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 /* RFC: This is from the original, I doubt that this is necessary at all:
983 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
984 * retransmitted past LOST markings in the first place? I'm not fully sure
985 * about undo and end of connection cases, which can cause R without L?
987 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
989 if ((tp
->retransmit_skb_hint
!= NULL
) &&
990 before(TCP_SKB_CB(skb
)->seq
,
991 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
992 tp
->retransmit_skb_hint
= NULL
;
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 /* This procedure tags the retransmission queue when SACKs arrive.
1007 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1008 * Packets in queue with these bits set are counted in variables
1009 * sacked_out, retrans_out and lost_out, correspondingly.
1011 * Valid combinations are:
1012 * Tag InFlight Description
1013 * 0 1 - orig segment is in flight.
1014 * S 0 - nothing flies, orig reached receiver.
1015 * L 0 - nothing flies, orig lost by net.
1016 * R 2 - both orig and retransmit are in flight.
1017 * L|R 1 - orig is lost, retransmit is in flight.
1018 * S|R 1 - orig reached receiver, retrans is still in flight.
1019 * (L|S|R is logically valid, it could occur when L|R is sacked,
1020 * but it is equivalent to plain S and code short-curcuits it to S.
1021 * L|S is logically invalid, it would mean -1 packet in flight 8))
1023 * These 6 states form finite state machine, controlled by the following events:
1024 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1025 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1026 * 3. Loss detection event of one of three flavors:
1027 * A. Scoreboard estimator decided the packet is lost.
1028 * A'. Reno "three dupacks" marks head of queue lost.
1029 * A''. Its FACK modfication, head until snd.fack is lost.
1030 * B. SACK arrives sacking data transmitted after never retransmitted
1031 * hole was sent out.
1032 * C. SACK arrives sacking SND.NXT at the moment, when the
1033 * segment was retransmitted.
1034 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1036 * It is pleasant to note, that state diagram turns out to be commutative,
1037 * so that we are allowed not to be bothered by order of our actions,
1038 * when multiple events arrive simultaneously. (see the function below).
1040 * Reordering detection.
1041 * --------------------
1042 * Reordering metric is maximal distance, which a packet can be displaced
1043 * in packet stream. With SACKs we can estimate it:
1045 * 1. SACK fills old hole and the corresponding segment was not
1046 * ever retransmitted -> reordering. Alas, we cannot use it
1047 * when segment was retransmitted.
1048 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1049 * for retransmitted and already SACKed segment -> reordering..
1050 * Both of these heuristics are not used in Loss state, when we cannot
1051 * account for retransmits accurately.
1053 * SACK block validation.
1054 * ----------------------
1056 * SACK block range validation checks that the received SACK block fits to
1057 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1058 * Note that SND.UNA is not included to the range though being valid because
1059 * it means that the receiver is rather inconsistent with itself reporting
1060 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1061 * perfectly valid, however, in light of RFC2018 which explicitly states
1062 * that "SACK block MUST reflect the newest segment. Even if the newest
1063 * segment is going to be discarded ...", not that it looks very clever
1064 * in case of head skb. Due to potentional receiver driven attacks, we
1065 * choose to avoid immediate execution of a walk in write queue due to
1066 * reneging and defer head skb's loss recovery to standard loss recovery
1067 * procedure that will eventually trigger (nothing forbids us doing this).
1069 * Implements also blockage to start_seq wrap-around. Problem lies in the
1070 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1071 * there's no guarantee that it will be before snd_nxt (n). The problem
1072 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1075 * <- outs wnd -> <- wrapzone ->
1076 * u e n u_w e_w s n_w
1078 * |<------------+------+----- TCP seqno space --------------+---------->|
1079 * ...-- <2^31 ->| |<--------...
1080 * ...---- >2^31 ------>| |<--------...
1082 * Current code wouldn't be vulnerable but it's better still to discard such
1083 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1084 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1085 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1086 * equal to the ideal case (infinite seqno space without wrap caused issues).
1088 * With D-SACK the lower bound is extended to cover sequence space below
1089 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1090 * again, D-SACK block must not to go across snd_una (for the same reason as
1091 * for the normal SACK blocks, explained above). But there all simplicity
1092 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1093 * fully below undo_marker they do not affect behavior in anyway and can
1094 * therefore be safely ignored. In rare cases (which are more or less
1095 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1096 * fragmentation and packet reordering past skb's retransmission. To consider
1097 * them correctly, the acceptable range must be extended even more though
1098 * the exact amount is rather hard to quantify. However, tp->max_window can
1099 * be used as an exaggerated estimate.
1101 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1102 u32 start_seq
, u32 end_seq
)
1104 /* Too far in future, or reversed (interpretation is ambiguous) */
1105 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1108 /* Nasty start_seq wrap-around check (see comments above) */
1109 if (!before(start_seq
, tp
->snd_nxt
))
1112 /* In outstanding window? ...This is valid exit for D-SACKs too.
1113 * start_seq == snd_una is non-sensical (see comments above)
1115 if (after(start_seq
, tp
->snd_una
))
1118 if (!is_dsack
|| !tp
->undo_marker
)
1121 /* ...Then it's D-SACK, and must reside below snd_una completely */
1122 if (!after(end_seq
, tp
->snd_una
))
1125 if (!before(start_seq
, tp
->undo_marker
))
1129 if (!after(end_seq
, tp
->undo_marker
))
1132 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1133 * start_seq < undo_marker and end_seq >= undo_marker.
1135 return !before(start_seq
, end_seq
- tp
->max_window
);
1138 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1139 * Event "C". Later note: FACK people cheated me again 8), we have to account
1140 * for reordering! Ugly, but should help.
1142 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1143 * less than what is now known to be received by the other end (derived from
1144 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1145 * retransmitted skbs to avoid some costly processing per ACKs.
1147 static void tcp_mark_lost_retrans(struct sock
*sk
)
1149 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1150 struct tcp_sock
*tp
= tcp_sk(sk
);
1151 struct sk_buff
*skb
;
1153 u32 new_low_seq
= tp
->snd_nxt
;
1154 u32 received_upto
= tcp_highest_sack_seq(tp
);
1156 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1157 !after(received_upto
, tp
->lost_retrans_low
) ||
1158 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1161 tcp_for_write_queue(skb
, sk
) {
1162 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1164 if (skb
== tcp_send_head(sk
))
1166 if (cnt
== tp
->retrans_out
)
1168 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1171 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1174 if (after(received_upto
, ack_seq
) &&
1176 !before(received_upto
,
1177 ack_seq
+ tp
->reordering
* tp
->mss_cache
))) {
1178 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1179 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1181 /* clear lost hint */
1182 tp
->retransmit_skb_hint
= NULL
;
1184 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1185 tp
->lost_out
+= tcp_skb_pcount(skb
);
1186 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1188 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1190 if (before(ack_seq
, new_low_seq
))
1191 new_low_seq
= ack_seq
;
1192 cnt
+= tcp_skb_pcount(skb
);
1196 if (tp
->retrans_out
)
1197 tp
->lost_retrans_low
= new_low_seq
;
1200 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1201 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1204 struct tcp_sock
*tp
= tcp_sk(sk
);
1205 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1206 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1209 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1212 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1213 } else if (num_sacks
> 1) {
1214 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1215 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1217 if (!after(end_seq_0
, end_seq_1
) &&
1218 !before(start_seq_0
, start_seq_1
)) {
1221 NET_INC_STATS_BH(sock_net(sk
),
1222 LINUX_MIB_TCPDSACKOFORECV
);
1226 /* D-SACK for already forgotten data... Do dumb counting. */
1228 !after(end_seq_0
, prior_snd_una
) &&
1229 after(end_seq_0
, tp
->undo_marker
))
1235 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1236 * the incoming SACK may not exactly match but we can find smaller MSS
1237 * aligned portion of it that matches. Therefore we might need to fragment
1238 * which may fail and creates some hassle (caller must handle error case
1241 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1242 u32 start_seq
, u32 end_seq
)
1245 unsigned int pkt_len
;
1247 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1248 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1250 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1251 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1253 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1256 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1258 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1259 err
= tcp_fragment(sk
, skb
, pkt_len
, skb_shinfo(skb
)->gso_size
);
1267 static int tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1268 int *reord
, int dup_sack
, int fack_count
)
1270 struct tcp_sock
*tp
= tcp_sk(sk
);
1271 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1274 /* Account D-SACK for retransmitted packet. */
1275 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1276 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1278 if (sacked
& TCPCB_SACKED_ACKED
)
1279 *reord
= min(fack_count
, *reord
);
1282 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1283 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1286 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1287 if (sacked
& TCPCB_SACKED_RETRANS
) {
1288 /* If the segment is not tagged as lost,
1289 * we do not clear RETRANS, believing
1290 * that retransmission is still in flight.
1292 if (sacked
& TCPCB_LOST
) {
1293 TCP_SKB_CB(skb
)->sacked
&=
1294 ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1295 tp
->lost_out
-= tcp_skb_pcount(skb
);
1296 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1298 /* clear lost hint */
1299 tp
->retransmit_skb_hint
= NULL
;
1302 if (!(sacked
& TCPCB_RETRANS
)) {
1303 /* New sack for not retransmitted frame,
1304 * which was in hole. It is reordering.
1306 if (before(TCP_SKB_CB(skb
)->seq
,
1307 tcp_highest_sack_seq(tp
)))
1308 *reord
= min(fack_count
, *reord
);
1310 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1311 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1312 flag
|= FLAG_ONLY_ORIG_SACKED
;
1315 if (sacked
& TCPCB_LOST
) {
1316 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1317 tp
->lost_out
-= tcp_skb_pcount(skb
);
1319 /* clear lost hint */
1320 tp
->retransmit_skb_hint
= NULL
;
1324 TCP_SKB_CB(skb
)->sacked
|= TCPCB_SACKED_ACKED
;
1325 flag
|= FLAG_DATA_SACKED
;
1326 tp
->sacked_out
+= tcp_skb_pcount(skb
);
1328 fack_count
+= tcp_skb_pcount(skb
);
1330 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1331 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1332 before(TCP_SKB_CB(skb
)->seq
,
1333 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1334 tp
->lost_cnt_hint
+= tcp_skb_pcount(skb
);
1336 if (fack_count
> tp
->fackets_out
)
1337 tp
->fackets_out
= fack_count
;
1339 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
1340 tcp_advance_highest_sack(sk
, skb
);
1343 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1344 * frames and clear it. undo_retrans is decreased above, L|R frames
1345 * are accounted above as well.
1347 if (dup_sack
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)) {
1348 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1349 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1350 tp
->retransmit_skb_hint
= NULL
;
1356 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1357 struct tcp_sack_block
*next_dup
,
1358 u32 start_seq
, u32 end_seq
,
1359 int dup_sack_in
, int *fack_count
,
1360 int *reord
, int *flag
)
1362 tcp_for_write_queue_from(skb
, sk
) {
1364 int dup_sack
= dup_sack_in
;
1366 if (skb
== tcp_send_head(sk
))
1369 /* queue is in-order => we can short-circuit the walk early */
1370 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1373 if ((next_dup
!= NULL
) &&
1374 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1375 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1376 next_dup
->start_seq
,
1383 in_sack
= tcp_match_skb_to_sack(sk
, skb
, start_seq
,
1385 if (unlikely(in_sack
< 0))
1389 *flag
|= tcp_sacktag_one(skb
, sk
, reord
, dup_sack
,
1392 *fack_count
+= tcp_skb_pcount(skb
);
1397 /* Avoid all extra work that is being done by sacktag while walking in
1400 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1401 u32 skip_to_seq
, int *fack_count
)
1403 tcp_for_write_queue_from(skb
, sk
) {
1404 if (skb
== tcp_send_head(sk
))
1407 if (!before(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1410 *fack_count
+= tcp_skb_pcount(skb
);
1415 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1417 struct tcp_sack_block
*next_dup
,
1419 int *fack_count
, int *reord
,
1422 if (next_dup
== NULL
)
1425 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1426 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
, fack_count
);
1427 skb
= tcp_sacktag_walk(skb
, sk
, NULL
,
1428 next_dup
->start_seq
, next_dup
->end_seq
,
1429 1, fack_count
, reord
, flag
);
1435 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1437 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1441 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1444 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1445 struct tcp_sock
*tp
= tcp_sk(sk
);
1446 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1447 TCP_SKB_CB(ack_skb
)->sacked
);
1448 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1449 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1450 struct tcp_sack_block
*cache
;
1451 struct sk_buff
*skb
;
1452 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1454 int reord
= tp
->packets_out
;
1456 int found_dup_sack
= 0;
1459 int first_sack_index
;
1461 if (!tp
->sacked_out
) {
1462 if (WARN_ON(tp
->fackets_out
))
1463 tp
->fackets_out
= 0;
1464 tcp_highest_sack_reset(sk
);
1467 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1468 num_sacks
, prior_snd_una
);
1470 flag
|= FLAG_DSACKING_ACK
;
1472 /* Eliminate too old ACKs, but take into
1473 * account more or less fresh ones, they can
1474 * contain valid SACK info.
