2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #include <linux/module.h>
66 #include <linux/sysctl.h>
67 #include <linux/kernel.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
75 int sysctl_tcp_timestamps __read_mostly
= 1;
76 int sysctl_tcp_window_scaling __read_mostly
= 1;
77 int sysctl_tcp_sack __read_mostly
= 1;
78 int sysctl_tcp_fack __read_mostly
= 1;
79 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
80 int sysctl_tcp_ecn __read_mostly
;
81 int sysctl_tcp_dsack __read_mostly
= 1;
82 int sysctl_tcp_app_win __read_mostly
= 31;
83 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
85 int sysctl_tcp_stdurg __read_mostly
;
86 int sysctl_tcp_rfc1337 __read_mostly
;
87 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
88 int sysctl_tcp_frto __read_mostly
= 2;
89 int sysctl_tcp_frto_response __read_mostly
;
90 int sysctl_tcp_nometrics_save __read_mostly
;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
93 int sysctl_tcp_abc __read_mostly
;
95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
101 #define FLAG_ECE 0x40 /* ECE in this ACK */
102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
107 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
108 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
110 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
111 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
112 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
113 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
114 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
116 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
117 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
119 /* Adapt the MSS value used to make delayed ack decision to the
122 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
124 struct inet_connection_sock
*icsk
= inet_csk(sk
);
125 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
128 icsk
->icsk_ack
.last_seg_size
= 0;
130 /* skb->len may jitter because of SACKs, even if peer
131 * sends good full-sized frames.
133 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
134 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
135 icsk
->icsk_ack
.rcv_mss
= len
;
137 /* Otherwise, we make more careful check taking into account,
138 * that SACKs block is variable.
140 * "len" is invariant segment length, including TCP header.
142 len
+= skb
->data
- skb_transport_header(skb
);
143 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
144 /* If PSH is not set, packet should be
145 * full sized, provided peer TCP is not badly broken.
146 * This observation (if it is correct 8)) allows
147 * to handle super-low mtu links fairly.
149 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
150 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
151 /* Subtract also invariant (if peer is RFC compliant),
152 * tcp header plus fixed timestamp option length.
153 * Resulting "len" is MSS free of SACK jitter.
155 len
-= tcp_sk(sk
)->tcp_header_len
;
156 icsk
->icsk_ack
.last_seg_size
= len
;
158 icsk
->icsk_ack
.rcv_mss
= len
;
162 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
163 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
164 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
168 static void tcp_incr_quickack(struct sock
*sk
)
170 struct inet_connection_sock
*icsk
= inet_csk(sk
);
171 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
175 if (quickacks
> icsk
->icsk_ack
.quick
)
176 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
179 void tcp_enter_quickack_mode(struct sock
*sk
)
181 struct inet_connection_sock
*icsk
= inet_csk(sk
);
182 tcp_incr_quickack(sk
);
183 icsk
->icsk_ack
.pingpong
= 0;
184 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
187 /* Send ACKs quickly, if "quick" count is not exhausted
188 * and the session is not interactive.
191 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
193 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
194 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
197 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
199 if (tp
->ecn_flags
& TCP_ECN_OK
)
200 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
203 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
205 if (tcp_hdr(skb
)->cwr
)
206 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
209 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
211 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
214 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
216 if (tp
->ecn_flags
& TCP_ECN_OK
) {
217 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
218 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
219 /* Funny extension: if ECT is not set on a segment,
220 * it is surely retransmit. It is not in ECN RFC,
221 * but Linux follows this rule. */
222 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
223 tcp_enter_quickack_mode((struct sock
*)tp
);
227 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
229 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
230 tp
->ecn_flags
&= ~TCP_ECN_OK
;
233 static inline void TCP_ECN_rcv_syn(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 int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
241 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
246 /* Buffer size and advertised window tuning.
248 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
251 static void tcp_fixup_sndbuf(struct sock
*sk
)
253 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
254 sizeof(struct sk_buff
);
256 if (sk
->sk_sndbuf
< 3 * sndmem
)
257 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
260 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
262 * All tcp_full_space() is split to two parts: "network" buffer, allocated
263 * forward and advertised in receiver window (tp->rcv_wnd) and
264 * "application buffer", required to isolate scheduling/application
265 * latencies from network.
266 * window_clamp is maximal advertised window. It can be less than
267 * tcp_full_space(), in this case tcp_full_space() - window_clamp
268 * is reserved for "application" buffer. The less window_clamp is
269 * the smoother our behaviour from viewpoint of network, but the lower
270 * throughput and the higher sensitivity of the connection to losses. 8)
272 * rcv_ssthresh is more strict window_clamp used at "slow start"
273 * phase to predict further behaviour of this connection.
274 * It is used for two goals:
275 * - to enforce header prediction at sender, even when application
276 * requires some significant "application buffer". It is check #1.
277 * - to prevent pruning of receive queue because of misprediction
278 * of receiver window. Check #2.
280 * The scheme does not work when sender sends good segments opening
281 * window and then starts to feed us spaghetti. But it should work
282 * in common situations. Otherwise, we have to rely on queue collapsing.
285 /* Slow part of check#2. */
286 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
288 struct tcp_sock
*tp
= tcp_sk(sk
);
290 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
291 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
293 while (tp
->rcv_ssthresh
<= window
) {
294 if (truesize
<= skb
->len
)
295 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
303 static void tcp_grow_window(struct sock
*sk
, struct sk_buff
*skb
)
305 struct tcp_sock
*tp
= tcp_sk(sk
);
308 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
309 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
310 !tcp_memory_pressure
) {
313 /* Check #2. Increase window, if skb with such overhead
314 * will fit to rcvbuf in future.
316 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
317 incr
= 2 * tp
->advmss
;
319 incr
= __tcp_grow_window(sk
, skb
);
322 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
324 inet_csk(sk
)->icsk_ack
.quick
|= 1;
329 /* 3. Tuning rcvbuf, when connection enters established state. */
331 static void tcp_fixup_rcvbuf(struct sock
*sk
)
333 struct tcp_sock
*tp
= tcp_sk(sk
);
334 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
336 /* Try to select rcvbuf so that 4 mss-sized segments
337 * will fit to window and corresponding skbs will fit to our rcvbuf.
338 * (was 3; 4 is minimum to allow fast retransmit to work.)
340 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
342 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
343 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
346 /* 4. Try to fixup all. It is made immediately after connection enters
349 static void tcp_init_buffer_space(struct sock
*sk
)
351 struct tcp_sock
*tp
= tcp_sk(sk
);
354 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
355 tcp_fixup_rcvbuf(sk
);
356 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
357 tcp_fixup_sndbuf(sk
);
359 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
361 maxwin
= tcp_full_space(sk
);
363 if (tp
->window_clamp
>= maxwin
) {
364 tp
->window_clamp
= maxwin
;
366 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
367 tp
->window_clamp
= max(maxwin
-
368 (maxwin
>> sysctl_tcp_app_win
),
372 /* Force reservation of one segment. */
373 if (sysctl_tcp_app_win
&&
374 tp
->window_clamp
> 2 * tp
->advmss
&&
375 tp
->window_clamp
+ tp
->advmss
> maxwin
)
376 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
378 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
379 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
382 /* 5. Recalculate window clamp after socket hit its memory bounds. */
383 static void tcp_clamp_window(struct sock
*sk
)
385 struct tcp_sock
*tp
= tcp_sk(sk
);
386 struct inet_connection_sock
*icsk
= inet_csk(sk
);
388 icsk
->icsk_ack
.quick
= 0;
390 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
391 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
392 !tcp_memory_pressure
&&
393 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
394 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
397 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
398 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
401 /* Initialize RCV_MSS value.
402 * RCV_MSS is an our guess about MSS used by the peer.
403 * We haven't any direct information about the MSS.
404 * It's better to underestimate the RCV_MSS rather than overestimate.
405 * Overestimations make us ACKing less frequently than needed.
406 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
408 void tcp_initialize_rcv_mss(struct sock
*sk
)
410 struct tcp_sock
*tp
= tcp_sk(sk
);
411 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
413 hint
= min(hint
, tp
->rcv_wnd
/ 2);
414 hint
= min(hint
, TCP_MIN_RCVMSS
);
415 hint
= max(hint
, TCP_MIN_MSS
);
417 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
420 /* Receiver "autotuning" code.
422 * The algorithm for RTT estimation w/o timestamps is based on
423 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
424 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
426 * More detail on this code can be found at
427 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
428 * though this reference is out of date. A new paper
431 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
433 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
439 if (new_sample
!= 0) {
440 /* If we sample in larger samples in the non-timestamp
441 * case, we could grossly overestimate the RTT especially
442 * with chatty applications or bulk transfer apps which
443 * are stalled on filesystem I/O.
445 * Also, since we are only going for a minimum in the
446 * non-timestamp case, we do not smooth things out
447 * else with timestamps disabled convergence takes too
451 m
-= (new_sample
>> 3);
453 } else if (m
< new_sample
)
456 /* No previous measure. */
460 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
461 tp
->rcv_rtt_est
.rtt
= new_sample
;
464 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
466 if (tp
->rcv_rtt_est
.time
== 0)
468 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
470 tcp_rcv_rtt_update(tp
, jiffies
- tp
->rcv_rtt_est
.time
, 1);
473 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
474 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
477 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
478 const struct sk_buff
*skb
)
480 struct tcp_sock
*tp
= tcp_sk(sk
);
481 if (tp
->rx_opt
.rcv_tsecr
&&
482 (TCP_SKB_CB(skb
)->end_seq
-
483 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
484 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
488 * This function should be called every time data is copied to user space.
489 * It calculates the appropriate TCP receive buffer space.
491 void tcp_rcv_space_adjust(struct sock
*sk
)
493 struct tcp_sock
*tp
= tcp_sk(sk
);
497 if (tp
->rcvq_space
.time
== 0)
500 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
501 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
504 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
506 space
= max(tp
->rcvq_space
.space
, space
);
508 if (tp
->rcvq_space
.space
!= space
) {
511 tp
->rcvq_space
.space
= space
;
513 if (sysctl_tcp_moderate_rcvbuf
&&
514 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
515 int new_clamp
= space
;
517 /* Receive space grows, normalize in order to
518 * take into account packet headers and sk_buff
519 * structure overhead.
524 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
525 16 + sizeof(struct sk_buff
));
526 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
529 space
= min(space
, sysctl_tcp_rmem
[2]);
530 if (space
> sk
->sk_rcvbuf
) {
531 sk
->sk_rcvbuf
= space
;
533 /* Make the window clamp follow along. */
534 tp
->window_clamp
= new_clamp
;
540 tp
->rcvq_space
.seq
= tp
->copied_seq
;
541 tp
->rcvq_space
.time
= tcp_time_stamp
;
544 /* There is something which you must keep in mind when you analyze the
545 * behavior of the tp->ato delayed ack timeout interval. When a
546 * connection starts up, we want to ack as quickly as possible. The
547 * problem is that "good" TCP's do slow start at the beginning of data
548 * transmission. The means that until we send the first few ACK's the
549 * sender will sit on his end and only queue most of his data, because
550 * he can only send snd_cwnd unacked packets at any given time. For
551 * each ACK we send, he increments snd_cwnd and transmits more of his
554 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
556 struct tcp_sock
*tp
= tcp_sk(sk
);
557 struct inet_connection_sock
*icsk
= inet_csk(sk
);
560 inet_csk_schedule_ack(sk
);
562 tcp_measure_rcv_mss(sk
, skb
);
564 tcp_rcv_rtt_measure(tp
);
566 now
= tcp_time_stamp
;
568 if (!icsk
->icsk_ack
.ato
) {
569 /* The _first_ data packet received, initialize
570 * delayed ACK engine.
572 tcp_incr_quickack(sk
);
573 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
575 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
577 if (m
<= TCP_ATO_MIN
/ 2) {
578 /* The fastest case is the first. */
579 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
580 } else if (m
< icsk
->icsk_ack
.ato
) {
581 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
582 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
583 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
584 } else if (m
> icsk
->icsk_rto
) {
585 /* Too long gap. Apparently sender failed to
586 * restart window, so that we send ACKs quickly.
588 tcp_incr_quickack(sk
);
592 icsk
->icsk_ack
.lrcvtime
= now
;
594 TCP_ECN_check_ce(tp
, skb
);
597 tcp_grow_window(sk
, skb
);
600 static u32
tcp_rto_min(struct sock
*sk
)
602 struct dst_entry
*dst
= __sk_dst_get(sk
);
603 u32 rto_min
= TCP_RTO_MIN
;
605 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
606 rto_min
= dst_metric_rtt(dst
, RTAX_RTO_MIN
);
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
= (tp
->srtt
>> 3) + tp
->rttvar
;
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.
709 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
710 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
713 /* Save metrics learned by this TCP session.
714 This function is called only, when TCP finishes successfully
715 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
717 void tcp_update_metrics(struct sock
*sk
)
719 struct tcp_sock
*tp
= tcp_sk(sk
);
720 struct dst_entry
*dst
= __sk_dst_get(sk
);
722 if (sysctl_tcp_nometrics_save
)
727 if (dst
&& (dst
->flags
& DST_HOST
)) {
728 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
732 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
733 /* This session failed to estimate rtt. Why?
734 * Probably, no packets returned in time.
737 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
738 dst
->metrics
[RTAX_RTT
- 1] = 0;
742 rtt
= dst_metric_rtt(dst
, RTAX_RTT
);
745 /* If newly calculated rtt larger than stored one,
746 * store new one. Otherwise, use EWMA. Remember,
747 * rtt overestimation is always better than underestimation.
749 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
751 set_dst_metric_rtt(dst
, RTAX_RTT
, tp
->srtt
);
753 set_dst_metric_rtt(dst
, RTAX_RTT
, rtt
- (m
>> 3));
756 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
761 /* Scale deviation to rttvar fixed point */
766 var
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
770 var
-= (var
- m
) >> 2;
772 set_dst_metric_rtt(dst
, RTAX_RTTVAR
, var
);
775 if (tp
->snd_ssthresh
>= 0xFFFF) {
776 /* Slow start still did not finish. */
777 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
778 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
779 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
780 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
781 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
782 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
783 dst
->metrics
[RTAX_CWND
- 1] = tp
->snd_cwnd
;
784 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
785 icsk
->icsk_ca_state
== TCP_CA_Open
) {
786 /* Cong. avoidance phase, cwnd is reliable. */
787 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
788 dst
->metrics
[RTAX_SSTHRESH
-1] =
789 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
790 if (!dst_metric_locked(dst
, RTAX_CWND
))
791 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_cwnd
) >> 1;
793 /* Else slow start did not finish, cwnd is non-sense,
794 ssthresh may be also invalid.
796 if (!dst_metric_locked(dst
, RTAX_CWND
))
797 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_ssthresh
) >> 1;
798 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
799 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
800 tp
->snd_ssthresh
> dst_metric(dst
, RTAX_SSTHRESH
))
801 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
804 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
805 if (dst_metric(dst
, RTAX_REORDERING
) < tp
->reordering
&&
806 tp
->reordering
!= sysctl_tcp_reordering
)
807 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
812 /* Numbers are taken from RFC3390.
814 * John Heffner states:
816 * The RFC specifies a window of no more than 4380 bytes
817 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
818 * is a bit misleading because they use a clamp at 4380 bytes
819 * rather than use a multiplier in the relevant range.
821 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
823 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
826 if (tp
->mss_cache
> 1460)
829 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
831 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
834 /* Set slow start threshold and cwnd not falling to slow start */
835 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
837 struct tcp_sock
*tp
= tcp_sk(sk
);
838 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
840 tp
->prior_ssthresh
= 0;
842 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
845 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
846 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
847 tcp_packets_in_flight(tp
) + 1U);
848 tp
->snd_cwnd_cnt
= 0;
849 tp
->high_seq
= tp
->snd_nxt
;
850 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
851 TCP_ECN_queue_cwr(tp
);
853 tcp_set_ca_state(sk
, TCP_CA_CWR
);
858 * Packet counting of FACK is based on in-order assumptions, therefore TCP
859 * disables it when reordering is detected
861 static void tcp_disable_fack(struct tcp_sock
*tp
)
863 /* RFC3517 uses different metric in lost marker => reset on change */
865 tp
->lost_skb_hint
= NULL
;
866 tp
->rx_opt
.sack_ok
&= ~2;
869 /* Take a notice that peer is sending D-SACKs */
870 static void tcp_dsack_seen(struct tcp_sock
*tp
)
872 tp
->rx_opt
.sack_ok
|= 4;
875 /* Initialize metrics on socket. */
877 static void tcp_init_metrics(struct sock
*sk
)
879 struct tcp_sock
*tp
= tcp_sk(sk
);
880 struct dst_entry
*dst
= __sk_dst_get(sk
);
887 if (dst_metric_locked(dst
, RTAX_CWND
))
888 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
889 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
890 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
891 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
892 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
894 if (dst_metric(dst
, RTAX_REORDERING
) &&
895 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
896 tcp_disable_fack(tp
);
897 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
900 if (dst_metric(dst
, RTAX_RTT
) == 0)
903 if (!tp
->srtt
&& dst_metric_rtt(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
906 /* Initial rtt is determined from SYN,SYN-ACK.
