2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #include <linux/module.h>
66 #include <linux/sysctl.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71 #include <asm/unaligned.h>
72 #include <net/netdma.h>
74 int sysctl_tcp_timestamps __read_mostly
= 1;
75 int sysctl_tcp_window_scaling __read_mostly
= 1;
76 int sysctl_tcp_sack __read_mostly
= 1;
77 int sysctl_tcp_fack __read_mostly
= 1;
78 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
79 int sysctl_tcp_ecn __read_mostly
;
80 int sysctl_tcp_dsack __read_mostly
= 1;
81 int sysctl_tcp_app_win __read_mostly
= 31;
82 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
84 int sysctl_tcp_stdurg __read_mostly
;
85 int sysctl_tcp_rfc1337 __read_mostly
;
86 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
87 int sysctl_tcp_frto __read_mostly
= 2;
88 int sysctl_tcp_frto_response __read_mostly
;
89 int sysctl_tcp_nometrics_save __read_mostly
;
91 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
92 int sysctl_tcp_abc __read_mostly
;
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
104 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
105 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
106 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
107 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
118 /* Adapt the MSS value used to make delayed ack decision to the
121 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
123 struct inet_connection_sock
*icsk
= inet_csk(sk
);
124 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
127 icsk
->icsk_ack
.last_seg_size
= 0;
129 /* skb->len may jitter because of SACKs, even if peer
130 * sends good full-sized frames.
132 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
133 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
134 icsk
->icsk_ack
.rcv_mss
= len
;
136 /* Otherwise, we make more careful check taking into account,
137 * that SACKs block is variable.
139 * "len" is invariant segment length, including TCP header.
141 len
+= skb
->data
- skb_transport_header(skb
);
142 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
143 /* If PSH is not set, packet should be
144 * full sized, provided peer TCP is not badly broken.
145 * This observation (if it is correct 8)) allows
146 * to handle super-low mtu links fairly.
148 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
149 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
150 /* Subtract also invariant (if peer is RFC compliant),
151 * tcp header plus fixed timestamp option length.
152 * Resulting "len" is MSS free of SACK jitter.
154 len
-= tcp_sk(sk
)->tcp_header_len
;
155 icsk
->icsk_ack
.last_seg_size
= len
;
157 icsk
->icsk_ack
.rcv_mss
= len
;
161 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
162 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
163 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
167 static void tcp_incr_quickack(struct sock
*sk
)
169 struct inet_connection_sock
*icsk
= inet_csk(sk
);
170 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
174 if (quickacks
> icsk
->icsk_ack
.quick
)
175 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
178 void tcp_enter_quickack_mode(struct sock
*sk
)
180 struct inet_connection_sock
*icsk
= inet_csk(sk
);
181 tcp_incr_quickack(sk
);
182 icsk
->icsk_ack
.pingpong
= 0;
183 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
186 /* Send ACKs quickly, if "quick" count is not exhausted
187 * and the session is not interactive.
190 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
192 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
193 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
196 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
198 if (tp
->ecn_flags
& TCP_ECN_OK
)
199 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
202 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
204 if (tcp_hdr(skb
)->cwr
)
205 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
208 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
210 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
213 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
215 if (tp
->ecn_flags
& TCP_ECN_OK
) {
216 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
217 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
218 /* Funny extension: if ECT is not set on a segment,
219 * it is surely retransmit. It is not in ECN RFC,
220 * but Linux follows this rule. */
221 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
222 tcp_enter_quickack_mode((struct sock
*)tp
);
226 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
228 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
229 tp
->ecn_flags
&= ~TCP_ECN_OK
;
232 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
234 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
235 tp
->ecn_flags
&= ~TCP_ECN_OK
;
238 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
240 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
245 /* Buffer size and advertised window tuning.
247 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
250 static void tcp_fixup_sndbuf(struct sock
*sk
)
252 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
253 sizeof(struct sk_buff
);
255 if (sk
->sk_sndbuf
< 3 * sndmem
)
256 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
259 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261 * All tcp_full_space() is split to two parts: "network" buffer, allocated
262 * forward and advertised in receiver window (tp->rcv_wnd) and
263 * "application buffer", required to isolate scheduling/application
264 * latencies from network.
265 * window_clamp is maximal advertised window. It can be less than
266 * tcp_full_space(), in this case tcp_full_space() - window_clamp
267 * is reserved for "application" buffer. The less window_clamp is
268 * the smoother our behaviour from viewpoint of network, but the lower
269 * throughput and the higher sensitivity of the connection to losses. 8)
271 * rcv_ssthresh is more strict window_clamp used at "slow start"
272 * phase to predict further behaviour of this connection.
273 * It is used for two goals:
274 * - to enforce header prediction at sender, even when application
275 * requires some significant "application buffer". It is check #1.
276 * - to prevent pruning of receive queue because of misprediction
277 * of receiver window. Check #2.
279 * The scheme does not work when sender sends good segments opening
280 * window and then starts to feed us spaghetti. But it should work
281 * in common situations. Otherwise, we have to rely on queue collapsing.
284 /* Slow part of check#2. */
285 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
287 struct tcp_sock
*tp
= tcp_sk(sk
);
289 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
290 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
292 while (tp
->rcv_ssthresh
<= window
) {
293 if (truesize
<= skb
->len
)
294 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
302 static void tcp_grow_window(struct sock
*sk
, struct sk_buff
*skb
)
304 struct tcp_sock
*tp
= tcp_sk(sk
);
307 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
308 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
309 !tcp_memory_pressure
) {
312 /* Check #2. Increase window, if skb with such overhead
313 * will fit to rcvbuf in future.
315 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
316 incr
= 2 * tp
->advmss
;
318 incr
= __tcp_grow_window(sk
, skb
);
321 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
323 inet_csk(sk
)->icsk_ack
.quick
|= 1;
328 /* 3. Tuning rcvbuf, when connection enters established state. */
330 static void tcp_fixup_rcvbuf(struct sock
*sk
)
332 struct tcp_sock
*tp
= tcp_sk(sk
);
333 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
335 /* Try to select rcvbuf so that 4 mss-sized segments
336 * will fit to window and corresponding skbs will fit to our rcvbuf.
337 * (was 3; 4 is minimum to allow fast retransmit to work.)
339 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
341 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
342 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
345 /* 4. Try to fixup all. It is made immediately after connection enters
348 static void tcp_init_buffer_space(struct sock
*sk
)
350 struct tcp_sock
*tp
= tcp_sk(sk
);
353 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
354 tcp_fixup_rcvbuf(sk
);
355 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
356 tcp_fixup_sndbuf(sk
);
358 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
360 maxwin
= tcp_full_space(sk
);
362 if (tp
->window_clamp
>= maxwin
) {
363 tp
->window_clamp
= maxwin
;
365 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
366 tp
->window_clamp
= max(maxwin
-
367 (maxwin
>> sysctl_tcp_app_win
),
371 /* Force reservation of one segment. */
372 if (sysctl_tcp_app_win
&&
373 tp
->window_clamp
> 2 * tp
->advmss
&&
374 tp
->window_clamp
+ tp
->advmss
> maxwin
)
375 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
377 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
378 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
381 /* 5. Recalculate window clamp after socket hit its memory bounds. */
382 static void tcp_clamp_window(struct sock
*sk
)
384 struct tcp_sock
*tp
= tcp_sk(sk
);
385 struct inet_connection_sock
*icsk
= inet_csk(sk
);
387 icsk
->icsk_ack
.quick
= 0;
389 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
390 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
391 !tcp_memory_pressure
&&
392 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
393 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
396 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
397 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
400 /* Initialize RCV_MSS value.
401 * RCV_MSS is an our guess about MSS used by the peer.
402 * We haven't any direct information about the MSS.
403 * It's better to underestimate the RCV_MSS rather than overestimate.
404 * Overestimations make us ACKing less frequently than needed.
405 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407 void tcp_initialize_rcv_mss(struct sock
*sk
)
409 struct tcp_sock
*tp
= tcp_sk(sk
);
410 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
412 hint
= min(hint
, tp
->rcv_wnd
/ 2);
413 hint
= min(hint
, TCP_MIN_RCVMSS
);
414 hint
= max(hint
, TCP_MIN_MSS
);
416 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
419 /* Receiver "autotuning" code.
421 * The algorithm for RTT estimation w/o timestamps is based on
422 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
423 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425 * More detail on this code can be found at
426 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
427 * though this reference is out of date. A new paper
430 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
432 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
438 if (new_sample
!= 0) {
439 /* If we sample in larger samples in the non-timestamp
440 * case, we could grossly overestimate the RTT especially
441 * with chatty applications or bulk transfer apps which
442 * are stalled on filesystem I/O.
444 * Also, since we are only going for a minimum in the
445 * non-timestamp case, we do not smooth things out
446 * else with timestamps disabled convergence takes too
450 m
-= (new_sample
>> 3);
452 } else if (m
< new_sample
)
455 /* No previous measure. */
459 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
460 tp
->rcv_rtt_est
.rtt
= new_sample
;
463 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
465 if (tp
->rcv_rtt_est
.time
== 0)
467 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
469 tcp_rcv_rtt_update(tp
, jiffies
- tp
->rcv_rtt_est
.time
, 1);
472 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
473 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
476 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
477 const struct sk_buff
*skb
)
479 struct tcp_sock
*tp
= tcp_sk(sk
);
480 if (tp
->rx_opt
.rcv_tsecr
&&
481 (TCP_SKB_CB(skb
)->end_seq
-
482 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
483 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
487 * This function should be called every time data is copied to user space.
488 * It calculates the appropriate TCP receive buffer space.
490 void tcp_rcv_space_adjust(struct sock
*sk
)
492 struct tcp_sock
*tp
= tcp_sk(sk
);
496 if (tp
->rcvq_space
.time
== 0)
499 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
500 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
503 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
505 space
= max(tp
->rcvq_space
.space
, space
);
507 if (tp
->rcvq_space
.space
!= space
) {
510 tp
->rcvq_space
.space
= space
;
512 if (sysctl_tcp_moderate_rcvbuf
&&
513 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
514 int new_clamp
= space
;
516 /* Receive space grows, normalize in order to
517 * take into account packet headers and sk_buff
518 * structure overhead.
523 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
524 16 + sizeof(struct sk_buff
));
525 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
528 space
= min(space
, sysctl_tcp_rmem
[2]);
529 if (space
> sk
->sk_rcvbuf
) {
530 sk
->sk_rcvbuf
= space
;
532 /* Make the window clamp follow along. */
533 tp
->window_clamp
= new_clamp
;
539 tp
->rcvq_space
.seq
= tp
->copied_seq
;
540 tp
->rcvq_space
.time
= tcp_time_stamp
;
543 /* There is something which you must keep in mind when you analyze the
544 * behavior of the tp->ato delayed ack timeout interval. When a
545 * connection starts up, we want to ack as quickly as possible. The
546 * problem is that "good" TCP's do slow start at the beginning of data
547 * transmission. The means that until we send the first few ACK's the
548 * sender will sit on his end and only queue most of his data, because
549 * he can only send snd_cwnd unacked packets at any given time. For
550 * each ACK we send, he increments snd_cwnd and transmits more of his
553 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
555 struct tcp_sock
*tp
= tcp_sk(sk
);
556 struct inet_connection_sock
*icsk
= inet_csk(sk
);
559 inet_csk_schedule_ack(sk
);
561 tcp_measure_rcv_mss(sk
, skb
);
563 tcp_rcv_rtt_measure(tp
);
565 now
= tcp_time_stamp
;
567 if (!icsk
->icsk_ack
.ato
) {
568 /* The _first_ data packet received, initialize
569 * delayed ACK engine.
571 tcp_incr_quickack(sk
);
572 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
574 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
576 if (m
<= TCP_ATO_MIN
/ 2) {
577 /* The fastest case is the first. */
578 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
579 } else if (m
< icsk
->icsk_ack
.ato
) {
580 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
581 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
582 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
583 } else if (m
> icsk
->icsk_rto
) {
584 /* Too long gap. Apparently sender failed to
585 * restart window, so that we send ACKs quickly.
587 tcp_incr_quickack(sk
);
591 icsk
->icsk_ack
.lrcvtime
= now
;
593 TCP_ECN_check_ce(tp
, skb
);
596 tcp_grow_window(sk
, skb
);
599 static u32
tcp_rto_min(struct sock
*sk
)
601 struct dst_entry
*dst
= __sk_dst_get(sk
);
602 u32 rto_min
= TCP_RTO_MIN
;
604 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
605 rto_min
= dst_metric_rtt(dst
, RTAX_RTO_MIN
);
609 /* Called to compute a smoothed rtt estimate. The data fed to this
610 * routine either comes from timestamps, or from segments that were
611 * known _not_ to have been retransmitted [see Karn/Partridge
612 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
613 * piece by Van Jacobson.
614 * NOTE: the next three routines used to be one big routine.
615 * To save cycles in the RFC 1323 implementation it was better to break
616 * it up into three procedures. -- erics
618 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
620 struct tcp_sock
*tp
= tcp_sk(sk
);
621 long m
= mrtt
; /* RTT */
623 /* The following amusing code comes from Jacobson's
624 * article in SIGCOMM '88. Note that rtt and mdev
625 * are scaled versions of rtt and mean deviation.
626 * This is designed to be as fast as possible
627 * m stands for "measurement".
629 * On a 1990 paper the rto value is changed to:
630 * RTO = rtt + 4 * mdev
632 * Funny. This algorithm seems to be very broken.
633 * These formulae increase RTO, when it should be decreased, increase
634 * too slowly, when it should be increased quickly, decrease too quickly
635 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
636 * does not matter how to _calculate_ it. Seems, it was trap
637 * that VJ failed to avoid. 8)
642 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
643 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
645 m
= -m
; /* m is now abs(error) */
646 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
647 /* This is similar to one of Eifel findings.
648 * Eifel blocks mdev updates when rtt decreases.
649 * This solution is a bit different: we use finer gain
650 * for mdev in this case (alpha*beta).
651 * Like Eifel it also prevents growth of rto,
652 * but also it limits too fast rto decreases,
653 * happening in pure Eifel.
658 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
660 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
661 if (tp
->mdev
> tp
->mdev_max
) {
662 tp
->mdev_max
= tp
->mdev
;
663 if (tp
->mdev_max
> tp
->rttvar
)
664 tp
->rttvar
= tp
->mdev_max
;
666 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
667 if (tp
->mdev_max
< tp
->rttvar
)
668 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
669 tp
->rtt_seq
= tp
->snd_nxt
;
670 tp
->mdev_max
= tcp_rto_min(sk
);
673 /* no previous measure. */
674 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
675 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
676 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
677 tp
->rtt_seq
= tp
->snd_nxt
;
681 /* Calculate rto without backoff. This is the second half of Van Jacobson's
682 * routine referred to above.
684 static inline void tcp_set_rto(struct sock
*sk
)
686 const struct tcp_sock
*tp
= tcp_sk(sk
);
687 /* Old crap is replaced with new one. 8)
690 * 1. If rtt variance happened to be less 50msec, it is hallucination.
691 * It cannot be less due to utterly erratic ACK generation made
692 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
693 * to do with delayed acks, because at cwnd>2 true delack timeout
694 * is invisible. Actually, Linux-2.4 also generates erratic
695 * ACKs in some circumstances.
697 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
699 /* 2. Fixups made earlier cannot be right.
700 * If we do not estimate RTO correctly without them,
701 * all the algo is pure shit and should be replaced
702 * with correct one. It is exactly, which we pretend to do.
706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707 * guarantees that rto is higher.
709 static inline void tcp_bound_rto(struct sock
*sk
)
711 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
712 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
715 /* Save metrics learned by this TCP session.
