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/slab.h>
66 #include <linux/module.h>
67 #include <linux/sysctl.h>
68 #include <linux/kernel.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
74 #include <net/netdma.h>
76 int sysctl_tcp_timestamps __read_mostly
= 1;
77 int sysctl_tcp_window_scaling __read_mostly
= 1;
78 int sysctl_tcp_sack __read_mostly
= 1;
79 int sysctl_tcp_fack __read_mostly
= 1;
80 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
81 EXPORT_SYMBOL(sysctl_tcp_reordering
);
82 int sysctl_tcp_ecn __read_mostly
= 2;
83 EXPORT_SYMBOL(sysctl_tcp_ecn
);
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 int sysctl_tcp_stdurg __read_mostly
;
90 int sysctl_tcp_rfc1337 __read_mostly
;
91 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
92 int sysctl_tcp_frto __read_mostly
= 2;
93 int sysctl_tcp_frto_response __read_mostly
;
94 int sysctl_tcp_nometrics_save __read_mostly
;
96 int sysctl_tcp_thin_dupack __read_mostly
;
98 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
99 int sysctl_tcp_abc __read_mostly
;
101 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
102 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
103 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
104 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
105 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
106 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
107 #define FLAG_ECE 0x40 /* ECE in this ACK */
108 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
114 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
122 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
123 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
130 struct inet_connection_sock
*icsk
= inet_csk(sk
);
131 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
134 icsk
->icsk_ack
.last_seg_size
= 0;
136 /* skb->len may jitter because of SACKs, even if peer
137 * sends good full-sized frames.
139 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
140 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
141 icsk
->icsk_ack
.rcv_mss
= len
;
143 /* Otherwise, we make more careful check taking into account,
144 * that SACKs block is variable.
146 * "len" is invariant segment length, including TCP header.
148 len
+= skb
->data
- skb_transport_header(skb
);
149 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
150 /* If PSH is not set, packet should be
151 * full sized, provided peer TCP is not badly broken.
152 * This observation (if it is correct 8)) allows
153 * to handle super-low mtu links fairly.
155 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
156 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
157 /* Subtract also invariant (if peer is RFC compliant),
158 * tcp header plus fixed timestamp option length.
159 * Resulting "len" is MSS free of SACK jitter.
161 len
-= tcp_sk(sk
)->tcp_header_len
;
162 icsk
->icsk_ack
.last_seg_size
= len
;
164 icsk
->icsk_ack
.rcv_mss
= len
;
168 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
169 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
170 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
174 static void tcp_incr_quickack(struct sock
*sk
)
176 struct inet_connection_sock
*icsk
= inet_csk(sk
);
177 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
181 if (quickacks
> icsk
->icsk_ack
.quick
)
182 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
185 static void tcp_enter_quickack_mode(struct sock
*sk
)
187 struct inet_connection_sock
*icsk
= inet_csk(sk
);
188 tcp_incr_quickack(sk
);
189 icsk
->icsk_ack
.pingpong
= 0;
190 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
193 /* Send ACKs quickly, if "quick" count is not exhausted
194 * and the session is not interactive.
197 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
199 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
200 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
205 if (tp
->ecn_flags
& TCP_ECN_OK
)
206 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
211 if (tcp_hdr(skb
)->cwr
)
212 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
217 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
220 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
222 if (tp
->ecn_flags
& TCP_ECN_OK
) {
223 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
224 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
225 /* Funny extension: if ECT is not set on a segment,
226 * it is surely retransmit. It is not in ECN RFC,
227 * but Linux follows this rule. */
228 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
229 tcp_enter_quickack_mode((struct sock
*)tp
);
233 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
235 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
236 tp
->ecn_flags
&= ~TCP_ECN_OK
;
239 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
241 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
242 tp
->ecn_flags
&= ~TCP_ECN_OK
;
245 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
247 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
252 /* Buffer size and advertised window tuning.
254 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
257 static void tcp_fixup_sndbuf(struct sock
*sk
)
259 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
260 sizeof(struct sk_buff
);
262 if (sk
->sk_sndbuf
< 3 * sndmem
) {
263 sk
->sk_sndbuf
= 3 * sndmem
;
264 if (sk
->sk_sndbuf
> sysctl_tcp_wmem
[2])
265 sk
->sk_sndbuf
= sysctl_tcp_wmem
[2];
269 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
271 * All tcp_full_space() is split to two parts: "network" buffer, allocated
272 * forward and advertised in receiver window (tp->rcv_wnd) and
273 * "application buffer", required to isolate scheduling/application
274 * latencies from network.
275 * window_clamp is maximal advertised window. It can be less than
276 * tcp_full_space(), in this case tcp_full_space() - window_clamp
277 * is reserved for "application" buffer. The less window_clamp is
278 * the smoother our behaviour from viewpoint of network, but the lower
279 * throughput and the higher sensitivity of the connection to losses. 8)
281 * rcv_ssthresh is more strict window_clamp used at "slow start"
282 * phase to predict further behaviour of this connection.
283 * It is used for two goals:
284 * - to enforce header prediction at sender, even when application
285 * requires some significant "application buffer". It is check #1.
286 * - to prevent pruning of receive queue because of misprediction
287 * of receiver window. Check #2.
289 * The scheme does not work when sender sends good segments opening
290 * window and then starts to feed us spaghetti. But it should work
291 * in common situations. Otherwise, we have to rely on queue collapsing.
294 /* Slow part of check#2. */
295 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
297 struct tcp_sock
*tp
= tcp_sk(sk
);
299 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
300 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
302 while (tp
->rcv_ssthresh
<= window
) {
303 if (truesize
<= skb
->len
)
304 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
312 static void tcp_grow_window(struct sock
*sk
, struct sk_buff
*skb
)
314 struct tcp_sock
*tp
= tcp_sk(sk
);
317 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
318 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
319 !tcp_memory_pressure
) {
322 /* Check #2. Increase window, if skb with such overhead
323 * will fit to rcvbuf in future.
325 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
326 incr
= 2 * tp
->advmss
;
328 incr
= __tcp_grow_window(sk
, skb
);
331 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
333 inet_csk(sk
)->icsk_ack
.quick
|= 1;
338 /* 3. Tuning rcvbuf, when connection enters established state. */
340 static void tcp_fixup_rcvbuf(struct sock
*sk
)
342 struct tcp_sock
*tp
= tcp_sk(sk
);
343 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
345 /* Try to select rcvbuf so that 4 mss-sized segments
346 * will fit to window and corresponding skbs will fit to our rcvbuf.
347 * (was 3; 4 is minimum to allow fast retransmit to work.)
349 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
351 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
352 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
355 /* 4. Try to fixup all. It is made immediately after connection enters
358 static void tcp_init_buffer_space(struct sock
*sk
)
360 struct tcp_sock
*tp
= tcp_sk(sk
);
363 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
364 tcp_fixup_rcvbuf(sk
);
365 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
366 tcp_fixup_sndbuf(sk
);
368 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
370 maxwin
= tcp_full_space(sk
);
372 if (tp
->window_clamp
>= maxwin
) {
373 tp
->window_clamp
= maxwin
;
375 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
376 tp
->window_clamp
= max(maxwin
-
377 (maxwin
>> sysctl_tcp_app_win
),
381 /* Force reservation of one segment. */
382 if (sysctl_tcp_app_win
&&
383 tp
->window_clamp
> 2 * tp
->advmss
&&
384 tp
->window_clamp
+ tp
->advmss
> maxwin
)
385 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
387 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
388 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
391 /* 5. Recalculate window clamp after socket hit its memory bounds. */
392 static void tcp_clamp_window(struct sock
*sk
)
394 struct tcp_sock
*tp
= tcp_sk(sk
);
395 struct inet_connection_sock
*icsk
= inet_csk(sk
);
397 icsk
->icsk_ack
.quick
= 0;
399 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
400 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
401 !tcp_memory_pressure
&&
402 atomic_long_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
403 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
406 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
407 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
410 /* Initialize RCV_MSS value.
411 * RCV_MSS is an our guess about MSS used by the peer.
412 * We haven't any direct information about the MSS.
413 * It's better to underestimate the RCV_MSS rather than overestimate.
414 * Overestimations make us ACKing less frequently than needed.
415 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
417 void tcp_initialize_rcv_mss(struct sock
*sk
)
419 struct tcp_sock
*tp
= tcp_sk(sk
);
420 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
422 hint
= min(hint
, tp
->rcv_wnd
/ 2);
423 hint
= min(hint
, TCP_MSS_DEFAULT
);
424 hint
= max(hint
, TCP_MIN_MSS
);
426 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
428 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
430 /* Receiver "autotuning" code.
432 * The algorithm for RTT estimation w/o timestamps is based on
433 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
434 * <http://public.lanl.gov/radiant/pubs.html#DRS>
436 * More detail on this code can be found at
437 * <http://staff.psc.edu/jheffner/>,
438 * though this reference is out of date. A new paper
441 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
443 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
449 if (new_sample
!= 0) {
450 /* If we sample in larger samples in the non-timestamp
451 * case, we could grossly overestimate the RTT especially
452 * with chatty applications or bulk transfer apps which
453 * are stalled on filesystem I/O.
455 * Also, since we are only going for a minimum in the
456 * non-timestamp case, we do not smooth things out
457 * else with timestamps disabled convergence takes too
461 m
-= (new_sample
>> 3);
463 } else if (m
< new_sample
)
466 /* No previous measure. */
470 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
471 tp
->rcv_rtt_est
.rtt
= new_sample
;
474 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
476 if (tp
->rcv_rtt_est
.time
== 0)
478 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
480 tcp_rcv_rtt_update(tp
, jiffies
- tp
->rcv_rtt_est
.time
, 1);
483 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
484 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
487 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
488 const struct sk_buff
*skb
)
490 struct tcp_sock
*tp
= tcp_sk(sk
);
491 if (tp
->rx_opt
.rcv_tsecr
&&
492 (TCP_SKB_CB(skb
)->end_seq
-
493 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
494 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
498 * This function should be called every time data is copied to user space.
499 * It calculates the appropriate TCP receive buffer space.
501 void tcp_rcv_space_adjust(struct sock
*sk
)
503 struct tcp_sock
*tp
= tcp_sk(sk
);
507 if (tp
->rcvq_space
.time
== 0)
510 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
511 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
514 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
516 space
= max(tp
->rcvq_space
.space
, space
);
518 if (tp
->rcvq_space
.space
!= space
) {
521 tp
->rcvq_space
.space
= space
;
523 if (sysctl_tcp_moderate_rcvbuf
&&
524 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
525 int new_clamp
= space
;
527 /* Receive space grows, normalize in order to
528 * take into account packet headers and sk_buff
529 * structure overhead.
534 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
535 16 + sizeof(struct sk_buff
));
536 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
539 space
= min(space
, sysctl_tcp_rmem
[2]);
540 if (space
> sk
->sk_rcvbuf
) {
541 sk
->sk_rcvbuf
= space
;
543 /* Make the window clamp follow along. */
544 tp
->window_clamp
= new_clamp
;
550 tp
->rcvq_space
.seq
= tp
->copied_seq
;
551 tp
->rcvq_space
.time
= tcp_time_stamp
;
554 /* There is something which you must keep in mind when you analyze the
555 * behavior of the tp->ato delayed ack timeout interval. When a
556 * connection starts up, we want to ack as quickly as possible. The
557 * problem is that "good" TCP's do slow start at the beginning of data
558 * transmission. The means that until we send the first few ACK's the
559 * sender will sit on his end and only queue most of his data, because
560 * he can only send snd_cwnd unacked packets at any given time. For
561 * each ACK we send, he increments snd_cwnd and transmits more of his
564 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
566 struct tcp_sock
*tp
= tcp_sk(sk
);
567 struct inet_connection_sock
*icsk
= inet_csk(sk
);
570 inet_csk_schedule_ack(sk
);
572 tcp_measure_rcv_mss(sk
, skb
);
574 tcp_rcv_rtt_measure(tp
);
576 now
= tcp_time_stamp
;
578 if (!icsk
->icsk_ack
.ato
) {
579 /* The _first_ data packet received, initialize
580 * delayed ACK engine.
582 tcp_incr_quickack(sk
);
583 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
585 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
587 if (m
<= TCP_ATO_MIN
/ 2) {
588 /* The fastest case is the first. */
589 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
590 } else if (m
< icsk
->icsk_ack
.ato
) {
591 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
592 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
593 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
594 } else if (m
> icsk
->icsk_rto
) {
595 /* Too long gap. Apparently sender failed to
596 * restart window, so that we send ACKs quickly.
598 tcp_incr_quickack(sk
);
602 icsk
->icsk_ack
.lrcvtime
= now
;
604 TCP_ECN_check_ce(tp
, skb
);
607 tcp_grow_window(sk
, skb
);
610 /* Called to compute a smoothed rtt estimate. The data fed to this
611 * routine either comes from timestamps, or from segments that were
612 * known _not_ to have been retransmitted [see Karn/Partridge
613 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
614 * piece by Van Jacobson.
615 * NOTE: the next three routines used to be one big routine.
616 * To save cycles in the RFC 1323 implementation it was better to break
617 * it up into three procedures. -- erics
619 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
621 struct tcp_sock
*tp
= tcp_sk(sk
);
622 long m
= mrtt
; /* RTT */
624 /* The following amusing code comes from Jacobson's
625 * article in SIGCOMM '88. Note that rtt and mdev
626 * are scaled versions of rtt and mean deviation.
627 * This is designed to be as fast as possible
628 * m stands for "measurement".
630 * On a 1990 paper the rto value is changed to:
631 * RTO = rtt + 4 * mdev
633 * Funny. This algorithm seems to be very broken.
634 * These formulae increase RTO, when it should be decreased, increase
635 * too slowly, when it should be increased quickly, decrease too quickly
636 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
637 * does not matter how to _calculate_ it. Seems, it was trap
638 * that VJ failed to avoid. 8)
643 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
644 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
646 m
= -m
; /* m is now abs(error) */
647 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
648 /* This is similar to one of Eifel findings.
649 * Eifel blocks mdev updates when rtt decreases.
650 * This solution is a bit different: we use finer gain
651 * for mdev in this case (alpha*beta).
652 * Like Eifel it also prevents growth of rto,
653 * but also it limits too fast rto decreases,
654 * happening in pure Eifel.
659 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
661 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
662 if (tp
->mdev
> tp
->mdev_max
) {
663 tp
->mdev_max
= tp
->mdev
;
664 if (tp
->mdev_max
> tp
->rttvar
)
665 tp
->rttvar
= tp
->mdev_max
;
667 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
668 if (tp
->mdev_max
< tp
->rttvar
)
669 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
670 tp
->rtt_seq
= tp
->snd_nxt
;
671 tp
->mdev_max
= tcp_rto_min(sk
);
674 /* no previous measure. */
675 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
676 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
677 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
678 tp
->rtt_seq
= tp
->snd_nxt
;
682 /* Calculate rto without backoff. This is the second half of Van Jacobson's
683 * routine referred to above.
685 static inline void tcp_set_rto(struct sock
*sk
)
687 const struct tcp_sock
*tp
= tcp_sk(sk
);
688 /* Old crap is replaced with new one. 8)
691 * 1. If rtt variance happened to be less 50msec, it is hallucination.
692 * It cannot be less due to utterly erratic ACK generation made
693 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
694 * to do with delayed acks, because at cwnd>2 true delack timeout
695 * is invisible. Actually, Linux-2.4 also generates erratic
696 * ACKs in some circumstances.
698 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
700 /* 2. Fixups made earlier cannot be right.
701 * If we do not estimate RTO correctly without them,
702 * all the algo is pure shit and should be replaced
703 * 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.
712 /* Save metrics learned by this TCP session.
713 This function is called only, when TCP finishes successfully
714 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
716 void tcp_update_metrics(struct sock
*sk
)
718 struct tcp_sock
*tp
= tcp_sk(sk
);
719 struct dst_entry
*dst
= __sk_dst_get(sk
);
721 if (sysctl_tcp_nometrics_save
)
726 if (dst
&& (dst
->flags
& DST_HOST
)) {
727 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
731 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
732 /* This session failed to estimate rtt. Why?
733 * Probably, no packets returned in time.
736 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
737 dst_metric_set(dst
, RTAX_RTT
, 0);
741 rtt
= dst_metric_rtt(dst
, RTAX_RTT
);
744 /* If newly calculated rtt larger than stored one,
745 * store new one. Otherwise, use EWMA. Remember,
746 * rtt overestimation is always better than underestimation.
748 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
750 set_dst_metric_rtt(dst
, RTAX_RTT
, tp
->srtt
);
752 set_dst_metric_rtt(dst
, RTAX_RTT
, rtt
- (m
>> 3));
755 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
760 /* Scale deviation to rttvar fixed point */
765 var
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
769 var
-= (var
- m
) >> 2;
771 set_dst_metric_rtt(dst
, RTAX_RTTVAR
, var
);
774 if (tcp_in_initial_slowstart(tp
)) {
775 /* Slow start still did not finish. */
776 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
777 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
778 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
779 dst_metric_set(dst
, RTAX_SSTHRESH
, tp
->snd_cwnd
>> 1);
780 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
781 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
782 dst_metric_set(dst
, RTAX_CWND
, tp
->snd_cwnd
);
783 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
784 icsk
->icsk_ca_state
== TCP_CA_Open
) {
785 /* Cong. avoidance phase, cwnd is reliable. */
786 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
787 dst_metric_set(dst
, RTAX_SSTHRESH
,
788 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
));
789 if (!dst_metric_locked(dst
, RTAX_CWND
))
790 dst_metric_set(dst
, RTAX_CWND
,
791 (dst_metric(dst
, RTAX_CWND
) +
794 /* Else slow start did not finish, cwnd is non-sense,
795 ssthresh may be also invalid.