907 * The segment is small and rtt may appear much
908 * less than real one. Use per-dst memory
909 * to make it more realistic.
911 * A bit of theory. RTT is time passed after "normal" sized packet
912 * is sent until it is ACKed. In normal circumstances sending small
913 * packets force peer to delay ACKs and calculation is correct too.
914 * The algorithm is adaptive and, provided we follow specs, it
915 * NEVER underestimate RTT. BUT! If peer tries to make some clever
916 * tricks sort of "quick acks" for time long enough to decrease RTT
917 * to low value, and then abruptly stops to do it and starts to delay
918 * ACKs, wait for troubles.
920 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
921 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
922 tp
->rtt_seq
= tp
->snd_nxt
;
924 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
925 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
926 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
929 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
931 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
932 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
936 /* Play conservative. If timestamps are not
937 * supported, TCP will fail to recalculate correct
938 * rtt, if initial rto is too small. FORGET ALL AND RESET!
940 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
942 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
943 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
947 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
950 struct tcp_sock
*tp
= tcp_sk(sk
);
951 if (metric
> tp
->reordering
) {
954 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
956 /* This exciting event is worth to be remembered. 8) */
958 mib_idx
= LINUX_MIB_TCPTSREORDER
;
959 else if (tcp_is_reno(tp
))
960 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
961 else if (tcp_is_fack(tp
))
962 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
964 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
966 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
967 #if FASTRETRANS_DEBUG > 1
968 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
969 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
973 tp
->undo_marker
? tp
->undo_retrans
: 0);
975 tcp_disable_fack(tp
);
979 /* This must be called before lost_out is incremented */
980 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
982 if ((tp
->retransmit_skb_hint
== NULL
) ||
983 before(TCP_SKB_CB(skb
)->seq
,
984 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
985 tp
->retransmit_skb_hint
= skb
;
988 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
989 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
992 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
994 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
995 tcp_verify_retransmit_hint(tp
, skb
);
997 tp
->lost_out
+= tcp_skb_pcount(skb
);
998 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1002 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
1003 struct sk_buff
*skb
)
1005 tcp_verify_retransmit_hint(tp
, skb
);
1007 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1008 tp
->lost_out
+= tcp_skb_pcount(skb
);
1009 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1013 /* This procedure tags the retransmission queue when SACKs arrive.
1015 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1016 * Packets in queue with these bits set are counted in variables
1017 * sacked_out, retrans_out and lost_out, correspondingly.
1019 * Valid combinations are:
1020 * Tag InFlight Description
1021 * 0 1 - orig segment is in flight.
1022 * S 0 - nothing flies, orig reached receiver.
1023 * L 0 - nothing flies, orig lost by net.
1024 * R 2 - both orig and retransmit are in flight.
1025 * L|R 1 - orig is lost, retransmit is in flight.
1026 * S|R 1 - orig reached receiver, retrans is still in flight.
1027 * (L|S|R is logically valid, it could occur when L|R is sacked,
1028 * but it is equivalent to plain S and code short-curcuits it to S.
1029 * L|S is logically invalid, it would mean -1 packet in flight 8))
1031 * These 6 states form finite state machine, controlled by the following events:
1032 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1033 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1034 * 3. Loss detection event of one of three flavors:
1035 * A. Scoreboard estimator decided the packet is lost.
1036 * A'. Reno "three dupacks" marks head of queue lost.
1037 * A''. Its FACK modfication, head until snd.fack is lost.
1038 * B. SACK arrives sacking data transmitted after never retransmitted
1039 * hole was sent out.
1040 * C. SACK arrives sacking SND.NXT at the moment, when the
1041 * segment was retransmitted.
1042 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1044 * It is pleasant to note, that state diagram turns out to be commutative,
1045 * so that we are allowed not to be bothered by order of our actions,
1046 * when multiple events arrive simultaneously. (see the function below).
1048 * Reordering detection.
1049 * --------------------
1050 * Reordering metric is maximal distance, which a packet can be displaced
1051 * in packet stream. With SACKs we can estimate it:
1053 * 1. SACK fills old hole and the corresponding segment was not
1054 * ever retransmitted -> reordering. Alas, we cannot use it
1055 * when segment was retransmitted.
1056 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1057 * for retransmitted and already SACKed segment -> reordering..
1058 * Both of these heuristics are not used in Loss state, when we cannot
1059 * account for retransmits accurately.
1061 * SACK block validation.
1062 * ----------------------
1064 * SACK block range validation checks that the received SACK block fits to
1065 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1066 * Note that SND.UNA is not included to the range though being valid because
1067 * it means that the receiver is rather inconsistent with itself reporting
1068 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1069 * perfectly valid, however, in light of RFC2018 which explicitly states
1070 * that "SACK block MUST reflect the newest segment. Even if the newest
1071 * segment is going to be discarded ...", not that it looks very clever
1072 * in case of head skb. Due to potentional receiver driven attacks, we
1073 * choose to avoid immediate execution of a walk in write queue due to
1074 * reneging and defer head skb's loss recovery to standard loss recovery
1075 * procedure that will eventually trigger (nothing forbids us doing this).
1077 * Implements also blockage to start_seq wrap-around. Problem lies in the
1078 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1079 * there's no guarantee that it will be before snd_nxt (n). The problem
1080 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1083 * <- outs wnd -> <- wrapzone ->
1084 * u e n u_w e_w s n_w
1086 * |<------------+------+----- TCP seqno space --------------+---------->|
1087 * ...-- <2^31 ->| |<--------...
1088 * ...---- >2^31 ------>| |<--------...
1090 * Current code wouldn't be vulnerable but it's better still to discard such
1091 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1092 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1093 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1094 * equal to the ideal case (infinite seqno space without wrap caused issues).
1096 * With D-SACK the lower bound is extended to cover sequence space below
1097 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1098 * again, D-SACK block must not to go across snd_una (for the same reason as
1099 * for the normal SACK blocks, explained above). But there all simplicity
1100 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1101 * fully below undo_marker they do not affect behavior in anyway and can
1102 * therefore be safely ignored. In rare cases (which are more or less
1103 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1104 * fragmentation and packet reordering past skb's retransmission. To consider
1105 * them correctly, the acceptable range must be extended even more though
1106 * the exact amount is rather hard to quantify. However, tp->max_window can
1107 * be used as an exaggerated estimate.
1109 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1110 u32 start_seq
, u32 end_seq
)
1112 /* Too far in future, or reversed (interpretation is ambiguous) */
1113 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1116 /* Nasty start_seq wrap-around check (see comments above) */
1117 if (!before(start_seq
, tp
->snd_nxt
))
1120 /* In outstanding window? ...This is valid exit for D-SACKs too.
1121 * start_seq == snd_una is non-sensical (see comments above)
1123 if (after(start_seq
, tp
->snd_una
))
1126 if (!is_dsack
|| !tp
->undo_marker
)
1129 /* ...Then it's D-SACK, and must reside below snd_una completely */
1130 if (!after(end_seq
, tp
->snd_una
))
1133 if (!before(start_seq
, tp
->undo_marker
))
1137 if (!after(end_seq
, tp
->undo_marker
))
1140 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1141 * start_seq < undo_marker and end_seq >= undo_marker.
1143 return !before(start_seq
, end_seq
- tp
->max_window
);
1146 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1147 * Event "C". Later note: FACK people cheated me again 8), we have to account
1148 * for reordering! Ugly, but should help.
1150 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1151 * less than what is now known to be received by the other end (derived from
1152 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1153 * retransmitted skbs to avoid some costly processing per ACKs.
1155 static void tcp_mark_lost_retrans(struct sock
*sk
)
1157 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1158 struct tcp_sock
*tp
= tcp_sk(sk
);
1159 struct sk_buff
*skb
;
1161 u32 new_low_seq
= tp
->snd_nxt
;
1162 u32 received_upto
= tcp_highest_sack_seq(tp
);
1164 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1165 !after(received_upto
, tp
->lost_retrans_low
) ||
1166 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1169 tcp_for_write_queue(skb
, sk
) {
1170 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1172 if (skb
== tcp_send_head(sk
))
1174 if (cnt
== tp
->retrans_out
)
1176 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1179 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1182 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1183 * constraint here (see above) but figuring out that at
1184 * least tp->reordering SACK blocks reside between ack_seq
1185 * and received_upto is not easy task to do cheaply with
1186 * the available datastructures.
1188 * Whether FACK should check here for tp->reordering segs
1189 * in-between one could argue for either way (it would be
1190 * rather simple to implement as we could count fack_count
1191 * during the walk and do tp->fackets_out - fack_count).
1193 if (after(received_upto
, ack_seq
)) {
1194 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1195 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1197 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1198 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1200 if (before(ack_seq
, new_low_seq
))
1201 new_low_seq
= ack_seq
;
1202 cnt
+= tcp_skb_pcount(skb
);
1206 if (tp
->retrans_out
)
1207 tp
->lost_retrans_low
= new_low_seq
;
1210 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1211 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1214 struct tcp_sock
*tp
= tcp_sk(sk
);
1215 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1216 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1219 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1222 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1223 } else if (num_sacks
> 1) {
1224 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1225 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1227 if (!after(end_seq_0
, end_seq_1
) &&
1228 !before(start_seq_0
, start_seq_1
)) {
1231 NET_INC_STATS_BH(sock_net(sk
),
1232 LINUX_MIB_TCPDSACKOFORECV
);
1236 /* D-SACK for already forgotten data... Do dumb counting. */
1238 !after(end_seq_0
, prior_snd_una
) &&
1239 after(end_seq_0
, tp
->undo_marker
))
1245 struct tcp_sacktag_state
{
1251 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1252 * the incoming SACK may not exactly match but we can find smaller MSS
1253 * aligned portion of it that matches. Therefore we might need to fragment
1254 * which may fail and creates some hassle (caller must handle error case
1257 * FIXME: this could be merged to shift decision code
1259 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1260 u32 start_seq
, u32 end_seq
)
1263 unsigned int pkt_len
;
1266 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1267 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1269 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1270 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1271 mss
= tcp_skb_mss(skb
);
1272 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1275 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1279 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1284 /* Round if necessary so that SACKs cover only full MSSes
1285 * and/or the remaining small portion (if present)
1287 if (pkt_len
> mss
) {
1288 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1289 if (!in_sack
&& new_len
< pkt_len
) {
1291 if (new_len
> skb
->len
)
1296 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1304 static u8
tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1305 struct tcp_sacktag_state
*state
,
1306 int dup_sack
, int pcount
)
1308 struct tcp_sock
*tp
= tcp_sk(sk
);
1309 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1310 int fack_count
= state
->fack_count
;
1312 /* Account D-SACK for retransmitted packet. */
1313 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1314 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1316 if (sacked
& TCPCB_SACKED_ACKED
)
1317 state
->reord
= min(fack_count
, state
->reord
);
1320 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1321 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1324 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1325 if (sacked
& TCPCB_SACKED_RETRANS
) {
1326 /* If the segment is not tagged as lost,
1327 * we do not clear RETRANS, believing
1328 * that retransmission is still in flight.
1330 if (sacked
& TCPCB_LOST
) {
1331 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1332 tp
->lost_out
-= pcount
;
1333 tp
->retrans_out
-= pcount
;
1336 if (!(sacked
& TCPCB_RETRANS
)) {
1337 /* New sack for not retransmitted frame,
1338 * which was in hole. It is reordering.
1340 if (before(TCP_SKB_CB(skb
)->seq
,
1341 tcp_highest_sack_seq(tp
)))
1342 state
->reord
= min(fack_count
,
1345 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1346 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1347 state
->flag
|= FLAG_ONLY_ORIG_SACKED
;
1350 if (sacked
& TCPCB_LOST
) {
1351 sacked
&= ~TCPCB_LOST
;
1352 tp
->lost_out
-= pcount
;
1356 sacked
|= TCPCB_SACKED_ACKED
;
1357 state
->flag
|= FLAG_DATA_SACKED
;
1358 tp
->sacked_out
+= pcount
;
1360 fack_count
+= pcount
;
1362 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1363 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1364 before(TCP_SKB_CB(skb
)->seq
,
1365 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1366 tp
->lost_cnt_hint
+= pcount
;
1368 if (fack_count
> tp
->fackets_out
)
1369 tp
->fackets_out
= fack_count
;
1372 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1373 * frames and clear it. undo_retrans is decreased above, L|R frames
1374 * are accounted above as well.
1376 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1377 sacked
&= ~TCPCB_SACKED_RETRANS
;
1378 tp
->retrans_out
-= pcount
;
1384 static int tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1385 struct tcp_sacktag_state
*state
,
1386 unsigned int pcount
, int shifted
, int mss
,
1389 struct tcp_sock
*tp
= tcp_sk(sk
);
1390 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1394 /* Tweak before seqno plays */
1395 if (!tcp_is_fack(tp
) && tcp_is_sack(tp
) && tp
->lost_skb_hint
&&
1396 !before(TCP_SKB_CB(tp
->lost_skb_hint
)->seq
, TCP_SKB_CB(skb
)->seq
))
1397 tp
->lost_cnt_hint
+= pcount
;
1399 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1400 TCP_SKB_CB(skb
)->seq
+= shifted
;
1402 skb_shinfo(prev
)->gso_segs
+= pcount
;
1403 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1404 skb_shinfo(skb
)->gso_segs
-= pcount
;
1406 /* When we're adding to gso_segs == 1, gso_size will be zero,
1407 * in theory this shouldn't be necessary but as long as DSACK
1408 * code can come after this skb later on it's better to keep
1409 * setting gso_size to something.
1411 if (!skb_shinfo(prev
)->gso_size
) {
1412 skb_shinfo(prev
)->gso_size
= mss
;
1413 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1416 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1417 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1418 skb_shinfo(skb
)->gso_size
= 0;
1419 skb_shinfo(skb
)->gso_type
= 0;
1422 /* We discard results */
1423 tcp_sacktag_one(skb
, sk
, state
, dup_sack
, pcount
);
1425 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1426 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1429 BUG_ON(!tcp_skb_pcount(skb
));
1430 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1434 /* Whole SKB was eaten :-) */
1436 if (skb
== tp
->retransmit_skb_hint
)
1437 tp
->retransmit_skb_hint
= prev
;
1438 if (skb
== tp
->scoreboard_skb_hint
)
1439 tp
->scoreboard_skb_hint
= prev
;
1440 if (skb
== tp
->lost_skb_hint
) {
1441 tp
->lost_skb_hint
= prev
;
1442 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1445 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(prev
)->flags
;
1446 if (skb
== tcp_highest_sack(sk
))
1447 tcp_advance_highest_sack(sk
, skb
);
1449 tcp_unlink_write_queue(skb
, sk
);
1450 sk_wmem_free_skb(sk
, skb
);
1452 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1457 /* I wish gso_size would have a bit more sane initialization than
1458 * something-or-zero which complicates things
1460 static int tcp_skb_seglen(struct sk_buff
*skb
)
1462 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1465 /* Shifting pages past head area doesn't work */
1466 static int skb_can_shift(struct sk_buff
*skb
)
1468 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1471 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1474 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1475 struct tcp_sacktag_state
*state
,
1476 u32 start_seq
, u32 end_seq
,
1479 struct tcp_sock
*tp
= tcp_sk(sk
);
1480 struct sk_buff
*prev
;
1486 if (!sk_can_gso(sk
))
1489 /* Normally R but no L won't result in plain S */
1491 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1493 if (!skb_can_shift(skb
))
1495 /* This frame is about to be dropped (was ACKed). */
1496 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1499 /* Can only happen with delayed DSACK + discard craziness */
1500 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1502 prev
= tcp_write_queue_prev(sk
, skb
);
1504 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1507 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1508 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1512 pcount
= tcp_skb_pcount(skb
);
1513 mss
= tcp_skb_seglen(skb
);
1515 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1516 * drop this restriction as unnecessary
1518 if (mss
!= tcp_skb_seglen(prev
))
1521 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1523 /* CHECKME: This is non-MSS split case only?, this will
1524 * cause skipped skbs due to advancing loop btw, original
1525 * has that feature too
1527 if (tcp_skb_pcount(skb
) <= 1)
1530 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1532 /* TODO: head merge to next could be attempted here
1533 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1534 * though it might not be worth of the additional hassle
1536 * ...we can probably just fallback to what was done
1537 * previously. We could try merging non-SACKed ones
1538 * as well but it probably isn't going to buy off
1539 * because later SACKs might again split them, and
1540 * it would make skb timestamp tracking considerably
1546 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1548 BUG_ON(len
> skb
->len
);
1550 /* MSS boundaries should be honoured or else pcount will
1551 * severely break even though it makes things bit trickier.
