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 int 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
;
1128 u32 new_low_seq
= tp
->snd_nxt
;
1129 u32 received_upto
= TCP_SKB_CB(tp
->highest_sack
)->end_seq
;
1131 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1132 !after(received_upto
, tp
->lost_retrans_low
) ||
1133 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1136 tcp_for_write_queue(skb
, sk
) {
1137 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1139 if (skb
== tcp_send_head(sk
))
1141 if (cnt
== tp
->retrans_out
)
1143 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1146 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1149 if (after(received_upto
, ack_seq
) &&
1151 !before(received_upto
,
1152 ack_seq
+ tp
->reordering
* tp
->mss_cache
))) {
1153 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1154 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1156 /* clear lost hint */
1157 tp
->retransmit_skb_hint
= NULL
;
1159 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1160 tp
->lost_out
+= tcp_skb_pcount(skb
);
1161 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1162 flag
|= FLAG_DATA_SACKED
;
1163 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT
);
1166 if (before(ack_seq
, new_low_seq
))
1167 new_low_seq
= ack_seq
;
1168 cnt
+= tcp_skb_pcount(skb
);
1172 if (tp
->retrans_out
)
1173 tp
->lost_retrans_low
= new_low_seq
;
1178 static int tcp_check_dsack(struct tcp_sock
*tp
, struct sk_buff
*ack_skb
,
1179 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1182 u32 start_seq_0
= ntohl(get_unaligned(&sp
[0].start_seq
));
1183 u32 end_seq_0
= ntohl(get_unaligned(&sp
[0].end_seq
));
1186 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1189 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV
);
1190 } else if (num_sacks
> 1) {
1191 u32 end_seq_1
= ntohl(get_unaligned(&sp
[1].end_seq
));
1192 u32 start_seq_1
= ntohl(get_unaligned(&sp
[1].start_seq
));
1194 if (!after(end_seq_0
, end_seq_1
) &&
1195 !before(start_seq_0
, start_seq_1
)) {
1198 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV
);
1202 /* D-SACK for already forgotten data... Do dumb counting. */
1204 !after(end_seq_0
, prior_snd_una
) &&
1205 after(end_seq_0
, tp
->undo_marker
))
1211 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1212 * the incoming SACK may not exactly match but we can find smaller MSS
1213 * aligned portion of it that matches. Therefore we might need to fragment
1214 * which may fail and creates some hassle (caller must handle error case
1217 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1218 u32 start_seq
, u32 end_seq
)
1221 unsigned int pkt_len
;
1223 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1224 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1226 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1227 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1229 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1232 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1234 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1235 err
= tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
);
1243 static int tcp_sacktag_one(struct sk_buff
*skb
, struct tcp_sock
*tp
,
1244 int *reord
, int dup_sack
, int fack_count
)
1246 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1249 /* Account D-SACK for retransmitted packet. */
1250 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1251 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1253 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
) &&
1254 (sacked
& TCPCB_SACKED_ACKED
))
1255 *reord
= min(fack_count
, *reord
);
1258 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1259 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1262 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1263 if (sacked
& TCPCB_SACKED_RETRANS
) {
1264 /* If the segment is not tagged as lost,
1265 * we do not clear RETRANS, believing
1266 * that retransmission is still in flight.
1268 if (sacked
& TCPCB_LOST
) {
1269 TCP_SKB_CB(skb
)->sacked
&=
1270 ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1271 tp
->lost_out
-= tcp_skb_pcount(skb
);
1272 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1274 /* clear lost hint */
1275 tp
->retransmit_skb_hint
= NULL
;
1278 if (!(sacked
& TCPCB_RETRANS
)) {
1279 /* New sack for not retransmitted frame,
1280 * which was in hole. It is reordering.
1282 if (before(TCP_SKB_CB(skb
)->seq
,
1283 tcp_highest_sack_seq(tp
)))
1284 *reord
= min(fack_count
, *reord
);
1286 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1287 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1288 flag
|= FLAG_ONLY_ORIG_SACKED
;
1291 if (sacked
& TCPCB_LOST
) {
1292 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1293 tp
->lost_out
-= tcp_skb_pcount(skb
);
1295 /* clear lost hint */
1296 tp
->retransmit_skb_hint
= NULL
;
1300 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1301 flag
|= FLAG_DATA_SACKED
;
1302 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1304 fack_count
+= tcp_skb_pcount(skb
);
1306 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1307 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1308 before(TCP_SKB_CB(skb
)->seq
,
1309 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1310 tp
->lost_cnt_hint
+= tcp_skb_pcount(skb
);
1312 if (fack_count
> tp
->fackets_out
)
1313 tp
->fackets_out
= fack_count
;
1315 if (after(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
1316 tp
->highest_sack
= skb
;
1319 if (dup_sack
&& (sacked
& TCPCB_RETRANS
))
1320 *reord
= min(fack_count
, *reord
);
1323 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1324 * frames and clear it. undo_retrans is decreased above, L|R frames
1325 * are accounted above as well.
1327 if (dup_sack
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)) {
1328 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1329 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1330 tp
->retransmit_skb_hint
= NULL
;
1336 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1337 struct tcp_sack_block
*next_dup
,
1338 u32 start_seq
, u32 end_seq
,
1339 int dup_sack_in
, int *fack_count
,
1340 int *reord
, int *flag
)
1342 struct tcp_sock
*tp
= tcp_sk(sk
);
1344 tcp_for_write_queue_from(skb
, sk
) {
1346 int dup_sack
= dup_sack_in
;
1348 if (skb
== tcp_send_head(sk
))
1351 /* queue is in-order => we can short-circuit the walk early */
1352 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1355 if ((next_dup
!= NULL
) &&
1356 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1357 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1358 next_dup
->start_seq
,
1365 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
, end_seq
);
1366 if (unlikely(in_sack
< 0))
1370 *flag
|= tcp_sacktag_one(skb
, tp
, reord
, dup_sack
, *fack_count
);
1372 *fack_count
+= tcp_skb_pcount(skb
);
1377 /* Avoid all extra work that is being done by sacktag while walking in
1380 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1383 tcp_for_write_queue_from(skb
, sk
) {
1384 if (skb
== tcp_send_head(sk
))
1387 if (before(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1393 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1395 struct tcp_sack_block
*next_dup
,
1397 int *fack_count
, int *reord
,
1400 if (next_dup
== NULL
)
1403 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1404 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
);
1405 tcp_sacktag_walk(skb
, sk
, NULL
,
1406 next_dup
->start_seq
, next_dup
->end_seq
,
1407 1, fack_count
, reord
, flag
);
1413 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1415 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1419 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
, u32 prior_snd_una
)
1421 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1422 struct tcp_sock
*tp
= tcp_sk(sk
);
1423 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1424 TCP_SKB_CB(ack_skb
)->sacked
);
1425 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1426 struct tcp_sack_block sp
[4];
1427 struct tcp_sack_block
*cache
;
1428 struct sk_buff
*skb
;
1429 int num_sacks
= (ptr
[1] - TCPOLEN_SACK_BASE
)>>3;
1431 int reord
= tp
->packets_out
;
1433 int found_dup_sack
= 0;
1436 int first_sack_index
;
1438 if (!tp
->sacked_out
) {
1439 if (WARN_ON(tp
->fackets_out
))
1440 tp
->fackets_out
= 0;
1441 tp
->highest_sack
= tcp_write_queue_head(sk
);
1444 found_dup_sack
= tcp_check_dsack(tp
, ack_skb
, sp_wire
,
1445 num_sacks
, prior_snd_una
);
1447 flag
|= FLAG_DSACKING_ACK
;
1449 /* Eliminate too old ACKs, but take into
1450 * account more or less fresh ones, they can
1451 * contain valid SACK info.
1453 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1456 if (!tp
->packets_out
)
1460 first_sack_index
= 0;
1461 for (i
= 0; i
< num_sacks
; i
++) {
1462 int dup_sack
= !i
&& found_dup_sack
;
1464 sp
[used_sacks
].start_seq
= ntohl(get_unaligned(&sp_wire
[i
].start_seq
));
1465 sp
[used_sacks
].end_seq
= ntohl(get_unaligned(&sp_wire
[i
].end_seq
));
1467 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1468 sp
[used_sacks
].start_seq
,
1469 sp
[used_sacks
].end_seq
)) {
1471 if (!tp
->undo_marker
)
1472 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO
);
1474 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD
);
1476 /* Don't count olds caused by ACK reordering */
1477 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1478 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1480 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD
);
1483 first_sack_index
= -1;
1487 /* Ignore very old stuff early */
1488 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1494 /* order SACK blocks to allow in order walk of the retrans queue */
1495 for (i
= used_sacks
- 1; i
> 0; i
--) {
1496 for (j
= 0; j
< i
; j
++){
1497 if (after(sp
[j
].start_seq
, sp
[j
+1].start_seq
)) {
1498 struct tcp_sack_block tmp
;
1504 /* Track where the first SACK block goes to */
1505 if (j
== first_sack_index
)
1506 first_sack_index
= j
+1;
1511 skb
= tcp_write_queue_head(sk
);
1515 if (!tp
->sacked_out
) {
1516 /* It's already past, so skip checking against it */
1517 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1519 cache
= tp
->recv_sack_cache
;
1520 /* Skip empty blocks in at head of the cache */
1521 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1526 while (i
< used_sacks
) {
1527 u32 start_seq
= sp
[i
].start_seq
;
1528 u32 end_seq
= sp
[i
].end_seq
;
1529 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1530 struct tcp_sack_block
*next_dup
= NULL
;
1532 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1533 next_dup
= &sp
[i
+ 1];
1535 /* Event "B" in the comment above. */
1536 if (after(end_seq
, tp
->high_seq
))
1537 flag
|= FLAG_DATA_LOST
;
1539 /* Skip too early cached blocks */
1540 while (tcp_sack_cache_ok(tp
, cache
) &&
1541 !before(start_seq
, cache
->end_seq
))
1544 /* Can skip some work by looking recv_sack_cache? */
1545 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1546 after(end_seq
, cache
->start_seq
)) {
1549 if (before(start_seq
, cache
->start_seq
)) {
1550 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1551 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
,
1552 cache
->start_seq
, dup_sack
,
1553 &fack_count
, &reord
, &flag
);
1556 /* Rest of the block already fully processed? */
1557 if (!after(end_seq
, cache
->end_seq
))
1560 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
, cache
->end_seq
,
1561 &fack_count
, &reord
, &flag
);
1563 /* ...tail remains todo... */
1564 if (TCP_SKB_CB(tp
->highest_sack
)->end_seq
== cache
->end_seq
) {
1565 /* ...but better entrypoint exists! */
1566 skb
= tcp_write_queue_next(sk
, tp
->highest_sack
);
1567 fack_count
= tp
->fackets_out
;
1572 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
);
1573 /* Check overlap against next cached too (past this one already) */
1578 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1579 skb
= tcp_write_queue_next(sk
, tp
->highest_sack
);
1580 fack_count
= tp
->fackets_out
;
1582 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
);
1585 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1586 dup_sack
, &fack_count
, &reord
, &flag
);
1589 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1590 * due to in-order walk
1592 if (after(end_seq
, tp
->frto_highmark
))
1593 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1598 /* Clear the head of the cache sack blocks so we can skip it next time */
1599 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1600 tp
->recv_sack_cache
[i
].start_seq
= 0;
1601 tp
->recv_sack_cache
[i
].end_seq
= 0;
1603 for (j
= 0; j
< used_sacks
; j
++)
1604 tp
->recv_sack_cache
[i
++] = sp
[j
];
1606 flag
|= tcp_mark_lost_retrans(sk
);
1608 tcp_verify_left_out(tp
);
1610 if ((reord
< tp
->fackets_out
) &&
1611 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1612 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1613 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1617 #if FASTRETRANS_DEBUG > 0
1618 BUG_TRAP((int)tp
->sacked_out
>= 0);
1619 BUG_TRAP((int)tp
->lost_out
>= 0);
1620 BUG_TRAP((int)tp
->retrans_out
>= 0);
1621 BUG_TRAP((int)tcp_packets_in_flight(tp
) >= 0);
1626 /* If we receive more dupacks than we expected counting segments
1627 * in assumption of absent reordering, interpret this as reordering.
1628 * The only another reason could be bug in receiver TCP.