797 if (!dst_metric_locked(dst
, RTAX_CWND
))
798 dst_metric_set(dst
, RTAX_CWND
,
799 (dst_metric(dst
, RTAX_CWND
) +
800 tp
->snd_ssthresh
) >> 1);
801 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
802 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
803 tp
->snd_ssthresh
> dst_metric(dst
, RTAX_SSTHRESH
))
804 dst_metric_set(dst
, RTAX_SSTHRESH
, tp
->snd_ssthresh
);
807 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
808 if (dst_metric(dst
, RTAX_REORDERING
) < tp
->reordering
&&
809 tp
->reordering
!= sysctl_tcp_reordering
)
810 dst_metric_set(dst
, RTAX_REORDERING
, tp
->reordering
);
815 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
817 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
820 cwnd
= rfc3390_bytes_to_packets(tp
->mss_cache
);
821 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
824 /* Set slow start threshold and cwnd not falling to slow start */
825 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
827 struct tcp_sock
*tp
= tcp_sk(sk
);
828 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
830 tp
->prior_ssthresh
= 0;
832 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
835 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
836 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
837 tcp_packets_in_flight(tp
) + 1U);
838 tp
->snd_cwnd_cnt
= 0;
839 tp
->high_seq
= tp
->snd_nxt
;
840 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
841 TCP_ECN_queue_cwr(tp
);
843 tcp_set_ca_state(sk
, TCP_CA_CWR
);
848 * Packet counting of FACK is based on in-order assumptions, therefore TCP
849 * disables it when reordering is detected
851 static void tcp_disable_fack(struct tcp_sock
*tp
)
853 /* RFC3517 uses different metric in lost marker => reset on change */
855 tp
->lost_skb_hint
= NULL
;
856 tp
->rx_opt
.sack_ok
&= ~2;
859 /* Take a notice that peer is sending D-SACKs */
860 static void tcp_dsack_seen(struct tcp_sock
*tp
)
862 tp
->rx_opt
.sack_ok
|= 4;
865 /* Initialize metrics on socket. */
867 static void tcp_init_metrics(struct sock
*sk
)
869 struct tcp_sock
*tp
= tcp_sk(sk
);
870 struct dst_entry
*dst
= __sk_dst_get(sk
);
877 if (dst_metric_locked(dst
, RTAX_CWND
))
878 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
879 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
880 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
881 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
882 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
884 if (dst_metric(dst
, RTAX_REORDERING
) &&
885 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
886 tcp_disable_fack(tp
);
887 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
890 if (dst_metric(dst
, RTAX_RTT
) == 0)
893 if (!tp
->srtt
&& dst_metric_rtt(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
896 /* Initial rtt is determined from SYN,SYN-ACK.
897 * The segment is small and rtt may appear much
898 * less than real one. Use per-dst memory
899 * to make it more realistic.
901 * A bit of theory. RTT is time passed after "normal" sized packet
902 * is sent until it is ACKed. In normal circumstances sending small
903 * packets force peer to delay ACKs and calculation is correct too.
904 * The algorithm is adaptive and, provided we follow specs, it
905 * NEVER underestimate RTT. BUT! If peer tries to make some clever
906 * tricks sort of "quick acks" for time long enough to decrease RTT
907 * to low value, and then abruptly stops to do it and starts to delay
908 * ACKs, wait for troubles.
910 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
911 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
912 tp
->rtt_seq
= tp
->snd_nxt
;
914 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
915 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
916 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
919 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
923 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
924 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
928 /* Play conservative. If timestamps are not
929 * supported, TCP will fail to recalculate correct
930 * rtt, if initial rto is too small. FORGET ALL AND RESET!
932 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
934 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
935 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
940 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
943 struct tcp_sock
*tp
= tcp_sk(sk
);
944 if (metric
> tp
->reordering
) {
947 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
949 /* This exciting event is worth to be remembered. 8) */
951 mib_idx
= LINUX_MIB_TCPTSREORDER
;
952 else if (tcp_is_reno(tp
))
953 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
954 else if (tcp_is_fack(tp
))
955 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
957 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
959 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
960 #if FASTRETRANS_DEBUG > 1
961 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
962 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
966 tp
->undo_marker
? tp
->undo_retrans
: 0);
968 tcp_disable_fack(tp
);
972 /* This must be called before lost_out is incremented */
973 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
975 if ((tp
->retransmit_skb_hint
== NULL
) ||
976 before(TCP_SKB_CB(skb
)->seq
,
977 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
978 tp
->retransmit_skb_hint
= skb
;
981 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
982 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
985 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
987 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
988 tcp_verify_retransmit_hint(tp
, skb
);
990 tp
->lost_out
+= tcp_skb_pcount(skb
);
991 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
995 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
998 tcp_verify_retransmit_hint(tp
, skb
);
1000 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1001 tp
->lost_out
+= tcp_skb_pcount(skb
);
1002 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1006 /* This procedure tags the retransmission queue when SACKs arrive.
1008 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1009 * Packets in queue with these bits set are counted in variables
1010 * sacked_out, retrans_out and lost_out, correspondingly.
1012 * Valid combinations are:
1013 * Tag InFlight Description
1014 * 0 1 - orig segment is in flight.
1015 * S 0 - nothing flies, orig reached receiver.
1016 * L 0 - nothing flies, orig lost by net.
1017 * R 2 - both orig and retransmit are in flight.
1018 * L|R 1 - orig is lost, retransmit is in flight.
1019 * S|R 1 - orig reached receiver, retrans is still in flight.
1020 * (L|S|R is logically valid, it could occur when L|R is sacked,
1021 * but it is equivalent to plain S and code short-curcuits it to S.
1022 * L|S is logically invalid, it would mean -1 packet in flight 8))
1024 * These 6 states form finite state machine, controlled by the following events:
1025 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1026 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1027 * 3. Loss detection event of one of three flavors:
1028 * A. Scoreboard estimator decided the packet is lost.
1029 * A'. Reno "three dupacks" marks head of queue lost.
1030 * A''. Its FACK modfication, head until snd.fack is lost.
1031 * B. SACK arrives sacking data transmitted after never retransmitted
1032 * hole was sent out.
1033 * C. SACK arrives sacking SND.NXT at the moment, when the
1034 * segment was retransmitted.
1035 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1037 * It is pleasant to note, that state diagram turns out to be commutative,
1038 * so that we are allowed not to be bothered by order of our actions,
1039 * when multiple events arrive simultaneously. (see the function below).
1041 * Reordering detection.
1042 * --------------------
1043 * Reordering metric is maximal distance, which a packet can be displaced
1044 * in packet stream. With SACKs we can estimate it:
1046 * 1. SACK fills old hole and the corresponding segment was not
1047 * ever retransmitted -> reordering. Alas, we cannot use it
1048 * when segment was retransmitted.
1049 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1050 * for retransmitted and already SACKed segment -> reordering..
1051 * Both of these heuristics are not used in Loss state, when we cannot
1052 * account for retransmits accurately.
1054 * SACK block validation.
1055 * ----------------------
1057 * SACK block range validation checks that the received SACK block fits to
1058 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1059 * Note that SND.UNA is not included to the range though being valid because
1060 * it means that the receiver is rather inconsistent with itself reporting
1061 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1062 * perfectly valid, however, in light of RFC2018 which explicitly states
1063 * that "SACK block MUST reflect the newest segment. Even if the newest
1064 * segment is going to be discarded ...", not that it looks very clever
1065 * in case of head skb. Due to potentional receiver driven attacks, we
1066 * choose to avoid immediate execution of a walk in write queue due to
1067 * reneging and defer head skb's loss recovery to standard loss recovery
1068 * procedure that will eventually trigger (nothing forbids us doing this).
1070 * Implements also blockage to start_seq wrap-around. Problem lies in the
1071 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1072 * there's no guarantee that it will be before snd_nxt (n). The problem
1073 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1076 * <- outs wnd -> <- wrapzone ->
1077 * u e n u_w e_w s n_w
1079 * |<------------+------+----- TCP seqno space --------------+---------->|
1080 * ...-- <2^31 ->| |<--------...
1081 * ...---- >2^31 ------>| |<--------...
1083 * Current code wouldn't be vulnerable but it's better still to discard such
1084 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1085 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1086 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1087 * equal to the ideal case (infinite seqno space without wrap caused issues).
1089 * With D-SACK the lower bound is extended to cover sequence space below
1090 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1091 * again, D-SACK block must not to go across snd_una (for the same reason as
1092 * for the normal SACK blocks, explained above). But there all simplicity
1093 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1094 * fully below undo_marker they do not affect behavior in anyway and can
1095 * therefore be safely ignored. In rare cases (which are more or less
1096 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1097 * fragmentation and packet reordering past skb's retransmission. To consider
1098 * them correctly, the acceptable range must be extended even more though
1099 * the exact amount is rather hard to quantify. However, tp->max_window can
1100 * be used as an exaggerated estimate.
1102 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1103 u32 start_seq
, u32 end_seq
)
1105 /* Too far in future, or reversed (interpretation is ambiguous) */
1106 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1109 /* Nasty start_seq wrap-around check (see comments above) */
1110 if (!before(start_seq
, tp
->snd_nxt
))
1113 /* In outstanding window? ...This is valid exit for D-SACKs too.
1114 * start_seq == snd_una is non-sensical (see comments above)
1116 if (after(start_seq
, tp
->snd_una
))
1119 if (!is_dsack
|| !tp
->undo_marker
)
1122 /* ...Then it's D-SACK, and must reside below snd_una completely */
1123 if (!after(end_seq
, tp
->snd_una
))
1126 if (!before(start_seq
, tp
->undo_marker
))
1130 if (!after(end_seq
, tp
->undo_marker
))
1133 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1134 * start_seq < undo_marker and end_seq >= undo_marker.
1136 return !before(start_seq
, end_seq
- tp
->max_window
);
1139 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1140 * Event "C". Later note: FACK people cheated me again 8), we have to account
1141 * for reordering! Ugly, but should help.
1143 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1144 * less than what is now known to be received by the other end (derived from
1145 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1146 * retransmitted skbs to avoid some costly processing per ACKs.
1148 static void tcp_mark_lost_retrans(struct sock
*sk
)
1150 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1151 struct tcp_sock
*tp
= tcp_sk(sk
);
1152 struct sk_buff
*skb
;
1154 u32 new_low_seq
= tp
->snd_nxt
;
1155 u32 received_upto
= tcp_highest_sack_seq(tp
);
1157 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1158 !after(received_upto
, tp
->lost_retrans_low
) ||
1159 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1162 tcp_for_write_queue(skb
, sk
) {
1163 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1165 if (skb
== tcp_send_head(sk
))
1167 if (cnt
== tp
->retrans_out
)
1169 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1172 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1175 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1176 * constraint here (see above) but figuring out that at
1177 * least tp->reordering SACK blocks reside between ack_seq
1178 * and received_upto is not easy task to do cheaply with
1179 * the available datastructures.
1181 * Whether FACK should check here for tp->reordering segs
1182 * in-between one could argue for either way (it would be
1183 * rather simple to implement as we could count fack_count
1184 * during the walk and do tp->fackets_out - fack_count).
1186 if (after(received_upto
, ack_seq
)) {
1187 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1188 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1190 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1191 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1193 if (before(ack_seq
, new_low_seq
))
1194 new_low_seq
= ack_seq
;
1195 cnt
+= tcp_skb_pcount(skb
);
1199 if (tp
->retrans_out
)
1200 tp
->lost_retrans_low
= new_low_seq
;
1203 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1204 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1207 struct tcp_sock
*tp
= tcp_sk(sk
);
1208 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1209 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1212 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1215 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1216 } else if (num_sacks
> 1) {
1217 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1218 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1220 if (!after(end_seq_0
, end_seq_1
) &&
1221 !before(start_seq_0
, start_seq_1
)) {
1224 NET_INC_STATS_BH(sock_net(sk
),
1225 LINUX_MIB_TCPDSACKOFORECV
);
1229 /* D-SACK for already forgotten data... Do dumb counting. */
1231 !after(end_seq_0
, prior_snd_una
) &&
1232 after(end_seq_0
, tp
->undo_marker
))
1238 struct tcp_sacktag_state
{
1244 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1245 * the incoming SACK may not exactly match but we can find smaller MSS
1246 * aligned portion of it that matches. Therefore we might need to fragment
1247 * which may fail and creates some hassle (caller must handle error case
1250 * FIXME: this could be merged to shift decision code
1252 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1253 u32 start_seq
, u32 end_seq
)
1256 unsigned int pkt_len
;
1259 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1260 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1262 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1263 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1264 mss
= tcp_skb_mss(skb
);
1265 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1268 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1272 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1277 /* Round if necessary so that SACKs cover only full MSSes
1278 * and/or the remaining small portion (if present)
1280 if (pkt_len
> mss
) {
1281 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1282 if (!in_sack
&& new_len
< pkt_len
) {
1284 if (new_len
> skb
->len
)
1289 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1297 static u8
tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1298 struct tcp_sacktag_state
*state
,
1299 int dup_sack
, int pcount
)
1301 struct tcp_sock
*tp
= tcp_sk(sk
);
1302 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1303 int fack_count
= state
->fack_count
;
1305 /* Account D-SACK for retransmitted packet. */
1306 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1307 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1309 if (sacked
& TCPCB_SACKED_ACKED
)
1310 state
->reord
= min(fack_count
, state
->reord
);
1313 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1314 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1317 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1318 if (sacked
& TCPCB_SACKED_RETRANS
) {
1319 /* If the segment is not tagged as lost,
1320 * we do not clear RETRANS, believing
1321 * that retransmission is still in flight.
1323 if (sacked
& TCPCB_LOST
) {
1324 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1325 tp
->lost_out
-= pcount
;
1326 tp
->retrans_out
-= pcount
;
1329 if (!(sacked
& TCPCB_RETRANS
)) {
1330 /* New sack for not retransmitted frame,
1331 * which was in hole. It is reordering.
1333 if (before(TCP_SKB_CB(skb
)->seq
,
1334 tcp_highest_sack_seq(tp
)))
1335 state
->reord
= min(fack_count
,
1338 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1339 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1340 state
->flag
|= FLAG_ONLY_ORIG_SACKED
;
1343 if (sacked
& TCPCB_LOST
) {
1344 sacked
&= ~TCPCB_LOST
;
1345 tp
->lost_out
-= pcount
;
1349 sacked
|= TCPCB_SACKED_ACKED
;
1350 state
->flag
|= FLAG_DATA_SACKED
;
1351 tp
->sacked_out
+= pcount
;
1353 fack_count
+= pcount
;
1355 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1356 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1357 before(TCP_SKB_CB(skb
)->seq
,
1358 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1359 tp
->lost_cnt_hint
+= pcount
;
1361 if (fack_count
> tp
->fackets_out
)
1362 tp
->fackets_out
= fack_count
;
1365 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1366 * frames and clear it. undo_retrans is decreased above, L|R frames
1367 * are accounted above as well.
1369 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1370 sacked
&= ~TCPCB_SACKED_RETRANS
;
1371 tp
->retrans_out
-= pcount
;
1377 static int tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1378 struct tcp_sacktag_state
*state
,
1379 unsigned int pcount
, int shifted
, int mss
,
1382 struct tcp_sock
*tp
= tcp_sk(sk
);
1383 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1387 /* Tweak before seqno plays */
1388 if (!tcp_is_fack(tp
) && tcp_is_sack(tp
) && tp
->lost_skb_hint
&&
1389 !before(TCP_SKB_CB(tp
->lost_skb_hint
)->seq
, TCP_SKB_CB(skb
)->seq
))
1390 tp
->lost_cnt_hint
+= pcount
;
1392 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1393 TCP_SKB_CB(skb
)->seq
+= shifted
;
1395 skb_shinfo(prev
)->gso_segs
+= pcount
;
1396 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1397 skb_shinfo(skb
)->gso_segs
-= pcount
;
1399 /* When we're adding to gso_segs == 1, gso_size will be zero,
1400 * in theory this shouldn't be necessary but as long as DSACK
1401 * code can come after this skb later on it's better to keep
1402 * setting gso_size to something.
