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
) ||
1579 (skb
== tcp_send_head(sk
)) ||
1580 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1581 (mss
!= tcp_shift_mss(skb
)))
1585 if (skb_shift(prev
, skb
, len
)) {
1586 pcount
+= tcp_skb_pcount(skb
);
1587 tcp_shifted_skb(sk
, prev
, skb
, tcp_skb_pcount(skb
), len
,
1588 *fack_count
, reord
, flag
, mss
);
1592 *fack_count
+= pcount
;
1599 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1603 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1604 struct tcp_sack_block
*next_dup
,
1605 u32 start_seq
, u32 end_seq
,
1606 int dup_sack_in
, int *fack_count
,
1607 int *reord
, int *flag
)
1609 struct tcp_sock
*tp
= tcp_sk(sk
);
1610 struct sk_buff
*tmp
;
1612 tcp_for_write_queue_from(skb
, sk
) {
1614 int dup_sack
= dup_sack_in
;
1616 if (skb
== tcp_send_head(sk
))
1619 /* queue is in-order => we can short-circuit the walk early */
1620 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1623 if ((next_dup
!= NULL
) &&
1624 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1625 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1626 next_dup
->start_seq
,
1632 /* skb reference here is a bit tricky to get right, since
1633 * shifting can eat and free both this skb and the next,
1634 * so not even _safe variant of the loop is enough.
1637 tmp
= tcp_shift_skb_data(sk
, skb
, start_seq
,
1639 fack_count
, reord
, flag
);
1648 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1654 if (unlikely(in_sack
< 0))
1658 *flag
|= tcp_sacktag_one(skb
, sk
, reord
, dup_sack
,
1660 &(TCP_SKB_CB(skb
)->sacked
),
1661 tcp_skb_pcount(skb
));
1663 if (!before(TCP_SKB_CB(skb
)->seq
,
1664 tcp_highest_sack_seq(tp
)))
1665 tcp_advance_highest_sack(sk
, skb
);
1668 *fack_count
+= tcp_skb_pcount(skb
);
1673 /* Avoid all extra work that is being done by sacktag while walking in
1676 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1677 u32 skip_to_seq
, int *fack_count
)
1679 tcp_for_write_queue_from(skb
, sk
) {
1680 if (skb
== tcp_send_head(sk
))
1683 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1686 *fack_count
+= tcp_skb_pcount(skb
);
1691 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1693 struct tcp_sack_block
*next_dup
,
1695 int *fack_count
, int *reord
,
1698 if (next_dup
== NULL
)
1701 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1702 skb
= tcp_sacktag_skip(skb
, sk
, next_dup
->start_seq
, fack_count
);
1703 skb
= tcp_sacktag_walk(skb
, sk
, NULL
,
1704 next_dup
->start_seq
, next_dup
->end_seq
,
1705 1, fack_count
, reord
, flag
);
1711 static int tcp_sack_cache_ok(struct tcp_sock
*tp
, struct tcp_sack_block
*cache
)
1713 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1717 tcp_sacktag_write_queue(struct sock
*sk
, struct sk_buff
*ack_skb
,
1720 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1721 struct tcp_sock
*tp
= tcp_sk(sk
);
1722 unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1723 TCP_SKB_CB(ack_skb
)->sacked
);
1724 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1725 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1726 struct tcp_sack_block
*cache
;
1727 struct sk_buff
*skb
;
1728 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1730 int reord
= tp
->packets_out
;
1732 int found_dup_sack
= 0;
1735 int first_sack_index
;
1737 if (!tp
->sacked_out
) {
1738 if (WARN_ON(tp
->fackets_out
))
1739 tp
->fackets_out
= 0;
1740 tcp_highest_sack_reset(sk
);
1743 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1744 num_sacks
, prior_snd_una
);
1746 flag
|= FLAG_DSACKING_ACK
;
1748 /* Eliminate too old ACKs, but take into
1749 * account more or less fresh ones, they can
1750 * contain valid SACK info.
1752 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1755 if (!tp
->packets_out
)
1759 first_sack_index
= 0;
1760 for (i
= 0; i
< num_sacks
; i
++) {
1761 int dup_sack
= !i
&& found_dup_sack
;
1763 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1764 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1766 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1767 sp
[used_sacks
].start_seq
,
1768 sp
[used_sacks
].end_seq
)) {
1772 if (!tp
->undo_marker
)
1773 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1775 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1777 /* Don't count olds caused by ACK reordering */
1778 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1779 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1781 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1784 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1786 first_sack_index
= -1;
1790 /* Ignore very old stuff early */
1791 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1797 /* order SACK blocks to allow in order walk of the retrans queue */
1798 for (i
= used_sacks
- 1; i
> 0; i
--) {
1799 for (j
= 0; j
< i
; j
++) {
1800 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1801 struct tcp_sack_block tmp
;
1807 /* Track where the first SACK block goes to */
1808 if (j
== first_sack_index
)
1809 first_sack_index
= j
+ 1;
1814 skb
= tcp_write_queue_head(sk
);
1818 if (!tp
->sacked_out
) {
1819 /* It's already past, so skip checking against it */
1820 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1822 cache
= tp
->recv_sack_cache
;
1823 /* Skip empty blocks in at head of the cache */
1824 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1829 while (i
< used_sacks
) {
1830 u32 start_seq
= sp
[i
].start_seq
;
1831 u32 end_seq
= sp
[i
].end_seq
;
1832 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1833 struct tcp_sack_block
*next_dup
= NULL
;
1835 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1836 next_dup
= &sp
[i
+ 1];
1838 /* Event "B" in the comment above. */
1839 if (after(end_seq
, tp
->high_seq
))
1840 flag
|= FLAG_DATA_LOST
;
1842 /* Skip too early cached blocks */
1843 while (tcp_sack_cache_ok(tp
, cache
) &&
1844 !before(start_seq
, cache
->end_seq
))
1847 /* Can skip some work by looking recv_sack_cache? */
1848 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1849 after(end_seq
, cache
->start_seq
)) {
1852 if (before(start_seq
, cache
->start_seq
)) {
1853 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
,
1855 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1858 dup_sack
, &fack_count
,
1862 /* Rest of the block already fully processed? */
1863 if (!after(end_seq
, cache
->end_seq
))
1866 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1868 &fack_count
, &reord
,
1871 /* ...tail remains todo... */
1872 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1873 /* ...but better entrypoint exists! */
1874 skb
= tcp_highest_sack(sk
);
1877 fack_count
= tp
->fackets_out
;
1882 skb
= tcp_sacktag_skip(skb
, sk
, cache
->end_seq
,
1884 /* Check overlap against next cached too (past this one already) */
1889 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1890 skb
= tcp_highest_sack(sk
);
1893 fack_count
= tp
->fackets_out
;
1895 skb
= tcp_sacktag_skip(skb
, sk
, start_seq
, &fack_count
);
1898 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, start_seq
, end_seq
,
1899 dup_sack
, &fack_count
, &reord
, &flag
);
1902 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1903 * due to in-order walk
1905 if (after(end_seq
, tp
->frto_highmark
))
1906 flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1911 /* Clear the head of the cache sack blocks so we can skip it next time */
1912 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1913 tp
->recv_sack_cache
[i
].start_seq
= 0;
1914 tp
->recv_sack_cache
[i
].end_seq
= 0;
1916 for (j
= 0; j
< used_sacks
; j
++)
1917 tp
->recv_sack_cache
[i
++] = sp
[j
];
1919 tcp_mark_lost_retrans(sk
);
1921 tcp_verify_left_out(tp
);
1923 if ((reord
< tp
->fackets_out
) &&
1924 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1925 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1926 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
1930 #if FASTRETRANS_DEBUG > 0
1931 WARN_ON((int)tp
->sacked_out
< 0);
1932 WARN_ON((int)tp
->lost_out
< 0);
1933 WARN_ON((int)tp
->retrans_out
< 0);
1934 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1939 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1940 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1942 static int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1946 holes
= max(tp
->lost_out
, 1U);
1947 holes
= min(holes
, tp
->packets_out
);
1949 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1950 tp
->sacked_out
= tp
->packets_out
- holes
;
1956 /* If we receive more dupacks than we expected counting segments
1957 * in assumption of absent reordering, interpret this as reordering.
1958 * The only another reason could be bug in receiver TCP.
1960 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1962 struct tcp_sock
*tp
= tcp_sk(sk
);
1963 if (tcp_limit_reno_sacked(tp
))
1964 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1967 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1969 static void tcp_add_reno_sack(struct sock
*sk
)
1971 struct tcp_sock
*tp
= tcp_sk(sk
);
1973 tcp_check_reno_reordering(sk
, 0);
1974 tcp_verify_left_out(tp
);
1977 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1979 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1981 struct tcp_sock
*tp
= tcp_sk(sk
);
1984 /* One ACK acked hole. The rest eat duplicate ACKs. */
1985 if (acked
- 1 >= tp
->sacked_out
)
1988 tp
->sacked_out
-= acked
- 1;
1990 tcp_check_reno_reordering(sk
, acked
);
1991 tcp_verify_left_out(tp
);
1994 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1999 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
2001 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
2004 /* F-RTO can only be used if TCP has never retransmitted anything other than
2005 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2007 int tcp_use_frto(struct sock
*sk
)
2009 const struct tcp_sock
*tp
= tcp_sk(sk
);
2010 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2011 struct sk_buff
*skb
;
2013 if (!sysctl_tcp_frto
)
2016 /* MTU probe and F-RTO won't really play nicely along currently */
2017 if (icsk
->icsk_mtup
.probe_size
)
2020 if (tcp_is_sackfrto(tp
))
2023 /* Avoid expensive walking of rexmit queue if possible */
2024 if (tp
->retrans_out
> 1)
2027 skb
= tcp_write_queue_head(sk
);
2028 if (tcp_skb_is_last(sk
, skb
))
2030 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
2031 tcp_for_write_queue_from(skb
, sk
) {
2032 if (skb
== tcp_send_head(sk
))
2034 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2036 /* Short-circuit when first non-SACKed skb has been checked */
2037 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2043 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2044 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2045 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2046 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2047 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2048 * bits are handled if the Loss state is really to be entered (in
2049 * tcp_enter_frto_loss).
2051 * Do like tcp_enter_loss() would; when RTO expires the second time it
2053 * "Reduce ssthresh if it has not yet been made inside this window."
2055 void tcp_enter_frto(struct sock
*sk
)
2057 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2058 struct tcp_sock
*tp
= tcp_sk(sk
);
2059 struct sk_buff
*skb
;
2061 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
2062 tp
->snd_una
== tp
->high_seq
||
2063 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
2064 !icsk
->icsk_retransmits
)) {
2065 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2066 /* Our state is too optimistic in ssthresh() call because cwnd
2067 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2068 * recovery has not yet completed. Pattern would be this: RTO,
2069 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2071 * RFC4138 should be more specific on what to do, even though
2072 * RTO is quite unlikely to occur after the first Cumulative ACK
2073 * due to back-off and complexity of triggering events ...
2075 if (tp
->frto_counter
) {
2077 stored_cwnd
= tp
->snd_cwnd
;
2079 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2080 tp
->snd_cwnd
= stored_cwnd
;
2082 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2084 /* ... in theory, cong.control module could do "any tricks" in
2085 * ssthresh(), which means that ca_state, lost bits and lost_out
2086 * counter would have to be faked before the call occurs. We
2087 * consider that too expensive, unlikely and hacky, so modules
2088 * using these in ssthresh() must deal these incompatibility
2089 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2091 tcp_ca_event(sk
, CA_EVENT_FRTO
);
2094 tp
->undo_marker
= tp
->snd_una
;
2095 tp
->undo_retrans
= 0;
2097 skb
= tcp_write_queue_head(sk
);
2098 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2099 tp
->undo_marker
= 0;
2100 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2101 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2102 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2104 tcp_verify_left_out(tp
);
2106 /* Too bad if TCP was application limited */
2107 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2109 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2110 * The last condition is necessary at least in tp->frto_counter case.
2112 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
2113 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
2114 after(tp
->high_seq
, tp
->snd_una
)) {
2115 tp
->frto_highmark
= tp
->high_seq
;
2117 tp
->frto_highmark
= tp
->snd_nxt
;
2119 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
2120 tp
->high_seq
= tp
->snd_nxt
;
2121 tp
->frto_counter
= 1;
2124 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2125 * which indicates that we should follow the traditional RTO recovery,
2126 * i.e. mark everything lost and do go-back-N retransmission.
2128 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
2130 struct tcp_sock
*tp
= tcp_sk(sk
);
2131 struct sk_buff
*skb
;
2134 tp
->retrans_out
= 0;
2135 if (tcp_is_reno(tp
))
2136 tcp_reset_reno_sack(tp
);
2138 tcp_for_write_queue(skb
, sk
) {
2139 if (skb
== tcp_send_head(sk
))
2142 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2144 * Count the retransmission made on RTO correctly (only when
2145 * waiting for the first ACK and did not get it)...