1552 * Optimize common case to avoid most of the divides
1554 mss
= tcp_skb_mss(skb
);
1556 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1557 * drop this restriction as unnecessary
1559 if (mss
!= tcp_skb_seglen(prev
))
1564 } else if (len
< mss
) {
1572 if (!skb_shift(prev
, skb
, len
))
1574 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1577 /* Hole filled allows collapsing with the next as well, this is very
1578 * useful when hole on every nth skb pattern happens
1580 if (prev
== tcp_write_queue_tail(sk
))
1582 skb
= tcp_write_queue_next(sk
, prev
);
1584 if (!skb_can_shift(skb
) ||
1585 (skb
== tcp_send_head(sk
)) ||
1586 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1587 (mss
!= tcp_skb_seglen(skb
)))
1591 if (skb_shift(prev
, skb
, len
)) {
1592 pcount
+= tcp_skb_pcount(skb
);
1593 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1597 state
->fack_count
+= pcount
;
1604 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1608 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1609 struct tcp_sack_block
*next_dup
,
1610 struct tcp_sacktag_state
*state
,
1611 u32 start_seq
, u32 end_seq
,
1614 struct tcp_sock
*tp
= tcp_sk(sk
);
1615 struct sk_buff
*tmp
;
1617 tcp_for_write_queue_from(skb
, sk
) {
1619 int dup_sack
= dup_sack_in
;
1621 if (skb
== tcp_send_head(sk
))
1624 /* queue is in-order => we can short-circuit the walk early */
1625 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1628 if ((next_dup
!= NULL
) &&
1629 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1630 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1631 next_dup
->start_seq
,
1637 /* skb reference here is a bit tricky to get right, since
1638 * shifting can eat and free both this skb and the next,
1639 * so not even _safe variant of the loop is enough.
1642 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1643 start_seq
, end_seq
, dup_sack
);
1652 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1658 if (unlikely(in_sack
< 0))
1662 TCP_SKB_CB(skb
)->sacked
= tcp_sacktag_one(skb
, sk
,
1665 tcp_skb_pcount(skb
));
1667 if (!before(TCP_SKB_CB(skb
)->seq
,
1668 tcp_highest_sack_seq(tp
)))
1669 tcp_advance_highest_sack(sk
, skb
);
1672 state
->fack_count
+= tcp_skb_pcount(skb
);
1677 /* Avoid all extra work that is being done by sacktag while walking in
1680 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1681 struct tcp_sacktag_state
*state
,
1684 tcp_for_write_queue_from(skb
, sk
) {
1685 if (skb
== tcp_send_head(sk
))
1688 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1691 state
->fack_count
+= tcp_skb_pcount(skb
);
1696 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1698 struct tcp_sack_block
*next_dup
,
1699 struct tcp_sacktag_state
*state
,
1702 if (next_dup
== NULL
)
1705 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1706 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1707 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1708 next_dup
->start_seq
, next_dup
->end_seq
,
1715 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1717 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1721 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1724 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1725 struct tcp_sock
*tp
= tcp_sk(sk
);
1726 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1727 TCP_SKB_CB(ack_skb
)->sacked
);
1728 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1729 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1730 struct tcp_sack_block
*cache
;
1731 struct tcp_sacktag_state state
;
1732 struct sk_buff
*skb
;
1733 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1735 int found_dup_sack
= 0;
1737 int first_sack_index
;
1740 state
.reord
= tp
->packets_out
;
1742 if (!tp
->sacked_out
) {
1743 if (WARN_ON(tp
->fackets_out
))
1744 tp
->fackets_out
= 0;
1745 tcp_highest_sack_reset(sk
);
1748 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1749 num_sacks
, prior_snd_una
);
1751 state
.flag
|= FLAG_DSACKING_ACK
;
1753 /* Eliminate too old ACKs, but take into
1754 * account more or less fresh ones, they can
1755 * contain valid SACK info.
1757 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1760 if (!tp
->packets_out
)
1764 first_sack_index
= 0;
1765 for (i
= 0; i
< num_sacks
; i
++) {
1766 int dup_sack
= !i
&& found_dup_sack
;
1768 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1769 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1771 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1772 sp
[used_sacks
].start_seq
,
1773 sp
[used_sacks
].end_seq
)) {
1777 if (!tp
->undo_marker
)
1778 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1780 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1782 /* Don't count olds caused by ACK reordering */
1783 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1784 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1786 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1789 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1791 first_sack_index
= -1;
1795 /* Ignore very old stuff early */
1796 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1802 /* order SACK blocks to allow in order walk of the retrans queue */
1803 for (i
= used_sacks
- 1; i
> 0; i
--) {
1804 for (j
= 0; j
< i
; j
++) {
1805 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1806 swap(sp
[j
], sp
[j
+ 1]);
1808 /* Track where the first SACK block goes to */
1809 if (j
== first_sack_index
)
1810 first_sack_index
= j
+ 1;
1815 skb
= tcp_write_queue_head(sk
);
1816 state
.fack_count
= 0;
1819 if (!tp
->sacked_out
) {
1820 /* It's already past, so skip checking against it */
1821 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1823 cache
= tp
->recv_sack_cache
;
1824 /* Skip empty blocks in at head of the cache */
1825 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1830 while (i
< used_sacks
) {
1831 u32 start_seq
= sp
[i
].start_seq
;
1832 u32 end_seq
= sp
[i
].end_seq
;
1833 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1834 struct tcp_sack_block
*next_dup
= NULL
;
1836 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1837 next_dup
= &sp
[i
+ 1];
1839 /* Event "B" in the comment above. */
1840 if (after(end_seq
, tp
->high_seq
))
1841 state
.flag
|= FLAG_DATA_LOST
;
1843 /* Skip too early cached blocks */
1844 while (tcp_sack_cache_ok(tp
, cache
) &&
1845 !before(start_seq
, cache
->end_seq
))
1848 /* Can skip some work by looking recv_sack_cache? */
1849 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1850 after(end_seq
, cache
->start_seq
)) {
1853 if (before(start_seq
, cache
->start_seq
)) {
1854 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1856 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1863 /* Rest of the block already fully processed? */
1864 if (!after(end_seq
, cache
->end_seq
))
1867 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1871 /* ...tail remains todo... */
1872 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1873 /* ...but better entrypoint exists! */
1874 skb
= tcp_highest_sack(sk
);
1877 state
.fack_count
= tp
->fackets_out
;
1882 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1883 /* Check overlap against next cached too (past this one already) */
1888 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1889 skb
= tcp_highest_sack(sk
);
1892 state
.fack_count
= tp
->fackets_out
;
1894 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1897 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1898 start_seq
, end_seq
, dup_sack
);
1901 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1902 * due to in-order walk
1904 if (after(end_seq
, tp
->frto_highmark
))
1905 state
.flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1910 /* Clear the head of the cache sack blocks so we can skip it next time */
1911 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1912 tp
->recv_sack_cache
[i
].start_seq
= 0;
1913 tp
->recv_sack_cache
[i
].end_seq
= 0;
1915 for (j
= 0; j
< used_sacks
; j
++)
1916 tp
->recv_sack_cache
[i
++] = sp
[j
];
1918 tcp_mark_lost_retrans(sk
);
1920 tcp_verify_left_out(tp
);
1922 if ((state
.reord
< tp
->fackets_out
) &&
1923 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1924 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1925 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1929 #if FASTRETRANS_DEBUG > 0
1930 WARN_ON((int)tp
->sacked_out
< 0);
1931 WARN_ON((int)tp
->lost_out
< 0);
1932 WARN_ON((int)tp
->retrans_out
< 0);
1933 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1938 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1939 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1941 static int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1945 holes
= max(tp
->lost_out
, 1U);
1946 holes
= min(holes
, tp
->packets_out
);
1948 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1949 tp
->sacked_out
= tp
->packets_out
- holes
;
1955 /* If we receive more dupacks than we expected counting segments
1956 * in assumption of absent reordering, interpret this as reordering.
1957 * The only another reason could be bug in receiver TCP.
1959 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1961 struct tcp_sock
*tp
= tcp_sk(sk
);
1962 if (tcp_limit_reno_sacked(tp
))
1963 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1966 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1968 static void tcp_add_reno_sack(struct sock
*sk
)
1970 struct tcp_sock
*tp
= tcp_sk(sk
);
1972 tcp_check_reno_reordering(sk
, 0);
1973 tcp_verify_left_out(tp
);
1976 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1978 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1980 struct tcp_sock
*tp
= tcp_sk(sk
);
1983 /* One ACK acked hole. The rest eat duplicate ACKs. */
1984 if (acked
- 1 >= tp
->sacked_out
)
1987 tp
->sacked_out
-= acked
- 1;
1989 tcp_check_reno_reordering(sk
, acked
);
1990 tcp_verify_left_out(tp
);
1993 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1998 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
2000 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
2003 /* F-RTO can only be used if TCP has never retransmitted anything other than
2004 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2006 int tcp_use_frto(struct sock
*sk
)
2008 const struct tcp_sock
*tp
= tcp_sk(sk
);
2009 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2010 struct sk_buff
*skb
;
2012 if (!sysctl_tcp_frto
)
2015 /* MTU probe and F-RTO won't really play nicely along currently */
2016 if (icsk
->icsk_mtup
.probe_size
)
2019 if (tcp_is_sackfrto(tp
))
2022 /* Avoid expensive walking of rexmit queue if possible */
2023 if (tp
->retrans_out
> 1)
2026 skb
= tcp_write_queue_head(sk
);
2027 if (tcp_skb_is_last(sk
, skb
))
2029 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
2030 tcp_for_write_queue_from(skb
, sk
) {
2031 if (skb
== tcp_send_head(sk
))
2033 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2035 /* Short-circuit when first non-SACKed skb has been checked */
2036 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2042 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2043 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2044 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2045 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2046 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2047 * bits are handled if the Loss state is really to be entered (in
2048 * tcp_enter_frto_loss).
2050 * Do like tcp_enter_loss() would; when RTO expires the second time it
2052 * "Reduce ssthresh if it has not yet been made inside this window."
2054 void tcp_enter_frto(struct sock
*sk
)
2056 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2057 struct tcp_sock
*tp
= tcp_sk(sk
);
2058 struct sk_buff
*skb
;
2060 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
2061 tp
->snd_una
== tp
->high_seq
||
2062 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
2063 !icsk
->icsk_retransmits
)) {
2064 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2065 /* Our state is too optimistic in ssthresh() call because cwnd
2066 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2067 * recovery has not yet completed. Pattern would be this: RTO,
2068 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2070 * RFC4138 should be more specific on what to do, even though
2071 * RTO is quite unlikely to occur after the first Cumulative ACK
2072 * due to back-off and complexity of triggering events ...
2074 if (tp
->frto_counter
) {
2076 stored_cwnd
= tp
->snd_cwnd
;
2078 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2079 tp
->snd_cwnd
= stored_cwnd
;
2081 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2083 /* ... in theory, cong.control module could do "any tricks" in
2084 * ssthresh(), which means that ca_state, lost bits and lost_out
2085 * counter would have to be faked before the call occurs. We
2086 * consider that too expensive, unlikely and hacky, so modules
2087 * using these in ssthresh() must deal these incompatibility
2088 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2090 tcp_ca_event(sk
, CA_EVENT_FRTO
);
2093 tp
->undo_marker
= tp
->snd_una
;
2094 tp
->undo_retrans
= 0;
2096 skb
= tcp_write_queue_head(sk
);
2097 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2098 tp
->undo_marker
= 0;
2099 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2100 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2101 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2103 tcp_verify_left_out(tp
);
2105 /* Too bad if TCP was application limited */
2106 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2108 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2109 * The last condition is necessary at least in tp->frto_counter case.
2111 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
2112 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
2113 after(tp
->high_seq
, tp
->snd_una
)) {
2114 tp
->frto_highmark
= tp
->high_seq
;
2116 tp
->frto_highmark
= tp
->snd_nxt
;
2118 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
2119 tp
->high_seq
= tp
->snd_nxt
;
2120 tp
->frto_counter
= 1;
2123 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2124 * which indicates that we should follow the traditional RTO recovery,
2125 * i.e. mark everything lost and do go-back-N retransmission.
2127 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
2129 struct tcp_sock
*tp
= tcp_sk(sk
);
2130 struct sk_buff
*skb
;
2133 tp
->retrans_out
= 0;
2134 if (tcp_is_reno(tp
))
2135 tcp_reset_reno_sack(tp
);
2137 tcp_for_write_queue(skb
, sk
) {
2138 if (skb
== tcp_send_head(sk
))
2141 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2143 * Count the retransmission made on RTO correctly (only when
2144 * waiting for the first ACK and did not get it)...
2146 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
2147 /* For some reason this R-bit might get cleared? */
2148 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
2149 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2150 /* ...enter this if branch just for the first segment */
2151 flag
|= FLAG_DATA_ACKED
;
2153 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2154 tp
->undo_marker
= 0;
2155 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2158 /* Marking forward transmissions that were made after RTO lost
2159 * can cause unnecessary retransmissions in some scenarios,
2160 * SACK blocks will mitigate that in some but not in all cases.
2161 * We used to not mark them but it was causing break-ups with
2162 * receivers that do only in-order receival.
2164 * TODO: we could detect presence of such receiver and select
2165 * different behavior per flow.
2167 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2168 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2169 tp
->lost_out
+= tcp_skb_pcount(skb
);
2170 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2173 tcp_verify_left_out(tp
);
2175 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2176 tp
->snd_cwnd_cnt
= 0;
2177 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2178 tp
->frto_counter
= 0;
2179 tp
->bytes_acked
= 0;
2181 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2182 sysctl_tcp_reordering
);
2183 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2184 tp
->high_seq
= tp
->snd_nxt
;
2185 TCP_ECN_queue_cwr(tp
);
2187 tcp_clear_all_retrans_hints(tp
);
2190 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2192 tp
->retrans_out
= 0;
2195 tp
->undo_marker
= 0;
2196 tp
->undo_retrans
= 0;
2199 void tcp_clear_retrans(struct tcp_sock
*tp
)
2201 tcp_clear_retrans_partial(tp
);
2203 tp
->fackets_out
= 0;
2207 /* Enter Loss state. If "how" is not zero, forget all SACK information
2208 * and reset tags completely, otherwise preserve SACKs. If receiver
2209 * dropped its ofo queue, we will know this due to reneging detection.
2211 void tcp_enter_loss(struct sock
*sk
, int how
)
2213 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2214 struct tcp_sock
*tp
= tcp_sk(sk
);
2215 struct sk_buff
*skb
;
2217 /* Reduce ssthresh if it has not yet been made inside this window. */
2218 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2219 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2220 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2221 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2222 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2225 tp
->snd_cwnd_cnt
= 0;
2226 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2228 tp
->bytes_acked
= 0;
2229 tcp_clear_retrans_partial(tp
);
2231 if (tcp_is_reno(tp
))
2232 tcp_reset_reno_sack(tp
);
2235 /* Push undo marker, if it was plain RTO and nothing
2236 * was retransmitted. */
2237 tp
->undo_marker
= tp
->snd_una
;
2240 tp
->fackets_out
= 0;
2242 tcp_clear_all_retrans_hints(tp
);
2244 tcp_for_write_queue(skb
, sk
) {
2245 if (skb
== tcp_send_head(sk
))
2248 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2249 tp
->undo_marker
= 0;
2250 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2251 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2252 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2253 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2254 tp
->lost_out
+= tcp_skb_pcount(skb
);
2255 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2258 tcp_verify_left_out(tp
);
2260 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2261 sysctl_tcp_reordering
);
2262 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2263 tp
->high_seq
= tp
->snd_nxt
;
2264 TCP_ECN_queue_cwr(tp
);
2265 /* Abort F-RTO algorithm if one is in progress */
2266 tp
->frto_counter
= 0;
2269 /* If ACK arrived pointing to a remembered SACK, it means that our
2270 * remembered SACKs do not reflect real state of receiver i.e.