1404 if (!skb_shinfo(prev
)->gso_size
) {
1405 skb_shinfo(prev
)->gso_size
= mss
;
1406 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1409 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1410 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1411 skb_shinfo(skb
)->gso_size
= 0;
1412 skb_shinfo(skb
)->gso_type
= 0;
1415 /* We discard results */
1416 tcp_sacktag_one(skb
, sk
, state
, dup_sack
, pcount
);
1418 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1419 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1422 BUG_ON(!tcp_skb_pcount(skb
));
1423 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1427 /* Whole SKB was eaten :-) */
1429 if (skb
== tp
->retransmit_skb_hint
)
1430 tp
->retransmit_skb_hint
= prev
;
1431 if (skb
== tp
->scoreboard_skb_hint
)
1432 tp
->scoreboard_skb_hint
= prev
;
1433 if (skb
== tp
->lost_skb_hint
) {
1434 tp
->lost_skb_hint
= prev
;
1435 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1438 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(prev
)->flags
;
1439 if (skb
== tcp_highest_sack(sk
))
1440 tcp_advance_highest_sack(sk
, skb
);
1442 tcp_unlink_write_queue(skb
, sk
);
1443 sk_wmem_free_skb(sk
, skb
);
1445 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1450 /* I wish gso_size would have a bit more sane initialization than
1451 * something-or-zero which complicates things
1453 static int tcp_skb_seglen(struct sk_buff
*skb
)
1455 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1458 /* Shifting pages past head area doesn't work */
1459 static int skb_can_shift(struct sk_buff
*skb
)
1461 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1464 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1467 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1468 struct tcp_sacktag_state
*state
,
1469 u32 start_seq
, u32 end_seq
,
1472 struct tcp_sock
*tp
= tcp_sk(sk
);
1473 struct sk_buff
*prev
;
1479 if (!sk_can_gso(sk
))
1482 /* Normally R but no L won't result in plain S */
1484 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1486 if (!skb_can_shift(skb
))
1488 /* This frame is about to be dropped (was ACKed). */
1489 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1492 /* Can only happen with delayed DSACK + discard craziness */
1493 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1495 prev
= tcp_write_queue_prev(sk
, skb
);
1497 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1500 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1501 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1505 pcount
= tcp_skb_pcount(skb
);
1506 mss
= tcp_skb_seglen(skb
);
1508 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1509 * drop this restriction as unnecessary
1511 if (mss
!= tcp_skb_seglen(prev
))
1514 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1516 /* CHECKME: This is non-MSS split case only?, this will
1517 * cause skipped skbs due to advancing loop btw, original
1518 * has that feature too
1520 if (tcp_skb_pcount(skb
) <= 1)
1523 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1525 /* TODO: head merge to next could be attempted here
1526 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1527 * though it might not be worth of the additional hassle
1529 * ...we can probably just fallback to what was done
1530 * previously. We could try merging non-SACKed ones
1531 * as well but it probably isn't going to buy off
1532 * because later SACKs might again split them, and
1533 * it would make skb timestamp tracking considerably
1539 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1541 BUG_ON(len
> skb
->len
);
1543 /* MSS boundaries should be honoured or else pcount will
1544 * severely break even though it makes things bit trickier.
1545 * Optimize common case to avoid most of the divides
1547 mss
= tcp_skb_mss(skb
);
1549 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1550 * drop this restriction as unnecessary
1552 if (mss
!= tcp_skb_seglen(prev
))
1557 } else if (len
< mss
) {
1565 if (!skb_shift(prev
, skb
, len
))
1567 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1570 /* Hole filled allows collapsing with the next as well, this is very
1571 * useful when hole on every nth skb pattern happens
1573 if (prev
== tcp_write_queue_tail(sk
))
1575 skb
= tcp_write_queue_next(sk
, prev
);
1577 if (!skb_can_shift(skb
) ||
1578 (skb
== tcp_send_head(sk
)) ||
1579 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1580 (mss
!= tcp_skb_seglen(skb
)))
1584 if (skb_shift(prev
, skb
, len
)) {
1585 pcount
+= tcp_skb_pcount(skb
);
1586 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1590 state
->fack_count
+= pcount
;
1597 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1601 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1602 struct tcp_sack_block
*next_dup
,
1603 struct tcp_sacktag_state
*state
,
1604 u32 start_seq
, u32 end_seq
,
1607 struct tcp_sock
*tp
= tcp_sk(sk
);
1608 struct sk_buff
*tmp
;
1610 tcp_for_write_queue_from(skb
, sk
) {
1612 int dup_sack
= dup_sack_in
;
1614 if (skb
== tcp_send_head(sk
))
1617 /* queue is in-order => we can short-circuit the walk early */
1618 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1621 if ((next_dup
!= NULL
) &&
1622 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1623 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1624 next_dup
->start_seq
,
1630 /* skb reference here is a bit tricky to get right, since
1631 * shifting can eat and free both this skb and the next,
1632 * so not even _safe variant of the loop is enough.
1635 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1636 start_seq
, end_seq
, dup_sack
);
1645 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1651 if (unlikely(in_sack
< 0))
1655 TCP_SKB_CB(skb
)->sacked
= tcp_sacktag_one(skb
, sk
,
1658 tcp_skb_pcount(skb
));
1660 if (!before(TCP_SKB_CB(skb
)->seq
,
1661 tcp_highest_sack_seq(tp
)))
1662 tcp_advance_highest_sack(sk
, skb
);
1665 state
->fack_count
+= tcp_skb_pcount(skb
);
1670 /* Avoid all extra work that is being done by sacktag while walking in
1673 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1674 struct tcp_sacktag_state
*state
,
1677 tcp_for_write_queue_from(skb
, sk
) {
1678 if (skb
== tcp_send_head(sk
))
1681 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1684 state
->fack_count
+= tcp_skb_pcount(skb
);
1689 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1691 struct tcp_sack_block
*next_dup
,
1692 struct tcp_sacktag_state
*state
,
1695 if (next_dup
== NULL
)
1698 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1699 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1700 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1701 next_dup
->start_seq
, next_dup
->end_seq
,
1708 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1710 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1714 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1717 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1718 struct tcp_sock
*tp
= tcp_sk(sk
);
1719 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1720 TCP_SKB_CB(ack_skb
)->sacked
);
1721 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1722 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1723 struct tcp_sack_block
*cache
;
1724 struct tcp_sacktag_state state
;
1725 struct sk_buff
*skb
;
1726 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1728 int found_dup_sack
= 0;
1730 int first_sack_index
;
1733 state
.reord
= tp
->packets_out
;
1735 if (!tp
->sacked_out
) {
1736 if (WARN_ON(tp
->fackets_out
))
1737 tp
->fackets_out
= 0;
1738 tcp_highest_sack_reset(sk
);
1741 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1742 num_sacks
, prior_snd_una
);
1744 state
.flag
|= FLAG_DSACKING_ACK
;
1746 /* Eliminate too old ACKs, but take into
1747 * account more or less fresh ones, they can
1748 * contain valid SACK info.
1750 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1753 if (!tp
->packets_out
)
1757 first_sack_index
= 0;
1758 for (i
= 0; i
< num_sacks
; i
++) {
1759 int dup_sack
= !i
&& found_dup_sack
;
1761 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1762 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1764 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1765 sp
[used_sacks
].start_seq
,
1766 sp
[used_sacks
].end_seq
)) {
1770 if (!tp
->undo_marker
)
1771 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1773 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1775 /* Don't count olds caused by ACK reordering */
1776 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1777 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1779 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1782 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1784 first_sack_index
= -1;
1788 /* Ignore very old stuff early */
1789 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1795 /* order SACK blocks to allow in order walk of the retrans queue */
1796 for (i
= used_sacks
- 1; i
> 0; i
--) {
1797 for (j
= 0; j
< i
; j
++) {
1798 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1799 swap(sp
[j
], sp
[j
+ 1]);
1801 /* Track where the first SACK block goes to */
1802 if (j
== first_sack_index
)
1803 first_sack_index
= j
+ 1;
1808 skb
= tcp_write_queue_head(sk
);
1809 state
.fack_count
= 0;
1812 if (!tp
->sacked_out
) {
1813 /* It's already past, so skip checking against it */
1814 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1816 cache
= tp
->recv_sack_cache
;
1817 /* Skip empty blocks in at head of the cache */
1818 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1823 while (i
< used_sacks
) {
1824 u32 start_seq
= sp
[i
].start_seq
;
1825 u32 end_seq
= sp
[i
].end_seq
;
1826 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1827 struct tcp_sack_block
*next_dup
= NULL
;
1829 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1830 next_dup
= &sp
[i
+ 1];
1832 /* Event "B" in the comment above. */
1833 if (after(end_seq
, tp
->high_seq
))
1834 state
.flag
|= FLAG_DATA_LOST
;
1836 /* Skip too early cached blocks */
1837 while (tcp_sack_cache_ok(tp
, cache
) &&
1838 !before(start_seq
, cache
->end_seq
))
1841 /* Can skip some work by looking recv_sack_cache? */
1842 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1843 after(end_seq
, cache
->start_seq
)) {
1846 if (before(start_seq
, cache
->start_seq
)) {
1847 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1849 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1856 /* Rest of the block already fully processed? */
1857 if (!after(end_seq
, cache
->end_seq
))
1860 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1864 /* ...tail remains todo... */
1865 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1866 /* ...but better entrypoint exists! */
1867 skb
= tcp_highest_sack(sk
);
1870 state
.fack_count
= tp
->fackets_out
;
1875 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1876 /* Check overlap against next cached too (past this one already) */
1881 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1882 skb
= tcp_highest_sack(sk
);
1885 state
.fack_count
= tp
->fackets_out
;
1887 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1890 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1891 start_seq
, end_seq
, dup_sack
);
1894 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1895 * due to in-order walk
1897 if (after(end_seq
, tp
->frto_highmark
))
1898 state
.flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1903 /* Clear the head of the cache sack blocks so we can skip it next time */
1904 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1905 tp
->recv_sack_cache
[i
].start_seq
= 0;
1906 tp
->recv_sack_cache
[i
].end_seq
= 0;
1908 for (j
= 0; j
< used_sacks
; j
++)
1909 tp
->recv_sack_cache
[i
++] = sp
[j
];
1911 tcp_mark_lost_retrans(sk
);
1913 tcp_verify_left_out(tp
);
1915 if ((state
.reord
< tp
->fackets_out
) &&
1916 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1917 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1918 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1922 #if FASTRETRANS_DEBUG > 0
1923 WARN_ON((int)tp
->sacked_out
< 0);
1924 WARN_ON((int)tp
->lost_out
< 0);
1925 WARN_ON((int)tp
->retrans_out
< 0);
1926 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1931 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1932 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1934 static int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1938 holes
= max(tp
->lost_out
, 1U);
1939 holes
= min(holes
, tp
->packets_out
);
1941 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1942 tp
->sacked_out
= tp
->packets_out
- holes
;
1948 /* If we receive more dupacks than we expected counting segments
1949 * in assumption of absent reordering, interpret this as reordering.
1950 * The only another reason could be bug in receiver TCP.
1952 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1954 struct tcp_sock
*tp
= tcp_sk(sk
);
1955 if (tcp_limit_reno_sacked(tp
))
1956 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1959 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1961 static void tcp_add_reno_sack(struct sock
*sk
)
1963 struct tcp_sock
*tp
= tcp_sk(sk
);
1965 tcp_check_reno_reordering(sk
, 0);
1966 tcp_verify_left_out(tp
);
1969 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1971 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1973 struct tcp_sock
*tp
= tcp_sk(sk
);
1976 /* One ACK acked hole. The rest eat duplicate ACKs. */
1977 if (acked
- 1 >= tp
->sacked_out
)
1980 tp
->sacked_out
-= acked
- 1;
1982 tcp_check_reno_reordering(sk
, acked
);
1983 tcp_verify_left_out(tp
);
1986 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1991 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
1993 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
1996 /* F-RTO can only be used if TCP has never retransmitted anything other than
1997 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1999 int tcp_use_frto(struct sock
*sk
)
2001 const struct tcp_sock
*tp
= tcp_sk(sk
);
2002 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2003 struct sk_buff
*skb
;
2005 if (!sysctl_tcp_frto
)
2008 /* MTU probe and F-RTO won't really play nicely along currently */
2009 if (icsk
->icsk_mtup
.probe_size
)
2012 if (tcp_is_sackfrto(tp
))
2015 /* Avoid expensive walking of rexmit queue if possible */
2016 if (tp
->retrans_out
> 1)
2019 skb
= tcp_write_queue_head(sk
);
2020 if (tcp_skb_is_last(sk
, skb
))
2022 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
2023 tcp_for_write_queue_from(skb
, sk
) {
2024 if (skb
== tcp_send_head(sk
))
2026 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2028 /* Short-circuit when first non-SACKed skb has been checked */
2029 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2035 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2036 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2037 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2038 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2039 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2040 * bits are handled if the Loss state is really to be entered (in
2041 * tcp_enter_frto_loss).
2043 * Do like tcp_enter_loss() would; when RTO expires the second time it
2045 * "Reduce ssthresh if it has not yet been made inside this window."
2047 void tcp_enter_frto(struct sock
*sk
)
2049 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2050 struct tcp_sock
*tp
= tcp_sk(sk
);
2051 struct sk_buff
*skb
;
2053 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
2054 tp
->snd_una
== tp
->high_seq
||
2055 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
2056 !icsk
->icsk_retransmits
)) {
2057 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2058 /* Our state is too optimistic in ssthresh() call because cwnd
2059 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2060 * recovery has not yet completed. Pattern would be this: RTO,
2061 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2063 * RFC4138 should be more specific on what to do, even though
2064 * RTO is quite unlikely to occur after the first Cumulative ACK
2065 * due to back-off and complexity of triggering events ...
2067 if (tp
->frto_counter
) {
2069 stored_cwnd
= tp
->snd_cwnd
;
2071 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2072 tp
->snd_cwnd
= stored_cwnd
;
2074 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2076 /* ... in theory, cong.control module could do "any tricks" in
2077 * ssthresh(), which means that ca_state, lost bits and lost_out
2078 * counter would have to be faked before the call occurs. We
2079 * consider that too expensive, unlikely and hacky, so modules
2080 * using these in ssthresh() must deal these incompatibility
2081 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2083 tcp_ca_event(sk
, CA_EVENT_FRTO
);
2086 tp
->undo_marker
= tp
->snd_una
;
2087 tp
->undo_retrans
= 0;
2089 skb
= tcp_write_queue_head(sk
);
2090 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2091 tp
->undo_marker
= 0;
2092 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2093 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2094 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2096 tcp_verify_left_out(tp
);
2098 /* Too bad if TCP was application limited */
2099 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2101 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2102 * The last condition is necessary at least in tp->frto_counter case.
2104 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
2105 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
2106 after(tp
->high_seq
, tp
->snd_una
)) {
2107 tp
->frto_highmark
= tp
->high_seq
;
2109 tp
->frto_highmark
= tp
->snd_nxt
;
2111 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
2112 tp
->high_seq
= tp
->snd_nxt
;
2113 tp
->frto_counter
= 1;
2116 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2117 * which indicates that we should follow the traditional RTO recovery,
2118 * i.e. mark everything lost and do go-back-N retransmission.
2120 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
2122 struct tcp_sock
*tp
= tcp_sk(sk
);
2123 struct sk_buff
*skb
;
2126 tp
->retrans_out
= 0;
2127 if (tcp_is_reno(tp
))
2128 tcp_reset_reno_sack(tp
);
2130 tcp_for_write_queue(skb
, sk
) {
2131 if (skb
== tcp_send_head(sk
))
2134 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2136 * Count the retransmission made on RTO correctly (only when
2137 * waiting for the first ACK and did not get it)...
2139 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
2140 /* For some reason this R-bit might get cleared? */
2141 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
2142 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2143 /* ...enter this if branch just for the first segment */
2144 flag
|= FLAG_DATA_ACKED
;
2146 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2147 tp
->undo_marker
= 0;
2148 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2151 /* Marking forward transmissions that were made after RTO lost
2152 * can cause unnecessary retransmissions in some scenarios,
2153 * SACK blocks will mitigate that in some but not in all cases.
2154 * We used to not mark them but it was causing break-ups with
2155 * receivers that do only in-order receival.
2157 * TODO: we could detect presence of such receiver and select
2158 * different behavior per flow.
2160 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2161 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2162 tp
->lost_out
+= tcp_skb_pcount(skb
);
2163 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2166 tcp_verify_left_out(tp
);
2168 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2169 tp
->snd_cwnd_cnt
= 0;
2170 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2171 tp
->frto_counter
= 0;
2172 tp
->bytes_acked
= 0;
2174 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2175 sysctl_tcp_reordering
);
2176 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2177 tp
->high_seq
= tp
->snd_nxt
;
2178 TCP_ECN_queue_cwr(tp
);
2180 tcp_clear_all_retrans_hints(tp
);
2183 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2185 tp
->retrans_out
= 0;
2188 tp
->undo_marker
= 0;
2189 tp
->undo_retrans
= 0;
2192 void tcp_clear_retrans(struct tcp_sock
*tp
)
2194 tcp_clear_retrans_partial(tp
);
2196 tp
->fackets_out
= 0;
2200 /* Enter Loss state. If "how" is not zero, forget all SACK information
2201 * and reset tags completely, otherwise preserve SACKs. If receiver
2202 * dropped its ofo queue, we will know this due to reneging detection.
