2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #include <linux/slab.h>
66 #include <linux/module.h>
67 #include <linux/sysctl.h>
68 #include <linux/kernel.h>
71 #include <net/inet_common.h>
72 #include <linux/ipsec.h>
73 #include <asm/unaligned.h>
74 #include <net/netdma.h>
76 int sysctl_tcp_timestamps __read_mostly
= 1;
77 int sysctl_tcp_window_scaling __read_mostly
= 1;
78 int sysctl_tcp_sack __read_mostly
= 1;
79 int sysctl_tcp_fack __read_mostly
= 1;
80 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
81 EXPORT_SYMBOL(sysctl_tcp_reordering
);
82 int sysctl_tcp_ecn __read_mostly
= 2;
83 EXPORT_SYMBOL(sysctl_tcp_ecn
);
84 int sysctl_tcp_dsack __read_mostly
= 1;
85 int sysctl_tcp_app_win __read_mostly
= 31;
86 int sysctl_tcp_adv_win_scale __read_mostly
= 2;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 int sysctl_tcp_stdurg __read_mostly
;
90 int sysctl_tcp_rfc1337 __read_mostly
;
91 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
92 int sysctl_tcp_frto __read_mostly
= 2;
93 int sysctl_tcp_frto_response __read_mostly
;
94 int sysctl_tcp_nometrics_save __read_mostly
;
96 int sysctl_tcp_thin_dupack __read_mostly
;
98 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
99 int sysctl_tcp_abc __read_mostly
;
101 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
102 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
103 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
104 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
105 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
106 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
107 #define FLAG_ECE 0x40 /* ECE in this ACK */
108 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
109 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
110 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
111 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
112 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
115 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
116 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
117 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
118 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
119 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 /* Adapt the MSS value used to make delayed ack decision to the
127 static void tcp_measure_rcv_mss(struct sock
*sk
, const struct sk_buff
*skb
)
129 struct inet_connection_sock
*icsk
= inet_csk(sk
);
130 const unsigned int lss
= icsk
->icsk_ack
.last_seg_size
;
133 icsk
->icsk_ack
.last_seg_size
= 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
138 len
= skb_shinfo(skb
)->gso_size
? : skb
->len
;
139 if (len
>= icsk
->icsk_ack
.rcv_mss
) {
140 icsk
->icsk_ack
.rcv_mss
= len
;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len
+= skb
->data
- skb_transport_header(skb
);
148 if (len
>= TCP_MSS_DEFAULT
+ sizeof(struct tcphdr
) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len
>= TCP_MIN_MSS
+ sizeof(struct tcphdr
) &&
155 !(tcp_flag_word(tcp_hdr(skb
)) & TCP_REMNANT
))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len
-= tcp_sk(sk
)->tcp_header_len
;
161 icsk
->icsk_ack
.last_seg_size
= len
;
163 icsk
->icsk_ack
.rcv_mss
= len
;
167 if (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
)
168 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED2
;
169 icsk
->icsk_ack
.pending
|= ICSK_ACK_PUSHED
;
173 static void tcp_incr_quickack(struct sock
*sk
)
175 struct inet_connection_sock
*icsk
= inet_csk(sk
);
176 unsigned quickacks
= tcp_sk(sk
)->rcv_wnd
/ (2 * icsk
->icsk_ack
.rcv_mss
);
180 if (quickacks
> icsk
->icsk_ack
.quick
)
181 icsk
->icsk_ack
.quick
= min(quickacks
, TCP_MAX_QUICKACKS
);
184 static void tcp_enter_quickack_mode(struct sock
*sk
)
186 struct inet_connection_sock
*icsk
= inet_csk(sk
);
187 tcp_incr_quickack(sk
);
188 icsk
->icsk_ack
.pingpong
= 0;
189 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static inline int tcp_in_quickack_mode(const struct sock
*sk
)
198 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
199 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
202 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
204 if (tp
->ecn_flags
& TCP_ECN_OK
)
205 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
208 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
210 if (tcp_hdr(skb
)->cwr
)
211 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
214 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
216 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
219 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
221 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
224 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
225 case INET_ECN_NOT_ECT
:
226 /* Funny extension: if ECT is not set on a segment,
227 * and we already seen ECT on a previous segment,
228 * it is probably a retransmit.
230 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
231 tcp_enter_quickack_mode((struct sock
*)tp
);
234 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
237 tp
->ecn_flags
|= TCP_ECN_SEEN
;
241 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
243 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
244 tp
->ecn_flags
&= ~TCP_ECN_OK
;
247 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
249 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
250 tp
->ecn_flags
&= ~TCP_ECN_OK
;
253 static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
255 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
260 /* Buffer size and advertised window tuning.
262 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
265 static void tcp_fixup_sndbuf(struct sock
*sk
)
267 int sndmem
= SKB_TRUESIZE(tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
);
269 sndmem
*= TCP_INIT_CWND
;
270 if (sk
->sk_sndbuf
< sndmem
)
271 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
274 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
276 * All tcp_full_space() is split to two parts: "network" buffer, allocated
277 * forward and advertised in receiver window (tp->rcv_wnd) and
278 * "application buffer", required to isolate scheduling/application
279 * latencies from network.
280 * window_clamp is maximal advertised window. It can be less than
281 * tcp_full_space(), in this case tcp_full_space() - window_clamp
282 * is reserved for "application" buffer. The less window_clamp is
283 * the smoother our behaviour from viewpoint of network, but the lower
284 * throughput and the higher sensitivity of the connection to losses. 8)
286 * rcv_ssthresh is more strict window_clamp used at "slow start"
287 * phase to predict further behaviour of this connection.
288 * It is used for two goals:
289 * - to enforce header prediction at sender, even when application
290 * requires some significant "application buffer". It is check #1.
291 * - to prevent pruning of receive queue because of misprediction
292 * of receiver window. Check #2.
294 * The scheme does not work when sender sends good segments opening
295 * window and then starts to feed us spaghetti. But it should work
296 * in common situations. Otherwise, we have to rely on queue collapsing.
299 /* Slow part of check#2. */
300 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
302 struct tcp_sock
*tp
= tcp_sk(sk
);
304 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
305 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
307 while (tp
->rcv_ssthresh
<= window
) {
308 if (truesize
<= skb
->len
)
309 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
317 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
319 struct tcp_sock
*tp
= tcp_sk(sk
);
322 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
323 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
324 !sk_under_memory_pressure(sk
)) {
327 /* Check #2. Increase window, if skb with such overhead
328 * will fit to rcvbuf in future.
330 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
331 incr
= 2 * tp
->advmss
;
333 incr
= __tcp_grow_window(sk
, skb
);
336 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
338 inet_csk(sk
)->icsk_ack
.quick
|= 1;
343 /* 3. Tuning rcvbuf, when connection enters established state. */
345 static void tcp_fixup_rcvbuf(struct sock
*sk
)
347 u32 mss
= tcp_sk(sk
)->advmss
;
348 u32 icwnd
= TCP_DEFAULT_INIT_RCVWND
;
351 /* Limit to 10 segments if mss <= 1460,
352 * or 14600/mss segments, with a minimum of two segments.
355 icwnd
= max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
357 rcvmem
= SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
);
358 while (tcp_win_from_space(rcvmem
) < mss
)
363 if (sk
->sk_rcvbuf
< rcvmem
)
364 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
367 /* 4. Try to fixup all. It is made immediately after connection enters
370 static void tcp_init_buffer_space(struct sock
*sk
)
372 struct tcp_sock
*tp
= tcp_sk(sk
);
375 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
376 tcp_fixup_rcvbuf(sk
);
377 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
378 tcp_fixup_sndbuf(sk
);
380 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
382 maxwin
= tcp_full_space(sk
);
384 if (tp
->window_clamp
>= maxwin
) {
385 tp
->window_clamp
= maxwin
;
387 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
388 tp
->window_clamp
= max(maxwin
-
389 (maxwin
>> sysctl_tcp_app_win
),
393 /* Force reservation of one segment. */
394 if (sysctl_tcp_app_win
&&
395 tp
->window_clamp
> 2 * tp
->advmss
&&
396 tp
->window_clamp
+ tp
->advmss
> maxwin
)
397 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
399 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
400 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
403 /* 5. Recalculate window clamp after socket hit its memory bounds. */
404 static void tcp_clamp_window(struct sock
*sk
)
406 struct tcp_sock
*tp
= tcp_sk(sk
);
407 struct inet_connection_sock
*icsk
= inet_csk(sk
);
409 icsk
->icsk_ack
.quick
= 0;
411 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
412 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
413 !sk_under_memory_pressure(sk
) &&
414 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
415 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
418 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
419 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
422 /* Initialize RCV_MSS value.
423 * RCV_MSS is an our guess about MSS used by the peer.
424 * We haven't any direct information about the MSS.
425 * It's better to underestimate the RCV_MSS rather than overestimate.
426 * Overestimations make us ACKing less frequently than needed.
427 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
429 void tcp_initialize_rcv_mss(struct sock
*sk
)
431 const struct tcp_sock
*tp
= tcp_sk(sk
);
432 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
434 hint
= min(hint
, tp
->rcv_wnd
/ 2);
435 hint
= min(hint
, TCP_MSS_DEFAULT
);
436 hint
= max(hint
, TCP_MIN_MSS
);
438 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
440 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
442 /* Receiver "autotuning" code.
444 * The algorithm for RTT estimation w/o timestamps is based on
445 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
446 * <http://public.lanl.gov/radiant/pubs.html#DRS>
448 * More detail on this code can be found at
449 * <http://staff.psc.edu/jheffner/>,
450 * though this reference is out of date. A new paper
453 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
455 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
461 if (new_sample
!= 0) {
462 /* If we sample in larger samples in the non-timestamp
463 * case, we could grossly overestimate the RTT especially
464 * with chatty applications or bulk transfer apps which
465 * are stalled on filesystem I/O.
467 * Also, since we are only going for a minimum in the
468 * non-timestamp case, we do not smooth things out
469 * else with timestamps disabled convergence takes too
473 m
-= (new_sample
>> 3);
475 } else if (m
< new_sample
)
478 /* No previous measure. */
482 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
483 tp
->rcv_rtt_est
.rtt
= new_sample
;
486 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
488 if (tp
->rcv_rtt_est
.time
== 0)
490 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
492 tcp_rcv_rtt_update(tp
, jiffies
- tp
->rcv_rtt_est
.time
, 1);
495 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
496 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
499 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
500 const struct sk_buff
*skb
)
502 struct tcp_sock
*tp
= tcp_sk(sk
);
503 if (tp
->rx_opt
.rcv_tsecr
&&
504 (TCP_SKB_CB(skb
)->end_seq
-
505 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
506 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
510 * This function should be called every time data is copied to user space.
511 * It calculates the appropriate TCP receive buffer space.
513 void tcp_rcv_space_adjust(struct sock
*sk
)
515 struct tcp_sock
*tp
= tcp_sk(sk
);
519 if (tp
->rcvq_space
.time
== 0)
522 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
523 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
526 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
528 space
= max(tp
->rcvq_space
.space
, space
);
530 if (tp
->rcvq_space
.space
!= space
) {
533 tp
->rcvq_space
.space
= space
;
535 if (sysctl_tcp_moderate_rcvbuf
&&
536 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
537 int new_clamp
= space
;
539 /* Receive space grows, normalize in order to
540 * take into account packet headers and sk_buff
541 * structure overhead.
546 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
547 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
550 space
= min(space
, sysctl_tcp_rmem
[2]);
551 if (space
> sk
->sk_rcvbuf
) {
552 sk
->sk_rcvbuf
= space
;
554 /* Make the window clamp follow along. */
555 tp
->window_clamp
= new_clamp
;
561 tp
->rcvq_space
.seq
= tp
->copied_seq
;
562 tp
->rcvq_space
.time
= tcp_time_stamp
;
565 /* There is something which you must keep in mind when you analyze the
566 * behavior of the tp->ato delayed ack timeout interval. When a
567 * connection starts up, we want to ack as quickly as possible. The
568 * problem is that "good" TCP's do slow start at the beginning of data
569 * transmission. The means that until we send the first few ACK's the
570 * sender will sit on his end and only queue most of his data, because
571 * he can only send snd_cwnd unacked packets at any given time. For
572 * each ACK we send, he increments snd_cwnd and transmits more of his
575 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
577 struct tcp_sock
*tp
= tcp_sk(sk
);
578 struct inet_connection_sock
*icsk
= inet_csk(sk
);
581 inet_csk_schedule_ack(sk
);
583 tcp_measure_rcv_mss(sk
, skb
);
585 tcp_rcv_rtt_measure(tp
);
587 now
= tcp_time_stamp
;
589 if (!icsk
->icsk_ack
.ato
) {
590 /* The _first_ data packet received, initialize
591 * delayed ACK engine.
593 tcp_incr_quickack(sk
);
594 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
596 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
598 if (m
<= TCP_ATO_MIN
/ 2) {
599 /* The fastest case is the first. */
600 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
601 } else if (m
< icsk
->icsk_ack
.ato
) {
602 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
603 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
604 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
605 } else if (m
> icsk
->icsk_rto
) {
606 /* Too long gap. Apparently sender failed to
607 * restart window, so that we send ACKs quickly.
609 tcp_incr_quickack(sk
);
613 icsk
->icsk_ack
.lrcvtime
= now
;
615 TCP_ECN_check_ce(tp
, skb
);
618 tcp_grow_window(sk
, skb
);
621 /* Called to compute a smoothed rtt estimate. The data fed to this
622 * routine either comes from timestamps, or from segments that were
623 * known _not_ to have been retransmitted [see Karn/Partridge
624 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
625 * piece by Van Jacobson.
626 * NOTE: the next three routines used to be one big routine.
627 * To save cycles in the RFC 1323 implementation it was better to break
628 * it up into three procedures. -- erics
630 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
632 struct tcp_sock
*tp
= tcp_sk(sk
);
633 long m
= mrtt
; /* RTT */
635 /* The following amusing code comes from Jacobson's
636 * article in SIGCOMM '88. Note that rtt and mdev
637 * are scaled versions of rtt and mean deviation.
638 * This is designed to be as fast as possible
639 * m stands for "measurement".
641 * On a 1990 paper the rto value is changed to:
642 * RTO = rtt + 4 * mdev
644 * Funny. This algorithm seems to be very broken.
645 * These formulae increase RTO, when it should be decreased, increase
646 * too slowly, when it should be increased quickly, decrease too quickly
647 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
648 * does not matter how to _calculate_ it. Seems, it was trap
649 * that VJ failed to avoid. 8)
654 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
655 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
657 m
= -m
; /* m is now abs(error) */
658 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
659 /* This is similar to one of Eifel findings.
660 * Eifel blocks mdev updates when rtt decreases.
661 * This solution is a bit different: we use finer gain
662 * for mdev in this case (alpha*beta).
663 * Like Eifel it also prevents growth of rto,
664 * but also it limits too fast rto decreases,
665 * happening in pure Eifel.
670 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
672 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
673 if (tp
->mdev
> tp
->mdev_max
) {
674 tp
->mdev_max
= tp
->mdev
;
675 if (tp
->mdev_max
> tp
->rttvar
)
676 tp
->rttvar
= tp
->mdev_max
;
678 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
679 if (tp
->mdev_max
< tp
->rttvar
)
680 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
681 tp
->rtt_seq
= tp
->snd_nxt
;
682 tp
->mdev_max
= tcp_rto_min(sk
);
685 /* no previous measure. */
686 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
687 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
688 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
689 tp
->rtt_seq
= tp
->snd_nxt
;
693 /* Calculate rto without backoff. This is the second half of Van Jacobson's
694 * routine referred to above.
696 static inline void tcp_set_rto(struct sock
*sk
)
698 const struct tcp_sock
*tp
= tcp_sk(sk
);
699 /* Old crap is replaced with new one. 8)
702 * 1. If rtt variance happened to be less 50msec, it is hallucination.
703 * It cannot be less due to utterly erratic ACK generation made
704 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
705 * to do with delayed acks, because at cwnd>2 true delack timeout
706 * is invisible. Actually, Linux-2.4 also generates erratic
707 * ACKs in some circumstances.
709 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
711 /* 2. Fixups made earlier cannot be right.
712 * If we do not estimate RTO correctly without them,
713 * all the algo is pure shit and should be replaced
714 * with correct one. It is exactly, which we pretend to do.
717 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
718 * guarantees that rto is higher.
723 /* Save metrics learned by this TCP session.
724 This function is called only, when TCP finishes successfully
725 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
727 void tcp_update_metrics(struct sock
*sk
)
729 struct tcp_sock
*tp
= tcp_sk(sk
);
730 struct dst_entry
*dst
= __sk_dst_get(sk
);
732 if (sysctl_tcp_nometrics_save
)
737 if (dst
&& (dst
->flags
& DST_HOST
)) {
738 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
742 if (icsk
->icsk_backoff
|| !tp
->srtt
) {
743 /* This session failed to estimate rtt. Why?
744 * Probably, no packets returned in time.
747 if (!(dst_metric_locked(dst
, RTAX_RTT
)))
748 dst_metric_set(dst
, RTAX_RTT
, 0);
752 rtt
= dst_metric_rtt(dst
, RTAX_RTT
);
755 /* If newly calculated rtt larger than stored one,
756 * store new one. Otherwise, use EWMA. Remember,
757 * rtt overestimation is always better than underestimation.
759 if (!(dst_metric_locked(dst
, RTAX_RTT
))) {
761 set_dst_metric_rtt(dst
, RTAX_RTT
, tp
->srtt
);
763 set_dst_metric_rtt(dst
, RTAX_RTT
, rtt
- (m
>> 3));
766 if (!(dst_metric_locked(dst
, RTAX_RTTVAR
))) {
771 /* Scale deviation to rttvar fixed point */
776 var
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
780 var
-= (var
- m
) >> 2;
782 set_dst_metric_rtt(dst
, RTAX_RTTVAR
, var
);
785 if (tcp_in_initial_slowstart(tp
)) {
786 /* Slow start still did not finish. */
787 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
788 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
789 (tp
->snd_cwnd
>> 1) > dst_metric(dst
, RTAX_SSTHRESH
))
790 dst_metric_set(dst
, RTAX_SSTHRESH
, tp
->snd_cwnd
>> 1);
791 if (!dst_metric_locked(dst
, RTAX_CWND
) &&
792 tp
->snd_cwnd
> dst_metric(dst
, RTAX_CWND
))
793 dst_metric_set(dst
, RTAX_CWND
, tp
->snd_cwnd
);
794 } else if (tp
->snd_cwnd
> tp
->snd_ssthresh
&&
795 icsk
->icsk_ca_state
== TCP_CA_Open
) {
796 /* Cong. avoidance phase, cwnd is reliable. */
797 if (!dst_metric_locked(dst
, RTAX_SSTHRESH
))
798 dst_metric_set(dst
, RTAX_SSTHRESH
,
799 max(tp
->snd_cwnd
>> 1, tp
->snd_ssthresh
));
800 if (!dst_metric_locked(dst
, RTAX_CWND
))
801 dst_metric_set(dst
, RTAX_CWND
,
802 (dst_metric(dst
, RTAX_CWND
) +
805 /* Else slow start did not finish, cwnd is non-sense,
806 ssthresh may be also invalid.