2271 * receiver _host_ is heavily congested (or buggy).
2273 * Do processing similar to RTO timeout.
2275 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2277 if (flag
& FLAG_SACK_RENEGING
) {
2278 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2279 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2281 tcp_enter_loss(sk
, 1);
2282 icsk
->icsk_retransmits
++;
2283 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2284 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2285 icsk
->icsk_rto
, TCP_RTO_MAX
);
2291 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2293 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2296 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2297 * counter when SACK is enabled (without SACK, sacked_out is used for
2300 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2301 * segments up to the highest received SACK block so far and holes in
2304 * With reordering, holes may still be in flight, so RFC3517 recovery
2305 * uses pure sacked_out (total number of SACKed segments) even though
2306 * it violates the RFC that uses duplicate ACKs, often these are equal
2307 * but when e.g. out-of-window ACKs or packet duplication occurs,
2308 * they differ. Since neither occurs due to loss, TCP should really
2311 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
2313 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2316 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2318 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2321 static inline int tcp_head_timedout(struct sock
*sk
)
2323 struct tcp_sock
*tp
= tcp_sk(sk
);
2325 return tp
->packets_out
&&
2326 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2329 /* Linux NewReno/SACK/FACK/ECN state machine.
2330 * --------------------------------------
2332 * "Open" Normal state, no dubious events, fast path.
2333 * "Disorder" In all the respects it is "Open",
2334 * but requires a bit more attention. It is entered when
2335 * we see some SACKs or dupacks. It is split of "Open"
2336 * mainly to move some processing from fast path to slow one.
2337 * "CWR" CWND was reduced due to some Congestion Notification event.
2338 * It can be ECN, ICMP source quench, local device congestion.
2339 * "Recovery" CWND was reduced, we are fast-retransmitting.
2340 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2342 * tcp_fastretrans_alert() is entered:
2343 * - each incoming ACK, if state is not "Open"
2344 * - when arrived ACK is unusual, namely:
2349 * Counting packets in flight is pretty simple.
2351 * in_flight = packets_out - left_out + retrans_out
2353 * packets_out is SND.NXT-SND.UNA counted in packets.
2355 * retrans_out is number of retransmitted segments.
2357 * left_out is number of segments left network, but not ACKed yet.
2359 * left_out = sacked_out + lost_out
2361 * sacked_out: Packets, which arrived to receiver out of order
2362 * and hence not ACKed. With SACKs this number is simply
2363 * amount of SACKed data. Even without SACKs
2364 * it is easy to give pretty reliable estimate of this number,
2365 * counting duplicate ACKs.
2367 * lost_out: Packets lost by network. TCP has no explicit
2368 * "loss notification" feedback from network (for now).
2369 * It means that this number can be only _guessed_.
2370 * Actually, it is the heuristics to predict lossage that
2371 * distinguishes different algorithms.
2373 * F.e. after RTO, when all the queue is considered as lost,
2374 * lost_out = packets_out and in_flight = retrans_out.
2376 * Essentially, we have now two algorithms counting
2379 * FACK: It is the simplest heuristics. As soon as we decided
2380 * that something is lost, we decide that _all_ not SACKed
2381 * packets until the most forward SACK are lost. I.e.
2382 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2383 * It is absolutely correct estimate, if network does not reorder
2384 * packets. And it loses any connection to reality when reordering
2385 * takes place. We use FACK by default until reordering
2386 * is suspected on the path to this destination.
2388 * NewReno: when Recovery is entered, we assume that one segment
2389 * is lost (classic Reno). While we are in Recovery and
2390 * a partial ACK arrives, we assume that one more packet
2391 * is lost (NewReno). This heuristics are the same in NewReno
2394 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2395 * deflation etc. CWND is real congestion window, never inflated, changes
2396 * only according to classic VJ rules.
2398 * Really tricky (and requiring careful tuning) part of algorithm
2399 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2400 * The first determines the moment _when_ we should reduce CWND and,
2401 * hence, slow down forward transmission. In fact, it determines the moment
2402 * when we decide that hole is caused by loss, rather than by a reorder.
2404 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2405 * holes, caused by lost packets.
2407 * And the most logically complicated part of algorithm is undo
2408 * heuristics. We detect false retransmits due to both too early
2409 * fast retransmit (reordering) and underestimated RTO, analyzing
2410 * timestamps and D-SACKs. When we detect that some segments were
2411 * retransmitted by mistake and CWND reduction was wrong, we undo
2412 * window reduction and abort recovery phase. This logic is hidden
2413 * inside several functions named tcp_try_undo_<something>.
2416 /* This function decides, when we should leave Disordered state
2417 * and enter Recovery phase, reducing congestion window.
2419 * Main question: may we further continue forward transmission
2420 * with the same cwnd?
2422 static int tcp_time_to_recover(struct sock
*sk
)
2424 struct tcp_sock
*tp
= tcp_sk(sk
);
2427 /* Do not perform any recovery during F-RTO algorithm */
2428 if (tp
->frto_counter
)
2431 /* Trick#1: The loss is proven. */
2435 /* Not-A-Trick#2 : Classic rule... */
2436 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2439 /* Trick#3 : when we use RFC2988 timer restart, fast
2440 * retransmit can be triggered by timeout of queue head.
2442 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2445 /* Trick#4: It is still not OK... But will it be useful to delay
2448 packets_out
= tp
->packets_out
;
2449 if (packets_out
<= tp
->reordering
&&
2450 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2451 !tcp_may_send_now(sk
)) {
2452 /* We have nothing to send. This connection is limited
2453 * either by receiver window or by application.
2461 /* New heuristics: it is possible only after we switched to restart timer
2462 * each time when something is ACKed. Hence, we can detect timed out packets
2463 * during fast retransmit without falling to slow start.
2465 * Usefulness of this as is very questionable, since we should know which of
2466 * the segments is the next to timeout which is relatively expensive to find
2467 * in general case unless we add some data structure just for that. The
2468 * current approach certainly won't find the right one too often and when it
2469 * finally does find _something_ it usually marks large part of the window
2470 * right away (because a retransmission with a larger timestamp blocks the
2471 * loop from advancing). -ij
2473 static void tcp_timeout_skbs(struct sock
*sk
)
2475 struct tcp_sock
*tp
= tcp_sk(sk
);
2476 struct sk_buff
*skb
;
2478 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2481 skb
= tp
->scoreboard_skb_hint
;
2482 if (tp
->scoreboard_skb_hint
== NULL
)
2483 skb
= tcp_write_queue_head(sk
);
2485 tcp_for_write_queue_from(skb
, sk
) {
2486 if (skb
== tcp_send_head(sk
))
2488 if (!tcp_skb_timedout(sk
, skb
))
2491 tcp_skb_mark_lost(tp
, skb
);
2494 tp
->scoreboard_skb_hint
= skb
;
2496 tcp_verify_left_out(tp
);
2499 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2500 * is against sacked "cnt", otherwise it's against facked "cnt"
2502 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2504 struct tcp_sock
*tp
= tcp_sk(sk
);
2505 struct sk_buff
*skb
;
2510 WARN_ON(packets
> tp
->packets_out
);
2511 if (tp
->lost_skb_hint
) {
2512 skb
= tp
->lost_skb_hint
;
2513 cnt
= tp
->lost_cnt_hint
;
2515 skb
= tcp_write_queue_head(sk
);
2519 tcp_for_write_queue_from(skb
, sk
) {
2520 if (skb
== tcp_send_head(sk
))
2522 /* TODO: do this better */
2523 /* this is not the most efficient way to do this... */
2524 tp
->lost_skb_hint
= skb
;
2525 tp
->lost_cnt_hint
= cnt
;
2527 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2531 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2532 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2533 cnt
+= tcp_skb_pcount(skb
);
2535 if (cnt
> packets
) {
2536 if (tcp_is_sack(tp
) || (oldcnt
>= packets
))
2539 mss
= skb_shinfo(skb
)->gso_size
;
2540 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2546 tcp_skb_mark_lost(tp
, skb
);
2548 tcp_verify_left_out(tp
);
2551 /* Account newly detected lost packet(s) */
2553 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2555 struct tcp_sock
*tp
= tcp_sk(sk
);
2557 if (tcp_is_reno(tp
)) {
2558 tcp_mark_head_lost(sk
, 1);
2559 } else if (tcp_is_fack(tp
)) {
2560 int lost
= tp
->fackets_out
- tp
->reordering
;
2563 tcp_mark_head_lost(sk
, lost
);
2565 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2566 if (sacked_upto
< fast_rexmit
)
2567 sacked_upto
= fast_rexmit
;
2568 tcp_mark_head_lost(sk
, sacked_upto
);
2571 tcp_timeout_skbs(sk
);
2574 /* CWND moderation, preventing bursts due to too big ACKs
2575 * in dubious situations.
2577 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2579 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2580 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2581 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2584 /* Lower bound on congestion window is slow start threshold
2585 * unless congestion avoidance choice decides to overide it.
2587 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2589 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2591 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2594 /* Decrease cwnd each second ack. */
2595 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2597 struct tcp_sock
*tp
= tcp_sk(sk
);
2598 int decr
= tp
->snd_cwnd_cnt
+ 1;
2600 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2601 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2602 tp
->snd_cwnd_cnt
= decr
& 1;
2605 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2606 tp
->snd_cwnd
-= decr
;
2608 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2609 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2613 /* Nothing was retransmitted or returned timestamp is less
2614 * than timestamp of the first retransmission.
2616 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2618 return !tp
->retrans_stamp
||
2619 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2620 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2623 /* Undo procedures. */
2625 #if FASTRETRANS_DEBUG > 1
2626 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2628 struct tcp_sock
*tp
= tcp_sk(sk
);
2629 struct inet_sock
*inet
= inet_sk(sk
);
2631 if (sk
->sk_family
== AF_INET
) {
2632 printk(KERN_DEBUG
"Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2634 &inet
->daddr
, ntohs(inet
->dport
),
2635 tp
->snd_cwnd
, tcp_left_out(tp
),
2636 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2639 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2640 else if (sk
->sk_family
== AF_INET6
) {
2641 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2642 printk(KERN_DEBUG
"Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2644 &np
->daddr
, ntohs(inet
->dport
),
2645 tp
->snd_cwnd
, tcp_left_out(tp
),
2646 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2652 #define DBGUNDO(x...) do { } while (0)
2655 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2657 struct tcp_sock
*tp
= tcp_sk(sk
);
2659 if (tp
->prior_ssthresh
) {
2660 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2662 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2663 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2665 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2667 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2668 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2669 TCP_ECN_withdraw_cwr(tp
);
2672 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2674 tcp_moderate_cwnd(tp
);
2675 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2678 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2680 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2683 /* People celebrate: "We love our President!" */
2684 static int tcp_try_undo_recovery(struct sock
*sk
)
2686 struct tcp_sock
*tp
= tcp_sk(sk
);
2688 if (tcp_may_undo(tp
)) {
2691 /* Happy end! We did not retransmit anything
2692 * or our original transmission succeeded.
2694 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2695 tcp_undo_cwr(sk
, 1);
2696 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2697 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2699 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2701 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2702 tp
->undo_marker
= 0;
2704 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2705 /* Hold old state until something *above* high_seq
2706 * is ACKed. For Reno it is MUST to prevent false
2707 * fast retransmits (RFC2582). SACK TCP is safe. */
2708 tcp_moderate_cwnd(tp
);
2711 tcp_set_ca_state(sk
, TCP_CA_Open
);
2715 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2716 static void tcp_try_undo_dsack(struct sock
*sk
)
2718 struct tcp_sock
*tp
= tcp_sk(sk
);
2720 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2721 DBGUNDO(sk
, "D-SACK");
2722 tcp_undo_cwr(sk
, 1);
2723 tp
->undo_marker
= 0;
2724 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2728 /* Undo during fast recovery after partial ACK. */
2730 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2732 struct tcp_sock
*tp
= tcp_sk(sk
);
2733 /* Partial ACK arrived. Force Hoe's retransmit. */
2734 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2736 if (tcp_may_undo(tp
)) {
2737 /* Plain luck! Hole if filled with delayed
2738 * packet, rather than with a retransmit.
2740 if (tp
->retrans_out
== 0)
2741 tp
->retrans_stamp
= 0;
2743 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2746 tcp_undo_cwr(sk
, 0);
2747 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2749 /* So... Do not make Hoe's retransmit yet.
2750 * If the first packet was delayed, the rest
2751 * ones are most probably delayed as well.
2758 /* Undo during loss recovery after partial ACK. */
2759 static int tcp_try_undo_loss(struct sock
*sk
)
2761 struct tcp_sock
*tp
= tcp_sk(sk
);
2763 if (tcp_may_undo(tp
)) {
2764 struct sk_buff
*skb
;
2765 tcp_for_write_queue(skb
, sk
) {
2766 if (skb
== tcp_send_head(sk
))
2768 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2771 tcp_clear_all_retrans_hints(tp
);
2773 DBGUNDO(sk
, "partial loss");
2775 tcp_undo_cwr(sk
, 1);
2776 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2777 inet_csk(sk
)->icsk_retransmits
= 0;
2778 tp
->undo_marker
= 0;
2779 if (tcp_is_sack(tp
))
2780 tcp_set_ca_state(sk
, TCP_CA_Open
);
2786 static inline void tcp_complete_cwr(struct sock
*sk
)
2788 struct tcp_sock
*tp
= tcp_sk(sk
);
2789 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2790 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2791 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2794 static void tcp_try_keep_open(struct sock
*sk
)
2796 struct tcp_sock
*tp
= tcp_sk(sk
);
2797 int state
= TCP_CA_Open
;
2799 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2800 state
= TCP_CA_Disorder
;
2802 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2803 tcp_set_ca_state(sk
, state
);
2804 tp
->high_seq
= tp
->snd_nxt
;
2808 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2810 struct tcp_sock
*tp
= tcp_sk(sk
);
2812 tcp_verify_left_out(tp
);
2814 if (!tp
->frto_counter
&& tp
->retrans_out
== 0)
2815 tp
->retrans_stamp
= 0;
2817 if (flag
& FLAG_ECE
)
2818 tcp_enter_cwr(sk
, 1);
2820 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2821 tcp_try_keep_open(sk
);
2822 tcp_moderate_cwnd(tp
);
2824 tcp_cwnd_down(sk
, flag
);
2828 static void tcp_mtup_probe_failed(struct sock
*sk
)
2830 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2832 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2833 icsk
->icsk_mtup
.probe_size
= 0;
2836 static void tcp_mtup_probe_success(struct sock
*sk
)
2838 struct tcp_sock
*tp
= tcp_sk(sk
);
2839 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2841 /* FIXME: breaks with very large cwnd */
2842 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2843 tp
->snd_cwnd
= tp
->snd_cwnd
*
2844 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2845 icsk
->icsk_mtup
.probe_size
;
2846 tp
->snd_cwnd_cnt
= 0;
2847 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2848 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2850 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2851 icsk
->icsk_mtup
.probe_size
= 0;
2852 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2855 /* Do a simple retransmit without using the backoff mechanisms in
2856 * tcp_timer. This is used for path mtu discovery.
2857 * The socket is already locked here.
2859 void tcp_simple_retransmit(struct sock
*sk
)
2861 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2862 struct tcp_sock
*tp
= tcp_sk(sk
);
2863 struct sk_buff
*skb
;
2864 unsigned int mss
= tcp_current_mss(sk
);
2865 u32 prior_lost
= tp
->lost_out
;
2867 tcp_for_write_queue(skb
, sk
) {
2868 if (skb
== tcp_send_head(sk
))
2870 if (tcp_skb_seglen(skb
) > mss
&&
2871 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2872 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2873 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2874 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2876 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2880 tcp_clear_retrans_hints_partial(tp
);
2882 if (prior_lost
== tp
->lost_out
)
2885 if (tcp_is_reno(tp
))
2886 tcp_limit_reno_sacked(tp
);
2888 tcp_verify_left_out(tp
);
2890 /* Don't muck with the congestion window here.