716 This function is called only, when TCP finishes successfully
717 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
719 void tcp_update_metrics(struct sock
*sk
)
721 struct tcp_sock
*tp
= tcp_sk(sk
);
722 struct dst_entry
*dst
= __sk_dst_get(sk
);
724 if (sysctl_tcp_nometrics_save
)
729 if (dst
&& (dst
->flags
& DST_HOST
)) {
730 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
734 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
735 /* This session failed to estimate rtt. Why?
736 * Probably, no packets returned in time.
739 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
740 dst
->metrics
[RTAX_RTT
- 1] = 0;
744 rtt
= dst_metric_rtt(dst
, RTAX_RTT
);
747 /* If newly calculated rtt larger than stored one,
748 * store new one. Otherwise, use EWMA. Remember,
749 * rtt overestimation is always better than underestimation.
751 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
753 set_dst_metric_rtt(dst
, RTAX_RTT
, tp
->srtt
);
755 set_dst_metric_rtt(dst
, RTAX_RTT
, rtt
- (m
>> 3));
758 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
763 /* Scale deviation to rttvar fixed point */
768 var
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
772 var
-= (var
- m
) >> 2;
774 set_dst_metric_rtt(dst
, RTAX_RTTVAR
, var
);
777 if (tp
->snd_ssthresh
>= 0xFFFF) {
778 /* Slow start still did not finish. */
779 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
780 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
781 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
782 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
783 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
784 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
785 dst
->metrics
[RTAX_CWND
- 1] = tp
->snd_cwnd
;
786 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
787 icsk
->icsk_ca_state
== TCP_CA_Open
) {
788 /* Cong. avoidance phase, cwnd is reliable. */
789 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
790 dst
->metrics
[RTAX_SSTHRESH
-1] =
791 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
792 if (!dst_metric_locked(dst
, RTAX_CWND
))
793 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_cwnd
) >> 1;
795 /* Else slow start did not finish, cwnd is non-sense,
796 ssthresh may be also invalid.
798 if (!dst_metric_locked(dst
, RTAX_CWND
))
799 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_ssthresh
) >> 1;
800 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
801 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
802 tp
->snd_ssthresh
> dst_metric(dst
, RTAX_SSTHRESH
))
803 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
806 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
807 if (dst_metric(dst
, RTAX_REORDERING
) < tp
->reordering
&&
808 tp
->reordering
!= sysctl_tcp_reordering
)
809 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
814 /* Numbers are taken from RFC3390.
816 * John Heffner states:
818 * The RFC specifies a window of no more than 4380 bytes
819 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
820 * is a bit misleading because they use a clamp at 4380 bytes
821 * rather than use a multiplier in the relevant range.
823 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
825 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
828 if (tp
->mss_cache
> 1460)
831 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
833 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
836 /* Set slow start threshold and cwnd not falling to slow start */
837 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
839 struct tcp_sock
*tp
= tcp_sk(sk
);
840 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
842 tp
->prior_ssthresh
= 0;
844 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
847 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
848 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
849 tcp_packets_in_flight(tp
) + 1U);
850 tp
->snd_cwnd_cnt
= 0;
851 tp
->high_seq
= tp
->snd_nxt
;
852 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
853 TCP_ECN_queue_cwr(tp
);
855 tcp_set_ca_state(sk
, TCP_CA_CWR
);
860 * Packet counting of FACK is based on in-order assumptions, therefore TCP
861 * disables it when reordering is detected
863 static void tcp_disable_fack(struct tcp_sock
*tp
)
865 /* RFC3517 uses different metric in lost marker => reset on change */
867 tp
->lost_skb_hint
= NULL
;
868 tp
->rx_opt
.sack_ok
&= ~2;
871 /* Take a notice that peer is sending D-SACKs */
872 static void tcp_dsack_seen(struct tcp_sock
*tp
)
874 tp
->rx_opt
.sack_ok
|= 4;
877 /* Initialize metrics on socket. */
879 static void tcp_init_metrics(struct sock
*sk
)
881 struct tcp_sock
*tp
= tcp_sk(sk
);
882 struct dst_entry
*dst
= __sk_dst_get(sk
);
889 if (dst_metric_locked(dst
, RTAX_CWND
))
890 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
891 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
892 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
893 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
894 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
896 if (dst_metric(dst
, RTAX_REORDERING
) &&
897 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
898 tcp_disable_fack(tp
);
899 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
902 if (dst_metric(dst
, RTAX_RTT
) == 0)
905 if (!tp
->srtt
&& dst_metric_rtt(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
908 /* Initial rtt is determined from SYN,SYN-ACK.
909 * The segment is small and rtt may appear much
910 * less than real one. Use per-dst memory
911 * to make it more realistic.
913 * A bit of theory. RTT is time passed after "normal" sized packet
914 * is sent until it is ACKed. In normal circumstances sending small
915 * packets force peer to delay ACKs and calculation is correct too.
916 * The algorithm is adaptive and, provided we follow specs, it
917 * NEVER underestimate RTT. BUT! If peer tries to make some clever
918 * tricks sort of "quick acks" for time long enough to decrease RTT
919 * to low value, and then abruptly stops to do it and starts to delay
920 * ACKs, wait for troubles.
922 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
923 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
924 tp
->rtt_seq
= tp
->snd_nxt
;
926 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
927 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
928 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
932 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
934 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
935 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
939 /* Play conservative. If timestamps are not
940 * supported, TCP will fail to recalculate correct
941 * rtt, if initial rto is too small. FORGET ALL AND RESET!
943 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
945 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
946 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
950 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
953 struct tcp_sock
*tp
= tcp_sk(sk
);
954 if (metric
> tp
->reordering
) {
957 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
959 /* This exciting event is worth to be remembered. 8) */
961 mib_idx
= LINUX_MIB_TCPTSREORDER
;
962 else if (tcp_is_reno(tp
))
963 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
964 else if (tcp_is_fack(tp
))
965 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
967 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
969 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
970 #if FASTRETRANS_DEBUG > 1
971 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
972 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
976 tp
->undo_marker
? tp
->undo_retrans
: 0);
978 tcp_disable_fack(tp
);
982 /* This must be called before lost_out is incremented */
983 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
985 if ((tp
->retransmit_skb_hint
== NULL
) ||
986 before(TCP_SKB_CB(skb
)->seq
,
987 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
988 tp
->retransmit_skb_hint
= skb
;
991 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
992 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
995 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
997 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
998 tcp_verify_retransmit_hint(tp
, skb
);
1000 tp
->lost_out
+= tcp_skb_pcount(skb
);
1001 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1005 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1007 tcp_verify_retransmit_hint(tp
, skb
);
1009 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1010 tp
->lost_out
+= tcp_skb_pcount(skb
);
1011 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1015 /* This procedure tags the retransmission queue when SACKs arrive.
1017 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1018 * Packets in queue with these bits set are counted in variables
1019 * sacked_out, retrans_out and lost_out, correspondingly.
1021 * Valid combinations are:
1022 * Tag InFlight Description
1023 * 0 1 - orig segment is in flight.
1024 * S 0 - nothing flies, orig reached receiver.
1025 * L 0 - nothing flies, orig lost by net.
1026 * R 2 - both orig and retransmit are in flight.
1027 * L|R 1 - orig is lost, retransmit is in flight.
1028 * S|R 1 - orig reached receiver, retrans is still in flight.
1029 * (L|S|R is logically valid, it could occur when L|R is sacked,
1030 * but it is equivalent to plain S and code short-curcuits it to S.
1031 * L|S is logically invalid, it would mean -1 packet in flight 8))
1033 * These 6 states form finite state machine, controlled by the following events:
1034 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1035 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1036 * 3. Loss detection event of one of three flavors:
1037 * A. Scoreboard estimator decided the packet is lost.
1038 * A'. Reno "three dupacks" marks head of queue lost.
1039 * A''. Its FACK modfication, head until snd.fack is lost.
1040 * B. SACK arrives sacking data transmitted after never retransmitted
1041 * hole was sent out.
1042 * C. SACK arrives sacking SND.NXT at the moment, when the
1043 * segment was retransmitted.
1044 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1046 * It is pleasant to note, that state diagram turns out to be commutative,
1047 * so that we are allowed not to be bothered by order of our actions,
1048 * when multiple events arrive simultaneously. (see the function below).
1050 * Reordering detection.
1051 * --------------------
1052 * Reordering metric is maximal distance, which a packet can be displaced
1053 * in packet stream. With SACKs we can estimate it:
1055 * 1. SACK fills old hole and the corresponding segment was not
1056 * ever retransmitted -> reordering. Alas, we cannot use it
1057 * when segment was retransmitted.
1058 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1059 * for retransmitted and already SACKed segment -> reordering..
1060 * Both of these heuristics are not used in Loss state, when we cannot
1061 * account for retransmits accurately.
1063 * SACK block validation.
1064 * ----------------------
1066 * SACK block range validation checks that the received SACK block fits to
1067 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1068 * Note that SND.UNA is not included to the range though being valid because
1069 * it means that the receiver is rather inconsistent with itself reporting
1070 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1071 * perfectly valid, however, in light of RFC2018 which explicitly states
1072 * that "SACK block MUST reflect the newest segment. Even if the newest
1073 * segment is going to be discarded ...", not that it looks very clever
1074 * in case of head skb. Due to potentional receiver driven attacks, we
1075 * choose to avoid immediate execution of a walk in write queue due to
1076 * reneging and defer head skb's loss recovery to standard loss recovery
1077 * procedure that will eventually trigger (nothing forbids us doing this).
1079 * Implements also blockage to start_seq wrap-around. Problem lies in the
1080 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1081 * there's no guarantee that it will be before snd_nxt (n). The problem
1082 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1085 * <- outs wnd -> <- wrapzone ->
1086 * u e n u_w e_w s n_w
1088 * |<------------+------+----- TCP seqno space --------------+---------->|
1089 * ...-- <2^31 ->| |<--------...
1090 * ...---- >2^31 ------>| |<--------...
1092 * Current code wouldn't be vulnerable but it's better still to discard such
1093 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1094 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1095 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1096 * equal to the ideal case (infinite seqno space without wrap caused issues).
1098 * With D-SACK the lower bound is extended to cover sequence space below
1099 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1100 * again, D-SACK block must not to go across snd_una (for the same reason as
1101 * for the normal SACK blocks, explained above). But there all simplicity
1102 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1103 * fully below undo_marker they do not affect behavior in anyway and can
1104 * therefore be safely ignored. In rare cases (which are more or less
1105 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1106 * fragmentation and packet reordering past skb's retransmission. To consider
1107 * them correctly, the acceptable range must be extended even more though
1108 * the exact amount is rather hard to quantify. However, tp->max_window can
1109 * be used as an exaggerated estimate.
1111 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1112 u32 start_seq
, u32 end_seq
)
1114 /* Too far in future, or reversed (interpretation is ambiguous) */
1115 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1118 /* Nasty start_seq wrap-around check (see comments above) */
1119 if (!before(start_seq
, tp
->snd_nxt
))
1122 /* In outstanding window? ...This is valid exit for D-SACKs too.
1123 * start_seq == snd_una is non-sensical (see comments above)
1125 if (after(start_seq
, tp
->snd_una
))
1128 if (!is_dsack
|| !tp
->undo_marker
)
1131 /* ...Then it's D-SACK, and must reside below snd_una completely */
1132 if (!after(end_seq
, tp
->snd_una
))
1135 if (!before(start_seq
, tp
->undo_marker
))
1139 if (!after(end_seq
, tp
->undo_marker
))
1142 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1143 * start_seq < undo_marker and end_seq >= undo_marker.
1145 return !before(start_seq
, end_seq
- tp
->max_window
);
1148 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1149 * Event "C". Later note: FACK people cheated me again 8), we have to account
1150 * for reordering! Ugly, but should help.
1152 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1153 * less than what is now known to be received by the other end (derived from
1154 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1155 * retransmitted skbs to avoid some costly processing per ACKs.
1157 static void tcp_mark_lost_retrans(struct sock
*sk
)
1159 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1160 struct tcp_sock
*tp
= tcp_sk(sk
);
1161 struct sk_buff
*skb
;
1163 u32 new_low_seq
= tp
->snd_nxt
;
1164 u32 received_upto
= tcp_highest_sack_seq(tp
);
1166 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1167 !after(received_upto
, tp
->lost_retrans_low
) ||
1168 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1171 tcp_for_write_queue(skb
, sk
) {
1172 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1174 if (skb
== tcp_send_head(sk
))
1176 if (cnt
== tp
->retrans_out
)
1178 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1181 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1184 if (after(received_upto
, ack_seq
) &&
1186 !before(received_upto
,
1187 ack_seq
+ tp
->reordering
* tp
->mss_cache
))) {
1188 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1189 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1191 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1192 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1194 if (before(ack_seq
, new_low_seq
))
1195 new_low_seq
= ack_seq
;
1196 cnt
+= tcp_skb_pcount(skb
);
1200 if (tp
->retrans_out
)
1201 tp
->lost_retrans_low
= new_low_seq
;
1204 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1205 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1208 struct tcp_sock
*tp
= tcp_sk(sk
);
1209 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1210 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1213 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1216 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1217 } else if (num_sacks
> 1) {
1218 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1219 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1221 if (!after(end_seq_0
, end_seq_1
) &&
1222 !before(start_seq_0
, start_seq_1
)) {
1225 NET_INC_STATS_BH(sock_net(sk
),
1226 LINUX_MIB_TCPDSACKOFORECV
);
1230 /* D-SACK for already forgotten data... Do dumb counting. */
1232 !after(end_seq_0
, prior_snd_una
) &&
1233 after(end_seq_0
, tp
->undo_marker
))
1239 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1240 * the incoming SACK may not exactly match but we can find smaller MSS
1241 * aligned portion of it that matches. Therefore we might need to fragment
1242 * which may fail and creates some hassle (caller must handle error case
1245 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1246 u32 start_seq
, u32 end_seq
)
1249 unsigned int pkt_len
;
1251 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1252 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1254 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1255 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1257 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1260 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1262 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1263 err
= tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
);
1271 static int tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1272 int *reord
, int dup_sack
, int fack_count
)
1274 struct tcp_sock
*tp
= tcp_sk(sk
);
1275 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1278 /* Account D-SACK for retransmitted packet. */
1279 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1280 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1282 if (sacked
& TCPCB_SACKED_ACKED
)
1283 *reord
= min(fack_count
, *reord
);
1286 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1287 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1290 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1291 if (sacked
& TCPCB_SACKED_RETRANS
) {
1292 /* If the segment is not tagged as lost,
1293 * we do not clear RETRANS, believing
1294 * that retransmission is still in flight.
1296 if (sacked
& TCPCB_LOST
) {
1297 TCP_SKB_CB(skb
)->sacked
&=
1298 ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1299 tp
->lost_out
-= tcp_skb_pcount(skb
);
1300 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1303 if (!(sacked
& TCPCB_RETRANS
)) {
1304 /* New sack for not retransmitted frame,
1305 * which was in hole. It is reordering.
1307 if (before(TCP_SKB_CB(skb
)->seq
,
1308 tcp_highest_sack_seq(tp
)))
1309 *reord
= min(fack_count
, *reord
);
1311 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1312 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1313 flag
|= FLAG_ONLY_ORIG_SACKED
;
1316 if (sacked
& TCPCB_LOST
) {
1317 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1318 tp
->lost_out
-= tcp_skb_pcount(skb
);
1322 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1323 flag
|= FLAG_DATA_SACKED
;
1324 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1326 fack_count
+= tcp_skb_pcount(skb
);
1328 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1329 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1330 before(TCP_SKB_CB(skb
)->seq
,
1331 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1332 tp
->lost_cnt_hint
+= tcp_skb_pcount(skb
);
1334 if (fack_count
> tp
->fackets_out
)
1335 tp
->fackets_out
= fack_count
;
1337 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
1338 tcp_advance_highest_sack(sk
, skb
);
1341 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1342 * frames and clear it. undo_retrans is decreased above, L|R frames
1343 * are accounted above as well.
