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
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
73 #include <net/inet_common.h>
74 #include <linux/ipsec.h>
75 #include <asm/unaligned.h>
76 #include <net/netdma.h>
78 int sysctl_tcp_timestamps __read_mostly
= 1;
79 int sysctl_tcp_window_scaling __read_mostly
= 1;
80 int sysctl_tcp_sack __read_mostly
= 1;
81 int sysctl_tcp_fack __read_mostly
= 1;
82 int sysctl_tcp_reordering __read_mostly
= TCP_FASTRETRANS_THRESH
;
83 EXPORT_SYMBOL(sysctl_tcp_reordering
);
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
= 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale
);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit
= 100;
92 int sysctl_tcp_stdurg __read_mostly
;
93 int sysctl_tcp_rfc1337 __read_mostly
;
94 int sysctl_tcp_max_orphans __read_mostly
= NR_FILE
;
95 int sysctl_tcp_frto __read_mostly
= 2;
97 int sysctl_tcp_thin_dupack __read_mostly
;
99 int sysctl_tcp_moderate_rcvbuf __read_mostly
= 1;
100 int sysctl_tcp_early_retrans __read_mostly
= 3;
102 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
103 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
108 #define FLAG_ECE 0x40 /* ECE in this ACK */
109 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
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 int 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 bool tcp_in_quickack_mode(const struct sock
*sk
)
198 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
200 return icsk
->icsk_ack
.quick
&& !icsk
->icsk_ack
.pingpong
;
203 static inline void TCP_ECN_queue_cwr(struct tcp_sock
*tp
)
205 if (tp
->ecn_flags
& TCP_ECN_OK
)
206 tp
->ecn_flags
|= TCP_ECN_QUEUE_CWR
;
209 static inline void TCP_ECN_accept_cwr(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
211 if (tcp_hdr(skb
)->cwr
)
212 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
215 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock
*tp
)
217 tp
->ecn_flags
&= ~TCP_ECN_DEMAND_CWR
;
220 static inline void TCP_ECN_check_ce(struct tcp_sock
*tp
, const struct sk_buff
*skb
)
222 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
225 switch (TCP_SKB_CB(skb
)->ip_dsfield
& INET_ECN_MASK
) {
226 case INET_ECN_NOT_ECT
:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
231 if (tp
->ecn_flags
& TCP_ECN_SEEN
)
232 tcp_enter_quickack_mode((struct sock
*)tp
);
235 if (!(tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock
*)tp
);
238 tp
->ecn_flags
|= TCP_ECN_DEMAND_CWR
;
242 tp
->ecn_flags
|= TCP_ECN_SEEN
;
246 static inline void TCP_ECN_rcv_synack(struct tcp_sock
*tp
, const struct tcphdr
*th
)
248 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| th
->cwr
))
249 tp
->ecn_flags
&= ~TCP_ECN_OK
;
252 static inline void TCP_ECN_rcv_syn(struct tcp_sock
*tp
, const struct tcphdr
*th
)
254 if ((tp
->ecn_flags
& TCP_ECN_OK
) && (!th
->ece
|| !th
->cwr
))
255 tp
->ecn_flags
&= ~TCP_ECN_OK
;
258 static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock
*tp
, const struct tcphdr
*th
)
260 if (th
->ece
&& !th
->syn
&& (tp
->ecn_flags
& TCP_ECN_OK
))
265 /* Buffer size and advertised window tuning.
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
270 static void tcp_fixup_sndbuf(struct sock
*sk
)
272 int sndmem
= SKB_TRUESIZE(tcp_sk(sk
)->rx_opt
.mss_clamp
+ MAX_TCP_HEADER
);
274 sndmem
*= TCP_INIT_CWND
;
275 if (sk
->sk_sndbuf
< sndmem
)
276 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
279 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
281 * All tcp_full_space() is split to two parts: "network" buffer, allocated
282 * forward and advertised in receiver window (tp->rcv_wnd) and
283 * "application buffer", required to isolate scheduling/application
284 * latencies from network.
285 * window_clamp is maximal advertised window. It can be less than
286 * tcp_full_space(), in this case tcp_full_space() - window_clamp
287 * is reserved for "application" buffer. The less window_clamp is
288 * the smoother our behaviour from viewpoint of network, but the lower
289 * throughput and the higher sensitivity of the connection to losses. 8)
291 * rcv_ssthresh is more strict window_clamp used at "slow start"
292 * phase to predict further behaviour of this connection.
293 * It is used for two goals:
294 * - to enforce header prediction at sender, even when application
295 * requires some significant "application buffer". It is check #1.
296 * - to prevent pruning of receive queue because of misprediction
297 * of receiver window. Check #2.
299 * The scheme does not work when sender sends good segments opening
300 * window and then starts to feed us spaghetti. But it should work
301 * in common situations. Otherwise, we have to rely on queue collapsing.
304 /* Slow part of check#2. */
305 static int __tcp_grow_window(const struct sock
*sk
, const struct sk_buff
*skb
)
307 struct tcp_sock
*tp
= tcp_sk(sk
);
309 int truesize
= tcp_win_from_space(skb
->truesize
) >> 1;
310 int window
= tcp_win_from_space(sysctl_tcp_rmem
[2]) >> 1;
312 while (tp
->rcv_ssthresh
<= window
) {
313 if (truesize
<= skb
->len
)
314 return 2 * inet_csk(sk
)->icsk_ack
.rcv_mss
;
322 static void tcp_grow_window(struct sock
*sk
, const struct sk_buff
*skb
)
324 struct tcp_sock
*tp
= tcp_sk(sk
);
327 if (tp
->rcv_ssthresh
< tp
->window_clamp
&&
328 (int)tp
->rcv_ssthresh
< tcp_space(sk
) &&
329 !sk_under_memory_pressure(sk
)) {
332 /* Check #2. Increase window, if skb with such overhead
333 * will fit to rcvbuf in future.
335 if (tcp_win_from_space(skb
->truesize
) <= skb
->len
)
336 incr
= 2 * tp
->advmss
;
338 incr
= __tcp_grow_window(sk
, skb
);
341 incr
= max_t(int, incr
, 2 * skb
->len
);
342 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
+ incr
,
344 inet_csk(sk
)->icsk_ack
.quick
|= 1;
349 /* 3. Tuning rcvbuf, when connection enters established state. */
351 static void tcp_fixup_rcvbuf(struct sock
*sk
)
353 u32 mss
= tcp_sk(sk
)->advmss
;
354 u32 icwnd
= TCP_DEFAULT_INIT_RCVWND
;
357 /* Limit to 10 segments if mss <= 1460,
358 * or 14600/mss segments, with a minimum of two segments.
361 icwnd
= max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
363 rcvmem
= 2 * SKB_TRUESIZE(mss
+ MAX_TCP_HEADER
);
367 if (sk
->sk_rcvbuf
< rcvmem
)
368 sk
->sk_rcvbuf
= min(rcvmem
, sysctl_tcp_rmem
[2]);
371 /* 4. Try to fixup all. It is made immediately after connection enters
374 void tcp_init_buffer_space(struct sock
*sk
)
376 struct tcp_sock
*tp
= tcp_sk(sk
);
379 if (!(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
))
380 tcp_fixup_rcvbuf(sk
);
381 if (!(sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
))
382 tcp_fixup_sndbuf(sk
);
384 tp
->rcvq_space
.space
= tp
->rcv_wnd
;
386 maxwin
= tcp_full_space(sk
);
388 if (tp
->window_clamp
>= maxwin
) {
389 tp
->window_clamp
= maxwin
;
391 if (sysctl_tcp_app_win
&& maxwin
> 4 * tp
->advmss
)
392 tp
->window_clamp
= max(maxwin
-
393 (maxwin
>> sysctl_tcp_app_win
),
397 /* Force reservation of one segment. */
398 if (sysctl_tcp_app_win
&&
399 tp
->window_clamp
> 2 * tp
->advmss
&&
400 tp
->window_clamp
+ tp
->advmss
> maxwin
)
401 tp
->window_clamp
= max(2 * tp
->advmss
, maxwin
- tp
->advmss
);
403 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, tp
->window_clamp
);
404 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
407 /* 5. Recalculate window clamp after socket hit its memory bounds. */
408 static void tcp_clamp_window(struct sock
*sk
)
410 struct tcp_sock
*tp
= tcp_sk(sk
);
411 struct inet_connection_sock
*icsk
= inet_csk(sk
);
413 icsk
->icsk_ack
.quick
= 0;
415 if (sk
->sk_rcvbuf
< sysctl_tcp_rmem
[2] &&
416 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
) &&
417 !sk_under_memory_pressure(sk
) &&
418 sk_memory_allocated(sk
) < sk_prot_mem_limits(sk
, 0)) {
419 sk
->sk_rcvbuf
= min(atomic_read(&sk
->sk_rmem_alloc
),
422 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
)
423 tp
->rcv_ssthresh
= min(tp
->window_clamp
, 2U * tp
->advmss
);
426 /* Initialize RCV_MSS value.
427 * RCV_MSS is an our guess about MSS used by the peer.
428 * We haven't any direct information about the MSS.
429 * It's better to underestimate the RCV_MSS rather than overestimate.
430 * Overestimations make us ACKing less frequently than needed.
431 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
433 void tcp_initialize_rcv_mss(struct sock
*sk
)
435 const struct tcp_sock
*tp
= tcp_sk(sk
);
436 unsigned int hint
= min_t(unsigned int, tp
->advmss
, tp
->mss_cache
);
438 hint
= min(hint
, tp
->rcv_wnd
/ 2);
439 hint
= min(hint
, TCP_MSS_DEFAULT
);
440 hint
= max(hint
, TCP_MIN_MSS
);
442 inet_csk(sk
)->icsk_ack
.rcv_mss
= hint
;
444 EXPORT_SYMBOL(tcp_initialize_rcv_mss
);
446 /* Receiver "autotuning" code.
448 * The algorithm for RTT estimation w/o timestamps is based on
449 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
450 * <http://public.lanl.gov/radiant/pubs.html#DRS>
452 * More detail on this code can be found at
453 * <http://staff.psc.edu/jheffner/>,
454 * though this reference is out of date. A new paper
457 static void tcp_rcv_rtt_update(struct tcp_sock
*tp
, u32 sample
, int win_dep
)
459 u32 new_sample
= tp
->rcv_rtt_est
.rtt
;
465 if (new_sample
!= 0) {
466 /* If we sample in larger samples in the non-timestamp
467 * case, we could grossly overestimate the RTT especially
468 * with chatty applications or bulk transfer apps which
469 * are stalled on filesystem I/O.
471 * Also, since we are only going for a minimum in the
472 * non-timestamp case, we do not smooth things out
473 * else with timestamps disabled convergence takes too
477 m
-= (new_sample
>> 3);
485 /* No previous measure. */
489 if (tp
->rcv_rtt_est
.rtt
!= new_sample
)
490 tp
->rcv_rtt_est
.rtt
= new_sample
;
493 static inline void tcp_rcv_rtt_measure(struct tcp_sock
*tp
)
495 if (tp
->rcv_rtt_est
.time
== 0)
497 if (before(tp
->rcv_nxt
, tp
->rcv_rtt_est
.seq
))
499 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rcv_rtt_est
.time
, 1);
502 tp
->rcv_rtt_est
.seq
= tp
->rcv_nxt
+ tp
->rcv_wnd
;
503 tp
->rcv_rtt_est
.time
= tcp_time_stamp
;
506 static inline void tcp_rcv_rtt_measure_ts(struct sock
*sk
,
507 const struct sk_buff
*skb
)
509 struct tcp_sock
*tp
= tcp_sk(sk
);
510 if (tp
->rx_opt
.rcv_tsecr
&&
511 (TCP_SKB_CB(skb
)->end_seq
-
512 TCP_SKB_CB(skb
)->seq
>= inet_csk(sk
)->icsk_ack
.rcv_mss
))
513 tcp_rcv_rtt_update(tp
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
, 0);
517 * This function should be called every time data is copied to user space.
518 * It calculates the appropriate TCP receive buffer space.
520 void tcp_rcv_space_adjust(struct sock
*sk
)
522 struct tcp_sock
*tp
= tcp_sk(sk
);
526 if (tp
->rcvq_space
.time
== 0)
529 time
= tcp_time_stamp
- tp
->rcvq_space
.time
;
530 if (time
< (tp
->rcv_rtt_est
.rtt
>> 3) || tp
->rcv_rtt_est
.rtt
== 0)
533 space
= 2 * (tp
->copied_seq
- tp
->rcvq_space
.seq
);
535 space
= max(tp
->rcvq_space
.space
, space
);
537 if (tp
->rcvq_space
.space
!= space
) {
540 tp
->rcvq_space
.space
= space
;
542 if (sysctl_tcp_moderate_rcvbuf
&&
543 !(sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
)) {
544 int new_clamp
= space
;
546 /* Receive space grows, normalize in order to
547 * take into account packet headers and sk_buff
548 * structure overhead.
553 rcvmem
= SKB_TRUESIZE(tp
->advmss
+ MAX_TCP_HEADER
);
554 while (tcp_win_from_space(rcvmem
) < tp
->advmss
)
557 space
= min(space
, sysctl_tcp_rmem
[2]);
558 if (space
> sk
->sk_rcvbuf
) {
559 sk
->sk_rcvbuf
= space
;
561 /* Make the window clamp follow along. */
562 tp
->window_clamp
= new_clamp
;
568 tp
->rcvq_space
.seq
= tp
->copied_seq
;
569 tp
->rcvq_space
.time
= tcp_time_stamp
;
572 /* There is something which you must keep in mind when you analyze the
573 * behavior of the tp->ato delayed ack timeout interval. When a
574 * connection starts up, we want to ack as quickly as possible. The
575 * problem is that "good" TCP's do slow start at the beginning of data
576 * transmission. The means that until we send the first few ACK's the
577 * sender will sit on his end and only queue most of his data, because
578 * he can only send snd_cwnd unacked packets at any given time. For
579 * each ACK we send, he increments snd_cwnd and transmits more of his
582 static void tcp_event_data_recv(struct sock
*sk
, struct sk_buff
*skb
)
584 struct tcp_sock
*tp
= tcp_sk(sk
);
585 struct inet_connection_sock
*icsk
= inet_csk(sk
);
588 inet_csk_schedule_ack(sk
);
590 tcp_measure_rcv_mss(sk
, skb
);
592 tcp_rcv_rtt_measure(tp
);
594 now
= tcp_time_stamp
;
596 if (!icsk
->icsk_ack
.ato
) {
597 /* The _first_ data packet received, initialize
598 * delayed ACK engine.
600 tcp_incr_quickack(sk
);
601 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
603 int m
= now
- icsk
->icsk_ack
.lrcvtime
;
605 if (m
<= TCP_ATO_MIN
/ 2) {
606 /* The fastest case is the first. */
607 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + TCP_ATO_MIN
/ 2;
608 } else if (m
< icsk
->icsk_ack
.ato
) {
609 icsk
->icsk_ack
.ato
= (icsk
->icsk_ack
.ato
>> 1) + m
;
610 if (icsk
->icsk_ack
.ato
> icsk
->icsk_rto
)
611 icsk
->icsk_ack
.ato
= icsk
->icsk_rto
;
612 } else if (m
> icsk
->icsk_rto
) {
613 /* Too long gap. Apparently sender failed to
614 * restart window, so that we send ACKs quickly.
616 tcp_incr_quickack(sk
);
620 icsk
->icsk_ack
.lrcvtime
= now
;
622 TCP_ECN_check_ce(tp
, skb
);
625 tcp_grow_window(sk
, skb
);
628 /* Called to compute a smoothed rtt estimate. The data fed to this
629 * routine either comes from timestamps, or from segments that were
630 * known _not_ to have been retransmitted [see Karn/Partridge
631 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
632 * piece by Van Jacobson.
633 * NOTE: the next three routines used to be one big routine.
634 * To save cycles in the RFC 1323 implementation it was better to break
635 * it up into three procedures. -- erics
637 static void tcp_rtt_estimator(struct sock
*sk
, const __u32 mrtt
)
639 struct tcp_sock
*tp
= tcp_sk(sk
);
640 long m
= mrtt
; /* RTT */
642 /* The following amusing code comes from Jacobson's
643 * article in SIGCOMM '88. Note that rtt and mdev
644 * are scaled versions of rtt and mean deviation.
645 * This is designed to be as fast as possible
646 * m stands for "measurement".
648 * On a 1990 paper the rto value is changed to:
649 * RTO = rtt + 4 * mdev
651 * Funny. This algorithm seems to be very broken.
652 * These formulae increase RTO, when it should be decreased, increase
653 * too slowly, when it should be increased quickly, decrease too quickly
654 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
655 * does not matter how to _calculate_ it. Seems, it was trap
656 * that VJ failed to avoid. 8)
661 m
-= (tp
->srtt
>> 3); /* m is now error in rtt est */
662 tp
->srtt
+= m
; /* rtt = 7/8 rtt + 1/8 new */
664 m
= -m
; /* m is now abs(error) */
665 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
666 /* This is similar to one of Eifel findings.
667 * Eifel blocks mdev updates when rtt decreases.
668 * This solution is a bit different: we use finer gain
669 * for mdev in this case (alpha*beta).
670 * Like Eifel it also prevents growth of rto,
671 * but also it limits too fast rto decreases,
672 * happening in pure Eifel.
677 m
-= (tp
->mdev
>> 2); /* similar update on mdev */
679 tp
->mdev
+= m
; /* mdev = 3/4 mdev + 1/4 new */
680 if (tp
->mdev
> tp
->mdev_max
) {
681 tp
->mdev_max
= tp
->mdev
;
682 if (tp
->mdev_max
> tp
->rttvar
)
683 tp
->rttvar
= tp
->mdev_max
;
685 if (after(tp
->snd_una
, tp
->rtt_seq
)) {
686 if (tp
->mdev_max
< tp
->rttvar
)
687 tp
->rttvar
-= (tp
->rttvar
- tp
->mdev_max
) >> 2;
688 tp
->rtt_seq
= tp
->snd_nxt
;
689 tp
->mdev_max
= tcp_rto_min(sk
);
692 /* no previous measure. */
693 tp
->srtt
= m
<< 3; /* take the measured time to be rtt */
694 tp
->mdev
= m
<< 1; /* make sure rto = 3*rtt */
695 tp
->mdev_max
= tp
->rttvar
= max(tp
->mdev
, tcp_rto_min(sk
));
696 tp
->rtt_seq
= tp
->snd_nxt
;
700 /* Calculate rto without backoff. This is the second half of Van Jacobson's
701 * routine referred to above.
703 void tcp_set_rto(struct sock
*sk
)
705 const struct tcp_sock
*tp
= tcp_sk(sk
);
706 /* Old crap is replaced with new one. 8)
709 * 1. If rtt variance happened to be less 50msec, it is hallucination.
710 * It cannot be less due to utterly erratic ACK generation made
711 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
712 * to do with delayed acks, because at cwnd>2 true delack timeout
713 * is invisible. Actually, Linux-2.4 also generates erratic
714 * ACKs in some circumstances.
716 inet_csk(sk
)->icsk_rto
= __tcp_set_rto(tp
);
718 /* 2. Fixups made earlier cannot be right.
719 * If we do not estimate RTO correctly without them,
720 * all the algo is pure shit and should be replaced
721 * with correct one. It is exactly, which we pretend to do.
724 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
725 * guarantees that rto is higher.
730 __u32
tcp_init_cwnd(const struct tcp_sock
*tp
, const struct dst_entry
*dst
)
732 __u32 cwnd
= (dst
? dst_metric(dst
, RTAX_INITCWND
) : 0);
735 cwnd
= TCP_INIT_CWND
;
736 return min_t(__u32
, cwnd
, tp
->snd_cwnd_clamp
);
740 * Packet counting of FACK is based on in-order assumptions, therefore TCP
741 * disables it when reordering is detected
743 void tcp_disable_fack(struct tcp_sock
*tp
)
745 /* RFC3517 uses different metric in lost marker => reset on change */
747 tp
->lost_skb_hint
= NULL
;
748 tp
->rx_opt
.sack_ok
&= ~TCP_FACK_ENABLED
;
751 /* Take a notice that peer is sending D-SACKs */
752 static void tcp_dsack_seen(struct tcp_sock
*tp
)
754 tp
->rx_opt
.sack_ok
|= TCP_DSACK_SEEN
;
757 static void tcp_update_reordering(struct sock
*sk
, const int metric
,
760 struct tcp_sock
*tp
= tcp_sk(sk
);
761 if (metric
> tp
->reordering
) {
764 tp
->reordering
= min(TCP_MAX_REORDERING
, metric
);
766 /* This exciting event is worth to be remembered. 8) */
768 mib_idx
= LINUX_MIB_TCPTSREORDER
;
769 else if (tcp_is_reno(tp
))
770 mib_idx
= LINUX_MIB_TCPRENOREORDER
;
771 else if (tcp_is_fack(tp
))
772 mib_idx
= LINUX_MIB_TCPFACKREORDER
;
774 mib_idx
= LINUX_MIB_TCPSACKREORDER
;
776 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
777 #if FASTRETRANS_DEBUG > 1
778 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
779 tp
->rx_opt
.sack_ok
, inet_csk(sk
)->icsk_ca_state
,
783 tp
->undo_marker
? tp
->undo_retrans
: 0);
785 tcp_disable_fack(tp
);
789 tcp_disable_early_retrans(tp
);
792 /* This must be called before lost_out is incremented */
793 static void tcp_verify_retransmit_hint(struct tcp_sock
*tp
, struct sk_buff
*skb
)
795 if ((tp
->retransmit_skb_hint
== NULL
) ||
796 before(TCP_SKB_CB(skb
)->seq
,
797 TCP_SKB_CB(tp
->retransmit_skb_hint
)->seq
))
798 tp
->retransmit_skb_hint
= skb
;
801 after(TCP_SKB_CB(skb
)->end_seq
, tp
->retransmit_high
))
802 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
805 static void tcp_skb_mark_lost(struct tcp_sock
*tp
, struct sk_buff
*skb
)
807 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
808 tcp_verify_retransmit_hint(tp
, skb
);
810 tp
->lost_out
+= tcp_skb_pcount(skb
);
811 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
815 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock
*tp
,
818 tcp_verify_retransmit_hint(tp
, skb
);
820 if (!(TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_ACKED
))) {
821 tp
->lost_out
+= tcp_skb_pcount(skb
);
822 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
826 /* This procedure tags the retransmission queue when SACKs arrive.
828 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
829 * Packets in queue with these bits set are counted in variables
830 * sacked_out, retrans_out and lost_out, correspondingly.
832 * Valid combinations are:
833 * Tag InFlight Description
834 * 0 1 - orig segment is in flight.
835 * S 0 - nothing flies, orig reached receiver.
836 * L 0 - nothing flies, orig lost by net.
837 * R 2 - both orig and retransmit are in flight.