2204 void tcp_enter_loss(struct sock
*sk
, int how
)
2206 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2207 struct tcp_sock
*tp
= tcp_sk(sk
);
2208 struct sk_buff
*skb
;
2210 /* Reduce ssthresh if it has not yet been made inside this window. */
2211 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2212 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2213 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2214 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2215 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2218 tp
->snd_cwnd_cnt
= 0;
2219 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2221 tp
->bytes_acked
= 0;
2222 tcp_clear_retrans_partial(tp
);
2224 if (tcp_is_reno(tp
))
2225 tcp_reset_reno_sack(tp
);
2228 /* Push undo marker, if it was plain RTO and nothing
2229 * was retransmitted. */
2230 tp
->undo_marker
= tp
->snd_una
;
2233 tp
->fackets_out
= 0;
2235 tcp_clear_all_retrans_hints(tp
);
2237 tcp_for_write_queue(skb
, sk
) {
2238 if (skb
== tcp_send_head(sk
))
2241 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2242 tp
->undo_marker
= 0;
2243 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2244 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2245 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2246 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2247 tp
->lost_out
+= tcp_skb_pcount(skb
);
2248 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2251 tcp_verify_left_out(tp
);
2253 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2254 sysctl_tcp_reordering
);
2255 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2256 tp
->high_seq
= tp
->snd_nxt
;
2257 TCP_ECN_queue_cwr(tp
);
2258 /* Abort F-RTO algorithm if one is in progress */
2259 tp
->frto_counter
= 0;
2262 /* If ACK arrived pointing to a remembered SACK, it means that our
2263 * remembered SACKs do not reflect real state of receiver i.e.
2264 * receiver _host_ is heavily congested (or buggy).
2266 * Do processing similar to RTO timeout.
2268 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2270 if (flag
& FLAG_SACK_RENEGING
) {
2271 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2272 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2274 tcp_enter_loss(sk
, 1);
2275 icsk
->icsk_retransmits
++;
2276 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2277 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2278 icsk
->icsk_rto
, TCP_RTO_MAX
);
2284 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2286 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2289 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2290 * counter when SACK is enabled (without SACK, sacked_out is used for
2293 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2294 * segments up to the highest received SACK block so far and holes in
2297 * With reordering, holes may still be in flight, so RFC3517 recovery
2298 * uses pure sacked_out (total number of SACKed segments) even though
2299 * it violates the RFC that uses duplicate ACKs, often these are equal
2300 * but when e.g. out-of-window ACKs or packet duplication occurs,
2301 * they differ. Since neither occurs due to loss, TCP should really
2304 static inline int tcp_dupack_heuristics(struct tcp_sock
*tp
)
2306 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2309 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2311 return tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
;
2314 static inline int tcp_head_timedout(struct sock
*sk
)
2316 struct tcp_sock
*tp
= tcp_sk(sk
);
2318 return tp
->packets_out
&&
2319 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2322 /* Linux NewReno/SACK/FACK/ECN state machine.
2323 * --------------------------------------
2325 * "Open" Normal state, no dubious events, fast path.
2326 * "Disorder" In all the respects it is "Open",
2327 * but requires a bit more attention. It is entered when
2328 * we see some SACKs or dupacks. It is split of "Open"
2329 * mainly to move some processing from fast path to slow one.
2330 * "CWR" CWND was reduced due to some Congestion Notification event.
2331 * It can be ECN, ICMP source quench, local device congestion.
2332 * "Recovery" CWND was reduced, we are fast-retransmitting.
2333 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2335 * tcp_fastretrans_alert() is entered:
2336 * - each incoming ACK, if state is not "Open"
2337 * - when arrived ACK is unusual, namely:
2342 * Counting packets in flight is pretty simple.
2344 * in_flight = packets_out - left_out + retrans_out
2346 * packets_out is SND.NXT-SND.UNA counted in packets.
2348 * retrans_out is number of retransmitted segments.
2350 * left_out is number of segments left network, but not ACKed yet.
2352 * left_out = sacked_out + lost_out
2354 * sacked_out: Packets, which arrived to receiver out of order
2355 * and hence not ACKed. With SACKs this number is simply
2356 * amount of SACKed data. Even without SACKs
2357 * it is easy to give pretty reliable estimate of this number,
2358 * counting duplicate ACKs.
2360 * lost_out: Packets lost by network. TCP has no explicit
2361 * "loss notification" feedback from network (for now).
2362 * It means that this number can be only _guessed_.
2363 * Actually, it is the heuristics to predict lossage that
2364 * distinguishes different algorithms.
2366 * F.e. after RTO, when all the queue is considered as lost,
2367 * lost_out = packets_out and in_flight = retrans_out.
2369 * Essentially, we have now two algorithms counting
2372 * FACK: It is the simplest heuristics. As soon as we decided
2373 * that something is lost, we decide that _all_ not SACKed
2374 * packets until the most forward SACK are lost. I.e.
2375 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2376 * It is absolutely correct estimate, if network does not reorder
2377 * packets. And it loses any connection to reality when reordering
2378 * takes place. We use FACK by default until reordering
2379 * is suspected on the path to this destination.
2381 * NewReno: when Recovery is entered, we assume that one segment
2382 * is lost (classic Reno). While we are in Recovery and
2383 * a partial ACK arrives, we assume that one more packet
2384 * is lost (NewReno). This heuristics are the same in NewReno
2387 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2388 * deflation etc. CWND is real congestion window, never inflated, changes
2389 * only according to classic VJ rules.
2391 * Really tricky (and requiring careful tuning) part of algorithm
2392 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2393 * The first determines the moment _when_ we should reduce CWND and,
2394 * hence, slow down forward transmission. In fact, it determines the moment
2395 * when we decide that hole is caused by loss, rather than by a reorder.
2397 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2398 * holes, caused by lost packets.
2400 * And the most logically complicated part of algorithm is undo
2401 * heuristics. We detect false retransmits due to both too early
2402 * fast retransmit (reordering) and underestimated RTO, analyzing
2403 * timestamps and D-SACKs. When we detect that some segments were
2404 * retransmitted by mistake and CWND reduction was wrong, we undo
2405 * window reduction and abort recovery phase. This logic is hidden
2406 * inside several functions named tcp_try_undo_<something>.
2409 /* This function decides, when we should leave Disordered state
2410 * and enter Recovery phase, reducing congestion window.
2412 * Main question: may we further continue forward transmission
2413 * with the same cwnd?
2415 static int tcp_time_to_recover(struct sock
*sk
)
2417 struct tcp_sock
*tp
= tcp_sk(sk
);
2420 /* Do not perform any recovery during F-RTO algorithm */
2421 if (tp
->frto_counter
)
2424 /* Trick#1: The loss is proven. */
2428 /* Not-A-Trick#2 : Classic rule... */
2429 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2432 /* Trick#3 : when we use RFC2988 timer restart, fast
2433 * retransmit can be triggered by timeout of queue head.
2435 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2438 /* Trick#4: It is still not OK... But will it be useful to delay
2441 packets_out
= tp
->packets_out
;
2442 if (packets_out
<= tp
->reordering
&&
2443 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2444 !tcp_may_send_now(sk
)) {
2445 /* We have nothing to send. This connection is limited
2446 * either by receiver window or by application.
2451 /* If a thin stream is detected, retransmit after first
2452 * received dupack. Employ only if SACK is supported in order
2453 * to avoid possible corner-case series of spurious retransmissions
2454 * Use only if there are no unsent data.
2456 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2457 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2458 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2464 /* New heuristics: it is possible only after we switched to restart timer
2465 * each time when something is ACKed. Hence, we can detect timed out packets
2466 * during fast retransmit without falling to slow start.
2468 * Usefulness of this as is very questionable, since we should know which of
2469 * the segments is the next to timeout which is relatively expensive to find
2470 * in general case unless we add some data structure just for that. The
2471 * current approach certainly won't find the right one too often and when it
2472 * finally does find _something_ it usually marks large part of the window
2473 * right away (because a retransmission with a larger timestamp blocks the
2474 * loop from advancing). -ij
2476 static void tcp_timeout_skbs(struct sock
*sk
)
2478 struct tcp_sock
*tp
= tcp_sk(sk
);
2479 struct sk_buff
*skb
;
2481 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2484 skb
= tp
->scoreboard_skb_hint
;
2485 if (tp
->scoreboard_skb_hint
== NULL
)
2486 skb
= tcp_write_queue_head(sk
);
2488 tcp_for_write_queue_from(skb
, sk
) {
2489 if (skb
== tcp_send_head(sk
))
2491 if (!tcp_skb_timedout(sk
, skb
))
2494 tcp_skb_mark_lost(tp
, skb
);
2497 tp
->scoreboard_skb_hint
= skb
;
2499 tcp_verify_left_out(tp
);
2502 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2503 * is against sacked "cnt", otherwise it's against facked "cnt"
2505 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2507 struct tcp_sock
*tp
= tcp_sk(sk
);
2508 struct sk_buff
*skb
;
2513 WARN_ON(packets
> tp
->packets_out
);
2514 if (tp
->lost_skb_hint
) {
2515 skb
= tp
->lost_skb_hint
;
2516 cnt
= tp
->lost_cnt_hint
;
2517 /* Head already handled? */
2518 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2521 skb
= tcp_write_queue_head(sk
);
2525 tcp_for_write_queue_from(skb
, sk
) {
2526 if (skb
== tcp_send_head(sk
))
2528 /* TODO: do this better */
2529 /* this is not the most efficient way to do this... */
2530 tp
->lost_skb_hint
= skb
;
2531 tp
->lost_cnt_hint
= cnt
;
2533 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2537 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2538 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2539 cnt
+= tcp_skb_pcount(skb
);
2541 if (cnt
> packets
) {
2542 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2543 (oldcnt
>= packets
))
2546 mss
= skb_shinfo(skb
)->gso_size
;
2547 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2553 tcp_skb_mark_lost(tp
, skb
);
2558 tcp_verify_left_out(tp
);
2561 /* Account newly detected lost packet(s) */
2563 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2565 struct tcp_sock
*tp
= tcp_sk(sk
);
2567 if (tcp_is_reno(tp
)) {
2568 tcp_mark_head_lost(sk
, 1, 1);
2569 } else if (tcp_is_fack(tp
)) {
2570 int lost
= tp
->fackets_out
- tp
->reordering
;
2573 tcp_mark_head_lost(sk
, lost
, 0);
2575 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2576 if (sacked_upto
>= 0)
2577 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2578 else if (fast_rexmit
)
2579 tcp_mark_head_lost(sk
, 1, 1);
2582 tcp_timeout_skbs(sk
);
2585 /* CWND moderation, preventing bursts due to too big ACKs
2586 * in dubious situations.
2588 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2590 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2591 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2592 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2595 /* Lower bound on congestion window is slow start threshold
2596 * unless congestion avoidance choice decides to overide it.
2598 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2600 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2602 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2605 /* Decrease cwnd each second ack. */
2606 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2608 struct tcp_sock
*tp
= tcp_sk(sk
);
2609 int decr
= tp
->snd_cwnd_cnt
+ 1;
2611 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2612 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2613 tp
->snd_cwnd_cnt
= decr
& 1;
2616 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2617 tp
->snd_cwnd
-= decr
;
2619 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2620 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2624 /* Nothing was retransmitted or returned timestamp is less
2625 * than timestamp of the first retransmission.
2627 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2629 return !tp
->retrans_stamp
||
2630 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2631 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2634 /* Undo procedures. */
2636 #if FASTRETRANS_DEBUG > 1
2637 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2639 struct tcp_sock
*tp
= tcp_sk(sk
);
2640 struct inet_sock
*inet
= inet_sk(sk
);
2642 if (sk
->sk_family
== AF_INET
) {
2643 printk(KERN_DEBUG
"Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2645 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2646 tp
->snd_cwnd
, tcp_left_out(tp
),
2647 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2650 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2651 else if (sk
->sk_family
== AF_INET6
) {
2652 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2653 printk(KERN_DEBUG
"Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2655 &np
->daddr
, ntohs(inet
->inet_dport
),
2656 tp
->snd_cwnd
, tcp_left_out(tp
),
2657 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2663 #define DBGUNDO(x...) do { } while (0)
2666 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2668 struct tcp_sock
*tp
= tcp_sk(sk
);
2670 if (tp
->prior_ssthresh
) {
2671 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2673 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2674 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2676 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2678 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2679 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2680 TCP_ECN_withdraw_cwr(tp
);
2683 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2685 tcp_moderate_cwnd(tp
);
2686 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2689 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2691 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2694 /* People celebrate: "We love our President!" */
2695 static int tcp_try_undo_recovery(struct sock
*sk
)
2697 struct tcp_sock
*tp
= tcp_sk(sk
);
2699 if (tcp_may_undo(tp
)) {
2702 /* Happy end! We did not retransmit anything
2703 * or our original transmission succeeded.
2705 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2706 tcp_undo_cwr(sk
, 1);
2707 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2708 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2710 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2712 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2713 tp
->undo_marker
= 0;
2715 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2716 /* Hold old state until something *above* high_seq
2717 * is ACKed. For Reno it is MUST to prevent false
2718 * fast retransmits (RFC2582). SACK TCP is safe. */
2719 tcp_moderate_cwnd(tp
);
2722 tcp_set_ca_state(sk
, TCP_CA_Open
);
2726 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2727 static void tcp_try_undo_dsack(struct sock
*sk
)
2729 struct tcp_sock
*tp
= tcp_sk(sk
);
2731 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2732 DBGUNDO(sk
, "D-SACK");
2733 tcp_undo_cwr(sk
, 1);
2734 tp
->undo_marker
= 0;
2735 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2739 /* We can clear retrans_stamp when there are no retransmissions in the
2740 * window. It would seem that it is trivially available for us in
2741 * tp->retrans_out, however, that kind of assumptions doesn't consider
2742 * what will happen if errors occur when sending retransmission for the
2743 * second time. ...It could the that such segment has only
2744 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2745 * the head skb is enough except for some reneging corner cases that
2746 * are not worth the effort.
2748 * Main reason for all this complexity is the fact that connection dying
2749 * time now depends on the validity of the retrans_stamp, in particular,
2750 * that successive retransmissions of a segment must not advance
2751 * retrans_stamp under any conditions.
2753 static int tcp_any_retrans_done(struct sock
*sk
)
2755 struct tcp_sock
*tp
= tcp_sk(sk
);
2756 struct sk_buff
*skb
;
2758 if (tp
->retrans_out
)
2761 skb
= tcp_write_queue_head(sk
);
2762 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2768 /* Undo during fast recovery after partial ACK. */
2770 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2772 struct tcp_sock
*tp
= tcp_sk(sk
);
2773 /* Partial ACK arrived. Force Hoe's retransmit. */
2774 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2776 if (tcp_may_undo(tp
)) {
2777 /* Plain luck! Hole if filled with delayed
2778 * packet, rather than with a retransmit.
2780 if (!tcp_any_retrans_done(sk
))
2781 tp
->retrans_stamp
= 0;
2783 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2786 tcp_undo_cwr(sk
, 0);
2787 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2789 /* So... Do not make Hoe's retransmit yet.
2790 * If the first packet was delayed, the rest
2791 * ones are most probably delayed as well.
2798 /* Undo during loss recovery after partial ACK. */
2799 static int tcp_try_undo_loss(struct sock
*sk
)
2801 struct tcp_sock
*tp
= tcp_sk(sk
);
2803 if (tcp_may_undo(tp
)) {
2804 struct sk_buff
*skb
;
2805 tcp_for_write_queue(skb
, sk
) {
2806 if (skb
== tcp_send_head(sk
))
2808 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2811 tcp_clear_all_retrans_hints(tp
);
2813 DBGUNDO(sk
, "partial loss");
2815 tcp_undo_cwr(sk
, 1);
2816 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2817 inet_csk(sk
)->icsk_retransmits
= 0;
2818 tp
->undo_marker
= 0;
2819 if (tcp_is_sack(tp
))
2820 tcp_set_ca_state(sk
, TCP_CA_Open
);
2826 static inline void tcp_complete_cwr(struct sock
*sk
)
2828 struct tcp_sock
*tp
= tcp_sk(sk
);
2829 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2830 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2831 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2834 static void tcp_try_keep_open(struct sock
*sk
)
2836 struct tcp_sock
*tp
= tcp_sk(sk
);
2837 int state
= TCP_CA_Open
;
2839 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
) || tp
->undo_marker
)
2840 state
= TCP_CA_Disorder
;
2842 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2843 tcp_set_ca_state(sk
, state
);
2844 tp
->high_seq
= tp
->snd_nxt
;
2848 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2850 struct tcp_sock
*tp
= tcp_sk(sk
);
2852 tcp_verify_left_out(tp
);
2854 if (!tp
->frto_counter
&& !tcp_any_retrans_done(sk
))
2855 tp
->retrans_stamp
= 0;
2857 if (flag
& FLAG_ECE
)
2858 tcp_enter_cwr(sk
, 1);
2860 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2861 tcp_try_keep_open(sk
);
2862 tcp_moderate_cwnd(tp
);
2864 tcp_cwnd_down(sk
, flag
);
2868 static void tcp_mtup_probe_failed(struct sock
*sk
)
2870 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2872 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2873 icsk
->icsk_mtup
.probe_size
= 0;
2876 static void tcp_mtup_probe_success(struct sock
*sk
)
2878 struct tcp_sock
*tp
= tcp_sk(sk
);
2879 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2881 /* FIXME: breaks with very large cwnd */
2882 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2883 tp
->snd_cwnd
= tp
->snd_cwnd
*
2884 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2885 icsk
->icsk_mtup
.probe_size
;
2886 tp
->snd_cwnd_cnt
= 0;
2887 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2888 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2890 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2891 icsk
->icsk_mtup
.probe_size
= 0;
2892 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2895 /* Do a simple retransmit without using the backoff mechanisms in
2896 * tcp_timer. This is used for path mtu discovery.
