2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #include <linux/module.h>
66 #include <linux/sysctl.h>
69 #include <net/inet_common.h>
70 #include <linux/ipsec.h>
71 #include <asm/unaligned.h>
72 #include <net/netdma.h>
74 int sysctl_tcp_timestamps __read_mostly
= 1;
75 int sysctl_tcp_window_scaling __read_mostly
= 1;
76 int sysctl_tcp_sack __read_mostly
= 1;
77 int sysctl_tcp_fack __read_mostly
= 1;
78 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
79 int sysctl_tcp_ecn __read_mostly
;
80 int sysctl_tcp_dsack __read_mostly
= 1;
81 int sysctl_tcp_app_win __read_mostly
= 31;
82 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
84 int sysctl_tcp_stdurg __read_mostly
;
85 int sysctl_tcp_rfc1337 __read_mostly
;
86 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
87 int sysctl_tcp_frto __read_mostly
= 2;
88 int sysctl_tcp_frto_response __read_mostly
;
89 int sysctl_tcp_nometrics_save __read_mostly
;
91 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
92 int sysctl_tcp_abc __read_mostly
;
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
103 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
104 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
105 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
106 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
107 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
115 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
118 /* Adapt the MSS value used to make delayed ack decision to the
121 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
123 struct inet_connection_sock
*icsk
= inet_csk(sk
);
124 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
127 icsk
->icsk_ack
.last_seg_size
= 0;
129 /* skb->len may jitter because of SACKs, even if peer
130 * sends good full-sized frames.
132 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
133 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
134 icsk
->icsk_ack
.rcv_mss
= len
;
136 /* Otherwise, we make more careful check taking into account,
137 * that SACKs block is variable.
139 * "len" is invariant segment length, including TCP header.
141 len
+= skb
->data
- skb_transport_header(skb
);
142 if (len
>= TCP_MIN_RCVMSS
+ sizeof(struct tcphdr
) ||
143 /* If PSH is not set, packet should be
144 * full sized, provided peer TCP is not badly broken.
145 * This observation (if it is correct 8)) allows
146 * to handle super-low mtu links fairly.
148 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
149 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
150 /* Subtract also invariant (if peer is RFC compliant),
151 * tcp header plus fixed timestamp option length.
152 * Resulting "len" is MSS free of SACK jitter.
154 len
-= tcp_sk(sk
)->tcp_header_len
;
155 icsk
->icsk_ack
.last_seg_size
= len
;
157 icsk
->icsk_ack
.rcv_mss
= len
;
161 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
162 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
163 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
167 static void tcp_incr_quickack(struct sock
*sk
)
169 struct inet_connection_sock
*icsk
= inet_csk(sk
);
170 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
174 if (quickacks
> icsk
->icsk_ack
.quick
)
175 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
178 void tcp_enter_quickack_mode(struct sock
*sk
)
180 struct inet_connection_sock
*icsk
= inet_csk(sk
);
181 tcp_incr_quickack(sk
);
182 icsk
->icsk_ack
.pingpong
= 0;
183 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
186 /* Send ACKs quickly, if "quick" count is not exhausted
187 * and the session is not interactive.
190 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
192 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
193 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
196 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
198 if (tp
->ecn_flags
& TCP_ECN_OK
)
199 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
202 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, struct sk_buff
*skb
)
204 if (tcp_hdr(skb
)->cwr
)
205 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
208 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
210 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
213 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, struct sk_buff
*skb
)
215 if (tp
->ecn_flags
& TCP_ECN_OK
) {
216 if (INET_ECN_is_ce(TCP_SKB_CB(skb
)->flags
))
217 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
218 /* Funny extension: if ECT is not set on a segment,
219 * it is surely retransmit. It is not in ECN RFC,
220 * but Linux follows this rule. */
221 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb
)->flags
)))
222 tcp_enter_quickack_mode((struct sock
*)tp
);
226 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, struct tcphdr
*th
)
228 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
229 tp
->ecn_flags
&= ~TCP_ECN_OK
;
232 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, struct tcphdr
*th
)
234 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
235 tp
->ecn_flags
&= ~TCP_ECN_OK
;
238 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock
*tp
, struct tcphdr
*th
)
240 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
245 /* Buffer size and advertised window tuning.
247 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
250 static void tcp_fixup_sndbuf(struct sock
*sk
)
252 int sndmem
= tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
+ 16 +
253 sizeof(struct sk_buff
);
255 if (sk
->sk_sndbuf
< 3 * sndmem
)
256 sk
->sk_sndbuf
= min(3 * sndmem
, sysctl_tcp_wmem
[2]);
259 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261 * All tcp_full_space() is split to two parts: "network" buffer, allocated
262 * forward and advertised in receiver window (tp->rcv_wnd) and
263 * "application buffer", required to isolate scheduling/application
264 * latencies from network.
265 * window_clamp is maximal advertised window. It can be less than
266 * tcp_full_space(), in this case tcp_full_space() - window_clamp
267 * is reserved for "application" buffer. The less window_clamp is
268 * the smoother our behaviour from viewpoint of network, but the lower
269 * throughput and the higher sensitivity of the connection to losses. 8)
271 * rcv_ssthresh is more strict window_clamp used at "slow start"
272 * phase to predict further behaviour of this connection.
273 * It is used for two goals:
274 * - to enforce header prediction at sender, even when application
275 * requires some significant "application buffer". It is check #1.
276 * - to prevent pruning of receive queue because of misprediction
277 * of receiver window. Check #2.
279 * The scheme does not work when sender sends good segments opening
280 * window and then starts to feed us spaghetti. But it should work
281 * in common situations. Otherwise, we have to rely on queue collapsing.
284 /* Slow part of check#2. */
285 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
287 struct tcp_sock
*tp
= tcp_sk(sk
);
289 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
290 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
292 while (tp
->rcv_ssthresh
<= window
) {
293 if (truesize
<= skb
->len
)
294 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
302 static void tcp_grow_window(struct sock
*sk
, struct sk_buff
*skb
)
304 struct tcp_sock
*tp
= tcp_sk(sk
);
307 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
308 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
309 !tcp_memory_pressure
) {
312 /* Check #2. Increase window, if skb with such overhead
313 * will fit to rcvbuf in future.
315 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
316 incr
= 2 * tp
->advmss
;
318 incr
= __tcp_grow_window(sk
, skb
);
321 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
323 inet_csk(sk
)->icsk_ack
.quick
|= 1;
328 /* 3. Tuning rcvbuf, when connection enters established state. */
330 static void tcp_fixup_rcvbuf(struct sock
*sk
)
332 struct tcp_sock
*tp
= tcp_sk(sk
);
333 int rcvmem
= tp
->advmss
+ MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
335 /* Try to select rcvbuf so that 4 mss-sized segments
336 * will fit to window and corresponding skbs will fit to our rcvbuf.
337 * (was 3; 4 is minimum to allow fast retransmit to work.)
339 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
341 if (sk
->sk_rcvbuf
< 4 * rcvmem
)
342 sk
->sk_rcvbuf
= min(4 * rcvmem
, sysctl_tcp_rmem
[2]);
345 /* 4. Try to fixup all. It is made immediately after connection enters
348 static void tcp_init_buffer_space(struct sock
*sk
)
350 struct tcp_sock
*tp
= tcp_sk(sk
);
353 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
354 tcp_fixup_rcvbuf(sk
);
355 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
356 tcp_fixup_sndbuf(sk
);
358 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
360 maxwin
= tcp_full_space(sk
);
362 if (tp
->window_clamp
>= maxwin
) {
363 tp
->window_clamp
= maxwin
;
365 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
366 tp
->window_clamp
= max(maxwin
-
367 (maxwin
>> sysctl_tcp_app_win
),
371 /* Force reservation of one segment. */
372 if (sysctl_tcp_app_win
&&
373 tp
->window_clamp
> 2 * tp
->advmss
&&
374 tp
->window_clamp
+ tp
->advmss
> maxwin
)
375 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
377 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
378 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
381 /* 5. Recalculate window clamp after socket hit its memory bounds. */
382 static void tcp_clamp_window(struct sock
*sk
)
384 struct tcp_sock
*tp
= tcp_sk(sk
);
385 struct inet_connection_sock
*icsk
= inet_csk(sk
);
387 icsk
->icsk_ack
.quick
= 0;
389 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
390 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
391 !tcp_memory_pressure
&&
392 atomic_read(&tcp_memory_allocated
) < sysctl_tcp_mem
[0]) {
393 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
396 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
397 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
400 /* Initialize RCV_MSS value.
401 * RCV_MSS is an our guess about MSS used by the peer.
402 * We haven't any direct information about the MSS.
403 * It's better to underestimate the RCV_MSS rather than overestimate.
404 * Overestimations make us ACKing less frequently than needed.
405 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407 void tcp_initialize_rcv_mss(struct sock
*sk
)
409 struct tcp_sock
*tp
= tcp_sk(sk
);
410 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
412 hint
= min(hint
, tp
->rcv_wnd
/ 2);
413 hint
= min(hint
, TCP_MIN_RCVMSS
);
414 hint
= max(hint
, TCP_MIN_MSS
);
416 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
419 /* Receiver "autotuning" code.
421 * The algorithm for RTT estimation w/o timestamps is based on
422 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
423 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425 * More detail on this code can be found at
426 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
427 * though this reference is out of date. A new paper
430 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
432 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
438 if (new_sample
!= 0) {
439 /* If we sample in larger samples in the non-timestamp
440 * case, we could grossly overestimate the RTT especially
441 * with chatty applications or bulk transfer apps which
442 * are stalled on filesystem I/O.
444 * Also, since we are only going for a minimum in the
445 * non-timestamp case, we do not smooth things out
446 * else with timestamps disabled convergence takes too
450 m
-= (new_sample
>> 3);
452 } else if (m
< new_sample
)
455 /* No previous measure. */
459 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
460 tp
->rcv_rtt_est
.rtt
= new_sample
;
463 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
465 if (tp
->rcv_rtt_est
.time
== 0)
467 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
469 tcp_rcv_rtt_update(tp
, jiffies
- tp
->rcv_rtt_est
.time
, 1);
472 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
473 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
476 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
477 const struct sk_buff
*skb
)
479 struct tcp_sock
*tp
= tcp_sk(sk
);
480 if (tp
->rx_opt
.rcv_tsecr
&&
481 (TCP_SKB_CB(skb
)->end_seq
-
482 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
483 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
487 * This function should be called every time data is copied to user space.
488 * It calculates the appropriate TCP receive buffer space.
490 void tcp_rcv_space_adjust(struct sock
*sk
)
492 struct tcp_sock
*tp
= tcp_sk(sk
);
496 if (tp
->rcvq_space
.time
== 0)
499 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
500 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
503 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
505 space
= max(tp
->rcvq_space
.space
, space
);
507 if (tp
->rcvq_space
.space
!= space
) {
510 tp
->rcvq_space
.space
= space
;
512 if (sysctl_tcp_moderate_rcvbuf
&&
513 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
514 int new_clamp
= space
;
516 /* Receive space grows, normalize in order to
517 * take into account packet headers and sk_buff
518 * structure overhead.
523 rcvmem
= (tp
->advmss
+ MAX_TCP_HEADER
+
524 16 + sizeof(struct sk_buff
));
525 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
528 space
= min(space
, sysctl_tcp_rmem
[2]);
529 if (space
> sk
->sk_rcvbuf
) {
530 sk
->sk_rcvbuf
= space
;
532 /* Make the window clamp follow along. */
533 tp
->window_clamp
= new_clamp
;
539 tp
->rcvq_space
.seq
= tp
->copied_seq
;
540 tp
->rcvq_space
.time
= tcp_time_stamp
;
543 /* There is something which you must keep in mind when you analyze the
544 * behavior of the tp->ato delayed ack timeout interval. When a
545 * connection starts up, we want to ack as quickly as possible. The
546 * problem is that "good" TCP's do slow start at the beginning of data
547 * transmission. The means that until we send the first few ACK's the
548 * sender will sit on his end and only queue most of his data, because
549 * he can only send snd_cwnd unacked packets at any given time. For
550 * each ACK we send, he increments snd_cwnd and transmits more of his
553 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
555 struct tcp_sock
*tp
= tcp_sk(sk
);
556 struct inet_connection_sock
*icsk
= inet_csk(sk
);
559 inet_csk_schedule_ack(sk
);
561 tcp_measure_rcv_mss(sk
, skb
);
563 tcp_rcv_rtt_measure(tp
);
565 now
= tcp_time_stamp
;
567 if (!icsk
->icsk_ack
.ato
) {
568 /* The _first_ data packet received, initialize
569 * delayed ACK engine.
571 tcp_incr_quickack(sk
);
572 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
574 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
576 if (m
<= TCP_ATO_MIN
/ 2) {
577 /* The fastest case is the first. */
578 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
579 } else if (m
< icsk
->icsk_ack
.ato
) {
580 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
581 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
582 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
583 } else if (m
> icsk
->icsk_rto
) {
584 /* Too long gap. Apparently sender failed to
585 * restart window, so that we send ACKs quickly.
587 tcp_incr_quickack(sk
);
591 icsk
->icsk_ack
.lrcvtime
= now
;
593 TCP_ECN_check_ce(tp
, skb
);
596 tcp_grow_window(sk
, skb
);
599 static u32
tcp_rto_min(struct sock
*sk
)
601 struct dst_entry
*dst
= __sk_dst_get(sk
);
602 u32 rto_min
= TCP_RTO_MIN
;
604 if (dst
&& dst_metric_locked(dst
, RTAX_RTO_MIN
))
605 rto_min
= dst_metric_rtt(dst
, RTAX_RTO_MIN
);
609 /* Called to compute a smoothed rtt estimate. The data fed to this
610 * routine either comes from timestamps, or from segments that were
611 * known _not_ to have been retransmitted [see Karn/Partridge
612 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
613 * piece by Van Jacobson.
614 * NOTE: the next three routines used to be one big routine.
615 * To save cycles in the RFC 1323 implementation it was better to break
616 * it up into three procedures. -- erics
618 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
620 struct tcp_sock
*tp
= tcp_sk(sk
);
621 long m
= mrtt
; /* RTT */
623 /* The following amusing code comes from Jacobson's
624 * article in SIGCOMM '88. Note that rtt and mdev
625 * are scaled versions of rtt and mean deviation.
626 * This is designed to be as fast as possible
627 * m stands for "measurement".
629 * On a 1990 paper the rto value is changed to:
630 * RTO = rtt + 4 * mdev
632 * Funny. This algorithm seems to be very broken.
633 * These formulae increase RTO, when it should be decreased, increase
634 * too slowly, when it should be increased quickly, decrease too quickly
635 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
636 * does not matter how to _calculate_ it. Seems, it was trap
637 * that VJ failed to avoid. 8)
642 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
643 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
645 m
= -m
; /* m is now abs(error) */
646 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
647 /* This is similar to one of Eifel findings.
648 * Eifel blocks mdev updates when rtt decreases.
649 * This solution is a bit different: we use finer gain
650 * for mdev in this case (alpha*beta).
651 * Like Eifel it also prevents growth of rto,
652 * but also it limits too fast rto decreases,
653 * happening in pure Eifel.
658 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
660 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
661 if (tp
->mdev
> tp
->mdev_max
) {
662 tp
->mdev_max
= tp
->mdev
;
663 if (tp
->mdev_max
> tp
->rttvar
)
664 tp
->rttvar
= tp
->mdev_max
;
666 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
667 if (tp
->mdev_max
< tp
->rttvar
)
668 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
669 tp
->rtt_seq
= tp
->snd_nxt
;
670 tp
->mdev_max
= tcp_rto_min(sk
);
673 /* no previous measure. */
674 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
675 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
676 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
677 tp
->rtt_seq
= tp
->snd_nxt
;
681 /* Calculate rto without backoff. This is the second half of Van Jacobson's
682 * routine referred to above.
684 static inline void tcp_set_rto(struct sock
*sk
)
686 const struct tcp_sock
*tp
= tcp_sk(sk
);
687 /* Old crap is replaced with new one. 8)
690 * 1. If rtt variance happened to be less 50msec, it is hallucination.
691 * It cannot be less due to utterly erratic ACK generation made
692 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
693 * to do with delayed acks, because at cwnd>2 true delack timeout
694 * is invisible. Actually, Linux-2.4 also generates erratic
695 * ACKs in some circumstances.
697 inet_csk(sk
)->icsk_rto
= (tp
->srtt
>> 3) + tp
->rttvar
;
699 /* 2. Fixups made earlier cannot be right.
700 * If we do not estimate RTO correctly without them,
701 * all the algo is pure shit and should be replaced
702 * with correct one. It is exactly, which we pretend to do.
706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707 * guarantees that rto is higher.
709 static inline void tcp_bound_rto(struct sock
*sk
)
711 if (inet_csk(sk
)->icsk_rto
> TCP_RTO_MAX
)
712 inet_csk(sk
)->icsk_rto
= TCP_RTO_MAX
;
715 /* Save metrics learned by this TCP session.
