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 (sacked
& TCPCB_SACKED_ACKED
)
1250 *reord
= min(fack_count
, *reord
);
1253 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1254 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1257 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1258 if (sacked
& TCPCB_SACKED_RETRANS
) {
1259 /* If the segment is not tagged as lost,
1260 * we do not clear RETRANS, believing
1261 * that retransmission is still in flight.
1263 if (sacked
& TCPCB_LOST
) {
1264 TCP_SKB_CB(skb
)->sacked
&=
1265 ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1266 tp
->lost_out
-= tcp_skb_pcount(skb
);
1267 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1269 /* clear lost hint */
1270 tp
->retransmit_skb_hint
= NULL
;
1273 if (!(sacked
& TCPCB_RETRANS
)) {
1274 /* New sack for not retransmitted frame,
1275 * which was in hole. It is reordering.
1277 if (before(TCP_SKB_CB(skb
)->seq
,
1278 tcp_highest_sack_seq(tp
)))
1279 *reord
= min(fack_count
, *reord
);
1281 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1282 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1283 flag
|= FLAG_ONLY_ORIG_SACKED
;
1286 if (sacked
& TCPCB_LOST
) {
1287 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1288 tp
->lost_out
-= tcp_skb_pcount(skb
);
1290 /* clear lost hint */
1291 tp
->retransmit_skb_hint
= NULL
;
1295 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1296 flag
|= FLAG_DATA_SACKED
;
1297 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1299 fack_count
+= tcp_skb_pcount(skb
);
1301 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1302 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1303 before(TCP_SKB_CB(skb
)->seq
,
1304 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1305 tp
->lost_cnt_hint
+= tcp_skb_pcount(skb
);
1307 if (fack_count
> tp
->fackets_out
)
1308 tp
->fackets_out
= fack_count
;
1310 if (after(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
1311 tp
->highest_sack
= skb
;
1314 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1315 * frames and clear it. undo_retrans is decreased above, L|R frames
1316 * are accounted above as well.
1318 if (dup_sack
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)) {
1319 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1320 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1321 tp
->retransmit_skb_hint
= NULL
;
1327 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1328 struct tcp_sack_block
*next_dup
,
1329 u32 start_seq
, u32 end_seq
,
1330 int dup_sack_in
, int *fack_count
,
1331 int *reord
, int *flag
)
1333 struct tcp_sock
*tp
= tcp_sk(sk
);
1335 tcp_for_write_queue_from(skb
, sk
) {
1337 int dup_sack
= dup_sack_in
;
1339 if (skb
== tcp_send_head(sk
))
1342 /* queue is in-order => we can short-circuit the walk early */
1343 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1346 if ((next_dup
!= NULL
) &&
1347 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1348 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1349 next_dup
->start_seq
,
1356 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
, end_seq
);
1357 if (unlikely(in_sack
< 0))
1361 *flag
|= tcp_sacktag_one(skb
, tp
, reord
, dup_sack
, *fack_count
);
1363 *fack_count
+= tcp_skb_pcount(skb
);
1368 /* Avoid all extra work that is being done by sacktag while walking in
1371 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1374 tcp_for_write_queue_from(skb
, sk
) {
1375 if (skb
== tcp_send_head(sk
))
1378 if (!before(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1384 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1386 struct tcp_sack_block
*next_dup
,
1388 int *fack_count
, int *reord
,
1391 if (next_dup
== NULL
)
1394 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1395 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
);
1396 tcp_sacktag_walk(skb
, sk
, NULL
,
1397 next_dup
->start_seq
, next_dup
->end_seq
,
1398 1, fack_count
, reord
, flag
);
1404 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1406 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1410 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
1412 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1413 struct tcp_sock
*tp
= tcp_sk(sk
);
1414 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1415 TCP_SKB_CB(ack_skb
)->sacked
);
1416 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1417 struct tcp_sack_block sp
[4];
1418 struct tcp_sack_block
*cache
;
1419 struct sk_buff
*skb
;
1420 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
1422 int reord
= tp
->packets_out
;
1424 int found_dup_sack
= 0;
1427 int first_sack_index
;
1429 if (!tp
->sacked_out
) {
1430 if (WARN_ON(tp
->fackets_out
))
1431 tp
->fackets_out
= 0;
1432 tp
->highest_sack
= tcp_write_queue_head(sk
);
1435 found_dup_sack
= tcp_check_dsack(tp
, ack_skb
, sp_wire
,
1436 num_sacks
, prior_snd_una
);
1438 flag
|= FLAG_DSACKING_ACK
;
1440 /* Eliminate too old ACKs, but take into
1441 * account more or less fresh ones, they can
1442 * contain valid SACK info.
1444 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1447 if (!tp
->packets_out
)
1451 first_sack_index
= 0;
1452 for (i
= 0; i
< num_sacks
; i
++) {
1453 int dup_sack
= !i
&& found_dup_sack
;
1455 sp
[used_sacks
].start_seq
= ntohl(get_unaligned(&sp_wire
[i
].start_seq
));
1456 sp
[used_sacks
].end_seq
= ntohl(get_unaligned(&sp_wire
[i
].end_seq
));
1458 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1459 sp
[used_sacks
].start_seq
,
1460 sp
[used_sacks
].end_seq
)) {
1462 if (!tp
->undo_marker
)
1463 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO
);
1465 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD
);
1467 /* Don't count olds caused by ACK reordering */
1468 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1469 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1471 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD
);
1474 first_sack_index
= -1;
1478 /* Ignore very old stuff early */
1479 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1485 /* order SACK blocks to allow in order walk of the retrans queue */
1486 for (i
= used_sacks
- 1; i
> 0; i
--) {
1487 for (j
= 0; j
< i
; j
++){
1488 if (after(sp
[j
].start_seq
, sp
[j
+1].start_seq
)) {
1489 struct tcp_sack_block tmp
;
1495 /* Track where the first SACK block goes to */
1496 if (j
== first_sack_index
)
1497 first_sack_index
= j
+1;
1502 skb
= tcp_write_queue_head(sk
);
1506 if (!tp
->sacked_out
) {
1507 /* It's already past, so skip checking against it */
1508 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1510 cache
= tp
->recv_sack_cache
;
1511 /* Skip empty blocks in at head of the cache */
1512 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1517 while (i
< used_sacks
) {
1518 u32 start_seq
= sp
[i
].start_seq
;
1519 u32 end_seq
= sp
[i
].end_seq
;
1520 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1521 struct tcp_sack_block
*next_dup
= NULL
;
1523 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1524 next_dup
= &sp
[i
+ 1];
1526 /* Event "B" in the comment above. */
1527 if (after(end_seq
, tp
->high_seq
))
1528 flag
|= FLAG_DATA_LOST
;
1530 /* Skip too early cached blocks */
1531 while (tcp_sack_cache_ok(tp
, cache
) &&
1532 !before(start_seq
, cache
->end_seq
))
1535 /* Can skip some work by looking recv_sack_cache? */
1536 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1537 after(end_seq
, cache
->start_seq
)) {
1540 if (before(start_seq
, cache
->start_seq
)) {
1541 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1542 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
,
1543 cache
->start_seq
, dup_sack
,
1544 &fack_count
, &reord
, &flag
);
1547 /* Rest of the block already fully processed? */
1548 if (!after(end_seq
, cache
->end_seq
))
1551 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
, cache
->end_seq
,
1552 &fack_count
, &reord
, &flag
);
1554 /* ...tail remains todo... */
1555 if (TCP_SKB_CB(tp
->highest_sack
)->end_seq
== cache
->end_seq
) {
1556 /* ...but better entrypoint exists! */
1557 skb
= tcp_write_queue_next(sk
, tp
->highest_sack
);
1558 fack_count
= tp
->fackets_out
;
1563 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
);
1564 /* Check overlap against next cached too (past this one already) */
1569 if (tp
->sacked_out
&& !before(start_seq
, tcp_highest_sack_seq(tp
))) {
1570 skb
= tcp_write_queue_next(sk
, tp
->highest_sack
);
1571 fack_count
= tp
->fackets_out
;
1573 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1576 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1577 dup_sack
, &fack_count
, &reord
, &flag
);
1580 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1581 * due to in-order walk
1583 if (after(end_seq
, tp
->frto_highmark
))
1584 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1589 /* Clear the head of the cache sack blocks so we can skip it next time */
1590 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1591 tp
->recv_sack_cache
[i
].start_seq
= 0;
1592 tp
->recv_sack_cache
[i
].end_seq
= 0;
1594 for (j
= 0; j
< used_sacks
; j
++)
1595 tp
->recv_sack_cache
[i
++] = sp
[j
];
1597 tcp_mark_lost_retrans(sk
);
1599 tcp_verify_left_out(tp
);
1601 if ((reord
< tp
->fackets_out
) &&
1602 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1603 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1604 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1608 #if FASTRETRANS_DEBUG > 0
1609 BUG_TRAP((int)tp
->sacked_out
>= 0);
1610 BUG_TRAP((int)tp
->lost_out
>= 0);
1611 BUG_TRAP((int)tp
->retrans_out
>= 0);
1612 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1617 /* If we receive more dupacks than we expected counting segments
1618 * in assumption of absent reordering, interpret this as reordering.
1619 * The only another reason could be bug in receiver TCP.
1621 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1623 struct tcp_sock
*tp
= tcp_sk(sk
);
1626 holes
= max(tp
->lost_out
, 1U);
1627 holes
= min(holes
, tp
->packets_out
);
1629 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1630 tp
->sacked_out
= tp
->packets_out
- holes
;
1631 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1635 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1637 static void tcp_add_reno_sack(struct sock
*sk
)
1639 struct tcp_sock
*tp
= tcp_sk(sk
);
1641 tcp_check_reno_reordering(sk
, 0);
1642 tcp_verify_left_out(tp
);
1645 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1647 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1649 struct tcp_sock
*tp
= tcp_sk(sk
);
1652 /* One ACK acked hole. The rest eat duplicate ACKs. */
1653 if (acked
-1 >= tp
->sacked_out
)
1656 tp
->sacked_out
-= acked
-1;
1658 tcp_check_reno_reordering(sk
, acked
);
1659 tcp_verify_left_out(tp
);
1662 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1667 /* F-RTO can only be used if TCP has never retransmitted anything other than
1668 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1670 int tcp_use_frto(struct sock
*sk
)
1672 const struct tcp_sock
*tp
= tcp_sk(sk
);
1673 struct sk_buff
*skb
;
1675 if (!sysctl_tcp_frto
)
1681 /* Avoid expensive walking of rexmit queue if possible */
1682 if (tp
->retrans_out
> 1)
1685 skb
= tcp_write_queue_head(sk
);
1686 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1687 tcp_for_write_queue_from(skb
, sk
) {
1688 if (skb
== tcp_send_head(sk
))
1690 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1692 /* Short-circuit when first non-SACKed skb has been checked */
1693 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1699 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1700 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1701 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1702 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1703 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1704 * bits are handled if the Loss state is really to be entered (in
1705 * tcp_enter_frto_loss).
1707 * Do like tcp_enter_loss() would; when RTO expires the second time it
1709 * "Reduce ssthresh if it has not yet been made inside this window."
1711 void tcp_enter_frto(struct sock
*sk
)
1713 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1714 struct tcp_sock
*tp
= tcp_sk(sk
);
1715 struct sk_buff
*skb
;
1717 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1718 tp
->snd_una
== tp
->high_seq
||
1719 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1720 !icsk
->icsk_retransmits
)) {
1721 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1722 /* Our state is too optimistic in ssthresh() call because cwnd
1723 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1724 * recovery has not yet completed. Pattern would be this: RTO,
1725 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1727 * RFC4138 should be more specific on what to do, even though
1728 * RTO is quite unlikely to occur after the first Cumulative ACK
1729 * due to back-off and complexity of triggering events ...