2897 * The socket is already locked here.
2899 void tcp_simple_retransmit(struct sock
*sk
)
2901 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2902 struct tcp_sock
*tp
= tcp_sk(sk
);
2903 struct sk_buff
*skb
;
2904 unsigned int mss
= tcp_current_mss(sk
);
2905 u32 prior_lost
= tp
->lost_out
;
2907 tcp_for_write_queue(skb
, sk
) {
2908 if (skb
== tcp_send_head(sk
))
2910 if (tcp_skb_seglen(skb
) > mss
&&
2911 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2912 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2913 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2914 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2916 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2920 tcp_clear_retrans_hints_partial(tp
);
2922 if (prior_lost
== tp
->lost_out
)
2925 if (tcp_is_reno(tp
))
2926 tcp_limit_reno_sacked(tp
);
2928 tcp_verify_left_out(tp
);
2930 /* Don't muck with the congestion window here.
2931 * Reason is that we do not increase amount of _data_
2932 * in network, but units changed and effective
2933 * cwnd/ssthresh really reduced now.
2935 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2936 tp
->high_seq
= tp
->snd_nxt
;
2937 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2938 tp
->prior_ssthresh
= 0;
2939 tp
->undo_marker
= 0;
2940 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2942 tcp_xmit_retransmit_queue(sk
);
2944 EXPORT_SYMBOL(tcp_simple_retransmit
);
2946 /* Process an event, which can update packets-in-flight not trivially.
2947 * Main goal of this function is to calculate new estimate for left_out,
2948 * taking into account both packets sitting in receiver's buffer and
2949 * packets lost by network.
2951 * Besides that it does CWND reduction, when packet loss is detected
2952 * and changes state of machine.
2954 * It does _not_ decide what to send, it is made in function
2955 * tcp_xmit_retransmit_queue().
2957 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2959 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2960 struct tcp_sock
*tp
= tcp_sk(sk
);
2961 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2962 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2963 (tcp_fackets_out(tp
) > tp
->reordering
));
2964 int fast_rexmit
= 0, mib_idx
;
2966 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2968 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2969 tp
->fackets_out
= 0;
2971 /* Now state machine starts.
2972 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2973 if (flag
& FLAG_ECE
)
2974 tp
->prior_ssthresh
= 0;
2976 /* B. In all the states check for reneging SACKs. */
2977 if (tcp_check_sack_reneging(sk
, flag
))
2980 /* C. Process data loss notification, provided it is valid. */
2981 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2982 before(tp
->snd_una
, tp
->high_seq
) &&
2983 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2984 tp
->fackets_out
> tp
->reordering
) {
2985 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
, 0);
2986 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2989 /* D. Check consistency of the current state. */
2990 tcp_verify_left_out(tp
);
2992 /* E. Check state exit conditions. State can be terminated
2993 * when high_seq is ACKed. */
2994 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2995 WARN_ON(tp
->retrans_out
!= 0);
2996 tp
->retrans_stamp
= 0;
2997 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2998 switch (icsk
->icsk_ca_state
) {
3000 icsk
->icsk_retransmits
= 0;
3001 if (tcp_try_undo_recovery(sk
))
3006 /* CWR is to be held something *above* high_seq
3007 * is ACKed for CWR bit to reach receiver. */
3008 if (tp
->snd_una
!= tp
->high_seq
) {
3009 tcp_complete_cwr(sk
);
3010 tcp_set_ca_state(sk
, TCP_CA_Open
);
3014 case TCP_CA_Disorder
:
3015 tcp_try_undo_dsack(sk
);
3016 if (!tp
->undo_marker
||
3017 /* For SACK case do not Open to allow to undo
3018 * catching for all duplicate ACKs. */
3019 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
3020 tp
->undo_marker
= 0;
3021 tcp_set_ca_state(sk
, TCP_CA_Open
);
3025 case TCP_CA_Recovery
:
3026 if (tcp_is_reno(tp
))
3027 tcp_reset_reno_sack(tp
);
3028 if (tcp_try_undo_recovery(sk
))
3030 tcp_complete_cwr(sk
);
3035 /* F. Process state. */
3036 switch (icsk
->icsk_ca_state
) {
3037 case TCP_CA_Recovery
:
3038 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
3039 if (tcp_is_reno(tp
) && is_dupack
)
3040 tcp_add_reno_sack(sk
);
3042 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
3045 if (flag
& FLAG_DATA_ACKED
)
3046 icsk
->icsk_retransmits
= 0;
3047 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
3048 tcp_reset_reno_sack(tp
);
3049 if (!tcp_try_undo_loss(sk
)) {
3050 tcp_moderate_cwnd(tp
);
3051 tcp_xmit_retransmit_queue(sk
);
3054 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
3056 /* Loss is undone; fall through to processing in Open state. */
3058 if (tcp_is_reno(tp
)) {
3059 if (flag
& FLAG_SND_UNA_ADVANCED
)
3060 tcp_reset_reno_sack(tp
);
3062 tcp_add_reno_sack(sk
);
3065 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
3066 tcp_try_undo_dsack(sk
);
3068 if (!tcp_time_to_recover(sk
)) {
3069 tcp_try_to_open(sk
, flag
);
3073 /* MTU probe failure: don't reduce cwnd */
3074 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3075 icsk
->icsk_mtup
.probe_size
&&
3076 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3077 tcp_mtup_probe_failed(sk
);
3078 /* Restores the reduction we did in tcp_mtup_probe() */
3080 tcp_simple_retransmit(sk
);
3084 /* Otherwise enter Recovery state */
3086 if (tcp_is_reno(tp
))
3087 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
3089 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
3091 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3093 tp
->high_seq
= tp
->snd_nxt
;
3094 tp
->prior_ssthresh
= 0;
3095 tp
->undo_marker
= tp
->snd_una
;
3096 tp
->undo_retrans
= tp
->retrans_out
;
3098 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
3099 if (!(flag
& FLAG_ECE
))
3100 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
3101 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
3102 TCP_ECN_queue_cwr(tp
);
3105 tp
->bytes_acked
= 0;
3106 tp
->snd_cwnd_cnt
= 0;
3107 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
3111 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3112 tcp_update_scoreboard(sk
, fast_rexmit
);
3113 tcp_cwnd_down(sk
, flag
);
3114 tcp_xmit_retransmit_queue(sk
);
3117 static void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
3119 tcp_rtt_estimator(sk
, seq_rtt
);
3121 inet_csk(sk
)->icsk_backoff
= 0;
3124 /* Read draft-ietf-tcplw-high-performance before mucking
3125 * with this code. (Supersedes RFC1323)
3127 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3129 /* RTTM Rule: A TSecr value received in a segment is used to
3130 * update the averaged RTT measurement only if the segment
3131 * acknowledges some new data, i.e., only if it advances the
3132 * left edge of the send window.
3134 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3135 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3137 * Changed: reset backoff as soon as we see the first valid sample.
3138 * If we do not, we get strongly overestimated rto. With timestamps
3139 * samples are accepted even from very old segments: f.e., when rtt=1
3140 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3141 * answer arrives rto becomes 120 seconds! If at least one of segments
3142 * in window is lost... Voila. --ANK (010210)
3144 struct tcp_sock
*tp
= tcp_sk(sk
);
3146 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
3149 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3151 /* We don't have a timestamp. Can only use
3152 * packets that are not retransmitted to determine
3153 * rtt estimates. Also, we must not reset the
3154 * backoff for rto until we get a non-retransmitted
3155 * packet. This allows us to deal with a situation
3156 * where the network delay has increased suddenly.
3157 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3160 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3163 tcp_valid_rtt_meas(sk
, seq_rtt
);
3166 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3169 const struct tcp_sock
*tp
= tcp_sk(sk
);
3170 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3171 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3172 tcp_ack_saw_tstamp(sk
, flag
);
3173 else if (seq_rtt
>= 0)
3174 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3177 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3179 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3180 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3181 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3184 /* Restart timer after forward progress on connection.
3185 * RFC2988 recommends to restart timer to now+rto.
3187 static void tcp_rearm_rto(struct sock
*sk
)
3189 struct tcp_sock
*tp
= tcp_sk(sk
);
3191 if (!tp
->packets_out
) {
3192 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3194 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3195 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3199 /* If we get here, the whole TSO packet has not been acked. */
3200 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3202 struct tcp_sock
*tp
= tcp_sk(sk
);
3205 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3207 packets_acked
= tcp_skb_pcount(skb
);
3208 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3210 packets_acked
-= tcp_skb_pcount(skb
);
3212 if (packets_acked
) {
3213 BUG_ON(tcp_skb_pcount(skb
) == 0);
3214 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3217 return packets_acked
;
3220 /* Remove acknowledged frames from the retransmission queue. If our packet
3221 * is before the ack sequence we can discard it as it's confirmed to have
3222 * arrived at the other end.
3224 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3227 struct tcp_sock
*tp
= tcp_sk(sk
);
3228 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3229 struct sk_buff
*skb
;
3230 u32 now
= tcp_time_stamp
;
3231 int fully_acked
= 1;
3234 u32 reord
= tp
->packets_out
;
3235 u32 prior_sacked
= tp
->sacked_out
;
3237 s32 ca_seq_rtt
= -1;
3238 ktime_t last_ackt
= net_invalid_timestamp();
3240 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3241 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3243 u8 sacked
= scb
->sacked
;
3245 /* Determine how many packets and what bytes were acked, tso and else */
3246 if (after(scb
->end_seq
, tp
->snd_una
)) {
3247 if (tcp_skb_pcount(skb
) == 1 ||
3248 !after(tp
->snd_una
, scb
->seq
))
3251 acked_pcount
= tcp_tso_acked(sk
, skb
);
3257 acked_pcount
= tcp_skb_pcount(skb
);
3260 if (sacked
& TCPCB_RETRANS
) {
3261 if (sacked
& TCPCB_SACKED_RETRANS
)
3262 tp
->retrans_out
-= acked_pcount
;
3263 flag
|= FLAG_RETRANS_DATA_ACKED
;
3266 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3267 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3269 ca_seq_rtt
= now
- scb
->when
;
3270 last_ackt
= skb
->tstamp
;
3272 seq_rtt
= ca_seq_rtt
;
3274 if (!(sacked
& TCPCB_SACKED_ACKED
))
3275 reord
= min(pkts_acked
, reord
);
3278 if (sacked
& TCPCB_SACKED_ACKED
)
3279 tp
->sacked_out
-= acked_pcount
;
3280 if (sacked
& TCPCB_LOST
)
3281 tp
->lost_out
-= acked_pcount
;
3283 tp
->packets_out
-= acked_pcount
;
3284 pkts_acked
+= acked_pcount
;
3286 /* Initial outgoing SYN's get put onto the write_queue
3287 * just like anything else we transmit. It is not
3288 * true data, and if we misinform our callers that
3289 * this ACK acks real data, we will erroneously exit
3290 * connection startup slow start one packet too
3291 * quickly. This is severely frowned upon behavior.
3293 if (!(scb
->flags
& TCPHDR_SYN
)) {
3294 flag
|= FLAG_DATA_ACKED
;
3296 flag
|= FLAG_SYN_ACKED
;
3297 tp
->retrans_stamp
= 0;
3303 tcp_unlink_write_queue(skb
, sk
);
3304 sk_wmem_free_skb(sk
, skb
);
3305 tp
->scoreboard_skb_hint
= NULL
;
3306 if (skb
== tp
->retransmit_skb_hint
)
3307 tp
->retransmit_skb_hint
= NULL
;
3308 if (skb
== tp
->lost_skb_hint
)
3309 tp
->lost_skb_hint
= NULL
;
3312 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3313 tp
->snd_up
= tp
->snd_una
;
3315 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3316 flag
|= FLAG_SACK_RENEGING
;
3318 if (flag
& FLAG_ACKED
) {
3319 const struct tcp_congestion_ops
*ca_ops
3320 = inet_csk(sk
)->icsk_ca_ops
;
3322 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3323 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3324 tcp_mtup_probe_success(sk
);
3327 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3330 if (tcp_is_reno(tp
)) {
3331 tcp_remove_reno_sacks(sk
, pkts_acked
);
3335 /* Non-retransmitted hole got filled? That's reordering */
3336 if (reord
< prior_fackets
)
3337 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3339 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3340 prior_sacked
- tp
->sacked_out
;
3341 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3344 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3346 if (ca_ops
->pkts_acked
) {
3349 /* Is the ACK triggering packet unambiguous? */
3350 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3351 /* High resolution needed and available? */
3352 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3353 !ktime_equal(last_ackt
,
3354 net_invalid_timestamp()))
3355 rtt_us
= ktime_us_delta(ktime_get_real(),
3357 else if (ca_seq_rtt
> 0)
3358 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3361 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3365 #if FASTRETRANS_DEBUG > 0
3366 WARN_ON((int)tp
->sacked_out
< 0);
3367 WARN_ON((int)tp
->lost_out
< 0);
3368 WARN_ON((int)tp
->retrans_out
< 0);
3369 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3370 icsk
= inet_csk(sk
);
3372 printk(KERN_DEBUG
"Leak l=%u %d\n",
3373 tp
->lost_out
, icsk
->icsk_ca_state
);
3376 if (tp
->sacked_out
) {
3377 printk(KERN_DEBUG
"Leak s=%u %d\n",
3378 tp
->sacked_out
, icsk
->icsk_ca_state
);
3381 if (tp
->retrans_out
) {
3382 printk(KERN_DEBUG
"Leak r=%u %d\n",
3383 tp
->retrans_out
, icsk
->icsk_ca_state
);
3384 tp
->retrans_out
= 0;
3391 static void tcp_ack_probe(struct sock
*sk
)
3393 const struct tcp_sock
*tp
= tcp_sk(sk
);
3394 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3396 /* Was it a usable window open? */
3398 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3399 icsk
->icsk_backoff
= 0;
3400 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3401 /* Socket must be waked up by subsequent tcp_data_snd_check().
3402 * This function is not for random using!
3405 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3406 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3411 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3413 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3414 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3417 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3419 const struct tcp_sock
*tp
= tcp_sk(sk
);
3420 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3421 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3424 /* Check that window update is acceptable.
3425 * The function assumes that snd_una<=ack<=snd_next.
3427 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3428 const u32 ack
, const u32 ack_seq
,
3431 return after(ack
, tp
->snd_una
) ||
3432 after(ack_seq
, tp
->snd_wl1
) ||
3433 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3436 /* Update our send window.
3438 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3439 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3441 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3444 struct tcp_sock
*tp
= tcp_sk(sk
);
3446 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3448 if (likely(!tcp_hdr(skb
)->syn
))
3449 nwin
<<= tp
->rx_opt
.snd_wscale
;
3451 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3452 flag
|= FLAG_WIN_UPDATE
;
3453 tcp_update_wl(tp
, ack_seq
);
3455 if (tp
->snd_wnd
!= nwin
) {
3458 /* Note, it is the only place, where
3459 * fast path is recovered for sending TCP.
3462 tcp_fast_path_check(sk
);
3464 if (nwin
> tp
->max_window
) {
3465 tp
->max_window
= nwin
;
3466 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3476 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3477 * continue in congestion avoidance.
3479 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3481 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3482 tp
->snd_cwnd_cnt
= 0;
3483 tp
->bytes_acked
= 0;
3484 TCP_ECN_queue_cwr(tp
);
3485 tcp_moderate_cwnd(tp
);
3488 /* A conservative spurious RTO response algorithm: reduce cwnd using
3489 * rate halving and continue in congestion avoidance.
3491 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3493 tcp_enter_cwr(sk
, 0);
3496 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3498 if (flag
& FLAG_ECE
)
3499 tcp_ratehalving_spur_to_response(sk
);
3501 tcp_undo_cwr(sk
, 1);
3504 /* F-RTO spurious RTO detection algorithm (RFC4138)
3506 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3507 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3508 * window (but not to or beyond highest sequence sent before RTO):
3509 * On First ACK, send two new segments out.
3510 * On Second ACK, RTO was likely spurious. Do spurious response (response
3511 * algorithm is not part of the F-RTO detection algorithm
3512 * given in RFC4138 but can be selected separately).
3513 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3514 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3515 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3516 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3518 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3519 * original window even after we transmit two new data segments.
3522 * on first step, wait until first cumulative ACK arrives, then move to
3523 * the second step. In second step, the next ACK decides.