1476 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1479 if (!tp
->packets_out
)
1483 first_sack_index
= 0;
1484 for (i
= 0; i
< num_sacks
; i
++) {
1485 int dup_sack
= !i
&& found_dup_sack
;
1487 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1488 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1490 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1491 sp
[used_sacks
].start_seq
,
1492 sp
[used_sacks
].end_seq
)) {
1496 if (!tp
->undo_marker
)
1497 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1499 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1501 /* Don't count olds caused by ACK reordering */
1502 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1503 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1505 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1508 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1510 first_sack_index
= -1;
1514 /* Ignore very old stuff early */
1515 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1521 /* order SACK blocks to allow in order walk of the retrans queue */
1522 for (i
= used_sacks
- 1; i
> 0; i
--) {
1523 for (j
= 0; j
< i
; j
++) {
1524 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1525 struct tcp_sack_block tmp
;
1531 /* Track where the first SACK block goes to */
1532 if (j
== first_sack_index
)
1533 first_sack_index
= j
+ 1;
1538 skb
= tcp_write_queue_head(sk
);
1542 if (!tp
->sacked_out
) {
1543 /* It's already past, so skip checking against it */
1544 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1546 cache
= tp
->recv_sack_cache
;
1547 /* Skip empty blocks in at head of the cache */
1548 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1553 while (i
< used_sacks
) {
1554 u32 start_seq
= sp
[i
].start_seq
;
1555 u32 end_seq
= sp
[i
].end_seq
;
1556 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1557 struct tcp_sack_block
*next_dup
= NULL
;
1559 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1560 next_dup
= &sp
[i
+ 1];
1562 /* Event "B" in the comment above. */
1563 if (after(end_seq
, tp
->high_seq
))
1564 flag
|= FLAG_DATA_LOST
;
1566 /* Skip too early cached blocks */
1567 while (tcp_sack_cache_ok(tp
, cache
) &&
1568 !before(start_seq
, cache
->end_seq
))
1571 /* Can skip some work by looking recv_sack_cache? */
1572 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1573 after(end_seq
, cache
->start_seq
)) {
1576 if (before(start_seq
, cache
->start_seq
)) {
1577 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
,
1579 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1582 dup_sack
, &fack_count
,
1586 /* Rest of the block already fully processed? */
1587 if (!after(end_seq
, cache
->end_seq
))
1590 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1592 &fack_count
, &reord
,
1595 /* ...tail remains todo... */
1596 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1597 /* ...but better entrypoint exists! */
1598 skb
= tcp_highest_sack(sk
);
1601 fack_count
= tp
->fackets_out
;
1606 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
,
1608 /* Check overlap against next cached too (past this one already) */
1613 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1614 skb
= tcp_highest_sack(sk
);
1617 fack_count
= tp
->fackets_out
;
1619 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
, &fack_count
);
1622 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1623 dup_sack
, &fack_count
, &reord
, &flag
);
1626 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1627 * due to in-order walk
1629 if (after(end_seq
, tp
->frto_highmark
))
1630 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1635 /* Clear the head of the cache sack blocks so we can skip it next time */
1636 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1637 tp
->recv_sack_cache
[i
].start_seq
= 0;
1638 tp
->recv_sack_cache
[i
].end_seq
= 0;
1640 for (j
= 0; j
< used_sacks
; j
++)
1641 tp
->recv_sack_cache
[i
++] = sp
[j
];
1643 tcp_mark_lost_retrans(sk
);
1645 tcp_verify_left_out(tp
);
1647 if ((reord
< tp
->fackets_out
) &&
1648 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1649 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1650 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1654 #if FASTRETRANS_DEBUG > 0
1655 WARN_ON((int)tp
->sacked_out
< 0);
1656 WARN_ON((int)tp
->lost_out
< 0);
1657 WARN_ON((int)tp
->retrans_out
< 0);
1658 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1663 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1664 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1666 int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1670 holes
= max(tp
->lost_out
, 1U);
1671 holes
= min(holes
, tp
->packets_out
);
1673 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1674 tp
->sacked_out
= tp
->packets_out
- holes
;
1680 /* If we receive more dupacks than we expected counting segments
1681 * in assumption of absent reordering, interpret this as reordering.
1682 * The only another reason could be bug in receiver TCP.
1684 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1686 struct tcp_sock
*tp
= tcp_sk(sk
);
1687 if (tcp_limit_reno_sacked(tp
))
1688 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1691 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1693 static void tcp_add_reno_sack(struct sock
*sk
)
1695 struct tcp_sock
*tp
= tcp_sk(sk
);
1697 tcp_check_reno_reordering(sk
, 0);
1698 tcp_verify_left_out(tp
);
1701 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1703 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1705 struct tcp_sock
*tp
= tcp_sk(sk
);
1708 /* One ACK acked hole. The rest eat duplicate ACKs. */
1709 if (acked
- 1 >= tp
->sacked_out
)
1712 tp
->sacked_out
-= acked
- 1;
1714 tcp_check_reno_reordering(sk
, acked
);
1715 tcp_verify_left_out(tp
);
1718 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1723 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
1725 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
1728 /* F-RTO can only be used if TCP has never retransmitted anything other than
1729 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1731 int tcp_use_frto(struct sock
*sk
)
1733 const struct tcp_sock
*tp
= tcp_sk(sk
);
1734 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1735 struct sk_buff
*skb
;
1737 if (!sysctl_tcp_frto
)
1740 /* MTU probe and F-RTO won't really play nicely along currently */
1741 if (icsk
->icsk_mtup
.probe_size
)
1744 if (tcp_is_sackfrto(tp
))
1747 /* Avoid expensive walking of rexmit queue if possible */
1748 if (tp
->retrans_out
> 1)
1751 skb
= tcp_write_queue_head(sk
);
1752 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
1753 tcp_for_write_queue_from(skb
, sk
) {
1754 if (skb
== tcp_send_head(sk
))
1756 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1758 /* Short-circuit when first non-SACKed skb has been checked */
1759 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1765 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1766 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1767 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1768 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1769 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1770 * bits are handled if the Loss state is really to be entered (in
1771 * tcp_enter_frto_loss).
1773 * Do like tcp_enter_loss() would; when RTO expires the second time it
1775 * "Reduce ssthresh if it has not yet been made inside this window."
1777 void tcp_enter_frto(struct sock
*sk
)
1779 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1780 struct tcp_sock
*tp
= tcp_sk(sk
);
1781 struct sk_buff
*skb
;
1783 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
1784 tp
->snd_una
== tp
->high_seq
||
1785 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
1786 !icsk
->icsk_retransmits
)) {
1787 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1788 /* Our state is too optimistic in ssthresh() call because cwnd
1789 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
1790 * recovery has not yet completed. Pattern would be this: RTO,
1791 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1793 * RFC4138 should be more specific on what to do, even though
1794 * RTO is quite unlikely to occur after the first Cumulative ACK
1795 * due to back-off and complexity of triggering events ...
1797 if (tp
->frto_counter
) {
1799 stored_cwnd
= tp
->snd_cwnd
;
1801 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1802 tp
->snd_cwnd
= stored_cwnd
;
1804 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1806 /* ... in theory, cong.control module could do "any tricks" in
1807 * ssthresh(), which means that ca_state, lost bits and lost_out
1808 * counter would have to be faked before the call occurs. We
1809 * consider that too expensive, unlikely and hacky, so modules
1810 * using these in ssthresh() must deal these incompatibility
1811 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1813 tcp_ca_event(sk
, CA_EVENT_FRTO
);
1816 tp
->undo_marker
= tp
->snd_una
;
1817 tp
->undo_retrans
= 0;
1819 skb
= tcp_write_queue_head(sk
);
1820 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1821 tp
->undo_marker
= 0;
1822 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
1823 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1824 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1826 tcp_verify_left_out(tp
);
1828 /* Too bad if TCP was application limited */
1829 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
1831 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1832 * The last condition is necessary at least in tp->frto_counter case.
1834 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
1835 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
1836 after(tp
->high_seq
, tp
->snd_una
)) {
1837 tp
->frto_highmark
= tp
->high_seq
;
1839 tp
->frto_highmark
= tp
->snd_nxt
;
1841 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
1842 tp
->high_seq
= tp
->snd_nxt
;
1843 tp
->frto_counter
= 1;
1846 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1847 * which indicates that we should follow the traditional RTO recovery,
1848 * i.e. mark everything lost and do go-back-N retransmission.
1850 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
1852 struct tcp_sock
*tp
= tcp_sk(sk
);
1853 struct sk_buff
*skb
;
1856 tp
->retrans_out
= 0;
1857 if (tcp_is_reno(tp
))
1858 tcp_reset_reno_sack(tp
);
1860 tcp_for_write_queue(skb
, sk
) {
1861 if (skb
== tcp_send_head(sk
))
1864 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
1866 * Count the retransmission made on RTO correctly (only when
1867 * waiting for the first ACK and did not get it)...
1869 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
1870 /* For some reason this R-bit might get cleared? */
1871 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1872 tp
->retrans_out
+= tcp_skb_pcount(skb
);
1873 /* ...enter this if branch just for the first segment */
1874 flag
|= FLAG_DATA_ACKED
;
1876 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1877 tp
->undo_marker
= 0;
1878 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1881 /* Marking forward transmissions that were made after RTO lost
1882 * can cause unnecessary retransmissions in some scenarios,
1883 * SACK blocks will mitigate that in some but not in all cases.
1884 * We used to not mark them but it was causing break-ups with
1885 * receivers that do only in-order receival.
1887 * TODO: we could detect presence of such receiver and select
1888 * different behavior per flow.
1890 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
1891 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1892 tp
->lost_out
+= tcp_skb_pcount(skb
);
1895 tcp_verify_left_out(tp
);
1897 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
1898 tp
->snd_cwnd_cnt
= 0;
1899 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1900 tp
->frto_counter
= 0;
1901 tp
->bytes_acked
= 0;
1903 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1904 sysctl_tcp_reordering
);
1905 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1906 tp
->high_seq
= tp
->snd_nxt
;
1907 TCP_ECN_queue_cwr(tp
);
1909 tcp_clear_all_retrans_hints(tp
);
1912 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1914 tp
->retrans_out
= 0;
1917 tp
->undo_marker
= 0;
1918 tp
->undo_retrans
= 0;
1921 void tcp_clear_retrans(struct tcp_sock
*tp
)
1923 tcp_clear_retrans_partial(tp
);
1925 tp
->fackets_out
= 0;
1929 /* Enter Loss state. If "how" is not zero, forget all SACK information
1930 * and reset tags completely, otherwise preserve SACKs. If receiver
1931 * dropped its ofo queue, we will know this due to reneging detection.
1933 void tcp_enter_loss(struct sock
*sk
, int how
)
1935 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1936 struct tcp_sock
*tp
= tcp_sk(sk
);
1937 struct sk_buff
*skb
;
1939 /* Reduce ssthresh if it has not yet been made inside this window. */
1940 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
1941 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1942 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1943 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1944 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1947 tp
->snd_cwnd_cnt
= 0;
1948 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1950 tp
->bytes_acked
= 0;
1951 tcp_clear_retrans_partial(tp
);
1953 if (tcp_is_reno(tp
))
1954 tcp_reset_reno_sack(tp
);
1957 /* Push undo marker, if it was plain RTO and nothing
1958 * was retransmitted. */
1959 tp
->undo_marker
= tp
->snd_una
;
1962 tp
->fackets_out
= 0;
1964 tcp_clear_all_retrans_hints(tp
);
1966 tcp_for_write_queue(skb
, sk
) {
1967 if (skb
== tcp_send_head(sk
))
1970 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1971 tp
->undo_marker
= 0;
1972 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1973 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1974 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1975 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1976 tp
->lost_out
+= tcp_skb_pcount(skb
);
1979 tcp_verify_left_out(tp
);
1981 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1982 sysctl_tcp_reordering
);
1983 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1984 tp
->high_seq
= tp
->snd_nxt
;
1985 TCP_ECN_queue_cwr(tp
);
1986 /* Abort F-RTO algorithm if one is in progress */
1987 tp
->frto_counter
= 0;
1990 /* If ACK arrived pointing to a remembered SACK, it means that our
1991 * remembered SACKs do not reflect real state of receiver i.e.
1992 * receiver _host_ is heavily congested (or buggy).
1994 * Do processing similar to RTO timeout.
1996 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1998 if (flag
& FLAG_SACK_RENEGING
) {
1999 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2000 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2002 tcp_enter_loss(sk
, 1);
2003 icsk
->icsk_retransmits
++;
2004 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2005 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2006 icsk
->icsk_rto
, TCP_RTO_MAX
);
2012 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2014 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2017 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2018 * counter when SACK is enabled (without SACK, sacked_out is used for
2021 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2022 * segments up to the highest received SACK block so far and holes in
2025 * With reordering, holes may still be in flight, so RFC3517 recovery
2026 * uses pure sacked_out (total number of SACKed segments) even though
2027 * it violates the RFC that uses duplicate ACKs, often these are equal
2028 * but when e.g. out-of-window ACKs or packet duplication occurs,
2029 * they differ. Since neither occurs due to loss, TCP should really
2032 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
2034 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2037 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2039 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2042 static inline int tcp_head_timedout(struct sock
*sk
)
2044 struct tcp_sock
*tp
= tcp_sk(sk
);
2046 return tp
->packets_out
&&
2047 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2050 /* Linux NewReno/SACK/FACK/ECN state machine.