1731 if (tp
->frto_counter
) {
1733 stored_cwnd
= tp
->snd_cwnd
;
1735 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1736 tp
->snd_cwnd
= stored_cwnd
;
1738 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1740 /* ... in theory, cong.control module could do "any tricks" in
1741 * ssthresh(), which means that ca_state, lost bits and lost_out
1742 * counter would have to be faked before the call occurs. We
1743 * consider that too expensive, unlikely and hacky, so modules
1744 * using these in ssthresh() must deal these incompatibility
1745 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1747 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1750 tp
->undo_marker
= tp
->snd_una
;
1751 tp
->undo_retrans
= 0;
1753 skb
= tcp_write_queue_head(sk
);
1754 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1755 tp
->undo_marker
= 0;
1756 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1757 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1758 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1760 tcp_verify_left_out(tp
);
1762 /* Too bad if TCP was application limited */
1763 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1765 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1766 * The last condition is necessary at least in tp->frto_counter case.
1768 if (IsSackFrto() && (tp
->frto_counter
||
1769 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1770 after(tp
->high_seq
, tp
->snd_una
)) {
1771 tp
->frto_highmark
= tp
->high_seq
;
1773 tp
->frto_highmark
= tp
->snd_nxt
;
1775 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1776 tp
->high_seq
= tp
->snd_nxt
;
1777 tp
->frto_counter
= 1;
1780 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1781 * which indicates that we should follow the traditional RTO recovery,
1782 * i.e. mark everything lost and do go-back-N retransmission.
1784 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1786 struct tcp_sock
*tp
= tcp_sk(sk
);
1787 struct sk_buff
*skb
;
1790 tp
->retrans_out
= 0;
1791 if (tcp_is_reno(tp
))
1792 tcp_reset_reno_sack(tp
);
1794 tcp_for_write_queue(skb
, sk
) {
1795 if (skb
== tcp_send_head(sk
))
1798 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1800 * Count the retransmission made on RTO correctly (only when
1801 * waiting for the first ACK and did not get it)...
1803 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1804 /* For some reason this R-bit might get cleared? */
1805 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1806 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1807 /* ...enter this if branch just for the first segment */
1808 flag
|= FLAG_DATA_ACKED
;
1810 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1811 tp
->undo_marker
= 0;
1812 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1815 /* Don't lost mark skbs that were fwd transmitted after RTO */
1816 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) &&
1817 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
)) {
1818 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1819 tp
->lost_out
+= tcp_skb_pcount(skb
);
1822 tcp_verify_left_out(tp
);
1824 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1825 tp
->snd_cwnd_cnt
= 0;
1826 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1827 tp
->frto_counter
= 0;
1828 tp
->bytes_acked
= 0;
1830 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1831 sysctl_tcp_reordering
);
1832 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1833 tp
->high_seq
= tp
->frto_highmark
;
1834 TCP_ECN_queue_cwr(tp
);
1836 tcp_clear_retrans_hints_partial(tp
);
1839 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1841 tp
->retrans_out
= 0;
1844 tp
->undo_marker
= 0;
1845 tp
->undo_retrans
= 0;
1848 void tcp_clear_retrans(struct tcp_sock
*tp
)
1850 tcp_clear_retrans_partial(tp
);
1852 tp
->fackets_out
= 0;
1856 /* Enter Loss state. If "how" is not zero, forget all SACK information
1857 * and reset tags completely, otherwise preserve SACKs. If receiver
1858 * dropped its ofo queue, we will know this due to reneging detection.
1860 void tcp_enter_loss(struct sock
*sk
, int how
)
1862 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1863 struct tcp_sock
*tp
= tcp_sk(sk
);
1864 struct sk_buff
*skb
;
1866 /* Reduce ssthresh if it has not yet been made inside this window. */
1867 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1868 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1869 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1870 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1871 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1874 tp
->snd_cwnd_cnt
= 0;
1875 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1877 tp
->bytes_acked
= 0;
1878 tcp_clear_retrans_partial(tp
);
1880 if (tcp_is_reno(tp
))
1881 tcp_reset_reno_sack(tp
);
1884 /* Push undo marker, if it was plain RTO and nothing
1885 * was retransmitted. */
1886 tp
->undo_marker
= tp
->snd_una
;
1887 tcp_clear_retrans_hints_partial(tp
);
1890 tp
->fackets_out
= 0;
1891 tcp_clear_all_retrans_hints(tp
);
1894 tcp_for_write_queue(skb
, sk
) {
1895 if (skb
== tcp_send_head(sk
))
1898 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1899 tp
->undo_marker
= 0;
1900 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1901 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1902 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1903 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1904 tp
->lost_out
+= tcp_skb_pcount(skb
);
1907 tcp_verify_left_out(tp
);
1909 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1910 sysctl_tcp_reordering
);
1911 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1912 tp
->high_seq
= tp
->snd_nxt
;
1913 TCP_ECN_queue_cwr(tp
);
1914 /* Abort F-RTO algorithm if one is in progress */
1915 tp
->frto_counter
= 0;
1918 static int tcp_check_sack_reneging(struct sock
*sk
)
1920 struct sk_buff
*skb
;
1922 /* If ACK arrived pointing to a remembered SACK,
1923 * it means that our remembered SACKs do not reflect
1924 * real state of receiver i.e.
1925 * receiver _host_ is heavily congested (or buggy).
1926 * Do processing similar to RTO timeout.
1928 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1929 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1930 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1931 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1933 tcp_enter_loss(sk
, 1);
1934 icsk
->icsk_retransmits
++;
1935 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1936 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1937 icsk
->icsk_rto
, TCP_RTO_MAX
);
1943 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1945 return tcp_is_reno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1948 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1949 * counter when SACK is enabled (without SACK, sacked_out is used for
1952 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1953 * segments up to the highest received SACK block so far and holes in
1956 * With reordering, holes may still be in flight, so RFC3517 recovery
1957 * uses pure sacked_out (total number of SACKed segments) even though
1958 * it violates the RFC that uses duplicate ACKs, often these are equal
1959 * but when e.g. out-of-window ACKs or packet duplication occurs,
1960 * they differ. Since neither occurs due to loss, TCP should really
1963 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
1965 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1968 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1970 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1973 static inline int tcp_head_timedout(struct sock
*sk
)
1975 struct tcp_sock
*tp
= tcp_sk(sk
);
1977 return tp
->packets_out
&&
1978 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1981 /* Linux NewReno/SACK/FACK/ECN state machine.
1982 * --------------------------------------
1984 * "Open" Normal state, no dubious events, fast path.
1985 * "Disorder" In all the respects it is "Open",
1986 * but requires a bit more attention. It is entered when
1987 * we see some SACKs or dupacks. It is split of "Open"
1988 * mainly to move some processing from fast path to slow one.
1989 * "CWR" CWND was reduced due to some Congestion Notification event.
1990 * It can be ECN, ICMP source quench, local device congestion.
1991 * "Recovery" CWND was reduced, we are fast-retransmitting.
1992 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1994 * tcp_fastretrans_alert() is entered:
1995 * - each incoming ACK, if state is not "Open"
1996 * - when arrived ACK is unusual, namely:
2001 * Counting packets in flight is pretty simple.
2003 * in_flight = packets_out - left_out + retrans_out
2005 * packets_out is SND.NXT-SND.UNA counted in packets.
2007 * retrans_out is number of retransmitted segments.
2009 * left_out is number of segments left network, but not ACKed yet.
2011 * left_out = sacked_out + lost_out
2013 * sacked_out: Packets, which arrived to receiver out of order
2014 * and hence not ACKed. With SACKs this number is simply
2015 * amount of SACKed data. Even without SACKs
2016 * it is easy to give pretty reliable estimate of this number,
2017 * counting duplicate ACKs.
2019 * lost_out: Packets lost by network. TCP has no explicit
2020 * "loss notification" feedback from network (for now).
2021 * It means that this number can be only _guessed_.
2022 * Actually, it is the heuristics to predict lossage that
2023 * distinguishes different algorithms.
2025 * F.e. after RTO, when all the queue is considered as lost,
2026 * lost_out = packets_out and in_flight = retrans_out.
2028 * Essentially, we have now two algorithms counting
2031 * FACK: It is the simplest heuristics. As soon as we decided
2032 * that something is lost, we decide that _all_ not SACKed
2033 * packets until the most forward SACK are lost. I.e.
2034 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2035 * It is absolutely correct estimate, if network does not reorder
2036 * packets. And it loses any connection to reality when reordering
2037 * takes place. We use FACK by default until reordering
2038 * is suspected on the path to this destination.
2040 * NewReno: when Recovery is entered, we assume that one segment
2041 * is lost (classic Reno). While we are in Recovery and
2042 * a partial ACK arrives, we assume that one more packet
2043 * is lost (NewReno). This heuristics are the same in NewReno
2046 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2047 * deflation etc. CWND is real congestion window, never inflated, changes
2048 * only according to classic VJ rules.
2050 * Really tricky (and requiring careful tuning) part of algorithm
2051 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2052 * The first determines the moment _when_ we should reduce CWND and,
2053 * hence, slow down forward transmission. In fact, it determines the moment
2054 * when we decide that hole is caused by loss, rather than by a reorder.
2056 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2057 * holes, caused by lost packets.
2059 * And the most logically complicated part of algorithm is undo
2060 * heuristics. We detect false retransmits due to both too early
2061 * fast retransmit (reordering) and underestimated RTO, analyzing
2062 * timestamps and D-SACKs. When we detect that some segments were
2063 * retransmitted by mistake and CWND reduction was wrong, we undo
2064 * window reduction and abort recovery phase. This logic is hidden
2065 * inside several functions named tcp_try_undo_<something>.
2068 /* This function decides, when we should leave Disordered state
2069 * and enter Recovery phase, reducing congestion window.
2071 * Main question: may we further continue forward transmission
2072 * with the same cwnd?
2074 static int tcp_time_to_recover(struct sock
*sk
)
2076 struct tcp_sock
*tp
= tcp_sk(sk
);
2079 /* Do not perform any recovery during F-RTO algorithm */
2080 if (tp
->frto_counter
)
2083 /* Trick#1: The loss is proven. */
2087 /* Not-A-Trick#2 : Classic rule... */
2088 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2091 /* Trick#3 : when we use RFC2988 timer restart, fast
2092 * retransmit can be triggered by timeout of queue head.
2094 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2097 /* Trick#4: It is still not OK... But will it be useful to delay
2100 packets_out
= tp
->packets_out
;
2101 if (packets_out
<= tp
->reordering
&&
2102 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2103 !tcp_may_send_now(sk
)) {
2104 /* We have nothing to send. This connection is limited
2105 * either by receiver window or by application.
2113 /* RFC: This is from the original, I doubt that this is necessary at all:
2114 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2115 * retransmitted past LOST markings in the first place? I'm not fully sure
2116 * about undo and end of connection cases, which can cause R without L?