838 * L|R 1 - orig is lost, retransmit is in flight.
839 * S|R 1 - orig reached receiver, retrans is still in flight.
840 * (L|S|R is logically valid, it could occur when L|R is sacked,
841 * but it is equivalent to plain S and code short-curcuits it to S.
842 * L|S is logically invalid, it would mean -1 packet in flight 8))
844 * These 6 states form finite state machine, controlled by the following events:
845 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
846 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
847 * 3. Loss detection event of two flavors:
848 * A. Scoreboard estimator decided the packet is lost.
849 * A'. Reno "three dupacks" marks head of queue lost.
850 * A''. Its FACK modification, head until snd.fack is lost.
851 * B. SACK arrives sacking SND.NXT at the moment, when the
852 * segment was retransmitted.
853 * 4. D-SACK added new rule: D-SACK changes any tag to S.
855 * It is pleasant to note, that state diagram turns out to be commutative,
856 * so that we are allowed not to be bothered by order of our actions,
857 * when multiple events arrive simultaneously. (see the function below).
859 * Reordering detection.
860 * --------------------
861 * Reordering metric is maximal distance, which a packet can be displaced
862 * in packet stream. With SACKs we can estimate it:
864 * 1. SACK fills old hole and the corresponding segment was not
865 * ever retransmitted -> reordering. Alas, we cannot use it
866 * when segment was retransmitted.
867 * 2. The last flaw is solved with D-SACK. D-SACK arrives
868 * for retransmitted and already SACKed segment -> reordering..
869 * Both of these heuristics are not used in Loss state, when we cannot
870 * account for retransmits accurately.
872 * SACK block validation.
873 * ----------------------
875 * SACK block range validation checks that the received SACK block fits to
876 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
877 * Note that SND.UNA is not included to the range though being valid because
878 * it means that the receiver is rather inconsistent with itself reporting
879 * SACK reneging when it should advance SND.UNA. Such SACK block this is
880 * perfectly valid, however, in light of RFC2018 which explicitly states
881 * that "SACK block MUST reflect the newest segment. Even if the newest
882 * segment is going to be discarded ...", not that it looks very clever
883 * in case of head skb. Due to potentional receiver driven attacks, we
884 * choose to avoid immediate execution of a walk in write queue due to
885 * reneging and defer head skb's loss recovery to standard loss recovery
886 * procedure that will eventually trigger (nothing forbids us doing this).
888 * Implements also blockage to start_seq wrap-around. Problem lies in the
889 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
890 * there's no guarantee that it will be before snd_nxt (n). The problem
891 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
894 * <- outs wnd -> <- wrapzone ->
895 * u e n u_w e_w s n_w
897 * |<------------+------+----- TCP seqno space --------------+---------->|
898 * ...-- <2^31 ->| |<--------...
899 * ...---- >2^31 ------>| |<--------...
901 * Current code wouldn't be vulnerable but it's better still to discard such
902 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
903 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
904 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
905 * equal to the ideal case (infinite seqno space without wrap caused issues).
907 * With D-SACK the lower bound is extended to cover sequence space below
908 * SND.UNA down to undo_marker, which is the last point of interest. Yet
909 * again, D-SACK block must not to go across snd_una (for the same reason as
910 * for the normal SACK blocks, explained above). But there all simplicity
911 * ends, TCP might receive valid D-SACKs below that. As long as they reside
912 * fully below undo_marker they do not affect behavior in anyway and can
913 * therefore be safely ignored. In rare cases (which are more or less
914 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
915 * fragmentation and packet reordering past skb's retransmission. To consider
916 * them correctly, the acceptable range must be extended even more though
917 * the exact amount is rather hard to quantify. However, tp->max_window can
918 * be used as an exaggerated estimate.
920 static bool tcp_is_sackblock_valid(struct tcp_sock
*tp
, bool is_dsack
,
921 u32 start_seq
, u32 end_seq
)
923 /* Too far in future, or reversed (interpretation is ambiguous) */
924 if (after(end_seq
, tp
->snd_nxt
) || !before(start_seq
, end_seq
))
927 /* Nasty start_seq wrap-around check (see comments above) */
928 if (!before(start_seq
, tp
->snd_nxt
))
931 /* In outstanding window? ...This is valid exit for D-SACKs too.
932 * start_seq == snd_una is non-sensical (see comments above)
934 if (after(start_seq
, tp
->snd_una
))
937 if (!is_dsack
|| !tp
->undo_marker
)
940 /* ...Then it's D-SACK, and must reside below snd_una completely */
941 if (after(end_seq
, tp
->snd_una
))
944 if (!before(start_seq
, tp
->undo_marker
))
948 if (!after(end_seq
, tp
->undo_marker
))
951 /* Undo_marker boundary crossing (overestimates a lot). Known already:
952 * start_seq < undo_marker and end_seq >= undo_marker.
954 return !before(start_seq
, end_seq
- tp
->max_window
);
957 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
958 * Event "B". Later note: FACK people cheated me again 8), we have to account
959 * for reordering! Ugly, but should help.
961 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
962 * less than what is now known to be received by the other end (derived from
963 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
964 * retransmitted skbs to avoid some costly processing per ACKs.
966 static void tcp_mark_lost_retrans(struct sock
*sk
)
968 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
969 struct tcp_sock
*tp
= tcp_sk(sk
);
972 u32 new_low_seq
= tp
->snd_nxt
;
973 u32 received_upto
= tcp_highest_sack_seq(tp
);
975 if (!tcp_is_fack(tp
) || !tp
->retrans_out
||
976 !after(received_upto
, tp
->lost_retrans_low
) ||
977 icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
980 tcp_for_write_queue(skb
, sk
) {
981 u32 ack_seq
= TCP_SKB_CB(skb
)->ack_seq
;
983 if (skb
== tcp_send_head(sk
))
985 if (cnt
== tp
->retrans_out
)
987 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
990 if (!(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
))
993 /* TODO: We would like to get rid of tcp_is_fack(tp) only
994 * constraint here (see above) but figuring out that at
995 * least tp->reordering SACK blocks reside between ack_seq
996 * and received_upto is not easy task to do cheaply with
997 * the available datastructures.
999 * Whether FACK should check here for tp->reordering segs
1000 * in-between one could argue for either way (it would be
1001 * rather simple to implement as we could count fack_count
1002 * during the walk and do tp->fackets_out - fack_count).
1004 if (after(received_upto
, ack_seq
)) {
1005 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
1006 tp
->retrans_out
-= tcp_skb_pcount(skb
);
1008 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
1009 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSTRETRANSMIT
);
1011 if (before(ack_seq
, new_low_seq
))
1012 new_low_seq
= ack_seq
;
1013 cnt
+= tcp_skb_pcount(skb
);
1017 if (tp
->retrans_out
)
1018 tp
->lost_retrans_low
= new_low_seq
;
1021 static bool tcp_check_dsack(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1022 struct tcp_sack_block_wire
*sp
, int num_sacks
,
1025 struct tcp_sock
*tp
= tcp_sk(sk
);
1026 u32 start_seq_0
= get_unaligned_be32(&sp
[0].start_seq
);
1027 u32 end_seq_0
= get_unaligned_be32(&sp
[0].end_seq
);
1028 bool dup_sack
= false;
1030 if (before(start_seq_0
, TCP_SKB_CB(ack_skb
)->ack_seq
)) {
1033 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKRECV
);
1034 } else if (num_sacks
> 1) {
1035 u32 end_seq_1
= get_unaligned_be32(&sp
[1].end_seq
);
1036 u32 start_seq_1
= get_unaligned_be32(&sp
[1].start_seq
);
1038 if (!after(end_seq_0
, end_seq_1
) &&
1039 !before(start_seq_0
, start_seq_1
)) {
1042 NET_INC_STATS_BH(sock_net(sk
),
1043 LINUX_MIB_TCPDSACKOFORECV
);
1047 /* D-SACK for already forgotten data... Do dumb counting. */
1048 if (dup_sack
&& tp
->undo_marker
&& tp
->undo_retrans
&&
1049 !after(end_seq_0
, prior_snd_una
) &&
1050 after(end_seq_0
, tp
->undo_marker
))
1056 struct tcp_sacktag_state
{
1062 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1063 * the incoming SACK may not exactly match but we can find smaller MSS
1064 * aligned portion of it that matches. Therefore we might need to fragment
1065 * which may fail and creates some hassle (caller must handle error case
1068 * FIXME: this could be merged to shift decision code
1070 static int tcp_match_skb_to_sack(struct sock
*sk
, struct sk_buff
*skb
,
1071 u32 start_seq
, u32 end_seq
)
1075 unsigned int pkt_len
;
1078 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1079 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1081 if (tcp_skb_pcount(skb
) > 1 && !in_sack
&&
1082 after(TCP_SKB_CB(skb
)->end_seq
, start_seq
)) {
1083 mss
= tcp_skb_mss(skb
);
1084 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1087 pkt_len
= start_seq
- TCP_SKB_CB(skb
)->seq
;
1091 pkt_len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1096 /* Round if necessary so that SACKs cover only full MSSes
1097 * and/or the remaining small portion (if present)
1099 if (pkt_len
> mss
) {
1100 unsigned int new_len
= (pkt_len
/ mss
) * mss
;
1101 if (!in_sack
&& new_len
< pkt_len
) {
1103 if (new_len
> skb
->len
)
1108 err
= tcp_fragment(sk
, skb
, pkt_len
, mss
);
1116 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1117 static u8
tcp_sacktag_one(struct sock
*sk
,
1118 struct tcp_sacktag_state
*state
, u8 sacked
,
1119 u32 start_seq
, u32 end_seq
,
1120 bool dup_sack
, int pcount
)
1122 struct tcp_sock
*tp
= tcp_sk(sk
);
1123 int fack_count
= state
->fack_count
;
1125 /* Account D-SACK for retransmitted packet. */
1126 if (dup_sack
&& (sacked
& TCPCB_RETRANS
)) {
1127 if (tp
->undo_marker
&& tp
->undo_retrans
&&
1128 after(end_seq
, tp
->undo_marker
))
1130 if (sacked
& TCPCB_SACKED_ACKED
)
1131 state
->reord
= min(fack_count
, state
->reord
);
1134 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1135 if (!after(end_seq
, tp
->snd_una
))
1138 if (!(sacked
& TCPCB_SACKED_ACKED
)) {
1139 if (sacked
& TCPCB_SACKED_RETRANS
) {
1140 /* If the segment is not tagged as lost,
1141 * we do not clear RETRANS, believing
1142 * that retransmission is still in flight.
1144 if (sacked
& TCPCB_LOST
) {
1145 sacked
&= ~(TCPCB_LOST
|TCPCB_SACKED_RETRANS
);
1146 tp
->lost_out
-= pcount
;
1147 tp
->retrans_out
-= pcount
;
1150 if (!(sacked
& TCPCB_RETRANS
)) {
1151 /* New sack for not retransmitted frame,
1152 * which was in hole. It is reordering.
1154 if (before(start_seq
,
1155 tcp_highest_sack_seq(tp
)))
1156 state
->reord
= min(fack_count
,
1158 if (!after(end_seq
, tp
->high_seq
))
1159 state
->flag
|= FLAG_ORIG_SACK_ACKED
;
1162 if (sacked
& TCPCB_LOST
) {
1163 sacked
&= ~TCPCB_LOST
;
1164 tp
->lost_out
-= pcount
;
1168 sacked
|= TCPCB_SACKED_ACKED
;
1169 state
->flag
|= FLAG_DATA_SACKED
;
1170 tp
->sacked_out
+= pcount
;
1172 fack_count
+= pcount
;
1174 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1175 if (!tcp_is_fack(tp
) && (tp
->lost_skb_hint
!= NULL
) &&
1176 before(start_seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
))
1177 tp
->lost_cnt_hint
+= pcount
;
1179 if (fack_count
> tp
->fackets_out
)
1180 tp
->fackets_out
= fack_count
;
1183 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1184 * frames and clear it. undo_retrans is decreased above, L|R frames
1185 * are accounted above as well.
1187 if (dup_sack
&& (sacked
& TCPCB_SACKED_RETRANS
)) {
1188 sacked
&= ~TCPCB_SACKED_RETRANS
;
1189 tp
->retrans_out
-= pcount
;
1195 /* Shift newly-SACKed bytes from this skb to the immediately previous
1196 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1198 static bool tcp_shifted_skb(struct sock
*sk
, struct sk_buff
*skb
,
1199 struct tcp_sacktag_state
*state
,
1200 unsigned int pcount
, int shifted
, int mss
,
1203 struct tcp_sock
*tp
= tcp_sk(sk
);
1204 struct sk_buff
*prev
= tcp_write_queue_prev(sk
, skb
);
1205 u32 start_seq
= TCP_SKB_CB(skb
)->seq
; /* start of newly-SACKed */
1206 u32 end_seq
= start_seq
+ shifted
; /* end of newly-SACKed */
1210 /* Adjust counters and hints for the newly sacked sequence
1211 * range but discard the return value since prev is already
1212 * marked. We must tag the range first because the seq
1213 * advancement below implicitly advances
1214 * tcp_highest_sack_seq() when skb is highest_sack.
1216 tcp_sacktag_one(sk
, state
, TCP_SKB_CB(skb
)->sacked
,
1217 start_seq
, end_seq
, dup_sack
, pcount
);
1219 if (skb
== tp
->lost_skb_hint
)
1220 tp
->lost_cnt_hint
+= pcount
;
1222 TCP_SKB_CB(prev
)->end_seq
+= shifted
;
1223 TCP_SKB_CB(skb
)->seq
+= shifted
;
1225 skb_shinfo(prev
)->gso_segs
+= pcount
;
1226 BUG_ON(skb_shinfo(skb
)->gso_segs
< pcount
);
1227 skb_shinfo(skb
)->gso_segs
-= pcount
;
1229 /* When we're adding to gso_segs == 1, gso_size will be zero,
1230 * in theory this shouldn't be necessary but as long as DSACK
1231 * code can come after this skb later on it's better to keep
1232 * setting gso_size to something.
1234 if (!skb_shinfo(prev
)->gso_size
) {
1235 skb_shinfo(prev
)->gso_size
= mss
;
1236 skb_shinfo(prev
)->gso_type
= sk
->sk_gso_type
;
1239 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1240 if (skb_shinfo(skb
)->gso_segs
<= 1) {
1241 skb_shinfo(skb
)->gso_size
= 0;
1242 skb_shinfo(skb
)->gso_type
= 0;
1245 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1246 TCP_SKB_CB(prev
)->sacked
|= (TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
);
1249 BUG_ON(!tcp_skb_pcount(skb
));
1250 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTED
);
1254 /* Whole SKB was eaten :-) */
1256 if (skb
== tp
->retransmit_skb_hint
)
1257 tp
->retransmit_skb_hint
= prev
;
1258 if (skb
== tp
->lost_skb_hint
) {
1259 tp
->lost_skb_hint
= prev
;
1260 tp
->lost_cnt_hint
-= tcp_skb_pcount(prev
);
1263 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(prev
)->tcp_flags
;
1264 if (skb
== tcp_highest_sack(sk
))
1265 tcp_advance_highest_sack(sk
, skb
);
1267 tcp_unlink_write_queue(skb
, sk
);
1268 sk_wmem_free_skb(sk
, skb
);
1270 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKMERGED
);
1275 /* I wish gso_size would have a bit more sane initialization than
1276 * something-or-zero which complicates things
1278 static int tcp_skb_seglen(const struct sk_buff
*skb
)
1280 return tcp_skb_pcount(skb
) == 1 ? skb
->len
: tcp_skb_mss(skb
);
1283 /* Shifting pages past head area doesn't work */
1284 static int skb_can_shift(const struct sk_buff
*skb
)
1286 return !skb_headlen(skb
) && skb_is_nonlinear(skb
);
1289 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1292 static struct sk_buff
*tcp_shift_skb_data(struct sock
*sk
, struct sk_buff
*skb
,
1293 struct tcp_sacktag_state
*state
,
1294 u32 start_seq
, u32 end_seq
,
1297 struct tcp_sock
*tp
= tcp_sk(sk
);
1298 struct sk_buff
*prev
;
1304 if (!sk_can_gso(sk
))
1307 /* Normally R but no L won't result in plain S */
1309 (TCP_SKB_CB(skb
)->sacked
& (TCPCB_LOST
|TCPCB_SACKED_RETRANS
)) == TCPCB_SACKED_RETRANS
)
1311 if (!skb_can_shift(skb
))
1313 /* This frame is about to be dropped (was ACKed). */
1314 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
1317 /* Can only happen with delayed DSACK + discard craziness */
1318 if (unlikely(skb
== tcp_write_queue_head(sk
)))
1320 prev
= tcp_write_queue_prev(sk
, skb
);
1322 if ((TCP_SKB_CB(prev
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
)
1325 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
) &&
1326 !before(end_seq
, TCP_SKB_CB(skb
)->end_seq
);
1330 pcount
= tcp_skb_pcount(skb
);
1331 mss
= tcp_skb_seglen(skb
);
1333 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1334 * drop this restriction as unnecessary
1336 if (mss
!= tcp_skb_seglen(prev
))
1339 if (!after(TCP_SKB_CB(skb
)->end_seq
, start_seq
))
1341 /* CHECKME: This is non-MSS split case only?, this will
1342 * cause skipped skbs due to advancing loop btw, original
1343 * has that feature too
1345 if (tcp_skb_pcount(skb
) <= 1)
1348 in_sack
= !after(start_seq
, TCP_SKB_CB(skb
)->seq
);
1350 /* TODO: head merge to next could be attempted here
1351 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1352 * though it might not be worth of the additional hassle
1354 * ...we can probably just fallback to what was done
1355 * previously. We could try merging non-SACKed ones
1356 * as well but it probably isn't going to buy off
1357 * because later SACKs might again split them, and
1358 * it would make skb timestamp tracking considerably
1364 len
= end_seq
- TCP_SKB_CB(skb
)->seq
;
1366 BUG_ON(len
> skb
->len
);
1368 /* MSS boundaries should be honoured or else pcount will
1369 * severely break even though it makes things bit trickier.
1370 * Optimize common case to avoid most of the divides
1372 mss
= tcp_skb_mss(skb
);
1374 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1375 * drop this restriction as unnecessary
1377 if (mss
!= tcp_skb_seglen(prev
))
1382 } else if (len
< mss
) {
1390 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1391 if (!after(TCP_SKB_CB(skb
)->seq
+ len
, tp
->snd_una
))
1394 if (!skb_shift(prev
, skb
, len
))
1396 if (!tcp_shifted_skb(sk
, skb
, state
, pcount
, len
, mss
, dup_sack
))
1399 /* Hole filled allows collapsing with the next as well, this is very
1400 * useful when hole on every nth skb pattern happens
1402 if (prev
== tcp_write_queue_tail(sk
))
1404 skb
= tcp_write_queue_next(sk
, prev
);
1406 if (!skb_can_shift(skb
) ||
1407 (skb
== tcp_send_head(sk
)) ||
1408 ((TCP_SKB_CB(skb
)->sacked
& TCPCB_TAGBITS
) != TCPCB_SACKED_ACKED
) ||
1409 (mss
!= tcp_skb_seglen(skb
)))
1413 if (skb_shift(prev
, skb
, len
)) {
1414 pcount
+= tcp_skb_pcount(skb
);
1415 tcp_shifted_skb(sk
, skb
, state
, tcp_skb_pcount(skb
), len
, mss
, 0);
1419 state
->fack_count
+= pcount
;
1426 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_SACKSHIFTFALLBACK
);
1430 static struct sk_buff
*tcp_sacktag_walk(struct sk_buff
*skb
, struct sock
*sk
,
1431 struct tcp_sack_block
*next_dup
,
1432 struct tcp_sacktag_state
*state
,
1433 u32 start_seq
, u32 end_seq
,
1436 struct tcp_sock
*tp
= tcp_sk(sk
);
1437 struct sk_buff
*tmp
;
1439 tcp_for_write_queue_from(skb
, sk
) {
1441 bool dup_sack
= dup_sack_in
;
1443 if (skb
== tcp_send_head(sk
))
1446 /* queue is in-order => we can short-circuit the walk early */
1447 if (!before(TCP_SKB_CB(skb
)->seq
, end_seq
))
1450 if ((next_dup
!= NULL
) &&
1451 before(TCP_SKB_CB(skb
)->seq
, next_dup
->end_seq
)) {
1452 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1453 next_dup
->start_seq
,
1459 /* skb reference here is a bit tricky to get right, since
1460 * shifting can eat and free both this skb and the next,
1461 * so not even _safe variant of the loop is enough.
1464 tmp
= tcp_shift_skb_data(sk
, skb
, state
,
1465 start_seq
, end_seq
, dup_sack
);
1474 in_sack
= tcp_match_skb_to_sack(sk
, skb
,
1480 if (unlikely(in_sack
< 0))
1484 TCP_SKB_CB(skb
)->sacked
=
1487 TCP_SKB_CB(skb
)->sacked
,
1488 TCP_SKB_CB(skb
)->seq
,
1489 TCP_SKB_CB(skb
)->end_seq
,
1491 tcp_skb_pcount(skb
));
1493 if (!before(TCP_SKB_CB(skb
)->seq
,
1494 tcp_highest_sack_seq(tp
)))
1495 tcp_advance_highest_sack(sk
, skb
);
1498 state
->fack_count
+= tcp_skb_pcount(skb
);
1503 /* Avoid all extra work that is being done by sacktag while walking in
1506 static struct sk_buff
*tcp_sacktag_skip(struct sk_buff
*skb
, struct sock
*sk
,
1507 struct tcp_sacktag_state
*state
,
1510 tcp_for_write_queue_from(skb
, sk
) {
1511 if (skb
== tcp_send_head(sk
))
1514 if (after(TCP_SKB_CB(skb
)->end_seq
, skip_to_seq
))
1517 state
->fack_count
+= tcp_skb_pcount(skb
);
1522 static struct sk_buff
*tcp_maybe_skipping_dsack(struct sk_buff
*skb
,
1524 struct tcp_sack_block
*next_dup
,
1525 struct tcp_sacktag_state
*state
,
1528 if (next_dup
== NULL
)
1531 if (before(next_dup
->start_seq
, skip_to_seq
)) {
1532 skb
= tcp_sacktag_skip(skb
, sk
, state
, next_dup
->start_seq
);
1533 skb
= tcp_sacktag_walk(skb
, sk
, NULL
, state
,
1534 next_dup
->start_seq
, next_dup
->end_seq
,
1541 static int tcp_sack_cache_ok(const struct tcp_sock
*tp
, const struct tcp_sack_block
*cache
)
1543 return cache
< tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1547 tcp_sacktag_write_queue(struct sock
*sk
, const struct sk_buff
*ack_skb
,
1550 struct tcp_sock
*tp
= tcp_sk(sk
);
1551 const unsigned char *ptr
= (skb_transport_header(ack_skb
) +
1552 TCP_SKB_CB(ack_skb
)->sacked
);
1553 struct tcp_sack_block_wire
*sp_wire
= (struct tcp_sack_block_wire
*)(ptr
+2);
1554 struct tcp_sack_block sp
[TCP_NUM_SACKS
];
1555 struct tcp_sack_block
*cache
;
1556 struct tcp_sacktag_state state
;
1557 struct sk_buff
*skb
;
1558 int num_sacks
= min(TCP_NUM_SACKS
, (ptr
[1] - TCPOLEN_SACK_BASE
) >> 3);
1560 bool found_dup_sack
= false;
1562 int first_sack_index
;
1565 state
.reord
= tp
->packets_out
;
1567 if (!tp
->sacked_out
) {
1568 if (WARN_ON(tp
->fackets_out
))
1569 tp
->fackets_out
= 0;
1570 tcp_highest_sack_reset(sk
);
1573 found_dup_sack
= tcp_check_dsack(sk
, ack_skb
, sp_wire
,
1574 num_sacks
, prior_snd_una
);
1576 state
.flag
|= FLAG_DSACKING_ACK
;
1578 /* Eliminate too old ACKs, but take into
1579 * account more or less fresh ones, they can
1580 * contain valid SACK info.