808 if (!dst_metric_locked(dst
, RTAX_CWND
))
809 dst_metric_set(dst
, RTAX_CWND
,
810 (dst_metric(dst
, RTAX_CWND
) +
811 tp
->snd_ssthresh
) >> 1);
812 if (dst_metric(dst
, RTAX_SSTHRESH
) &&
813 !dst_metric_locked(dst
, RTAX_SSTHRESH
) &&
814 tp
->snd_ssthresh
> dst_metric(dst
, RTAX_SSTHRESH
))
815 dst_metric_set(dst
, RTAX_SSTHRESH
, tp
->snd_ssthresh
);
818 if (!dst_metric_locked(dst
, RTAX_REORDERING
)) {
819 if (dst_metric(dst
, RTAX_REORDERING
) < tp
->reordering
&&
820 tp
->reordering
!= sysctl_tcp_reordering
)
821 dst_metric_set(dst
, RTAX_REORDERING
, tp
->reordering
);
826 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
828 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
831 cwnd
= TCP_INIT_CWND
;
832 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
835 /* Set slow start threshold and cwnd not falling to slow start */
836 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
838 struct tcp_sock
*tp
= tcp_sk(sk
);
839 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
841 tp
->prior_ssthresh
= 0;
843 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
846 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
847 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
848 tcp_packets_in_flight(tp
) + 1U);
849 tp
->snd_cwnd_cnt
= 0;
850 tp
->high_seq
= tp
->snd_nxt
;
851 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
852 TCP_ECN_queue_cwr(tp
);
854 tcp_set_ca_state(sk
, TCP_CA_CWR
);
859 * Packet counting of FACK is based on in-order assumptions, therefore TCP
860 * disables it when reordering is detected
862 static void tcp_disable_fack(struct tcp_sock
*tp
)
864 /* RFC3517 uses different metric in lost marker => reset on change */
866 tp
->lost_skb_hint
= NULL
;
867 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
870 /* Take a notice that peer is sending D-SACKs */
871 static void tcp_dsack_seen(struct tcp_sock
*tp
)
873 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
876 /* Initialize metrics on socket. */
878 static void tcp_init_metrics(struct sock
*sk
)
880 struct tcp_sock
*tp
= tcp_sk(sk
);
881 struct dst_entry
*dst
= __sk_dst_get(sk
);
888 if (dst_metric_locked(dst
, RTAX_CWND
))
889 tp
->snd_cwnd_clamp
= dst_metric(dst
, RTAX_CWND
);
890 if (dst_metric(dst
, RTAX_SSTHRESH
)) {
891 tp
->snd_ssthresh
= dst_metric(dst
, RTAX_SSTHRESH
);
892 if (tp
->snd_ssthresh
> tp
->snd_cwnd_clamp
)
893 tp
->snd_ssthresh
= tp
->snd_cwnd_clamp
;
895 /* ssthresh may have been reduced unnecessarily during.
896 * 3WHS. Restore it back to its initial default.
898 tp
->snd_ssthresh
= TCP_INFINITE_SSTHRESH
;
900 if (dst_metric(dst
, RTAX_REORDERING
) &&
901 tp
->reordering
!= dst_metric(dst
, RTAX_REORDERING
)) {
902 tcp_disable_fack(tp
);
903 tp
->reordering
= dst_metric(dst
, RTAX_REORDERING
);
906 if (dst_metric(dst
, RTAX_RTT
) == 0 || tp
->srtt
== 0)
909 /* Initial rtt is determined from SYN,SYN-ACK.
910 * The segment is small and rtt may appear much
911 * less than real one. Use per-dst memory
912 * to make it more realistic.
914 * A bit of theory. RTT is time passed after "normal" sized packet
915 * is sent until it is ACKed. In normal circumstances sending small
916 * packets force peer to delay ACKs and calculation is correct too.
917 * The algorithm is adaptive and, provided we follow specs, it
918 * NEVER underestimate RTT. BUT! If peer tries to make some clever
919 * tricks sort of "quick acks" for time long enough to decrease RTT
920 * to low value, and then abruptly stops to do it and starts to delay
921 * ACKs, wait for troubles.
923 if (dst_metric_rtt(dst
, RTAX_RTT
) > tp
->srtt
) {
924 tp
->srtt
= dst_metric_rtt(dst
, RTAX_RTT
);
925 tp
->rtt_seq
= tp
->snd_nxt
;
927 if (dst_metric_rtt(dst
, RTAX_RTTVAR
) > tp
->mdev
) {
928 tp
->mdev
= dst_metric_rtt(dst
, RTAX_RTTVAR
);
929 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
934 /* RFC2988bis: We've failed to get a valid RTT sample from
935 * 3WHS. This is most likely due to retransmission,
936 * including spurious one. Reset the RTO back to 3secs
937 * from the more aggressive 1sec to avoid more spurious
940 tp
->mdev
= tp
->mdev_max
= tp
->rttvar
= TCP_TIMEOUT_FALLBACK
;
941 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_FALLBACK
;
943 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
944 * retransmitted. In light of RFC2988bis' more aggressive 1sec
945 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
946 * retransmission has occurred.
948 if (tp
->total_retrans
> 1)
951 tp
->snd_cwnd
= tcp_init_cwnd(tp
, dst
);
952 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
955 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
958 struct tcp_sock
*tp
= tcp_sk(sk
);
959 if (metric
> tp
->reordering
) {
962 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
964 /* This exciting event is worth to be remembered. 8) */
966 mib_idx
= LINUX_MIB_TCPTSREORDER
;
967 else if (tcp_is_reno(tp
))
968 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
969 else if (tcp_is_fack(tp
))
970 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
972 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
974 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
975 #if FASTRETRANS_DEBUG > 1
976 printk(KERN_DEBUG
"Disorder%d %d %u f%u s%u rr%d\n",
977 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
981 tp
->undo_marker
? tp
->undo_retrans
: 0);
983 tcp_disable_fack(tp
);
987 /* This must be called before lost_out is incremented */
988 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
990 if ((tp
->retransmit_skb_hint
== NULL
) ||
991 before(TCP_SKB_CB(skb
)->seq
,
992 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
993 tp
->retransmit_skb_hint
= skb
;
996 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
997 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1000 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1002 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1003 tcp_verify_retransmit_hint(tp
, skb
);
1005 tp
->lost_out
+= tcp_skb_pcount(skb
);
1006 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1010 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
1011 struct sk_buff
*skb
)
1013 tcp_verify_retransmit_hint(tp
, skb
);
1015 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
1016 tp
->lost_out
+= tcp_skb_pcount(skb
);
1017 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1021 /* This procedure tags the retransmission queue when SACKs arrive.
1023 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1024 * Packets in queue with these bits set are counted in variables
1025 * sacked_out, retrans_out and lost_out, correspondingly.
1027 * Valid combinations are:
1028 * Tag InFlight Description
1029 * 0 1 - orig segment is in flight.
1030 * S 0 - nothing flies, orig reached receiver.
1031 * L 0 - nothing flies, orig lost by net.
1032 * R 2 - both orig and retransmit are in flight.
1033 * L|R 1 - orig is lost, retransmit is in flight.
1034 * S|R 1 - orig reached receiver, retrans is still in flight.
1035 * (L|S|R is logically valid, it could occur when L|R is sacked,
1036 * but it is equivalent to plain S and code short-curcuits it to S.
1037 * L|S is logically invalid, it would mean -1 packet in flight 8))
1039 * These 6 states form finite state machine, controlled by the following events:
1040 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1041 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1042 * 3. Loss detection event of two flavors:
1043 * A. Scoreboard estimator decided the packet is lost.
1044 * A'. Reno "three dupacks" marks head of queue lost.
1045 * A''. Its FACK modification, head until snd.fack is lost.
1046 * B. SACK arrives sacking SND.NXT at the moment, when the
1047 * segment was retransmitted.
1048 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1050 * It is pleasant to note, that state diagram turns out to be commutative,
1051 * so that we are allowed not to be bothered by order of our actions,
1052 * when multiple events arrive simultaneously. (see the function below).
1054 * Reordering detection.
1055 * --------------------
1056 * Reordering metric is maximal distance, which a packet can be displaced
1057 * in packet stream. With SACKs we can estimate it:
1059 * 1. SACK fills old hole and the corresponding segment was not
1060 * ever retransmitted -> reordering. Alas, we cannot use it
1061 * when segment was retransmitted.
1062 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1063 * for retransmitted and already SACKed segment -> reordering..
1064 * Both of these heuristics are not used in Loss state, when we cannot
1065 * account for retransmits accurately.
1067 * SACK block validation.
1068 * ----------------------
1070 * SACK block range validation checks that the received SACK block fits to
1071 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1072 * Note that SND.UNA is not included to the range though being valid because
1073 * it means that the receiver is rather inconsistent with itself reporting
1074 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1075 * perfectly valid, however, in light of RFC2018 which explicitly states
1076 * that "SACK block MUST reflect the newest segment. Even if the newest
1077 * segment is going to be discarded ...", not that it looks very clever
1078 * in case of head skb. Due to potentional receiver driven attacks, we
1079 * choose to avoid immediate execution of a walk in write queue due to
1080 * reneging and defer head skb's loss recovery to standard loss recovery
1081 * procedure that will eventually trigger (nothing forbids us doing this).
1083 * Implements also blockage to start_seq wrap-around. Problem lies in the
1084 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1085 * there's no guarantee that it will be before snd_nxt (n). The problem
1086 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1089 * <- outs wnd -> <- wrapzone ->
1090 * u e n u_w e_w s n_w
1092 * |<------------+------+----- TCP seqno space --------------+---------->|
1093 * ...-- <2^31 ->| |<--------...
1094 * ...---- >2^31 ------>| |<--------...
1096 * Current code wouldn't be vulnerable but it's better still to discard such
1097 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1098 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1099 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1100 * equal to the ideal case (infinite seqno space without wrap caused issues).
1102 * With D-SACK the lower bound is extended to cover sequence space below
1103 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1104 * again, D-SACK block must not to go across snd_una (for the same reason as
1105 * for the normal SACK blocks, explained above). But there all simplicity
1106 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1107 * fully below undo_marker they do not affect behavior in anyway and can
1108 * therefore be safely ignored. In rare cases (which are more or less
1109 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1110 * fragmentation and packet reordering past skb's retransmission. To consider
1111 * them correctly, the acceptable range must be extended even more though
1112 * the exact amount is rather hard to quantify. However, tp->max_window can
1113 * be used as an exaggerated estimate.
1115 static int tcp_is_sackblock_valid(struct tcp_sock
*tp
, int is_dsack
,
1116 u32 start_seq
, u32 end_seq
)
1118 /* Too far in future, or reversed (interpretation is ambiguous) */
1119 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
1122 /* Nasty start_seq wrap-around check (see comments above) */
1123 if (!before(start_seq
, tp
->snd_nxt
))
1126 /* In outstanding window? ...This is valid exit for D-SACKs too.
1127 * start_seq == snd_una is non-sensical (see comments above)
1129 if (after(start_seq
, tp
->snd_una
))
1132 if (!is_dsack
|| !tp
->undo_marker
)
1135 /* ...Then it's D-SACK, and must reside below snd_una completely */
1136 if (after(end_seq
, tp
->snd_una
))
1139 if (!before(start_seq
, tp
->undo_marker
))
1143 if (!after(end_seq
, tp
->undo_marker
))
1146 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1147 * start_seq < undo_marker and end_seq >= undo_marker.
1149 return !before(start_seq
, end_seq
- tp
->max_window
);
1152 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1153 * Event "B". Later note: FACK people cheated me again 8), we have to account
1154 * for reordering! Ugly, but should help.
1156 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1157 * less than what is now known to be received by the other end (derived from
1158 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1159 * retransmitted skbs to avoid some costly processing per ACKs.
1161 static void tcp_mark_lost_retrans(struct sock
*sk
)
1163 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1164 struct tcp_sock
*tp
= tcp_sk(sk
);
1165 struct sk_buff
*skb
;
1167 u32 new_low_seq
= tp
->snd_nxt
;
1168 u32 received_upto
= tcp_highest_sack_seq(tp
);
1170 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
1171 !after(received_upto
, tp
->lost_retrans_low
) ||
1172 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
1175 tcp_for_write_queue(skb
, sk
) {
1176 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
1178 if (skb
== tcp_send_head(sk
))
1180 if (cnt
== tp
->retrans_out
)
1182 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1185 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
1188 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1189 * constraint here (see above) but figuring out that at
1190 * least tp->reordering SACK blocks reside between ack_seq
1191 * and received_upto is not easy task to do cheaply with
1192 * the available datastructures.
1194 * Whether FACK should check here for tp->reordering segs
1195 * in-between one could argue for either way (it would be
1196 * rather simple to implement as we could count fack_count
1197 * during the walk and do tp->fackets_out - fack_count).
1199 if (after(received_upto
, ack_seq
)) {
1200 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1201 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1203 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1204 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1206 if (before(ack_seq
, new_low_seq
))
1207 new_low_seq
= ack_seq
;
1208 cnt
+= tcp_skb_pcount(skb
);
1212 if (tp
->retrans_out
)
1213 tp
->lost_retrans_low
= new_low_seq
;
1216 static int tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1217 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1220 struct tcp_sock
*tp
= tcp_sk(sk
);
1221 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1222 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1225 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1228 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1229 } else if (num_sacks
> 1) {
1230 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1231 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1233 if (!after(end_seq_0
, end_seq_1
) &&
1234 !before(start_seq_0
, start_seq_1
)) {
1237 NET_INC_STATS_BH(sock_net(sk
),
1238 LINUX_MIB_TCPDSACKOFORECV
);
1242 /* D-SACK for already forgotten data... Do dumb counting. */
1243 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1244 !after(end_seq_0
, prior_snd_una
) &&
1245 after(end_seq_0
, tp
->undo_marker
))
1251 struct tcp_sacktag_state
{
1257 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1258 * the incoming SACK may not exactly match but we can find smaller MSS
1259 * aligned portion of it that matches. Therefore we might need to fragment
1260 * which may fail and creates some hassle (caller must handle error case
1263 * FIXME: this could be merged to shift decision code
1265 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1266 u32 start_seq
, u32 end_seq
)
1269 unsigned int pkt_len
;
1272 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1273 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1275 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1276 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1277 mss
= tcp_skb_mss(skb
);
1278 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1281 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1285 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1290 /* Round if necessary so that SACKs cover only full MSSes
1291 * and/or the remaining small portion (if present)
1293 if (pkt_len
> mss
) {
1294 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1295 if (!in_sack
&& new_len
< pkt_len
) {
1297 if (new_len
> skb
->len
)
1302 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1310 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1311 static u8
tcp_sacktag_one(struct sock
*sk
,
1312 struct tcp_sacktag_state
*state
, u8 sacked
,
1313 u32 start_seq
, u32 end_seq
,
1314 int dup_sack
, int pcount
)
1316 struct tcp_sock
*tp
= tcp_sk(sk
);
1317 int fack_count
= state
->fack_count
;
1319 /* Account D-SACK for retransmitted packet. */
1320 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1321 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1322 after(end_seq
, tp
->undo_marker
))
1324 if (sacked
& TCPCB_SACKED_ACKED
)
1325 state
->reord
= min(fack_count
, state
->reord
);
1328 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1329 if (!after(end_seq
, tp
->snd_una
))
1332 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1333 if (sacked
& TCPCB_SACKED_RETRANS
) {
1334 /* If the segment is not tagged as lost,
1335 * we do not clear RETRANS, believing
1336 * that retransmission is still in flight.
1338 if (sacked
& TCPCB_LOST
) {
1339 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1340 tp
->lost_out
-= pcount
;
1341 tp
->retrans_out
-= pcount
;
1344 if (!(sacked
& TCPCB_RETRANS
)) {
1345 /* New sack for not retransmitted frame,
1346 * which was in hole. It is reordering.
1348 if (before(start_seq
,
1349 tcp_highest_sack_seq(tp
)))
1350 state
->reord
= min(fack_count
,
1353 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1354 if (!after(end_seq
, tp
->frto_highmark
))
1355 state
->flag
|= FLAG_ONLY_ORIG_SACKED
;
1358 if (sacked
& TCPCB_LOST
) {
1359 sacked
&= ~TCPCB_LOST
;
1360 tp
->lost_out
-= pcount
;
1364 sacked
|= TCPCB_SACKED_ACKED
;
1365 state
->flag
|= FLAG_DATA_SACKED
;
1366 tp
->sacked_out
+= pcount
;
1368 fack_count
+= pcount
;
1370 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1371 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1372 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1373 tp
->lost_cnt_hint
+= pcount
;
1375 if (fack_count
> tp
->fackets_out
)
1376 tp
->fackets_out
= fack_count
;
1379 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1380 * frames and clear it. undo_retrans is decreased above, L|R frames
1381 * are accounted above as well.
1383 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1384 sacked
&= ~TCPCB_SACKED_RETRANS
;
1385 tp
->retrans_out
-= pcount
;
1391 /* Shift newly-SACKed bytes from this skb to the immediately previous
1392 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1394 static int tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1395 struct tcp_sacktag_state
*state
,
1396 unsigned int pcount
, int shifted
, int mss
,
1399 struct tcp_sock
*tp
= tcp_sk(sk
);
1400 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1401 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1402 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1406 /* Adjust counters and hints for the newly sacked sequence
1407 * range but discard the return value since prev is already
1408 * marked. We must tag the range first because the seq
1409 * advancement below implicitly advances
1410 * tcp_highest_sack_seq() when skb is highest_sack.
1412 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1413 start_seq
, end_seq
, dup_sack
, pcount
);
1415 if (skb
== tp
->lost_skb_hint
)
1416 tp
->lost_cnt_hint
+= pcount
;
1418 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1419 TCP_SKB_CB(skb
)->seq
+= shifted
;
1421 skb_shinfo(prev
)->gso_segs
+= pcount
;
1422 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1423 skb_shinfo(skb
)->gso_segs
-= pcount
;
1425 /* When we're adding to gso_segs == 1, gso_size will be zero,
1426 * in theory this shouldn't be necessary but as long as DSACK
1427 * code can come after this skb later on it's better to keep
1428 * setting gso_size to something.
