnd_cnt
= decr
&1;
2255 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2256 tp
->snd_cwnd
-= decr
;
2258 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
2259 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2263 /* Nothing was retransmitted or returned timestamp is less
2264 * than timestamp of the first retransmission.
2266 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2268 return !tp
->retrans_stamp
||
2269 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2270 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
2273 /* Undo procedures. */
2275 #if FASTRETRANS_DEBUG > 1
2276 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2278 struct tcp_sock
*tp
= tcp_sk(sk
);
2279 struct inet_sock
*inet
= inet_sk(sk
);
2281 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2283 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2284 tp
->snd_cwnd
, tcp_left_out(tp
),
2285 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2289 #define DBGUNDO(x...) do { } while (0)
2292 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2294 struct tcp_sock
*tp
= tcp_sk(sk
);
2296 if (tp
->prior_ssthresh
) {
2297 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2299 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2300 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2302 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
2304 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2305 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2306 TCP_ECN_withdraw_cwr(tp
);
2309 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2311 tcp_moderate_cwnd(tp
);
2312 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2314 /* There is something screwy going on with the retrans hints after
2316 tcp_clear_all_retrans_hints(tp
);
2319 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2321 return tp
->undo_marker
&&
2322 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2325 /* People celebrate: "We love our President!" */
2326 static int tcp_try_undo_recovery(struct sock
*sk
)
2328 struct tcp_sock
*tp
= tcp_sk(sk
);
2330 if (tcp_may_undo(tp
)) {
2331 /* Happy end! We did not retransmit anything
2332 * or our original transmission succeeded.
2334 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2335 tcp_undo_cwr(sk
, 1);
2336 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2337 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2339 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
2340 tp
->undo_marker
= 0;
2342 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2343 /* Hold old state until something *above* high_seq
2344 * is ACKed. For Reno it is MUST to prevent false
2345 * fast retransmits (RFC2582). SACK TCP is safe. */
2346 tcp_moderate_cwnd(tp
);
2349 tcp_set_ca_state(sk
, TCP_CA_Open
);
2353 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2354 static void tcp_try_undo_dsack(struct sock
*sk
)
2356 struct tcp_sock
*tp
= tcp_sk(sk
);
2358 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2359 DBGUNDO(sk
, "D-SACK");
2360 tcp_undo_cwr(sk
, 1);
2361 tp
->undo_marker
= 0;
2362 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2366 /* Undo during fast recovery after partial ACK. */
2368 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2370 struct tcp_sock
*tp
= tcp_sk(sk
);
2371 /* Partial ACK arrived. Force Hoe's retransmit. */
2372 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2374 if (tcp_may_undo(tp
)) {
2375 /* Plain luck! Hole if filled with delayed
2376 * packet, rather than with a retransmit.
2378 if (tp
->retrans_out
== 0)
2379 tp
->retrans_stamp
= 0;
2381 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2384 tcp_undo_cwr(sk
, 0);
2385 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2387 /* So... Do not make Hoe's retransmit yet.
2388 * If the first packet was delayed, the rest
2389 * ones are most probably delayed as well.
2396 /* Undo during loss recovery after partial ACK. */
2397 static int tcp_try_undo_loss(struct sock
*sk
)
2399 struct tcp_sock
*tp
= tcp_sk(sk
);
2401 if (tcp_may_undo(tp
)) {
2402 struct sk_buff
*skb
;
2403 tcp_for_write_queue(skb
, sk
) {
2404 if (skb
== tcp_send_head(sk
))
2406 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2409 tcp_clear_all_retrans_hints(tp
);
2411 DBGUNDO(sk
, "partial loss");
2413 tcp_undo_cwr(sk
, 1);
2414 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2415 inet_csk(sk
)->icsk_retransmits
= 0;
2416 tp
->undo_marker
= 0;
2417 if (tcp_is_sack(tp
))
2418 tcp_set_ca_state(sk
, TCP_CA_Open
);
2424 static inline void tcp_complete_cwr(struct sock
*sk
)
2426 struct tcp_sock
*tp
= tcp_sk(sk
);
2427 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2428 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2429 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2432 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2434 struct tcp_sock
*tp
= tcp_sk(sk
);
2436 tcp_verify_left_out(tp
);
2438 if (tp
->retrans_out
== 0)
2439 tp
->retrans_stamp
= 0;
2442 tcp_enter_cwr(sk
, 1);
2444 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2445 int state
= TCP_CA_Open
;
2447 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2448 state
= TCP_CA_Disorder
;
2450 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2451 tcp_set_ca_state(sk
, state
);
2452 tp
->high_seq
= tp
->snd_nxt
;
2454 tcp_moderate_cwnd(tp
);
2456 tcp_cwnd_down(sk
, flag
);
2460 static void tcp_mtup_probe_failed(struct sock
*sk
)
2462 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2464 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2465 icsk
->icsk_mtup
.probe_size
= 0;
2468 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2470 struct tcp_sock
*tp
= tcp_sk(sk
);
2471 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2473 /* FIXME: breaks with very large cwnd */
2474 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2475 tp
->snd_cwnd
= tp
->snd_cwnd
*
2476 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2477 icsk
->icsk_mtup
.probe_size
;
2478 tp
->snd_cwnd_cnt
= 0;
2479 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2480 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2482 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2483 icsk
->icsk_mtup
.probe_size
= 0;
2484 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2488 /* Process an event, which can update packets-in-flight not trivially.
2489 * Main goal of this function is to calculate new estimate for left_out,
2490 * taking into account both packets sitting in receiver's buffer and
2491 * packets lost by network.
2493 * Besides that it does CWND reduction, when packet loss is detected
2494 * and changes state of machine.
2496 * It does _not_ decide what to send, it is made in function
2497 * tcp_xmit_retransmit_queue().
2500 tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2502 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2503 struct tcp_sock
*tp
= tcp_sk(sk
);
2504 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2505 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2506 (tcp_fackets_out(tp
) > tp
->reordering
));
2507 int fast_rexmit
= 0;
2509 /* Some technical things:
2510 * 1. Reno does not count dupacks (sacked_out) automatically. */
2511 if (!tp
->packets_out
)
2514 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2515 tp
->fackets_out
= 0;
2517 /* Now state machine starts.
2518 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2520 tp
->prior_ssthresh
= 0;
2522 /* B. In all the states check for reneging SACKs. */
2523 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2526 /* C. Process data loss notification, provided it is valid. */
2527 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2528 before(tp
->snd_una
, tp
->high_seq
) &&
2529 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2530 tp
->fackets_out
> tp
->reordering
) {
2531 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, 0);
2532 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2535 /* D. Check consistency of the current state. */
2536 tcp_verify_left_out(tp
);
2538 /* E. Check state exit conditions. State can be terminated
2539 * when high_seq is ACKed. */
2540 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2541 BUG_TRAP(tp
->retrans_out
== 0);
2542 tp
->retrans_stamp
= 0;
2543 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2544 switch (icsk
->icsk_ca_state
) {
2546 icsk
->icsk_retransmits
= 0;
2547 if (tcp_try_undo_recovery(sk
))
2552 /* CWR is to be held something *above* high_seq
2553 * is ACKed for CWR bit to reach receiver. */
2554 if (tp
->snd_una
!= tp
->high_seq
) {
2555 tcp_complete_cwr(sk
);
2556 tcp_set_ca_state(sk
, TCP_CA_Open
);
2560 case TCP_CA_Disorder
:
2561 tcp_try_undo_dsack(sk
);
2562 if (!tp
->undo_marker
||
2563 /* For SACK case do not Open to allow to undo
2564 * catching for all duplicate ACKs. */
2565 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2566 tp
->undo_marker
= 0;
2567 tcp_set_ca_state(sk
, TCP_CA_Open
);
2571 case TCP_CA_Recovery
:
2572 if (tcp_is_reno(tp
))
2573 tcp_reset_reno_sack(tp
);
2574 if (tcp_try_undo_recovery(sk
))
2576 tcp_complete_cwr(sk
);
2581 /* F. Process state. */
2582 switch (icsk
->icsk_ca_state
) {
2583 case TCP_CA_Recovery
:
2584 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2585 if (tcp_is_reno(tp
) && is_dupack
)
2586 tcp_add_reno_sack(sk
);
2588 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2591 if (flag
&FLAG_DATA_ACKED
)
2592 icsk
->icsk_retransmits
= 0;
2593 if (!tcp_try_undo_loss(sk
)) {
2594 tcp_moderate_cwnd(tp
);
2595 tcp_xmit_retransmit_queue(sk
);
2598 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2600 /* Loss is undone; fall through to processing in Open state. */
2602 if (tcp_is_reno(tp
)) {
2603 if (flag
& FLAG_SND_UNA_ADVANCED
)
2604 tcp_reset_reno_sack(tp
);
2606 tcp_add_reno_sack(sk
);
2609 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2610 tcp_try_undo_dsack(sk
);
2612 if (!tcp_time_to_recover(sk
)) {
2613 tcp_try_to_open(sk
, flag
);
2617 /* MTU probe failure: don't reduce cwnd */
2618 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2619 icsk
->icsk_mtup
.probe_size
&&
2620 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2621 tcp_mtup_probe_failed(sk
);
2622 /* Restores the reduction we did in tcp_mtup_probe() */
2624 tcp_simple_retransmit(sk
);
2628 /* Otherwise enter Recovery state */
2630 if (tcp_is_reno(tp
))
2631 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2633 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2635 tp
->high_seq
= tp
->snd_nxt
;
2636 tp
->prior_ssthresh
= 0;
2637 tp
->undo_marker
= tp
->snd_una
;
2638 tp
->undo_retrans
= tp
->retrans_out
;
2640 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2641 if (!(flag
&FLAG_ECE
))
2642 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2643 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2644 TCP_ECN_queue_cwr(tp
);
2647 tp
->bytes_acked
= 0;
2648 tp
->snd_cwnd_cnt
= 0;
2649 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2653 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2654 tcp_update_scoreboard(sk
, fast_rexmit
);
2655 tcp_cwnd_down(sk
, flag
);
2656 tcp_xmit_retransmit_queue(sk
);
2659 /* Read draft-ietf-tcplw-high-performance before mucking
2660 * with this code. (Supersedes RFC1323)
2662 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2664 /* RTTM Rule: A TSecr value received in a segment is used to
2665 * update the averaged RTT measurement only if the segment
2666 * acknowledges some new data, i.e., only if it advances the
2667 * left edge of the send window.