1630 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1632 struct tcp_sock
*tp
= tcp_sk(sk
);
1635 holes
= max(tp
->lost_out
, 1U);
1636 holes
= min(holes
, tp
->packets_out
);
1638 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1639 tp
->sacked_out
= tp
->packets_out
- holes
;
1640 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1644 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1646 static void tcp_add_reno_sack(struct sock
*sk
)
1648 struct tcp_sock
*tp
= tcp_sk(sk
);
1650 tcp_check_reno_reordering(sk
, 0);
1651 tcp_verify_left_out(tp
);
1654 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1656 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1658 struct tcp_sock
*tp
= tcp_sk(sk
);
1661 /* One ACK acked hole. The rest eat duplicate ACKs. */
1662 if (acked
-1 >= tp
->sacked_out
)
1665 tp
->sacked_out
-= acked
-1;
1667 tcp_check_reno_reordering(sk
, acked
);
1668 tcp_verify_left_out(tp
);
1671 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1676 /* F-RTO can only be used if TCP has never retransmitted anything other than
1677 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1679 int tcp_use_frto(struct sock
*sk
)
1681 const struct tcp_sock
*tp
= tcp_sk(sk
);
1682 struct sk_buff
*skb
;
1684 if (!sysctl_tcp_frto
)
1690 /* Avoid expensive walking of rexmit queue if possible */
1691 if (tp
->retrans_out
> 1)
1694 skb
= tcp_write_queue_head(sk
);
1695 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1696 tcp_for_write_queue_from(skb
, sk
) {
1697 if (skb
== tcp_send_head(sk
))
1699 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1701 /* Short-circuit when first non-SACKed skb has been checked */
1702 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
))
1708 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1709 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1710 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1711 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1712 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1713 * bits are handled if the Loss state is really to be entered (in
1714 * tcp_enter_frto_loss).
1716 * Do like tcp_enter_loss() would; when RTO expires the second time it
1718 * "Reduce ssthresh if it has not yet been made inside this window."
1720 void tcp_enter_frto(struct sock
*sk
)
1722 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1723 struct tcp_sock
*tp
= tcp_sk(sk
);
1724 struct sk_buff
*skb
;
1726 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1727 tp
->snd_una
== tp
->high_seq
||
1728 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1729 !icsk
->icsk_retransmits
)) {
1730 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1731 /* Our state is too optimistic in ssthresh() call because cwnd
1732 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1733 * recovery has not yet completed. Pattern would be this: RTO,
1734 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1736 * RFC4138 should be more specific on what to do, even though
1737 * RTO is quite unlikely to occur after the first Cumulative ACK
1738 * due to back-off and complexity of triggering events ...
1740 if (tp
->frto_counter
) {
1742 stored_cwnd
= tp
->snd_cwnd
;
1744 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1745 tp
->snd_cwnd
= stored_cwnd
;
1747 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1749 /* ... in theory, cong.control module could do "any tricks" in
1750 * ssthresh(), which means that ca_state, lost bits and lost_out
1751 * counter would have to be faked before the call occurs. We
1752 * consider that too expensive, unlikely and hacky, so modules
1753 * using these in ssthresh() must deal these incompatibility
1754 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1756 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1759 tp
->undo_marker
= tp
->snd_una
;
1760 tp
->undo_retrans
= 0;
1762 skb
= tcp_write_queue_head(sk
);
1763 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1764 tp
->undo_marker
= 0;
1765 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1766 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1767 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1769 tcp_verify_left_out(tp
);
1771 /* Too bad if TCP was application limited */
1772 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1774 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1775 * The last condition is necessary at least in tp->frto_counter case.
1777 if (IsSackFrto() && (tp
->frto_counter
||
1778 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1779 after(tp
->high_seq
, tp
->snd_una
)) {
1780 tp
->frto_highmark
= tp
->high_seq
;
1782 tp
->frto_highmark
= tp
->snd_nxt
;
1784 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1785 tp
->high_seq
= tp
->snd_nxt
;
1786 tp
->frto_counter
= 1;
1789 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1790 * which indicates that we should follow the traditional RTO recovery,
1791 * i.e. mark everything lost and do go-back-N retransmission.
1793 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1795 struct tcp_sock
*tp
= tcp_sk(sk
);
1796 struct sk_buff
*skb
;
1799 tp
->retrans_out
= 0;
1800 if (tcp_is_reno(tp
))
1801 tcp_reset_reno_sack(tp
);
1803 tcp_for_write_queue(skb
, sk
) {
1804 if (skb
== tcp_send_head(sk
))
1807 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1809 * Count the retransmission made on RTO correctly (only when
1810 * waiting for the first ACK and did not get it)...
1812 if ((tp
->frto_counter
== 1) && !(flag
&FLAG_DATA_ACKED
)) {
1813 /* For some reason this R-bit might get cleared? */
1814 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1815 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1816 /* ...enter this if branch just for the first segment */
1817 flag
|= FLAG_DATA_ACKED
;
1819 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1820 tp
->undo_marker
= 0;
1821 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1824 /* Don't lost mark skbs that were fwd transmitted after RTO */
1825 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) &&
1826 !after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
)) {
1827 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1828 tp
->lost_out
+= tcp_skb_pcount(skb
);
1831 tcp_verify_left_out(tp
);
1833 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1834 tp
->snd_cwnd_cnt
= 0;
1835 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1836 tp
->frto_counter
= 0;
1837 tp
->bytes_acked
= 0;
1839 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1840 sysctl_tcp_reordering
);
1841 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1842 tp
->high_seq
= tp
->frto_highmark
;
1843 TCP_ECN_queue_cwr(tp
);
1845 tcp_clear_retrans_hints_partial(tp
);
1848 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1850 tp
->retrans_out
= 0;
1853 tp
->undo_marker
= 0;
1854 tp
->undo_retrans
= 0;
1857 void tcp_clear_retrans(struct tcp_sock
*tp
)
1859 tcp_clear_retrans_partial(tp
);
1861 tp
->fackets_out
= 0;
1865 /* Enter Loss state. If "how" is not zero, forget all SACK information
1866 * and reset tags completely, otherwise preserve SACKs. If receiver
1867 * dropped its ofo queue, we will know this due to reneging detection.
1869 void tcp_enter_loss(struct sock
*sk
, int how
)
1871 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1872 struct tcp_sock
*tp
= tcp_sk(sk
);
1873 struct sk_buff
*skb
;
1875 /* Reduce ssthresh if it has not yet been made inside this window. */
1876 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1877 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1878 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1879 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1880 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1883 tp
->snd_cwnd_cnt
= 0;
1884 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1886 tp
->bytes_acked
= 0;
1887 tcp_clear_retrans_partial(tp
);
1889 if (tcp_is_reno(tp
))
1890 tcp_reset_reno_sack(tp
);
1893 /* Push undo marker, if it was plain RTO and nothing
1894 * was retransmitted. */
1895 tp
->undo_marker
= tp
->snd_una
;
1896 tcp_clear_retrans_hints_partial(tp
);
1899 tp
->fackets_out
= 0;
1900 tcp_clear_all_retrans_hints(tp
);
1903 tcp_for_write_queue(skb
, sk
) {
1904 if (skb
== tcp_send_head(sk
))
1907 if (TCP_SKB_CB(skb
)->sacked
&TCPCB_RETRANS
)
1908 tp
->undo_marker
= 0;
1909 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1910 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1911 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1912 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1913 tp
->lost_out
+= tcp_skb_pcount(skb
);
1916 tcp_verify_left_out(tp
);
1918 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1919 sysctl_tcp_reordering
);
1920 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1921 tp
->high_seq
= tp
->snd_nxt
;
1922 TCP_ECN_queue_cwr(tp
);
1923 /* Abort F-RTO algorithm if one is in progress */
1924 tp
->frto_counter
= 0;
1927 static int tcp_check_sack_reneging(struct sock
*sk
)
1929 struct sk_buff
*skb
;
1931 /* If ACK arrived pointing to a remembered SACK,
1932 * it means that our remembered SACKs do not reflect
1933 * real state of receiver i.e.
1934 * receiver _host_ is heavily congested (or buggy).
1935 * Do processing similar to RTO timeout.
1937 if ((skb
= tcp_write_queue_head(sk
)) != NULL
&&
1938 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1939 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1940 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING
);
1942 tcp_enter_loss(sk
, 1);
1943 icsk
->icsk_retransmits
++;
1944 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1945 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1946 icsk
->icsk_rto
, TCP_RTO_MAX
);
1952 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
1954 return tcp_is_reno(tp
) ? tp
->sacked_out
+1 : tp
->fackets_out
;
1957 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1958 * counter when SACK is enabled (without SACK, sacked_out is used for
1961 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1962 * segments up to the highest received SACK block so far and holes in
1965 * With reordering, holes may still be in flight, so RFC3517 recovery
1966 * uses pure sacked_out (total number of SACKed segments) even though
1967 * it violates the RFC that uses duplicate ACKs, often these are equal
1968 * but when e.g. out-of-window ACKs or packet duplication occurs,
1969 * they differ. Since neither occurs due to loss, TCP should really
1972 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
1974 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1977 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
1979 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
1982 static inline int tcp_head_timedout(struct sock
*sk
)
1984 struct tcp_sock
*tp
= tcp_sk(sk
);
1986 return tp
->packets_out
&&
1987 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
1990 /* Linux NewReno/SACK/FACK/ECN state machine.
1991 * --------------------------------------
1993 * "Open" Normal state, no dubious events, fast path.
1994 * "Disorder" In all the respects it is "Open",
1995 * but requires a bit more attention. It is entered when
1996 * we see some SACKs or dupacks. It is split of "Open"
1997 * mainly to move some processing from fast path to slow one.
1998 * "CWR" CWND was reduced due to some Congestion Notification event.
1999 * It can be ECN, ICMP source quench, local device congestion.
2000 * "Recovery" CWND was reduced, we are fast-retransmitting.
2001 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2003 * tcp_fastretrans_alert() is entered:
2004 * - each incoming ACK, if state is not "Open"
2005 * - when arrived ACK is unusual, namely:
2010 * Counting packets in flight is pretty simple.
2012 * in_flight = packets_out - left_out + retrans_out
2014 * packets_out is SND.NXT-SND.UNA counted in packets.
2016 * retrans_out is number of retransmitted segments.
2018 * left_out is number of segments left network, but not ACKed yet.
2020 * left_out = sacked_out + lost_out
2022 * sacked_out: Packets, which arrived to receiver out of order
2023 * and hence not ACKed. With SACKs this number is simply
2024 * amount of SACKed data. Even without SACKs
2025 * it is easy to give pretty reliable estimate of this number,
2026 * counting duplicate ACKs.
2028 * lost_out: Packets lost by network. TCP has no explicit
2029 * "loss notification" feedback from network (for now).
2030 * It means that this number can be only _guessed_.
2031 * Actually, it is the heuristics to predict lossage that
2032 * distinguishes different algorithms.
2034 * F.e. after RTO, when all the queue is considered as lost,
2035 * lost_out = packets_out and in_flight = retrans_out.
2037 * Essentially, we have now two algorithms counting
2040 * FACK: It is the simplest heuristics. As soon as we decided
2041 * that something is lost, we decide that _all_ not SACKed
2042 * packets until the most forward SACK are lost. I.e.
2043 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2044 * It is absolutely correct estimate, if network does not reorder
2045 * packets. And it loses any connection to reality when reordering
2046 * takes place. We use FACK by default until reordering
2047 * is suspected on the path to this destination.
2049 * NewReno: when Recovery is entered, we assume that one segment
2050 * is lost (classic Reno). While we are in Recovery and
2051 * a partial ACK arrives, we assume that one more packet
2052 * is lost (NewReno). This heuristics are the same in NewReno
2055 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2056 * deflation etc. CWND is real congestion window, never inflated, changes
2057 * only according to classic VJ rules.
2059 * Really tricky (and requiring careful tuning) part of algorithm
2060 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2061 * The first determines the moment _when_ we should reduce CWND and,
2062 * hence, slow down forward transmission. In fact, it determines the moment
2063 * when we decide that hole is caused by loss, rather than by a reorder.
2065 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2066 * holes, caused by lost packets.
2068 * And the most logically complicated part of algorithm is undo
2069 * heuristics. We detect false retransmits due to both too early
2070 * fast retransmit (reordering) and underestimated RTO, analyzing
2071 * timestamps and D-SACKs. When we detect that some segments were
2072 * retransmitted by mistake and CWND reduction was wrong, we undo
2073 * window reduction and abort recovery phase. This logic is hidden
2074 * inside several functions named tcp_try_undo_<something>.
2077 /* This function decides, when we should leave Disordered state
2078 * and enter Recovery phase, reducing congestion window.
2080 * Main question: may we further continue forward transmission
2081 * with the same cwnd?
2083 static int tcp_time_to_recover(struct sock
*sk
)
2085 struct tcp_sock
*tp
= tcp_sk(sk
);
2088 /* Do not perform any recovery during F-RTO algorithm */
2089 if (tp
->frto_counter
)
2092 /* Trick#1: The loss is proven. */
2096 /* Not-A-Trick#2 : Classic rule... */
2097 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2100 /* Trick#3 : when we use RFC2988 timer restart, fast
2101 * retransmit can be triggered by timeout of queue head.
2103 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2106 /* Trick#4: It is still not OK... But will it be useful to delay
2109 packets_out
= tp
->packets_out
;
2110 if (packets_out
<= tp
->reordering
&&
2111 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2112 !tcp_may_send_now(sk
)) {
2113 /* We have nothing to send. This connection is limited
2114 * either by receiver window or by application.