2118 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
2119 struct sk_buff
*skb
)
2121 if ((tp
->retransmit_skb_hint
!= NULL
) &&
2122 before(TCP_SKB_CB(skb
)->seq
,
2123 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
2124 tp
->retransmit_skb_hint
= NULL
;
2127 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2128 * is against sacked "cnt", otherwise it's against facked "cnt"
2130 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int fast_rexmit
)
2132 struct tcp_sock
*tp
= tcp_sk(sk
);
2133 struct sk_buff
*skb
;
2136 BUG_TRAP(packets
<= tp
->packets_out
);
2137 if (tp
->lost_skb_hint
) {
2138 skb
= tp
->lost_skb_hint
;
2139 cnt
= tp
->lost_cnt_hint
;
2141 skb
= tcp_write_queue_head(sk
);
2145 tcp_for_write_queue_from(skb
, sk
) {
2146 if (skb
== tcp_send_head(sk
))
2148 /* TODO: do this better */
2149 /* this is not the most efficient way to do this... */
2150 tp
->lost_skb_hint
= skb
;
2151 tp
->lost_cnt_hint
= cnt
;
2153 if (tcp_is_fack(tp
) ||
2154 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2155 cnt
+= tcp_skb_pcount(skb
);
2157 if (((!fast_rexmit
|| (tp
->lost_out
> 0)) && (cnt
> packets
)) ||
2158 after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2160 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2161 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2162 tp
->lost_out
+= tcp_skb_pcount(skb
);
2163 tcp_verify_retransmit_hint(tp
, skb
);
2166 tcp_verify_left_out(tp
);
2169 /* Account newly detected lost packet(s) */
2171 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2173 struct tcp_sock
*tp
= tcp_sk(sk
);
2175 if (tcp_is_reno(tp
)) {
2176 tcp_mark_head_lost(sk
, 1, fast_rexmit
);
2177 } else if (tcp_is_fack(tp
)) {
2178 int lost
= tp
->fackets_out
- tp
->reordering
;
2181 tcp_mark_head_lost(sk
, lost
, fast_rexmit
);
2183 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2184 if (sacked_upto
< 0)
2186 tcp_mark_head_lost(sk
, sacked_upto
, fast_rexmit
);
2189 /* New heuristics: it is possible only after we switched
2190 * to restart timer each time when something is ACKed.
2191 * Hence, we can detect timed out packets during fast
2192 * retransmit without falling to slow start.
2194 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2195 struct sk_buff
*skb
;
2197 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2198 : tcp_write_queue_head(sk
);
2200 tcp_for_write_queue_from(skb
, sk
) {
2201 if (skb
== tcp_send_head(sk
))
2203 if (!tcp_skb_timedout(sk
, skb
))
2206 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2207 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2208 tp
->lost_out
+= tcp_skb_pcount(skb
);
2209 tcp_verify_retransmit_hint(tp
, skb
);
2213 tp
->scoreboard_skb_hint
= skb
;
2215 tcp_verify_left_out(tp
);
2219 /* CWND moderation, preventing bursts due to too big ACKs
2220 * in dubious situations.
2222 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2224 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2225 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
2226 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2229 /* Lower bound on congestion window is slow start threshold
2230 * unless congestion avoidance choice decides to overide it.
2232 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2234 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2236 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2239 /* Decrease cwnd each second ack. */
2240 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2242 struct tcp_sock
*tp
= tcp_sk(sk
);
2243 int decr
= tp
->snd_cwnd_cnt
+ 1;
2245 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
2246 (tcp_is_reno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
2247 tp
->snd_cwnd_cnt
= decr
&1;
2250 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2251 tp
->snd_cwnd
-= decr
;
2253 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
2254 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2258 /* Nothing was retransmitted or returned timestamp is less
2259 * than timestamp of the first retransmission.
2261 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2263 return !tp
->retrans_stamp
||
2264 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2265 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
2268 /* Undo procedures. */
2270 #if FASTRETRANS_DEBUG > 1
2271 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2273 struct tcp_sock
*tp
= tcp_sk(sk
);
2274 struct inet_sock
*inet
= inet_sk(sk
);
2276 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2278 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2279 tp
->snd_cwnd
, tcp_left_out(tp
),
2280 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2284 #define DBGUNDO(x...) do { } while (0)
2287 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2289 struct tcp_sock
*tp
= tcp_sk(sk
);
2291 if (tp
->prior_ssthresh
) {
2292 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2294 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2295 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2297 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
2299 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2300 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2301 TCP_ECN_withdraw_cwr(tp
);
2304 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2306 tcp_moderate_cwnd(tp
);
2307 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2309 /* There is something screwy going on with the retrans hints after
2311 tcp_clear_all_retrans_hints(tp
);
2314 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2316 return tp
->undo_marker
&&
2317 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2320 /* People celebrate: "We love our President!" */
2321 static int tcp_try_undo_recovery(struct sock
*sk
)
2323 struct tcp_sock
*tp
= tcp_sk(sk
);
2325 if (tcp_may_undo(tp
)) {
2326 /* Happy end! We did not retransmit anything
2327 * or our original transmission succeeded.
2329 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2330 tcp_undo_cwr(sk
, 1);
2331 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2332 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2334 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
2335 tp
->undo_marker
= 0;
2337 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2338 /* Hold old state until something *above* high_seq
2339 * is ACKed. For Reno it is MUST to prevent false
2340 * fast retransmits (RFC2582). SACK TCP is safe. */
2341 tcp_moderate_cwnd(tp
);
2344 tcp_set_ca_state(sk
, TCP_CA_Open
);
2348 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2349 static void tcp_try_undo_dsack(struct sock
*sk
)
2351 struct tcp_sock
*tp
= tcp_sk(sk
);
2353 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2354 DBGUNDO(sk
, "D-SACK");
2355 tcp_undo_cwr(sk
, 1);
2356 tp
->undo_marker
= 0;
2357 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2361 /* Undo during fast recovery after partial ACK. */
2363 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2365 struct tcp_sock
*tp
= tcp_sk(sk
);
2366 /* Partial ACK arrived. Force Hoe's retransmit. */
2367 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2369 if (tcp_may_undo(tp
)) {
2370 /* Plain luck! Hole if filled with delayed
2371 * packet, rather than with a retransmit.
2373 if (tp
->retrans_out
== 0)
2374 tp
->retrans_stamp
= 0;
2376 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2379 tcp_undo_cwr(sk
, 0);
2380 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2382 /* So... Do not make Hoe's retransmit yet.
2383 * If the first packet was delayed, the rest
2384 * ones are most probably delayed as well.
2391 /* Undo during loss recovery after partial ACK. */
2392 static int tcp_try_undo_loss(struct sock
*sk
)
2394 struct tcp_sock
*tp
= tcp_sk(sk
);
2396 if (tcp_may_undo(tp
)) {
2397 struct sk_buff
*skb
;
2398 tcp_for_write_queue(skb
, sk
) {
2399 if (skb
== tcp_send_head(sk
))
2401 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2404 tcp_clear_all_retrans_hints(tp
);
2406 DBGUNDO(sk
, "partial loss");
2408 tcp_undo_cwr(sk
, 1);
2409 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2410 inet_csk(sk
)->icsk_retransmits
= 0;
2411 tp
->undo_marker
= 0;
2412 if (tcp_is_sack(tp
))
2413 tcp_set_ca_state(sk
, TCP_CA_Open
);
2419 static inline void tcp_complete_cwr(struct sock
*sk
)
2421 struct tcp_sock
*tp
= tcp_sk(sk
);
2422 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2423 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2424 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2427 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2429 struct tcp_sock
*tp
= tcp_sk(sk
);
2431 tcp_verify_left_out(tp
);
2433 if (tp
->retrans_out
== 0)
2434 tp
->retrans_stamp
= 0;
2437 tcp_enter_cwr(sk
, 1);
2439 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2440 int state
= TCP_CA_Open
;
2442 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2443 state
= TCP_CA_Disorder
;
2445 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2446 tcp_set_ca_state(sk
, state
);
2447 tp
->high_seq
= tp
->snd_nxt
;
2449 tcp_moderate_cwnd(tp
);
2451 tcp_cwnd_down(sk
, flag
);
2455 static void tcp_mtup_probe_failed(struct sock
*sk
)
2457 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2459 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2460 icsk
->icsk_mtup
.probe_size
= 0;
2463 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2465 struct tcp_sock
*tp
= tcp_sk(sk
);
2466 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2468 /* FIXME: breaks with very large cwnd */
2469 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2470 tp
->snd_cwnd
= tp
->snd_cwnd
*
2471 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2472 icsk
->icsk_mtup
.probe_size
;
2473 tp
->snd_cwnd_cnt
= 0;
2474 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2475 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2477 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2478 icsk
->icsk_mtup
.probe_size
= 0;
2479 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2483 /* Process an event, which can update packets-in-flight not trivially.
2484 * Main goal of this function is to calculate new estimate for left_out,
2485 * taking into account both packets sitting in receiver's buffer and
2486 * packets lost by network.
2488 * Besides that it does CWND reduction, when packet loss is detected
2489 * and changes state of machine.
2491 * It does _not_ decide what to send, it is made in function
2492 * tcp_xmit_retransmit_queue().
2495 tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2497 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2498 struct tcp_sock
*tp
= tcp_sk(sk
);
2499 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2500 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2501 (tcp_fackets_out(tp
) > tp
->reordering
));
2502 int fast_rexmit
= 0;
2504 /* Some technical things:
2505 * 1. Reno does not count dupacks (sacked_out) automatically. */
2506 if (!tp
->packets_out
)
2509 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2510 tp
->fackets_out
= 0;
2512 /* Now state machine starts.
2513 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2515 tp
->prior_ssthresh
= 0;
2517 /* B. In all the states check for reneging SACKs. */
2518 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2521 /* C. Process data loss notification, provided it is valid. */
2522 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2523 before(tp
->snd_una
, tp
->high_seq
) &&
2524 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2525 tp
->fackets_out
> tp
->reordering
) {
2526 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, 0);
2527 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2530 /* D. Check consistency of the current state. */
2531 tcp_verify_left_out(tp
);
2533 /* E. Check state exit conditions. State can be terminated
2534 * when high_seq is ACKed. */
2535 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2536 BUG_TRAP(tp
->retrans_out
== 0);
2537 tp
->retrans_stamp
= 0;
2538 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2539 switch (icsk
->icsk_ca_state
) {
2541 icsk
->icsk_retransmits
= 0;
2542 if (tcp_try_undo_recovery(sk
))
2547 /* CWR is to be held something *above* high_seq
2548 * is ACKed for CWR bit to reach receiver. */
2549 if (tp
->snd_una
!= tp
->high_seq
) {
2550 tcp_complete_cwr(sk
);
2551 tcp_set_ca_state(sk
, TCP_CA_Open
);
2555 case TCP_CA_Disorder
:
2556 tcp_try_undo_dsack(sk
);
2557 if (!tp
->undo_marker
||
2558 /* For SACK case do not Open to allow to undo
2559 * catching for all duplicate ACKs. */
2560 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2561 tp
->undo_marker
= 0;
2562 tcp_set_ca_state(sk
, TCP_CA_Open
);
2566 case TCP_CA_Recovery
:
2567 if (tcp_is_reno(tp
))
2568 tcp_reset_reno_sack(tp
);
2569 if (tcp_try_undo_recovery(sk
))
2571 tcp_complete_cwr(sk
);
2576 /* F. Process state. */
2577 switch (icsk
->icsk_ca_state
) {
2578 case TCP_CA_Recovery
:
2579 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2580 if (tcp_is_reno(tp
) && is_dupack
)
2581 tcp_add_reno_sack(sk
);
2583 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2586 if (flag
&FLAG_DATA_ACKED
)
2587 icsk
->icsk_retransmits
= 0;
2588 if (!tcp_try_undo_loss(sk
)) {
2589 tcp_moderate_cwnd(tp
);
2590 tcp_xmit_retransmit_queue(sk
);
2593 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2595 /* Loss is undone; fall through to processing in Open state. */
2597 if (tcp_is_reno(tp
)) {
2598 if (flag
& FLAG_SND_UNA_ADVANCED
)
2599 tcp_reset_reno_sack(tp
);
2601 tcp_add_reno_sack(sk
);
2604 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2605 tcp_try_undo_dsack(sk
);
2607 if (!tcp_time_to_recover(sk
)) {
2608 tcp_try_to_open(sk
, flag
);
2612 /* MTU probe failure: don't reduce cwnd */
2613 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2614 icsk
->icsk_mtup
.probe_size
&&
2615 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2616 tcp_mtup_probe_failed(sk
);
2617 /* Restores the reduction we did in tcp_mtup_probe() */
2619 tcp_simple_retransmit(sk
);
2623 /* Otherwise enter Recovery state */
2625 if (tcp_is_reno(tp
))
2626 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2628 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2630 tp
->high_seq
= tp
->snd_nxt
;
2631 tp
->prior_ssthresh
= 0;
2632 tp
->undo_marker
= tp
->snd_una
;
2633 tp
->undo_retrans
= tp
->retrans_out
;
2635 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2636 if (!(flag
&FLAG_ECE
))
2637 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2638 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2639 TCP_ECN_queue_cwr(tp
);
2642 tp
->bytes_acked
= 0;
2643 tp
->snd_cwnd_cnt
= 0;
2644 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2648 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2649 tcp_update_scoreboard(sk
, fast_rexmit
);
2650 tcp_cwnd_down(sk
, flag
);
2651 tcp_xmit_retransmit_queue(sk
);
2654 /* Read draft-ietf-tcplw-high-performance before mucking
2655 * with this code. (Supersedes RFC1323)
2657 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2659 /* RTTM Rule: A TSecr value received in a segment is used to
2660 * update the averaged RTT measurement only if the segment
2661 * acknowledges some new data, i.e., only if it advances the
2662 * left edge of the send window.