2669 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2670 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2672 * Changed: reset backoff as soon as we see the first valid sample.
2673 * If we do not, we get strongly overestimated rto. With timestamps
2674 * samples are accepted even from very old segments: f.e., when rtt=1
2675 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2676 * answer arrives rto becomes 120 seconds! If at least one of segments
2677 * in window is lost... Voila. --ANK (010210)
2679 struct tcp_sock
*tp
= tcp_sk(sk
);
2680 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2681 tcp_rtt_estimator(sk
, seq_rtt
);
2683 inet_csk(sk
)->icsk_backoff
= 0;
2687 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2689 /* We don't have a timestamp. Can only use
2690 * packets that are not retransmitted to determine
2691 * rtt estimates. Also, we must not reset the
2692 * backoff for rto until we get a non-retransmitted
2693 * packet. This allows us to deal with a situation
2694 * where the network delay has increased suddenly.
2695 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2698 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2701 tcp_rtt_estimator(sk
, seq_rtt
);
2703 inet_csk(sk
)->icsk_backoff
= 0;
2707 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2710 const struct tcp_sock
*tp
= tcp_sk(sk
);
2711 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2712 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2713 tcp_ack_saw_tstamp(sk
, flag
);
2714 else if (seq_rtt
>= 0)
2715 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2718 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
,
2719 u32 in_flight
, int good
)
2721 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2722 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
, good
);
2723 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2726 /* Restart timer after forward progress on connection.
2727 * RFC2988 recommends to restart timer to now+rto.
2729 static void tcp_rearm_rto(struct sock
*sk
)
2731 struct tcp_sock
*tp
= tcp_sk(sk
);
2733 if (!tp
->packets_out
) {
2734 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2736 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2740 /* If we get here, the whole TSO packet has not been acked. */
2741 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2743 struct tcp_sock
*tp
= tcp_sk(sk
);
2746 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2748 packets_acked
= tcp_skb_pcount(skb
);
2749 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2751 packets_acked
-= tcp_skb_pcount(skb
);
2753 if (packets_acked
) {
2754 BUG_ON(tcp_skb_pcount(skb
) == 0);
2755 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2758 return packets_acked
;
2761 /* Remove acknowledged frames from the retransmission queue. If our packet
2762 * is before the ack sequence we can discard it as it's confirmed to have
2763 * arrived at the other end.
2765 static int tcp_clean_rtx_queue(struct sock
*sk
, s32
*seq_rtt_p
,
2768 struct tcp_sock
*tp
= tcp_sk(sk
);
2769 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2770 struct sk_buff
*skb
;
2771 u32 now
= tcp_time_stamp
;
2772 int fully_acked
= 1;
2774 int prior_packets
= tp
->packets_out
;
2776 u32 reord
= tp
->packets_out
;
2778 s32 ca_seq_rtt
= -1;
2779 ktime_t last_ackt
= net_invalid_timestamp();
2781 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2782 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2785 u8 sacked
= scb
->sacked
;
2787 /* Determine how many packets and what bytes were acked, tso and else */
2788 if (after(scb
->end_seq
, tp
->snd_una
)) {
2789 if (tcp_skb_pcount(skb
) == 1 ||
2790 !after(tp
->snd_una
, scb
->seq
))
2793 packets_acked
= tcp_tso_acked(sk
, skb
);
2798 end_seq
= tp
->snd_una
;
2800 packets_acked
= tcp_skb_pcount(skb
);
2801 end_seq
= scb
->end_seq
;
2804 /* MTU probing checks */
2805 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2806 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2807 tcp_mtup_probe_success(sk
, skb
);
2811 if (sacked
& TCPCB_RETRANS
) {
2812 if (sacked
& TCPCB_SACKED_RETRANS
)
2813 tp
->retrans_out
-= packets_acked
;
2814 flag
|= FLAG_RETRANS_DATA_ACKED
;
2817 if ((flag
& FLAG_DATA_ACKED
) ||
2818 (packets_acked
> 1))
2819 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2821 ca_seq_rtt
= now
- scb
->when
;
2822 last_ackt
= skb
->tstamp
;
2824 seq_rtt
= ca_seq_rtt
;
2826 if (!(sacked
& TCPCB_SACKED_ACKED
))
2827 reord
= min(cnt
, reord
);
2830 if (sacked
& TCPCB_SACKED_ACKED
)
2831 tp
->sacked_out
-= packets_acked
;
2832 if (sacked
& TCPCB_LOST
)
2833 tp
->lost_out
-= packets_acked
;
2835 if ((sacked
& TCPCB_URG
) && tp
->urg_mode
&&
2836 !before(end_seq
, tp
->snd_up
))
2839 ca_seq_rtt
= now
- scb
->when
;
2840 last_ackt
= skb
->tstamp
;
2842 seq_rtt
= ca_seq_rtt
;
2844 reord
= min(cnt
, reord
);
2846 tp
->packets_out
-= packets_acked
;
2847 cnt
+= packets_acked
;
2849 /* Initial outgoing SYN's get put onto the write_queue
2850 * just like anything else we transmit. It is not
2851 * true data, and if we misinform our callers that
2852 * this ACK acks real data, we will erroneously exit
2853 * connection startup slow start one packet too
2854 * quickly. This is severely frowned upon behavior.
2856 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2857 flag
|= FLAG_DATA_ACKED
;
2859 flag
|= FLAG_SYN_ACKED
;
2860 tp
->retrans_stamp
= 0;
2866 tcp_unlink_write_queue(skb
, sk
);
2867 sk_stream_free_skb(sk
, skb
);
2868 tcp_clear_all_retrans_hints(tp
);
2871 if (flag
& FLAG_ACKED
) {
2872 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2873 const struct tcp_congestion_ops
*ca_ops
2874 = inet_csk(sk
)->icsk_ca_ops
;
2876 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2879 if (tcp_is_reno(tp
)) {
2880 tcp_remove_reno_sacks(sk
, pkts_acked
);
2882 /* Non-retransmitted hole got filled? That's reordering */
2883 if (reord
< prior_fackets
)
2884 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2887 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2889 if (ca_ops
->pkts_acked
) {
2892 /* Is the ACK triggering packet unambiguous? */
2893 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2894 /* High resolution needed and available? */
2895 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2896 !ktime_equal(last_ackt
,
2897 net_invalid_timestamp()))
2898 rtt_us
= ktime_us_delta(ktime_get_real(),
2900 else if (ca_seq_rtt
> 0)
2901 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
2904 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2908 #if FASTRETRANS_DEBUG > 0
2909 BUG_TRAP((int)tp
->sacked_out
>= 0);
2910 BUG_TRAP((int)tp
->lost_out
>= 0);
2911 BUG_TRAP((int)tp
->retrans_out
>= 0);
2912 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2913 icsk
= inet_csk(sk
);
2915 printk(KERN_DEBUG
"Leak l=%u %d\n",
2916 tp
->lost_out
, icsk
->icsk_ca_state
);
2919 if (tp
->sacked_out
) {
2920 printk(KERN_DEBUG
"Leak s=%u %d\n",
2921 tp
->sacked_out
, icsk
->icsk_ca_state
);
2924 if (tp
->retrans_out
) {
2925 printk(KERN_DEBUG
"Leak r=%u %d\n",
2926 tp
->retrans_out
, icsk
->icsk_ca_state
);
2927 tp
->retrans_out
= 0;
2931 *seq_rtt_p
= seq_rtt
;
2935 static void tcp_ack_probe(struct sock
*sk
)
2937 const struct tcp_sock
*tp
= tcp_sk(sk
);
2938 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2940 /* Was it a usable window open? */
2942 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2943 tp
->snd_una
+ tp
->snd_wnd
)) {
2944 icsk
->icsk_backoff
= 0;
2945 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2946 /* Socket must be waked up by subsequent tcp_data_snd_check().