2051 * --------------------------------------
2053 * "Open" Normal state, no dubious events, fast path.
2054 * "Disorder" In all the respects it is "Open",
2055 * but requires a bit more attention. It is entered when
2056 * we see some SACKs or dupacks. It is split of "Open"
2057 * mainly to move some processing from fast path to slow one.
2058 * "CWR" CWND was reduced due to some Congestion Notification event.
2059 * It can be ECN, ICMP source quench, local device congestion.
2060 * "Recovery" CWND was reduced, we are fast-retransmitting.
2061 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2063 * tcp_fastretrans_alert() is entered:
2064 * - each incoming ACK, if state is not "Open"
2065 * - when arrived ACK is unusual, namely:
2070 * Counting packets in flight is pretty simple.
2072 * in_flight = packets_out - left_out + retrans_out
2074 * packets_out is SND.NXT-SND.UNA counted in packets.
2076 * retrans_out is number of retransmitted segments.
2078 * left_out is number of segments left network, but not ACKed yet.
2080 * left_out = sacked_out + lost_out
2082 * sacked_out: Packets, which arrived to receiver out of order
2083 * and hence not ACKed. With SACKs this number is simply
2084 * amount of SACKed data. Even without SACKs
2085 * it is easy to give pretty reliable estimate of this number,
2086 * counting duplicate ACKs.
2088 * lost_out: Packets lost by network. TCP has no explicit
2089 * "loss notification" feedback from network (for now).
2090 * It means that this number can be only _guessed_.
2091 * Actually, it is the heuristics to predict lossage that
2092 * distinguishes different algorithms.
2094 * F.e. after RTO, when all the queue is considered as lost,
2095 * lost_out = packets_out and in_flight = retrans_out.
2097 * Essentially, we have now two algorithms counting
2100 * FACK: It is the simplest heuristics. As soon as we decided
2101 * that something is lost, we decide that _all_ not SACKed
2102 * packets until the most forward SACK are lost. I.e.
2103 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2104 * It is absolutely correct estimate, if network does not reorder
2105 * packets. And it loses any connection to reality when reordering
2106 * takes place. We use FACK by default until reordering
2107 * is suspected on the path to this destination.
2109 * NewReno: when Recovery is entered, we assume that one segment
2110 * is lost (classic Reno). While we are in Recovery and
2111 * a partial ACK arrives, we assume that one more packet
2112 * is lost (NewReno). This heuristics are the same in NewReno
2115 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2116 * deflation etc. CWND is real congestion window, never inflated, changes
2117 * only according to classic VJ rules.
2119 * Really tricky (and requiring careful tuning) part of algorithm
2120 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2121 * The first determines the moment _when_ we should reduce CWND and,
2122 * hence, slow down forward transmission. In fact, it determines the moment
2123 * when we decide that hole is caused by loss, rather than by a reorder.
2125 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2126 * holes, caused by lost packets.
2128 * And the most logically complicated part of algorithm is undo
2129 * heuristics. We detect false retransmits due to both too early
2130 * fast retransmit (reordering) and underestimated RTO, analyzing
2131 * timestamps and D-SACKs. When we detect that some segments were
2132 * retransmitted by mistake and CWND reduction was wrong, we undo
2133 * window reduction and abort recovery phase. This logic is hidden
2134 * inside several functions named tcp_try_undo_<something>.
2137 /* This function decides, when we should leave Disordered state
2138 * and enter Recovery phase, reducing congestion window.
2140 * Main question: may we further continue forward transmission
2141 * with the same cwnd?
2143 static int tcp_time_to_recover(struct sock
*sk
)
2145 struct tcp_sock
*tp
= tcp_sk(sk
);
2148 /* Do not perform any recovery during F-RTO algorithm */
2149 if (tp
->frto_counter
)
2152 /* Trick#1: The loss is proven. */
2156 /* Not-A-Trick#2 : Classic rule... */
2157 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2160 /* Trick#3 : when we use RFC2988 timer restart, fast
2161 * retransmit can be triggered by timeout of queue head.
2163 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2166 /* Trick#4: It is still not OK... But will it be useful to delay
2169 packets_out
= tp
->packets_out
;
2170 if (packets_out
<= tp
->reordering
&&
2171 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2172 !tcp_may_send_now(sk
)) {
2173 /* We have nothing to send. This connection is limited
2174 * either by receiver window or by application.
2182 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2183 * is against sacked "cnt", otherwise it's against facked "cnt"
2185 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2187 struct tcp_sock
*tp
= tcp_sk(sk
);
2188 struct sk_buff
*skb
;
2193 WARN_ON(packets
> tp
->packets_out
);
2194 if (tp
->lost_skb_hint
) {
2195 skb
= tp
->lost_skb_hint
;
2196 cnt
= tp
->lost_cnt_hint
;
2198 skb
= tcp_write_queue_head(sk
);
2202 tcp_for_write_queue_from(skb
, sk
) {
2203 if (skb
== tcp_send_head(sk
))
2205 /* TODO: do this better */
2206 /* this is not the most efficient way to do this... */
2207 tp
->lost_skb_hint
= skb
;
2208 tp
->lost_cnt_hint
= cnt
;
2210 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2214 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2215 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2216 cnt
+= tcp_skb_pcount(skb
);
2218 if (cnt
> packets
) {
2219 if (tcp_is_sack(tp
) || (oldcnt
>= packets
))
2222 mss
= skb_shinfo(skb
)->gso_size
;
2223 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2229 tcp_skb_mark_lost(tp
, skb
);
2231 tcp_verify_left_out(tp
);
2234 /* Account newly detected lost packet(s) */
2236 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2238 struct tcp_sock
*tp
= tcp_sk(sk
);
2240 if (tcp_is_reno(tp
)) {
2241 tcp_mark_head_lost(sk
, 1);
2242 } else if (tcp_is_fack(tp
)) {
2243 int lost
= tp
->fackets_out
- tp
->reordering
;
2246 tcp_mark_head_lost(sk
, lost
);
2248 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2249 if (sacked_upto
< fast_rexmit
)
2250 sacked_upto
= fast_rexmit
;
2251 tcp_mark_head_lost(sk
, sacked_upto
);
2254 /* New heuristics: it is possible only after we switched
2255 * to restart timer each time when something is ACKed.
2256 * Hence, we can detect timed out packets during fast
2257 * retransmit without falling to slow start.
2259 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2260 struct sk_buff
*skb
;
2262 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2263 : tcp_write_queue_head(sk
);
2265 tcp_for_write_queue_from(skb
, sk
) {
2266 if (skb
== tcp_send_head(sk
))
2268 if (!tcp_skb_timedout(sk
, skb
))
2271 tcp_skb_mark_lost(tp
, skb
);
2274 tp
->scoreboard_skb_hint
= skb
;
2276 tcp_verify_left_out(tp
);
2280 /* CWND moderation, preventing bursts due to too big ACKs
2281 * in dubious situations.
2283 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2285 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2286 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2287 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2290 /* Lower bound on congestion window is slow start threshold
2291 * unless congestion avoidance choice decides to overide it.
2293 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2295 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2297 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2300 /* Decrease cwnd each second ack. */
2301 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2303 struct tcp_sock
*tp
= tcp_sk(sk
);
2304 int decr
= tp
->snd_cwnd_cnt
+ 1;
2306 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2307 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2308 tp
->snd_cwnd_cnt
= decr
& 1;
2311 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2312 tp
->snd_cwnd
-= decr
;
2314 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2315 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2319 /* Nothing was retransmitted or returned timestamp is less
2320 * than timestamp of the first retransmission.
2322 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2324 return !tp
->retrans_stamp
||
2325 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2326 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2329 /* Undo procedures. */
2331 #if FASTRETRANS_DEBUG > 1
2332 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2334 struct tcp_sock
*tp
= tcp_sk(sk
);
2335 struct inet_sock
*inet
= inet_sk(sk
);
2337 if (sk
->sk_family
== AF_INET
) {
2338 printk(KERN_DEBUG
"Undo %s " NIPQUAD_FMT
"/%u c%u l%u ss%u/%u p%u\n",
2340 NIPQUAD(inet
->daddr
), ntohs(inet
->dport
),
2341 tp
->snd_cwnd
, tcp_left_out(tp
),
2342 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2345 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2346 else if (sk
->sk_family
== AF_INET6
) {
2347 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2348 printk(KERN_DEBUG
"Undo %s " NIP6_FMT
"/%u c%u l%u ss%u/%u p%u\n",
2350 NIP6(np
->daddr
), ntohs(inet
->dport
),
2351 tp
->snd_cwnd
, tcp_left_out(tp
),
2352 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2358 #define DBGUNDO(x...) do { } while (0)
2361 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2363 struct tcp_sock
*tp
= tcp_sk(sk
);
2365 if (tp
->prior_ssthresh
) {
2366 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2368 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2369 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2371 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2373 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2374 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2375 TCP_ECN_withdraw_cwr(tp
);
2378 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2380 tcp_moderate_cwnd(tp
);
2381 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2383 /* There is something screwy going on with the retrans hints after
2385 tcp_clear_all_retrans_hints(tp
);
2388 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2390 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2393 /* People celebrate: "We love our President!" */
2394 static int tcp_try_undo_recovery(struct sock
*sk
)
2396 struct tcp_sock
*tp
= tcp_sk(sk
);
2398 if (tcp_may_undo(tp
)) {
2401 /* Happy end! We did not retransmit anything
2402 * or our original transmission succeeded.
2404 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2405 tcp_undo_cwr(sk
, 1);
2406 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2407 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2409 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2411 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2412 tp
->undo_marker
= 0;
2414 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2415 /* Hold old state until something *above* high_seq
2416 * is ACKed. For Reno it is MUST to prevent false
2417 * fast retransmits (RFC2582). SACK TCP is safe. */
2418 tcp_moderate_cwnd(tp
);
2421 tcp_set_ca_state(sk
, TCP_CA_Open
);
2425 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2426 static void tcp_try_undo_dsack(struct sock
*sk
)
2428 struct tcp_sock
*tp
= tcp_sk(sk
);
2430 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2431 DBGUNDO(sk
, "D-SACK");
2432 tcp_undo_cwr(sk
, 1);
2433 tp
->undo_marker
= 0;
2434 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2438 /* Undo during fast recovery after partial ACK. */
2440 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2442 struct tcp_sock
*tp
= tcp_sk(sk
);
2443 /* Partial ACK arrived. Force Hoe's retransmit. */
2444 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2446 if (tcp_may_undo(tp
)) {
2447 /* Plain luck! Hole if filled with delayed
2448 * packet, rather than with a retransmit.
2450 if (tp
->retrans_out
== 0)
2451 tp
->retrans_stamp
= 0;
2453 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2456 tcp_undo_cwr(sk
, 0);
2457 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2459 /* So... Do not make Hoe's retransmit yet.
2460 * If the first packet was delayed, the rest
2461 * ones are most probably delayed as well.
2468 /* Undo during loss recovery after partial ACK. */
2469 static int tcp_try_undo_loss(struct sock
*sk
)
2471 struct tcp_sock
*tp
= tcp_sk(sk
);
2473 if (tcp_may_undo(tp
)) {
2474 struct sk_buff
*skb
;
2475 tcp_for_write_queue(skb
, sk
) {
2476 if (skb
== tcp_send_head(sk
))
2478 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2481 tcp_clear_all_retrans_hints(tp
);
2483 DBGUNDO(sk
, "partial loss");
2485 tcp_undo_cwr(sk
, 1);
2486 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2487 inet_csk(sk
)->icsk_retransmits
= 0;
2488 tp
->undo_marker
= 0;
2489 if (tcp_is_sack(tp
))
2490 tcp_set_ca_state(sk
, TCP_CA_Open
);
2496 static inline void tcp_complete_cwr(struct sock
*sk
)
2498 struct tcp_sock
*tp
= tcp_sk(sk
);
2499 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2500 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2501 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2504 static void tcp_try_keep_open(struct sock
*sk
)
2506 struct tcp_sock
*tp
= tcp_sk(sk
);
2507 int state
= TCP_CA_Open
;
2509 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2510 state
= TCP_CA_Disorder
;
2512 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2513 tcp_set_ca_state(sk
, state
);
2514 tp
->high_seq
= tp
->snd_nxt
;
2518 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2520 struct tcp_sock
*tp
= tcp_sk(sk
);
2522 tcp_verify_left_out(tp
);
2524 if (!tp
->frto_counter
&& tp
->retrans_out
== 0)
2525 tp
->retrans_stamp
= 0;
2527 if (flag
& FLAG_ECE
)
2528 tcp_enter_cwr(sk
, 1);
2530 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2531 tcp_try_keep_open(sk
);
2532 tcp_moderate_cwnd(tp
);
2534 tcp_cwnd_down(sk
, flag
);
2538 static void tcp_mtup_probe_failed(struct sock
*sk
)
2540 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2542 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2543 icsk
->icsk_mtup
.probe_size
= 0;
2546 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2548 struct tcp_sock
*tp
= tcp_sk(sk
);
2549 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2551 /* FIXME: breaks with very large cwnd */
2552 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2553 tp
->snd_cwnd
= tp
->snd_cwnd
*
2554 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2555 icsk
->icsk_mtup
.probe_size
;
2556 tp
->snd_cwnd_cnt
= 0;
2557 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2558 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2560 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2561 icsk
->icsk_mtup
.probe_size
= 0;
2562 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2565 /* Process an event, which can update packets-in-flight not trivially.