1430 if (!skb_shinfo(prev
)->gso_size
) {
1431 skb_shinfo(prev
)->gso_size
= mss
;
1432 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1435 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1436 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1437 skb_shinfo(skb
)->gso_size
= 0;
1438 skb_shinfo(skb
)->gso_type
= 0;
1441 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1442 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1445 BUG_ON(!tcp_skb_pcount(skb
));
1446 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1450 /* Whole SKB was eaten :-) */
1452 if (skb
== tp
->retransmit_skb_hint
)
1453 tp
->retransmit_skb_hint
= prev
;
1454 if (skb
== tp
->scoreboard_skb_hint
)
1455 tp
->scoreboard_skb_hint
= prev
;
1456 if (skb
== tp
->lost_skb_hint
) {
1457 tp
->lost_skb_hint
= prev
;
1458 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1461 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(prev
)->tcp_flags
;
1462 if (skb
== tcp_highest_sack(sk
))
1463 tcp_advance_highest_sack(sk
, skb
);
1465 tcp_unlink_write_queue(skb
, sk
);
1466 sk_wmem_free_skb(sk
, skb
);
1468 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1473 /* I wish gso_size would have a bit more sane initialization than
1474 * something-or-zero which complicates things
1476 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1478 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1481 /* Shifting pages past head area doesn't work */
1482 static int skb_can_shift(const struct sk_buff
*skb
)
1484 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1487 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1490 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1491 struct tcp_sacktag_state
*state
,
1492 u32 start_seq
, u32 end_seq
,
1495 struct tcp_sock
*tp
= tcp_sk(sk
);
1496 struct sk_buff
*prev
;
1502 if (!sk_can_gso(sk
))
1505 /* Normally R but no L won't result in plain S */
1507 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1509 if (!skb_can_shift(skb
))
1511 /* This frame is about to be dropped (was ACKed). */
1512 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1515 /* Can only happen with delayed DSACK + discard craziness */
1516 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1518 prev
= tcp_write_queue_prev(sk
, skb
);
1520 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1523 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1524 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1528 pcount
= tcp_skb_pcount(skb
);
1529 mss
= tcp_skb_seglen(skb
);
1531 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1532 * drop this restriction as unnecessary
1534 if (mss
!= tcp_skb_seglen(prev
))
1537 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1539 /* CHECKME: This is non-MSS split case only?, this will
1540 * cause skipped skbs due to advancing loop btw, original
1541 * has that feature too
1543 if (tcp_skb_pcount(skb
) <= 1)
1546 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1548 /* TODO: head merge to next could be attempted here
1549 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1550 * though it might not be worth of the additional hassle
1552 * ...we can probably just fallback to what was done
1553 * previously. We could try merging non-SACKed ones
1554 * as well but it probably isn't going to buy off
1555 * because later SACKs might again split them, and
1556 * it would make skb timestamp tracking considerably
1562 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1564 BUG_ON(len
> skb
->len
);
1566 /* MSS boundaries should be honoured or else pcount will
1567 * severely break even though it makes things bit trickier.
1568 * Optimize common case to avoid most of the divides
1570 mss
= tcp_skb_mss(skb
);
1572 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1573 * drop this restriction as unnecessary
1575 if (mss
!= tcp_skb_seglen(prev
))
1580 } else if (len
< mss
) {
1588 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1589 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1592 if (!skb_shift(prev
, skb
, len
))
1594 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1597 /* Hole filled allows collapsing with the next as well, this is very
1598 * useful when hole on every nth skb pattern happens
1600 if (prev
== tcp_write_queue_tail(sk
))
1602 skb
= tcp_write_queue_next(sk
, prev
);
1604 if (!skb_can_shift(skb
) ||
1605 (skb
== tcp_send_head(sk
)) ||
1606 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1607 (mss
!= tcp_skb_seglen(skb
)))
1611 if (skb_shift(prev
, skb
, len
)) {
1612 pcount
+= tcp_skb_pcount(skb
);
1613 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1617 state
->fack_count
+= pcount
;
1624 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1628 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1629 struct tcp_sack_block
*next_dup
,
1630 struct tcp_sacktag_state
*state
,
1631 u32 start_seq
, u32 end_seq
,
1634 struct tcp_sock
*tp
= tcp_sk(sk
);
1635 struct sk_buff
*tmp
;
1637 tcp_for_write_queue_from(skb
, sk
) {
1639 int dup_sack
= dup_sack_in
;
1641 if (skb
== tcp_send_head(sk
))
1644 /* queue is in-order => we can short-circuit the walk early */
1645 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1648 if ((next_dup
!= NULL
) &&
1649 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1650 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1651 next_dup
->start_seq
,
1657 /* skb reference here is a bit tricky to get right, since
1658 * shifting can eat and free both this skb and the next,
1659 * so not even _safe variant of the loop is enough.
1662 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1663 start_seq
, end_seq
, dup_sack
);
1672 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1678 if (unlikely(in_sack
< 0))
1682 TCP_SKB_CB(skb
)->sacked
=
1685 TCP_SKB_CB(skb
)->sacked
,
1686 TCP_SKB_CB(skb
)->seq
,
1687 TCP_SKB_CB(skb
)->end_seq
,
1689 tcp_skb_pcount(skb
));
1691 if (!before(TCP_SKB_CB(skb
)->seq
,
1692 tcp_highest_sack_seq(tp
)))
1693 tcp_advance_highest_sack(sk
, skb
);
1696 state
->fack_count
+= tcp_skb_pcount(skb
);
1701 /* Avoid all extra work that is being done by sacktag while walking in
1704 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1705 struct tcp_sacktag_state
*state
,
1708 tcp_for_write_queue_from(skb
, sk
) {
1709 if (skb
== tcp_send_head(sk
))
1712 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1715 state
->fack_count
+= tcp_skb_pcount(skb
);
1720 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1722 struct tcp_sack_block
*next_dup
,
1723 struct tcp_sacktag_state
*state
,
1726 if (next_dup
== NULL
)
1729 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1730 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1731 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1732 next_dup
->start_seq
, next_dup
->end_seq
,
1739 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1741 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1745 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1748 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1749 struct tcp_sock
*tp
= tcp_sk(sk
);
1750 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1751 TCP_SKB_CB(ack_skb
)->sacked
);
1752 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1753 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1754 struct tcp_sack_block
*cache
;
1755 struct tcp_sacktag_state state
;
1756 struct sk_buff
*skb
;
1757 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1759 int found_dup_sack
= 0;
1761 int first_sack_index
;
1764 state
.reord
= tp
->packets_out
;
1766 if (!tp
->sacked_out
) {
1767 if (WARN_ON(tp
->fackets_out
))
1768 tp
->fackets_out
= 0;
1769 tcp_highest_sack_reset(sk
);
1772 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1773 num_sacks
, prior_snd_una
);
1775 state
.flag
|= FLAG_DSACKING_ACK
;
1777 /* Eliminate too old ACKs, but take into
1778 * account more or less fresh ones, they can
1779 * contain valid SACK info.
1781 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1784 if (!tp
->packets_out
)
1788 first_sack_index
= 0;
1789 for (i
= 0; i
< num_sacks
; i
++) {
1790 int dup_sack
= !i
&& found_dup_sack
;
1792 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1793 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1795 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1796 sp
[used_sacks
].start_seq
,
1797 sp
[used_sacks
].end_seq
)) {
1801 if (!tp
->undo_marker
)
1802 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1804 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1806 /* Don't count olds caused by ACK reordering */
1807 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1808 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1810 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1813 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1815 first_sack_index
= -1;
1819 /* Ignore very old stuff early */
1820 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1826 /* order SACK blocks to allow in order walk of the retrans queue */
1827 for (i
= used_sacks
- 1; i
> 0; i
--) {
1828 for (j
= 0; j
< i
; j
++) {
1829 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1830 swap(sp
[j
], sp
[j
+ 1]);
1832 /* Track where the first SACK block goes to */
1833 if (j
== first_sack_index
)
1834 first_sack_index
= j
+ 1;
1839 skb
= tcp_write_queue_head(sk
);
1840 state
.fack_count
= 0;
1843 if (!tp
->sacked_out
) {
1844 /* It's already past, so skip checking against it */
1845 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1847 cache
= tp
->recv_sack_cache
;
1848 /* Skip empty blocks in at head of the cache */
1849 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1854 while (i
< used_sacks
) {
1855 u32 start_seq
= sp
[i
].start_seq
;
1856 u32 end_seq
= sp
[i
].end_seq
;
1857 int dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1858 struct tcp_sack_block
*next_dup
= NULL
;
1860 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1861 next_dup
= &sp
[i
+ 1];
1863 /* Skip too early cached blocks */
1864 while (tcp_sack_cache_ok(tp
, cache
) &&
1865 !before(start_seq
, cache
->end_seq
))
1868 /* Can skip some work by looking recv_sack_cache? */
1869 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1870 after(end_seq
, cache
->start_seq
)) {
1873 if (before(start_seq
, cache
->start_seq
)) {
1874 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1876 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1883 /* Rest of the block already fully processed? */
1884 if (!after(end_seq
, cache
->end_seq
))
1887 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1891 /* ...tail remains todo... */
1892 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1893 /* ...but better entrypoint exists! */
1894 skb
= tcp_highest_sack(sk
);
1897 state
.fack_count
= tp
->fackets_out
;
1902 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1903 /* Check overlap against next cached too (past this one already) */
1908 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1909 skb
= tcp_highest_sack(sk
);
1912 state
.fack_count
= tp
->fackets_out
;
1914 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1917 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1918 start_seq
, end_seq
, dup_sack
);
1921 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1922 * due to in-order walk
1924 if (after(end_seq
, tp
->frto_highmark
))
1925 state
.flag
&= ~FLAG_ONLY_ORIG_SACKED
;
1930 /* Clear the head of the cache sack blocks so we can skip it next time */
1931 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1932 tp
->recv_sack_cache
[i
].start_seq
= 0;
1933 tp
->recv_sack_cache
[i
].end_seq
= 0;
1935 for (j
= 0; j
< used_sacks
; j
++)
1936 tp
->recv_sack_cache
[i
++] = sp
[j
];
1938 tcp_mark_lost_retrans(sk
);
1940 tcp_verify_left_out(tp
);
1942 if ((state
.reord
< tp
->fackets_out
) &&
1943 ((icsk
->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
) &&
1944 (!tp
->frto_highmark
|| after(tp
->snd_una
, tp
->frto_highmark
)))
1945 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1949 #if FASTRETRANS_DEBUG > 0
1950 WARN_ON((int)tp
->sacked_out
< 0);
1951 WARN_ON((int)tp
->lost_out
< 0);
1952 WARN_ON((int)tp
->retrans_out
< 0);
1953 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1958 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1959 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
1961 static int tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1965 holes
= max(tp
->lost_out
, 1U);
1966 holes
= min(holes
, tp
->packets_out
);
1968 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1969 tp
->sacked_out
= tp
->packets_out
- holes
;
1975 /* If we receive more dupacks than we expected counting segments
1976 * in assumption of absent reordering, interpret this as reordering.
1977 * The only another reason could be bug in receiver TCP.
1979 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1981 struct tcp_sock
*tp
= tcp_sk(sk
);
1982 if (tcp_limit_reno_sacked(tp
))
1983 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1986 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1988 static void tcp_add_reno_sack(struct sock
*sk
)
1990 struct tcp_sock
*tp
= tcp_sk(sk
);
1992 tcp_check_reno_reordering(sk
, 0);
1993 tcp_verify_left_out(tp
);
1996 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1998 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
2000 struct tcp_sock
*tp
= tcp_sk(sk
);
2003 /* One ACK acked hole. The rest eat duplicate ACKs. */
2004 if (acked
- 1 >= tp
->sacked_out
)
2007 tp
->sacked_out
-= acked
- 1;
2009 tcp_check_reno_reordering(sk
, acked
);
2010 tcp_verify_left_out(tp
);
2013 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
2018 static int tcp_is_sackfrto(const struct tcp_sock
*tp
)
2020 return (sysctl_tcp_frto
== 0x2) && !tcp_is_reno(tp
);
2023 /* F-RTO can only be used if TCP has never retransmitted anything other than
2024 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2026 int tcp_use_frto(struct sock
*sk
)
2028 const struct tcp_sock
*tp
= tcp_sk(sk
);
2029 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2030 struct sk_buff
*skb
;
2032 if (!sysctl_tcp_frto
)
2035 /* MTU probe and F-RTO won't really play nicely along currently */
2036 if (icsk
->icsk_mtup
.probe_size
)
2039 if (tcp_is_sackfrto(tp
))
2042 /* Avoid expensive walking of rexmit queue if possible */
2043 if (tp
->retrans_out
> 1)
2046 skb
= tcp_write_queue_head(sk
);
2047 if (tcp_skb_is_last(sk
, skb
))
2049 skb
= tcp_write_queue_next(sk
, skb
); /* Skips head */
2050 tcp_for_write_queue_from(skb
, sk
) {
2051 if (skb
== tcp_send_head(sk
))
2053 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2055 /* Short-circuit when first non-SACKed skb has been checked */
2056 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2062 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2063 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2064 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2065 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2066 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2067 * bits are handled if the Loss state is really to be entered (in
2068 * tcp_enter_frto_loss).
2070 * Do like tcp_enter_loss() would; when RTO expires the second time it
2072 * "Reduce ssthresh if it has not yet been made inside this window."
2074 void tcp_enter_frto(struct sock
*sk
)
2076 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2077 struct tcp_sock
*tp
= tcp_sk(sk
);
2078 struct sk_buff
*skb
;
2080 if ((!tp
->frto_counter
&& icsk
->icsk_ca_state
<= TCP_CA_Disorder
) ||
2081 tp
->snd_una
== tp
->high_seq
||
2082 ((icsk
->icsk_ca_state
== TCP_CA_Loss
|| tp
->frto_counter
) &&
2083 !icsk
->icsk_retransmits
)) {
2084 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2085 /* Our state is too optimistic in ssthresh() call because cwnd
2086 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2087 * recovery has not yet completed. Pattern would be this: RTO,
2088 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2090 * RFC4138 should be more specific on what to do, even though
2091 * RTO is quite unlikely to occur after the first Cumulative ACK
2092 * due to back-off and complexity of triggering events ...
2094 if (tp
->frto_counter
) {
2096 stored_cwnd
= tp
->snd_cwnd
;
2098 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2099 tp
->snd_cwnd
= stored_cwnd
;
2101 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2103 /* ... in theory, cong.control module could do "any tricks" in
2104 * ssthresh(), which means that ca_state, lost bits and lost_out
2105 * counter would have to be faked before the call occurs. We
2106 * consider that too expensive, unlikely and hacky, so modules
2107 * using these in ssthresh() must deal these incompatibility
2108 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2110 tcp_ca_event(sk
, CA_EVENT_FRTO
);
2113 tp
->undo_marker
= tp
->snd_una
;
2114 tp
->undo_retrans
= 0;
2116 skb
= tcp_write_queue_head(sk
);
2117 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2118 tp
->undo_marker
= 0;
2119 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2120 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2121 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2123 tcp_verify_left_out(tp
);
2125 /* Too bad if TCP was application limited */
2126 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2128 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2129 * The last condition is necessary at least in tp->frto_counter case.
2131 if (tcp_is_sackfrto(tp
) && (tp
->frto_counter
||
2132 ((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Recovery
|TCPF_CA_Loss
))) &&
2133 after(tp
->high_seq
, tp
->snd_una
)) {
2134 tp
->frto_highmark
= tp
->high_seq
;
2136 tp
->frto_highmark
= tp
->snd_nxt
;
2138 tcp_set_ca_state(sk
, TCP_CA_Disorder
);
2139 tp
->high_seq
= tp
->snd_nxt
;
2140 tp
->frto_counter
= 1;
2143 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2144 * which indicates that we should follow the traditional RTO recovery,
2145 * i.e. mark everything lost and do go-back-N retransmission.
2147 static void tcp_enter_frto_loss(struct sock
*sk
, int allowed_segments
, int flag
)
2149 struct tcp_sock
*tp
= tcp_sk(sk
);
2150 struct sk_buff
*skb
;
2153 tp
->retrans_out
= 0;
2154 if (tcp_is_reno(tp
))
2155 tcp_reset_reno_sack(tp
);
2157 tcp_for_write_queue(skb
, sk
) {
2158 if (skb
== tcp_send_head(sk
))
2161 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2163 * Count the retransmission made on RTO correctly (only when
2164 * waiting for the first ACK and did not get it)...
2166 if ((tp
->frto_counter
== 1) && !(flag
& FLAG_DATA_ACKED
)) {
2167 /* For some reason this R-bit might get cleared? */
2168 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
2169 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2170 /* ...enter this if branch just for the first segment */
2171 flag
|= FLAG_DATA_ACKED
;
2173 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2174 tp
->undo_marker
= 0;
2175 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2178 /* Marking forward transmissions that were made after RTO lost
2179 * can cause unnecessary retransmissions in some scenarios,
2180 * SACK blocks will mitigate that in some but not in all cases.
2181 * We used to not mark them but it was causing break-ups with
2182 * receivers that do only in-order receival.
2184 * TODO: we could detect presence of such receiver and select
2185 * different behavior per flow.
2187 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2188 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2189 tp
->lost_out
+= tcp_skb_pcount(skb
);
2190 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2193 tcp_verify_left_out(tp
);
2195 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + allowed_segments
;
2196 tp
->snd_cwnd_cnt
= 0;
2197 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2198 tp
->frto_counter
= 0;
2199 tp
->bytes_acked
= 0;
2201 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2202 sysctl_tcp_reordering
);
2203 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2204 tp
->high_seq
= tp
->snd_nxt
;
2205 TCP_ECN_queue_cwr(tp
);
2207 tcp_clear_all_retrans_hints(tp
);
2210 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
2212 tp
->retrans_out
= 0;
2215 tp
->undo_marker
= 0;
2216 tp
->undo_retrans
= 0;
2219 void tcp_clear_retrans(struct tcp_sock
*tp
)
2221 tcp_clear_retrans_partial(tp
);
2223 tp
->fackets_out
= 0;
2227 /* Enter Loss state. If "how" is not zero, forget all SACK information
2228 * and reset tags completely, otherwise preserve SACKs. If receiver
2229 * dropped its ofo queue, we will know this due to reneging detection.