1582 if (before(TCP_SKB_CB(ack_skb
)->ack_seq
, prior_snd_una
- tp
->max_window
))
1585 if (!tp
->packets_out
)
1589 first_sack_index
= 0;
1590 for (i
= 0; i
< num_sacks
; i
++) {
1591 bool dup_sack
= !i
&& found_dup_sack
;
1593 sp
[used_sacks
].start_seq
= get_unaligned_be32(&sp_wire
[i
].start_seq
);
1594 sp
[used_sacks
].end_seq
= get_unaligned_be32(&sp_wire
[i
].end_seq
);
1596 if (!tcp_is_sackblock_valid(tp
, dup_sack
,
1597 sp
[used_sacks
].start_seq
,
1598 sp
[used_sacks
].end_seq
)) {
1602 if (!tp
->undo_marker
)
1603 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDNOUNDO
;
1605 mib_idx
= LINUX_MIB_TCPDSACKIGNOREDOLD
;
1607 /* Don't count olds caused by ACK reordering */
1608 if ((TCP_SKB_CB(ack_skb
)->ack_seq
!= tp
->snd_una
) &&
1609 !after(sp
[used_sacks
].end_seq
, tp
->snd_una
))
1611 mib_idx
= LINUX_MIB_TCPSACKDISCARD
;
1614 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
1616 first_sack_index
= -1;
1620 /* Ignore very old stuff early */
1621 if (!after(sp
[used_sacks
].end_seq
, prior_snd_una
))
1627 /* order SACK blocks to allow in order walk of the retrans queue */
1628 for (i
= used_sacks
- 1; i
> 0; i
--) {
1629 for (j
= 0; j
< i
; j
++) {
1630 if (after(sp
[j
].start_seq
, sp
[j
+ 1].start_seq
)) {
1631 swap(sp
[j
], sp
[j
+ 1]);
1633 /* Track where the first SACK block goes to */
1634 if (j
== first_sack_index
)
1635 first_sack_index
= j
+ 1;
1640 skb
= tcp_write_queue_head(sk
);
1641 state
.fack_count
= 0;
1644 if (!tp
->sacked_out
) {
1645 /* It's already past, so skip checking against it */
1646 cache
= tp
->recv_sack_cache
+ ARRAY_SIZE(tp
->recv_sack_cache
);
1648 cache
= tp
->recv_sack_cache
;
1649 /* Skip empty blocks in at head of the cache */
1650 while (tcp_sack_cache_ok(tp
, cache
) && !cache
->start_seq
&&
1655 while (i
< used_sacks
) {
1656 u32 start_seq
= sp
[i
].start_seq
;
1657 u32 end_seq
= sp
[i
].end_seq
;
1658 bool dup_sack
= (found_dup_sack
&& (i
== first_sack_index
));
1659 struct tcp_sack_block
*next_dup
= NULL
;
1661 if (found_dup_sack
&& ((i
+ 1) == first_sack_index
))
1662 next_dup
= &sp
[i
+ 1];
1664 /* Skip too early cached blocks */
1665 while (tcp_sack_cache_ok(tp
, cache
) &&
1666 !before(start_seq
, cache
->end_seq
))
1669 /* Can skip some work by looking recv_sack_cache? */
1670 if (tcp_sack_cache_ok(tp
, cache
) && !dup_sack
&&
1671 after(end_seq
, cache
->start_seq
)) {
1674 if (before(start_seq
, cache
->start_seq
)) {
1675 skb
= tcp_sacktag_skip(skb
, sk
, &state
,
1677 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
,
1684 /* Rest of the block already fully processed? */
1685 if (!after(end_seq
, cache
->end_seq
))
1688 skb
= tcp_maybe_skipping_dsack(skb
, sk
, next_dup
,
1692 /* ...tail remains todo... */
1693 if (tcp_highest_sack_seq(tp
) == cache
->end_seq
) {
1694 /* ...but better entrypoint exists! */
1695 skb
= tcp_highest_sack(sk
);
1698 state
.fack_count
= tp
->fackets_out
;
1703 skb
= tcp_sacktag_skip(skb
, sk
, &state
, cache
->end_seq
);
1704 /* Check overlap against next cached too (past this one already) */
1709 if (!before(start_seq
, tcp_highest_sack_seq(tp
))) {
1710 skb
= tcp_highest_sack(sk
);
1713 state
.fack_count
= tp
->fackets_out
;
1715 skb
= tcp_sacktag_skip(skb
, sk
, &state
, start_seq
);
1718 skb
= tcp_sacktag_walk(skb
, sk
, next_dup
, &state
,
1719 start_seq
, end_seq
, dup_sack
);
1725 /* Clear the head of the cache sack blocks so we can skip it next time */
1726 for (i
= 0; i
< ARRAY_SIZE(tp
->recv_sack_cache
) - used_sacks
; i
++) {
1727 tp
->recv_sack_cache
[i
].start_seq
= 0;
1728 tp
->recv_sack_cache
[i
].end_seq
= 0;
1730 for (j
= 0; j
< used_sacks
; j
++)
1731 tp
->recv_sack_cache
[i
++] = sp
[j
];
1733 tcp_mark_lost_retrans(sk
);
1735 tcp_verify_left_out(tp
);
1737 if ((state
.reord
< tp
->fackets_out
) &&
1738 ((inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Loss
) || tp
->undo_marker
))
1739 tcp_update_reordering(sk
, tp
->fackets_out
- state
.reord
, 0);
1743 #if FASTRETRANS_DEBUG > 0
1744 WARN_ON((int)tp
->sacked_out
< 0);
1745 WARN_ON((int)tp
->lost_out
< 0);
1746 WARN_ON((int)tp
->retrans_out
< 0);
1747 WARN_ON((int)tcp_packets_in_flight(tp
) < 0);
1752 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1753 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1755 static bool tcp_limit_reno_sacked(struct tcp_sock
*tp
)
1759 holes
= max(tp
->lost_out
, 1U);
1760 holes
= min(holes
, tp
->packets_out
);
1762 if ((tp
->sacked_out
+ holes
) > tp
->packets_out
) {
1763 tp
->sacked_out
= tp
->packets_out
- holes
;
1769 /* If we receive more dupacks than we expected counting segments
1770 * in assumption of absent reordering, interpret this as reordering.
1771 * The only another reason could be bug in receiver TCP.
1773 static void tcp_check_reno_reordering(struct sock
*sk
, const int addend
)
1775 struct tcp_sock
*tp
= tcp_sk(sk
);
1776 if (tcp_limit_reno_sacked(tp
))
1777 tcp_update_reordering(sk
, tp
->packets_out
+ addend
, 0);
1780 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1782 static void tcp_add_reno_sack(struct sock
*sk
)
1784 struct tcp_sock
*tp
= tcp_sk(sk
);
1786 tcp_check_reno_reordering(sk
, 0);
1787 tcp_verify_left_out(tp
);
1790 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1792 static void tcp_remove_reno_sacks(struct sock
*sk
, int acked
)
1794 struct tcp_sock
*tp
= tcp_sk(sk
);
1797 /* One ACK acked hole. The rest eat duplicate ACKs. */
1798 if (acked
- 1 >= tp
->sacked_out
)
1801 tp
->sacked_out
-= acked
- 1;
1803 tcp_check_reno_reordering(sk
, acked
);
1804 tcp_verify_left_out(tp
);
1807 static inline void tcp_reset_reno_sack(struct tcp_sock
*tp
)
1812 static void tcp_clear_retrans_partial(struct tcp_sock
*tp
)
1814 tp
->retrans_out
= 0;
1817 tp
->undo_marker
= 0;
1818 tp
->undo_retrans
= 0;
1821 void tcp_clear_retrans(struct tcp_sock
*tp
)
1823 tcp_clear_retrans_partial(tp
);
1825 tp
->fackets_out
= 0;
1829 /* Enter Loss state. If "how" is not zero, forget all SACK information
1830 * and reset tags completely, otherwise preserve SACKs. If receiver
1831 * dropped its ofo queue, we will know this due to reneging detection.
1833 void tcp_enter_loss(struct sock
*sk
, int how
)
1835 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1836 struct tcp_sock
*tp
= tcp_sk(sk
);
1837 struct sk_buff
*skb
;
1838 bool new_recovery
= false;
1840 /* Reduce ssthresh if it has not yet been made inside this window. */
1841 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
||
1842 !after(tp
->high_seq
, tp
->snd_una
) ||
1843 (icsk
->icsk_ca_state
== TCP_CA_Loss
&& !icsk
->icsk_retransmits
)) {
1844 new_recovery
= true;
1845 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
1846 tp
->snd_ssthresh
= icsk
->icsk_ca_ops
->ssthresh(sk
);
1847 tcp_ca_event(sk
, CA_EVENT_LOSS
);
1850 tp
->snd_cwnd_cnt
= 0;
1851 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1853 tcp_clear_retrans_partial(tp
);
1855 if (tcp_is_reno(tp
))
1856 tcp_reset_reno_sack(tp
);
1858 tp
->undo_marker
= tp
->snd_una
;
1861 tp
->fackets_out
= 0;
1863 tcp_clear_all_retrans_hints(tp
);
1865 tcp_for_write_queue(skb
, sk
) {
1866 if (skb
== tcp_send_head(sk
))
1869 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_RETRANS
)
1870 tp
->undo_marker
= 0;
1871 TCP_SKB_CB(skb
)->sacked
&= (~TCPCB_TAGBITS
)|TCPCB_SACKED_ACKED
;
1872 if (!(TCP_SKB_CB(skb
)->sacked
&TCPCB_SACKED_ACKED
) || how
) {
1873 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_ACKED
;
1874 TCP_SKB_CB(skb
)->sacked
|= TCPCB_LOST
;
1875 tp
->lost_out
+= tcp_skb_pcount(skb
);
1876 tp
->retransmit_high
= TCP_SKB_CB(skb
)->end_seq
;
1879 tcp_verify_left_out(tp
);
1881 tp
->reordering
= min_t(unsigned int, tp
->reordering
,
1882 sysctl_tcp_reordering
);
1883 tcp_set_ca_state(sk
, TCP_CA_Loss
);
1884 tp
->high_seq
= tp
->snd_nxt
;
1885 TCP_ECN_queue_cwr(tp
);
1887 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1888 * loss recovery is underway except recurring timeout(s) on
1889 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1891 tp
->frto
= sysctl_tcp_frto
&&
1892 (new_recovery
|| icsk
->icsk_retransmits
) &&
1893 !inet_csk(sk
)->icsk_mtup
.probe_size
;
1896 /* If ACK arrived pointing to a remembered SACK, it means that our
1897 * remembered SACKs do not reflect real state of receiver i.e.
1898 * receiver _host_ is heavily congested (or buggy).
1900 * Do processing similar to RTO timeout.
1902 static bool tcp_check_sack_reneging(struct sock
*sk
, int flag
)
1904 if (flag
& FLAG_SACK_RENEGING
) {
1905 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1906 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSACKRENEGING
);
1908 tcp_enter_loss(sk
, 1);
1909 icsk
->icsk_retransmits
++;
1910 tcp_retransmit_skb(sk
, tcp_write_queue_head(sk
));
1911 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
1912 icsk
->icsk_rto
, TCP_RTO_MAX
);
1918 static inline int tcp_fackets_out(const struct tcp_sock
*tp
)
1920 return tcp_is_reno(tp
) ? tp
->sacked_out
+ 1 : tp
->fackets_out
;
1923 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1924 * counter when SACK is enabled (without SACK, sacked_out is used for
1927 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1928 * segments up to the highest received SACK block so far and holes in
1931 * With reordering, holes may still be in flight, so RFC3517 recovery
1932 * uses pure sacked_out (total number of SACKed segments) even though
1933 * it violates the RFC that uses duplicate ACKs, often these are equal
1934 * but when e.g. out-of-window ACKs or packet duplication occurs,
1935 * they differ. Since neither occurs due to loss, TCP should really
1938 static inline int tcp_dupack_heuristics(const struct tcp_sock
*tp
)
1940 return tcp_is_fack(tp
) ? tp
->fackets_out
: tp
->sacked_out
+ 1;
1943 static bool tcp_pause_early_retransmit(struct sock
*sk
, int flag
)
1945 struct tcp_sock
*tp
= tcp_sk(sk
);
1946 unsigned long delay
;
1948 /* Delay early retransmit and entering fast recovery for
1949 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1950 * available, or RTO is scheduled to fire first.
1952 if (sysctl_tcp_early_retrans
< 2 || sysctl_tcp_early_retrans
> 3 ||
1953 (flag
& FLAG_ECE
) || !tp
->srtt
)
1956 delay
= max_t(unsigned long, (tp
->srtt
>> 5), msecs_to_jiffies(2));
1957 if (!time_after(inet_csk(sk
)->icsk_timeout
, (jiffies
+ delay
)))
1960 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_EARLY_RETRANS
, delay
,
1965 /* Linux NewReno/SACK/FACK/ECN state machine.
1966 * --------------------------------------
1968 * "Open" Normal state, no dubious events, fast path.
1969 * "Disorder" In all the respects it is "Open",
1970 * but requires a bit more attention. It is entered when
1971 * we see some SACKs or dupacks. It is split of "Open"
1972 * mainly to move some processing from fast path to slow one.
1973 * "CWR" CWND was reduced due to some Congestion Notification event.
1974 * It can be ECN, ICMP source quench, local device congestion.
1975 * "Recovery" CWND was reduced, we are fast-retransmitting.
1976 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1978 * tcp_fastretrans_alert() is entered:
1979 * - each incoming ACK, if state is not "Open"
1980 * - when arrived ACK is unusual, namely:
1985 * Counting packets in flight is pretty simple.
1987 * in_flight = packets_out - left_out + retrans_out
1989 * packets_out is SND.NXT-SND.UNA counted in packets.
1991 * retrans_out is number of retransmitted segments.
1993 * left_out is number of segments left network, but not ACKed yet.
1995 * left_out = sacked_out + lost_out
1997 * sacked_out: Packets, which arrived to receiver out of order
1998 * and hence not ACKed. With SACKs this number is simply
1999 * amount of SACKed data. Even without SACKs
2000 * it is easy to give pretty reliable estimate of this number,
2001 * counting duplicate ACKs.
2003 * lost_out: Packets lost by network. TCP has no explicit
2004 * "loss notification" feedback from network (for now).
2005 * It means that this number can be only _guessed_.
2006 * Actually, it is the heuristics to predict lossage that
2007 * distinguishes different algorithms.
2009 * F.e. after RTO, when all the queue is considered as lost,
2010 * lost_out = packets_out and in_flight = retrans_out.
2012 * Essentially, we have now two algorithms counting
2015 * FACK: It is the simplest heuristics. As soon as we decided
2016 * that something is lost, we decide that _all_ not SACKed
2017 * packets until the most forward SACK are lost. I.e.
2018 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2019 * It is absolutely correct estimate, if network does not reorder
2020 * packets. And it loses any connection to reality when reordering
2021 * takes place. We use FACK by default until reordering
2022 * is suspected on the path to this destination.
2024 * NewReno: when Recovery is entered, we assume that one segment
2025 * is lost (classic Reno). While we are in Recovery and
2026 * a partial ACK arrives, we assume that one more packet
2027 * is lost (NewReno). This heuristics are the same in NewReno
2030 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2031 * deflation etc. CWND is real congestion window, never inflated, changes
2032 * only according to classic VJ rules.
2034 * Really tricky (and requiring careful tuning) part of algorithm
2035 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2036 * The first determines the moment _when_ we should reduce CWND and,
2037 * hence, slow down forward transmission. In fact, it determines the moment
2038 * when we decide that hole is caused by loss, rather than by a reorder.
2040 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2041 * holes, caused by lost packets.
2043 * And the most logically complicated part of algorithm is undo
2044 * heuristics. We detect false retransmits due to both too early
2045 * fast retransmit (reordering) and underestimated RTO, analyzing
2046 * timestamps and D-SACKs. When we detect that some segments were
2047 * retransmitted by mistake and CWND reduction was wrong, we undo
2048 * window reduction and abort recovery phase. This logic is hidden
2049 * inside several functions named tcp_try_undo_<something>.
2052 /* This function decides, when we should leave Disordered state
2053 * and enter Recovery phase, reducing congestion window.
2055 * Main question: may we further continue forward transmission
2056 * with the same cwnd?
2058 static bool tcp_time_to_recover(struct sock
*sk
, int flag
)
2060 struct tcp_sock
*tp
= tcp_sk(sk
);
2063 /* Trick#1: The loss is proven. */
2067 /* Not-A-Trick#2 : Classic rule... */
2068 if (tcp_dupack_heuristics(tp
) > tp
->reordering
)
2071 /* Trick#4: It is still not OK... But will it be useful to delay
2074 packets_out
= tp
->packets_out
;
2075 if (packets_out
<= tp
->reordering
&&
2076 tp
->sacked_out
>= max_t(__u32
, packets_out
/2, sysctl_tcp_reordering
) &&
2077 !tcp_may_send_now(sk
)) {
2078 /* We have nothing to send. This connection is limited
2079 * either by receiver window or by application.
2084 /* If a thin stream is detected, retransmit after first
2085 * received dupack. Employ only if SACK is supported in order
2086 * to avoid possible corner-case series of spurious retransmissions
2087 * Use only if there are no unsent data.
2089 if ((tp
->thin_dupack
|| sysctl_tcp_thin_dupack
) &&
2090 tcp_stream_is_thin(tp
) && tcp_dupack_heuristics(tp
) > 1 &&
2091 tcp_is_sack(tp
) && !tcp_send_head(sk
))
2094 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2095 * retransmissions due to small network reorderings, we implement
2096 * Mitigation A.3 in the RFC and delay the retransmission for a short
2097 * interval if appropriate.
2099 if (tp
->do_early_retrans
&& !tp
->retrans_out
&& tp
->sacked_out
&&
2100 (tp
->packets_out
>= (tp
->sacked_out
+ 1) && tp
->packets_out
< 4) &&
2101 !tcp_may_send_now(sk
))
2102 return !tcp_pause_early_retransmit(sk
, flag
);
2107 /* Detect loss in event "A" above by marking head of queue up as lost.
2108 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2109 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2110 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2111 * the maximum SACKed segments to pass before reaching this limit.
2113 static void tcp_mark_head_lost(struct sock
*sk
, int packets
, int mark_head
)
2115 struct tcp_sock
*tp
= tcp_sk(sk
);
2116 struct sk_buff
*skb
;
2120 /* Use SACK to deduce losses of new sequences sent during recovery */
2121 const u32 loss_high
= tcp_is_sack(tp
) ? tp
->snd_nxt
: tp
->high_seq
;
2123 WARN_ON(packets
> tp
->packets_out
);
2124 if (tp
->lost_skb_hint
) {
2125 skb
= tp
->lost_skb_hint
;
2126 cnt
= tp
->lost_cnt_hint
;
2127 /* Head already handled? */
2128 if (mark_head
&& skb
!= tcp_write_queue_head(sk
))
2131 skb
= tcp_write_queue_head(sk
);
2135 tcp_for_write_queue_from(skb
, sk
) {
2136 if (skb
== tcp_send_head(sk
))
2138 /* TODO: do this better */
2139 /* this is not the most efficient way to do this... */
2140 tp
->lost_skb_hint
= skb
;
2141 tp
->lost_cnt_hint
= cnt
;
2143 if (after(TCP_SKB_CB(skb
)->end_seq
, loss_high
))
2147 if (tcp_is_fack(tp
) || tcp_is_reno(tp
) ||
2148 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
2149 cnt
+= tcp_skb_pcount(skb
);
2151 if (cnt
> packets
) {
2152 if ((tcp_is_sack(tp
) && !tcp_is_fack(tp
)) ||
2153 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
) ||
2154 (oldcnt
>= packets
))
2157 mss
= skb_shinfo(skb
)->gso_size
;
2158 err
= tcp_fragment(sk
, skb
, (packets
- oldcnt
) * mss
, mss
);
2164 tcp_skb_mark_lost(tp
, skb
);
2169 tcp_verify_left_out(tp
);
2172 /* Account newly detected lost packet(s) */
2174 static void tcp_update_scoreboard(struct sock
*sk
, int fast_rexmit
)
2176 struct tcp_sock
*tp
= tcp_sk(sk
);
2178 if (tcp_is_reno(tp
)) {
2179 tcp_mark_head_lost(sk
, 1, 1);
2180 } else if (tcp_is_fack(tp
)) {
2181 int lost
= tp
->fackets_out
- tp
->reordering
;
2184 tcp_mark_head_lost(sk
, lost
, 0);
2186 int sacked_upto
= tp
->sacked_out
- tp
->reordering
;
2187 if (sacked_upto
>= 0)
2188 tcp_mark_head_lost(sk
, sacked_upto
, 0);
2189 else if (fast_rexmit
)
2190 tcp_mark_head_lost(sk
, 1, 1);
2194 /* CWND moderation, preventing bursts due to too big ACKs
2195 * in dubious situations.
2197 static inline void tcp_moderate_cwnd(struct tcp_sock
*tp
)
2199 tp
->snd_cwnd
= min(tp
->snd_cwnd
,
2200 tcp_packets_in_flight(tp
) + tcp_max_burst(tp
));
2201 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2204 /* Nothing was retransmitted or returned timestamp is less
2205 * than timestamp of the first retransmission.