2147 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
2148 /* For some reason this R-bit might get cleared? */
2149 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
2150 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2151 /* ...enter this if branch just for the first segment */
2152 flag
|= FLAG_DATA_ACKED
;
2154 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2155 tp
->undo_marker
= 0;
2156 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2159 /* Marking forward transmissions that were made after RTO lost
2160 * can cause unnecessary retransmissions in some scenarios,
2161 * SACK blocks will mitigate that in some but not in all cases.
2162 * We used to not mark them but it was causing break-ups with
2163 * receivers that do only in-order receival.
2165 * TODO: we could detect presence of such receiver and select
2166 * different behavior per flow.
2168 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2169 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2170 tp
->lost_out
+= tcp_skb_pcount(skb
);
2171 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2174 tcp_verify_left_out(tp
);
2176 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2177 tp
->snd_cwnd_cnt
= 0;
2178 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2179 tp
->frto_counter
= 0;
2180 tp
->bytes_acked
= 0;
2182 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2183 sysctl_tcp_reordering
);
2184 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2185 tp
->high_seq
= tp
->snd_nxt
;
2186 TCP_ECN_queue_cwr(tp
);
2188 tcp_clear_all_retrans_hints(tp
);
2191 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2193 tp
->retrans_out
= 0;
2196 tp
->undo_marker
= 0;
2197 tp
->undo_retrans
= 0;
2200 void tcp_clear_retrans(struct tcp_sock
*tp
)
2202 tcp_clear_retrans_partial(tp
);
2204 tp
->fackets_out
= 0;
2208 /* Enter Loss state. If "how" is not zero, forget all SACK information
2209 * and reset tags completely, otherwise preserve SACKs. If receiver
2210 * dropped its ofo queue, we will know this due to reneging detection.
2212 void tcp_enter_loss(struct sock
*sk
, int how
)
2214 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2215 struct tcp_sock
*tp
= tcp_sk(sk
);
2216 struct sk_buff
*skb
;
2218 /* Reduce ssthresh if it has not yet been made inside this window. */
2219 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2220 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2221 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2222 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2223 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2226 tp
->snd_cwnd_cnt
= 0;
2227 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2229 tp
->bytes_acked
= 0;
2230 tcp_clear_retrans_partial(tp
);
2232 if (tcp_is_reno(tp
))
2233 tcp_reset_reno_sack(tp
);
2236 /* Push undo marker, if it was plain RTO and nothing
2237 * was retransmitted. */
2238 tp
->undo_marker
= tp
->snd_una
;
2241 tp
->fackets_out
= 0;
2243 tcp_clear_all_retrans_hints(tp
);
2245 tcp_for_write_queue(skb
, sk
) {
2246 if (skb
== tcp_send_head(sk
))
2249 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2250 tp
->undo_marker
= 0;
2251 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2252 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2253 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2254 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2255 tp
->lost_out
+= tcp_skb_pcount(skb
);
2256 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2259 tcp_verify_left_out(tp
);
2261 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2262 sysctl_tcp_reordering
);
2263 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2264 tp
->high_seq
= tp
->snd_nxt
;
2265 TCP_ECN_queue_cwr(tp
);
2266 /* Abort F-RTO algorithm if one is in progress */
2267 tp
->frto_counter
= 0;
2270 /* If ACK arrived pointing to a remembered SACK, it means that our
2271 * remembered SACKs do not reflect real state of receiver i.e.
2272 * receiver _host_ is heavily congested (or buggy).
2274 * Do processing similar to RTO timeout.
2276 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2278 if (flag
& FLAG_SACK_RENEGING
) {
2279 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2280 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2282 tcp_enter_loss(sk
, 1);
2283 icsk
->icsk_retransmits
++;
2284 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2285 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2286 icsk
->icsk_rto
, TCP_RTO_MAX
);
2292 static inline int tcp_fackets_out(struct tcp_sock
*tp
)
2294 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2297 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2298 * counter when SACK is enabled (without SACK, sacked_out is used for
2301 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2302 * segments up to the highest received SACK block so far and holes in
2305 * With reordering, holes may still be in flight, so RFC3517 recovery
2306 * uses pure sacked_out (total number of SACKed segments) even though
2307 * it violates the RFC that uses duplicate ACKs, often these are equal
2308 * but when e.g. out-of-window ACKs or packet duplication occurs,
2309 * they differ. Since neither occurs due to loss, TCP should really
2312 static inline int tcp_dupack_heurestics(struct tcp_sock
*tp
)
2314 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2317 static inline int tcp_skb_timedout(struct sock
*sk
, struct sk_buff
*skb
)
2319 return (tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
);
2322 static inline int tcp_head_timedout(struct sock
*sk
)
2324 struct tcp_sock
*tp
= tcp_sk(sk
);
2326 return tp
->packets_out
&&
2327 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2330 /* Linux NewReno/SACK/FACK/ECN state machine.
2331 * --------------------------------------
2333 * "Open" Normal state, no dubious events, fast path.
2334 * "Disorder" In all the respects it is "Open",
2335 * but requires a bit more attention. It is entered when
2336 * we see some SACKs or dupacks. It is split of "Open"
2337 * mainly to move some processing from fast path to slow one.
2338 * "CWR" CWND was reduced due to some Congestion Notification event.
2339 * It can be ECN, ICMP source quench, local device congestion.
2340 * "Recovery" CWND was reduced, we are fast-retransmitting.
2341 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2343 * tcp_fastretrans_alert() is entered:
2344 * - each incoming ACK, if state is not "Open"
2345 * - when arrived ACK is unusual, namely:
2350 * Counting packets in flight is pretty simple.
2352 * in_flight = packets_out - left_out + retrans_out
2354 * packets_out is SND.NXT-SND.UNA counted in packets.
2356 * retrans_out is number of retransmitted segments.
2358 * left_out is number of segments left network, but not ACKed yet.
2360 * left_out = sacked_out + lost_out
2362 * sacked_out: Packets, which arrived to receiver out of order
2363 * and hence not ACKed. With SACKs this number is simply
2364 * amount of SACKed data. Even without SACKs
2365 * it is easy to give pretty reliable estimate of this number,
2366 * counting duplicate ACKs.
2368 * lost_out: Packets lost by network. TCP has no explicit
2369 * "loss notification" feedback from network (for now).
2370 * It means that this number can be only _guessed_.
2371 * Actually, it is the heuristics to predict lossage that
2372 * distinguishes different algorithms.
2374 * F.e. after RTO, when all the queue is considered as lost,
2375 * lost_out = packets_out and in_flight = retrans_out.
2377 * Essentially, we have now two algorithms counting
2380 * FACK: It is the simplest heuristics. As soon as we decided
2381 * that something is lost, we decide that _all_ not SACKed
2382 * packets until the most forward SACK are lost. I.e.
2383 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2384 * It is absolutely correct estimate, if network does not reorder
2385 * packets. And it loses any connection to reality when reordering
2386 * takes place. We use FACK by default until reordering
2387 * is suspected on the path to this destination.
2389 * NewReno: when Recovery is entered, we assume that one segment
2390 * is lost (classic Reno). While we are in Recovery and
2391 * a partial ACK arrives, we assume that one more packet
2392 * is lost (NewReno). This heuristics are the same in NewReno
2395 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2396 * deflation etc. CWND is real congestion window, never inflated, changes
2397 * only according to classic VJ rules.
2399 * Really tricky (and requiring careful tuning) part of algorithm
2400 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2401 * The first determines the moment _when_ we should reduce CWND and,
2402 * hence, slow down forward transmission. In fact, it determines the moment
2403 * when we decide that hole is caused by loss, rather than by a reorder.
2405 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2406 * holes, caused by lost packets.
2408 * And the most logically complicated part of algorithm is undo
2409 * heuristics. We detect false retransmits due to both too early
2410 * fast retransmit (reordering) and underestimated RTO, analyzing
2411 * timestamps and D-SACKs. When we detect that some segments were
2412 * retransmitted by mistake and CWND reduction was wrong, we undo
2413 * window reduction and abort recovery phase. This logic is hidden
2414 * inside several functions named tcp_try_undo_<something>.
2417 /* This function decides, when we should leave Disordered state
2418 * and enter Recovery phase, reducing congestion window.
2420 * Main question: may we further continue forward transmission
2421 * with the same cwnd?
2423 static int tcp_time_to_recover(struct sock
*sk
)
2425 struct tcp_sock
*tp
= tcp_sk(sk
);
2428 /* Do not perform any recovery during F-RTO algorithm */
2429 if (tp
->frto_counter
)
2432 /* Trick#1: The loss is proven. */
2436 /* Not-A-Trick#2 : Classic rule... */
2437 if (tcp_dupack_heurestics(tp
) > tp
->reordering
)
2440 /* Trick#3 : when we use RFC2988 timer restart, fast
2441 * retransmit can be triggered by timeout of queue head.
2443 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2446 /* Trick#4: It is still not OK... But will it be useful to delay
2449 packets_out
= tp
->packets_out
;
2450 if (packets_out
<= tp
->reordering
&&
2451 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2452 !tcp_may_send_now(sk
)) {
2453 /* We have nothing to send. This connection is limited
2454 * either by receiver window or by application.
2462 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2463 * is against sacked "cnt", otherwise it's against facked "cnt"
2465 static void tcp_mark_head_lost(struct sock
*sk
, int packets
)
2467 struct tcp_sock
*tp
= tcp_sk(sk
);
2468 struct sk_buff
*skb
;
2473 WARN_ON(packets
> tp
->packets_out
);
2474 if (tp
->lost_skb_hint
) {
2475 skb
= tp
->lost_skb_hint
;
2476 cnt
= tp
->lost_cnt_hint
;
2478 skb
= tcp_write_queue_head(sk
);
2482 tcp_for_write_queue_from(skb
, sk
) {
2483 if (skb
== tcp_send_head(sk
))
2485 /* TODO: do this better */
2486 /* this is not the most efficient way to do this... */
2487 tp
->lost_skb_hint
= skb
;
2488 tp
->lost_cnt_hint
= cnt
;
2490 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->high_seq
))
2494 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2495 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2496 cnt
+= tcp_skb_pcount(skb
);
2498 if (cnt
> packets
) {
2499 if (tcp_is_sack(tp
) || (oldcnt
>= packets
))
2502 mss
= skb_shinfo(skb
)->gso_size
;
2503 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2509 tcp_skb_mark_lost(tp
, skb
);
2511 tcp_verify_left_out(tp
);
2514 /* Account newly detected lost packet(s) */
2516 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2518 struct tcp_sock
*tp
= tcp_sk(sk
);
2520 if (tcp_is_reno(tp
)) {
2521 tcp_mark_head_lost(sk
, 1);
2522 } else if (tcp_is_fack(tp
)) {
2523 int lost
= tp
->fackets_out
- tp
->reordering
;
2526 tcp_mark_head_lost(sk
, lost
);
2528 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2529 if (sacked_upto
< fast_rexmit
)
2530 sacked_upto
= fast_rexmit
;
2531 tcp_mark_head_lost(sk
, sacked_upto
);
2534 /* New heuristics: it is possible only after we switched
2535 * to restart timer each time when something is ACKed.
2536 * Hence, we can detect timed out packets during fast
2537 * retransmit without falling to slow start.
2539 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
)) {
2540 struct sk_buff
*skb
;
2542 skb
= tp
->scoreboard_skb_hint
? tp
->scoreboard_skb_hint
2543 : tcp_write_queue_head(sk
);
2545 tcp_for_write_queue_from(skb
, sk
) {
2546 if (skb
== tcp_send_head(sk
))
2548 if (!tcp_skb_timedout(sk
, skb
))
2551 tcp_skb_mark_lost(tp
, skb
);
2554 tp
->scoreboard_skb_hint
= skb
;
2556 tcp_verify_left_out(tp
);
2560 /* CWND moderation, preventing bursts due to too big ACKs
2561 * in dubious situations.
2563 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2565 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2566 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2567 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2570 /* Lower bound on congestion window is slow start threshold
2571 * unless congestion avoidance choice decides to overide it.
2573 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2575 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2577 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2580 /* Decrease cwnd each second ack. */
2581 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2583 struct tcp_sock
*tp
= tcp_sk(sk
);
2584 int decr
= tp
->snd_cwnd_cnt
+ 1;
2586 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2587 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2588 tp
->snd_cwnd_cnt
= decr
& 1;
2591 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2592 tp
->snd_cwnd
-= decr
;
2594 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2595 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2599 /* Nothing was retransmitted or returned timestamp is less
2600 * than timestamp of the first retransmission.