2231 void tcp_enter_loss(struct sock
*sk
, int how
)
2233 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2234 struct tcp_sock
*tp
= tcp_sk(sk
);
2235 struct sk_buff
*skb
;
2237 /* Reduce ssthresh if it has not yet been made inside this window. */
2238 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
|| tp
->snd_una
== tp
->high_seq
||
2239 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
2240 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2241 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
2242 tcp_ca_event(sk
, CA_EVENT_LOSS
);
2245 tp
->snd_cwnd_cnt
= 0;
2246 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2248 tp
->bytes_acked
= 0;
2249 tcp_clear_retrans_partial(tp
);
2251 if (tcp_is_reno(tp
))
2252 tcp_reset_reno_sack(tp
);
2255 /* Push undo marker, if it was plain RTO and nothing
2256 * was retransmitted. */
2257 tp
->undo_marker
= tp
->snd_una
;
2260 tp
->fackets_out
= 0;
2262 tcp_clear_all_retrans_hints(tp
);
2264 tcp_for_write_queue(skb
, sk
) {
2265 if (skb
== tcp_send_head(sk
))
2268 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
2269 tp
->undo_marker
= 0;
2270 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
2271 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
2272 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
2273 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
2274 tp
->lost_out
+= tcp_skb_pcount(skb
);
2275 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
2278 tcp_verify_left_out(tp
);
2280 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
2281 sysctl_tcp_reordering
);
2282 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2283 tp
->high_seq
= tp
->snd_nxt
;
2284 TCP_ECN_queue_cwr(tp
);
2285 /* Abort F-RTO algorithm if one is in progress */
2286 tp
->frto_counter
= 0;
2289 /* If ACK arrived pointing to a remembered SACK, it means that our
2290 * remembered SACKs do not reflect real state of receiver i.e.
2291 * receiver _host_ is heavily congested (or buggy).
2293 * Do processing similar to RTO timeout.
2295 static int tcp_check_sack_reneging(struct sock
*sk
, int flag
)
2297 if (flag
& FLAG_SACK_RENEGING
) {
2298 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2299 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
2301 tcp_enter_loss(sk
, 1);
2302 icsk
->icsk_retransmits
++;
2303 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
2304 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2305 icsk
->icsk_rto
, TCP_RTO_MAX
);
2311 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
2313 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
2316 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2317 * counter when SACK is enabled (without SACK, sacked_out is used for
2320 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2321 * segments up to the highest received SACK block so far and holes in
2324 * With reordering, holes may still be in flight, so RFC3517 recovery
2325 * uses pure sacked_out (total number of SACKed segments) even though
2326 * it violates the RFC that uses duplicate ACKs, often these are equal
2327 * but when e.g. out-of-window ACKs or packet duplication occurs,
2328 * they differ. Since neither occurs due to loss, TCP should really
2331 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
2333 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
2336 static inline int tcp_skb_timedout(const struct sock
*sk
,
2337 const struct sk_buff
*skb
)
2339 return tcp_time_stamp
- TCP_SKB_CB(skb
)->when
> inet_csk(sk
)->icsk_rto
;
2342 static inline int tcp_head_timedout(const struct sock
*sk
)
2344 const struct tcp_sock
*tp
= tcp_sk(sk
);
2346 return tp
->packets_out
&&
2347 tcp_skb_timedout(sk
, tcp_write_queue_head(sk
));
2350 /* Linux NewReno/SACK/FACK/ECN state machine.
2351 * --------------------------------------
2353 * "Open" Normal state, no dubious events, fast path.
2354 * "Disorder" In all the respects it is "Open",
2355 * but requires a bit more attention. It is entered when
2356 * we see some SACKs or dupacks. It is split of "Open"
2357 * mainly to move some processing from fast path to slow one.
2358 * "CWR" CWND was reduced due to some Congestion Notification event.
2359 * It can be ECN, ICMP source quench, local device congestion.
2360 * "Recovery" CWND was reduced, we are fast-retransmitting.
2361 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2363 * tcp_fastretrans_alert() is entered:
2364 * - each incoming ACK, if state is not "Open"
2365 * - when arrived ACK is unusual, namely:
2370 * Counting packets in flight is pretty simple.
2372 * in_flight = packets_out - left_out + retrans_out
2374 * packets_out is SND.NXT-SND.UNA counted in packets.
2376 * retrans_out is number of retransmitted segments.
2378 * left_out is number of segments left network, but not ACKed yet.
2380 * left_out = sacked_out + lost_out
2382 * sacked_out: Packets, which arrived to receiver out of order
2383 * and hence not ACKed. With SACKs this number is simply
2384 * amount of SACKed data. Even without SACKs
2385 * it is easy to give pretty reliable estimate of this number,
2386 * counting duplicate ACKs.
2388 * lost_out: Packets lost by network. TCP has no explicit
2389 * "loss notification" feedback from network (for now).
2390 * It means that this number can be only _guessed_.
2391 * Actually, it is the heuristics to predict lossage that
2392 * distinguishes different algorithms.
2394 * F.e. after RTO, when all the queue is considered as lost,
2395 * lost_out = packets_out and in_flight = retrans_out.
2397 * Essentially, we have now two algorithms counting
2400 * FACK: It is the simplest heuristics. As soon as we decided
2401 * that something is lost, we decide that _all_ not SACKed
2402 * packets until the most forward SACK are lost. I.e.
2403 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2404 * It is absolutely correct estimate, if network does not reorder
2405 * packets. And it loses any connection to reality when reordering
2406 * takes place. We use FACK by default until reordering
2407 * is suspected on the path to this destination.
2409 * NewReno: when Recovery is entered, we assume that one segment
2410 * is lost (classic Reno). While we are in Recovery and
2411 * a partial ACK arrives, we assume that one more packet
2412 * is lost (NewReno). This heuristics are the same in NewReno
2415 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2416 * deflation etc. CWND is real congestion window, never inflated, changes
2417 * only according to classic VJ rules.
2419 * Really tricky (and requiring careful tuning) part of algorithm
2420 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2421 * The first determines the moment _when_ we should reduce CWND and,
2422 * hence, slow down forward transmission. In fact, it determines the moment
2423 * when we decide that hole is caused by loss, rather than by a reorder.
2425 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2426 * holes, caused by lost packets.
2428 * And the most logically complicated part of algorithm is undo
2429 * heuristics. We detect false retransmits due to both too early
2430 * fast retransmit (reordering) and underestimated RTO, analyzing
2431 * timestamps and D-SACKs. When we detect that some segments were
2432 * retransmitted by mistake and CWND reduction was wrong, we undo
2433 * window reduction and abort recovery phase. This logic is hidden
2434 * inside several functions named tcp_try_undo_<something>.
2437 /* This function decides, when we should leave Disordered state
2438 * and enter Recovery phase, reducing congestion window.
2440 * Main question: may we further continue forward transmission
2441 * with the same cwnd?
2443 static int tcp_time_to_recover(struct sock
*sk
)
2445 struct tcp_sock
*tp
= tcp_sk(sk
);
2448 /* Do not perform any recovery during F-RTO algorithm */
2449 if (tp
->frto_counter
)
2452 /* Trick#1: The loss is proven. */
2456 /* Not-A-Trick#2 : Classic rule... */
2457 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2460 /* Trick#3 : when we use RFC2988 timer restart, fast
2461 * retransmit can be triggered by timeout of queue head.
2463 if (tcp_is_fack(tp
) && tcp_head_timedout(sk
))
2466 /* Trick#4: It is still not OK... But will it be useful to delay
2469 packets_out
= tp
->packets_out
;
2470 if (packets_out
<= tp
->reordering
&&
2471 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2472 !tcp_may_send_now(sk
)) {
2473 /* We have nothing to send. This connection is limited
2474 * either by receiver window or by application.
2479 /* If a thin stream is detected, retransmit after first
2480 * received dupack. Employ only if SACK is supported in order
2481 * to avoid possible corner-case series of spurious retransmissions
2482 * Use only if there are no unsent data.
2484 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2485 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2486 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2492 /* New heuristics: it is possible only after we switched to restart timer
2493 * each time when something is ACKed. Hence, we can detect timed out packets
2494 * during fast retransmit without falling to slow start.
2496 * Usefulness of this as is very questionable, since we should know which of
2497 * the segments is the next to timeout which is relatively expensive to find
2498 * in general case unless we add some data structure just for that. The
2499 * current approach certainly won't find the right one too often and when it
2500 * finally does find _something_ it usually marks large part of the window
2501 * right away (because a retransmission with a larger timestamp blocks the
2502 * loop from advancing). -ij
2504 static void tcp_timeout_skbs(struct sock
*sk
)
2506 struct tcp_sock
*tp
= tcp_sk(sk
);
2507 struct sk_buff
*skb
;
2509 if (!tcp_is_fack(tp
) || !tcp_head_timedout(sk
))
2512 skb
= tp
->scoreboard_skb_hint
;
2513 if (tp
->scoreboard_skb_hint
== NULL
)
2514 skb
= tcp_write_queue_head(sk
);
2516 tcp_for_write_queue_from(skb
, sk
) {
2517 if (skb
== tcp_send_head(sk
))
2519 if (!tcp_skb_timedout(sk
, skb
))
2522 tcp_skb_mark_lost(tp
, skb
);
2525 tp
->scoreboard_skb_hint
= skb
;
2527 tcp_verify_left_out(tp
);
2530 /* Detect loss in event "A" above by marking head of queue up as lost.
2531 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2532 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2533 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2534 * the maximum SACKed segments to pass before reaching this limit.
2536 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2538 struct tcp_sock
*tp
= tcp_sk(sk
);
2539 struct sk_buff
*skb
;
2543 /* Use SACK to deduce losses of new sequences sent during recovery */
2544 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2546 WARN_ON(packets
> tp
->packets_out
);
2547 if (tp
->lost_skb_hint
) {
2548 skb
= tp
->lost_skb_hint
;
2549 cnt
= tp
->lost_cnt_hint
;
2550 /* Head already handled? */
2551 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2554 skb
= tcp_write_queue_head(sk
);
2558 tcp_for_write_queue_from(skb
, sk
) {
2559 if (skb
== tcp_send_head(sk
))
2561 /* TODO: do this better */
2562 /* this is not the most efficient way to do this... */
2563 tp
->lost_skb_hint
= skb
;
2564 tp
->lost_cnt_hint
= cnt
;
2566 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2570 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2571 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2572 cnt
+= tcp_skb_pcount(skb
);
2574 if (cnt
> packets
) {
2575 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2576 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2577 (oldcnt
>= packets
))
2580 mss
= skb_shinfo(skb
)->gso_size
;
2581 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2587 tcp_skb_mark_lost(tp
, skb
);
2592 tcp_verify_left_out(tp
);
2595 /* Account newly detected lost packet(s) */
2597 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2599 struct tcp_sock
*tp
= tcp_sk(sk
);
2601 if (tcp_is_reno(tp
)) {
2602 tcp_mark_head_lost(sk
, 1, 1);
2603 } else if (tcp_is_fack(tp
)) {
2604 int lost
= tp
->fackets_out
- tp
->reordering
;
2607 tcp_mark_head_lost(sk
, lost
, 0);
2609 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2610 if (sacked_upto
>= 0)
2611 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2612 else if (fast_rexmit
)
2613 tcp_mark_head_lost(sk
, 1, 1);
2616 tcp_timeout_skbs(sk
);
2619 /* CWND moderation, preventing bursts due to too big ACKs
2620 * in dubious situations.
2622 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2624 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2625 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2626 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2629 /* Lower bound on congestion window is slow start threshold
2630 * unless congestion avoidance choice decides to overide it.
2632 static inline u32
tcp_cwnd_min(const struct sock
*sk
)
2634 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
2636 return ca_ops
->min_cwnd
? ca_ops
->min_cwnd(sk
) : tcp_sk(sk
)->snd_ssthresh
;
2639 /* Decrease cwnd each second ack. */
2640 static void tcp_cwnd_down(struct sock
*sk
, int flag
)
2642 struct tcp_sock
*tp
= tcp_sk(sk
);
2643 int decr
= tp
->snd_cwnd_cnt
+ 1;
2645 if ((flag
& (FLAG_ANY_PROGRESS
| FLAG_DSACKING_ACK
)) ||
2646 (tcp_is_reno(tp
) && !(flag
& FLAG_NOT_DUP
))) {
2647 tp
->snd_cwnd_cnt
= decr
& 1;
2650 if (decr
&& tp
->snd_cwnd
> tcp_cwnd_min(sk
))
2651 tp
->snd_cwnd
-= decr
;
2653 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tcp_packets_in_flight(tp
) + 1);
2654 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2658 /* Nothing was retransmitted or returned timestamp is less
2659 * than timestamp of the first retransmission.
2661 static inline int tcp_packet_delayed(const struct tcp_sock
*tp
)
2663 return !tp
->retrans_stamp
||
2664 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2665 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2668 /* Undo procedures. */
2670 #if FASTRETRANS_DEBUG > 1
2671 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2673 struct tcp_sock
*tp
= tcp_sk(sk
);
2674 struct inet_sock
*inet
= inet_sk(sk
);
2676 if (sk
->sk_family
== AF_INET
) {
2677 printk(KERN_DEBUG
"Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2679 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2680 tp
->snd_cwnd
, tcp_left_out(tp
),
2681 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2684 #if IS_ENABLED(CONFIG_IPV6)
2685 else if (sk
->sk_family
== AF_INET6
) {
2686 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2687 printk(KERN_DEBUG
"Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2689 &np
->daddr
, ntohs(inet
->inet_dport
),
2690 tp
->snd_cwnd
, tcp_left_out(tp
),
2691 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2697 #define DBGUNDO(x...) do { } while (0)
2700 static void tcp_undo_cwr(struct sock
*sk
, const bool undo_ssthresh
)
2702 struct tcp_sock
*tp
= tcp_sk(sk
);
2704 if (tp
->prior_ssthresh
) {
2705 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2707 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2708 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2710 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2712 if (undo_ssthresh
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2713 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2714 TCP_ECN_withdraw_cwr(tp
);
2717 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2719 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2722 static inline int tcp_may_undo(const struct tcp_sock
*tp
)
2724 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2727 /* People celebrate: "We love our President!" */
2728 static int tcp_try_undo_recovery(struct sock
*sk
)
2730 struct tcp_sock
*tp
= tcp_sk(sk
);
2732 if (tcp_may_undo(tp
)) {
2735 /* Happy end! We did not retransmit anything
2736 * or our original transmission succeeded.
2738 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2739 tcp_undo_cwr(sk
, true);
2740 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2741 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2743 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2745 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2746 tp
->undo_marker
= 0;
2748 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2749 /* Hold old state until something *above* high_seq
2750 * is ACKed. For Reno it is MUST to prevent false
2751 * fast retransmits (RFC2582). SACK TCP is safe. */
2752 tcp_moderate_cwnd(tp
);
2755 tcp_set_ca_state(sk
, TCP_CA_Open
);
2759 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2760 static void tcp_try_undo_dsack(struct sock
*sk
)
2762 struct tcp_sock
*tp
= tcp_sk(sk
);
2764 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2765 DBGUNDO(sk
, "D-SACK");
2766 tcp_undo_cwr(sk
, true);
2767 tp
->undo_marker
= 0;
2768 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2772 /* We can clear retrans_stamp when there are no retransmissions in the
2773 * window. It would seem that it is trivially available for us in
2774 * tp->retrans_out, however, that kind of assumptions doesn't consider
2775 * what will happen if errors occur when sending retransmission for the
2776 * second time. ...It could the that such segment has only
2777 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2778 * the head skb is enough except for some reneging corner cases that
2779 * are not worth the effort.
2781 * Main reason for all this complexity is the fact that connection dying
2782 * time now depends on the validity of the retrans_stamp, in particular,
2783 * that successive retransmissions of a segment must not advance
2784 * retrans_stamp under any conditions.
2786 static int tcp_any_retrans_done(const struct sock
*sk
)
2788 const struct tcp_sock
*tp
= tcp_sk(sk
);
2789 struct sk_buff
*skb
;
2791 if (tp
->retrans_out
)
2794 skb
= tcp_write_queue_head(sk
);
2795 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2801 /* Undo during fast recovery after partial ACK. */
2803 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2805 struct tcp_sock
*tp
= tcp_sk(sk
);
2806 /* Partial ACK arrived. Force Hoe's retransmit. */
2807 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2809 if (tcp_may_undo(tp
)) {
2810 /* Plain luck! Hole if filled with delayed
2811 * packet, rather than with a retransmit.
2813 if (!tcp_any_retrans_done(sk
))
2814 tp
->retrans_stamp
= 0;
2816 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2819 tcp_undo_cwr(sk
, false);
2820 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2822 /* So... Do not make Hoe's retransmit yet.
2823 * If the first packet was delayed, the rest
2824 * ones are most probably delayed as well.
2831 /* Undo during loss recovery after partial ACK. */
2832 static int tcp_try_undo_loss(struct sock
*sk
)
2834 struct tcp_sock
*tp
= tcp_sk(sk
);
2836 if (tcp_may_undo(tp
)) {
2837 struct sk_buff
*skb
;
2838 tcp_for_write_queue(skb
, sk
) {
2839 if (skb
== tcp_send_head(sk
))
2841 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2844 tcp_clear_all_retrans_hints(tp
);
2846 DBGUNDO(sk
, "partial loss");
2848 tcp_undo_cwr(sk
, true);
2849 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2850 inet_csk(sk
)->icsk_retransmits
= 0;
2851 tp
->undo_marker
= 0;
2852 if (tcp_is_sack(tp
))
2853 tcp_set_ca_state(sk
, TCP_CA_Open
);
2859 static inline void tcp_complete_cwr(struct sock
*sk
)
2861 struct tcp_sock
*tp
= tcp_sk(sk
);
2863 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2864 if (tp
->undo_marker
) {
2865 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
)
2866 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2868 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2869 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2871 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2874 static void tcp_try_keep_open(struct sock
*sk
)
2876 struct tcp_sock
*tp
= tcp_sk(sk
);
2877 int state
= TCP_CA_Open
;
2879 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2880 state
= TCP_CA_Disorder
;
2882 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2883 tcp_set_ca_state(sk
, state
);
2884 tp
->high_seq
= tp
->snd_nxt
;
2888 static void tcp_try_to_open(struct sock
*sk
, int flag
)
2890 struct tcp_sock
*tp
= tcp_sk(sk
);
2892 tcp_verify_left_out(tp
);
2894 if (!tp
->frto_counter
&& !tcp_any_retrans_done(sk
))
2895 tp
->retrans_stamp
= 0;
2897 if (flag
& FLAG_ECE
)
2898 tcp_enter_cwr(sk
, 1);
2900 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2901 tcp_try_keep_open(sk
);
2902 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2903 tcp_moderate_cwnd(tp
);
2905 tcp_cwnd_down(sk
, flag
);
2909 static void tcp_mtup_probe_failed(struct sock
*sk
)
2911 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2913 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2914 icsk
->icsk_mtup
.probe_size
= 0;
2917 static void tcp_mtup_probe_success(struct sock
*sk
)
2919 struct tcp_sock
*tp
= tcp_sk(sk
);
2920 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2922 /* FIXME: breaks with very large cwnd */
2923 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2924 tp
->snd_cwnd
= tp
->snd_cwnd
*
2925 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2926 icsk
->icsk_mtup
.probe_size
;
2927 tp
->snd_cwnd_cnt
= 0;
2928 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2929 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2931 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2932 icsk
->icsk_mtup
.probe_size
= 0;
2933 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2936 /* Do a simple retransmit without using the backoff mechanisms in
2937 * tcp_timer. This is used for path mtu discovery.