3525 * F-RTO is implemented (mainly) in four functions:
3526 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3527 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3528 * called when tcp_use_frto() showed green light
3529 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3530 * - tcp_enter_frto_loss() is called if there is not enough evidence
3531 * to prove that the RTO is indeed spurious. It transfers the control
3532 * from F-RTO to the conventional RTO recovery
3534 static int tcp_process_frto(struct sock
*sk
, int flag
)
3536 struct tcp_sock
*tp
= tcp_sk(sk
);
3538 tcp_verify_left_out(tp
);
3540 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3541 if (flag
& FLAG_DATA_ACKED
)
3542 inet_csk(sk
)->icsk_retransmits
= 0;
3544 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3545 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3546 tp
->undo_marker
= 0;
3548 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3549 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3553 if (!tcp_is_sackfrto(tp
)) {
3554 /* RFC4138 shortcoming in step 2; should also have case c):
3555 * ACK isn't duplicate nor advances window, e.g., opposite dir
3558 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3561 if (!(flag
& FLAG_DATA_ACKED
)) {
3562 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3567 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3568 /* Prevent sending of new data. */
3569 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3570 tcp_packets_in_flight(tp
));
3574 if ((tp
->frto_counter
>= 2) &&
3575 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3576 ((flag
& FLAG_DATA_SACKED
) &&
3577 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3578 /* RFC4138 shortcoming (see comment above) */
3579 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3580 (flag
& FLAG_NOT_DUP
))
3583 tcp_enter_frto_loss(sk
, 3, flag
);
3588 if (tp
->frto_counter
== 1) {
3589 /* tcp_may_send_now needs to see updated state */
3590 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3591 tp
->frto_counter
= 2;
3593 if (!tcp_may_send_now(sk
))
3594 tcp_enter_frto_loss(sk
, 2, flag
);
3598 switch (sysctl_tcp_frto_response
) {
3600 tcp_undo_spur_to_response(sk
, flag
);
3603 tcp_conservative_spur_to_response(tp
);
3606 tcp_ratehalving_spur_to_response(sk
);
3609 tp
->frto_counter
= 0;
3610 tp
->undo_marker
= 0;
3611 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3616 /* This routine deals with incoming acks, but not outgoing ones. */
3617 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3619 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3620 struct tcp_sock
*tp
= tcp_sk(sk
);
3621 u32 prior_snd_una
= tp
->snd_una
;
3622 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3623 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3624 u32 prior_in_flight
;
3629 /* If the ack is older than previous acks
3630 * then we can probably ignore it.
3632 if (before(ack
, prior_snd_una
))
3635 /* If the ack includes data we haven't sent yet, discard
3636 * this segment (RFC793 Section 3.9).
3638 if (after(ack
, tp
->snd_nxt
))
3641 if (after(ack
, prior_snd_una
))
3642 flag
|= FLAG_SND_UNA_ADVANCED
;
3644 if (sysctl_tcp_abc
) {
3645 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3646 tp
->bytes_acked
+= ack
- prior_snd_una
;
3647 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3648 /* we assume just one segment left network */
3649 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3653 prior_fackets
= tp
->fackets_out
;
3654 prior_in_flight
= tcp_packets_in_flight(tp
);
3656 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3657 /* Window is constant, pure forward advance.
3658 * No more checks are required.
3659 * Note, we use the fact that SND.UNA>=SND.WL2.
3661 tcp_update_wl(tp
, ack_seq
);
3663 flag
|= FLAG_WIN_UPDATE
;
3665 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3667 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3669 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3672 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3674 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3676 if (TCP_SKB_CB(skb
)->sacked
)
3677 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3679 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3682 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3685 /* We passed data and got it acked, remove any soft error
3686 * log. Something worked...
3688 sk
->sk_err_soft
= 0;
3689 icsk
->icsk_probes_out
= 0;
3690 tp
->rcv_tstamp
= tcp_time_stamp
;
3691 prior_packets
= tp
->packets_out
;
3695 /* See if we can take anything off of the retransmit queue. */
3696 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3698 if (tp
->frto_counter
)
3699 frto_cwnd
= tcp_process_frto(sk
, flag
);
3700 /* Guarantee sacktag reordering detection against wrap-arounds */
3701 if (before(tp
->frto_highmark
, tp
->snd_una
))
3702 tp
->frto_highmark
= 0;
3704 if (tcp_ack_is_dubious(sk
, flag
)) {
3705 /* Advance CWND, if state allows this. */
3706 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3707 tcp_may_raise_cwnd(sk
, flag
))
3708 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3709 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3712 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3713 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3716 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3717 dst_confirm(__sk_dst_get(sk
));
3722 /* If this ack opens up a zero window, clear backoff. It was
3723 * being used to time the probes, and is probably far higher than
3724 * it needs to be for normal retransmission.
3726 if (tcp_send_head(sk
))
3731 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3735 if (TCP_SKB_CB(skb
)->sacked
) {
3736 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3737 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3738 tcp_try_keep_open(sk
);
3741 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3745 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3746 * But, this can also be called on packets in the established flow when
3747 * the fast version below fails.
3749 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3750 u8
**hvpp
, int estab
)
3753 struct tcphdr
*th
= tcp_hdr(skb
);
3754 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3756 ptr
= (unsigned char *)(th
+ 1);
3757 opt_rx
->saw_tstamp
= 0;
3759 while (length
> 0) {
3760 int opcode
= *ptr
++;
3766 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3771 if (opsize
< 2) /* "silly options" */
3773 if (opsize
> length
)
3774 return; /* don't parse partial options */
3777 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3778 u16 in_mss
= get_unaligned_be16(ptr
);
3780 if (opt_rx
->user_mss
&&
3781 opt_rx
->user_mss
< in_mss
)
3782 in_mss
= opt_rx
->user_mss
;
3783 opt_rx
->mss_clamp
= in_mss
;
3788 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3789 !estab
&& sysctl_tcp_window_scaling
) {
3790 __u8 snd_wscale
= *(__u8
*)ptr
;
3791 opt_rx
->wscale_ok
= 1;
3792 if (snd_wscale
> 14) {
3793 if (net_ratelimit())
3794 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3795 "scaling value %d >14 received.\n",
3799 opt_rx
->snd_wscale
= snd_wscale
;
3802 case TCPOPT_TIMESTAMP
:
3803 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3804 ((estab
&& opt_rx
->tstamp_ok
) ||
3805 (!estab
&& sysctl_tcp_timestamps
))) {
3806 opt_rx
->saw_tstamp
= 1;
3807 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3808 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3811 case TCPOPT_SACK_PERM
:
3812 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3813 !estab
&& sysctl_tcp_sack
) {
3814 opt_rx
->sack_ok
= 1;
3815 tcp_sack_reset(opt_rx
);
3820 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3821 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3823 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3826 #ifdef CONFIG_TCP_MD5SIG
3829 * The MD5 Hash has already been
3830 * checked (see tcp_v{4,6}_do_rcv()).
3835 /* This option is variable length.
3838 case TCPOLEN_COOKIE_BASE
:
3839 /* not yet implemented */
3841 case TCPOLEN_COOKIE_PAIR
:
3842 /* not yet implemented */
3844 case TCPOLEN_COOKIE_MIN
+0:
3845 case TCPOLEN_COOKIE_MIN
+2:
3846 case TCPOLEN_COOKIE_MIN
+4:
3847 case TCPOLEN_COOKIE_MIN
+6:
3848 case TCPOLEN_COOKIE_MAX
:
3849 /* 16-bit multiple */
3850 opt_rx
->cookie_plus
= opsize
;
3865 EXPORT_SYMBOL(tcp_parse_options
);
3867 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3869 __be32
*ptr
= (__be32
*)(th
+ 1);
3871 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3872 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3873 tp
->rx_opt
.saw_tstamp
= 1;
3875 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3877 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3883 /* Fast parse options. This hopes to only see timestamps.
3884 * If it is wrong it falls back on tcp_parse_options().
3886 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3887 struct tcp_sock
*tp
, u8
**hvpp
)
3889 /* In the spirit of fast parsing, compare doff directly to constant
3890 * values. Because equality is used, short doff can be ignored here.
3892 if (th
->doff
== (sizeof(*th
) / 4)) {
3893 tp
->rx_opt
.saw_tstamp
= 0;
3895 } else if (tp
->rx_opt
.tstamp_ok
&&
3896 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3897 if (tcp_parse_aligned_timestamp(tp
, th
))
3900 tcp_parse_options(skb
, &tp
->rx_opt
, hvpp
, 1);
3904 #ifdef CONFIG_TCP_MD5SIG
3906 * Parse MD5 Signature option
3908 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3910 int length
= (th
->doff
<< 2) - sizeof (*th
);
3911 u8
*ptr
= (u8
*)(th
+ 1);
3913 /* If the TCP option is too short, we can short cut */
3914 if (length
< TCPOLEN_MD5SIG
)
3917 while (length
> 0) {
3918 int opcode
= *ptr
++;
3929 if (opsize
< 2 || opsize
> length
)
3931 if (opcode
== TCPOPT_MD5SIG
)
3932 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3939 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3942 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3944 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3945 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3948 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3950 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3951 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3952 * extra check below makes sure this can only happen
3953 * for pure ACK frames. -DaveM
3955 * Not only, also it occurs for expired timestamps.
3958 if (tcp_paws_check(&tp
->rx_opt
, 0))
3959 tcp_store_ts_recent(tp
);
3963 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3965 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3966 * it can pass through stack. So, the following predicate verifies that
3967 * this segment is not used for anything but congestion avoidance or
3968 * fast retransmit. Moreover, we even are able to eliminate most of such
3969 * second order effects, if we apply some small "replay" window (~RTO)
3970 * to timestamp space.
3972 * All these measures still do not guarantee that we reject wrapped ACKs
3973 * on networks with high bandwidth, when sequence space is recycled fastly,
3974 * but it guarantees that such events will be very rare and do not affect
3975 * connection seriously. This doesn't look nice, but alas, PAWS is really
3978 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3979 * states that events when retransmit arrives after original data are rare.
3980 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3981 * the biggest problem on large power networks even with minor reordering.
3982 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3983 * up to bandwidth of 18Gigabit/sec. 8) ]
3986 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3988 struct tcp_sock
*tp
= tcp_sk(sk
);
3989 struct tcphdr
*th
= tcp_hdr(skb
);
3990 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3991 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3993 return (/* 1. Pure ACK with correct sequence number. */
3994 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3996 /* 2. ... and duplicate ACK. */
3997 ack
== tp
->snd_una
&&
3999 /* 3. ... and does not update window. */
4000 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
4002 /* 4. ... and sits in replay window. */
4003 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
4006 static inline int tcp_paws_discard(const struct sock
*sk
,
4007 const struct sk_buff
*skb
)
4009 const struct tcp_sock
*tp
= tcp_sk(sk
);
4011 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
4012 !tcp_disordered_ack(sk
, skb
);
4015 /* Check segment sequence number for validity.
4017 * Segment controls are considered valid, if the segment
4018 * fits to the window after truncation to the window. Acceptability
4019 * of data (and SYN, FIN, of course) is checked separately.
4020 * See tcp_data_queue(), for example.
4022 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4023 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4024 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4025 * (borrowed from freebsd)
4028 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4030 return !before(end_seq
, tp
->rcv_wup
) &&
4031 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4034 /* When we get a reset we do this. */
4035 static void tcp_reset(struct sock
*sk
)
4037 /* We want the right error as BSD sees it (and indeed as we do). */
4038 switch (sk
->sk_state
) {
4040 sk
->sk_err
= ECONNREFUSED
;
4042 case TCP_CLOSE_WAIT
:
4048 sk
->sk_err
= ECONNRESET
;
4050 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4053 if (!sock_flag(sk
, SOCK_DEAD
))
4054 sk
->sk_error_report(sk
);
4060 * Process the FIN bit. This now behaves as it is supposed to work
4061 * and the FIN takes effect when it is validly part of sequence
4062 * space. Not before when we get holes.
4064 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4065 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4068 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4069 * close and we go into CLOSING (and later onto TIME-WAIT)
4071 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4073 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
4075 struct tcp_sock
*tp
= tcp_sk(sk
);
4077 inet_csk_schedule_ack(sk
);
4079 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4080 sock_set_flag(sk
, SOCK_DONE
);
4082 switch (sk
->sk_state
) {
4084 case TCP_ESTABLISHED
:
4085 /* Move to CLOSE_WAIT */
4086 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4087 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4090 case TCP_CLOSE_WAIT
:
4092 /* Received a retransmission of the FIN, do
4097 /* RFC793: Remain in the LAST-ACK state. */
4101 /* This case occurs when a simultaneous close
4102 * happens, we must ack the received FIN and
4103 * enter the CLOSING state.
4106 tcp_set_state(sk
, TCP_CLOSING
);
4109 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4111 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4114 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4115 * cases we should never reach this piece of code.
4117 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
4118 __func__
, sk
->sk_state
);
4122 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4123 * Probably, we should reset in this case. For now drop them.
4125 __skb_queue_purge(&tp
->out_of_order_queue
);
4126 if (tcp_is_sack(tp
))
4127 tcp_sack_reset(&tp
->rx_opt
);
4130 if (!sock_flag(sk
, SOCK_DEAD
)) {
4131 sk
->sk_state_change(sk
);
4133 /* Do not send POLL_HUP for half duplex close. */
4134 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4135 sk
->sk_state
== TCP_CLOSE
)
4136 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4138 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4142 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4145 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4146 if (before(seq
, sp
->start_seq
))
4147 sp
->start_seq
= seq
;
4148 if (after(end_seq
, sp
->end_seq
))
4149 sp
->end_seq
= end_seq
;
4155 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4157 struct tcp_sock
*tp
= tcp_sk(sk
);
4159 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4162 if (before(seq
, tp
->rcv_nxt
))
4163 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4165 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4167 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4169 tp
->rx_opt
.dsack
= 1;
4170 tp
->duplicate_sack
[0].start_seq
= seq
;
4171 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4175 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4177 struct tcp_sock
*tp
= tcp_sk(sk
);
4179 if (!tp
->rx_opt
.dsack
)
4180 tcp_dsack_set(sk
, seq
, end_seq
);
4182 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4185 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
4187 struct tcp_sock
*tp
= tcp_sk(sk
);
4189 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4190 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4191 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4192 tcp_enter_quickack_mode(sk
);
4194 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4195 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4197 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4198 end_seq
= tp
->rcv_nxt
;
4199 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4206 /* These routines update the SACK block as out-of-order packets arrive or
4207 * in-order packets close up the sequence space.
4209 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4212 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4213 struct tcp_sack_block
*swalk
= sp
+ 1;
4215 /* See if the recent change to the first SACK eats into
4216 * or hits the sequence space of other SACK blocks, if so coalesce.
4218 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4219 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4222 /* Zap SWALK, by moving every further SACK up by one slot.
4223 * Decrease num_sacks.
4225 tp
->rx_opt
.num_sacks
--;
4226 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4230 this_sack
++, swalk
++;
4234 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4236 struct tcp_sock
*tp
= tcp_sk(sk
);
4237 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4238 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4244 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4245 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4246 /* Rotate this_sack to the first one. */
4247 for (; this_sack
> 0; this_sack
--, sp
--)
4248 swap(*sp
, *(sp
- 1));
4250 tcp_sack_maybe_coalesce(tp
);
4255 /* Could not find an adjacent existing SACK, build a new one,
4256 * put it at the front, and shift everyone else down. We
4257 * always know there is at least one SACK present already here.
4259 * If the sack array is full, forget about the last one.
4261 if (this_sack
>= TCP_NUM_SACKS
) {
4263 tp
->rx_opt
.num_sacks
--;
4266 for (; this_sack
> 0; this_sack
--, sp
--)
4270 /* Build the new head SACK, and we're done. */
4271 sp
->start_seq
= seq
;
4272 sp
->end_seq
= end_seq
;
4273 tp
->rx_opt
.num_sacks
++;
4276 /* RCV.NXT advances, some SACKs should be eaten. */
4278 static void tcp_sack_remove(struct tcp_sock
*tp
)
4280 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4281 int num_sacks
= tp
->rx_opt
.num_sacks
;
4284 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4285 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4286 tp
->rx_opt
.num_sacks
= 0;
4290 for (this_sack
= 0; this_sack
< num_sacks
;) {
4291 /* Check if the start of the sack is covered by RCV.NXT. */
4292 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4295 /* RCV.NXT must cover all the block! */
4296 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4298 /* Zap this SACK, by moving forward any other SACKS. */
4299 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4300 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4307 tp
->rx_opt
.num_sacks
= num_sacks
;
4310 /* This one checks to see if we can put data from the
4311 * out_of_order queue into the receive_queue.
4313 static void tcp_ofo_queue(struct sock
*sk
)
4315 struct tcp_sock
*tp
= tcp_sk(sk
);
4316 __u32 dsack_high
= tp
->rcv_nxt
;
4317 struct sk_buff
*skb
;
4319 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4320 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4323 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4324 __u32 dsack
= dsack_high
;
4325 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4326 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4327 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4330 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4331 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4332 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4336 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4337 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4338 TCP_SKB_CB(skb
)->end_seq
);
4340 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4341 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4342 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4343 if (tcp_hdr(skb
)->fin
)
4344 tcp_fin(skb
, sk
, tcp_hdr(skb
));
4348 static int tcp_prune_ofo_queue(struct sock
*sk
);
4349 static int tcp_prune_queue(struct sock
*sk
);
4351 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
4353 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4354 !sk_rmem_schedule(sk
, size
)) {
4356 if (tcp_prune_queue(sk
) < 0)
4359 if (!sk_rmem_schedule(sk
, size
)) {
4360 if (!tcp_prune_ofo_queue(sk
))
4363 if (!sk_rmem_schedule(sk
, size
))
4370 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4372 struct tcphdr
*th
= tcp_hdr(skb
);
4373 struct tcp_sock
*tp
= tcp_sk(sk
);
4376 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4380 __skb_pull(skb
, th
->doff
* 4);
4382 TCP_ECN_accept_cwr(tp
, skb
);
4384 tp
->rx_opt
.dsack
= 0;
4386 /* Queue data for delivery to the user.