2602 static inline int tcp_packet_delayed(struct tcp_sock
*tp
)
2604 return !tp
->retrans_stamp
||
2605 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2606 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2609 /* Undo procedures. */
2611 #if FASTRETRANS_DEBUG > 1
2612 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2614 struct tcp_sock
*tp
= tcp_sk(sk
);
2615 struct inet_sock
*inet
= inet_sk(sk
);
2617 if (sk
->sk_family
== AF_INET
) {
2618 printk(KERN_DEBUG
"Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2620 &inet
->daddr
, ntohs(inet
->dport
),
2621 tp
->snd_cwnd
, tcp_left_out(tp
),
2622 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2625 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2626 else if (sk
->sk_family
== AF_INET6
) {
2627 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2628 printk(KERN_DEBUG
"Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2630 &np
->daddr
, ntohs(inet
->dport
),
2631 tp
->snd_cwnd
, tcp_left_out(tp
),
2632 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2638 #define DBGUNDO(x...) do { } while (0)
2641 static void tcp_undo_cwr(struct sock
*sk
, const int undo
)
2643 struct tcp_sock
*tp
= tcp_sk(sk
);
2645 if (tp
->prior_ssthresh
) {
2646 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2648 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2649 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2651 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2653 if (undo
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2654 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2655 TCP_ECN_withdraw_cwr(tp
);
2658 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2660 tcp_moderate_cwnd(tp
);
2661 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2664 static inline int tcp_may_undo(struct tcp_sock
*tp
)
2666 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2669 /* People celebrate: "We love our President!" */
2670 static int tcp_try_undo_recovery(struct sock
*sk
)
2672 struct tcp_sock
*tp
= tcp_sk(sk
);
2674 if (tcp_may_undo(tp
)) {
2677 /* Happy end! We did not retransmit anything
2678 * or our original transmission succeeded.
2680 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2681 tcp_undo_cwr(sk
, 1);
2682 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2683 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2685 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2687 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2688 tp
->undo_marker
= 0;
2690 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2691 /* Hold old state until something *above* high_seq
2692 * is ACKed. For Reno it is MUST to prevent false
2693 * fast retransmits (RFC2582). SACK TCP is safe. */
2694 tcp_moderate_cwnd(tp
);
2697 tcp_set_ca_state(sk
, TCP_CA_Open
);
2701 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2702 static void tcp_try_undo_dsack(struct sock
*sk
)
2704 struct tcp_sock
*tp
= tcp_sk(sk
);
2706 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2707 DBGUNDO(sk
, "D-SACK");
2708 tcp_undo_cwr(sk
, 1);
2709 tp
->undo_marker
= 0;
2710 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2714 /* Undo during fast recovery after partial ACK. */
2716 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2718 struct tcp_sock
*tp
= tcp_sk(sk
);
2719 /* Partial ACK arrived. Force Hoe's retransmit. */
2720 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2722 if (tcp_may_undo(tp
)) {
2723 /* Plain luck! Hole if filled with delayed
2724 * packet, rather than with a retransmit.
2726 if (tp
->retrans_out
== 0)
2727 tp
->retrans_stamp
= 0;
2729 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2732 tcp_undo_cwr(sk
, 0);
2733 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2735 /* So... Do not make Hoe's retransmit yet.
2736 * If the first packet was delayed, the rest
2737 * ones are most probably delayed as well.
2744 /* Undo during loss recovery after partial ACK. */
2745 static int tcp_try_undo_loss(struct sock
*sk
)
2747 struct tcp_sock
*tp
= tcp_sk(sk
);
2749 if (tcp_may_undo(tp
)) {
2750 struct sk_buff
*skb
;
2751 tcp_for_write_queue(skb
, sk
) {
2752 if (skb
== tcp_send_head(sk
))
2754 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2757 tcp_clear_all_retrans_hints(tp
);
2759 DBGUNDO(sk
, "partial loss");
2761 tcp_undo_cwr(sk
, 1);
2762 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2763 inet_csk(sk
)->icsk_retransmits
= 0;
2764 tp
->undo_marker
= 0;
2765 if (tcp_is_sack(tp
))
2766 tcp_set_ca_state(sk
, TCP_CA_Open
);
2772 static inline void tcp_complete_cwr(struct sock
*sk
)
2774 struct tcp_sock
*tp
= tcp_sk(sk
);
2775 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2776 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2777 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2780 static void tcp_try_keep_open(struct sock
*sk
)
2782 struct tcp_sock
*tp
= tcp_sk(sk
);
2783 int state
= TCP_CA_Open
;
2785 if (tcp_left_out(tp
) || tp
->retrans_out
|| tp
->undo_marker
)
2786 state
= TCP_CA_Disorder
;
2788 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2789 tcp_set_ca_state(sk
, state
);
2790 tp
->high_seq
= tp
->snd_nxt
;
2794 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2796 struct tcp_sock
*tp
= tcp_sk(sk
);
2798 tcp_verify_left_out(tp
);
2800 if (!tp
->frto_counter
&& tp
->retrans_out
== 0)
2801 tp
->retrans_stamp
= 0;
2803 if (flag
& FLAG_ECE
)
2804 tcp_enter_cwr(sk
, 1);
2806 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2807 tcp_try_keep_open(sk
);
2808 tcp_moderate_cwnd(tp
);
2810 tcp_cwnd_down(sk
, flag
);
2814 static void tcp_mtup_probe_failed(struct sock
*sk
)
2816 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2818 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2819 icsk
->icsk_mtup
.probe_size
= 0;
2822 static void tcp_mtup_probe_success(struct sock
*sk
, struct sk_buff
*skb
)
2824 struct tcp_sock
*tp
= tcp_sk(sk
);
2825 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2827 /* FIXME: breaks with very large cwnd */
2828 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2829 tp
->snd_cwnd
= tp
->snd_cwnd
*
2830 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2831 icsk
->icsk_mtup
.probe_size
;
2832 tp
->snd_cwnd_cnt
= 0;
2833 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2834 tp
->rcv_ssthresh
= tcp_current_ssthresh(sk
);
2836 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2837 icsk
->icsk_mtup
.probe_size
= 0;
2838 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2841 /* Do a simple retransmit without using the backoff mechanisms in
2842 * tcp_timer. This is used for path mtu discovery.
2843 * The socket is already locked here.
2845 void tcp_simple_retransmit(struct sock
*sk
)
2847 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2848 struct tcp_sock
*tp
= tcp_sk(sk
);
2849 struct sk_buff
*skb
;
2850 unsigned int mss
= tcp_current_mss(sk
, 0);
2851 u32 prior_lost
= tp
->lost_out
;
2853 tcp_for_write_queue(skb
, sk
) {
2854 if (skb
== tcp_send_head(sk
))
2856 if (skb
->len
> mss
&&
2857 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2858 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2859 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2860 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2862 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2866 tcp_clear_retrans_hints_partial(tp
);
2868 if (prior_lost
== tp
->lost_out
)
2871 if (tcp_is_reno(tp
))
2872 tcp_limit_reno_sacked(tp
);
2874 tcp_verify_left_out(tp
);
2876 /* Don't muck with the congestion window here.
2877 * Reason is that we do not increase amount of _data_
2878 * in network, but units changed and effective
2879 * cwnd/ssthresh really reduced now.
2881 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2882 tp
->high_seq
= tp
->snd_nxt
;
2883 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2884 tp
->prior_ssthresh
= 0;
2885 tp
->undo_marker
= 0;
2886 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2888 tcp_xmit_retransmit_queue(sk
);
2891 /* Process an event, which can update packets-in-flight not trivially.
2892 * Main goal of this function is to calculate new estimate for left_out,
2893 * taking into account both packets sitting in receiver's buffer and
2894 * packets lost by network.
2896 * Besides that it does CWND reduction, when packet loss is detected
2897 * and changes state of machine.
2899 * It does _not_ decide what to send, it is made in function
2900 * tcp_xmit_retransmit_queue().
2902 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
, int flag
)
2904 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2905 struct tcp_sock
*tp
= tcp_sk(sk
);
2906 int is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
2907 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2908 (tcp_fackets_out(tp
) > tp
->reordering
));
2909 int fast_rexmit
= 0, mib_idx
;
2911 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2913 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2914 tp
->fackets_out
= 0;
2916 /* Now state machine starts.
2917 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2918 if (flag
& FLAG_ECE
)
2919 tp
->prior_ssthresh
= 0;
2921 /* B. In all the states check for reneging SACKs. */
2922 if (tcp_check_sack_reneging(sk
, flag
))
2925 /* C. Process data loss notification, provided it is valid. */
2926 if (tcp_is_fack(tp
) && (flag
& FLAG_DATA_LOST
) &&
2927 before(tp
->snd_una
, tp
->high_seq
) &&
2928 icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2929 tp
->fackets_out
> tp
->reordering
) {
2930 tcp_mark_head_lost(sk
, tp
->fackets_out
- tp
->reordering
);
2931 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSS
);
2934 /* D. Check consistency of the current state. */
2935 tcp_verify_left_out(tp
);
2937 /* E. Check state exit conditions. State can be terminated
2938 * when high_seq is ACKed. */
2939 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2940 WARN_ON(tp
->retrans_out
!= 0);
2941 tp
->retrans_stamp
= 0;
2942 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2943 switch (icsk
->icsk_ca_state
) {
2945 icsk
->icsk_retransmits
= 0;
2946 if (tcp_try_undo_recovery(sk
))
2951 /* CWR is to be held something *above* high_seq
2952 * is ACKed for CWR bit to reach receiver. */
2953 if (tp
->snd_una
!= tp
->high_seq
) {
2954 tcp_complete_cwr(sk
);
2955 tcp_set_ca_state(sk
, TCP_CA_Open
);
2959 case TCP_CA_Disorder
:
2960 tcp_try_undo_dsack(sk
);
2961 if (!tp
->undo_marker
||
2962 /* For SACK case do not Open to allow to undo
2963 * catching for all duplicate ACKs. */
2964 tcp_is_reno(tp
) || tp
->snd_una
!= tp
->high_seq
) {
2965 tp
->undo_marker
= 0;
2966 tcp_set_ca_state(sk
, TCP_CA_Open
);
2970 case TCP_CA_Recovery
:
2971 if (tcp_is_reno(tp
))
2972 tcp_reset_reno_sack(tp
);
2973 if (tcp_try_undo_recovery(sk
))
2975 tcp_complete_cwr(sk
);
2980 /* F. Process state. */
2981 switch (icsk
->icsk_ca_state
) {
2982 case TCP_CA_Recovery
:
2983 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2984 if (tcp_is_reno(tp
) && is_dupack
)
2985 tcp_add_reno_sack(sk
);
2987 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2990 if (flag
& FLAG_DATA_ACKED
)
2991 icsk
->icsk_retransmits
= 0;
2992 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
2993 tcp_reset_reno_sack(tp
);
2994 if (!tcp_try_undo_loss(sk
)) {
2995 tcp_moderate_cwnd(tp
);
2996 tcp_xmit_retransmit_queue(sk
);
2999 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
3001 /* Loss is undone; fall through to processing in Open state. */
3003 if (tcp_is_reno(tp
)) {
3004 if (flag
& FLAG_SND_UNA_ADVANCED
)
3005 tcp_reset_reno_sack(tp
);
3007 tcp_add_reno_sack(sk
);
3010 if (icsk
->icsk_ca_state
== TCP_CA_Disorder
)
3011 tcp_try_undo_dsack(sk
);
3013 if (!tcp_time_to_recover(sk
)) {
3014 tcp_try_to_open(sk
, flag
);
3018 /* MTU probe failure: don't reduce cwnd */
3019 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3020 icsk
->icsk_mtup
.probe_size
&&
3021 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3022 tcp_mtup_probe_failed(sk
);
3023 /* Restores the reduction we did in tcp_mtup_probe() */
3025 tcp_simple_retransmit(sk
);
3029 /* Otherwise enter Recovery state */
3031 if (tcp_is_reno(tp
))
3032 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
3034 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
3036 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3038 tp
->high_seq
= tp
->snd_nxt
;
3039 tp
->prior_ssthresh
= 0;
3040 tp
->undo_marker
= tp
->snd_una
;
3041 tp
->undo_retrans
= tp
->retrans_out
;
3043 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
3044 if (!(flag
& FLAG_ECE
))
3045 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
3046 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
3047 TCP_ECN_queue_cwr(tp
);
3050 tp
->bytes_acked
= 0;
3051 tp
->snd_cwnd_cnt
= 0;
3052 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
3056 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3057 tcp_update_scoreboard(sk
, fast_rexmit
);
3058 tcp_cwnd_down(sk
, flag
);
3059 tcp_xmit_retransmit_queue(sk
);
3062 /* Read draft-ietf-tcplw-high-performance before mucking
3063 * with this code. (Supersedes RFC1323)
3065 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3067 /* RTTM Rule: A TSecr value received in a segment is used to
3068 * update the averaged RTT measurement only if the segment
3069 * acknowledges some new data, i.e., only if it advances the
3070 * left edge of the send window.