1345 if (dup_sack
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)) {
1346 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1347 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1353 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1354 struct tcp_sack_block
*next_dup
,
1355 u32 start_seq
, u32 end_seq
,
1356 int dup_sack_in
, int *fack_count
,
1357 int *reord
, int *flag
)
1359 tcp_for_write_queue_from(skb
, sk
) {
1361 int dup_sack
= dup_sack_in
;
1363 if (skb
== tcp_send_head(sk
))
1366 /* queue is in-order => we can short-circuit the walk early */
1367 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1370 if ((next_dup
!= NULL
) &&
1371 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1372 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1373 next_dup
->start_seq
,
1380 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
,
1382 if (unlikely(in_sack
< 0))
1386 *flag
|= tcp_sacktag_one(skb
, sk
, reord
, dup_sack
,
1389 *fack_count
+= tcp_skb_pcount(skb
);
1394 /* Avoid all extra work that is being done by sacktag while walking in
1397 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1398 u32 skip_to_seq
, int *fack_count
)
1400 tcp_for_write_queue_from(skb
, sk
) {
1401 if (skb
== tcp_send_head(sk
))
1404 if (!before(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1407 *fack_count
+= tcp_skb_pcount(skb
);
1412 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1414 struct tcp_sack_block
*next_dup
,
1416 int *fack_count
, int *reord
,
1419 if (next_dup
== NULL
)
1422 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1423 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
, fack_count
);
1424 skb
= tcp_sacktag_walk(skb
, sk
, NULL
,
1425 next_dup
->start_seq
, next_dup
->end_seq
,
1426 1, fack_count
, reord
, flag
);
1432 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1434 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1438 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1441 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1442 struct tcp_sock
*tp
= tcp_sk(sk
);
1443 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1444 TCP_SKB_CB(ack_skb
)->sacked
);
1445 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1446 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1447 struct tcp_sack_block
*cache
;
1448 struct sk_buff
*skb
;
1449 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1451 int reord
= tp
->packets_out
;
1453 int found_dup_sack
= 0;
1456 int first_sack_index
;
1458 if (!tp
->sacked_out
) {
1459 if (WARN_ON(tp
->fackets_out
))
1460 tp
->fackets_out
= 0;
1461 tcp_highest_sack_reset(sk
);
1464 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1465 num_sacks
, prior_snd_una
);
1467 flag
|= FLAG_DSACKING_ACK
;
1469 /* Eliminate too old ACKs, but take into
1470 * account more or less fresh ones, they can
1471 * contain valid SACK info.
1473 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1476 if (!tp
->packets_out
)
1480 first_sack_index
= 0;
1481 for (i
= 0; i
< num_sacks
; i
++) {
1482 int dup_sack
= !i
&& found_dup_sack
;
1484 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1485 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1487 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1488 sp
[used_sacks
].start_seq
,
1489 sp
[used_sacks
].end_seq
)) {
1493 if (!tp
->undo_marker
)
1494 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1496 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1498 /* Don't count olds caused by ACK reordering */
1499 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1500 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1502 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1505 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1507 first_sack_index
= -1;
1511 /* Ignore very old stuff early */
1512 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1518 /* order SACK blocks to allow in order walk of the retrans queue */
1519 for (i
= used_sacks
- 1; i
> 0; i
--) {
1520 for (j
= 0; j
< i
; j
++) {
1521 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1522 struct tcp_sack_block tmp
;
1528 /* Track where the first SACK block goes to */
1529 if (j
== first_sack_index
)
1530 first_sack_index
= j
+ 1;
1535 skb
= tcp_write_queue_head(sk
);
1539 if (!tp
->sacked_out
) {
1540 /* It's already past, so skip checking against it */
1541 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1543 cache
= tp
->recv_sack_cache
;
1544 /* Skip empty blocks in at head of the cache */
1545 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1550 while (i
< used_sacks
) {
1551 u32 start_seq
= sp
[i
].start_seq
;
1552 u32 end_seq
= sp
[i
].end_seq
;
1553 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1554 struct tcp_sack_block
*next_dup
= NULL
;
1556 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1557 next_dup
= &sp
[i
+ 1];
1559 /* Event "B" in the comment above. */
1560 if (after(end_seq
, tp
->high_seq
))
1561 flag
|= FLAG_DATA_LOST
;
1563 /* Skip too early cached blocks */
1564 while (tcp_sack_cache_ok(tp
, cache
) &&
1565 !before(start_seq
, cache
->end_seq
))
1568 /* Can skip some work by looking recv_sack_cache? */
1569 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1570 after(end_seq
, cache
->start_seq
)) {
1573 if (before(start_seq
, cache
->start_seq
)) {
1574 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
,
1576 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1579 dup_sack
, &fack_count
,
1583 /* Rest of the block already fully processed? */
1584 if (!after(end_seq
, cache
->end_seq
))
1587 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1589 &fack_count
, &reord
,
1592 /* ...tail remains todo... */
1593 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1594 /* ...but better entrypoint exists! */
1595 skb
= tcp_highest_sack(sk
);
1598 fack_count
= tp
->fackets_out
;
1603 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
,
1605 /* Check overlap against next cached too (past this one already) */
1610 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1611 skb
= tcp_highest_sack(sk
);
1614 fack_count
= tp
->fackets_out
;
1616 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
, &fack_count
);
1619 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1620 dup_sack
, &fack_count
, &reord
, &flag
);
1623 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1624 * due to in-order walk
1626 if (after(end_seq
, tp
->frto_highmark
))
1627 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1632 /* Clear the head of the cache sack blocks so we can skip it next time */
1633 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1634 tp
->recv_sack_cache
[i
].start_seq
= 0;
1635 tp
->recv_sack_cache
[i
].end_seq
= 0;
1637 for (j
= 0; j
< used_sacks
; j
++)
1638 tp
->recv_sack_cache
[i
++] = sp
[j
];
1640 tcp_mark_lost_retrans(sk
);
1642 tcp_verify_left_out(tp
);
1644 if ((reord
< tp
->fackets_out
) &&
1645 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1646 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1647 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1651 #if FASTRETRANS_DEBUG > 0
1652 WARN_ON((int)tp
->sacked_out
< 0);
1653 WARN_ON((int)tp
->lost_out
< 0);
1654 WARN_ON((int)tp
->retrans_out
< 0);
1655 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1660 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1661 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1663 int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1667 holes
= max(tp
->lost_out
, 1U);
1668 holes
= min(holes
, tp
->packets_out
);
1670 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1671 tp
->sacked_out
= tp
->packets_out
- holes
;
1677 /* If we receive more dupacks than we expected counting segments
1678 * in assumption of absent reordering, interpret this as reordering.
1679 * The only another reason could be bug in receiver TCP.
1681 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1683 struct tcp_sock
*tp
= tcp_sk(sk
);
1684 if (tcp_limit_reno_sacked(tp
))
1685 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1688 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1690 static void tcp_add_reno_sack(struct sock
*sk
)
1692 struct tcp_sock
*tp
= tcp_sk(sk
);
1694 tcp_check_reno_reordering(sk
, 0);
1695 tcp_verify_left_out(tp
);
1698 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1700 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1702 struct tcp_sock
*tp
= tcp_sk(sk
);
1705 /* One ACK acked hole. The rest eat duplicate ACKs. */
1706 if (acked
- 1 >= tp
->sacked_out
)
1709 tp
->sacked_out
-= acked
- 1;
1711 tcp_check_reno_reordering(sk
, acked
);
1712 tcp_verify_left_out(tp
);
1715 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1720 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
1722 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
1725 /* F-RTO can only be used if TCP has never retransmitted anything other than
1726 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1728 int tcp_use_frto(struct sock
*sk
)
1730 const struct tcp_sock
*tp
= tcp_sk(sk
);
1731 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1732 struct sk_buff
*skb
;
1734 if (!sysctl_tcp_frto
)
1737 /* MTU probe and F-RTO won't really play nicely along currently */
1738 if (icsk
->icsk_mtup
.probe_size
)
1741 if (tcp_is_sackfrto(tp
))
1744 /* Avoid expensive walking of rexmit queue if possible */
1745 if (tp
->retrans_out
> 1)
1748 skb
= tcp_write_queue_head(sk
);
1749 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1750 tcp_for_write_queue_from(skb
, sk
) {
1751 if (skb
== tcp_send_head(sk
))
1753 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1755 /* Short-circuit when first non-SACKed skb has been checked */
1756 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1762 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1763 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1764 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1765 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1766 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1767 * bits are handled if the Loss state is really to be entered (in
1768 * tcp_enter_frto_loss).
1770 * Do like tcp_enter_loss() would; when RTO expires the second time it
1772 * "Reduce ssthresh if it has not yet been made inside this window."
1774 void tcp_enter_frto(struct sock
*sk
)
1776 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1777 struct tcp_sock
*tp
= tcp_sk(sk
);
1778 struct sk_buff
*skb
;
1780 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1781 tp
->snd_una
== tp
->high_seq
||
1782 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1783 !icsk
->icsk_retransmits
)) {
1784 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1785 /* Our state is too optimistic in ssthresh() call because cwnd
1786 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1787 * recovery has not yet completed. Pattern would be this: RTO,
1788 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1790 * RFC4138 should be more specific on what to do, even though
1791 * RTO is quite unlikely to occur after the first Cumulative ACK
1792 * due to back-off and complexity of triggering events ...
1794 if (tp
->frto_counter
) {
1796 stored_cwnd
= tp
->snd_cwnd
;
1798 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1799 tp
->snd_cwnd
= stored_cwnd
;
1801 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1803 /* ... in theory, cong.control module could do "any tricks" in
1804 * ssthresh(), which means that ca_state, lost bits and lost_out
1805 * counter would have to be faked before the call occurs. We
1806 * consider that too expensive, unlikely and hacky, so modules
1807 * using these in ssthresh() must deal these incompatibility
1808 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1810 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1813 tp
->undo_marker
= tp
->snd_una
;
1814 tp
->undo_retrans
= 0;
1816 skb
= tcp_write_queue_head(sk
);
1817 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1818 tp
->undo_marker
= 0;
1819 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1820 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1821 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1823 tcp_verify_left_out(tp
);
1825 /* Too bad if TCP was application limited */
1826 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1828 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1829 * The last condition is necessary at least in tp->frto_counter case.
1831 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
1832 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1833 after(tp
->high_seq
, tp
->snd_una
)) {
1834 tp
->frto_highmark
= tp
->high_seq
;
1836 tp
->frto_highmark
= tp
->snd_nxt
;
1838 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1839 tp
->high_seq
= tp
->snd_nxt
;
1840 tp
->frto_counter
= 1;
1843 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1844 * which indicates that we should follow the traditional RTO recovery,
1845 * i.e. mark everything lost and do go-back-N retransmission.
1847 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1849 struct tcp_sock
*tp
= tcp_sk(sk
);
1850 struct sk_buff
*skb
;
1853 tp
->retrans_out
= 0;
1854 if (tcp_is_reno(tp
))
1855 tcp_reset_reno_sack(tp
);
1857 tcp_for_write_queue(skb
, sk
) {
1858 if (skb
== tcp_send_head(sk
))
1861 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1863 * Count the retransmission made on RTO correctly (only when
1864 * waiting for the first ACK and did not get it)...
1866 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
1867 /* For some reason this R-bit might get cleared? */
1868 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1869 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1870 /* ...enter this if branch just for the first segment */
1871 flag
|= FLAG_DATA_ACKED
;
1873 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1874 tp
->undo_marker
= 0;
1875 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1878 /* Marking forward transmissions that were made after RTO lost
1879 * can cause unnecessary retransmissions in some scenarios,
1880 * SACK blocks will mitigate that in some but not in all cases.
1881 * We used to not mark them but it was causing break-ups with
1882 * receivers that do only in-order receival.
1884 * TODO: we could detect presence of such receiver and select
1885 * different behavior per flow.
1887 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1888 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1889 tp
->lost_out
+= tcp_skb_pcount(skb
);
1890 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1893 tcp_verify_left_out(tp
);
1895 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1896 tp
->snd_cwnd_cnt
= 0;
1897 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1898 tp
->frto_counter
= 0;
1899 tp
->bytes_acked
= 0;
1901 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1902 sysctl_tcp_reordering
);
1903 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1904 tp
->high_seq
= tp
->snd_nxt
;
1905 TCP_ECN_queue_cwr(tp
);
1907 tcp_clear_all_retrans_hints(tp
);
1910 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1912 tp
->retrans_out
= 0;
1915 tp
->undo_marker
= 0;
1916 tp
->undo_retrans
= 0;
1919 void tcp_clear_retrans(struct tcp_sock
*tp
)
1921 tcp_clear_retrans_partial(tp
);
1923 tp
->fackets_out
= 0;
1927 /* Enter Loss state. If "how" is not zero, forget all SACK information
1928 * and reset tags completely, otherwise preserve SACKs. If receiver
1929 * dropped its ofo queue, we will know this due to reneging detection.
1931 void tcp_enter_loss(struct sock
*sk
, int how
)
1933 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1934 struct tcp_sock
*tp
= tcp_sk(sk
);
1935 struct sk_buff
*skb
;
1937 /* Reduce ssthresh if it has not yet been made inside this window. */
1938 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1939 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1940 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1941 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1942 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1945 tp
->snd_cwnd_cnt
= 0;
1946 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1948 tp
->bytes_acked
= 0;
1949 tcp_clear_retrans_partial(tp
);
1951 if (tcp_is_reno(tp
))
1952 tcp_reset_reno_sack(tp
);
1955 /* Push undo marker, if it was plain RTO and nothing
1956 * was retransmitted. */
1957 tp
->undo_marker
= tp
->snd_una
;
1960 tp
->fackets_out
= 0;
1962 tcp_clear_all_retrans_hints(tp
);
1964 tcp_for_write_queue(skb
, sk
) {
1965 if (skb
== tcp_send_head(sk
))
1968 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1969 tp
->undo_marker
= 0;
1970 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1971 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1972 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1973 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1974 tp
->lost_out
+= tcp_skb_pcount(skb
);
1975 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1978 tcp_verify_left_out(tp
);
1980 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1981 sysctl_tcp_reordering
);
1982 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1983 tp
->high_seq
= tp
->snd_nxt
;
1984 TCP_ECN_queue_cwr(tp
);
1985 /* Abort F-RTO algorithm if one is in progress */
1986 tp
->frto_counter
= 0;
1989 /* If ACK arrived pointing to a remembered SACK, it means that our
1990 * remembered SACKs do not reflect real state of receiver i.e.
1991 * receiver _host_ is heavily congested (or buggy).
1993 * Do processing similar to RTO timeout.
1995 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1997 if (flag
& FLAG_SACK_RENEGING
) {
1998 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1999 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2001 tcp_enter_loss(sk
, 1);
2002 icsk
->icsk_retransmits
++;
2003 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2004 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2005 icsk
->icsk_rto
, TCP_RTO_MAX
);
2011 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2013 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2016 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2017 * counter when SACK is enabled (without SACK, sacked_out is used for
2020 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2021 * segments up to the highest received SACK block so far and holes in
2024 * With reordering, holes may still be in flight, so RFC3517 recovery
2025 * uses pure sacked_out (total number of SACKed segments) even though
2026 * it violates the RFC that uses duplicate ACKs, often these are equal
2027 * but when e.g. out-of-window ACKs or packet duplication occurs,
2028 * they differ. Since neither occurs due to loss, TCP should really
2031 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
2033 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2036 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2038 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2041 static inline int tcp_head_timedout(struct sock
*sk
)
2043 struct tcp_sock
*tp
= tcp_sk(sk
);
2045 return tp
->packets_out
&&
2046 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2049 /* Linux NewReno/SACK/FACK/ECN state machine.