2891 * Reason is that we do not increase amount of _data_
2892 * in network, but units changed and effective
2893 * cwnd/ssthresh really reduced now.
2895 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2896 tp
->high_seq
= tp
->snd_nxt
;
2897 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2898 tp
->prior_ssthresh
= 0;
2899 tp
->undo_marker
= 0;
2900 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2902 tcp_xmit_retransmit_queue(sk
);
2905 /* Process an event, which can update packets-in-flight not trivially.
2906 * Main goal of this function is to calculate new estimate for left_out,
2907 * taking into account both packets sitting in receiver's buffer and
2908 * packets lost by network.
2910 * Besides that it does CWND reduction, when packet loss is detected
2911 * and changes state of machine.
2913 * It does _not_ decide what to send, it is made in function
2914 * tcp_xmit_retransmit_queue().
2916 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2918 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2919 struct tcp_sock
*tp
= tcp_sk(sk
);
2920 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2921 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2922 (tcp_fackets_out(tp
) > tp
->reordering
));
2923 int fast_rexmit
= 0, mib_idx
;
2925 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2927 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2928 tp
->fackets_out
= 0;
2930 /* Now state machine starts.
2931 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2932 if (flag
& FLAG_ECE
)
2933 tp
->prior_ssthresh
= 0;
2935 /* B. In all the states check for reneging SACKs. */
2936 if (tcp_check_sack_reneging(sk
, flag
))
2939 /* C. Process data loss notification, provided it is valid. */
2940 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2941 before(tp
->snd_una
, tp
->high_seq
) &&
2942 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2943 tp
->fackets_out
> tp
->reordering
) {
2944 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2945 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2948 /* D. Check consistency of the current state. */
2949 tcp_verify_left_out(tp
);
2951 /* E. Check state exit conditions. State can be terminated
2952 * when high_seq is ACKed. */
2953 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2954 WARN_ON(tp
->retrans_out
!= 0);
2955 tp
->retrans_stamp
= 0;
2956 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2957 switch (icsk
->icsk_ca_state
) {
2959 icsk
->icsk_retransmits
= 0;
2960 if (tcp_try_undo_recovery(sk
))
2965 /* CWR is to be held something *above* high_seq
2966 * is ACKed for CWR bit to reach receiver. */
2967 if (tp
->snd_una
!= tp
->high_seq
) {
2968 tcp_complete_cwr(sk
);
2969 tcp_set_ca_state(sk
, TCP_CA_Open
);
2973 case TCP_CA_Disorder
:
2974 tcp_try_undo_dsack(sk
);
2975 if (!tp
->undo_marker
||
2976 /* For SACK case do not Open to allow to undo
2977 * catching for all duplicate ACKs. */
2978 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2979 tp
->undo_marker
= 0;
2980 tcp_set_ca_state(sk
, TCP_CA_Open
);
2984 case TCP_CA_Recovery
:
2985 if (tcp_is_reno(tp
))
2986 tcp_reset_reno_sack(tp
);
2987 if (tcp_try_undo_recovery(sk
))
2989 tcp_complete_cwr(sk
);
2994 /* F. Process state. */
2995 switch (icsk
->icsk_ca_state
) {
2996 case TCP_CA_Recovery
:
2997 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2998 if (tcp_is_reno(tp
) && is_dupack
)
2999 tcp_add_reno_sack(sk
);
3001 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
3004 if (flag
& FLAG_DATA_ACKED
)
3005 icsk
->icsk_retransmits
= 0;
3006 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
3007 tcp_reset_reno_sack(tp
);
3008 if (!tcp_try_undo_loss(sk
)) {
3009 tcp_moderate_cwnd(tp
);
3010 tcp_xmit_retransmit_queue(sk
);
3013 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
3015 /* Loss is undone; fall through to processing in Open state. */
3017 if (tcp_is_reno(tp
)) {
3018 if (flag
& FLAG_SND_UNA_ADVANCED
)
3019 tcp_reset_reno_sack(tp
);
3021 tcp_add_reno_sack(sk
);
3024 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
3025 tcp_try_undo_dsack(sk
);
3027 if (!tcp_time_to_recover(sk
)) {
3028 tcp_try_to_open(sk
, flag
);
3032 /* MTU probe failure: don't reduce cwnd */
3033 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3034 icsk
->icsk_mtup
.probe_size
&&
3035 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3036 tcp_mtup_probe_failed(sk
);
3037 /* Restores the reduction we did in tcp_mtup_probe() */
3039 tcp_simple_retransmit(sk
);
3043 /* Otherwise enter Recovery state */
3045 if (tcp_is_reno(tp
))
3046 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
3048 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
3050 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3052 tp
->high_seq
= tp
->snd_nxt
;
3053 tp
->prior_ssthresh
= 0;
3054 tp
->undo_marker
= tp
->snd_una
;
3055 tp
->undo_retrans
= tp
->retrans_out
;
3057 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
3058 if (!(flag
& FLAG_ECE
))
3059 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
3060 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
3061 TCP_ECN_queue_cwr(tp
);
3064 tp
->bytes_acked
= 0;
3065 tp
->snd_cwnd_cnt
= 0;
3066 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
3070 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3071 tcp_update_scoreboard(sk
, fast_rexmit
);
3072 tcp_cwnd_down(sk
, flag
);
3073 tcp_xmit_retransmit_queue(sk
);
3076 static void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
3078 tcp_rtt_estimator(sk
, seq_rtt
);
3080 inet_csk(sk
)->icsk_backoff
= 0;
3083 /* Read draft-ietf-tcplw-high-performance before mucking
3084 * with this code. (Supersedes RFC1323)
3086 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3088 /* RTTM Rule: A TSecr value received in a segment is used to
3089 * update the averaged RTT measurement only if the segment
3090 * acknowledges some new data, i.e., only if it advances the
3091 * left edge of the send window.
3093 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3094 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3096 * Changed: reset backoff as soon as we see the first valid sample.
3097 * If we do not, we get strongly overestimated rto. With timestamps
3098 * samples are accepted even from very old segments: f.e., when rtt=1
3099 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3100 * answer arrives rto becomes 120 seconds! If at least one of segments
3101 * in window is lost... Voila. --ANK (010210)
3103 struct tcp_sock
*tp
= tcp_sk(sk
);
3105 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
3108 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3110 /* We don't have a timestamp. Can only use
3111 * packets that are not retransmitted to determine
3112 * rtt estimates. Also, we must not reset the
3113 * backoff for rto until we get a non-retransmitted
3114 * packet. This allows us to deal with a situation
3115 * where the network delay has increased suddenly.
3116 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3119 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3122 tcp_valid_rtt_meas(sk
, seq_rtt
);
3125 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3128 const struct tcp_sock
*tp
= tcp_sk(sk
);
3129 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3130 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3131 tcp_ack_saw_tstamp(sk
, flag
);
3132 else if (seq_rtt
>= 0)
3133 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3136 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3138 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3139 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3140 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3143 /* Restart timer after forward progress on connection.
3144 * RFC2988 recommends to restart timer to now+rto.
3146 static void tcp_rearm_rto(struct sock
*sk
)
3148 struct tcp_sock
*tp
= tcp_sk(sk
);
3150 if (!tp
->packets_out
) {
3151 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3153 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3154 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3158 /* If we get here, the whole TSO packet has not been acked. */
3159 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3161 struct tcp_sock
*tp
= tcp_sk(sk
);
3164 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3166 packets_acked
= tcp_skb_pcount(skb
);
3167 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3169 packets_acked
-= tcp_skb_pcount(skb
);
3171 if (packets_acked
) {
3172 BUG_ON(tcp_skb_pcount(skb
) == 0);
3173 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3176 return packets_acked
;
3179 /* Remove acknowledged frames from the retransmission queue. If our packet
3180 * is before the ack sequence we can discard it as it's confirmed to have
3181 * arrived at the other end.
3183 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3186 struct tcp_sock
*tp
= tcp_sk(sk
);
3187 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3188 struct sk_buff
*skb
;
3189 u32 now
= tcp_time_stamp
;
3190 int fully_acked
= 1;
3193 u32 reord
= tp
->packets_out
;
3194 u32 prior_sacked
= tp
->sacked_out
;
3196 s32 ca_seq_rtt
= -1;
3197 ktime_t last_ackt
= net_invalid_timestamp();
3199 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3200 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3202 u8 sacked
= scb
->sacked
;
3204 /* Determine how many packets and what bytes were acked, tso and else */
3205 if (after(scb
->end_seq
, tp
->snd_una
)) {
3206 if (tcp_skb_pcount(skb
) == 1 ||
3207 !after(tp
->snd_una
, scb
->seq
))
3210 acked_pcount
= tcp_tso_acked(sk
, skb
);
3216 acked_pcount
= tcp_skb_pcount(skb
);
3219 if (sacked
& TCPCB_RETRANS
) {
3220 if (sacked
& TCPCB_SACKED_RETRANS
)
3221 tp
->retrans_out
-= acked_pcount
;
3222 flag
|= FLAG_RETRANS_DATA_ACKED
;
3225 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3226 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3228 ca_seq_rtt
= now
- scb
->when
;
3229 last_ackt
= skb
->tstamp
;
3231 seq_rtt
= ca_seq_rtt
;
3233 if (!(sacked
& TCPCB_SACKED_ACKED
))
3234 reord
= min(pkts_acked
, reord
);
3237 if (sacked
& TCPCB_SACKED_ACKED
)
3238 tp
->sacked_out
-= acked_pcount
;
3239 if (sacked
& TCPCB_LOST
)
3240 tp
->lost_out
-= acked_pcount
;
3242 tp
->packets_out
-= acked_pcount
;
3243 pkts_acked
+= acked_pcount
;
3245 /* Initial outgoing SYN's get put onto the write_queue
3246 * just like anything else we transmit. It is not
3247 * true data, and if we misinform our callers that
3248 * this ACK acks real data, we will erroneously exit
3249 * connection startup slow start one packet too
3250 * quickly. This is severely frowned upon behavior.
3252 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
3253 flag
|= FLAG_DATA_ACKED
;
3255 flag
|= FLAG_SYN_ACKED
;
3256 tp
->retrans_stamp
= 0;
3262 tcp_unlink_write_queue(skb
, sk
);
3263 sk_wmem_free_skb(sk
, skb
);
3264 tp
->scoreboard_skb_hint
= NULL
;
3265 if (skb
== tp
->retransmit_skb_hint
)
3266 tp
->retransmit_skb_hint
= NULL
;
3267 if (skb
== tp
->lost_skb_hint
)
3268 tp
->lost_skb_hint
= NULL
;
3271 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3272 tp
->snd_up
= tp
->snd_una
;
3274 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3275 flag
|= FLAG_SACK_RENEGING
;
3277 if (flag
& FLAG_ACKED
) {
3278 const struct tcp_congestion_ops
*ca_ops
3279 = inet_csk(sk
)->icsk_ca_ops
;
3281 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3282 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3283 tcp_mtup_probe_success(sk
);
3286 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3289 if (tcp_is_reno(tp
)) {
3290 tcp_remove_reno_sacks(sk
, pkts_acked
);
3294 /* Non-retransmitted hole got filled? That's reordering */
3295 if (reord
< prior_fackets
)
3296 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3298 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3299 prior_sacked
- tp
->sacked_out
;
3300 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3303 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3305 if (ca_ops
->pkts_acked
) {
3308 /* Is the ACK triggering packet unambiguous? */
3309 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3310 /* High resolution needed and available? */
3311 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3312 !ktime_equal(last_ackt
,
3313 net_invalid_timestamp()))
3314 rtt_us
= ktime_us_delta(ktime_get_real(),
3316 else if (ca_seq_rtt
> 0)
3317 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3320 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3324 #if FASTRETRANS_DEBUG > 0
3325 WARN_ON((int)tp
->sacked_out
< 0);
3326 WARN_ON((int)tp
->lost_out
< 0);
3327 WARN_ON((int)tp
->retrans_out
< 0);
3328 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3329 icsk
= inet_csk(sk
);
3331 printk(KERN_DEBUG
"Leak l=%u %d\n",
3332 tp
->lost_out
, icsk
->icsk_ca_state
);
3335 if (tp
->sacked_out
) {
3336 printk(KERN_DEBUG
"Leak s=%u %d\n",
3337 tp
->sacked_out
, icsk
->icsk_ca_state
);
3340 if (tp
->retrans_out
) {
3341 printk(KERN_DEBUG
"Leak r=%u %d\n",
3342 tp
->retrans_out
, icsk
->icsk_ca_state
);
3343 tp
->retrans_out
= 0;
3350 static void tcp_ack_probe(struct sock
*sk
)
3352 const struct tcp_sock
*tp
= tcp_sk(sk
);
3353 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3355 /* Was it a usable window open? */
3357 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3358 icsk
->icsk_backoff
= 0;
3359 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3360 /* Socket must be waked up by subsequent tcp_data_snd_check().
3361 * This function is not for random using!
3364 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3365 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3370 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3372 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3373 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3376 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3378 const struct tcp_sock
*tp
= tcp_sk(sk
);
3379 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3380 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3383 /* Check that window update is acceptable.
3384 * The function assumes that snd_una<=ack<=snd_next.
3386 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3387 const u32 ack
, const u32 ack_seq
,
3390 return (after(ack
, tp
->snd_una
) ||
3391 after(ack_seq
, tp
->snd_wl1
) ||
3392 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3395 /* Update our send window.
3397 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3398 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3400 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3403 struct tcp_sock
*tp
= tcp_sk(sk
);
3405 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3407 if (likely(!tcp_hdr(skb
)->syn
))
3408 nwin
<<= tp
->rx_opt
.snd_wscale
;
3410 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3411 flag
|= FLAG_WIN_UPDATE
;
3412 tcp_update_wl(tp
, ack_seq
);
3414 if (tp
->snd_wnd
!= nwin
) {
3417 /* Note, it is the only place, where
3418 * fast path is recovered for sending TCP.
3421 tcp_fast_path_check(sk
);
3423 if (nwin
> tp
->max_window
) {
3424 tp
->max_window
= nwin
;
3425 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3435 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3436 * continue in congestion avoidance.
3438 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3440 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3441 tp
->snd_cwnd_cnt
= 0;
3442 tp
->bytes_acked
= 0;
3443 TCP_ECN_queue_cwr(tp
);
3444 tcp_moderate_cwnd(tp
);
3447 /* A conservative spurious RTO response algorithm: reduce cwnd using
3448 * rate halving and continue in congestion avoidance.
3450 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3452 tcp_enter_cwr(sk
, 0);
3455 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3457 if (flag
& FLAG_ECE
)
3458 tcp_ratehalving_spur_to_response(sk
);
3460 tcp_undo_cwr(sk
, 1);
3463 /* F-RTO spurious RTO detection algorithm (RFC4138)
3465 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3466 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3467 * window (but not to or beyond highest sequence sent before RTO):
3468 * On First ACK, send two new segments out.
3469 * On Second ACK, RTO was likely spurious. Do spurious response (response
3470 * algorithm is not part of the F-RTO detection algorithm
3471 * given in RFC4138 but can be selected separately).
3472 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3473 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3474 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3475 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3477 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3478 * original window even after we transmit two new data segments.
3481 * on first step, wait until first cumulative ACK arrives, then move to
3482 * the second step. In second step, the next ACK decides.