716 This function is called only, when TCP finishes successfully
717 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
719 void tcp_update_metrics(struct sock
*sk
)
721 struct tcp_sock
*tp
= tcp_sk(sk
);
722 struct dst_entry
*dst
= __sk_dst_get(sk
);
724 if (sysctl_tcp_nometrics_save
)
729 if (dst
&& (dst
->flags
& DST_HOST
)) {
730 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
734 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
735 /* This session failed to estimate rtt. Why?
736 * Probably, no packets returned in time.
739 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
740 dst
->metrics
[RTAX_RTT
- 1] = 0;
744 rtt
= dst_metric_rtt(dst
, RTAX_RTT
);
747 /* If newly calculated rtt larger than stored one,
748 * store new one. Otherwise, use EWMA. Remember,
749 * rtt overestimation is always better than underestimation.
751 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
753 set_dst_metric_rtt(dst
, RTAX_RTT
, tp
->srtt
);
755 set_dst_metric_rtt(dst
, RTAX_RTT
, rtt
- (m
>> 3));
758 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
763 /* Scale deviation to rttvar fixed point */
768 var
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
772 var
-= (var
- m
) >> 2;
774 set_dst_metric_rtt(dst
, RTAX_RTTVAR
, var
);
777 if (tp
->snd_ssthresh
>= 0xFFFF) {
778 /* Slow start still did not finish. */
779 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
780 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
781 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
782 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_cwnd
>> 1;
783 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
784 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
785 dst
->metrics
[RTAX_CWND
- 1] = tp
->snd_cwnd
;
786 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
787 icsk
->icsk_ca_state
== TCP_CA_Open
) {
788 /* Cong. avoidance phase, cwnd is reliable. */
789 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
790 dst
->metrics
[RTAX_SSTHRESH
-1] =
791 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
);
792 if (!dst_metric_locked(dst
, RTAX_CWND
))
793 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_cwnd
) >> 1;
795 /* Else slow start did not finish, cwnd is non-sense,
796 ssthresh may be also invalid.
798 if (!dst_metric_locked(dst
, RTAX_CWND
))
799 dst
->metrics
[RTAX_CWND
-1] = (dst_metric(dst
, RTAX_CWND
) + tp
->snd_ssthresh
) >> 1;
800 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
801 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
802 tp
->snd_ssthresh
> dst_metric(dst
, RTAX_SSTHRESH
))
803 dst
->metrics
[RTAX_SSTHRESH
-1] = tp
->snd_ssthresh
;
806 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
807 if (dst_metric(dst
, RTAX_REORDERING
) < tp
->reordering
&&
808 tp
->reordering
!= sysctl_tcp_reordering
)
809 dst
->metrics
[RTAX_REORDERING
-1] = tp
->reordering
;
814 /* Numbers are taken from RFC3390.
816 * John Heffner states:
818 * The RFC specifies a window of no more than 4380 bytes
819 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
820 * is a bit misleading because they use a clamp at 4380 bytes
821 * rather than use a multiplier in the relevant range.
823 __u32
tcp_init_cwnd(struct tcp_sock
*tp
, struct dst_entry
*dst
)
825 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
828 if (tp
->mss_cache
> 1460)
831 cwnd
= (tp
->mss_cache
> 1095) ? 3 : 4;
833 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
836 /* Set slow start threshold and cwnd not falling to slow start */
837 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
839 struct tcp_sock
*tp
= tcp_sk(sk
);
840 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
842 tp
->prior_ssthresh
= 0;
844 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
847 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
848 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
849 tcp_packets_in_flight(tp
) + 1U);
850 tp
->snd_cwnd_cnt
= 0;
851 tp
->high_seq
= tp
->snd_nxt
;
852 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
853 TCP_ECN_queue_cwr(tp
);
855 tcp_set_ca_state(sk
, TCP_CA_CWR
);
860 * Packet counting of FACK is based on in-order assumptions, therefore TCP
861 * disables it when reordering is detected
863 static void tcp_disable_fack(struct tcp_sock
*tp
)
865 /* RFC3517 uses different metric in lost marker => reset on change */
867 tp
->lost_skb_hint
= NULL
;
868 tp
->rx_opt
.sack_ok
&= ~2;
871 /* Take a notice that peer is sending D-SACKs */
872 static void tcp_dsack_seen(struct tcp_sock
*tp
)
874 tp
->rx_opt
.sack_ok
|= 4;
877 /* Initialize metrics on socket. */
879 static void tcp_init_metrics(struct sock
*sk
)
881 struct tcp_sock
*tp
= tcp_sk(sk
);
882 struct dst_entry
*dst
= __sk_dst_get(sk
);
889 if (dst_metric_locked(dst
, RTAX_CWND
))
890 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
891 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
892 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
893 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
894 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
896 if (dst_metric(dst
, RTAX_REORDERING
) &&
897 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
898 tcp_disable_fack(tp
);
899 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
902 if (dst_metric(dst
, RTAX_RTT
) == 0)
905 if (!tp
->srtt
&& dst_metric_rtt(dst
, RTAX_RTT
) < (TCP_TIMEOUT_INIT
<< 3))
908 /* Initial rtt is determined from SYN,SYN-ACK.
909 * The segment is small and rtt may appear much
910 * less than real one. Use per-dst memory
911 * to make it more realistic.
913 * A bit of theory. RTT is time passed after "normal" sized packet
914 * is sent until it is ACKed. In normal circumstances sending small
915 * packets force peer to delay ACKs and calculation is correct too.
916 * The algorithm is adaptive and, provided we follow specs, it
917 * NEVER underestimate RTT. BUT! If peer tries to make some clever
918 * tricks sort of "quick acks" for time long enough to decrease RTT
919 * to low value, and then abruptly stops to do it and starts to delay
920 * ACKs, wait for troubles.
922 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
923 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
924 tp
->rtt_seq
= tp
->snd_nxt
;
926 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
927 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
928 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
932 if (inet_csk(sk
)->icsk_rto
< TCP_TIMEOUT_INIT
&& !tp
->rx_opt
.saw_tstamp
)
934 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
935 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
939 /* Play conservative. If timestamps are not
940 * supported, TCP will fail to recalculate correct
941 * rtt, if initial rto is too small. FORGET ALL AND RESET!
943 if (!tp
->rx_opt
.saw_tstamp
&& tp
->srtt
) {
945 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_INIT
;
946 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
950 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
953 struct tcp_sock
*tp
= tcp_sk(sk
);
954 if (metric
> tp
->reordering
) {
957 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
959 /* This exciting event is worth to be remembered. 8) */
961 mib_idx
= LINUX_MIB_TCPTSREORDER
;
962 else if (tcp_is_reno(tp
))
963 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
964 else if (tcp_is_fack(tp
))
965 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
967 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
969 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
970 #if FASTRETRANS_DEBUG > 1
971 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
972 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
976 tp
->undo_marker
? tp
->undo_retrans
: 0);
978 tcp_disable_fack(tp
);
982 /* This must be called before lost_out is incremented */
983 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
985 if ((tp
->retransmit_skb_hint
== NULL
) ||
986 before(TCP_SKB_CB(skb
)->seq
,
987 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
988 tp
->retransmit_skb_hint
= skb
;
991 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
992 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
995 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
997 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
998 tcp_verify_retransmit_hint(tp
, skb
);
1000 tp
->lost_out
+= tcp_skb_pcount(skb
);
1001 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1005 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
1006 struct sk_buff
*skb
)
1008 tcp_verify_retransmit_hint(tp
, skb
);
1010 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1011 tp
->lost_out
+= tcp_skb_pcount(skb
);
1012 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1016 /* This procedure tags the retransmission queue when SACKs arrive.
1018 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1019 * Packets in queue with these bits set are counted in variables
1020 * sacked_out, retrans_out and lost_out, correspondingly.
1022 * Valid combinations are:
1023 * Tag InFlight Description
1024 * 0 1 - orig segment is in flight.
1025 * S 0 - nothing flies, orig reached receiver.
1026 * L 0 - nothing flies, orig lost by net.
1027 * R 2 - both orig and retransmit are in flight.
1028 * L|R 1 - orig is lost, retransmit is in flight.
1029 * S|R 1 - orig reached receiver, retrans is still in flight.
1030 * (L|S|R is logically valid, it could occur when L|R is sacked,
1031 * but it is equivalent to plain S and code short-curcuits it to S.
1032 * L|S is logically invalid, it would mean -1 packet in flight 8))
1034 * These 6 states form finite state machine, controlled by the following events:
1035 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1036 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1037 * 3. Loss detection event of one of three flavors:
1038 * A. Scoreboard estimator decided the packet is lost.
1039 * A'. Reno "three dupacks" marks head of queue lost.
1040 * A''. Its FACK modfication, head until snd.fack is lost.
1041 * B. SACK arrives sacking data transmitted after never retransmitted
1042 * hole was sent out.
1043 * C. SACK arrives sacking SND.NXT at the moment, when the
1044 * segment was retransmitted.
1045 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1047 * It is pleasant to note, that state diagram turns out to be commutative,
1048 * so that we are allowed not to be bothered by order of our actions,
1049 * when multiple events arrive simultaneously. (see the function below).
1051 * Reordering detection.
1052 * --------------------
1053 * Reordering metric is maximal distance, which a packet can be displaced
1054 * in packet stream. With SACKs we can estimate it:
1056 * 1. SACK fills old hole and the corresponding segment was not
1057 * ever retransmitted -> reordering. Alas, we cannot use it
1058 * when segment was retransmitted.
1059 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1060 * for retransmitted and already SACKed segment -> reordering..
1061 * Both of these heuristics are not used in Loss state, when we cannot
1062 * account for retransmits accurately.
1064 * SACK block validation.
1065 * ----------------------
1067 * SACK block range validation checks that the received SACK block fits to
1068 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1069 * Note that SND.UNA is not included to the range though being valid because
1070 * it means that the receiver is rather inconsistent with itself reporting
1071 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1072 * perfectly valid, however, in light of RFC2018 which explicitly states
1073 * that "SACK block MUST reflect the newest segment. Even if the newest
1074 * segment is going to be discarded ...", not that it looks very clever
1075 * in case of head skb. Due to potentional receiver driven attacks, we
1076 * choose to avoid immediate execution of a walk in write queue due to
1077 * reneging and defer head skb's loss recovery to standard loss recovery
1078 * procedure that will eventually trigger (nothing forbids us doing this).
1080 * Implements also blockage to start_seq wrap-around. Problem lies in the
1081 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1082 * there's no guarantee that it will be before snd_nxt (n). The problem
1083 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1086 * <- outs wnd -> <- wrapzone ->
1087 * u e n u_w e_w s n_w
1089 * |<------------+------+----- TCP seqno space --------------+---------->|
1090 * ...-- <2^31 ->| |<--------...
1091 * ...---- >2^31 ------>| |<--------...
1093 * Current code wouldn't be vulnerable but it's better still to discard such
1094 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1095 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1096 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1097 * equal to the ideal case (infinite seqno space without wrap caused issues).
1099 * With D-SACK the lower bound is extended to cover sequence space below
1100 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1101 * again, D-SACK block must not to go across snd_una (for the same reason as
1102 * for the normal SACK blocks, explained above). But there all simplicity
1103 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1104 * fully below undo_marker they do not affect behavior in anyway and can
1105 * therefore be safely ignored. In rare cases (which are more or less
1106 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1107 * fragmentation and packet reordering past skb's retransmission. To consider
1108 * them correctly, the acceptable range must be extended even more though
1109 * the exact amount is rather hard to quantify. However, tp->max_window can
1110 * be used as an exaggerated estimate.
1112 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1113 u32 start_seq
, u32 end_seq
)
1115 /* Too far in future, or reversed (interpretation is ambiguous) */
1116 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1119 /* Nasty start_seq wrap-around check (see comments above) */
1120 if (!before(start_seq
, tp
->snd_nxt
))
1123 /* In outstanding window? ...This is valid exit for D-SACKs too.
1124 * start_seq == snd_una is non-sensical (see comments above)
1126 if (after(start_seq
, tp
->snd_una
))
1129 if (!is_dsack
|| !tp
->undo_marker
)
1132 /* ...Then it's D-SACK, and must reside below snd_una completely */
1133 if (!after(end_seq
, tp
->snd_una
))
1136 if (!before(start_seq
, tp
->undo_marker
))
1140 if (!after(end_seq
, tp
->undo_marker
))
1143 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1144 * start_seq < undo_marker and end_seq >= undo_marker.
1146 return !before(start_seq
, end_seq
- tp
->max_window
);
1149 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1150 * Event "C". Later note: FACK people cheated me again 8), we have to account
1151 * for reordering! Ugly, but should help.
1153 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1154 * less than what is now known to be received by the other end (derived from
1155 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1156 * retransmitted skbs to avoid some costly processing per ACKs.
1158 static void tcp_mark_lost_retrans(struct sock
*sk
)
1160 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1161 struct tcp_sock
*tp
= tcp_sk(sk
);
1162 struct sk_buff
*skb
;
1164 u32 new_low_seq
= tp
->snd_nxt
;
1165 u32 received_upto
= tcp_highest_sack_seq(tp
);
1167 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1168 !after(received_upto
, tp
->lost_retrans_low
) ||
1169 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1172 tcp_for_write_queue(skb
, sk
) {
1173 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1175 if (skb
== tcp_send_head(sk
))
1177 if (cnt
== tp
->retrans_out
)
1179 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1182 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1185 if (after(received_upto
, ack_seq
) &&
1187 !before(received_upto
,
1188 ack_seq
+ tp
->reordering
* tp
->mss_cache
))) {
1189 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1190 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1192 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1193 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1195 if (before(ack_seq
, new_low_seq
))
1196 new_low_seq
= ack_seq
;
1197 cnt
+= tcp_skb_pcount(skb
);
1201 if (tp
->retrans_out
)
1202 tp
->lost_retrans_low
= new_low_seq
;
1205 static int tcp_check_dsack(struct sock
*sk
, struct sk_buff
*ack_skb
,
1206 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1209 struct tcp_sock
*tp
= tcp_sk(sk
);
1210 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1211 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1214 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1217 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1218 } else if (num_sacks
> 1) {
1219 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1220 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1222 if (!after(end_seq_0
, end_seq_1
) &&
1223 !before(start_seq_0
, start_seq_1
)) {
1226 NET_INC_STATS_BH(sock_net(sk
),
1227 LINUX_MIB_TCPDSACKOFORECV
);
1231 /* D-SACK for already forgotten data... Do dumb counting. */
1233 !after(end_seq_0
, prior_snd_una
) &&
1234 after(end_seq_0
, tp
->undo_marker
))
1240 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1241 * the incoming SACK may not exactly match but we can find smaller MSS
1242 * aligned portion of it that matches. Therefore we might need to fragment
1243 * which may fail and creates some hassle (caller must handle error case
1246 * FIXME: this could be merged to shift decision code
1248 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1249 u32 start_seq
, u32 end_seq
)
1252 unsigned int pkt_len
;
1255 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1256 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1258 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1259 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1260 mss
= tcp_skb_mss(skb
);
1261 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1264 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1268 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1273 /* Round if necessary so that SACKs cover only full MSSes
1274 * and/or the remaining small portion (if present)
1276 if (pkt_len
> mss
) {
1277 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1278 if (!in_sack
&& new_len
< pkt_len
) {
1280 if (new_len
> skb
->len
)
1285 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1293 static int tcp_sacktag_one(struct sk_buff
*skb
, struct sock
*sk
,
1294 int *reord
, int dup_sack
, int fack_count
,
1295 u8
*sackedto
, int pcount
)
1297 struct tcp_sock
*tp
= tcp_sk(sk
);
1298 u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
1301 /* Account D-SACK for retransmitted packet. */
1302 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1303 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->undo_marker
))
1305 if (sacked
& TCPCB_SACKED_ACKED
)
1306 *reord
= min(fack_count
, *reord
);
1309 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1310 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1313 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1314 if (sacked
& TCPCB_SACKED_RETRANS
) {
1315 /* If the segment is not tagged as lost,
1316 * we do not clear RETRANS, believing
1317 * that retransmission is still in flight.
1319 if (sacked
& TCPCB_LOST
) {
1320 *sackedto
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1321 tp
->lost_out
-= pcount
;
1322 tp
->retrans_out
-= pcount
;
1325 if (!(sacked
& TCPCB_RETRANS
)) {
1326 /* New sack for not retransmitted frame,
1327 * which was in hole. It is reordering.