2947 * This function is not for random using!
2950 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2951 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2956 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2958 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2959 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2962 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2964 const struct tcp_sock
*tp
= tcp_sk(sk
);
2965 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2966 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2969 /* Check that window update is acceptable.
2970 * The function assumes that snd_una<=ack<=snd_next.
2972 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2973 const u32 ack_seq
, const u32 nwin
)
2975 return (after(ack
, tp
->snd_una
) ||
2976 after(ack_seq
, tp
->snd_wl1
) ||
2977 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2980 /* Update our send window.
2982 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2983 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2985 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2988 struct tcp_sock
*tp
= tcp_sk(sk
);
2990 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2992 if (likely(!tcp_hdr(skb
)->syn
))
2993 nwin
<<= tp
->rx_opt
.snd_wscale
;
2995 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2996 flag
|= FLAG_WIN_UPDATE
;
2997 tcp_update_wl(tp
, ack
, ack_seq
);
2999 if (tp
->snd_wnd
!= nwin
) {
3002 /* Note, it is the only place, where
3003 * fast path is recovered for sending TCP.
3006 tcp_fast_path_check(sk
);
3008 if (nwin
> tp
->max_window
) {
3009 tp
->max_window
= nwin
;
3010 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3020 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3021 * continue in congestion avoidance.
3023 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3025 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3026 tp
->snd_cwnd_cnt
= 0;
3027 tp
->bytes_acked
= 0;
3028 TCP_ECN_queue_cwr(tp
);
3029 tcp_moderate_cwnd(tp
);
3032 /* A conservative spurious RTO response algorithm: reduce cwnd using
3033 * rate halving and continue in congestion avoidance.
3035 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3037 tcp_enter_cwr(sk
, 0);
3040 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3043 tcp_ratehalving_spur_to_response(sk
);
3045 tcp_undo_cwr(sk
, 1);
3048 /* F-RTO spurious RTO detection algorithm (RFC4138)
3050 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3051 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3052 * window (but not to or beyond highest sequence sent before RTO):
3053 * On First ACK, send two new segments out.
3054 * On Second ACK, RTO was likely spurious. Do spurious response (response
3055 * algorithm is not part of the F-RTO detection algorithm
3056 * given in RFC4138 but can be selected separately).
3057 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3058 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3059 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3060 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3062 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3063 * original window even after we transmit two new data segments.
3066 * on first step, wait until first cumulative ACK arrives, then move to
3067 * the second step. In second step, the next ACK decides.
3069 * F-RTO is implemented (mainly) in four functions:
3070 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3071 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3072 * called when tcp_use_frto() showed green light
3073 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3074 * - tcp_enter_frto_loss() is called if there is not enough evidence
3075 * to prove that the RTO is indeed spurious. It transfers the control
3076 * from F-RTO to the conventional RTO recovery
3078 static int tcp_process_frto(struct sock
*sk
, int flag
)
3080 struct tcp_sock
*tp
= tcp_sk(sk
);
3082 tcp_verify_left_out(tp
);
3084 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3085 if (flag
&FLAG_DATA_ACKED
)
3086 inet_csk(sk
)->icsk_retransmits
= 0;
3088 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3089 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3090 tp
->undo_marker
= 0;
3092 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3093 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3097 if (!IsSackFrto() || tcp_is_reno(tp
)) {
3098 /* RFC4138 shortcoming in step 2; should also have case c):
3099 * ACK isn't duplicate nor advances window, e.g., opposite dir
3102 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3105 if (!(flag
&FLAG_DATA_ACKED
)) {
3106 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3111 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3112 /* Prevent sending of new data. */
3113 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3114 tcp_packets_in_flight(tp
));
3118 if ((tp
->frto_counter
>= 2) &&
3119 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
3120 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
3121 /* RFC4138 shortcoming (see comment above) */
3122 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3125 tcp_enter_frto_loss(sk
, 3, flag
);
3130 if (tp
->frto_counter
== 1) {
3131 /* tcp_may_send_now needs to see updated state */
3132 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3133 tp
->frto_counter
= 2;
3135 if (!tcp_may_send_now(sk
))
3136 tcp_enter_frto_loss(sk
, 2, flag
);
3140 switch (sysctl_tcp_frto_response
) {
3142 tcp_undo_spur_to_response(sk
, flag
);
3145 tcp_conservative_spur_to_response(tp
);
3148 tcp_ratehalving_spur_to_response(sk
);
3151 tp
->frto_counter
= 0;
3152 tp
->undo_marker
= 0;
3153 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS
);
3158 /* This routine deals with incoming acks, but not outgoing ones. */
3159 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3161 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3162 struct tcp_sock
*tp
= tcp_sk(sk
);
3163 u32 prior_snd_una
= tp
->snd_una
;
3164 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3165 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3166 u32 prior_in_flight
;
3172 /* If the ack is newer than sent or older than previous acks
3173 * then we can probably ignore it.
3175 if (after(ack
, tp
->snd_nxt
))
3176 goto uninteresting_ack
;
3178 if (before(ack
, prior_snd_una
))
3181 if (after(ack
, prior_snd_una
))
3182 flag
|= FLAG_SND_UNA_ADVANCED
;
3184 if (sysctl_tcp_abc
) {
3185 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3186 tp
->bytes_acked
+= ack
- prior_snd_una
;
3187 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3188 /* we assume just one segment left network */
3189 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
3192 prior_fackets
= tp
->fackets_out
;
3193 prior_in_flight
= tcp_packets_in_flight(tp
);
3195 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3196 /* Window is constant, pure forward advance.
3197 * No more checks are required.
3198 * Note, we use the fact that SND.UNA>=SND.WL2.
3200 tcp_update_wl(tp
, ack
, ack_seq
);
3202 flag
|= FLAG_WIN_UPDATE
;
3204 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3206 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
3208 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3211 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
3213 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3215 if (TCP_SKB_CB(skb
)->sacked
)
3216 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3218 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3221 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3224 /* We passed data and got it acked, remove any soft error
3225 * log. Something worked...
3227 sk
->sk_err_soft
= 0;
3228 tp
->rcv_tstamp
= tcp_time_stamp
;
3229 prior_packets
= tp
->packets_out
;
3233 /* See if we can take anything off of the retransmit queue. */
3234 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
, prior_fackets
);
3236 if (tp
->frto_counter
)
3237 frto_cwnd
= tcp_process_frto(sk
, flag
);
3238 /* Guarantee sacktag reordering detection against wrap-arounds */
3239 if (before(tp
->frto_highmark
, tp
->snd_una
))
3240 tp
->frto_highmark
= 0;
3242 if (tcp_ack_is_dubious(sk
, flag
)) {
3243 /* Advance CWND, if state allows this. */
3244 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3245 tcp_may_raise_cwnd(sk
, flag
))
3246 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 0);
3247 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
, flag
);
3249 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3250 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 1);
3253 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
3254 dst_confirm(sk
->sk_dst_cache
);
3259 icsk
->icsk_probes_out
= 0;
3261 /* If this ack opens up a zero window, clear backoff. It was
3262 * being used to time the probes, and is probably far higher than
3263 * it needs to be for normal retransmission.
3265 if (tcp_send_head(sk
))
3270 if (TCP_SKB_CB(skb
)->sacked
)
3271 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3274 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3279 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3280 * But, this can also be called on packets in the established flow when
3281 * the fast version below fails.