3072 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3073 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3075 * Changed: reset backoff as soon as we see the first valid sample.
3076 * If we do not, we get strongly overestimated rto. With timestamps
3077 * samples are accepted even from very old segments: f.e., when rtt=1
3078 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3079 * answer arrives rto becomes 120 seconds! If at least one of segments
3080 * in window is lost... Voila. --ANK (010210)
3082 struct tcp_sock
*tp
= tcp_sk(sk
);
3083 const __u32 seq_rtt
= tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
;
3084 tcp_rtt_estimator(sk
, seq_rtt
);
3086 inet_csk(sk
)->icsk_backoff
= 0;
3090 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3092 /* We don't have a timestamp. Can only use
3093 * packets that are not retransmitted to determine
3094 * rtt estimates. Also, we must not reset the
3095 * backoff for rto until we get a non-retransmitted
3096 * packet. This allows us to deal with a situation
3097 * where the network delay has increased suddenly.
3098 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3101 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3104 tcp_rtt_estimator(sk
, seq_rtt
);
3106 inet_csk(sk
)->icsk_backoff
= 0;
3110 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3113 const struct tcp_sock
*tp
= tcp_sk(sk
);
3114 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3115 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3116 tcp_ack_saw_tstamp(sk
, flag
);
3117 else if (seq_rtt
>= 0)
3118 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3121 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3123 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3124 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3125 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3128 /* Restart timer after forward progress on connection.
3129 * RFC2988 recommends to restart timer to now+rto.
3131 static void tcp_rearm_rto(struct sock
*sk
)
3133 struct tcp_sock
*tp
= tcp_sk(sk
);
3135 if (!tp
->packets_out
) {
3136 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3138 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3139 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3143 /* If we get here, the whole TSO packet has not been acked. */
3144 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3146 struct tcp_sock
*tp
= tcp_sk(sk
);
3149 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3151 packets_acked
= tcp_skb_pcount(skb
);
3152 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3154 packets_acked
-= tcp_skb_pcount(skb
);
3156 if (packets_acked
) {
3157 BUG_ON(tcp_skb_pcount(skb
) == 0);
3158 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3161 return packets_acked
;
3164 /* Remove acknowledged frames from the retransmission queue. If our packet
3165 * is before the ack sequence we can discard it as it's confirmed to have
3166 * arrived at the other end.
3168 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3171 struct tcp_sock
*tp
= tcp_sk(sk
);
3172 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3173 struct sk_buff
*skb
;
3174 u32 now
= tcp_time_stamp
;
3175 int fully_acked
= 1;
3178 u32 reord
= tp
->packets_out
;
3179 u32 prior_sacked
= tp
->sacked_out
;
3181 s32 ca_seq_rtt
= -1;
3182 ktime_t last_ackt
= net_invalid_timestamp();
3184 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3185 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3188 u8 sacked
= scb
->sacked
;
3190 /* Determine how many packets and what bytes were acked, tso and else */
3191 if (after(scb
->end_seq
, tp
->snd_una
)) {
3192 if (tcp_skb_pcount(skb
) == 1 ||
3193 !after(tp
->snd_una
, scb
->seq
))
3196 acked_pcount
= tcp_tso_acked(sk
, skb
);
3201 end_seq
= tp
->snd_una
;
3203 acked_pcount
= tcp_skb_pcount(skb
);
3204 end_seq
= scb
->end_seq
;
3207 /* MTU probing checks */
3208 if (fully_acked
&& icsk
->icsk_mtup
.probe_size
&&
3209 !after(tp
->mtu_probe
.probe_seq_end
, scb
->end_seq
)) {
3210 tcp_mtup_probe_success(sk
, skb
);
3213 if (sacked
& TCPCB_RETRANS
) {
3214 if (sacked
& TCPCB_SACKED_RETRANS
)
3215 tp
->retrans_out
-= acked_pcount
;
3216 flag
|= FLAG_RETRANS_DATA_ACKED
;
3219 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3220 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3222 ca_seq_rtt
= now
- scb
->when
;
3223 last_ackt
= skb
->tstamp
;
3225 seq_rtt
= ca_seq_rtt
;
3227 if (!(sacked
& TCPCB_SACKED_ACKED
))
3228 reord
= min(pkts_acked
, reord
);
3231 if (sacked
& TCPCB_SACKED_ACKED
)
3232 tp
->sacked_out
-= acked_pcount
;
3233 if (sacked
& TCPCB_LOST
)
3234 tp
->lost_out
-= acked_pcount
;
3236 tp
->packets_out
-= acked_pcount
;
3237 pkts_acked
+= acked_pcount
;
3239 /* Initial outgoing SYN's get put onto the write_queue
3240 * just like anything else we transmit. It is not
3241 * true data, and if we misinform our callers that
3242 * this ACK acks real data, we will erroneously exit
3243 * connection startup slow start one packet too
3244 * quickly. This is severely frowned upon behavior.
3246 if (!(scb
->flags
& TCPCB_FLAG_SYN
)) {
3247 flag
|= FLAG_DATA_ACKED
;
3249 flag
|= FLAG_SYN_ACKED
;
3250 tp
->retrans_stamp
= 0;
3256 tcp_unlink_write_queue(skb
, sk
);
3257 sk_wmem_free_skb(sk
, skb
);
3258 tp
->scoreboard_skb_hint
= NULL
;
3259 if (skb
== tp
->retransmit_skb_hint
)
3260 tp
->retransmit_skb_hint
= NULL
;
3261 if (skb
== tp
->lost_skb_hint
)
3262 tp
->lost_skb_hint
= NULL
;
3265 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3266 tp
->snd_up
= tp
->snd_una
;
3268 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3269 flag
|= FLAG_SACK_RENEGING
;
3271 if (flag
& FLAG_ACKED
) {
3272 const struct tcp_congestion_ops
*ca_ops
3273 = inet_csk(sk
)->icsk_ca_ops
;
3275 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3278 if (tcp_is_reno(tp
)) {
3279 tcp_remove_reno_sacks(sk
, pkts_acked
);
3281 /* Non-retransmitted hole got filled? That's reordering */
3282 if (reord
< prior_fackets
)
3283 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3285 /* No need to care for underflows here because
3286 * the lost_skb_hint gets NULLed if we're past it
3287 * (or something non-trivial happened)
3289 if (tcp_is_fack(tp
))
3290 tp
->lost_cnt_hint
-= pkts_acked
;
3292 tp
->lost_cnt_hint
-= prior_sacked
- tp
->sacked_out
;
3295 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3297 if (ca_ops
->pkts_acked
) {
3300 /* Is the ACK triggering packet unambiguous? */
3301 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3302 /* High resolution needed and available? */
3303 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3304 !ktime_equal(last_ackt
,
3305 net_invalid_timestamp()))
3306 rtt_us
= ktime_us_delta(ktime_get_real(),
3308 else if (ca_seq_rtt
> 0)
3309 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3312 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3316 #if FASTRETRANS_DEBUG > 0
3317 WARN_ON((int)tp
->sacked_out
< 0);
3318 WARN_ON((int)tp
->lost_out
< 0);
3319 WARN_ON((int)tp
->retrans_out
< 0);
3320 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3321 icsk
= inet_csk(sk
);
3323 printk(KERN_DEBUG
"Leak l=%u %d\n",
3324 tp
->lost_out
, icsk
->icsk_ca_state
);
3327 if (tp
->sacked_out
) {
3328 printk(KERN_DEBUG
"Leak s=%u %d\n",
3329 tp
->sacked_out
, icsk
->icsk_ca_state
);
3332 if (tp
->retrans_out
) {
3333 printk(KERN_DEBUG
"Leak r=%u %d\n",
3334 tp
->retrans_out
, icsk
->icsk_ca_state
);
3335 tp
->retrans_out
= 0;
3342 static void tcp_ack_probe(struct sock
*sk
)
3344 const struct tcp_sock
*tp
= tcp_sk(sk
);
3345 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3347 /* Was it a usable window open? */
3349 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3350 icsk
->icsk_backoff
= 0;
3351 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3352 /* Socket must be waked up by subsequent tcp_data_snd_check().
3353 * This function is not for random using!
3356 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3357 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3362 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3364 return (!(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3365 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
);
3368 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3370 const struct tcp_sock
*tp
= tcp_sk(sk
);
3371 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3372 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3375 /* Check that window update is acceptable.
3376 * The function assumes that snd_una<=ack<=snd_next.
3378 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3379 const u32 ack
, const u32 ack_seq
,
3382 return (after(ack
, tp
->snd_una
) ||
3383 after(ack_seq
, tp
->snd_wl1
) ||
3384 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
));
3387 /* Update our send window.
3389 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3390 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3392 static int tcp_ack_update_window(struct sock
*sk
, struct sk_buff
*skb
, u32 ack
,
3395 struct tcp_sock
*tp
= tcp_sk(sk
);
3397 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3399 if (likely(!tcp_hdr(skb
)->syn
))
3400 nwin
<<= tp
->rx_opt
.snd_wscale
;
3402 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3403 flag
|= FLAG_WIN_UPDATE
;
3404 tcp_update_wl(tp
, ack
, ack_seq
);
3406 if (tp
->snd_wnd
!= nwin
) {
3409 /* Note, it is the only place, where
3410 * fast path is recovered for sending TCP.
3413 tcp_fast_path_check(sk
);
3415 if (nwin
> tp
->max_window
) {
3416 tp
->max_window
= nwin
;
3417 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3427 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3428 * continue in congestion avoidance.
3430 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3432 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3433 tp
->snd_cwnd_cnt
= 0;
3434 tp
->bytes_acked
= 0;
3435 TCP_ECN_queue_cwr(tp
);
3436 tcp_moderate_cwnd(tp
);
3439 /* A conservative spurious RTO response algorithm: reduce cwnd using
3440 * rate halving and continue in congestion avoidance.
3442 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3444 tcp_enter_cwr(sk
, 0);
3447 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3449 if (flag
& FLAG_ECE
)
3450 tcp_ratehalving_spur_to_response(sk
);
3452 tcp_undo_cwr(sk
, 1);
3455 /* F-RTO spurious RTO detection algorithm (RFC4138)
3457 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3458 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3459 * window (but not to or beyond highest sequence sent before RTO):
3460 * On First ACK, send two new segments out.
3461 * On Second ACK, RTO was likely spurious. Do spurious response (response
3462 * algorithm is not part of the F-RTO detection algorithm
3463 * given in RFC4138 but can be selected separately).
3464 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3465 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3466 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3467 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3469 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3470 * original window even after we transmit two new data segments.
3473 * on first step, wait until first cumulative ACK arrives, then move to
3474 * the second step. In second step, the next ACK decides.
3476 * F-RTO is implemented (mainly) in four functions:
3477 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3478 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3479 * called when tcp_use_frto() showed green light
3480 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3481 * - tcp_enter_frto_loss() is called if there is not enough evidence
3482 * to prove that the RTO is indeed spurious. It transfers the control
3483 * from F-RTO to the conventional RTO recovery
3485 static int tcp_process_frto(struct sock
*sk
, int flag
)
3487 struct tcp_sock
*tp
= tcp_sk(sk
);
3489 tcp_verify_left_out(tp
);
3491 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3492 if (flag
& FLAG_DATA_ACKED
)
3493 inet_csk(sk
)->icsk_retransmits
= 0;
3495 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3496 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3497 tp
->undo_marker
= 0;
3499 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3500 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3504 if (!tcp_is_sackfrto(tp
)) {
3505 /* RFC4138 shortcoming in step 2; should also have case c):
3506 * ACK isn't duplicate nor advances window, e.g., opposite dir
3509 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3512 if (!(flag
& FLAG_DATA_ACKED
)) {
3513 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3518 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3519 /* Prevent sending of new data. */
3520 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3521 tcp_packets_in_flight(tp
));
3525 if ((tp
->frto_counter
>= 2) &&
3526 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3527 ((flag
& FLAG_DATA_SACKED
) &&
3528 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3529 /* RFC4138 shortcoming (see comment above) */
3530 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3531 (flag
& FLAG_NOT_DUP
))
3534 tcp_enter_frto_loss(sk
, 3, flag
);
3539 if (tp
->frto_counter
== 1) {
3540 /* tcp_may_send_now needs to see updated state */
3541 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3542 tp
->frto_counter
= 2;
3544 if (!tcp_may_send_now(sk
))
3545 tcp_enter_frto_loss(sk
, 2, flag
);
3549 switch (sysctl_tcp_frto_response
) {
3551 tcp_undo_spur_to_response(sk
, flag
);
3554 tcp_conservative_spur_to_response(tp
);
3557 tcp_ratehalving_spur_to_response(sk
);
3560 tp
->frto_counter
= 0;
3561 tp
->undo_marker
= 0;
3562 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3567 /* This routine deals with incoming acks, but not outgoing ones. */
3568 static int tcp_ack(struct sock
*sk
, struct sk_buff
*skb
, int flag
)
3570 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3571 struct tcp_sock
*tp
= tcp_sk(sk
);
3572 u32 prior_snd_una
= tp
->snd_una
;
3573 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3574 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3575 u32 prior_in_flight
;
3580 /* If the ack is newer than sent or older than previous acks
3581 * then we can probably ignore it.