1329 if (before(TCP_SKB_CB(skb
)->seq
,
1330 tcp_highest_sack_seq(tp
)))
1331 *reord
= min(fack_count
, *reord
);
1333 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1334 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->frto_highmark
))
1335 flag
|= FLAG_ONLY_ORIG_SACKED
;
1338 if (sacked
& TCPCB_LOST
) {
1339 *sackedto
&= ~TCPCB_LOST
;
1340 tp
->lost_out
-= pcount
;
1344 *sackedto
|= TCPCB_SACKED_ACKED
;
1345 flag
|= FLAG_DATA_SACKED
;
1346 tp
->sacked_out
+= pcount
;
1348 fack_count
+= pcount
;
1350 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1351 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1352 before(TCP_SKB_CB(skb
)->seq
,
1353 TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1354 tp
->lost_cnt_hint
+= pcount
;
1356 if (fack_count
> tp
->fackets_out
)
1357 tp
->fackets_out
= fack_count
;
1360 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1361 * frames and clear it. undo_retrans is decreased above, L|R frames
1362 * are accounted above as well.
1364 if (dup_sack
&& (*sackedto
& TCPCB_SACKED_RETRANS
)) {
1365 *sackedto
&= ~TCPCB_SACKED_RETRANS
;
1366 tp
->retrans_out
-= pcount
;
1372 static int tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*prev
,
1373 struct sk_buff
*skb
, unsigned int pcount
,
1374 int shifted
, int fack_count
, int *reord
,
1377 struct tcp_sock
*tp
= tcp_sk(sk
);
1378 u8 dummy_sacked
= TCP_SKB_CB(skb
)->sacked
; /* We discard results */
1382 /* Tweak before seqno plays */
1383 if (!tcp_is_fack(tp
) && tcp_is_sack(tp
) && tp
->lost_skb_hint
&&
1384 !before(TCP_SKB_CB(tp
->lost_skb_hint
)->seq
, TCP_SKB_CB(skb
)->seq
))
1385 tp
->lost_cnt_hint
+= pcount
;
1387 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1388 TCP_SKB_CB(skb
)->seq
+= shifted
;
1390 skb_shinfo(prev
)->gso_segs
+= pcount
;
1391 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1392 skb_shinfo(skb
)->gso_segs
-= pcount
;
1394 /* When we're adding to gso_segs == 1, gso_size will be zero,
1395 * in theory this shouldn't be necessary but as long as DSACK
1396 * code can come after this skb later on it's better to keep
1397 * setting gso_size to something.
1399 if (!skb_shinfo(prev
)->gso_size
) {
1400 skb_shinfo(prev
)->gso_size
= mss
;
1401 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1404 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1405 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1406 skb_shinfo(skb
)->gso_size
= 0;
1407 skb_shinfo(skb
)->gso_type
= 0;
1410 *flag
|= tcp_sacktag_one(skb
, sk
, reord
, 0, fack_count
, &dummy_sacked
,
1413 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1414 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1417 BUG_ON(!tcp_skb_pcount(skb
));
1418 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1422 /* Whole SKB was eaten :-) */
1424 if (skb
== tp
->retransmit_skb_hint
)
1425 tp
->retransmit_skb_hint
= prev
;
1426 if (skb
== tp
->scoreboard_skb_hint
)
1427 tp
->scoreboard_skb_hint
= prev
;
1428 if (skb
== tp
->lost_skb_hint
) {
1429 tp
->lost_skb_hint
= prev
;
1430 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1433 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(prev
)->flags
;
1434 if (skb
== tcp_highest_sack(sk
))
1435 tcp_advance_highest_sack(sk
, skb
);
1437 tcp_unlink_write_queue(skb
, sk
);
1438 sk_wmem_free_skb(sk
, skb
);
1440 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1445 /* I wish gso_size would have a bit more sane initialization than
1446 * something-or-zero which complicates things
1448 static int tcp_shift_mss(struct sk_buff
*skb
)
1450 int mss
= tcp_skb_mss(skb
);
1458 /* Shifting pages past head area doesn't work */
1459 static int skb_can_shift(struct sk_buff
*skb
)
1461 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1464 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1467 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1468 u32 start_seq
, u32 end_seq
,
1469 int dup_sack
, int *fack_count
,
1470 int *reord
, int *flag
)
1472 struct tcp_sock
*tp
= tcp_sk(sk
);
1473 struct sk_buff
*prev
;
1479 if (!sk_can_gso(sk
))
1482 /* Normally R but no L won't result in plain S */
1484 (TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) == TCPCB_SACKED_RETRANS
)
1486 if (!skb_can_shift(skb
))
1488 /* This frame is about to be dropped (was ACKed). */
1489 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1492 /* Can only happen with delayed DSACK + discard craziness */
1493 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1495 prev
= tcp_write_queue_prev(sk
, skb
);
1497 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1500 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1501 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1505 pcount
= tcp_skb_pcount(skb
);
1506 mss
= tcp_shift_mss(skb
);
1508 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1509 * drop this restriction as unnecessary
1511 if (mss
!= tcp_shift_mss(prev
))
1514 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1516 /* CHECKME: This is non-MSS split case only?, this will
1517 * cause skipped skbs due to advancing loop btw, original
1518 * has that feature too
1520 if (tcp_skb_pcount(skb
) <= 1)
1523 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1525 /* TODO: head merge to next could be attempted here
1526 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1527 * though it might not be worth of the additional hassle
1529 * ...we can probably just fallback to what was done
1530 * previously. We could try merging non-SACKed ones
1531 * as well but it probably isn't going to buy off
1532 * because later SACKs might again split them, and
1533 * it would make skb timestamp tracking considerably
1539 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1541 BUG_ON(len
> skb
->len
);
1543 /* MSS boundaries should be honoured or else pcount will
1544 * severely break even though it makes things bit trickier.
1545 * Optimize common case to avoid most of the divides
1547 mss
= tcp_skb_mss(skb
);
1549 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1550 * drop this restriction as unnecessary
1552 if (mss
!= tcp_shift_mss(prev
))
1557 } else if (len
< mss
) {
1565 if (!skb_shift(prev
, skb
, len
))
1567 if (!tcp_shifted_skb(sk
, prev
, skb
, pcount
, len
, *fack_count
, reord
,
1571 /* Hole filled allows collapsing with the next as well, this is very
1572 * useful when hole on every nth skb pattern happens
1574 if (prev
== tcp_write_queue_tail(sk
))
1576 skb
= tcp_write_queue_next(sk
, prev
);
1578 if (!skb_can_shift(skb
))
1580 if (skb
== tcp_send_head(sk
))
1582 if ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1586 if (skb_shift(prev
, skb
, len
)) {
1587 pcount
+= tcp_skb_pcount(skb
);
1588 tcp_shifted_skb(sk
, prev
, skb
, tcp_skb_pcount(skb
), len
,
1589 *fack_count
, reord
, flag
, mss
);
1593 *fack_count
+= pcount
;
1600 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1604 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1605 struct tcp_sack_block
*next_dup
,
1606 u32 start_seq
, u32 end_seq
,
1607 int dup_sack_in
, int *fack_count
,
1608 int *reord
, int *flag
)
1610 struct tcp_sock
*tp
= tcp_sk(sk
);
1611 struct sk_buff
*tmp
;
1613 tcp_for_write_queue_from(skb
, sk
) {
1615 int dup_sack
= dup_sack_in
;
1617 if (skb
== tcp_send_head(sk
))
1620 /* queue is in-order => we can short-circuit the walk early */
1621 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1624 if ((next_dup
!= NULL
) &&
1625 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1626 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1627 next_dup
->start_seq
,
1633 /* skb reference here is a bit tricky to get right, since
1634 * shifting can eat and free both this skb and the next,
1635 * so not even _safe variant of the loop is enough.
1638 tmp
= tcp_shift_skb_data(sk
, skb
, start_seq
,
1640 fack_count
, reord
, flag
);
1649 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1655 if (unlikely(in_sack
< 0))
1659 *flag
|= tcp_sacktag_one(skb
, sk
, reord
, dup_sack
,
1661 &(TCP_SKB_CB(skb
)->sacked
),
1662 tcp_skb_pcount(skb
));
1664 if (!before(TCP_SKB_CB(skb
)->seq
,
1665 tcp_highest_sack_seq(tp
)))
1666 tcp_advance_highest_sack(sk
, skb
);
1669 *fack_count
+= tcp_skb_pcount(skb
);
1674 /* Avoid all extra work that is being done by sacktag while walking in
1677 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1678 u32 skip_to_seq
, int *fack_count
)
1680 tcp_for_write_queue_from(skb
, sk
) {
1681 if (skb
== tcp_send_head(sk
))
1684 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1687 *fack_count
+= tcp_skb_pcount(skb
);
1692 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1694 struct tcp_sack_block
*next_dup
,
1696 int *fack_count
, int *reord
,
1699 if (next_dup
== NULL
)
1702 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1703 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
, fack_count
);
1704 skb
= tcp_sacktag_walk(skb
, sk
, NULL
,
1705 next_dup
->start_seq
, next_dup
->end_seq
,
1706 1, fack_count
, reord
, flag
);
1712 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1714 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1718 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1721 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1722 struct tcp_sock
*tp
= tcp_sk(sk
);
1723 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1724 TCP_SKB_CB(ack_skb
)->sacked
);
1725 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1726 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1727 struct tcp_sack_block
*cache
;
1728 struct sk_buff
*skb
;
1729 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1731 int reord
= tp
->packets_out
;
1733 int found_dup_sack
= 0;
1736 int first_sack_index
;
1738 if (!tp
->sacked_out
) {
1739 if (WARN_ON(tp
->fackets_out
))
1740 tp
->fackets_out
= 0;
1741 tcp_highest_sack_reset(sk
);
1744 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1745 num_sacks
, prior_snd_una
);
1747 flag
|= FLAG_DSACKING_ACK
;
1749 /* Eliminate too old ACKs, but take into
1750 * account more or less fresh ones, they can
1751 * contain valid SACK info.
1753 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1756 if (!tp
->packets_out
)
1760 first_sack_index
= 0;
1761 for (i
= 0; i
< num_sacks
; i
++) {
1762 int dup_sack
= !i
&& found_dup_sack
;
1764 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1765 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1767 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1768 sp
[used_sacks
].start_seq
,
1769 sp
[used_sacks
].end_seq
)) {
1773 if (!tp
->undo_marker
)
1774 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1776 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1778 /* Don't count olds caused by ACK reordering */
1779 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1780 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1782 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1785 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1787 first_sack_index
= -1;
1791 /* Ignore very old stuff early */
1792 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1798 /* order SACK blocks to allow in order walk of the retrans queue */
1799 for (i
= used_sacks
- 1; i
> 0; i
--) {
1800 for (j
= 0; j
< i
; j
++) {
1801 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1802 struct tcp_sack_block tmp
;
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
);
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 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
, start_seq
,
1856 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1859 dup_sack
, &fack_count
,
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
,
1869 &fack_count
, &reord
,
1872 /* ...tail remains todo... */
1873 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1874 /* ...but better entrypoint exists! */
1875 skb
= tcp_highest_sack(sk
);
1878 fack_count
= tp
->fackets_out
;
1883 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
,
1885 /* Check overlap against next cached too (past this one already) */
1890 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1891 skb
= tcp_highest_sack(sk
);
1894 fack_count
= tp
->fackets_out
;
1896 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
, &fack_count
);
1899 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1900 dup_sack
, &fack_count
, &reord
, &flag
);
1903 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1904 * due to in-order walk
1906 if (after(end_seq
, tp
->frto_highmark
))
1907 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1912 /* Clear the head of the cache sack blocks so we can skip it next time */
1913 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1914 tp
->recv_sack_cache
[i
].start_seq
= 0;
1915 tp
->recv_sack_cache
[i
].end_seq
= 0;
1917 for (j
= 0; j
< used_sacks
; j
++)
1918 tp
->recv_sack_cache
[i
++] = sp
[j
];
1920 tcp_mark_lost_retrans(sk
);
1922 tcp_verify_left_out(tp
);
1924 if ((reord
< tp
->fackets_out
) &&
1925 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1926 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1927 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1931 #if FASTRETRANS_DEBUG > 0
1932 WARN_ON((int)tp
->sacked_out
< 0);
1933 WARN_ON((int)tp
->lost_out
< 0);
1934 WARN_ON((int)tp
->retrans_out
< 0);
1935 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1940 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1941 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1943 static int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1947 holes
= max(tp
->lost_out
, 1U);
1948 holes
= min(holes
, tp
->packets_out
);
1950 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1951 tp
->sacked_out
= tp
->packets_out
- holes
;
1957 /* If we receive more dupacks than we expected counting segments
1958 * in assumption of absent reordering, interpret this as reordering.
1959 * The only another reason could be bug in receiver TCP.
1961 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1963 struct tcp_sock
*tp
= tcp_sk(sk
);
1964 if (tcp_limit_reno_sacked(tp
))
1965 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1968 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1970 static void tcp_add_reno_sack(struct sock
*sk
)
1972 struct tcp_sock
*tp
= tcp_sk(sk
);
1974 tcp_check_reno_reordering(sk
, 0);
1975 tcp_verify_left_out(tp
);
1978 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1980 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1982 struct tcp_sock
*tp
= tcp_sk(sk
);
1985 /* One ACK acked hole. The rest eat duplicate ACKs. */
1986 if (acked
- 1 >= tp
->sacked_out
)
1989 tp
->sacked_out
-= acked
- 1;
1991 tcp_check_reno_reordering(sk
, acked
);
1992 tcp_verify_left_out(tp
);
1995 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
2000 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
2002 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
2005 /* F-RTO can only be used if TCP has never retransmitted anything other than
2006 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2008 int tcp_use_frto(struct sock
*sk
)
2010 const struct tcp_sock
*tp
= tcp_sk(sk
);
2011 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2012 struct sk_buff
*skb
;
2014 if (!sysctl_tcp_frto
)
2017 /* MTU probe and F-RTO won't really play nicely along currently */
2018 if (icsk
->icsk_mtup
.probe_size
)
2021 if (tcp_is_sackfrto(tp
))
2024 /* Avoid expensive walking of rexmit queue if possible */
2025 if (tp
->retrans_out
> 1)
2028 skb
= tcp_write_queue_head(sk
);
2029 if (tcp_skb_is_last(sk
, skb
))
2031 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
2032 tcp_for_write_queue_from(skb
, sk
) {
2033 if (skb
== tcp_send_head(sk
))
2035 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2037 /* Short-circuit when first non-SACKed skb has been checked */
2038 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2044 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2045 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2046 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2047 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2048 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2049 * bits are handled if the Loss state is really to be entered (in
2050 * tcp_enter_frto_loss).
2052 * Do like tcp_enter_loss() would; when RTO expires the second time it
2054 * "Reduce ssthresh if it has not yet been made inside this window."
2056 void tcp_enter_frto(struct sock
*sk
)
2058 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2059 struct tcp_sock
*tp
= tcp_sk(sk
);
2060 struct sk_buff
*skb
;
2062 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
2063 tp
->snd_una
== tp
->high_seq
||
2064 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
2065 !icsk
->icsk_retransmits
)) {
2066 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2067 /* Our state is too optimistic in ssthresh() call because cwnd
2068 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2069 * recovery has not yet completed. Pattern would be this: RTO,
2070 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2072 * RFC4138 should be more specific on what to do, even though
2073 * RTO is quite unlikely to occur after the first Cumulative ACK
2074 * due to back-off and complexity of triggering events ...
2076 if (tp
->frto_counter
) {
2078 stored_cwnd
= tp
->snd_cwnd
;
2080 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2081 tp
->snd_cwnd
= stored_cwnd
;
2083 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2085 /* ... in theory, cong.control module could do "any tricks" in
2086 * ssthresh(), which means that ca_state, lost bits and lost_out
2087 * counter would have to be faked before the call occurs. We
2088 * consider that too expensive, unlikely and hacky, so modules
2089 * using these in ssthresh() must deal these incompatibility
2090 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2092 tcp_ca_event(sk
, CA_EVENT_FRTO
);
2095 tp
->undo_marker
= tp
->snd_una
;
2096 tp
->undo_retrans
= 0;
2098 skb
= tcp_write_queue_head(sk
);
2099 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2100 tp
->undo_marker
= 0;
2101 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2102 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2103 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2105 tcp_verify_left_out(tp
);
2107 /* Too bad if TCP was application limited */
2108 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2110 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2111 * The last condition is necessary at least in tp->frto_counter case.
2113 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
2114 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
2115 after(tp
->high_seq
, tp
->snd_una
)) {
2116 tp
->frto_highmark
= tp
->high_seq
;
2118 tp
->frto_highmark
= tp
->snd_nxt
;
2120 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
2121 tp
->high_seq
= tp
->snd_nxt
;
2122 tp
->frto_counter
= 1;
2125 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2126 * which indicates that we should follow the traditional RTO recovery,
2127 * i.e. mark everything lost and do go-back-N retransmission.
2129 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
2131 struct tcp_sock
*tp
= tcp_sk(sk
);
2132 struct sk_buff
*skb
;
2135 tp
->retrans_out
= 0;
2136 if (tcp_is_reno(tp
))
2137 tcp_reset_reno_sack(tp
);
2139 tcp_for_write_queue(skb
, sk
) {
2140 if (skb
== tcp_send_head(sk
))
2143 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2145 * Count the retransmission made on RTO correctly (only when
2146 * waiting for the first ACK and did not get it)...