3283 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
3286 struct tcphdr
*th
= tcp_hdr(skb
);
3287 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
3289 ptr
= (unsigned char *)(th
+ 1);
3290 opt_rx
->saw_tstamp
= 0;
3292 while (length
> 0) {
3299 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3304 if (opsize
< 2) /* "silly options" */
3306 if (opsize
> length
)
3307 return; /* don't parse partial options */
3310 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3311 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
3313 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
3314 in_mss
= opt_rx
->user_mss
;
3315 opt_rx
->mss_clamp
= in_mss
;
3320 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
3321 if (sysctl_tcp_window_scaling
) {
3322 __u8 snd_wscale
= *(__u8
*) ptr
;
3323 opt_rx
->wscale_ok
= 1;
3324 if (snd_wscale
> 14) {
3325 if (net_ratelimit())
3326 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3327 "scaling value %d >14 received.\n",
3331 opt_rx
->snd_wscale
= snd_wscale
;
3334 case TCPOPT_TIMESTAMP
:
3335 if (opsize
==TCPOLEN_TIMESTAMP
) {
3336 if ((estab
&& opt_rx
->tstamp_ok
) ||
3337 (!estab
&& sysctl_tcp_timestamps
)) {
3338 opt_rx
->saw_tstamp
= 1;
3339 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
3340 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
3344 case TCPOPT_SACK_PERM
:
3345 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
3346 if (sysctl_tcp_sack
) {
3347 opt_rx
->sack_ok
= 1;
3348 tcp_sack_reset(opt_rx
);
3354 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3355 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3357 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3360 #ifdef CONFIG_TCP_MD5SIG
3363 * The MD5 Hash has already been
3364 * checked (see tcp_v{4,6}_do_rcv()).
3376 /* Fast parse options. This hopes to only see timestamps.
3377 * If it is wrong it falls back on tcp_parse_options().
3379 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3380 struct tcp_sock
*tp
)
3382 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3383 tp
->rx_opt
.saw_tstamp
= 0;
3385 } else if (tp
->rx_opt
.tstamp_ok
&&
3386 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3387 __be32
*ptr
= (__be32
*)(th
+ 1);
3388 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3389 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3390 tp
->rx_opt
.saw_tstamp
= 1;
3392 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3394 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3398 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3402 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3404 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3405 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3408 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3410 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3411 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3412 * extra check below makes sure this can only happen
3413 * for pure ACK frames. -DaveM
3415 * Not only, also it occurs for expired timestamps.
3418 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3419 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3420 tcp_store_ts_recent(tp
);
3424 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3426 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3427 * it can pass through stack. So, the following predicate verifies that
3428 * this segment is not used for anything but congestion avoidance or
3429 * fast retransmit. Moreover, we even are able to eliminate most of such
3430 * second order effects, if we apply some small "replay" window (~RTO)
3431 * to timestamp space.
3433 * All these measures still do not guarantee that we reject wrapped ACKs
3434 * on networks with high bandwidth, when sequence space is recycled fastly,
3435 * but it guarantees that such events will be very rare and do not affect
3436 * connection seriously. This doesn't look nice, but alas, PAWS is really
3439 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3440 * states that events when retransmit arrives after original data are rare.
3441 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3442 * the biggest problem on large power networks even with minor reordering.
3443 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3444 * up to bandwidth of 18Gigabit/sec. 8) ]
3447 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3449 struct tcp_sock
*tp
= tcp_sk(sk
);
3450 struct tcphdr
*th
= tcp_hdr(skb
);
3451 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3452 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3454 return (/* 1. Pure ACK with correct sequence number. */
3455 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3457 /* 2. ... and duplicate ACK. */
3458 ack
== tp
->snd_una
&&
3460 /* 3. ... and does not update window. */
3461 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3463 /* 4. ... and sits in replay window. */
3464 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3467 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3469 const struct tcp_sock
*tp
= tcp_sk(sk
);
3470 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3471 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3472 !tcp_disordered_ack(sk
, skb
));
3475 /* Check segment sequence number for validity.
3477 * Segment controls are considered valid, if the segment
3478 * fits to the window after truncation to the window. Acceptability
3479 * of data (and SYN, FIN, of course) is checked separately.
3480 * See tcp_data_queue(), for example.
3482 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3483 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3484 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3485 * (borrowed from freebsd)
3488 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3490 return !before(end_seq
, tp
->rcv_wup
) &&
3491 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3494 /* When we get a reset we do this. */
3495 static void tcp_reset(struct sock
*sk
)
3497 /* We want the right error as BSD sees it (and indeed as we do). */
3498 switch (sk
->sk_state
) {
3500 sk
->sk_err
= ECONNREFUSED
;
3502 case TCP_CLOSE_WAIT
:
3508 sk
->sk_err
= ECONNRESET
;
3511 if (!sock_flag(sk
, SOCK_DEAD
))
3512 sk
->sk_error_report(sk
);
3518 * Process the FIN bit. This now behaves as it is supposed to work
3519 * and the FIN takes effect when it is validly part of sequence
3520 * space. Not before when we get holes.
3522 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3523 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3526 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3527 * close and we go into CLOSING (and later onto TIME-WAIT)
3529 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3531 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3533 struct tcp_sock
*tp
= tcp_sk(sk
);
3535 inet_csk_schedule_ack(sk
);
3537 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3538 sock_set_flag(sk
, SOCK_DONE
);
3540 switch (sk
->sk_state
) {
3542 case TCP_ESTABLISHED
:
3543 /* Move to CLOSE_WAIT */
3544 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3545 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3548 case TCP_CLOSE_WAIT
:
3550 /* Received a retransmission of the FIN, do
3555 /* RFC793: Remain in the LAST-ACK state. */
3559 /* This case occurs when a simultaneous close
3560 * happens, we must ack the received FIN and
3561 * enter the CLOSING state.
3564 tcp_set_state(sk
, TCP_CLOSING
);
3567 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3569 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3572 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3573 * cases we should never reach this piece of code.
3575 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3576 __FUNCTION__
, sk
->sk_state
);
3580 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3581 * Probably, we should reset in this case. For now drop them.
3583 __skb_queue_purge(&tp
->out_of_order_queue
);
3584 if (tcp_is_sack(tp
))
3585 tcp_sack_reset(&tp
->rx_opt
);
3586 sk_stream_mem_reclaim(sk
);
3588 if (!sock_flag(sk
, SOCK_DEAD
)) {
3589 sk
->sk_state_change(sk
);
3591 /* Do not send POLL_HUP for half duplex close. */
3592 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3593 sk
->sk_state
== TCP_CLOSE
)
3594 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3596 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3600 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3602 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3603 if (before(seq
, sp
->start_seq
))
3604 sp
->start_seq
= seq
;
3605 if (after(end_seq
, sp
->end_seq
))
3606 sp
->end_seq
= end_seq
;
3612 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3614 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3615 if (before(seq
, tp
->rcv_nxt
))
3616 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3618 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3620 tp
->rx_opt
.dsack
= 1;
3621 tp
->duplicate_sack
[0].start_seq
= seq
;
3622 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3623 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3627 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3629 if (!tp
->rx_opt
.dsack
)
3630 tcp_dsack_set(tp
, seq
, end_seq
);
3632 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3635 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3637 struct tcp_sock
*tp
= tcp_sk(sk
);
3639 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3640 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3641 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3642 tcp_enter_quickack_mode(sk
);
3644 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3645 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3647 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3648 end_seq
= tp
->rcv_nxt
;
3649 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3656 /* These routines update the SACK block as out-of-order packets arrive or
3657 * in-order packets close up the sequence space.
3659 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3662 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3663 struct tcp_sack_block
*swalk
= sp
+1;
3665 /* See if the recent change to the first SACK eats into
3666 * or hits the sequence space of other SACK blocks, if so coalesce.
3668 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3669 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3672 /* Zap SWALK, by moving every further SACK up by one slot.
3673 * Decrease num_sacks.
3675 tp
->rx_opt
.num_sacks
--;
3676 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3677 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3681 this_sack
++, swalk
++;
3685 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3689 tmp
= sack1
->start_seq
;
3690 sack1
->start_seq
= sack2
->start_seq
;
3691 sack2
->start_seq
= tmp
;
3693 tmp
= sack1
->end_seq
;
3694 sack1
->end_seq
= sack2
->end_seq
;
3695 sack2
->end_seq
= tmp
;
3698 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3700 struct tcp_sock
*tp
= tcp_sk(sk
);
3701 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3702 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3708 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3709 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3710 /* Rotate this_sack to the first one. */
3711 for (; this_sack
>0; this_sack
--, sp
--)
3712 tcp_sack_swap(sp
, sp
-1);
3714 tcp_sack_maybe_coalesce(tp
);
3719 /* Could not find an adjacent existing SACK, build a new one,
3720 * put it at the front, and shift everyone else down. We
3721 * always know there is at least one SACK present already here.
3723 * If the sack array is full, forget about the last one.