3583 if (after(ack
, tp
->snd_nxt
))
3584 goto uninteresting_ack
;
3586 if (before(ack
, prior_snd_una
))
3589 if (after(ack
, prior_snd_una
))
3590 flag
|= FLAG_SND_UNA_ADVANCED
;
3592 if (sysctl_tcp_abc
) {
3593 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3594 tp
->bytes_acked
+= ack
- prior_snd_una
;
3595 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3596 /* we assume just one segment left network */
3597 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3601 prior_fackets
= tp
->fackets_out
;
3602 prior_in_flight
= tcp_packets_in_flight(tp
);
3604 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3605 /* Window is constant, pure forward advance.
3606 * No more checks are required.
3607 * Note, we use the fact that SND.UNA>=SND.WL2.
3609 tcp_update_wl(tp
, ack
, ack_seq
);
3611 flag
|= FLAG_WIN_UPDATE
;
3613 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3615 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3617 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3620 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3622 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3624 if (TCP_SKB_CB(skb
)->sacked
)
3625 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3627 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3630 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3633 /* We passed data and got it acked, remove any soft error
3634 * log. Something worked...
3636 sk
->sk_err_soft
= 0;
3637 icsk
->icsk_probes_out
= 0;
3638 tp
->rcv_tstamp
= tcp_time_stamp
;
3639 prior_packets
= tp
->packets_out
;
3643 /* See if we can take anything off of the retransmit queue. */
3644 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3646 if (tp
->frto_counter
)
3647 frto_cwnd
= tcp_process_frto(sk
, flag
);
3648 /* Guarantee sacktag reordering detection against wrap-arounds */
3649 if (before(tp
->frto_highmark
, tp
->snd_una
))
3650 tp
->frto_highmark
= 0;
3652 if (tcp_ack_is_dubious(sk
, flag
)) {
3653 /* Advance CWND, if state allows this. */
3654 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3655 tcp_may_raise_cwnd(sk
, flag
))
3656 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3657 tcp_fastretrans_alert(sk
, prior_packets
- tp
->packets_out
,
3660 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3661 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3664 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3665 dst_confirm(sk
->sk_dst_cache
);
3670 /* If this ack opens up a zero window, clear backoff. It was
3671 * being used to time the probes, and is probably far higher than
3672 * it needs to be for normal retransmission.
3674 if (tcp_send_head(sk
))
3679 if (TCP_SKB_CB(skb
)->sacked
) {
3680 tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3681 if (icsk
->icsk_ca_state
== TCP_CA_Open
)
3682 tcp_try_keep_open(sk
);
3686 SOCK_DEBUG(sk
, "Ack %u out of %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3690 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3691 * But, this can also be called on packets in the established flow when
3692 * the fast version below fails.
3694 void tcp_parse_options(struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3698 struct tcphdr
*th
= tcp_hdr(skb
);
3699 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3701 ptr
= (unsigned char *)(th
+ 1);
3702 opt_rx
->saw_tstamp
= 0;
3704 while (length
> 0) {
3705 int opcode
= *ptr
++;
3711 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3716 if (opsize
< 2) /* "silly options" */
3718 if (opsize
> length
)
3719 return; /* don't parse partial options */
3722 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3723 u16 in_mss
= get_unaligned_be16(ptr
);
3725 if (opt_rx
->user_mss
&&
3726 opt_rx
->user_mss
< in_mss
)
3727 in_mss
= opt_rx
->user_mss
;
3728 opt_rx
->mss_clamp
= in_mss
;
3733 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3734 !estab
&& sysctl_tcp_window_scaling
) {
3735 __u8 snd_wscale
= *(__u8
*)ptr
;
3736 opt_rx
->wscale_ok
= 1;
3737 if (snd_wscale
> 14) {
3738 if (net_ratelimit())
3739 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3740 "scaling value %d >14 received.\n",
3744 opt_rx
->snd_wscale
= snd_wscale
;
3747 case TCPOPT_TIMESTAMP
:
3748 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3749 ((estab
&& opt_rx
->tstamp_ok
) ||
3750 (!estab
&& sysctl_tcp_timestamps
))) {
3751 opt_rx
->saw_tstamp
= 1;
3752 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3753 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3756 case TCPOPT_SACK_PERM
:
3757 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3758 !estab
&& sysctl_tcp_sack
) {
3759 opt_rx
->sack_ok
= 1;
3760 tcp_sack_reset(opt_rx
);
3765 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3766 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3768 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3771 #ifdef CONFIG_TCP_MD5SIG
3774 * The MD5 Hash has already been
3775 * checked (see tcp_v{4,6}_do_rcv()).
3787 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, struct tcphdr
*th
)
3789 __be32
*ptr
= (__be32
*)(th
+ 1);
3791 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3792 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3793 tp
->rx_opt
.saw_tstamp
= 1;
3795 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3797 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3803 /* Fast parse options. This hopes to only see timestamps.
3804 * If it is wrong it falls back on tcp_parse_options().
3806 static int tcp_fast_parse_options(struct sk_buff
*skb
, struct tcphdr
*th
,
3807 struct tcp_sock
*tp
)
3809 if (th
->doff
== sizeof(struct tcphdr
) >> 2) {
3810 tp
->rx_opt
.saw_tstamp
= 0;
3812 } else if (tp
->rx_opt
.tstamp_ok
&&
3813 th
->doff
== (sizeof(struct tcphdr
)>>2)+(TCPOLEN_TSTAMP_ALIGNED
>>2)) {
3814 if (tcp_parse_aligned_timestamp(tp
, th
))
3817 tcp_parse_options(skb
, &tp
->rx_opt
, 1);
3821 #ifdef CONFIG_TCP_MD5SIG
3823 * Parse MD5 Signature option
3825 u8
*tcp_parse_md5sig_option(struct tcphdr
*th
)
3827 int length
= (th
->doff
<< 2) - sizeof (*th
);
3828 u8
*ptr
= (u8
*)(th
+ 1);
3830 /* If the TCP option is too short, we can short cut */
3831 if (length
< TCPOLEN_MD5SIG
)
3834 while (length
> 0) {
3835 int opcode
= *ptr
++;
3846 if (opsize
< 2 || opsize
> length
)
3848 if (opcode
== TCPOPT_MD5SIG
)
3858 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
3860 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3861 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3864 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3866 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3867 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3868 * extra check below makes sure this can only happen
3869 * for pure ACK frames. -DaveM
3871 * Not only, also it occurs for expired timestamps.
3874 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) >= 0 ||
3875 get_seconds() >= tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
)
3876 tcp_store_ts_recent(tp
);
3880 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3882 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3883 * it can pass through stack. So, the following predicate verifies that
3884 * this segment is not used for anything but congestion avoidance or
3885 * fast retransmit. Moreover, we even are able to eliminate most of such
3886 * second order effects, if we apply some small "replay" window (~RTO)
3887 * to timestamp space.
3889 * All these measures still do not guarantee that we reject wrapped ACKs
3890 * on networks with high bandwidth, when sequence space is recycled fastly,
3891 * but it guarantees that such events will be very rare and do not affect
3892 * connection seriously. This doesn't look nice, but alas, PAWS is really
3895 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3896 * states that events when retransmit arrives after original data are rare.
3897 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3898 * the biggest problem on large power networks even with minor reordering.
3899 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3900 * up to bandwidth of 18Gigabit/sec. 8) ]
3903 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3905 struct tcp_sock
*tp
= tcp_sk(sk
);
3906 struct tcphdr
*th
= tcp_hdr(skb
);
3907 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3908 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3910 return (/* 1. Pure ACK with correct sequence number. */
3911 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3913 /* 2. ... and duplicate ACK. */
3914 ack
== tp
->snd_una
&&
3916 /* 3. ... and does not update window. */
3917 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3919 /* 4. ... and sits in replay window. */
3920 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3923 static inline int tcp_paws_discard(const struct sock
*sk
,
3924 const struct sk_buff
*skb
)
3926 const struct tcp_sock
*tp
= tcp_sk(sk
);
3927 return ((s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) > TCP_PAWS_WINDOW
&&
3928 get_seconds() < tp
->rx_opt
.ts_recent_stamp
+ TCP_PAWS_24DAYS
&&
3929 !tcp_disordered_ack(sk
, skb
));
3932 /* Check segment sequence number for validity.
3934 * Segment controls are considered valid, if the segment
3935 * fits to the window after truncation to the window. Acceptability
3936 * of data (and SYN, FIN, of course) is checked separately.
3937 * See tcp_data_queue(), for example.
3939 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3940 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3941 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3942 * (borrowed from freebsd)
3945 static inline int tcp_sequence(struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3947 return !before(end_seq
, tp
->rcv_wup
) &&
3948 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3951 /* When we get a reset we do this. */
3952 static void tcp_reset(struct sock
*sk
)
3954 /* We want the right error as BSD sees it (and indeed as we do). */
3955 switch (sk
->sk_state
) {
3957 sk
->sk_err
= ECONNREFUSED
;
3959 case TCP_CLOSE_WAIT
:
3965 sk
->sk_err
= ECONNRESET
;
3968 if (!sock_flag(sk
, SOCK_DEAD
))
3969 sk
->sk_error_report(sk
);
3975 * Process the FIN bit. This now behaves as it is supposed to work
3976 * and the FIN takes effect when it is validly part of sequence
3977 * space. Not before when we get holes.
3979 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3980 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3983 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3984 * close and we go into CLOSING (and later onto TIME-WAIT)
3986 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3988 static void tcp_fin(struct sk_buff
*skb
, struct sock
*sk
, struct tcphdr
*th
)
3990 struct tcp_sock
*tp
= tcp_sk(sk
);
3992 inet_csk_schedule_ack(sk
);
3994 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3995 sock_set_flag(sk
, SOCK_DONE
);
3997 switch (sk
->sk_state
) {
3999 case TCP_ESTABLISHED
:
4000 /* Move to CLOSE_WAIT */
4001 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4002 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4005 case TCP_CLOSE_WAIT
:
4007 /* Received a retransmission of the FIN, do
4012 /* RFC793: Remain in the LAST-ACK state. */
4016 /* This case occurs when a simultaneous close
4017 * happens, we must ack the received FIN and
4018 * enter the CLOSING state.
4021 tcp_set_state(sk
, TCP_CLOSING
);
4024 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4026 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4029 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4030 * cases we should never reach this piece of code.
4032 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
4033 __func__
, sk
->sk_state
);
4037 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4038 * Probably, we should reset in this case. For now drop them.
4040 __skb_queue_purge(&tp
->out_of_order_queue
);
4041 if (tcp_is_sack(tp
))
4042 tcp_sack_reset(&tp
->rx_opt
);
4045 if (!sock_flag(sk
, SOCK_DEAD
)) {
4046 sk
->sk_state_change(sk
);
4048 /* Do not send POLL_HUP for half duplex close. */
4049 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4050 sk
->sk_state
== TCP_CLOSE
)
4051 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4053 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4057 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4060 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4061 if (before(seq
, sp
->start_seq
))
4062 sp
->start_seq
= seq
;
4063 if (after(end_seq
, sp
->end_seq
))
4064 sp
->end_seq
= end_seq
;
4070 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4072 struct tcp_sock
*tp
= tcp_sk(sk
);
4074 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4077 if (before(seq
, tp
->rcv_nxt
))
4078 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4080 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4082 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4084 tp
->rx_opt
.dsack
= 1;
4085 tp
->duplicate_sack
[0].start_seq
= seq
;
4086 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4087 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ 1;
4091 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4093 struct tcp_sock
*tp
= tcp_sk(sk
);
4095 if (!tp
->rx_opt
.dsack
)
4096 tcp_dsack_set(sk
, seq
, end_seq
);
4098 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4101 static void tcp_send_dupack(struct sock
*sk
, struct sk_buff
*skb
)
4103 struct tcp_sock
*tp
= tcp_sk(sk
);
4105 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4106 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4107 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4108 tcp_enter_quickack_mode(sk
);
4110 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4111 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4113 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4114 end_seq
= tp
->rcv_nxt
;
4115 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4122 /* These routines update the SACK block as out-of-order packets arrive or
4123 * in-order packets close up the sequence space.
4125 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4128 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4129 struct tcp_sack_block
*swalk
= sp
+ 1;
4131 /* See if the recent change to the first SACK eats into
4132 * or hits the sequence space of other SACK blocks, if so coalesce.