3484 * F-RTO is implemented (mainly) in four functions:
3485 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3486 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3487 * called when tcp_use_frto() showed green light
3488 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3489 * - tcp_enter_frto_loss() is called if there is not enough evidence
3490 * to prove that the RTO is indeed spurious. It transfers the control
3491 * from F-RTO to the conventional RTO recovery
3493 static int tcp_process_frto(struct sock
*sk
, int flag
)
3495 struct tcp_sock
*tp
= tcp_sk(sk
);
3497 tcp_verify_left_out(tp
);
3499 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3500 if (flag
& FLAG_DATA_ACKED
)
3501 inet_csk(sk
)->icsk_retransmits
= 0;
3503 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3504 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3505 tp
->undo_marker
= 0;
3507 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3508 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3512 if (!tcp_is_sackfrto(tp
)) {
3513 /* RFC4138 shortcoming in step 2; should also have case c):
3514 * ACK isn't duplicate nor advances window, e.g., opposite dir
3517 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3520 if (!(flag
& FLAG_DATA_ACKED
)) {
3521 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3526 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3527 /* Prevent sending of new data. */
3528 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3529 tcp_packets_in_flight(tp
));
3533 if ((tp
->frto_counter
>= 2) &&
3534 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3535 ((flag
& FLAG_DATA_SACKED
) &&
3536 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3537 /* RFC4138 shortcoming (see comment above) */
3538 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3539 (flag
& FLAG_NOT_DUP
))
3542 tcp_enter_frto_loss(sk
, 3, flag
);
3547 if (tp
->frto_counter
== 1) {
3548 /* tcp_may_send_now needs to see updated state */
3549 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3550 tp
->frto_counter
= 2;
3552 if (!tcp_may_send_now(sk
))
3553 tcp_enter_frto_loss(sk
, 2, flag
);
3557 switch (sysctl_tcp_frto_response
) {
3559 tcp_undo_spur_to_response(sk
, flag
);
3562 tcp_conservative_spur_to_response(tp
);
3565 tcp_ratehalving_spur_to_response(sk
);
3568 tp
->frto_counter
= 0;
3569 tp
->undo_marker
= 0;
3570 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3575 /* This routine deals with incoming acks, but not outgoing ones. */
3576 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3578 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3579 struct tcp_sock
*tp
= tcp_sk(sk
);
3580 u32 prior_snd_una
= tp
->snd_una
;
3581 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3582 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3583 u32 prior_in_flight
;
3588 /* If the ack is older than previous acks
3589 * then we can probably ignore it.
3591 if (before(ack
, prior_snd_una
))
3594 /* If the ack includes data we haven't sent yet, discard
3595 * this segment (RFC793 Section 3.9).
3597 if (after(ack
, tp
->snd_nxt
))
3600 if (after(ack
, prior_snd_una
))
3601 flag
|= FLAG_SND_UNA_ADVANCED
;
3603 if (sysctl_tcp_abc
) {
3604 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3605 tp
->bytes_acked
+= ack
- prior_snd_una
;
3606 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3607 /* we assume just one segment left network */
3608 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3612 prior_fackets
= tp
->fackets_out
;
3613 prior_in_flight
= tcp_packets_in_flight(tp
);
3615 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3616 /* Window is constant, pure forward advance.
3617 * No more checks are required.
3618 * Note, we use the fact that SND.UNA>=SND.WL2.
3620 tcp_update_wl(tp
, ack_seq
);
3622 flag
|= FLAG_WIN_UPDATE
;
3624 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3626 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3628 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3631 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3633 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3635 if (TCP_SKB_CB(skb
)->sacked
)
3636 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3638 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3641 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3644 /* We passed data and got it acked, remove any soft error
3645 * log. Something worked...
3647 sk
->sk_err_soft
= 0;
3648 icsk
->icsk_probes_out
= 0;
3649 tp
->rcv_tstamp
= tcp_time_stamp
;
3650 prior_packets
= tp
->packets_out
;
3654 /* See if we can take anything off of the retransmit queue. */
3655 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3657 if (tp
->frto_counter
)
3658 frto_cwnd
= tcp_process_frto(sk
, flag
);
3659 /* Guarantee sacktag reordering detection against wrap-arounds */
3660 if (before(tp
->frto_highmark
, tp
->snd_una
))
3661 tp
->frto_highmark
= 0;
3663 if (tcp_ack_is_dubious(sk
, flag
)) {
3664 /* Advance CWND, if state allows this. */
3665 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3666 tcp_may_raise_cwnd(sk
, flag
))
3667 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3668 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3671 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3672 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3675 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3676 dst_confirm(sk
->sk_dst_cache
);
3681 /* If this ack opens up a zero window, clear backoff. It was
3682 * being used to time the probes, and is probably far higher than
3683 * it needs to be for normal retransmission.
3685 if (tcp_send_head(sk
))
3690 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3694 if (TCP_SKB_CB(skb
)->sacked
) {
3695 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3696 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3697 tcp_try_keep_open(sk
);
3700 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3704 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3705 * But, this can also be called on packets in the established flow when
3706 * the fast version below fails.
3708 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3712 struct tcphdr
*th
= tcp_hdr(skb
);
3713 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3715 ptr
= (unsigned char *)(th
+ 1);
3716 opt_rx
->saw_tstamp
= 0;
3718 while (length
> 0) {
3719 int opcode
= *ptr
++;
3725 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3730 if (opsize
< 2) /* "silly options" */
3732 if (opsize
> length
)
3733 return; /* don't parse partial options */
3736 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3737 u16 in_mss
= get_unaligned_be16(ptr
);
3739 if (opt_rx
->user_mss
&&
3740 opt_rx
->user_mss
< in_mss
)
3741 in_mss
= opt_rx
->user_mss
;
3742 opt_rx
->mss_clamp
= in_mss
;
3747 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3748 !estab
&& sysctl_tcp_window_scaling
) {
3749 __u8 snd_wscale
= *(__u8
*)ptr
;
3750 opt_rx
->wscale_ok
= 1;
3751 if (snd_wscale
> 14) {
3752 if (net_ratelimit())
3753 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3754 "scaling value %d >14 received.\n",
3758 opt_rx
->snd_wscale
= snd_wscale
;
3761 case TCPOPT_TIMESTAMP
:
3762 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3763 ((estab
&& opt_rx
->tstamp_ok
) ||
3764 (!estab
&& sysctl_tcp_timestamps
))) {
3765 opt_rx
->saw_tstamp
= 1;
3766 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3767 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3770 case TCPOPT_SACK_PERM
:
3771 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3772 !estab
&& sysctl_tcp_sack
) {
3773 opt_rx
->sack_ok
= 1;
3774 tcp_sack_reset(opt_rx
);
3779 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3780 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3782 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3785 #ifdef CONFIG_TCP_MD5SIG
3788 * The MD5 Hash has already been
3789 * checked (see tcp_v{4,6}_do_rcv()).
3801 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3803 __be32
*ptr
= (__be32
*)(th
+ 1);
3805 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3806 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3807 tp
->rx_opt
.saw_tstamp
= 1;
3809 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3811 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3817 /* Fast parse options. This hopes to only see timestamps.
3818 * If it is wrong it falls back on tcp_parse_options().
3820 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3821 struct tcp_sock
*tp
)
3823 if (th
->doff
== sizeof(struct tcphdr
) >> 2) {
3824 tp
->rx_opt
.saw_tstamp
= 0;
3826 } else if (tp
->rx_opt
.tstamp_ok
&&
3827 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3828 if (tcp_parse_aligned_timestamp(tp
, th
))
3831 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3835 #ifdef CONFIG_TCP_MD5SIG
3837 * Parse MD5 Signature option
3839 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3841 int length
= (th
->doff
<< 2) - sizeof (*th
);
3842 u8
*ptr
= (u8
*)(th
+ 1);
3844 /* If the TCP option is too short, we can short cut */
3845 if (length
< TCPOLEN_MD5SIG
)
3848 while (length
> 0) {
3849 int opcode
= *ptr
++;
3860 if (opsize
< 2 || opsize
> length
)
3862 if (opcode
== TCPOPT_MD5SIG
)
3872 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3874 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3875 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3878 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3880 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3881 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3882 * extra check below makes sure this can only happen
3883 * for pure ACK frames. -DaveM
3885 * Not only, also it occurs for expired timestamps.
3888 if (tcp_paws_check(&tp
->rx_opt
, 0))
3889 tcp_store_ts_recent(tp
);
3893 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3895 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3896 * it can pass through stack. So, the following predicate verifies that
3897 * this segment is not used for anything but congestion avoidance or
3898 * fast retransmit. Moreover, we even are able to eliminate most of such
3899 * second order effects, if we apply some small "replay" window (~RTO)
3900 * to timestamp space.
3902 * All these measures still do not guarantee that we reject wrapped ACKs
3903 * on networks with high bandwidth, when sequence space is recycled fastly,
3904 * but it guarantees that such events will be very rare and do not affect
3905 * connection seriously. This doesn't look nice, but alas, PAWS is really
3908 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3909 * states that events when retransmit arrives after original data are rare.
3910 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3911 * the biggest problem on large power networks even with minor reordering.
3912 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3913 * up to bandwidth of 18Gigabit/sec. 8) ]
3916 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3918 struct tcp_sock
*tp
= tcp_sk(sk
);
3919 struct tcphdr
*th
= tcp_hdr(skb
);
3920 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3921 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3923 return (/* 1. Pure ACK with correct sequence number. */
3924 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3926 /* 2. ... and duplicate ACK. */
3927 ack
== tp
->snd_una
&&
3929 /* 3. ... and does not update window. */
3930 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3932 /* 4. ... and sits in replay window. */
3933 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3936 static inline int tcp_paws_discard(const struct sock
*sk
,
3937 const struct sk_buff
*skb
)
3939 const struct tcp_sock
*tp
= tcp_sk(sk
);
3941 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3942 !tcp_disordered_ack(sk
, skb
);
3945 /* Check segment sequence number for validity.
3947 * Segment controls are considered valid, if the segment
3948 * fits to the window after truncation to the window. Acceptability
3949 * of data (and SYN, FIN, of course) is checked separately.
3950 * See tcp_data_queue(), for example.
3952 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3953 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3954 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3955 * (borrowed from freebsd)
3958 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3960 return !before(end_seq
, tp
->rcv_wup
) &&
3961 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3964 /* When we get a reset we do this. */
3965 static void tcp_reset(struct sock
*sk
)
3967 /* We want the right error as BSD sees it (and indeed as we do). */
3968 switch (sk
->sk_state
) {
3970 sk
->sk_err
= ECONNREFUSED
;
3972 case TCP_CLOSE_WAIT
:
3978 sk
->sk_err
= ECONNRESET
;
3981 if (!sock_flag(sk
, SOCK_DEAD
))
3982 sk
->sk_error_report(sk
);
3988 * Process the FIN bit. This now behaves as it is supposed to work
3989 * and the FIN takes effect when it is validly part of sequence
3990 * space. Not before when we get holes.
3992 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3993 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3996 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3997 * close and we go into CLOSING (and later onto TIME-WAIT)
3999 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4001 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
4003 struct tcp_sock
*tp
= tcp_sk(sk
);
4005 inet_csk_schedule_ack(sk
);
4007 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4008 sock_set_flag(sk
, SOCK_DONE
);
4010 switch (sk
->sk_state
) {
4012 case TCP_ESTABLISHED
:
4013 /* Move to CLOSE_WAIT */
4014 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4015 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4018 case TCP_CLOSE_WAIT
:
4020 /* Received a retransmission of the FIN, do
4025 /* RFC793: Remain in the LAST-ACK state. */
4029 /* This case occurs when a simultaneous close
4030 * happens, we must ack the received FIN and
4031 * enter the CLOSING state.
4034 tcp_set_state(sk
, TCP_CLOSING
);
4037 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4039 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4042 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4043 * cases we should never reach this piece of code.
4045 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
4046 __func__
, sk
->sk_state
);
4050 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4051 * Probably, we should reset in this case. For now drop them.
4053 __skb_queue_purge(&tp
->out_of_order_queue
);
4054 if (tcp_is_sack(tp
))
4055 tcp_sack_reset(&tp
->rx_opt
);
4058 if (!sock_flag(sk
, SOCK_DEAD
)) {
4059 sk
->sk_state_change(sk
);
4061 /* Do not send POLL_HUP for half duplex close. */
4062 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4063 sk
->sk_state
== TCP_CLOSE
)
4064 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4066 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4070 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4073 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4074 if (before(seq
, sp
->start_seq
))
4075 sp
->start_seq
= seq
;
4076 if (after(end_seq
, sp
->end_seq
))
4077 sp
->end_seq
= end_seq
;
4083 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4085 struct tcp_sock
*tp
= tcp_sk(sk
);
4087 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4090 if (before(seq
, tp
->rcv_nxt
))
4091 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4093 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4095 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4097 tp
->rx_opt
.dsack
= 1;
4098 tp
->duplicate_sack
[0].start_seq
= seq
;
4099 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4103 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4105 struct tcp_sock
*tp
= tcp_sk(sk
);
4107 if (!tp
->rx_opt
.dsack
)
4108 tcp_dsack_set(sk
, seq
, end_seq
);
4110 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4113 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
4115 struct tcp_sock
*tp
= tcp_sk(sk
);
4117 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4118 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4119 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4120 tcp_enter_quickack_mode(sk
);
4122 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4123 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4125 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4126 end_seq
= tp
->rcv_nxt
;
4127 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4134 /* These routines update the SACK block as out-of-order packets arrive or
4135 * in-order packets close up the sequence space.
4137 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4140 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4141 struct tcp_sack_block
*swalk
= sp
+ 1;
4143 /* See if the recent change to the first SACK eats into
4144 * or hits the sequence space of other SACK blocks, if so coalesce.
4146 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4147 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4150 /* Zap SWALK, by moving every further SACK up by one slot.
4151 * Decrease num_sacks.
4153 tp
->rx_opt
.num_sacks
--;
4154 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4158 this_sack
++, swalk
++;
4162 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4164 struct tcp_sock
*tp
= tcp_sk(sk
);
4165 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4166 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4172 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4173 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4174 /* Rotate this_sack to the first one. */
4175 for (; this_sack
> 0; this_sack
--, sp
--)
4176 swap(*sp
, *(sp
- 1));
4178 tcp_sack_maybe_coalesce(tp
);
4183 /* Could not find an adjacent existing SACK, build a new one,
4184 * put it at the front, and shift everyone else down. We
4185 * always know there is at least one SACK present already here.
4187 * If the sack array is full, forget about the last one.
4189 if (this_sack
>= TCP_NUM_SACKS
) {
4191 tp
->rx_opt
.num_sacks
--;
4194 for (; this_sack
> 0; this_sack
--, sp
--)
4198 /* Build the new head SACK, and we're done. */
4199 sp
->start_seq
= seq
;
4200 sp
->end_seq
= end_seq
;
4201 tp
->rx_opt
.num_sacks
++;
4204 /* RCV.NXT advances, some SACKs should be eaten. */
4206 static void tcp_sack_remove(struct tcp_sock
*tp
)
4208 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4209 int num_sacks
= tp
->rx_opt
.num_sacks
;
4212 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4213 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4214 tp
->rx_opt
.num_sacks
= 0;
4218 for (this_sack
= 0; this_sack
< num_sacks
;) {
4219 /* Check if the start of the sack is covered by RCV.NXT. */
4220 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4223 /* RCV.NXT must cover all the block! */
4224 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4226 /* Zap this SACK, by moving forward any other SACKS. */
4227 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4228 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4235 tp
->rx_opt
.num_sacks
= num_sacks
;
4238 /* This one checks to see if we can put data from the
4239 * out_of_order queue into the receive_queue.
4241 static void tcp_ofo_queue(struct sock
*sk
)
4243 struct tcp_sock
*tp
= tcp_sk(sk
);
4244 __u32 dsack_high
= tp
->rcv_nxt
;
4245 struct sk_buff
*skb
;
4247 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4248 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4251 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4252 __u32 dsack
= dsack_high
;
4253 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4254 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4255 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4258 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4259 SOCK_DEBUG(sk
, "ofo packet was already received \n");
4260 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4264 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4265 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4266 TCP_SKB_CB(skb
)->end_seq
);
4268 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4269 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4270 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4271 if (tcp_hdr(skb
)->fin
)
4272 tcp_fin(skb
, sk
, tcp_hdr(skb
));
4276 static int tcp_prune_ofo_queue(struct sock
*sk
);
4277 static int tcp_prune_queue(struct sock
*sk
);
4279 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
4281 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4282 !sk_rmem_schedule(sk
, size
)) {
4284 if (tcp_prune_queue(sk
) < 0)
4287 if (!sk_rmem_schedule(sk
, size
)) {
4288 if (!tcp_prune_ofo_queue(sk
))
4291 if (!sk_rmem_schedule(sk
, size
))
4298 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4300 struct tcphdr
*th
= tcp_hdr(skb
);
4301 struct tcp_sock
*tp
= tcp_sk(sk
);
4304 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4307 __skb_pull(skb
, th
->doff
* 4);
4309 TCP_ECN_accept_cwr(tp
, skb
);
4311 tp
->rx_opt
.dsack
= 0;
4313 /* Queue data for delivery to the user.