2148 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
2149 /* For some reason this R-bit might get cleared? */
2150 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
2151 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2152 /* ...enter this if branch just for the first segment */
2153 flag
|= FLAG_DATA_ACKED
;
2155 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2156 tp
->undo_marker
= 0;
2157 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2160 /* Marking forward transmissions that were made after RTO lost
2161 * can cause unnecessary retransmissions in some scenarios,
2162 * SACK blocks will mitigate that in some but not in all cases.
2163 * We used to not mark them but it was causing break-ups with
2164 * receivers that do only in-order receival.
2166 * TODO: we could detect presence of such receiver and select
2167 * different behavior per flow.
2169 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2170 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2171 tp
->lost_out
+= tcp_skb_pcount(skb
);
2172 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2175 tcp_verify_left_out(tp
);
2177 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2178 tp
->snd_cwnd_cnt
= 0;
2179 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2180 tp
->frto_counter
= 0;
2181 tp
->bytes_acked
= 0;
2183 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2184 sysctl_tcp_reordering
);
2185 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2186 tp
->high_seq
= tp
->snd_nxt
;
2187 TCP_ECN_queue_cwr(tp
);
2189 tcp_clear_all_retrans_hints(tp
);
2192 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2194 tp
->retrans_out
= 0;
2197 tp
->undo_marker
= 0;
2198 tp
->undo_retrans
= 0;
2201 void tcp_clear_retrans(struct tcp_sock
*tp
)
2203 tcp_clear_retrans_partial(tp
);
2205 tp
->fackets_out
= 0;
2209 /* Enter Loss state. If "how" is not zero, forget all SACK information
2210 * and reset tags completely, otherwise preserve SACKs. If receiver
2211 * dropped its ofo queue, we will know this due to reneging detection.
2213 void tcp_enter_loss(struct sock
*sk
, int how
)
2215 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2216 struct tcp_sock
*tp
= tcp_sk(sk
);
2217 struct sk_buff
*skb
;
2219 /* Reduce ssthresh if it has not yet been made inside this window. */
2220 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2221 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2222 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2223 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2224 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2227 tp
->snd_cwnd_cnt
= 0;
2228 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2230 tp
->bytes_acked
= 0;
2231 tcp_clear_retrans_partial(tp
);
2233 if (tcp_is_reno(tp
))
2234 tcp_reset_reno_sack(tp
);
2237 /* Push undo marker, if it was plain RTO and nothing
2238 * was retransmitted. */
2239 tp
->undo_marker
= tp
->snd_una
;
2242 tp
->fackets_out
= 0;
2244 tcp_clear_all_retrans_hints(tp
);
2246 tcp_for_write_queue(skb
, sk
) {
2247 if (skb
== tcp_send_head(sk
))
2250 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2251 tp
->undo_marker
= 0;
2252 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2253 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2254 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2255 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2256 tp
->lost_out
+= tcp_skb_pcount(skb
);
2257 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2260 tcp_verify_left_out(tp
);
2262 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2263 sysctl_tcp_reordering
);
2264 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2265 tp
->high_seq
= tp
->snd_nxt
;
2266 TCP_ECN_queue_cwr(tp
);
2267 /* Abort F-RTO algorithm if one is in progress */
2268 tp
->frto_counter
= 0;
2271 /* If ACK arrived pointing to a remembered SACK, it means that our
2272 * remembered SACKs do not reflect real state of receiver i.e.
2273 * receiver _host_ is heavily congested (or buggy).
2275 * Do processing similar to RTO timeout.
2277 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2279 if (flag
& FLAG_SACK_RENEGING
) {
2280 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2281 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2283 tcp_enter_loss(sk
, 1);
2284 icsk
->icsk_retransmits
++;
2285 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2286 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2287 icsk
->icsk_rto
, TCP_RTO_MAX
);
2293 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2295 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2298 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2299 * counter when SACK is enabled (without SACK, sacked_out is used for
2302 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2303 * segments up to the highest received SACK block so far and holes in
2306 * With reordering, holes may still be in flight, so RFC3517 recovery
2307 * uses pure sacked_out (total number of SACKed segments) even though
2308 * it violates the RFC that uses duplicate ACKs, often these are equal
2309 * but when e.g. out-of-window ACKs or packet duplication occurs,
2310 * they differ. Since neither occurs due to loss, TCP should really
2313 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
2315 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2318 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2320 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2323 static inline int tcp_head_timedout(struct sock
*sk
)
2325 struct tcp_sock
*tp
= tcp_sk(sk
);
2327 return tp
->packets_out
&&
2328 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2331 /* Linux NewReno/SACK/FACK/ECN state machine.
2332 * --------------------------------------
2334 * "Open" Normal state, no dubious events, fast path.
2335 * "Disorder" In all the respects it is "Open",
2336 * but requires a bit more attention. It is entered when
2337 * we see some SACKs or dupacks. It is split of "Open"
2338 * mainly to move some processing from fast path to slow one.
2339 * "CWR" CWND was reduced due to some Congestion Notification event.
2340 * It can be ECN, ICMP source quench, local device congestion.
2341 * "Recovery" CWND was reduced, we are fast-retransmitting.
2342 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2344 * tcp_fastretrans_alert() is entered:
2345 * - each incoming ACK, if state is not "Open"
2346 * - when arrived ACK is unusual, namely:
2351 * Counting packets in flight is pretty simple.
2353 * in_flight = packets_out - left_out + retrans_out
2355 * packets_out is SND.NXT-SND.UNA counted in packets.
2357 * retrans_out is number of retransmitted segments.
2359 * left_out is number of segments left network, but not ACKed yet.
2361 * left_out = sacked_out + lost_out
2363 * sacked_out: Packets, which arrived to receiver out of order
2364 * and hence not ACKed. With SACKs this number is simply
2365 * amount of SACKed data. Even without SACKs
2366 * it is easy to give pretty reliable estimate of this number,
2367 * counting duplicate ACKs.
2369 * lost_out: Packets lost by network. TCP has no explicit
2370 * "loss notification" feedback from network (for now).
2371 * It means that this number can be only _guessed_.
2372 * Actually, it is the heuristics to predict lossage that
2373 * distinguishes different algorithms.
2375 * F.e. after RTO, when all the queue is considered as lost,
2376 * lost_out = packets_out and in_flight = retrans_out.
2378 * Essentially, we have now two algorithms counting
2381 * FACK: It is the simplest heuristics. As soon as we decided
2382 * that something is lost, we decide that _all_ not SACKed
2383 * packets until the most forward SACK are lost. I.e.
2384 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2385 * It is absolutely correct estimate, if network does not reorder
2386 * packets. And it loses any connection to reality when reordering
2387 * takes place. We use FACK by default until reordering
2388 * is suspected on the path to this destination.
2390 * NewReno: when Recovery is entered, we assume that one segment
2391 * is lost (classic Reno). While we are in Recovery and
2392 * a partial ACK arrives, we assume that one more packet
2393 * is lost (NewReno). This heuristics are the same in NewReno
2396 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2397 * deflation etc. CWND is real congestion window, never inflated, changes
2398 * only according to classic VJ rules.
2400 * Really tricky (and requiring careful tuning) part of algorithm
2401 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2402 * The first determines the moment _when_ we should reduce CWND and,
2403 * hence, slow down forward transmission. In fact, it determines the moment
2404 * when we decide that hole is caused by loss, rather than by a reorder.
2406 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2407 * holes, caused by lost packets.
2409 * And the most logically complicated part of algorithm is undo
2410 * heuristics. We detect false retransmits due to both too early
2411 * fast retransmit (reordering) and underestimated RTO, analyzing
2412 * timestamps and D-SACKs. When we detect that some segments were
2413 * retransmitted by mistake and CWND reduction was wrong, we undo
2414 * window reduction and abort recovery phase. This logic is hidden
2415 * inside several functions named tcp_try_undo_<something>.
2418 /* This function decides, when we should leave Disordered state
2419 * and enter Recovery phase, reducing congestion window.
2421 * Main question: may we further continue forward transmission
2422 * with the same cwnd?
2424 static int tcp_time_to_recover(struct sock
*sk
)
2426 struct tcp_sock
*tp
= tcp_sk(sk
);
2429 /* Do not perform any recovery during F-RTO algorithm */
2430 if (tp
->frto_counter
)
2433 /* Trick#1: The loss is proven. */
2437 /* Not-A-Trick#2 : Classic rule... */
2438 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2441 /* Trick#3 : when we use RFC2988 timer restart, fast
2442 * retransmit can be triggered by timeout of queue head.
2444 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2447 /* Trick#4: It is still not OK... But will it be useful to delay
2450 packets_out
= tp
->packets_out
;
2451 if (packets_out
<= tp
->reordering
&&
2452 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2453 !tcp_may_send_now(sk
)) {
2454 /* We have nothing to send. This connection is limited
2455 * either by receiver window or by application.
2463 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2464 * is against sacked "cnt", otherwise it's against facked "cnt"
2466 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2468 struct tcp_sock
*tp
= tcp_sk(sk
);
2469 struct sk_buff
*skb
;
2474 WARN_ON(packets
> tp
->packets_out
);
2475 if (tp
->lost_skb_hint
) {
2476 skb
= tp
->lost_skb_hint
;
2477 cnt
= tp
->lost_cnt_hint
;
2479 skb
= tcp_write_queue_head(sk
);
2483 tcp_for_write_queue_from(skb
, sk
) {
2484 if (skb
== tcp_send_head(sk
))
2486 /* TODO: do this better */
2487 /* this is not the most efficient way to do this... */
2488 tp
->lost_skb_hint
= skb
;
2489 tp
->lost_cnt_hint
= cnt
;
2491 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2495 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2496 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2497 cnt
+= tcp_skb_pcount(skb
);
2499 if (cnt
> packets
) {
2500 if (tcp_is_sack(tp
) || (oldcnt
>= packets
))
2503 mss
= skb_shinfo(skb
)->gso_size
;
2504 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2510 tcp_skb_mark_lost(tp
, skb
);
2512 tcp_verify_left_out(tp
);
2515 /* Account newly detected lost packet(s) */
2517 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2519 struct tcp_sock
*tp
= tcp_sk(sk
);
2521 if (tcp_is_reno(tp
)) {
2522 tcp_mark_head_lost(sk
, 1);
2523 } else if (tcp_is_fack(tp
)) {
2524 int lost
= tp
->fackets_out
- tp
->reordering
;
2527 tcp_mark_head_lost(sk
, lost
);
2529 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2530 if (sacked_upto
< fast_rexmit
)
2531 sacked_upto
= fast_rexmit
;
2532 tcp_mark_head_lost(sk
, sacked_upto
);
2535 /* New heuristics: it is possible only after we switched
2536 * to restart timer each time when something is ACKed.
2537 * Hence, we can detect timed out packets during fast
2538 * retransmit without falling to slow start.
2540 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2541 struct sk_buff
*skb
;
2543 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2544 : tcp_write_queue_head(sk
);
2546 tcp_for_write_queue_from(skb
, sk
) {
2547 if (skb
== tcp_send_head(sk
))
2549 if (!tcp_skb_timedout(sk
, skb
))
2552 tcp_skb_mark_lost(tp
, skb
);
2555 tp
->scoreboard_skb_hint
= skb
;
2557 tcp_verify_left_out(tp
);
2561 /* CWND moderation, preventing bursts due to too big ACKs
2562 * in dubious situations.
2564 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2566 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2567 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2568 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2571 /* Lower bound on congestion window is slow start threshold
2572 * unless congestion avoidance choice decides to overide it.
2574 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2576 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2578 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2581 /* Decrease cwnd each second ack. */
2582 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2584 struct tcp_sock
*tp
= tcp_sk(sk
);
2585 int decr
= tp
->snd_cwnd_cnt
+ 1;
2587 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2588 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2589 tp
->snd_cwnd_cnt
= decr
& 1;
2592 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2593 tp
->snd_cwnd
-= decr
;
2595 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2596 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2600 /* Nothing was retransmitted or returned timestamp is less
2601 * than timestamp of the first retransmission.
2603 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2605 return !tp
->retrans_stamp
||
2606 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2607 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2610 /* Undo procedures. */
2612 #if FASTRETRANS_DEBUG > 1
2613 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2615 struct tcp_sock
*tp
= tcp_sk(sk
);
2616 struct inet_sock
*inet
= inet_sk(sk
);
2618 if (sk
->sk_family
== AF_INET
) {
2619 printk(KERN_DEBUG
"Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2621 &inet
->daddr
, ntohs(inet
->dport
),
2622 tp
->snd_cwnd
, tcp_left_out(tp
),
2623 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2626 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2627 else if (sk
->sk_family
== AF_INET6
) {
2628 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2629 printk(KERN_DEBUG
"Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2631 &np
->daddr
, ntohs(inet
->dport
),
2632 tp
->snd_cwnd
, tcp_left_out(tp
),
2633 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2639 #define DBGUNDO(x...) do { } while (0)
2642 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2644 struct tcp_sock
*tp
= tcp_sk(sk
);
2646 if (tp
->prior_ssthresh
) {
2647 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2649 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2650 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2652 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2654 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2655 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2656 TCP_ECN_withdraw_cwr(tp
);
2659 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2661 tcp_moderate_cwnd(tp
);
2662 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2665 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2667 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2670 /* People celebrate: "We love our President!" */
2671 static int tcp_try_undo_recovery(struct sock
*sk
)
2673 struct tcp_sock
*tp
= tcp_sk(sk
);
2675 if (tcp_may_undo(tp
)) {
2678 /* Happy end! We did not retransmit anything
2679 * or our original transmission succeeded.
2681 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2682 tcp_undo_cwr(sk
, 1);
2683 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2684 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2686 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2688 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2689 tp
->undo_marker
= 0;
2691 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2692 /* Hold old state until something *above* high_seq
2693 * is ACKed. For Reno it is MUST to prevent false
2694 * fast retransmits (RFC2582). SACK TCP is safe. */
2695 tcp_moderate_cwnd(tp
);
2698 tcp_set_ca_state(sk
, TCP_CA_Open
);
2702 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2703 static void tcp_try_undo_dsack(struct sock
*sk
)
2705 struct tcp_sock
*tp
= tcp_sk(sk
);
2707 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2708 DBGUNDO(sk
, "D-SACK");
2709 tcp_undo_cwr(sk
, 1);
2710 tp
->undo_marker
= 0;
2711 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2715 /* Undo during fast recovery after partial ACK. */
2717 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2719 struct tcp_sock
*tp
= tcp_sk(sk
);
2720 /* Partial ACK arrived. Force Hoe's retransmit. */
2721 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2723 if (tcp_may_undo(tp
)) {
2724 /* Plain luck! Hole if filled with delayed
2725 * packet, rather than with a retransmit.
2727 if (tp
->retrans_out
== 0)
2728 tp
->retrans_stamp
= 0;
2730 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2733 tcp_undo_cwr(sk
, 0);
2734 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2736 /* So... Do not make Hoe's retransmit yet.
2737 * If the first packet was delayed, the rest
2738 * ones are most probably delayed as well.
2745 /* Undo during loss recovery after partial ACK. */
2746 static int tcp_try_undo_loss(struct sock
*sk
)
2748 struct tcp_sock
*tp
= tcp_sk(sk
);
2750 if (tcp_may_undo(tp
)) {
2751 struct sk_buff
*skb
;
2752 tcp_for_write_queue(skb
, sk
) {
2753 if (skb
== tcp_send_head(sk
))
2755 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2758 tcp_clear_all_retrans_hints(tp
);
2760 DBGUNDO(sk
, "partial loss");
2762 tcp_undo_cwr(sk
, 1);
2763 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2764 inet_csk(sk
)->icsk_retransmits
= 0;
2765 tp
->undo_marker
= 0;
2766 if (tcp_is_sack(tp
))
2767 tcp_set_ca_state(sk
, TCP_CA_Open
);
2773 static inline void tcp_complete_cwr(struct sock
*sk
)
2775 struct tcp_sock
*tp
= tcp_sk(sk
);
2776 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2777 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2778 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2781 static void tcp_try_keep_open(struct sock
*sk
)
2783 struct tcp_sock
*tp
= tcp_sk(sk
);
2784 int state
= TCP_CA_Open
;
2786 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2787 state
= TCP_CA_Disorder
;
2789 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2790 tcp_set_ca_state(sk
, state
);
2791 tp
->high_seq
= tp
->snd_nxt
;
2795 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2797 struct tcp_sock
*tp
= tcp_sk(sk
);
2799 tcp_verify_left_out(tp
);
2801 if (!tp
->frto_counter
&& tp
->retrans_out
== 0)
2802 tp
->retrans_stamp
= 0;
2804 if (flag
& FLAG_ECE
)
2805 tcp_enter_cwr(sk
, 1);
2807 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2808 tcp_try_keep_open(sk
);
2809 tcp_moderate_cwnd(tp
);
2811 tcp_cwnd_down(sk
, flag
);
2815 static void tcp_mtup_probe_failed(struct sock
*sk
)
2817 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2819 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2820 icsk
->icsk_mtup
.probe_size
= 0;
2823 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2825 struct tcp_sock
*tp
= tcp_sk(sk
);
2826 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2828 /* FIXME: breaks with very large cwnd */
2829 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2830 tp
->snd_cwnd
= tp
->snd_cwnd
*
2831 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2832 icsk
->icsk_mtup
.probe_size
;
2833 tp
->snd_cwnd_cnt
= 0;
2834 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2835 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2837 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2838 icsk
->icsk_mtup
.probe_size
= 0;
2839 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2842 /* Do a simple retransmit without using the backoff mechanisms in
2843 * tcp_timer. This is used for path mtu discovery.