4134 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4135 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4138 /* Zap SWALK, by moving every further SACK up by one slot.
4139 * Decrease num_sacks.
4141 tp
->rx_opt
.num_sacks
--;
4142 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
4144 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4148 this_sack
++, swalk
++;
4152 static inline void tcp_sack_swap(struct tcp_sack_block
*sack1
,
4153 struct tcp_sack_block
*sack2
)
4157 tmp
= sack1
->start_seq
;
4158 sack1
->start_seq
= sack2
->start_seq
;
4159 sack2
->start_seq
= tmp
;
4161 tmp
= sack1
->end_seq
;
4162 sack1
->end_seq
= sack2
->end_seq
;
4163 sack2
->end_seq
= tmp
;
4166 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4168 struct tcp_sock
*tp
= tcp_sk(sk
);
4169 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4170 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4176 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4177 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4178 /* Rotate this_sack to the first one. */
4179 for (; this_sack
> 0; this_sack
--, sp
--)
4180 tcp_sack_swap(sp
, sp
- 1);
4182 tcp_sack_maybe_coalesce(tp
);
4187 /* Could not find an adjacent existing SACK, build a new one,
4188 * put it at the front, and shift everyone else down. We
4189 * always know there is at least one SACK present already here.
4191 * If the sack array is full, forget about the last one.
4193 if (this_sack
>= TCP_NUM_SACKS
) {
4195 tp
->rx_opt
.num_sacks
--;
4198 for (; this_sack
> 0; this_sack
--, sp
--)
4202 /* Build the new head SACK, and we're done. */
4203 sp
->start_seq
= seq
;
4204 sp
->end_seq
= end_seq
;
4205 tp
->rx_opt
.num_sacks
++;
4206 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
4209 /* RCV.NXT advances, some SACKs should be eaten. */
4211 static void tcp_sack_remove(struct tcp_sock
*tp
)
4213 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4214 int num_sacks
= tp
->rx_opt
.num_sacks
;
4217 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4218 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4219 tp
->rx_opt
.num_sacks
= 0;
4220 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.dsack
;
4224 for (this_sack
= 0; this_sack
< num_sacks
;) {
4225 /* Check if the start of the sack is covered by RCV.NXT. */
4226 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4229 /* RCV.NXT must cover all the block! */
4230 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4232 /* Zap this SACK, by moving forward any other SACKS. */
4233 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4234 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4241 if (num_sacks
!= tp
->rx_opt
.num_sacks
) {
4242 tp
->rx_opt
.num_sacks
= num_sacks
;
4243 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
+
4248 /* This one checks to see if we can put data from the
4249 * out_of_order queue into the receive_queue.
4251 static void tcp_ofo_queue(struct sock
*sk
)
4253 struct tcp_sock
*tp
= tcp_sk(sk
);
4254 __u32 dsack_high
= tp
->rcv_nxt
;
4255 struct sk_buff
*skb
;
4257 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4258 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4261 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4262 __u32 dsack
= dsack_high
;
4263 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4264 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4265 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4268 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4269 SOCK_DEBUG(sk
, "ofo packet was already received \n");
4270 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4274 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4275 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4276 TCP_SKB_CB(skb
)->end_seq
);
4278 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4279 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4280 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4281 if (tcp_hdr(skb
)->fin
)
4282 tcp_fin(skb
, sk
, tcp_hdr(skb
));
4286 static int tcp_prune_ofo_queue(struct sock
*sk
);
4287 static int tcp_prune_queue(struct sock
*sk
);
4289 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
4291 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4292 !sk_rmem_schedule(sk
, size
)) {
4294 if (tcp_prune_queue(sk
) < 0)
4297 if (!sk_rmem_schedule(sk
, size
)) {
4298 if (!tcp_prune_ofo_queue(sk
))
4301 if (!sk_rmem_schedule(sk
, size
))
4308 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4310 struct tcphdr
*th
= tcp_hdr(skb
);
4311 struct tcp_sock
*tp
= tcp_sk(sk
);
4314 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4317 __skb_pull(skb
, th
->doff
* 4);
4319 TCP_ECN_accept_cwr(tp
, skb
);
4321 if (tp
->rx_opt
.dsack
) {
4322 tp
->rx_opt
.dsack
= 0;
4323 tp
->rx_opt
.eff_sacks
= tp
->rx_opt
.num_sacks
;
4326 /* Queue data for delivery to the user.
4327 * Packets in sequence go to the receive queue.
4328 * Out of sequence packets to the out_of_order_queue.
4330 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4331 if (tcp_receive_window(tp
) == 0)
4334 /* Ok. In sequence. In window. */
4335 if (tp
->ucopy
.task
== current
&&
4336 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4337 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4338 int chunk
= min_t(unsigned int, skb
->len
,
4341 __set_current_state(TASK_RUNNING
);
4344 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4345 tp
->ucopy
.len
-= chunk
;
4346 tp
->copied_seq
+= chunk
;
4347 eaten
= (chunk
== skb
->len
&& !th
->fin
);
4348 tcp_rcv_space_adjust(sk
);
4356 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4359 skb_set_owner_r(skb
, sk
);
4360 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4362 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4364 tcp_event_data_recv(sk
, skb
);
4366 tcp_fin(skb
, sk
, th
);
4368 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4371 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4372 * gap in queue is filled.
4374 if (skb_queue_empty(&tp
->out_of_order_queue
))
4375 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4378 if (tp
->rx_opt
.num_sacks
)
4379 tcp_sack_remove(tp
);
4381 tcp_fast_path_check(sk
);
4385 else if (!sock_flag(sk
, SOCK_DEAD
))
4386 sk
->sk_data_ready(sk
, 0);
4390 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4391 /* A retransmit, 2nd most common case. Force an immediate ack. */
4392 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4393 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4396 tcp_enter_quickack_mode(sk
);
4397 inet_csk_schedule_ack(sk
);
4403 /* Out of window. F.e. zero window probe. */
4404 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4407 tcp_enter_quickack_mode(sk
);
4409 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4410 /* Partial packet, seq < rcv_next < end_seq */
4411 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4412 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4413 TCP_SKB_CB(skb
)->end_seq
);
4415 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4417 /* If window is closed, drop tail of packet. But after
4418 * remembering D-SACK for its head made in previous line.
4420 if (!tcp_receive_window(tp
))
4425 TCP_ECN_check_ce(tp
, skb
);
4427 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4430 /* Disable header prediction. */
4432 inet_csk_schedule_ack(sk
);
4434 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4435 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4437 skb_set_owner_r(skb
, sk
);
4439 if (!skb_peek(&tp
->out_of_order_queue
)) {
4440 /* Initial out of order segment, build 1 SACK. */
4441 if (tcp_is_sack(tp
)) {
4442 tp
->rx_opt
.num_sacks
= 1;
4443 tp
->rx_opt
.dsack
= 0;
4444 tp
->rx_opt
.eff_sacks
= 1;
4445 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4446 tp
->selective_acks
[0].end_seq
=
4447 TCP_SKB_CB(skb
)->end_seq
;
4449 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4451 struct sk_buff
*skb1
= tp
->out_of_order_queue
.prev
;
4452 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4453 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4455 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4456 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4458 if (!tp
->rx_opt
.num_sacks
||
4459 tp
->selective_acks
[0].end_seq
!= seq
)
4462 /* Common case: data arrive in order after hole. */
4463 tp
->selective_acks
[0].end_seq
= end_seq
;
4467 /* Find place to insert this segment. */
4469 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4471 } while ((skb1
= skb1
->prev
) !=
4472 (struct sk_buff
*)&tp
->out_of_order_queue
);
4474 /* Do skb overlap to previous one? */
4475 if (skb1
!= (struct sk_buff
*)&tp
->out_of_order_queue
&&
4476 before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4477 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4478 /* All the bits are present. Drop. */
4480 tcp_dsack_set(sk
, seq
, end_seq
);
4483 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4484 /* Partial overlap. */
4485 tcp_dsack_set(sk
, seq
,
4486 TCP_SKB_CB(skb1
)->end_seq
);
4491 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4493 /* And clean segments covered by new one as whole. */
4494 while ((skb1
= skb
->next
) !=
4495 (struct sk_buff
*)&tp
->out_of_order_queue
&&
4496 after(end_seq
, TCP_SKB_CB(skb1
)->seq
)) {
4497 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4498 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4502 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4503 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4504 TCP_SKB_CB(skb1
)->end_seq
);
4509 if (tcp_is_sack(tp
))
4510 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4514 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4515 struct sk_buff_head
*list
)
4517 struct sk_buff
*next
= skb
->next
;
4519 __skb_unlink(skb
, list
);
4521 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4526 /* Collapse contiguous sequence of skbs head..tail with
4527 * sequence numbers start..end.
4528 * Segments with FIN/SYN are not collapsed (only because this
4532 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4533 struct sk_buff
*head
, struct sk_buff
*tail
,
4536 struct sk_buff
*skb
;
4538 /* First, check that queue is collapsible and find
4539 * the point where collapsing can be useful. */
4540 for (skb
= head
; skb
!= tail
;) {
4541 /* No new bits? It is possible on ofo queue. */
4542 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4543 skb
= tcp_collapse_one(sk
, skb
, list
);
4547 /* The first skb to collapse is:
4549 * - bloated or contains data before "start" or
4550 * overlaps to the next one.
4552 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4553 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4554 before(TCP_SKB_CB(skb
)->seq
, start
) ||
4555 (skb
->next
!= tail
&&
4556 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
->next
)->seq
)))
4559 /* Decided to skip this, advance start seq. */
4560 start
= TCP_SKB_CB(skb
)->end_seq
;
4563 if (skb
== tail
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4566 while (before(start
, end
)) {
4567 struct sk_buff
*nskb
;
4568 unsigned int header
= skb_headroom(skb
);
4569 int copy
= SKB_MAX_ORDER(header
, 0);
4571 /* Too big header? This can happen with IPv6. */
4574 if (end
- start
< copy
)
4576 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4580 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4581 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4583 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4585 skb_reserve(nskb
, header
);
4586 memcpy(nskb
->head
, skb
->head
, header
);
4587 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4588 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4589 __skb_queue_before(list
, skb
, nskb
);
4590 skb_set_owner_r(nskb
, sk
);
4592 /* Copy data, releasing collapsed skbs. */
4594 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4595 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4599 size
= min(copy
, size
);
4600 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4602 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4606 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4607 skb
= tcp_collapse_one(sk
, skb
, list
);
4609 tcp_hdr(skb
)->syn
||
4617 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4618 * and tcp_collapse() them until all the queue is collapsed.
4620 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4622 struct tcp_sock
*tp
= tcp_sk(sk
);
4623 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4624 struct sk_buff
*head
;
4630 start
= TCP_SKB_CB(skb
)->seq
;
4631 end
= TCP_SKB_CB(skb
)->end_seq
;
4637 /* Segment is terminated when we see gap or when
4638 * we are at the end of all the queue. */
4639 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
||
4640 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4641 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4642 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4643 head
, skb
, start
, end
);
4645 if (skb
== (struct sk_buff
*)&tp
->out_of_order_queue
)
4647 /* Start new segment */
4648 start
= TCP_SKB_CB(skb
)->seq
;
4649 end
= TCP_SKB_CB(skb
)->end_seq
;
4651 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4652 start
= TCP_SKB_CB(skb
)->seq
;
4653 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4654 end
= TCP_SKB_CB(skb
)->end_seq
;
4660 * Purge the out-of-order queue.
4661 * Return true if queue was pruned.
4663 static int tcp_prune_ofo_queue(struct sock
*sk
)
4665 struct tcp_sock
*tp
= tcp_sk(sk
);
4668 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4669 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4670 __skb_queue_purge(&tp
->out_of_order_queue
);
4672 /* Reset SACK state. A conforming SACK implementation will
4673 * do the same at a timeout based retransmit. When a connection
4674 * is in a sad state like this, we care only about integrity
4675 * of the connection not performance.
4677 if (tp
->rx_opt
.sack_ok
)
4678 tcp_sack_reset(&tp
->rx_opt
);
4685 /* Reduce allocated memory if we can, trying to get
4686 * the socket within its memory limits again.
4688 * Return less than zero if we should start dropping frames
4689 * until the socket owning process reads some of the data
4690 * to stabilize the situation.
4692 static int tcp_prune_queue(struct sock
*sk
)
4694 struct tcp_sock
*tp
= tcp_sk(sk
);
4696 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4698 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4700 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4701 tcp_clamp_window(sk
);
4702 else if (tcp_memory_pressure
)
4703 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4705 tcp_collapse_ofo_queue(sk
);
4706 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4707 sk
->sk_receive_queue
.next
,
4708 (struct sk_buff
*)&sk
->sk_receive_queue
,
4709 tp
->copied_seq
, tp
->rcv_nxt
);
4712 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4715 /* Collapsing did not help, destructive actions follow.