3725 if (this_sack
>= 4) {
3727 tp
->rx_opt
.num_sacks
--;
3730 for (; this_sack
> 0; this_sack
--, sp
--)
3734 /* Build the new head SACK, and we're done. */
3735 sp
->start_seq
= seq
;
3736 sp
->end_seq
= end_seq
;
3737 tp
->rx_opt
.num_sacks
++;
3738 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3741 /* RCV.NXT advances, some SACKs should be eaten. */
3743 static void tcp_sack_remove(struct tcp_sock
*tp
)
3745 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3746 int num_sacks
= tp
->rx_opt
.num_sacks
;
3749 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3750 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3751 tp
->rx_opt
.num_sacks
= 0;
3752 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3756 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3757 /* Check if the start of the sack is covered by RCV.NXT. */
3758 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3761 /* RCV.NXT must cover all the block! */
3762 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3764 /* Zap this SACK, by moving forward any other SACKS. */
3765 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3766 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3773 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3774 tp
->rx_opt
.num_sacks
= num_sacks
;
3775 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3779 /* This one checks to see if we can put data from the
3780 * out_of_order queue into the receive_queue.
3782 static void tcp_ofo_queue(struct sock
*sk
)
3784 struct tcp_sock
*tp
= tcp_sk(sk
);
3785 __u32 dsack_high
= tp
->rcv_nxt
;
3786 struct sk_buff
*skb
;
3788 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3789 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3792 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3793 __u32 dsack
= dsack_high
;
3794 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3795 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3796 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3799 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3800 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3801 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3805 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3806 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3807 TCP_SKB_CB(skb
)->end_seq
);
3809 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3810 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3811 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3812 if (tcp_hdr(skb
)->fin
)
3813 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3817 static int tcp_prune_queue(struct sock
*sk
);
3819 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3821 struct tcphdr
*th
= tcp_hdr(skb
);
3822 struct tcp_sock
*tp
= tcp_sk(sk
);
3825 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3828 __skb_pull(skb
, th
->doff
*4);
3830 TCP_ECN_accept_cwr(tp
, skb
);
3832 if (tp
->rx_opt
.dsack
) {
3833 tp
->rx_opt
.dsack
= 0;
3834 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3835 4 - tp
->rx_opt
.tstamp_ok
);
3838 /* Queue data for delivery to the user.
3839 * Packets in sequence go to the receive queue.
3840 * Out of sequence packets to the out_of_order_queue.
3842 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3843 if (tcp_receive_window(tp
) == 0)
3846 /* Ok. In sequence. In window. */
3847 if (tp
->ucopy
.task
== current
&&
3848 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3849 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3850 int chunk
= min_t(unsigned int, skb
->len
,
3853 __set_current_state(TASK_RUNNING
);
3856 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3857 tp
->ucopy
.len
-= chunk
;
3858 tp
->copied_seq
+= chunk
;
3859 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3860 tcp_rcv_space_adjust(sk
);
3868 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3869 !sk_stream_rmem_schedule(sk
, skb
))) {
3870 if (tcp_prune_queue(sk
) < 0 ||
3871 !sk_stream_rmem_schedule(sk
, skb
))
3874 sk_stream_set_owner_r(skb
, sk
);
3875 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3877 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3879 tcp_event_data_recv(sk
, skb
);
3881 tcp_fin(skb
, sk
, th
);
3883 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3886 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3887 * gap in queue is filled.
3889 if (skb_queue_empty(&tp
->out_of_order_queue
))
3890 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3893 if (tp
->rx_opt
.num_sacks
)
3894 tcp_sack_remove(tp
);
3896 tcp_fast_path_check(sk
);
3900 else if (!sock_flag(sk
, SOCK_DEAD
))
3901 sk
->sk_data_ready(sk
, 0);
3905 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3906 /* A retransmit, 2nd most common case. Force an immediate ack. */
3907 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3908 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3911 tcp_enter_quickack_mode(sk
);
3912 inet_csk_schedule_ack(sk
);
3918 /* Out of window. F.e. zero window probe. */
3919 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3922 tcp_enter_quickack_mode(sk
);
3924 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3925 /* Partial packet, seq < rcv_next < end_seq */
3926 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3927 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3928 TCP_SKB_CB(skb
)->end_seq
);
3930 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3932 /* If window is closed, drop tail of packet. But after
3933 * remembering D-SACK for its head made in previous line.
3935 if (!tcp_receive_window(tp
))
3940 TCP_ECN_check_ce(tp
, skb
);
3942 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3943 !sk_stream_rmem_schedule(sk
, skb
)) {
3944 if (tcp_prune_queue(sk
) < 0 ||
3945 !sk_stream_rmem_schedule(sk
, skb
))
3949 /* Disable header prediction. */
3951 inet_csk_schedule_ack(sk
);
3953 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3954 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3956 sk_stream_set_owner_r(skb
, sk
);
3958 if (!skb_peek(&tp
->out_of_order_queue
)) {
3959 /* Initial out of order segment, build 1 SACK. */
3960 if (tcp_is_sack(tp
)) {
3961 tp
->rx_opt
.num_sacks
= 1;
3962 tp
->rx_opt
.dsack
= 0;
3963 tp
->rx_opt
.eff_sacks
= 1;
3964 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3965 tp
->selective_acks
[0].end_seq
=
3966 TCP_SKB_CB(skb
)->end_seq
;
3968 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3970 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3971 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3972 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3974 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3975 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3977 if (!tp
->rx_opt
.num_sacks
||
3978 tp
->selective_acks
[0].end_seq
!= seq
)
3981 /* Common case: data arrive in order after hole. */
3982 tp
->selective_acks
[0].end_seq
= end_seq
;
3986 /* Find place to insert this segment. */
3988 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3990 } while ((skb1
= skb1
->prev
) !=
3991 (struct sk_buff
*)&tp
->out_of_order_queue
);
3993 /* Do skb overlap to previous one? */
3994 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3995 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3996 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3997 /* All the bits are present. Drop. */
3999 tcp_dsack_set(tp
, seq
, end_seq
);
4002 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4003 /* Partial overlap. */
4004 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4009 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
4011 /* And clean segments covered by new one as whole. */
4012 while ((skb1
= skb
->next
) !=
4013 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4014 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4015 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4016 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
4019 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4020 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
4025 if (tcp_is_sack(tp
))
4026 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4030 /* Collapse contiguous sequence of skbs head..tail with
4031 * sequence numbers start..end.
4032 * Segments with FIN/SYN are not collapsed (only because this
4036 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4037 struct sk_buff
*head
, struct sk_buff
*tail
,
4040 struct sk_buff
*skb
;
4042 /* First, check that queue is collapsible and find
4043 * the point where collapsing can be useful. */
4044 for (skb
= head
; skb
!= tail
; ) {
4045 /* No new bits? It is possible on ofo queue. */
4046 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4047 struct sk_buff
*next
= skb
->next
;
4048 __skb_unlink(skb
, list
);
4050 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
4055 /* The first skb to collapse is:
4057 * - bloated or contains data before "start" or
4058 * overlaps to the next one.
4060 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4061 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4062 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4063 (skb
->next
!= tail
&&
4064 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4067 /* Decided to skip this, advance start seq. */
4068 start
= TCP_SKB_CB(skb
)->end_seq
;
4071 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4074 while (before(start
, end
)) {
4075 struct sk_buff
*nskb
;
4076 unsigned int header
= skb_headroom(skb
);
4077 int copy
= SKB_MAX_ORDER(header
, 0);
4079 /* Too big header? This can happen with IPv6. */
4082 if (end
-start
< copy
)
4084 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
4088 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4089 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4091 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4093 skb_reserve(nskb
, header
);
4094 memcpy(nskb
->head
, skb
->head
, header
);
4095 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4096 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4097 __skb_insert(nskb
, skb
->prev
, skb
, list
);
4098 sk_stream_set_owner_r(nskb
, sk
);
4100 /* Copy data, releasing collapsed skbs. */
4102 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4103 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4107 size
= min(copy
, size
);
4108 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4110 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4114 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4115 struct sk_buff
*next
= skb
->next
;
4116 __skb_unlink(skb
, list
);
4118 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
4121 tcp_hdr(skb
)->syn
||
4129 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4130 * and tcp_collapse() them until all the queue is collapsed.
4132 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4134 struct tcp_sock
*tp
= tcp_sk(sk
);
4135 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4136 struct sk_buff
*head
;
4142 start
= TCP_SKB_CB(skb
)->seq
;
4143 end
= TCP_SKB_CB(skb
)->end_seq
;
4149 /* Segment is terminated when we see gap or when
4150 * we are at the end of all the queue. */
4151 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4152 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4153 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4154 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4155 head
, skb
, start
, end
);
4157 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4159 /* Start new segment */
4160 start
= TCP_SKB_CB(skb
)->seq
;
4161 end
= TCP_SKB_CB(skb
)->end_seq
;
4163 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4164 start
= TCP_SKB_CB(skb
)->seq
;
4165 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4166 end
= TCP_SKB_CB(skb
)->end_seq
;
4171 /* Reduce allocated memory if we can, trying to get
4172 * the socket within its memory limits again.