2844 * The socket is already locked here.
2846 void tcp_simple_retransmit(struct sock
*sk
)
2848 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2849 struct tcp_sock
*tp
= tcp_sk(sk
);
2850 struct sk_buff
*skb
;
2851 unsigned int mss
= tcp_current_mss(sk
, 0);
2852 u32 prior_lost
= tp
->lost_out
;
2854 tcp_for_write_queue(skb
, sk
) {
2855 if (skb
== tcp_send_head(sk
))
2857 if (skb
->len
> mss
&&
2858 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2859 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2860 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2861 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2863 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2867 tcp_clear_retrans_hints_partial(tp
);
2869 if (prior_lost
== tp
->lost_out
)
2872 if (tcp_is_reno(tp
))
2873 tcp_limit_reno_sacked(tp
);
2875 tcp_verify_left_out(tp
);
2877 /* Don't muck with the congestion window here.
2878 * Reason is that we do not increase amount of _data_
2879 * in network, but units changed and effective
2880 * cwnd/ssthresh really reduced now.
2882 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2883 tp
->high_seq
= tp
->snd_nxt
;
2884 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2885 tp
->prior_ssthresh
= 0;
2886 tp
->undo_marker
= 0;
2887 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2889 tcp_xmit_retransmit_queue(sk
);
2892 /* Process an event, which can update packets-in-flight not trivially.
2893 * Main goal of this function is to calculate new estimate for left_out,
2894 * taking into account both packets sitting in receiver's buffer and
2895 * packets lost by network.
2897 * Besides that it does CWND reduction, when packet loss is detected
2898 * and changes state of machine.
2900 * It does _not_ decide what to send, it is made in function
2901 * tcp_xmit_retransmit_queue().
2903 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2905 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2906 struct tcp_sock
*tp
= tcp_sk(sk
);
2907 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2908 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2909 (tcp_fackets_out(tp
) > tp
->reordering
));
2910 int fast_rexmit
= 0, mib_idx
;
2912 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2914 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2915 tp
->fackets_out
= 0;
2917 /* Now state machine starts.
2918 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2919 if (flag
& FLAG_ECE
)
2920 tp
->prior_ssthresh
= 0;
2922 /* B. In all the states check for reneging SACKs. */
2923 if (tcp_check_sack_reneging(sk
, flag
))
2926 /* C. Process data loss notification, provided it is valid. */
2927 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2928 before(tp
->snd_una
, tp
->high_seq
) &&
2929 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2930 tp
->fackets_out
> tp
->reordering
) {
2931 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2932 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2935 /* D. Check consistency of the current state. */
2936 tcp_verify_left_out(tp
);
2938 /* E. Check state exit conditions. State can be terminated
2939 * when high_seq is ACKed. */
2940 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2941 WARN_ON(tp
->retrans_out
!= 0);
2942 tp
->retrans_stamp
= 0;
2943 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2944 switch (icsk
->icsk_ca_state
) {
2946 icsk
->icsk_retransmits
= 0;
2947 if (tcp_try_undo_recovery(sk
))
2952 /* CWR is to be held something *above* high_seq
2953 * is ACKed for CWR bit to reach receiver. */
2954 if (tp
->snd_una
!= tp
->high_seq
) {
2955 tcp_complete_cwr(sk
);
2956 tcp_set_ca_state(sk
, TCP_CA_Open
);
2960 case TCP_CA_Disorder
:
2961 tcp_try_undo_dsack(sk
);
2962 if (!tp
->undo_marker
||
2963 /* For SACK case do not Open to allow to undo
2964 * catching for all duplicate ACKs. */
2965 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2966 tp
->undo_marker
= 0;
2967 tcp_set_ca_state(sk
, TCP_CA_Open
);
2971 case TCP_CA_Recovery
:
2972 if (tcp_is_reno(tp
))
2973 tcp_reset_reno_sack(tp
);
2974 if (tcp_try_undo_recovery(sk
))
2976 tcp_complete_cwr(sk
);
2981 /* F. Process state. */
2982 switch (icsk
->icsk_ca_state
) {
2983 case TCP_CA_Recovery
:
2984 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2985 if (tcp_is_reno(tp
) && is_dupack
)
2986 tcp_add_reno_sack(sk
);
2988 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2991 if (flag
& FLAG_DATA_ACKED
)
2992 icsk
->icsk_retransmits
= 0;
2993 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
2994 tcp_reset_reno_sack(tp
);
2995 if (!tcp_try_undo_loss(sk
)) {
2996 tcp_moderate_cwnd(tp
);
2997 tcp_xmit_retransmit_queue(sk
);
3000 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
3002 /* Loss is undone; fall through to processing in Open state. */
3004 if (tcp_is_reno(tp
)) {
3005 if (flag
& FLAG_SND_UNA_ADVANCED
)
3006 tcp_reset_reno_sack(tp
);
3008 tcp_add_reno_sack(sk
);
3011 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
3012 tcp_try_undo_dsack(sk
);
3014 if (!tcp_time_to_recover(sk
)) {
3015 tcp_try_to_open(sk
, flag
);
3019 /* MTU probe failure: don't reduce cwnd */
3020 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3021 icsk
->icsk_mtup
.probe_size
&&
3022 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3023 tcp_mtup_probe_failed(sk
);
3024 /* Restores the reduction we did in tcp_mtup_probe() */
3026 tcp_simple_retransmit(sk
);
3030 /* Otherwise enter Recovery state */
3032 if (tcp_is_reno(tp
))
3033 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
3035 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
3037 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3039 tp
->high_seq
= tp
->snd_nxt
;
3040 tp
->prior_ssthresh
= 0;
3041 tp
->undo_marker
= tp
->snd_una
;
3042 tp
->undo_retrans
= tp
->retrans_out
;
3044 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
3045 if (!(flag
& FLAG_ECE
))
3046 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
3047 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
3048 TCP_ECN_queue_cwr(tp
);
3051 tp
->bytes_acked
= 0;
3052 tp
->snd_cwnd_cnt
= 0;
3053 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
3057 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3058 tcp_update_scoreboard(sk
, fast_rexmit
);
3059 tcp_cwnd_down(sk
, flag
);
3060 tcp_xmit_retransmit_queue(sk
);
3063 /* Read draft-ietf-tcplw-high-performance before mucking
3064 * with this code. (Supersedes RFC1323)
3066 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3068 /* RTTM Rule: A TSecr value received in a segment is used to
3069 * update the averaged RTT measurement only if the segment
3070 * acknowledges some new data, i.e., only if it advances the
3071 * left edge of the send window.
3073 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3074 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3076 * Changed: reset backoff as soon as we see the first valid sample.
3077 * If we do not, we get strongly overestimated rto. With timestamps
3078 * samples are accepted even from very old segments: f.e., when rtt=1
3079 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3080 * answer arrives rto becomes 120 seconds! If at least one of segments
3081 * in window is lost... Voila. --ANK (010210)
3083 struct tcp_sock
*tp
= tcp_sk(sk
);
3084 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
3085 tcp_rtt_estimator(sk
, seq_rtt
);
3087 inet_csk(sk
)->icsk_backoff
= 0;
3091 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3093 /* We don't have a timestamp. Can only use
3094 * packets that are not retransmitted to determine
3095 * rtt estimates. Also, we must not reset the
3096 * backoff for rto until we get a non-retransmitted
3097 * packet. This allows us to deal with a situation
3098 * where the network delay has increased suddenly.
3099 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3102 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3105 tcp_rtt_estimator(sk
, seq_rtt
);
3107 inet_csk(sk
)->icsk_backoff
= 0;
3111 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3114 const struct tcp_sock
*tp
= tcp_sk(sk
);
3115 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3116 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3117 tcp_ack_saw_tstamp(sk
, flag
);
3118 else if (seq_rtt
>= 0)
3119 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3122 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3124 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3125 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3126 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3129 /* Restart timer after forward progress on connection.
3130 * RFC2988 recommends to restart timer to now+rto.
3132 static void tcp_rearm_rto(struct sock
*sk
)
3134 struct tcp_sock
*tp
= tcp_sk(sk
);
3136 if (!tp
->packets_out
) {
3137 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3139 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3140 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3144 /* If we get here, the whole TSO packet has not been acked. */
3145 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3147 struct tcp_sock
*tp
= tcp_sk(sk
);
3150 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3152 packets_acked
= tcp_skb_pcount(skb
);
3153 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3155 packets_acked
-= tcp_skb_pcount(skb
);
3157 if (packets_acked
) {
3158 BUG_ON(tcp_skb_pcount(skb
) == 0);
3159 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3162 return packets_acked
;
3165 /* Remove acknowledged frames from the retransmission queue. If our packet
3166 * is before the ack sequence we can discard it as it's confirmed to have
3167 * arrived at the other end.
3169 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3172 struct tcp_sock
*tp
= tcp_sk(sk
);
3173 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3174 struct sk_buff
*skb
;
3175 u32 now
= tcp_time_stamp
;
3176 int fully_acked
= 1;
3179 u32 reord
= tp
->packets_out
;
3180 u32 prior_sacked
= tp
->sacked_out
;
3182 s32 ca_seq_rtt
= -1;
3183 ktime_t last_ackt
= net_invalid_timestamp();
3185 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3186 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3189 u8 sacked
= scb
->sacked
;
3191 /* Determine how many packets and what bytes were acked, tso and else */
3192 if (after(scb
->end_seq
, tp
->snd_una
)) {
3193 if (tcp_skb_pcount(skb
) == 1 ||
3194 !after(tp
->snd_una
, scb
->seq
))
3197 acked_pcount
= tcp_tso_acked(sk
, skb
);
3202 end_seq
= tp
->snd_una
;
3204 acked_pcount
= tcp_skb_pcount(skb
);
3205 end_seq
= scb
->end_seq
;
3208 /* MTU probing checks */
3209 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
3210 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
3211 tcp_mtup_probe_success(sk
, skb
);
3214 if (sacked
& TCPCB_RETRANS
) {
3215 if (sacked
& TCPCB_SACKED_RETRANS
)
3216 tp
->retrans_out
-= acked_pcount
;
3217 flag
|= FLAG_RETRANS_DATA_ACKED
;
3220 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3221 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3223 ca_seq_rtt
= now
- scb
->when
;
3224 last_ackt
= skb
->tstamp
;
3226 seq_rtt
= ca_seq_rtt
;
3228 if (!(sacked
& TCPCB_SACKED_ACKED
))
3229 reord
= min(pkts_acked
, reord
);
3232 if (sacked
& TCPCB_SACKED_ACKED
)
3233 tp
->sacked_out
-= acked_pcount
;
3234 if (sacked
& TCPCB_LOST
)
3235 tp
->lost_out
-= acked_pcount
;
3237 tp
->packets_out
-= acked_pcount
;
3238 pkts_acked
+= acked_pcount
;
3240 /* Initial outgoing SYN's get put onto the write_queue
3241 * just like anything else we transmit. It is not
3242 * true data, and if we misinform our callers that
3243 * this ACK acks real data, we will erroneously exit
3244 * connection startup slow start one packet too
3245 * quickly. This is severely frowned upon behavior.
3247 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
3248 flag
|= FLAG_DATA_ACKED
;
3250 flag
|= FLAG_SYN_ACKED
;
3251 tp
->retrans_stamp
= 0;
3257 tcp_unlink_write_queue(skb
, sk
);
3258 sk_wmem_free_skb(sk
, skb
);
3259 tp
->scoreboard_skb_hint
= NULL
;
3260 if (skb
== tp
->retransmit_skb_hint
)
3261 tp
->retransmit_skb_hint
= NULL
;
3262 if (skb
== tp
->lost_skb_hint
)
3263 tp
->lost_skb_hint
= NULL
;
3266 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3267 tp
->snd_up
= tp
->snd_una
;
3269 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3270 flag
|= FLAG_SACK_RENEGING
;
3272 if (flag
& FLAG_ACKED
) {
3273 const struct tcp_congestion_ops
*ca_ops
3274 = inet_csk(sk
)->icsk_ca_ops
;
3276 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3279 if (tcp_is_reno(tp
)) {
3280 tcp_remove_reno_sacks(sk
, pkts_acked
);
3282 /* Non-retransmitted hole got filled? That's reordering */
3283 if (reord
< prior_fackets
)
3284 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3286 /* No need to care for underflows here because
3287 * the lost_skb_hint gets NULLed if we're past it
3288 * (or something non-trivial happened)
3290 if (tcp_is_fack(tp
))
3291 tp
->lost_cnt_hint
-= pkts_acked
;
3293 tp
->lost_cnt_hint
-= prior_sacked
- tp
->sacked_out
;
3296 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3298 if (ca_ops
->pkts_acked
) {
3301 /* Is the ACK triggering packet unambiguous? */
3302 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3303 /* High resolution needed and available? */
3304 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3305 !ktime_equal(last_ackt
,
3306 net_invalid_timestamp()))
3307 rtt_us
= ktime_us_delta(ktime_get_real(),
3309 else if (ca_seq_rtt
> 0)
3310 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3313 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3317 #if FASTRETRANS_DEBUG > 0
3318 WARN_ON((int)tp
->sacked_out
< 0);
3319 WARN_ON((int)tp
->lost_out
< 0);
3320 WARN_ON((int)tp
->retrans_out
< 0);
3321 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3322 icsk
= inet_csk(sk
);
3324 printk(KERN_DEBUG
"Leak l=%u %d\n",
3325 tp
->lost_out
, icsk
->icsk_ca_state
);
3328 if (tp
->sacked_out
) {
3329 printk(KERN_DEBUG
"Leak s=%u %d\n",
3330 tp
->sacked_out
, icsk
->icsk_ca_state
);
3333 if (tp
->retrans_out
) {
3334 printk(KERN_DEBUG
"Leak r=%u %d\n",
3335 tp
->retrans_out
, icsk
->icsk_ca_state
);
3336 tp
->retrans_out
= 0;
3343 static void tcp_ack_probe(struct sock
*sk
)
3345 const struct tcp_sock
*tp
= tcp_sk(sk
);
3346 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3348 /* Was it a usable window open? */
3350 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3351 icsk
->icsk_backoff
= 0;
3352 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3353 /* Socket must be waked up by subsequent tcp_data_snd_check().
3354 * This function is not for random using!
3357 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3358 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3363 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3365 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3366 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3369 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3371 const struct tcp_sock
*tp
= tcp_sk(sk
);
3372 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3373 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3376 /* Check that window update is acceptable.
3377 * The function assumes that snd_una<=ack<=snd_next.
3379 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3380 const u32 ack
, const u32 ack_seq
,
3383 return (after(ack
, tp
->snd_una
) ||
3384 after(ack_seq
, tp
->snd_wl1
) ||
3385 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3388 /* Update our send window.
3390 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3391 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3393 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3396 struct tcp_sock
*tp
= tcp_sk(sk
);
3398 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3400 if (likely(!tcp_hdr(skb
)->syn
))
3401 nwin
<<= tp
->rx_opt
.snd_wscale
;
3403 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3404 flag
|= FLAG_WIN_UPDATE
;
3405 tcp_update_wl(tp
, ack
, ack_seq
);
3407 if (tp
->snd_wnd
!= nwin
) {
3410 /* Note, it is the only place, where
3411 * fast path is recovered for sending TCP.
3414 tcp_fast_path_check(sk
);
3416 if (nwin
> tp
->max_window
) {
3417 tp
->max_window
= nwin
;
3418 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3428 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3429 * continue in congestion avoidance.
3431 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3433 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3434 tp
->snd_cwnd_cnt
= 0;
3435 tp
->bytes_acked
= 0;
3436 TCP_ECN_queue_cwr(tp
);
3437 tcp_moderate_cwnd(tp
);
3440 /* A conservative spurious RTO response algorithm: reduce cwnd using
3441 * rate halving and continue in congestion avoidance.
3443 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3445 tcp_enter_cwr(sk
, 0);
3448 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3450 if (flag
& FLAG_ECE
)
3451 tcp_ratehalving_spur_to_response(sk
);
3453 tcp_undo_cwr(sk
, 1);
3456 /* F-RTO spurious RTO detection algorithm (RFC4138)
3458 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3459 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3460 * window (but not to or beyond highest sequence sent before RTO):
3461 * On First ACK, send two new segments out.