2938 * The socket is already locked here.
2940 void tcp_simple_retransmit(struct sock
*sk
)
2942 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2943 struct tcp_sock
*tp
= tcp_sk(sk
);
2944 struct sk_buff
*skb
;
2945 unsigned int mss
= tcp_current_mss(sk
);
2946 u32 prior_lost
= tp
->lost_out
;
2948 tcp_for_write_queue(skb
, sk
) {
2949 if (skb
== tcp_send_head(sk
))
2951 if (tcp_skb_seglen(skb
) > mss
&&
2952 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2953 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2954 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2955 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2957 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2961 tcp_clear_retrans_hints_partial(tp
);
2963 if (prior_lost
== tp
->lost_out
)
2966 if (tcp_is_reno(tp
))
2967 tcp_limit_reno_sacked(tp
);
2969 tcp_verify_left_out(tp
);
2971 /* Don't muck with the congestion window here.
2972 * Reason is that we do not increase amount of _data_
2973 * in network, but units changed and effective
2974 * cwnd/ssthresh really reduced now.
2976 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2977 tp
->high_seq
= tp
->snd_nxt
;
2978 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2979 tp
->prior_ssthresh
= 0;
2980 tp
->undo_marker
= 0;
2981 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2983 tcp_xmit_retransmit_queue(sk
);
2985 EXPORT_SYMBOL(tcp_simple_retransmit
);
2987 /* This function implements the PRR algorithm, specifcally the PRR-SSRB
2988 * (proportional rate reduction with slow start reduction bound) as described in
2989 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
2990 * It computes the number of packets to send (sndcnt) based on packets newly
2992 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2993 * cwnd reductions across a full RTT.
2994 * 2) If packets in flight is lower than ssthresh (such as due to excess
2995 * losses and/or application stalls), do not perform any further cwnd
2996 * reductions, but instead slow start up to ssthresh.
2998 static void tcp_update_cwnd_in_recovery(struct sock
*sk
, int newly_acked_sacked
,
2999 int fast_rexmit
, int flag
)
3001 struct tcp_sock
*tp
= tcp_sk(sk
);
3003 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
3005 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
3006 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
3008 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
3010 sndcnt
= min_t(int, delta
,
3011 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
3012 newly_acked_sacked
) + 1);
3015 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
3016 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
3019 /* Process an event, which can update packets-in-flight not trivially.
3020 * Main goal of this function is to calculate new estimate for left_out,
3021 * taking into account both packets sitting in receiver's buffer and
3022 * packets lost by network.
3024 * Besides that it does CWND reduction, when packet loss is detected
3025 * and changes state of machine.
3027 * It does _not_ decide what to send, it is made in function
3028 * tcp_xmit_retransmit_queue().
3030 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
,
3031 int newly_acked_sacked
, bool is_dupack
,
3034 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3035 struct tcp_sock
*tp
= tcp_sk(sk
);
3036 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
3037 (tcp_fackets_out(tp
) > tp
->reordering
));
3038 int fast_rexmit
= 0, mib_idx
;
3040 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
3042 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
3043 tp
->fackets_out
= 0;
3045 /* Now state machine starts.
3046 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3047 if (flag
& FLAG_ECE
)
3048 tp
->prior_ssthresh
= 0;
3050 /* B. In all the states check for reneging SACKs. */
3051 if (tcp_check_sack_reneging(sk
, flag
))
3054 /* C. Check consistency of the current state. */
3055 tcp_verify_left_out(tp
);
3057 /* D. Check state exit conditions. State can be terminated
3058 * when high_seq is ACKed. */
3059 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
3060 WARN_ON(tp
->retrans_out
!= 0);
3061 tp
->retrans_stamp
= 0;
3062 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
3063 switch (icsk
->icsk_ca_state
) {
3065 icsk
->icsk_retransmits
= 0;
3066 if (tcp_try_undo_recovery(sk
))
3071 /* CWR is to be held something *above* high_seq
3072 * is ACKed for CWR bit to reach receiver. */
3073 if (tp
->snd_una
!= tp
->high_seq
) {
3074 tcp_complete_cwr(sk
);
3075 tcp_set_ca_state(sk
, TCP_CA_Open
);
3079 case TCP_CA_Recovery
:
3080 if (tcp_is_reno(tp
))
3081 tcp_reset_reno_sack(tp
);
3082 if (tcp_try_undo_recovery(sk
))
3084 tcp_complete_cwr(sk
);
3089 /* E. Process state. */
3090 switch (icsk
->icsk_ca_state
) {
3091 case TCP_CA_Recovery
:
3092 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
3093 if (tcp_is_reno(tp
) && is_dupack
)
3094 tcp_add_reno_sack(sk
);
3096 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
3099 if (flag
& FLAG_DATA_ACKED
)
3100 icsk
->icsk_retransmits
= 0;
3101 if (tcp_is_reno(tp
) && flag
& FLAG_SND_UNA_ADVANCED
)
3102 tcp_reset_reno_sack(tp
);
3103 if (!tcp_try_undo_loss(sk
)) {
3104 tcp_moderate_cwnd(tp
);
3105 tcp_xmit_retransmit_queue(sk
);
3108 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
3110 /* Loss is undone; fall through to processing in Open state. */
3112 if (tcp_is_reno(tp
)) {
3113 if (flag
& FLAG_SND_UNA_ADVANCED
)
3114 tcp_reset_reno_sack(tp
);
3116 tcp_add_reno_sack(sk
);
3119 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
3120 tcp_try_undo_dsack(sk
);
3122 if (!tcp_time_to_recover(sk
)) {
3123 tcp_try_to_open(sk
, flag
);
3127 /* MTU probe failure: don't reduce cwnd */
3128 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
3129 icsk
->icsk_mtup
.probe_size
&&
3130 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
3131 tcp_mtup_probe_failed(sk
);
3132 /* Restores the reduction we did in tcp_mtup_probe() */
3134 tcp_simple_retransmit(sk
);
3138 /* Otherwise enter Recovery state */
3140 if (tcp_is_reno(tp
))
3141 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
3143 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
3145 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3147 tp
->high_seq
= tp
->snd_nxt
;
3148 tp
->prior_ssthresh
= 0;
3149 tp
->undo_marker
= tp
->snd_una
;
3150 tp
->undo_retrans
= tp
->retrans_out
;
3152 if (icsk
->icsk_ca_state
< TCP_CA_CWR
) {
3153 if (!(flag
& FLAG_ECE
))
3154 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
3155 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
3156 TCP_ECN_queue_cwr(tp
);
3159 tp
->bytes_acked
= 0;
3160 tp
->snd_cwnd_cnt
= 0;
3161 tp
->prior_cwnd
= tp
->snd_cwnd
;
3162 tp
->prr_delivered
= 0;
3164 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
3168 if (do_lost
|| (tcp_is_fack(tp
) && tcp_head_timedout(sk
)))
3169 tcp_update_scoreboard(sk
, fast_rexmit
);
3170 tp
->prr_delivered
+= newly_acked_sacked
;
3171 tcp_update_cwnd_in_recovery(sk
, newly_acked_sacked
, fast_rexmit
, flag
);
3172 tcp_xmit_retransmit_queue(sk
);
3175 void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
3177 tcp_rtt_estimator(sk
, seq_rtt
);
3179 inet_csk(sk
)->icsk_backoff
= 0;
3181 EXPORT_SYMBOL(tcp_valid_rtt_meas
);
3183 /* Read draft-ietf-tcplw-high-performance before mucking
3184 * with this code. (Supersedes RFC1323)
3186 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
3188 /* RTTM Rule: A TSecr value received in a segment is used to
3189 * update the averaged RTT measurement only if the segment
3190 * acknowledges some new data, i.e., only if it advances the
3191 * left edge of the send window.
3193 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3194 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3196 * Changed: reset backoff as soon as we see the first valid sample.
3197 * If we do not, we get strongly overestimated rto. With timestamps
3198 * samples are accepted even from very old segments: f.e., when rtt=1
3199 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3200 * answer arrives rto becomes 120 seconds! If at least one of segments
3201 * in window is lost... Voila. --ANK (010210)
3203 struct tcp_sock
*tp
= tcp_sk(sk
);
3205 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
3208 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
3210 /* We don't have a timestamp. Can only use
3211 * packets that are not retransmitted to determine
3212 * rtt estimates. Also, we must not reset the
3213 * backoff for rto until we get a non-retransmitted
3214 * packet. This allows us to deal with a situation
3215 * where the network delay has increased suddenly.
3216 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3219 if (flag
& FLAG_RETRANS_DATA_ACKED
)
3222 tcp_valid_rtt_meas(sk
, seq_rtt
);
3225 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
3228 const struct tcp_sock
*tp
= tcp_sk(sk
);
3229 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3230 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
3231 tcp_ack_saw_tstamp(sk
, flag
);
3232 else if (seq_rtt
>= 0)
3233 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
3236 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
3238 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3239 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
3240 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
3243 /* Restart timer after forward progress on connection.
3244 * RFC2988 recommends to restart timer to now+rto.
3246 static void tcp_rearm_rto(struct sock
*sk
)
3248 const struct tcp_sock
*tp
= tcp_sk(sk
);
3250 if (!tp
->packets_out
) {
3251 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
3253 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3254 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3258 /* If we get here, the whole TSO packet has not been acked. */
3259 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
3261 struct tcp_sock
*tp
= tcp_sk(sk
);
3264 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
3266 packets_acked
= tcp_skb_pcount(skb
);
3267 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3269 packets_acked
-= tcp_skb_pcount(skb
);
3271 if (packets_acked
) {
3272 BUG_ON(tcp_skb_pcount(skb
) == 0);
3273 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
3276 return packets_acked
;
3279 /* Remove acknowledged frames from the retransmission queue. If our packet
3280 * is before the ack sequence we can discard it as it's confirmed to have
3281 * arrived at the other end.
3283 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
3286 struct tcp_sock
*tp
= tcp_sk(sk
);
3287 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3288 struct sk_buff
*skb
;
3289 u32 now
= tcp_time_stamp
;
3290 int fully_acked
= 1;
3293 u32 reord
= tp
->packets_out
;
3294 u32 prior_sacked
= tp
->sacked_out
;
3296 s32 ca_seq_rtt
= -1;
3297 ktime_t last_ackt
= net_invalid_timestamp();
3299 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
3300 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
3302 u8 sacked
= scb
->sacked
;
3304 /* Determine how many packets and what bytes were acked, tso and else */
3305 if (after(scb
->end_seq
, tp
->snd_una
)) {
3306 if (tcp_skb_pcount(skb
) == 1 ||
3307 !after(tp
->snd_una
, scb
->seq
))
3310 acked_pcount
= tcp_tso_acked(sk
, skb
);
3316 acked_pcount
= tcp_skb_pcount(skb
);
3319 if (sacked
& TCPCB_RETRANS
) {
3320 if (sacked
& TCPCB_SACKED_RETRANS
)
3321 tp
->retrans_out
-= acked_pcount
;
3322 flag
|= FLAG_RETRANS_DATA_ACKED
;
3325 if ((flag
& FLAG_DATA_ACKED
) || (acked_pcount
> 1))
3326 flag
|= FLAG_NONHEAD_RETRANS_ACKED
;
3328 ca_seq_rtt
= now
- scb
->when
;
3329 last_ackt
= skb
->tstamp
;
3331 seq_rtt
= ca_seq_rtt
;
3333 if (!(sacked
& TCPCB_SACKED_ACKED
))
3334 reord
= min(pkts_acked
, reord
);
3337 if (sacked
& TCPCB_SACKED_ACKED
)
3338 tp
->sacked_out
-= acked_pcount
;
3339 if (sacked
& TCPCB_LOST
)
3340 tp
->lost_out
-= acked_pcount
;
3342 tp
->packets_out
-= acked_pcount
;
3343 pkts_acked
+= acked_pcount
;
3345 /* Initial outgoing SYN's get put onto the write_queue
3346 * just like anything else we transmit. It is not
3347 * true data, and if we misinform our callers that
3348 * this ACK acks real data, we will erroneously exit
3349 * connection startup slow start one packet too
3350 * quickly. This is severely frowned upon behavior.
3352 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3353 flag
|= FLAG_DATA_ACKED
;
3355 flag
|= FLAG_SYN_ACKED
;
3356 tp
->retrans_stamp
= 0;
3362 tcp_unlink_write_queue(skb
, sk
);
3363 sk_wmem_free_skb(sk
, skb
);
3364 tp
->scoreboard_skb_hint
= NULL
;
3365 if (skb
== tp
->retransmit_skb_hint
)
3366 tp
->retransmit_skb_hint
= NULL
;
3367 if (skb
== tp
->lost_skb_hint
)
3368 tp
->lost_skb_hint
= NULL
;
3371 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3372 tp
->snd_up
= tp
->snd_una
;
3374 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3375 flag
|= FLAG_SACK_RENEGING
;
3377 if (flag
& FLAG_ACKED
) {
3378 const struct tcp_congestion_ops
*ca_ops
3379 = inet_csk(sk
)->icsk_ca_ops
;
3381 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3382 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3383 tcp_mtup_probe_success(sk
);
3386 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3389 if (tcp_is_reno(tp
)) {
3390 tcp_remove_reno_sacks(sk
, pkts_acked
);
3394 /* Non-retransmitted hole got filled? That's reordering */
3395 if (reord
< prior_fackets
)
3396 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3398 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3399 prior_sacked
- tp
->sacked_out
;
3400 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3403 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3405 if (ca_ops
->pkts_acked
) {
3408 /* Is the ACK triggering packet unambiguous? */
3409 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3410 /* High resolution needed and available? */
3411 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3412 !ktime_equal(last_ackt
,
3413 net_invalid_timestamp()))
3414 rtt_us
= ktime_us_delta(ktime_get_real(),
3416 else if (ca_seq_rtt
>= 0)
3417 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3420 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3424 #if FASTRETRANS_DEBUG > 0
3425 WARN_ON((int)tp
->sacked_out
< 0);
3426 WARN_ON((int)tp
->lost_out
< 0);
3427 WARN_ON((int)tp
->retrans_out
< 0);
3428 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3429 icsk
= inet_csk(sk
);
3431 printk(KERN_DEBUG
"Leak l=%u %d\n",
3432 tp
->lost_out
, icsk
->icsk_ca_state
);
3435 if (tp
->sacked_out
) {
3436 printk(KERN_DEBUG
"Leak s=%u %d\n",
3437 tp
->sacked_out
, icsk
->icsk_ca_state
);
3440 if (tp
->retrans_out
) {
3441 printk(KERN_DEBUG
"Leak r=%u %d\n",
3442 tp
->retrans_out
, icsk
->icsk_ca_state
);
3443 tp
->retrans_out
= 0;
3450 static void tcp_ack_probe(struct sock
*sk
)
3452 const struct tcp_sock
*tp
= tcp_sk(sk
);
3453 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3455 /* Was it a usable window open? */
3457 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3458 icsk
->icsk_backoff
= 0;
3459 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3460 /* Socket must be waked up by subsequent tcp_data_snd_check().
3461 * This function is not for random using!
3464 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3465 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3470 static inline int tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3472 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3473 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3476 static inline int tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3478 const struct tcp_sock
*tp
= tcp_sk(sk
);
3479 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3480 !((1 << inet_csk(sk
)->icsk_ca_state
) & (TCPF_CA_Recovery
| TCPF_CA_CWR
));
3483 /* Check that window update is acceptable.
3484 * The function assumes that snd_una<=ack<=snd_next.
3486 static inline int tcp_may_update_window(const struct tcp_sock
*tp
,
3487 const u32 ack
, const u32 ack_seq
,
3490 return after(ack
, tp
->snd_una
) ||
3491 after(ack_seq
, tp
->snd_wl1
) ||
3492 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3495 /* Update our send window.
3497 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3498 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3500 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3503 struct tcp_sock
*tp
= tcp_sk(sk
);
3505 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3507 if (likely(!tcp_hdr(skb
)->syn
))
3508 nwin
<<= tp
->rx_opt
.snd_wscale
;
3510 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3511 flag
|= FLAG_WIN_UPDATE
;
3512 tcp_update_wl(tp
, ack_seq
);
3514 if (tp
->snd_wnd
!= nwin
) {
3517 /* Note, it is the only place, where
3518 * fast path is recovered for sending TCP.
3521 tcp_fast_path_check(sk
);
3523 if (nwin
> tp
->max_window
) {
3524 tp
->max_window
= nwin
;
3525 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3535 /* A very conservative spurious RTO response algorithm: reduce cwnd and
3536 * continue in congestion avoidance.
3538 static void tcp_conservative_spur_to_response(struct tcp_sock
*tp
)
3540 tp
->snd_cwnd
= min(tp
->snd_cwnd
, tp
->snd_ssthresh
);
3541 tp
->snd_cwnd_cnt
= 0;
3542 tp
->bytes_acked
= 0;
3543 TCP_ECN_queue_cwr(tp
);
3544 tcp_moderate_cwnd(tp
);
3547 /* A conservative spurious RTO response algorithm: reduce cwnd using
3548 * rate halving and continue in congestion avoidance.
3550 static void tcp_ratehalving_spur_to_response(struct sock
*sk
)
3552 tcp_enter_cwr(sk
, 0);
3555 static void tcp_undo_spur_to_response(struct sock
*sk
, int flag
)
3557 if (flag
& FLAG_ECE
)
3558 tcp_ratehalving_spur_to_response(sk
);
3560 tcp_undo_cwr(sk
, true);
3563 /* F-RTO spurious RTO detection algorithm (RFC4138)
3565 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3566 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3567 * window (but not to or beyond highest sequence sent before RTO):
3568 * On First ACK, send two new segments out.
3569 * On Second ACK, RTO was likely spurious. Do spurious response (response
3570 * algorithm is not part of the F-RTO detection algorithm
3571 * given in RFC4138 but can be selected separately).