2664 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2665 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2667 * Changed: reset backoff as soon as we see the first valid sample.
2668 * If we do not, we get strongly overestimated rto. With timestamps
2669 * samples are accepted even from very old segments: f.e., when rtt=1
2670 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2671 * answer arrives rto becomes 120 seconds! If at least one of segments
2672 * in window is lost... Voila. --ANK (010210)
2674 struct tcp_sock
*tp
= tcp_sk(sk
);
2675 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2676 tcp_rtt_estimator(sk
, seq_rtt
);
2678 inet_csk(sk
)->icsk_backoff
= 0;
2682 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2684 /* We don't have a timestamp. Can only use
2685 * packets that are not retransmitted to determine
2686 * rtt estimates. Also, we must not reset the
2687 * backoff for rto until we get a non-retransmitted
2688 * packet. This allows us to deal with a situation
2689 * where the network delay has increased suddenly.
2690 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2693 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2696 tcp_rtt_estimator(sk
, seq_rtt
);
2698 inet_csk(sk
)->icsk_backoff
= 0;
2702 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2705 const struct tcp_sock
*tp
= tcp_sk(sk
);
2706 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2707 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2708 tcp_ack_saw_tstamp(sk
, flag
);
2709 else if (seq_rtt
>= 0)
2710 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2713 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2715 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2716 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2717 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2720 /* Restart timer after forward progress on connection.
2721 * RFC2988 recommends to restart timer to now+rto.
2723 static void tcp_rearm_rto(struct sock
*sk
)
2725 struct tcp_sock
*tp
= tcp_sk(sk
);
2727 if (!tp
->packets_out
) {
2728 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2730 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2734 /* If we get here, the whole TSO packet has not been acked. */
2735 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2737 struct tcp_sock
*tp
= tcp_sk(sk
);
2740 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2742 packets_acked
= tcp_skb_pcount(skb
);
2743 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2745 packets_acked
-= tcp_skb_pcount(skb
);
2747 if (packets_acked
) {
2748 BUG_ON(tcp_skb_pcount(skb
) == 0);
2749 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2752 return packets_acked
;
2755 /* Remove acknowledged frames from the retransmission queue. If our packet
2756 * is before the ack sequence we can discard it as it's confirmed to have
2757 * arrived at the other end.
2759 static int tcp_clean_rtx_queue(struct sock
*sk
, s32
*seq_rtt_p
,
2762 struct tcp_sock
*tp
= tcp_sk(sk
);
2763 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2764 struct sk_buff
*skb
;
2765 u32 now
= tcp_time_stamp
;
2766 int fully_acked
= 1;
2768 int prior_packets
= tp
->packets_out
;
2770 u32 reord
= tp
->packets_out
;
2772 s32 ca_seq_rtt
= -1;
2773 ktime_t last_ackt
= net_invalid_timestamp();
2775 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2776 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2779 u8 sacked
= scb
->sacked
;
2781 /* Determine how many packets and what bytes were acked, tso and else */
2782 if (after(scb
->end_seq
, tp
->snd_una
)) {
2783 if (tcp_skb_pcount(skb
) == 1 ||
2784 !after(tp
->snd_una
, scb
->seq
))
2787 packets_acked
= tcp_tso_acked(sk
, skb
);
2792 end_seq
= tp
->snd_una
;
2794 packets_acked
= tcp_skb_pcount(skb
);
2795 end_seq
= scb
->end_seq
;
2798 /* MTU probing checks */
2799 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2800 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2801 tcp_mtup_probe_success(sk
, skb
);
2805 if (sacked
& TCPCB_RETRANS
) {
2806 if (sacked
& TCPCB_SACKED_RETRANS
)
2807 tp
->retrans_out
-= packets_acked
;
2808 flag
|= FLAG_RETRANS_DATA_ACKED
;
2811 if ((flag
& FLAG_DATA_ACKED
) ||
2812 (packets_acked
> 1))
2813 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2815 ca_seq_rtt
= now
- scb
->when
;
2816 last_ackt
= skb
->tstamp
;
2818 seq_rtt
= ca_seq_rtt
;
2820 if (!(sacked
& TCPCB_SACKED_ACKED
))
2821 reord
= min(cnt
, reord
);
2824 if (sacked
& TCPCB_SACKED_ACKED
)
2825 tp
->sacked_out
-= packets_acked
;
2826 if (sacked
& TCPCB_LOST
)
2827 tp
->lost_out
-= packets_acked
;
2829 if ((sacked
& TCPCB_URG
) && tp
->urg_mode
&&
2830 !before(end_seq
, tp
->snd_up
))
2833 ca_seq_rtt
= now
- scb
->when
;
2834 last_ackt
= skb
->tstamp
;
2836 seq_rtt
= ca_seq_rtt
;
2838 reord
= min(cnt
, reord
);
2840 tp
->packets_out
-= packets_acked
;
2841 cnt
+= packets_acked
;
2843 /* Initial outgoing SYN's get put onto the write_queue
2844 * just like anything else we transmit. It is not
2845 * true data, and if we misinform our callers that
2846 * this ACK acks real data, we will erroneously exit
2847 * connection startup slow start one packet too
2848 * quickly. This is severely frowned upon behavior.
2850 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2851 flag
|= FLAG_DATA_ACKED
;
2853 flag
|= FLAG_SYN_ACKED
;
2854 tp
->retrans_stamp
= 0;
2860 tcp_unlink_write_queue(skb
, sk
);
2861 sk_stream_free_skb(sk
, skb
);
2862 tcp_clear_all_retrans_hints(tp
);
2865 if (flag
& FLAG_ACKED
) {
2866 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2867 const struct tcp_congestion_ops
*ca_ops
2868 = inet_csk(sk
)->icsk_ca_ops
;
2870 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2873 if (tcp_is_reno(tp
)) {
2874 tcp_remove_reno_sacks(sk
, pkts_acked
);
2876 /* Non-retransmitted hole got filled? That's reordering */
2877 if (reord
< prior_fackets
)
2878 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2881 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2883 if (ca_ops
->pkts_acked
) {
2886 /* Is the ACK triggering packet unambiguous? */
2887 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2888 /* High resolution needed and available? */
2889 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2890 !ktime_equal(last_ackt
,
2891 net_invalid_timestamp()))
2892 rtt_us
= ktime_us_delta(ktime_get_real(),
2894 else if (ca_seq_rtt
> 0)
2895 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
2898 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2902 #if FASTRETRANS_DEBUG > 0
2903 BUG_TRAP((int)tp
->sacked_out
>= 0);
2904 BUG_TRAP((int)tp
->lost_out
>= 0);
2905 BUG_TRAP((int)tp
->retrans_out
>= 0);
2906 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2907 icsk
= inet_csk(sk
);
2909 printk(KERN_DEBUG
"Leak l=%u %d\n",
2910 tp
->lost_out
, icsk
->icsk_ca_state
);
2913 if (tp
->sacked_out
) {
2914 printk(KERN_DEBUG
"Leak s=%u %d\n",
2915 tp
->sacked_out
, icsk
->icsk_ca_state
);
2918 if (tp
->retrans_out
) {
2919 printk(KERN_DEBUG
"Leak r=%u %d\n",
2920 tp
->retrans_out
, icsk
->icsk_ca_state
);
2921 tp
->retrans_out
= 0;
2925 *seq_rtt_p
= seq_rtt
;
2929 static void tcp_ack_probe(struct sock
*sk
)
2931 const struct tcp_sock
*tp
= tcp_sk(sk
);
2932 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2934 /* Was it a usable window open? */
2936 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2937 tp
->snd_una
+ tp
->snd_wnd
)) {
2938 icsk
->icsk_backoff
= 0;
2939 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2940 /* Socket must be waked up by subsequent tcp_data_snd_check().
2941 * This function is not for random using!
2944 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2945 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2950 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2952 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2953 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2956 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2958 const struct tcp_sock
*tp
= tcp_sk(sk
);
2959 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2960 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2963 /* Check that window update is acceptable.
2964 * The function assumes that snd_una<=ack<=snd_next.
2966 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2967 const u32 ack_seq
, const u32 nwin
)
2969 return (after(ack
, tp
->snd_una
) ||
2970 after(ack_seq
, tp
->snd_wl1
) ||
2971 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2974 /* Update our send window.
2976 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2977 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2979 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2982 struct tcp_sock
*tp
= tcp_sk(sk
);
2984 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2986 if (likely(!tcp_hdr(skb
)->syn
))
2987 nwin
<<= tp
->rx_opt
.snd_wscale
;
2989 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
2990 flag
|= FLAG_WIN_UPDATE
;
2991 tcp_update_wl(tp
, ack
, ack_seq
);
2993 if (tp
->snd_wnd
!= nwin
) {
2996 /* Note, it is the only place, where
2997 * fast path is recovered for sending TCP.
3000 tcp_fast_path_check(sk
);
3002 if (nwin
> tp
->max_window
) {
3003 tp
->max_window
= nwin
;
3004 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3014 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3015 * continue in congestion avoidance.
3017 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3019 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3020 tp
->snd_cwnd_cnt
= 0;
3021 tp
->bytes_acked
= 0;
3022 TCP_ECN_queue_cwr(tp
);
3023 tcp_moderate_cwnd(tp
);
3026 /* A conservative spurious RTO response algorithm: reduce cwnd using
3027 * rate halving and continue in congestion avoidance.
3029 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3031 tcp_enter_cwr(sk
, 0);
3034 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3037 tcp_ratehalving_spur_to_response(sk
);
3039 tcp_undo_cwr(sk
, 1);
3042 /* F-RTO spurious RTO detection algorithm (RFC4138)
3044 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3045 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3046 * window (but not to or beyond highest sequence sent before RTO):
3047 * On First ACK, send two new segments out.
3048 * On Second ACK, RTO was likely spurious. Do spurious response (response
3049 * algorithm is not part of the F-RTO detection algorithm
3050 * given in RFC4138 but can be selected separately).
3051 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3052 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3053 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3054 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3056 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3057 * original window even after we transmit two new data segments.
3060 * on first step, wait until first cumulative ACK arrives, then move to
3061 * the second step. In second step, the next ACK decides.