3462 * On Second ACK, RTO was likely spurious. Do spurious response (response
3463 * algorithm is not part of the F-RTO detection algorithm
3464 * given in RFC4138 but can be selected separately).
3465 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3466 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3467 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3468 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3470 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3471 * original window even after we transmit two new data segments.
3474 * on first step, wait until first cumulative ACK arrives, then move to
3475 * the second step. In second step, the next ACK decides.
3477 * F-RTO is implemented (mainly) in four functions:
3478 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3479 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3480 * called when tcp_use_frto() showed green light
3481 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3482 * - tcp_enter_frto_loss() is called if there is not enough evidence
3483 * to prove that the RTO is indeed spurious. It transfers the control
3484 * from F-RTO to the conventional RTO recovery
3486 static int tcp_process_frto(struct sock
*sk
, int flag
)
3488 struct tcp_sock
*tp
= tcp_sk(sk
);
3490 tcp_verify_left_out(tp
);
3492 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3493 if (flag
& FLAG_DATA_ACKED
)
3494 inet_csk(sk
)->icsk_retransmits
= 0;
3496 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3497 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3498 tp
->undo_marker
= 0;
3500 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3501 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3505 if (!tcp_is_sackfrto(tp
)) {
3506 /* RFC4138 shortcoming in step 2; should also have case c):
3507 * ACK isn't duplicate nor advances window, e.g., opposite dir
3510 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3513 if (!(flag
& FLAG_DATA_ACKED
)) {
3514 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3519 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3520 /* Prevent sending of new data. */
3521 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3522 tcp_packets_in_flight(tp
));
3526 if ((tp
->frto_counter
>= 2) &&
3527 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3528 ((flag
& FLAG_DATA_SACKED
) &&
3529 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3530 /* RFC4138 shortcoming (see comment above) */
3531 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3532 (flag
& FLAG_NOT_DUP
))
3535 tcp_enter_frto_loss(sk
, 3, flag
);
3540 if (tp
->frto_counter
== 1) {
3541 /* tcp_may_send_now needs to see updated state */
3542 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3543 tp
->frto_counter
= 2;
3545 if (!tcp_may_send_now(sk
))
3546 tcp_enter_frto_loss(sk
, 2, flag
);
3550 switch (sysctl_tcp_frto_response
) {
3552 tcp_undo_spur_to_response(sk
, flag
);
3555 tcp_conservative_spur_to_response(tp
);
3558 tcp_ratehalving_spur_to_response(sk
);
3561 tp
->frto_counter
= 0;
3562 tp
->undo_marker
= 0;
3563 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3568 /* This routine deals with incoming acks, but not outgoing ones. */
3569 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3571 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3572 struct tcp_sock
*tp
= tcp_sk(sk
);
3573 u32 prior_snd_una
= tp
->snd_una
;
3574 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3575 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3576 u32 prior_in_flight
;
3581 /* If the ack is newer than sent or older than previous acks
3582 * then we can probably ignore it.
3584 if (after(ack
, tp
->snd_nxt
))
3585 goto uninteresting_ack
;
3587 if (before(ack
, prior_snd_una
))
3590 if (after(ack
, prior_snd_una
))
3591 flag
|= FLAG_SND_UNA_ADVANCED
;
3593 if (sysctl_tcp_abc
) {
3594 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3595 tp
->bytes_acked
+= ack
- prior_snd_una
;
3596 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3597 /* we assume just one segment left network */
3598 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3602 prior_fackets
= tp
->fackets_out
;
3603 prior_in_flight
= tcp_packets_in_flight(tp
);
3605 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3606 /* Window is constant, pure forward advance.
3607 * No more checks are required.
3608 * Note, we use the fact that SND.UNA>=SND.WL2.
3610 tcp_update_wl(tp
, ack
, ack_seq
);
3612 flag
|= FLAG_WIN_UPDATE
;
3614 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3616 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3618 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3621 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3623 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3625 if (TCP_SKB_CB(skb
)->sacked
)
3626 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3628 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3631 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3634 /* We passed data and got it acked, remove any soft error
3635 * log. Something worked...
3637 sk
->sk_err_soft
= 0;
3638 icsk
->icsk_probes_out
= 0;
3639 tp
->rcv_tstamp
= tcp_time_stamp
;
3640 prior_packets
= tp
->packets_out
;
3644 /* See if we can take anything off of the retransmit queue. */
3645 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3647 if (tp
->frto_counter
)
3648 frto_cwnd
= tcp_process_frto(sk
, flag
);
3649 /* Guarantee sacktag reordering detection against wrap-arounds */
3650 if (before(tp
->frto_highmark
, tp
->snd_una
))
3651 tp
->frto_highmark
= 0;
3653 if (tcp_ack_is_dubious(sk
, flag
)) {
3654 /* Advance CWND, if state allows this. */
3655 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3656 tcp_may_raise_cwnd(sk
, flag
))
3657 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3658 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3661 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3662 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3665 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3666 dst_confirm(sk
->sk_dst_cache
);
3671 /* If this ack opens up a zero window, clear backoff. It was
3672 * being used to time the probes, and is probably far higher than
3673 * it needs to be for normal retransmission.
3675 if (tcp_send_head(sk
))
3680 if (TCP_SKB_CB(skb
)->sacked
) {
3681 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3682 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3683 tcp_try_keep_open(sk
);
3687 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3691 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3692 * But, this can also be called on packets in the established flow when
3693 * the fast version below fails.
3695 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3699 struct tcphdr
*th
= tcp_hdr(skb
);
3700 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3702 ptr
= (unsigned char *)(th
+ 1);
3703 opt_rx
->saw_tstamp
= 0;
3705 while (length
> 0) {
3706 int opcode
= *ptr
++;
3712 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3717 if (opsize
< 2) /* "silly options" */
3719 if (opsize
> length
)
3720 return; /* don't parse partial options */
3723 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3724 u16 in_mss
= get_unaligned_be16(ptr
);
3726 if (opt_rx
->user_mss
&&
3727 opt_rx
->user_mss
< in_mss
)
3728 in_mss
= opt_rx
->user_mss
;
3729 opt_rx
->mss_clamp
= in_mss
;
3734 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3735 !estab
&& sysctl_tcp_window_scaling
) {
3736 __u8 snd_wscale
= *(__u8
*)ptr
;
3737 opt_rx
->wscale_ok
= 1;
3738 if (snd_wscale
> 14) {
3739 if (net_ratelimit())
3740 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3741 "scaling value %d >14 received.\n",
3745 opt_rx
->snd_wscale
= snd_wscale
;
3748 case TCPOPT_TIMESTAMP
:
3749 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3750 ((estab
&& opt_rx
->tstamp_ok
) ||
3751 (!estab
&& sysctl_tcp_timestamps
))) {
3752 opt_rx
->saw_tstamp
= 1;
3753 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3754 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3757 case TCPOPT_SACK_PERM
:
3758 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3759 !estab
&& sysctl_tcp_sack
) {
3760 opt_rx
->sack_ok
= 1;
3761 tcp_sack_reset(opt_rx
);
3766 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3767 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3769 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3772 #ifdef CONFIG_TCP_MD5SIG
3775 * The MD5 Hash has already been
3776 * checked (see tcp_v{4,6}_do_rcv()).
3788 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3790 __be32
*ptr
= (__be32
*)(th
+ 1);
3792 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3793 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3794 tp
->rx_opt
.saw_tstamp
= 1;
3796 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3798 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3804 /* Fast parse options. This hopes to only see timestamps.
3805 * If it is wrong it falls back on tcp_parse_options().
3807 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3808 struct tcp_sock
*tp
)
3810 if (th
->doff
== sizeof(struct tcphdr
) >> 2) {
3811 tp
->rx_opt
.saw_tstamp
= 0;
3813 } else if (tp
->rx_opt
.tstamp_ok
&&
3814 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3815 if (tcp_parse_aligned_timestamp(tp
, th
))
3818 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3822 #ifdef CONFIG_TCP_MD5SIG
3824 * Parse MD5 Signature option
3826 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3828 int length
= (th
->doff
<< 2) - sizeof (*th
);
3829 u8
*ptr
= (u8
*)(th
+ 1);
3831 /* If the TCP option is too short, we can short cut */
3832 if (length
< TCPOLEN_MD5SIG
)
3835 while (length
> 0) {
3836 int opcode
= *ptr
++;
3847 if (opsize
< 2 || opsize
> length
)
3849 if (opcode
== TCPOPT_MD5SIG
)
3859 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3861 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3862 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3865 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3867 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3868 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3869 * extra check below makes sure this can only happen
3870 * for pure ACK frames. -DaveM
3872 * Not only, also it occurs for expired timestamps.
3875 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3876 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3877 tcp_store_ts_recent(tp
);
3881 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3883 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3884 * it can pass through stack. So, the following predicate verifies that
3885 * this segment is not used for anything but congestion avoidance or
3886 * fast retransmit. Moreover, we even are able to eliminate most of such
3887 * second order effects, if we apply some small "replay" window (~RTO)
3888 * to timestamp space.
3890 * All these measures still do not guarantee that we reject wrapped ACKs
3891 * on networks with high bandwidth, when sequence space is recycled fastly,
3892 * but it guarantees that such events will be very rare and do not affect
3893 * connection seriously. This doesn't look nice, but alas, PAWS is really
3896 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3897 * states that events when retransmit arrives after original data are rare.
3898 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3899 * the biggest problem on large power networks even with minor reordering.
3900 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3901 * up to bandwidth of 18Gigabit/sec. 8) ]
3904 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3906 struct tcp_sock
*tp
= tcp_sk(sk
);
3907 struct tcphdr
*th
= tcp_hdr(skb
);
3908 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3909 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3911 return (/* 1. Pure ACK with correct sequence number. */
3912 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3914 /* 2. ... and duplicate ACK. */
3915 ack
== tp
->snd_una
&&
3917 /* 3. ... and does not update window. */
3918 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3920 /* 4. ... and sits in replay window. */
3921 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3924 static inline int tcp_paws_discard(const struct sock
*sk
,
3925 const struct sk_buff
*skb
)
3927 const struct tcp_sock
*tp
= tcp_sk(sk
);
3928 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3929 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3930 !tcp_disordered_ack(sk
, skb
));
3933 /* Check segment sequence number for validity.
3935 * Segment controls are considered valid, if the segment
3936 * fits to the window after truncation to the window. Acceptability
3937 * of data (and SYN, FIN, of course) is checked separately.
3938 * See tcp_data_queue(), for example.
3940 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3941 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3942 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3943 * (borrowed from freebsd)
3946 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3948 return !before(end_seq
, tp
->rcv_wup
) &&
3949 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3952 /* When we get a reset we do this. */
3953 static void tcp_reset(struct sock
*sk
)
3955 /* We want the right error as BSD sees it (and indeed as we do). */
3956 switch (sk
->sk_state
) {
3958 sk
->sk_err
= ECONNREFUSED
;
3960 case TCP_CLOSE_WAIT
:
3966 sk
->sk_err
= ECONNRESET
;
3969 if (!sock_flag(sk
, SOCK_DEAD
))
3970 sk
->sk_error_report(sk
);
3976 * Process the FIN bit. This now behaves as it is supposed to work
3977 * and the FIN takes effect when it is validly part of sequence
3978 * space. Not before when we get holes.
3980 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3981 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3984 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3985 * close and we go into CLOSING (and later onto TIME-WAIT)
3987 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3989 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3991 struct tcp_sock
*tp
= tcp_sk(sk
);
3993 inet_csk_schedule_ack(sk
);
3995 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3996 sock_set_flag(sk
, SOCK_DONE
);
3998 switch (sk
->sk_state
) {
4000 case TCP_ESTABLISHED
:
4001 /* Move to CLOSE_WAIT */
4002 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4003 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4006 case TCP_CLOSE_WAIT
:
4008 /* Received a retransmission of the FIN, do
4013 /* RFC793: Remain in the LAST-ACK state. */
4017 /* This case occurs when a simultaneous close
4018 * happens, we must ack the received FIN and
4019 * enter the CLOSING state.
4022 tcp_set_state(sk
, TCP_CLOSING
);
4025 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4027 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4030 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4031 * cases we should never reach this piece of code.
4033 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
4034 __func__
, sk
->sk_state
);
4038 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4039 * Probably, we should reset in this case. For now drop them.
4041 __skb_queue_purge(&tp
->out_of_order_queue
);
4042 if (tcp_is_sack(tp
))
4043 tcp_sack_reset(&tp
->rx_opt
);
4046 if (!sock_flag(sk
, SOCK_DEAD
)) {
4047 sk
->sk_state_change(sk
);
4049 /* Do not send POLL_HUP for half duplex close. */
4050 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4051 sk
->sk_state
== TCP_CLOSE
)
4052 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4054 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4058 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4061 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4062 if (before(seq
, sp
->start_seq
))
4063 sp
->start_seq
= seq
;
4064 if (after(end_seq
, sp
->end_seq
))
4065 sp
->end_seq
= end_seq
;
4071 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4073 struct tcp_sock
*tp
= tcp_sk(sk
);
4075 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4078 if (before(seq
, tp
->rcv_nxt
))
4079 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4081 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4083 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4085 tp
->rx_opt
.dsack
= 1;
4086 tp
->duplicate_sack
[0].start_seq
= seq
;
4087 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4088 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ 1;
4092 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4094 struct tcp_sock
*tp
= tcp_sk(sk
);
4096 if (!tp
->rx_opt
.dsack
)
4097 tcp_dsack_set(sk
, seq
, end_seq
);
4099 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4102 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
4104 struct tcp_sock
*tp
= tcp_sk(sk
);
4106 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4107 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4108 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4109 tcp_enter_quickack_mode(sk
);
4111 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4112 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4114 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4115 end_seq
= tp
->rcv_nxt
;
4116 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4123 /* These routines update the SACK block as out-of-order packets arrive or
4124 * in-order packets close up the sequence space.
4126 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4129 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4130 struct tcp_sack_block
*swalk
= sp
+ 1;
4132 /* See if the recent change to the first SACK eats into
4133 * or hits the sequence space of other SACK blocks, if so coalesce.
4135 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4136 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4139 /* Zap SWALK, by moving every further SACK up by one slot.
4140 * Decrease num_sacks.
4142 tp
->rx_opt
.num_sacks
--;
4143 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
4145 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4149 this_sack
++, swalk
++;
4153 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
,
4154 struct tcp_sack_block
*sack2
)
4158 tmp
= sack1
->start_seq
;
4159 sack1
->start_seq
= sack2
->start_seq
;
4160 sack2
->start_seq
= tmp
;
4162 tmp
= sack1
->end_seq
;
4163 sack1
->end_seq
= sack2
->end_seq
;
4164 sack2
->end_seq
= tmp
;
4167 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4169 struct tcp_sock
*tp
= tcp_sk(sk
);
4170 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4171 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4177 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4178 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4179 /* Rotate this_sack to the first one. */
4180 for (; this_sack
> 0; this_sack
--, sp
--)
4181 tcp_sack_swap(sp
, sp
- 1);
4183 tcp_sack_maybe_coalesce(tp
);
4188 /* Could not find an adjacent existing SACK, build a new one,
4189 * put it at the front, and shift everyone else down. We
4190 * always know there is at least one SACK present already here.
4192 * If the sack array is full, forget about the last one.
4194 if (this_sack
>= TCP_NUM_SACKS
) {
4196 tp
->rx_opt
.num_sacks
--;
4199 for (; this_sack
> 0; this_sack
--, sp
--)
4203 /* Build the new head SACK, and we're done. */
4204 sp
->start_seq
= seq
;
4205 sp
->end_seq
= end_seq
;
4206 tp
->rx_opt
.num_sacks
++;
4207 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
4210 /* RCV.NXT advances, some SACKs should be eaten. */
4212 static void tcp_sack_remove(struct tcp_sock
*tp
)
4214 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4215 int num_sacks
= tp
->rx_opt
.num_sacks
;
4218 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4219 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4220 tp
->rx_opt
.num_sacks
= 0;
4221 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
4225 for (this_sack
= 0; this_sack
< num_sacks
;) {
4226 /* Check if the start of the sack is covered by RCV.NXT. */
4227 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4230 /* RCV.NXT must cover all the block! */
4231 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4233 /* Zap this SACK, by moving forward any other SACKS. */
4234 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4235 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4242 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
4243 tp
->rx_opt
.num_sacks
= num_sacks
;
4244 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
4249 /* This one checks to see if we can put data from the
4250 * out_of_order queue into the receive_queue.