2566 * Main goal of this function is to calculate new estimate for left_out,
2567 * taking into account both packets sitting in receiver's buffer and
2568 * packets lost by network.
2570 * Besides that it does CWND reduction, when packet loss is detected
2571 * and changes state of machine.
2573 * It does _not_ decide what to send, it is made in function
2574 * tcp_xmit_retransmit_queue().
2576 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2578 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2579 struct tcp_sock
*tp
= tcp_sk(sk
);
2580 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2581 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2582 (tcp_fackets_out(tp
) > tp
->reordering
));
2583 int fast_rexmit
= 0, mib_idx
;
2585 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2587 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2588 tp
->fackets_out
= 0;
2590 /* Now state machine starts.
2591 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2592 if (flag
& FLAG_ECE
)
2593 tp
->prior_ssthresh
= 0;
2595 /* B. In all the states check for reneging SACKs. */
2596 if (tcp_check_sack_reneging(sk
, flag
))
2599 /* C. Process data loss notification, provided it is valid. */
2600 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2601 before(tp
->snd_una
, tp
->high_seq
) &&
2602 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2603 tp
->fackets_out
> tp
->reordering
) {
2604 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2605 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2608 /* D. Check consistency of the current state. */
2609 tcp_verify_left_out(tp
);
2611 /* E. Check state exit conditions. State can be terminated
2612 * when high_seq is ACKed. */
2613 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2614 WARN_ON(tp
->retrans_out
!= 0);
2615 tp
->retrans_stamp
= 0;
2616 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2617 switch (icsk
->icsk_ca_state
) {
2619 icsk
->icsk_retransmits
= 0;
2620 if (tcp_try_undo_recovery(sk
))
2625 /* CWR is to be held something *above* high_seq
2626 * is ACKed for CWR bit to reach receiver. */
2627 if (tp
->snd_una
!= tp
->high_seq
) {
2628 tcp_complete_cwr(sk
);
2629 tcp_set_ca_state(sk
, TCP_CA_Open
);
2633 case TCP_CA_Disorder
:
2634 tcp_try_undo_dsack(sk
);
2635 if (!tp
->undo_marker
||
2636 /* For SACK case do not Open to allow to undo
2637 * catching for all duplicate ACKs. */
2638 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2639 tp
->undo_marker
= 0;
2640 tcp_set_ca_state(sk
, TCP_CA_Open
);
2644 case TCP_CA_Recovery
:
2645 if (tcp_is_reno(tp
))
2646 tcp_reset_reno_sack(tp
);
2647 if (tcp_try_undo_recovery(sk
))
2649 tcp_complete_cwr(sk
);
2654 /* F. Process state. */
2655 switch (icsk
->icsk_ca_state
) {
2656 case TCP_CA_Recovery
:
2657 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2658 if (tcp_is_reno(tp
) && is_dupack
)
2659 tcp_add_reno_sack(sk
);
2661 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2664 if (flag
& FLAG_DATA_ACKED
)
2665 icsk
->icsk_retransmits
= 0;
2666 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
2667 tcp_reset_reno_sack(tp
);
2668 if (!tcp_try_undo_loss(sk
)) {
2669 tcp_moderate_cwnd(tp
);
2670 tcp_xmit_retransmit_queue(sk
);
2673 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2675 /* Loss is undone; fall through to processing in Open state. */
2677 if (tcp_is_reno(tp
)) {
2678 if (flag
& FLAG_SND_UNA_ADVANCED
)
2679 tcp_reset_reno_sack(tp
);
2681 tcp_add_reno_sack(sk
);
2684 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
2685 tcp_try_undo_dsack(sk
);
2687 if (!tcp_time_to_recover(sk
)) {
2688 tcp_try_to_open(sk
, flag
);
2692 /* MTU probe failure: don't reduce cwnd */
2693 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2694 icsk
->icsk_mtup
.probe_size
&&
2695 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2696 tcp_mtup_probe_failed(sk
);
2697 /* Restores the reduction we did in tcp_mtup_probe() */
2699 tcp_simple_retransmit(sk
);
2703 /* Otherwise enter Recovery state */
2705 if (tcp_is_reno(tp
))
2706 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2708 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2710 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2712 tp
->high_seq
= tp
->snd_nxt
;
2713 tp
->prior_ssthresh
= 0;
2714 tp
->undo_marker
= tp
->snd_una
;
2715 tp
->undo_retrans
= tp
->retrans_out
;
2717 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
2718 if (!(flag
& FLAG_ECE
))
2719 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2720 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2721 TCP_ECN_queue_cwr(tp
);
2724 tp
->bytes_acked
= 0;
2725 tp
->snd_cwnd_cnt
= 0;
2726 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2730 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
2731 tcp_update_scoreboard(sk
, fast_rexmit
);
2732 tcp_cwnd_down(sk
, flag
);
2733 tcp_xmit_retransmit_queue(sk
);
2736 /* Read draft-ietf-tcplw-high-performance before mucking
2737 * with this code. (Supersedes RFC1323)
2739 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2741 /* RTTM Rule: A TSecr value received in a segment is used to
2742 * update the averaged RTT measurement only if the segment
2743 * acknowledges some new data, i.e., only if it advances the
2744 * left edge of the send window.
2746 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2747 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2749 * Changed: reset backoff as soon as we see the first valid sample.
2750 * If we do not, we get strongly overestimated rto. With timestamps
2751 * samples are accepted even from very old segments: f.e., when rtt=1
2752 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2753 * answer arrives rto becomes 120 seconds! If at least one of segments
2754 * in window is lost... Voila. --ANK (010210)
2756 struct tcp_sock
*tp
= tcp_sk(sk
);
2757 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
2758 tcp_rtt_estimator(sk
, seq_rtt
);
2760 inet_csk(sk
)->icsk_backoff
= 0;
2764 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2766 /* We don't have a timestamp. Can only use
2767 * packets that are not retransmitted to determine
2768 * rtt estimates. Also, we must not reset the
2769 * backoff for rto until we get a non-retransmitted
2770 * packet. This allows us to deal with a situation
2771 * where the network delay has increased suddenly.
2772 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2775 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2778 tcp_rtt_estimator(sk
, seq_rtt
);
2780 inet_csk(sk
)->icsk_backoff
= 0;
2784 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2787 const struct tcp_sock
*tp
= tcp_sk(sk
);
2788 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2789 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2790 tcp_ack_saw_tstamp(sk
, flag
);
2791 else if (seq_rtt
>= 0)
2792 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2795 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2797 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2798 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2799 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2802 /* Restart timer after forward progress on connection.
2803 * RFC2988 recommends to restart timer to now+rto.
2805 static void tcp_rearm_rto(struct sock
*sk
)
2807 struct tcp_sock
*tp
= tcp_sk(sk
);
2809 if (!tp
->packets_out
) {
2810 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2812 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2813 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2817 /* If we get here, the whole TSO packet has not been acked. */
2818 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2820 struct tcp_sock
*tp
= tcp_sk(sk
);
2823 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2825 packets_acked
= tcp_skb_pcount(skb
);
2826 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2828 packets_acked
-= tcp_skb_pcount(skb
);
2830 if (packets_acked
) {
2831 BUG_ON(tcp_skb_pcount(skb
) == 0);
2832 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2835 return packets_acked
;
2838 /* Remove acknowledged frames from the retransmission queue. If our packet
2839 * is before the ack sequence we can discard it as it's confirmed to have
2840 * arrived at the other end.
2842 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
)
2844 struct tcp_sock
*tp
= tcp_sk(sk
);
2845 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2846 struct sk_buff
*skb
;
2847 u32 now
= tcp_time_stamp
;
2848 int fully_acked
= 1;
2851 u32 reord
= tp
->packets_out
;
2853 s32 ca_seq_rtt
= -1;
2854 ktime_t last_ackt
= net_invalid_timestamp();
2856 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2857 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2860 u8 sacked
= scb
->sacked
;
2862 /* Determine how many packets and what bytes were acked, tso and else */
2863 if (after(scb
->end_seq
, tp
->snd_una
)) {
2864 if (tcp_skb_pcount(skb
) == 1 ||
2865 !after(tp
->snd_una
, scb
->seq
))
2868 acked_pcount
= tcp_tso_acked(sk
, skb
);
2873 end_seq
= tp
->snd_una
;
2875 acked_pcount
= tcp_skb_pcount(skb
);
2876 end_seq
= scb
->end_seq
;
2879 /* MTU probing checks */
2880 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
2881 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
2882 tcp_mtup_probe_success(sk
, skb
);
2885 if (sacked
& TCPCB_RETRANS
) {
2886 if (sacked
& TCPCB_SACKED_RETRANS
)
2887 tp
->retrans_out
-= acked_pcount
;
2888 flag
|= FLAG_RETRANS_DATA_ACKED
;
2891 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
2892 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
2894 ca_seq_rtt
= now
- scb
->when
;
2895 last_ackt
= skb
->tstamp
;
2897 seq_rtt
= ca_seq_rtt
;
2899 if (!(sacked
& TCPCB_SACKED_ACKED
))
2900 reord
= min(pkts_acked
, reord
);
2903 if (sacked
& TCPCB_SACKED_ACKED
)
2904 tp
->sacked_out
-= acked_pcount
;
2905 if (sacked
& TCPCB_LOST
)
2906 tp
->lost_out
-= acked_pcount
;
2908 if (unlikely(tp
->urg_mode
&& !before(end_seq
, tp
->snd_up
)))
2911 tp
->packets_out
-= acked_pcount
;
2912 pkts_acked
+= acked_pcount
;
2914 /* Initial outgoing SYN's get put onto the write_queue
2915 * just like anything else we transmit. It is not
2916 * true data, and if we misinform our callers that
2917 * this ACK acks real data, we will erroneously exit
2918 * connection startup slow start one packet too
2919 * quickly. This is severely frowned upon behavior.
2921 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
2922 flag
|= FLAG_DATA_ACKED
;
2924 flag
|= FLAG_SYN_ACKED
;
2925 tp
->retrans_stamp
= 0;
2931 tcp_unlink_write_queue(skb
, sk
);
2932 sk_wmem_free_skb(sk
, skb
);
2933 tcp_clear_all_retrans_hints(tp
);
2936 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2937 flag
|= FLAG_SACK_RENEGING
;
2939 if (flag
& FLAG_ACKED
) {
2940 const struct tcp_congestion_ops
*ca_ops
2941 = inet_csk(sk
)->icsk_ca_ops
;
2943 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
2946 if (tcp_is_reno(tp
)) {
2947 tcp_remove_reno_sacks(sk
, pkts_acked
);
2949 /* Non-retransmitted hole got filled? That's reordering */
2950 if (reord
< prior_fackets
)
2951 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
2954 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
2956 if (ca_ops
->pkts_acked
) {
2959 /* Is the ACK triggering packet unambiguous? */
2960 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
2961 /* High resolution needed and available? */
2962 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
2963 !ktime_equal(last_ackt
,
2964 net_invalid_timestamp()))
2965 rtt_us
= ktime_us_delta(ktime_get_real(),
2967 else if (ca_seq_rtt
> 0)
2968 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
2971 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
2975 #if FASTRETRANS_DEBUG > 0
2976 WARN_ON((int)tp
->sacked_out
< 0);
2977 WARN_ON((int)tp
->lost_out
< 0);
2978 WARN_ON((int)tp
->retrans_out
< 0);
2979 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
2980 icsk
= inet_csk(sk
);
2982 printk(KERN_DEBUG
"Leak l=%u %d\n",
2983 tp
->lost_out
, icsk
->icsk_ca_state
);
2986 if (tp
->sacked_out
) {
2987 printk(KERN_DEBUG
"Leak s=%u %d\n",
2988 tp
->sacked_out
, icsk
->icsk_ca_state
);
2991 if (tp
->retrans_out
) {
2992 printk(KERN_DEBUG
"Leak r=%u %d\n",
2993 tp
->retrans_out
, icsk
->icsk_ca_state
);
2994 tp
->retrans_out
= 0;
3001 static void tcp_ack_probe(struct sock
*sk
)
3003 const struct tcp_sock
*tp
= tcp_sk(sk
);
3004 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3006 /* Was it a usable window open? */
3008 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3009 icsk
->icsk_backoff
= 0;
3010 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3011 /* Socket must be waked up by subsequent tcp_data_snd_check().
3012 * This function is not for random using!
3015 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3016 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3021 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3023 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3024 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3027 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3029 const struct tcp_sock
*tp
= tcp_sk(sk
);
3030 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3031 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3034 /* Check that window update is acceptable.
3035 * The function assumes that snd_una<=ack<=snd_next.
3037 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3038 const u32 ack
, const u32 ack_seq
,
3041 return (after(ack
, tp
->snd_una
) ||
3042 after(ack_seq
, tp
->snd_wl1
) ||
3043 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3046 /* Update our send window.