3063 * F-RTO is implemented (mainly) in four functions:
3064 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3065 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3066 * called when tcp_use_frto() showed green light
3067 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3068 * - tcp_enter_frto_loss() is called if there is not enough evidence
3069 * to prove that the RTO is indeed spurious. It transfers the control
3070 * from F-RTO to the conventional RTO recovery
3072 static int tcp_process_frto(struct sock
*sk
, int flag
)
3074 struct tcp_sock
*tp
= tcp_sk(sk
);
3076 tcp_verify_left_out(tp
);
3078 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3079 if (flag
&FLAG_DATA_ACKED
)
3080 inet_csk(sk
)->icsk_retransmits
= 0;
3082 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3083 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3084 tp
->undo_marker
= 0;
3086 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3087 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3091 if (!IsSackFrto() || tcp_is_reno(tp
)) {
3092 /* RFC4138 shortcoming in step 2; should also have case c):
3093 * ACK isn't duplicate nor advances window, e.g., opposite dir
3096 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3099 if (!(flag
&FLAG_DATA_ACKED
)) {
3100 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3105 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3106 /* Prevent sending of new data. */
3107 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3108 tcp_packets_in_flight(tp
));
3112 if ((tp
->frto_counter
>= 2) &&
3113 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
3114 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
3115 /* RFC4138 shortcoming (see comment above) */
3116 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3119 tcp_enter_frto_loss(sk
, 3, flag
);
3124 if (tp
->frto_counter
== 1) {
3125 /* tcp_may_send_now needs to see updated state */
3126 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3127 tp
->frto_counter
= 2;
3129 if (!tcp_may_send_now(sk
))
3130 tcp_enter_frto_loss(sk
, 2, flag
);
3134 switch (sysctl_tcp_frto_response
) {
3136 tcp_undo_spur_to_response(sk
, flag
);
3139 tcp_conservative_spur_to_response(tp
);
3142 tcp_ratehalving_spur_to_response(sk
);
3145 tp
->frto_counter
= 0;
3146 tp
->undo_marker
= 0;
3147 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS
);
3152 /* This routine deals with incoming acks, but not outgoing ones. */
3153 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3155 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3156 struct tcp_sock
*tp
= tcp_sk(sk
);
3157 u32 prior_snd_una
= tp
->snd_una
;
3158 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3159 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3160 u32 prior_in_flight
;
3166 /* If the ack is newer than sent or older than previous acks
3167 * then we can probably ignore it.
3169 if (after(ack
, tp
->snd_nxt
))
3170 goto uninteresting_ack
;
3172 if (before(ack
, prior_snd_una
))
3175 if (after(ack
, prior_snd_una
))
3176 flag
|= FLAG_SND_UNA_ADVANCED
;
3178 if (sysctl_tcp_abc
) {
3179 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3180 tp
->bytes_acked
+= ack
- prior_snd_una
;
3181 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3182 /* we assume just one segment left network */
3183 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
3186 prior_fackets
= tp
->fackets_out
;
3187 prior_in_flight
= tcp_packets_in_flight(tp
);
3189 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3190 /* Window is constant, pure forward advance.
3191 * No more checks are required.
3192 * Note, we use the fact that SND.UNA>=SND.WL2.
3194 tcp_update_wl(tp
, ack
, ack_seq
);
3196 flag
|= FLAG_WIN_UPDATE
;
3198 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3200 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
3202 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3205 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
3207 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3209 if (TCP_SKB_CB(skb
)->sacked
)
3210 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3212 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3215 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3218 /* We passed data and got it acked, remove any soft error
3219 * log. Something worked...
3221 sk
->sk_err_soft
= 0;
3222 tp
->rcv_tstamp
= tcp_time_stamp
;
3223 prior_packets
= tp
->packets_out
;
3227 /* See if we can take anything off of the retransmit queue. */
3228 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
, prior_fackets
);
3230 if (tp
->frto_counter
)
3231 frto_cwnd
= tcp_process_frto(sk
, flag
);
3232 /* Guarantee sacktag reordering detection against wrap-arounds */
3233 if (before(tp
->frto_highmark
, tp
->snd_una
))
3234 tp
->frto_highmark
= 0;
3236 if (tcp_ack_is_dubious(sk
, flag
)) {
3237 /* Advance CWND, if state allows this. */
3238 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3239 tcp_may_raise_cwnd(sk
, flag
))
3240 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3241 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
, flag
);
3243 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3244 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3247 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
3248 dst_confirm(sk
->sk_dst_cache
);
3253 icsk
->icsk_probes_out
= 0;
3255 /* If this ack opens up a zero window, clear backoff. It was
3256 * being used to time the probes, and is probably far higher than
3257 * it needs to be for normal retransmission.
3259 if (tcp_send_head(sk
))
3264 if (TCP_SKB_CB(skb
)->sacked
)
3265 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3268 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3273 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3274 * But, this can also be called on packets in the established flow when
3275 * the fast version below fails.
3277 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
3280 struct tcphdr
*th
= tcp_hdr(skb
);
3281 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
3283 ptr
= (unsigned char *)(th
+ 1);
3284 opt_rx
->saw_tstamp
= 0;
3286 while (length
> 0) {
3293 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3298 if (opsize
< 2) /* "silly options" */
3300 if (opsize
> length
)
3301 return; /* don't parse partial options */
3304 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3305 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
3307 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
3308 in_mss
= opt_rx
->user_mss
;
3309 opt_rx
->mss_clamp
= in_mss
;
3314 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
3315 if (sysctl_tcp_window_scaling
) {
3316 __u8 snd_wscale
= *(__u8
*) ptr
;
3317 opt_rx
->wscale_ok
= 1;
3318 if (snd_wscale
> 14) {
3319 if (net_ratelimit())
3320 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3321 "scaling value %d >14 received.\n",
3325 opt_rx
->snd_wscale
= snd_wscale
;
3328 case TCPOPT_TIMESTAMP
:
3329 if (opsize
==TCPOLEN_TIMESTAMP
) {
3330 if ((estab
&& opt_rx
->tstamp_ok
) ||
3331 (!estab
&& sysctl_tcp_timestamps
)) {
3332 opt_rx
->saw_tstamp
= 1;
3333 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
3334 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
3338 case TCPOPT_SACK_PERM
:
3339 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
3340 if (sysctl_tcp_sack
) {
3341 opt_rx
->sack_ok
= 1;
3342 tcp_sack_reset(opt_rx
);
3348 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3349 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3351 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3354 #ifdef CONFIG_TCP_MD5SIG
3357 * The MD5 Hash has already been
3358 * checked (see tcp_v{4,6}_do_rcv()).
3370 /* Fast parse options. This hopes to only see timestamps.
3371 * If it is wrong it falls back on tcp_parse_options().
3373 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3374 struct tcp_sock
*tp
)
3376 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3377 tp
->rx_opt
.saw_tstamp
= 0;
3379 } else if (tp
->rx_opt
.tstamp_ok
&&
3380 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3381 __be32
*ptr
= (__be32
*)(th
+ 1);
3382 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3383 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3384 tp
->rx_opt
.saw_tstamp
= 1;
3386 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3388 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3392 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3396 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3398 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3399 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3402 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3404 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3405 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3406 * extra check below makes sure this can only happen
3407 * for pure ACK frames. -DaveM
3409 * Not only, also it occurs for expired timestamps.
3412 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3413 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3414 tcp_store_ts_recent(tp
);
3418 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3420 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3421 * it can pass through stack. So, the following predicate verifies that
3422 * this segment is not used for anything but congestion avoidance or
3423 * fast retransmit. Moreover, we even are able to eliminate most of such
3424 * second order effects, if we apply some small "replay" window (~RTO)
3425 * to timestamp space.
3427 * All these measures still do not guarantee that we reject wrapped ACKs
3428 * on networks with high bandwidth, when sequence space is recycled fastly,
3429 * but it guarantees that such events will be very rare and do not affect
3430 * connection seriously. This doesn't look nice, but alas, PAWS is really
3433 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3434 * states that events when retransmit arrives after original data are rare.
3435 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3436 * the biggest problem on large power networks even with minor reordering.
3437 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3438 * up to bandwidth of 18Gigabit/sec. 8) ]
3441 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3443 struct tcp_sock
*tp
= tcp_sk(sk
);
3444 struct tcphdr
*th
= tcp_hdr(skb
);
3445 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3446 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3448 return (/* 1. Pure ACK with correct sequence number. */
3449 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3451 /* 2. ... and duplicate ACK. */
3452 ack
== tp
->snd_una
&&
3454 /* 3. ... and does not update window. */
3455 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3457 /* 4. ... and sits in replay window. */
3458 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3461 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3463 const struct tcp_sock
*tp
= tcp_sk(sk
);
3464 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3465 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3466 !tcp_disordered_ack(sk
, skb
));
3469 /* Check segment sequence number for validity.
3471 * Segment controls are considered valid, if the segment
3472 * fits to the window after truncation to the window. Acceptability
3473 * of data (and SYN, FIN, of course) is checked separately.
3474 * See tcp_data_queue(), for example.
3476 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3477 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3478 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3479 * (borrowed from freebsd)
3482 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3484 return !before(end_seq
, tp
->rcv_wup
) &&
3485 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3488 /* When we get a reset we do this. */
3489 static void tcp_reset(struct sock
*sk
)
3491 /* We want the right error as BSD sees it (and indeed as we do). */
3492 switch (sk
->sk_state
) {
3494 sk
->sk_err
= ECONNREFUSED
;
3496 case TCP_CLOSE_WAIT
:
3502 sk
->sk_err
= ECONNRESET
;
3505 if (!sock_flag(sk
, SOCK_DEAD
))
3506 sk
->sk_error_report(sk
);
3512 * Process the FIN bit. This now behaves as it is supposed to work
3513 * and the FIN takes effect when it is validly part of sequence
3514 * space. Not before when we get holes.
3516 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3517 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3520 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3521 * close and we go into CLOSING (and later onto TIME-WAIT)
3523 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3525 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3527 struct tcp_sock
*tp
= tcp_sk(sk
);
3529 inet_csk_schedule_ack(sk
);
3531 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3532 sock_set_flag(sk
, SOCK_DONE
);
3534 switch (sk
->sk_state
) {
3536 case TCP_ESTABLISHED
:
3537 /* Move to CLOSE_WAIT */
3538 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3539 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3542 case TCP_CLOSE_WAIT
:
3544 /* Received a retransmission of the FIN, do
3549 /* RFC793: Remain in the LAST-ACK state. */
3553 /* This case occurs when a simultaneous close
3554 * happens, we must ack the received FIN and
3555 * enter the CLOSING state.
3558 tcp_set_state(sk
, TCP_CLOSING
);
3561 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3563 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3566 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3567 * cases we should never reach this piece of code.
3569 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3570 __FUNCTION__
, sk
->sk_state
);
3574 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3575 * Probably, we should reset in this case. For now drop them.
3577 __skb_queue_purge(&tp
->out_of_order_queue
);
3578 if (tcp_is_sack(tp
))
3579 tcp_sack_reset(&tp
->rx_opt
);
3580 sk_stream_mem_reclaim(sk
);
3582 if (!sock_flag(sk
, SOCK_DEAD
)) {
3583 sk
->sk_state_change(sk
);
3585 /* Do not send POLL_HUP for half duplex close. */
3586 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3587 sk
->sk_state
== TCP_CLOSE
)
3588 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3590 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3594 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3596 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3597 if (before(seq
, sp
->start_seq
))
3598 sp
->start_seq
= seq
;
3599 if (after(end_seq
, sp
->end_seq
))
3600 sp
->end_seq
= end_seq
;
3606 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3608 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3609 if (before(seq
, tp
->rcv_nxt
))
3610 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3612 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3614 tp
->rx_opt
.dsack
= 1;
3615 tp
->duplicate_sack
[0].start_seq
= seq
;
3616 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3617 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3621 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3623 if (!tp
->rx_opt
.dsack
)
3624 tcp_dsack_set(tp
, seq
, end_seq
);
3626 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3629 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3631 struct tcp_sock
*tp
= tcp_sk(sk
);
3633 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3634 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3635 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3636 tcp_enter_quickack_mode(sk
);
3638 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3639 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3641 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3642 end_seq
= tp
->rcv_nxt
;
3643 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3650 /* These routines update the SACK block as out-of-order packets arrive or
3651 * in-order packets close up the sequence space.
3653 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3656 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3657 struct tcp_sack_block
*swalk
= sp
+1;
3659 /* See if the recent change to the first SACK eats into
3660 * or hits the sequence space of other SACK blocks, if so coalesce.
3662 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3663 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3666 /* Zap SWALK, by moving every further SACK up by one slot.
3667 * Decrease num_sacks.