2122 /* RFC: This is from the original, I doubt that this is necessary at all:
2123 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2124 * retransmitted past LOST markings in the first place? I'm not fully sure
2125 * about undo and end of connection cases, which can cause R without L?
2127 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
,
2128 struct sk_buff
*skb
)
2130 if ((tp
->retransmit_skb_hint
!= NULL
) &&
2131 before(TCP_SKB_CB(skb
)->seq
,
2132 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
2133 tp
->retransmit_skb_hint
= NULL
;
2136 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2137 * is against sacked "cnt", otherwise it's against facked "cnt"
2139 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int fast_rexmit
)
2141 struct tcp_sock
*tp
= tcp_sk(sk
);
2142 struct sk_buff
*skb
;
2145 BUG_TRAP(packets
<= tp
->packets_out
);
2146 if (tp
->lost_skb_hint
) {
2147 skb
= tp
->lost_skb_hint
;
2148 cnt
= tp
->lost_cnt_hint
;
2150 skb
= tcp_write_queue_head(sk
);
2154 tcp_for_write_queue_from(skb
, sk
) {
2155 if (skb
== tcp_send_head(sk
))
2157 /* TODO: do this better */
2158 /* this is not the most efficient way to do this... */
2159 tp
->lost_skb_hint
= skb
;
2160 tp
->lost_cnt_hint
= cnt
;
2162 if (tcp_is_fack(tp
) ||
2163 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2164 cnt
+= tcp_skb_pcount(skb
);
2166 if (((!fast_rexmit
|| (tp
->lost_out
> 0)) && (cnt
> packets
)) ||
2167 after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2169 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2170 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2171 tp
->lost_out
+= tcp_skb_pcount(skb
);
2172 tcp_verify_retransmit_hint(tp
, skb
);
2175 tcp_verify_left_out(tp
);
2178 /* Account newly detected lost packet(s) */
2180 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2182 struct tcp_sock
*tp
= tcp_sk(sk
);
2184 if (tcp_is_reno(tp
)) {
2185 tcp_mark_head_lost(sk
, 1, fast_rexmit
);
2186 } else if (tcp_is_fack(tp
)) {
2187 int lost
= tp
->fackets_out
- tp
->reordering
;
2190 tcp_mark_head_lost(sk
, lost
, fast_rexmit
);
2192 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2193 if (sacked_upto
< 0)
2195 tcp_mark_head_lost(sk
, sacked_upto
, fast_rexmit
);
2198 /* New heuristics: it is possible only after we switched
2199 * to restart timer each time when something is ACKed.
2200 * Hence, we can detect timed out packets during fast
2201 * retransmit without falling to slow start.
2203 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2204 struct sk_buff
*skb
;
2206 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2207 : tcp_write_queue_head(sk
);
2209 tcp_for_write_queue_from(skb
, sk
) {
2210 if (skb
== tcp_send_head(sk
))
2212 if (!tcp_skb_timedout(sk
, skb
))
2215 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_SACKED_ACKED
|TCPCB_LOST
))) {
2216 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2217 tp
->lost_out
+= tcp_skb_pcount(skb
);
2218 tcp_verify_retransmit_hint(tp
, skb
);
2222 tp
->scoreboard_skb_hint
= skb
;
2224 tcp_verify_left_out(tp
);
2228 /* CWND moderation, preventing bursts due to too big ACKs
2229 * in dubious situations.
2231 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2233 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2234 tcp_packets_in_flight(tp
)+tcp_max_burst(tp
));
2235 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2238 /* Lower bound on congestion window is slow start threshold
2239 * unless congestion avoidance choice decides to overide it.
2241 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2243 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2245 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2248 /* Decrease cwnd each second ack. */
2249 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2251 struct tcp_sock
*tp
= tcp_sk(sk
);
2252 int decr
= tp
->snd_cwnd_cnt
+ 1;
2254 if ((flag
&(FLAG_ANY_PROGRESS
|FLAG_DSACKING_ACK
)) ||
2255 (tcp_is_reno(tp
) && !(flag
&FLAG_NOT_DUP
))) {
2256 tp
->snd_cwnd_cnt
= decr
&1;
2259 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2260 tp
->snd_cwnd
-= decr
;
2262 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
)+1);
2263 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2267 /* Nothing was retransmitted or returned timestamp is less
2268 * than timestamp of the first retransmission.
2270 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2272 return !tp
->retrans_stamp
||
2273 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2274 (__s32
)(tp
->rx_opt
.rcv_tsecr
- tp
->retrans_stamp
) < 0);
2277 /* Undo procedures. */
2279 #if FASTRETRANS_DEBUG > 1
2280 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2282 struct tcp_sock
*tp
= tcp_sk(sk
);
2283 struct inet_sock
*inet
= inet_sk(sk
);
2285 printk(KERN_DEBUG
"Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2287 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2288 tp
->snd_cwnd
, tcp_left_out(tp
),
2289 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2293 #define DBGUNDO(x...) do { } while (0)
2296 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2298 struct tcp_sock
*tp
= tcp_sk(sk
);
2300 if (tp
->prior_ssthresh
) {
2301 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2303 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2304 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2306 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<<1);
2308 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2309 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2310 TCP_ECN_withdraw_cwr(tp
);
2313 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2315 tcp_moderate_cwnd(tp
);
2316 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2318 /* There is something screwy going on with the retrans hints after
2320 tcp_clear_all_retrans_hints(tp
);
2323 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2325 return tp
->undo_marker
&&
2326 (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2329 /* People celebrate: "We love our President!" */
2330 static int tcp_try_undo_recovery(struct sock
*sk
)
2332 struct tcp_sock
*tp
= tcp_sk(sk
);
2334 if (tcp_may_undo(tp
)) {
2335 /* Happy end! We did not retransmit anything
2336 * or our original transmission succeeded.
2338 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2339 tcp_undo_cwr(sk
, 1);
2340 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2341 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2343 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO
);
2344 tp
->undo_marker
= 0;
2346 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2347 /* Hold old state until something *above* high_seq
2348 * is ACKed. For Reno it is MUST to prevent false
2349 * fast retransmits (RFC2582). SACK TCP is safe. */
2350 tcp_moderate_cwnd(tp
);
2353 tcp_set_ca_state(sk
, TCP_CA_Open
);
2357 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2358 static void tcp_try_undo_dsack(struct sock
*sk
)
2360 struct tcp_sock
*tp
= tcp_sk(sk
);
2362 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2363 DBGUNDO(sk
, "D-SACK");
2364 tcp_undo_cwr(sk
, 1);
2365 tp
->undo_marker
= 0;
2366 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO
);
2370 /* Undo during fast recovery after partial ACK. */
2372 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2374 struct tcp_sock
*tp
= tcp_sk(sk
);
2375 /* Partial ACK arrived. Force Hoe's retransmit. */
2376 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2378 if (tcp_may_undo(tp
)) {
2379 /* Plain luck! Hole if filled with delayed
2380 * packet, rather than with a retransmit.
2382 if (tp
->retrans_out
== 0)
2383 tp
->retrans_stamp
= 0;
2385 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2388 tcp_undo_cwr(sk
, 0);
2389 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO
);
2391 /* So... Do not make Hoe's retransmit yet.
2392 * If the first packet was delayed, the rest
2393 * ones are most probably delayed as well.
2400 /* Undo during loss recovery after partial ACK. */
2401 static int tcp_try_undo_loss(struct sock
*sk
)
2403 struct tcp_sock
*tp
= tcp_sk(sk
);
2405 if (tcp_may_undo(tp
)) {
2406 struct sk_buff
*skb
;
2407 tcp_for_write_queue(skb
, sk
) {
2408 if (skb
== tcp_send_head(sk
))
2410 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2413 tcp_clear_all_retrans_hints(tp
);
2415 DBGUNDO(sk
, "partial loss");
2417 tcp_undo_cwr(sk
, 1);
2418 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO
);
2419 inet_csk(sk
)->icsk_retransmits
= 0;
2420 tp
->undo_marker
= 0;
2421 if (tcp_is_sack(tp
))
2422 tcp_set_ca_state(sk
, TCP_CA_Open
);
2428 static inline void tcp_complete_cwr(struct sock
*sk
)
2430 struct tcp_sock
*tp
= tcp_sk(sk
);
2431 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2432 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2433 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2436 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2438 struct tcp_sock
*tp
= tcp_sk(sk
);
2440 tcp_verify_left_out(tp
);
2442 if (tp
->retrans_out
== 0)
2443 tp
->retrans_stamp
= 0;
2446 tcp_enter_cwr(sk
, 1);
2448 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2449 int state
= TCP_CA_Open
;
2451 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2452 state
= TCP_CA_Disorder
;
2454 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2455 tcp_set_ca_state(sk
, state
);
2456 tp
->high_seq
= tp
->snd_nxt
;
2458 tcp_moderate_cwnd(tp
);
2460 tcp_cwnd_down(sk
, flag
);
2464 static void tcp_mtup_probe_failed(struct sock
*sk
)
2466 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2468 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2469 icsk
->icsk_mtup
.probe_size
= 0;
2472 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2474 struct tcp_sock
*tp
= tcp_sk(sk
);
2475 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2477 /* FIXME: breaks with very large cwnd */
2478 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2479 tp
->snd_cwnd
= tp
->snd_cwnd
*
2480 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2481 icsk
->icsk_mtup
.probe_size
;
2482 tp
->snd_cwnd_cnt
= 0;
2483 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2484 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2486 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2487 icsk
->icsk_mtup
.probe_size
= 0;
2488 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2492 /* Process an event, which can update packets-in-flight not trivially.
2493 * Main goal of this function is to calculate new estimate for left_out,
2494 * taking into account both packets sitting in receiver's buffer and
2495 * packets lost by network.
2497 * Besides that it does CWND reduction, when packet loss is detected
2498 * and changes state of machine.
2500 * It does _not_ decide what to send, it is made in function
2501 * tcp_xmit_retransmit_queue().
2504 tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2506 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2507 struct tcp_sock
*tp
= tcp_sk(sk
);
2508 int is_dupack
= !(flag
&(FLAG_SND_UNA_ADVANCED
|FLAG_NOT_DUP
));
2509 int do_lost
= is_dupack
|| ((flag
&FLAG_DATA_SACKED
) &&
2510 (tcp_fackets_out(tp
) > tp
->reordering
));
2511 int fast_rexmit
= 0;
2513 /* Some technical things:
2514 * 1. Reno does not count dupacks (sacked_out) automatically. */
2515 if (!tp
->packets_out
)
2518 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2519 tp
->fackets_out
= 0;
2521 /* Now state machine starts.
2522 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2524 tp
->prior_ssthresh
= 0;
2526 /* B. In all the states check for reneging SACKs. */
2527 if (tp
->sacked_out
&& tcp_check_sack_reneging(sk
))
2530 /* C. Process data loss notification, provided it is valid. */
2531 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2532 before(tp
->snd_una
, tp
->high_seq
) &&
2533 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2534 tp
->fackets_out
> tp
->reordering
) {
2535 tcp_mark_head_lost(sk
, tp
->fackets_out
-tp
->reordering
, 0);
2536 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS
);
2539 /* D. Check consistency of the current state. */
2540 tcp_verify_left_out(tp
);
2542 /* E. Check state exit conditions. State can be terminated
2543 * when high_seq is ACKed. */
2544 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2545 BUG_TRAP(tp
->retrans_out
== 0);
2546 tp
->retrans_stamp
= 0;
2547 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2548 switch (icsk
->icsk_ca_state
) {
2550 icsk
->icsk_retransmits
= 0;
2551 if (tcp_try_undo_recovery(sk
))
2556 /* CWR is to be held something *above* high_seq
2557 * is ACKed for CWR bit to reach receiver. */
2558 if (tp
->snd_una
!= tp
->high_seq
) {
2559 tcp_complete_cwr(sk
);
2560 tcp_set_ca_state(sk
, TCP_CA_Open
);
2564 case TCP_CA_Disorder
:
2565 tcp_try_undo_dsack(sk
);
2566 if (!tp
->undo_marker
||
2567 /* For SACK case do not Open to allow to undo
2568 * catching for all duplicate ACKs. */
2569 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2570 tp
->undo_marker
= 0;
2571 tcp_set_ca_state(sk
, TCP_CA_Open
);
2575 case TCP_CA_Recovery
:
2576 if (tcp_is_reno(tp
))
2577 tcp_reset_reno_sack(tp
);
2578 if (tcp_try_undo_recovery(sk
))
2580 tcp_complete_cwr(sk
);
2585 /* F. Process state. */
2586 switch (icsk
->icsk_ca_state
) {
2587 case TCP_CA_Recovery
:
2588 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2589 if (tcp_is_reno(tp
) && is_dupack
)
2590 tcp_add_reno_sack(sk
);
2592 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2595 if (flag
&FLAG_DATA_ACKED
)
2596 icsk
->icsk_retransmits
= 0;
2597 if (!tcp_try_undo_loss(sk
)) {
2598 tcp_moderate_cwnd(tp
);
2599 tcp_xmit_retransmit_queue(sk
);
2602 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2604 /* Loss is undone; fall through to processing in Open state. */
2606 if (tcp_is_reno(tp
)) {
2607 if (flag
& FLAG_SND_UNA_ADVANCED
)
2608 tcp_reset_reno_sack(tp
);
2610 tcp_add_reno_sack(sk
);
2613 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2614 tcp_try_undo_dsack(sk
);
2616 if (!tcp_time_to_recover(sk
)) {
2617 tcp_try_to_open(sk
, flag
);
2621 /* MTU probe failure: don't reduce cwnd */
2622 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2623 icsk
->icsk_mtup
.probe_size
&&
2624 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2625 tcp_mtup_probe_failed(sk
);
2626 /* Restores the reduction we did in tcp_mtup_probe() */
2628 tcp_simple_retransmit(sk
);
2632 /* Otherwise enter Recovery state */
2634 if (tcp_is_reno(tp
))
2635 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY
);
2637 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY
);
2639 tp
->high_seq
= tp
->snd_nxt
;
2640 tp
->prior_ssthresh
= 0;
2641 tp
->undo_marker
= tp
->snd_una
;
2642 tp
->undo_retrans
= tp
->retrans_out
;
2644 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2645 if (!(flag
&FLAG_ECE
))
2646 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2647 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2648 TCP_ECN_queue_cwr(tp
);
2651 tp
->bytes_acked
= 0;
2652 tp
->snd_cwnd_cnt
= 0;
2653 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2657 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2658 tcp_update_scoreboard(sk
, fast_rexmit
);
2659 tcp_cwnd_down(sk
, flag
);
2660 tcp_xmit_retransmit_queue(sk
);
2663 /* Read draft-ietf-tcplw-high-performance before mucking
2664 * with this code. (Supersedes RFC1323)
2666 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2668 /* RTTM Rule: A TSecr value received in a segment is used to
2669 * update the averaged RTT measurement only if the segment
2670 * acknowledges some new data, i.e., only if it advances the
2671 * left edge of the send window.