4174 * Return less than zero if we should start dropping frames
4175 * until the socket owning process reads some of the data
4176 * to stabilize the situation.
4178 static int tcp_prune_queue(struct sock
*sk
)
4180 struct tcp_sock
*tp
= tcp_sk(sk
);
4182 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4184 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
4186 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4187 tcp_clamp_window(sk
);
4188 else if (tcp_memory_pressure
)
4189 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4191 tcp_collapse_ofo_queue(sk
);
4192 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4193 sk
->sk_receive_queue
.next
,
4194 (struct sk_buff
*)&sk
->sk_receive_queue
,
4195 tp
->copied_seq
, tp
->rcv_nxt
);
4196 sk_stream_mem_reclaim(sk
);
4198 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4201 /* Collapsing did not help, destructive actions follow.
4202 * This must not ever occur. */
4204 /* First, purge the out_of_order queue. */
4205 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4206 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
4207 __skb_queue_purge(&tp
->out_of_order_queue
);
4209 /* Reset SACK state. A conforming SACK implementation will
4210 * do the same at a timeout based retransmit. When a connection
4211 * is in a sad state like this, we care only about integrity
4212 * of the connection not performance.
4214 if (tcp_is_sack(tp
))
4215 tcp_sack_reset(&tp
->rx_opt
);
4216 sk_stream_mem_reclaim(sk
);
4219 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4222 /* If we are really being abused, tell the caller to silently
4223 * drop receive data on the floor. It will get retransmitted
4224 * and hopefully then we'll have sufficient space.
4226 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
4228 /* Massive buffer overcommit. */
4234 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4235 * As additional protections, we do not touch cwnd in retransmission phases,
4236 * and if application hit its sndbuf limit recently.
4238 void tcp_cwnd_application_limited(struct sock
*sk
)
4240 struct tcp_sock
*tp
= tcp_sk(sk
);
4242 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4243 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4244 /* Limited by application or receiver window. */
4245 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4246 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4247 if (win_used
< tp
->snd_cwnd
) {
4248 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4249 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4251 tp
->snd_cwnd_used
= 0;
4253 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4256 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4258 struct tcp_sock
*tp
= tcp_sk(sk
);
4260 /* If the user specified a specific send buffer setting, do
4263 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4266 /* If we are under global TCP memory pressure, do not expand. */
4267 if (tcp_memory_pressure
)
4270 /* If we are under soft global TCP memory pressure, do not expand. */
4271 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4274 /* If we filled the congestion window, do not expand. */
4275 if (tp
->packets_out
>= tp
->snd_cwnd
)
4281 /* When incoming ACK allowed to free some skb from write_queue,
4282 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4283 * on the exit from tcp input handler.
4285 * PROBLEM: sndbuf expansion does not work well with largesend.
4287 static void tcp_new_space(struct sock
*sk
)
4289 struct tcp_sock
*tp
= tcp_sk(sk
);
4291 if (tcp_should_expand_sndbuf(sk
)) {
4292 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4293 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4294 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4295 tp
->reordering
+ 1);
4296 sndmem
*= 2*demanded
;
4297 if (sndmem
> sk
->sk_sndbuf
)
4298 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4299 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4302 sk
->sk_write_space(sk
);
4305 static void tcp_check_space(struct sock
*sk
)
4307 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4308 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4309 if (sk
->sk_socket
&&
4310 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4315 static inline void tcp_data_snd_check(struct sock
*sk
)
4317 tcp_push_pending_frames(sk
);
4318 tcp_check_space(sk
);
4322 * Check if sending an ack is needed.
4324 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4326 struct tcp_sock
*tp
= tcp_sk(sk
);
4328 /* More than one full frame received... */
4329 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4330 /* ... and right edge of window advances far enough.
4331 * (tcp_recvmsg() will send ACK otherwise). Or...
4333 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4334 /* We ACK each frame or... */
4335 tcp_in_quickack_mode(sk
) ||
4336 /* We have out of order data. */
4338 skb_peek(&tp
->out_of_order_queue
))) {
4339 /* Then ack it now */
4342 /* Else, send delayed ack. */
4343 tcp_send_delayed_ack(sk
);
4347 static inline void tcp_ack_snd_check(struct sock
*sk
)
4349 if (!inet_csk_ack_scheduled(sk
)) {
4350 /* We sent a data segment already. */
4353 __tcp_ack_snd_check(sk
, 1);
4357 * This routine is only called when we have urgent data
4358 * signaled. Its the 'slow' part of tcp_urg. It could be
4359 * moved inline now as tcp_urg is only called from one
4360 * place. We handle URGent data wrong. We have to - as
4361 * BSD still doesn't use the correction from RFC961.
4362 * For 1003.1g we should support a new option TCP_STDURG to permit
4363 * either form (or just set the sysctl tcp_stdurg).
4366 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4368 struct tcp_sock
*tp
= tcp_sk(sk
);
4369 u32 ptr
= ntohs(th
->urg_ptr
);
4371 if (ptr
&& !sysctl_tcp_stdurg
)
4373 ptr
+= ntohl(th
->seq
);
4375 /* Ignore urgent data that we've already seen and read. */
4376 if (after(tp
->copied_seq
, ptr
))
4379 /* Do not replay urg ptr.
4381 * NOTE: interesting situation not covered by specs.
4382 * Misbehaving sender may send urg ptr, pointing to segment,
4383 * which we already have in ofo queue. We are not able to fetch
4384 * such data and will stay in TCP_URG_NOTYET until will be eaten
4385 * by recvmsg(). Seems, we are not obliged to handle such wicked
4386 * situations. But it is worth to think about possibility of some
4387 * DoSes using some hypothetical application level deadlock.
4389 if (before(ptr
, tp
->rcv_nxt
))
4392 /* Do we already have a newer (or duplicate) urgent pointer? */
4393 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4396 /* Tell the world about our new urgent pointer. */
4399 /* We may be adding urgent data when the last byte read was
4400 * urgent. To do this requires some care. We cannot just ignore
4401 * tp->copied_seq since we would read the last urgent byte again
4402 * as data, nor can we alter copied_seq until this data arrives
4403 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4405 * NOTE. Double Dutch. Rendering to plain English: author of comment
4406 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4407 * and expect that both A and B disappear from stream. This is _wrong_.
4408 * Though this happens in BSD with high probability, this is occasional.
4409 * Any application relying on this is buggy. Note also, that fix "works"
4410 * only in this artificial test. Insert some normal data between A and B and we will
4411 * decline of BSD again. Verdict: it is better to remove to trap
4414 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4415 !sock_flag(sk
, SOCK_URGINLINE
) &&
4416 tp
->copied_seq
!= tp
->rcv_nxt
) {
4417 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4419 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4420 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4425 tp
->urg_data
= TCP_URG_NOTYET
;
4428 /* Disable header prediction. */
4432 /* This is the 'fast' part of urgent handling. */
4433 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4435 struct tcp_sock
*tp
= tcp_sk(sk
);
4437 /* Check if we get a new urgent pointer - normally not. */
4439 tcp_check_urg(sk
,th
);
4441 /* Do we wait for any urgent data? - normally not... */
4442 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4443 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4446 /* Is the urgent pointer pointing into this packet? */
4447 if (ptr
< skb
->len
) {
4449 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4451 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4452 if (!sock_flag(sk
, SOCK_DEAD
))
4453 sk
->sk_data_ready(sk
, 0);
4458 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4460 struct tcp_sock
*tp
= tcp_sk(sk
);
4461 int chunk
= skb
->len
- hlen
;
4465 if (skb_csum_unnecessary(skb
))
4466 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4468 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4472 tp
->ucopy
.len
-= chunk
;
4473 tp
->copied_seq
+= chunk
;
4474 tcp_rcv_space_adjust(sk
);
4481 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4485 if (sock_owned_by_user(sk
)) {
4487 result
= __tcp_checksum_complete(skb
);
4490 result
= __tcp_checksum_complete(skb
);
4495 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4497 return !skb_csum_unnecessary(skb
) &&
4498 __tcp_checksum_complete_user(sk
, skb
);
4501 #ifdef CONFIG_NET_DMA
4502 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4504 struct tcp_sock
*tp
= tcp_sk(sk
);
4505 int chunk
= skb
->len
- hlen
;
4507 int copied_early
= 0;
4509 if (tp
->ucopy
.wakeup
)
4512 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4513 tp
->ucopy
.dma_chan
= get_softnet_dma();
4515 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4517 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4518 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4523 tp
->ucopy
.dma_cookie
= dma_cookie
;
4526 tp
->ucopy
.len
-= chunk
;
4527 tp
->copied_seq
+= chunk
;
4528 tcp_rcv_space_adjust(sk
);
4530 if ((tp
->ucopy
.len
== 0) ||
4531 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4532 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4533 tp
->ucopy
.wakeup
= 1;
4534 sk
->sk_data_ready(sk
, 0);
4536 } else if (chunk
> 0) {
4537 tp
->ucopy
.wakeup
= 1;
4538 sk
->sk_data_ready(sk
, 0);
4541 return copied_early
;
4543 #endif /* CONFIG_NET_DMA */
4546 * TCP receive function for the ESTABLISHED state.