2207 static inline bool tcp_packet_delayed(const struct tcp_sock
*tp
)
2209 return !tp
->retrans_stamp
||
2210 (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
2211 before(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
));
2214 /* Undo procedures. */
2216 #if FASTRETRANS_DEBUG > 1
2217 static void DBGUNDO(struct sock
*sk
, const char *msg
)
2219 struct tcp_sock
*tp
= tcp_sk(sk
);
2220 struct inet_sock
*inet
= inet_sk(sk
);
2222 if (sk
->sk_family
== AF_INET
) {
2223 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2225 &inet
->inet_daddr
, ntohs(inet
->inet_dport
),
2226 tp
->snd_cwnd
, tcp_left_out(tp
),
2227 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2230 #if IS_ENABLED(CONFIG_IPV6)
2231 else if (sk
->sk_family
== AF_INET6
) {
2232 struct ipv6_pinfo
*np
= inet6_sk(sk
);
2233 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2235 &np
->daddr
, ntohs(inet
->inet_dport
),
2236 tp
->snd_cwnd
, tcp_left_out(tp
),
2237 tp
->snd_ssthresh
, tp
->prior_ssthresh
,
2243 #define DBGUNDO(x...) do { } while (0)
2246 static void tcp_undo_cwr(struct sock
*sk
, const bool undo_ssthresh
)
2248 struct tcp_sock
*tp
= tcp_sk(sk
);
2250 if (tp
->prior_ssthresh
) {
2251 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2253 if (icsk
->icsk_ca_ops
->undo_cwnd
)
2254 tp
->snd_cwnd
= icsk
->icsk_ca_ops
->undo_cwnd(sk
);
2256 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
<< 1);
2258 if (undo_ssthresh
&& tp
->prior_ssthresh
> tp
->snd_ssthresh
) {
2259 tp
->snd_ssthresh
= tp
->prior_ssthresh
;
2260 TCP_ECN_withdraw_cwr(tp
);
2263 tp
->snd_cwnd
= max(tp
->snd_cwnd
, tp
->snd_ssthresh
);
2265 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2268 static inline bool tcp_may_undo(const struct tcp_sock
*tp
)
2270 return tp
->undo_marker
&& (!tp
->undo_retrans
|| tcp_packet_delayed(tp
));
2273 /* People celebrate: "We love our President!" */
2274 static bool tcp_try_undo_recovery(struct sock
*sk
)
2276 struct tcp_sock
*tp
= tcp_sk(sk
);
2278 if (tcp_may_undo(tp
)) {
2281 /* Happy end! We did not retransmit anything
2282 * or our original transmission succeeded.
2284 DBGUNDO(sk
, inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
? "loss" : "retrans");
2285 tcp_undo_cwr(sk
, true);
2286 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Loss
)
2287 mib_idx
= LINUX_MIB_TCPLOSSUNDO
;
2289 mib_idx
= LINUX_MIB_TCPFULLUNDO
;
2291 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2292 tp
->undo_marker
= 0;
2294 if (tp
->snd_una
== tp
->high_seq
&& tcp_is_reno(tp
)) {
2295 /* Hold old state until something *above* high_seq
2296 * is ACKed. For Reno it is MUST to prevent false
2297 * fast retransmits (RFC2582). SACK TCP is safe. */
2298 tcp_moderate_cwnd(tp
);
2301 tcp_set_ca_state(sk
, TCP_CA_Open
);
2305 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2306 static void tcp_try_undo_dsack(struct sock
*sk
)
2308 struct tcp_sock
*tp
= tcp_sk(sk
);
2310 if (tp
->undo_marker
&& !tp
->undo_retrans
) {
2311 DBGUNDO(sk
, "D-SACK");
2312 tcp_undo_cwr(sk
, true);
2313 tp
->undo_marker
= 0;
2314 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPDSACKUNDO
);
2318 /* We can clear retrans_stamp when there are no retransmissions in the
2319 * window. It would seem that it is trivially available for us in
2320 * tp->retrans_out, however, that kind of assumptions doesn't consider
2321 * what will happen if errors occur when sending retransmission for the
2322 * second time. ...It could the that such segment has only
2323 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2324 * the head skb is enough except for some reneging corner cases that
2325 * are not worth the effort.
2327 * Main reason for all this complexity is the fact that connection dying
2328 * time now depends on the validity of the retrans_stamp, in particular,
2329 * that successive retransmissions of a segment must not advance
2330 * retrans_stamp under any conditions.
2332 static bool tcp_any_retrans_done(const struct sock
*sk
)
2334 const struct tcp_sock
*tp
= tcp_sk(sk
);
2335 struct sk_buff
*skb
;
2337 if (tp
->retrans_out
)
2340 skb
= tcp_write_queue_head(sk
);
2341 if (unlikely(skb
&& TCP_SKB_CB(skb
)->sacked
& TCPCB_EVER_RETRANS
))
2347 /* Undo during fast recovery after partial ACK. */
2349 static int tcp_try_undo_partial(struct sock
*sk
, int acked
)
2351 struct tcp_sock
*tp
= tcp_sk(sk
);
2352 /* Partial ACK arrived. Force Hoe's retransmit. */
2353 int failed
= tcp_is_reno(tp
) || (tcp_fackets_out(tp
) > tp
->reordering
);
2355 if (tcp_may_undo(tp
)) {
2356 /* Plain luck! Hole if filled with delayed
2357 * packet, rather than with a retransmit.
2359 if (!tcp_any_retrans_done(sk
))
2360 tp
->retrans_stamp
= 0;
2362 tcp_update_reordering(sk
, tcp_fackets_out(tp
) + acked
, 1);
2365 tcp_undo_cwr(sk
, false);
2366 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPARTIALUNDO
);
2368 /* So... Do not make Hoe's retransmit yet.
2369 * If the first packet was delayed, the rest
2370 * ones are most probably delayed as well.
2377 /* Undo during loss recovery after partial ACK or using F-RTO. */
2378 static bool tcp_try_undo_loss(struct sock
*sk
, bool frto_undo
)
2380 struct tcp_sock
*tp
= tcp_sk(sk
);
2382 if (frto_undo
|| tcp_may_undo(tp
)) {
2383 struct sk_buff
*skb
;
2384 tcp_for_write_queue(skb
, sk
) {
2385 if (skb
== tcp_send_head(sk
))
2387 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_LOST
;
2390 tcp_clear_all_retrans_hints(tp
);
2392 DBGUNDO(sk
, "partial loss");
2394 tcp_undo_cwr(sk
, true);
2395 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSUNDO
);
2397 NET_INC_STATS_BH(sock_net(sk
),
2398 LINUX_MIB_TCPSPURIOUSRTOS
);
2399 inet_csk(sk
)->icsk_retransmits
= 0;
2400 tp
->undo_marker
= 0;
2401 if (frto_undo
|| tcp_is_sack(tp
))
2402 tcp_set_ca_state(sk
, TCP_CA_Open
);
2408 /* The cwnd reduction in CWR and Recovery use the PRR algorithm
2409 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2410 * It computes the number of packets to send (sndcnt) based on packets newly
2412 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2413 * cwnd reductions across a full RTT.
2414 * 2) If packets in flight is lower than ssthresh (such as due to excess
2415 * losses and/or application stalls), do not perform any further cwnd
2416 * reductions, but instead slow start up to ssthresh.
2418 static void tcp_init_cwnd_reduction(struct sock
*sk
, const bool set_ssthresh
)
2420 struct tcp_sock
*tp
= tcp_sk(sk
);
2422 tp
->high_seq
= tp
->snd_nxt
;
2423 tp
->tlp_high_seq
= 0;
2424 tp
->snd_cwnd_cnt
= 0;
2425 tp
->prior_cwnd
= tp
->snd_cwnd
;
2426 tp
->prr_delivered
= 0;
2429 tp
->snd_ssthresh
= inet_csk(sk
)->icsk_ca_ops
->ssthresh(sk
);
2430 TCP_ECN_queue_cwr(tp
);
2433 static void tcp_cwnd_reduction(struct sock
*sk
, int newly_acked_sacked
,
2436 struct tcp_sock
*tp
= tcp_sk(sk
);
2438 int delta
= tp
->snd_ssthresh
- tcp_packets_in_flight(tp
);
2440 tp
->prr_delivered
+= newly_acked_sacked
;
2441 if (tcp_packets_in_flight(tp
) > tp
->snd_ssthresh
) {
2442 u64 dividend
= (u64
)tp
->snd_ssthresh
* tp
->prr_delivered
+
2444 sndcnt
= div_u64(dividend
, tp
->prior_cwnd
) - tp
->prr_out
;
2446 sndcnt
= min_t(int, delta
,
2447 max_t(int, tp
->prr_delivered
- tp
->prr_out
,
2448 newly_acked_sacked
) + 1);
2451 sndcnt
= max(sndcnt
, (fast_rexmit
? 1 : 0));
2452 tp
->snd_cwnd
= tcp_packets_in_flight(tp
) + sndcnt
;
2455 static inline void tcp_end_cwnd_reduction(struct sock
*sk
)
2457 struct tcp_sock
*tp
= tcp_sk(sk
);
2459 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2460 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_CWR
||
2461 (tp
->undo_marker
&& tp
->snd_ssthresh
< TCP_INFINITE_SSTHRESH
)) {
2462 tp
->snd_cwnd
= tp
->snd_ssthresh
;
2463 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2465 tcp_ca_event(sk
, CA_EVENT_COMPLETE_CWR
);
2468 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2469 void tcp_enter_cwr(struct sock
*sk
, const int set_ssthresh
)
2471 struct tcp_sock
*tp
= tcp_sk(sk
);
2473 tp
->prior_ssthresh
= 0;
2474 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2475 tp
->undo_marker
= 0;
2476 tcp_init_cwnd_reduction(sk
, set_ssthresh
);
2477 tcp_set_ca_state(sk
, TCP_CA_CWR
);
2481 static void tcp_try_keep_open(struct sock
*sk
)
2483 struct tcp_sock
*tp
= tcp_sk(sk
);
2484 int state
= TCP_CA_Open
;
2486 if (tcp_left_out(tp
) || tcp_any_retrans_done(sk
))
2487 state
= TCP_CA_Disorder
;
2489 if (inet_csk(sk
)->icsk_ca_state
!= state
) {
2490 tcp_set_ca_state(sk
, state
);
2491 tp
->high_seq
= tp
->snd_nxt
;
2495 static void tcp_try_to_open(struct sock
*sk
, int flag
, int newly_acked_sacked
)
2497 struct tcp_sock
*tp
= tcp_sk(sk
);
2499 tcp_verify_left_out(tp
);
2501 if (!tcp_any_retrans_done(sk
))
2502 tp
->retrans_stamp
= 0;
2504 if (flag
& FLAG_ECE
)
2505 tcp_enter_cwr(sk
, 1);
2507 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_CWR
) {
2508 tcp_try_keep_open(sk
);
2509 if (inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2510 tcp_moderate_cwnd(tp
);
2512 tcp_cwnd_reduction(sk
, newly_acked_sacked
, 0);
2516 static void tcp_mtup_probe_failed(struct sock
*sk
)
2518 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2520 icsk
->icsk_mtup
.search_high
= icsk
->icsk_mtup
.probe_size
- 1;
2521 icsk
->icsk_mtup
.probe_size
= 0;
2524 static void tcp_mtup_probe_success(struct sock
*sk
)
2526 struct tcp_sock
*tp
= tcp_sk(sk
);
2527 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2529 /* FIXME: breaks with very large cwnd */
2530 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2531 tp
->snd_cwnd
= tp
->snd_cwnd
*
2532 tcp_mss_to_mtu(sk
, tp
->mss_cache
) /
2533 icsk
->icsk_mtup
.probe_size
;
2534 tp
->snd_cwnd_cnt
= 0;
2535 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
2536 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2538 icsk
->icsk_mtup
.search_low
= icsk
->icsk_mtup
.probe_size
;
2539 icsk
->icsk_mtup
.probe_size
= 0;
2540 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
2543 /* Do a simple retransmit without using the backoff mechanisms in
2544 * tcp_timer. This is used for path mtu discovery.
2545 * The socket is already locked here.
2547 void tcp_simple_retransmit(struct sock
*sk
)
2549 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2550 struct tcp_sock
*tp
= tcp_sk(sk
);
2551 struct sk_buff
*skb
;
2552 unsigned int mss
= tcp_current_mss(sk
);
2553 u32 prior_lost
= tp
->lost_out
;
2555 tcp_for_write_queue(skb
, sk
) {
2556 if (skb
== tcp_send_head(sk
))
2558 if (tcp_skb_seglen(skb
) > mss
&&
2559 !(TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)) {
2560 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2561 TCP_SKB_CB(skb
)->sacked
&= ~TCPCB_SACKED_RETRANS
;
2562 tp
->retrans_out
-= tcp_skb_pcount(skb
);
2564 tcp_skb_mark_lost_uncond_verify(tp
, skb
);
2568 tcp_clear_retrans_hints_partial(tp
);
2570 if (prior_lost
== tp
->lost_out
)
2573 if (tcp_is_reno(tp
))
2574 tcp_limit_reno_sacked(tp
);
2576 tcp_verify_left_out(tp
);
2578 /* Don't muck with the congestion window here.
2579 * Reason is that we do not increase amount of _data_
2580 * in network, but units changed and effective
2581 * cwnd/ssthresh really reduced now.
2583 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
) {
2584 tp
->high_seq
= tp
->snd_nxt
;
2585 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
2586 tp
->prior_ssthresh
= 0;
2587 tp
->undo_marker
= 0;
2588 tcp_set_ca_state(sk
, TCP_CA_Loss
);
2590 tcp_xmit_retransmit_queue(sk
);
2592 EXPORT_SYMBOL(tcp_simple_retransmit
);
2594 static void tcp_enter_recovery(struct sock
*sk
, bool ece_ack
)
2596 struct tcp_sock
*tp
= tcp_sk(sk
);
2599 if (tcp_is_reno(tp
))
2600 mib_idx
= LINUX_MIB_TCPRENORECOVERY
;
2602 mib_idx
= LINUX_MIB_TCPSACKRECOVERY
;
2604 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2606 tp
->prior_ssthresh
= 0;
2607 tp
->undo_marker
= tp
->snd_una
;
2608 tp
->undo_retrans
= tp
->retrans_out
;
2610 if (inet_csk(sk
)->icsk_ca_state
< TCP_CA_CWR
) {
2612 tp
->prior_ssthresh
= tcp_current_ssthresh(sk
);
2613 tcp_init_cwnd_reduction(sk
, true);
2615 tcp_set_ca_state(sk
, TCP_CA_Recovery
);
2618 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2619 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2621 static void tcp_process_loss(struct sock
*sk
, int flag
, bool is_dupack
)
2623 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2624 struct tcp_sock
*tp
= tcp_sk(sk
);
2625 bool recovered
= !before(tp
->snd_una
, tp
->high_seq
);
2627 if (tp
->frto
) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2628 if (flag
& FLAG_ORIG_SACK_ACKED
) {
2629 /* Step 3.b. A timeout is spurious if not all data are
2630 * lost, i.e., never-retransmitted data are (s)acked.
2632 tcp_try_undo_loss(sk
, true);
2635 if (after(tp
->snd_nxt
, tp
->high_seq
) &&
2636 (flag
& FLAG_DATA_SACKED
|| is_dupack
)) {
2637 tp
->frto
= 0; /* Loss was real: 2nd part of step 3.a */
2638 } else if (flag
& FLAG_SND_UNA_ADVANCED
&& !recovered
) {
2639 tp
->high_seq
= tp
->snd_nxt
;
2640 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
),
2642 if (after(tp
->snd_nxt
, tp
->high_seq
))
2643 return; /* Step 2.b */
2649 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2650 icsk
->icsk_retransmits
= 0;
2651 tcp_try_undo_recovery(sk
);
2654 if (flag
& FLAG_DATA_ACKED
)
2655 icsk
->icsk_retransmits
= 0;
2656 if (tcp_is_reno(tp
)) {
2657 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2658 * delivered. Lower inflight to clock out (re)tranmissions.
2660 if (after(tp
->snd_nxt
, tp
->high_seq
) && is_dupack
)
2661 tcp_add_reno_sack(sk
);
2662 else if (flag
& FLAG_SND_UNA_ADVANCED
)
2663 tcp_reset_reno_sack(tp
);
2665 if (tcp_try_undo_loss(sk
, false))
2667 tcp_xmit_retransmit_queue(sk
);
2670 /* Process an event, which can update packets-in-flight not trivially.
2671 * Main goal of this function is to calculate new estimate for left_out,
2672 * taking into account both packets sitting in receiver's buffer and
2673 * packets lost by network.
2675 * Besides that it does CWND reduction, when packet loss is detected
2676 * and changes state of machine.
2678 * It does _not_ decide what to send, it is made in function
2679 * tcp_xmit_retransmit_queue().
2681 static void tcp_fastretrans_alert(struct sock
*sk
, int pkts_acked
,
2682 int prior_sacked
, int prior_packets
,
2683 bool is_dupack
, int flag
)
2685 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2686 struct tcp_sock
*tp
= tcp_sk(sk
);
2687 int do_lost
= is_dupack
|| ((flag
& FLAG_DATA_SACKED
) &&
2688 (tcp_fackets_out(tp
) > tp
->reordering
));
2689 int newly_acked_sacked
= 0;
2690 int fast_rexmit
= 0;
2692 if (WARN_ON(!tp
->packets_out
&& tp
->sacked_out
))
2694 if (WARN_ON(!tp
->sacked_out
&& tp
->fackets_out
))
2695 tp
->fackets_out
= 0;
2697 /* Now state machine starts.
2698 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2699 if (flag
& FLAG_ECE
)
2700 tp
->prior_ssthresh
= 0;
2702 /* B. In all the states check for reneging SACKs. */
2703 if (tcp_check_sack_reneging(sk
, flag
))
2706 /* C. Check consistency of the current state. */
2707 tcp_verify_left_out(tp
);
2709 /* D. Check state exit conditions. State can be terminated
2710 * when high_seq is ACKed. */
2711 if (icsk
->icsk_ca_state
== TCP_CA_Open
) {
2712 WARN_ON(tp
->retrans_out
!= 0);
2713 tp
->retrans_stamp
= 0;
2714 } else if (!before(tp
->snd_una
, tp
->high_seq
)) {
2715 switch (icsk
->icsk_ca_state
) {
2717 /* CWR is to be held something *above* high_seq
2718 * is ACKed for CWR bit to reach receiver. */
2719 if (tp
->snd_una
!= tp
->high_seq
) {
2720 tcp_end_cwnd_reduction(sk
);
2721 tcp_set_ca_state(sk
, TCP_CA_Open
);
2725 case TCP_CA_Recovery
:
2726 if (tcp_is_reno(tp
))
2727 tcp_reset_reno_sack(tp
);
2728 if (tcp_try_undo_recovery(sk
))
2730 tcp_end_cwnd_reduction(sk
);
2735 /* E. Process state. */
2736 switch (icsk
->icsk_ca_state
) {
2737 case TCP_CA_Recovery
:
2738 if (!(flag
& FLAG_SND_UNA_ADVANCED
)) {
2739 if (tcp_is_reno(tp
) && is_dupack
)
2740 tcp_add_reno_sack(sk
);
2742 do_lost
= tcp_try_undo_partial(sk
, pkts_acked
);
2743 newly_acked_sacked
= prior_packets
- tp
->packets_out
+
2744 tp
->sacked_out
- prior_sacked
;
2747 tcp_process_loss(sk
, flag
, is_dupack
);
2748 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
2750 /* Fall through to processing in Open state. */
2752 if (tcp_is_reno(tp
)) {
2753 if (flag
& FLAG_SND_UNA_ADVANCED
)
2754 tcp_reset_reno_sack(tp
);
2756 tcp_add_reno_sack(sk
);
2758 newly_acked_sacked
= prior_packets
- tp
->packets_out
+
2759 tp
->sacked_out
- prior_sacked
;
2761 if (icsk
->icsk_ca_state
<= TCP_CA_Disorder
)
2762 tcp_try_undo_dsack(sk
);
2764 if (!tcp_time_to_recover(sk
, flag
)) {
2765 tcp_try_to_open(sk
, flag
, newly_acked_sacked
);
2769 /* MTU probe failure: don't reduce cwnd */
2770 if (icsk
->icsk_ca_state
< TCP_CA_CWR
&&
2771 icsk
->icsk_mtup
.probe_size
&&
2772 tp
->snd_una
== tp
->mtu_probe
.probe_seq_start
) {
2773 tcp_mtup_probe_failed(sk
);
2774 /* Restores the reduction we did in tcp_mtup_probe() */
2776 tcp_simple_retransmit(sk
);
2780 /* Otherwise enter Recovery state */
2781 tcp_enter_recovery(sk
, (flag
& FLAG_ECE
));
2786 tcp_update_scoreboard(sk
, fast_rexmit
);
2787 tcp_cwnd_reduction(sk
, newly_acked_sacked
, fast_rexmit
);
2788 tcp_xmit_retransmit_queue(sk
);
2791 void tcp_valid_rtt_meas(struct sock
*sk
, u32 seq_rtt
)
2793 tcp_rtt_estimator(sk
, seq_rtt
);
2795 inet_csk(sk
)->icsk_backoff
= 0;
2797 EXPORT_SYMBOL(tcp_valid_rtt_meas
);
2799 /* Read draft-ietf-tcplw-high-performance before mucking
2800 * with this code. (Supersedes RFC1323)
2802 static void tcp_ack_saw_tstamp(struct sock
*sk
, int flag
)
2804 /* RTTM Rule: A TSecr value received in a segment is used to
2805 * update the averaged RTT measurement only if the segment
2806 * acknowledges some new data, i.e., only if it advances the
2807 * left edge of the send window.
2809 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2810 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2812 * Changed: reset backoff as soon as we see the first valid sample.
2813 * If we do not, we get strongly overestimated rto. With timestamps
2814 * samples are accepted even from very old segments: f.e., when rtt=1
2815 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2816 * answer arrives rto becomes 120 seconds! If at least one of segments
2817 * in window is lost... Voila. --ANK (010210)
2819 struct tcp_sock
*tp
= tcp_sk(sk
);
2821 tcp_valid_rtt_meas(sk
, tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
2824 static void tcp_ack_no_tstamp(struct sock
*sk
, u32 seq_rtt
, int flag
)
2826 /* We don't have a timestamp. Can only use
2827 * packets that are not retransmitted to determine
2828 * rtt estimates. Also, we must not reset the
2829 * backoff for rto until we get a non-retransmitted
2830 * packet. This allows us to deal with a situation
2831 * where the network delay has increased suddenly.
2832 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2835 if (flag
& FLAG_RETRANS_DATA_ACKED
)
2838 tcp_valid_rtt_meas(sk
, seq_rtt
);
2841 static inline void tcp_ack_update_rtt(struct sock
*sk
, const int flag
,
2844 const struct tcp_sock
*tp
= tcp_sk(sk
);
2845 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2846 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
)
2847 tcp_ack_saw_tstamp(sk
, flag
);
2848 else if (seq_rtt
>= 0)
2849 tcp_ack_no_tstamp(sk
, seq_rtt
, flag
);
2852 static void tcp_cong_avoid(struct sock
*sk
, u32 ack
, u32 in_flight
)
2854 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2855 icsk
->icsk_ca_ops
->cong_avoid(sk
, ack
, in_flight
);
2856 tcp_sk(sk
)->snd_cwnd_stamp
= tcp_time_stamp
;
2859 /* Restart timer after forward progress on connection.