2673 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2674 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2676 * Changed: reset backoff as soon as we see the first valid sample.
2677 * If we do not, we get strongly overestimated rto. With timestamps
2678 * samples are accepted even from very old segments: f.e., when rtt=1
2679 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2680 * answer arrives rto becomes 120 seconds! If at least one of segments
2681 * in window is lost... Voila. --ANK (010210)
2683 struct tcp_sock
*tp
= tcp_sk(sk
);
2684 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2685 tcp_rtt_estimator(sk
, seq_rtt
);
2687 inet_csk(sk
)->icsk_backoff
= 0;
2691 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2693 /* We don't have a timestamp. Can only use
2694 * packets that are not retransmitted to determine
2695 * rtt estimates. Also, we must not reset the
2696 * backoff for rto until we get a non-retransmitted
2697 * packet. This allows us to deal with a situation
2698 * where the network delay has increased suddenly.
2699 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2702 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2705 tcp_rtt_estimator(sk
, seq_rtt
);
2707 inet_csk(sk
)->icsk_backoff
= 0;
2711 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2714 const struct tcp_sock
*tp
= tcp_sk(sk
);
2715 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2716 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2717 tcp_ack_saw_tstamp(sk
, flag
);
2718 else if (seq_rtt
>= 0)
2719 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2722 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
,
2723 u32 in_flight
, int good
)
2725 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2726 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
, good
);
2727 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2730 /* Restart timer after forward progress on connection.
2731 * RFC2988 recommends to restart timer to now+rto.
2733 static void tcp_rearm_rto(struct sock
*sk
)
2735 struct tcp_sock
*tp
= tcp_sk(sk
);
2737 if (!tp
->packets_out
) {
2738 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2740 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2744 /* If we get here, the whole TSO packet has not been acked. */
2745 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2747 struct tcp_sock
*tp
= tcp_sk(sk
);
2750 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2752 packets_acked
= tcp_skb_pcount(skb
);
2753 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2755 packets_acked
-= tcp_skb_pcount(skb
);
2757 if (packets_acked
) {
2758 BUG_ON(tcp_skb_pcount(skb
) == 0);
2759 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2762 return packets_acked
;
2765 /* Remove acknowledged frames from the retransmission queue. If our packet
2766 * is before the ack sequence we can discard it as it's confirmed to have
2767 * arrived at the other end.
2769 static int tcp_clean_rtx_queue(struct sock
*sk
, s32
*seq_rtt_p
,
2772 struct tcp_sock
*tp
= tcp_sk(sk
);
2773 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2774 struct sk_buff
*skb
;
2775 u32 now
= tcp_time_stamp
;
2776 int fully_acked
= 1;
2778 int prior_packets
= tp
->packets_out
;
2780 u32 reord
= tp
->packets_out
;
2782 s32 ca_seq_rtt
= -1;
2783 ktime_t last_ackt
= net_invalid_timestamp();
2785 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2786 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2789 u8 sacked
= scb
->sacked
;
2791 /* Determine how many packets and what bytes were acked, tso and else */
2792 if (after(scb
->end_seq
, tp
->snd_una
)) {
2793 if (tcp_skb_pcount(skb
) == 1 ||
2794 !after(tp
->snd_una
, scb
->seq
))
2797 packets_acked
= tcp_tso_acked(sk
, skb
);
2802 end_seq
= tp
->snd_una
;
2804 packets_acked
= tcp_skb_pcount(skb
);
2805 end_seq
= scb
->end_seq
;
2808 /* MTU probing checks */
2809 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2810 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2811 tcp_mtup_probe_success(sk
, skb
);
2815 if (sacked
& TCPCB_RETRANS
) {
2816 if (sacked
& TCPCB_SACKED_RETRANS
)
2817 tp
->retrans_out
-= packets_acked
;
2818 flag
|= FLAG_RETRANS_DATA_ACKED
;
2821 if ((flag
& FLAG_DATA_ACKED
) ||
2822 (packets_acked
> 1))
2823 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2825 ca_seq_rtt
= now
- scb
->when
;
2826 last_ackt
= skb
->tstamp
;
2828 seq_rtt
= ca_seq_rtt
;
2830 if (!(sacked
& TCPCB_SACKED_ACKED
))
2831 reord
= min(cnt
, reord
);
2834 if (sacked
& TCPCB_SACKED_ACKED
)
2835 tp
->sacked_out
-= packets_acked
;
2836 if (sacked
& TCPCB_LOST
)
2837 tp
->lost_out
-= packets_acked
;
2839 if ((sacked
& TCPCB_URG
) && tp
->urg_mode
&&
2840 !before(end_seq
, tp
->snd_up
))
2843 ca_seq_rtt
= now
- scb
->when
;
2844 last_ackt
= skb
->tstamp
;
2846 seq_rtt
= ca_seq_rtt
;
2848 reord
= min(cnt
, reord
);
2850 tp
->packets_out
-= packets_acked
;
2851 cnt
+= packets_acked
;
2853 /* Initial outgoing SYN's get put onto the write_queue
2854 * just like anything else we transmit. It is not
2855 * true data, and if we misinform our callers that
2856 * this ACK acks real data, we will erroneously exit
2857 * connection startup slow start one packet too
2858 * quickly. This is severely frowned upon behavior.
2860 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2861 flag
|= FLAG_DATA_ACKED
;
2863 flag
|= FLAG_SYN_ACKED
;
2864 tp
->retrans_stamp
= 0;
2870 tcp_unlink_write_queue(skb
, sk
);
2871 sk_stream_free_skb(sk
, skb
);
2872 tcp_clear_all_retrans_hints(tp
);
2875 if (flag
& FLAG_ACKED
) {
2876 u32 pkts_acked
= prior_packets
- tp
->packets_out
;
2877 const struct tcp_congestion_ops
*ca_ops
2878 = inet_csk(sk
)->icsk_ca_ops
;
2880 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2883 if (tcp_is_reno(tp
)) {
2884 tcp_remove_reno_sacks(sk
, pkts_acked
);
2886 /* Non-retransmitted hole got filled? That's reordering */
2887 if (reord
< prior_fackets
)
2888 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2891 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2893 if (ca_ops
->pkts_acked
) {
2896 /* Is the ACK triggering packet unambiguous? */
2897 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2898 /* High resolution needed and available? */
2899 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2900 !ktime_equal(last_ackt
,
2901 net_invalid_timestamp()))
2902 rtt_us
= ktime_us_delta(ktime_get_real(),
2904 else if (ca_seq_rtt
> 0)
2905 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
2908 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2912 #if FASTRETRANS_DEBUG > 0
2913 BUG_TRAP((int)tp
->sacked_out
>= 0);
2914 BUG_TRAP((int)tp
->lost_out
>= 0);
2915 BUG_TRAP((int)tp
->retrans_out
>= 0);
2916 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2917 icsk
= inet_csk(sk
);
2919 printk(KERN_DEBUG
"Leak l=%u %d\n",
2920 tp
->lost_out
, icsk
->icsk_ca_state
);
2923 if (tp
->sacked_out
) {
2924 printk(KERN_DEBUG
"Leak s=%u %d\n",
2925 tp
->sacked_out
, icsk
->icsk_ca_state
);
2928 if (tp
->retrans_out
) {
2929 printk(KERN_DEBUG
"Leak r=%u %d\n",
2930 tp
->retrans_out
, icsk
->icsk_ca_state
);
2931 tp
->retrans_out
= 0;
2935 *seq_rtt_p
= seq_rtt
;
2939 static void tcp_ack_probe(struct sock
*sk
)
2941 const struct tcp_sock
*tp
= tcp_sk(sk
);
2942 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2944 /* Was it a usable window open? */
2946 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
,
2947 tp
->snd_una
+ tp
->snd_wnd
)) {
2948 icsk
->icsk_backoff
= 0;
2949 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
2950 /* Socket must be waked up by subsequent tcp_data_snd_check().
2951 * This function is not for random using!
2954 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2955 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2960 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
2962 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
2963 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
2966 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
2968 const struct tcp_sock
*tp
= tcp_sk(sk
);
2969 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
2970 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
2973 /* Check that window update is acceptable.
2974 * The function assumes that snd_una<=ack<=snd_next.
2976 static inline int tcp_may_update_window(const struct tcp_sock
*tp
, const u32 ack
,
2977 const u32 ack_seq
, const u32 nwin
)
2979 return (after(ack
, tp
->snd_una
) ||
2980 after(ack_seq
, tp
->snd_wl1
) ||
2981 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
2984 /* Update our send window.
2986 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2987 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2989 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
2992 struct tcp_sock
*tp
= tcp_sk(sk
);
2994 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
2996 if (likely(!tcp_hdr(skb
)->syn
))
2997 nwin
<<= tp
->rx_opt
.snd_wscale
;
2999 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3000 flag
|= FLAG_WIN_UPDATE
;
3001 tcp_update_wl(tp
, ack
, ack_seq
);
3003 if (tp
->snd_wnd
!= nwin
) {
3006 /* Note, it is the only place, where
3007 * fast path is recovered for sending TCP.
3010 tcp_fast_path_check(sk
);
3012 if (nwin
> tp
->max_window
) {
3013 tp
->max_window
= nwin
;
3014 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3024 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3025 * continue in congestion avoidance.
3027 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3029 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3030 tp
->snd_cwnd_cnt
= 0;
3031 tp
->bytes_acked
= 0;
3032 TCP_ECN_queue_cwr(tp
);
3033 tcp_moderate_cwnd(tp
);
3036 /* A conservative spurious RTO response algorithm: reduce cwnd using
3037 * rate halving and continue in congestion avoidance.
3039 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3041 tcp_enter_cwr(sk
, 0);
3044 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3047 tcp_ratehalving_spur_to_response(sk
);
3049 tcp_undo_cwr(sk
, 1);
3052 /* F-RTO spurious RTO detection algorithm (RFC4138)
3054 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3055 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3056 * window (but not to or beyond highest sequence sent before RTO):
3057 * On First ACK, send two new segments out.
3058 * On Second ACK, RTO was likely spurious. Do spurious response (response
3059 * algorithm is not part of the F-RTO detection algorithm
3060 * given in RFC4138 but can be selected separately).
3061 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3062 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3063 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3064 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3066 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3067 * original window even after we transmit two new data segments.
3070 * on first step, wait until first cumulative ACK arrives, then move to
3071 * the second step. In second step, the next ACK decides.