4548 * It is split into a fast path and a slow path. The fast path is
4550 * - A zero window was announced from us - zero window probing
4551 * is only handled properly in the slow path.
4552 * - Out of order segments arrived.
4553 * - Urgent data is expected.
4554 * - There is no buffer space left
4555 * - Unexpected TCP flags/window values/header lengths are received
4556 * (detected by checking the TCP header against pred_flags)
4557 * - Data is sent in both directions. Fast path only supports pure senders
4558 * or pure receivers (this means either the sequence number or the ack
4559 * value must stay constant)
4560 * - Unexpected TCP option.
4562 * When these conditions are not satisfied it drops into a standard
4563 * receive procedure patterned after RFC793 to handle all cases.
4564 * The first three cases are guaranteed by proper pred_flags setting,
4565 * the rest is checked inline. Fast processing is turned on in
4566 * tcp_data_queue when everything is OK.
4568 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4569 struct tcphdr
*th
, unsigned len
)
4571 struct tcp_sock
*tp
= tcp_sk(sk
);
4574 * Header prediction.
4575 * The code loosely follows the one in the famous
4576 * "30 instruction TCP receive" Van Jacobson mail.
4578 * Van's trick is to deposit buffers into socket queue
4579 * on a device interrupt, to call tcp_recv function
4580 * on the receive process context and checksum and copy
4581 * the buffer to user space. smart...
4583 * Our current scheme is not silly either but we take the
4584 * extra cost of the net_bh soft interrupt processing...
4585 * We do checksum and copy also but from device to kernel.
4588 tp
->rx_opt
.saw_tstamp
= 0;
4590 /* pred_flags is 0xS?10 << 16 + snd_wnd
4591 * if header_prediction is to be made
4592 * 'S' will always be tp->tcp_header_len >> 2
4593 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4594 * turn it off (when there are holes in the receive
4595 * space for instance)
4596 * PSH flag is ignored.
4599 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4600 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4601 int tcp_header_len
= tp
->tcp_header_len
;
4603 /* Timestamp header prediction: tcp_header_len
4604 * is automatically equal to th->doff*4 due to pred_flags
4608 /* Check timestamp */
4609 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4610 __be32
*ptr
= (__be32
*)(th
+ 1);
4612 /* No? Slow path! */
4613 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4614 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4617 tp
->rx_opt
.saw_tstamp
= 1;
4619 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4621 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4623 /* If PAWS failed, check it more carefully in slow path */
4624 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4627 /* DO NOT update ts_recent here, if checksum fails
4628 * and timestamp was corrupted part, it will result
4629 * in a hung connection since we will drop all
4630 * future packets due to the PAWS test.
4634 if (len
<= tcp_header_len
) {
4635 /* Bulk data transfer: sender */
4636 if (len
== tcp_header_len
) {
4637 /* Predicted packet is in window by definition.
4638 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4639 * Hence, check seq<=rcv_wup reduces to:
4641 if (tcp_header_len
==
4642 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4643 tp
->rcv_nxt
== tp
->rcv_wup
)
4644 tcp_store_ts_recent(tp
);
4646 /* We know that such packets are checksummed
4649 tcp_ack(sk
, skb
, 0);
4651 tcp_data_snd_check(sk
);
4653 } else { /* Header too small */
4654 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4659 int copied_early
= 0;
4661 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4662 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4663 #ifdef CONFIG_NET_DMA
4664 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4669 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4670 __set_current_state(TASK_RUNNING
);
4672 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4676 /* Predicted packet is in window by definition.
4677 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4678 * Hence, check seq<=rcv_wup reduces to:
4680 if (tcp_header_len
==
4681 (sizeof(struct tcphdr
) +
4682 TCPOLEN_TSTAMP_ALIGNED
) &&
4683 tp
->rcv_nxt
== tp
->rcv_wup
)
4684 tcp_store_ts_recent(tp
);
4686 tcp_rcv_rtt_measure_ts(sk
, skb
);
4688 __skb_pull(skb
, tcp_header_len
);
4689 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4690 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4693 tcp_cleanup_rbuf(sk
, skb
->len
);
4696 if (tcp_checksum_complete_user(sk
, skb
))
4699 /* Predicted packet is in window by definition.
4700 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4701 * Hence, check seq<=rcv_wup reduces to:
4703 if (tcp_header_len
==
4704 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4705 tp
->rcv_nxt
== tp
->rcv_wup
)
4706 tcp_store_ts_recent(tp
);
4708 tcp_rcv_rtt_measure_ts(sk
, skb
);
4710 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4713 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4715 /* Bulk data transfer: receiver */
4716 __skb_pull(skb
,tcp_header_len
);
4717 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4718 sk_stream_set_owner_r(skb
, sk
);
4719 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4722 tcp_event_data_recv(sk
, skb
);
4724 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4725 /* Well, only one small jumplet in fast path... */
4726 tcp_ack(sk
, skb
, FLAG_DATA
);
4727 tcp_data_snd_check(sk
);
4728 if (!inet_csk_ack_scheduled(sk
))
4732 __tcp_ack_snd_check(sk
, 0);
4734 #ifdef CONFIG_NET_DMA
4736 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4742 sk
->sk_data_ready(sk
, 0);
4748 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4752 * RFC1323: H1. Apply PAWS check first.
4754 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4755 tcp_paws_discard(sk
, skb
)) {
4757 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4758 tcp_send_dupack(sk
, skb
);
4761 /* Resets are accepted even if PAWS failed.
4763 ts_recent update must be made after we are sure
4764 that the packet is in window.
4769 * Standard slow path.
4772 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4773 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4774 * (RST) segments are validated by checking their SEQ-fields."
4775 * And page 69: "If an incoming segment is not acceptable,
4776 * an acknowledgment should be sent in reply (unless the RST bit
4777 * is set, if so drop the segment and return)".
4780 tcp_send_dupack(sk
, skb
);
4789 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4791 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4792 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4793 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4800 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4802 tcp_rcv_rtt_measure_ts(sk
, skb
);
4804 /* Process urgent data. */
4805 tcp_urg(sk
, skb
, th
);
4807 /* step 7: process the segment text */
4808 tcp_data_queue(sk
, skb
);
4810 tcp_data_snd_check(sk
);
4811 tcp_ack_snd_check(sk
);
4815 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4822 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4823 struct tcphdr
*th
, unsigned len
)
4825 struct tcp_sock
*tp
= tcp_sk(sk
);
4826 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4827 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4829 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4833 * "If the state is SYN-SENT then
4834 * first check the ACK bit
4835 * If the ACK bit is set
4836 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4837 * a reset (unless the RST bit is set, if so drop
4838 * the segment and return)"
4840 * We do not send data with SYN, so that RFC-correct
4843 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4844 goto reset_and_undo
;
4846 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4847 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4849 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4850 goto reset_and_undo
;
4853 /* Now ACK is acceptable.
4855 * "If the RST bit is set
4856 * If the ACK was acceptable then signal the user "error:
4857 * connection reset", drop the segment, enter CLOSED state,
4858 * delete TCB, and return."
4867 * "fifth, if neither of the SYN or RST bits is set then
4868 * drop the segment and return."
4874 goto discard_and_undo
;
4877 * "If the SYN bit is on ...
4878 * are acceptable then ...
4879 * (our SYN has been ACKed), change the connection
4880 * state to ESTABLISHED..."
4883 TCP_ECN_rcv_synack(tp
, th
);
4885 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4886 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4888 /* Ok.. it's good. Set up sequence numbers and
4889 * move to established.
4891 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4892 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4894 /* RFC1323: The window in SYN & SYN/ACK segments is
4897 tp
->snd_wnd
= ntohs(th
->window
);
4898 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4900 if (!tp
->rx_opt
.wscale_ok
) {
4901 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4902 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4905 if (tp
->rx_opt
.saw_tstamp
) {
4906 tp
->rx_opt
.tstamp_ok
= 1;
4907 tp
->tcp_header_len
=
4908 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4909 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4910 tcp_store_ts_recent(tp
);
4912 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4915 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
4916 tcp_enable_fack(tp
);
4919 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4920 tcp_initialize_rcv_mss(sk
);
4922 /* Remember, tcp_poll() does not lock socket!