4252 static void tcp_ofo_queue(struct sock
*sk
)
4254 struct tcp_sock
*tp
= tcp_sk(sk
);
4255 __u32 dsack_high
= tp
->rcv_nxt
;
4256 struct sk_buff
*skb
;
4258 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4259 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4262 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4263 __u32 dsack
= dsack_high
;
4264 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4265 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4266 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4269 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4270 SOCK_DEBUG(sk
, "ofo packet was already received \n");
4271 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4275 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4276 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4277 TCP_SKB_CB(skb
)->end_seq
);
4279 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4280 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4281 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4282 if (tcp_hdr(skb
)->fin
)
4283 tcp_fin(skb
, sk
, tcp_hdr(skb
));
4287 static int tcp_prune_ofo_queue(struct sock
*sk
);
4288 static int tcp_prune_queue(struct sock
*sk
);
4290 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
4292 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4293 !sk_rmem_schedule(sk
, size
)) {
4295 if (tcp_prune_queue(sk
) < 0)
4298 if (!sk_rmem_schedule(sk
, size
)) {
4299 if (!tcp_prune_ofo_queue(sk
))
4302 if (!sk_rmem_schedule(sk
, size
))
4309 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4311 struct tcphdr
*th
= tcp_hdr(skb
);
4312 struct tcp_sock
*tp
= tcp_sk(sk
);
4315 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4318 __skb_pull(skb
, th
->doff
* 4);
4320 TCP_ECN_accept_cwr(tp
, skb
);
4322 if (tp
->rx_opt
.dsack
) {
4323 tp
->rx_opt
.dsack
= 0;
4324 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
;
4327 /* Queue data for delivery to the user.
4328 * Packets in sequence go to the receive queue.
4329 * Out of sequence packets to the out_of_order_queue.
4331 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4332 if (tcp_receive_window(tp
) == 0)
4335 /* Ok. In sequence. In window. */
4336 if (tp
->ucopy
.task
== current
&&
4337 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4338 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4339 int chunk
= min_t(unsigned int, skb
->len
,
4342 __set_current_state(TASK_RUNNING
);
4345 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4346 tp
->ucopy
.len
-= chunk
;
4347 tp
->copied_seq
+= chunk
;
4348 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4349 tcp_rcv_space_adjust(sk
);
4357 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4360 skb_set_owner_r(skb
, sk
);
4361 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4363 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4365 tcp_event_data_recv(sk
, skb
);
4367 tcp_fin(skb
, sk
, th
);
4369 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4372 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4373 * gap in queue is filled.
4375 if (skb_queue_empty(&tp
->out_of_order_queue
))
4376 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4379 if (tp
->rx_opt
.num_sacks
)
4380 tcp_sack_remove(tp
);
4382 tcp_fast_path_check(sk
);
4386 else if (!sock_flag(sk
, SOCK_DEAD
))
4387 sk
->sk_data_ready(sk
, 0);
4391 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4392 /* A retransmit, 2nd most common case. Force an immediate ack. */
4393 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4394 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4397 tcp_enter_quickack_mode(sk
);
4398 inet_csk_schedule_ack(sk
);
4404 /* Out of window. F.e. zero window probe. */
4405 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4408 tcp_enter_quickack_mode(sk
);
4410 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4411 /* Partial packet, seq < rcv_next < end_seq */
4412 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4413 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4414 TCP_SKB_CB(skb
)->end_seq
);
4416 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4418 /* If window is closed, drop tail of packet. But after
4419 * remembering D-SACK for its head made in previous line.
4421 if (!tcp_receive_window(tp
))
4426 TCP_ECN_check_ce(tp
, skb
);
4428 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4431 /* Disable header prediction. */
4433 inet_csk_schedule_ack(sk
);
4435 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4436 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4438 skb_set_owner_r(skb
, sk
);
4440 if (!skb_peek(&tp
->out_of_order_queue
)) {
4441 /* Initial out of order segment, build 1 SACK. */
4442 if (tcp_is_sack(tp
)) {
4443 tp
->rx_opt
.num_sacks
= 1;
4444 tp
->rx_opt
.dsack
= 0;
4445 tp
->rx_opt
.eff_sacks
= 1;
4446 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4447 tp
->selective_acks
[0].end_seq
=
4448 TCP_SKB_CB(skb
)->end_seq
;
4450 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4452 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
4453 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4454 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4456 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4457 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4459 if (!tp
->rx_opt
.num_sacks
||
4460 tp
->selective_acks
[0].end_seq
!= seq
)
4463 /* Common case: data arrive in order after hole. */
4464 tp
->selective_acks
[0].end_seq
= end_seq
;
4468 /* Find place to insert this segment. */
4470 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4472 } while ((skb1
= skb1
->prev
) !=
4473 (struct sk_buff
*)&tp
->out_of_order_queue
);
4475 /* Do skb overlap to previous one? */
4476 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
4477 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4478 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4479 /* All the bits are present. Drop. */
4481 tcp_dsack_set(sk
, seq
, end_seq
);
4484 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4485 /* Partial overlap. */
4486 tcp_dsack_set(sk
, seq
,
4487 TCP_SKB_CB(skb1
)->end_seq
);
4492 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4494 /* And clean segments covered by new one as whole. */
4495 while ((skb1
= skb
->next
) !=
4496 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4497 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4498 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4499 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4503 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4504 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4505 TCP_SKB_CB(skb1
)->end_seq
);
4510 if (tcp_is_sack(tp
))
4511 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4515 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4516 struct sk_buff_head
*list
)
4518 struct sk_buff
*next
= skb
->next
;
4520 __skb_unlink(skb
, list
);
4522 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4527 /* Collapse contiguous sequence of skbs head..tail with
4528 * sequence numbers start..end.
4529 * Segments with FIN/SYN are not collapsed (only because this
4533 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4534 struct sk_buff
*head
, struct sk_buff
*tail
,
4537 struct sk_buff
*skb
;
4539 /* First, check that queue is collapsible and find
4540 * the point where collapsing can be useful. */
4541 for (skb
= head
; skb
!= tail
;) {
4542 /* No new bits? It is possible on ofo queue. */
4543 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4544 skb
= tcp_collapse_one(sk
, skb
, list
);
4548 /* The first skb to collapse is:
4550 * - bloated or contains data before "start" or
4551 * overlaps to the next one.
4553 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4554 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4555 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4556 (skb
->next
!= tail
&&
4557 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4560 /* Decided to skip this, advance start seq. */
4561 start
= TCP_SKB_CB(skb
)->end_seq
;
4564 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4567 while (before(start
, end
)) {
4568 struct sk_buff
*nskb
;
4569 unsigned int header
= skb_headroom(skb
);
4570 int copy
= SKB_MAX_ORDER(header
, 0);
4572 /* Too big header? This can happen with IPv6. */
4575 if (end
- start
< copy
)
4577 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4581 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4582 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4584 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4586 skb_reserve(nskb
, header
);
4587 memcpy(nskb
->head
, skb
->head
, header
);
4588 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4589 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4590 __skb_queue_before(list
, skb
, nskb
);
4591 skb_set_owner_r(nskb
, sk
);
4593 /* Copy data, releasing collapsed skbs. */
4595 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4596 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4600 size
= min(copy
, size
);
4601 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4603 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4607 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4608 skb
= tcp_collapse_one(sk
, skb
, list
);
4610 tcp_hdr(skb
)->syn
||
4618 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4619 * and tcp_collapse() them until all the queue is collapsed.
4621 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4623 struct tcp_sock
*tp
= tcp_sk(sk
);
4624 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4625 struct sk_buff
*head
;
4631 start
= TCP_SKB_CB(skb
)->seq
;
4632 end
= TCP_SKB_CB(skb
)->end_seq
;
4638 /* Segment is terminated when we see gap or when
4639 * we are at the end of all the queue. */
4640 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4641 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4642 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4643 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4644 head
, skb
, start
, end
);
4646 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4648 /* Start new segment */
4649 start
= TCP_SKB_CB(skb
)->seq
;
4650 end
= TCP_SKB_CB(skb
)->end_seq
;
4652 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4653 start
= TCP_SKB_CB(skb
)->seq
;
4654 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4655 end
= TCP_SKB_CB(skb
)->end_seq
;
4661 * Purge the out-of-order queue.
4662 * Return true if queue was pruned.
4664 static int tcp_prune_ofo_queue(struct sock
*sk
)
4666 struct tcp_sock
*tp
= tcp_sk(sk
);
4669 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4670 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4671 __skb_queue_purge(&tp
->out_of_order_queue
);
4673 /* Reset SACK state. A conforming SACK implementation will
4674 * do the same at a timeout based retransmit. When a connection
4675 * is in a sad state like this, we care only about integrity
4676 * of the connection not performance.
4678 if (tp
->rx_opt
.sack_ok
)
4679 tcp_sack_reset(&tp
->rx_opt
);
4686 /* Reduce allocated memory if we can, trying to get
4687 * the socket within its memory limits again.
4689 * Return less than zero if we should start dropping frames
4690 * until the socket owning process reads some of the data
4691 * to stabilize the situation.
4693 static int tcp_prune_queue(struct sock
*sk
)
4695 struct tcp_sock
*tp
= tcp_sk(sk
);
4697 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4699 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4701 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4702 tcp_clamp_window(sk
);
4703 else if (tcp_memory_pressure
)
4704 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4706 tcp_collapse_ofo_queue(sk
);
4707 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4708 sk
->sk_receive_queue
.next
,
4709 (struct sk_buff
*)&sk
->sk_receive_queue
,
4710 tp
->copied_seq
, tp
->rcv_nxt
);
4713 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4716 /* Collapsing did not help, destructive actions follow.
4717 * This must not ever occur. */
4719 tcp_prune_ofo_queue(sk
);
4721 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4724 /* If we are really being abused, tell the caller to silently
4725 * drop receive data on the floor. It will get retransmitted
4726 * and hopefully then we'll have sufficient space.
4728 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4730 /* Massive buffer overcommit. */
4735 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4736 * As additional protections, we do not touch cwnd in retransmission phases,
4737 * and if application hit its sndbuf limit recently.
4739 void tcp_cwnd_application_limited(struct sock
*sk
)
4741 struct tcp_sock
*tp
= tcp_sk(sk
);
4743 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4744 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4745 /* Limited by application or receiver window. */
4746 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4747 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4748 if (win_used
< tp
->snd_cwnd
) {
4749 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4750 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4752 tp
->snd_cwnd_used
= 0;
4754 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4757 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4759 struct tcp_sock
*tp
= tcp_sk(sk
);
4761 /* If the user specified a specific send buffer setting, do
4764 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4767 /* If we are under global TCP memory pressure, do not expand. */
4768 if (tcp_memory_pressure
)
4771 /* If we are under soft global TCP memory pressure, do not expand. */
4772 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4775 /* If we filled the congestion window, do not expand. */
4776 if (tp
->packets_out
>= tp
->snd_cwnd
)
4782 /* When incoming ACK allowed to free some skb from write_queue,
4783 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4784 * on the exit from tcp input handler.
4786 * PROBLEM: sndbuf expansion does not work well with largesend.
4788 static void tcp_new_space(struct sock
*sk
)
4790 struct tcp_sock
*tp
= tcp_sk(sk
);
4792 if (tcp_should_expand_sndbuf(sk
)) {
4793 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4794 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
4795 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4796 tp
->reordering
+ 1);
4797 sndmem
*= 2 * demanded
;
4798 if (sndmem
> sk
->sk_sndbuf
)
4799 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4800 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4803 sk
->sk_write_space(sk
);
4806 static void tcp_check_space(struct sock
*sk
)
4808 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4809 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4810 if (sk
->sk_socket
&&
4811 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4816 static inline void tcp_data_snd_check(struct sock
*sk
)
4818 tcp_push_pending_frames(sk
);
4819 tcp_check_space(sk
);
4823 * Check if sending an ack is needed.
4825 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4827 struct tcp_sock
*tp
= tcp_sk(sk
);
4829 /* More than one full frame received... */
4830 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4831 /* ... and right edge of window advances far enough.
4832 * (tcp_recvmsg() will send ACK otherwise). Or...
4834 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4835 /* We ACK each frame or... */
4836 tcp_in_quickack_mode(sk
) ||
4837 /* We have out of order data. */
4838 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4839 /* Then ack it now */
4842 /* Else, send delayed ack. */
4843 tcp_send_delayed_ack(sk
);
4847 static inline void tcp_ack_snd_check(struct sock
*sk
)
4849 if (!inet_csk_ack_scheduled(sk
)) {
4850 /* We sent a data segment already. */
4853 __tcp_ack_snd_check(sk
, 1);
4857 * This routine is only called when we have urgent data
4858 * signaled. Its the 'slow' part of tcp_urg. It could be
4859 * moved inline now as tcp_urg is only called from one
4860 * place. We handle URGent data wrong. We have to - as
4861 * BSD still doesn't use the correction from RFC961.
4862 * For 1003.1g we should support a new option TCP_STDURG to permit
4863 * either form (or just set the sysctl tcp_stdurg).
4866 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4868 struct tcp_sock
*tp
= tcp_sk(sk
);
4869 u32 ptr
= ntohs(th
->urg_ptr
);
4871 if (ptr
&& !sysctl_tcp_stdurg
)
4873 ptr
+= ntohl(th
->seq
);
4875 /* Ignore urgent data that we've already seen and read. */
4876 if (after(tp
->copied_seq
, ptr
))
4879 /* Do not replay urg ptr.
4881 * NOTE: interesting situation not covered by specs.
4882 * Misbehaving sender may send urg ptr, pointing to segment,
4883 * which we already have in ofo queue. We are not able to fetch
4884 * such data and will stay in TCP_URG_NOTYET until will be eaten
4885 * by recvmsg(). Seems, we are not obliged to handle such wicked
4886 * situations. But it is worth to think about possibility of some
4887 * DoSes using some hypothetical application level deadlock.
4889 if (before(ptr
, tp
->rcv_nxt
))
4892 /* Do we already have a newer (or duplicate) urgent pointer? */
4893 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4896 /* Tell the world about our new urgent pointer. */
4899 /* We may be adding urgent data when the last byte read was
4900 * urgent. To do this requires some care. We cannot just ignore
4901 * tp->copied_seq since we would read the last urgent byte again
4902 * as data, nor can we alter copied_seq until this data arrives
4903 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4905 * NOTE. Double Dutch. Rendering to plain English: author of comment
4906 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4907 * and expect that both A and B disappear from stream. This is _wrong_.
4908 * Though this happens in BSD with high probability, this is occasional.
4909 * Any application relying on this is buggy. Note also, that fix "works"
4910 * only in this artificial test. Insert some normal data between A and B and we will
4911 * decline of BSD again. Verdict: it is better to remove to trap
4914 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4915 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4916 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4918 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4919 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4924 tp
->urg_data
= TCP_URG_NOTYET
;
4927 /* Disable header prediction. */
4931 /* This is the 'fast' part of urgent handling. */
4932 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4934 struct tcp_sock
*tp
= tcp_sk(sk
);
4936 /* Check if we get a new urgent pointer - normally not. */
4938 tcp_check_urg(sk
, th
);
4940 /* Do we wait for any urgent data? - normally not... */
4941 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4942 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4945 /* Is the urgent pointer pointing into this packet? */
4946 if (ptr
< skb
->len
) {
4948 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4950 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4951 if (!sock_flag(sk
, SOCK_DEAD
))
4952 sk
->sk_data_ready(sk
, 0);
4957 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4959 struct tcp_sock
*tp
= tcp_sk(sk
);
4960 int chunk
= skb
->len
- hlen
;
4964 if (skb_csum_unnecessary(skb
))
4965 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4967 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4971 tp
->ucopy
.len
-= chunk
;
4972 tp
->copied_seq
+= chunk
;
4973 tcp_rcv_space_adjust(sk
);
4980 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4981 struct sk_buff
*skb
)
4985 if (sock_owned_by_user(sk
)) {
4987 result
= __tcp_checksum_complete(skb
);
4990 result
= __tcp_checksum_complete(skb
);
4995 static inline int tcp_checksum_complete_user(struct sock
*sk
,
4996 struct sk_buff
*skb
)
4998 return !skb_csum_unnecessary(skb
) &&
4999 __tcp_checksum_complete_user(sk
, skb
);
5002 #ifdef CONFIG_NET_DMA
5003 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5006 struct tcp_sock
*tp
= tcp_sk(sk
);
5007 int chunk
= skb
->len
- hlen
;
5009 int copied_early
= 0;
5011 if (tp
->ucopy
.wakeup
)
5014 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5015 tp
->ucopy
.dma_chan
= get_softnet_dma();
5017 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5019 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5021 tp
->ucopy
.iov
, chunk
,
5022 tp
->ucopy
.pinned_list
);
5027 tp
->ucopy
.dma_cookie
= dma_cookie
;
5030 tp
->ucopy
.len
-= chunk
;
5031 tp
->copied_seq
+= chunk
;
5032 tcp_rcv_space_adjust(sk
);
5034 if ((tp
->ucopy
.len
== 0) ||
5035 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5036 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5037 tp
->ucopy
.wakeup
= 1;
5038 sk
->sk_data_ready(sk
, 0);
5040 } else if (chunk
> 0) {
5041 tp
->ucopy
.wakeup
= 1;
5042 sk
->sk_data_ready(sk
, 0);
5045 return copied_early
;
5047 #endif /* CONFIG_NET_DMA */
5049 /* Does PAWS and seqno based validation of an incoming segment, flags will
5050 * play significant role here.