3073 * F-RTO is implemented (mainly) in four functions:
3074 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3075 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3076 * called when tcp_use_frto() showed green light
3077 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3078 * - tcp_enter_frto_loss() is called if there is not enough evidence
3079 * to prove that the RTO is indeed spurious. It transfers the control
3080 * from F-RTO to the conventional RTO recovery
3082 static int tcp_process_frto(struct sock
*sk
, int flag
)
3084 struct tcp_sock
*tp
= tcp_sk(sk
);
3086 tcp_verify_left_out(tp
);
3088 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3089 if (flag
&FLAG_DATA_ACKED
)
3090 inet_csk(sk
)->icsk_retransmits
= 0;
3092 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3093 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3094 tp
->undo_marker
= 0;
3096 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3097 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3101 if (!IsSackFrto() || tcp_is_reno(tp
)) {
3102 /* RFC4138 shortcoming in step 2; should also have case c):
3103 * ACK isn't duplicate nor advances window, e.g., opposite dir
3106 if (!(flag
&FLAG_ANY_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3109 if (!(flag
&FLAG_DATA_ACKED
)) {
3110 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3115 if (!(flag
&FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3116 /* Prevent sending of new data. */
3117 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3118 tcp_packets_in_flight(tp
));
3122 if ((tp
->frto_counter
>= 2) &&
3123 (!(flag
&FLAG_FORWARD_PROGRESS
) ||
3124 ((flag
&FLAG_DATA_SACKED
) && !(flag
&FLAG_ONLY_ORIG_SACKED
)))) {
3125 /* RFC4138 shortcoming (see comment above) */
3126 if (!(flag
&FLAG_FORWARD_PROGRESS
) && (flag
&FLAG_NOT_DUP
))
3129 tcp_enter_frto_loss(sk
, 3, flag
);
3134 if (tp
->frto_counter
== 1) {
3135 /* tcp_may_send_now needs to see updated state */
3136 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3137 tp
->frto_counter
= 2;
3139 if (!tcp_may_send_now(sk
))
3140 tcp_enter_frto_loss(sk
, 2, flag
);
3144 switch (sysctl_tcp_frto_response
) {
3146 tcp_undo_spur_to_response(sk
, flag
);
3149 tcp_conservative_spur_to_response(tp
);
3152 tcp_ratehalving_spur_to_response(sk
);
3155 tp
->frto_counter
= 0;
3156 tp
->undo_marker
= 0;
3157 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS
);
3162 /* This routine deals with incoming acks, but not outgoing ones. */
3163 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3165 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3166 struct tcp_sock
*tp
= tcp_sk(sk
);
3167 u32 prior_snd_una
= tp
->snd_una
;
3168 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3169 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3170 u32 prior_in_flight
;
3176 /* If the ack is newer than sent or older than previous acks
3177 * then we can probably ignore it.
3179 if (after(ack
, tp
->snd_nxt
))
3180 goto uninteresting_ack
;
3182 if (before(ack
, prior_snd_una
))
3185 if (after(ack
, prior_snd_una
))
3186 flag
|= FLAG_SND_UNA_ADVANCED
;
3188 if (sysctl_tcp_abc
) {
3189 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3190 tp
->bytes_acked
+= ack
- prior_snd_una
;
3191 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3192 /* we assume just one segment left network */
3193 tp
->bytes_acked
+= min(ack
- prior_snd_una
, tp
->mss_cache
);
3196 prior_fackets
= tp
->fackets_out
;
3197 prior_in_flight
= tcp_packets_in_flight(tp
);
3199 if (!(flag
&FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3200 /* Window is constant, pure forward advance.
3201 * No more checks are required.
3202 * Note, we use the fact that SND.UNA>=SND.WL2.
3204 tcp_update_wl(tp
, ack
, ack_seq
);
3206 flag
|= FLAG_WIN_UPDATE
;
3208 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3210 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS
);
3212 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3215 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS
);
3217 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3219 if (TCP_SKB_CB(skb
)->sacked
)
3220 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3222 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3225 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3228 /* We passed data and got it acked, remove any soft error
3229 * log. Something worked...
3231 sk
->sk_err_soft
= 0;
3232 tp
->rcv_tstamp
= tcp_time_stamp
;
3233 prior_packets
= tp
->packets_out
;
3237 /* See if we can take anything off of the retransmit queue. */
3238 flag
|= tcp_clean_rtx_queue(sk
, &seq_rtt
, prior_fackets
);
3240 if (tp
->frto_counter
)
3241 frto_cwnd
= tcp_process_frto(sk
, flag
);
3242 /* Guarantee sacktag reordering detection against wrap-arounds */
3243 if (before(tp
->frto_highmark
, tp
->snd_una
))
3244 tp
->frto_highmark
= 0;
3246 if (tcp_ack_is_dubious(sk
, flag
)) {
3247 /* Advance CWND, if state allows this. */
3248 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3249 tcp_may_raise_cwnd(sk
, flag
))
3250 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 0);
3251 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
, flag
);
3253 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3254 tcp_cong_avoid(sk
, ack
, prior_in_flight
, 1);
3257 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
&FLAG_NOT_DUP
))
3258 dst_confirm(sk
->sk_dst_cache
);
3263 icsk
->icsk_probes_out
= 0;
3265 /* If this ack opens up a zero window, clear backoff. It was
3266 * being used to time the probes, and is probably far higher than
3267 * it needs to be for normal retransmission.
3269 if (tcp_send_head(sk
))
3274 if (TCP_SKB_CB(skb
)->sacked
)
3275 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3278 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3283 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3284 * But, this can also be called on packets in the established flow when
3285 * the fast version below fails.
3287 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
, int estab
)
3290 struct tcphdr
*th
= tcp_hdr(skb
);
3291 int length
=(th
->doff
*4)-sizeof(struct tcphdr
);
3293 ptr
= (unsigned char *)(th
+ 1);
3294 opt_rx
->saw_tstamp
= 0;
3296 while (length
> 0) {
3303 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3308 if (opsize
< 2) /* "silly options" */
3310 if (opsize
> length
)
3311 return; /* don't parse partial options */
3314 if (opsize
==TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3315 u16 in_mss
= ntohs(get_unaligned((__be16
*)ptr
));
3317 if (opt_rx
->user_mss
&& opt_rx
->user_mss
< in_mss
)
3318 in_mss
= opt_rx
->user_mss
;
3319 opt_rx
->mss_clamp
= in_mss
;
3324 if (opsize
==TCPOLEN_WINDOW
&& th
->syn
&& !estab
)
3325 if (sysctl_tcp_window_scaling
) {
3326 __u8 snd_wscale
= *(__u8
*) ptr
;
3327 opt_rx
->wscale_ok
= 1;
3328 if (snd_wscale
> 14) {
3329 if (net_ratelimit())
3330 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3331 "scaling value %d >14 received.\n",
3335 opt_rx
->snd_wscale
= snd_wscale
;
3338 case TCPOPT_TIMESTAMP
:
3339 if (opsize
==TCPOLEN_TIMESTAMP
) {
3340 if ((estab
&& opt_rx
->tstamp_ok
) ||
3341 (!estab
&& sysctl_tcp_timestamps
)) {
3342 opt_rx
->saw_tstamp
= 1;
3343 opt_rx
->rcv_tsval
= ntohl(get_unaligned((__be32
*)ptr
));
3344 opt_rx
->rcv_tsecr
= ntohl(get_unaligned((__be32
*)(ptr
+4)));
3348 case TCPOPT_SACK_PERM
:
3349 if (opsize
==TCPOLEN_SACK_PERM
&& th
->syn
&& !estab
) {
3350 if (sysctl_tcp_sack
) {
3351 opt_rx
->sack_ok
= 1;
3352 tcp_sack_reset(opt_rx
);
3358 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3359 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3361 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3364 #ifdef CONFIG_TCP_MD5SIG
3367 * The MD5 Hash has already been
3368 * checked (see tcp_v{4,6}_do_rcv()).
3380 /* Fast parse options. This hopes to only see timestamps.
3381 * If it is wrong it falls back on tcp_parse_options().
3383 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3384 struct tcp_sock
*tp
)
3386 if (th
->doff
== sizeof(struct tcphdr
)>>2) {
3387 tp
->rx_opt
.saw_tstamp
= 0;
3389 } else if (tp
->rx_opt
.tstamp_ok
&&
3390 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3391 __be32
*ptr
= (__be32
*)(th
+ 1);
3392 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3393 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3394 tp
->rx_opt
.saw_tstamp
= 1;
3396 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3398 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3402 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3406 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3408 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3409 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3412 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3414 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3415 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3416 * extra check below makes sure this can only happen
3417 * for pure ACK frames. -DaveM
3419 * Not only, also it occurs for expired timestamps.
3422 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3423 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3424 tcp_store_ts_recent(tp
);
3428 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3430 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3431 * it can pass through stack. So, the following predicate verifies that
3432 * this segment is not used for anything but congestion avoidance or
3433 * fast retransmit. Moreover, we even are able to eliminate most of such
3434 * second order effects, if we apply some small "replay" window (~RTO)
3435 * to timestamp space.
3437 * All these measures still do not guarantee that we reject wrapped ACKs
3438 * on networks with high bandwidth, when sequence space is recycled fastly,
3439 * but it guarantees that such events will be very rare and do not affect
3440 * connection seriously. This doesn't look nice, but alas, PAWS is really
3443 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3444 * states that events when retransmit arrives after original data are rare.
3445 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3446 * the biggest problem on large power networks even with minor reordering.
3447 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3448 * up to bandwidth of 18Gigabit/sec. 8) ]
3451 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3453 struct tcp_sock
*tp
= tcp_sk(sk
);
3454 struct tcphdr
*th
= tcp_hdr(skb
);
3455 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3456 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3458 return (/* 1. Pure ACK with correct sequence number. */
3459 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3461 /* 2. ... and duplicate ACK. */
3462 ack
== tp
->snd_una
&&
3464 /* 3. ... and does not update window. */
3465 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3467 /* 4. ... and sits in replay window. */
3468 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3471 static inline int tcp_paws_discard(const struct sock
*sk
, const struct sk_buff
*skb
)
3473 const struct tcp_sock
*tp
= tcp_sk(sk
);
3474 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3475 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3476 !tcp_disordered_ack(sk
, skb
));
3479 /* Check segment sequence number for validity.
3481 * Segment controls are considered valid, if the segment
3482 * fits to the window after truncation to the window. Acceptability
3483 * of data (and SYN, FIN, of course) is checked separately.
3484 * See tcp_data_queue(), for example.
3486 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3487 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3488 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3489 * (borrowed from freebsd)
3492 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3494 return !before(end_seq
, tp
->rcv_wup
) &&
3495 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3498 /* When we get a reset we do this. */
3499 static void tcp_reset(struct sock
*sk
)
3501 /* We want the right error as BSD sees it (and indeed as we do). */
3502 switch (sk
->sk_state
) {
3504 sk
->sk_err
= ECONNREFUSED
;
3506 case TCP_CLOSE_WAIT
:
3512 sk
->sk_err
= ECONNRESET
;
3515 if (!sock_flag(sk
, SOCK_DEAD
))
3516 sk
->sk_error_report(sk
);
3522 * Process the FIN bit. This now behaves as it is supposed to work
3523 * and the FIN takes effect when it is validly part of sequence
3524 * space. Not before when we get holes.
3526 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3527 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3530 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3531 * close and we go into CLOSING (and later onto TIME-WAIT)
3533 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3535 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3537 struct tcp_sock
*tp
= tcp_sk(sk
);
3539 inet_csk_schedule_ack(sk
);
3541 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3542 sock_set_flag(sk
, SOCK_DONE
);
3544 switch (sk
->sk_state
) {
3546 case TCP_ESTABLISHED
:
3547 /* Move to CLOSE_WAIT */
3548 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3549 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3552 case TCP_CLOSE_WAIT
:
3554 /* Received a retransmission of the FIN, do
3559 /* RFC793: Remain in the LAST-ACK state. */
3563 /* This case occurs when a simultaneous close
3564 * happens, we must ack the received FIN and
3565 * enter the CLOSING state.
3568 tcp_set_state(sk
, TCP_CLOSING
);
3571 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3573 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3576 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3577 * cases we should never reach this piece of code.
3579 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3580 __FUNCTION__
, sk
->sk_state
);
3584 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3585 * Probably, we should reset in this case. For now drop them.
3587 __skb_queue_purge(&tp
->out_of_order_queue
);
3588 if (tcp_is_sack(tp
))
3589 tcp_sack_reset(&tp
->rx_opt
);
3590 sk_stream_mem_reclaim(sk
);
3592 if (!sock_flag(sk
, SOCK_DEAD
)) {
3593 sk
->sk_state_change(sk
);
3595 /* Do not send POLL_HUP for half duplex close. */
3596 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3597 sk
->sk_state
== TCP_CLOSE
)
3598 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3600 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3604 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
, u32 end_seq
)
3606 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3607 if (before(seq
, sp
->start_seq
))
3608 sp
->start_seq
= seq
;
3609 if (after(end_seq
, sp
->end_seq
))
3610 sp
->end_seq
= end_seq
;
3616 static void tcp_dsack_set(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3618 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3619 if (before(seq
, tp
->rcv_nxt
))
3620 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT
);
3622 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT
);
3624 tp
->rx_opt
.dsack
= 1;
3625 tp
->duplicate_sack
[0].start_seq
= seq
;
3626 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3627 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ 1, 4 - tp
->rx_opt
.tstamp_ok
);
3631 static void tcp_dsack_extend(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3633 if (!tp
->rx_opt
.dsack
)
3634 tcp_dsack_set(tp
, seq
, end_seq
);
3636 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3639 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3641 struct tcp_sock
*tp
= tcp_sk(sk
);
3643 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3644 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3645 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3646 tcp_enter_quickack_mode(sk
);
3648 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3649 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3651 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3652 end_seq
= tp
->rcv_nxt
;
3653 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3660 /* These routines update the SACK block as out-of-order packets arrive or
3661 * in-order packets close up the sequence space.