2050 * --------------------------------------
2052 * "Open" Normal state, no dubious events, fast path.
2053 * "Disorder" In all the respects it is "Open",
2054 * but requires a bit more attention. It is entered when
2055 * we see some SACKs or dupacks. It is split of "Open"
2056 * mainly to move some processing from fast path to slow one.
2057 * "CWR" CWND was reduced due to some Congestion Notification event.
2058 * It can be ECN, ICMP source quench, local device congestion.
2059 * "Recovery" CWND was reduced, we are fast-retransmitting.
2060 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2062 * tcp_fastretrans_alert() is entered:
2063 * - each incoming ACK, if state is not "Open"
2064 * - when arrived ACK is unusual, namely:
2069 * Counting packets in flight is pretty simple.
2071 * in_flight = packets_out - left_out + retrans_out
2073 * packets_out is SND.NXT-SND.UNA counted in packets.
2075 * retrans_out is number of retransmitted segments.
2077 * left_out is number of segments left network, but not ACKed yet.
2079 * left_out = sacked_out + lost_out
2081 * sacked_out: Packets, which arrived to receiver out of order
2082 * and hence not ACKed. With SACKs this number is simply
2083 * amount of SACKed data. Even without SACKs
2084 * it is easy to give pretty reliable estimate of this number,
2085 * counting duplicate ACKs.
2087 * lost_out: Packets lost by network. TCP has no explicit
2088 * "loss notification" feedback from network (for now).
2089 * It means that this number can be only _guessed_.
2090 * Actually, it is the heuristics to predict lossage that
2091 * distinguishes different algorithms.
2093 * F.e. after RTO, when all the queue is considered as lost,
2094 * lost_out = packets_out and in_flight = retrans_out.
2096 * Essentially, we have now two algorithms counting
2099 * FACK: It is the simplest heuristics. As soon as we decided
2100 * that something is lost, we decide that _all_ not SACKed
2101 * packets until the most forward SACK are lost. I.e.
2102 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2103 * It is absolutely correct estimate, if network does not reorder
2104 * packets. And it loses any connection to reality when reordering
2105 * takes place. We use FACK by default until reordering
2106 * is suspected on the path to this destination.
2108 * NewReno: when Recovery is entered, we assume that one segment
2109 * is lost (classic Reno). While we are in Recovery and
2110 * a partial ACK arrives, we assume that one more packet
2111 * is lost (NewReno). This heuristics are the same in NewReno
2114 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2115 * deflation etc. CWND is real congestion window, never inflated, changes
2116 * only according to classic VJ rules.
2118 * Really tricky (and requiring careful tuning) part of algorithm
2119 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2120 * The first determines the moment _when_ we should reduce CWND and,
2121 * hence, slow down forward transmission. In fact, it determines the moment
2122 * when we decide that hole is caused by loss, rather than by a reorder.
2124 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2125 * holes, caused by lost packets.
2127 * And the most logically complicated part of algorithm is undo
2128 * heuristics. We detect false retransmits due to both too early
2129 * fast retransmit (reordering) and underestimated RTO, analyzing
2130 * timestamps and D-SACKs. When we detect that some segments were
2131 * retransmitted by mistake and CWND reduction was wrong, we undo
2132 * window reduction and abort recovery phase. This logic is hidden
2133 * inside several functions named tcp_try_undo_<something>.
2136 /* This function decides, when we should leave Disordered state
2137 * and enter Recovery phase, reducing congestion window.
2139 * Main question: may we further continue forward transmission
2140 * with the same cwnd?
2142 static int tcp_time_to_recover(struct sock
*sk
)
2144 struct tcp_sock
*tp
= tcp_sk(sk
);
2147 /* Do not perform any recovery during F-RTO algorithm */
2148 if (tp
->frto_counter
)
2151 /* Trick#1: The loss is proven. */
2155 /* Not-A-Trick#2 : Classic rule... */
2156 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2159 /* Trick#3 : when we use RFC2988 timer restart, fast
2160 * retransmit can be triggered by timeout of queue head.
2162 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2165 /* Trick#4: It is still not OK... But will it be useful to delay
2168 packets_out
= tp
->packets_out
;
2169 if (packets_out
<= tp
->reordering
&&
2170 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2171 !tcp_may_send_now(sk
)) {
2172 /* We have nothing to send. This connection is limited
2173 * either by receiver window or by application.
2181 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2182 * is against sacked "cnt", otherwise it's against facked "cnt"
2184 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2186 struct tcp_sock
*tp
= tcp_sk(sk
);
2187 struct sk_buff
*skb
;
2192 WARN_ON(packets
> tp
->packets_out
);
2193 if (tp
->lost_skb_hint
) {
2194 skb
= tp
->lost_skb_hint
;
2195 cnt
= tp
->lost_cnt_hint
;
2197 skb
= tcp_write_queue_head(sk
);
2201 tcp_for_write_queue_from(skb
, sk
) {
2202 if (skb
== tcp_send_head(sk
))
2204 /* TODO: do this better */
2205 /* this is not the most efficient way to do this... */
2206 tp
->lost_skb_hint
= skb
;
2207 tp
->lost_cnt_hint
= cnt
;
2209 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2213 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2214 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2215 cnt
+= tcp_skb_pcount(skb
);
2217 if (cnt
> packets
) {
2218 if (tcp_is_sack(tp
) || (oldcnt
>= packets
))
2221 mss
= skb_shinfo(skb
)->gso_size
;
2222 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2228 tcp_skb_mark_lost(tp
, skb
);
2230 tcp_verify_left_out(tp
);
2233 /* Account newly detected lost packet(s) */
2235 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2237 struct tcp_sock
*tp
= tcp_sk(sk
);
2239 if (tcp_is_reno(tp
)) {
2240 tcp_mark_head_lost(sk
, 1);
2241 } else if (tcp_is_fack(tp
)) {
2242 int lost
= tp
->fackets_out
- tp
->reordering
;
2245 tcp_mark_head_lost(sk
, lost
);
2247 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2248 if (sacked_upto
< fast_rexmit
)
2249 sacked_upto
= fast_rexmit
;
2250 tcp_mark_head_lost(sk
, sacked_upto
);
2253 /* New heuristics: it is possible only after we switched
2254 * to restart timer each time when something is ACKed.
2255 * Hence, we can detect timed out packets during fast
2256 * retransmit without falling to slow start.
2258 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2259 struct sk_buff
*skb
;
2261 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2262 : tcp_write_queue_head(sk
);
2264 tcp_for_write_queue_from(skb
, sk
) {
2265 if (skb
== tcp_send_head(sk
))
2267 if (!tcp_skb_timedout(sk
, skb
))
2270 tcp_skb_mark_lost(tp
, skb
);
2273 tp
->scoreboard_skb_hint
= skb
;
2275 tcp_verify_left_out(tp
);
2279 /* CWND moderation, preventing bursts due to too big ACKs
2280 * in dubious situations.
2282 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2284 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2285 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2286 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2289 /* Lower bound on congestion window is slow start threshold
2290 * unless congestion avoidance choice decides to overide it.
2292 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2294 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2296 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2299 /* Decrease cwnd each second ack. */
2300 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2302 struct tcp_sock
*tp
= tcp_sk(sk
);
2303 int decr
= tp
->snd_cwnd_cnt
+ 1;
2305 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2306 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2307 tp
->snd_cwnd_cnt
= decr
& 1;
2310 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2311 tp
->snd_cwnd
-= decr
;
2313 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2314 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2318 /* Nothing was retransmitted or returned timestamp is less
2319 * than timestamp of the first retransmission.
2321 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2323 return !tp
->retrans_stamp
||
2324 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2325 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2328 /* Undo procedures. */
2330 #if FASTRETRANS_DEBUG > 1
2331 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2333 struct tcp_sock
*tp
= tcp_sk(sk
);
2334 struct inet_sock
*inet
= inet_sk(sk
);
2336 if (sk
->sk_family
== AF_INET
) {
2337 printk(KERN_DEBUG
"Undo %s " NIPQUAD_FMT
"/%u c%u l%u ss%u/%u p%u\n",
2339 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2340 tp
->snd_cwnd
, tcp_left_out(tp
),
2341 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2344 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2345 else if (sk
->sk_family
== AF_INET6
) {
2346 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2347 printk(KERN_DEBUG
"Undo %s " NIP6_FMT
"/%u c%u l%u ss%u/%u p%u\n",
2349 NIP6(np
->daddr
), ntohs(inet
->dport
),
2350 tp
->snd_cwnd
, tcp_left_out(tp
),
2351 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2357 #define DBGUNDO(x...) do { } while (0)
2360 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2362 struct tcp_sock
*tp
= tcp_sk(sk
);
2364 if (tp
->prior_ssthresh
) {
2365 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2367 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2368 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2370 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2372 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2373 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2374 TCP_ECN_withdraw_cwr(tp
);
2377 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2379 tcp_moderate_cwnd(tp
);
2380 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2383 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2385 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2388 /* People celebrate: "We love our President!" */
2389 static int tcp_try_undo_recovery(struct sock
*sk
)
2391 struct tcp_sock
*tp
= tcp_sk(sk
);
2393 if (tcp_may_undo(tp
)) {
2396 /* Happy end! We did not retransmit anything
2397 * or our original transmission succeeded.
2399 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2400 tcp_undo_cwr(sk
, 1);
2401 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2402 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2404 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2406 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2407 tp
->undo_marker
= 0;
2409 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2410 /* Hold old state until something *above* high_seq
2411 * is ACKed. For Reno it is MUST to prevent false
2412 * fast retransmits (RFC2582). SACK TCP is safe. */
2413 tcp_moderate_cwnd(tp
);
2416 tcp_set_ca_state(sk
, TCP_CA_Open
);
2420 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2421 static void tcp_try_undo_dsack(struct sock
*sk
)
2423 struct tcp_sock
*tp
= tcp_sk(sk
);
2425 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2426 DBGUNDO(sk
, "D-SACK");
2427 tcp_undo_cwr(sk
, 1);
2428 tp
->undo_marker
= 0;
2429 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2433 /* Undo during fast recovery after partial ACK. */
2435 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2437 struct tcp_sock
*tp
= tcp_sk(sk
);
2438 /* Partial ACK arrived. Force Hoe's retransmit. */
2439 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2441 if (tcp_may_undo(tp
)) {
2442 /* Plain luck! Hole if filled with delayed
2443 * packet, rather than with a retransmit.
2445 if (tp
->retrans_out
== 0)
2446 tp
->retrans_stamp
= 0;
2448 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2451 tcp_undo_cwr(sk
, 0);
2452 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2454 /* So... Do not make Hoe's retransmit yet.
2455 * If the first packet was delayed, the rest
2456 * ones are most probably delayed as well.
2463 /* Undo during loss recovery after partial ACK. */
2464 static int tcp_try_undo_loss(struct sock
*sk
)
2466 struct tcp_sock
*tp
= tcp_sk(sk
);
2468 if (tcp_may_undo(tp
)) {
2469 struct sk_buff
*skb
;
2470 tcp_for_write_queue(skb
, sk
) {
2471 if (skb
== tcp_send_head(sk
))
2473 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2476 tcp_clear_all_retrans_hints(tp
);
2478 DBGUNDO(sk
, "partial loss");
2480 tcp_undo_cwr(sk
, 1);
2481 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2482 inet_csk(sk
)->icsk_retransmits
= 0;
2483 tp
->undo_marker
= 0;
2484 if (tcp_is_sack(tp
))
2485 tcp_set_ca_state(sk
, TCP_CA_Open
);
2491 static inline void tcp_complete_cwr(struct sock
*sk
)
2493 struct tcp_sock
*tp
= tcp_sk(sk
);
2494 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2495 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2496 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2499 static void tcp_try_keep_open(struct sock
*sk
)
2501 struct tcp_sock
*tp
= tcp_sk(sk
);
2502 int state
= TCP_CA_Open
;
2504 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2505 state
= TCP_CA_Disorder
;
2507 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2508 tcp_set_ca_state(sk
, state
);
2509 tp
->high_seq
= tp
->snd_nxt
;
2513 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2515 struct tcp_sock
*tp
= tcp_sk(sk
);
2517 tcp_verify_left_out(tp
);
2519 if (!tp
->frto_counter
&& tp
->retrans_out
== 0)
2520 tp
->retrans_stamp
= 0;
2522 if (flag
& FLAG_ECE
)
2523 tcp_enter_cwr(sk
, 1);
2525 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2526 tcp_try_keep_open(sk
);
2527 tcp_moderate_cwnd(tp
);
2529 tcp_cwnd_down(sk
, flag
);
2533 static void tcp_mtup_probe_failed(struct sock
*sk
)
2535 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2537 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2538 icsk
->icsk_mtup
.probe_size
= 0;
2541 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2543 struct tcp_sock
*tp
= tcp_sk(sk
);
2544 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2546 /* FIXME: breaks with very large cwnd */
2547 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2548 tp
->snd_cwnd
= tp
->snd_cwnd
*
2549 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2550 icsk
->icsk_mtup
.probe_size
;
2551 tp
->snd_cwnd_cnt
= 0;
2552 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2553 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2555 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2556 icsk
->icsk_mtup
.probe_size
= 0;
2557 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2560 /* Process an event, which can update packets-in-flight not trivially.
2561 * Main goal of this function is to calculate new estimate for left_out,
2562 * taking into account both packets sitting in receiver's buffer and
2563 * packets lost by network.
2565 * Besides that it does CWND reduction, when packet loss is detected
2566 * and changes state of machine.
2568 * It does _not_ decide what to send, it is made in function
2569 * tcp_xmit_retransmit_queue().
2571 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2573 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2574 struct tcp_sock
*tp
= tcp_sk(sk
);
2575 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2576 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2577 (tcp_fackets_out(tp
) > tp
->reordering
));
2578 int fast_rexmit
= 0, mib_idx
;
2580 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2582 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2583 tp
->fackets_out
= 0;
2585 /* Now state machine starts.