3572 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3573 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3574 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3575 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3577 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3578 * original window even after we transmit two new data segments.
3581 * on first step, wait until first cumulative ACK arrives, then move to
3582 * the second step. In second step, the next ACK decides.
3584 * F-RTO is implemented (mainly) in four functions:
3585 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3586 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3587 * called when tcp_use_frto() showed green light
3588 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3589 * - tcp_enter_frto_loss() is called if there is not enough evidence
3590 * to prove that the RTO is indeed spurious. It transfers the control
3591 * from F-RTO to the conventional RTO recovery
3593 static int tcp_process_frto(struct sock
*sk
, int flag
)
3595 struct tcp_sock
*tp
= tcp_sk(sk
);
3597 tcp_verify_left_out(tp
);
3599 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3600 if (flag
& FLAG_DATA_ACKED
)
3601 inet_csk(sk
)->icsk_retransmits
= 0;
3603 if ((flag
& FLAG_NONHEAD_RETRANS_ACKED
) ||
3604 ((tp
->frto_counter
>= 2) && (flag
& FLAG_RETRANS_DATA_ACKED
)))
3605 tp
->undo_marker
= 0;
3607 if (!before(tp
->snd_una
, tp
->frto_highmark
)) {
3608 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 2 : 3), flag
);
3612 if (!tcp_is_sackfrto(tp
)) {
3613 /* RFC4138 shortcoming in step 2; should also have case c):
3614 * ACK isn't duplicate nor advances window, e.g., opposite dir
3617 if (!(flag
& FLAG_ANY_PROGRESS
) && (flag
& FLAG_NOT_DUP
))
3620 if (!(flag
& FLAG_DATA_ACKED
)) {
3621 tcp_enter_frto_loss(sk
, (tp
->frto_counter
== 1 ? 0 : 3),
3626 if (!(flag
& FLAG_DATA_ACKED
) && (tp
->frto_counter
== 1)) {
3627 /* Prevent sending of new data. */
3628 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
3629 tcp_packets_in_flight(tp
));
3633 if ((tp
->frto_counter
>= 2) &&
3634 (!(flag
& FLAG_FORWARD_PROGRESS
) ||
3635 ((flag
& FLAG_DATA_SACKED
) &&
3636 !(flag
& FLAG_ONLY_ORIG_SACKED
)))) {
3637 /* RFC4138 shortcoming (see comment above) */
3638 if (!(flag
& FLAG_FORWARD_PROGRESS
) &&
3639 (flag
& FLAG_NOT_DUP
))
3642 tcp_enter_frto_loss(sk
, 3, flag
);
3647 if (tp
->frto_counter
== 1) {
3648 /* tcp_may_send_now needs to see updated state */
3649 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + 2;
3650 tp
->frto_counter
= 2;
3652 if (!tcp_may_send_now(sk
))
3653 tcp_enter_frto_loss(sk
, 2, flag
);
3657 switch (sysctl_tcp_frto_response
) {
3659 tcp_undo_spur_to_response(sk
, flag
);
3662 tcp_conservative_spur_to_response(tp
);
3665 tcp_ratehalving_spur_to_response(sk
);
3668 tp
->frto_counter
= 0;
3669 tp
->undo_marker
= 0;
3670 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSPURIOUSRTOS
);
3675 /* This routine deals with incoming acks, but not outgoing ones. */
3676 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3678 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3679 struct tcp_sock
*tp
= tcp_sk(sk
);
3680 u32 prior_snd_una
= tp
->snd_una
;
3681 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3682 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3683 bool is_dupack
= false;
3684 u32 prior_in_flight
;
3687 int prior_sacked
= tp
->sacked_out
;
3689 int newly_acked_sacked
= 0;
3692 /* If the ack is older than previous acks
3693 * then we can probably ignore it.
3695 if (before(ack
, prior_snd_una
))
3698 /* If the ack includes data we haven't sent yet, discard
3699 * this segment (RFC793 Section 3.9).
3701 if (after(ack
, tp
->snd_nxt
))
3704 if (after(ack
, prior_snd_una
))
3705 flag
|= FLAG_SND_UNA_ADVANCED
;
3707 if (sysctl_tcp_abc
) {
3708 if (icsk
->icsk_ca_state
< TCP_CA_CWR
)
3709 tp
->bytes_acked
+= ack
- prior_snd_una
;
3710 else if (icsk
->icsk_ca_state
== TCP_CA_Loss
)
3711 /* we assume just one segment left network */
3712 tp
->bytes_acked
+= min(ack
- prior_snd_una
,
3716 prior_fackets
= tp
->fackets_out
;
3717 prior_in_flight
= tcp_packets_in_flight(tp
);
3719 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3720 /* Window is constant, pure forward advance.
3721 * No more checks are required.
3722 * Note, we use the fact that SND.UNA>=SND.WL2.
3724 tcp_update_wl(tp
, ack_seq
);
3726 flag
|= FLAG_WIN_UPDATE
;
3728 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3730 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3732 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3735 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3737 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3739 if (TCP_SKB_CB(skb
)->sacked
)
3740 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3742 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3745 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3748 /* We passed data and got it acked, remove any soft error
3749 * log. Something worked...
3751 sk
->sk_err_soft
= 0;
3752 icsk
->icsk_probes_out
= 0;
3753 tp
->rcv_tstamp
= tcp_time_stamp
;
3754 prior_packets
= tp
->packets_out
;
3758 /* See if we can take anything off of the retransmit queue. */
3759 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3761 pkts_acked
= prior_packets
- tp
->packets_out
;
3762 newly_acked_sacked
= (prior_packets
- prior_sacked
) -
3763 (tp
->packets_out
- tp
->sacked_out
);
3765 if (tp
->frto_counter
)
3766 frto_cwnd
= tcp_process_frto(sk
, flag
);
3767 /* Guarantee sacktag reordering detection against wrap-arounds */
3768 if (before(tp
->frto_highmark
, tp
->snd_una
))
3769 tp
->frto_highmark
= 0;
3771 if (tcp_ack_is_dubious(sk
, flag
)) {
3772 /* Advance CWND, if state allows this. */
3773 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
&&
3774 tcp_may_raise_cwnd(sk
, flag
))
3775 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3776 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3777 tcp_fastretrans_alert(sk
, pkts_acked
, newly_acked_sacked
,
3780 if ((flag
& FLAG_DATA_ACKED
) && !frto_cwnd
)
3781 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3784 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
))
3785 dst_confirm(__sk_dst_get(sk
));
3790 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3791 if (flag
& FLAG_DSACKING_ACK
)
3792 tcp_fastretrans_alert(sk
, pkts_acked
, newly_acked_sacked
,
3794 /* If this ack opens up a zero window, clear backoff. It was
3795 * being used to time the probes, and is probably far higher than
3796 * it needs to be for normal retransmission.
3798 if (tcp_send_head(sk
))
3803 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3807 /* If data was SACKed, tag it and see if we should send more data.
3808 * If data was DSACKed, see if we can undo a cwnd reduction.
3810 if (TCP_SKB_CB(skb
)->sacked
) {
3811 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3812 newly_acked_sacked
= tp
->sacked_out
- prior_sacked
;
3813 tcp_fastretrans_alert(sk
, pkts_acked
, newly_acked_sacked
,
3817 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3821 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3822 * But, this can also be called on packets in the established flow when
3823 * the fast version below fails.
3825 void tcp_parse_options(const struct sk_buff
*skb
, struct tcp_options_received
*opt_rx
,
3826 const u8
**hvpp
, int estab
)
3828 const unsigned char *ptr
;
3829 const struct tcphdr
*th
= tcp_hdr(skb
);
3830 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3832 ptr
= (const unsigned char *)(th
+ 1);
3833 opt_rx
->saw_tstamp
= 0;
3835 while (length
> 0) {
3836 int opcode
= *ptr
++;
3842 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3847 if (opsize
< 2) /* "silly options" */
3849 if (opsize
> length
)
3850 return; /* don't parse partial options */
3853 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3854 u16 in_mss
= get_unaligned_be16(ptr
);
3856 if (opt_rx
->user_mss
&&
3857 opt_rx
->user_mss
< in_mss
)
3858 in_mss
= opt_rx
->user_mss
;
3859 opt_rx
->mss_clamp
= in_mss
;
3864 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3865 !estab
&& sysctl_tcp_window_scaling
) {
3866 __u8 snd_wscale
= *(__u8
*)ptr
;
3867 opt_rx
->wscale_ok
= 1;
3868 if (snd_wscale
> 14) {
3869 if (net_ratelimit())
3870 printk(KERN_INFO
"tcp_parse_options: Illegal window "
3871 "scaling value %d >14 received.\n",
3875 opt_rx
->snd_wscale
= snd_wscale
;
3878 case TCPOPT_TIMESTAMP
:
3879 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3880 ((estab
&& opt_rx
->tstamp_ok
) ||
3881 (!estab
&& sysctl_tcp_timestamps
))) {
3882 opt_rx
->saw_tstamp
= 1;
3883 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3884 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3887 case TCPOPT_SACK_PERM
:
3888 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3889 !estab
&& sysctl_tcp_sack
) {
3890 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3891 tcp_sack_reset(opt_rx
);
3896 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3897 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3899 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3902 #ifdef CONFIG_TCP_MD5SIG
3905 * The MD5 Hash has already been
3906 * checked (see tcp_v{4,6}_do_rcv()).
3911 /* This option is variable length.
3914 case TCPOLEN_COOKIE_BASE
:
3915 /* not yet implemented */
3917 case TCPOLEN_COOKIE_PAIR
:
3918 /* not yet implemented */
3920 case TCPOLEN_COOKIE_MIN
+0:
3921 case TCPOLEN_COOKIE_MIN
+2:
3922 case TCPOLEN_COOKIE_MIN
+4:
3923 case TCPOLEN_COOKIE_MIN
+6:
3924 case TCPOLEN_COOKIE_MAX
:
3925 /* 16-bit multiple */
3926 opt_rx
->cookie_plus
= opsize
;
3941 EXPORT_SYMBOL(tcp_parse_options
);
3943 static int tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3945 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3947 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3948 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3949 tp
->rx_opt
.saw_tstamp
= 1;
3951 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3953 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
);
3959 /* Fast parse options. This hopes to only see timestamps.
3960 * If it is wrong it falls back on tcp_parse_options().
3962 static int tcp_fast_parse_options(const struct sk_buff
*skb
,
3963 const struct tcphdr
*th
,
3964 struct tcp_sock
*tp
, const u8
**hvpp
)
3966 /* In the spirit of fast parsing, compare doff directly to constant
3967 * values. Because equality is used, short doff can be ignored here.
3969 if (th
->doff
== (sizeof(*th
) / 4)) {
3970 tp
->rx_opt
.saw_tstamp
= 0;
3972 } else if (tp
->rx_opt
.tstamp_ok
&&
3973 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3974 if (tcp_parse_aligned_timestamp(tp
, th
))
3977 tcp_parse_options(skb
, &tp
->rx_opt
, hvpp
, 1);
3981 #ifdef CONFIG_TCP_MD5SIG
3983 * Parse MD5 Signature option
3985 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3987 int length
= (th
->doff
<< 2) - sizeof(*th
);
3988 const u8
*ptr
= (const u8
*)(th
+ 1);
3990 /* If the TCP option is too short, we can short cut */
3991 if (length
< TCPOLEN_MD5SIG
)
3994 while (length
> 0) {
3995 int opcode
= *ptr
++;
4006 if (opsize
< 2 || opsize
> length
)
4008 if (opcode
== TCPOPT_MD5SIG
)
4009 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
4016 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
4019 static inline void tcp_store_ts_recent(struct tcp_sock
*tp
)
4021 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
4022 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
4025 static inline void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
4027 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
4028 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
4029 * extra check below makes sure this can only happen
4030 * for pure ACK frames. -DaveM
4032 * Not only, also it occurs for expired timestamps.
4035 if (tcp_paws_check(&tp
->rx_opt
, 0))
4036 tcp_store_ts_recent(tp
);
4040 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4042 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4043 * it can pass through stack. So, the following predicate verifies that
4044 * this segment is not used for anything but congestion avoidance or
4045 * fast retransmit. Moreover, we even are able to eliminate most of such
4046 * second order effects, if we apply some small "replay" window (~RTO)
4047 * to timestamp space.
4049 * All these measures still do not guarantee that we reject wrapped ACKs
4050 * on networks with high bandwidth, when sequence space is recycled fastly,
4051 * but it guarantees that such events will be very rare and do not affect
4052 * connection seriously. This doesn't look nice, but alas, PAWS is really
4055 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4056 * states that events when retransmit arrives after original data are rare.
4057 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4058 * the biggest problem on large power networks even with minor reordering.
4059 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4060 * up to bandwidth of 18Gigabit/sec. 8) ]
4063 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
4065 const struct tcp_sock
*tp
= tcp_sk(sk
);
4066 const struct tcphdr
*th
= tcp_hdr(skb
);
4067 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4068 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
4070 return (/* 1. Pure ACK with correct sequence number. */
4071 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
4073 /* 2. ... and duplicate ACK. */
4074 ack
== tp
->snd_una
&&
4076 /* 3. ... and does not update window. */
4077 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
4079 /* 4. ... and sits in replay window. */
4080 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
4083 static inline int tcp_paws_discard(const struct sock
*sk
,
4084 const struct sk_buff
*skb
)
4086 const struct tcp_sock
*tp
= tcp_sk(sk
);
4088 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
4089 !tcp_disordered_ack(sk
, skb
);
4092 /* Check segment sequence number for validity.
4094 * Segment controls are considered valid, if the segment
4095 * fits to the window after truncation to the window. Acceptability
4096 * of data (and SYN, FIN, of course) is checked separately.
4097 * See tcp_data_queue(), for example.
4099 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4100 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4101 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4102 * (borrowed from freebsd)
4105 static inline int tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
4107 return !before(end_seq
, tp
->rcv_wup
) &&
4108 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
4111 /* When we get a reset we do this. */
4112 static void tcp_reset(struct sock
*sk
)
4114 /* We want the right error as BSD sees it (and indeed as we do). */
4115 switch (sk
->sk_state
) {
4117 sk
->sk_err
= ECONNREFUSED
;
4119 case TCP_CLOSE_WAIT
:
4125 sk
->sk_err
= ECONNRESET
;
4127 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4130 if (!sock_flag(sk
, SOCK_DEAD
))
4131 sk
->sk_error_report(sk
);
4137 * Process the FIN bit. This now behaves as it is supposed to work
4138 * and the FIN takes effect when it is validly part of sequence
4139 * space. Not before when we get holes.
4141 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4142 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4145 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4146 * close and we go into CLOSING (and later onto TIME-WAIT)
4148 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4150 static void tcp_fin(struct sock
*sk
)
4152 struct tcp_sock
*tp
= tcp_sk(sk
);
4154 inet_csk_schedule_ack(sk
);
4156 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
4157 sock_set_flag(sk
, SOCK_DONE
);
4159 switch (sk
->sk_state
) {
4161 case TCP_ESTABLISHED
:
4162 /* Move to CLOSE_WAIT */
4163 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
4164 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
4167 case TCP_CLOSE_WAIT
:
4169 /* Received a retransmission of the FIN, do
4174 /* RFC793: Remain in the LAST-ACK state. */
4178 /* This case occurs when a simultaneous close
4179 * happens, we must ack the received FIN and
4180 * enter the CLOSING state.
4183 tcp_set_state(sk
, TCP_CLOSING
);
4186 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4188 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
4191 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4192 * cases we should never reach this piece of code.
4194 printk(KERN_ERR
"%s: Impossible, sk->sk_state=%d\n",
4195 __func__
, sk
->sk_state
);
4199 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4200 * Probably, we should reset in this case. For now drop them.
4202 __skb_queue_purge(&tp
->out_of_order_queue
);
4203 if (tcp_is_sack(tp
))
4204 tcp_sack_reset(&tp
->rx_opt
);
4207 if (!sock_flag(sk
, SOCK_DEAD
)) {
4208 sk
->sk_state_change(sk
);
4210 /* Do not send POLL_HUP for half duplex close. */
4211 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
4212 sk
->sk_state
== TCP_CLOSE
)
4213 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
4215 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
4219 static inline int tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
4222 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
4223 if (before(seq
, sp
->start_seq
))
4224 sp
->start_seq
= seq
;
4225 if (after(end_seq
, sp
->end_seq
))
4226 sp
->end_seq
= end_seq
;
4232 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
4234 struct tcp_sock
*tp
= tcp_sk(sk
);
4236 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4239 if (before(seq
, tp
->rcv_nxt
))
4240 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
4242 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
4244 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
4246 tp
->rx_opt
.dsack
= 1;
4247 tp
->duplicate_sack
[0].start_seq
= seq
;
4248 tp
->duplicate_sack
[0].end_seq
= end_seq
;
4252 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
4254 struct tcp_sock
*tp
= tcp_sk(sk
);
4256 if (!tp
->rx_opt
.dsack
)
4257 tcp_dsack_set(sk
, seq
, end_seq
);
4259 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
4262 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
4264 struct tcp_sock
*tp
= tcp_sk(sk
);
4266 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
4267 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4268 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4269 tcp_enter_quickack_mode(sk
);
4271 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
4272 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4274 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
4275 end_seq
= tp
->rcv_nxt
;
4276 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
4283 /* These routines update the SACK block as out-of-order packets arrive or
4284 * in-order packets close up the sequence space.
4286 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
4289 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4290 struct tcp_sack_block
*swalk
= sp
+ 1;
4292 /* See if the recent change to the first SACK eats into
4293 * or hits the sequence space of other SACK blocks, if so coalesce.
4295 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
4296 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
4299 /* Zap SWALK, by moving every further SACK up by one slot.
4300 * Decrease num_sacks.
4302 tp
->rx_opt
.num_sacks
--;
4303 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
4307 this_sack
++, swalk
++;
4311 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
4313 struct tcp_sock
*tp
= tcp_sk(sk
);
4314 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4315 int cur_sacks
= tp
->rx_opt
.num_sacks
;
4321 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
4322 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
4323 /* Rotate this_sack to the first one. */
4324 for (; this_sack
> 0; this_sack
--, sp
--)
4325 swap(*sp
, *(sp
- 1));
4327 tcp_sack_maybe_coalesce(tp
);
4332 /* Could not find an adjacent existing SACK, build a new one,
4333 * put it at the front, and shift everyone else down. We
4334 * always know there is at least one SACK present already here.
4336 * If the sack array is full, forget about the last one.