3048 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3049 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3051 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3054 struct tcp_sock
*tp
= tcp_sk(sk
);
3056 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3058 if (likely(!tcp_hdr(skb
)->syn
))
3059 nwin
<<= tp
->rx_opt
.snd_wscale
;
3061 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3062 flag
|= FLAG_WIN_UPDATE
;
3063 tcp_update_wl(tp
, ack
, ack_seq
);
3065 if (tp
->snd_wnd
!= nwin
) {
3068 /* Note, it is the only place, where
3069 * fast path is recovered for sending TCP.
3072 tcp_fast_path_check(sk
);
3074 if (nwin
> tp
->max_window
) {
3075 tp
->max_window
= nwin
;
3076 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3086 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3087 * continue in congestion avoidance.
3089 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3091 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3092 tp
->snd_cwnd_cnt
= 0;
3093 tp
->bytes_acked
= 0;
3094 TCP_ECN_queue_cwr(tp
);
3095 tcp_moderate_cwnd(tp
);
3098 /* A conservative spurious RTO response algorithm: reduce cwnd using
3099 * rate halving and continue in congestion avoidance.
3101 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3103 tcp_enter_cwr(sk
, 0);
3106 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3108 if (flag
& FLAG_ECE
)
3109 tcp_ratehalving_spur_to_response(sk
);
3111 tcp_undo_cwr(sk
, 1);
3114 /* F-RTO spurious RTO detection algorithm (RFC4138)
3116 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3117 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3118 * window (but not to or beyond highest sequence sent before RTO):
3119 * On First ACK, send two new segments out.
3120 * On Second ACK, RTO was likely spurious. Do spurious response (response
3121 * algorithm is not part of the F-RTO detection algorithm
3122 * given in RFC4138 but can be selected separately).
3123 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3124 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3125 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3126 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3128 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3129 * original window even after we transmit two new data segments.
3132 * on first step, wait until first cumulative ACK arrives, then move to
3133 * the second step. In second step, the next ACK decides.
3135 * F-RTO is implemented (mainly) in four functions:
3136 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3137 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3138 * called when tcp_use_frto() showed green light
3139 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3140 * - tcp_enter_frto_loss() is called if there is not enough evidence
3141 * to prove that the RTO is indeed spurious. It transfers the control
3142 * from F-RTO to the conventional RTO recovery
3144 static int tcp_process_frto(struct sock
*sk
, int flag
)
3146 struct tcp_sock
*tp
= tcp_sk(sk
);
3148 tcp_verify_left_out(tp
);
3150 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3151 if (flag
& FLAG_DATA_ACKED
)
3152 inet_csk(sk
)->icsk_retransmits
= 0;
3154 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3155 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3156 tp
->undo_marker
= 0;
3158 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3159 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3163 if (!tcp_is_sackfrto(tp
)) {
3164 /* RFC4138 shortcoming in step 2; should also have case c):
3165 * ACK isn't duplicate nor advances window, e.g., opposite dir
3168 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3171 if (!(flag
& FLAG_DATA_ACKED
)) {
3172 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3177 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3178 /* Prevent sending of new data. */
3179 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3180 tcp_packets_in_flight(tp
));
3184 if ((tp
->frto_counter
>= 2) &&
3185 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3186 ((flag
& FLAG_DATA_SACKED
) &&
3187 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3188 /* RFC4138 shortcoming (see comment above) */
3189 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3190 (flag
& FLAG_NOT_DUP
))
3193 tcp_enter_frto_loss(sk
, 3, flag
);
3198 if (tp
->frto_counter
== 1) {
3199 /* tcp_may_send_now needs to see updated state */
3200 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3201 tp
->frto_counter
= 2;
3203 if (!tcp_may_send_now(sk
))
3204 tcp_enter_frto_loss(sk
, 2, flag
);
3208 switch (sysctl_tcp_frto_response
) {
3210 tcp_undo_spur_to_response(sk
, flag
);
3213 tcp_conservative_spur_to_response(tp
);
3216 tcp_ratehalving_spur_to_response(sk
);
3219 tp
->frto_counter
= 0;
3220 tp
->undo_marker
= 0;
3221 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3226 /* This routine deals with incoming acks, but not outgoing ones. */
3227 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3229 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3230 struct tcp_sock
*tp
= tcp_sk(sk
);
3231 u32 prior_snd_una
= tp
->snd_una
;
3232 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3233 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3234 u32 prior_in_flight
;
3239 /* If the ack is newer than sent or older than previous acks
3240 * then we can probably ignore it.
3242 if (after(ack
, tp
->snd_nxt
))
3243 goto uninteresting_ack
;
3245 if (before(ack
, prior_snd_una
))
3248 if (after(ack
, prior_snd_una
))
3249 flag
|= FLAG_SND_UNA_ADVANCED
;
3251 if (sysctl_tcp_abc
) {
3252 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3253 tp
->bytes_acked
+= ack
- prior_snd_una
;
3254 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3255 /* we assume just one segment left network */
3256 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3260 prior_fackets
= tp
->fackets_out
;
3261 prior_in_flight
= tcp_packets_in_flight(tp
);
3263 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3264 /* Window is constant, pure forward advance.
3265 * No more checks are required.
3266 * Note, we use the fact that SND.UNA>=SND.WL2.
3268 tcp_update_wl(tp
, ack
, ack_seq
);
3270 flag
|= FLAG_WIN_UPDATE
;
3272 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3274 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3276 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3279 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3281 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3283 if (TCP_SKB_CB(skb
)->sacked
)
3284 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3286 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3289 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3292 /* We passed data and got it acked, remove any soft error
3293 * log. Something worked...
3295 sk
->sk_err_soft
= 0;
3296 icsk
->icsk_probes_out
= 0;
3297 tp
->rcv_tstamp
= tcp_time_stamp
;
3298 prior_packets
= tp
->packets_out
;
3302 /* See if we can take anything off of the retransmit queue. */
3303 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
);
3305 if (tp
->frto_counter
)
3306 frto_cwnd
= tcp_process_frto(sk
, flag
);
3307 /* Guarantee sacktag reordering detection against wrap-arounds */
3308 if (before(tp
->frto_highmark
, tp
->snd_una
))
3309 tp
->frto_highmark
= 0;
3311 if (tcp_ack_is_dubious(sk
, flag
)) {
3312 /* Advance CWND, if state allows this. */
3313 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3314 tcp_may_raise_cwnd(sk
, flag
))
3315 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3316 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3319 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3320 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3323 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3324 dst_confirm(sk
->sk_dst_cache
);
3329 /* If this ack opens up a zero window, clear backoff. It was
3330 * being used to time the probes, and is probably far higher than
3331 * it needs to be for normal retransmission.
3333 if (tcp_send_head(sk
))
3338 if (TCP_SKB_CB(skb
)->sacked
) {
3339 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3340 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3341 tcp_try_keep_open(sk
);
3345 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3349 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3350 * But, this can also be called on packets in the established flow when
3351 * the fast version below fails.
3353 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3357 struct tcphdr
*th
= tcp_hdr(skb
);
3358 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3360 ptr
= (unsigned char *)(th
+ 1);
3361 opt_rx
->saw_tstamp
= 0;
3363 while (length
> 0) {
3364 int opcode
= *ptr
++;
3370 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3375 if (opsize
< 2) /* "silly options" */
3377 if (opsize
> length
)
3378 return; /* don't parse partial options */
3381 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3382 u16 in_mss
= get_unaligned_be16(ptr
);
3384 if (opt_rx
->user_mss
&&
3385 opt_rx
->user_mss
< in_mss
)
3386 in_mss
= opt_rx
->user_mss
;
3387 opt_rx
->mss_clamp
= in_mss
;
3392 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3393 !estab
&& sysctl_tcp_window_scaling
) {
3394 __u8 snd_wscale
= *(__u8
*)ptr
;
3395 opt_rx
->wscale_ok
= 1;
3396 if (snd_wscale
> 14) {
3397 if (net_ratelimit())
3398 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3399 "scaling value %d >14 received.\n",
3403 opt_rx
->snd_wscale
= snd_wscale
;
3406 case TCPOPT_TIMESTAMP
:
3407 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3408 ((estab
&& opt_rx
->tstamp_ok
) ||
3409 (!estab
&& sysctl_tcp_timestamps
))) {
3410 opt_rx
->saw_tstamp
= 1;
3411 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3412 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3415 case TCPOPT_SACK_PERM
:
3416 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3417 !estab
&& sysctl_tcp_sack
) {
3418 opt_rx
->sack_ok
= 1;
3419 tcp_sack_reset(opt_rx
);
3424 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3425 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3427 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3430 #ifdef CONFIG_TCP_MD5SIG
3433 * The MD5 Hash has already been
3434 * checked (see tcp_v{4,6}_do_rcv()).
3446 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3448 __be32
*ptr
= (__be32
*)(th
+ 1);
3450 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3451 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3452 tp
->rx_opt
.saw_tstamp
= 1;
3454 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3456 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3462 /* Fast parse options. This hopes to only see timestamps.
3463 * If it is wrong it falls back on tcp_parse_options().
3465 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3466 struct tcp_sock
*tp
)
3468 if (th
->doff
== sizeof(struct tcphdr
) >> 2) {
3469 tp
->rx_opt
.saw_tstamp
= 0;
3471 } else if (tp
->rx_opt
.tstamp_ok
&&
3472 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3473 if (tcp_parse_aligned_timestamp(tp
, th
))
3476 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3480 #ifdef CONFIG_TCP_MD5SIG
3482 * Parse MD5 Signature option
3484 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3486 int length
= (th
->doff
<< 2) - sizeof (*th
);
3487 u8
*ptr
= (u8
*)(th
+ 1);
3489 /* If the TCP option is too short, we can short cut */
3490 if (length
< TCPOLEN_MD5SIG
)
3493 while (length
> 0) {
3494 int opcode
= *ptr
++;
3505 if (opsize
< 2 || opsize
> length
)
3507 if (opcode
== TCPOPT_MD5SIG
)
3517 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3519 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3520 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3523 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3525 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3526 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3527 * extra check below makes sure this can only happen
3528 * for pure ACK frames. -DaveM
3530 * Not only, also it occurs for expired timestamps.
3533 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3534 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3535 tcp_store_ts_recent(tp
);
3539 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3541 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3542 * it can pass through stack. So, the following predicate verifies that
3543 * this segment is not used for anything but congestion avoidance or
3544 * fast retransmit. Moreover, we even are able to eliminate most of such
3545 * second order effects, if we apply some small "replay" window (~RTO)
3546 * to timestamp space.
3548 * All these measures still do not guarantee that we reject wrapped ACKs
3549 * on networks with high bandwidth, when sequence space is recycled fastly,
3550 * but it guarantees that such events will be very rare and do not affect
3551 * connection seriously. This doesn't look nice, but alas, PAWS is really
3554 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3555 * states that events when retransmit arrives after original data are rare.
3556 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3557 * the biggest problem on large power networks even with minor reordering.
3558 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3559 * up to bandwidth of 18Gigabit/sec. 8) ]
3562 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3564 struct tcp_sock
*tp
= tcp_sk(sk
);
3565 struct tcphdr
*th
= tcp_hdr(skb
);
3566 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3567 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3569 return (/* 1. Pure ACK with correct sequence number. */
3570 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3572 /* 2. ... and duplicate ACK. */
3573 ack
== tp
->snd_una
&&
3575 /* 3. ... and does not update window. */
3576 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3578 /* 4. ... and sits in replay window. */
3579 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3582 static inline int tcp_paws_discard(const struct sock
*sk
,
3583 const struct sk_buff
*skb
)
3585 const struct tcp_sock
*tp
= tcp_sk(sk
);
3586 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3587 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3588 !tcp_disordered_ack(sk
, skb
));
3591 /* Check segment sequence number for validity.
3593 * Segment controls are considered valid, if the segment
3594 * fits to the window after truncation to the window. Acceptability
3595 * of data (and SYN, FIN, of course) is checked separately.
3596 * See tcp_data_queue(), for example.
3598 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3599 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3600 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3601 * (borrowed from freebsd)
3604 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3606 return !before(end_seq
, tp
->rcv_wup
) &&
3607 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3610 /* When we get a reset we do this. */
3611 static void tcp_reset(struct sock
*sk
)
3613 /* We want the right error as BSD sees it (and indeed as we do). */
3614 switch (sk
->sk_state
) {
3616 sk
->sk_err
= ECONNREFUSED
;
3618 case TCP_CLOSE_WAIT
:
3624 sk
->sk_err
= ECONNRESET
;
3627 if (!sock_flag(sk
, SOCK_DEAD
))
3628 sk
->sk_error_report(sk
);
3634 * Process the FIN bit. This now behaves as it is supposed to work
3635 * and the FIN takes effect when it is validly part of sequence
3636 * space. Not before when we get holes.
3638 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3639 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3642 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3643 * close and we go into CLOSING (and later onto TIME-WAIT)
3645 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3647 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3649 struct tcp_sock
*tp
= tcp_sk(sk
);
3651 inet_csk_schedule_ack(sk
);
3653 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3654 sock_set_flag(sk
, SOCK_DONE
);
3656 switch (sk
->sk_state
) {
3658 case TCP_ESTABLISHED
:
3659 /* Move to CLOSE_WAIT */
3660 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3661 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3664 case TCP_CLOSE_WAIT
:
3666 /* Received a retransmission of the FIN, do
3671 /* RFC793: Remain in the LAST-ACK state. */
3675 /* This case occurs when a simultaneous close
3676 * happens, we must ack the received FIN and
3677 * enter the CLOSING state.