2586 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2587 if (flag
& FLAG_ECE
)
2588 tp
->prior_ssthresh
= 0;
2590 /* B. In all the states check for reneging SACKs. */
2591 if (tcp_check_sack_reneging(sk
, flag
))
2594 /* C. Process data loss notification, provided it is valid. */
2595 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2596 before(tp
->snd_una
, tp
->high_seq
) &&
2597 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2598 tp
->fackets_out
> tp
->reordering
) {
2599 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2600 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2603 /* D. Check consistency of the current state. */
2604 tcp_verify_left_out(tp
);
2606 /* E. Check state exit conditions. State can be terminated
2607 * when high_seq is ACKed. */
2608 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2609 WARN_ON(tp
->retrans_out
!= 0);
2610 tp
->retrans_stamp
= 0;
2611 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2612 switch (icsk
->icsk_ca_state
) {
2614 icsk
->icsk_retransmits
= 0;
2615 if (tcp_try_undo_recovery(sk
))
2620 /* CWR is to be held something *above* high_seq
2621 * is ACKed for CWR bit to reach receiver. */
2622 if (tp
->snd_una
!= tp
->high_seq
) {
2623 tcp_complete_cwr(sk
);
2624 tcp_set_ca_state(sk
, TCP_CA_Open
);
2628 case TCP_CA_Disorder
:
2629 tcp_try_undo_dsack(sk
);
2630 if (!tp
->undo_marker
||
2631 /* For SACK case do not Open to allow to undo
2632 * catching for all duplicate ACKs. */
2633 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2634 tp
->undo_marker
= 0;
2635 tcp_set_ca_state(sk
, TCP_CA_Open
);
2639 case TCP_CA_Recovery
:
2640 if (tcp_is_reno(tp
))
2641 tcp_reset_reno_sack(tp
);
2642 if (tcp_try_undo_recovery(sk
))
2644 tcp_complete_cwr(sk
);
2649 /* F. Process state. */
2650 switch (icsk
->icsk_ca_state
) {
2651 case TCP_CA_Recovery
:
2652 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2653 if (tcp_is_reno(tp
) && is_dupack
)
2654 tcp_add_reno_sack(sk
);
2656 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2659 if (flag
& FLAG_DATA_ACKED
)
2660 icsk
->icsk_retransmits
= 0;
2661 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
2662 tcp_reset_reno_sack(tp
);
2663 if (!tcp_try_undo_loss(sk
)) {
2664 tcp_moderate_cwnd(tp
);
2665 tcp_xmit_retransmit_queue(sk
);
2668 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2670 /* Loss is undone; fall through to processing in Open state. */
2672 if (tcp_is_reno(tp
)) {
2673 if (flag
& FLAG_SND_UNA_ADVANCED
)
2674 tcp_reset_reno_sack(tp
);
2676 tcp_add_reno_sack(sk
);
2679 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2680 tcp_try_undo_dsack(sk
);
2682 if (!tcp_time_to_recover(sk
)) {
2683 tcp_try_to_open(sk
, flag
);
2687 /* MTU probe failure: don't reduce cwnd */
2688 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2689 icsk
->icsk_mtup
.probe_size
&&
2690 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2691 tcp_mtup_probe_failed(sk
);
2692 /* Restores the reduction we did in tcp_mtup_probe() */
2694 tcp_simple_retransmit(sk
);
2698 /* Otherwise enter Recovery state */
2700 if (tcp_is_reno(tp
))
2701 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2703 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2705 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2707 tp
->high_seq
= tp
->snd_nxt
;
2708 tp
->prior_ssthresh
= 0;
2709 tp
->undo_marker
= tp
->snd_una
;
2710 tp
->undo_retrans
= tp
->retrans_out
;
2712 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2713 if (!(flag
& FLAG_ECE
))
2714 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2715 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2716 TCP_ECN_queue_cwr(tp
);
2719 tp
->bytes_acked
= 0;
2720 tp
->snd_cwnd_cnt
= 0;
2721 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2725 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2726 tcp_update_scoreboard(sk
, fast_rexmit
);
2727 tcp_cwnd_down(sk
, flag
);
2728 tcp_xmit_retransmit_queue(sk
);
2731 /* Read draft-ietf-tcplw-high-performance before mucking
2732 * with this code. (Supersedes RFC1323)
2734 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2736 /* RTTM Rule: A TSecr value received in a segment is used to
2737 * update the averaged RTT measurement only if the segment
2738 * acknowledges some new data, i.e., only if it advances the
2739 * left edge of the send window.
2741 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2742 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2744 * Changed: reset backoff as soon as we see the first valid sample.
2745 * If we do not, we get strongly overestimated rto. With timestamps
2746 * samples are accepted even from very old segments: f.e., when rtt=1
2747 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2748 * answer arrives rto becomes 120 seconds! If at least one of segments
2749 * in window is lost... Voila. --ANK (010210)
2751 struct tcp_sock
*tp
= tcp_sk(sk
);
2752 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2753 tcp_rtt_estimator(sk
, seq_rtt
);
2755 inet_csk(sk
)->icsk_backoff
= 0;
2759 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2761 /* We don't have a timestamp. Can only use
2762 * packets that are not retransmitted to determine
2763 * rtt estimates. Also, we must not reset the
2764 * backoff for rto until we get a non-retransmitted
2765 * packet. This allows us to deal with a situation
2766 * where the network delay has increased suddenly.
2767 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2770 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2773 tcp_rtt_estimator(sk
, seq_rtt
);
2775 inet_csk(sk
)->icsk_backoff
= 0;
2779 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2782 const struct tcp_sock
*tp
= tcp_sk(sk
);
2783 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2784 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2785 tcp_ack_saw_tstamp(sk
, flag
);
2786 else if (seq_rtt
>= 0)
2787 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2790 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2792 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2793 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2794 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2797 /* Restart timer after forward progress on connection.
2798 * RFC2988 recommends to restart timer to now+rto.
2800 static void tcp_rearm_rto(struct sock
*sk
)
2802 struct tcp_sock
*tp
= tcp_sk(sk
);
2804 if (!tp
->packets_out
) {
2805 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2807 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2808 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2812 /* If we get here, the whole TSO packet has not been acked. */
2813 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2815 struct tcp_sock
*tp
= tcp_sk(sk
);
2818 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2820 packets_acked
= tcp_skb_pcount(skb
);
2821 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2823 packets_acked
-= tcp_skb_pcount(skb
);
2825 if (packets_acked
) {
2826 BUG_ON(tcp_skb_pcount(skb
) == 0);
2827 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2830 return packets_acked
;
2833 /* Remove acknowledged frames from the retransmission queue. If our packet
2834 * is before the ack sequence we can discard it as it's confirmed to have
2835 * arrived at the other end.
2837 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
)
2839 struct tcp_sock
*tp
= tcp_sk(sk
);
2840 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2841 struct sk_buff
*skb
;
2842 u32 now
= tcp_time_stamp
;
2843 int fully_acked
= 1;
2846 u32 reord
= tp
->packets_out
;
2848 s32 ca_seq_rtt
= -1;
2849 ktime_t last_ackt
= net_invalid_timestamp();
2851 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2852 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2855 u8 sacked
= scb
->sacked
;
2857 /* Determine how many packets and what bytes were acked, tso and else */
2858 if (after(scb
->end_seq
, tp
->snd_una
)) {
2859 if (tcp_skb_pcount(skb
) == 1 ||
2860 !after(tp
->snd_una
, scb
->seq
))
2863 acked_pcount
= tcp_tso_acked(sk
, skb
);
2868 end_seq
= tp
->snd_una
;
2870 acked_pcount
= tcp_skb_pcount(skb
);
2871 end_seq
= scb
->end_seq
;
2874 /* MTU probing checks */
2875 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2876 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2877 tcp_mtup_probe_success(sk
, skb
);
2880 if (sacked
& TCPCB_RETRANS
) {
2881 if (sacked
& TCPCB_SACKED_RETRANS
)
2882 tp
->retrans_out
-= acked_pcount
;
2883 flag
|= FLAG_RETRANS_DATA_ACKED
;
2886 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
2887 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2889 ca_seq_rtt
= now
- scb
->when
;
2890 last_ackt
= skb
->tstamp
;
2892 seq_rtt
= ca_seq_rtt
;
2894 if (!(sacked
& TCPCB_SACKED_ACKED
))
2895 reord
= min(pkts_acked
, reord
);
2898 if (sacked
& TCPCB_SACKED_ACKED
)
2899 tp
->sacked_out
-= acked_pcount
;
2900 if (sacked
& TCPCB_LOST
)
2901 tp
->lost_out
-= acked_pcount
;
2903 if (unlikely(tp
->urg_mode
&& !before(end_seq
, tp
->snd_up
)))
2906 tp
->packets_out
-= acked_pcount
;
2907 pkts_acked
+= acked_pcount
;
2909 /* Initial outgoing SYN's get put onto the write_queue
2910 * just like anything else we transmit. It is not
2911 * true data, and if we misinform our callers that
2912 * this ACK acks real data, we will erroneously exit
2913 * connection startup slow start one packet too
2914 * quickly. This is severely frowned upon behavior.
2916 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2917 flag
|= FLAG_DATA_ACKED
;
2919 flag
|= FLAG_SYN_ACKED
;
2920 tp
->retrans_stamp
= 0;
2926 tcp_unlink_write_queue(skb
, sk
);
2927 sk_wmem_free_skb(sk
, skb
);
2928 tcp_clear_all_retrans_hints(tp
);
2931 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2932 flag
|= FLAG_SACK_RENEGING
;
2934 if (flag
& FLAG_ACKED
) {
2935 const struct tcp_congestion_ops
*ca_ops
2936 = inet_csk(sk
)->icsk_ca_ops
;
2938 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2941 if (tcp_is_reno(tp
)) {
2942 tcp_remove_reno_sacks(sk
, pkts_acked
);
2944 /* Non-retransmitted hole got filled? That's reordering */
2945 if (reord
< prior_fackets
)
2946 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2949 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2951 if (ca_ops
->pkts_acked
) {
2954 /* Is the ACK triggering packet unambiguous? */
2955 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2956 /* High resolution needed and available? */
2957 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2958 !ktime_equal(last_ackt
,
2959 net_invalid_timestamp()))
2960 rtt_us
= ktime_us_delta(ktime_get_real(),
2962 else if (ca_seq_rtt
> 0)
2963 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
2966 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2970 #if FASTRETRANS_DEBUG > 0
2971 WARN_ON((int)tp
->sacked_out
< 0);
2972 WARN_ON((int)tp
->lost_out
< 0);
2973 WARN_ON((int)tp
->retrans_out
< 0);
2974 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2975 icsk
= inet_csk(sk
);
2977 printk(KERN_DEBUG
"Leak l=%u %d\n",
2978 tp
->lost_out
, icsk
->icsk_ca_state
);
2981 if (tp
->sacked_out
) {
2982 printk(KERN_DEBUG
"Leak s=%u %d\n",
2983 tp
->sacked_out
, icsk
->icsk_ca_state
);
2986 if (tp
->retrans_out
) {
2987 printk(KERN_DEBUG
"Leak r=%u %d\n",
2988 tp
->retrans_out
, icsk
->icsk_ca_state
);
2989 tp
->retrans_out
= 0;
2996 static void tcp_ack_probe(struct sock
*sk
)
2998 const struct tcp_sock
*tp
= tcp_sk(sk
);
2999 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3001 /* Was it a usable window open? */
3003 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3004 icsk
->icsk_backoff
= 0;
3005 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3006 /* Socket must be waked up by subsequent tcp_data_snd_check().
3007 * This function is not for random using!
3010 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3011 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3016 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3018 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3019 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3022 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3024 const struct tcp_sock
*tp
= tcp_sk(sk
);
3025 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3026 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3029 /* Check that window update is acceptable.
3030 * The function assumes that snd_una<=ack<=snd_next.
3032 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3033 const u32 ack
, const u32 ack_seq
,
3036 return (after(ack
, tp
->snd_una
) ||
3037 after(ack_seq
, tp
->snd_wl1
) ||
3038 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3041 /* Update our send window.
3043 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3044 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3046 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3049 struct tcp_sock
*tp
= tcp_sk(sk
);
3051 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3053 if (likely(!tcp_hdr(skb
)->syn
))
3054 nwin
<<= tp
->rx_opt
.snd_wscale
;
3056 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3057 flag
|= FLAG_WIN_UPDATE
;
3058 tcp_update_wl(tp
, ack
, ack_seq
);
3060 if (tp
->snd_wnd
!= nwin
) {
3063 /* Note, it is the only place, where
3064 * fast path is recovered for sending TCP.
3067 tcp_fast_path_check(sk
);
3069 if (nwin
> tp
->max_window
) {
3070 tp
->max_window
= nwin
;
3071 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3081 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3082 * continue in congestion avoidance.
3084 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3086 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3087 tp
->snd_cwnd_cnt
= 0;
3088 tp
->bytes_acked
= 0;
3089 TCP_ECN_queue_cwr(tp
);
3090 tcp_moderate_cwnd(tp
);
3093 /* A conservative spurious RTO response algorithm: reduce cwnd using
3094 * rate halving and continue in congestion avoidance.
3096 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3098 tcp_enter_cwr(sk
, 0);
3101 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3103 if (flag
& FLAG_ECE
)
3104 tcp_ratehalving_spur_to_response(sk
);
3106 tcp_undo_cwr(sk
, 1);
3109 /* F-RTO spurious RTO detection algorithm (RFC4138)
3111 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3112 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3113 * window (but not to or beyond highest sequence sent before RTO):
3114 * On First ACK, send two new segments out.
3115 * On Second ACK, RTO was likely spurious. Do spurious response (response
3116 * algorithm is not part of the F-RTO detection algorithm
3117 * given in RFC4138 but can be selected separately).
3118 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3119 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3120 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3121 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3123 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3124 * original window even after we transmit two new data segments.
3127 * on first step, wait until first cumulative ACK arrives, then move to
3128 * the second step. In second step, the next ACK decides.
3130 * F-RTO is implemented (mainly) in four functions:
3131 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3132 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3133 * called when tcp_use_frto() showed green light
3134 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3135 * - tcp_enter_frto_loss() is called if there is not enough evidence
3136 * to prove that the RTO is indeed spurious. It transfers the control
3137 * from F-RTO to the conventional RTO recovery
3139 static int tcp_process_frto(struct sock
*sk
, int flag
)
3141 struct tcp_sock
*tp
= tcp_sk(sk
);
3143 tcp_verify_left_out(tp
);
3145 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3146 if (flag
& FLAG_DATA_ACKED
)
3147 inet_csk(sk
)->icsk_retransmits
= 0;
3149 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3150 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3151 tp
->undo_marker
= 0;
3153 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3154 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3158 if (!tcp_is_sackfrto(tp
)) {
3159 /* RFC4138 shortcoming in step 2; should also have case c):
3160 * ACK isn't duplicate nor advances window, e.g., opposite dir
3163 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3166 if (!(flag
& FLAG_DATA_ACKED
)) {
3167 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3172 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3173 /* Prevent sending of new data. */
3174 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3175 tcp_packets_in_flight(tp
));
3179 if ((tp
->frto_counter
>= 2) &&
3180 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3181 ((flag
& FLAG_DATA_SACKED
) &&
3182 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3183 /* RFC4138 shortcoming (see comment above) */
3184 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3185 (flag
& FLAG_NOT_DUP
))
3188 tcp_enter_frto_loss(sk
, 3, flag
);
3193 if (tp
->frto_counter
== 1) {
3194 /* tcp_may_send_now needs to see updated state */
3195 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3196 tp
->frto_counter
= 2;
3198 if (!tcp_may_send_now(sk
))
3199 tcp_enter_frto_loss(sk
, 2, flag
);
3203 switch (sysctl_tcp_frto_response
) {
3205 tcp_undo_spur_to_response(sk
, flag
);
3208 tcp_conservative_spur_to_response(tp
);
3211 tcp_ratehalving_spur_to_response(sk
);
3214 tp
->frto_counter
= 0;
3215 tp
->undo_marker
= 0;
3216 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3221 /* This routine deals with incoming acks, but not outgoing ones. */
3222 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3224 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3225 struct tcp_sock
*tp
= tcp_sk(sk
);
3226 u32 prior_snd_una
= tp
->snd_una
;
3227 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3228 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3229 u32 prior_in_flight
;
3234 /* If the ack is newer than sent or older than previous acks
3235 * then we can probably ignore it.
3237 if (after(ack
, tp
->snd_nxt
))
3238 goto uninteresting_ack
;
3240 if (before(ack
, prior_snd_una
))
3243 if (after(ack
, prior_snd_una
))
3244 flag
|= FLAG_SND_UNA_ADVANCED
;
3246 if (sysctl_tcp_abc
) {
3247 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3248 tp
->bytes_acked
+= ack
- prior_snd_una
;
3249 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3250 /* we assume just one segment left network */
3251 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3255 prior_fackets
= tp
->fackets_out
;
3256 prior_in_flight
= tcp_packets_in_flight(tp
);
3258 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3259 /* Window is constant, pure forward advance.
3260 * No more checks are required.