4716 * This must not ever occur. */
4718 tcp_prune_ofo_queue(sk
);
4720 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4723 /* If we are really being abused, tell the caller to silently
4724 * drop receive data on the floor. It will get retransmitted
4725 * and hopefully then we'll have sufficient space.
4727 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4729 /* Massive buffer overcommit. */
4734 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4735 * As additional protections, we do not touch cwnd in retransmission phases,
4736 * and if application hit its sndbuf limit recently.
4738 void tcp_cwnd_application_limited(struct sock
*sk
)
4740 struct tcp_sock
*tp
= tcp_sk(sk
);
4742 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4743 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4744 /* Limited by application or receiver window. */
4745 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4746 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4747 if (win_used
< tp
->snd_cwnd
) {
4748 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4749 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4751 tp
->snd_cwnd_used
= 0;
4753 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4756 static int tcp_should_expand_sndbuf(struct sock
*sk
)
4758 struct tcp_sock
*tp
= tcp_sk(sk
);
4760 /* If the user specified a specific send buffer setting, do
4763 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4766 /* If we are under global TCP memory pressure, do not expand. */
4767 if (tcp_memory_pressure
)
4770 /* If we are under soft global TCP memory pressure, do not expand. */
4771 if (atomic_read(&tcp_memory_allocated
) >= sysctl_tcp_mem
[0])
4774 /* If we filled the congestion window, do not expand. */
4775 if (tp
->packets_out
>= tp
->snd_cwnd
)
4781 /* When incoming ACK allowed to free some skb from write_queue,
4782 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4783 * on the exit from tcp input handler.
4785 * PROBLEM: sndbuf expansion does not work well with largesend.
4787 static void tcp_new_space(struct sock
*sk
)
4789 struct tcp_sock
*tp
= tcp_sk(sk
);
4791 if (tcp_should_expand_sndbuf(sk
)) {
4792 int sndmem
= max_t(u32
, tp
->rx_opt
.mss_clamp
, tp
->mss_cache
) +
4793 MAX_TCP_HEADER
+ 16 + sizeof(struct sk_buff
);
4794 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4795 tp
->reordering
+ 1);
4796 sndmem
*= 2 * demanded
;
4797 if (sndmem
> sk
->sk_sndbuf
)
4798 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4799 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4802 sk
->sk_write_space(sk
);
4805 static void tcp_check_space(struct sock
*sk
)
4807 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4808 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4809 if (sk
->sk_socket
&&
4810 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4815 static inline void tcp_data_snd_check(struct sock
*sk
)
4817 tcp_push_pending_frames(sk
);
4818 tcp_check_space(sk
);
4822 * Check if sending an ack is needed.
4824 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4826 struct tcp_sock
*tp
= tcp_sk(sk
);
4828 /* More than one full frame received... */
4829 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
4830 /* ... and right edge of window advances far enough.
4831 * (tcp_recvmsg() will send ACK otherwise). Or...
4833 && __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4834 /* We ACK each frame or... */
4835 tcp_in_quickack_mode(sk
) ||
4836 /* We have out of order data. */
4837 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4838 /* Then ack it now */
4841 /* Else, send delayed ack. */
4842 tcp_send_delayed_ack(sk
);
4846 static inline void tcp_ack_snd_check(struct sock
*sk
)
4848 if (!inet_csk_ack_scheduled(sk
)) {
4849 /* We sent a data segment already. */
4852 __tcp_ack_snd_check(sk
, 1);
4856 * This routine is only called when we have urgent data
4857 * signaled. Its the 'slow' part of tcp_urg. It could be
4858 * moved inline now as tcp_urg is only called from one
4859 * place. We handle URGent data wrong. We have to - as
4860 * BSD still doesn't use the correction from RFC961.
4861 * For 1003.1g we should support a new option TCP_STDURG to permit
4862 * either form (or just set the sysctl tcp_stdurg).
4865 static void tcp_check_urg(struct sock
*sk
, struct tcphdr
*th
)
4867 struct tcp_sock
*tp
= tcp_sk(sk
);
4868 u32 ptr
= ntohs(th
->urg_ptr
);
4870 if (ptr
&& !sysctl_tcp_stdurg
)
4872 ptr
+= ntohl(th
->seq
);
4874 /* Ignore urgent data that we've already seen and read. */
4875 if (after(tp
->copied_seq
, ptr
))
4878 /* Do not replay urg ptr.
4880 * NOTE: interesting situation not covered by specs.
4881 * Misbehaving sender may send urg ptr, pointing to segment,
4882 * which we already have in ofo queue. We are not able to fetch
4883 * such data and will stay in TCP_URG_NOTYET until will be eaten
4884 * by recvmsg(). Seems, we are not obliged to handle such wicked
4885 * situations. But it is worth to think about possibility of some
4886 * DoSes using some hypothetical application level deadlock.
4888 if (before(ptr
, tp
->rcv_nxt
))
4891 /* Do we already have a newer (or duplicate) urgent pointer? */
4892 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4895 /* Tell the world about our new urgent pointer. */
4898 /* We may be adding urgent data when the last byte read was
4899 * urgent. To do this requires some care. We cannot just ignore
4900 * tp->copied_seq since we would read the last urgent byte again
4901 * as data, nor can we alter copied_seq until this data arrives
4902 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4904 * NOTE. Double Dutch. Rendering to plain English: author of comment
4905 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4906 * and expect that both A and B disappear from stream. This is _wrong_.
4907 * Though this happens in BSD with high probability, this is occasional.
4908 * Any application relying on this is buggy. Note also, that fix "works"
4909 * only in this artificial test. Insert some normal data between A and B and we will
4910 * decline of BSD again. Verdict: it is better to remove to trap
4913 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4914 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4915 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4917 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4918 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4923 tp
->urg_data
= TCP_URG_NOTYET
;
4926 /* Disable header prediction. */
4930 /* This is the 'fast' part of urgent handling. */
4931 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, struct tcphdr
*th
)
4933 struct tcp_sock
*tp
= tcp_sk(sk
);
4935 /* Check if we get a new urgent pointer - normally not. */
4937 tcp_check_urg(sk
, th
);
4939 /* Do we wait for any urgent data? - normally not... */
4940 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4941 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4944 /* Is the urgent pointer pointing into this packet? */
4945 if (ptr
< skb
->len
) {
4947 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4949 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4950 if (!sock_flag(sk
, SOCK_DEAD
))
4951 sk
->sk_data_ready(sk
, 0);
4956 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4958 struct tcp_sock
*tp
= tcp_sk(sk
);
4959 int chunk
= skb
->len
- hlen
;
4963 if (skb_csum_unnecessary(skb
))
4964 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4966 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4970 tp
->ucopy
.len
-= chunk
;
4971 tp
->copied_seq
+= chunk
;
4972 tcp_rcv_space_adjust(sk
);
4979 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4980 struct sk_buff
*skb
)
4984 if (sock_owned_by_user(sk
)) {
4986 result
= __tcp_checksum_complete(skb
);
4989 result
= __tcp_checksum_complete(skb
);
4994 static inline int tcp_checksum_complete_user(struct sock
*sk
,
4995 struct sk_buff
*skb
)
4997 return !skb_csum_unnecessary(skb
) &&
4998 __tcp_checksum_complete_user(sk
, skb
);
5001 #ifdef CONFIG_NET_DMA
5002 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5005 struct tcp_sock
*tp
= tcp_sk(sk
);
5006 int chunk
= skb
->len
- hlen
;
5008 int copied_early
= 0;
5010 if (tp
->ucopy
.wakeup
)
5013 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5014 tp
->ucopy
.dma_chan
= get_softnet_dma();
5016 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5018 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5020 tp
->ucopy
.iov
, chunk
,
5021 tp
->ucopy
.pinned_list
);
5026 tp
->ucopy
.dma_cookie
= dma_cookie
;
5029 tp
->ucopy
.len
-= chunk
;
5030 tp
->copied_seq
+= chunk
;
5031 tcp_rcv_space_adjust(sk
);
5033 if ((tp
->ucopy
.len
== 0) ||
5034 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5035 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5036 tp
->ucopy
.wakeup
= 1;
5037 sk
->sk_data_ready(sk
, 0);
5039 } else if (chunk
> 0) {
5040 tp
->ucopy
.wakeup
= 1;
5041 sk
->sk_data_ready(sk
, 0);
5044 return copied_early
;
5046 #endif /* CONFIG_NET_DMA */
5048 /* Does PAWS and seqno based validation of an incoming segment, flags will
5049 * play significant role here.
5051 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5052 struct tcphdr
*th
, int syn_inerr
)
5054 struct tcp_sock
*tp
= tcp_sk(sk
);
5056 /* RFC1323: H1. Apply PAWS check first. */
5057 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
5058 tcp_paws_discard(sk
, skb
)) {
5060 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5061 tcp_send_dupack(sk
, skb
);
5064 /* Reset is accepted even if it did not pass PAWS. */
5067 /* Step 1: check sequence number */
5068 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5069 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5070 * (RST) segments are validated by checking their SEQ-fields."
5071 * And page 69: "If an incoming segment is not acceptable,
5072 * an acknowledgment should be sent in reply (unless the RST
5073 * bit is set, if so drop the segment and return)".
5076 tcp_send_dupack(sk
, skb
);
5080 /* Step 2: check RST bit */
5086 /* ts_recent update must be made after we are sure that the packet
5089 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5091 /* step 3: check security and precedence [ignored] */
5093 /* step 4: Check for a SYN in window. */
5094 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5096 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5097 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5110 * TCP receive function for the ESTABLISHED state.
5112 * It is split into a fast path and a slow path. The fast path is
5114 * - A zero window was announced from us - zero window probing
5115 * is only handled properly in the slow path.
5116 * - Out of order segments arrived.
5117 * - Urgent data is expected.
5118 * - There is no buffer space left
5119 * - Unexpected TCP flags/window values/header lengths are received
5120 * (detected by checking the TCP header against pred_flags)
5121 * - Data is sent in both directions. Fast path only supports pure senders
5122 * or pure receivers (this means either the sequence number or the ack
5123 * value must stay constant)
5124 * - Unexpected TCP option.
5126 * When these conditions are not satisfied it drops into a standard
5127 * receive procedure patterned after RFC793 to handle all cases.
5128 * The first three cases are guaranteed by proper pred_flags setting,
5129 * the rest is checked inline. Fast processing is turned on in
5130 * tcp_data_queue when everything is OK.
5132 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5133 struct tcphdr
*th
, unsigned len
)
5135 struct tcp_sock
*tp
= tcp_sk(sk
);
5139 * Header prediction.
5140 * The code loosely follows the one in the famous
5141 * "30 instruction TCP receive" Van Jacobson mail.
5143 * Van's trick is to deposit buffers into socket queue
5144 * on a device interrupt, to call tcp_recv function
5145 * on the receive process context and checksum and copy
5146 * the buffer to user space. smart...
5148 * Our current scheme is not silly either but we take the
5149 * extra cost of the net_bh soft interrupt processing...
5150 * We do checksum and copy also but from device to kernel.
5153 tp
->rx_opt
.saw_tstamp
= 0;
5155 /* pred_flags is 0xS?10 << 16 + snd_wnd
5156 * if header_prediction is to be made
5157 * 'S' will always be tp->tcp_header_len >> 2
5158 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5159 * turn it off (when there are holes in the receive
5160 * space for instance)
5161 * PSH flag is ignored.
5164 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5165 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
5166 int tcp_header_len
= tp
->tcp_header_len
;
5168 /* Timestamp header prediction: tcp_header_len
5169 * is automatically equal to th->doff*4 due to pred_flags
5173 /* Check timestamp */
5174 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5175 /* No? Slow path! */
5176 if (!tcp_parse_aligned_timestamp(tp
, th
))
5179 /* If PAWS failed, check it more carefully in slow path */
5180 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5183 /* DO NOT update ts_recent here, if checksum fails
5184 * and timestamp was corrupted part, it will result
5185 * in a hung connection since we will drop all
5186 * future packets due to the PAWS test.
5190 if (len
<= tcp_header_len
) {
5191 /* Bulk data transfer: sender */
5192 if (len
== tcp_header_len
) {
5193 /* Predicted packet is in window by definition.
5194 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5195 * Hence, check seq<=rcv_wup reduces to:
5197 if (tcp_header_len
==
5198 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5199 tp
->rcv_nxt
== tp
->rcv_wup
)
5200 tcp_store_ts_recent(tp
);
5202 /* We know that such packets are checksummed
5205 tcp_ack(sk
, skb
, 0);
5207 tcp_data_snd_check(sk
);
5209 } else { /* Header too small */
5210 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5215 int copied_early
= 0;
5217 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5218 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5219 #ifdef CONFIG_NET_DMA
5220 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5225 if (tp
->ucopy
.task
== current
&&
5226 sock_owned_by_user(sk
) && !copied_early
) {
5227 __set_current_state(TASK_RUNNING
);
5229 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5233 /* Predicted packet is in window by definition.