5052 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5053 struct tcphdr
*th
, int syn_inerr
)
5055 struct tcp_sock
*tp
= tcp_sk(sk
);
5057 /* RFC1323: H1. Apply PAWS check first. */
5058 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5059 tcp_paws_discard(sk
, skb
)) {
5061 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5062 tcp_send_dupack(sk
, skb
);
5065 /* Reset is accepted even if it did not pass PAWS. */
5068 /* Step 1: check sequence number */
5069 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5070 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5071 * (RST) segments are validated by checking their SEQ-fields."
5072 * And page 69: "If an incoming segment is not acceptable,
5073 * an acknowledgment should be sent in reply (unless the RST
5074 * bit is set, if so drop the segment and return)".
5077 tcp_send_dupack(sk
, skb
);
5081 /* Step 2: check RST bit */
5087 /* ts_recent update must be made after we are sure that the packet
5090 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5092 /* step 3: check security and precedence [ignored] */
5094 /* step 4: Check for a SYN in window. */
5095 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5097 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5098 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5111 * TCP receive function for the ESTABLISHED state.
5113 * It is split into a fast path and a slow path. The fast path is
5115 * - A zero window was announced from us - zero window probing
5116 * is only handled properly in the slow path.
5117 * - Out of order segments arrived.
5118 * - Urgent data is expected.
5119 * - There is no buffer space left
5120 * - Unexpected TCP flags/window values/header lengths are received
5121 * (detected by checking the TCP header against pred_flags)
5122 * - Data is sent in both directions. Fast path only supports pure senders
5123 * or pure receivers (this means either the sequence number or the ack
5124 * value must stay constant)
5125 * - Unexpected TCP option.
5127 * When these conditions are not satisfied it drops into a standard
5128 * receive procedure patterned after RFC793 to handle all cases.
5129 * The first three cases are guaranteed by proper pred_flags setting,
5130 * the rest is checked inline. Fast processing is turned on in
5131 * tcp_data_queue when everything is OK.
5133 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5134 struct tcphdr
*th
, unsigned len
)
5136 struct tcp_sock
*tp
= tcp_sk(sk
);
5140 * Header prediction.
5141 * The code loosely follows the one in the famous
5142 * "30 instruction TCP receive" Van Jacobson mail.
5144 * Van's trick is to deposit buffers into socket queue
5145 * on a device interrupt, to call tcp_recv function
5146 * on the receive process context and checksum and copy
5147 * the buffer to user space. smart...
5149 * Our current scheme is not silly either but we take the
5150 * extra cost of the net_bh soft interrupt processing...
5151 * We do checksum and copy also but from device to kernel.
5154 tp
->rx_opt
.saw_tstamp
= 0;
5156 /* pred_flags is 0xS?10 << 16 + snd_wnd
5157 * if header_prediction is to be made
5158 * 'S' will always be tp->tcp_header_len >> 2
5159 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5160 * turn it off (when there are holes in the receive
5161 * space for instance)
5162 * PSH flag is ignored.
5165 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5166 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
5167 int tcp_header_len
= tp
->tcp_header_len
;
5169 /* Timestamp header prediction: tcp_header_len
5170 * is automatically equal to th->doff*4 due to pred_flags
5174 /* Check timestamp */
5175 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5176 /* No? Slow path! */
5177 if (!tcp_parse_aligned_timestamp(tp
, th
))
5180 /* If PAWS failed, check it more carefully in slow path */
5181 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5184 /* DO NOT update ts_recent here, if checksum fails
5185 * and timestamp was corrupted part, it will result
5186 * in a hung connection since we will drop all
5187 * future packets due to the PAWS test.
5191 if (len
<= tcp_header_len
) {
5192 /* Bulk data transfer: sender */
5193 if (len
== tcp_header_len
) {
5194 /* Predicted packet is in window by definition.
5195 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5196 * Hence, check seq<=rcv_wup reduces to:
5198 if (tcp_header_len
==
5199 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5200 tp
->rcv_nxt
== tp
->rcv_wup
)
5201 tcp_store_ts_recent(tp
);
5203 /* We know that such packets are checksummed
5206 tcp_ack(sk
, skb
, 0);
5208 tcp_data_snd_check(sk
);
5210 } else { /* Header too small */
5211 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5216 int copied_early
= 0;
5218 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5219 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5220 #ifdef CONFIG_NET_DMA
5221 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5226 if (tp
->ucopy
.task
== current
&&
5227 sock_owned_by_user(sk
) && !copied_early
) {
5228 __set_current_state(TASK_RUNNING
);
5230 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5234 /* Predicted packet is in window by definition.
5235 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5236 * Hence, check seq<=rcv_wup reduces to:
5238 if (tcp_header_len
==
5239 (sizeof(struct tcphdr
) +
5240 TCPOLEN_TSTAMP_ALIGNED
) &&
5241 tp
->rcv_nxt
== tp
->rcv_wup
)
5242 tcp_store_ts_recent(tp
);
5244 tcp_rcv_rtt_measure_ts(sk
, skb
);
5246 __skb_pull(skb
, tcp_header_len
);
5247 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5248 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5251 tcp_cleanup_rbuf(sk
, skb
->len
);
5254 if (tcp_checksum_complete_user(sk
, skb
))
5257 /* Predicted packet is in window by definition.
5258 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5259 * Hence, check seq<=rcv_wup reduces to:
5261 if (tcp_header_len
==
5262 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5263 tp
->rcv_nxt
== tp
->rcv_wup
)
5264 tcp_store_ts_recent(tp
);
5266 tcp_rcv_rtt_measure_ts(sk
, skb
);
5268 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5271 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5273 /* Bulk data transfer: receiver */
5274 __skb_pull(skb
, tcp_header_len
);
5275 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
5276 skb_set_owner_r(skb
, sk
);
5277 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5280 tcp_event_data_recv(sk
, skb
);
5282 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5283 /* Well, only one small jumplet in fast path... */
5284 tcp_ack(sk
, skb
, FLAG_DATA
);
5285 tcp_data_snd_check(sk
);
5286 if (!inet_csk_ack_scheduled(sk
))
5290 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5291 __tcp_ack_snd_check(sk
, 0);
5293 #ifdef CONFIG_NET_DMA
5295 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5301 sk
->sk_data_ready(sk
, 0);
5307 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5311 * Standard slow path.
5314 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
5320 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5322 tcp_rcv_rtt_measure_ts(sk
, skb
);
5324 /* Process urgent data. */
5325 tcp_urg(sk
, skb
, th
);
5327 /* step 7: process the segment text */
5328 tcp_data_queue(sk
, skb
);
5330 tcp_data_snd_check(sk
);
5331 tcp_ack_snd_check(sk
);
5335 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5342 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5343 struct tcphdr
*th
, unsigned len
)
5345 struct tcp_sock
*tp
= tcp_sk(sk
);
5346 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5347 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5349 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
5353 * "If the state is SYN-SENT then
5354 * first check the ACK bit
5355 * If the ACK bit is set
5356 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5357 * a reset (unless the RST bit is set, if so drop
5358 * the segment and return)"
5360 * We do not send data with SYN, so that RFC-correct
5363 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5364 goto reset_and_undo
;
5366 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5367 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5369 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5370 goto reset_and_undo
;
5373 /* Now ACK is acceptable.
5375 * "If the RST bit is set
5376 * If the ACK was acceptable then signal the user "error:
5377 * connection reset", drop the segment, enter CLOSED state,
5378 * delete TCB, and return."
5387 * "fifth, if neither of the SYN or RST bits is set then
5388 * drop the segment and return."
5394 goto discard_and_undo
;
5397 * "If the SYN bit is on ...
5398 * are acceptable then ...
5399 * (our SYN has been ACKed), change the connection
5400 * state to ESTABLISHED..."
5403 TCP_ECN_rcv_synack(tp
, th
);
5405 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5406 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5408 /* Ok.. it's good. Set up sequence numbers and
5409 * move to established.
5411 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5412 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5414 /* RFC1323: The window in SYN & SYN/ACK segments is
5417 tp
->snd_wnd
= ntohs(th
->window
);
5418 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
5420 if (!tp
->rx_opt
.wscale_ok
) {
5421 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5422 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5425 if (tp
->rx_opt
.saw_tstamp
) {
5426 tp
->rx_opt
.tstamp_ok
= 1;
5427 tp
->tcp_header_len
=
5428 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5429 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5430 tcp_store_ts_recent(tp
);
5432 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5435 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5436 tcp_enable_fack(tp
);
5439 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5440 tcp_initialize_rcv_mss(sk
);
5442 /* Remember, tcp_poll() does not lock socket!
5443 * Change state from SYN-SENT only after copied_seq
5444 * is initialized. */
5445 tp
->copied_seq
= tp
->rcv_nxt
;
5447 tcp_set_state(sk
, TCP_ESTABLISHED
);
5449 security_inet_conn_established(sk
, skb
);
5451 /* Make sure socket is routed, for correct metrics. */
5452 icsk
->icsk_af_ops
->rebuild_header(sk
);
5454 tcp_init_metrics(sk
);
5456 tcp_init_congestion_control(sk
);
5458 /* Prevent spurious tcp_cwnd_restart() on first data
5461 tp
->lsndtime
= tcp_time_stamp
;
5463 tcp_init_buffer_space(sk
);
5465 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5466 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5468 if (!tp
->rx_opt
.snd_wscale
)
5469 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5473 if (!sock_flag(sk
, SOCK_DEAD
)) {
5474 sk
->sk_state_change(sk
);
5475 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5478 if (sk
->sk_write_pending
||
5479 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5480 icsk
->icsk_ack
.pingpong
) {
5481 /* Save one ACK. Data will be ready after
5482 * several ticks, if write_pending is set.
5484 * It may be deleted, but with this feature tcpdumps
5485 * look so _wonderfully_ clever, that I was not able
5486 * to stand against the temptation 8) --ANK
5488 inet_csk_schedule_ack(sk
);
5489 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5490 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5491 tcp_incr_quickack(sk
);
5492 tcp_enter_quickack_mode(sk
);
5493 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5494 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5505 /* No ACK in the segment */
5509 * "If the RST bit is set
5511 * Otherwise (no ACK) drop the segment and return."
5514 goto discard_and_undo
;
5518 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5519 tcp_paws_check(&tp
->rx_opt
, 0))
5520 goto discard_and_undo
;
5523 /* We see SYN without ACK. It is attempt of
5524 * simultaneous connect with crossed SYNs.
5525 * Particularly, it can be connect to self.
5527 tcp_set_state(sk
, TCP_SYN_RECV
);
5529 if (tp
->rx_opt
.saw_tstamp
) {
5530 tp
->rx_opt
.tstamp_ok
= 1;
5531 tcp_store_ts_recent(tp
);
5532 tp
->tcp_header_len
=
5533 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5535 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5538 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5539 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5541 /* RFC1323: The window in SYN & SYN/ACK segments is
5544 tp
->snd_wnd
= ntohs(th
->window
);
5545 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5546 tp
->max_window
= tp
->snd_wnd
;
5548 TCP_ECN_rcv_syn(tp
, th
);
5551 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5552 tcp_initialize_rcv_mss(sk
);
5554 tcp_send_synack(sk
);
5556 /* Note, we could accept data and URG from this segment.
5557 * There are no obstacles to make this.
5559 * However, if we ignore data in ACKless segments sometimes,
5560 * we have no reasons to accept it sometimes.
5561 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5562 * is not flawless. So, discard packet for sanity.
5563 * Uncomment this return to process the data.
5570 /* "fifth, if neither of the SYN or RST bits is set then
5571 * drop the segment and return."
5575 tcp_clear_options(&tp
->rx_opt
);
5576 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5580 tcp_clear_options(&tp
->rx_opt
);
5581 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5586 * This function implements the receiving procedure of RFC 793 for
5587 * all states except ESTABLISHED and TIME_WAIT.
5588 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5589 * address independent.
5592 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5593 struct tcphdr
*th
, unsigned len
)
5595 struct tcp_sock
*tp
= tcp_sk(sk
);
5596 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5600 tp
->rx_opt
.saw_tstamp
= 0;
5602 switch (sk
->sk_state
) {
5614 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5617 /* Now we have several options: In theory there is
5618 * nothing else in the frame. KA9Q has an option to
5619 * send data with the syn, BSD accepts data with the
5620 * syn up to the [to be] advertised window and
5621 * Solaris 2.1 gives you a protocol error. For now
5622 * we just ignore it, that fits the spec precisely
5623 * and avoids incompatibilities. It would be nice in
5624 * future to drop through and process the data.
5626 * Now that TTCP is starting to be used we ought to
5628 * But, this leaves one open to an easy denial of
5629 * service attack, and SYN cookies can't defend
5630 * against this problem. So, we drop the data
5631 * in the interest of security over speed unless
5632 * it's still in use.
5640 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5644 /* Do step6 onward by hand. */
5645 tcp_urg(sk
, skb
, th
);
5647 tcp_data_snd_check(sk
);
5651 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5655 /* step 5: check the ACK field */
5657 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5659 switch (sk
->sk_state
) {
5662 tp
->copied_seq
= tp
->rcv_nxt
;
5664 tcp_set_state(sk
, TCP_ESTABLISHED
);
5665 sk
->sk_state_change(sk
);
5667 /* Note, that this wakeup is only for marginal
5668 * crossed SYN case. Passively open sockets
5669 * are not waked up, because sk->sk_sleep ==
5670 * NULL and sk->sk_socket == NULL.
5674 SOCK_WAKE_IO
, POLL_OUT
);
5676 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5677 tp
->snd_wnd
= ntohs(th
->window
) <<
5678 tp
->rx_opt
.snd_wscale
;
5679 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5680 TCP_SKB_CB(skb
)->seq
);
5682 /* tcp_ack considers this ACK as duplicate
5683 * and does not calculate rtt.
5684 * Fix it at least with timestamps.
5686 if (tp
->rx_opt
.saw_tstamp
&&
5687 tp
->rx_opt
.rcv_tsecr
&& !tp
->srtt
)
5688 tcp_ack_saw_tstamp(sk
, 0);
5690 if (tp
->rx_opt
.tstamp_ok
)
5691 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5693 /* Make sure socket is routed, for
5696 icsk
->icsk_af_ops
->rebuild_header(sk
);
5698 tcp_init_metrics(sk
);
5700 tcp_init_congestion_control(sk
);
5702 /* Prevent spurious tcp_cwnd_restart() on
5703 * first data packet.
5705 tp
->lsndtime
= tcp_time_stamp
;
5708 tcp_initialize_rcv_mss(sk
);
5709 tcp_init_buffer_space(sk
);
5710 tcp_fast_path_on(tp
);
5717 if (tp
->snd_una
== tp
->write_seq
) {
5718 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5719 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5720 dst_confirm(sk
->sk_dst_cache
);
5722 if (!sock_flag(sk
, SOCK_DEAD
))
5723 /* Wake up lingering close() */
5724 sk
->sk_state_change(sk
);
5728 if (tp
->linger2
< 0 ||
5729 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5730 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5732 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5736 tmo
= tcp_fin_time(sk
);
5737 if (tmo
> TCP_TIMEWAIT_LEN
) {
5738 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5739 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5740 /* Bad case. We could lose such FIN otherwise.
5741 * It is not a big problem, but it looks confusing
5742 * and not so rare event. We still can lose it now,
5743 * if it spins in bh_lock_sock(), but it is really
5746 inet_csk_reset_keepalive_timer(sk
, tmo
);
5748 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5756 if (tp
->snd_una
== tp
->write_seq
) {
5757 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5763 if (tp
->snd_una
== tp
->write_seq
) {
5764 tcp_update_metrics(sk
);
5773 /* step 6: check the URG bit */
5774 tcp_urg(sk
, skb
, th
);
5776 /* step 7: process the segment text */
5777 switch (sk
->sk_state
) {
5778 case TCP_CLOSE_WAIT
:
5781 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5785 /* RFC 793 says to queue data in these states,
5786 * RFC 1122 says we MUST send a reset.
5787 * BSD 4.4 also does reset.
5789 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5790 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5791 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5792 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5798 case TCP_ESTABLISHED
:
5799 tcp_data_queue(sk
, skb
);
5804 /* tcp_data could move socket to TIME-WAIT */
5805 if (sk
->sk_state
!= TCP_CLOSE
) {
5806 tcp_data_snd_check(sk
);
5807 tcp_ack_snd_check(sk
);
5817 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5818 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5819 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
5820 EXPORT_SYMBOL(tcp_parse_options
);
5821 #ifdef CONFIG_TCP_MD5SIG
5822 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
5824 EXPORT_SYMBOL(tcp_rcv_established
);
5825 EXPORT_SYMBOL(tcp_rcv_state_process
);
5826 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);