3261 * Note, we use the fact that SND.UNA>=SND.WL2.
3263 tcp_update_wl(tp
, ack
, ack_seq
);
3265 flag
|= FLAG_WIN_UPDATE
;
3267 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3269 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3271 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3274 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3276 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3278 if (TCP_SKB_CB(skb
)->sacked
)
3279 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3281 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3284 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3287 /* We passed data and got it acked, remove any soft error
3288 * log. Something worked...
3290 sk
->sk_err_soft
= 0;
3291 icsk
->icsk_probes_out
= 0;
3292 tp
->rcv_tstamp
= tcp_time_stamp
;
3293 prior_packets
= tp
->packets_out
;
3297 /* See if we can take anything off of the retransmit queue. */
3298 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
);
3300 if (tp
->frto_counter
)
3301 frto_cwnd
= tcp_process_frto(sk
, flag
);
3302 /* Guarantee sacktag reordering detection against wrap-arounds */
3303 if (before(tp
->frto_highmark
, tp
->snd_una
))
3304 tp
->frto_highmark
= 0;
3306 if (tcp_ack_is_dubious(sk
, flag
)) {
3307 /* Advance CWND, if state allows this. */
3308 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3309 tcp_may_raise_cwnd(sk
, flag
))
3310 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3311 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3314 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3315 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3318 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3319 dst_confirm(sk
->sk_dst_cache
);
3324 /* If this ack opens up a zero window, clear backoff. It was
3325 * being used to time the probes, and is probably far higher than
3326 * it needs to be for normal retransmission.
3328 if (tcp_send_head(sk
))
3333 if (TCP_SKB_CB(skb
)->sacked
) {
3334 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3335 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3336 tcp_try_keep_open(sk
);
3340 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3344 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3345 * But, this can also be called on packets in the established flow when
3346 * the fast version below fails.
3348 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3352 struct tcphdr
*th
= tcp_hdr(skb
);
3353 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3355 ptr
= (unsigned char *)(th
+ 1);
3356 opt_rx
->saw_tstamp
= 0;
3358 while (length
> 0) {
3359 int opcode
= *ptr
++;
3365 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3370 if (opsize
< 2) /* "silly options" */
3372 if (opsize
> length
)
3373 return; /* don't parse partial options */
3376 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3377 u16 in_mss
= get_unaligned_be16(ptr
);
3379 if (opt_rx
->user_mss
&&
3380 opt_rx
->user_mss
< in_mss
)
3381 in_mss
= opt_rx
->user_mss
;
3382 opt_rx
->mss_clamp
= in_mss
;
3387 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3388 !estab
&& sysctl_tcp_window_scaling
) {
3389 __u8 snd_wscale
= *(__u8
*)ptr
;
3390 opt_rx
->wscale_ok
= 1;
3391 if (snd_wscale
> 14) {
3392 if (net_ratelimit())
3393 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3394 "scaling value %d >14 received.\n",
3398 opt_rx
->snd_wscale
= snd_wscale
;
3401 case TCPOPT_TIMESTAMP
:
3402 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3403 ((estab
&& opt_rx
->tstamp_ok
) ||
3404 (!estab
&& sysctl_tcp_timestamps
))) {
3405 opt_rx
->saw_tstamp
= 1;
3406 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3407 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3410 case TCPOPT_SACK_PERM
:
3411 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3412 !estab
&& sysctl_tcp_sack
) {
3413 opt_rx
->sack_ok
= 1;
3414 tcp_sack_reset(opt_rx
);
3419 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3420 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3422 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3425 #ifdef CONFIG_TCP_MD5SIG
3428 * The MD5 Hash has already been
3429 * checked (see tcp_v{4,6}_do_rcv()).
3441 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3443 __be32
*ptr
= (__be32
*)(th
+ 1);
3445 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3446 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3447 tp
->rx_opt
.saw_tstamp
= 1;
3449 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3451 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3457 /* Fast parse options. This hopes to only see timestamps.
3458 * If it is wrong it falls back on tcp_parse_options().
3460 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3461 struct tcp_sock
*tp
)
3463 if (th
->doff
== sizeof(struct tcphdr
) >> 2) {
3464 tp
->rx_opt
.saw_tstamp
= 0;
3466 } else if (tp
->rx_opt
.tstamp_ok
&&
3467 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3468 if (tcp_parse_aligned_timestamp(tp
, th
))
3471 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3475 #ifdef CONFIG_TCP_MD5SIG
3477 * Parse MD5 Signature option
3479 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3481 int length
= (th
->doff
<< 2) - sizeof (*th
);
3482 u8
*ptr
= (u8
*)(th
+ 1);
3484 /* If the TCP option is too short, we can short cut */
3485 if (length
< TCPOLEN_MD5SIG
)
3488 while (length
> 0) {
3489 int opcode
= *ptr
++;
3500 if (opsize
< 2 || opsize
> length
)
3502 if (opcode
== TCPOPT_MD5SIG
)
3512 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3514 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3515 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3518 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3520 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3521 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3522 * extra check below makes sure this can only happen
3523 * for pure ACK frames. -DaveM
3525 * Not only, also it occurs for expired timestamps.
3528 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3529 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3530 tcp_store_ts_recent(tp
);
3534 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3536 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3537 * it can pass through stack. So, the following predicate verifies that
3538 * this segment is not used for anything but congestion avoidance or
3539 * fast retransmit. Moreover, we even are able to eliminate most of such
3540 * second order effects, if we apply some small "replay" window (~RTO)
3541 * to timestamp space.
3543 * All these measures still do not guarantee that we reject wrapped ACKs
3544 * on networks with high bandwidth, when sequence space is recycled fastly,
3545 * but it guarantees that such events will be very rare and do not affect
3546 * connection seriously. This doesn't look nice, but alas, PAWS is really
3549 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3550 * states that events when retransmit arrives after original data are rare.
3551 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3552 * the biggest problem on large power networks even with minor reordering.
3553 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3554 * up to bandwidth of 18Gigabit/sec. 8) ]
3557 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3559 struct tcp_sock
*tp
= tcp_sk(sk
);
3560 struct tcphdr
*th
= tcp_hdr(skb
);
3561 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3562 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3564 return (/* 1. Pure ACK with correct sequence number. */
3565 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3567 /* 2. ... and duplicate ACK. */
3568 ack
== tp
->snd_una
&&
3570 /* 3. ... and does not update window. */
3571 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3573 /* 4. ... and sits in replay window. */
3574 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3577 static inline int tcp_paws_discard(const struct sock
*sk
,
3578 const struct sk_buff
*skb
)
3580 const struct tcp_sock
*tp
= tcp_sk(sk
);
3581 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3582 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3583 !tcp_disordered_ack(sk
, skb
));
3586 /* Check segment sequence number for validity.
3588 * Segment controls are considered valid, if the segment
3589 * fits to the window after truncation to the window. Acceptability
3590 * of data (and SYN, FIN, of course) is checked separately.
3591 * See tcp_data_queue(), for example.
3593 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3594 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3595 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3596 * (borrowed from freebsd)
3599 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3601 return !before(end_seq
, tp
->rcv_wup
) &&
3602 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3605 /* When we get a reset we do this. */
3606 static void tcp_reset(struct sock
*sk
)
3608 /* We want the right error as BSD sees it (and indeed as we do). */
3609 switch (sk
->sk_state
) {
3611 sk
->sk_err
= ECONNREFUSED
;
3613 case TCP_CLOSE_WAIT
:
3619 sk
->sk_err
= ECONNRESET
;
3622 if (!sock_flag(sk
, SOCK_DEAD
))
3623 sk
->sk_error_report(sk
);
3629 * Process the FIN bit. This now behaves as it is supposed to work
3630 * and the FIN takes effect when it is validly part of sequence
3631 * space. Not before when we get holes.
3633 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3634 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3637 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3638 * close and we go into CLOSING (and later onto TIME-WAIT)
3640 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3642 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3644 struct tcp_sock
*tp
= tcp_sk(sk
);
3646 inet_csk_schedule_ack(sk
);
3648 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3649 sock_set_flag(sk
, SOCK_DONE
);
3651 switch (sk
->sk_state
) {
3653 case TCP_ESTABLISHED
:
3654 /* Move to CLOSE_WAIT */
3655 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3656 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3659 case TCP_CLOSE_WAIT
:
3661 /* Received a retransmission of the FIN, do
3666 /* RFC793: Remain in the LAST-ACK state. */
3670 /* This case occurs when a simultaneous close
3671 * happens, we must ack the received FIN and
3672 * enter the CLOSING state.
3675 tcp_set_state(sk
, TCP_CLOSING
);
3678 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3680 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3683 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3684 * cases we should never reach this piece of code.
3686 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3687 __func__
, sk
->sk_state
);
3691 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3692 * Probably, we should reset in this case. For now drop them.
3694 __skb_queue_purge(&tp
->out_of_order_queue
);
3695 if (tcp_is_sack(tp
))
3696 tcp_sack_reset(&tp
->rx_opt
);
3699 if (!sock_flag(sk
, SOCK_DEAD
)) {
3700 sk
->sk_state_change(sk
);
3702 /* Do not send POLL_HUP for half duplex close. */
3703 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3704 sk
->sk_state
== TCP_CLOSE
)
3705 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3707 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3711 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3714 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3715 if (before(seq
, sp
->start_seq
))
3716 sp
->start_seq
= seq
;
3717 if (after(end_seq
, sp
->end_seq
))
3718 sp
->end_seq
= end_seq
;
3724 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3726 struct tcp_sock
*tp
= tcp_sk(sk
);
3728 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3731 if (before(seq
, tp
->rcv_nxt
))
3732 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3734 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3736 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3738 tp
->rx_opt
.dsack
= 1;
3739 tp
->duplicate_sack
[0].start_seq
= seq
;
3740 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3741 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ 1;
3745 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3747 struct tcp_sock
*tp
= tcp_sk(sk
);
3749 if (!tp
->rx_opt
.dsack
)
3750 tcp_dsack_set(sk
, seq
, end_seq
);
3752 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3755 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3757 struct tcp_sock
*tp
= tcp_sk(sk
);
3759 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3760 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3761 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3762 tcp_enter_quickack_mode(sk
);
3764 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3765 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3767 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3768 end_seq
= tp
->rcv_nxt
;
3769 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3776 /* These routines update the SACK block as out-of-order packets arrive or
3777 * in-order packets close up the sequence space.
3779 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3782 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3783 struct tcp_sack_block
*swalk
= sp
+ 1;
3785 /* See if the recent change to the first SACK eats into
3786 * or hits the sequence space of other SACK blocks, if so coalesce.
3788 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3789 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3792 /* Zap SWALK, by moving every further SACK up by one slot.
3793 * Decrease num_sacks.
3795 tp
->rx_opt
.num_sacks
--;
3796 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
3798 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3802 this_sack
++, swalk
++;
3806 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
,
3807 struct tcp_sack_block
*sack2
)
3811 tmp
= sack1
->start_seq
;
3812 sack1
->start_seq
= sack2
->start_seq
;
3813 sack2
->start_seq
= tmp
;
3815 tmp
= sack1
->end_seq
;
3816 sack1
->end_seq
= sack2
->end_seq
;
3817 sack2
->end_seq
= tmp
;
3820 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3822 struct tcp_sock
*tp
= tcp_sk(sk
);
3823 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3824 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3830 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3831 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3832 /* Rotate this_sack to the first one. */
3833 for (; this_sack
> 0; this_sack
--, sp
--)
3834 tcp_sack_swap(sp
, sp
- 1);
3836 tcp_sack_maybe_coalesce(tp
);
3841 /* Could not find an adjacent existing SACK, build a new one,
3842 * put it at the front, and shift everyone else down. We
3843 * always know there is at least one SACK present already here.
3845 * If the sack array is full, forget about the last one.
3847 if (this_sack
>= TCP_NUM_SACKS
) {
3849 tp
->rx_opt
.num_sacks
--;
3852 for (; this_sack
> 0; this_sack
--, sp
--)
3856 /* Build the new head SACK, and we're done. */
3857 sp
->start_seq
= seq
;
3858 sp
->end_seq
= end_seq
;
3859 tp
->rx_opt
.num_sacks
++;
3860 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
3863 /* RCV.NXT advances, some SACKs should be eaten. */
3865 static void tcp_sack_remove(struct tcp_sock
*tp
)
3867 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3868 int num_sacks
= tp
->rx_opt
.num_sacks
;
3871 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3872 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3873 tp
->rx_opt
.num_sacks
= 0;
3874 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3878 for (this_sack
= 0; this_sack
< num_sacks
;) {
3879 /* Check if the start of the sack is covered by RCV.NXT. */
3880 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3883 /* RCV.NXT must cover all the block! */
3884 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
3886 /* Zap this SACK, by moving forward any other SACKS. */
3887 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3888 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3895 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3896 tp
->rx_opt
.num_sacks
= num_sacks
;
3897 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
3902 /* This one checks to see if we can put data from the
3903 * out_of_order queue into the receive_queue.
3905 static void tcp_ofo_queue(struct sock
*sk
)
3907 struct tcp_sock
*tp
= tcp_sk(sk
);
3908 __u32 dsack_high
= tp
->rcv_nxt
;
3909 struct sk_buff
*skb
;
3911 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3912 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3915 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3916 __u32 dsack
= dsack_high
;
3917 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3918 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3919 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
3922 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3923 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3924 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3928 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3929 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3930 TCP_SKB_CB(skb
)->end_seq
);
3932 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3933 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3934 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3935 if (tcp_hdr(skb
)->fin
)
3936 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3940 static int tcp_prune_ofo_queue(struct sock
*sk
);
3941 static int tcp_prune_queue(struct sock
*sk
);
3943 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
3945 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3946 !sk_rmem_schedule(sk
, size
)) {
3948 if (tcp_prune_queue(sk
) < 0)
3951 if (!sk_rmem_schedule(sk
, size
)) {
3952 if (!tcp_prune_ofo_queue(sk
))
3955 if (!sk_rmem_schedule(sk
, size
))
3962 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3964 struct tcphdr
*th
= tcp_hdr(skb
);
3965 struct tcp_sock
*tp
= tcp_sk(sk
);
3968 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3971 __skb_pull(skb
, th
->doff
* 4);
3973 TCP_ECN_accept_cwr(tp
, skb
);
3975 if (tp
->rx_opt
.dsack
) {
3976 tp
->rx_opt
.dsack
= 0;
3977 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
;
3980 /* Queue data for delivery to the user.
3981 * Packets in sequence go to the receive queue.
3982 * Out of sequence packets to the out_of_order_queue.
3984 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3985 if (tcp_receive_window(tp
) == 0)
3988 /* Ok. In sequence. In window. */
3989 if (tp
->ucopy
.task
== current
&&
3990 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3991 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3992 int chunk
= min_t(unsigned int, skb
->len
,
3995 __set_current_state(TASK_RUNNING
);
3998 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
3999 tp
->ucopy
.len
-= chunk
;
4000 tp
->copied_seq
+= chunk
;
4001 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4002 tcp_rcv_space_adjust(sk
);
4010 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4013 skb_set_owner_r(skb
, sk
);
4014 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4016 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4018 tcp_event_data_recv(sk
, skb
);
4020 tcp_fin(skb
, sk
, th
);
4022 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4025 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4026 * gap in queue is filled.