3663 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3666 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3667 struct tcp_sack_block
*swalk
= sp
+1;
3669 /* See if the recent change to the first SACK eats into
3670 * or hits the sequence space of other SACK blocks, if so coalesce.
3672 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
; ) {
3673 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3676 /* Zap SWALK, by moving every further SACK up by one slot.
3677 * Decrease num_sacks.
3679 tp
->rx_opt
.num_sacks
--;
3680 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3681 for (i
=this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3685 this_sack
++, swalk
++;
3689 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
, struct tcp_sack_block
*sack2
)
3693 tmp
= sack1
->start_seq
;
3694 sack1
->start_seq
= sack2
->start_seq
;
3695 sack2
->start_seq
= tmp
;
3697 tmp
= sack1
->end_seq
;
3698 sack1
->end_seq
= sack2
->end_seq
;
3699 sack2
->end_seq
= tmp
;
3702 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3704 struct tcp_sock
*tp
= tcp_sk(sk
);
3705 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3706 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3712 for (this_sack
=0; this_sack
<cur_sacks
; this_sack
++, sp
++) {
3713 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3714 /* Rotate this_sack to the first one. */
3715 for (; this_sack
>0; this_sack
--, sp
--)
3716 tcp_sack_swap(sp
, sp
-1);
3718 tcp_sack_maybe_coalesce(tp
);
3723 /* Could not find an adjacent existing SACK, build a new one,
3724 * put it at the front, and shift everyone else down. We
3725 * always know there is at least one SACK present already here.
3727 * If the sack array is full, forget about the last one.
3729 if (this_sack
>= 4) {
3731 tp
->rx_opt
.num_sacks
--;
3734 for (; this_sack
> 0; this_sack
--, sp
--)
3738 /* Build the new head SACK, and we're done. */
3739 sp
->start_seq
= seq
;
3740 sp
->end_seq
= end_seq
;
3741 tp
->rx_opt
.num_sacks
++;
3742 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3745 /* RCV.NXT advances, some SACKs should be eaten. */
3747 static void tcp_sack_remove(struct tcp_sock
*tp
)
3749 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3750 int num_sacks
= tp
->rx_opt
.num_sacks
;
3753 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3754 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3755 tp
->rx_opt
.num_sacks
= 0;
3756 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3760 for (this_sack
= 0; this_sack
< num_sacks
; ) {
3761 /* Check if the start of the sack is covered by RCV.NXT. */
3762 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3765 /* RCV.NXT must cover all the block! */
3766 BUG_TRAP(!before(tp
->rcv_nxt
, sp
->end_seq
));
3768 /* Zap this SACK, by moving forward any other SACKS. */
3769 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3770 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3777 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3778 tp
->rx_opt
.num_sacks
= num_sacks
;
3779 tp
->rx_opt
.eff_sacks
= min(tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
, 4 - tp
->rx_opt
.tstamp_ok
);
3783 /* This one checks to see if we can put data from the
3784 * out_of_order queue into the receive_queue.
3786 static void tcp_ofo_queue(struct sock
*sk
)
3788 struct tcp_sock
*tp
= tcp_sk(sk
);
3789 __u32 dsack_high
= tp
->rcv_nxt
;
3790 struct sk_buff
*skb
;
3792 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3793 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3796 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3797 __u32 dsack
= dsack_high
;
3798 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3799 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3800 tcp_dsack_extend(tp
, TCP_SKB_CB(skb
)->seq
, dsack
);
3803 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3804 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3805 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3809 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3810 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3811 TCP_SKB_CB(skb
)->end_seq
);
3813 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3814 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3815 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3816 if (tcp_hdr(skb
)->fin
)
3817 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3821 static int tcp_prune_queue(struct sock
*sk
);
3823 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3825 struct tcphdr
*th
= tcp_hdr(skb
);
3826 struct tcp_sock
*tp
= tcp_sk(sk
);
3829 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3832 __skb_pull(skb
, th
->doff
*4);
3834 TCP_ECN_accept_cwr(tp
, skb
);
3836 if (tp
->rx_opt
.dsack
) {
3837 tp
->rx_opt
.dsack
= 0;
3838 tp
->rx_opt
.eff_sacks
= min_t(unsigned int, tp
->rx_opt
.num_sacks
,
3839 4 - tp
->rx_opt
.tstamp_ok
);
3842 /* Queue data for delivery to the user.
3843 * Packets in sequence go to the receive queue.
3844 * Out of sequence packets to the out_of_order_queue.
3846 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3847 if (tcp_receive_window(tp
) == 0)
3850 /* Ok. In sequence. In window. */
3851 if (tp
->ucopy
.task
== current
&&
3852 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3853 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3854 int chunk
= min_t(unsigned int, skb
->len
,
3857 __set_current_state(TASK_RUNNING
);
3860 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3861 tp
->ucopy
.len
-= chunk
;
3862 tp
->copied_seq
+= chunk
;
3863 eaten
= (chunk
== skb
->len
&& !th
->fin
);
3864 tcp_rcv_space_adjust(sk
);
3872 (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3873 !sk_stream_rmem_schedule(sk
, skb
))) {
3874 if (tcp_prune_queue(sk
) < 0 ||
3875 !sk_stream_rmem_schedule(sk
, skb
))
3878 sk_stream_set_owner_r(skb
, sk
);
3879 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3881 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3883 tcp_event_data_recv(sk
, skb
);
3885 tcp_fin(skb
, sk
, th
);
3887 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
3890 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3891 * gap in queue is filled.
3893 if (skb_queue_empty(&tp
->out_of_order_queue
))
3894 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
3897 if (tp
->rx_opt
.num_sacks
)
3898 tcp_sack_remove(tp
);
3900 tcp_fast_path_check(sk
);
3904 else if (!sock_flag(sk
, SOCK_DEAD
))
3905 sk
->sk_data_ready(sk
, 0);
3909 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3910 /* A retransmit, 2nd most common case. Force an immediate ack. */
3911 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST
);
3912 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3915 tcp_enter_quickack_mode(sk
);
3916 inet_csk_schedule_ack(sk
);
3922 /* Out of window. F.e. zero window probe. */
3923 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
3926 tcp_enter_quickack_mode(sk
);
3928 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3929 /* Partial packet, seq < rcv_next < end_seq */
3930 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
3931 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3932 TCP_SKB_CB(skb
)->end_seq
);
3934 tcp_dsack_set(tp
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
3936 /* If window is closed, drop tail of packet. But after
3937 * remembering D-SACK for its head made in previous line.
3939 if (!tcp_receive_window(tp
))
3944 TCP_ECN_check_ce(tp
, skb
);
3946 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3947 !sk_stream_rmem_schedule(sk
, skb
)) {
3948 if (tcp_prune_queue(sk
) < 0 ||
3949 !sk_stream_rmem_schedule(sk
, skb
))
3953 /* Disable header prediction. */
3955 inet_csk_schedule_ack(sk
);
3957 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
3958 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
3960 sk_stream_set_owner_r(skb
, sk
);
3962 if (!skb_peek(&tp
->out_of_order_queue
)) {
3963 /* Initial out of order segment, build 1 SACK. */
3964 if (tcp_is_sack(tp
)) {
3965 tp
->rx_opt
.num_sacks
= 1;
3966 tp
->rx_opt
.dsack
= 0;
3967 tp
->rx_opt
.eff_sacks
= 1;
3968 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
3969 tp
->selective_acks
[0].end_seq
=
3970 TCP_SKB_CB(skb
)->end_seq
;
3972 __skb_queue_head(&tp
->out_of_order_queue
,skb
);
3974 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
3975 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3976 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3978 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
3979 __skb_append(skb1
, skb
, &tp
->out_of_order_queue
);
3981 if (!tp
->rx_opt
.num_sacks
||
3982 tp
->selective_acks
[0].end_seq
!= seq
)
3985 /* Common case: data arrive in order after hole. */
3986 tp
->selective_acks
[0].end_seq
= end_seq
;
3990 /* Find place to insert this segment. */
3992 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
3994 } while ((skb1
= skb1
->prev
) !=
3995 (struct sk_buff
*)&tp
->out_of_order_queue
);
3997 /* Do skb overlap to previous one? */
3998 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
3999 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4000 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4001 /* All the bits are present. Drop. */
4003 tcp_dsack_set(tp
, seq
, end_seq
);
4006 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4007 /* Partial overlap. */
4008 tcp_dsack_set(tp
, seq
, TCP_SKB_CB(skb1
)->end_seq
);
4013 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
4015 /* And clean segments covered by new one as whole. */
4016 while ((skb1
= skb
->next
) !=
4017 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4018 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4019 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4020 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, end_seq
);
4023 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4024 tcp_dsack_extend(tp
, TCP_SKB_CB(skb1
)->seq
, TCP_SKB_CB(skb1
)->end_seq
);
4029 if (tcp_is_sack(tp
))
4030 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4034 /* Collapse contiguous sequence of skbs head..tail with
4035 * sequence numbers start..end.
4036 * Segments with FIN/SYN are not collapsed (only because this
4040 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4041 struct sk_buff
*head
, struct sk_buff
*tail
,
4044 struct sk_buff
*skb
;
4046 /* First, check that queue is collapsible and find
4047 * the point where collapsing can be useful. */
4048 for (skb
= head
; skb
!= tail
; ) {
4049 /* No new bits? It is possible on ofo queue. */
4050 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4051 struct sk_buff
*next
= skb
->next
;
4052 __skb_unlink(skb
, list
);
4054 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
4059 /* The first skb to collapse is:
4061 * - bloated or contains data before "start" or
4062 * overlaps to the next one.
4064 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4065 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4066 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4067 (skb
->next
!= tail
&&
4068 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4071 /* Decided to skip this, advance start seq. */
4072 start
= TCP_SKB_CB(skb
)->end_seq
;
4075 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4078 while (before(start
, end
)) {
4079 struct sk_buff
*nskb
;
4080 unsigned int header
= skb_headroom(skb
);
4081 int copy
= SKB_MAX_ORDER(header
, 0);
4083 /* Too big header? This can happen with IPv6. */
4086 if (end
-start
< copy
)
4088 nskb
= alloc_skb(copy
+header
, GFP_ATOMIC
);
4092 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4093 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4095 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4097 skb_reserve(nskb
, header
);
4098 memcpy(nskb
->head
, skb
->head
, header
);
4099 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4100 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4101 __skb_insert(nskb
, skb
->prev
, skb
, list
);
4102 sk_stream_set_owner_r(nskb
, sk
);
4104 /* Copy data, releasing collapsed skbs. */
4106 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4107 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4111 size
= min(copy
, size
);
4112 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4114 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4118 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4119 struct sk_buff
*next
= skb
->next
;
4120 __skb_unlink(skb
, list
);
4122 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED
);
4125 tcp_hdr(skb
)->syn
||
4133 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4134 * and tcp_collapse() them until all the queue is collapsed.
4136 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4138 struct tcp_sock
*tp
= tcp_sk(sk
);
4139 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4140 struct sk_buff
*head
;
4146 start
= TCP_SKB_CB(skb
)->seq
;
4147 end
= TCP_SKB_CB(skb
)->end_seq
;
4153 /* Segment is terminated when we see gap or when
4154 * we are at the end of all the queue. */
4155 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4156 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4157 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4158 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4159 head
, skb
, start
, end
);
4161 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4163 /* Start new segment */
4164 start
= TCP_SKB_CB(skb
)->seq
;
4165 end
= TCP_SKB_CB(skb
)->end_seq
;
4167 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4168 start
= TCP_SKB_CB(skb
)->seq
;
4169 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4170 end
= TCP_SKB_CB(skb
)->end_seq
;
4175 /* Reduce allocated memory if we can, trying to get
4176 * the socket within its memory limits again.
4178 * Return less than zero if we should start dropping frames
4179 * until the socket owning process reads some of the data
4180 * to stabilize the situation.
4182 static int tcp_prune_queue(struct sock
*sk
)
4184 struct tcp_sock
*tp
= tcp_sk(sk
);
4186 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4188 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED
);
4190 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4191 tcp_clamp_window(sk
);
4192 else if (tcp_memory_pressure
)
4193 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4195 tcp_collapse_ofo_queue(sk
);
4196 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4197 sk
->sk_receive_queue
.next
,
4198 (struct sk_buff
*)&sk
->sk_receive_queue
,
4199 tp
->copied_seq
, tp
->rcv_nxt
);
4200 sk_stream_mem_reclaim(sk
);
4202 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4205 /* Collapsing did not help, destructive actions follow.