2860 * RFC2988 recommends to restart timer to now+rto.
2862 void tcp_rearm_rto(struct sock
*sk
)
2864 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2865 struct tcp_sock
*tp
= tcp_sk(sk
);
2867 /* If the retrans timer is currently being used by Fast Open
2868 * for SYN-ACK retrans purpose, stay put.
2870 if (tp
->fastopen_rsk
)
2873 if (!tp
->packets_out
) {
2874 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_RETRANS
);
2876 u32 rto
= inet_csk(sk
)->icsk_rto
;
2877 /* Offset the time elapsed after installing regular RTO */
2878 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
2879 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
2880 struct sk_buff
*skb
= tcp_write_queue_head(sk
);
2881 const u32 rto_time_stamp
= TCP_SKB_CB(skb
)->when
+ rto
;
2882 s32 delta
= (s32
)(rto_time_stamp
- tcp_time_stamp
);
2883 /* delta may not be positive if the socket is locked
2884 * when the retrans timer fires and is rescheduled.
2889 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
, rto
,
2894 /* This function is called when the delayed ER timer fires. TCP enters
2895 * fast recovery and performs fast-retransmit.
2897 void tcp_resume_early_retransmit(struct sock
*sk
)
2899 struct tcp_sock
*tp
= tcp_sk(sk
);
2903 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2904 if (!tp
->do_early_retrans
)
2907 tcp_enter_recovery(sk
, false);
2908 tcp_update_scoreboard(sk
, 1);
2909 tcp_xmit_retransmit_queue(sk
);
2912 /* If we get here, the whole TSO packet has not been acked. */
2913 static u32
tcp_tso_acked(struct sock
*sk
, struct sk_buff
*skb
)
2915 struct tcp_sock
*tp
= tcp_sk(sk
);
2918 BUG_ON(!after(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
));
2920 packets_acked
= tcp_skb_pcount(skb
);
2921 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2923 packets_acked
-= tcp_skb_pcount(skb
);
2925 if (packets_acked
) {
2926 BUG_ON(tcp_skb_pcount(skb
) == 0);
2927 BUG_ON(!before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
));
2930 return packets_acked
;
2933 /* Remove acknowledged frames from the retransmission queue. If our packet
2934 * is before the ack sequence we can discard it as it's confirmed to have
2935 * arrived at the other end.
2937 static int tcp_clean_rtx_queue(struct sock
*sk
, int prior_fackets
,
2940 struct tcp_sock
*tp
= tcp_sk(sk
);
2941 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2942 struct sk_buff
*skb
;
2943 u32 now
= tcp_time_stamp
;
2944 int fully_acked
= true;
2947 u32 reord
= tp
->packets_out
;
2948 u32 prior_sacked
= tp
->sacked_out
;
2950 s32 ca_seq_rtt
= -1;
2951 ktime_t last_ackt
= net_invalid_timestamp();
2953 while ((skb
= tcp_write_queue_head(sk
)) && skb
!= tcp_send_head(sk
)) {
2954 struct tcp_skb_cb
*scb
= TCP_SKB_CB(skb
);
2956 u8 sacked
= scb
->sacked
;
2958 /* Determine how many packets and what bytes were acked, tso and else */
2959 if (after(scb
->end_seq
, tp
->snd_una
)) {
2960 if (tcp_skb_pcount(skb
) == 1 ||
2961 !after(tp
->snd_una
, scb
->seq
))
2964 acked_pcount
= tcp_tso_acked(sk
, skb
);
2968 fully_acked
= false;
2970 acked_pcount
= tcp_skb_pcount(skb
);
2973 if (sacked
& TCPCB_RETRANS
) {
2974 if (sacked
& TCPCB_SACKED_RETRANS
)
2975 tp
->retrans_out
-= acked_pcount
;
2976 flag
|= FLAG_RETRANS_DATA_ACKED
;
2980 ca_seq_rtt
= now
- scb
->when
;
2981 last_ackt
= skb
->tstamp
;
2983 seq_rtt
= ca_seq_rtt
;
2985 if (!(sacked
& TCPCB_SACKED_ACKED
))
2986 reord
= min(pkts_acked
, reord
);
2987 if (!after(scb
->end_seq
, tp
->high_seq
))
2988 flag
|= FLAG_ORIG_SACK_ACKED
;
2991 if (sacked
& TCPCB_SACKED_ACKED
)
2992 tp
->sacked_out
-= acked_pcount
;
2993 if (sacked
& TCPCB_LOST
)
2994 tp
->lost_out
-= acked_pcount
;
2996 tp
->packets_out
-= acked_pcount
;
2997 pkts_acked
+= acked_pcount
;
2999 /* Initial outgoing SYN's get put onto the write_queue
3000 * just like anything else we transmit. It is not
3001 * true data, and if we misinform our callers that
3002 * this ACK acks real data, we will erroneously exit
3003 * connection startup slow start one packet too
3004 * quickly. This is severely frowned upon behavior.
3006 if (!(scb
->tcp_flags
& TCPHDR_SYN
)) {
3007 flag
|= FLAG_DATA_ACKED
;
3009 flag
|= FLAG_SYN_ACKED
;
3010 tp
->retrans_stamp
= 0;
3016 tcp_unlink_write_queue(skb
, sk
);
3017 sk_wmem_free_skb(sk
, skb
);
3018 if (skb
== tp
->retransmit_skb_hint
)
3019 tp
->retransmit_skb_hint
= NULL
;
3020 if (skb
== tp
->lost_skb_hint
)
3021 tp
->lost_skb_hint
= NULL
;
3024 if (likely(between(tp
->snd_up
, prior_snd_una
, tp
->snd_una
)))
3025 tp
->snd_up
= tp
->snd_una
;
3027 if (skb
&& (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
3028 flag
|= FLAG_SACK_RENEGING
;
3030 if (flag
& FLAG_ACKED
) {
3031 const struct tcp_congestion_ops
*ca_ops
3032 = inet_csk(sk
)->icsk_ca_ops
;
3034 if (unlikely(icsk
->icsk_mtup
.probe_size
&&
3035 !after(tp
->mtu_probe
.probe_seq_end
, tp
->snd_una
))) {
3036 tcp_mtup_probe_success(sk
);
3039 tcp_ack_update_rtt(sk
, flag
, seq_rtt
);
3042 if (tcp_is_reno(tp
)) {
3043 tcp_remove_reno_sacks(sk
, pkts_acked
);
3047 /* Non-retransmitted hole got filled? That's reordering */
3048 if (reord
< prior_fackets
)
3049 tcp_update_reordering(sk
, tp
->fackets_out
- reord
, 0);
3051 delta
= tcp_is_fack(tp
) ? pkts_acked
:
3052 prior_sacked
- tp
->sacked_out
;
3053 tp
->lost_cnt_hint
-= min(tp
->lost_cnt_hint
, delta
);
3056 tp
->fackets_out
-= min(pkts_acked
, tp
->fackets_out
);
3058 if (ca_ops
->pkts_acked
) {
3061 /* Is the ACK triggering packet unambiguous? */
3062 if (!(flag
& FLAG_RETRANS_DATA_ACKED
)) {
3063 /* High resolution needed and available? */
3064 if (ca_ops
->flags
& TCP_CONG_RTT_STAMP
&&
3065 !ktime_equal(last_ackt
,
3066 net_invalid_timestamp()))
3067 rtt_us
= ktime_us_delta(ktime_get_real(),
3069 else if (ca_seq_rtt
>= 0)
3070 rtt_us
= jiffies_to_usecs(ca_seq_rtt
);
3073 ca_ops
->pkts_acked(sk
, pkts_acked
, rtt_us
);
3077 #if FASTRETRANS_DEBUG > 0
3078 WARN_ON((int)tp
->sacked_out
< 0);
3079 WARN_ON((int)tp
->lost_out
< 0);
3080 WARN_ON((int)tp
->retrans_out
< 0);
3081 if (!tp
->packets_out
&& tcp_is_sack(tp
)) {
3082 icsk
= inet_csk(sk
);
3084 pr_debug("Leak l=%u %d\n",
3085 tp
->lost_out
, icsk
->icsk_ca_state
);
3088 if (tp
->sacked_out
) {
3089 pr_debug("Leak s=%u %d\n",
3090 tp
->sacked_out
, icsk
->icsk_ca_state
);
3093 if (tp
->retrans_out
) {
3094 pr_debug("Leak r=%u %d\n",
3095 tp
->retrans_out
, icsk
->icsk_ca_state
);
3096 tp
->retrans_out
= 0;
3103 static void tcp_ack_probe(struct sock
*sk
)
3105 const struct tcp_sock
*tp
= tcp_sk(sk
);
3106 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3108 /* Was it a usable window open? */
3110 if (!after(TCP_SKB_CB(tcp_send_head(sk
))->end_seq
, tcp_wnd_end(tp
))) {
3111 icsk
->icsk_backoff
= 0;
3112 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_PROBE0
);
3113 /* Socket must be waked up by subsequent tcp_data_snd_check().
3114 * This function is not for random using!
3117 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3118 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3123 static inline bool tcp_ack_is_dubious(const struct sock
*sk
, const int flag
)
3125 return !(flag
& FLAG_NOT_DUP
) || (flag
& FLAG_CA_ALERT
) ||
3126 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
;
3129 static inline bool tcp_may_raise_cwnd(const struct sock
*sk
, const int flag
)
3131 const struct tcp_sock
*tp
= tcp_sk(sk
);
3132 return (!(flag
& FLAG_ECE
) || tp
->snd_cwnd
< tp
->snd_ssthresh
) &&
3133 !tcp_in_cwnd_reduction(sk
);
3136 /* Check that window update is acceptable.
3137 * The function assumes that snd_una<=ack<=snd_next.
3139 static inline bool tcp_may_update_window(const struct tcp_sock
*tp
,
3140 const u32 ack
, const u32 ack_seq
,
3143 return after(ack
, tp
->snd_una
) ||
3144 after(ack_seq
, tp
->snd_wl1
) ||
3145 (ack_seq
== tp
->snd_wl1
&& nwin
> tp
->snd_wnd
);
3148 /* Update our send window.
3150 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3151 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3153 static int tcp_ack_update_window(struct sock
*sk
, const struct sk_buff
*skb
, u32 ack
,
3156 struct tcp_sock
*tp
= tcp_sk(sk
);
3158 u32 nwin
= ntohs(tcp_hdr(skb
)->window
);
3160 if (likely(!tcp_hdr(skb
)->syn
))
3161 nwin
<<= tp
->rx_opt
.snd_wscale
;
3163 if (tcp_may_update_window(tp
, ack
, ack_seq
, nwin
)) {
3164 flag
|= FLAG_WIN_UPDATE
;
3165 tcp_update_wl(tp
, ack_seq
);
3167 if (tp
->snd_wnd
!= nwin
) {
3170 /* Note, it is the only place, where
3171 * fast path is recovered for sending TCP.
3174 tcp_fast_path_check(sk
);
3176 if (nwin
> tp
->max_window
) {
3177 tp
->max_window
= nwin
;
3178 tcp_sync_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
);
3188 /* RFC 5961 7 [ACK Throttling] */
3189 static void tcp_send_challenge_ack(struct sock
*sk
)
3191 /* unprotected vars, we dont care of overwrites */
3192 static u32 challenge_timestamp
;
3193 static unsigned int challenge_count
;
3194 u32 now
= jiffies
/ HZ
;
3196 if (now
!= challenge_timestamp
) {
3197 challenge_timestamp
= now
;
3198 challenge_count
= 0;
3200 if (++challenge_count
<= sysctl_tcp_challenge_ack_limit
) {
3201 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPCHALLENGEACK
);
3206 static void tcp_store_ts_recent(struct tcp_sock
*tp
)
3208 tp
->rx_opt
.ts_recent
= tp
->rx_opt
.rcv_tsval
;
3209 tp
->rx_opt
.ts_recent_stamp
= get_seconds();
3212 static void tcp_replace_ts_recent(struct tcp_sock
*tp
, u32 seq
)
3214 if (tp
->rx_opt
.saw_tstamp
&& !after(seq
, tp
->rcv_wup
)) {
3215 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3216 * extra check below makes sure this can only happen
3217 * for pure ACK frames. -DaveM
3219 * Not only, also it occurs for expired timestamps.
3222 if (tcp_paws_check(&tp
->rx_opt
, 0))
3223 tcp_store_ts_recent(tp
);
3227 /* This routine deals with acks during a TLP episode.
3228 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3230 static void tcp_process_tlp_ack(struct sock
*sk
, u32 ack
, int flag
)
3232 struct tcp_sock
*tp
= tcp_sk(sk
);
3233 bool is_tlp_dupack
= (ack
== tp
->tlp_high_seq
) &&
3234 !(flag
& (FLAG_SND_UNA_ADVANCED
|
3235 FLAG_NOT_DUP
| FLAG_DATA_SACKED
));
3237 /* Mark the end of TLP episode on receiving TLP dupack or when
3238 * ack is after tlp_high_seq.
3240 if (is_tlp_dupack
) {
3241 tp
->tlp_high_seq
= 0;
3245 if (after(ack
, tp
->tlp_high_seq
)) {
3246 tp
->tlp_high_seq
= 0;
3247 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3248 if (!(flag
& FLAG_DSACKING_ACK
)) {
3249 tcp_init_cwnd_reduction(sk
, true);
3250 tcp_set_ca_state(sk
, TCP_CA_CWR
);
3251 tcp_end_cwnd_reduction(sk
);
3252 tcp_set_ca_state(sk
, TCP_CA_Open
);
3253 NET_INC_STATS_BH(sock_net(sk
),
3254 LINUX_MIB_TCPLOSSPROBERECOVERY
);
3259 /* This routine deals with incoming acks, but not outgoing ones. */
3260 static int tcp_ack(struct sock
*sk
, const struct sk_buff
*skb
, int flag
)
3262 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3263 struct tcp_sock
*tp
= tcp_sk(sk
);
3264 u32 prior_snd_una
= tp
->snd_una
;
3265 u32 ack_seq
= TCP_SKB_CB(skb
)->seq
;
3266 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3267 bool is_dupack
= false;
3268 u32 prior_in_flight
;
3270 int prior_packets
= tp
->packets_out
;
3271 int prior_sacked
= tp
->sacked_out
;
3273 int previous_packets_out
= 0;
3275 /* If the ack is older than previous acks
3276 * then we can probably ignore it.
3278 if (before(ack
, prior_snd_una
)) {
3279 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3280 if (before(ack
, prior_snd_una
- tp
->max_window
)) {
3281 tcp_send_challenge_ack(sk
);
3287 /* If the ack includes data we haven't sent yet, discard
3288 * this segment (RFC793 Section 3.9).
3290 if (after(ack
, tp
->snd_nxt
))
3293 if (icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
3294 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
3297 if (after(ack
, prior_snd_una
))
3298 flag
|= FLAG_SND_UNA_ADVANCED
;
3300 prior_fackets
= tp
->fackets_out
;
3301 prior_in_flight
= tcp_packets_in_flight(tp
);
3303 /* ts_recent update must be made after we are sure that the packet
3306 if (flag
& FLAG_UPDATE_TS_RECENT
)
3307 tcp_replace_ts_recent(tp
, TCP_SKB_CB(skb
)->seq
);
3309 if (!(flag
& FLAG_SLOWPATH
) && after(ack
, prior_snd_una
)) {
3310 /* Window is constant, pure forward advance.
3311 * No more checks are required.
3312 * Note, we use the fact that SND.UNA>=SND.WL2.
3314 tcp_update_wl(tp
, ack_seq
);
3316 flag
|= FLAG_WIN_UPDATE
;
3318 tcp_ca_event(sk
, CA_EVENT_FAST_ACK
);
3320 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPACKS
);
3322 if (ack_seq
!= TCP_SKB_CB(skb
)->end_seq
)
3325 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPPUREACKS
);
3327 flag
|= tcp_ack_update_window(sk
, skb
, ack
, ack_seq
);
3329 if (TCP_SKB_CB(skb
)->sacked
)
3330 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3332 if (TCP_ECN_rcv_ecn_echo(tp
, tcp_hdr(skb
)))
3335 tcp_ca_event(sk
, CA_EVENT_SLOW_ACK
);
3338 /* We passed data and got it acked, remove any soft error
3339 * log. Something worked...
3341 sk
->sk_err_soft
= 0;
3342 icsk
->icsk_probes_out
= 0;
3343 tp
->rcv_tstamp
= tcp_time_stamp
;
3347 /* See if we can take anything off of the retransmit queue. */
3348 previous_packets_out
= tp
->packets_out
;
3349 flag
|= tcp_clean_rtx_queue(sk
, prior_fackets
, prior_snd_una
);
3351 pkts_acked
= previous_packets_out
- tp
->packets_out
;
3353 if (tcp_ack_is_dubious(sk
, flag
)) {
3354 /* Advance CWND, if state allows this. */
3355 if ((flag
& FLAG_DATA_ACKED
) && tcp_may_raise_cwnd(sk
, flag
))
3356 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3357 is_dupack
= !(flag
& (FLAG_SND_UNA_ADVANCED
| FLAG_NOT_DUP
));
3358 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3359 prior_packets
, is_dupack
, flag
);
3361 if (flag
& FLAG_DATA_ACKED
)
3362 tcp_cong_avoid(sk
, ack
, prior_in_flight
);
3365 if (tp
->tlp_high_seq
)
3366 tcp_process_tlp_ack(sk
, ack
, flag
);
3368 if ((flag
& FLAG_FORWARD_PROGRESS
) || !(flag
& FLAG_NOT_DUP
)) {
3369 struct dst_entry
*dst
= __sk_dst_get(sk
);
3374 if (icsk
->icsk_pending
== ICSK_TIME_RETRANS
)
3375 tcp_schedule_loss_probe(sk
);
3379 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3380 if (flag
& FLAG_DSACKING_ACK
)
3381 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3382 prior_packets
, is_dupack
, flag
);
3383 /* If this ack opens up a zero window, clear backoff. It was
3384 * being used to time the probes, and is probably far higher than
3385 * it needs to be for normal retransmission.
3387 if (tcp_send_head(sk
))
3390 if (tp
->tlp_high_seq
)
3391 tcp_process_tlp_ack(sk
, ack
, flag
);
3395 SOCK_DEBUG(sk
, "Ack %u after %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3399 /* If data was SACKed, tag it and see if we should send more data.
3400 * If data was DSACKed, see if we can undo a cwnd reduction.
3402 if (TCP_SKB_CB(skb
)->sacked
) {
3403 flag
|= tcp_sacktag_write_queue(sk
, skb
, prior_snd_una
);
3404 tcp_fastretrans_alert(sk
, pkts_acked
, prior_sacked
,
3405 prior_packets
, is_dupack
, flag
);
3408 SOCK_DEBUG(sk
, "Ack %u before %u:%u\n", ack
, tp
->snd_una
, tp
->snd_nxt
);
3412 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3413 * But, this can also be called on packets in the established flow when
3414 * the fast version below fails.
3416 void tcp_parse_options(const struct sk_buff
*skb
,
3417 struct tcp_options_received
*opt_rx
, int estab
,
3418 struct tcp_fastopen_cookie
*foc
)
3420 const unsigned char *ptr
;
3421 const struct tcphdr
*th
= tcp_hdr(skb
);
3422 int length
= (th
->doff
* 4) - sizeof(struct tcphdr
);
3424 ptr
= (const unsigned char *)(th
+ 1);
3425 opt_rx
->saw_tstamp
= 0;
3427 while (length
> 0) {
3428 int opcode
= *ptr
++;
3434 case TCPOPT_NOP
: /* Ref: RFC 793 section 3.1 */
3439 if (opsize
< 2) /* "silly options" */
3441 if (opsize
> length
)
3442 return; /* don't parse partial options */
3445 if (opsize
== TCPOLEN_MSS
&& th
->syn
&& !estab
) {
3446 u16 in_mss
= get_unaligned_be16(ptr
);
3448 if (opt_rx
->user_mss
&&
3449 opt_rx
->user_mss
< in_mss
)
3450 in_mss
= opt_rx
->user_mss
;
3451 opt_rx
->mss_clamp
= in_mss
;
3456 if (opsize
== TCPOLEN_WINDOW
&& th
->syn
&&
3457 !estab
&& sysctl_tcp_window_scaling
) {
3458 __u8 snd_wscale
= *(__u8
*)ptr
;
3459 opt_rx
->wscale_ok
= 1;
3460 if (snd_wscale
> 14) {
3461 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3466 opt_rx
->snd_wscale
= snd_wscale
;
3469 case TCPOPT_TIMESTAMP
:
3470 if ((opsize
== TCPOLEN_TIMESTAMP
) &&
3471 ((estab
&& opt_rx
->tstamp_ok
) ||
3472 (!estab
&& sysctl_tcp_timestamps
))) {
3473 opt_rx
->saw_tstamp
= 1;
3474 opt_rx
->rcv_tsval
= get_unaligned_be32(ptr
);
3475 opt_rx
->rcv_tsecr
= get_unaligned_be32(ptr
+ 4);
3478 case TCPOPT_SACK_PERM
:
3479 if (opsize
== TCPOLEN_SACK_PERM
&& th
->syn
&&
3480 !estab
&& sysctl_tcp_sack
) {
3481 opt_rx
->sack_ok
= TCP_SACK_SEEN
;
3482 tcp_sack_reset(opt_rx
);
3487 if ((opsize
>= (TCPOLEN_SACK_BASE
+ TCPOLEN_SACK_PERBLOCK
)) &&
3488 !((opsize
- TCPOLEN_SACK_BASE
) % TCPOLEN_SACK_PERBLOCK
) &&
3490 TCP_SKB_CB(skb
)->sacked
= (ptr
- 2) - (unsigned char *)th
;
3493 #ifdef CONFIG_TCP_MD5SIG
3496 * The MD5 Hash has already been
3497 * checked (see tcp_v{4,6}_do_rcv()).
3502 /* Fast Open option shares code 254 using a
3503 * 16 bits magic number. It's valid only in
3504 * SYN or SYN-ACK with an even size.