3669 tp
->rx_opt
.num_sacks
--;
3670 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3671 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3675 this_sack
++, swalk
++;
3679 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3683 tmp
= sack1
->start_seq
;
3684 sack1
->start_seq
= sack2
->start_seq
;
3685 sack2
->start_seq
= tmp
;
3687 tmp
= sack1
->end_seq
;
3688 sack1
->end_seq
= sack2
->end_seq
;
3689 sack2
->end_seq
= tmp
;
3692 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3694 struct tcp_sock
*tp
= tcp_sk(sk
);
3695 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3696 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3702 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3703 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3704 /* Rotate this_sack to the first one. */
3705 for (; this_sack
>0; this_sack
--, sp
--)
3706 tcp_sack_swap(sp
, sp
-1);
3708 tcp_sack_maybe_coalesce(tp
);
3713 /* Could not find an adjacent existing SACK, build a new one,
3714 * put it at the front, and shift everyone else down. We
3715 * always know there is at least one SACK present already here.
3717 * If the sack array is full, forget about the last one.
3719 if (this_sack
>= 4) {
3721 tp
->rx_opt
.num_sacks
--;
3724 for (; this_sack
> 0; this_sack
--, sp
--)
3728 /* Build the new head SACK, and we're done. */
3729 sp
->start_seq
= seq
;
3730 sp
->end_seq
= end_seq
;
3731 tp
->rx_opt
.num_sacks
++;
3732 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3735 /* RCV.NXT advances, some SACKs should be eaten. */
3737 static void tcp_sack_remove(struct tcp_sock
*tp
)
3739 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3740 int num_sacks
= tp
->rx_opt
.num_sacks
;
3743 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3744 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3745 tp
->rx_opt
.num_sacks
= 0;
3746 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3750 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3751 /* Check if the start of the sack is covered by RCV.NXT. */
3752 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3755 /* RCV.NXT must cover all the block! */
3756 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3758 /* Zap this SACK, by moving forward any other SACKS. */
3759 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3760 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3767 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3768 tp
->rx_opt
.num_sacks
= num_sacks
;
3769 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3773 /* This one checks to see if we can put data from the
3774 * out_of_order queue into the receive_queue.
3776 static void tcp_ofo_queue(struct sock
*sk
)
3778 struct tcp_sock
*tp
= tcp_sk(sk
);
3779 __u32 dsack_high
= tp
->rcv_nxt
;
3780 struct sk_buff
*skb
;
3782 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3783 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3786 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3787 __u32 dsack
= dsack_high
;
3788 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3789 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3790 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3793 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3794 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3795 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3799 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3800 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3801 TCP_SKB_CB(skb
)->end_seq
);
3803 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3804 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3805 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3806 if (tcp_hdr(skb
)->fin
)
3807 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3811 static int tcp_prune_queue(struct sock
*sk
);
3813 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3815 struct tcphdr
*th
= tcp_hdr(skb
);
3816 struct tcp_sock
*tp
= tcp_sk(sk
);
3819 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3822 __skb_pull(skb
, th
->doff
*4);
3824 TCP_ECN_accept_cwr(tp
, skb
);
3826 if (tp
->rx_opt
.dsack
) {
3827 tp
->rx_opt
.dsack
= 0;
3828 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3829 4 - tp
->rx_opt
.tstamp_ok
);
3832 /* Queue data for delivery to the user.
3833 * Packets in sequence go to the receive queue.
3834 * Out of sequence packets to the out_of_order_queue.
3836 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3837 if (tcp_receive_window(tp
) == 0)
3840 /* Ok. In sequence. In window. */
3841 if (tp
->ucopy
.task
== current
&&
3842 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3843 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3844 int chunk
= min_t(unsigned int, skb
->len
,
3847 __set_current_state(TASK_RUNNING
);
3850 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3851 tp
->ucopy
.len
-= chunk
;
3852 tp
->copied_seq
+= chunk
;
3853 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3854 tcp_rcv_space_adjust(sk
);
3862 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3863 !sk_stream_rmem_schedule(sk
, skb
))) {
3864 if (tcp_prune_queue(sk
) < 0 ||
3865 !sk_stream_rmem_schedule(sk
, skb
))
3868 sk_stream_set_owner_r(skb
, sk
);
3869 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3871 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3873 tcp_event_data_recv(sk
, skb
);
3875 tcp_fin(skb
, sk
, th
);
3877 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3880 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3881 * gap in queue is filled.
3883 if (skb_queue_empty(&tp
->out_of_order_queue
))
3884 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3887 if (tp
->rx_opt
.num_sacks
)
3888 tcp_sack_remove(tp
);
3890 tcp_fast_path_check(sk
);
3894 else if (!sock_flag(sk
, SOCK_DEAD
))
3895 sk
->sk_data_ready(sk
, 0);
3899 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3900 /* A retransmit, 2nd most common case. Force an immediate ack. */
3901 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3902 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3905 tcp_enter_quickack_mode(sk
);
3906 inet_csk_schedule_ack(sk
);
3912 /* Out of window. F.e. zero window probe. */
3913 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3916 tcp_enter_quickack_mode(sk
);
3918 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3919 /* Partial packet, seq < rcv_next < end_seq */
3920 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3921 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3922 TCP_SKB_CB(skb
)->end_seq
);
3924 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3926 /* If window is closed, drop tail of packet. But after
3927 * remembering D-SACK for its head made in previous line.
3929 if (!tcp_receive_window(tp
))
3934 TCP_ECN_check_ce(tp
, skb
);
3936 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3937 !sk_stream_rmem_schedule(sk
, skb
)) {
3938 if (tcp_prune_queue(sk
) < 0 ||
3939 !sk_stream_rmem_schedule(sk
, skb
))
3943 /* Disable header prediction. */
3945 inet_csk_schedule_ack(sk
);
3947 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3948 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3950 sk_stream_set_owner_r(skb
, sk
);
3952 if (!skb_peek(&tp
->out_of_order_queue
)) {
3953 /* Initial out of order segment, build 1 SACK. */
3954 if (tcp_is_sack(tp
)) {
3955 tp
->rx_opt
.num_sacks
= 1;
3956 tp
->rx_opt
.dsack
= 0;
3957 tp
->rx_opt
.eff_sacks
= 1;
3958 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3959 tp
->selective_acks
[0].end_seq
=
3960 TCP_SKB_CB(skb
)->end_seq
;
3962 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3964 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3965 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3966 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3968 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3969 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3971 if (!tp
->rx_opt
.num_sacks
||
3972 tp
->selective_acks
[0].end_seq
!= seq
)
3975 /* Common case: data arrive in order after hole. */
3976 tp
->selective_acks
[0].end_seq
= end_seq
;
3980 /* Find place to insert this segment. */
3982 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3984 } while ((skb1
= skb1
->prev
) !=
3985 (struct sk_buff
*)&tp
->out_of_order_queue
);
3987 /* Do skb overlap to previous one? */
3988 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3989 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3990 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
3991 /* All the bits are present. Drop. */
3993 tcp_dsack_set(tp
, seq
, end_seq
);
3996 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
3997 /* Partial overlap. */
3998 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4003 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
4005 /* And clean segments covered by new one as whole. */
4006 while ((skb1
= skb
->next
) !=
4007 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4008 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4009 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4010 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
4013 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4014 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
4019 if (tcp_is_sack(tp
))
4020 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4024 /* Collapse contiguous sequence of skbs head..tail with
4025 * sequence numbers start..end.
4026 * Segments with FIN/SYN are not collapsed (only because this
4030 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4031 struct sk_buff
*head
, struct sk_buff
*tail
,
4034 struct sk_buff
*skb
;
4036 /* First, check that queue is collapsible and find
4037 * the point where collapsing can be useful. */
4038 for (skb
= head
; skb
!= tail
; ) {
4039 /* No new bits? It is possible on ofo queue. */
4040 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4041 struct sk_buff
*next
= skb
->next
;
4042 __skb_unlink(skb
, list
);
4044 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
4049 /* The first skb to collapse is:
4051 * - bloated or contains data before "start" or
4052 * overlaps to the next one.
4054 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4055 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4056 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4057 (skb
->next
!= tail
&&
4058 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4061 /* Decided to skip this, advance start seq. */
4062 start
= TCP_SKB_CB(skb
)->end_seq
;
4065 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4068 while (before(start
, end
)) {
4069 struct sk_buff
*nskb
;
4070 unsigned int header
= skb_headroom(skb
);
4071 int copy
= SKB_MAX_ORDER(header
, 0);
4073 /* Too big header? This can happen with IPv6. */
4076 if (end
-start
< copy
)
4078 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
4082 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4083 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4085 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4087 skb_reserve(nskb
, header
);
4088 memcpy(nskb
->head
, skb
->head
, header
);
4089 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4090 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4091 __skb_insert(nskb
, skb
->prev
, skb
, list
);
4092 sk_stream_set_owner_r(nskb
, sk
);
4094 /* Copy data, releasing collapsed skbs. */
4096 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4097 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4101 size
= min(copy
, size
);
4102 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4104 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4108 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4109 struct sk_buff
*next
= skb
->next
;
4110 __skb_unlink(skb
, list
);
4112 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
4115 tcp_hdr(skb
)->syn
||
4123 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4124 * and tcp_collapse() them until all the queue is collapsed.
4126 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4128 struct tcp_sock
*tp
= tcp_sk(sk
);
4129 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4130 struct sk_buff
*head
;
4136 start
= TCP_SKB_CB(skb
)->seq
;
4137 end
= TCP_SKB_CB(skb
)->end_seq
;
4143 /* Segment is terminated when we see gap or when
4144 * we are at the end of all the queue. */
4145 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4146 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4147 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4148 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4149 head
, skb
, start
, end
);
4151 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4153 /* Start new segment */
4154 start
= TCP_SKB_CB(skb
)->seq
;
4155 end
= TCP_SKB_CB(skb
)->end_seq
;
4157 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4158 start
= TCP_SKB_CB(skb
)->seq
;
4159 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4160 end
= TCP_SKB_CB(skb
)->end_seq
;
4165 /* Reduce allocated memory if we can, trying to get
4166 * the socket within its memory limits again.
4168 * Return less than zero if we should start dropping frames
4169 * until the socket owning process reads some of the data
4170 * to stabilize the situation.
4172 static int tcp_prune_queue(struct sock
*sk
)
4174 struct tcp_sock
*tp
= tcp_sk(sk
);
4176 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4178 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
4180 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4181 tcp_clamp_window(sk
);
4182 else if (tcp_memory_pressure
)
4183 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4185 tcp_collapse_ofo_queue(sk
);
4186 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4187 sk
->sk_receive_queue
.next
,
4188 (struct sk_buff
*)&sk
->sk_receive_queue
,
4189 tp
->copied_seq
, tp
->rcv_nxt
);
4190 sk_stream_mem_reclaim(sk
);
4192 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4195 /* Collapsing did not help, destructive actions follow.
4196 * This must not ever occur. */
4198 /* First, purge the out_of_order queue. */
4199 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4200 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
4201 __skb_queue_purge(&tp
->out_of_order_queue
);
4203 /* Reset SACK state. A conforming SACK implementation will
4204 * do the same at a timeout based retransmit. When a connection
4205 * is in a sad state like this, we care only about integrity
4206 * of the connection not performance.
4208 if (tcp_is_sack(tp
))
4209 tcp_sack_reset(&tp
->rx_opt
);
4210 sk_stream_mem_reclaim(sk
);
4213 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4216 /* If we are really being abused, tell the caller to silently
4217 * drop receive data on the floor. It will get retransmitted
4218 * and hopefully then we'll have sufficient space.
4220 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
4222 /* Massive buffer overcommit. */
4228 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4229 * As additional protections, we do not touch cwnd in retransmission phases,
4230 * and if application hit its sndbuf limit recently.