4206 * This must not ever occur. */
4208 /* First, purge the out_of_order queue. */
4209 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4210 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED
);
4211 __skb_queue_purge(&tp
->out_of_order_queue
);
4213 /* Reset SACK state. A conforming SACK implementation will
4214 * do the same at a timeout based retransmit. When a connection
4215 * is in a sad state like this, we care only about integrity
4216 * of the connection not performance.
4218 if (tcp_is_sack(tp
))
4219 tcp_sack_reset(&tp
->rx_opt
);
4220 sk_stream_mem_reclaim(sk
);
4223 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4226 /* If we are really being abused, tell the caller to silently
4227 * drop receive data on the floor. It will get retransmitted
4228 * and hopefully then we'll have sufficient space.
4230 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED
);
4232 /* Massive buffer overcommit. */
4238 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4239 * As additional protections, we do not touch cwnd in retransmission phases,
4240 * and if application hit its sndbuf limit recently.
4242 void tcp_cwnd_application_limited(struct sock
*sk
)
4244 struct tcp_sock
*tp
= tcp_sk(sk
);
4246 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4247 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4248 /* Limited by application or receiver window. */
4249 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4250 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4251 if (win_used
< tp
->snd_cwnd
) {
4252 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4253 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4255 tp
->snd_cwnd_used
= 0;
4257 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4260 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4262 struct tcp_sock
*tp
= tcp_sk(sk
);
4264 /* If the user specified a specific send buffer setting, do
4267 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4270 /* If we are under global TCP memory pressure, do not expand. */
4271 if (tcp_memory_pressure
)
4274 /* If we are under soft global TCP memory pressure, do not expand. */
4275 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4278 /* If we filled the congestion window, do not expand. */
4279 if (tp
->packets_out
>= tp
->snd_cwnd
)
4285 /* When incoming ACK allowed to free some skb from write_queue,
4286 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4287 * on the exit from tcp input handler.
4289 * PROBLEM: sndbuf expansion does not work well with largesend.
4291 static void tcp_new_space(struct sock
*sk
)
4293 struct tcp_sock
*tp
= tcp_sk(sk
);
4295 if (tcp_should_expand_sndbuf(sk
)) {
4296 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4297 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4298 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4299 tp
->reordering
+ 1);
4300 sndmem
*= 2*demanded
;
4301 if (sndmem
> sk
->sk_sndbuf
)
4302 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4303 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4306 sk
->sk_write_space(sk
);
4309 static void tcp_check_space(struct sock
*sk
)
4311 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4312 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4313 if (sk
->sk_socket
&&
4314 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4319 static inline void tcp_data_snd_check(struct sock
*sk
)
4321 tcp_push_pending_frames(sk
);
4322 tcp_check_space(sk
);
4326 * Check if sending an ack is needed.
4328 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4330 struct tcp_sock
*tp
= tcp_sk(sk
);
4332 /* More than one full frame received... */
4333 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4334 /* ... and right edge of window advances far enough.
4335 * (tcp_recvmsg() will send ACK otherwise). Or...
4337 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4338 /* We ACK each frame or... */
4339 tcp_in_quickack_mode(sk
) ||
4340 /* We have out of order data. */
4342 skb_peek(&tp
->out_of_order_queue
))) {
4343 /* Then ack it now */
4346 /* Else, send delayed ack. */
4347 tcp_send_delayed_ack(sk
);
4351 static inline void tcp_ack_snd_check(struct sock
*sk
)
4353 if (!inet_csk_ack_scheduled(sk
)) {
4354 /* We sent a data segment already. */
4357 __tcp_ack_snd_check(sk
, 1);
4361 * This routine is only called when we have urgent data
4362 * signaled. Its the 'slow' part of tcp_urg. It could be
4363 * moved inline now as tcp_urg is only called from one
4364 * place. We handle URGent data wrong. We have to - as
4365 * BSD still doesn't use the correction from RFC961.
4366 * For 1003.1g we should support a new option TCP_STDURG to permit
4367 * either form (or just set the sysctl tcp_stdurg).
4370 static void tcp_check_urg(struct sock
* sk
, struct tcphdr
* th
)
4372 struct tcp_sock
*tp
= tcp_sk(sk
);
4373 u32 ptr
= ntohs(th
->urg_ptr
);
4375 if (ptr
&& !sysctl_tcp_stdurg
)
4377 ptr
+= ntohl(th
->seq
);
4379 /* Ignore urgent data that we've already seen and read. */
4380 if (after(tp
->copied_seq
, ptr
))
4383 /* Do not replay urg ptr.
4385 * NOTE: interesting situation not covered by specs.
4386 * Misbehaving sender may send urg ptr, pointing to segment,
4387 * which we already have in ofo queue. We are not able to fetch
4388 * such data and will stay in TCP_URG_NOTYET until will be eaten
4389 * by recvmsg(). Seems, we are not obliged to handle such wicked
4390 * situations. But it is worth to think about possibility of some
4391 * DoSes using some hypothetical application level deadlock.
4393 if (before(ptr
, tp
->rcv_nxt
))
4396 /* Do we already have a newer (or duplicate) urgent pointer? */
4397 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4400 /* Tell the world about our new urgent pointer. */
4403 /* We may be adding urgent data when the last byte read was
4404 * urgent. To do this requires some care. We cannot just ignore
4405 * tp->copied_seq since we would read the last urgent byte again
4406 * as data, nor can we alter copied_seq until this data arrives
4407 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4409 * NOTE. Double Dutch. Rendering to plain English: author of comment
4410 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4411 * and expect that both A and B disappear from stream. This is _wrong_.
4412 * Though this happens in BSD with high probability, this is occasional.
4413 * Any application relying on this is buggy. Note also, that fix "works"
4414 * only in this artificial test. Insert some normal data between A and B and we will
4415 * decline of BSD again. Verdict: it is better to remove to trap
4418 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4419 !sock_flag(sk
, SOCK_URGINLINE
) &&
4420 tp
->copied_seq
!= tp
->rcv_nxt
) {
4421 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4423 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4424 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4429 tp
->urg_data
= TCP_URG_NOTYET
;
4432 /* Disable header prediction. */
4436 /* This is the 'fast' part of urgent handling. */
4437 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4439 struct tcp_sock
*tp
= tcp_sk(sk
);
4441 /* Check if we get a new urgent pointer - normally not. */
4443 tcp_check_urg(sk
,th
);
4445 /* Do we wait for any urgent data? - normally not... */
4446 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4447 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4450 /* Is the urgent pointer pointing into this packet? */
4451 if (ptr
< skb
->len
) {
4453 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4455 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4456 if (!sock_flag(sk
, SOCK_DEAD
))
4457 sk
->sk_data_ready(sk
, 0);
4462 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4464 struct tcp_sock
*tp
= tcp_sk(sk
);
4465 int chunk
= skb
->len
- hlen
;
4469 if (skb_csum_unnecessary(skb
))
4470 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4472 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4476 tp
->ucopy
.len
-= chunk
;
4477 tp
->copied_seq
+= chunk
;
4478 tcp_rcv_space_adjust(sk
);
4485 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4489 if (sock_owned_by_user(sk
)) {
4491 result
= __tcp_checksum_complete(skb
);
4494 result
= __tcp_checksum_complete(skb
);
4499 static inline int tcp_checksum_complete_user(struct sock
*sk
, struct sk_buff
*skb
)
4501 return !skb_csum_unnecessary(skb
) &&
4502 __tcp_checksum_complete_user(sk
, skb
);
4505 #ifdef CONFIG_NET_DMA
4506 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4508 struct tcp_sock
*tp
= tcp_sk(sk
);
4509 int chunk
= skb
->len
- hlen
;
4511 int copied_early
= 0;
4513 if (tp
->ucopy
.wakeup
)
4516 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4517 tp
->ucopy
.dma_chan
= get_softnet_dma();
4519 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4521 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4522 skb
, hlen
, tp
->ucopy
.iov
, chunk
, tp
->ucopy
.pinned_list
);
4527 tp
->ucopy
.dma_cookie
= dma_cookie
;
4530 tp
->ucopy
.len
-= chunk
;
4531 tp
->copied_seq
+= chunk
;
4532 tcp_rcv_space_adjust(sk
);
4534 if ((tp
->ucopy
.len
== 0) ||
4535 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4536 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4537 tp
->ucopy
.wakeup
= 1;
4538 sk
->sk_data_ready(sk
, 0);
4540 } else if (chunk
> 0) {
4541 tp
->ucopy
.wakeup
= 1;
4542 sk
->sk_data_ready(sk
, 0);
4545 return copied_early
;
4547 #endif /* CONFIG_NET_DMA */
4550 * TCP receive function for the ESTABLISHED state.
4552 * It is split into a fast path and a slow path. The fast path is
4554 * - A zero window was announced from us - zero window probing
4555 * is only handled properly in the slow path.
4556 * - Out of order segments arrived.
4557 * - Urgent data is expected.
4558 * - There is no buffer space left
4559 * - Unexpected TCP flags/window values/header lengths are received
4560 * (detected by checking the TCP header against pred_flags)
4561 * - Data is sent in both directions. Fast path only supports pure senders
4562 * or pure receivers (this means either the sequence number or the ack
4563 * value must stay constant)
4564 * - Unexpected TCP option.
4566 * When these conditions are not satisfied it drops into a standard
4567 * receive procedure patterned after RFC793 to handle all cases.
4568 * The first three cases are guaranteed by proper pred_flags setting,
4569 * the rest is checked inline. Fast processing is turned on in
4570 * tcp_data_queue when everything is OK.
4572 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4573 struct tcphdr
*th
, unsigned len
)
4575 struct tcp_sock
*tp
= tcp_sk(sk
);
4578 * Header prediction.
4579 * The code loosely follows the one in the famous
4580 * "30 instruction TCP receive" Van Jacobson mail.
4582 * Van's trick is to deposit buffers into socket queue
4583 * on a device interrupt, to call tcp_recv function
4584 * on the receive process context and checksum and copy
4585 * the buffer to user space. smart...
4587 * Our current scheme is not silly either but we take the
4588 * extra cost of the net_bh soft interrupt processing...
4589 * We do checksum and copy also but from device to kernel.
4592 tp
->rx_opt
.saw_tstamp
= 0;
4594 /* pred_flags is 0xS?10 << 16 + snd_wnd
4595 * if header_prediction is to be made
4596 * 'S' will always be tp->tcp_header_len >> 2
4597 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4598 * turn it off (when there are holes in the receive
4599 * space for instance)
4600 * PSH flag is ignored.
4603 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4604 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4605 int tcp_header_len
= tp
->tcp_header_len
;
4607 /* Timestamp header prediction: tcp_header_len
4608 * is automatically equal to th->doff*4 due to pred_flags
4612 /* Check timestamp */
4613 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4614 __be32
*ptr
= (__be32
*)(th
+ 1);
4616 /* No? Slow path! */
4617 if (*ptr
!= htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
4618 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
))
4621 tp
->rx_opt
.saw_tstamp
= 1;
4623 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
4625 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
4627 /* If PAWS failed, check it more carefully in slow path */
4628 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4631 /* DO NOT update ts_recent here, if checksum fails
4632 * and timestamp was corrupted part, it will result
4633 * in a hung connection since we will drop all
4634 * future packets due to the PAWS test.
4638 if (len
<= tcp_header_len
) {
4639 /* Bulk data transfer: sender */
4640 if (len
== tcp_header_len
) {
4641 /* Predicted packet is in window by definition.
4642 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4643 * Hence, check seq<=rcv_wup reduces to:
4645 if (tcp_header_len
==
4646 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4647 tp
->rcv_nxt
== tp
->rcv_wup
)
4648 tcp_store_ts_recent(tp
);
4650 /* We know that such packets are checksummed
4653 tcp_ack(sk
, skb
, 0);
4655 tcp_data_snd_check(sk
);
4657 } else { /* Header too small */
4658 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4663 int copied_early
= 0;
4665 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4666 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4667 #ifdef CONFIG_NET_DMA
4668 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4673 if (tp
->ucopy
.task
== current
&& sock_owned_by_user(sk
) && !copied_early
) {
4674 __set_current_state(TASK_RUNNING
);
4676 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4680 /* Predicted packet is in window by definition.
4681 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4682 * Hence, check seq<=rcv_wup reduces to:
4684 if (tcp_header_len
==
4685 (sizeof(struct tcphdr
) +
4686 TCPOLEN_TSTAMP_ALIGNED
) &&
4687 tp
->rcv_nxt
== tp
->rcv_wup
)
4688 tcp_store_ts_recent(tp
);
4690 tcp_rcv_rtt_measure_ts(sk
, skb
);
4692 __skb_pull(skb
, tcp_header_len
);
4693 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4694 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER
);
4697 tcp_cleanup_rbuf(sk
, skb
->len
);
4700 if (tcp_checksum_complete_user(sk
, skb
))
4703 /* Predicted packet is in window by definition.