4338 if (this_sack
>= TCP_NUM_SACKS
) {
4340 tp
->rx_opt
.num_sacks
--;
4343 for (; this_sack
> 0; this_sack
--, sp
--)
4347 /* Build the new head SACK, and we're done. */
4348 sp
->start_seq
= seq
;
4349 sp
->end_seq
= end_seq
;
4350 tp
->rx_opt
.num_sacks
++;
4353 /* RCV.NXT advances, some SACKs should be eaten. */
4355 static void tcp_sack_remove(struct tcp_sock
*tp
)
4357 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
4358 int num_sacks
= tp
->rx_opt
.num_sacks
;
4361 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4362 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
4363 tp
->rx_opt
.num_sacks
= 0;
4367 for (this_sack
= 0; this_sack
< num_sacks
;) {
4368 /* Check if the start of the sack is covered by RCV.NXT. */
4369 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
4372 /* RCV.NXT must cover all the block! */
4373 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
4375 /* Zap this SACK, by moving forward any other SACKS. */
4376 for (i
=this_sack
+1; i
< num_sacks
; i
++)
4377 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
4384 tp
->rx_opt
.num_sacks
= num_sacks
;
4387 /* This one checks to see if we can put data from the
4388 * out_of_order queue into the receive_queue.
4390 static void tcp_ofo_queue(struct sock
*sk
)
4392 struct tcp_sock
*tp
= tcp_sk(sk
);
4393 __u32 dsack_high
= tp
->rcv_nxt
;
4394 struct sk_buff
*skb
;
4396 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
4397 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
4400 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
4401 __u32 dsack
= dsack_high
;
4402 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
4403 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
4404 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
4407 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4408 SOCK_DEBUG(sk
, "ofo packet was already received\n");
4409 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4413 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
4414 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4415 TCP_SKB_CB(skb
)->end_seq
);
4417 __skb_unlink(skb
, &tp
->out_of_order_queue
);
4418 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4419 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4420 if (tcp_hdr(skb
)->fin
)
4425 static int tcp_prune_ofo_queue(struct sock
*sk
);
4426 static int tcp_prune_queue(struct sock
*sk
);
4428 static inline int tcp_try_rmem_schedule(struct sock
*sk
, unsigned int size
)
4430 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
4431 !sk_rmem_schedule(sk
, size
)) {
4433 if (tcp_prune_queue(sk
) < 0)
4436 if (!sk_rmem_schedule(sk
, size
)) {
4437 if (!tcp_prune_ofo_queue(sk
))
4440 if (!sk_rmem_schedule(sk
, size
))
4447 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4449 const struct tcphdr
*th
= tcp_hdr(skb
);
4450 struct tcp_sock
*tp
= tcp_sk(sk
);
4453 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4457 __skb_pull(skb
, th
->doff
* 4);
4459 TCP_ECN_accept_cwr(tp
, skb
);
4461 tp
->rx_opt
.dsack
= 0;
4463 /* Queue data for delivery to the user.
4464 * Packets in sequence go to the receive queue.
4465 * Out of sequence packets to the out_of_order_queue.
4467 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4468 if (tcp_receive_window(tp
) == 0)
4471 /* Ok. In sequence. In window. */
4472 if (tp
->ucopy
.task
== current
&&
4473 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4474 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4475 int chunk
= min_t(unsigned int, skb
->len
,
4478 __set_current_state(TASK_RUNNING
);
4481 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4482 tp
->ucopy
.len
-= chunk
;
4483 tp
->copied_seq
+= chunk
;
4484 eaten
= (chunk
== skb
->len
);
4485 tcp_rcv_space_adjust(sk
);
4493 tcp_try_rmem_schedule(sk
, skb
->truesize
))
4496 skb_set_owner_r(skb
, sk
);
4497 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4499 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4501 tcp_event_data_recv(sk
, skb
);
4505 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4508 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4509 * gap in queue is filled.
4511 if (skb_queue_empty(&tp
->out_of_order_queue
))
4512 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4515 if (tp
->rx_opt
.num_sacks
)
4516 tcp_sack_remove(tp
);
4518 tcp_fast_path_check(sk
);
4522 else if (!sock_flag(sk
, SOCK_DEAD
))
4523 sk
->sk_data_ready(sk
, 0);
4527 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4528 /* A retransmit, 2nd most common case. Force an immediate ack. */
4529 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4530 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4533 tcp_enter_quickack_mode(sk
);
4534 inet_csk_schedule_ack(sk
);
4540 /* Out of window. F.e. zero window probe. */
4541 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4544 tcp_enter_quickack_mode(sk
);
4546 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4547 /* Partial packet, seq < rcv_next < end_seq */
4548 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4549 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4550 TCP_SKB_CB(skb
)->end_seq
);
4552 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4554 /* If window is closed, drop tail of packet. But after
4555 * remembering D-SACK for its head made in previous line.
4557 if (!tcp_receive_window(tp
))
4562 TCP_ECN_check_ce(tp
, skb
);
4564 if (tcp_try_rmem_schedule(sk
, skb
->truesize
))
4567 /* Disable header prediction. */
4569 inet_csk_schedule_ack(sk
);
4571 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4572 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4574 skb_set_owner_r(skb
, sk
);
4576 if (!skb_peek(&tp
->out_of_order_queue
)) {
4577 /* Initial out of order segment, build 1 SACK. */
4578 if (tcp_is_sack(tp
)) {
4579 tp
->rx_opt
.num_sacks
= 1;
4580 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4581 tp
->selective_acks
[0].end_seq
=
4582 TCP_SKB_CB(skb
)->end_seq
;
4584 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4586 struct sk_buff
*skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4587 u32 seq
= TCP_SKB_CB(skb
)->seq
;
4588 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4590 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4591 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4593 if (!tp
->rx_opt
.num_sacks
||
4594 tp
->selective_acks
[0].end_seq
!= seq
)
4597 /* Common case: data arrive in order after hole. */
4598 tp
->selective_acks
[0].end_seq
= end_seq
;
4602 /* Find place to insert this segment. */
4604 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4606 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4610 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4613 /* Do skb overlap to previous one? */
4614 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4615 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4616 /* All the bits are present. Drop. */
4618 tcp_dsack_set(sk
, seq
, end_seq
);
4621 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4622 /* Partial overlap. */
4623 tcp_dsack_set(sk
, seq
,
4624 TCP_SKB_CB(skb1
)->end_seq
);
4626 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4630 skb1
= skb_queue_prev(
4631 &tp
->out_of_order_queue
,
4636 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4638 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4640 /* And clean segments covered by new one as whole. */
4641 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4642 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4644 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4646 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4647 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4651 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4652 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4653 TCP_SKB_CB(skb1
)->end_seq
);
4658 if (tcp_is_sack(tp
))
4659 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4663 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4664 struct sk_buff_head
*list
)
4666 struct sk_buff
*next
= NULL
;
4668 if (!skb_queue_is_last(list
, skb
))
4669 next
= skb_queue_next(list
, skb
);
4671 __skb_unlink(skb
, list
);
4673 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4678 /* Collapse contiguous sequence of skbs head..tail with
4679 * sequence numbers start..end.
4681 * If tail is NULL, this means until the end of the list.
4683 * Segments with FIN/SYN are not collapsed (only because this
4687 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4688 struct sk_buff
*head
, struct sk_buff
*tail
,
4691 struct sk_buff
*skb
, *n
;
4694 /* First, check that queue is collapsible and find
4695 * the point where collapsing can be useful. */
4699 skb_queue_walk_from_safe(list
, skb
, n
) {
4702 /* No new bits? It is possible on ofo queue. */
4703 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4704 skb
= tcp_collapse_one(sk
, skb
, list
);
4710 /* The first skb to collapse is:
4712 * - bloated or contains data before "start" or
4713 * overlaps to the next one.
4715 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4716 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4717 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4718 end_of_skbs
= false;
4722 if (!skb_queue_is_last(list
, skb
)) {
4723 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4725 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4726 end_of_skbs
= false;
4731 /* Decided to skip this, advance start seq. */
4732 start
= TCP_SKB_CB(skb
)->end_seq
;
4734 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4737 while (before(start
, end
)) {
4738 struct sk_buff
*nskb
;
4739 unsigned int header
= skb_headroom(skb
);
4740 int copy
= SKB_MAX_ORDER(header
, 0);
4742 /* Too big header? This can happen with IPv6. */
4745 if (end
- start
< copy
)
4747 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4751 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4752 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4754 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4756 skb_reserve(nskb
, header
);
4757 memcpy(nskb
->head
, skb
->head
, header
);
4758 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4759 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4760 __skb_queue_before(list
, skb
, nskb
);
4761 skb_set_owner_r(nskb
, sk
);
4763 /* Copy data, releasing collapsed skbs. */
4765 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4766 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4770 size
= min(copy
, size
);
4771 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4773 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4777 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4778 skb
= tcp_collapse_one(sk
, skb
, list
);
4781 tcp_hdr(skb
)->syn
||
4789 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4790 * and tcp_collapse() them until all the queue is collapsed.
4792 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4794 struct tcp_sock
*tp
= tcp_sk(sk
);
4795 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4796 struct sk_buff
*head
;
4802 start
= TCP_SKB_CB(skb
)->seq
;
4803 end
= TCP_SKB_CB(skb
)->end_seq
;
4807 struct sk_buff
*next
= NULL
;
4809 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4810 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4813 /* Segment is terminated when we see gap or when
4814 * we are at the end of all the queue. */
4816 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4817 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4818 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4819 head
, skb
, start
, end
);
4823 /* Start new segment */
4824 start
= TCP_SKB_CB(skb
)->seq
;
4825 end
= TCP_SKB_CB(skb
)->end_seq
;
4827 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4828 start
= TCP_SKB_CB(skb
)->seq
;
4829 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4830 end
= TCP_SKB_CB(skb
)->end_seq
;
4836 * Purge the out-of-order queue.
4837 * Return true if queue was pruned.
4839 static int tcp_prune_ofo_queue(struct sock
*sk
)
4841 struct tcp_sock
*tp
= tcp_sk(sk
);
4844 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4845 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4846 __skb_queue_purge(&tp
->out_of_order_queue
);
4848 /* Reset SACK state. A conforming SACK implementation will
4849 * do the same at a timeout based retransmit. When a connection
4850 * is in a sad state like this, we care only about integrity
4851 * of the connection not performance.
4853 if (tp
->rx_opt
.sack_ok
)
4854 tcp_sack_reset(&tp
->rx_opt
);
4861 /* Reduce allocated memory if we can, trying to get
4862 * the socket within its memory limits again.
4864 * Return less than zero if we should start dropping frames
4865 * until the socket owning process reads some of the data
4866 * to stabilize the situation.
4868 static int tcp_prune_queue(struct sock
*sk
)
4870 struct tcp_sock
*tp
= tcp_sk(sk
);
4872 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4874 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4876 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4877 tcp_clamp_window(sk
);
4878 else if (sk_under_memory_pressure(sk
))
4879 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4881 tcp_collapse_ofo_queue(sk
);
4882 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4883 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4884 skb_peek(&sk
->sk_receive_queue
),
4886 tp
->copied_seq
, tp
->rcv_nxt
);
4889 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4892 /* Collapsing did not help, destructive actions follow.
4893 * This must not ever occur. */
4895 tcp_prune_ofo_queue(sk
);
4897 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4900 /* If we are really being abused, tell the caller to silently
4901 * drop receive data on the floor. It will get retransmitted
4902 * and hopefully then we'll have sufficient space.
4904 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4906 /* Massive buffer overcommit. */
4911 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4912 * As additional protections, we do not touch cwnd in retransmission phases,
4913 * and if application hit its sndbuf limit recently.
4915 void tcp_cwnd_application_limited(struct sock
*sk
)
4917 struct tcp_sock
*tp
= tcp_sk(sk
);
4919 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4920 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4921 /* Limited by application or receiver window. */
4922 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4923 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4924 if (win_used
< tp
->snd_cwnd
) {
4925 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4926 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4928 tp
->snd_cwnd_used
= 0;
4930 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4933 static int tcp_should_expand_sndbuf(const struct sock
*sk
)
4935 const struct tcp_sock
*tp
= tcp_sk(sk
);
4937 /* If the user specified a specific send buffer setting, do
4940 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4943 /* If we are under global TCP memory pressure, do not expand. */
4944 if (sk_under_memory_pressure(sk
))
4947 /* If we are under soft global TCP memory pressure, do not expand. */
4948 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4951 /* If we filled the congestion window, do not expand. */
4952 if (tp
->packets_out
>= tp
->snd_cwnd
)
4958 /* When incoming ACK allowed to free some skb from write_queue,
4959 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4960 * on the exit from tcp input handler.
4962 * PROBLEM: sndbuf expansion does not work well with largesend.
4964 static void tcp_new_space(struct sock
*sk
)
4966 struct tcp_sock
*tp
= tcp_sk(sk
);
4968 if (tcp_should_expand_sndbuf(sk
)) {
4969 int sndmem
= SKB_TRUESIZE(max_t(u32
,
4970 tp
->rx_opt
.mss_clamp
,
4973 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4974 tp
->reordering
+ 1);
4975 sndmem
*= 2 * demanded
;
4976 if (sndmem
> sk
->sk_sndbuf
)
4977 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4978 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4981 sk
->sk_write_space(sk
);
4984 static void tcp_check_space(struct sock
*sk
)
4986 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4987 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4988 if (sk
->sk_socket
&&
4989 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4994 static inline void tcp_data_snd_check(struct sock
*sk
)
4996 tcp_push_pending_frames(sk
);
4997 tcp_check_space(sk
);
5001 * Check if sending an ack is needed.
5003 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
5005 struct tcp_sock
*tp
= tcp_sk(sk
);
5007 /* More than one full frame received... */
5008 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
5009 /* ... and right edge of window advances far enough.
5010 * (tcp_recvmsg() will send ACK otherwise). Or...
5012 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
5013 /* We ACK each frame or... */
5014 tcp_in_quickack_mode(sk
) ||
5015 /* We have out of order data. */
5016 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
5017 /* Then ack it now */
5020 /* Else, send delayed ack. */
5021 tcp_send_delayed_ack(sk
);
5025 static inline void tcp_ack_snd_check(struct sock
*sk
)
5027 if (!inet_csk_ack_scheduled(sk
)) {
5028 /* We sent a data segment already. */
5031 __tcp_ack_snd_check(sk
, 1);
5035 * This routine is only called when we have urgent data
5036 * signaled. Its the 'slow' part of tcp_urg. It could be
5037 * moved inline now as tcp_urg is only called from one
5038 * place. We handle URGent data wrong. We have to - as
5039 * BSD still doesn't use the correction from RFC961.
5040 * For 1003.1g we should support a new option TCP_STDURG to permit
5041 * either form (or just set the sysctl tcp_stdurg).
5044 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
5046 struct tcp_sock
*tp
= tcp_sk(sk
);
5047 u32 ptr
= ntohs(th
->urg_ptr
);
5049 if (ptr
&& !sysctl_tcp_stdurg
)
5051 ptr
+= ntohl(th
->seq
);
5053 /* Ignore urgent data that we've already seen and read. */
5054 if (after(tp
->copied_seq
, ptr
))
5057 /* Do not replay urg ptr.
5059 * NOTE: interesting situation not covered by specs.
5060 * Misbehaving sender may send urg ptr, pointing to segment,
5061 * which we already have in ofo queue. We are not able to fetch
5062 * such data and will stay in TCP_URG_NOTYET until will be eaten
5063 * by recvmsg(). Seems, we are not obliged to handle such wicked
5064 * situations. But it is worth to think about possibility of some
5065 * DoSes using some hypothetical application level deadlock.
5067 if (before(ptr
, tp
->rcv_nxt
))
5070 /* Do we already have a newer (or duplicate) urgent pointer? */
5071 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
5074 /* Tell the world about our new urgent pointer. */
5077 /* We may be adding urgent data when the last byte read was
5078 * urgent. To do this requires some care. We cannot just ignore
5079 * tp->copied_seq since we would read the last urgent byte again
5080 * as data, nor can we alter copied_seq until this data arrives
5081 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5083 * NOTE. Double Dutch. Rendering to plain English: author of comment
5084 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5085 * and expect that both A and B disappear from stream. This is _wrong_.
5086 * Though this happens in BSD with high probability, this is occasional.
5087 * Any application relying on this is buggy. Note also, that fix "works"
5088 * only in this artificial test. Insert some normal data between A and B and we will
5089 * decline of BSD again. Verdict: it is better to remove to trap
5092 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
5093 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
5094 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
5096 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5097 __skb_unlink(skb
, &sk
->sk_receive_queue
);
5102 tp
->urg_data
= TCP_URG_NOTYET
;
5105 /* Disable header prediction. */
5109 /* This is the 'fast' part of urgent handling. */
5110 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
5112 struct tcp_sock
*tp
= tcp_sk(sk
);
5114 /* Check if we get a new urgent pointer - normally not. */
5116 tcp_check_urg(sk
, th
);
5118 /* Do we wait for any urgent data? - normally not... */
5119 if (tp
->urg_data
== TCP_URG_NOTYET
) {
5120 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
5123 /* Is the urgent pointer pointing into this packet? */
5124 if (ptr
< skb
->len
) {
5126 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
5128 tp
->urg_data
= TCP_URG_VALID
| tmp
;
5129 if (!sock_flag(sk
, SOCK_DEAD
))
5130 sk
->sk_data_ready(sk
, 0);
5135 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
5137 struct tcp_sock
*tp
= tcp_sk(sk
);
5138 int chunk
= skb
->len
- hlen
;
5142 if (skb_csum_unnecessary(skb
))
5143 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
5145 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
5149 tp
->ucopy
.len
-= chunk
;
5150 tp
->copied_seq
+= chunk
;
5151 tcp_rcv_space_adjust(sk
);
5158 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
5159 struct sk_buff
*skb
)
5163 if (sock_owned_by_user(sk
)) {
5165 result
= __tcp_checksum_complete(skb
);
5168 result
= __tcp_checksum_complete(skb
);
5173 static inline int tcp_checksum_complete_user(struct sock
*sk
,
5174 struct sk_buff
*skb
)
5176 return !skb_csum_unnecessary(skb
) &&
5177 __tcp_checksum_complete_user(sk
, skb
);
5180 #ifdef CONFIG_NET_DMA
5181 static int tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
5184 struct tcp_sock
*tp
= tcp_sk(sk
);
5185 int chunk
= skb
->len
- hlen
;
5187 int copied_early
= 0;
5189 if (tp
->ucopy
.wakeup
)
5192 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
5193 tp
->ucopy
.dma_chan
= dma_find_channel(DMA_MEMCPY
);
5195 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
5197 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
5199 tp
->ucopy
.iov
, chunk
,
5200 tp
->ucopy
.pinned_list
);
5205 tp
->ucopy
.dma_cookie
= dma_cookie
;
5208 tp
->ucopy
.len
-= chunk
;
5209 tp
->copied_seq
+= chunk
;
5210 tcp_rcv_space_adjust(sk
);
5212 if ((tp
->ucopy
.len
== 0) ||
5213 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
5214 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
5215 tp
->ucopy
.wakeup
= 1;
5216 sk
->sk_data_ready(sk
, 0);
5218 } else if (chunk
> 0) {
5219 tp
->ucopy
.wakeup
= 1;
5220 sk
->sk_data_ready(sk
, 0);
5223 return copied_early
;
5225 #endif /* CONFIG_NET_DMA */
5227 /* Does PAWS and seqno based validation of an incoming segment, flags will
5228 * play significant role here.