3680 tcp_set_state(sk
, TCP_CLOSING
);
3683 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3685 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3688 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3689 * cases we should never reach this piece of code.
3691 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
3692 __func__
, sk
->sk_state
);
3696 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3697 * Probably, we should reset in this case. For now drop them.
3699 __skb_queue_purge(&tp
->out_of_order_queue
);
3700 if (tcp_is_sack(tp
))
3701 tcp_sack_reset(&tp
->rx_opt
);
3704 if (!sock_flag(sk
, SOCK_DEAD
)) {
3705 sk
->sk_state_change(sk
);
3707 /* Do not send POLL_HUP for half duplex close. */
3708 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3709 sk
->sk_state
== TCP_CLOSE
)
3710 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3712 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3716 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3719 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3720 if (before(seq
, sp
->start_seq
))
3721 sp
->start_seq
= seq
;
3722 if (after(end_seq
, sp
->end_seq
))
3723 sp
->end_seq
= end_seq
;
3729 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3731 struct tcp_sock
*tp
= tcp_sk(sk
);
3733 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3736 if (before(seq
, tp
->rcv_nxt
))
3737 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3739 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3741 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3743 tp
->rx_opt
.dsack
= 1;
3744 tp
->duplicate_sack
[0].start_seq
= seq
;
3745 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3746 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ 1;
3750 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3752 struct tcp_sock
*tp
= tcp_sk(sk
);
3754 if (!tp
->rx_opt
.dsack
)
3755 tcp_dsack_set(sk
, seq
, end_seq
);
3757 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3760 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
3762 struct tcp_sock
*tp
= tcp_sk(sk
);
3764 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3765 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3766 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3767 tcp_enter_quickack_mode(sk
);
3769 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3770 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3772 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3773 end_seq
= tp
->rcv_nxt
;
3774 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3781 /* These routines update the SACK block as out-of-order packets arrive or
3782 * in-order packets close up the sequence space.
3784 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3787 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3788 struct tcp_sack_block
*swalk
= sp
+ 1;
3790 /* See if the recent change to the first SACK eats into
3791 * or hits the sequence space of other SACK blocks, if so coalesce.
3793 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3794 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3797 /* Zap SWALK, by moving every further SACK up by one slot.
3798 * Decrease num_sacks.
3800 tp
->rx_opt
.num_sacks
--;
3801 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
3803 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3807 this_sack
++, swalk
++;
3811 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
,
3812 struct tcp_sack_block
*sack2
)
3816 tmp
= sack1
->start_seq
;
3817 sack1
->start_seq
= sack2
->start_seq
;
3818 sack2
->start_seq
= tmp
;
3820 tmp
= sack1
->end_seq
;
3821 sack1
->end_seq
= sack2
->end_seq
;
3822 sack2
->end_seq
= tmp
;
3825 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3827 struct tcp_sock
*tp
= tcp_sk(sk
);
3828 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3829 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3835 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3836 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3837 /* Rotate this_sack to the first one. */
3838 for (; this_sack
> 0; this_sack
--, sp
--)
3839 tcp_sack_swap(sp
, sp
- 1);
3841 tcp_sack_maybe_coalesce(tp
);
3846 /* Could not find an adjacent existing SACK, build a new one,
3847 * put it at the front, and shift everyone else down. We
3848 * always know there is at least one SACK present already here.
3850 * If the sack array is full, forget about the last one.
3852 if (this_sack
>= TCP_NUM_SACKS
) {
3854 tp
->rx_opt
.num_sacks
--;
3857 for (; this_sack
> 0; this_sack
--, sp
--)
3861 /* Build the new head SACK, and we're done. */
3862 sp
->start_seq
= seq
;
3863 sp
->end_seq
= end_seq
;
3864 tp
->rx_opt
.num_sacks
++;
3865 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
3868 /* RCV.NXT advances, some SACKs should be eaten. */
3870 static void tcp_sack_remove(struct tcp_sock
*tp
)
3872 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3873 int num_sacks
= tp
->rx_opt
.num_sacks
;
3876 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3877 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3878 tp
->rx_opt
.num_sacks
= 0;
3879 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
3883 for (this_sack
= 0; this_sack
< num_sacks
;) {
3884 /* Check if the start of the sack is covered by RCV.NXT. */
3885 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3888 /* RCV.NXT must cover all the block! */
3889 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
3891 /* Zap this SACK, by moving forward any other SACKS. */
3892 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3893 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3900 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
3901 tp
->rx_opt
.num_sacks
= num_sacks
;
3902 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
3907 /* This one checks to see if we can put data from the
3908 * out_of_order queue into the receive_queue.
3910 static void tcp_ofo_queue(struct sock
*sk
)
3912 struct tcp_sock
*tp
= tcp_sk(sk
);
3913 __u32 dsack_high
= tp
->rcv_nxt
;
3914 struct sk_buff
*skb
;
3916 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3917 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3920 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3921 __u32 dsack
= dsack_high
;
3922 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3923 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3924 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
3927 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3928 SOCK_DEBUG(sk
, "ofo packet was already received \n");
3929 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3933 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3934 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3935 TCP_SKB_CB(skb
)->end_seq
);
3937 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3938 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3939 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3940 if (tcp_hdr(skb
)->fin
)
3941 tcp_fin(skb
, sk
, tcp_hdr(skb
));
3945 static int tcp_prune_ofo_queue(struct sock
*sk
);
3946 static int tcp_prune_queue(struct sock
*sk
);
3948 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
3950 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3951 !sk_rmem_schedule(sk
, size
)) {
3953 if (tcp_prune_queue(sk
) < 0)
3956 if (!sk_rmem_schedule(sk
, size
)) {
3957 if (!tcp_prune_ofo_queue(sk
))
3960 if (!sk_rmem_schedule(sk
, size
))
3967 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
3969 struct tcphdr
*th
= tcp_hdr(skb
);
3970 struct tcp_sock
*tp
= tcp_sk(sk
);
3973 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
3976 __skb_pull(skb
, th
->doff
* 4);
3978 TCP_ECN_accept_cwr(tp
, skb
);
3980 if (tp
->rx_opt
.dsack
) {
3981 tp
->rx_opt
.dsack
= 0;
3982 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
;
3985 /* Queue data for delivery to the user.
3986 * Packets in sequence go to the receive queue.
3987 * Out of sequence packets to the out_of_order_queue.
3989 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
3990 if (tcp_receive_window(tp
) == 0)
3993 /* Ok. In sequence. In window. */
3994 if (tp
->ucopy
.task
== current
&&
3995 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
3996 sock_owned_by_user(sk
) && !tp
->urg_data
) {
3997 int chunk
= min_t(unsigned int, skb
->len
,
4000 __set_current_state(TASK_RUNNING
);
4003 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4004 tp
->ucopy
.len
-= chunk
;
4005 tp
->copied_seq
+= chunk
;
4006 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4007 tcp_rcv_space_adjust(sk
);
4015 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4018 skb_set_owner_r(skb
, sk
);
4019 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4021 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4023 tcp_event_data_recv(sk
, skb
);
4025 tcp_fin(skb
, sk
, th
);
4027 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4030 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4031 * gap in queue is filled.
4033 if (skb_queue_empty(&tp
->out_of_order_queue
))
4034 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4037 if (tp
->rx_opt
.num_sacks
)
4038 tcp_sack_remove(tp
);
4040 tcp_fast_path_check(sk
);
4044 else if (!sock_flag(sk
, SOCK_DEAD
))
4045 sk
->sk_data_ready(sk
, 0);
4049 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4050 /* A retransmit, 2nd most common case. Force an immediate ack. */
4051 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4052 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4055 tcp_enter_quickack_mode(sk
);
4056 inet_csk_schedule_ack(sk
);
4062 /* Out of window. F.e. zero window probe. */
4063 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4066 tcp_enter_quickack_mode(sk
);
4068 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4069 /* Partial packet, seq < rcv_next < end_seq */
4070 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4071 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4072 TCP_SKB_CB(skb
)->end_seq
);
4074 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4076 /* If window is closed, drop tail of packet. But after
4077 * remembering D-SACK for its head made in previous line.
4079 if (!tcp_receive_window(tp
))
4084 TCP_ECN_check_ce(tp
, skb
);
4086 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4089 /* Disable header prediction. */
4091 inet_csk_schedule_ack(sk
);
4093 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4094 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4096 skb_set_owner_r(skb
, sk
);
4098 if (!skb_peek(&tp
->out_of_order_queue
)) {
4099 /* Initial out of order segment, build 1 SACK. */
4100 if (tcp_is_sack(tp
)) {
4101 tp
->rx_opt
.num_sacks
= 1;
4102 tp
->rx_opt
.dsack
= 0;
4103 tp
->rx_opt
.eff_sacks
= 1;
4104 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4105 tp
->selective_acks
[0].end_seq
=
4106 TCP_SKB_CB(skb
)->end_seq
;
4108 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4110 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
4111 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4112 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4114 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4115 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4117 if (!tp
->rx_opt
.num_sacks
||
4118 tp
->selective_acks
[0].end_seq
!= seq
)
4121 /* Common case: data arrive in order after hole. */
4122 tp
->selective_acks
[0].end_seq
= end_seq
;
4126 /* Find place to insert this segment. */
4128 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4130 } while ((skb1
= skb1
->prev
) !=
4131 (struct sk_buff
*)&tp
->out_of_order_queue
);
4133 /* Do skb overlap to previous one? */
4134 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
4135 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4136 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4137 /* All the bits are present. Drop. */
4139 tcp_dsack_set(sk
, seq
, end_seq
);
4142 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4143 /* Partial overlap. */
4144 tcp_dsack_set(sk
, seq
,
4145 TCP_SKB_CB(skb1
)->end_seq
);
4150 __skb_insert(skb
, skb1
, skb1
->next
, &tp
->out_of_order_queue
);
4152 /* And clean segments covered by new one as whole. */
4153 while ((skb1
= skb
->next
) !=
4154 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4155 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4156 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4157 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4161 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4162 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4163 TCP_SKB_CB(skb1
)->end_seq
);
4168 if (tcp_is_sack(tp
))
4169 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4173 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4174 struct sk_buff_head
*list
)
4176 struct sk_buff
*next
= skb
->next
;
4178 __skb_unlink(skb
, list
);
4180 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4185 /* Collapse contiguous sequence of skbs head..tail with
4186 * sequence numbers start..end.
4187 * Segments with FIN/SYN are not collapsed (only because this
4191 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4192 struct sk_buff
*head
, struct sk_buff
*tail
,
4195 struct sk_buff
*skb
;
4197 /* First, check that queue is collapsible and find
4198 * the point where collapsing can be useful. */
4199 for (skb
= head
; skb
!= tail
;) {
4200 /* No new bits? It is possible on ofo queue. */
4201 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4202 skb
= tcp_collapse_one(sk
, skb
, list
);
4206 /* The first skb to collapse is:
4208 * - bloated or contains data before "start" or
4209 * overlaps to the next one.
4211 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4212 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4213 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4214 (skb
->next
!= tail
&&
4215 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4218 /* Decided to skip this, advance start seq. */
4219 start
= TCP_SKB_CB(skb
)->end_seq
;
4222 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4225 while (before(start
, end
)) {
4226 struct sk_buff
*nskb
;
4227 unsigned int header
= skb_headroom(skb
);
4228 int copy
= SKB_MAX_ORDER(header
, 0);
4230 /* Too big header? This can happen with IPv6. */
4233 if (end
- start
< copy
)
4235 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4239 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4240 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4242 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4244 skb_reserve(nskb
, header
);
4245 memcpy(nskb
->head
, skb
->head
, header
);
4246 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4247 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4248 __skb_insert(nskb
, skb
->prev
, skb
, list
);
4249 skb_set_owner_r(nskb
, sk
);
4251 /* Copy data, releasing collapsed skbs. */
4253 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4254 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4258 size
= min(copy
, size
);
4259 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4261 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4265 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4266 skb
= tcp_collapse_one(sk
, skb
, list
);
4268 tcp_hdr(skb
)->syn
||
4276 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4277 * and tcp_collapse() them until all the queue is collapsed.
4279 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4281 struct tcp_sock
*tp
= tcp_sk(sk
);
4282 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4283 struct sk_buff
*head
;
4289 start
= TCP_SKB_CB(skb
)->seq
;
4290 end
= TCP_SKB_CB(skb
)->end_seq
;
4296 /* Segment is terminated when we see gap or when
4297 * we are at the end of all the queue. */
4298 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4299 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4300 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4301 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4302 head
, skb
, start
, end
);
4304 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4306 /* Start new segment */
4307 start
= TCP_SKB_CB(skb
)->seq
;
4308 end
= TCP_SKB_CB(skb
)->end_seq
;
4310 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4311 start
= TCP_SKB_CB(skb
)->seq
;
4312 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4313 end
= TCP_SKB_CB(skb
)->end_seq
;
4319 * Purge the out-of-order queue.