4704 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4705 * Hence, check seq<=rcv_wup reduces to:
4707 if (tcp_header_len
==
4708 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4709 tp
->rcv_nxt
== tp
->rcv_wup
)
4710 tcp_store_ts_recent(tp
);
4712 tcp_rcv_rtt_measure_ts(sk
, skb
);
4714 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4717 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS
);
4719 /* Bulk data transfer: receiver */
4720 __skb_pull(skb
,tcp_header_len
);
4721 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4722 sk_stream_set_owner_r(skb
, sk
);
4723 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4726 tcp_event_data_recv(sk
, skb
);
4728 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4729 /* Well, only one small jumplet in fast path... */
4730 tcp_ack(sk
, skb
, FLAG_DATA
);
4731 tcp_data_snd_check(sk
);
4732 if (!inet_csk_ack_scheduled(sk
))
4736 __tcp_ack_snd_check(sk
, 0);
4738 #ifdef CONFIG_NET_DMA
4740 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4746 sk
->sk_data_ready(sk
, 0);
4752 if (len
< (th
->doff
<<2) || tcp_checksum_complete_user(sk
, skb
))
4756 * RFC1323: H1. Apply PAWS check first.
4758 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4759 tcp_paws_discard(sk
, skb
)) {
4761 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
4762 tcp_send_dupack(sk
, skb
);
4765 /* Resets are accepted even if PAWS failed.
4767 ts_recent update must be made after we are sure
4768 that the packet is in window.
4773 * Standard slow path.
4776 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4777 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4778 * (RST) segments are validated by checking their SEQ-fields."
4779 * And page 69: "If an incoming segment is not acceptable,
4780 * an acknowledgment should be sent in reply (unless the RST bit
4781 * is set, if so drop the segment and return)".
4784 tcp_send_dupack(sk
, skb
);
4793 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4795 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4796 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4797 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
4804 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4806 tcp_rcv_rtt_measure_ts(sk
, skb
);
4808 /* Process urgent data. */
4809 tcp_urg(sk
, skb
, th
);
4811 /* step 7: process the segment text */
4812 tcp_data_queue(sk
, skb
);
4814 tcp_data_snd_check(sk
);
4815 tcp_ack_snd_check(sk
);
4819 TCP_INC_STATS_BH(TCP_MIB_INERRS
);
4826 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4827 struct tcphdr
*th
, unsigned len
)
4829 struct tcp_sock
*tp
= tcp_sk(sk
);
4830 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4831 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
4833 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
4837 * "If the state is SYN-SENT then
4838 * first check the ACK bit
4839 * If the ACK bit is set
4840 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4841 * a reset (unless the RST bit is set, if so drop
4842 * the segment and return)"
4844 * We do not send data with SYN, so that RFC-correct
4847 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
4848 goto reset_and_undo
;
4850 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
4851 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
4853 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED
);
4854 goto reset_and_undo
;
4857 /* Now ACK is acceptable.
4859 * "If the RST bit is set
4860 * If the ACK was acceptable then signal the user "error:
4861 * connection reset", drop the segment, enter CLOSED state,
4862 * delete TCB, and return."
4871 * "fifth, if neither of the SYN or RST bits is set then
4872 * drop the segment and return."
4878 goto discard_and_undo
;
4881 * "If the SYN bit is on ...
4882 * are acceptable then ...
4883 * (our SYN has been ACKed), change the connection
4884 * state to ESTABLISHED..."
4887 TCP_ECN_rcv_synack(tp
, th
);
4889 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
4890 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4892 /* Ok.. it's good. Set up sequence numbers and
4893 * move to established.
4895 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
4896 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
4898 /* RFC1323: The window in SYN & SYN/ACK segments is
4901 tp
->snd_wnd
= ntohs(th
->window
);
4902 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
4904 if (!tp
->rx_opt
.wscale_ok
) {
4905 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
4906 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
4909 if (tp
->rx_opt
.saw_tstamp
) {
4910 tp
->rx_opt
.tstamp_ok
= 1;
4911 tp
->tcp_header_len
=
4912 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
4913 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
4914 tcp_store_ts_recent(tp
);
4916 tp
->tcp_header_len
= sizeof(struct tcphdr
);
4919 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
4920 tcp_enable_fack(tp
);
4923 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
4924 tcp_initialize_rcv_mss(sk
);
4926 /* Remember, tcp_poll() does not lock socket!
4927 * Change state from SYN-SENT only after copied_seq
4928 * is initialized. */
4929 tp
->copied_seq
= tp
->rcv_nxt
;
4931 tcp_set_state(sk
, TCP_ESTABLISHED
);
4933 security_inet_conn_established(sk
, skb
);
4935 /* Make sure socket is routed, for correct metrics. */
4936 icsk
->icsk_af_ops
->rebuild_header(sk
);
4938 tcp_init_metrics(sk
);
4940 tcp_init_congestion_control(sk
);
4942 /* Prevent spurious tcp_cwnd_restart() on first data
4945 tp
->lsndtime
= tcp_time_stamp
;
4947 tcp_init_buffer_space(sk
);
4949 if (sock_flag(sk
, SOCK_KEEPOPEN
))
4950 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
4952 if (!tp
->rx_opt
.snd_wscale
)
4953 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
4957 if (!sock_flag(sk
, SOCK_DEAD
)) {
4958 sk
->sk_state_change(sk
);
4959 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
4962 if (sk
->sk_write_pending
||
4963 icsk
->icsk_accept_queue
.rskq_defer_accept
||
4964 icsk
->icsk_ack
.pingpong
) {
4965 /* Save one ACK. Data will be ready after
4966 * several ticks, if write_pending is set.
4968 * It may be deleted, but with this feature tcpdumps
4969 * look so _wonderfully_ clever, that I was not able
4970 * to stand against the temptation 8) --ANK
4972 inet_csk_schedule_ack(sk
);
4973 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
4974 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
4975 tcp_incr_quickack(sk
);
4976 tcp_enter_quickack_mode(sk
);
4977 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
4978 TCP_DELACK_MAX
, TCP_RTO_MAX
);
4989 /* No ACK in the segment */
4993 * "If the RST bit is set
4995 * Otherwise (no ACK) drop the segment and return."
4998 goto discard_and_undo
;
5002 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&& tcp_paws_check(&tp
->rx_opt
, 0))
5003 goto discard_and_undo
;
5006 /* We see SYN without ACK. It is attempt of
5007 * simultaneous connect with crossed SYNs.
5008 * Particularly, it can be connect to self.
5010 tcp_set_state(sk
, TCP_SYN_RECV
);
5012 if (tp
->rx_opt
.saw_tstamp
) {
5013 tp
->rx_opt
.tstamp_ok
= 1;
5014 tcp_store_ts_recent(tp
);
5015 tp
->tcp_header_len
=
5016 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5018 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5021 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5022 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5024 /* RFC1323: The window in SYN & SYN/ACK segments is
5027 tp
->snd_wnd
= ntohs(th
->window
);
5028 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5029 tp
->max_window
= tp
->snd_wnd
;
5031 TCP_ECN_rcv_syn(tp
, th
);
5034 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5035 tcp_initialize_rcv_mss(sk
);
5038 tcp_send_synack(sk
);
5040 /* Note, we could accept data and URG from this segment.
5041 * There are no obstacles to make this.
5043 * However, if we ignore data in ACKless segments sometimes,
5044 * we have no reasons to accept it sometimes.
5045 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5046 * is not flawless. So, discard packet for sanity.
5047 * Uncomment this return to process the data.
5054 /* "fifth, if neither of the SYN or RST bits is set then
5055 * drop the segment and return."
5059 tcp_clear_options(&tp
->rx_opt
);
5060 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5064 tcp_clear_options(&tp
->rx_opt
);
5065 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5071 * This function implements the receiving procedure of RFC 793 for
5072 * all states except ESTABLISHED and TIME_WAIT.
5073 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5074 * address independent.
5077 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5078 struct tcphdr
*th
, unsigned len
)
5080 struct tcp_sock
*tp
= tcp_sk(sk
);
5081 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5084 tp
->rx_opt
.saw_tstamp
= 0;
5086 switch (sk
->sk_state
) {
5098 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5101 /* Now we have several options: In theory there is
5102 * nothing else in the frame. KA9Q has an option to
5103 * send data with the syn, BSD accepts data with the
5104 * syn up to the [to be] advertised window and
5105 * Solaris 2.1 gives you a protocol error. For now
5106 * we just ignore it, that fits the spec precisely
5107 * and avoids incompatibilities. It would be nice in
5108 * future to drop through and process the data.
5110 * Now that TTCP is starting to be used we ought to
5112 * But, this leaves one open to an easy denial of
5113 * service attack, and SYN cookies can't defend
5114 * against this problem. So, we drop the data
5115 * in the interest of security over speed unless
5116 * it's still in use.
5124 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5128 /* Do step6 onward by hand. */
5129 tcp_urg(sk
, skb
, th
);
5131 tcp_data_snd_check(sk
);
5135 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5136 tcp_paws_discard(sk
, skb
)) {
5138 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED
);
5139 tcp_send_dupack(sk
, skb
);
5142 /* Reset is accepted even if it did not pass PAWS. */
5145 /* step 1: check sequence number */
5146 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5148 tcp_send_dupack(sk
, skb
);
5152 /* step 2: check RST bit */
5158 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5160 /* step 3: check security and precedence [ignored] */
5164 * Check for a SYN in window.
5166 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5167 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN
);
5172 /* step 5: check the ACK field */
5174 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5176 switch (sk
->sk_state
) {
5179 tp
->copied_seq
= tp
->rcv_nxt
;
5181 tcp_set_state(sk
, TCP_ESTABLISHED
);
5182 sk
->sk_state_change(sk
);
5184 /* Note, that this wakeup is only for marginal
5185 * crossed SYN case. Passively open sockets
5186 * are not waked up, because sk->sk_sleep ==
5187 * NULL and sk->sk_socket == NULL.
5191 SOCK_WAKE_IO
, POLL_OUT
);
5193 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5194 tp
->snd_wnd
= ntohs(th
->window
) <<
5195 tp
->rx_opt
.snd_wscale
;
5196 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5197 TCP_SKB_CB(skb
)->seq
);
5199 /* tcp_ack considers this ACK as duplicate
5200 * and does not calculate rtt.
5201 * Fix it at least with timestamps.
5203 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5205 tcp_ack_saw_tstamp(sk
, 0);
5207 if (tp
->rx_opt
.tstamp_ok
)
5208 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5210 /* Make sure socket is routed, for
5213 icsk
->icsk_af_ops
->rebuild_header(sk
);
5215 tcp_init_metrics(sk
);
5217 tcp_init_congestion_control(sk
);
5219 /* Prevent spurious tcp_cwnd_restart() on
5220 * first data packet.
5222 tp
->lsndtime
= tcp_time_stamp
;
5225 tcp_initialize_rcv_mss(sk
);
5226 tcp_init_buffer_space(sk
);
5227 tcp_fast_path_on(tp
);
5234 if (tp
->snd_una
== tp
->write_seq
) {
5235 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5236 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5237 dst_confirm(sk
->sk_dst_cache
);
5239 if (!sock_flag(sk
, SOCK_DEAD
))
5240 /* Wake up lingering close() */
5241 sk
->sk_state_change(sk
);
5245 if (tp
->linger2
< 0 ||
5246 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5247 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5249 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5253 tmo
= tcp_fin_time(sk
);
5254 if (tmo
> TCP_TIMEWAIT_LEN
) {
5255 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5256 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5257 /* Bad case. We could lose such FIN otherwise.
5258 * It is not a big problem, but it looks confusing
5259 * and not so rare event. We still can lose it now,
5260 * if it spins in bh_lock_sock(), but it is really
5263 inet_csk_reset_keepalive_timer(sk
, tmo
);
5265 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5273 if (tp
->snd_una
== tp
->write_seq
) {
5274 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5280 if (tp
->snd_una
== tp
->write_seq
) {
5281 tcp_update_metrics(sk
);
5290 /* step 6: check the URG bit */
5291 tcp_urg(sk
, skb
, th
);
5293 /* step 7: process the segment text */
5294 switch (sk
->sk_state
) {
5295 case TCP_CLOSE_WAIT
:
5298 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5302 /* RFC 793 says to queue data in these states,
5303 * RFC 1122 says we MUST send a reset.
5304 * BSD 4.4 also does reset.
5306 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5307 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5308 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5309 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA
);
5315 case TCP_ESTABLISHED
:
5316 tcp_data_queue(sk
, skb
);
5321 /* tcp_data could move socket to TIME-WAIT */
5322 if (sk
->sk_state
!= TCP_CLOSE
) {
5323 tcp_data_snd_check(sk
);
5324 tcp_ack_snd_check(sk
);
5334 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5335 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5336 EXPORT_SYMBOL(tcp_parse_options
);
5337 EXPORT_SYMBOL(tcp_rcv_established
);
5338 EXPORT_SYMBOL(tcp_rcv_state_process
);
5339 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);