5230 static int tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
5231 const struct tcphdr
*th
, int syn_inerr
)
5233 const u8
*hash_location
;
5234 struct tcp_sock
*tp
= tcp_sk(sk
);
5236 /* RFC1323: H1. Apply PAWS check first. */
5237 if (tcp_fast_parse_options(skb
, th
, tp
, &hash_location
) &&
5238 tp
->rx_opt
.saw_tstamp
&&
5239 tcp_paws_discard(sk
, skb
)) {
5241 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
5242 tcp_send_dupack(sk
, skb
);
5245 /* Reset is accepted even if it did not pass PAWS. */
5248 /* Step 1: check sequence number */
5249 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
5250 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5251 * (RST) segments are validated by checking their SEQ-fields."
5252 * And page 69: "If an incoming segment is not acceptable,
5253 * an acknowledgment should be sent in reply (unless the RST
5254 * bit is set, if so drop the segment and return)".
5257 tcp_send_dupack(sk
, skb
);
5261 /* Step 2: check RST bit */
5267 /* ts_recent update must be made after we are sure that the packet
5270 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
5272 /* step 3: check security and precedence [ignored] */
5274 /* step 4: Check for a SYN in window. */
5275 if (th
->syn
&& !before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
5277 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5278 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONSYN
);
5291 * TCP receive function for the ESTABLISHED state.
5293 * It is split into a fast path and a slow path. The fast path is
5295 * - A zero window was announced from us - zero window probing
5296 * is only handled properly in the slow path.
5297 * - Out of order segments arrived.
5298 * - Urgent data is expected.
5299 * - There is no buffer space left
5300 * - Unexpected TCP flags/window values/header lengths are received
5301 * (detected by checking the TCP header against pred_flags)
5302 * - Data is sent in both directions. Fast path only supports pure senders
5303 * or pure receivers (this means either the sequence number or the ack
5304 * value must stay constant)
5305 * - Unexpected TCP option.
5307 * When these conditions are not satisfied it drops into a standard
5308 * receive procedure patterned after RFC793 to handle all cases.
5309 * The first three cases are guaranteed by proper pred_flags setting,
5310 * the rest is checked inline. Fast processing is turned on in
5311 * tcp_data_queue when everything is OK.
5313 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5314 const struct tcphdr
*th
, unsigned int len
)
5316 struct tcp_sock
*tp
= tcp_sk(sk
);
5320 * Header prediction.
5321 * The code loosely follows the one in the famous
5322 * "30 instruction TCP receive" Van Jacobson mail.
5324 * Van's trick is to deposit buffers into socket queue
5325 * on a device interrupt, to call tcp_recv function
5326 * on the receive process context and checksum and copy
5327 * the buffer to user space. smart...
5329 * Our current scheme is not silly either but we take the
5330 * extra cost of the net_bh soft interrupt processing...
5331 * We do checksum and copy also but from device to kernel.
5334 tp
->rx_opt
.saw_tstamp
= 0;
5336 /* pred_flags is 0xS?10 << 16 + snd_wnd
5337 * if header_prediction is to be made
5338 * 'S' will always be tp->tcp_header_len >> 2
5339 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5340 * turn it off (when there are holes in the receive
5341 * space for instance)
5342 * PSH flag is ignored.
5345 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5346 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5347 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5348 int tcp_header_len
= tp
->tcp_header_len
;
5350 /* Timestamp header prediction: tcp_header_len
5351 * is automatically equal to th->doff*4 due to pred_flags
5355 /* Check timestamp */
5356 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5357 /* No? Slow path! */
5358 if (!tcp_parse_aligned_timestamp(tp
, th
))
5361 /* If PAWS failed, check it more carefully in slow path */
5362 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5365 /* DO NOT update ts_recent here, if checksum fails
5366 * and timestamp was corrupted part, it will result
5367 * in a hung connection since we will drop all
5368 * future packets due to the PAWS test.
5372 if (len
<= tcp_header_len
) {
5373 /* Bulk data transfer: sender */
5374 if (len
== tcp_header_len
) {
5375 /* Predicted packet is in window by definition.
5376 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5377 * Hence, check seq<=rcv_wup reduces to:
5379 if (tcp_header_len
==
5380 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5381 tp
->rcv_nxt
== tp
->rcv_wup
)
5382 tcp_store_ts_recent(tp
);
5384 /* We know that such packets are checksummed
5387 tcp_ack(sk
, skb
, 0);
5389 tcp_data_snd_check(sk
);
5391 } else { /* Header too small */
5392 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5397 int copied_early
= 0;
5399 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5400 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5401 #ifdef CONFIG_NET_DMA
5402 if (tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5407 if (tp
->ucopy
.task
== current
&&
5408 sock_owned_by_user(sk
) && !copied_early
) {
5409 __set_current_state(TASK_RUNNING
);
5411 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5415 /* Predicted packet is in window by definition.
5416 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5417 * Hence, check seq<=rcv_wup reduces to:
5419 if (tcp_header_len
==
5420 (sizeof(struct tcphdr
) +
5421 TCPOLEN_TSTAMP_ALIGNED
) &&
5422 tp
->rcv_nxt
== tp
->rcv_wup
)
5423 tcp_store_ts_recent(tp
);
5425 tcp_rcv_rtt_measure_ts(sk
, skb
);
5427 __skb_pull(skb
, tcp_header_len
);
5428 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5429 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5432 tcp_cleanup_rbuf(sk
, skb
->len
);
5435 if (tcp_checksum_complete_user(sk
, skb
))
5438 /* Predicted packet is in window by definition.
5439 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5440 * Hence, check seq<=rcv_wup reduces to:
5442 if (tcp_header_len
==
5443 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5444 tp
->rcv_nxt
== tp
->rcv_wup
)
5445 tcp_store_ts_recent(tp
);
5447 tcp_rcv_rtt_measure_ts(sk
, skb
);
5449 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5452 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5454 /* Bulk data transfer: receiver */
5455 __skb_pull(skb
, tcp_header_len
);
5456 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
5457 skb_set_owner_r(skb
, sk
);
5458 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5461 tcp_event_data_recv(sk
, skb
);
5463 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5464 /* Well, only one small jumplet in fast path... */
5465 tcp_ack(sk
, skb
, FLAG_DATA
);
5466 tcp_data_snd_check(sk
);
5467 if (!inet_csk_ack_scheduled(sk
))
5471 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5472 __tcp_ack_snd_check(sk
, 0);
5474 #ifdef CONFIG_NET_DMA
5476 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5482 sk
->sk_data_ready(sk
, 0);
5488 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5492 * Standard slow path.
5495 res
= tcp_validate_incoming(sk
, skb
, th
, 1);
5500 if (th
->ack
&& tcp_ack(sk
, skb
, FLAG_SLOWPATH
) < 0)
5503 tcp_rcv_rtt_measure_ts(sk
, skb
);
5505 /* Process urgent data. */
5506 tcp_urg(sk
, skb
, th
);
5508 /* step 7: process the segment text */
5509 tcp_data_queue(sk
, skb
);
5511 tcp_data_snd_check(sk
);
5512 tcp_ack_snd_check(sk
);
5516 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5522 EXPORT_SYMBOL(tcp_rcv_established
);
5524 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5525 const struct tcphdr
*th
, unsigned int len
)
5527 const u8
*hash_location
;
5528 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5529 struct tcp_sock
*tp
= tcp_sk(sk
);
5530 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
5531 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5533 tcp_parse_options(skb
, &tp
->rx_opt
, &hash_location
, 0);
5537 * "If the state is SYN-SENT then
5538 * first check the ACK bit
5539 * If the ACK bit is set
5540 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5541 * a reset (unless the RST bit is set, if so drop
5542 * the segment and return)"
5544 * We do not send data with SYN, so that RFC-correct
5547 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_nxt
)
5548 goto reset_and_undo
;
5550 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5551 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5553 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5554 goto reset_and_undo
;
5557 /* Now ACK is acceptable.
5559 * "If the RST bit is set
5560 * If the ACK was acceptable then signal the user "error:
5561 * connection reset", drop the segment, enter CLOSED state,
5562 * delete TCB, and return."
5571 * "fifth, if neither of the SYN or RST bits is set then
5572 * drop the segment and return."
5578 goto discard_and_undo
;
5581 * "If the SYN bit is on ...
5582 * are acceptable then ...
5583 * (our SYN has been ACKed), change the connection
5584 * state to ESTABLISHED..."
5587 TCP_ECN_rcv_synack(tp
, th
);
5589 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5590 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5592 /* Ok.. it's good. Set up sequence numbers and
5593 * move to established.
5595 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5596 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5598 /* RFC1323: The window in SYN & SYN/ACK segments is
5601 tp
->snd_wnd
= ntohs(th
->window
);
5602 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5604 if (!tp
->rx_opt
.wscale_ok
) {
5605 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5606 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5609 if (tp
->rx_opt
.saw_tstamp
) {
5610 tp
->rx_opt
.tstamp_ok
= 1;
5611 tp
->tcp_header_len
=
5612 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5613 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5614 tcp_store_ts_recent(tp
);
5616 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5619 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5620 tcp_enable_fack(tp
);
5623 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5624 tcp_initialize_rcv_mss(sk
);
5626 /* Remember, tcp_poll() does not lock socket!
5627 * Change state from SYN-SENT only after copied_seq
5628 * is initialized. */
5629 tp
->copied_seq
= tp
->rcv_nxt
;
5632 cvp
->cookie_pair_size
> 0 &&
5633 tp
->rx_opt
.cookie_plus
> 0) {
5634 int cookie_size
= tp
->rx_opt
.cookie_plus
5635 - TCPOLEN_COOKIE_BASE
;
5636 int cookie_pair_size
= cookie_size
5637 + cvp
->cookie_desired
;
5639 /* A cookie extension option was sent and returned.
5640 * Note that each incoming SYNACK replaces the
5641 * Responder cookie. The initial exchange is most
5642 * fragile, as protection against spoofing relies
5643 * entirely upon the sequence and timestamp (above).
5644 * This replacement strategy allows the correct pair to
5645 * pass through, while any others will be filtered via
5646 * Responder verification later.
5648 if (sizeof(cvp
->cookie_pair
) >= cookie_pair_size
) {
5649 memcpy(&cvp
->cookie_pair
[cvp
->cookie_desired
],
5650 hash_location
, cookie_size
);
5651 cvp
->cookie_pair_size
= cookie_pair_size
;
5656 tcp_set_state(sk
, TCP_ESTABLISHED
);
5658 security_inet_conn_established(sk
, skb
);
5660 /* Make sure socket is routed, for correct metrics. */
5661 icsk
->icsk_af_ops
->rebuild_header(sk
);
5663 tcp_init_metrics(sk
);
5665 tcp_init_congestion_control(sk
);
5667 /* Prevent spurious tcp_cwnd_restart() on first data
5670 tp
->lsndtime
= tcp_time_stamp
;
5672 tcp_init_buffer_space(sk
);
5674 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5675 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5677 if (!tp
->rx_opt
.snd_wscale
)
5678 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5682 if (!sock_flag(sk
, SOCK_DEAD
)) {
5683 sk
->sk_state_change(sk
);
5684 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5687 if (sk
->sk_write_pending
||
5688 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5689 icsk
->icsk_ack
.pingpong
) {
5690 /* Save one ACK. Data will be ready after
5691 * several ticks, if write_pending is set.
5693 * It may be deleted, but with this feature tcpdumps
5694 * look so _wonderfully_ clever, that I was not able
5695 * to stand against the temptation 8) --ANK
5697 inet_csk_schedule_ack(sk
);
5698 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5699 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
5700 tcp_incr_quickack(sk
);
5701 tcp_enter_quickack_mode(sk
);
5702 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5703 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5714 /* No ACK in the segment */
5718 * "If the RST bit is set
5720 * Otherwise (no ACK) drop the segment and return."
5723 goto discard_and_undo
;
5727 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5728 tcp_paws_reject(&tp
->rx_opt
, 0))
5729 goto discard_and_undo
;
5732 /* We see SYN without ACK. It is attempt of
5733 * simultaneous connect with crossed SYNs.
5734 * Particularly, it can be connect to self.
5736 tcp_set_state(sk
, TCP_SYN_RECV
);
5738 if (tp
->rx_opt
.saw_tstamp
) {
5739 tp
->rx_opt
.tstamp_ok
= 1;
5740 tcp_store_ts_recent(tp
);
5741 tp
->tcp_header_len
=
5742 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5744 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5747 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5748 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5750 /* RFC1323: The window in SYN & SYN/ACK segments is
5753 tp
->snd_wnd
= ntohs(th
->window
);
5754 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5755 tp
->max_window
= tp
->snd_wnd
;
5757 TCP_ECN_rcv_syn(tp
, th
);
5760 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5761 tcp_initialize_rcv_mss(sk
);
5763 tcp_send_synack(sk
);
5765 /* Note, we could accept data and URG from this segment.
5766 * There are no obstacles to make this.
5768 * However, if we ignore data in ACKless segments sometimes,
5769 * we have no reasons to accept it sometimes.
5770 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5771 * is not flawless. So, discard packet for sanity.
5772 * Uncomment this return to process the data.
5779 /* "fifth, if neither of the SYN or RST bits is set then
5780 * drop the segment and return."
5784 tcp_clear_options(&tp
->rx_opt
);
5785 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5789 tcp_clear_options(&tp
->rx_opt
);
5790 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5795 * This function implements the receiving procedure of RFC 793 for
5796 * all states except ESTABLISHED and TIME_WAIT.
5797 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5798 * address independent.
5801 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5802 const struct tcphdr
*th
, unsigned int len
)
5804 struct tcp_sock
*tp
= tcp_sk(sk
);
5805 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5809 tp
->rx_opt
.saw_tstamp
= 0;
5811 switch (sk
->sk_state
) {
5825 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5828 /* Now we have several options: In theory there is
5829 * nothing else in the frame. KA9Q has an option to
5830 * send data with the syn, BSD accepts data with the
5831 * syn up to the [to be] advertised window and
5832 * Solaris 2.1 gives you a protocol error. For now
5833 * we just ignore it, that fits the spec precisely
5834 * and avoids incompatibilities. It would be nice in
5835 * future to drop through and process the data.
5837 * Now that TTCP is starting to be used we ought to
5839 * But, this leaves one open to an easy denial of
5840 * service attack, and SYN cookies can't defend
5841 * against this problem. So, we drop the data
5842 * in the interest of security over speed unless
5843 * it's still in use.
5851 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5855 /* Do step6 onward by hand. */
5856 tcp_urg(sk
, skb
, th
);
5858 tcp_data_snd_check(sk
);
5862 res
= tcp_validate_incoming(sk
, skb
, th
, 0);
5866 /* step 5: check the ACK field */
5868 int acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
) > 0;
5870 switch (sk
->sk_state
) {
5873 tp
->copied_seq
= tp
->rcv_nxt
;
5875 tcp_set_state(sk
, TCP_ESTABLISHED
);
5876 sk
->sk_state_change(sk
);
5878 /* Note, that this wakeup is only for marginal
5879 * crossed SYN case. Passively open sockets
5880 * are not waked up, because sk->sk_sleep ==
5881 * NULL and sk->sk_socket == NULL.
5885 SOCK_WAKE_IO
, POLL_OUT
);
5887 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5888 tp
->snd_wnd
= ntohs(th
->window
) <<
5889 tp
->rx_opt
.snd_wscale
;
5890 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5892 if (tp
->rx_opt
.tstamp_ok
)
5893 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5895 /* Make sure socket is routed, for
5898 icsk
->icsk_af_ops
->rebuild_header(sk
);
5900 tcp_init_metrics(sk
);
5902 tcp_init_congestion_control(sk
);
5904 /* Prevent spurious tcp_cwnd_restart() on
5905 * first data packet.
5907 tp
->lsndtime
= tcp_time_stamp
;
5910 tcp_initialize_rcv_mss(sk
);
5911 tcp_init_buffer_space(sk
);
5912 tcp_fast_path_on(tp
);
5919 if (tp
->snd_una
== tp
->write_seq
) {
5920 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5921 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5922 dst_confirm(__sk_dst_get(sk
));
5924 if (!sock_flag(sk
, SOCK_DEAD
))
5925 /* Wake up lingering close() */
5926 sk
->sk_state_change(sk
);
5930 if (tp
->linger2
< 0 ||
5931 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5932 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5934 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5938 tmo
= tcp_fin_time(sk
);
5939 if (tmo
> TCP_TIMEWAIT_LEN
) {
5940 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5941 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5942 /* Bad case. We could lose such FIN otherwise.
5943 * It is not a big problem, but it looks confusing
5944 * and not so rare event. We still can lose it now,
5945 * if it spins in bh_lock_sock(), but it is really
5948 inet_csk_reset_keepalive_timer(sk
, tmo
);
5950 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5958 if (tp
->snd_una
== tp
->write_seq
) {
5959 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5965 if (tp
->snd_una
== tp
->write_seq
) {
5966 tcp_update_metrics(sk
);
5975 /* step 6: check the URG bit */
5976 tcp_urg(sk
, skb
, th
);
5978 /* step 7: process the segment text */
5979 switch (sk
->sk_state
) {
5980 case TCP_CLOSE_WAIT
:
5983 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5987 /* RFC 793 says to queue data in these states,
5988 * RFC 1122 says we MUST send a reset.
5989 * BSD 4.4 also does reset.
5991 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5992 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5993 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5994 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
6000 case TCP_ESTABLISHED
:
6001 tcp_data_queue(sk
, skb
);
6006 /* tcp_data could move socket to TIME-WAIT */
6007 if (sk
->sk_state
!= TCP_CLOSE
) {
6008 tcp_data_snd_check(sk
);
6009 tcp_ack_snd_check(sk
);
6018 EXPORT_SYMBOL(tcp_rcv_state_process
);