4320 * Return true if queue was pruned.
4322 static int tcp_prune_ofo_queue(struct sock
*sk
)
4324 struct tcp_sock
*tp
= tcp_sk(sk
);
4327 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4328 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4329 __skb_queue_purge(&tp
->out_of_order_queue
);
4331 /* Reset SACK state. A conforming SACK implementation will
4332 * do the same at a timeout based retransmit. When a connection
4333 * is in a sad state like this, we care only about integrity
4334 * of the connection not performance.
4336 if (tp
->rx_opt
.sack_ok
)
4337 tcp_sack_reset(&tp
->rx_opt
);
4344 /* Reduce allocated memory if we can, trying to get
4345 * the socket within its memory limits again.
4347 * Return less than zero if we should start dropping frames
4348 * until the socket owning process reads some of the data
4349 * to stabilize the situation.
4351 static int tcp_prune_queue(struct sock
*sk
)
4353 struct tcp_sock
*tp
= tcp_sk(sk
);
4355 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4357 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4359 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4360 tcp_clamp_window(sk
);
4361 else if (tcp_memory_pressure
)
4362 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4364 tcp_collapse_ofo_queue(sk
);
4365 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4366 sk
->sk_receive_queue
.next
,
4367 (struct sk_buff
*)&sk
->sk_receive_queue
,
4368 tp
->copied_seq
, tp
->rcv_nxt
);
4371 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4374 /* Collapsing did not help, destructive actions follow.
4375 * This must not ever occur. */
4377 tcp_prune_ofo_queue(sk
);
4379 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4382 /* If we are really being abused, tell the caller to silently
4383 * drop receive data on the floor. It will get retransmitted
4384 * and hopefully then we'll have sufficient space.
4386 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4388 /* Massive buffer overcommit. */
4393 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4394 * As additional protections, we do not touch cwnd in retransmission phases,
4395 * and if application hit its sndbuf limit recently.
4397 void tcp_cwnd_application_limited(struct sock
*sk
)
4399 struct tcp_sock
*tp
= tcp_sk(sk
);
4401 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4402 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4403 /* Limited by application or receiver window. */
4404 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4405 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4406 if (win_used
< tp
->snd_cwnd
) {
4407 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4408 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4410 tp
->snd_cwnd_used
= 0;
4412 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4415 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4417 struct tcp_sock
*tp
= tcp_sk(sk
);
4419 /* If the user specified a specific send buffer setting, do
4422 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4425 /* If we are under global TCP memory pressure, do not expand. */
4426 if (tcp_memory_pressure
)
4429 /* If we are under soft global TCP memory pressure, do not expand. */
4430 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4433 /* If we filled the congestion window, do not expand. */
4434 if (tp
->packets_out
>= tp
->snd_cwnd
)
4440 /* When incoming ACK allowed to free some skb from write_queue,
4441 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4442 * on the exit from tcp input handler.
4444 * PROBLEM: sndbuf expansion does not work well with largesend.
4446 static void tcp_new_space(struct sock
*sk
)
4448 struct tcp_sock
*tp
= tcp_sk(sk
);
4450 if (tcp_should_expand_sndbuf(sk
)) {
4451 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4452 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
),
4453 demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4454 tp
->reordering
+ 1);
4455 sndmem
*= 2 * demanded
;
4456 if (sndmem
> sk
->sk_sndbuf
)
4457 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4458 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4461 sk
->sk_write_space(sk
);
4464 static void tcp_check_space(struct sock
*sk
)
4466 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4467 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4468 if (sk
->sk_socket
&&
4469 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4474 static inline void tcp_data_snd_check(struct sock
*sk
)
4476 tcp_push_pending_frames(sk
);
4477 tcp_check_space(sk
);
4481 * Check if sending an ack is needed.
4483 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4485 struct tcp_sock
*tp
= tcp_sk(sk
);
4487 /* More than one full frame received... */
4488 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4489 /* ... and right edge of window advances far enough.
4490 * (tcp_recvmsg() will send ACK otherwise). Or...
4492 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4493 /* We ACK each frame or... */
4494 tcp_in_quickack_mode(sk
) ||
4495 /* We have out of order data. */
4496 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4497 /* Then ack it now */
4500 /* Else, send delayed ack. */
4501 tcp_send_delayed_ack(sk
);
4505 static inline void tcp_ack_snd_check(struct sock
*sk
)
4507 if (!inet_csk_ack_scheduled(sk
)) {
4508 /* We sent a data segment already. */
4511 __tcp_ack_snd_check(sk
, 1);
4515 * This routine is only called when we have urgent data
4516 * signaled. Its the 'slow' part of tcp_urg. It could be
4517 * moved inline now as tcp_urg is only called from one
4518 * place. We handle URGent data wrong. We have to - as
4519 * BSD still doesn't use the correction from RFC961.
4520 * For 1003.1g we should support a new option TCP_STDURG to permit
4521 * either form (or just set the sysctl tcp_stdurg).
4524 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4526 struct tcp_sock
*tp
= tcp_sk(sk
);
4527 u32 ptr
= ntohs(th
->urg_ptr
);
4529 if (ptr
&& !sysctl_tcp_stdurg
)
4531 ptr
+= ntohl(th
->seq
);
4533 /* Ignore urgent data that we've already seen and read. */
4534 if (after(tp
->copied_seq
, ptr
))
4537 /* Do not replay urg ptr.
4539 * NOTE: interesting situation not covered by specs.
4540 * Misbehaving sender may send urg ptr, pointing to segment,
4541 * which we already have in ofo queue. We are not able to fetch
4542 * such data and will stay in TCP_URG_NOTYET until will be eaten
4543 * by recvmsg(). Seems, we are not obliged to handle such wicked
4544 * situations. But it is worth to think about possibility of some
4545 * DoSes using some hypothetical application level deadlock.
4547 if (before(ptr
, tp
->rcv_nxt
))
4550 /* Do we already have a newer (or duplicate) urgent pointer? */
4551 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4554 /* Tell the world about our new urgent pointer. */
4557 /* We may be adding urgent data when the last byte read was
4558 * urgent. To do this requires some care. We cannot just ignore
4559 * tp->copied_seq since we would read the last urgent byte again
4560 * as data, nor can we alter copied_seq until this data arrives
4561 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4563 * NOTE. Double Dutch. Rendering to plain English: author of comment
4564 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4565 * and expect that both A and B disappear from stream. This is _wrong_.
4566 * Though this happens in BSD with high probability, this is occasional.
4567 * Any application relying on this is buggy. Note also, that fix "works"
4568 * only in this artificial test. Insert some normal data between A and B and we will
4569 * decline of BSD again. Verdict: it is better to remove to trap
4572 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4573 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4574 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4576 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4577 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4582 tp
->urg_data
= TCP_URG_NOTYET
;
4585 /* Disable header prediction. */
4589 /* This is the 'fast' part of urgent handling. */
4590 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4592 struct tcp_sock
*tp
= tcp_sk(sk
);
4594 /* Check if we get a new urgent pointer - normally not. */
4596 tcp_check_urg(sk
, th
);
4598 /* Do we wait for any urgent data? - normally not... */
4599 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4600 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4603 /* Is the urgent pointer pointing into this packet? */
4604 if (ptr
< skb
->len
) {
4606 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4608 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4609 if (!sock_flag(sk
, SOCK_DEAD
))
4610 sk
->sk_data_ready(sk
, 0);
4615 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4617 struct tcp_sock
*tp
= tcp_sk(sk
);
4618 int chunk
= skb
->len
- hlen
;
4622 if (skb_csum_unnecessary(skb
))
4623 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4625 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4629 tp
->ucopy
.len
-= chunk
;
4630 tp
->copied_seq
+= chunk
;
4631 tcp_rcv_space_adjust(sk
);
4638 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4639 struct sk_buff
*skb
)
4643 if (sock_owned_by_user(sk
)) {
4645 result
= __tcp_checksum_complete(skb
);
4648 result
= __tcp_checksum_complete(skb
);
4653 static inline int tcp_checksum_complete_user(struct sock
*sk
,
4654 struct sk_buff
*skb
)
4656 return !skb_csum_unnecessary(skb
) &&
4657 __tcp_checksum_complete_user(sk
, skb
);
4660 #ifdef CONFIG_NET_DMA
4661 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4664 struct tcp_sock
*tp
= tcp_sk(sk
);
4665 int chunk
= skb
->len
- hlen
;
4667 int copied_early
= 0;
4669 if (tp
->ucopy
.wakeup
)
4672 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4673 tp
->ucopy
.dma_chan
= get_softnet_dma();
4675 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4677 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4679 tp
->ucopy
.iov
, chunk
,
4680 tp
->ucopy
.pinned_list
);
4685 tp
->ucopy
.dma_cookie
= dma_cookie
;
4688 tp
->ucopy
.len
-= chunk
;
4689 tp
->copied_seq
+= chunk
;
4690 tcp_rcv_space_adjust(sk
);
4692 if ((tp
->ucopy
.len
== 0) ||
4693 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4694 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4695 tp
->ucopy
.wakeup
= 1;
4696 sk
->sk_data_ready(sk
, 0);
4698 } else if (chunk
> 0) {
4699 tp
->ucopy
.wakeup
= 1;
4700 sk
->sk_data_ready(sk
, 0);
4703 return copied_early
;
4705 #endif /* CONFIG_NET_DMA */
4707 /* Does PAWS and seqno based validation of an incoming segment, flags will
4708 * play significant role here.
4710 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
4711 struct tcphdr
*th
, int syn_inerr
)
4713 struct tcp_sock
*tp
= tcp_sk(sk
);
4715 /* RFC1323: H1. Apply PAWS check first. */
4716 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4717 tcp_paws_discard(sk
, skb
)) {
4719 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
4720 tcp_send_dupack(sk
, skb
);
4723 /* Reset is accepted even if it did not pass PAWS. */
4726 /* Step 1: check sequence number */
4727 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4728 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4729 * (RST) segments are validated by checking their SEQ-fields."
4730 * And page 69: "If an incoming segment is not acceptable,
4731 * an acknowledgment should be sent in reply (unless the RST
4732 * bit is set, if so drop the segment and return)".
4735 tcp_send_dupack(sk
, skb
);
4739 /* Step 2: check RST bit */
4745 /* ts_recent update must be made after we are sure that the packet
4748 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
4750 /* step 3: check security and precedence [ignored] */
4752 /* step 4: Check for a SYN in window. */
4753 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4755 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4756 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
4769 * TCP receive function for the ESTABLISHED state.
4771 * It is split into a fast path and a slow path. The fast path is
4773 * - A zero window was announced from us - zero window probing
4774 * is only handled properly in the slow path.
4775 * - Out of order segments arrived.
4776 * - Urgent data is expected.
4777 * - There is no buffer space left
4778 * - Unexpected TCP flags/window values/header lengths are received
4779 * (detected by checking the TCP header against pred_flags)
4780 * - Data is sent in both directions. Fast path only supports pure senders
4781 * or pure receivers (this means either the sequence number or the ack
4782 * value must stay constant)
4783 * - Unexpected TCP option.
4785 * When these conditions are not satisfied it drops into a standard
4786 * receive procedure patterned after RFC793 to handle all cases.
4787 * The first three cases are guaranteed by proper pred_flags setting,
4788 * the rest is checked inline. Fast processing is turned on in
4789 * tcp_data_queue when everything is OK.
4791 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
4792 struct tcphdr
*th
, unsigned len
)
4794 struct tcp_sock
*tp
= tcp_sk(sk
);
4798 * Header prediction.
4799 * The code loosely follows the one in the famous
4800 * "30 instruction TCP receive" Van Jacobson mail.
4802 * Van's trick is to deposit buffers into socket queue
4803 * on a device interrupt, to call tcp_recv function
4804 * on the receive process context and checksum and copy
4805 * the buffer to user space. smart...
4807 * Our current scheme is not silly either but we take the
4808 * extra cost of the net_bh soft interrupt processing...
4809 * We do checksum and copy also but from device to kernel.
4812 tp
->rx_opt
.saw_tstamp
= 0;
4814 /* pred_flags is 0xS?10 << 16 + snd_wnd
4815 * if header_prediction is to be made
4816 * 'S' will always be tp->tcp_header_len >> 2
4817 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4818 * turn it off (when there are holes in the receive
4819 * space for instance)
4820 * PSH flag is ignored.
4823 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
4824 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4825 int tcp_header_len
= tp
->tcp_header_len
;
4827 /* Timestamp header prediction: tcp_header_len
4828 * is automatically equal to th->doff*4 due to pred_flags
4832 /* Check timestamp */
4833 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
4834 /* No? Slow path! */
4835 if (!tcp_parse_aligned_timestamp(tp
, th
))
4838 /* If PAWS failed, check it more carefully in slow path */
4839 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
4842 /* DO NOT update ts_recent here, if checksum fails
4843 * and timestamp was corrupted part, it will result
4844 * in a hung connection since we will drop all
4845 * future packets due to the PAWS test.
4849 if (len
<= tcp_header_len
) {
4850 /* Bulk data transfer: sender */
4851 if (len
== tcp_header_len
) {
4852 /* Predicted packet is in window by definition.