4387 * Packets in sequence go to the receive queue.
4388 * Out of sequence packets to the out_of_order_queue.
4390 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4391 if (tcp_receive_window(tp
) == 0)
4394 /* Ok. In sequence. In window. */
4395 if (tp
->ucopy
.task
== current
&&
4396 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4397 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4398 int chunk
= min_t(unsigned int, skb
->len
,
4401 __set_current_state(TASK_RUNNING
);
4404 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4405 tp
->ucopy
.len
-= chunk
;
4406 tp
->copied_seq
+= chunk
;
4407 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4408 tcp_rcv_space_adjust(sk
);
4416 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4419 skb_set_owner_r(skb
, sk
);
4420 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4422 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4424 tcp_event_data_recv(sk
, skb
);
4426 tcp_fin(skb
, sk
, th
);
4428 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4431 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4432 * gap in queue is filled.
4434 if (skb_queue_empty(&tp
->out_of_order_queue
))
4435 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4438 if (tp
->rx_opt
.num_sacks
)
4439 tcp_sack_remove(tp
);
4441 tcp_fast_path_check(sk
);
4445 else if (!sock_flag(sk
, SOCK_DEAD
))
4446 sk
->sk_data_ready(sk
, 0);
4450 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4451 /* A retransmit, 2nd most common case. Force an immediate ack. */
4452 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4453 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4456 tcp_enter_quickack_mode(sk
);
4457 inet_csk_schedule_ack(sk
);
4463 /* Out of window. F.e. zero window probe. */
4464 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4467 tcp_enter_quickack_mode(sk
);
4469 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4470 /* Partial packet, seq < rcv_next < end_seq */
4471 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4472 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4473 TCP_SKB_CB(skb
)->end_seq
);
4475 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4477 /* If window is closed, drop tail of packet. But after
4478 * remembering D-SACK for its head made in previous line.
4480 if (!tcp_receive_window(tp
))
4485 TCP_ECN_check_ce(tp
, skb
);
4487 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4490 /* Disable header prediction. */
4492 inet_csk_schedule_ack(sk
);
4494 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4495 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4497 skb_set_owner_r(skb
, sk
);
4499 if (!skb_peek(&tp
->out_of_order_queue
)) {
4500 /* Initial out of order segment, build 1 SACK. */
4501 if (tcp_is_sack(tp
)) {
4502 tp
->rx_opt
.num_sacks
= 1;
4503 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4504 tp
->selective_acks
[0].end_seq
=
4505 TCP_SKB_CB(skb
)->end_seq
;
4507 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4509 struct sk_buff
*skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4510 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4511 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4513 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4514 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4516 if (!tp
->rx_opt
.num_sacks
||
4517 tp
->selective_acks
[0].end_seq
!= seq
)
4520 /* Common case: data arrive in order after hole. */
4521 tp
->selective_acks
[0].end_seq
= end_seq
;
4525 /* Find place to insert this segment. */
4527 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4529 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4533 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4536 /* Do skb overlap to previous one? */
4537 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4538 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4539 /* All the bits are present. Drop. */
4541 tcp_dsack_set(sk
, seq
, end_seq
);
4544 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4545 /* Partial overlap. */
4546 tcp_dsack_set(sk
, seq
,
4547 TCP_SKB_CB(skb1
)->end_seq
);
4549 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4553 skb1
= skb_queue_prev(
4554 &tp
->out_of_order_queue
,
4559 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4561 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4563 /* And clean segments covered by new one as whole. */
4564 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4565 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4567 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4569 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4570 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4574 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4575 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4576 TCP_SKB_CB(skb1
)->end_seq
);
4581 if (tcp_is_sack(tp
))
4582 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4586 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4587 struct sk_buff_head
*list
)
4589 struct sk_buff
*next
= NULL
;
4591 if (!skb_queue_is_last(list
, skb
))
4592 next
= skb_queue_next(list
, skb
);
4594 __skb_unlink(skb
, list
);
4596 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4601 /* Collapse contiguous sequence of skbs head..tail with
4602 * sequence numbers start..end.
4604 * If tail is NULL, this means until the end of the list.
4606 * Segments with FIN/SYN are not collapsed (only because this
4610 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4611 struct sk_buff
*head
, struct sk_buff
*tail
,
4614 struct sk_buff
*skb
, *n
;
4617 /* First, check that queue is collapsible and find
4618 * the point where collapsing can be useful. */
4622 skb_queue_walk_from_safe(list
, skb
, n
) {
4625 /* No new bits? It is possible on ofo queue. */
4626 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4627 skb
= tcp_collapse_one(sk
, skb
, list
);
4633 /* The first skb to collapse is:
4635 * - bloated or contains data before "start" or
4636 * overlaps to the next one.
4638 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4639 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4640 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4641 end_of_skbs
= false;
4645 if (!skb_queue_is_last(list
, skb
)) {
4646 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4648 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4649 end_of_skbs
= false;
4654 /* Decided to skip this, advance start seq. */
4655 start
= TCP_SKB_CB(skb
)->end_seq
;
4657 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4660 while (before(start
, end
)) {
4661 struct sk_buff
*nskb
;
4662 unsigned int header
= skb_headroom(skb
);
4663 int copy
= SKB_MAX_ORDER(header
, 0);
4665 /* Too big header? This can happen with IPv6. */
4668 if (end
- start
< copy
)
4670 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4674 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4675 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4677 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4679 skb_reserve(nskb
, header
);
4680 memcpy(nskb
->head
, skb
->head
, header
);
4681 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4682 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4683 __skb_queue_before(list
, skb
, nskb
);
4684 skb_set_owner_r(nskb
, sk
);
4686 /* Copy data, releasing collapsed skbs. */
4688 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4689 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4693 size
= min(copy
, size
);
4694 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4696 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4700 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4701 skb
= tcp_collapse_one(sk
, skb
, list
);
4704 tcp_hdr(skb
)->syn
||
4712 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4713 * and tcp_collapse() them until all the queue is collapsed.
4715 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4717 struct tcp_sock
*tp
= tcp_sk(sk
);
4718 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4719 struct sk_buff
*head
;
4725 start
= TCP_SKB_CB(skb
)->seq
;
4726 end
= TCP_SKB_CB(skb
)->end_seq
;
4730 struct sk_buff
*next
= NULL
;
4732 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4733 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4736 /* Segment is terminated when we see gap or when
4737 * we are at the end of all the queue. */
4739 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4740 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4741 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4742 head
, skb
, start
, end
);
4746 /* Start new segment */
4747 start
= TCP_SKB_CB(skb
)->seq
;
4748 end
= TCP_SKB_CB(skb
)->end_seq
;
4750 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4751 start
= TCP_SKB_CB(skb
)->seq
;
4752 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4753 end
= TCP_SKB_CB(skb
)->end_seq
;
4759 * Purge the out-of-order queue.
4760 * Return true if queue was pruned.
4762 static int tcp_prune_ofo_queue(struct sock
*sk
)
4764 struct tcp_sock
*tp
= tcp_sk(sk
);
4767 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4768 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4769 __skb_queue_purge(&tp
->out_of_order_queue
);
4771 /* Reset SACK state. A conforming SACK implementation will
4772 * do the same at a timeout based retransmit. When a connection
4773 * is in a sad state like this, we care only about integrity
4774 * of the connection not performance.
4776 if (tp
->rx_opt
.sack_ok
)
4777 tcp_sack_reset(&tp
->rx_opt
);
4784 /* Reduce allocated memory if we can, trying to get
4785 * the socket within its memory limits again.
4787 * Return less than zero if we should start dropping frames
4788 * until the socket owning process reads some of the data
4789 * to stabilize the situation.
4791 static int tcp_prune_queue(struct sock
*sk
)
4793 struct tcp_sock
*tp
= tcp_sk(sk
);
4795 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4797 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4799 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4800 tcp_clamp_window(sk
);
4801 else if (tcp_memory_pressure
)
4802 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4804 tcp_collapse_ofo_queue(sk
);
4805 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4806 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4807 skb_peek(&sk
->sk_receive_queue
),
4809 tp
->copied_seq
, tp
->rcv_nxt
);
4812 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4815 /* Collapsing did not help, destructive actions follow.
4816 * This must not ever occur. */
4818 tcp_prune_ofo_queue(sk
);
4820 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4823 /* If we are really being abused, tell the caller to silently
4824 * drop receive data on the floor. It will get retransmitted
4825 * and hopefully then we'll have sufficient space.
4827 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4829 /* Massive buffer overcommit. */
4834 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4835 * As additional protections, we do not touch cwnd in retransmission phases,
4836 * and if application hit its sndbuf limit recently.
4838 void tcp_cwnd_application_limited(struct sock
*sk
)
4840 struct tcp_sock
*tp
= tcp_sk(sk
);
4842 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4843 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4844 /* Limited by application or receiver window. */
4845 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4846 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4847 if (win_used
< tp
->snd_cwnd
) {
4848 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4849 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4851 tp
->snd_cwnd_used
= 0;
4853 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4856 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4858 struct tcp_sock
*tp
= tcp_sk(sk
);
4860 /* If the user specified a specific send buffer setting, do
4863 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4866 /* If we are under global TCP memory pressure, do not expand. */
4867 if (tcp_memory_pressure
)
4870 /* If we are under soft global TCP memory pressure, do not expand. */
4871 if (atomic_long_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4874 /* If we filled the congestion window, do not expand. */
4875 if (tp
->packets_out
>= tp
->snd_cwnd
)
4881 /* When incoming ACK allowed to free some skb from write_queue,
4882 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4883 * on the exit from tcp input handler.
4885 * PROBLEM: sndbuf expansion does not work well with largesend.
4887 static void tcp_new_space(struct sock
*sk
)
4889 struct tcp_sock
*tp
= tcp_sk(sk
);
4891 if (tcp_should_expand_sndbuf(sk
)) {
4892 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4893 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
4894 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4895 tp
->reordering
+ 1);
4896 sndmem
*= 2 * demanded
;
4897 if (sndmem
> sk
->sk_sndbuf
)
4898 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4899 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4902 sk
->sk_write_space(sk
);
4905 static void tcp_check_space(struct sock
*sk
)
4907 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4908 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4909 if (sk
->sk_socket
&&
4910 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4915 static inline void tcp_data_snd_check(struct sock
*sk
)
4917 tcp_push_pending_frames(sk
);
4918 tcp_check_space(sk
);
4922 * Check if sending an ack is needed.
4924 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4926 struct tcp_sock
*tp
= tcp_sk(sk
);
4928 /* More than one full frame received... */
4929 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4930 /* ... and right edge of window advances far enough.
4931 * (tcp_recvmsg() will send ACK otherwise). Or...
4933 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4934 /* We ACK each frame or... */
4935 tcp_in_quickack_mode(sk
) ||
4936 /* We have out of order data. */
4937 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4938 /* Then ack it now */
4941 /* Else, send delayed ack. */
4942 tcp_send_delayed_ack(sk
);
4946 static inline void tcp_ack_snd_check(struct sock
*sk
)
4948 if (!inet_csk_ack_scheduled(sk
)) {
4949 /* We sent a data segment already. */
4952 __tcp_ack_snd_check(sk
, 1);
4956 * This routine is only called when we have urgent data
4957 * signaled. Its the 'slow' part of tcp_urg. It could be
4958 * moved inline now as tcp_urg is only called from one
4959 * place. We handle URGent data wrong. We have to - as
4960 * BSD still doesn't use the correction from RFC961.
4961 * For 1003.1g we should support a new option TCP_STDURG to permit
4962 * either form (or just set the sysctl tcp_stdurg).
4965 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4967 struct tcp_sock
*tp
= tcp_sk(sk
);
4968 u32 ptr
= ntohs(th
->urg_ptr
);
4970 if (ptr
&& !sysctl_tcp_stdurg
)
4972 ptr
+= ntohl(th
->seq
);
4974 /* Ignore urgent data that we've already seen and read. */
4975 if (after(tp
->copied_seq
, ptr
))
4978 /* Do not replay urg ptr.
4980 * NOTE: interesting situation not covered by specs.
4981 * Misbehaving sender may send urg ptr, pointing to segment,
4982 * which we already have in ofo queue. We are not able to fetch
4983 * such data and will stay in TCP_URG_NOTYET until will be eaten
4984 * by recvmsg(). Seems, we are not obliged to handle such wicked
4985 * situations. But it is worth to think about possibility of some
4986 * DoSes using some hypothetical application level deadlock.
4988 if (before(ptr
, tp
->rcv_nxt
))
4991 /* Do we already have a newer (or duplicate) urgent pointer? */
4992 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4995 /* Tell the world about our new urgent pointer. */
4998 /* We may be adding urgent data when the last byte read was
4999 * urgent. To do this requires some care. We cannot just ignore
5000 * tp->copied_seq since we would read the last urgent byte again
5001 * as data, nor can we alter copied_seq until this data arrives
5002 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5004 * NOTE. Double Dutch. Rendering to plain English: author of comment
5005 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5006 * and expect that both A and B disappear from stream. This is _wrong_.
5007 * Though this happens in BSD with high probability, this is occasional.
5008 * Any application relying on this is buggy. Note also, that fix "works"
5009 * only in this artificial test. Insert some normal data between A and B and we will
5010 * decline of BSD again. Verdict: it is better to remove to trap
5013 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5014 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5015 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5017 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5018 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5023 tp
->urg_data
= TCP_URG_NOTYET
;
5026 /* Disable header prediction. */
5030 /* This is the 'fast' part of urgent handling. */
5031 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
5033 struct tcp_sock
*tp
= tcp_sk(sk
);
5035 /* Check if we get a new urgent pointer - normally not. */
5037 tcp_check_urg(sk
, th
);
5039 /* Do we wait for any urgent data? - normally not... */
5040 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5041 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5044 /* Is the urgent pointer pointing into this packet? */
5045 if (ptr
< skb
->len
) {
5047 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5049 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5050 if (!sock_flag(sk
, SOCK_DEAD
))
5051 sk
->sk_data_ready(sk
, 0);
5056 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5058 struct tcp_sock
*tp
= tcp_sk(sk
);
5059 int chunk
= skb
->len
- hlen
;
5063 if (skb_csum_unnecessary(skb
))
5064 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
5066 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
5070 tp
->ucopy
.len
-= chunk
;
5071 tp
->copied_seq
+= chunk
;
5072 tcp_rcv_space_adjust(sk
);
5079 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
5080 struct sk_buff
*skb
)
5084 if (sock_owned_by_user(sk
)) {
5086 result
= __tcp_checksum_complete(skb
);
5089 result
= __tcp_checksum_complete(skb
);
5094 static inline int tcp_checksum_complete_user(struct sock
*sk
,
5095 struct sk_buff
*skb
)
5097 return !skb_csum_unnecessary(skb
) &&
5098 __tcp_checksum_complete_user(sk
, skb
);
5101 #ifdef CONFIG_NET_DMA
5102 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5105 struct tcp_sock
*tp
= tcp_sk(sk
);
5106 int chunk
= skb
->len
- hlen
;
5108 int copied_early
= 0;
5110 if (tp
->ucopy
.wakeup
)
5113 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5114 tp
->ucopy
.dma_chan
= dma_find_channel(DMA_MEMCPY
);
5116 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5118 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5120 tp
->ucopy
.iov
, chunk
,
5121 tp
->ucopy
.pinned_list
);
5126 tp
->ucopy
.dma_cookie
= dma_cookie
;
5129 tp
->ucopy
.len
-= chunk
;
5130 tp
->copied_seq
+= chunk
;
5131 tcp_rcv_space_adjust(sk
);
5133 if ((tp
->ucopy
.len
== 0) ||
5134 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5135 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5136 tp
->ucopy
.wakeup
= 1;
5137 sk
->sk_data_ready(sk
, 0);
5139 } else if (chunk
> 0) {
5140 tp
->ucopy
.wakeup
= 1;
5141 sk
->sk_data_ready(sk
, 0);
5144 return copied_early
;
5146 #endif /* CONFIG_NET_DMA */
5148 /* Does PAWS and seqno based validation of an incoming segment, flags will
5149 * play significant role here.
5151 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5152 struct tcphdr
*th
, int syn_inerr
)
5155 struct tcp_sock
*tp
= tcp_sk(sk
);
5157 /* RFC1323: H1. Apply PAWS check first. */
5158 if (tcp_fast_parse_options(skb
, th
, tp
, &hash_location
) &&
5159 tp
->rx_opt
.saw_tstamp
&&
5160 tcp_paws_discard(sk
, skb
)) {
5162 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5163 tcp_send_dupack(sk
, skb
);
5166 /* Reset is accepted even if it did not pass PAWS. */
5169 /* Step 1: check sequence number */
5170 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5171 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5172 * (RST) segments are validated by checking their SEQ-fields."