4923 * Change state from SYN-SENT only after copied_seq
4924 * is initialized. */
4925 tp
->copied_seq
= tp
->rcv_nxt
;
4927 tcp_set_state(sk
, TCP_ESTABLISHED
);
4929 security_inet_conn_established(sk
, skb
);
4931 /* Make sure socket is routed, for correct metrics. */
4932 icsk
->icsk_af_ops
->rebuild_header(sk
);
4934 tcp_init_metrics(sk
);
4936 tcp_init_congestion_control(sk
);
4938 /* Prevent spurious tcp_cwnd_restart() on first data
4941 tp
->lsndtime
= tcp_time_stamp
;
4943 tcp_init_buffer_space(sk
);
4945 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4946 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4948 if (!tp
->rx_opt
.snd_wscale
)
4949 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4953 if (!sock_flag(sk
, SOCK_DEAD
)) {
4954 sk
->sk_state_change(sk
);
4955 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
4958 if (sk
->sk_write_pending
||
4959 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4960 icsk
->icsk_ack
.pingpong
) {
4961 /* Save one ACK. Data will be ready after
4962 * several ticks, if write_pending is set.
4964 * It may be deleted, but with this feature tcpdumps
4965 * look so _wonderfully_ clever, that I was not able
4966 * to stand against the temptation 8) --ANK
4968 inet_csk_schedule_ack(sk
);
4969 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4970 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4971 tcp_incr_quickack(sk
);
4972 tcp_enter_quickack_mode(sk
);
4973 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4974 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4985 /* No ACK in the segment */
4989 * "If the RST bit is set
4991 * Otherwise (no ACK) drop the segment and return."
4994 goto discard_and_undo
;
4998 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4999 goto discard_and_undo
;
5002 /* We see SYN without ACK. It is attempt of
5003 * simultaneous connect with crossed SYNs.
5004 * Particularly, it can be connect to self.
5006 tcp_set_state(sk
, TCP_SYN_RECV
);
5008 if (tp
->rx_opt
.saw_tstamp
) {
5009 tp
->rx_opt
.tstamp_ok
= 1;
5010 tcp_store_ts_recent(tp
);
5011 tp
->tcp_header_len
=
5012 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5014 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5017 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5018 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5020 /* RFC1323: The window in SYN & SYN/ACK segments is
5023 tp
->snd_wnd
= ntohs(th
->window
);
5024 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5025 tp
->max_window
= tp
->snd_wnd
;
5027 TCP_ECN_rcv_syn(tp
, th
);
5030 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5031 tcp_initialize_rcv_mss(sk
);
5034 tcp_send_synack(sk
);
5036 /* Note, we could accept data and URG from this segment.
5037 * There are no obstacles to make this.
5039 * However, if we ignore data in ACKless segments sometimes,
5040 * we have no reasons to accept it sometimes.
5041 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5042 * is not flawless. So, discard packet for sanity.
5043 * Uncomment this return to process the data.
5050 /* "fifth, if neither of the SYN or RST bits is set then
5051 * drop the segment and return."
5055 tcp_clear_options(&tp
->rx_opt
);
5056 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5060 tcp_clear_options(&tp
->rx_opt
);
5061 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5067 * This function implements the receiving procedure of RFC 793 for
5068 * all states except ESTABLISHED and TIME_WAIT.
5069 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5070 * address independent.
5073 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5074 struct tcphdr
*th
, unsigned len
)
5076 struct tcp_sock
*tp
= tcp_sk(sk
);
5077 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5080 tp
->rx_opt
.saw_tstamp
= 0;
5082 switch (sk
->sk_state
) {
5094 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5097 /* Now we have several options: In theory there is
5098 * nothing else in the frame. KA9Q has an option to
5099 * send data with the syn, BSD accepts data with the
5100 * syn up to the [to be] advertised window and
5101 * Solaris 2.1 gives you a protocol error. For now
5102 * we just ignore it, that fits the spec precisely
5103 * and avoids incompatibilities. It would be nice in
5104 * future to drop through and process the data.
5106 * Now that TTCP is starting to be used we ought to
5108 * But, this leaves one open to an easy denial of
5109 * service attack, and SYN cookies can't defend
5110 * against this problem. So, we drop the data
5111 * in the interest of security over speed unless
5112 * it's still in use.
5120 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5124 /* Do step6 onward by hand. */
5125 tcp_urg(sk
, skb
, th
);
5127 tcp_data_snd_check(sk
);
5131 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5132 tcp_paws_discard(sk
, skb
)) {
5134 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
5135 tcp_send_dupack(sk
, skb
);
5138 /* Reset is accepted even if it did not pass PAWS. */
5141 /* step 1: check sequence number */
5142 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5144 tcp_send_dupack(sk
, skb
);
5148 /* step 2: check RST bit */
5154 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5156 /* step 3: check security and precedence [ignored] */
5160 * Check for a SYN in window.
5162 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5163 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
5168 /* step 5: check the ACK field */
5170 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5172 switch (sk
->sk_state
) {
5175 tp
->copied_seq
= tp
->rcv_nxt
;
5177 tcp_set_state(sk
, TCP_ESTABLISHED
);
5178 sk
->sk_state_change(sk
);
5180 /* Note, that this wakeup is only for marginal
5181 * crossed SYN case. Passively open sockets
5182 * are not waked up, because sk->sk_sleep ==
5183 * NULL and sk->sk_socket == NULL.
5187 SOCK_WAKE_IO
, POLL_OUT
);
5189 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5190 tp
->snd_wnd
= ntohs(th
->window
) <<
5191 tp
->rx_opt
.snd_wscale
;
5192 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5193 TCP_SKB_CB(skb
)->seq
);
5195 /* tcp_ack considers this ACK as duplicate
5196 * and does not calculate rtt.
5197 * Fix it at least with timestamps.
5199 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5201 tcp_ack_saw_tstamp(sk
, 0);
5203 if (tp
->rx_opt
.tstamp_ok
)
5204 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5206 /* Make sure socket is routed, for
5209 icsk
->icsk_af_ops
->rebuild_header(sk
);
5211 tcp_init_metrics(sk
);
5213 tcp_init_congestion_control(sk
);
5215 /* Prevent spurious tcp_cwnd_restart() on
5216 * first data packet.
5218 tp
->lsndtime
= tcp_time_stamp
;
5221 tcp_initialize_rcv_mss(sk
);
5222 tcp_init_buffer_space(sk
);
5223 tcp_fast_path_on(tp
);
5230 if (tp
->snd_una
== tp
->write_seq
) {
5231 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5232 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5233 dst_confirm(sk
->sk_dst_cache
);
5235 if (!sock_flag(sk
, SOCK_DEAD
))
5236 /* Wake up lingering close() */
5237 sk
->sk_state_change(sk
);
5241 if (tp
->linger2
< 0 ||
5242 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5243 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5245 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5249 tmo
= tcp_fin_time(sk
);
5250 if (tmo
> TCP_TIMEWAIT_LEN
) {
5251 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5252 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5253 /* Bad case. We could lose such FIN otherwise.
5254 * It is not a big problem, but it looks confusing
5255 * and not so rare event. We still can lose it now,
5256 * if it spins in bh_lock_sock(), but it is really
5259 inet_csk_reset_keepalive_timer(sk
, tmo
);
5261 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5269 if (tp
->snd_una
== tp
->write_seq
) {
5270 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5276 if (tp
->snd_una
== tp
->write_seq
) {
5277 tcp_update_metrics(sk
);
5286 /* step 6: check the URG bit */
5287 tcp_urg(sk
, skb
, th
);
5289 /* step 7: process the segment text */
5290 switch (sk
->sk_state
) {
5291 case TCP_CLOSE_WAIT
:
5294 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5298 /* RFC 793 says to queue data in these states,
5299 * RFC 1122 says we MUST send a reset.
5300 * BSD 4.4 also does reset.
5302 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5303 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5304 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5305 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5311 case TCP_ESTABLISHED
:
5312 tcp_data_queue(sk
, skb
);
5317 /* tcp_data could move socket to TIME-WAIT */
5318 if (sk
->sk_state
!= TCP_CLOSE
) {
5319 tcp_data_snd_check(sk
);
5320 tcp_ack_snd_check(sk
);
5330 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5331 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5332 EXPORT_SYMBOL(tcp_parse_options
);
5333 EXPORT_SYMBOL(tcp_rcv_established
);
5334 EXPORT_SYMBOL(tcp_rcv_state_process
);
5335 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);