4028 if (skb_queue_empty(&tp
->out_of_order_queue
))
4029 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4032 if (tp
->rx_opt
.num_sacks
)
4033 tcp_sack_remove(tp
);
4035 tcp_fast_path_check(sk
);
4039 else if (!sock_flag(sk
, SOCK_DEAD
))
4040 sk
->sk_data_ready(sk
, 0);
4044 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4045 /* A retransmit, 2nd most common case. Force an immediate ack. */
4046 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4047 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4050 tcp_enter_quickack_mode(sk
);
4051 inet_csk_schedule_ack(sk
);
4057 /* Out of window. F.e. zero window probe. */
4058 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4061 tcp_enter_quickack_mode(sk
);
4063 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4064 /* Partial packet, seq < rcv_next < end_seq */
4065 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4066 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4067 TCP_SKB_CB(skb
)->end_seq
);
4069 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4071 /* If window is closed, drop tail of packet. But after
4072 * remembering D-SACK for its head made in previous line.
4074 if (!tcp_receive_window(tp
))
4079 TCP_ECN_check_ce(tp
, skb
);
4081 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4084 /* Disable header prediction. */
4086 inet_csk_schedule_ack(sk
);
4088 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4089 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4091 skb_set_owner_r(skb
, sk
);
4093 if (!skb_peek(&tp
->out_of_order_queue
)) {
4094 /* Initial out of order segment, build 1 SACK. */
4095 if (tcp_is_sack(tp
)) {
4096 tp
->rx_opt
.num_sacks
= 1;
4097 tp
->rx_opt
.dsack
= 0;
4098 tp
->rx_opt
.eff_sacks
= 1;
4099 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4100 tp
->selective_acks
[0].end_seq
=
4101 TCP_SKB_CB(skb
)->end_seq
;
4103 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4105 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
4106 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4107 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4109 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4110 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4112 if (!tp
->rx_opt
.num_sacks
||
4113 tp
->selective_acks
[0].end_seq
!= seq
)
4116 /* Common case: data arrive in order after hole. */
4117 tp
->selective_acks
[0].end_seq
= end_seq
;
4121 /* Find place to insert this segment. */
4123 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4125 } while ((skb1
= skb1
->prev
) !=
4126 (struct sk_buff
*)&tp
->out_of_order_queue
);
4128 /* Do skb overlap to previous one? */
4129 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
4130 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4131 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4132 /* All the bits are present. Drop. */
4134 tcp_dsack_set(sk
, seq
, end_seq
);
4137 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4138 /* Partial overlap. */
4139 tcp_dsack_set(sk
, seq
,
4140 TCP_SKB_CB(skb1
)->end_seq
);
4145 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
4147 /* And clean segments covered by new one as whole. */
4148 while ((skb1
= skb
->next
) !=
4149 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4150 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4151 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4152 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4156 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4157 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4158 TCP_SKB_CB(skb1
)->end_seq
);
4163 if (tcp_is_sack(tp
))
4164 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4168 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4169 struct sk_buff_head
*list
)
4171 struct sk_buff
*next
= skb
->next
;
4173 __skb_unlink(skb
, list
);
4175 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4180 /* Collapse contiguous sequence of skbs head..tail with
4181 * sequence numbers start..end.
4182 * Segments with FIN/SYN are not collapsed (only because this
4186 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4187 struct sk_buff
*head
, struct sk_buff
*tail
,
4190 struct sk_buff
*skb
;
4192 /* First, check that queue is collapsible and find
4193 * the point where collapsing can be useful. */
4194 for (skb
= head
; skb
!= tail
;) {
4195 /* No new bits? It is possible on ofo queue. */
4196 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4197 skb
= tcp_collapse_one(sk
, skb
, list
);
4201 /* The first skb to collapse is:
4203 * - bloated or contains data before "start" or
4204 * overlaps to the next one.
4206 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4207 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4208 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4209 (skb
->next
!= tail
&&
4210 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4213 /* Decided to skip this, advance start seq. */
4214 start
= TCP_SKB_CB(skb
)->end_seq
;
4217 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4220 while (before(start
, end
)) {
4221 struct sk_buff
*nskb
;
4222 unsigned int header
= skb_headroom(skb
);
4223 int copy
= SKB_MAX_ORDER(header
, 0);
4225 /* Too big header? This can happen with IPv6. */
4228 if (end
- start
< copy
)
4230 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4234 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4235 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4237 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4239 skb_reserve(nskb
, header
);
4240 memcpy(nskb
->head
, skb
->head
, header
);
4241 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4242 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4243 __skb_insert(nskb
, skb
->prev
, skb
, list
);
4244 skb_set_owner_r(nskb
, sk
);
4246 /* Copy data, releasing collapsed skbs. */
4248 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4249 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4253 size
= min(copy
, size
);
4254 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4256 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4260 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4261 skb
= tcp_collapse_one(sk
, skb
, list
);
4263 tcp_hdr(skb
)->syn
||
4271 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4272 * and tcp_collapse() them until all the queue is collapsed.
4274 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4276 struct tcp_sock
*tp
= tcp_sk(sk
);
4277 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4278 struct sk_buff
*head
;
4284 start
= TCP_SKB_CB(skb
)->seq
;
4285 end
= TCP_SKB_CB(skb
)->end_seq
;
4291 /* Segment is terminated when we see gap or when
4292 * we are at the end of all the queue. */
4293 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4294 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4295 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4296 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4297 head
, skb
, start
, end
);
4299 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4301 /* Start new segment */
4302 start
= TCP_SKB_CB(skb
)->seq
;
4303 end
= TCP_SKB_CB(skb
)->end_seq
;
4305 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4306 start
= TCP_SKB_CB(skb
)->seq
;
4307 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4308 end
= TCP_SKB_CB(skb
)->end_seq
;
4314 * Purge the out-of-order queue.
4315 * Return true if queue was pruned.
4317 static int tcp_prune_ofo_queue(struct sock
*sk
)
4319 struct tcp_sock
*tp
= tcp_sk(sk
);
4322 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4323 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4324 __skb_queue_purge(&tp
->out_of_order_queue
);
4326 /* Reset SACK state. A conforming SACK implementation will
4327 * do the same at a timeout based retransmit. When a connection
4328 * is in a sad state like this, we care only about integrity
4329 * of the connection not performance.
4331 if (tp
->rx_opt
.sack_ok
)
4332 tcp_sack_reset(&tp
->rx_opt
);
4339 /* Reduce allocated memory if we can, trying to get
4340 * the socket within its memory limits again.
4342 * Return less than zero if we should start dropping frames
4343 * until the socket owning process reads some of the data
4344 * to stabilize the situation.
4346 static int tcp_prune_queue(struct sock
*sk
)
4348 struct tcp_sock
*tp
= tcp_sk(sk
);
4350 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4352 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4354 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4355 tcp_clamp_window(sk
);
4356 else if (tcp_memory_pressure
)
4357 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4359 tcp_collapse_ofo_queue(sk
);
4360 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4361 sk
->sk_receive_queue
.next
,
4362 (struct sk_buff
*)&sk
->sk_receive_queue
,
4363 tp
->copied_seq
, tp
->rcv_nxt
);
4366 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4369 /* Collapsing did not help, destructive actions follow.
4370 * This must not ever occur. */
4372 tcp_prune_ofo_queue(sk
);
4374 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4377 /* If we are really being abused, tell the caller to silently
4378 * drop receive data on the floor. It will get retransmitted
4379 * and hopefully then we'll have sufficient space.
4381 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4383 /* Massive buffer overcommit. */
4388 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4389 * As additional protections, we do not touch cwnd in retransmission phases,
4390 * and if application hit its sndbuf limit recently.
4392 void tcp_cwnd_application_limited(struct sock
*sk
)
4394 struct tcp_sock
*tp
= tcp_sk(sk
);
4396 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4397 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4398 /* Limited by application or receiver window. */
4399 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4400 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4401 if (win_used
< tp
->snd_cwnd
) {
4402 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4403 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4405 tp
->snd_cwnd_used
= 0;
4407 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4410 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4412 struct tcp_sock
*tp
= tcp_sk(sk
);
4414 /* If the user specified a specific send buffer setting, do
4417 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4420 /* If we are under global TCP memory pressure, do not expand. */
4421 if (tcp_memory_pressure
)
4424 /* If we are under soft global TCP memory pressure, do not expand. */
4425 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4428 /* If we filled the congestion window, do not expand. */
4429 if (tp
->packets_out
>= tp
->snd_cwnd
)
4435 /* When incoming ACK allowed to free some skb from write_queue,
4436 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4437 * on the exit from tcp input handler.
4439 * PROBLEM: sndbuf expansion does not work well with largesend.
4441 static void tcp_new_space(struct sock
*sk
)
4443 struct tcp_sock
*tp
= tcp_sk(sk
);
4445 if (tcp_should_expand_sndbuf(sk
)) {
4446 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4447 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4448 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4449 tp
->reordering
+ 1);
4450 sndmem
*= 2 * demanded
;
4451 if (sndmem
> sk
->sk_sndbuf
)
4452 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4453 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4456 sk
->sk_write_space(sk
);
4459 static void tcp_check_space(struct sock
*sk
)
4461 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4462 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4463 if (sk
->sk_socket
&&
4464 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4469 static inline void tcp_data_snd_check(struct sock
*sk
)
4471 tcp_push_pending_frames(sk
);
4472 tcp_check_space(sk
);
4476 * Check if sending an ack is needed.
4478 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4480 struct tcp_sock
*tp
= tcp_sk(sk
);
4482 /* More than one full frame received... */
4483 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4484 /* ... and right edge of window advances far enough.
4485 * (tcp_recvmsg() will send ACK otherwise). Or...
4487 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4488 /* We ACK each frame or... */
4489 tcp_in_quickack_mode(sk
) ||
4490 /* We have out of order data. */
4491 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4492 /* Then ack it now */
4495 /* Else, send delayed ack. */
4496 tcp_send_delayed_ack(sk
);
4500 static inline void tcp_ack_snd_check(struct sock
*sk
)
4502 if (!inet_csk_ack_scheduled(sk
)) {
4503 /* We sent a data segment already. */
4506 __tcp_ack_snd_check(sk
, 1);
4510 * This routine is only called when we have urgent data
4511 * signaled. Its the 'slow' part of tcp_urg. It could be
4512 * moved inline now as tcp_urg is only called from one
4513 * place. We handle URGent data wrong. We have to - as
4514 * BSD still doesn't use the correction from RFC961.
4515 * For 1003.1g we should support a new option TCP_STDURG to permit
4516 * either form (or just set the sysctl tcp_stdurg).
4519 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4521 struct tcp_sock
*tp
= tcp_sk(sk
);
4522 u32 ptr
= ntohs(th
->urg_ptr
);
4524 if (ptr
&& !sysctl_tcp_stdurg
)
4526 ptr
+= ntohl(th
->seq
);
4528 /* Ignore urgent data that we've already seen and read. */
4529 if (after(tp
->copied_seq
, ptr
))
4532 /* Do not replay urg ptr.
4534 * NOTE: interesting situation not covered by specs.
4535 * Misbehaving sender may send urg ptr, pointing to segment,
4536 * which we already have in ofo queue. We are not able to fetch
4537 * such data and will stay in TCP_URG_NOTYET until will be eaten
4538 * by recvmsg(). Seems, we are not obliged to handle such wicked
4539 * situations. But it is worth to think about possibility of some
4540 * DoSes using some hypothetical application level deadlock.
4542 if (before(ptr
, tp
->rcv_nxt
))
4545 /* Do we already have a newer (or duplicate) urgent pointer? */
4546 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4549 /* Tell the world about our new urgent pointer. */
4552 /* We may be adding urgent data when the last byte read was
4553 * urgent. To do this requires some care. We cannot just ignore
4554 * tp->copied_seq since we would read the last urgent byte again
4555 * as data, nor can we alter copied_seq until this data arrives
4556 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4558 * NOTE. Double Dutch. Rendering to plain English: author of comment
4559 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4560 * and expect that both A and B disappear from stream. This is _wrong_.
4561 * Though this happens in BSD with high probability, this is occasional.
4562 * Any application relying on this is buggy. Note also, that fix "works"
4563 * only in this artificial test. Insert some normal data between A and B and we will
4564 * decline of BSD again. Verdict: it is better to remove to trap
4567 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4568 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4569 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4571 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4572 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4577 tp
->urg_data
= TCP_URG_NOTYET
;
4580 /* Disable header prediction. */
4584 /* This is the 'fast' part of urgent handling. */
4585 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4587 struct tcp_sock
*tp
= tcp_sk(sk
);
4589 /* Check if we get a new urgent pointer - normally not. */
4591 tcp_check_urg(sk
, th
);
4593 /* Do we wait for any urgent data? - normally not... */
4594 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4595 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4598 /* Is the urgent pointer pointing into this packet? */
4599 if (ptr
< skb
->len
) {
4601 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4603 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4604 if (!sock_flag(sk
, SOCK_DEAD
))
4605 sk
->sk_data_ready(sk
, 0);
4610 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4612 struct tcp_sock
*tp
= tcp_sk(sk
);
4613 int chunk
= skb
->len
- hlen
;
4617 if (skb_csum_unnecessary(skb
))
4618 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4620 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4624 tp
->ucopy
.len
-= chunk
;
4625 tp
->copied_seq
+= chunk
;
4626 tcp_rcv_space_adjust(sk
);
4633 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4634 struct sk_buff
*skb
)
4638 if (sock_owned_by_user(sk
)) {
4640 result
= __tcp_checksum_complete(skb
);
4643 result
= __tcp_checksum_complete(skb
);
4648 static inline int tcp_checksum_complete_user(struct sock
*sk
,
4649 struct sk_buff
*skb
)
4651 return !skb_csum_unnecessary(skb
) &&
4652 __tcp_checksum_complete_user(sk
, skb
);
4655 #ifdef CONFIG_NET_DMA
4656 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4659 struct tcp_sock
*tp
= tcp_sk(sk
);
4660 int chunk
= skb
->len
- hlen
;
4662 int copied_early
= 0;
4664 if (tp
->ucopy
.wakeup
)
4667 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4668 tp
->ucopy
.dma_chan
= get_softnet_dma();
4670 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4672 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4674 tp
->ucopy
.iov
, chunk
,
4675 tp
->ucopy
.pinned_list
);
4680 tp
->ucopy
.dma_cookie
= dma_cookie
;
4683 tp
->ucopy
.len
-= chunk
;
4684 tp
->copied_seq
+= chunk
;
4685 tcp_rcv_space_adjust(sk
);
4687 if ((tp
->ucopy
.len
== 0) ||
4688 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4689 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4690 tp
->ucopy
.wakeup
= 1;
4691 sk
->sk_data_ready(sk
, 0);
4693 } else if (chunk
> 0) {
4694 tp
->ucopy
.wakeup
= 1;
4695 sk
->sk_data_ready(sk
, 0);
4698 return copied_early
;
4700 #endif /* CONFIG_NET_DMA */
4702 /* Does PAWS and seqno based validation of an incoming segment, flags will
4703 * play significant role here.
4705 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
4706 struct tcphdr
*th
, int syn_inerr
)
4708 struct tcp_sock
*tp
= tcp_sk(sk
);
4710 /* RFC1323: H1. Apply PAWS check first. */
4711 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4712 tcp_paws_discard(sk
, skb
)) {
4714 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
4715 tcp_send_dupack(sk
, skb
);
4718 /* Reset is accepted even if it did not pass PAWS. */
4721 /* Step 1: check sequence number */
4722 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4723 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4724 * (RST) segments are validated by checking their SEQ-fields."
4725 * And page 69: "If an incoming segment is not acceptable,
4726 * an acknowledgment should be sent in reply (unless the RST
4727 * bit is set, if so drop the segment and return)".
4730 tcp_send_dupack(sk
, skb
);
4734 /* Step 2: check RST bit */
4740 /* ts_recent update must be made after we are sure that the packet
4743 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4745 /* step 3: check security and precedence [ignored] */
4747 /* step 4: Check for a SYN in window. */
4748 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4750 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4751 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
4764 * TCP receive function for the ESTABLISHED state.