5173 * And page 69: "If an incoming segment is not acceptable,
5174 * an acknowledgment should be sent in reply (unless the RST
5175 * bit is set, if so drop the segment and return)".
5178 tcp_send_dupack(sk
, skb
);
5182 /* Step 2: check RST bit */
5188 /* ts_recent update must be made after we are sure that the packet
5191 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5193 /* step 3: check security and precedence [ignored] */
5195 /* step 4: Check for a SYN in window. */
5196 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5198 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5199 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5212 * TCP receive function for the ESTABLISHED state.
5214 * It is split into a fast path and a slow path. The fast path is
5216 * - A zero window was announced from us - zero window probing
5217 * is only handled properly in the slow path.
5218 * - Out of order segments arrived.
5219 * - Urgent data is expected.
5220 * - There is no buffer space left
5221 * - Unexpected TCP flags/window values/header lengths are received
5222 * (detected by checking the TCP header against pred_flags)
5223 * - Data is sent in both directions. Fast path only supports pure senders
5224 * or pure receivers (this means either the sequence number or the ack
5225 * value must stay constant)
5226 * - Unexpected TCP option.
5228 * When these conditions are not satisfied it drops into a standard
5229 * receive procedure patterned after RFC793 to handle all cases.
5230 * The first three cases are guaranteed by proper pred_flags setting,
5231 * the rest is checked inline. Fast processing is turned on in
5232 * tcp_data_queue when everything is OK.
5234 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5235 struct tcphdr
*th
, unsigned len
)
5237 struct tcp_sock
*tp
= tcp_sk(sk
);
5241 * Header prediction.
5242 * The code loosely follows the one in the famous
5243 * "30 instruction TCP receive" Van Jacobson mail.
5245 * Van's trick is to deposit buffers into socket queue
5246 * on a device interrupt, to call tcp_recv function
5247 * on the receive process context and checksum and copy
5248 * the buffer to user space. smart...
5250 * Our current scheme is not silly either but we take the
5251 * extra cost of the net_bh soft interrupt processing...
5252 * We do checksum and copy also but from device to kernel.
5255 tp
->rx_opt
.saw_tstamp
= 0;
5257 /* pred_flags is 0xS?10 << 16 + snd_wnd
5258 * if header_prediction is to be made
5259 * 'S' will always be tp->tcp_header_len >> 2
5260 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5261 * turn it off (when there are holes in the receive
5262 * space for instance)
5263 * PSH flag is ignored.
5266 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5267 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5268 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5269 int tcp_header_len
= tp
->tcp_header_len
;
5271 /* Timestamp header prediction: tcp_header_len
5272 * is automatically equal to th->doff*4 due to pred_flags
5276 /* Check timestamp */
5277 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5278 /* No? Slow path! */
5279 if (!tcp_parse_aligned_timestamp(tp
, th
))
5282 /* If PAWS failed, check it more carefully in slow path */
5283 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5286 /* DO NOT update ts_recent here, if checksum fails
5287 * and timestamp was corrupted part, it will result
5288 * in a hung connection since we will drop all
5289 * future packets due to the PAWS test.
5293 if (len
<= tcp_header_len
) {
5294 /* Bulk data transfer: sender */
5295 if (len
== tcp_header_len
) {
5296 /* Predicted packet is in window by definition.
5297 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5298 * Hence, check seq<=rcv_wup reduces to:
5300 if (tcp_header_len
==
5301 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5302 tp
->rcv_nxt
== tp
->rcv_wup
)
5303 tcp_store_ts_recent(tp
);
5305 /* We know that such packets are checksummed
5308 tcp_ack(sk
, skb
, 0);
5310 tcp_data_snd_check(sk
);
5312 } else { /* Header too small */
5313 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5318 int copied_early
= 0;
5320 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5321 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5322 #ifdef CONFIG_NET_DMA
5323 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5328 if (tp
->ucopy
.task
== current
&&
5329 sock_owned_by_user(sk
) && !copied_early
) {
5330 __set_current_state(TASK_RUNNING
);
5332 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5336 /* Predicted packet is in window by definition.
5337 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5338 * Hence, check seq<=rcv_wup reduces to:
5340 if (tcp_header_len
==
5341 (sizeof(struct tcphdr
) +
5342 TCPOLEN_TSTAMP_ALIGNED
) &&
5343 tp
->rcv_nxt
== tp
->rcv_wup
)
5344 tcp_store_ts_recent(tp
);
5346 tcp_rcv_rtt_measure_ts(sk
, skb
);
5348 __skb_pull(skb
, tcp_header_len
);
5349 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5350 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5353 tcp_cleanup_rbuf(sk
, skb
->len
);
5356 if (tcp_checksum_complete_user(sk
, skb
))
5359 /* Predicted packet is in window by definition.
5360 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5361 * Hence, check seq<=rcv_wup reduces to:
5363 if (tcp_header_len
==
5364 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5365 tp
->rcv_nxt
== tp
->rcv_wup
)
5366 tcp_store_ts_recent(tp
);
5368 tcp_rcv_rtt_measure_ts(sk
, skb
);
5370 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5373 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5375 /* Bulk data transfer: receiver */
5376 __skb_pull(skb
, tcp_header_len
);
5377 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
5378 skb_set_owner_r(skb
, sk
);
5379 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5382 tcp_event_data_recv(sk
, skb
);
5384 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5385 /* Well, only one small jumplet in fast path... */
5386 tcp_ack(sk
, skb
, FLAG_DATA
);
5387 tcp_data_snd_check(sk
);
5388 if (!inet_csk_ack_scheduled(sk
))
5392 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5393 __tcp_ack_snd_check(sk
, 0);
5395 #ifdef CONFIG_NET_DMA
5397 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5403 sk
->sk_data_ready(sk
, 0);
5409 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5413 * Standard slow path.
5416 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
5421 if (th
->ack
&& tcp_ack(sk
, skb
, FLAG_SLOWPATH
) < 0)
5424 tcp_rcv_rtt_measure_ts(sk
, skb
);
5426 /* Process urgent data. */
5427 tcp_urg(sk
, skb
, th
);
5429 /* step 7: process the segment text */
5430 tcp_data_queue(sk
, skb
);
5432 tcp_data_snd_check(sk
);
5433 tcp_ack_snd_check(sk
);
5437 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5443 EXPORT_SYMBOL(tcp_rcv_established
);
5445 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5446 struct tcphdr
*th
, unsigned len
)
5449 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5450 struct tcp_sock
*tp
= tcp_sk(sk
);
5451 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
5452 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5454 tcp_parse_options(skb
, &tp
->rx_opt
, &hash_location
, 0);
5458 * "If the state is SYN-SENT then
5459 * first check the ACK bit
5460 * If the ACK bit is set
5461 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5462 * a reset (unless the RST bit is set, if so drop
5463 * the segment and return)"
5465 * We do not send data with SYN, so that RFC-correct
5468 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5469 goto reset_and_undo
;
5471 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5472 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5474 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5475 goto reset_and_undo
;
5478 /* Now ACK is acceptable.
5480 * "If the RST bit is set
5481 * If the ACK was acceptable then signal the user "error:
5482 * connection reset", drop the segment, enter CLOSED state,
5483 * delete TCB, and return."
5492 * "fifth, if neither of the SYN or RST bits is set then
5493 * drop the segment and return."
5499 goto discard_and_undo
;
5502 * "If the SYN bit is on ...
5503 * are acceptable then ...
5504 * (our SYN has been ACKed), change the connection
5505 * state to ESTABLISHED..."
5508 TCP_ECN_rcv_synack(tp
, th
);
5510 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5511 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5513 /* Ok.. it's good. Set up sequence numbers and
5514 * move to established.
5516 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5517 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5519 /* RFC1323: The window in SYN & SYN/ACK segments is
5522 tp
->snd_wnd
= ntohs(th
->window
);
5523 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5525 if (!tp
->rx_opt
.wscale_ok
) {
5526 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5527 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5530 if (tp
->rx_opt
.saw_tstamp
) {
5531 tp
->rx_opt
.tstamp_ok
= 1;
5532 tp
->tcp_header_len
=
5533 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5534 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5535 tcp_store_ts_recent(tp
);
5537 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5540 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5541 tcp_enable_fack(tp
);
5544 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5545 tcp_initialize_rcv_mss(sk
);
5547 /* Remember, tcp_poll() does not lock socket!
5548 * Change state from SYN-SENT only after copied_seq
5549 * is initialized. */
5550 tp
->copied_seq
= tp
->rcv_nxt
;
5553 cvp
->cookie_pair_size
> 0 &&
5554 tp
->rx_opt
.cookie_plus
> 0) {
5555 int cookie_size
= tp
->rx_opt
.cookie_plus
5556 - TCPOLEN_COOKIE_BASE
;
5557 int cookie_pair_size
= cookie_size
5558 + cvp
->cookie_desired
;
5560 /* A cookie extension option was sent and returned.
5561 * Note that each incoming SYNACK replaces the
5562 * Responder cookie. The initial exchange is most
5563 * fragile, as protection against spoofing relies
5564 * entirely upon the sequence and timestamp (above).
5565 * This replacement strategy allows the correct pair to
5566 * pass through, while any others will be filtered via
5567 * Responder verification later.
5569 if (sizeof(cvp
->cookie_pair
) >= cookie_pair_size
) {
5570 memcpy(&cvp
->cookie_pair
[cvp
->cookie_desired
],
5571 hash_location
, cookie_size
);
5572 cvp
->cookie_pair_size
= cookie_pair_size
;
5577 tcp_set_state(sk
, TCP_ESTABLISHED
);
5579 security_inet_conn_established(sk
, skb
);
5581 /* Make sure socket is routed, for correct metrics. */
5582 icsk
->icsk_af_ops
->rebuild_header(sk
);
5584 tcp_init_metrics(sk
);
5586 tcp_init_congestion_control(sk
);
5588 /* Prevent spurious tcp_cwnd_restart() on first data
5591 tp
->lsndtime
= tcp_time_stamp
;
5593 tcp_init_buffer_space(sk
);
5595 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5596 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5598 if (!tp
->rx_opt
.snd_wscale
)
5599 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5603 if (!sock_flag(sk
, SOCK_DEAD
)) {
5604 sk
->sk_state_change(sk
);
5605 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5608 if (sk
->sk_write_pending
||
5609 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5610 icsk
->icsk_ack
.pingpong
) {
5611 /* Save one ACK. Data will be ready after
5612 * several ticks, if write_pending is set.
5614 * It may be deleted, but with this feature tcpdumps
5615 * look so _wonderfully_ clever, that I was not able
5616 * to stand against the temptation 8) --ANK
5618 inet_csk_schedule_ack(sk
);
5619 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5620 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5621 tcp_incr_quickack(sk
);
5622 tcp_enter_quickack_mode(sk
);
5623 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5624 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5635 /* No ACK in the segment */
5639 * "If the RST bit is set
5641 * Otherwise (no ACK) drop the segment and return."
5644 goto discard_and_undo
;
5648 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5649 tcp_paws_reject(&tp
->rx_opt
, 0))
5650 goto discard_and_undo
;
5653 /* We see SYN without ACK. It is attempt of
5654 * simultaneous connect with crossed SYNs.
5655 * Particularly, it can be connect to self.
5657 tcp_set_state(sk
, TCP_SYN_RECV
);
5659 if (tp
->rx_opt
.saw_tstamp
) {
5660 tp
->rx_opt
.tstamp_ok
= 1;
5661 tcp_store_ts_recent(tp
);
5662 tp
->tcp_header_len
=
5663 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5665 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5668 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5669 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5671 /* RFC1323: The window in SYN & SYN/ACK segments is
5674 tp
->snd_wnd
= ntohs(th
->window
);
5675 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5676 tp
->max_window
= tp
->snd_wnd
;
5678 TCP_ECN_rcv_syn(tp
, th
);
5681 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5682 tcp_initialize_rcv_mss(sk
);
5684 tcp_send_synack(sk
);
5686 /* Note, we could accept data and URG from this segment.
5687 * There are no obstacles to make this.
5689 * However, if we ignore data in ACKless segments sometimes,
5690 * we have no reasons to accept it sometimes.
5691 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5692 * is not flawless. So, discard packet for sanity.
5693 * Uncomment this return to process the data.
5700 /* "fifth, if neither of the SYN or RST bits is set then
5701 * drop the segment and return."
5705 tcp_clear_options(&tp
->rx_opt
);
5706 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5710 tcp_clear_options(&tp
->rx_opt
);
5711 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5716 * This function implements the receiving procedure of RFC 793 for
5717 * all states except ESTABLISHED and TIME_WAIT.
5718 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5719 * address independent.
5722 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5723 struct tcphdr
*th
, unsigned len
)
5725 struct tcp_sock
*tp
= tcp_sk(sk
);
5726 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5730 tp
->rx_opt
.saw_tstamp
= 0;
5732 switch (sk
->sk_state
) {
5744 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5747 /* Now we have several options: In theory there is
5748 * nothing else in the frame. KA9Q has an option to
5749 * send data with the syn, BSD accepts data with the
5750 * syn up to the [to be] advertised window and
5751 * Solaris 2.1 gives you a protocol error. For now
5752 * we just ignore it, that fits the spec precisely
5753 * and avoids incompatibilities. It would be nice in
5754 * future to drop through and process the data.
5756 * Now that TTCP is starting to be used we ought to
5758 * But, this leaves one open to an easy denial of
5759 * service attack, and SYN cookies can't defend
5760 * against this problem. So, we drop the data
5761 * in the interest of security over speed unless
5762 * it's still in use.
5770 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5774 /* Do step6 onward by hand. */
5775 tcp_urg(sk
, skb
, th
);
5777 tcp_data_snd_check(sk
);
5781 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5785 /* step 5: check the ACK field */
5787 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
) > 0;
5789 switch (sk
->sk_state
) {
5792 tp
->copied_seq
= tp
->rcv_nxt
;
5794 tcp_set_state(sk
, TCP_ESTABLISHED
);
5795 sk
->sk_state_change(sk
);
5797 /* Note, that this wakeup is only for marginal
5798 * crossed SYN case. Passively open sockets
5799 * are not waked up, because sk->sk_sleep ==
5800 * NULL and sk->sk_socket == NULL.
5804 SOCK_WAKE_IO
, POLL_OUT
);
5806 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5807 tp
->snd_wnd
= ntohs(th
->window
) <<
5808 tp
->rx_opt
.snd_wscale
;
5809 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5811 /* tcp_ack considers this ACK as duplicate
5812 * and does not calculate rtt.
5815 tcp_ack_update_rtt(sk
, 0, 0);
5817 if (tp
->rx_opt
.tstamp_ok
)
5818 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5820 /* Make sure socket is routed, for
5823 icsk
->icsk_af_ops
->rebuild_header(sk
);
5825 tcp_init_metrics(sk
);
5827 tcp_init_congestion_control(sk
);
5829 /* Prevent spurious tcp_cwnd_restart() on
5830 * first data packet.
5832 tp
->lsndtime
= tcp_time_stamp
;
5835 tcp_initialize_rcv_mss(sk
);
5836 tcp_init_buffer_space(sk
);
5837 tcp_fast_path_on(tp
);
5844 if (tp
->snd_una
== tp
->write_seq
) {
5845 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5846 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5847 dst_confirm(__sk_dst_get(sk
));
5849 if (!sock_flag(sk
, SOCK_DEAD
))
5850 /* Wake up lingering close() */
5851 sk
->sk_state_change(sk
);
5855 if (tp
->linger2
< 0 ||
5856 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5857 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5859 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5863 tmo
= tcp_fin_time(sk
);
5864 if (tmo
> TCP_TIMEWAIT_LEN
) {
5865 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5866 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5867 /* Bad case. We could lose such FIN otherwise.
5868 * It is not a big problem, but it looks confusing
5869 * and not so rare event. We still can lose it now,
5870 * if it spins in bh_lock_sock(), but it is really
5873 inet_csk_reset_keepalive_timer(sk
, tmo
);
5875 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5883 if (tp
->snd_una
== tp
->write_seq
) {
5884 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5890 if (tp
->snd_una
== tp
->write_seq
) {
5891 tcp_update_metrics(sk
);
5900 /* step 6: check the URG bit */
5901 tcp_urg(sk
, skb
, th
);
5903 /* step 7: process the segment text */
5904 switch (sk
->sk_state
) {
5905 case TCP_CLOSE_WAIT
:
5908 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5912 /* RFC 793 says to queue data in these states,
5913 * RFC 1122 says we MUST send a reset.
5914 * BSD 4.4 also does reset.
5916 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5917 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5918 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5919 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5925 case TCP_ESTABLISHED
:
5926 tcp_data_queue(sk
, skb
);
5931 /* tcp_data could move socket to TIME-WAIT */
5932 if (sk
->sk_state
!= TCP_CLOSE
) {
5933 tcp_data_snd_check(sk
);
5934 tcp_ack_snd_check(sk
);
5943 EXPORT_SYMBOL(tcp_rcv_state_process
);