3506 if (opsize
< TCPOLEN_EXP_FASTOPEN_BASE
||
3507 get_unaligned_be16(ptr
) != TCPOPT_FASTOPEN_MAGIC
||
3508 foc
== NULL
|| !th
->syn
|| (opsize
& 1))
3510 foc
->len
= opsize
- TCPOLEN_EXP_FASTOPEN_BASE
;
3511 if (foc
->len
>= TCP_FASTOPEN_COOKIE_MIN
&&
3512 foc
->len
<= TCP_FASTOPEN_COOKIE_MAX
)
3513 memcpy(foc
->val
, ptr
+ 2, foc
->len
);
3514 else if (foc
->len
!= 0)
3524 EXPORT_SYMBOL(tcp_parse_options
);
3526 static bool tcp_parse_aligned_timestamp(struct tcp_sock
*tp
, const struct tcphdr
*th
)
3528 const __be32
*ptr
= (const __be32
*)(th
+ 1);
3530 if (*ptr
== htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16)
3531 | (TCPOPT_TIMESTAMP
<< 8) | TCPOLEN_TIMESTAMP
)) {
3532 tp
->rx_opt
.saw_tstamp
= 1;
3534 tp
->rx_opt
.rcv_tsval
= ntohl(*ptr
);
3536 tp
->rx_opt
.rcv_tsecr
= ntohl(*ptr
) - tp
->tsoffset
;
3542 /* Fast parse options. This hopes to only see timestamps.
3543 * If it is wrong it falls back on tcp_parse_options().
3545 static bool tcp_fast_parse_options(const struct sk_buff
*skb
,
3546 const struct tcphdr
*th
, struct tcp_sock
*tp
)
3548 /* In the spirit of fast parsing, compare doff directly to constant
3549 * values. Because equality is used, short doff can be ignored here.
3551 if (th
->doff
== (sizeof(*th
) / 4)) {
3552 tp
->rx_opt
.saw_tstamp
= 0;
3554 } else if (tp
->rx_opt
.tstamp_ok
&&
3555 th
->doff
== ((sizeof(*th
) + TCPOLEN_TSTAMP_ALIGNED
) / 4)) {
3556 if (tcp_parse_aligned_timestamp(tp
, th
))
3560 tcp_parse_options(skb
, &tp
->rx_opt
, 1, NULL
);
3561 if (tp
->rx_opt
.saw_tstamp
)
3562 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
3567 #ifdef CONFIG_TCP_MD5SIG
3569 * Parse MD5 Signature option
3571 const u8
*tcp_parse_md5sig_option(const struct tcphdr
*th
)
3573 int length
= (th
->doff
<< 2) - sizeof(*th
);
3574 const u8
*ptr
= (const u8
*)(th
+ 1);
3576 /* If the TCP option is too short, we can short cut */
3577 if (length
< TCPOLEN_MD5SIG
)
3580 while (length
> 0) {
3581 int opcode
= *ptr
++;
3592 if (opsize
< 2 || opsize
> length
)
3594 if (opcode
== TCPOPT_MD5SIG
)
3595 return opsize
== TCPOLEN_MD5SIG
? ptr
: NULL
;
3602 EXPORT_SYMBOL(tcp_parse_md5sig_option
);
3605 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3607 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3608 * it can pass through stack. So, the following predicate verifies that
3609 * this segment is not used for anything but congestion avoidance or
3610 * fast retransmit. Moreover, we even are able to eliminate most of such
3611 * second order effects, if we apply some small "replay" window (~RTO)
3612 * to timestamp space.
3614 * All these measures still do not guarantee that we reject wrapped ACKs
3615 * on networks with high bandwidth, when sequence space is recycled fastly,
3616 * but it guarantees that such events will be very rare and do not affect
3617 * connection seriously. This doesn't look nice, but alas, PAWS is really
3620 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3621 * states that events when retransmit arrives after original data are rare.
3622 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3623 * the biggest problem on large power networks even with minor reordering.
3624 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3625 * up to bandwidth of 18Gigabit/sec. 8) ]
3628 static int tcp_disordered_ack(const struct sock
*sk
, const struct sk_buff
*skb
)
3630 const struct tcp_sock
*tp
= tcp_sk(sk
);
3631 const struct tcphdr
*th
= tcp_hdr(skb
);
3632 u32 seq
= TCP_SKB_CB(skb
)->seq
;
3633 u32 ack
= TCP_SKB_CB(skb
)->ack_seq
;
3635 return (/* 1. Pure ACK with correct sequence number. */
3636 (th
->ack
&& seq
== TCP_SKB_CB(skb
)->end_seq
&& seq
== tp
->rcv_nxt
) &&
3638 /* 2. ... and duplicate ACK. */
3639 ack
== tp
->snd_una
&&
3641 /* 3. ... and does not update window. */
3642 !tcp_may_update_window(tp
, ack
, seq
, ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
) &&
3644 /* 4. ... and sits in replay window. */
3645 (s32
)(tp
->rx_opt
.ts_recent
- tp
->rx_opt
.rcv_tsval
) <= (inet_csk(sk
)->icsk_rto
* 1024) / HZ
);
3648 static inline bool tcp_paws_discard(const struct sock
*sk
,
3649 const struct sk_buff
*skb
)
3651 const struct tcp_sock
*tp
= tcp_sk(sk
);
3653 return !tcp_paws_check(&tp
->rx_opt
, TCP_PAWS_WINDOW
) &&
3654 !tcp_disordered_ack(sk
, skb
);
3657 /* Check segment sequence number for validity.
3659 * Segment controls are considered valid, if the segment
3660 * fits to the window after truncation to the window. Acceptability
3661 * of data (and SYN, FIN, of course) is checked separately.
3662 * See tcp_data_queue(), for example.
3664 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3665 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3666 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3667 * (borrowed from freebsd)
3670 static inline bool tcp_sequence(const struct tcp_sock
*tp
, u32 seq
, u32 end_seq
)
3672 return !before(end_seq
, tp
->rcv_wup
) &&
3673 !after(seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
));
3676 /* When we get a reset we do this. */
3677 void tcp_reset(struct sock
*sk
)
3679 /* We want the right error as BSD sees it (and indeed as we do). */
3680 switch (sk
->sk_state
) {
3682 sk
->sk_err
= ECONNREFUSED
;
3684 case TCP_CLOSE_WAIT
:
3690 sk
->sk_err
= ECONNRESET
;
3692 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3695 if (!sock_flag(sk
, SOCK_DEAD
))
3696 sk
->sk_error_report(sk
);
3702 * Process the FIN bit. This now behaves as it is supposed to work
3703 * and the FIN takes effect when it is validly part of sequence
3704 * space. Not before when we get holes.
3706 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3707 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3710 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3711 * close and we go into CLOSING (and later onto TIME-WAIT)
3713 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3715 static void tcp_fin(struct sock
*sk
)
3717 struct tcp_sock
*tp
= tcp_sk(sk
);
3719 inet_csk_schedule_ack(sk
);
3721 sk
->sk_shutdown
|= RCV_SHUTDOWN
;
3722 sock_set_flag(sk
, SOCK_DONE
);
3724 switch (sk
->sk_state
) {
3726 case TCP_ESTABLISHED
:
3727 /* Move to CLOSE_WAIT */
3728 tcp_set_state(sk
, TCP_CLOSE_WAIT
);
3729 inet_csk(sk
)->icsk_ack
.pingpong
= 1;
3732 case TCP_CLOSE_WAIT
:
3734 /* Received a retransmission of the FIN, do
3739 /* RFC793: Remain in the LAST-ACK state. */
3743 /* This case occurs when a simultaneous close
3744 * happens, we must ack the received FIN and
3745 * enter the CLOSING state.
3748 tcp_set_state(sk
, TCP_CLOSING
);
3751 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3753 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
3756 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3757 * cases we should never reach this piece of code.
3759 pr_err("%s: Impossible, sk->sk_state=%d\n",
3760 __func__
, sk
->sk_state
);
3764 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3765 * Probably, we should reset in this case. For now drop them.
3767 __skb_queue_purge(&tp
->out_of_order_queue
);
3768 if (tcp_is_sack(tp
))
3769 tcp_sack_reset(&tp
->rx_opt
);
3772 if (!sock_flag(sk
, SOCK_DEAD
)) {
3773 sk
->sk_state_change(sk
);
3775 /* Do not send POLL_HUP for half duplex close. */
3776 if (sk
->sk_shutdown
== SHUTDOWN_MASK
||
3777 sk
->sk_state
== TCP_CLOSE
)
3778 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_HUP
);
3780 sk_wake_async(sk
, SOCK_WAKE_WAITD
, POLL_IN
);
3784 static inline bool tcp_sack_extend(struct tcp_sack_block
*sp
, u32 seq
,
3787 if (!after(seq
, sp
->end_seq
) && !after(sp
->start_seq
, end_seq
)) {
3788 if (before(seq
, sp
->start_seq
))
3789 sp
->start_seq
= seq
;
3790 if (after(end_seq
, sp
->end_seq
))
3791 sp
->end_seq
= end_seq
;
3797 static void tcp_dsack_set(struct sock
*sk
, u32 seq
, u32 end_seq
)
3799 struct tcp_sock
*tp
= tcp_sk(sk
);
3801 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3804 if (before(seq
, tp
->rcv_nxt
))
3805 mib_idx
= LINUX_MIB_TCPDSACKOLDSENT
;
3807 mib_idx
= LINUX_MIB_TCPDSACKOFOSENT
;
3809 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
3811 tp
->rx_opt
.dsack
= 1;
3812 tp
->duplicate_sack
[0].start_seq
= seq
;
3813 tp
->duplicate_sack
[0].end_seq
= end_seq
;
3817 static void tcp_dsack_extend(struct sock
*sk
, u32 seq
, u32 end_seq
)
3819 struct tcp_sock
*tp
= tcp_sk(sk
);
3821 if (!tp
->rx_opt
.dsack
)
3822 tcp_dsack_set(sk
, seq
, end_seq
);
3824 tcp_sack_extend(tp
->duplicate_sack
, seq
, end_seq
);
3827 static void tcp_send_dupack(struct sock
*sk
, const struct sk_buff
*skb
)
3829 struct tcp_sock
*tp
= tcp_sk(sk
);
3831 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
3832 before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
3833 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
3834 tcp_enter_quickack_mode(sk
);
3836 if (tcp_is_sack(tp
) && sysctl_tcp_dsack
) {
3837 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
3839 if (after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
))
3840 end_seq
= tp
->rcv_nxt
;
3841 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, end_seq
);
3848 /* These routines update the SACK block as out-of-order packets arrive or
3849 * in-order packets close up the sequence space.
3851 static void tcp_sack_maybe_coalesce(struct tcp_sock
*tp
)
3854 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3855 struct tcp_sack_block
*swalk
= sp
+ 1;
3857 /* See if the recent change to the first SACK eats into
3858 * or hits the sequence space of other SACK blocks, if so coalesce.
3860 for (this_sack
= 1; this_sack
< tp
->rx_opt
.num_sacks
;) {
3861 if (tcp_sack_extend(sp
, swalk
->start_seq
, swalk
->end_seq
)) {
3864 /* Zap SWALK, by moving every further SACK up by one slot.
3865 * Decrease num_sacks.
3867 tp
->rx_opt
.num_sacks
--;
3868 for (i
= this_sack
; i
< tp
->rx_opt
.num_sacks
; i
++)
3872 this_sack
++, swalk
++;
3876 static void tcp_sack_new_ofo_skb(struct sock
*sk
, u32 seq
, u32 end_seq
)
3878 struct tcp_sock
*tp
= tcp_sk(sk
);
3879 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3880 int cur_sacks
= tp
->rx_opt
.num_sacks
;
3886 for (this_sack
= 0; this_sack
< cur_sacks
; this_sack
++, sp
++) {
3887 if (tcp_sack_extend(sp
, seq
, end_seq
)) {
3888 /* Rotate this_sack to the first one. */
3889 for (; this_sack
> 0; this_sack
--, sp
--)
3890 swap(*sp
, *(sp
- 1));
3892 tcp_sack_maybe_coalesce(tp
);
3897 /* Could not find an adjacent existing SACK, build a new one,
3898 * put it at the front, and shift everyone else down. We
3899 * always know there is at least one SACK present already here.
3901 * If the sack array is full, forget about the last one.
3903 if (this_sack
>= TCP_NUM_SACKS
) {
3905 tp
->rx_opt
.num_sacks
--;
3908 for (; this_sack
> 0; this_sack
--, sp
--)
3912 /* Build the new head SACK, and we're done. */
3913 sp
->start_seq
= seq
;
3914 sp
->end_seq
= end_seq
;
3915 tp
->rx_opt
.num_sacks
++;
3918 /* RCV.NXT advances, some SACKs should be eaten. */
3920 static void tcp_sack_remove(struct tcp_sock
*tp
)
3922 struct tcp_sack_block
*sp
= &tp
->selective_acks
[0];
3923 int num_sacks
= tp
->rx_opt
.num_sacks
;
3926 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3927 if (skb_queue_empty(&tp
->out_of_order_queue
)) {
3928 tp
->rx_opt
.num_sacks
= 0;
3932 for (this_sack
= 0; this_sack
< num_sacks
;) {
3933 /* Check if the start of the sack is covered by RCV.NXT. */
3934 if (!before(tp
->rcv_nxt
, sp
->start_seq
)) {
3937 /* RCV.NXT must cover all the block! */
3938 WARN_ON(before(tp
->rcv_nxt
, sp
->end_seq
));
3940 /* Zap this SACK, by moving forward any other SACKS. */
3941 for (i
=this_sack
+1; i
< num_sacks
; i
++)
3942 tp
->selective_acks
[i
-1] = tp
->selective_acks
[i
];
3949 tp
->rx_opt
.num_sacks
= num_sacks
;
3952 /* This one checks to see if we can put data from the
3953 * out_of_order queue into the receive_queue.
3955 static void tcp_ofo_queue(struct sock
*sk
)
3957 struct tcp_sock
*tp
= tcp_sk(sk
);
3958 __u32 dsack_high
= tp
->rcv_nxt
;
3959 struct sk_buff
*skb
;
3961 while ((skb
= skb_peek(&tp
->out_of_order_queue
)) != NULL
) {
3962 if (after(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
3965 if (before(TCP_SKB_CB(skb
)->seq
, dsack_high
)) {
3966 __u32 dsack
= dsack_high
;
3967 if (before(TCP_SKB_CB(skb
)->end_seq
, dsack_high
))
3968 dsack_high
= TCP_SKB_CB(skb
)->end_seq
;
3969 tcp_dsack_extend(sk
, TCP_SKB_CB(skb
)->seq
, dsack
);
3972 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
3973 SOCK_DEBUG(sk
, "ofo packet was already received\n");
3974 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3978 SOCK_DEBUG(sk
, "ofo requeuing : rcv_next %X seq %X - %X\n",
3979 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
3980 TCP_SKB_CB(skb
)->end_seq
);
3982 __skb_unlink(skb
, &tp
->out_of_order_queue
);
3983 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
3984 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
3985 if (tcp_hdr(skb
)->fin
)
3990 static bool tcp_prune_ofo_queue(struct sock
*sk
);
3991 static int tcp_prune_queue(struct sock
*sk
);
3993 static int tcp_try_rmem_schedule(struct sock
*sk
, struct sk_buff
*skb
,
3996 if (atomic_read(&sk
->sk_rmem_alloc
) > sk
->sk_rcvbuf
||
3997 !sk_rmem_schedule(sk
, skb
, size
)) {
3999 if (tcp_prune_queue(sk
) < 0)
4002 if (!sk_rmem_schedule(sk
, skb
, size
)) {
4003 if (!tcp_prune_ofo_queue(sk
))
4006 if (!sk_rmem_schedule(sk
, skb
, size
))
4014 * tcp_try_coalesce - try to merge skb to prior one
4017 * @from: buffer to add in queue
4018 * @fragstolen: pointer to boolean
4020 * Before queueing skb @from after @to, try to merge them
4021 * to reduce overall memory use and queue lengths, if cost is small.
4022 * Packets in ofo or receive queues can stay a long time.
4023 * Better try to coalesce them right now to avoid future collapses.
4024 * Returns true if caller should free @from instead of queueing it
4026 static bool tcp_try_coalesce(struct sock
*sk
,
4028 struct sk_buff
*from
,
4033 *fragstolen
= false;
4035 if (tcp_hdr(from
)->fin
)
4038 /* Its possible this segment overlaps with prior segment in queue */
4039 if (TCP_SKB_CB(from
)->seq
!= TCP_SKB_CB(to
)->end_seq
)
4042 if (!skb_try_coalesce(to
, from
, fragstolen
, &delta
))
4045 atomic_add(delta
, &sk
->sk_rmem_alloc
);
4046 sk_mem_charge(sk
, delta
);
4047 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOALESCE
);
4048 TCP_SKB_CB(to
)->end_seq
= TCP_SKB_CB(from
)->end_seq
;
4049 TCP_SKB_CB(to
)->ack_seq
= TCP_SKB_CB(from
)->ack_seq
;
4053 static void tcp_data_queue_ofo(struct sock
*sk
, struct sk_buff
*skb
)
4055 struct tcp_sock
*tp
= tcp_sk(sk
);
4056 struct sk_buff
*skb1
;
4059 TCP_ECN_check_ce(tp
, skb
);
4061 if (unlikely(tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))) {
4062 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFODROP
);
4067 /* Disable header prediction. */
4069 inet_csk_schedule_ack(sk
);
4071 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOQUEUE
);
4072 SOCK_DEBUG(sk
, "out of order segment: rcv_next %X seq %X - %X\n",
4073 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4075 skb1
= skb_peek_tail(&tp
->out_of_order_queue
);
4077 /* Initial out of order segment, build 1 SACK. */
4078 if (tcp_is_sack(tp
)) {
4079 tp
->rx_opt
.num_sacks
= 1;
4080 tp
->selective_acks
[0].start_seq
= TCP_SKB_CB(skb
)->seq
;
4081 tp
->selective_acks
[0].end_seq
=
4082 TCP_SKB_CB(skb
)->end_seq
;
4084 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4088 seq
= TCP_SKB_CB(skb
)->seq
;
4089 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
4091 if (seq
== TCP_SKB_CB(skb1
)->end_seq
) {
4094 if (!tcp_try_coalesce(sk
, skb1
, skb
, &fragstolen
)) {
4095 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4097 kfree_skb_partial(skb
, fragstolen
);
4101 if (!tp
->rx_opt
.num_sacks
||
4102 tp
->selective_acks
[0].end_seq
!= seq
)
4105 /* Common case: data arrive in order after hole. */
4106 tp
->selective_acks
[0].end_seq
= end_seq
;
4110 /* Find place to insert this segment. */
4112 if (!after(TCP_SKB_CB(skb1
)->seq
, seq
))
4114 if (skb_queue_is_first(&tp
->out_of_order_queue
, skb1
)) {
4118 skb1
= skb_queue_prev(&tp
->out_of_order_queue
, skb1
);
4121 /* Do skb overlap to previous one? */
4122 if (skb1
&& before(seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4123 if (!after(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4124 /* All the bits are present. Drop. */
4125 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4128 tcp_dsack_set(sk
, seq
, end_seq
);
4131 if (after(seq
, TCP_SKB_CB(skb1
)->seq
)) {
4132 /* Partial overlap. */
4133 tcp_dsack_set(sk
, seq
,
4134 TCP_SKB_CB(skb1
)->end_seq
);
4136 if (skb_queue_is_first(&tp
->out_of_order_queue
,
4140 skb1
= skb_queue_prev(
4141 &tp
->out_of_order_queue
,
4146 __skb_queue_head(&tp
->out_of_order_queue
, skb
);
4148 __skb_queue_after(&tp
->out_of_order_queue
, skb1
, skb
);
4150 /* And clean segments covered by new one as whole. */
4151 while (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
)) {
4152 skb1
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4154 if (!after(end_seq
, TCP_SKB_CB(skb1
)->seq
))
4156 if (before(end_seq
, TCP_SKB_CB(skb1
)->end_seq
)) {
4157 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4161 __skb_unlink(skb1
, &tp
->out_of_order_queue
);
4162 tcp_dsack_extend(sk
, TCP_SKB_CB(skb1
)->seq
,
4163 TCP_SKB_CB(skb1
)->end_seq
);
4164 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPOFOMERGE
);
4169 if (tcp_is_sack(tp
))
4170 tcp_sack_new_ofo_skb(sk
, seq
, end_seq
);
4173 skb_set_owner_r(skb
, sk
);
4176 static int __must_check
tcp_queue_rcv(struct sock
*sk
, struct sk_buff
*skb
, int hdrlen
,
4180 struct sk_buff
*tail
= skb_peek_tail(&sk
->sk_receive_queue
);
4182 __skb_pull(skb
, hdrlen
);
4184 tcp_try_coalesce(sk
, tail
, skb
, fragstolen
)) ? 1 : 0;
4185 tcp_sk(sk
)->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4187 __skb_queue_tail(&sk
->sk_receive_queue
, skb
);
4188 skb_set_owner_r(skb
, sk
);
4193 int tcp_send_rcvq(struct sock
*sk
, struct msghdr
*msg
, size_t size
)
4195 struct sk_buff
*skb
= NULL
;
4202 skb
= alloc_skb(size
+ sizeof(*th
), sk
->sk_allocation
);
4206 if (tcp_try_rmem_schedule(sk
, skb
, size
+ sizeof(*th
)))
4209 th
= (struct tcphdr
*)skb_put(skb
, sizeof(*th
));
4210 skb_reset_transport_header(skb
);
4211 memset(th
, 0, sizeof(*th
));
4213 if (memcpy_fromiovec(skb_put(skb
, size
), msg
->msg_iov
, size
))
4216 TCP_SKB_CB(skb
)->seq
= tcp_sk(sk
)->rcv_nxt
;
4217 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ size
;
4218 TCP_SKB_CB(skb
)->ack_seq
= tcp_sk(sk
)->snd_una
- 1;
4220 if (tcp_queue_rcv(sk
, skb
, sizeof(*th
), &fragstolen
)) {
4221 WARN_ON_ONCE(fragstolen
); /* should not happen */
4232 static void tcp_data_queue(struct sock
*sk
, struct sk_buff
*skb
)
4234 const struct tcphdr
*th
= tcp_hdr(skb
);
4235 struct tcp_sock
*tp
= tcp_sk(sk
);
4237 bool fragstolen
= false;
4239 if (TCP_SKB_CB(skb
)->seq
== TCP_SKB_CB(skb
)->end_seq
)
4243 __skb_pull(skb
, th
->doff
* 4);
4245 TCP_ECN_accept_cwr(tp
, skb
);
4247 tp
->rx_opt
.dsack
= 0;
4249 /* Queue data for delivery to the user.
4250 * Packets in sequence go to the receive queue.
4251 * Out of sequence packets to the out_of_order_queue.
4253 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
) {
4254 if (tcp_receive_window(tp
) == 0)
4257 /* Ok. In sequence. In window. */
4258 if (tp
->ucopy
.task
== current
&&
4259 tp
->copied_seq
== tp
->rcv_nxt
&& tp
->ucopy
.len
&&
4260 sock_owned_by_user(sk
) && !tp
->urg_data
) {
4261 int chunk
= min_t(unsigned int, skb
->len
,
4264 __set_current_state(TASK_RUNNING
);
4267 if (!skb_copy_datagram_iovec(skb
, 0, tp
->ucopy
.iov
, chunk
)) {
4268 tp
->ucopy
.len
-= chunk
;
4269 tp
->copied_seq
+= chunk
;
4270 eaten
= (chunk
== skb
->len
);
4271 tcp_rcv_space_adjust(sk
);
4279 tcp_try_rmem_schedule(sk
, skb
, skb
->truesize
))
4282 eaten
= tcp_queue_rcv(sk
, skb
, 0, &fragstolen
);
4284 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
4286 tcp_event_data_recv(sk
, skb
);
4290 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4293 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4294 * gap in queue is filled.