4232 void tcp_cwnd_application_limited(struct sock
*sk
)
4234 struct tcp_sock
*tp
= tcp_sk(sk
);
4236 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4237 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4238 /* Limited by application or receiver window. */
4239 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4240 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4241 if (win_used
< tp
->snd_cwnd
) {
4242 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4243 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4245 tp
->snd_cwnd_used
= 0;
4247 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4250 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4252 struct tcp_sock
*tp
= tcp_sk(sk
);
4254 /* If the user specified a specific send buffer setting, do
4257 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4260 /* If we are under global TCP memory pressure, do not expand. */
4261 if (tcp_memory_pressure
)
4264 /* If we are under soft global TCP memory pressure, do not expand. */
4265 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4268 /* If we filled the congestion window, do not expand. */
4269 if (tp
->packets_out
>= tp
->snd_cwnd
)
4275 /* When incoming ACK allowed to free some skb from write_queue,
4276 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4277 * on the exit from tcp input handler.
4279 * PROBLEM: sndbuf expansion does not work well with largesend.
4281 static void tcp_new_space(struct sock
*sk
)
4283 struct tcp_sock
*tp
= tcp_sk(sk
);
4285 if (tcp_should_expand_sndbuf(sk
)) {
4286 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4287 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4288 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4289 tp
->reordering
+ 1);
4290 sndmem
*= 2*demanded
;
4291 if (sndmem
> sk
->sk_sndbuf
)
4292 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4293 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4296 sk
->sk_write_space(sk
);
4299 static void tcp_check_space(struct sock
*sk
)
4301 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4302 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4303 if (sk
->sk_socket
&&
4304 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4309 static inline void tcp_data_snd_check(struct sock
*sk
)
4311 tcp_push_pending_frames(sk
);
4312 tcp_check_space(sk
);
4316 * Check if sending an ack is needed.
4318 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4320 struct tcp_sock
*tp
= tcp_sk(sk
);
4322 /* More than one full frame received... */
4323 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4324 /* ... and right edge of window advances far enough.
4325 * (tcp_recvmsg() will send ACK otherwise). Or...
4327 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4328 /* We ACK each frame or... */
4329 tcp_in_quickack_mode(sk
) ||
4330 /* We have out of order data. */
4332 skb_peek(&tp
->out_of_order_queue
))) {
4333 /* Then ack it now */
4336 /* Else, send delayed ack. */
4337 tcp_send_delayed_ack(sk
);
4341 static inline void tcp_ack_snd_check(struct sock
*sk
)
4343 if (!inet_csk_ack_scheduled(sk
)) {
4344 /* We sent a data segment already. */
4347 __tcp_ack_snd_check(sk
, 1);
4351 * This routine is only called when we have urgent data
4352 * signaled. Its the 'slow' part of tcp_urg. It could be
4353 * moved inline now as tcp_urg is only called from one
4354 * place. We handle URGent data wrong. We have to - as
4355 * BSD still doesn't use the correction from RFC961.
4356 * For 1003.1g we should support a new option TCP_STDURG to permit
4357 * either form (or just set the sysctl tcp_stdurg).
4360 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4362 struct tcp_sock
*tp
= tcp_sk(sk
);
4363 u32 ptr
= ntohs(th
->urg_ptr
);
4365 if (ptr
&& !sysctl_tcp_stdurg
)
4367 ptr
+= ntohl(th
->seq
);
4369 /* Ignore urgent data that we've already seen and read. */
4370 if (after(tp
->copied_seq
, ptr
))
4373 /* Do not replay urg ptr.
4375 * NOTE: interesting situation not covered by specs.
4376 * Misbehaving sender may send urg ptr, pointing to segment,
4377 * which we already have in ofo queue. We are not able to fetch
4378 * such data and will stay in TCP_URG_NOTYET until will be eaten
4379 * by recvmsg(). Seems, we are not obliged to handle such wicked
4380 * situations. But it is worth to think about possibility of some
4381 * DoSes using some hypothetical application level deadlock.
4383 if (before(ptr
, tp
->rcv_nxt
))
4386 /* Do we already have a newer (or duplicate) urgent pointer? */
4387 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4390 /* Tell the world about our new urgent pointer. */
4393 /* We may be adding urgent data when the last byte read was
4394 * urgent. To do this requires some care. We cannot just ignore
4395 * tp->copied_seq since we would read the last urgent byte again
4396 * as data, nor can we alter copied_seq until this data arrives
4397 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4399 * NOTE. Double Dutch. Rendering to plain English: author of comment
4400 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4401 * and expect that both A and B disappear from stream. This is _wrong_.
4402 * Though this happens in BSD with high probability, this is occasional.
4403 * Any application relying on this is buggy. Note also, that fix "works"
4404 * only in this artificial test. Insert some normal data between A and B and we will
4405 * decline of BSD again. Verdict: it is better to remove to trap
4408 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4409 !sock_flag(sk
, SOCK_URGINLINE
) &&
4410 tp
->copied_seq
!= tp
->rcv_nxt
) {
4411 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4413 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4414 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4419 tp
->urg_data
= TCP_URG_NOTYET
;
4422 /* Disable header prediction. */
4426 /* This is the 'fast' part of urgent handling. */
4427 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4429 struct tcp_sock
*tp
= tcp_sk(sk
);
4431 /* Check if we get a new urgent pointer - normally not. */
4433 tcp_check_urg(sk
,th
);
4435 /* Do we wait for any urgent data? - normally not... */
4436 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4437 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4440 /* Is the urgent pointer pointing into this packet? */
4441 if (ptr
< skb
->len
) {
4443 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4445 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4446 if (!sock_flag(sk
, SOCK_DEAD
))
4447 sk
->sk_data_ready(sk
, 0);
4452 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4454 struct tcp_sock
*tp
= tcp_sk(sk
);
4455 int chunk
= skb
->len
- hlen
;
4459 if (skb_csum_unnecessary(skb
))
4460 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4462 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4466 tp
->ucopy
.len
-= chunk
;
4467 tp
->copied_seq
+= chunk
;
4468 tcp_rcv_space_adjust(sk
);
4475 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4479 if (sock_owned_by_user(sk
)) {
4481 result
= __tcp_checksum_complete(skb
);
4484 result
= __tcp_checksum_complete(skb
);
4489 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4491 return !skb_csum_unnecessary(skb
) &&
4492 __tcp_checksum_complete_user(sk
, skb
);
4495 #ifdef CONFIG_NET_DMA
4496 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4498 struct tcp_sock
*tp
= tcp_sk(sk
);
4499 int chunk
= skb
->len
- hlen
;
4501 int copied_early
= 0;
4503 if (tp
->ucopy
.wakeup
)
4506 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4507 tp
->ucopy
.dma_chan
= get_softnet_dma();
4509 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4511 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4512 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4517 tp
->ucopy
.dma_cookie
= dma_cookie
;
4520 tp
->ucopy
.len
-= chunk
;
4521 tp
->copied_seq
+= chunk
;
4522 tcp_rcv_space_adjust(sk
);
4524 if ((tp
->ucopy
.len
== 0) ||
4525 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4526 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4527 tp
->ucopy
.wakeup
= 1;
4528 sk
->sk_data_ready(sk
, 0);
4530 } else if (chunk
> 0) {
4531 tp
->ucopy
.wakeup
= 1;
4532 sk
->sk_data_ready(sk
, 0);
4535 return copied_early
;
4537 #endif /* CONFIG_NET_DMA */
4540 * TCP receive function for the ESTABLISHED state.
4542 * It is split into a fast path and a slow path. The fast path is
4544 * - A zero window was announced from us - zero window probing
4545 * is only handled properly in the slow path.
4546 * - Out of order segments arrived.
4547 * - Urgent data is expected.
4548 * - There is no buffer space left
4549 * - Unexpected TCP flags/window values/header lengths are received
4550 * (detected by checking the TCP header against pred_flags)
4551 * - Data is sent in both directions. Fast path only supports pure senders
4552 * or pure receivers (this means either the sequence number or the ack
4553 * value must stay constant)
4554 * - Unexpected TCP option.
4556 * When these conditions are not satisfied it drops into a standard
4557 * receive procedure patterned after RFC793 to handle all cases.
4558 * The first three cases are guaranteed by proper pred_flags setting,
4559 * the rest is checked inline. Fast processing is turned on in
4560 * tcp_data_queue when everything is OK.
4562 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4563 struct tcphdr
*th
, unsigned len
)
4565 struct tcp_sock
*tp
= tcp_sk(sk
);
4568 * Header prediction.
4569 * The code loosely follows the one in the famous
4570 * "30 instruction TCP receive" Van Jacobson mail.
4572 * Van's trick is to deposit buffers into socket queue
4573 * on a device interrupt, to call tcp_recv function
4574 * on the receive process context and checksum and copy
4575 * the buffer to user space. smart...
4577 * Our current scheme is not silly either but we take the
4578 * extra cost of the net_bh soft interrupt processing...
4579 * We do checksum and copy also but from device to kernel.
4582 tp
->rx_opt
.saw_tstamp
= 0;
4584 /* pred_flags is 0xS?10 << 16 + snd_wnd
4585 * if header_prediction is to be made
4586 * 'S' will always be tp->tcp_header_len >> 2
4587 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4588 * turn it off (when there are holes in the receive
4589 * space for instance)
4590 * PSH flag is ignored.
4593 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4594 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4595 int tcp_header_len
= tp
->tcp_header_len
;
4597 /* Timestamp header prediction: tcp_header_len
4598 * is automatically equal to th->doff*4 due to pred_flags
4602 /* Check timestamp */
4603 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4604 __be32
*ptr
= (__be32
*)(th
+ 1);
4606 /* No? Slow path! */
4607 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4608 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4611 tp
->rx_opt
.saw_tstamp
= 1;
4613 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4615 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4617 /* If PAWS failed, check it more carefully in slow path */
4618 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4621 /* DO NOT update ts_recent here, if checksum fails
4622 * and timestamp was corrupted part, it will result
4623 * in a hung connection since we will drop all
4624 * future packets due to the PAWS test.
4628 if (len
<= tcp_header_len
) {
4629 /* Bulk data transfer: sender */
4630 if (len
== tcp_header_len
) {
4631 /* Predicted packet is in window by definition.
4632 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4633 * Hence, check seq<=rcv_wup reduces to:
4635 if (tcp_header_len
==
4636 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4637 tp
->rcv_nxt
== tp
->rcv_wup
)
4638 tcp_store_ts_recent(tp
);
4640 /* We know that such packets are checksummed
4643 tcp_ack(sk
, skb
, 0);
4645 tcp_data_snd_check(sk
);
4647 } else { /* Header too small */
4648 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4653 int copied_early
= 0;
4655 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4656 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4657 #ifdef CONFIG_NET_DMA
4658 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4663 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4664 __set_current_state(TASK_RUNNING
);
4666 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4670 /* Predicted packet is in window by definition.
4671 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4672 * Hence, check seq<=rcv_wup reduces to:
4674 if (tcp_header_len
==
4675 (sizeof(struct tcphdr
) +
4676 TCPOLEN_TSTAMP_ALIGNED
) &&
4677 tp
->rcv_nxt
== tp
->rcv_wup
)
4678 tcp_store_ts_recent(tp
);
4680 tcp_rcv_rtt_measure_ts(sk
, skb
);
4682 __skb_pull(skb
, tcp_header_len
);
4683 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4684 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4687 tcp_cleanup_rbuf(sk
, skb
->len
);
4690 if (tcp_checksum_complete_user(sk
, skb
))
4693 /* Predicted packet is in window by definition.