4314 * Packets in sequence go to the receive queue.
4315 * Out of sequence packets to the out_of_order_queue.
4317 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4318 if (tcp_receive_window(tp
) == 0)
4321 /* Ok. In sequence. In window. */
4322 if (tp
->ucopy
.task
== current
&&
4323 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4324 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4325 int chunk
= min_t(unsigned int, skb
->len
,
4328 __set_current_state(TASK_RUNNING
);
4331 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4332 tp
->ucopy
.len
-= chunk
;
4333 tp
->copied_seq
+= chunk
;
4334 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4335 tcp_rcv_space_adjust(sk
);
4343 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4346 skb_set_owner_r(skb
, sk
);
4347 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4349 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4351 tcp_event_data_recv(sk
, skb
);
4353 tcp_fin(skb
, sk
, th
);
4355 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4358 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4359 * gap in queue is filled.
4361 if (skb_queue_empty(&tp
->out_of_order_queue
))
4362 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4365 if (tp
->rx_opt
.num_sacks
)
4366 tcp_sack_remove(tp
);
4368 tcp_fast_path_check(sk
);
4372 else if (!sock_flag(sk
, SOCK_DEAD
))
4373 sk
->sk_data_ready(sk
, 0);
4377 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4378 /* A retransmit, 2nd most common case. Force an immediate ack. */
4379 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4380 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4383 tcp_enter_quickack_mode(sk
);
4384 inet_csk_schedule_ack(sk
);
4390 /* Out of window. F.e. zero window probe. */
4391 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4394 tcp_enter_quickack_mode(sk
);
4396 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4397 /* Partial packet, seq < rcv_next < end_seq */
4398 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4399 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4400 TCP_SKB_CB(skb
)->end_seq
);
4402 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4404 /* If window is closed, drop tail of packet. But after
4405 * remembering D-SACK for its head made in previous line.
4407 if (!tcp_receive_window(tp
))
4412 TCP_ECN_check_ce(tp
, skb
);
4414 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4417 /* Disable header prediction. */
4419 inet_csk_schedule_ack(sk
);
4421 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4422 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4424 skb_set_owner_r(skb
, sk
);
4426 if (!skb_peek(&tp
->out_of_order_queue
)) {
4427 /* Initial out of order segment, build 1 SACK. */
4428 if (tcp_is_sack(tp
)) {
4429 tp
->rx_opt
.num_sacks
= 1;
4430 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4431 tp
->selective_acks
[0].end_seq
=
4432 TCP_SKB_CB(skb
)->end_seq
;
4434 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4436 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
4437 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4438 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4440 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4441 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4443 if (!tp
->rx_opt
.num_sacks
||
4444 tp
->selective_acks
[0].end_seq
!= seq
)
4447 /* Common case: data arrive in order after hole. */
4448 tp
->selective_acks
[0].end_seq
= end_seq
;
4452 /* Find place to insert this segment. */
4454 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4456 } while ((skb1
= skb1
->prev
) !=
4457 (struct sk_buff
*)&tp
->out_of_order_queue
);
4459 /* Do skb overlap to previous one? */
4460 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
4461 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4462 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4463 /* All the bits are present. Drop. */
4465 tcp_dsack_set(sk
, seq
, end_seq
);
4468 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4469 /* Partial overlap. */
4470 tcp_dsack_set(sk
, seq
,
4471 TCP_SKB_CB(skb1
)->end_seq
);
4476 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4478 /* And clean segments covered by new one as whole. */
4479 while ((skb1
= skb
->next
) !=
4480 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4481 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4482 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4483 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4487 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4488 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4489 TCP_SKB_CB(skb1
)->end_seq
);
4494 if (tcp_is_sack(tp
))
4495 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4499 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4500 struct sk_buff_head
*list
)
4502 struct sk_buff
*next
= skb
->next
;
4504 __skb_unlink(skb
, list
);
4506 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4511 /* Collapse contiguous sequence of skbs head..tail with
4512 * sequence numbers start..end.
4513 * Segments with FIN/SYN are not collapsed (only because this
4517 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4518 struct sk_buff
*head
, struct sk_buff
*tail
,
4521 struct sk_buff
*skb
;
4523 /* First, check that queue is collapsible and find
4524 * the point where collapsing can be useful. */
4525 for (skb
= head
; skb
!= tail
;) {
4526 /* No new bits? It is possible on ofo queue. */
4527 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4528 skb
= tcp_collapse_one(sk
, skb
, list
);
4532 /* The first skb to collapse is:
4534 * - bloated or contains data before "start" or
4535 * overlaps to the next one.
4537 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4538 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4539 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4540 (skb
->next
!= tail
&&
4541 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4544 /* Decided to skip this, advance start seq. */
4545 start
= TCP_SKB_CB(skb
)->end_seq
;
4548 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4551 while (before(start
, end
)) {
4552 struct sk_buff
*nskb
;
4553 unsigned int header
= skb_headroom(skb
);
4554 int copy
= SKB_MAX_ORDER(header
, 0);
4556 /* Too big header? This can happen with IPv6. */
4559 if (end
- start
< copy
)
4561 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4565 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4566 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4568 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4570 skb_reserve(nskb
, header
);
4571 memcpy(nskb
->head
, skb
->head
, header
);
4572 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4573 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4574 __skb_queue_before(list
, skb
, nskb
);
4575 skb_set_owner_r(nskb
, sk
);
4577 /* Copy data, releasing collapsed skbs. */
4579 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4580 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4584 size
= min(copy
, size
);
4585 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4587 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4591 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4592 skb
= tcp_collapse_one(sk
, skb
, list
);
4594 tcp_hdr(skb
)->syn
||
4602 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4603 * and tcp_collapse() them until all the queue is collapsed.
4605 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4607 struct tcp_sock
*tp
= tcp_sk(sk
);
4608 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4609 struct sk_buff
*head
;
4615 start
= TCP_SKB_CB(skb
)->seq
;
4616 end
= TCP_SKB_CB(skb
)->end_seq
;
4622 /* Segment is terminated when we see gap or when
4623 * we are at the end of all the queue. */
4624 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4625 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4626 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4627 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4628 head
, skb
, start
, end
);
4630 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4632 /* Start new segment */
4633 start
= TCP_SKB_CB(skb
)->seq
;
4634 end
= TCP_SKB_CB(skb
)->end_seq
;
4636 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4637 start
= TCP_SKB_CB(skb
)->seq
;
4638 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4639 end
= TCP_SKB_CB(skb
)->end_seq
;
4645 * Purge the out-of-order queue.
4646 * Return true if queue was pruned.
4648 static int tcp_prune_ofo_queue(struct sock
*sk
)
4650 struct tcp_sock
*tp
= tcp_sk(sk
);
4653 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4654 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4655 __skb_queue_purge(&tp
->out_of_order_queue
);
4657 /* Reset SACK state. A conforming SACK implementation will
4658 * do the same at a timeout based retransmit. When a connection
4659 * is in a sad state like this, we care only about integrity
4660 * of the connection not performance.
4662 if (tp
->rx_opt
.sack_ok
)
4663 tcp_sack_reset(&tp
->rx_opt
);
4670 /* Reduce allocated memory if we can, trying to get
4671 * the socket within its memory limits again.
4673 * Return less than zero if we should start dropping frames
4674 * until the socket owning process reads some of the data
4675 * to stabilize the situation.
4677 static int tcp_prune_queue(struct sock
*sk
)
4679 struct tcp_sock
*tp
= tcp_sk(sk
);
4681 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4683 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4685 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4686 tcp_clamp_window(sk
);
4687 else if (tcp_memory_pressure
)
4688 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4690 tcp_collapse_ofo_queue(sk
);
4691 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4692 sk
->sk_receive_queue
.next
,
4693 (struct sk_buff
*)&sk
->sk_receive_queue
,
4694 tp
->copied_seq
, tp
->rcv_nxt
);
4697 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4700 /* Collapsing did not help, destructive actions follow.
4701 * This must not ever occur. */
4703 tcp_prune_ofo_queue(sk
);
4705 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4708 /* If we are really being abused, tell the caller to silently
4709 * drop receive data on the floor. It will get retransmitted
4710 * and hopefully then we'll have sufficient space.
4712 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4714 /* Massive buffer overcommit. */
4719 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4720 * As additional protections, we do not touch cwnd in retransmission phases,
4721 * and if application hit its sndbuf limit recently.
4723 void tcp_cwnd_application_limited(struct sock
*sk
)
4725 struct tcp_sock
*tp
= tcp_sk(sk
);
4727 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4728 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4729 /* Limited by application or receiver window. */
4730 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4731 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4732 if (win_used
< tp
->snd_cwnd
) {
4733 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4734 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4736 tp
->snd_cwnd_used
= 0;
4738 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4741 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4743 struct tcp_sock
*tp
= tcp_sk(sk
);
4745 /* If the user specified a specific send buffer setting, do
4748 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4751 /* If we are under global TCP memory pressure, do not expand. */
4752 if (tcp_memory_pressure
)
4755 /* If we are under soft global TCP memory pressure, do not expand. */
4756 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4759 /* If we filled the congestion window, do not expand. */
4760 if (tp
->packets_out
>= tp
->snd_cwnd
)
4766 /* When incoming ACK allowed to free some skb from write_queue,
4767 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4768 * on the exit from tcp input handler.
4770 * PROBLEM: sndbuf expansion does not work well with largesend.
4772 static void tcp_new_space(struct sock
*sk
)
4774 struct tcp_sock
*tp
= tcp_sk(sk
);
4776 if (tcp_should_expand_sndbuf(sk
)) {
4777 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4778 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
4779 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4780 tp
->reordering
+ 1);
4781 sndmem
*= 2 * demanded
;
4782 if (sndmem
> sk
->sk_sndbuf
)
4783 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4784 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4787 sk
->sk_write_space(sk
);
4790 static void tcp_check_space(struct sock
*sk
)
4792 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4793 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4794 if (sk
->sk_socket
&&
4795 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4800 static inline void tcp_data_snd_check(struct sock
*sk
)
4802 tcp_push_pending_frames(sk
);
4803 tcp_check_space(sk
);
4807 * Check if sending an ack is needed.
4809 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4811 struct tcp_sock
*tp
= tcp_sk(sk
);
4813 /* More than one full frame received... */
4814 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4815 /* ... and right edge of window advances far enough.
4816 * (tcp_recvmsg() will send ACK otherwise). Or...
4818 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4819 /* We ACK each frame or... */
4820 tcp_in_quickack_mode(sk
) ||
4821 /* We have out of order data. */
4822 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4823 /* Then ack it now */
4826 /* Else, send delayed ack. */
4827 tcp_send_delayed_ack(sk
);
4831 static inline void tcp_ack_snd_check(struct sock
*sk
)
4833 if (!inet_csk_ack_scheduled(sk
)) {
4834 /* We sent a data segment already. */
4837 __tcp_ack_snd_check(sk
, 1);
4841 * This routine is only called when we have urgent data
4842 * signaled. Its the 'slow' part of tcp_urg. It could be
4843 * moved inline now as tcp_urg is only called from one
4844 * place. We handle URGent data wrong. We have to - as
4845 * BSD still doesn't use the correction from RFC961.
4846 * For 1003.1g we should support a new option TCP_STDURG to permit
4847 * either form (or just set the sysctl tcp_stdurg).
4850 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4852 struct tcp_sock
*tp
= tcp_sk(sk
);
4853 u32 ptr
= ntohs(th
->urg_ptr
);
4855 if (ptr
&& !sysctl_tcp_stdurg
)
4857 ptr
+= ntohl(th
->seq
);
4859 /* Ignore urgent data that we've already seen and read. */
4860 if (after(tp
->copied_seq
, ptr
))
4863 /* Do not replay urg ptr.
4865 * NOTE: interesting situation not covered by specs.
4866 * Misbehaving sender may send urg ptr, pointing to segment,
4867 * which we already have in ofo queue. We are not able to fetch
4868 * such data and will stay in TCP_URG_NOTYET until will be eaten
4869 * by recvmsg(). Seems, we are not obliged to handle such wicked
4870 * situations. But it is worth to think about possibility of some
4871 * DoSes using some hypothetical application level deadlock.
4873 if (before(ptr
, tp
->rcv_nxt
))
4876 /* Do we already have a newer (or duplicate) urgent pointer? */
4877 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4880 /* Tell the world about our new urgent pointer. */
4883 /* We may be adding urgent data when the last byte read was
4884 * urgent. To do this requires some care. We cannot just ignore
4885 * tp->copied_seq since we would read the last urgent byte again
4886 * as data, nor can we alter copied_seq until this data arrives
4887 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4889 * NOTE. Double Dutch. Rendering to plain English: author of comment
4890 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4891 * and expect that both A and B disappear from stream. This is _wrong_.
4892 * Though this happens in BSD with high probability, this is occasional.
4893 * Any application relying on this is buggy. Note also, that fix "works"
4894 * only in this artificial test. Insert some normal data between A and B and we will
4895 * decline of BSD again. Verdict: it is better to remove to trap
4898 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4899 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4900 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4902 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4903 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4908 tp
->urg_data
= TCP_URG_NOTYET
;
4911 /* Disable header prediction. */
4915 /* This is the 'fast' part of urgent handling. */
4916 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4918 struct tcp_sock
*tp
= tcp_sk(sk
);
4920 /* Check if we get a new urgent pointer - normally not. */
4922 tcp_check_urg(sk
, th
);
4924 /* Do we wait for any urgent data? - normally not... */
4925 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4926 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4929 /* Is the urgent pointer pointing into this packet? */
4930 if (ptr
< skb
->len
) {
4932 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4934 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4935 if (!sock_flag(sk
, SOCK_DEAD
))
4936 sk
->sk_data_ready(sk
, 0);
4941 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4943 struct tcp_sock
*tp
= tcp_sk(sk
);
4944 int chunk
= skb
->len
- hlen
;
4948 if (skb_csum_unnecessary(skb
))
4949 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4951 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4955 tp
->ucopy
.len
-= chunk
;
4956 tp
->copied_seq
+= chunk
;
4957 tcp_rcv_space_adjust(sk
);
4964 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4965 struct sk_buff
*skb
)
4969 if (sock_owned_by_user(sk
)) {
4971 result
= __tcp_checksum_complete(skb
);
4974 result
= __tcp_checksum_complete(skb
);
4979 static inline int tcp_checksum_complete_user(struct sock
*sk
,
4980 struct sk_buff
*skb
)
4982 return !skb_csum_unnecessary(skb
) &&
4983 __tcp_checksum_complete_user(sk
, skb
);
4986 #ifdef CONFIG_NET_DMA
4987 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4990 struct tcp_sock
*tp
= tcp_sk(sk
);
4991 int chunk
= skb
->len
- hlen
;
4993 int copied_early
= 0;
4995 if (tp
->ucopy
.wakeup
)
4998 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4999 tp
->ucopy
.dma_chan
= dma_find_channel(DMA_MEMCPY
);
5001 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5003 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5005 tp
->ucopy
.iov
, chunk
,
5006 tp
->ucopy
.pinned_list
);
5011 tp
->ucopy
.dma_cookie
= dma_cookie
;
5014 tp
->ucopy
.len
-= chunk
;
5015 tp
->copied_seq
+= chunk
;
5016 tcp_rcv_space_adjust(sk
);
5018 if ((tp
->ucopy
.len
== 0) ||
5019 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5020 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5021 tp
->ucopy
.wakeup
= 1;
5022 sk
->sk_data_ready(sk
, 0);
5024 } else if (chunk
> 0) {
5025 tp
->ucopy
.wakeup
= 1;
5026 sk
->sk_data_ready(sk
, 0);
5029 return copied_early
;
5031 #endif /* CONFIG_NET_DMA */
5033 /* Does PAWS and seqno based validation of an incoming segment, flags will
5034 * play significant role here.
5036 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5037 struct tcphdr
*th
, int syn_inerr
)
5039 struct tcp_sock
*tp
= tcp_sk(sk
);
5041 /* RFC1323: H1. Apply PAWS check first. */
5042 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5043 tcp_paws_discard(sk
, skb
)) {
5045 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5046 tcp_send_dupack(sk
, skb
);
5049 /* Reset is accepted even if it did not pass PAWS. */
5052 /* Step 1: check sequence number */
5053 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5054 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5055 * (RST) segments are validated by checking their SEQ-fields."