4296 if (skb_queue_empty(&tp
->out_of_order_queue
))
4297 inet_csk(sk
)->icsk_ack
.pingpong
= 0;
4300 if (tp
->rx_opt
.num_sacks
)
4301 tcp_sack_remove(tp
);
4303 tcp_fast_path_check(sk
);
4306 kfree_skb_partial(skb
, fragstolen
);
4307 if (!sock_flag(sk
, SOCK_DEAD
))
4308 sk
->sk_data_ready(sk
, 0);
4312 if (!after(TCP_SKB_CB(skb
)->end_seq
, tp
->rcv_nxt
)) {
4313 /* A retransmit, 2nd most common case. Force an immediate ack. */
4314 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_DELAYEDACKLOST
);
4315 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
);
4318 tcp_enter_quickack_mode(sk
);
4319 inet_csk_schedule_ack(sk
);
4325 /* Out of window. F.e. zero window probe. */
4326 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
+ tcp_receive_window(tp
)))
4329 tcp_enter_quickack_mode(sk
);
4331 if (before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
)) {
4332 /* Partial packet, seq < rcv_next < end_seq */
4333 SOCK_DEBUG(sk
, "partial packet: rcv_next %X seq %X - %X\n",
4334 tp
->rcv_nxt
, TCP_SKB_CB(skb
)->seq
,
4335 TCP_SKB_CB(skb
)->end_seq
);
4337 tcp_dsack_set(sk
, TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
);
4339 /* If window is closed, drop tail of packet. But after
4340 * remembering D-SACK for its head made in previous line.
4342 if (!tcp_receive_window(tp
))
4347 tcp_data_queue_ofo(sk
, skb
);
4350 static struct sk_buff
*tcp_collapse_one(struct sock
*sk
, struct sk_buff
*skb
,
4351 struct sk_buff_head
*list
)
4353 struct sk_buff
*next
= NULL
;
4355 if (!skb_queue_is_last(list
, skb
))
4356 next
= skb_queue_next(list
, skb
);
4358 __skb_unlink(skb
, list
);
4360 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRCVCOLLAPSED
);
4365 /* Collapse contiguous sequence of skbs head..tail with
4366 * sequence numbers start..end.
4368 * If tail is NULL, this means until the end of the list.
4370 * Segments with FIN/SYN are not collapsed (only because this
4374 tcp_collapse(struct sock
*sk
, struct sk_buff_head
*list
,
4375 struct sk_buff
*head
, struct sk_buff
*tail
,
4378 struct sk_buff
*skb
, *n
;
4381 /* First, check that queue is collapsible and find
4382 * the point where collapsing can be useful. */
4386 skb_queue_walk_from_safe(list
, skb
, n
) {
4389 /* No new bits? It is possible on ofo queue. */
4390 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4391 skb
= tcp_collapse_one(sk
, skb
, list
);
4397 /* The first skb to collapse is:
4399 * - bloated or contains data before "start" or
4400 * overlaps to the next one.
4402 if (!tcp_hdr(skb
)->syn
&& !tcp_hdr(skb
)->fin
&&
4403 (tcp_win_from_space(skb
->truesize
) > skb
->len
||
4404 before(TCP_SKB_CB(skb
)->seq
, start
))) {
4405 end_of_skbs
= false;
4409 if (!skb_queue_is_last(list
, skb
)) {
4410 struct sk_buff
*next
= skb_queue_next(list
, skb
);
4412 TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(next
)->seq
) {
4413 end_of_skbs
= false;
4418 /* Decided to skip this, advance start seq. */
4419 start
= TCP_SKB_CB(skb
)->end_seq
;
4421 if (end_of_skbs
|| tcp_hdr(skb
)->syn
|| tcp_hdr(skb
)->fin
)
4424 while (before(start
, end
)) {
4425 struct sk_buff
*nskb
;
4426 unsigned int header
= skb_headroom(skb
);
4427 int copy
= SKB_MAX_ORDER(header
, 0);
4429 /* Too big header? This can happen with IPv6. */
4432 if (end
- start
< copy
)
4434 nskb
= alloc_skb(copy
+ header
, GFP_ATOMIC
);
4438 skb_set_mac_header(nskb
, skb_mac_header(skb
) - skb
->head
);
4439 skb_set_network_header(nskb
, (skb_network_header(skb
) -
4441 skb_set_transport_header(nskb
, (skb_transport_header(skb
) -
4443 skb_reserve(nskb
, header
);
4444 memcpy(nskb
->head
, skb
->head
, header
);
4445 memcpy(nskb
->cb
, skb
->cb
, sizeof(skb
->cb
));
4446 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(nskb
)->end_seq
= start
;
4447 __skb_queue_before(list
, skb
, nskb
);
4448 skb_set_owner_r(nskb
, sk
);
4450 /* Copy data, releasing collapsed skbs. */
4452 int offset
= start
- TCP_SKB_CB(skb
)->seq
;
4453 int size
= TCP_SKB_CB(skb
)->end_seq
- start
;
4457 size
= min(copy
, size
);
4458 if (skb_copy_bits(skb
, offset
, skb_put(nskb
, size
), size
))
4460 TCP_SKB_CB(nskb
)->end_seq
+= size
;
4464 if (!before(start
, TCP_SKB_CB(skb
)->end_seq
)) {
4465 skb
= tcp_collapse_one(sk
, skb
, list
);
4468 tcp_hdr(skb
)->syn
||
4476 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4477 * and tcp_collapse() them until all the queue is collapsed.
4479 static void tcp_collapse_ofo_queue(struct sock
*sk
)
4481 struct tcp_sock
*tp
= tcp_sk(sk
);
4482 struct sk_buff
*skb
= skb_peek(&tp
->out_of_order_queue
);
4483 struct sk_buff
*head
;
4489 start
= TCP_SKB_CB(skb
)->seq
;
4490 end
= TCP_SKB_CB(skb
)->end_seq
;
4494 struct sk_buff
*next
= NULL
;
4496 if (!skb_queue_is_last(&tp
->out_of_order_queue
, skb
))
4497 next
= skb_queue_next(&tp
->out_of_order_queue
, skb
);
4500 /* Segment is terminated when we see gap or when
4501 * we are at the end of all the queue. */
4503 after(TCP_SKB_CB(skb
)->seq
, end
) ||
4504 before(TCP_SKB_CB(skb
)->end_seq
, start
)) {
4505 tcp_collapse(sk
, &tp
->out_of_order_queue
,
4506 head
, skb
, start
, end
);
4510 /* Start new segment */
4511 start
= TCP_SKB_CB(skb
)->seq
;
4512 end
= TCP_SKB_CB(skb
)->end_seq
;
4514 if (before(TCP_SKB_CB(skb
)->seq
, start
))
4515 start
= TCP_SKB_CB(skb
)->seq
;
4516 if (after(TCP_SKB_CB(skb
)->end_seq
, end
))
4517 end
= TCP_SKB_CB(skb
)->end_seq
;
4523 * Purge the out-of-order queue.
4524 * Return true if queue was pruned.
4526 static bool tcp_prune_ofo_queue(struct sock
*sk
)
4528 struct tcp_sock
*tp
= tcp_sk(sk
);
4531 if (!skb_queue_empty(&tp
->out_of_order_queue
)) {
4532 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_OFOPRUNED
);
4533 __skb_queue_purge(&tp
->out_of_order_queue
);
4535 /* Reset SACK state. A conforming SACK implementation will
4536 * do the same at a timeout based retransmit. When a connection
4537 * is in a sad state like this, we care only about integrity
4538 * of the connection not performance.
4540 if (tp
->rx_opt
.sack_ok
)
4541 tcp_sack_reset(&tp
->rx_opt
);
4548 /* Reduce allocated memory if we can, trying to get
4549 * the socket within its memory limits again.
4551 * Return less than zero if we should start dropping frames
4552 * until the socket owning process reads some of the data
4553 * to stabilize the situation.
4555 static int tcp_prune_queue(struct sock
*sk
)
4557 struct tcp_sock
*tp
= tcp_sk(sk
);
4559 SOCK_DEBUG(sk
, "prune_queue: c=%x\n", tp
->copied_seq
);
4561 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PRUNECALLED
);
4563 if (atomic_read(&sk
->sk_rmem_alloc
) >= sk
->sk_rcvbuf
)
4564 tcp_clamp_window(sk
);
4565 else if (sk_under_memory_pressure(sk
))
4566 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
, 4U * tp
->advmss
);
4568 tcp_collapse_ofo_queue(sk
);
4569 if (!skb_queue_empty(&sk
->sk_receive_queue
))
4570 tcp_collapse(sk
, &sk
->sk_receive_queue
,
4571 skb_peek(&sk
->sk_receive_queue
),
4573 tp
->copied_seq
, tp
->rcv_nxt
);
4576 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4579 /* Collapsing did not help, destructive actions follow.
4580 * This must not ever occur. */
4582 tcp_prune_ofo_queue(sk
);
4584 if (atomic_read(&sk
->sk_rmem_alloc
) <= sk
->sk_rcvbuf
)
4587 /* If we are really being abused, tell the caller to silently
4588 * drop receive data on the floor. It will get retransmitted
4589 * and hopefully then we'll have sufficient space.
4591 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_RCVPRUNED
);
4593 /* Massive buffer overcommit. */
4598 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4599 * As additional protections, we do not touch cwnd in retransmission phases,
4600 * and if application hit its sndbuf limit recently.
4602 void tcp_cwnd_application_limited(struct sock
*sk
)
4604 struct tcp_sock
*tp
= tcp_sk(sk
);
4606 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
4607 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
4608 /* Limited by application or receiver window. */
4609 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
4610 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
4611 if (win_used
< tp
->snd_cwnd
) {
4612 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
4613 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
4615 tp
->snd_cwnd_used
= 0;
4617 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4620 static bool tcp_should_expand_sndbuf(const struct sock
*sk
)
4622 const struct tcp_sock
*tp
= tcp_sk(sk
);
4624 /* If the user specified a specific send buffer setting, do
4627 if (sk
->sk_userlocks
& SOCK_SNDBUF_LOCK
)
4630 /* If we are under global TCP memory pressure, do not expand. */
4631 if (sk_under_memory_pressure(sk
))
4634 /* If we are under soft global TCP memory pressure, do not expand. */
4635 if (sk_memory_allocated(sk
) >= sk_prot_mem_limits(sk
, 0))
4638 /* If we filled the congestion window, do not expand. */
4639 if (tp
->packets_out
>= tp
->snd_cwnd
)
4645 /* When incoming ACK allowed to free some skb from write_queue,
4646 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4647 * on the exit from tcp input handler.
4649 * PROBLEM: sndbuf expansion does not work well with largesend.
4651 static void tcp_new_space(struct sock
*sk
)
4653 struct tcp_sock
*tp
= tcp_sk(sk
);
4655 if (tcp_should_expand_sndbuf(sk
)) {
4656 int sndmem
= SKB_TRUESIZE(max_t(u32
,
4657 tp
->rx_opt
.mss_clamp
,
4660 int demanded
= max_t(unsigned int, tp
->snd_cwnd
,
4661 tp
->reordering
+ 1);
4662 sndmem
*= 2 * demanded
;
4663 if (sndmem
> sk
->sk_sndbuf
)
4664 sk
->sk_sndbuf
= min(sndmem
, sysctl_tcp_wmem
[2]);
4665 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
4668 sk
->sk_write_space(sk
);
4671 static void tcp_check_space(struct sock
*sk
)
4673 if (sock_flag(sk
, SOCK_QUEUE_SHRUNK
)) {
4674 sock_reset_flag(sk
, SOCK_QUEUE_SHRUNK
);
4675 if (sk
->sk_socket
&&
4676 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
))
4681 static inline void tcp_data_snd_check(struct sock
*sk
)
4683 tcp_push_pending_frames(sk
);
4684 tcp_check_space(sk
);
4688 * Check if sending an ack is needed.
4690 static void __tcp_ack_snd_check(struct sock
*sk
, int ofo_possible
)
4692 struct tcp_sock
*tp
= tcp_sk(sk
);
4694 /* More than one full frame received... */
4695 if (((tp
->rcv_nxt
- tp
->rcv_wup
) > inet_csk(sk
)->icsk_ack
.rcv_mss
&&
4696 /* ... and right edge of window advances far enough.
4697 * (tcp_recvmsg() will send ACK otherwise). Or...
4699 __tcp_select_window(sk
) >= tp
->rcv_wnd
) ||
4700 /* We ACK each frame or... */
4701 tcp_in_quickack_mode(sk
) ||
4702 /* We have out of order data. */
4703 (ofo_possible
&& skb_peek(&tp
->out_of_order_queue
))) {
4704 /* Then ack it now */
4707 /* Else, send delayed ack. */
4708 tcp_send_delayed_ack(sk
);
4712 static inline void tcp_ack_snd_check(struct sock
*sk
)
4714 if (!inet_csk_ack_scheduled(sk
)) {
4715 /* We sent a data segment already. */
4718 __tcp_ack_snd_check(sk
, 1);
4722 * This routine is only called when we have urgent data
4723 * signaled. Its the 'slow' part of tcp_urg. It could be
4724 * moved inline now as tcp_urg is only called from one
4725 * place. We handle URGent data wrong. We have to - as
4726 * BSD still doesn't use the correction from RFC961.
4727 * For 1003.1g we should support a new option TCP_STDURG to permit
4728 * either form (or just set the sysctl tcp_stdurg).
4731 static void tcp_check_urg(struct sock
*sk
, const struct tcphdr
*th
)
4733 struct tcp_sock
*tp
= tcp_sk(sk
);
4734 u32 ptr
= ntohs(th
->urg_ptr
);
4736 if (ptr
&& !sysctl_tcp_stdurg
)
4738 ptr
+= ntohl(th
->seq
);
4740 /* Ignore urgent data that we've already seen and read. */
4741 if (after(tp
->copied_seq
, ptr
))
4744 /* Do not replay urg ptr.
4746 * NOTE: interesting situation not covered by specs.
4747 * Misbehaving sender may send urg ptr, pointing to segment,
4748 * which we already have in ofo queue. We are not able to fetch
4749 * such data and will stay in TCP_URG_NOTYET until will be eaten
4750 * by recvmsg(). Seems, we are not obliged to handle such wicked
4751 * situations. But it is worth to think about possibility of some
4752 * DoSes using some hypothetical application level deadlock.
4754 if (before(ptr
, tp
->rcv_nxt
))
4757 /* Do we already have a newer (or duplicate) urgent pointer? */
4758 if (tp
->urg_data
&& !after(ptr
, tp
->urg_seq
))
4761 /* Tell the world about our new urgent pointer. */
4764 /* We may be adding urgent data when the last byte read was
4765 * urgent. To do this requires some care. We cannot just ignore
4766 * tp->copied_seq since we would read the last urgent byte again
4767 * as data, nor can we alter copied_seq until this data arrives
4768 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4770 * NOTE. Double Dutch. Rendering to plain English: author of comment
4771 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4772 * and expect that both A and B disappear from stream. This is _wrong_.
4773 * Though this happens in BSD with high probability, this is occasional.
4774 * Any application relying on this is buggy. Note also, that fix "works"
4775 * only in this artificial test. Insert some normal data between A and B and we will
4776 * decline of BSD again. Verdict: it is better to remove to trap
4779 if (tp
->urg_seq
== tp
->copied_seq
&& tp
->urg_data
&&
4780 !sock_flag(sk
, SOCK_URGINLINE
) && tp
->copied_seq
!= tp
->rcv_nxt
) {
4781 struct sk_buff
*skb
= skb_peek(&sk
->sk_receive_queue
);
4783 if (skb
&& !before(tp
->copied_seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4784 __skb_unlink(skb
, &sk
->sk_receive_queue
);
4789 tp
->urg_data
= TCP_URG_NOTYET
;
4792 /* Disable header prediction. */
4796 /* This is the 'fast' part of urgent handling. */
4797 static void tcp_urg(struct sock
*sk
, struct sk_buff
*skb
, const struct tcphdr
*th
)
4799 struct tcp_sock
*tp
= tcp_sk(sk
);
4801 /* Check if we get a new urgent pointer - normally not. */
4803 tcp_check_urg(sk
, th
);
4805 /* Do we wait for any urgent data? - normally not... */
4806 if (tp
->urg_data
== TCP_URG_NOTYET
) {
4807 u32 ptr
= tp
->urg_seq
- ntohl(th
->seq
) + (th
->doff
* 4) -
4810 /* Is the urgent pointer pointing into this packet? */
4811 if (ptr
< skb
->len
) {
4813 if (skb_copy_bits(skb
, ptr
, &tmp
, 1))
4815 tp
->urg_data
= TCP_URG_VALID
| tmp
;
4816 if (!sock_flag(sk
, SOCK_DEAD
))
4817 sk
->sk_data_ready(sk
, 0);
4822 static int tcp_copy_to_iovec(struct sock
*sk
, struct sk_buff
*skb
, int hlen
)
4824 struct tcp_sock
*tp
= tcp_sk(sk
);
4825 int chunk
= skb
->len
- hlen
;
4829 if (skb_csum_unnecessary(skb
))
4830 err
= skb_copy_datagram_iovec(skb
, hlen
, tp
->ucopy
.iov
, chunk
);
4832 err
= skb_copy_and_csum_datagram_iovec(skb
, hlen
,
4836 tp
->ucopy
.len
-= chunk
;
4837 tp
->copied_seq
+= chunk
;
4838 tcp_rcv_space_adjust(sk
);
4845 static __sum16
__tcp_checksum_complete_user(struct sock
*sk
,
4846 struct sk_buff
*skb
)
4850 if (sock_owned_by_user(sk
)) {
4852 result
= __tcp_checksum_complete(skb
);
4855 result
= __tcp_checksum_complete(skb
);
4860 static inline bool tcp_checksum_complete_user(struct sock
*sk
,
4861 struct sk_buff
*skb
)
4863 return !skb_csum_unnecessary(skb
) &&
4864 __tcp_checksum_complete_user(sk
, skb
);
4867 #ifdef CONFIG_NET_DMA
4868 static bool tcp_dma_try_early_copy(struct sock
*sk
, struct sk_buff
*skb
,
4871 struct tcp_sock
*tp
= tcp_sk(sk
);
4872 int chunk
= skb
->len
- hlen
;
4874 bool copied_early
= false;
4876 if (tp
->ucopy
.wakeup
)
4879 if (!tp
->ucopy
.dma_chan
&& tp
->ucopy
.pinned_list
)
4880 tp
->ucopy
.dma_chan
= net_dma_find_channel();
4882 if (tp
->ucopy
.dma_chan
&& skb_csum_unnecessary(skb
)) {
4884 dma_cookie
= dma_skb_copy_datagram_iovec(tp
->ucopy
.dma_chan
,
4886 tp
->ucopy
.iov
, chunk
,
4887 tp
->ucopy
.pinned_list
);
4892 tp
->ucopy
.dma_cookie
= dma_cookie
;
4893 copied_early
= true;
4895 tp
->ucopy
.len
-= chunk
;
4896 tp
->copied_seq
+= chunk
;
4897 tcp_rcv_space_adjust(sk
);
4899 if ((tp
->ucopy
.len
== 0) ||
4900 (tcp_flag_word(tcp_hdr(skb
)) & TCP_FLAG_PSH
) ||
4901 (atomic_read(&sk
->sk_rmem_alloc
) > (sk
->sk_rcvbuf
>> 1))) {
4902 tp
->ucopy
.wakeup
= 1;
4903 sk
->sk_data_ready(sk
, 0);
4905 } else if (chunk
> 0) {
4906 tp
->ucopy
.wakeup
= 1;
4907 sk
->sk_data_ready(sk
, 0);
4910 return copied_early
;
4912 #endif /* CONFIG_NET_DMA */
4914 /* Does PAWS and seqno based validation of an incoming segment, flags will
4915 * play significant role here.
4917 static bool tcp_validate_incoming(struct sock
*sk
, struct sk_buff
*skb
,
4918 const struct tcphdr
*th
, int syn_inerr
)
4920 struct tcp_sock
*tp
= tcp_sk(sk
);
4922 /* RFC1323: H1. Apply PAWS check first. */
4923 if (tcp_fast_parse_options(skb
, th
, tp
) && tp
->rx_opt
.saw_tstamp
&&
4924 tcp_paws_discard(sk
, skb
)) {
4926 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSESTABREJECTED
);
4927 tcp_send_dupack(sk
, skb
);
4930 /* Reset is accepted even if it did not pass PAWS. */
4933 /* Step 1: check sequence number */
4934 if (!tcp_sequence(tp
, TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(skb
)->end_seq
)) {
4935 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4936 * (RST) segments are validated by checking their SEQ-fields."
4937 * And page 69: "If an incoming segment is not acceptable,
4938 * an acknowledgment should be sent in reply (unless the RST
4939 * bit is set, if so drop the segment and return)".
4944 tcp_send_dupack(sk
, skb
);
4949 /* Step 2: check RST bit */
4952 * If sequence number exactly matches RCV.NXT, then
4953 * RESET the connection
4955 * Send a challenge ACK
4957 if (TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
)
4960 tcp_send_challenge_ack(sk
);
4964 /* step 3: check security and precedence [ignored] */
4966 /* step 4: Check for a SYN
4967 * RFC 5691 4.2 : Send a challenge ack
4972 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
4973 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNCHALLENGE
);
4974 tcp_send_challenge_ack(sk
);
4986 * TCP receive function for the ESTABLISHED state.
4988 * It is split into a fast path and a slow path. The fast path is
4990 * - A zero window was announced from us - zero window probing
4991 * is only handled properly in the slow path.
4992 * - Out of order segments arrived.
4993 * - Urgent data is expected.
4994 * - There is no buffer space left
4995 * - Unexpected TCP flags/window values/header lengths are received
4996 * (detected by checking the TCP header against pred_flags)
4997 * - Data is sent in both directions. Fast path only supports pure senders
4998 * or pure receivers (this means either the sequence number or the ack
4999 * value must stay constant)
5000 * - Unexpected TCP option.
5002 * When these conditions are not satisfied it drops into a standard
5003 * receive procedure patterned after RFC793 to handle all cases.
5004 * The first three cases are guaranteed by proper pred_flags setting,
5005 * the rest is checked inline. Fast processing is turned on in
5006 * tcp_data_queue when everything is OK.