4694 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4695 * Hence, check seq<=rcv_wup reduces to:
4697 if (tcp_header_len
==
4698 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4699 tp
->rcv_nxt
== tp
->rcv_wup
)
4700 tcp_store_ts_recent(tp
);
4702 tcp_rcv_rtt_measure_ts(sk
, skb
);
4704 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4707 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4709 /* Bulk data transfer: receiver */
4710 __skb_pull(skb
,tcp_header_len
);
4711 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4712 sk_stream_set_owner_r(skb
, sk
);
4713 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4716 tcp_event_data_recv(sk
, skb
);
4718 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4719 /* Well, only one small jumplet in fast path... */
4720 tcp_ack(sk
, skb
, FLAG_DATA
);
4721 tcp_data_snd_check(sk
);
4722 if (!inet_csk_ack_scheduled(sk
))
4726 __tcp_ack_snd_check(sk
, 0);
4728 #ifdef CONFIG_NET_DMA
4730 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4736 sk
->sk_data_ready(sk
, 0);
4742 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4746 * RFC1323: H1. Apply PAWS check first.
4748 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4749 tcp_paws_discard(sk
, skb
)) {
4751 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4752 tcp_send_dupack(sk
, skb
);
4755 /* Resets are accepted even if PAWS failed.
4757 ts_recent update must be made after we are sure
4758 that the packet is in window.
4763 * Standard slow path.
4766 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4767 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4768 * (RST) segments are validated by checking their SEQ-fields."
4769 * And page 69: "If an incoming segment is not acceptable,
4770 * an acknowledgment should be sent in reply (unless the RST bit
4771 * is set, if so drop the segment and return)".
4774 tcp_send_dupack(sk
, skb
);
4783 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4785 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4786 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4787 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4794 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4796 tcp_rcv_rtt_measure_ts(sk
, skb
);
4798 /* Process urgent data. */
4799 tcp_urg(sk
, skb
, th
);
4801 /* step 7: process the segment text */
4802 tcp_data_queue(sk
, skb
);
4804 tcp_data_snd_check(sk
);
4805 tcp_ack_snd_check(sk
);
4809 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4816 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4817 struct tcphdr
*th
, unsigned len
)
4819 struct tcp_sock
*tp
= tcp_sk(sk
);
4820 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4821 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4823 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4827 * "If the state is SYN-SENT then
4828 * first check the ACK bit
4829 * If the ACK bit is set
4830 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4831 * a reset (unless the RST bit is set, if so drop
4832 * the segment and return)"
4834 * We do not send data with SYN, so that RFC-correct
4837 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4838 goto reset_and_undo
;
4840 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4841 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4843 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4844 goto reset_and_undo
;
4847 /* Now ACK is acceptable.
4849 * "If the RST bit is set
4850 * If the ACK was acceptable then signal the user "error:
4851 * connection reset", drop the segment, enter CLOSED state,
4852 * delete TCB, and return."
4861 * "fifth, if neither of the SYN or RST bits is set then
4862 * drop the segment and return."
4868 goto discard_and_undo
;
4871 * "If the SYN bit is on ...
4872 * are acceptable then ...
4873 * (our SYN has been ACKed), change the connection
4874 * state to ESTABLISHED..."
4877 TCP_ECN_rcv_synack(tp
, th
);
4879 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4880 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4882 /* Ok.. it's good. Set up sequence numbers and
4883 * move to established.
4885 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4886 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4888 /* RFC1323: The window in SYN & SYN/ACK segments is
4891 tp
->snd_wnd
= ntohs(th
->window
);
4892 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4894 if (!tp
->rx_opt
.wscale_ok
) {
4895 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4896 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4899 if (tp
->rx_opt
.saw_tstamp
) {
4900 tp
->rx_opt
.tstamp_ok
= 1;
4901 tp
->tcp_header_len
=
4902 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4903 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4904 tcp_store_ts_recent(tp
);
4906 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4909 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
4910 tcp_enable_fack(tp
);
4913 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4914 tcp_initialize_rcv_mss(sk
);
4916 /* Remember, tcp_poll() does not lock socket!
4917 * Change state from SYN-SENT only after copied_seq
4918 * is initialized. */
4919 tp
->copied_seq
= tp
->rcv_nxt
;
4921 tcp_set_state(sk
, TCP_ESTABLISHED
);
4923 security_inet_conn_established(sk
, skb
);
4925 /* Make sure socket is routed, for correct metrics. */
4926 icsk
->icsk_af_ops
->rebuild_header(sk
);
4928 tcp_init_metrics(sk
);
4930 tcp_init_congestion_control(sk
);
4932 /* Prevent spurious tcp_cwnd_restart() on first data
4935 tp
->lsndtime
= tcp_time_stamp
;
4937 tcp_init_buffer_space(sk
);
4939 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4940 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4942 if (!tp
->rx_opt
.snd_wscale
)
4943 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4947 if (!sock_flag(sk
, SOCK_DEAD
)) {
4948 sk
->sk_state_change(sk
);
4949 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
4952 if (sk
->sk_write_pending
||
4953 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4954 icsk
->icsk_ack
.pingpong
) {
4955 /* Save one ACK. Data will be ready after
4956 * several ticks, if write_pending is set.
4958 * It may be deleted, but with this feature tcpdumps
4959 * look so _wonderfully_ clever, that I was not able
4960 * to stand against the temptation 8) --ANK
4962 inet_csk_schedule_ack(sk
);
4963 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4964 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4965 tcp_incr_quickack(sk
);
4966 tcp_enter_quickack_mode(sk
);
4967 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4968 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4979 /* No ACK in the segment */
4983 * "If the RST bit is set
4985 * Otherwise (no ACK) drop the segment and return."
4988 goto discard_and_undo
;
4992 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
4993 goto discard_and_undo
;
4996 /* We see SYN without ACK. It is attempt of
4997 * simultaneous connect with crossed SYNs.
4998 * Particularly, it can be connect to self.
5000 tcp_set_state(sk
, TCP_SYN_RECV
);
5002 if (tp
->rx_opt
.saw_tstamp
) {
5003 tp
->rx_opt
.tstamp_ok
= 1;
5004 tcp_store_ts_recent(tp
);
5005 tp
->tcp_header_len
=
5006 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5008 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5011 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5012 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5014 /* RFC1323: The window in SYN & SYN/ACK segments is
5017 tp
->snd_wnd
= ntohs(th
->window
);
5018 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5019 tp
->max_window
= tp
->snd_wnd
;
5021 TCP_ECN_rcv_syn(tp
, th
);
5024 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5025 tcp_initialize_rcv_mss(sk
);
5028 tcp_send_synack(sk
);
5030 /* Note, we could accept data and URG from this segment.
5031 * There are no obstacles to make this.
5033 * However, if we ignore data in ACKless segments sometimes,
5034 * we have no reasons to accept it sometimes.
5035 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5036 * is not flawless. So, discard packet for sanity.
5037 * Uncomment this return to process the data.
5044 /* "fifth, if neither of the SYN or RST bits is set then
5045 * drop the segment and return."
5049 tcp_clear_options(&tp
->rx_opt
);
5050 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5054 tcp_clear_options(&tp
->rx_opt
);
5055 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5061 * This function implements the receiving procedure of RFC 793 for
5062 * all states except ESTABLISHED and TIME_WAIT.
5063 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5064 * address independent.
5067 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5068 struct tcphdr
*th
, unsigned len
)
5070 struct tcp_sock
*tp
= tcp_sk(sk
);
5071 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5074 tp
->rx_opt
.saw_tstamp
= 0;
5076 switch (sk
->sk_state
) {
5088 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5091 /* Now we have several options: In theory there is
5092 * nothing else in the frame. KA9Q has an option to
5093 * send data with the syn, BSD accepts data with the
5094 * syn up to the [to be] advertised window and
5095 * Solaris 2.1 gives you a protocol error. For now
5096 * we just ignore it, that fits the spec precisely
5097 * and avoids incompatibilities. It would be nice in
5098 * future to drop through and process the data.
5100 * Now that TTCP is starting to be used we ought to
5102 * But, this leaves one open to an easy denial of
5103 * service attack, and SYN cookies can't defend
5104 * against this problem. So, we drop the data
5105 * in the interest of security over speed unless
5106 * it's still in use.
5114 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5118 /* Do step6 onward by hand. */
5119 tcp_urg(sk
, skb
, th
);
5121 tcp_data_snd_check(sk
);
5125 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5126 tcp_paws_discard(sk
, skb
)) {
5128 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
5129 tcp_send_dupack(sk
, skb
);
5132 /* Reset is accepted even if it did not pass PAWS. */
5135 /* step 1: check sequence number */
5136 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5138 tcp_send_dupack(sk
, skb
);
5142 /* step 2: check RST bit */
5148 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5150 /* step 3: check security and precedence [ignored] */
5154 * Check for a SYN in window.
5156 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5157 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
5162 /* step 5: check the ACK field */
5164 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5166 switch (sk
->sk_state
) {
5169 tp
->copied_seq
= tp
->rcv_nxt
;
5171 tcp_set_state(sk
, TCP_ESTABLISHED
);
5172 sk
->sk_state_change(sk
);
5174 /* Note, that this wakeup is only for marginal
5175 * crossed SYN case. Passively open sockets
5176 * are not waked up, because sk->sk_sleep ==
5177 * NULL and sk->sk_socket == NULL.
5181 SOCK_WAKE_IO
, POLL_OUT
);
5183 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5184 tp
->snd_wnd
= ntohs(th
->window
) <<
5185 tp
->rx_opt
.snd_wscale
;
5186 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5187 TCP_SKB_CB(skb
)->seq
);
5189 /* tcp_ack considers this ACK as duplicate
5190 * and does not calculate rtt.
5191 * Fix it at least with timestamps.
5193 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5195 tcp_ack_saw_tstamp(sk
, 0);
5197 if (tp
->rx_opt
.tstamp_ok
)
5198 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5200 /* Make sure socket is routed, for
5203 icsk
->icsk_af_ops
->rebuild_header(sk
);
5205 tcp_init_metrics(sk
);
5207 tcp_init_congestion_control(sk
);
5209 /* Prevent spurious tcp_cwnd_restart() on
5210 * first data packet.
5212 tp
->lsndtime
= tcp_time_stamp
;
5215 tcp_initialize_rcv_mss(sk
);
5216 tcp_init_buffer_space(sk
);
5217 tcp_fast_path_on(tp
);
5224 if (tp
->snd_una
== tp
->write_seq
) {
5225 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5226 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5227 dst_confirm(sk
->sk_dst_cache
);
5229 if (!sock_flag(sk
, SOCK_DEAD
))
5230 /* Wake up lingering close() */
5231 sk
->sk_state_change(sk
);
5235 if (tp
->linger2
< 0 ||
5236 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5237 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5239 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5243 tmo
= tcp_fin_time(sk
);
5244 if (tmo
> TCP_TIMEWAIT_LEN
) {
5245 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5246 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5247 /* Bad case. We could lose such FIN otherwise.
5248 * It is not a big problem, but it looks confusing
5249 * and not so rare event. We still can lose it now,
5250 * if it spins in bh_lock_sock(), but it is really
5253 inet_csk_reset_keepalive_timer(sk
, tmo
);
5255 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5263 if (tp
->snd_una
== tp
->write_seq
) {
5264 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5270 if (tp
->snd_una
== tp
->write_seq
) {
5271 tcp_update_metrics(sk
);
5280 /* step 6: check the URG bit */
5281 tcp_urg(sk
, skb
, th
);
5283 /* step 7: process the segment text */
5284 switch (sk
->sk_state
) {
5285 case TCP_CLOSE_WAIT
:
5288 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5292 /* RFC 793 says to queue data in these states,
5293 * RFC 1122 says we MUST send a reset.
5294 * BSD 4.4 also does reset.
5296 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5297 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5298 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5299 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5305 case TCP_ESTABLISHED
:
5306 tcp_data_queue(sk
, skb
);
5311 /* tcp_data could move socket to TIME-WAIT */
5312 if (sk
->sk_state
!= TCP_CLOSE
) {
5313 tcp_data_snd_check(sk
);
5314 tcp_ack_snd_check(sk
);
5324 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5325 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5326 EXPORT_SYMBOL(tcp_parse_options
);
5327 EXPORT_SYMBOL(tcp_rcv_established
);
5328 EXPORT_SYMBOL(tcp_rcv_state_process
);
5329 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);