5056 * And page 69: "If an incoming segment is not acceptable,
5057 * an acknowledgment should be sent in reply (unless the RST
5058 * bit is set, if so drop the segment and return)".
5061 tcp_send_dupack(sk
, skb
);
5065 /* Step 2: check RST bit */
5071 /* ts_recent update must be made after we are sure that the packet
5074 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5076 /* step 3: check security and precedence [ignored] */
5078 /* step 4: Check for a SYN in window. */
5079 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5081 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5082 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5095 * TCP receive function for the ESTABLISHED state.
5097 * It is split into a fast path and a slow path. The fast path is
5099 * - A zero window was announced from us - zero window probing
5100 * is only handled properly in the slow path.
5101 * - Out of order segments arrived.
5102 * - Urgent data is expected.
5103 * - There is no buffer space left
5104 * - Unexpected TCP flags/window values/header lengths are received
5105 * (detected by checking the TCP header against pred_flags)
5106 * - Data is sent in both directions. Fast path only supports pure senders
5107 * or pure receivers (this means either the sequence number or the ack
5108 * value must stay constant)
5109 * - Unexpected TCP option.
5111 * When these conditions are not satisfied it drops into a standard
5112 * receive procedure patterned after RFC793 to handle all cases.
5113 * The first three cases are guaranteed by proper pred_flags setting,
5114 * the rest is checked inline. Fast processing is turned on in
5115 * tcp_data_queue when everything is OK.
5117 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5118 struct tcphdr
*th
, unsigned len
)
5120 struct tcp_sock
*tp
= tcp_sk(sk
);
5124 * Header prediction.
5125 * The code loosely follows the one in the famous
5126 * "30 instruction TCP receive" Van Jacobson mail.
5128 * Van's trick is to deposit buffers into socket queue
5129 * on a device interrupt, to call tcp_recv function
5130 * on the receive process context and checksum and copy
5131 * the buffer to user space. smart...
5133 * Our current scheme is not silly either but we take the
5134 * extra cost of the net_bh soft interrupt processing...
5135 * We do checksum and copy also but from device to kernel.
5138 tp
->rx_opt
.saw_tstamp
= 0;
5140 /* pred_flags is 0xS?10 << 16 + snd_wnd
5141 * if header_prediction is to be made
5142 * 'S' will always be tp->tcp_header_len >> 2
5143 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5144 * turn it off (when there are holes in the receive
5145 * space for instance)
5146 * PSH flag is ignored.
5149 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5150 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5151 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5152 int tcp_header_len
= tp
->tcp_header_len
;
5154 /* Timestamp header prediction: tcp_header_len
5155 * is automatically equal to th->doff*4 due to pred_flags
5159 /* Check timestamp */
5160 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5161 /* No? Slow path! */
5162 if (!tcp_parse_aligned_timestamp(tp
, th
))
5165 /* If PAWS failed, check it more carefully in slow path */
5166 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5169 /* DO NOT update ts_recent here, if checksum fails
5170 * and timestamp was corrupted part, it will result
5171 * in a hung connection since we will drop all
5172 * future packets due to the PAWS test.
5176 if (len
<= tcp_header_len
) {
5177 /* Bulk data transfer: sender */
5178 if (len
== tcp_header_len
) {
5179 /* Predicted packet is in window by definition.
5180 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5181 * Hence, check seq<=rcv_wup reduces to:
5183 if (tcp_header_len
==
5184 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5185 tp
->rcv_nxt
== tp
->rcv_wup
)
5186 tcp_store_ts_recent(tp
);
5188 /* We know that such packets are checksummed
5191 tcp_ack(sk
, skb
, 0);
5193 tcp_data_snd_check(sk
);
5195 } else { /* Header too small */
5196 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5201 int copied_early
= 0;
5203 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5204 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5205 #ifdef CONFIG_NET_DMA
5206 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5211 if (tp
->ucopy
.task
== current
&&
5212 sock_owned_by_user(sk
) && !copied_early
) {
5213 __set_current_state(TASK_RUNNING
);
5215 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5219 /* Predicted packet is in window by definition.
5220 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5221 * Hence, check seq<=rcv_wup reduces to:
5223 if (tcp_header_len
==
5224 (sizeof(struct tcphdr
) +
5225 TCPOLEN_TSTAMP_ALIGNED
) &&
5226 tp
->rcv_nxt
== tp
->rcv_wup
)
5227 tcp_store_ts_recent(tp
);
5229 tcp_rcv_rtt_measure_ts(sk
, skb
);
5231 __skb_pull(skb
, tcp_header_len
);
5232 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5233 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5236 tcp_cleanup_rbuf(sk
, skb
->len
);
5239 if (tcp_checksum_complete_user(sk
, skb
))
5242 /* Predicted packet is in window by definition.
5243 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5244 * Hence, check seq<=rcv_wup reduces to:
5246 if (tcp_header_len
==
5247 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5248 tp
->rcv_nxt
== tp
->rcv_wup
)
5249 tcp_store_ts_recent(tp
);
5251 tcp_rcv_rtt_measure_ts(sk
, skb
);
5253 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5256 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5258 /* Bulk data transfer: receiver */
5259 __skb_pull(skb
, tcp_header_len
);
5260 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
5261 skb_set_owner_r(skb
, sk
);
5262 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5265 tcp_event_data_recv(sk
, skb
);
5267 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5268 /* Well, only one small jumplet in fast path... */
5269 tcp_ack(sk
, skb
, FLAG_DATA
);
5270 tcp_data_snd_check(sk
);
5271 if (!inet_csk_ack_scheduled(sk
))
5275 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5276 __tcp_ack_snd_check(sk
, 0);
5278 #ifdef CONFIG_NET_DMA
5280 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5286 sk
->sk_data_ready(sk
, 0);
5292 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5296 * Standard slow path.
5299 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
5304 if (th
->ack
&& tcp_ack(sk
, skb
, FLAG_SLOWPATH
) < 0)
5307 tcp_rcv_rtt_measure_ts(sk
, skb
);
5309 /* Process urgent data. */
5310 tcp_urg(sk
, skb
, th
);
5312 /* step 7: process the segment text */
5313 tcp_data_queue(sk
, skb
);
5315 tcp_data_snd_check(sk
);
5316 tcp_ack_snd_check(sk
);
5320 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5327 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5328 struct tcphdr
*th
, unsigned len
)
5330 struct tcp_sock
*tp
= tcp_sk(sk
);
5331 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5332 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5334 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
5338 * "If the state is SYN-SENT then
5339 * first check the ACK bit
5340 * If the ACK bit is set
5341 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5342 * a reset (unless the RST bit is set, if so drop
5343 * the segment and return)"
5345 * We do not send data with SYN, so that RFC-correct
5348 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5349 goto reset_and_undo
;
5351 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5352 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5354 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5355 goto reset_and_undo
;
5358 /* Now ACK is acceptable.
5360 * "If the RST bit is set
5361 * If the ACK was acceptable then signal the user "error:
5362 * connection reset", drop the segment, enter CLOSED state,
5363 * delete TCB, and return."
5372 * "fifth, if neither of the SYN or RST bits is set then
5373 * drop the segment and return."
5379 goto discard_and_undo
;
5382 * "If the SYN bit is on ...
5383 * are acceptable then ...
5384 * (our SYN has been ACKed), change the connection
5385 * state to ESTABLISHED..."
5388 TCP_ECN_rcv_synack(tp
, th
);
5390 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5391 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5393 /* Ok.. it's good. Set up sequence numbers and
5394 * move to established.
5396 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5397 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5399 /* RFC1323: The window in SYN & SYN/ACK segments is
5402 tp
->snd_wnd
= ntohs(th
->window
);
5403 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5405 if (!tp
->rx_opt
.wscale_ok
) {
5406 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5407 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5410 if (tp
->rx_opt
.saw_tstamp
) {
5411 tp
->rx_opt
.tstamp_ok
= 1;
5412 tp
->tcp_header_len
=
5413 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5414 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5415 tcp_store_ts_recent(tp
);
5417 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5420 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5421 tcp_enable_fack(tp
);
5424 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5425 tcp_initialize_rcv_mss(sk
);
5427 /* Remember, tcp_poll() does not lock socket!
5428 * Change state from SYN-SENT only after copied_seq
5429 * is initialized. */
5430 tp
->copied_seq
= tp
->rcv_nxt
;
5432 tcp_set_state(sk
, TCP_ESTABLISHED
);
5434 security_inet_conn_established(sk
, skb
);
5436 /* Make sure socket is routed, for correct metrics. */
5437 icsk
->icsk_af_ops
->rebuild_header(sk
);
5439 tcp_init_metrics(sk
);
5441 tcp_init_congestion_control(sk
);
5443 /* Prevent spurious tcp_cwnd_restart() on first data
5446 tp
->lsndtime
= tcp_time_stamp
;
5448 tcp_init_buffer_space(sk
);
5450 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5451 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5453 if (!tp
->rx_opt
.snd_wscale
)
5454 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5458 if (!sock_flag(sk
, SOCK_DEAD
)) {
5459 sk
->sk_state_change(sk
);
5460 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5463 if (sk
->sk_write_pending
||
5464 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5465 icsk
->icsk_ack
.pingpong
) {
5466 /* Save one ACK. Data will be ready after
5467 * several ticks, if write_pending is set.
5469 * It may be deleted, but with this feature tcpdumps
5470 * look so _wonderfully_ clever, that I was not able
5471 * to stand against the temptation 8) --ANK
5473 inet_csk_schedule_ack(sk
);
5474 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5475 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5476 tcp_incr_quickack(sk
);
5477 tcp_enter_quickack_mode(sk
);
5478 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5479 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5490 /* No ACK in the segment */
5494 * "If the RST bit is set
5496 * Otherwise (no ACK) drop the segment and return."
5499 goto discard_and_undo
;
5503 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5504 tcp_paws_reject(&tp
->rx_opt
, 0))
5505 goto discard_and_undo
;
5508 /* We see SYN without ACK. It is attempt of
5509 * simultaneous connect with crossed SYNs.
5510 * Particularly, it can be connect to self.
5512 tcp_set_state(sk
, TCP_SYN_RECV
);
5514 if (tp
->rx_opt
.saw_tstamp
) {
5515 tp
->rx_opt
.tstamp_ok
= 1;
5516 tcp_store_ts_recent(tp
);
5517 tp
->tcp_header_len
=
5518 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5520 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5523 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5524 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5526 /* RFC1323: The window in SYN & SYN/ACK segments is
5529 tp
->snd_wnd
= ntohs(th
->window
);
5530 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5531 tp
->max_window
= tp
->snd_wnd
;
5533 TCP_ECN_rcv_syn(tp
, th
);
5536 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5537 tcp_initialize_rcv_mss(sk
);
5539 tcp_send_synack(sk
);
5541 /* Note, we could accept data and URG from this segment.
5542 * There are no obstacles to make this.
5544 * However, if we ignore data in ACKless segments sometimes,
5545 * we have no reasons to accept it sometimes.
5546 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5547 * is not flawless. So, discard packet for sanity.
5548 * Uncomment this return to process the data.
5555 /* "fifth, if neither of the SYN or RST bits is set then
5556 * drop the segment and return."
5560 tcp_clear_options(&tp
->rx_opt
);
5561 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5565 tcp_clear_options(&tp
->rx_opt
);
5566 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5571 * This function implements the receiving procedure of RFC 793 for
5572 * all states except ESTABLISHED and TIME_WAIT.
5573 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5574 * address independent.
5577 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5578 struct tcphdr
*th
, unsigned len
)
5580 struct tcp_sock
*tp
= tcp_sk(sk
);
5581 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5585 tp
->rx_opt
.saw_tstamp
= 0;
5587 switch (sk
->sk_state
) {
5599 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5602 /* Now we have several options: In theory there is
5603 * nothing else in the frame. KA9Q has an option to
5604 * send data with the syn, BSD accepts data with the
5605 * syn up to the [to be] advertised window and
5606 * Solaris 2.1 gives you a protocol error. For now
5607 * we just ignore it, that fits the spec precisely
5608 * and avoids incompatibilities. It would be nice in
5609 * future to drop through and process the data.
5611 * Now that TTCP is starting to be used we ought to
5613 * But, this leaves one open to an easy denial of
5614 * service attack, and SYN cookies can't defend
5615 * against this problem. So, we drop the data
5616 * in the interest of security over speed unless
5617 * it's still in use.
5625 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5629 /* Do step6 onward by hand. */
5630 tcp_urg(sk
, skb
, th
);
5632 tcp_data_snd_check(sk
);
5636 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5640 /* step 5: check the ACK field */
5642 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
) > 0;
5644 switch (sk
->sk_state
) {
5647 tp
->copied_seq
= tp
->rcv_nxt
;
5649 tcp_set_state(sk
, TCP_ESTABLISHED
);
5650 sk
->sk_state_change(sk
);
5652 /* Note, that this wakeup is only for marginal
5653 * crossed SYN case. Passively open sockets
5654 * are not waked up, because sk->sk_sleep ==
5655 * NULL and sk->sk_socket == NULL.
5659 SOCK_WAKE_IO
, POLL_OUT
);
5661 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5662 tp
->snd_wnd
= ntohs(th
->window
) <<
5663 tp
->rx_opt
.snd_wscale
;
5664 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5666 /* tcp_ack considers this ACK as duplicate
5667 * and does not calculate rtt.
5668 * Fix it at least with timestamps.
5670 if (tp
->rx_opt
.saw_tstamp
&&
5671 tp
->rx_opt
.rcv_tsecr
&& !tp
->srtt
)
5672 tcp_ack_saw_tstamp(sk
, 0);
5674 if (tp
->rx_opt
.tstamp_ok
)
5675 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5677 /* Make sure socket is routed, for
5680 icsk
->icsk_af_ops
->rebuild_header(sk
);
5682 tcp_init_metrics(sk
);
5684 tcp_init_congestion_control(sk
);
5686 /* Prevent spurious tcp_cwnd_restart() on
5687 * first data packet.
5689 tp
->lsndtime
= tcp_time_stamp
;
5692 tcp_initialize_rcv_mss(sk
);
5693 tcp_init_buffer_space(sk
);
5694 tcp_fast_path_on(tp
);
5701 if (tp
->snd_una
== tp
->write_seq
) {
5702 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5703 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5704 dst_confirm(sk
->sk_dst_cache
);
5706 if (!sock_flag(sk
, SOCK_DEAD
))
5707 /* Wake up lingering close() */
5708 sk
->sk_state_change(sk
);
5712 if (tp
->linger2
< 0 ||
5713 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5714 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5716 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5720 tmo
= tcp_fin_time(sk
);
5721 if (tmo
> TCP_TIMEWAIT_LEN
) {
5722 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5723 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5724 /* Bad case. We could lose such FIN otherwise.
5725 * It is not a big problem, but it looks confusing
5726 * and not so rare event. We still can lose it now,
5727 * if it spins in bh_lock_sock(), but it is really
5730 inet_csk_reset_keepalive_timer(sk
, tmo
);
5732 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5740 if (tp
->snd_una
== tp
->write_seq
) {
5741 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5747 if (tp
->snd_una
== tp
->write_seq
) {
5748 tcp_update_metrics(sk
);
5757 /* step 6: check the URG bit */
5758 tcp_urg(sk
, skb
, th
);
5760 /* step 7: process the segment text */
5761 switch (sk
->sk_state
) {
5762 case TCP_CLOSE_WAIT
:
5765 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5769 /* RFC 793 says to queue data in these states,
5770 * RFC 1122 says we MUST send a reset.
5771 * BSD 4.4 also does reset.
5773 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5774 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5775 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5776 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5782 case TCP_ESTABLISHED
:
5783 tcp_data_queue(sk
, skb
);
5788 /* tcp_data could move socket to TIME-WAIT */
5789 if (sk
->sk_state
!= TCP_CLOSE
) {
5790 tcp_data_snd_check(sk
);
5791 tcp_ack_snd_check(sk
);
5801 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5802 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5803 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
5804 EXPORT_SYMBOL(tcp_parse_options
);
5805 #ifdef CONFIG_TCP_MD5SIG
5806 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
5808 EXPORT_SYMBOL(tcp_rcv_established
);
5809 EXPORT_SYMBOL(tcp_rcv_state_process
);
5810 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);