4853 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4854 * Hence, check seq<=rcv_wup reduces to:
4856 if (tcp_header_len
==
4857 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4858 tp
->rcv_nxt
== tp
->rcv_wup
)
4859 tcp_store_ts_recent(tp
);
4861 /* We know that such packets are checksummed
4864 tcp_ack(sk
, skb
, 0);
4866 tcp_data_snd_check(sk
);
4868 } else { /* Header too small */
4869 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4874 int copied_early
= 0;
4876 if (tp
->copied_seq
== tp
->rcv_nxt
&&
4877 len
- tcp_header_len
<= tp
->ucopy
.len
) {
4878 #ifdef CONFIG_NET_DMA
4879 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
4884 if (tp
->ucopy
.task
== current
&&
4885 sock_owned_by_user(sk
) && !copied_early
) {
4886 __set_current_state(TASK_RUNNING
);
4888 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
4892 /* Predicted packet is in window by definition.
4893 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4894 * Hence, check seq<=rcv_wup reduces to:
4896 if (tcp_header_len
==
4897 (sizeof(struct tcphdr
) +
4898 TCPOLEN_TSTAMP_ALIGNED
) &&
4899 tp
->rcv_nxt
== tp
->rcv_wup
)
4900 tcp_store_ts_recent(tp
);
4902 tcp_rcv_rtt_measure_ts(sk
, skb
);
4904 __skb_pull(skb
, tcp_header_len
);
4905 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4906 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
4909 tcp_cleanup_rbuf(sk
, skb
->len
);
4912 if (tcp_checksum_complete_user(sk
, skb
))
4915 /* Predicted packet is in window by definition.
4916 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4917 * Hence, check seq<=rcv_wup reduces to:
4919 if (tcp_header_len
==
4920 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
4921 tp
->rcv_nxt
== tp
->rcv_wup
)
4922 tcp_store_ts_recent(tp
);
4924 tcp_rcv_rtt_measure_ts(sk
, skb
);
4926 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
4929 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
4931 /* Bulk data transfer: receiver */
4932 __skb_pull(skb
, tcp_header_len
);
4933 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4934 skb_set_owner_r(skb
, sk
);
4935 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4938 tcp_event_data_recv(sk
, skb
);
4940 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
4941 /* Well, only one small jumplet in fast path... */
4942 tcp_ack(sk
, skb
, FLAG_DATA
);
4943 tcp_data_snd_check(sk
);
4944 if (!inet_csk_ack_scheduled(sk
))
4948 __tcp_ack_snd_check(sk
, 0);
4950 #ifdef CONFIG_NET_DMA
4952 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
4958 sk
->sk_data_ready(sk
, 0);
4964 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
4968 * Standard slow path.
4971 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
4977 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
4979 tcp_rcv_rtt_measure_ts(sk
, skb
);
4981 /* Process urgent data. */
4982 tcp_urg(sk
, skb
, th
);
4984 /* step 7: process the segment text */
4985 tcp_data_queue(sk
, skb
);
4987 tcp_data_snd_check(sk
);
4988 tcp_ack_snd_check(sk
);
4992 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4999 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5000 struct tcphdr
*th
, unsigned len
)
5002 struct tcp_sock
*tp
= tcp_sk(sk
);
5003 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5004 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5006 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
5010 * "If the state is SYN-SENT then
5011 * first check the ACK bit
5012 * If the ACK bit is set
5013 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5014 * a reset (unless the RST bit is set, if so drop
5015 * the segment and return)"
5017 * We do not send data with SYN, so that RFC-correct
5020 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5021 goto reset_and_undo
;
5023 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5024 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5026 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5027 goto reset_and_undo
;
5030 /* Now ACK is acceptable.
5032 * "If the RST bit is set
5033 * If the ACK was acceptable then signal the user "error:
5034 * connection reset", drop the segment, enter CLOSED state,
5035 * delete TCB, and return."
5044 * "fifth, if neither of the SYN or RST bits is set then
5045 * drop the segment and return."
5051 goto discard_and_undo
;
5054 * "If the SYN bit is on ...
5055 * are acceptable then ...
5056 * (our SYN has been ACKed), change the connection
5057 * state to ESTABLISHED..."
5060 TCP_ECN_rcv_synack(tp
, th
);
5062 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5063 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5065 /* Ok.. it's good. Set up sequence numbers and
5066 * move to established.
5068 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5069 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5071 /* RFC1323: The window in SYN & SYN/ACK segments is
5074 tp
->snd_wnd
= ntohs(th
->window
);
5075 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
5077 if (!tp
->rx_opt
.wscale_ok
) {
5078 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5079 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5082 if (tp
->rx_opt
.saw_tstamp
) {
5083 tp
->rx_opt
.tstamp_ok
= 1;
5084 tp
->tcp_header_len
=
5085 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5086 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5087 tcp_store_ts_recent(tp
);
5089 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5092 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5093 tcp_enable_fack(tp
);
5096 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5097 tcp_initialize_rcv_mss(sk
);
5099 /* Remember, tcp_poll() does not lock socket!
5100 * Change state from SYN-SENT only after copied_seq
5101 * is initialized. */
5102 tp
->copied_seq
= tp
->rcv_nxt
;
5104 tcp_set_state(sk
, TCP_ESTABLISHED
);
5106 security_inet_conn_established(sk
, skb
);
5108 /* Make sure socket is routed, for correct metrics. */
5109 icsk
->icsk_af_ops
->rebuild_header(sk
);
5111 tcp_init_metrics(sk
);
5113 tcp_init_congestion_control(sk
);
5115 /* Prevent spurious tcp_cwnd_restart() on first data
5118 tp
->lsndtime
= tcp_time_stamp
;
5120 tcp_init_buffer_space(sk
);
5122 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5123 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5125 if (!tp
->rx_opt
.snd_wscale
)
5126 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5130 if (!sock_flag(sk
, SOCK_DEAD
)) {
5131 sk
->sk_state_change(sk
);
5132 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5135 if (sk
->sk_write_pending
||
5136 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5137 icsk
->icsk_ack
.pingpong
) {
5138 /* Save one ACK. Data will be ready after
5139 * several ticks, if write_pending is set.
5141 * It may be deleted, but with this feature tcpdumps
5142 * look so _wonderfully_ clever, that I was not able
5143 * to stand against the temptation 8) --ANK
5145 inet_csk_schedule_ack(sk
);
5146 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5147 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5148 tcp_incr_quickack(sk
);
5149 tcp_enter_quickack_mode(sk
);
5150 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5151 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5162 /* No ACK in the segment */
5166 * "If the RST bit is set
5168 * Otherwise (no ACK) drop the segment and return."
5171 goto discard_and_undo
;
5175 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5176 tcp_paws_check(&tp
->rx_opt
, 0))
5177 goto discard_and_undo
;
5180 /* We see SYN without ACK. It is attempt of
5181 * simultaneous connect with crossed SYNs.
5182 * Particularly, it can be connect to self.
5184 tcp_set_state(sk
, TCP_SYN_RECV
);
5186 if (tp
->rx_opt
.saw_tstamp
) {
5187 tp
->rx_opt
.tstamp_ok
= 1;
5188 tcp_store_ts_recent(tp
);
5189 tp
->tcp_header_len
=
5190 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5192 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5195 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5196 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5198 /* RFC1323: The window in SYN & SYN/ACK segments is
5201 tp
->snd_wnd
= ntohs(th
->window
);
5202 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5203 tp
->max_window
= tp
->snd_wnd
;
5205 TCP_ECN_rcv_syn(tp
, th
);
5208 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5209 tcp_initialize_rcv_mss(sk
);
5211 tcp_send_synack(sk
);
5213 /* Note, we could accept data and URG from this segment.
5214 * There are no obstacles to make this.
5216 * However, if we ignore data in ACKless segments sometimes,
5217 * we have no reasons to accept it sometimes.
5218 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5219 * is not flawless. So, discard packet for sanity.
5220 * Uncomment this return to process the data.
5227 /* "fifth, if neither of the SYN or RST bits is set then
5228 * drop the segment and return."
5232 tcp_clear_options(&tp
->rx_opt
);
5233 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5237 tcp_clear_options(&tp
->rx_opt
);
5238 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5243 * This function implements the receiving procedure of RFC 793 for
5244 * all states except ESTABLISHED and TIME_WAIT.
5245 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5246 * address independent.
5249 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5250 struct tcphdr
*th
, unsigned len
)
5252 struct tcp_sock
*tp
= tcp_sk(sk
);
5253 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5257 tp
->rx_opt
.saw_tstamp
= 0;
5259 switch (sk
->sk_state
) {
5271 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5274 /* Now we have several options: In theory there is
5275 * nothing else in the frame. KA9Q has an option to
5276 * send data with the syn, BSD accepts data with the
5277 * syn up to the [to be] advertised window and
5278 * Solaris 2.1 gives you a protocol error. For now
5279 * we just ignore it, that fits the spec precisely
5280 * and avoids incompatibilities. It would be nice in
5281 * future to drop through and process the data.
5283 * Now that TTCP is starting to be used we ought to
5285 * But, this leaves one open to an easy denial of
5286 * service attack, and SYN cookies can't defend
5287 * against this problem. So, we drop the data
5288 * in the interest of security over speed unless
5289 * it's still in use.
5297 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5301 /* Do step6 onward by hand. */
5302 tcp_urg(sk
, skb
, th
);
5304 tcp_data_snd_check(sk
);
5308 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5312 /* step 5: check the ACK field */
5314 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5316 switch (sk
->sk_state
) {
5319 tp
->copied_seq
= tp
->rcv_nxt
;
5321 tcp_set_state(sk
, TCP_ESTABLISHED
);
5322 sk
->sk_state_change(sk
);
5324 /* Note, that this wakeup is only for marginal
5325 * crossed SYN case. Passively open sockets
5326 * are not waked up, because sk->sk_sleep ==
5327 * NULL and sk->sk_socket == NULL.
5331 SOCK_WAKE_IO
, POLL_OUT
);
5333 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5334 tp
->snd_wnd
= ntohs(th
->window
) <<
5335 tp
->rx_opt
.snd_wscale
;
5336 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5337 TCP_SKB_CB(skb
)->seq
);
5339 /* tcp_ack considers this ACK as duplicate
5340 * and does not calculate rtt.
5341 * Fix it at least with timestamps.
5343 if (tp
->rx_opt
.saw_tstamp
&&
5344 tp
->rx_opt
.rcv_tsecr
&& !tp
->srtt
)
5345 tcp_ack_saw_tstamp(sk
, 0);
5347 if (tp
->rx_opt
.tstamp_ok
)
5348 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5350 /* Make sure socket is routed, for
5353 icsk
->icsk_af_ops
->rebuild_header(sk
);
5355 tcp_init_metrics(sk
);
5357 tcp_init_congestion_control(sk
);
5359 /* Prevent spurious tcp_cwnd_restart() on
5360 * first data packet.
5362 tp
->lsndtime
= tcp_time_stamp
;
5365 tcp_initialize_rcv_mss(sk
);
5366 tcp_init_buffer_space(sk
);
5367 tcp_fast_path_on(tp
);
5374 if (tp
->snd_una
== tp
->write_seq
) {
5375 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5376 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5377 dst_confirm(sk
->sk_dst_cache
);
5379 if (!sock_flag(sk
, SOCK_DEAD
))
5380 /* Wake up lingering close() */
5381 sk
->sk_state_change(sk
);
5385 if (tp
->linger2
< 0 ||
5386 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5387 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5389 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5393 tmo
= tcp_fin_time(sk
);
5394 if (tmo
> TCP_TIMEWAIT_LEN
) {
5395 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5396 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5397 /* Bad case. We could lose such FIN otherwise.
5398 * It is not a big problem, but it looks confusing
5399 * and not so rare event. We still can lose it now,
5400 * if it spins in bh_lock_sock(), but it is really
5403 inet_csk_reset_keepalive_timer(sk
, tmo
);
5405 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5413 if (tp
->snd_una
== tp
->write_seq
) {
5414 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5420 if (tp
->snd_una
== tp
->write_seq
) {
5421 tcp_update_metrics(sk
);
5430 /* step 6: check the URG bit */
5431 tcp_urg(sk
, skb
, th
);
5433 /* step 7: process the segment text */
5434 switch (sk
->sk_state
) {
5435 case TCP_CLOSE_WAIT
:
5438 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5442 /* RFC 793 says to queue data in these states,
5443 * RFC 1122 says we MUST send a reset.
5444 * BSD 4.4 also does reset.
5446 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5447 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5448 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5449 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5455 case TCP_ESTABLISHED
:
5456 tcp_data_queue(sk
, skb
);
5461 /* tcp_data could move socket to TIME-WAIT */
5462 if (sk
->sk_state
!= TCP_CLOSE
) {
5463 tcp_data_snd_check(sk
);
5464 tcp_ack_snd_check(sk
);
5474 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5475 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5476 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
5477 EXPORT_SYMBOL(tcp_parse_options
);
5478 #ifdef CONFIG_TCP_MD5SIG
5479 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
5481 EXPORT_SYMBOL(tcp_rcv_established
);
5482 EXPORT_SYMBOL(tcp_rcv_state_process
);
5483 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);