5234 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5235 * Hence, check seq<=rcv_wup reduces to:
5237 if (tcp_header_len
==
5238 (sizeof(struct tcphdr
) +
5239 TCPOLEN_TSTAMP_ALIGNED
) &&
5240 tp
->rcv_nxt
== tp
->rcv_wup
)
5241 tcp_store_ts_recent(tp
);
5243 tcp_rcv_rtt_measure_ts(sk
, skb
);
5245 __skb_pull(skb
, tcp_header_len
);
5246 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5247 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5250 tcp_cleanup_rbuf(sk
, skb
->len
);
5253 if (tcp_checksum_complete_user(sk
, skb
))
5256 /* Predicted packet is in window by definition.
5257 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5258 * Hence, check seq<=rcv_wup reduces to:
5260 if (tcp_header_len
==
5261 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5262 tp
->rcv_nxt
== tp
->rcv_wup
)
5263 tcp_store_ts_recent(tp
);
5265 tcp_rcv_rtt_measure_ts(sk
, skb
);
5267 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5270 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5272 /* Bulk data transfer: receiver */
5273 __skb_pull(skb
, tcp_header_len
);
5274 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
5275 skb_set_owner_r(skb
, sk
);
5276 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5279 tcp_event_data_recv(sk
, skb
);
5281 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5282 /* Well, only one small jumplet in fast path... */
5283 tcp_ack(sk
, skb
, FLAG_DATA
);
5284 tcp_data_snd_check(sk
);
5285 if (!inet_csk_ack_scheduled(sk
))
5289 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5290 __tcp_ack_snd_check(sk
, 0);
5292 #ifdef CONFIG_NET_DMA
5294 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5300 sk
->sk_data_ready(sk
, 0);
5306 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5310 * Standard slow path.
5313 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
5319 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5321 tcp_rcv_rtt_measure_ts(sk
, skb
);
5323 /* Process urgent data. */
5324 tcp_urg(sk
, skb
, th
);
5326 /* step 7: process the segment text */
5327 tcp_data_queue(sk
, skb
);
5329 tcp_data_snd_check(sk
);
5330 tcp_ack_snd_check(sk
);
5334 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5341 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5342 struct tcphdr
*th
, unsigned len
)
5344 struct tcp_sock
*tp
= tcp_sk(sk
);
5345 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5346 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5348 tcp_parse_options(skb
, &tp
->rx_opt
, 0);
5352 * "If the state is SYN-SENT then
5353 * first check the ACK bit
5354 * If the ACK bit is set
5355 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5356 * a reset (unless the RST bit is set, if so drop
5357 * the segment and return)"
5359 * We do not send data with SYN, so that RFC-correct
5362 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5363 goto reset_and_undo
;
5365 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5366 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5368 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5369 goto reset_and_undo
;
5372 /* Now ACK is acceptable.
5374 * "If the RST bit is set
5375 * If the ACK was acceptable then signal the user "error:
5376 * connection reset", drop the segment, enter CLOSED state,
5377 * delete TCB, and return."
5386 * "fifth, if neither of the SYN or RST bits is set then
5387 * drop the segment and return."
5393 goto discard_and_undo
;
5396 * "If the SYN bit is on ...
5397 * are acceptable then ...
5398 * (our SYN has been ACKed), change the connection
5399 * state to ESTABLISHED..."
5402 TCP_ECN_rcv_synack(tp
, th
);
5404 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5405 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5407 /* Ok.. it's good. Set up sequence numbers and
5408 * move to established.
5410 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5411 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5413 /* RFC1323: The window in SYN & SYN/ACK segments is
5416 tp
->snd_wnd
= ntohs(th
->window
);
5417 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
, TCP_SKB_CB(skb
)->seq
);
5419 if (!tp
->rx_opt
.wscale_ok
) {
5420 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5421 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5424 if (tp
->rx_opt
.saw_tstamp
) {
5425 tp
->rx_opt
.tstamp_ok
= 1;
5426 tp
->tcp_header_len
=
5427 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5428 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5429 tcp_store_ts_recent(tp
);
5431 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5434 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5435 tcp_enable_fack(tp
);
5438 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5439 tcp_initialize_rcv_mss(sk
);
5441 /* Remember, tcp_poll() does not lock socket!
5442 * Change state from SYN-SENT only after copied_seq
5443 * is initialized. */
5444 tp
->copied_seq
= tp
->rcv_nxt
;
5446 tcp_set_state(sk
, TCP_ESTABLISHED
);
5448 security_inet_conn_established(sk
, skb
);
5450 /* Make sure socket is routed, for correct metrics. */
5451 icsk
->icsk_af_ops
->rebuild_header(sk
);
5453 tcp_init_metrics(sk
);
5455 tcp_init_congestion_control(sk
);
5457 /* Prevent spurious tcp_cwnd_restart() on first data
5460 tp
->lsndtime
= tcp_time_stamp
;
5462 tcp_init_buffer_space(sk
);
5464 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5465 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5467 if (!tp
->rx_opt
.snd_wscale
)
5468 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5472 if (!sock_flag(sk
, SOCK_DEAD
)) {
5473 sk
->sk_state_change(sk
);
5474 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5477 if (sk
->sk_write_pending
||
5478 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5479 icsk
->icsk_ack
.pingpong
) {
5480 /* Save one ACK. Data will be ready after
5481 * several ticks, if write_pending is set.
5483 * It may be deleted, but with this feature tcpdumps
5484 * look so _wonderfully_ clever, that I was not able
5485 * to stand against the temptation 8) --ANK
5487 inet_csk_schedule_ack(sk
);
5488 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5489 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5490 tcp_incr_quickack(sk
);
5491 tcp_enter_quickack_mode(sk
);
5492 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5493 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5504 /* No ACK in the segment */
5508 * "If the RST bit is set
5510 * Otherwise (no ACK) drop the segment and return."
5513 goto discard_and_undo
;
5517 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5518 tcp_paws_check(&tp
->rx_opt
, 0))
5519 goto discard_and_undo
;
5522 /* We see SYN without ACK. It is attempt of
5523 * simultaneous connect with crossed SYNs.
5524 * Particularly, it can be connect to self.
5526 tcp_set_state(sk
, TCP_SYN_RECV
);
5528 if (tp
->rx_opt
.saw_tstamp
) {
5529 tp
->rx_opt
.tstamp_ok
= 1;
5530 tcp_store_ts_recent(tp
);
5531 tp
->tcp_header_len
=
5532 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5534 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5537 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5538 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5540 /* RFC1323: The window in SYN & SYN/ACK segments is
5543 tp
->snd_wnd
= ntohs(th
->window
);
5544 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5545 tp
->max_window
= tp
->snd_wnd
;
5547 TCP_ECN_rcv_syn(tp
, th
);
5550 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5551 tcp_initialize_rcv_mss(sk
);
5553 tcp_send_synack(sk
);
5555 /* Note, we could accept data and URG from this segment.
5556 * There are no obstacles to make this.
5558 * However, if we ignore data in ACKless segments sometimes,
5559 * we have no reasons to accept it sometimes.
5560 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5561 * is not flawless. So, discard packet for sanity.
5562 * Uncomment this return to process the data.
5569 /* "fifth, if neither of the SYN or RST bits is set then
5570 * drop the segment and return."
5574 tcp_clear_options(&tp
->rx_opt
);
5575 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5579 tcp_clear_options(&tp
->rx_opt
);
5580 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5585 * This function implements the receiving procedure of RFC 793 for
5586 * all states except ESTABLISHED and TIME_WAIT.
5587 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5588 * address independent.
5591 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5592 struct tcphdr
*th
, unsigned len
)
5594 struct tcp_sock
*tp
= tcp_sk(sk
);
5595 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5599 tp
->rx_opt
.saw_tstamp
= 0;
5601 switch (sk
->sk_state
) {
5613 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5616 /* Now we have several options: In theory there is
5617 * nothing else in the frame. KA9Q has an option to
5618 * send data with the syn, BSD accepts data with the
5619 * syn up to the [to be] advertised window and
5620 * Solaris 2.1 gives you a protocol error. For now
5621 * we just ignore it, that fits the spec precisely
5622 * and avoids incompatibilities. It would be nice in
5623 * future to drop through and process the data.
5625 * Now that TTCP is starting to be used we ought to
5627 * But, this leaves one open to an easy denial of
5628 * service attack, and SYN cookies can't defend
5629 * against this problem. So, we drop the data
5630 * in the interest of security over speed unless
5631 * it's still in use.
5639 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5643 /* Do step6 onward by hand. */
5644 tcp_urg(sk
, skb
, th
);
5646 tcp_data_snd_check(sk
);
5650 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5654 /* step 5: check the ACK field */
5656 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5658 switch (sk
->sk_state
) {
5661 tp
->copied_seq
= tp
->rcv_nxt
;
5663 tcp_set_state(sk
, TCP_ESTABLISHED
);
5664 sk
->sk_state_change(sk
);
5666 /* Note, that this wakeup is only for marginal
5667 * crossed SYN case. Passively open sockets
5668 * are not waked up, because sk->sk_sleep ==
5669 * NULL and sk->sk_socket == NULL.
5673 SOCK_WAKE_IO
, POLL_OUT
);
5675 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5676 tp
->snd_wnd
= ntohs(th
->window
) <<
5677 tp
->rx_opt
.snd_wscale
;
5678 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->ack_seq
,
5679 TCP_SKB_CB(skb
)->seq
);
5681 /* tcp_ack considers this ACK as duplicate
5682 * and does not calculate rtt.
5683 * Fix it at least with timestamps.
5685 if (tp
->rx_opt
.saw_tstamp
&&
5686 tp
->rx_opt
.rcv_tsecr
&& !tp
->srtt
)
5687 tcp_ack_saw_tstamp(sk
, 0);
5689 if (tp
->rx_opt
.tstamp_ok
)
5690 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5692 /* Make sure socket is routed, for
5695 icsk
->icsk_af_ops
->rebuild_header(sk
);
5697 tcp_init_metrics(sk
);
5699 tcp_init_congestion_control(sk
);
5701 /* Prevent spurious tcp_cwnd_restart() on
5702 * first data packet.
5704 tp
->lsndtime
= tcp_time_stamp
;
5707 tcp_initialize_rcv_mss(sk
);
5708 tcp_init_buffer_space(sk
);
5709 tcp_fast_path_on(tp
);
5716 if (tp
->snd_una
== tp
->write_seq
) {
5717 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5718 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5719 dst_confirm(sk
->sk_dst_cache
);
5721 if (!sock_flag(sk
, SOCK_DEAD
))
5722 /* Wake up lingering close() */
5723 sk
->sk_state_change(sk
);
5727 if (tp
->linger2
< 0 ||
5728 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5729 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5731 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5735 tmo
= tcp_fin_time(sk
);
5736 if (tmo
> TCP_TIMEWAIT_LEN
) {
5737 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5738 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5739 /* Bad case. We could lose such FIN otherwise.
5740 * It is not a big problem, but it looks confusing
5741 * and not so rare event. We still can lose it now,
5742 * if it spins in bh_lock_sock(), but it is really
5745 inet_csk_reset_keepalive_timer(sk
, tmo
);
5747 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5755 if (tp
->snd_una
== tp
->write_seq
) {
5756 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5762 if (tp
->snd_una
== tp
->write_seq
) {
5763 tcp_update_metrics(sk
);
5772 /* step 6: check the URG bit */
5773 tcp_urg(sk
, skb
, th
);
5775 /* step 7: process the segment text */
5776 switch (sk
->sk_state
) {
5777 case TCP_CLOSE_WAIT
:
5780 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5784 /* RFC 793 says to queue data in these states,
5785 * RFC 1122 says we MUST send a reset.
5786 * BSD 4.4 also does reset.
5788 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5789 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5790 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5791 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5797 case TCP_ESTABLISHED
:
5798 tcp_data_queue(sk
, skb
);
5803 /* tcp_data could move socket to TIME-WAIT */
5804 if (sk
->sk_state
!= TCP_CLOSE
) {
5805 tcp_data_snd_check(sk
);
5806 tcp_ack_snd_check(sk
);
5816 EXPORT_SYMBOL(sysctl_tcp_ecn
);
5817 EXPORT_SYMBOL(sysctl_tcp_reordering
);
5818 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
5819 EXPORT_SYMBOL(tcp_parse_options
);
5820 #ifdef CONFIG_TCP_MD5SIG
5821 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
5823 EXPORT_SYMBOL(tcp_rcv_established
);
5824 EXPORT_SYMBOL(tcp_rcv_state_process
);
5825 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);