4766 * It is split into a fast path and a slow path. The fast path is
4768 * - A zero window was announced from us - zero window probing
4769 * is only handled properly in the slow path.
4770 * - Out of order segments arrived.
4771 * - Urgent data is expected.
4772 * - There is no buffer space left
4773 * - Unexpected TCP flags/window values/header lengths are received
4774 * (detected by checking the TCP header against pred_flags)
4775 * - Data is sent in both directions. Fast path only supports pure senders
4776 * or pure receivers (this means either the sequence number or the ack
4777 * value must stay constant)
4778 * - Unexpected TCP option.
4780 * When these conditions are not satisfied it drops into a standard
4781 * receive procedure patterned after RFC793 to handle all cases.
4782 * The first three cases are guaranteed by proper pred_flags setting,
4783 * the rest is checked inline. Fast processing is turned on in
4784 * tcp_data_queue when everything is OK.
4786 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4787 struct tcphdr
*th
, unsigned len
)
4789 struct tcp_sock
*tp
= tcp_sk(sk
);
4793 * Header prediction.
4794 * The code loosely follows the one in the famous
4795 * "30 instruction TCP receive" Van Jacobson mail.
4797 * Van's trick is to deposit buffers into socket queue
4798 * on a device interrupt, to call tcp_recv function
4799 * on the receive process context and checksum and copy
4800 * the buffer to user space. smart...
4802 * Our current scheme is not silly either but we take the
4803 * extra cost of the net_bh soft interrupt processing...
4804 * We do checksum and copy also but from device to kernel.
4807 tp
->rx_opt
.saw_tstamp
= 0;
4809 /* pred_flags is 0xS?10 << 16 + snd_wnd
4810 * if header_prediction is to be made
4811 * 'S' will always be tp->tcp_header_len >> 2
4812 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4813 * turn it off (when there are holes in the receive
4814 * space for instance)
4815 * PSH flag is ignored.
4818 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4819 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4820 int tcp_header_len
= tp
->tcp_header_len
;
4822 /* Timestamp header prediction: tcp_header_len
4823 * is automatically equal to th->doff*4 due to pred_flags
4827 /* Check timestamp */
4828 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4829 /* No? Slow path! */
4830 if (!tcp_parse_aligned_timestamp(tp
, th
))
4833 /* If PAWS failed, check it more carefully in slow path */
4834 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4837 /* DO NOT update ts_recent here, if checksum fails
4838 * and timestamp was corrupted part, it will result
4839 * in a hung connection since we will drop all
4840 * future packets due to the PAWS test.
4844 if (len
<= tcp_header_len
) {
4845 /* Bulk data transfer: sender */
4846 if (len
== tcp_header_len
) {
4847 /* Predicted packet is in window by definition.
4848 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4849 * Hence, check seq<=rcv_wup reduces to:
4851 if (tcp_header_len
==
4852 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4853 tp
->rcv_nxt
== tp
->rcv_wup
)
4854 tcp_store_ts_recent(tp
);
4856 /* We know that such packets are checksummed
4859 tcp_ack(sk
, skb
, 0);
4861 tcp_data_snd_check(sk
);
4863 } else { /* Header too small */
4864 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4869 int copied_early
= 0;
4871 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4872 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4873 #ifdef CONFIG_NET_DMA
4874 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4879 if (tp
->ucopy
.task
== current
&&
4880 sock_owned_by_user(sk
) && !copied_early
) {
4881 __set_current_state(TASK_RUNNING
);
4883 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4887 /* Predicted packet is in window by definition.
4888 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4889 * Hence, check seq<=rcv_wup reduces to:
4891 if (tcp_header_len
==
4892 (sizeof(struct tcphdr
) +
4893 TCPOLEN_TSTAMP_ALIGNED
) &&
4894 tp
->rcv_nxt
== tp
->rcv_wup
)
4895 tcp_store_ts_recent(tp
);
4897 tcp_rcv_rtt_measure_ts(sk
, skb
);
4899 __skb_pull(skb
, tcp_header_len
);
4900 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4901 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
4904 tcp_cleanup_rbuf(sk
, skb
->len
);
4907 if (tcp_checksum_complete_user(sk
, skb
))
4910 /* Predicted packet is in window by definition.
4911 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4912 * Hence, check seq<=rcv_wup reduces to:
4914 if (tcp_header_len
==
4915 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4916 tp
->rcv_nxt
== tp
->rcv_wup
)
4917 tcp_store_ts_recent(tp
);
4919 tcp_rcv_rtt_measure_ts(sk
, skb
);
4921 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4924 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
4926 /* Bulk data transfer: receiver */
4927 __skb_pull(skb
, tcp_header_len
);
4928 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4929 skb_set_owner_r(skb
, sk
);
4930 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4933 tcp_event_data_recv(sk
, skb
);
4935 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4936 /* Well, only one small jumplet in fast path... */
4937 tcp_ack(sk
, skb
, FLAG_DATA
);
4938 tcp_data_snd_check(sk
);
4939 if (!inet_csk_ack_scheduled(sk
))
4943 __tcp_ack_snd_check(sk
, 0);
4945 #ifdef CONFIG_NET_DMA
4947 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4953 sk
->sk_data_ready(sk
, 0);
4959 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
4963 * Standard slow path.
4966 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
4972 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4974 tcp_rcv_rtt_measure_ts(sk
, skb
);
4976 /* Process urgent data. */
4977 tcp_urg(sk
, skb
, th
);
4979 /* step 7: process the segment text */
4980 tcp_data_queue(sk
, skb
);
4982 tcp_data_snd_check(sk
);
4983 tcp_ack_snd_check(sk
);
4987 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4994 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
4995 struct tcphdr
*th
, unsigned len
)
4997 struct tcp_sock
*tp
= tcp_sk(sk
);
4998 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4999 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5001 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
5005 * "If the state is SYN-SENT then
5006 * first check the ACK bit
5007 * If the ACK bit is set
5008 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5009 * a reset (unless the RST bit is set, if so drop
5010 * the segment and return)"
5012 * We do not send data with SYN, so that RFC-correct
5015 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5016 goto reset_and_undo
;
5018 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5019 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5021 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5022 goto reset_and_undo
;
5025 /* Now ACK is acceptable.
5027 * "If the RST bit is set
5028 * If the ACK was acceptable then signal the user "error:
5029 * connection reset", drop the segment, enter CLOSED state,
5030 * delete TCB, and return."
5039 * "fifth, if neither of the SYN or RST bits is set then
5040 * drop the segment and return."
5046 goto discard_and_undo
;
5049 * "If the SYN bit is on ...
5050 * are acceptable then ...
5051 * (our SYN has been ACKed), change the connection
5052 * state to ESTABLISHED..."
5055 TCP_ECN_rcv_synack(tp
, th
);
5057 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5058 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5060 /* Ok.. it's good. Set up sequence numbers and
5061 * move to established.
5063 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5064 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5066 /* RFC1323: The window in SYN & SYN/ACK segments is
5069 tp
->snd_wnd
= ntohs(th
->window
);
5070 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
5072 if (!tp
->rx_opt
.wscale_ok
) {
5073 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5074 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5077 if (tp
->rx_opt
.saw_tstamp
) {
5078 tp
->rx_opt
.tstamp_ok
= 1;
5079 tp
->tcp_header_len
=
5080 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5081 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5082 tcp_store_ts_recent(tp
);
5084 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5087 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5088 tcp_enable_fack(tp
);
5091 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5092 tcp_initialize_rcv_mss(sk
);
5094 /* Remember, tcp_poll() does not lock socket!
5095 * Change state from SYN-SENT only after copied_seq
5096 * is initialized. */
5097 tp
->copied_seq
= tp
->rcv_nxt
;
5099 tcp_set_state(sk
, TCP_ESTABLISHED
);
5101 security_inet_conn_established(sk
, skb
);
5103 /* Make sure socket is routed, for correct metrics. */
5104 icsk
->icsk_af_ops
->rebuild_header(sk
);
5106 tcp_init_metrics(sk
);
5108 tcp_init_congestion_control(sk
);
5110 /* Prevent spurious tcp_cwnd_restart() on first data
5113 tp
->lsndtime
= tcp_time_stamp
;
5115 tcp_init_buffer_space(sk
);
5117 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5118 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5120 if (!tp
->rx_opt
.snd_wscale
)
5121 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5125 if (!sock_flag(sk
, SOCK_DEAD
)) {
5126 sk
->sk_state_change(sk
);
5127 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5130 if (sk
->sk_write_pending
||
5131 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5132 icsk
->icsk_ack
.pingpong
) {
5133 /* Save one ACK. Data will be ready after
5134 * several ticks, if write_pending is set.
5136 * It may be deleted, but with this feature tcpdumps
5137 * look so _wonderfully_ clever, that I was not able
5138 * to stand against the temptation 8) --ANK
5140 inet_csk_schedule_ack(sk
);
5141 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5142 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5143 tcp_incr_quickack(sk
);
5144 tcp_enter_quickack_mode(sk
);
5145 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5146 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5157 /* No ACK in the segment */
5161 * "If the RST bit is set
5163 * Otherwise (no ACK) drop the segment and return."
5166 goto discard_and_undo
;
5170 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5171 tcp_paws_check(&tp
->rx_opt
, 0))
5172 goto discard_and_undo
;
5175 /* We see SYN without ACK. It is attempt of
5176 * simultaneous connect with crossed SYNs.
5177 * Particularly, it can be connect to self.
5179 tcp_set_state(sk
, TCP_SYN_RECV
);
5181 if (tp
->rx_opt
.saw_tstamp
) {
5182 tp
->rx_opt
.tstamp_ok
= 1;
5183 tcp_store_ts_recent(tp
);
5184 tp
->tcp_header_len
=
5185 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5187 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5190 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5191 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5193 /* RFC1323: The window in SYN & SYN/ACK segments is
5196 tp
->snd_wnd
= ntohs(th
->window
);
5197 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5198 tp
->max_window
= tp
->snd_wnd
;
5200 TCP_ECN_rcv_syn(tp
, th
);
5203 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5204 tcp_initialize_rcv_mss(sk
);
5206 tcp_send_synack(sk
);
5208 /* Note, we could accept data and URG from this segment.
5209 * There are no obstacles to make this.
5211 * However, if we ignore data in ACKless segments sometimes,
5212 * we have no reasons to accept it sometimes.
5213 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5214 * is not flawless. So, discard packet for sanity.
5215 * Uncomment this return to process the data.
5222 /* "fifth, if neither of the SYN or RST bits is set then
5223 * drop the segment and return."
5227 tcp_clear_options(&tp
->rx_opt
);
5228 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5232 tcp_clear_options(&tp
->rx_opt
);
5233 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5238 * This function implements the receiving procedure of RFC 793 for
5239 * all states except ESTABLISHED and TIME_WAIT.
5240 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5241 * address independent.
5244 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5245 struct tcphdr
*th
, unsigned len
)
5247 struct tcp_sock
*tp
= tcp_sk(sk
);
5248 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5252 tp
->rx_opt
.saw_tstamp
= 0;
5254 switch (sk
->sk_state
) {
5266 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5269 /* Now we have several options: In theory there is
5270 * nothing else in the frame. KA9Q has an option to
5271 * send data with the syn, BSD accepts data with the
5272 * syn up to the [to be] advertised window and
5273 * Solaris 2.1 gives you a protocol error. For now
5274 * we just ignore it, that fits the spec precisely
5275 * and avoids incompatibilities. It would be nice in
5276 * future to drop through and process the data.
5278 * Now that TTCP is starting to be used we ought to
5280 * But, this leaves one open to an easy denial of
5281 * service attack, and SYN cookies can't defend
5282 * against this problem. So, we drop the data
5283 * in the interest of security over speed unless
5284 * it's still in use.
5292 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5296 /* Do step6 onward by hand. */
5297 tcp_urg(sk
, skb
, th
);
5299 tcp_data_snd_check(sk
);
5303 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5307 /* step 5: check the ACK field */
5309 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5311 switch (sk
->sk_state
) {
5314 tp
->copied_seq
= tp
->rcv_nxt
;
5316 tcp_set_state(sk
, TCP_ESTABLISHED
);
5317 sk
->sk_state_change(sk
);
5319 /* Note, that this wakeup is only for marginal
5320 * crossed SYN case. Passively open sockets
5321 * are not waked up, because sk->sk_sleep ==
5322 * NULL and sk->sk_socket == NULL.
5326 SOCK_WAKE_IO
, POLL_OUT
);
5328 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5329 tp
->snd_wnd
= ntohs(th
->window
) <<
5330 tp
->rx_opt
.snd_wscale
;
5331 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5332 TCP_SKB_CB(skb
)->seq
);
5334 /* tcp_ack considers this ACK as duplicate
5335 * and does not calculate rtt.
5336 * Fix it at least with timestamps.
5338 if (tp
->rx_opt
.saw_tstamp
&&
5339 tp
->rx_opt
.rcv_tsecr
&& !tp
->srtt
)
5340 tcp_ack_saw_tstamp(sk
, 0);
5342 if (tp
->rx_opt
.tstamp_ok
)
5343 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5345 /* Make sure socket is routed, for
5348 icsk
->icsk_af_ops
->rebuild_header(sk
);
5350 tcp_init_metrics(sk
);
5352 tcp_init_congestion_control(sk
);
5354 /* Prevent spurious tcp_cwnd_restart() on
5355 * first data packet.
5357 tp
->lsndtime
= tcp_time_stamp
;
5360 tcp_initialize_rcv_mss(sk
);
5361 tcp_init_buffer_space(sk
);
5362 tcp_fast_path_on(tp
);
5369 if (tp
->snd_una
== tp
->write_seq
) {
5370 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5371 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5372 dst_confirm(sk
->sk_dst_cache
);
5374 if (!sock_flag(sk
, SOCK_DEAD
))
5375 /* Wake up lingering close() */
5376 sk
->sk_state_change(sk
);
5380 if (tp
->linger2
< 0 ||
5381 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5382 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5384 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5388 tmo
= tcp_fin_time(sk
);
5389 if (tmo
> TCP_TIMEWAIT_LEN
) {
5390 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5391 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5392 /* Bad case. We could lose such FIN otherwise.
5393 * It is not a big problem, but it looks confusing
5394 * and not so rare event. We still can lose it now,
5395 * if it spins in bh_lock_sock(), but it is really
5398 inet_csk_reset_keepalive_timer(sk
, tmo
);
5400 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5408 if (tp
->snd_una
== tp
->write_seq
) {
5409 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5415 if (tp
->snd_una
== tp
->write_seq
) {
5416 tcp_update_metrics(sk
);
5425 /* step 6: check the URG bit */
5426 tcp_urg(sk
, skb
, th
);
5428 /* step 7: process the segment text */
5429 switch (sk
->sk_state
) {
5430 case TCP_CLOSE_WAIT
:
5433 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5437 /* RFC 793 says to queue data in these states,
5438 * RFC 1122 says we MUST send a reset.
5439 * BSD 4.4 also does reset.
5441 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5442 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5443 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5444 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5450 case TCP_ESTABLISHED
:
5451 tcp_data_queue(sk
, skb
);
5456 /* tcp_data could move socket to TIME-WAIT */
5457 if (sk
->sk_state
!= TCP_CLOSE
) {
5458 tcp_data_snd_check(sk
);
5459 tcp_ack_snd_check(sk
);
5469 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5470 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5471 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
5472 EXPORT_SYMBOL(tcp_parse_options
);
5473 #ifdef CONFIG_TCP_MD5SIG
5474 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
5476 EXPORT_SYMBOL(tcp_rcv_established
);
5477 EXPORT_SYMBOL(tcp_rcv_state_process
);
5478 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);