5008 int tcp_rcv_established(struct sock
*sk
, struct sk_buff
*skb
,
5009 const struct tcphdr
*th
, unsigned int len
)
5011 struct tcp_sock
*tp
= tcp_sk(sk
);
5013 if (unlikely(sk
->sk_rx_dst
== NULL
))
5014 inet_csk(sk
)->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5016 * Header prediction.
5017 * The code loosely follows the one in the famous
5018 * "30 instruction TCP receive" Van Jacobson mail.
5020 * Van's trick is to deposit buffers into socket queue
5021 * on a device interrupt, to call tcp_recv function
5022 * on the receive process context and checksum and copy
5023 * the buffer to user space. smart...
5025 * Our current scheme is not silly either but we take the
5026 * extra cost of the net_bh soft interrupt processing...
5027 * We do checksum and copy also but from device to kernel.
5030 tp
->rx_opt
.saw_tstamp
= 0;
5032 /* pred_flags is 0xS?10 << 16 + snd_wnd
5033 * if header_prediction is to be made
5034 * 'S' will always be tp->tcp_header_len >> 2
5035 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5036 * turn it off (when there are holes in the receive
5037 * space for instance)
5038 * PSH flag is ignored.
5041 if ((tcp_flag_word(th
) & TCP_HP_BITS
) == tp
->pred_flags
&&
5042 TCP_SKB_CB(skb
)->seq
== tp
->rcv_nxt
&&
5043 !after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
)) {
5044 int tcp_header_len
= tp
->tcp_header_len
;
5046 /* Timestamp header prediction: tcp_header_len
5047 * is automatically equal to th->doff*4 due to pred_flags
5051 /* Check timestamp */
5052 if (tcp_header_len
== sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) {
5053 /* No? Slow path! */
5054 if (!tcp_parse_aligned_timestamp(tp
, th
))
5057 /* If PAWS failed, check it more carefully in slow path */
5058 if ((s32
)(tp
->rx_opt
.rcv_tsval
- tp
->rx_opt
.ts_recent
) < 0)
5061 /* DO NOT update ts_recent here, if checksum fails
5062 * and timestamp was corrupted part, it will result
5063 * in a hung connection since we will drop all
5064 * future packets due to the PAWS test.
5068 if (len
<= tcp_header_len
) {
5069 /* Bulk data transfer: sender */
5070 if (len
== tcp_header_len
) {
5071 /* Predicted packet is in window by definition.
5072 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5073 * Hence, check seq<=rcv_wup reduces to:
5075 if (tcp_header_len
==
5076 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5077 tp
->rcv_nxt
== tp
->rcv_wup
)
5078 tcp_store_ts_recent(tp
);
5080 /* We know that such packets are checksummed
5083 tcp_ack(sk
, skb
, 0);
5085 tcp_data_snd_check(sk
);
5087 } else { /* Header too small */
5088 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5093 int copied_early
= 0;
5094 bool fragstolen
= false;
5096 if (tp
->copied_seq
== tp
->rcv_nxt
&&
5097 len
- tcp_header_len
<= tp
->ucopy
.len
) {
5098 #ifdef CONFIG_NET_DMA
5099 if (tp
->ucopy
.task
== current
&&
5100 sock_owned_by_user(sk
) &&
5101 tcp_dma_try_early_copy(sk
, skb
, tcp_header_len
)) {
5106 if (tp
->ucopy
.task
== current
&&
5107 sock_owned_by_user(sk
) && !copied_early
) {
5108 __set_current_state(TASK_RUNNING
);
5110 if (!tcp_copy_to_iovec(sk
, skb
, tcp_header_len
))
5114 /* Predicted packet is in window by definition.
5115 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5116 * Hence, check seq<=rcv_wup reduces to:
5118 if (tcp_header_len
==
5119 (sizeof(struct tcphdr
) +
5120 TCPOLEN_TSTAMP_ALIGNED
) &&
5121 tp
->rcv_nxt
== tp
->rcv_wup
)
5122 tcp_store_ts_recent(tp
);
5124 tcp_rcv_rtt_measure_ts(sk
, skb
);
5126 __skb_pull(skb
, tcp_header_len
);
5127 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->end_seq
;
5128 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITSTOUSER
);
5131 tcp_cleanup_rbuf(sk
, skb
->len
);
5134 if (tcp_checksum_complete_user(sk
, skb
))
5137 if ((int)skb
->truesize
> sk
->sk_forward_alloc
)
5140 /* Predicted packet is in window by definition.
5141 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5142 * Hence, check seq<=rcv_wup reduces to:
5144 if (tcp_header_len
==
5145 (sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
) &&
5146 tp
->rcv_nxt
== tp
->rcv_wup
)
5147 tcp_store_ts_recent(tp
);
5149 tcp_rcv_rtt_measure_ts(sk
, skb
);
5151 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPHPHITS
);
5153 /* Bulk data transfer: receiver */
5154 eaten
= tcp_queue_rcv(sk
, skb
, tcp_header_len
,
5158 tcp_event_data_recv(sk
, skb
);
5160 if (TCP_SKB_CB(skb
)->ack_seq
!= tp
->snd_una
) {
5161 /* Well, only one small jumplet in fast path... */
5162 tcp_ack(sk
, skb
, FLAG_DATA
);
5163 tcp_data_snd_check(sk
);
5164 if (!inet_csk_ack_scheduled(sk
))
5168 if (!copied_early
|| tp
->rcv_nxt
!= tp
->rcv_wup
)
5169 __tcp_ack_snd_check(sk
, 0);
5171 #ifdef CONFIG_NET_DMA
5173 __skb_queue_tail(&sk
->sk_async_wait_queue
, skb
);
5177 kfree_skb_partial(skb
, fragstolen
);
5178 sk
->sk_data_ready(sk
, 0);
5184 if (len
< (th
->doff
<< 2) || tcp_checksum_complete_user(sk
, skb
))
5187 if (!th
->ack
&& !th
->rst
)
5191 * Standard slow path.
5194 if (!tcp_validate_incoming(sk
, skb
, th
, 1))
5198 if (tcp_ack(sk
, skb
, FLAG_SLOWPATH
| FLAG_UPDATE_TS_RECENT
) < 0)
5201 tcp_rcv_rtt_measure_ts(sk
, skb
);
5203 /* Process urgent data. */
5204 tcp_urg(sk
, skb
, th
);
5206 /* step 7: process the segment text */
5207 tcp_data_queue(sk
, skb
);
5209 tcp_data_snd_check(sk
);
5210 tcp_ack_snd_check(sk
);
5214 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_CSUMERRORS
);
5215 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_INERRS
);
5221 EXPORT_SYMBOL(tcp_rcv_established
);
5223 void tcp_finish_connect(struct sock
*sk
, struct sk_buff
*skb
)
5225 struct tcp_sock
*tp
= tcp_sk(sk
);
5226 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5228 tcp_set_state(sk
, TCP_ESTABLISHED
);
5231 icsk
->icsk_af_ops
->sk_rx_dst_set(sk
, skb
);
5232 security_inet_conn_established(sk
, skb
);
5235 /* Make sure socket is routed, for correct metrics. */
5236 icsk
->icsk_af_ops
->rebuild_header(sk
);
5238 tcp_init_metrics(sk
);
5240 tcp_init_congestion_control(sk
);
5242 /* Prevent spurious tcp_cwnd_restart() on first data
5245 tp
->lsndtime
= tcp_time_stamp
;
5247 tcp_init_buffer_space(sk
);
5249 if (sock_flag(sk
, SOCK_KEEPOPEN
))
5250 inet_csk_reset_keepalive_timer(sk
, keepalive_time_when(tp
));
5252 if (!tp
->rx_opt
.snd_wscale
)
5253 __tcp_fast_path_on(tp
, tp
->snd_wnd
);
5257 if (!sock_flag(sk
, SOCK_DEAD
)) {
5258 sk
->sk_state_change(sk
);
5259 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5263 static bool tcp_rcv_fastopen_synack(struct sock
*sk
, struct sk_buff
*synack
,
5264 struct tcp_fastopen_cookie
*cookie
)
5266 struct tcp_sock
*tp
= tcp_sk(sk
);
5267 struct sk_buff
*data
= tp
->syn_data
? tcp_write_queue_head(sk
) : NULL
;
5268 u16 mss
= tp
->rx_opt
.mss_clamp
;
5271 if (mss
== tp
->rx_opt
.user_mss
) {
5272 struct tcp_options_received opt
;
5274 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5275 tcp_clear_options(&opt
);
5276 opt
.user_mss
= opt
.mss_clamp
= 0;
5277 tcp_parse_options(synack
, &opt
, 0, NULL
);
5278 mss
= opt
.mss_clamp
;
5281 if (!tp
->syn_fastopen
) /* Ignore an unsolicited cookie */
5284 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5285 * the remote receives only the retransmitted (regular) SYNs: either
5286 * the original SYN-data or the corresponding SYN-ACK is lost.
5288 syn_drop
= (cookie
->len
<= 0 && data
&& tp
->total_retrans
);
5290 tcp_fastopen_cache_set(sk
, mss
, cookie
, syn_drop
);
5292 if (data
) { /* Retransmit unacked data in SYN */
5293 tcp_for_write_queue_from(data
, sk
) {
5294 if (data
== tcp_send_head(sk
) ||
5295 __tcp_retransmit_skb(sk
, data
))
5301 tp
->syn_data_acked
= tp
->syn_data
;
5305 static int tcp_rcv_synsent_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5306 const struct tcphdr
*th
, unsigned int len
)
5308 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5309 struct tcp_sock
*tp
= tcp_sk(sk
);
5310 struct tcp_fastopen_cookie foc
= { .len
= -1 };
5311 int saved_clamp
= tp
->rx_opt
.mss_clamp
;
5313 tcp_parse_options(skb
, &tp
->rx_opt
, 0, &foc
);
5314 if (tp
->rx_opt
.saw_tstamp
)
5315 tp
->rx_opt
.rcv_tsecr
-= tp
->tsoffset
;
5319 * "If the state is SYN-SENT then
5320 * first check the ACK bit
5321 * If the ACK bit is set
5322 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5323 * a reset (unless the RST bit is set, if so drop
5324 * the segment and return)"
5326 if (!after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_una
) ||
5327 after(TCP_SKB_CB(skb
)->ack_seq
, tp
->snd_nxt
))
5328 goto reset_and_undo
;
5330 if (tp
->rx_opt
.saw_tstamp
&& tp
->rx_opt
.rcv_tsecr
&&
5331 !between(tp
->rx_opt
.rcv_tsecr
, tp
->retrans_stamp
,
5333 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_PAWSACTIVEREJECTED
);
5334 goto reset_and_undo
;
5337 /* Now ACK is acceptable.
5339 * "If the RST bit is set
5340 * If the ACK was acceptable then signal the user "error:
5341 * connection reset", drop the segment, enter CLOSED state,
5342 * delete TCB, and return."
5351 * "fifth, if neither of the SYN or RST bits is set then
5352 * drop the segment and return."
5358 goto discard_and_undo
;
5361 * "If the SYN bit is on ...
5362 * are acceptable then ...
5363 * (our SYN has been ACKed), change the connection
5364 * state to ESTABLISHED..."
5367 TCP_ECN_rcv_synack(tp
, th
);
5369 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5370 tcp_ack(sk
, skb
, FLAG_SLOWPATH
);
5372 /* Ok.. it's good. Set up sequence numbers and
5373 * move to established.
5375 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5376 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5378 /* RFC1323: The window in SYN & SYN/ACK segments is
5381 tp
->snd_wnd
= ntohs(th
->window
);
5383 if (!tp
->rx_opt
.wscale_ok
) {
5384 tp
->rx_opt
.snd_wscale
= tp
->rx_opt
.rcv_wscale
= 0;
5385 tp
->window_clamp
= min(tp
->window_clamp
, 65535U);
5388 if (tp
->rx_opt
.saw_tstamp
) {
5389 tp
->rx_opt
.tstamp_ok
= 1;
5390 tp
->tcp_header_len
=
5391 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5392 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5393 tcp_store_ts_recent(tp
);
5395 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5398 if (tcp_is_sack(tp
) && sysctl_tcp_fack
)
5399 tcp_enable_fack(tp
);
5402 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5403 tcp_initialize_rcv_mss(sk
);
5405 /* Remember, tcp_poll() does not lock socket!
5406 * Change state from SYN-SENT only after copied_seq
5407 * is initialized. */
5408 tp
->copied_seq
= tp
->rcv_nxt
;
5412 tcp_finish_connect(sk
, skb
);
5414 if ((tp
->syn_fastopen
|| tp
->syn_data
) &&
5415 tcp_rcv_fastopen_synack(sk
, skb
, &foc
))
5418 if (sk
->sk_write_pending
||
5419 icsk
->icsk_accept_queue
.rskq_defer_accept
||
5420 icsk
->icsk_ack
.pingpong
) {
5421 /* Save one ACK. Data will be ready after
5422 * several ticks, if write_pending is set.
5424 * It may be deleted, but with this feature tcpdumps
5425 * look so _wonderfully_ clever, that I was not able
5426 * to stand against the temptation 8) --ANK
5428 inet_csk_schedule_ack(sk
);
5429 icsk
->icsk_ack
.lrcvtime
= tcp_time_stamp
;
5430 tcp_enter_quickack_mode(sk
);
5431 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
5432 TCP_DELACK_MAX
, TCP_RTO_MAX
);
5443 /* No ACK in the segment */
5447 * "If the RST bit is set
5449 * Otherwise (no ACK) drop the segment and return."
5452 goto discard_and_undo
;
5456 if (tp
->rx_opt
.ts_recent_stamp
&& tp
->rx_opt
.saw_tstamp
&&
5457 tcp_paws_reject(&tp
->rx_opt
, 0))
5458 goto discard_and_undo
;
5461 /* We see SYN without ACK. It is attempt of
5462 * simultaneous connect with crossed SYNs.
5463 * Particularly, it can be connect to self.
5465 tcp_set_state(sk
, TCP_SYN_RECV
);
5467 if (tp
->rx_opt
.saw_tstamp
) {
5468 tp
->rx_opt
.tstamp_ok
= 1;
5469 tcp_store_ts_recent(tp
);
5470 tp
->tcp_header_len
=
5471 sizeof(struct tcphdr
) + TCPOLEN_TSTAMP_ALIGNED
;
5473 tp
->tcp_header_len
= sizeof(struct tcphdr
);
5476 tp
->rcv_nxt
= TCP_SKB_CB(skb
)->seq
+ 1;
5477 tp
->rcv_wup
= TCP_SKB_CB(skb
)->seq
+ 1;
5479 /* RFC1323: The window in SYN & SYN/ACK segments is
5482 tp
->snd_wnd
= ntohs(th
->window
);
5483 tp
->snd_wl1
= TCP_SKB_CB(skb
)->seq
;
5484 tp
->max_window
= tp
->snd_wnd
;
5486 TCP_ECN_rcv_syn(tp
, th
);
5489 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
5490 tcp_initialize_rcv_mss(sk
);
5492 tcp_send_synack(sk
);
5494 /* Note, we could accept data and URG from this segment.
5495 * There are no obstacles to make this (except that we must
5496 * either change tcp_recvmsg() to prevent it from returning data
5497 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5499 * However, if we ignore data in ACKless segments sometimes,
5500 * we have no reasons to accept it sometimes.
5501 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5502 * is not flawless. So, discard packet for sanity.
5503 * Uncomment this return to process the data.
5510 /* "fifth, if neither of the SYN or RST bits is set then
5511 * drop the segment and return."
5515 tcp_clear_options(&tp
->rx_opt
);
5516 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5520 tcp_clear_options(&tp
->rx_opt
);
5521 tp
->rx_opt
.mss_clamp
= saved_clamp
;
5526 * This function implements the receiving procedure of RFC 793 for
5527 * all states except ESTABLISHED and TIME_WAIT.
5528 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5529 * address independent.
5532 int tcp_rcv_state_process(struct sock
*sk
, struct sk_buff
*skb
,
5533 const struct tcphdr
*th
, unsigned int len
)
5535 struct tcp_sock
*tp
= tcp_sk(sk
);
5536 struct inet_connection_sock
*icsk
= inet_csk(sk
);
5537 struct request_sock
*req
;
5541 tp
->rx_opt
.saw_tstamp
= 0;
5543 switch (sk
->sk_state
) {
5557 if (icsk
->icsk_af_ops
->conn_request(sk
, skb
) < 0)
5560 /* Now we have several options: In theory there is
5561 * nothing else in the frame. KA9Q has an option to
5562 * send data with the syn, BSD accepts data with the
5563 * syn up to the [to be] advertised window and
5564 * Solaris 2.1 gives you a protocol error. For now
5565 * we just ignore it, that fits the spec precisely
5566 * and avoids incompatibilities. It would be nice in
5567 * future to drop through and process the data.
5569 * Now that TTCP is starting to be used we ought to
5571 * But, this leaves one open to an easy denial of
5572 * service attack, and SYN cookies can't defend
5573 * against this problem. So, we drop the data
5574 * in the interest of security over speed unless
5575 * it's still in use.
5583 queued
= tcp_rcv_synsent_state_process(sk
, skb
, th
, len
);
5587 /* Do step6 onward by hand. */
5588 tcp_urg(sk
, skb
, th
);
5590 tcp_data_snd_check(sk
);
5594 req
= tp
->fastopen_rsk
;
5596 WARN_ON_ONCE(sk
->sk_state
!= TCP_SYN_RECV
&&
5597 sk
->sk_state
!= TCP_FIN_WAIT1
);
5599 if (tcp_check_req(sk
, skb
, req
, NULL
, true) == NULL
)
5603 if (!th
->ack
&& !th
->rst
)
5606 if (!tcp_validate_incoming(sk
, skb
, th
, 0))
5609 /* step 5: check the ACK field */
5610 acceptable
= tcp_ack(sk
, skb
, FLAG_SLOWPATH
|
5611 FLAG_UPDATE_TS_RECENT
) > 0;
5613 switch (sk
->sk_state
) {
5618 /* Once we leave TCP_SYN_RECV, we no longer need req
5622 tcp_synack_rtt_meas(sk
, req
);
5623 tp
->total_retrans
= req
->num_retrans
;
5625 reqsk_fastopen_remove(sk
, req
, false);
5627 /* Make sure socket is routed, for correct metrics. */
5628 icsk
->icsk_af_ops
->rebuild_header(sk
);
5629 tcp_init_congestion_control(sk
);
5632 tcp_init_buffer_space(sk
);
5633 tp
->copied_seq
= tp
->rcv_nxt
;
5636 tcp_set_state(sk
, TCP_ESTABLISHED
);
5637 sk
->sk_state_change(sk
);
5639 /* Note, that this wakeup is only for marginal crossed SYN case.
5640 * Passively open sockets are not waked up, because
5641 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5644 sk_wake_async(sk
, SOCK_WAKE_IO
, POLL_OUT
);
5646 tp
->snd_una
= TCP_SKB_CB(skb
)->ack_seq
;
5647 tp
->snd_wnd
= ntohs(th
->window
) << tp
->rx_opt
.snd_wscale
;
5648 tcp_init_wl(tp
, TCP_SKB_CB(skb
)->seq
);
5650 if (tp
->rx_opt
.tstamp_ok
)
5651 tp
->advmss
-= TCPOLEN_TSTAMP_ALIGNED
;
5654 /* Re-arm the timer because data may have been sent out.
5655 * This is similar to the regular data transmission case
5656 * when new data has just been ack'ed.
5658 * (TFO) - we could try to be more aggressive and
5659 * retransmitting any data sooner based on when they
5664 tcp_init_metrics(sk
);
5666 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5667 tp
->lsndtime
= tcp_time_stamp
;
5669 tcp_initialize_rcv_mss(sk
);
5670 tcp_fast_path_on(tp
);
5674 /* If we enter the TCP_FIN_WAIT1 state and we are a
5675 * Fast Open socket and this is the first acceptable
5676 * ACK we have received, this would have acknowledged
5677 * our SYNACK so stop the SYNACK timer.
5680 /* Return RST if ack_seq is invalid.
5681 * Note that RFC793 only says to generate a
5682 * DUPACK for it but for TCP Fast Open it seems
5683 * better to treat this case like TCP_SYN_RECV
5688 /* We no longer need the request sock. */
5689 reqsk_fastopen_remove(sk
, req
, false);
5692 if (tp
->snd_una
== tp
->write_seq
) {
5693 struct dst_entry
*dst
;
5695 tcp_set_state(sk
, TCP_FIN_WAIT2
);
5696 sk
->sk_shutdown
|= SEND_SHUTDOWN
;
5698 dst
= __sk_dst_get(sk
);
5702 if (!sock_flag(sk
, SOCK_DEAD
)) {
5703 /* Wake up lingering close() */
5704 sk
->sk_state_change(sk
);
5708 if (tp
->linger2
< 0 ||
5709 (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5710 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
))) {
5712 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5716 tmo
= tcp_fin_time(sk
);
5717 if (tmo
> TCP_TIMEWAIT_LEN
) {
5718 inet_csk_reset_keepalive_timer(sk
, tmo
- TCP_TIMEWAIT_LEN
);
5719 } else if (th
->fin
|| sock_owned_by_user(sk
)) {
5720 /* Bad case. We could lose such FIN otherwise.
5721 * It is not a big problem, but it looks confusing
5722 * and not so rare event. We still can lose it now,
5723 * if it spins in bh_lock_sock(), but it is really
5726 inet_csk_reset_keepalive_timer(sk
, tmo
);
5728 tcp_time_wait(sk
, TCP_FIN_WAIT2
, tmo
);
5736 if (tp
->snd_una
== tp
->write_seq
) {
5737 tcp_time_wait(sk
, TCP_TIME_WAIT
, 0);
5743 if (tp
->snd_una
== tp
->write_seq
) {
5744 tcp_update_metrics(sk
);
5751 /* step 6: check the URG bit */
5752 tcp_urg(sk
, skb
, th
);
5754 /* step 7: process the segment text */
5755 switch (sk
->sk_state
) {
5756 case TCP_CLOSE_WAIT
:
5759 if (!before(TCP_SKB_CB(skb
)->seq
, tp
->rcv_nxt
))
5763 /* RFC 793 says to queue data in these states,
5764 * RFC 1122 says we MUST send a reset.
5765 * BSD 4.4 also does reset.
5767 if (sk
->sk_shutdown
& RCV_SHUTDOWN
) {
5768 if (TCP_SKB_CB(skb
)->end_seq
!= TCP_SKB_CB(skb
)->seq
&&
5769 after(TCP_SKB_CB(skb
)->end_seq
- th
->fin
, tp
->rcv_nxt
)) {
5770 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPABORTONDATA
);
5776 case TCP_ESTABLISHED
:
5777 tcp_data_queue(sk
, skb
);
5782 /* tcp_data could move socket to TIME-WAIT */
5783 if (sk
->sk_state
!= TCP_CLOSE
) {
5784 tcp_data_snd_check(sk
);
5785 tcp_ack_snd_check(sk
);
5794 EXPORT_SYMBOL(tcp_rcv_state_process
);